Content uploaded by Ravilla Dilli
Author content
All content in this area was uploaded by Ravilla Dilli on Sep 04, 2021
Content may be subject to copyright.
Communications
in Computer and Information Science 147
Vinu V Das Gylson Thomas
Ford Lumban Gaol (Eds.)
Information Technology
and Mobile Communication
International Conference, AIM 2011
Nagpur, Maharashtra, India, April 21-22, 2011
Proceedings
13
Volume Editors
Vinu V Das
ACEEE, Trivandrum, Kerala, India
E-mail: vinuvdas@theaceee.org
Gylson Thomas
MES College of Engineering, Kuttippuram, Kerala, India
E-mail: gylson_thomas@yahoo.com
Ford Lumban Gaol
Binus University, Jakarta, Indonesia
E-mail: fordlg@gmail.com
ISSN 1865-0929 e-ISSN 1865-0937
ISBN 978-3-642-20572-9 e-ISBN 978-3-642-20573-6
DOI 10.1007/978-3-642-20573-6
Springer Heidelberg Dordrecht London NewYork
Library of Congress Control Number: 2011925374
CR Subject Classification (1998): C.2, D.2, H.4, H.3, I.2.11, K.4.4
© Springer-Verlag Berlin Heidelberg 2011
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in its current version, and permission for use must always be obtained from Springer.Violations are liable
to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general use.
Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
Preface
The International Conference on Advances in Information Technology and Mo-
bile Communication (AIM 2011) was sponsored and organized by The Associ-
ation of Computer Electronics and Electrical Engineers (ACEEE) and held at
Nagpur, Maharashtra, India during April 21-22, 2011.
The mission of the AIM International Conference is to bring together innova-
tive academics and industrial experts in the field of computer science, informa-
tion technology, computational engineering, mobile communication and security
to a common forum, where a constructive dialog on theoretical concepts, prac-
tical ideas and results of the state of the art can be developed. In addition, the
participants of the symposium have a chance to hear from renowned keynote
speakers. We would like to thank the Program Chairs, organization staff, and
the members of the ProgramCommittees for their hard work this year. We would
like to thank all our colleagues who served on different committees and acted as
reviewers to identify a set of high-quality research papers for AIM 2011.
The conference received 313 submissions overall. Only 92 papers were
accepted and registered for the AIM 2011 proceedings. We also thank Alfred
Hofmann, Janahanlal Stephen, and Gylson Thomas for the constant support
and guidance. We would like to express our gratitude to the Springer LNCS-
CCIS editorial team, especially Ms. Leonie Kunz, for producing such a wonderful
quality proceedings book.
February 2011 Vinu V. Das
AIM 2011 – Organization
Honorary Chairs
Shuvra Das University of Detroit Mercy, USA
Jiguo Yu Qufu Normal University, China
Technical Chairs
Sumeet Dua Louisiana Tech University, USA
Vijayakumar MG University, India
Amit Banerjee The Pennsylvania State University, USA
Technical Co-chairs
Natarajan Meghanathan Jackson State University, USA
Gylson Thomas MES College of Engineering, India
Hicham Elzabadani American University in Dubai
Shahrokh Valaee University of Toronto, Canada
General Chair
Janahanlal Stephen ILAHIA College of Engineering, India
Organizing Chair
Vinu V. Das The IDES
Organizing Co-chairs
T.S.B. Sudarshan BITS Pilani, India
Ford Lumban Gaol University of Indonesia
Publicity Chairs
Amlan Chakrabarti University of Calcutta, India
Prafulla Kumar Behera, Utkal University, India
VIII AIM 2011 – Organization
Publication Chairs
Vijayakumar NSS Engineering College, India
T.S.B. Sudarshan BITS Pilani, India
K.P. Soman Amritha University, India
N. Jaisankar VIT University, India
Program Committee Chairs
Harry E. Ruda University of Toronto, Canada
Deepak Laxmi Narasimha University of Malaya, Malaysia
N. Nagarajan Anna University, Coimbatore, India
Akash Rajak Krishna Institute of Engineering and
Technology, UP, India
M. Ayoub Khan CDAC, NOIDA, India
Table of Contents
Full Paper
Efficient Object Motion Prediction Using Adaptive Fuzzy Navigational
Environment .................................................... 1
Vijay S. Rajpurohit and M.M. Manohara Pai
An Efficient Protocol Using Smart Interval for Coordinated
Checkpointing ................................................... 6
Jagdish Makhijani, Manoj Kumar Niranjan, Mahesh Motwani,
A.K. Sachan, and Anil Rajput
Face Recognition System Using Discrete Wavelet Transform and Fast
PCA ........................................................... 13
K. Ramesha and K.B. Raja
Mining Indirect Association between Itemsets ....................... 19
B. Ramasubbareddy, A. Govardhan, and A. Ramamohanreddy
Reaction Attacks in the Matrix Scheme of NTRU Cryptosystem ....... 27
Rakesh Nayak, Jayaram Pradhan, and C.V. Sastry
Process Corner Analysis for Folding and Interpolating ADC ........... 33
Shruti Oza and N.M. Devashrayee
Fast Near-Lossless Image Compression with Tree Coding Having
Predictable Output Compression Size............................... 39
Soumik Banerjee and Debashish Chakroborty
Over Load Detection and Admission Control Policy in DRTDBS ....... 45
Nuparam and Udai Shanker
Wavelet Transform Based Image Registration and Image Fusion ....... 55
Manjusha Deshmukh and Sonal Gahankari
60 GHz Radio Channel Characteristics in an Indoor Environment for
Home Entertainment Networks .................................... 61
T. Rama Rao, S. Ramesh, and D. Murugesan
Improved Back Propagation Algorithm to Avoid Local Minima in
Multiplicative Neuron Model ...................................... 67
Kavita Burse, Manish Manoria, and Vishnu Pratap Singh Kirar
Technical White Paper on “Time and Frequency Synchronization in
OFDM” ........................................................ 74
Anagha Rathkanthiwar and Mridula Korde
X Table of Contents
Cell-ID Based Vehicle Locator and Real-Time Deactivator Using GSM
Network ........................................................ 82
Nilesh Dubey, Vandana Dubey, and Shivangi Bande
A Novel Design of Reconfigurable Architecture for Multistandard
Communication System ........................................... 87
T. Suresh and K.L. Shunmuganathan
Two Novel Long-Tail Pair Based Second Generation Current Conveyors
(CCII) .......................................................... 95
Amisha Naik and N.M. Devashrayee
Generating Testcases for Concurrent Systems Using UML State Chart
Diagram ........................................................ 100
Debashree Patnaik, Arup Abhinna Acharya, and Durga P. Mohapatra
Intelligent Agent Based Resource Sharing in Grid Computing .......... 106
V.V. Srinivas and V.V. Varadhan
Wideband Miniaturized Patch Antenna Design and Comparative
Analysis ........................................................ 111
Sanket Patel, Yogeshwar Kosta, Himanshu Soni, and Shobhit Patel
Predicting Number of Zombies in a DDoS Attack Using ANN Based
Scheme ......................................................... 117
B.B. Gupta, R.C. Joshi, M. Misra, A. Jain, S. Juyal,
R. Prabhakar, and A.K. Singh
A Novel Biometric Watermaking Approach Using LWT- SVD ......... 123
Meenakshi Arya and Rajesh Siddavatam
Detection and Prevention of Phishing Attack Using Dynamic
Wate rmarking ................................................... 132
Akhilendra Pratap Singh, Vimal Kumar, Sandeep Singh Sengar, and
Manoj Wairiya
A Search Tool Using Genetic Algorithm ............................ 138
M.K. Thanuja and C. Mala
Heterogeneous Data Mining Environment Based on DAM for Mobile
Computing Environments ......................................... 144
Ashutosh K. Dubey, Ganesh Raj Kushwaha, and Nishant Shrivastava
Selection of Views for Materialization Using Size and Query
Frequency ....................................................... 150
T.V. Vijay Kumar and Mohammad Haider
Key Validation Using Weighted-Edge Web of Trust Model ............ 156
Sumit Kumar, Nahar Singh, and Ashok Singh Sairam
Tab le of Cont ents XI
A Novel Reconfigurable Architecture for Enhancing Color Image Based
on Adaptive Saturation Feedback .................................. 162
M.C. Hanumantharaju, M. Ravishankar, D.R. Rameshbabu, and
S. Ramachandran
Signal Processing Approach for Prediction of Kink in Transmembrane
α-Helices ....................................................... 170
Jayakishan K. Meher, Nibedita Mishra, Pranab Kishor Mohapatra,
Mukesh Kumar Raval, Pramod Kumar Meher, and Gananath Dash
Cascaded H-Bridge Multilevel Boost Inverter without Inductors for
Electric/Hybrid Electric Vehicle Applications ........................ 178
S. Dhayanandh, A.P. Ramya Sri, S. Rajkumar, and N. Lavanya
Design of Microstrip Meandered Patch Antenna for Mobile
Communication .................................................. 184
Shobhit Patel, Jaymin Bhalani, Yogesh Kosta, and Sanket Patel
Building Gaussian Mixture Shadow Model for Removing Shadows in
Surveillance Videos............................................... 190
Archana Chougule and Pratap Halkarnikar
FAutoREDWithRED: To Increase the Overall Performance of Internet
Routers ......................................................... 196
K. Chitra and G. Padmavathi
Short Paper
Scan Based Sequential Circuit Testing Using DFT Advisor ............ 203
P. Reshma
Rate Adaptive Distributed Source-Channel Coding Using IRA Codes
for Wireless Sensor Networks ...................................... 207
Saikat Majumder and Shrish Verma
Web Cam Motion Detection Surveillance System Using Temporal
Difference and Optical Flow Detection with Multi Alerts .............. 214
V.D. Ambeth Kumar and M. Ramakrishan
Rts-Mirror: Real Time Synchronized Automated Rear Vision Mirror
System ......................................................... 222
KuldeepVerma,AnkitaAgarkar,andApoorvJoshi
Fuzzy Based PSO for Software Effort Estimation..................... 227
P.V.G.D. Prasad Reddy and CH.V.M.K. Hari
SLV: Sweep Line Voronoi Ad Hoc Routing Algorithm ................. 233
E. Rama Krishna, A. Venkat Reddy, N. Rambabu, and
G. Rajesh Kumar
XII Table of Contents
Hybrid Routing for Ad Hoc Wireless Networks ...................... 240
Ravilla Dilli, R.S. Murali Nath, and P. Chandra Shekar Reddy
Implementation of ARINC 429 16 Channel Transmitter Controller on
FPGA .......................................................... 245
Debasis Mukherjee, Niti Kumar, Kalyan Singh,
Hemanta Mondal, and B.V.R. Reddy
Segmentation of Image Using Watershed and Fast Level Set Methods ... 248
Minal M. Puranik and Shobha Krishnan
Tree Structured, Multi-hop Time Synchronization Approach in Wireless
Sensor Networks ................................................. 255
Surendra Rahamatkar, Ajay Agarwal, Praveen Sen, and Arun Yadav
A New Markov Chain Based Cost Evaluation Metric for Routing in
MANETs ....................................................... 259
Abhinav Tiwari, Nisha Wadhawan, and Neeraj Kumar
Texture Image Classification Using Gray Level Weight
Matrix (GLWM) ................................................. 263
R.S. Sabeenian and P.M. Dinesh
Formal Verification of IEEE802.11i WPA-GPG Authentication
Protocol ........................................................ 267
K.V. Krishnam Raju and V. Valli Kumari
Adaptive Steganography Based on Covariance and Dct ............... 273
N. Sathisha, Swetha Sreedharan, R. Ujwal, Kiran D’sa,
Aneeshwar R. Danda, K. Suresh Babu, K.B. Raja,
K.R. Venugopal, and L.M. Patnaik
Image Segmentation Using Grey Scale Weighted Average Method and
Type-2 Fuzzy Logic Systems ...................................... 277
Saikat Maity and Jaya Sil
Cluster Analysis and Pso for Software Cost Estimation ............... 281
Tegjyot Singh Sethi, CH.V.M.K. Hari, B.S.S. Kaushal, and
Abhishek Sharma
Controlling Crossover Probability in Case of a Genetic Algorithm ...... 287
Parama Bagchi and Shantanu Pal
A Qualitative Survey on Unicast Routing Algorithms in Delay Tolerant
Networks ....................................................... 291
Sushovan Patra, Anerudh Balaji, Sujoy Saha,
Amartya Mukherjee, and Subrata Nandi
Table of Contents XIII
Designing and Modeling of CMOS Low Noise Amplifier Using a
Composite MOSFET Model Working at Millimeter-Wave Band ........ 297
Adhira Raj, Karthigha Balamurugan, and M. Jayakumar
Designing Dependable Business Intelligence Solutions Using Agile Web
Services Mining Architectures ..................................... 301
A.V. Krishna Prasad, S. Ramakrishna, B. Padmaja Rani,
M. Upendra Kumar, and D. Shravani
A Modified Continuous Particle Swarm Optimization Algorithm for
Uncapacitated Facility Location Problem ........................... 305
Sujay Saha, Arnab Kole, and Kashinath Dey
Design of Hybrid Genetic Algorithm with Preferential Local Search for
Multiobjective Optimization Problems .............................. 312
J. Bhuvana and C. Aravindan
Synergy of Multi-agent Coordination Technique and Optimization
Techniques for Patient Scheduli .................................... 317
E. Grace Mary Kanaga and M.L. Valarmathi
Naive Bayes Approach for Website Classification ..................... 323
R. Rajalakshmi and C. Aravindan
Method to Improve the Efficiency of the Software by the Effective
Selection of the Test Cases from Test Suite Using Data Mining
Techniques ...................................................... 327
Lilly Raamesh and G.V. Uma
A Hybrid Intelligent Path Planning Approach to Cooperative Robots ... 332
K. Prasadh, Vinodh P. Vijayan, and Biju Paul
Towards Evaluating Resilience of SIP Server under Low Rate DoS
Attack .......................................................... 336
Abhishek Kumar, P. Shanthi Thilagam, Alwyn R. Pais,
Vishwas Sharma, and Kunal M. Sadalkar
Poster Paper
Performance Analysis of a Multi Window Stereo Algorithm on Small
Scale Distributed Systems: A Message Passing Environment ........... 340
Vijay S. Rajpurohit and M.M. Manohara Pai
Ant Colony Algorithm in MANET-Review and Alternate Approach for
Further Mo dificati on ............................................. 344
Jyoti Jain, Roopam Gupta, and T.K. Bandhopadhyay
XIV Table of Contents
RDCLRP-Route Discovery by Cross Layer Routing Protocol for Manet
Using Fuzzy Logic ............................................... 348
Mehajabeen Fatima, Roopam Gupta, and T.K. Bandhopadhyay
A New Classification Algorithm with GLCCM for the Altered
Fingerprints ..................................................... 352
R. Josphineleela and M. Ramakrishnan
Footprint Based Recognition System ............................... 358
V.D. Ambeth Kumar and M. Ramakrishan
An Efficient Approach for Data Replication in Distributed Database
Systems ........................................................ 368
Arun Kumar Yadav, Ajay Agarwal, and S. Rahmatkar
Unicast Quality of Service Routing in Mobile Ad Hoc Networks Based
on Neuro-fuzzy Agents............................................ 375
V.R. Budyal, S.S. Manvi, and S.G. Hiremath
Performance Comparison of Routing Protocols in Wireless Sensor
Networks ....................................................... 379
Geetika Ganda, Prachi, and Shaily Mittal
Security and Trust Management in MANET......................... 384
Akash Singh, Manish Maheshwari, Nikhil, and Neeraj Kumar
Analytical Parametric Evaluation of Dynamic Load Balancing
Algorithms in Distributed Systems ................................. 388
Mayuri A. Mehta and Devesh C. Jinwala
Wavelet Based Electrocardiogram Compression at Different
Quantization Levels .............................................. 392
A. Kumar and Ranjeet
Content Based Image Retrieval by Using an Integrated Matching
Technique Based on Most Similar Highest Priority Principle on the
Color and Texture Features of the Image Sub-blocks .................. 399
Ch. Kavitha, M. Babu Rao, B. Prabhakara Rao, and A. Govardhan
Understanding the Impact of Cache Performance on Multi-core
Architectures .................................................... 403
N. Ramasubramaniam, V.V. Srinivas, and P. Pavan Kumar
An Overview of Solution Approaches for Assignment Problem in
Wireless Telecommunication Network ............................... 407
K. Rajalakshmi and M. Hima Bindu
Design of Domain Specific Language for Web Services QoS Constraints
Definition ....................................................... 411
Monika Sikri
Tab le of Cont ents X V
Modified Auxiliary Channel Diffie Hellman Encrypted Key Exchange
Authentication Protocol .......................................... 417
Nitya Ramachandran and P. Yogesh
Bilateral Partitioning Based Character Recognition for Vehicle License
Plate ........................................................... 422
Siddhartha Choubey, G.R. Sinha, and Abha Choubey
Strategies for Parallelizing KMeans Data Clustering Algorithm ........ 427
S. Mohanavalli, S.M. Jaisakthi, and C. Aravindan
A Comparative Study of Different Queuing Models Used in Network
Routers for Congestion Avoidance .................................. 431
Narendran Rajagopalan and C. Mala
SAR Image Classification Using PCA and Texture Analysis ........... 435
Mandeep Singh and Gunjit Kaur
Performance of WiMAX/ IEEE 802.16 with Different Modulation and
Coding ......................................................... 440
Shubhangi R. Chaudhary
A Novel Stair-Case Replication (SCR) Based Fault Tolerance for MPI
Applications..................................................... 445
Sanjay Bansal, Sanjeev Sharma, and Ishita Trivedi
Evaluating Cloud Platforms- An Application Perspective.............. 449
Pankaj Deep Kaur and Inderveer Chana
An Intelligent Agent Based Temporal Action Status Access Control
Model for XML Information Management ........................... 454
N. Jaisankar and A. Kannan
An Economic Auction-Based Mechanism for Multi-service Overlay
Multicast Networks .............................................. 461
Mohammad Hossein Rezvani and Morteza Analoui
Server Virtualization: To Optimizing Messaging Services by Configuring
Front-End and Back-End Topology Using Exchange Server in Virtual
Environments ................................................... 468
R. Anand and T. Deenadayalan
Blind Source Separation for Convolutive Audio Mixing ............... 473
V. Jerine Rini Rosebell, D. Sugumar, Shindu, and Sherin
ICA Based Informed Source Separation for Digitally Watermarked
Audio Signals ................................................... 477
R. Sharanya, D. Sugumar, T.L. Sujithra, Susan Mary Bose, and
Divya Mary Koshy
XVI Table of Contents
Evaluation of Retrieval System Using Textural Features Based on
Wavelet Transform ............................................... 481
Lidiya Xavier and I. Thusnavis Bella Mary
Behavioural Level Watermarking Techniques for IP Identification Based
on Testing in SOC Design ......................................... 485
Newton david Raj, Josprakash, AntopremKumar, Daniel, and
Joshua Thomas
Rough Set Approach for Distributed Decision Tree and Attribute
Reduction in the Disseminated Environment ........................ 489
E. Chandra and P. Ajitha
MIMO and Smart Antenna Technologies for 3G and 4G .............. 493
Vanitha Rani Rentapalli and Zafer Jawed Khan
A GA-Artificial Neural Network Hybrid System for Financial Time
Series Forecasting ................................................ 499
Binoy B. Nair, S. Gnana Sai, A.N. Naveen, A. Lakshmi,
G.S. Venkatesh, and V.P. Mohandas
A Preemptive View Change for Fault Tolerant Agreement Using Single
Message Propagation ............................................. 507
Poonam Saini and Awadhesh Kumar Singh
A Model for Detection, Classification and Identification of Spam Mails
Using Decision Tree Algorithm .................................... 513
Hemant Pandey, Bhasker Pant, and Kumud Pant
Author Index .................................................. 517
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 1–5, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Efficient Object Motion Prediction Using Adaptive Fuzzy
Navigational Environment
Vijay S. Rajpurohit1 and M.M. Manohara Pai2
1 Gogte Institute of Institute of Technology, Belgaum, India
vijaysr2k@yahoo.com
2 Manipal Institute of Institute of Technology, Manipal, India
mm.pai@rediffmail.com
Abstract. This paper proposes an adaptive Fuzzy rule based motion prediction
algorithm for predicting the next instance position of a moving object. The pre-
diction algorithm is tested for real-life bench-marked data sets and compared
with existing motion prediction techniques. Results of the study indicate
that the performance of the predictor is comparable to the existing prediction
methods.
Keywords: Short Term Motion prediction, Rule base Optimization, Fuzzy Pre-
dictor algorithm, Adaptive navigational Environment.
1 Introduction
Short term Object motion prediction in a dynamic Robot navigation environment
refers to, the prediction of next instance position of a moving object based on the
previous history of its motion. Research literature has addressed solutions to the short
term object motion predictions with different methods such as, Curve fitting or Re-
gression methods, Neural network based approaches, Hidden Markov stochastic mod-
els, Bayesian Occupancy Filters, Extended Kalman Filter and Stochastic prediction
model[1][2][4]6[7]. The design of a navigational model in an automated mobile Ro-
bot system is influenced by its specific applications, the environment in which it op-
erates and the sensory system. Many navigational model representations have been
proposed, tested and implemented[3].
Based on the literature survey it is observed that i) The existing models lack flexi-
bility in handling the uncertainties of the real life situations. ii) Probabilistic models
sometimes fail to model the real-life uncertainties. iii) The existing prediction tech-
niques show poor response time due to their complex algorithmic structure. iv) Most
of the approaches validate the results with simulated data.
The present work provides a novel solution for short term motion prediction using
adaptive Fuzzy prediction technique. History of moving object motion positions are
captured in the form of Fuzzy rule base and the next instance object position is pre-
dicted using fuzzy inference process. Because of the multi valued nature of fuzzy
logic this approach enjoys high robustness in dealing with noisy and uncertain data.
2 V.S. Rajpurohit and M.M. Manohara Pai
However, direct implementation of the rule base is not suitable for real-life navigation
systems due to the formation of huge number of rules. To overcome this drawback
rule-base is optimized by adaptive navigational environment.
2 Fuzzy Rule Based Object Motion Prediction
The navigational
environmen
t
is modeled as Fuzzy World model [3] which can be
observed in most of the applications. The Fuzzy
representation
of the
environmen
t
is shown in Figure 1 with numerical notation
for each region.
Fig. 1. Division of
Navigation
Space into Fuzzy subsets of Range
an
d
Direction
In the rule-base formation phase, rules are defined and added to the
rule
base using
real-life data, expert knowledge base and a
simulator. At
time
t1,
the position (An-
gle and Range) of the moving object from the Robot is
read.
Using Fuzzification the
observed data is converted to Fuzzy value.
At
time
t2
(
t2 > t1
and
t2
−
t1 >
δ,
where δ is threshold time difference greater than
or
equal to 1 sec), the sensor reads
the position of the same object. The read
value
is converted to Fuzzy value. The same
process is followed at time t3 (
t3 >
t
2
and
t3
−
t2 = t2
−
t1)
to get the Fuzzy value
of the location of the same
ob
ject
under
observation.
A Fuzzy rule with the positions
of the moving object at
time
t1 and t2 as the
an
teceden
t
and the position of the ob-
ject at time t3 as
the
consequen
t
is formed and added to the rule-base. Each rule in
the rule-base is
represented as
I
F (R1 , θ1) and (R2, θ2)
T
H E N (R3,
θ
3)
where R1 and θ1
represe
n
t
the Range and the Angle respectively of the
ob
ject
at
time t1 , R2 and θ2
represen
t
the Range and the Angle respectively of
the
ob
jec
t
at
time t2, and R3 and θ3
represen
t
the Range and the Angle
respectiv
ely
of the object
at time
t3.
Efficient Object Motion Prediction Using Adaptive Fuzzy Navigational Environment 3
Similar rules are added to the rule-base for
differen
t
objects observed at
various
positions in the navigation environment
.
In the
implementation
phase of the
predic-
tor,
the Robot observes the
mo
v
ing
ob
jec
t
at time t1 and t2 and sends the data to the
Fuzzy predictor
algori
thm.
With the
application
of Fuzzy inference process, predic-
tion of the next
instance
position of the moving object is carried out. The complete
process of short
term
motion prediction is
represented
in Figure
2.
Fig. 2. Short term motion prediction
3 Rulebase Optimization Using Adaptive Navigational
Environment
To enhance the performance of the predictor algorithm, the basic navigation environ-
ment is altered such that at nearer distance only three, at moderate distance five
Fig. 3. Adaptive division of navigation space into Fuzzy regions
4 V.S. Rajpurohit and M.M. Manohara Pai
and at the far distance seven Fuzzy membership functions are defined for angular
subset by merging adjacent members in the angular subset(Fig. 3). By defining Adap-
tive navigational environment, the number of Fuzzy rules can be decreased as well as
the accuracy of the results can be further improved.
4 Experimental Results
The Fuzzy predictor
algorithm
is developed in
C++
language. The
algorithm
is
tested on 1.66 GHz machine in
VC++ environment.
The tests are carried out
for
real-life
benchmarked
datasets [5]. Figure 4 represents the
mo
vemen
t
of the ob-
jects from left to
righ
t
direction and the corresponding short
term
motion prediction
path. Pi and Ai
represen
t
the predicted and the actual
pa
t
h
traversed by the moving
object. Pi(G) and Ai(G)
represen
t
the predicted
goal
and the actual goal of the
object. A1 is the actual path observed and A1(G) is the actual goal reached by the
object
A1.
Fig. 4.
Prediction
graphs showing the few of the
path prediction solutions
for Short term
motion
prediction
Table 1. Comparison of Short term predictors
Short Term Predictor Relative Error Response time
in seconds
Neural Network predictor 6-17% 560x 10-3 sec
Bayesian Occupancy Filters 1-10% 100 x 10-3 sec
Extended Kalman Filter 1-20% 0.1 sec
Proposed Fuzzy Predictor Algorithm 1-10% 02 x 10-3 sec to
05 x 10-3 sec
Efficient Object Motion Prediction Using Adaptive Fuzzy Navigational Environment 5
Table 1 compares a few of the well known prediction techniques which are re-
implemented and compared with the developed Fuzzy predictor in respect of response
time and relative error. From the table it can be observed that the performance of the
predictor is comparable with regard to relative error but better than the other predic-
tion methods as far as response time is concerned.
5 Conclusion
In a dynamic navigation system the Robot has to avoid stationary and moving objects
to reach the final destination. Short Term motion prediction for moving objects in
such an environment is a challenging problem. This paper proposes a simplified ap-
proach for predicting the future position of a moving object using fuzzy inference
rules derived from expert knowledge. Fuzzy based prediction is more flexible, can
have more real life parameters, comparable to the existing approaches and suited for
real life situations. The results of the study indicate that, the Fuzzy predictor algo-
rithm gives comparable accuracy with quick response time when compared to existing
techniques.
Acknowledgments. The authors are thankful to the benchmark dataset provided by
EC Funded CAVIAR project, CMU Graphics lab and Motion capture web group.
References
1. Foka, A., Trahanias, P.E.: Predictive Autonomous Robot navigation. In: Proceedings of the
2002 IEEE/RSJ International Conference on Intelligent Robots and Systems, EFPL,
Lausanne, Switzerland, pp. 490–494 (October 2002)
2. Fayad, C., Web, P.: Optimized Fuzzy logic based algorithm for a mobile robot collision
avoidance in an unknown environment. In: 7th European Congress on Intelligent Tech-
niques & Soft Computing, Aachen, Germany, September 13-16 (1999)
3. Angelopoulou, E., Hong, T.-H., Wu, A.Y.: World Model Representa- tion for Mobile Ro-
bots. In: Proceedings of the Intelligent Vehicles 1992 Symposium, pp. 293–297 (1992)
4. Madhavan, R., Schlenoff, C.: Moving Object Prediction for Off-road Au- tonomous Navi-
gation. In: Proceedings of the SPIE Aerosense Conference, April 21-25 (2003)
5. Fisher, R., Santos-Victor, J., Crowley, J.: CAVIAR Video Sequence Ground Truth (2001),
http://homepages.inf.ed.ac.uk/rbf/CAVIAR/
6. Zhuang, H.-Z., Du, S.-X., Wu, T. -j.: On-line real-time path planning of mobile Robots in
dynamic uncertain environment. Journal of Zheing University Science A, 516–524 (2006)
7. Zhu, Q.: Hidden Markov Model for Dynamic Object Avoidance of Mobile Robot Naviga-
tion. IEEE Transactions on Robotics and Automation, 390–396 (1991)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 6–12, 2011.
© Springer-Verlag Berlin Heidelberg 2011
An Efficient Protocol Using Smart Interval for
Coordinated Checkpointing
Jagdish Makhijani1, Manoj Kumar Niranjan2, Mahesh Motwani3, A.K. Sachan4,
and Anil Rajput5
1 Rustamji Institute of Technology, BSF Academy, Tekanpur
j_makhijani@yahoo.com
2 Rustamji Institute of Technology, BSF Academy, Tekanpur
manoj_niranjan2000@yahoo.co.in
3 Rajiv Gandhi Technical University, Bhopal
mahesh.7@sify.com
4 Radharaman Institute of Technology & Science, Bhopal
sachanak_12@yahoo.com
5 Bhabha Engineering Research Institute, Bhopal
dranilrajput@hotmail.com
Abstract. Checkpointing using message logging is a very popular technique for
fault tolerance in distributed systems. The proposed protocol controls the lost
messages, orphan messages and also simplifies garbage collection which is not
available in most of existing protocols. In the protocol, all processes take
checkpoints at the end of their respective smart interval to form a global consis-
tent checkpoint. Since the checkpointing is allowed only within smart interval,
the protocol will minimize various overheads like checkpointing overhead,
message logging overhead etc.
Keywords: Distributed Systems, Checkpointing, Fault Tolerance and Message
Logging, Smart Interval.
1 Introduction
A distributed system consists of multiple autonomous computers that communicate
through a computer network in order to achieve a common goal. The distributed com-
puting systems generally tolerate undesired changes in their internal structure or ex-
ternal environment in regular working which can be referred to as faults. A Fault may
be a design or operational fault and may occur at once or many times. To make a sys-
tem fault tolerable, fault tolerance techniques such as Checkpointing may be used.
Checkpointing is the method of periodically recording the state of the system in stable
storage. Any such periodically saved state is called the checkpoint of the process [1].
A global state [2] of a distributed system is a set of individual process states, on per
process [1]. Checkpointing may be one of two types, i.e., independent and coordi-
nated checkpointing. In independent checkpointing, each process takes checkpoint
independently without requiring any synchronization when a checkpoint is taken [3].
An Efficient Protocol Using Smart Interval for Coordinated Checkpointing 7
In coordinated checkpointing, the processed coordinate their checkpointing action in
such a way that the set of local checkpoints taken is consistent [4,5,6].
The present work suggests a new coordinated checkpointing algorithm which ef-
fectively manages the lost and orphan messages. In this algorithm, each process takes
turn to act as checkpoint initiator. The checkpoint initiator sends messages to other
process to be prepare for checkpoint and then to take checkpoint. However, a process
has to maintain a log of received, send, and unacknowledged messages of the current
checkpointing interval. The initiator issues commit message to all other processes
after receiving checkpoint from all processes. The set of these checkpoints made per-
manent after issue of commit. If initiator does not receive all the local checkpoints, it
issues abort message for not making the tentative checkpoint permanent.
2 Existing Works
In the existing work, the initiator communicates with other processes to create a
checkpoint. In these old checkpointing protocols, if message communication can take
place after checkpoint request of initiator, the global checkpoint may be inconsistent.
This is shown in fig. 1 in which message m is sent by P0 after receiving a checkpoint
request from the initiator. If m reaches P1 before the checkpoint request, the check-
point will become inconsistent because checkpoint c1,x confirms that message m is
received from P0, while checkpoint c0,x says that it is not sent from P0. [11]
Fig. 1. Message communication between P0 and P1 causing inconsistent checkpoint
In another protocol, the message communication is allowed within a fixed time in-
terval only. This concept reduces message communications [7] which is beneficial in
decreasing the communication overhead. The main drawback of this protocol is the
fixation of a particular process as initiator process. Since a fixed process will act as
initiator in entire system execution, thus the probability of failure will be high.
In another checkpointing protocol, the process initiator is not fixed which reduces
the probability of failure of initiator. The drawback of this protocol is that the mes-
sage communication could be accomplished at any time i.e., there is no concept of
fixed time interval for message communication. Hence it increases communication
overhead and output commit latency [8].
The proposed protocol overcomes to these shortfalls. The proposed protocol uses
a fixed time interval for message communications which controls the message
Initiato
r
Checkpoint Request
P0
P1
C0,x
C1,x
m
8 J. Makhijani et al.
communication. This fixed time interval is called smart interval. This concept reduces
the communication overhead. The protocol also gives chance to every process to act
as initiator process which reduces the probability of failure of initiator.
3 System Model
Let us consider a system of ‘n’ processes, P0, P1, ……, Pn-1. The no. of processes ‘n’
is fixed for the duration of execution. Let the checkpoints be denoted as CPki, i.e.,
initial checkpoint CPk0 (i=0), first checkpoint CPk1 (i=1), second checkpoint CPk2
(i=2) and so on (here k is the process no.). The initial checkpoint is taken when the
system is being initialized. Each process maintains its own independent data struc-
tures, states and computations. Processes have no shared memory and no global clock.
All communications among processes are through message passing only. We are as-
suming followings:
1)
The underlying network guarantees reliable FIFO (First In First Out) delivery
of messages between any pair of processes. The assumption of FIFO delivery
assures the message synchronization.
2)
Each process takes turn to initiate checkpointing at regular interval. The initial
checkpoint (CPk0) is taken at the time of system initialization and initiated by
P0. The next checkpoint, i.e. first checkpoint (CPk1) will be initiated by P1 and
so on.
3)
The initiator process cannot be fail. If the initiator process fails, the global
checkpoint (which is always stored at initiator process) will be lost and the en-
tire system process will be collapsed.
The message communication will took place only in smart interval. The smart inter-
val is a specified time interval which is elapsed between the control messages for pre-
pare checkpoint and take checkpoint. If any process sends a message within smart
interval, it has to be logged and the process execution is continued. This enables han-
dling of lost messages. [10] The initiator process sends the control messages for pre-
pare checkpoint and take checkpoint to other processes.
4 Protocol Description
The checkpoint initiator process sends checkpoint-prepare-request-message to other
processes to start checkpointing. The other processes send their responses to the ini-
tiator process. If initiator process received replies from all processes within smart-
interval then it sends take-checkpoint-request-message and if initiator process does
not receive replies from any process within smart-interval then it will send abort-
checkpoint-request-message. The set of checkpoint of all processes received by initia-
tor process is called global checkpoint. A local checkpoint is denoted by CPki where k
is the process id and i is the checkpoint number. The ith global checkpoint is the set
CPi={CP0i, CP1i,………, CPn-1i} in a system of n processes. CPi is said to be consis-
tent if and only if ∀j,k∈[0,n-1]:j≠k ⇒(CPji→CPki) where → denotes the happened-
before relation described by Lamport in [9].
The maximum transmission delay to reach a message to destination is t. The T is
the checkpointing interval. Here T>3t, since checkpoint interval (T) is obviously
An Efficient Protocol Using Smart Interval for Coordinated Checkpointing 9
greater than smart-interval and the length of smart-interval is bound to be at least 3t to
survive the transmission delay of control messages (checkpoint-prepare-request-
message, response of checkpoint-prepare-request-message and take-checkpoint-
request-message and each transmission will take at least t) and to enable logging of
computational messages.
Fig. 2. Diagram showing message communication during smart interval
Now, let us define the following terms:
tprep=instant at which initiator process starts sending prepare request
trec=instant at which a process receives prepare request
ts=instant at which a computation message is sent
tr=instant at which a computation message is received
Ttrns=maximum transmission time for message including allowable delay (which is t)
Tag1=normal acknowledgement
Tag2=acknowledgement indicating that sender must keep this message logged
save_state (Pi)=procedure that saves the current state of process Pi
ack[]=keeps record of whether acknowledgement has come for the corresponding
message (ack[a] is set to 0 when ath message is sent and set to 1 when acknowledge-
ment with tag1 for ath message comes back, otherwise it is set to 0 if acknowledge-
ment with tag2 comes back for ath message.)
send(), receive()=functions for sending and receiving messages respectively.
s_id=sender id
5 Checkpointing Process
The checkpoint process starts at the time of system initialization. After T time interval
(which is decided by the programmer) of previous checkpoint, the next initiator
process starts the process of next checkpoint. The initiator process Pi sends check-
point-prepare-request-message to all other processes at tprep. On receiving checkpoint-
prepare-request-message, each process write tentative checkpoint after sending
response to the initiator.
10 J. Makhijani et al.
1)
Now, if initiator receives response from all processes, within (tprep+2*Ttrns), the
initiator process sends take-checkpoint-request-message to all processes. When
receiver receives take-checkpoint-request-message from initiator process, the
tentative checkpoint is made permanent. This will save the states of all proc-
esses which are responsible for preparing a global checkpoint. Now, suppose if
one or more process fails after responding to checkpoint-prepare-request-
message, then the tentative checkpoint (which is prepared in response to check-
point-prepare-request-message) is used to recover the failed process.
2)
Now suppose if one or more process fails to respond to checkpoint-prepare-
request-message, the initiator process sends abort-checkpoint-request-message
to all processes. On receiving this, the tentative checkpoint is deleted. The copy
of unacknowledged message keeps in a log in this case.
6 Algorithm
Step-I:
This step is executed at initiator process Pi
i. Send checkpoint-prepare-request-message to remaining processes at tprep for
(k+1)th checkpoint
ii. Remove (k-1)th checkpoint, if exist.
iii. Receive response from other processes within (tprep+2*Ttrns)
iv. If all processes respond positively then
Send take-checkpoint-request-message to all processes
Else (if even a single process does not respond positively or response does not ar-
rive to initiator process)
a. Send abort-checkpoint-request-message to all processes
b. Retain copies of unacknowledged messages in a log
Step-II:
This step is executed at other process Poth
i. Receive checkpoint-prepare-request-message from initiator at trec
ii. Send own response to initiator
iii. If response is positive then Call save_state(Poth) to write tentative-checkpoint
asynchronously
iv. Wait for decision of Pi till (trec+Ttrns+Ttrns)
v. If received decision is take-checkpoint-request-message then Change status of
tentative-checkpoint to permanent
Else
Delete tentative-checkpoint
vi. Delete messages whose acknowledgements have received. Log unacknow-
ledged messages.
Step-III:
This step is executed at any process Pany for receiving message
i. If ((checkpoint number in message)=(checkpoint number in Pany))
An Efficient Protocol Using Smart Interval for Coordinated Checkpointing 11
a. Send (tag1,s_id)
b. Receive(message)
ii. else if ((checkpoint number in message)>(checkpoint number in Pany))
a. save_state(Pany)
b. send(tag1,s_id)
c. receive(message)
iii. else if ((checkpoint number in message)<(checkpoint number in Pany))
a. send (tag2,s_id)
b. receive(message)
Step-IV:
This steps is executed at any process Pany for writing unacknowledged messages
i. for all k
if (ack[k]=0) then write kth message in buffer
7 Performance Results
The proposed algorithm is simulated in Windows Environment using Windows XP
Operating System and Visual Studio. It is assumed that coordinator process and net-
work will never fail. The result shows that a distributed system with proposed algo-
rithm will not fail, whether the total execution time may increase. This increase in
execution time will depend on no. of fault. Table-1 shows the summarized result:
Table 1. Performance Result of proposed algorithm
Total
Time to
Com-
plete
No. of
Process
Time to
Check
point
Checkpoint
Preperation
Time
Time to Com-
plete without
Error (with
algorithm)
No. of
Errors
occurred
Execution
Time In-
crease %
(with errors)
Execution Time
Increase %
check pointing
(with errors)
100 10 1 0.1 110 10 21.00 19.09
150 10 1 0.1 165 10 17.33 10.51
200 10 1 0.1 220 10 15.50 7.05
100 20 1 0.1 110 10 21.00 19.09
150 20 1 0.1 165 10 17.33 10.51
200 20 1 0.1 220 10 15.50 7.05
8 Conclusion
Consistent checkpointing is formed in a distributed manner by using proposed check-
pointing protocol. The checkpointing protocol of this paper also manages the unac-
knowledged messages to decrease the communication overhead. A global checkpoint
includes each and every checkpoint taken by a process. Hence the last global check-
point has to be retained. This decrease a lot of overhead since it required only 3*(n-1)
messages during checkpointing in ideal case for n number of processes. The proposed
protocol uses the concept of smart interval which allows message communication in a
fixed time interval. This provides faster message communication since the bandwidth
will be used for less no. of messages. The protocol minimizes output commit latency
12 J. Makhijani et al.
and simplifies garbage collection. The message communication may result high out-
put commit latency due to communication required and saving the message log to
stable storage. The proposed protocol allows message communication only in smart
interval. Since no. of messages is reduced, output commit latency is minimized. Gar-
bage collection is a technique of managing storage memory efficiently by removing
unwanted checkpoints from stable storage. In the proposed protocol, whenever initia-
tor process Pi sends checkpoint-prepare-request-message for (k+1)th checkpoint, the
protocol will automatically delete the (k-1)th global checkpoint which results simpli-
fied garbage collection. The protocol is useful in tolerating all types of software faults
occurred on non-initiator processes.
There is no provision for failure of initiator process which may be considered as a
drawback of our protocol.
References
1. Manivannan, D., Netzer, R.H.B., Singhal, M.: Finding Consistent Global Checkpoints in a
Distributed Computation. IEEE Trans. On Parallel & Distributed Systems 8(6), 623–627
(1997)
2. Tsai, J., Kuo, S.: Theoretical Analysis for Communication-Induced Checkpointing Proto-
cols with Rollback-Dependency Trackability. IEEE Trans. on Parallel & Distributed
Systems 9(10), 963–971 (1998)
3. Bhargava, B., Lian, S.R.: Independent Checkpointing and Concurrent Rollback for Recov-
ery in Distributed Systems-An Optimistic Approach. In: Proceeding of IEEE Symposium
on Reliable Distributed Systems, pp. 3–12 (1988)
4. Cao, G., Singhal, M.: On Coordinated Checkpointing in Distributed Systems. IEEE Trans-
actions on Parallel And Distributed Systems 9(12), 1213–1222 (1998)
5. Sharma, D.D., Pradhan, D.K.: An Efficient Coordinated Checkpointing Scheme for Multi-
computers. In: Proc. IEEE Workshop on Fault-Tolerant Parallel and Distributed Systems,
pp. 36–42 (June 1994)
6. Elnozahy, E.N., Johnson, D.B., Zwaenepoel, W.: The Performance of Consistent Check-
pointing. In: Proc. 11th Symp. Reliable Distributed Systems, pp. 39–47 (October 1992)
7. Subba Rao, C.D.V., Naidu, M.M.: A New, Efficient Coordinated Checkpointing Protocol
Combined with Selective Sender-Based Message Logging. In: IEEE/ACS International
Conference on Computer Systems and Applications, AICCSA 2008, pp. 444–447 (2008)
8. Neogy, S., Sinha, A., Das, P.K.: CCUML: A Checkpointing Protocol for Distributed Sys-
tem Processes. In: IEEE Transactions on TENCON 2004, IEEE Region 10 Conference,
November 21-24, vol. B, pp. 553–556 (2004)
9. Chandy, K.M., Lamport, L.: Distributed Snapshots: Determining Global States of Distrib-
uted Systems. ACM Trans. on Computer Systems 3, 63–75 (1985)
10. Subba Rao, C.D.V., Naidu, M.M.: A Survey of Error Recovery Techniques in Distributed
Systems. In: Proc. 28th Annual Convention and Exihibition of IEEE India Council,
pp. 284–289 (December 2002)
11. Mootaz Elnozahy, E.N., Alvisi, L., Wang, Y.-M., Johnson, D.B.: A Survey of Rollback-
Recovery Protocols in Message-Passing Systems. ACM Computing Surveys
(CSUR) 34(3), 375–408 (2002)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 13–18, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Face Recognition System Using Discrete Wavelet
Transform and Fast PCA
K. Ramesha1 and K.B. Raja2
1 Department of Telecommunication Engineering, Vemana Institute of Technology,
Koramangala, Bangalore-560034
rameshk13@yahoo.co.uk
2 Department of Electronics and Communication Engineering, University Visvesvaraya
College of Engineering, Bangalore
University, K.R. Circle, Bangalore-560001
Abstract. The face recognition system is used to create a national database for
the purpose of identity cards, voting in an electoral systems, bank transaction,
food distribution system, control over secured areas etc. In this paper we pro-
pose the Face Recognition System using Discrete Wavelet Transform and Fast
PCA (FRDF). The Discrete Wavelet Transform is applied on face images of Li-
bor Spacek database and only LL subband is considered. Fast Principal Com-
ponent Analysis using Gram-Schmidt orthogonalization process is applied to
generate coefficient vectors. The Euclidean Distance between test and database
face image coefficient vectors are computed for face recognition based on the
threshold value. It is observed that the face recognition rate is 100% and the
proposed algorithm for the computation of eigenvalues and eigenvectors im-
proves the computational efficiency as compared to Principal Component
Analysis (PCA) with same Mean Square Error (MSE).
Keywords: Face Recognition, Fast Principal Component Analysis, Discrete
Wavelet Transform, Eigenvalue, Eigenvector, Error Vector.
1 Introduction
Biometric authentication of a person has many advantages compared to the existing
traditional methods such as personal identity number, a key, a card etc. Biometric
mainly classified based on the characteristics of a person into two groups viz., physio-
logical characteristics i.e., Face, Fingerprint, Hand geometry, Hand vein, Iris, Retina,
DNA, Facial thermo gram and behavioral characteristics such as Signature, Keystroke
dynamics, Speech. The physiological biometrics has an advantage over behavioral
biometrics as the characteristics of a particular person do not change over a period of
years. The face recognition system attains lot of interest in these days due to its appli-
cations such as proof of identity card for access control to physical facilities, fight
against terrorism and crime, access control to services, law enforcement system and
content based video processing system. Fully automatic robust face recognition
is required due to the widespread use of photo identity for personal security and
14 K. Ramesha and K.B. Raja
identification. Automatic face recognition system is more reliable, effective and easy
to implement compared to other biometric systems such as fingerprint, hand geome-
try, hand vein, iris, retina, facial thermo gram, Signature, Keystroke dynamics, DNA
and Speech and it does not require special knowledge of a person or cooperation.
The process of face recognition normally consists of two steps. The first step is
face detection and localization, in which faces have to be found in the input image and
separated from the background. The second step is face feature extraction and recog-
nition. Challenges in face recognition research are face illumination variations, face
rotation, different hair style of a person wearing different kinds of spectacles, ageing
effects and facial expressions. Eigenface based analysis is a method about the features
of the whole face appearance and even to the whole sample set. Face recognition un-
der Pattern recognition, the most popular and necessary problem is dimensionality
reduction. Over the past few years, several face recognition systems based on PCA
have been proposed [1]. Instead of using N intensity values for an N pixel image, it is
feasible to specify an image by a set of M features, where M<<N. The selected fea-
tures must be able to uniquely represent the right class for their corresponding facial
images.
Kishore S Kinare and Bhirud [2] proposed Two Dimensional Principal Component
Analysis (2DPCA) on wavelet subband. Haar, Daubechies, Coiflet, Symlet, Bior-
thogonal and Reverse Biorthogonal wavelet transforms are used to extract image fea-
tures of facial images by decomposing face image in subbands of 1 to 8. By using
2DPCA and Euclidean Distance (ED) measures the features are analyzed. Daw-Tung
Lin [3] developed a method utilizing PCA to perform facial expression recognition
using Hierarchical Radial Basis Function Network for the facial expression classifica-
tions based on local feature extraction by PCA from lips and eyes images. Sheifali
Gupta et al., [4] provided a method for face recognition using PCA, an eigenface ap-
proach, in which a small set of characteristic pictures are used to describe the varia-
tion between face images. ED is used to match facial images. Vinod Pathangay and
Sukhendu Das [5] proposed the use of selective subands for PCA based frontal face
recognition with variations in illumination and expression. Subband face representa-
tion is evaluated using PCA. Kai Chen and Le Jun Zhao [6] presented real time face
recognition and tracking system to detect face for recognition and tracking. Hybrid
algorithm wavelet, PCA and Support Vector Machine is used for face recognition.
Face tracking is done by using meanshift and Kalman filter.
2 Proposed Face Recognition System Model
The face recognition approach based on DWT and Fast PCA is as shown in Figure 1.
2.1 Face Image Database
The face images are collected from Libor Spacek database for eleven persons from F1
to F11. The data set used for training and testing purposes contains male, female and
old person images of 180 * 200 pixels size. In this data set, twenty images of each
person without background with very minor variation in head turn, tilt and slant are
considered. The data set has images of small changes in face position, because images
Face Recognition System Using Discrete Wavelet Transform and Fast PCA 15
Fig. 1. FRDF Face Recognition System
have been acquired in speech mode with no variation in hair style and lighting. The
face image considered for testing may be from database or from out of database.
2.2 Two Dimensional DWT
The decomposition is applied at different levels repeatedly on low frequency channel
(LL) to obtain next level decomposition. The image is decomposed into four subbands
LL, LH, HL, and HH subbands by applying 2D DWT on face image. The LL subband
corresponds to low frequency components of an image and HL, LH and HH are high
frequency components of an image corresponds to vertical, horizontal and diagonal
subbands respectively. The LL subband we obtain is half the original image. Figure 2
shows the image decomposition based on wavelet scales. 2D DWT gives dimensional
reduction for less computational complexity, insensitive feature extraction, and mul-
tiresolution data approximation. The transform decomposes an image and hence dif-
ferent facial expressions are attenuated by removing high frequency components.
Wavelet coefficients are obtained by convolving a target function with wavelet ker-
nels and mathematically DWT can be given as in Equation (1)
)2()(
)2()(
*
,
*
,
)( ⎪
⎩
⎪
⎨
⎧
−=
−=
=∑
∑
qngnxa
qnhnxd
DWT p
pqp
p
pqp
nx (1
)
The coefficients dp,q gives the component details to the wavelet function, where as ap,q
gives approximation components of the image. The h(n) and g(n) in the Equation (1)
Database Face Images Test Face I
m
a
g
es
2D-Discrete Wavelet Transform
Euclidian Distance Classifier
LL Sub band
FPCA
Coefficient Vecto
r
Face image Match/Non-match
16 K. Ramesha and K.B. Raja
Fig. 2. Image decomposition based on wavelet scales
are functions, gives the coefficients of high pass and low pass filters respectively. The
parameters p and q refers to wavelet scale and translation factors.
2.3 FPCA
PCA based system is affected by non-convergence state of the algorithm and high
MSE. The major difference is for facial features such as eigenvalues and eigenvectors
extraction using Gram–Schmidt Orthogonalization method instead of eigenvalue
decomposition method in PCA. Fast PCA using Gram-Schmidt orthogonalization
process to find leading eigenvectors converges in little iteration without any initial
setting. Face recognition using FPCA decreases the decision time of a system, espe-
cially when high resolution images are used, hence FPCA is computationally more
efficient, easy to implement and generates same MSE as that of PCA.
2.4 Identify the Known/Unknown Face Image
The Euclidean distance between database set and test faces becomes error vector and
the average error vector becomes the threshold value for face recognition. The mini-
mum ED between database and test image are recorded which leads to Difference
Error Vector (DEV). If the value of DEV is less than the threshold value, then the face
image is concluded as match otherwise non-match.
3 Performance Analysis and Results
For testing and analysis purpose, 400 images of 20 persons are considered. First 10
images of each person are used to create a face database. The second set of 10 images
of each person is used as test images to determine recognition rate. For Non-
matching, the test images are from different persons other than the persons used to
create database. It is observed from the Table 1 that the average recognition rate is
100% in the case of proposed algorithm compared to the average recognition rate of
80% in the case of existing algorithm.
The CPU time increases exponentially as the the database size increases in the case
of PCA method [7] of face recogntion, whereas in the case of proposed FRDF the
CPU time variation is very low for larger database. The FRDF algorithm using Harr
Face Recognition System Using Discrete Wavelet Transform and Fast PCA 17
Table 1. FPCA and FRDF Face Recognition Rates
Recognition Rate
Database FPCA [8] FRDF
F1 90% 100%
F2 100% 100%
F3 100% 100%
F4 50% 100%
F5 90% 90%
F6 90% 100%
F7 100% 100%
F8 100% 100%
F9 90% 100%
F10 0% 100%
F11 90% 100%
wavelet, scale-1 requires only few leading principal components to achieve 100%
recognition accuracy with less computational time.
4 Conclusion
The face recognition is normally used to identify a person for natural security. The
FRDF algorithm is proposed in the paper. The face images from Libor Spacek data-
base is considered for training and testing. The DWT is applied on face and consid-
ered only LL subband by leaving other subbands. The FPCA using Gram–Schmidt
orthogonalization process is applied on LL subband to generate leading eigenvalues to
compute face features. The Euclidean distance is used to compare the face features of
database and test face images to obtain face match/non-match. It is observed that in
the proposed algorithm CPU time and recognition rate is improved compared to the
existing algorithm. In future different transformations may be used on large face data-
base for robust identification with minimum time consumption.
References
1. Nicholl, P., Amira, A.: DWT/PCA Face Recognition using Automatic Coefficient Selection.
In: 4th IEEE International Symposium on Electronic Design, Test and Applications, Hong
Kong, pp. 390–393 (2008)
2. Kinare, K.S., Bhirud, S.G.: Face Recognition based on Two-Dimensional PCA on Wavelet
Subband. Int. J. Recent Trends in Engineering 2(2), 51–54 (2009)
3. Lin, D.-T.: Facial Expression Classification using PCA and Hierarchical Radial Basis Func-
tion Network. J. Information Science and Engineering, 1033–1046 (2006)
4. Gupta, S., Sahu, O.P., Gupta, R., Goel, A.: A Bespoke Approach for Face Recognition
using PCA. Int. J. Computer Science and Engineering 2(2), 155–158 (2010)
18 K. Ramesha and K.B. Raja
5. Pathangay, V., Das, S.: Exploring the use of Selective Wavelet Subbands for PCA based
Face Recognition. In: National Conference on Image Processing, pp. 182–185. IISc, Banga-
lore (2005)
6. Chen, K., Zhao, L.J.: Robust Real Time Face Recognition and Tracking System. J. Com-
puter Science and Technology 9(2), 82–88 (2009)
7. Sharma, A., Paliwal, K.K.: Fast Principal Component Analysis using Fixed-Point Algo-
rithm. ELSEVIER I. J. Pattern Recognition Letters 28, 1151–1155 (2007)
8. Sajid, I., Ahmed, M.M., Taj, I.: Design and Implementation of a Face Recognition System
using Fast PCA. In: International Symposium on Computer Science and its Applications,
Hobort, ACT, pp. 126–130 (2008)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 19–26, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Mining Indirect Association between Itemsets
B. Ramasubbareddy1, A. Govardhan2, and A. Ramamohanreddy3
1 Associate Professor, Jyothishmathi Institute of Technology and Science, Karimnagar, India
rsreddyphd@gmail.com
2 Professor & Principal, JNTUH college of Engineering, Karimnagar, India
govardhan_cse@yahoo.co.in
3 Professor, S.V.U. College of Engineering, S.V. University, Tirupati, India
ramamohansvu@yahoo.com
Abstract. Discovering association rules is one of the important tasks in data
mining. While most of the existing algorithms are developed for efficient min-
ing of frequent patterns, it has been noted recently that some of the infrequent
patterns, such as negative associations and indirect associations, provide useful
insight into the data. Existing indirect association mining algorithms mine indi-
rect associations between items and require two join operations. But in this
paper, we propose an algorithm for mining the complete set of indirect associa-
tions between pair of items and itemsets which require only one join operation.
Keywords: Data mining, positive and negative association rules, indirect
association.
1 Introduction
Association rule mining is a data mining task that discovers associations among items
in a transactional database. Association rules have been extensively studied in the
literature for their usefulness in many application domains such as recommender sys-
tems, diagnosis decisions support, telecommunication, intrusion detection, etc. Effi-
cient discovery of such rules has been a major focus in the data mining research.
From the celebrated Apriori algorithm [1] there have been a remarkable number of
variants and improvements of association rule mining algorithms [2]. A typical exam-
ple of association rule mining application is the market basket analysis. In this exam-
ple, the behavior of the customers is studied with reference to buying different
products in a shopping store. The discovery of interesting patterns in this collection of
data can lead to important marketing and management strategic decisions. For in-
stance, if a customer buys bread, what are chances that customer buys milk as well?
Depending on some measure to represent the said chances of such an association,
marketing personnel can develop better planning of the shelf space in the store or can
base their discount strategies on such associations/correlations found in the data. All
the traditional association rule mining algorithms were developed to find positive as-
sociations between items.
In [9], a new class of patterns called indirect associations has been proposed and its
utilities have been examined in various application domains. Consider a pair of items
20 B. Ramasubbareddy, A. Govardhan, and A. Ramamohanreddy
X and Y that are rarely present together in the same transaction. If both items are
highly dependent on the presence of another itemset M, then the pair (X, Y) is said to
be indirectly associated via M. There are many advantages in mining indirect associa-
tions in large data sets. For example, an indirect association between a pair of words
in text documents can be used to classify query results into categories [9]. For in-
stance, the words coal and data can be indirectly associated via mining. If only the
word mining is used in a query, documents in both mining domains are returned. Dis-
covery of the indirect association between coal and data enables us to classify the
retrieved documents into coal mining and data mining. There are also potential appli-
cations of indirect associations in many other real-world domains, such as competitive
product analysis and stock market analysis [9].
This paper is structured as follows: the next section contains preliminaries about
Indirect Association Rules, In Section3, existing strategies for mining indirect associ-
ation rules are reviewed. The proposed algorithm is presented in Section 4 for
finding all valid indirect association rules for pairs of multiple itemsets. Section 5
contains conclusions and future work.
2 Basic Concepts and Terminology
Let I = {i1, i2,. . . , im} be a set of m items. A subset X
⊆
I is called an itemset. A k-
itemset is an itemset that contains k items. Let D = {T1, T2,. . . , Tn } be a set of n
transactions, called a transaction database, where each transaction Tj, j = 1, 2, . . . , n,
is a set of items such that Tj
⊆
I. Each transaction is associated with a unique identifi-
er, called its TID. A transaction T contains an itemset X if and only if X
⊆
T. The
support of an itemset X is the percentage of transactions in D containing X. An item-
set X in a transaction database D is called “frequent itemset” if its support is at least a
user-specified minimum support threshold viz., minsup. Accordingly, an infrequent
itemset is an itemset that is not a frequent itemset.
2.1 Negative Association Rules
An association rule is an implication of the form X ⇒ Y, where X ⊂ I, Y ⊂ I, and X ∩
Y = ∅. Here, X is called the antecedent and Y is called the consequent of the rule. The
confidence of an association rule X ⇒ Y is the conditional probability that a transac-
tion contains Y, given that it contains X. The support of rule X ⇒ Y is defined as:
sup(X ⇒ Y) = sup(X ∪ Y). Negative association was first pointed out by Brin et al. in
[5]. Since then, many techniques for mining negative associations have been developed
[8, 9, 10]. In the case of negative associations we are interested in finding itemsets that
have a very low probability of occurring together. That is, a negative association be-
tween two itemsets X and Y, denoted as X ⇒ or Y ⇒ , means that X and Y
appear very rarely in the same transaction. Mining negative association rules is compu-
tational intractable with a naive approach because billions of negative associations may
be found in a large database while almost all of them are extremely uninteresting. This
problem was addressed in [8] by combining previously discovered positive associa-
tions with domain knowledge to constrain the search space such that fewer but more
interesting negative rules are mined. A general framework for mining both positive and
˥Y
˥X
Mining Indirect Association between Itemsets 21
negative association rules of interest was presented in [10], in which no domain know-
ledge was requires, and the negative association rules were given in more concrete
expressions to indicate actual relationships between different itemsets. However, al-
though the sets of the positive and negative itemsets of interest in the database were
minimized in this framework, the search space for negative itemsets of interest was
still huge. Another problem was that it tended to produce too many negative associa-
tion rules, thus the practical application of this framework remained uncertain.
2.2 Indirect Association
Indirect association is closely related to negative association, they are both dealing
with itemsets that do not have sufficiently high support. Indirect associations provide
an effective way to detect interesting negative associations by discovering only “in-
frequent itempairs that are highly expected to be frequent” without using negative
items or domain knowledge.
Definition (Indirect Association). A pair of itemsets X and Y is indirectly associated
via a mediator M, if the following conditions hold:
1. sup( X, Y ) < ts (Itepair Support Condition)
2. There exists a non-empty set M such that
(a) sup( X ∪ M) ≥ tf , sup( Y ∪ M) ≥ tf ; (Mediator Support Condition)
(b) dep( X , M) ≥ td, dep( Y, M) ≥ td, (Mediator Dependence Condition)
where dep(P, Q) is a measure of the dependence between itemsets P and Q.
The thresholds above are called itemset pair support threshold (ts), mediator sup-
port threshold (tf ), and mediator dependence threshold (td), respectively. In practice, it
is reasonably to set tf ≥ ts
Condition 1 is needed because an indirect relationship between two items is signif-
icant only if both items rarely occur together in the same transaction. Otherwise, it
makes more sense to characterize the pair in terms of their direct association.
Condition 2(a) can be used to guarantee that the statistical significance of the media-
tor set. In particular, for market basket data, the support of an itemset affects the amount
of revenue generated and justifies the feasibility of a marketing decision. Moreover,
support has a nice downward closure property which allows us to prune the combina-
torial search space of the problem. Condition2 (b) ensures that only items that are highly
dependent on the presence of X and Y will be used to form the mediator set.
Over the years, many measures have been proposed to quantify the degree of de-
pendence between attributes of a dataset. From statistics, the Chi-Square test is often
used for this purpose. However, the drawback of this approach is that it does not
measure the strength of dependencies between items [19]. Furthermore, the
Chi-Square statistic depends on the number of transactions in the database. As a re-
sult, other statistical measures of association are often used, including Pearson’s Φ
coefficient, Goodman and Krushkal’s λ, Yule’s Q and Y coefficients, etc [16]. Interest
factor is another measure that has been used quite extensively to quantify the strength
of dependency among items [11, 12, 13].
Definition. Given a pair of itemsets, say X and Y, its’ IS measure can be computed
using the following equation:
22 B. Ramasubbareddy, A. Govardhan, and A. Ramamohanreddy
, ,
(1)
Where P denotes the probability that the given itemset appears in a transaction.
3 Related Work in Indirect Association Rule Mining
It is observed that automated document translation systems tend to produce lexicon
translation tables that are full of indirectly-associated words [15]. A lexicon transla-
tion table encodes the probability that two words from different languages being se-
mantically equivalent to another. The presence of indirect association can pollute the
resulting tables, thereby reducing the overall precision of the system. An iterative
strategy was proposed in [15] to clean up existing translation tables by finding only
the most probable translations for a given word.
The notion of internal and external measures of similarity between attributes of a data-
base relation was introduced in [14]. Internal similarity between two attributes X and Y is
a measure whose value depends only on the values of X and Y columns. Conversely, ex-
ternal measure takes into account data from other columns (called the probe attributes).
Their notion of probe attributes is similar to mediators for indirect association in [14].
However, their sole purpose of using probe attributes is to perform attribute clustering.
Indirect association is closely related to the notion of negative association rules
[17]. In both cases, we are dealing with itemsets that do not have sufficiently high
support. A negative association rule discovers what are the items a customer will not
likely buy given that he/she buys a certain set of other items. Typically, the number of
negative association rules can be prohibitively large and the majority of them are not
interesting to a data analyst. The use of domain knowledge, in the form of item tax-
onomy, was proposed in [17] to decide what constitutes an interesting negative asso-
ciation rule. The intuition here is that items belonging to the same parent node in
taxonomy are expected to have similar types of associations with other items. If the
observed support is significantly smaller than its expected value, then there is a nega-
tive association exists between the items. Again, unlike indirect association, these
types of regularities do not specifically look for mediating elements.
Another related area is the study of functional dependencies in relational databases.
Functional dependencies are relationships that exist between attributes of a relation.
However, the emphasis of functional dependencies is to find dependent and indepen-
dent attributes for applications such as semantic query optimization [18] and reverse
engineering [18].
In [20], authors proposed an efficient algorithm, called HI-mine, based on a new data
structure, called HI-Struct, for mining the complete set of indirect associations between
items. Experimental results show that HI-mine’s performance is significantly better than
that of the previously developed algorithm for mining indirect associations on both syn-
thetic and real world data sets over practical ranges of support specifications.
In [21], IAM algorithm proceeds in four phases: an initialization phase, a pruning
phase, a bridge itemset calculation phase, and a ranking phase. The purpose of the initia-
lization phase is to allocate the memory needed. The second phase is a process of prun-
ing for the purpose of minimizing the search space of problem. The threshold value of
pruning is min-sup(s). The third phase, the Bridge Itemset Calculation Phase, is the
Mining Indirect Association between Itemsets 23
most important for this algorithm. The last phase, a ranking phase, is mainly to finish
the ranking operation according to 'the closeness value in the linked vector C for the
purpose of providing decision makers the most useful indirect association rules.
4 Algorithm
The existing work done on generating indirect associations between pair of items on-
ly. In this paper, we propose a new method which generates indirect associations be-
tween pair of itemsets. This method contains two algorithms. Algorithm1 finds set of
all frequent itemsets and set of all Valid Candidates (VC). An itemset V is said to be
Valid candidate if sup (V) ≤ ts and all subsets of V are frequent. Algorithm 2 finds set
of all indirect association rules between pairs of itemsets.
Algorithm1. Finding Frequent(F) and ValidCandidates (VC)
Input: TDB- Transactional Database, ms, ts
Output: F- Frequent itemsets, VC- ValidCandidates
Method:
1. Find F1, the set all frequent 1-itemsets
2. for(K=2;Fk-1 != Φ ; K++)
3. { CK = FK-1 ⋈ FK-1
// Pruning infrequent itemsets
4. for each c ε CK {
5. if any sub-set of c is not a member of FK-1 then
6. CK = CK –{ c} }
// find support count for each itemset in CK
7. for each c in CK
8. if support(c )≥ ms then Fk = Fk U { c }
9. for each c in CK
10. if support(c )≤ ts then VC= VC U { c }
11. F= F U FK }
Algorithm2. Mining Indirect Association Rules
Input: F, VC, tf, td
Output: Indirect Association Rules
Method:
1. for each I (= X U Y) ε VC {
2. for each i ε F {
3. If (support(X U i ) ≥ tf && support ( Y U i )≥ tf )
4. If (dependency( X U i ) ≥ td && dependency (Y U i)≥ td )
5. IAR= IAR U (X,Y/i ) } }
• Line1, each I is infrequent but all subsets of I are frequent because all items of VC
passed Apriori property.
• Line2, i is a frequent itemset
• Line3, since X , Y and i are frequent itemsets then X U i and Y U i may be frequent
• Line4, IS measure is used to find the dependency between two itemsets.
24 B. Ramasubbareddy,
5 Experimental Res
u
To evaluate the performanc
synthetic transactional datab
mented on java platform. W
e
set and mediator and mediat
o
Data set consisting of 5
4
mediator dependence as 0.
4
96,15,36 and 20 respective
l
and mediator dependency
v
Fig. 1. Graph showing the me
d
for 5400 transactions
Figure 2 is generated b
y
0.2,0.25,0.3,0.35,0.4 medi
a
ber of rules generated 31,2
7
Fig. 2. Graph showing the me
d
for 12000 transactions
A. Govardhan, and A. Ramamohanreddy
u
lts and
P
erformance Evaluation
e of proposed algorithm experiments are performed on
ases containing 5400 and 12000 transactions each and i
mp
e
concentrate on mediator support which is a support of it
e
o
r dependency which is estimated by “Eq. (1)”.
4
00 transactions with mediator support as 0.2,0.25,0.3,0
4
,0.45,0.5,0.55 and the total number of rules generate
d
l
y. Figure 1 shows the graph showing the mediator sup
p
v
s. to
t
al number of rules.
d
iator support and mediator dependency vs. total number of
r
y
considering 12000 transactions with mediator suppo
r
a
tor dependence as 0.4,0.45,0.5,0.55,0.6 and the total n
u
7
,7,7,and 6 respectively.
d
iator support and mediator dependency vs. total number of
r
two
m
ple-
t
em-
.35;
d
as
p
ort
r
ules
r
t as
u
m-
r
ules
Mining Indirect Association between Itemsets 25
6 Conclusion and Future Work
Existing indirect association mining algorithms mine indirect associations between
items whilst indirect association rules are called negative indirect association rules if
its mediator set contains presence and absence of items. In this paper, we propose an
efficient algorithm to discover all indirect associations between itemsets. In indirect
association mining for itempairs, algorithms require two join operations. To overcome
this disadvantage we have proposed a new algorithm to mine indirect associations
between itempairs and itemsets. This algorithm features, performing only one join
operation and generating indirect associations for pair of items and itemsets. In future
we propose to elaborate this work by conducting experiments on large databases to
test the scalability and to generate negative indirect association rules. Threshold selec-
tion is another issue that needs further investigation.
References
1. Agarwal, R., Aggarwal, C., Prasad, V.V.V.: Depth first generation of long patterns. In:
Proceedings of ACM-SIGKDD International Conference on Knowledge Discovery and
Data Mining (2000)
2. Agarwal, R., Aggarwal, C., Prasad, V.V.V.: A tree projection algorithm for generation of
frequent itemsets. Journal of Parallel and Distributed Computing, Special Issue on High
Performance Data (2000)
3. Ramasubbareddy, B., Govardhan, A., Ramamohanreddy, A.: Mining Positive and Nega-
tive Association Rules. In: IEEE ICSE 2010, Hefeai, China (August 2010)
4. Ramasubbareddy, B., Govardhan, A., Ramamohanreddy, A.: An Approach for Mining
Positive and Negative Association Rules. In: Second International Joint Journal Confe-
rence in Computer, Electronics and Electrical, CEE (2010)
5. Brin, S., Motwani, R., Ullman, J., Tsur, S.: Dynamic itemset counting and implication
rules for market basket data. In: Proceedings of the International ACM SIGMOD Confe-
rence, Tucson, Arizona, USA, pp. 255–264 (May 1997)
6. Tan, P., Kumar, V.: Interestingness measures for association patterns: A perspective. In:
KDD 2000 Workshop on Postprocessing in Machine Learning and Data Mining, Boston,
MA (August 2000)
7. Tan, P., Kumar, V.: Mining indirect associations in web data. In: Kohavi, R., Masand, B.,
Spiliopoulou, M., Srivastava, J. (eds.) WebKDD 2001. LNCS (LNAI), vol. 2356,
p. 145. Springer, Heidelberg (2002)
8. Savasere, A., Omiecinski, E., Navathe, S.: Mining for strong negative associations in a
large database of customer transactions. In: Proceedings of the 14th International Confe-
rence on Data Engineering, Orlando, Florida, pp. 494–502 (February 1998)
9. Tan, P., Kumar, V., Srivastava, J.: Indirect association: mining higher order dependencies
in data. In: Proceedings of the 4th European Conference on Principles and Practice of
Knowledge Discovery in Databases, Lyon, France, pp. 632–637 (2000)
10. Wu, X., Zhang, C., Zhang, S.: Mining both positive and negative association rules. In:
Proceedings of the 19th International Conference on Machine Learning (ICML 2002),
Sydney, Australia, pp. 658–665 (July 2002)
26 B. Ramasubbareddy, A. Govardhan, and A. Ramamohanreddy
11. Brin, S., Motwani, R., Silverstein, C.: Beyond market baskets: Generalizing association
rules to correlations. In: Proc. ACM SIGMOD Intl. Conf. Management of Data, Tuscon,
AZ, pp. 265–276 (1997)
12. Brijs, T., Swinnen, G., Vanhoof, K., Wets, G.: Using association rules for product assort-
ment decisions: A case study. In: Proc.of the Fifth ACM SIGKDD Conf. on Knowledge
Discovery and Data Mining, San Diego, Calif., pp. 254–260 (August 1999)
13. Cooley, R., Clifton, C.: TopCat: Data mining for topic identification in a text corpus. In:
Żytkow, J.M., Rauch, J. (eds.) PKDD 1999. LNCS (LNAI), vol. 1704, pp. 174–183.
Springer, Heidelberg (1999)
14. Das, G., Mannila, H., Ronkainen, P.: Similarity of attributes by external probes. In: Proc.
of the Fourth ACM SIGKDD Intl. Conf. on Knowledge Discovery and Data Mining, New
york, NY, pp. 23–29 (1998)
15. Melamed, D.: Automatic construction of clean broad-coverage translation lexicons. In: 2nd
Conference of the Association for Machine Translation in the Americas, ATMA 1996
(1996)
16. Reynolds, H.T.: The Analysis of Cross-Classifications. Macmillan Publishing Co., New
York (1997)
17. Savasere, A., Omiecinski, E., Navathe, S.: Mining for strong negative associations in a
large database of customer transactions. In: Proceedings of the 14th International Confe-
rence on Data Engineering, Orlando, Florida, pp. 494–502 (February 1998)
18. Tari, Z., Bukhres, O., Stokes, J., Hammoudi, S.: The reengineering of relational databases
based on key and data correlations. In: Sspaccapietra, S., Maryanski, F. (eds.) Searching
for Semantics: Data Mining, Reverse Engineering, etc. Chapman and Hall, Boca Raton
(1993)
19. Winkler, R., Hays, W.: Statistics: Probabilty, Inference and Decision, 2nd edn. Holt, Rine-
hart &Winston, New York (1975)
20. Wan, Q., An, A.: An Efficient Approach to Mining Indirect Associations, pp. 1–26. Kluw-
er Academic Publishers, Boston
21. Li, L., Xu, F., Wang, H., She, C., Zhihua Fan, I.A.M.: An Algorithm of Indirect Associa-
tion Mining. In: Proceedings of the 2004 International Conference on Intelligent Mecha-
tronics and Automation Chengdu, China (August 2004)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 27–32, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Reaction Attacks in the Matrix Scheme of NTRU
Cryptosystem
Rakesh Nayak1, Jayaram Pradhan2, and C.V. Sastry3
1 Associate Professor in Department of IT,
Sri Vasavi Engineering College, Tadepalligudem,
Andhra Pradesh, India
nayakrakesh8@gmail.com
2 Professor, Computer Sciences, Behrampur University,
Behrampur, Odisha, India
jayarampradhan@hotmail.com
3 Professor, School of Computer Science and informatics,
Sreenidhi Institute of Science and Technology,
Hyderabad, Andhra Pradesh, India
cvsastry40@yahoo.co.in
Abstract. An attacker produces a sequence of encrypted messages E, each of
which has a distinct probability, however small, of decrypting into a valid mes-
sage and also a probability of decrypting into an invalid message. The smallest
modification that an attacker can make to the cipher-text and can still decrypt it
correctly, gives information about the private key used to encrypt the message.
In this paper we assume that the attacker knows or the gets hold of intermediate
text which arises during the process of decryption.
Keywords: Encryption, Decryption, Cipher Text, Private Key, Public Key.
1 Introduction
In a recent preprint, Hall, Goldberg, and Schneier [1] have proposed an attack against
several public key cryptosystems based on lattice problems. They call their attack a
Reaction Attack. In their paper they describe how to mount a Reaction Attack on a
number of cryptosystems, including those suggested by McEliece, Atjai-Dwork, and
Goldreich-Goldwasser-Halevi. Since the NTRU public key cryptosystem is also based
on an underlying lattice problem, it is natural that a Reaction Attack should exist for
NTRU. In this paper we explain how an NTRU Reaction Attack would work and we
describe a number of ways in which a user of the NTRU public key cryptosystem can
thwart such attacks. Our main point will be that before such an attack can begin to
yield useful information, it must deform the encrypted message sufficiently that the
attack can be detected by the NTRU user. We also note that such an attack can only
hope to succeed if a single private key is used for the decryption of a large number of
messages. This is a situation that will rarely arise in an NTRU based system, since it
is easy to generate a key, meaning that keys can be changed frequently.
28 R. Nayak, J. Pradhan, and C.V. Sastry
2 Mathematical Preliminaries
2.1 Modular Arithmetic
Let A be an n x n matrix defined as
Aa a
a a
And let p be an integer. Then we define A (Mod p) as
AMod paMod p aMod p
aMod p aMod p
If p is a positive integer, and A, B are two matrices then, A is said to be congruent to
B modulo m, if A – B is divisible by p denoted by a ≡ b modulo m or simply a ≡ b
(mod m).
A matrix and congruence with the same modulus may be added, subtracted, and
multiplied just as is done with matrix operations.
The properties of modular arithmetic on matrix [13,14] are
(i) [A (mod p) + B(mod p)](mod p) = (A + B)(mod p)
(ii) [A (mod p) * B (mod p)](mod p) = (A*B)(mod p)
(iii) A-1 (mod p) = (A (mod p))-1(mod p)
(iv) If A ≡ B (Mod p) and C ≡ D (Mod p) then Ax + Cy (Mod p) = Bx + Dy (Mod p)
for all integers x and y.
(v) If A ≡ B (mod p) then An ≡ Bn (Mod p) for any positive integer n
And the property [A*(A-1mod p)](Mod p) = I is already shown in [5].
2.2 NTRU Encryption on Matrix
The NTRU Crypto-system [2] is based on three parameters p, q and N where p is a
small prime number and q and p are relatively-prime and N is the degree of the poly-
nomial in the ring of polynomials. Recently the NTRU cryptosystem using a ring of
polynomials has been extended [5] for a more compact matrix formalism using mod-
ular arithmetic.
Bob chooses two matrices X and Y, where matrixes X is an invertible matrix
(modulo p). He keeps the matrices X and Y private and generates a public key H as
follows:
H=p Xq*Y(modulo q)
Where Xq is X-1 modulo q or X * Xq=I (mod q) When Alice wants to send a message
to Bob, she converts the message to the form of binary matrix M (which is of the
same order as X and Y). She uses Bob’s public key and generates the cipher text E as
follows:
Reaction Attacks in the Matrix Scheme of NTRU Cryptosystem 29
E = H * R + M modulo q, where R is a compatible matrix and serves to obscure
the original message M. Bob after receiving the encrypted message uses the follow-
ing procedure to decrypt the message:
A= X * E (modulo q)
= X * (H * R + M) (modulo q)
= X*(p Xq * Y * R + M (modulo q)
= p Y * R + X*M
Let B = A (modulo p) = X* M modulo p
Now C = Xp*X*M modulo p = M, the original message.
3 Motivation
In the Reaction Attacks described in [2], one starts with a valid encrypted message e
and creates small modifications e’ = e + ε. The attacker makes the modification ε
larger and larger until the modified message e’ causes a decryption error. By compar-
ing the ε’s that cause decryption errors to those that do not, the attacker gains infor-
mation about either the plaintext message m underlying e or about the private key
used to encrypt m.
An encrypted NTRU message has the form e ≡ φh + m (mod q). The smallest mod-
ification that an attacker can make to e and still have it sometimes decrypt correctly is
,say,is obtained by adding npXi to e i.e., e’ = e + npXi for some 0 ≤ i < N and some n
≥ 1. This will cause a decryption failure (a so-called wrap or gap failure) if some coef-
ficient of the intermediate decryption polynomial a = pφg + mf is within np of q/2 and
if Xi f has a corresponding +1 coefficient. The important point to observe is that for
the correct choice of n, the i’s which cause decryption failure for e + npXi will reflect
(with some shifting and possible duplication) the i’s for which the private key f has a
term of the form +Xi. Thus the attacker potentially gains information about the +1 bits
in the private key f. Similarly, using negative values for n may give information about
the −1 bits of f.
The Reaction Attack does not compromise the hard mathematical problem underly-
ing the NTRU PKCS, which is the problem of finding the shortest vector in a lattice
of high dimension. None-the-less, it is a potentially serious attack for implementations
of the NTRU PKCS in hardware or software.
4 Proposed Method
An attacker would try to figure out the private key X, from the encrypted text. The
smallest modification that an attacker can make to the encrypted text and still decrypt
correctly gives some information about the private key. Let E’ = E + np*m(i, j) for
some 1≤ i ≤ N, 1 ≤ j ≤N and for some n, where m(i, j) is N X N matrix with 1 at (i,
j)th position and the rest zero. In this paper we assume that the attacker knows or the
descriptor reveals the value of A which is equal to X*E.
Let E be the valid encrypted message that decrypts correctly. So we have all the
elements of the matrix A satisfy aμ < q/2 and aν > -q/2, where aμ is the largest and aν
is the smallest element in the encrypted message.
30 R. Nayak, J. Pradhan, and C.V. Sastry
We have ,, ,0
, ,1
m[i,j] is a matrix where only one entry is 1 and the rest of the elements are 0’s. We
know that E[i,j] lies in (-q/2, q/2).
Now calculate A’ as A’= X * E’(mod q) = X*[E + p*m(i, j)](mod q)
= [X*E + X* p*m(i, j)] (mod q)
= [X*E (mod q) + X* np*m(i, j) (mod q) ] (mod q)
= [A + X* p*m(i, j) (mod q) ] (mod q) ------- (i)
We know that X is a matrix with entries (-1, 0, 1) and m[i, j] is a matrix with only one
entry with 1 and rest of the elements are 0’s. p is the NTRU parameter and n is the
smallest integer which causes wrap failure. q being very large compared to all other
parameters, it is safe to assume that X* np*m(i, j) (mod q) = X* np*m(i, j).
So we can write (i) as A’ = [A + X* np*m(i, j)] (mod q)
= A (mod q) + X* np*m(i, j) (mod q).
Now find A’ – A = X* np*m(i, j) (mod q). ---- --- (ii)
If we absorb np in m(i,j), we can write the above equation as
A’ - A = X*Y(i,j) where Y(i,j) = np m(i,j)
The above procedure can be repeated for different values of i and j in m(i, j) such that
∑(A’-A) = X * ∑Y(i,j) with Det( ∑Y(i,j) ) = 1.
Now multiplying Inverse of ∑Y(i,j) to ∑(A’-A) we get the required private key.
4.1 Example
Let X= {{1, 0, 0}, {1, -1, 1}, {0, 0, -1}}, M={{0, 0, -1}, {0, -1, 1}, {-1, 1, -1}}, and
A={{0, 3, 1}, {1, 1, 6}, {1, -1, 3}} . We know that p=3, take n= 3. Now send different
Table 1. Calculation of A’ and (A-A’) for a given mi(i, j)
mi(i, j) = A’= A’-A=
{{1, 0, 0}, {0, 0, 0}, {0, 0, 0}} {{-3, 3, 1}, {-2, 1, 6}, {4, -1, 3}}
{{-3,0,0},{-
3,0,0},{3,0,0}}
{{0,1, 0}, {0, 0, 0}, {0, 0, 0}} {{0, 0, 1}, {1, -2, 6}, {1, 2, 3}}
{{0,-3,0},{0,-
3,0},{0,3,0}}
{{0, 0,1}, {0, 0, 0}, {0, 0, 0}} {{0, 3, -2}, {1, 1, 3}, {1, -1, 6}}
{{0,0,-3},{0,0,-
3},{0,0,3}}
{{0, 0, 0}, {1, 0, 0}, {0, 0, 0}} {{0, 3, 1}, {4, 1, 6}, {1, -1, 3}} {{0,0,0},{3,0,0},{0,0,0}}
{{0, 0, 0}, {0,1, 0}, {0, 0, 0}} {{0, 3, 1}, {1, 4, 6}, {1, -1, 3}} {{0,0,0},{0,3,0},{0,0,0}}
{{0, 0, 0}, {0, 0, 1}, {0, 0, 0}} {{0, 3, 1}, {1, 1, 9}, {1, -1, 3}} {{0,0,0},{0,0,3},{0,0,0}}
{{0, 0, 0}, {0, 0, 0}, {1, 0, 0}} {{0, 3, 1}, {-2, 1, 6}, {-2, -1, 3}}
{{0,0,0},{-3,0,0},{-
3,0,0}}
{{0, 0, 0}, {0, 0, 0}, {0, 1, 0}} {{0, 3, 1}, {1, -2, 6}, {1, -4, 3}}
{{0,0,0},{0,-3,0},{0,-
3,0}}
{{0, 0, 0}, {0, 0, 0}, {0, 0, 1}} {{0, 3, 1}, {1, 1, 3}, {1, -1, 0}}
{{0,0,0},{0,0,-3},{0,0,-
3}}
Reaction Attacks in the Matrix Scheme of NTRU Cryptosystem 31
Table 2. Calculation of (A’-A)/np for a given mi(i, j)
(A’-A)/np=
{{-1,0,0},{-1,0,0},{1,0,0}}
{{0,-1,0},{0,-1,0},{0,1,0}}
{{0,0,-1},{0,0,-1},{0,0,1}}
{{0,0,0},{1,0,0},{0,0,0}}
{{0,0,0},{0,1,0},{0,0,0}}
{{0,0,0},{0,0,1},{0,0,0}}
{{0,0,0},{-1,0,0},{-1,0,0}}
{{0,0,0},{0,-1,0},{0,-1,0}}
{{0,0,0},{0,0,-1},{0,0,-1}}
mi(i, j) and assume that somehow the attacker knows the value of corresponding A’
which is shown in table 1.
In a 3 X 3 binary matrix, there are 84 different matrixes we can get whose determi-
nant is 1. Any matrix with determinant 1 ensures that in its inverse there is no frac-
tional part. This small procedure finds the number of different matrix whose determi-
nant is 1 and its corresponding inverse.
i=1;
For[y=1,y<=512,y++,
s=IntegerDigits[y, 2, 9]; // IntegerDigits[n, b, len] gives a list of the base-b digits in
// the integer n and pads the list on the left with zeros to give
// a list of length len.
m=Partition[s, 3]; // Partition[list, n] partitions list into non-overlapping
// sublists of length n.
If[Det[m]==1,Print[i++," ",m," ",Inverse[m]];]; ] //Det[f] gives the determinant.
Let us take another set of mi’s {{1, 0, 0}, {0, 0, 0}, {0, 0,0}} + {{0, 0,1}, {0, 0,
0}, {0, 0,0}} + {{0, 0, 0}, {1, 0, 0}, {0, 0,0}} + {{0, 0, 0}, {0,1, 0}, {0, 0,0}} + {{0,
0, 0}, {0, 0, 0}, {0, 0,1}} = {1,0,1},{1,1,0},{0,0,1}}. Whose inverse is {{1,0,-1},{-
1,1,1},{0,0,1}}
Add the corresponding (A’-A)/np. We get {{-1,0,0},{-1,0,0},{1,0,0}} + {{0,0,-
1},{0,0,-1},{0,0,1}}
+ {{0,0,0},{1,0,0},{0,0,0}} + {{0,0,0},{0,1,0},{0,0,0}} + {{0,0,0},{0,0,-
1},{0,0,-1}} = {{-1,0,-1},{0,1,-2},{1,0,0}}
Now Σ [(A-A’)/np ].(Inverse mi’s) = {{-1,0,-1},{0,1,-2},{1,0,0}} . {{1,0,-1},{-
1,1,1},{0,0,1}}
= {{-1,0,0},{-1,1,-1},{1,0,-1}}. This is the required X.
5 Conclusion
The method described in this paper to obtain the private key is possible if in adver-
tently the attacker gets hold of the intermediate text before final decryption.
32 R. Nayak, J. Pradhan, and C.V. Sastry
References
[1] Hall, C., Goldberg, I., Schneier, B.: Reaction attacks against several public-key cryptosys-
tems (April 1999) (preprint), http://www.counterpane.com
[2] NTRU Cryptosystem, Technical Reports, the free encyclopedia. NTRU Cryptosystems
Inc. (2002), http://www.ntru.comWikipedia
[3] Silverman, J.H.: Wraps, Gaps, and Lattice Constants NTRU Cryptosystems Technical
Report (March 15, 2001)
[4] Hoffstein, J., Silverman, J.H.: Reaction Attacks Against the NTRU Public Key Cryptosys-
tem, NTRU Cryptosystems Technical Report (June 2000)
[5] Nayak, R., Sastry, C.V., Pradhan, J.: A matrix formulation for NTRU cryptosystem. In:
Proceedings 16th IEEE International Conference on Networks (ICON 2008), New Delhi,
December 12-14 (2008)
[6] Hoffstein, J., Pipher, J., Silverman, J.H.: NTRU: A High Speed Public Key Cryptosystem.
Presented At He Hump Session of Euro. Crypt. 1996 (1996) (preprint)
[7] Hoffstein, J., Lieman, D., Silverman, J.: Polynomial Rings and Efficient Public Key Au-
thentication. In: Blum, M., Lee, C.H. (eds.) Proceeding of the International Workshop on
Cryptographic Techniques and E-Commerce (CrypTEC 1999). City University of Hong
Kong Press (1999)
[8] El-Gamal, T.: A public key cryptosystem and a signature scheme based on discrete loga-
rithms. IEEE Transactions on Information Theory 31, 469–472 (1985)
[9] Hoffstein, J., Pipher, J., Silverman, J.: NTRU: A Ring Based Public Key Cryptosystem.
In: Buhler, J.P. (ed.) ANTS 1998. LNCS, vol. 1423, pp. 267–288. Springer, Heidelberg
(1998)
[10] Wells Jr., A.L.: A polynomial form for logarithms modulo a prime. IEEE Transactions on
Information Theory 30, 845–846 (1984)
[11] Diffie, W., Hellman, M.E.: New directions in cryptography. IEEE Information theory
June 23- 25 (1975); IEEE International Symposium on Information Theory, Sweden, June
21-24 (1976)
[12] Brassard, G., Bratley, P.: Fundamentals of Algorithm, PHI (1996)
[13] Horowitz, E., Sahani, S., Rajasekharan, S.: Fundamental of Computer Algorithm, Galgo-
tia (1998)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 33–38, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Process Corner Analysis for
Folding and Interpolating ADC
Shruti Oza1 and N.M. Devashrayee2
1 EC Department, Kalol Institute of Technology & Research Centre, Nr. Highway,
Kalol-382721, Gujarat, India
swajan_2004@yahoo.com
2 EC Department, IT-Nirma University of Science & Technology, S.G. Road,
Ahmedabad Gujarat, India
nalin_deepika@yahoo.com
Abstract. Folding and Interpolating ADC have been shown to be an effective
means of digitization of high bandwidth signals at intermediate resolution. The
paper designs Folding and Interpolating ADC using cascaded folding amplifier
to observe the effect of process variations. The primary circuit effects, resulted
from process variations that are liable to degrade the performance of ADC are
transistor mismatch, resistor mismatch and amplifier-comparator offsets. The
device matching in reference generation, folding amplifier, interpolation and
comparator offsets specify overall performance of ADC. Since the mismatches
are random, Monte Carlo Analysis is used to estimate the linearity performance.
In this paper the design is simulated using 0.35 μm, 3.3V to study the effect of
process corners.
Keywords: Cascaded Folding Amplifier, Comparator, Encoder, Folding and In-
terpolating ADC, Interpolation, Process Corners.
1 Introduction
ADC is one of the most important building blocks to transform analog signals to digi-
tal signal process systems. For high-speed application, Flash ADC is widely used.
However, N-bit Flash ADC needs
2
N
- 1 comparators, which consume large power
and occupy area. Folding architecture is an alternative approach to reduce the com-
plexity of Flash ADC. Before the outputs of the preamps are fed into comparators,
folding amplifiers are inserted. Folding amplifier combines the outputs of several
preamplifiers and generates folding waveforms, which contains information of those
preamplifiers. After folding processing, one comparator deals with more quantization
levels. Hence, the number of the comparators is reduced. The number of comparators
required for a Folding ADC decreases as the folding order increases. The architecture
still keep high conversion rate [1-4].
Another attractive feature of Folding and Interpolating ADC is that high-speed
sample and hold amplifier is optional due to parallel operation of fine and coarse
converter. However, in most of the papers coarse converter is similar to Flash ADC.
34 S. Oza and N.M. Devashrayee
Liu et al proposed design of coarse converter using folding circuit similar to fine
converter [4]. Wenzek suggested cascaded dummy differential amplifiers with similar
structure like those in the folding stage of the fine converter to overcome the different
latencies of coarse and fine converter [5].
Systematic and random variations in process are posing a major challenge to the
future high performance VLSI design. Variation in the process parameters such as
impurity concentration densities, oxide thickness and diffusion depths caused by non-
uniform conditions during deposition and/or during diffusions of the impurities.
Variations in the dimensions of the devices are due to limited resolution of the photo-
lithographic process [6-8].
The main focus over here is to observe effect of process parameter variations at
device level on the performance of Folding and Interpolation ADC. Section-2
describes design of Folding and Interpolating design with details of implemented
folding amplifier, interpolation technique, comparator section, coarse converter and
encoder. Section-3 discusses simulation results through Monte Carlo Analysis with
wide range of randomly chosen device parameter.
2 Design of Folding and Interpolation ADC
Process variations are due to manufacturing phenomena and static in nature. Process
variations are deviations from intended or designed values for the structural or electri-
cal parameters of concern. The process variations can be lot-to-lot, wafer-to- wafer
(inter process), die-to-die or within a die (intra process). The paper focuses on Front-
End of Line (FEOL) variations, which refers to the variations at device level. The
major sources of FOEL variations consist of transistor gate length and gate width
variations, gate oxide thickness variations, doping-related variations, etc. The thresh-
old voltage can vary due to changes in oxide thickness, substrate, polysilicon and
implant impurity levels and surface charge [6-8]. These process parameter variations
result in variation of power and delay and non-linearity in ADC. Therefore it is very
important to observe effect of these parameters on the performance of ADC. The
effects of process variation can be observed through typical, slow and fast transistors.
The concept of Folding ADC was first introduced by Arbel and Kurz in 1975. The
main motivation was the dramatic reduction of the number of comparators required in
the design. In Folding and Interpolating ADC, number of zero-crossing points Z is
determined by the following equation:
IFNZ FF ⋅⋅= (1)
Where F
N is the number of primary folding waveform, F
F is the folding factor and
I is the interpolating rate. The choice of F
N, F
F and I play very important role in
design of Folding ADC.
The aim of the fine converter inside the Folding and Interpolating ADC is to proc-
ess the input signal, either a sampled or an unsampled signal, and to resolve it into the
lower significant bits of the word length of the whole analog-to-digital converter. The
output signals of the fine converter are then synchronized with output signals of the
coarse converter and combined to yield the output of the complete converter. The fine
Process Corner Analysis for Folding and Interpolating ADC 35
converter is composed of a folding stage, an interpolation stage, a comparator stage,
and an encoder as shown in Fig. 1(a).
Figure 1(a) shows block diagram of implemented 6-bit Folding and Interpolating
converter. Total 64 zero crossing points are obtained by selecting folding factor
F
F=8, Interpolating Factor I=2 and Number of Folding block F
N=4. The Fine
converter uses cascaded folding amplifier with folding factor F
F=8. The encoder
generates 4 fine bits while 2 coarse bits are generated using pre-processing folding
amplifier.
Fig. 1. (a) Folding and Interpolating ADC (b) zero crossing points
The function of a folder is to separate the input signal range into several intervals.
By gradually developing higher folding factor using cascaded folding amplifier helps
in achieving a high folding degree while avoiding the high gain and/or bandwidth
requirements of each folder. The first stage is implemented with folding factor=4. The
goal of the folding amplifier is to overcome all the limitations of existing design,
utilization of all transistor and minimizing power and settling time. Figure 2 (a) shows
implanted folding amplifier with folding factor=4 for the first stage. The second stage
is shown in Figure 2 (b). The differential output of first stage is two signals FA_P and
FA_N, applied as input for the second stage. Such two folding amplifiers’ output are
applied to the second stage to generate FF=8 as shown in figure. The total power
required by cascaded folding amplifier is only 180uW.
Fig. 2. (a) First stage of Cascaded Folding Amplifier with FF=4 (b) Second Stage
36 S. Oza and N.M. Devashrayee
Interpolation is often employed to generate extra folding waveforms without in-
creasing number of folding amplifiers. There are basically two methods to interpolate
the folded signals, namely voltage-mode (resistive) interpolation and current-mode
interpolation. The voltage-mode interpolation can be implemented using a resistance
ladder. The advantage of a voltage-mode interpolation over current mode is its design
simplicity and low power operation. Figure 1 (b) shows 64 zero crossing points gen-
erated by folding and interpolation stages, using F
N= 4, F
F=8 (first stage F
F=4,
second stage F
F=2) and interpolating factor=2.
In order to achieve low power, high-speed operation of the design, comparator is
another important block. When a comparator must drive a significant amount of out-
put capacitance in very short times, it is advisable to follow the latch by circuits that
can quickly generate large amount of current. A high-speed comparator following
these principles is designed in Fig. 3(a). The first stage is a low gain, high bandwidth
preamplifier that drives a latch (Decision Circuit). The latch outputs are used to drive
an inverter (Output Stage). Figure 3(b) shows output of fine comparator (cyclic ther-
mometer) with process variations.
Fig. 3. (a) Voltage Comparator (b) Cyclic Code (output) of Fine Converter
The coarse converter divides input into several intervals and provides MSBs. In-
stead of using traditional Flash architecture for implementing coarse converter, fold-
ing circuit is used as pre-processing. This helps in reducing number of comparators
and encoder logic.
The cyclic output (Fig. 3(b)) can be easily converted into binary code. To convert
cyclic code into binary, logic shown in equations (2) can be used, comparing binary
and cyclic code. The encoder based on XOR-OR logic can be used to convert the
cyclic code into binary. The simulation results of the encoder-digital output with
process variations are shown in Fig 5(a).
Encoder logic:
B3= C7
B2= C7 ⊕C3
B1= C7 ⊕C5 + C3 ⊕C1
B0= C7 ⊕C6 + C5 ⊕C4+ C3 ⊕C2+ C1⊕ C0… (2)
Where B0 to B3 are 4-bit fine converter output and C0 to C7 are 8-bit fine comparator
output.
Process Corner Analysis for Folding and Interpolating ADC 37
3 Monte Carlo Analysis
The folding architecture imposes tight constraints on device matching as mismatch
results in shifting zero crossing points from the ideal values. Mismatch in first stage
of folding creates large performance degradation. For accurate zero crossings, folding
and interpolating stages must be well matched, which depends on differential pair
mismatch, current source mismatch and resistance mismatch.
The mismatch within a folding amplifier includes mismatch in the threshold volt-
age and beta, mismatch between various tail current sources and mismatch of slope of
two folding signals. The input offset voltage depends on load resistor mismatch, tran-
sistor dimensions mismatch and threshold voltage mismatch In this paper,
V
TH
is
varied randomly with 6% variation and µ0 is varied randomly with 5% on each side
from their standard value. During simulations, Monte Carlo runs are performed 100
times to get accurate results. Figure 4(a) shows results for output of folding amplifier.
Among all the corners, the worst-case error = 0.22%.
Fig. 4. Process Corner Analysis (a) for Folding Amplifier (b) for Interpolation Stage
Fig. 4(b) shows the simulated results for interpolated signal due to variations in
process parameters on each side from their standard values. For 100 Monte Carlo
runs, worst-case error is 0.108% interpolation, which is half than worst-case error for
folded signal. This indicates that interpolating helps in reducing non-linearity through
averaging.
Figure 5 (b) shows switching power of converter at 100MHz clock frequency,
1MHz input. The range of average power variation is 10.05-10.6mW and peak power
is 19.88-24.9mW. The dc power required is in range of 10.05mW-17.9mW.
Fig. 5. (a) Digital Output of encoder (b) Switching Power due to Process Variations at Clock
Frequency of 100MHz, Input Frequency of 1MHz
38 S. Oza and N.M. Devashrayee
4 Conclusion
In this paper, 6-bit low power Folding and Interpolating ADC is designed with mini-
mum comparators and hardware. The architecture uses novel cascaded folding ampli-
fier to achieve high folding factor. The design uses folding pre-processing circuit for
both coarse and fine converter. To achieve low power operation, folding block, com-
parator, encoder, and coarse converters are optimized. The design is simulated using
0.35μm technology at 3.3V. The effect of process variations is observed using 100
Monte Carlo runs. The worst-case error in zero crossings after folding stage is 0.22%.
The error is reduced due to resistive interpolation and is 0.108%. The worst-case error
at final stage (MSBS and LSBs) is 2% in LSB. No error (error in codes, spikes or
pulse duration/delay variations) is observed for remaining bits of converter. Due to
cyclic output of folding block comparison, number of comparators required is only 10
in case of 6-bit converter. The design also helps in reducing latency difference of
coarse and fine converter.
References
1. Nauta, B., Venes, A.: A 70MS/s110-mW 8-b CMOS Folding and Interpolating A/D Con-
verter. IEEE Journal Of Solid-State Circuits 30(12), 1302–1308 (1995)
2. Thirugnanam, R., Ha, D.S., Choi, S.S.: Design of a 4-bit 1.4 GSamples/s Low Power Fold-
ing ADC for DS-CDMA UWB Transceivers. In: IEEE International Conference on Ultra-
Wide Band, ICU 2005, pp. 536–541 (2005)
3. Kim, K.M., Yoon, K.S.: An 8-Bit CMOS Current-Mode Folding And Interpolation A/D
Converter With Three-Level Folding Amplifiers. In: IEEE 39th Midwest Symposium on
Circuits and Systems, vol. 1, pp. 201–204 (1996)
4. Liu, Z., Wang, Y., Jia, S., Ji, L., Zhang, X.: Low-Power CMOS Folding and Interpolating
ADC with a Fully-folding Technique. In: 7th International Conference on ASIC, ASICON
2007, pp. 265–268 (2007)
5. Lin, I.K.-L., Kemna, A., Hosticka, B.J.: Modular Low-Power, High-Speed CMOS Analog-
To-Digital Converter of Embedded Systems. Kluwer Academic Publishers, Dordrecht
(2003)
6. Bowman, K., et al.: Impact of die-to-die and within-die parameter fluctuations on the
maximum clock frequency distribution for gigascale integration. IEEE Journal of Solid-
State Circuits 37, 183–190 (2002)
7. Borkar, S.: Parameter Variations and Impact on Circuits & Microarchitecture. C2S2
MARCO review (2003)
8. http://www.britannica.com/bps/../18/../Chapter-2-Process-
Variations
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 39–44, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Fast Near-Lossless Image Compression with Tree
Coding Having Predictable Output Compression Size
Soumik Banerjee and Debashish Chakroborty
Department of Computer Sc. & Engineering.
St. Thomas’ College of Engg. & Tech., Kolkata, India
{soumik.stcet,sunnydeba}@gmail.com
Abstract. Image compression, in the present context of heavy network traffic,
is going through major research and development. Traditional entropy coding
techniques, for their high computational cost is becoming inappropriate. In this
paper a novel near-lossless image compression algorithm had been proposed
which follows simple tree encoding and prediction method for image encoding.
The prediction technique uses a simple summation process to retrieve image
data from residual samples. The algorithm had been tested on several gray-scale
standard test images, both continuous and discreet tone, and had produced
compression comparable to other state-of-the-art compression algorithms. The
output compressed file sizes had shown that they are independent of image data,
and depends only on the resolution of the image, an unique property that can
exploited for networking bandwidth utilization.
Keywords: Modelling – coding architecture, median filtering, residual samples,
tree encoding, progressive transmission.
1 Introduction
Image compression researches are conducted to optimize storage space requirements
for efficiency in storage and data-transfer over networks [1]. The lossless concept of
image compression [2] although produces errorless output, cannot provide excellent
compression. The lossy concept exploits the fact that human cognition is unable to
detect very small intensity variation in small areas of an image. Although this concept
provides excellent compression but sometimes loss of image information becomes
unacceptable, for instance in medical image processing or image archival where
machine based image processing is involved.
Recently, the research interest in this field had been diverted to development of
near-lossless compression methods. Algorithms like LOCO-1[5], JPEG-LS[3],
JBIG[4] and FELICS[7] have been developed on this concept. All these algorithms
follow the modeling-coding architecture.
Firstly, image data was modeled using prediction algorithm like DCT[8] or
DWT[9] and the deviation of predicted value from actual, called residual sample, was
stored. Then the residual samples were encoded using entropy encoding techniques
like Adaptive Huffman, Hierarchical Interpolation or Tree encoding[6]. These
40 S. Banerjee and D. Chakroborty
approaches involved intensive computation due to the statistical data manipulation
and encoding techniques employed by them. Each of the standards provides a faster
version, which results in a trade-off in compression quality.
In this paper a new compression technique for gray-scale images had been
proposed. The technique involves three basic steps: 1) image smoothening 2) tree
encoding of the image information and 3) representing intensity of tree using ASCII
character set for optimized storage requirement. The computational cost was low
since only linear search and subtraction operation was involved in image modeling
part. The simulated results on standard test images had given comparable or better
compression ratio than the standard algorithms.
The remaining paper had been organized as follows: section 2 discusses the
detailed algorithm, section 3 presents the simulation results and in section 4, the paper
is concluded.
2 Proposed Algorithm
2.1 Image Compression
The proposed algorithm complied with the modeling–coding architecture. The image
had been smoothened to emphasize intensity correlation (section 2.1.1). The image
modeling part had been done using residual calculation from maximum intensity in
blocks of the image (section 2.1.2). The encoding of thus produced residual samples
had been done using tree encoding (section 2.1.2). Consequently, data representation
completed the process (section 2.1.3).
2.1.1 Image Smoothening
The image had been smoothened using median filtering on 2X2 masks throughout the
image. This step ensured that the truncation of the image tree did not result in drastic
loss of image information since median filtering emphasized on statistical correlation
among pixels.
2.1.2 Modeling and Progressive Encoding of Image Data
The smoothened image had been encoded in a tree architecture using 4 children (one
block) at each level for next level node determination. In the proposed method,
consecutive blocks are selected and according to the following steps the image was
encoded:-
1. Calculate the maximum intensity value in the selected block.
2. Subtract all other intensities in the block from this value.
3. The last intensity will be involved in further tree encoding at the next level,
hence it is not stored at the current loop of execution.
4. The remaining three are stored using character representation (section 2.1.3).
The conceptual idea behind the encoding technique is explained in figure1.
Fast Near-Lossless Image Compression 41
Fig. 1.
As shown in the figure, at each level the resolution of the image gets halved. The
resultant tree (expanded to 2 levels) for a 256X256 image is shown in fig 2.
Fig. 2.
2.1.3 Residual Sample Coding
The maximum intensity, in each block, had been selected for quantization (refer to
2.1.2). Hence, residual samples were in the range of [0 – 255] since only 8 bit images
have considered. 8 bit ASCII character set defines 255 character symbol set. Of these,
three symbols were found unfit for the purpose of intensity representation. ASCII
symbols, corresponding to each intensity value, had been used in the sample
encoding. The last three intensities had been assigned 2-character hybrid symbols.
Also the subtraction for the local maximum intensity ensures that no negative value
appears, which otherwise would have to be represented using symbols thereby
reducing the resultant compression ratio.
Alternate Strategies for Achieving Greater Compression
The image tree formed can be truncated at several levels. Default strategy was to
truncate the tree at level 0 because of high correlation of intensity values in 2X2
256,256
128,256 128,128 256,128
64,128 64,64 128,64
42 S. Banerjee and D. Chakroborty
masking window. This concept was extended to level 1 tree truncation, which resulted
in a smoothened image with 4X4 masking. The second approach, although notably
increased the compression ratio, resulted in comparatively higher error of deviation
(refer section 3, table 1 and 2).
2.2 Image Decompression
The decompression of the compressed file was done following steps discussed below:
1. Generated the symbol table as discussed in section 2.1.3.
2. Read symbols from compressed file and found the corresponding intensity
values. Hence the image tree was remade.
3. Found the maximum in a parent-children block and subtract other values from
the maximum value to regenerate original image matrix.
4. Use the last level leaf node values to regenerate the entire image matrix, which
is the median filtered (2X2 masking) output of the original image.
3 Simulation Results
Some definitions that were used to study and compare results with other techniques:
Bits/Pixel (BPP) = (8 X Compressed File Size)/(Actual File Size) (1)
RMSE
∑∑,
,
(2)
The algorithm had been applied on standard gray-scale test images, both
continuous-tone and discreet-tone types. The compression percentage achieved had
varied from 96-98% for different test images. The test images were shown in figure 3.
Fig. 3. The test image set: (from left to right) lena(512X512); mandrill (256X256); discreet
(64X64); peppers(256X256); nasa(512X512); jet (128X128)
The error in output generated had been shown in table 2. The scale of reference
was Root Mean Square Error (RMSE), as defined above. The tabulation was done for
RMSE values acquired for level 0 and level 1 image tree truncation (refer to fig1).
The experimental results suggested that the output size was predictable. For a
512X512 8bpp image, the output size was 48KB exact in each instance for level 0
truncation of image tree. For level1 truncation the size was 12.0 KB, although this
Fast Near-Lossless Image Compression 43
Table 1. The comparative study of detailed results of the compression output in BPP of the
proposed algorithm, with level 0(L0) and level 1(L1) truncation with JPEG 2000
Image name Lena Mandrill Peppers Nasa Discreet Jet
JPEG 2000 0.46 0.68 0.41 0.28
0.56 0.12
PROPOSED
(L0) 0.39 0.42 0.39 0.24 0.48 0.38
PROPOSED
(L1) 0.09 0.13 0.10 0.11 0.05 0.09
Table 2. The RMSE result for level 0(L0) and level 1(L1) truncation of the original image tree
Image Lena Mandrill Peppers discreet Nasa Jet
RMSE (L 0) 15.35 12.33 17.36 25.09 5.54 6.46
RMSE (L 1) 20.24 17.35 21.54 29.09 7.82 10.98
increased the error level, as discussed in table 2. The predictability of the output
image size could be used as a great advantage for network utilization by helping in
pre-planning resource allocation.
4 Conclusion
In this paper a new approach of compressing image had been presented using the
modelling-coding architecture. The computation time was comparatively much lower
than the standard algorithms, which involved complex algorithm (DCT approach in
JPEG, DPCM in CALIC) for image modelling. The results showed encouraging
compression ratio and information loss was also within moderate range even for level 1
truncation strategy of the proposed method. The algorithm ensured fast encoding,
predictive output result and simple progressive transmission based decoding which
made it suitable for network data-transfer or fabrication on mobile devices.
References
[1] Salomon, D.: Data Compression The Complete Reference, 3rd edn. Springer Press,
Heidelberg
[2] Halder, A., Chakroborty, D.: An Efficient Lossless Image Compression Using Special
Character Replacement. In: ICCET 2010, Jodhpur, Rajasthan, India, November 13-14,
pp. E-62 – E-67 (2010)
44 S. Banerjee and D. Chakroborty
[3] Acharya, T., Tsai, P.S.: JPEG 2000 standard for image compression (2000)
[4] Wu, X.: Context-Based, Adaptive, Lossless Image Coding. IEEE Trans. Comm. 45(4)
[5] Weinberger, M., Seroussi, G., Sapiro, G.: The LOCO-I Lossless Image Compression
Algorithm
[6] Cai, H., Li, J.: Lossless Image Compression with Tree Coding of Magnitude Levels, 0-
7803-9332-5/05 ©2005 IEEE
[7] Howard, P.G., Vitter, J.S.: Fast and Efficient Lossless Image Compression. In: IEEE DCC
1993 (1993)
[8] Watson, A.B.: Image Compression Using the Discrete Cosine Transform, NASA Ames
Research Centre. Mathematica Journal 4(1), 81–88 (1994)
[9] Ansari, M.A., Anand, R.S.: DWT based Context Modelling of Medical Image
Compression. In: XXXII National Systems Conference, NSC 2008, December 17-19
(2008)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 45–54, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Over Load Detection and Admission Control Policy in
DRTDBS
Nuparam1 and Udai Shanker2
1 Computer Science and Engineering Department, FGIET,
Raebareli, 229001, U.P. India
2 Computer Science and Engineering Department, MMMEC,
Gorakhpur, 273010, U.P. India
{nrcua80,udaigkp}@gmail.com
Abstract. Today’s Real Time System (RTS) is characterized by managing large
volume of distributed data making real time distributed data processing a reality
[1, 2]. The demand for real time data services is increasing in many large scale
distributed real time applications. The transaction work load in DRTDBS may
not be balanced and the transaction access pattern may be time varying and
skewed. Hence, computation workload on large scale distributed system can
lead to large number of transaction deadline misses [3, 4]. Hence, efficient da-
tabase management algorithm and protocol for accessing and maintaining data
are required to satisfy timing constraints of transaction supported applications.
In this paper, an algorithm has been proposed for admission control in
DRTDBS consisting of local controller and global load balancer working at
each site, which decide whether to admit or reject the newly arrived transaction.
The simulation results show that the new algorithm successfully balances the
workload in DRTDBS [5, 6].
Keywords: Admission control, local controller, global load balancer, distrib-
uted data processing, computation workload.
1 Introduction
In recent year, we have seen the emergence of large scale distributed real time data-
base system, which embedded in advance traffic control, factory automation, global
environment control and nation-wide electrical power grid control. DRTDBS can re-
lieve the difficulty of developing data intensive real time applications by supporting
the logical and temporal consistence of interrelated database distributed over a com-
puter network via transaction management. They support transaction that has explicit
timing constraints which is expressed in the form of dead line. A transaction is con-
sidered to have finished executing if exactly one of two things occurs: either its pri-
mary task is completed (success fully completed) or its recovery block is completed
(safe termination). Committed transaction brings a profit to the system whereas a ter-
minated transaction brings no profit. The goal of over load detection and admission
control policy employed in the system is to maximize the profit [7, 8, 9]. The presence
46 Nuparam and U. Shanker
of multiple sites in the distributed environment raises issue that are not present in cen-
tralized system. In large scale distributed environment it is a challenging to provide
data services with guarantees, while still meeting temporal requirement of transaction.
One main difficulty lies in long and highly variable remote data access delays. Large
scale distributed real time database system utilizing wide geographical area have to
use a network that share by many participant for cost effectiveness. Second major
challenge involves the complex interaction among a large number of nodes, which can
incur unpredictable work load for each node. Transaction work load fluctuation causes
uneven distribution of workload among the sites even if on the average, all sites re-
ceive similar amount of workload. The third challenge is the data dependent nature of
transaction. End to end transaction access pattern may be time varying and skewed.
This can be achieved only by timely access to remote data and timely processed of
centralized data.
The system architecture consists of overload detection and admission control of
scheduling transaction which provide early notification of failure to submitted transac-
tion that are deemed not valuable or in capable of completing in time, when transac-
tion is submitted to the system. An admission control policy is employed to decide
whether to admit or reject that transaction. Once admitted, a transaction is guaranteed
to finish executing before its dead line. When any site is detecting overload then it
distributes the load to other site. In DRTDBS there are two type of transaction, global
and local. The global transactions are distributed real time transaction executed at
more than one site whereas the local transaction executes at generation site only.
2 Distributed Real Time Database Model
2.1 Real Time Database Model
We focus our study on medium scale distributed database since the load balancer need
full information from every site to make accurate decision. Several applications that
required distributed real time data services fall in that range for example a ship board
control system which control navigation and surveillance consist of 6 distributed con-
trol unit and 2 general control consoles located throughout the platform and linked
together via a ship wide redundant Ethernet to share distributed real time data and
coordinate the activity [10].
2.2 Distributed Real Time Database Model
The performance of the system is evaluated by developing two simulation models for
DRTDBS. The first one is for main memory resident DRTDBS which eliminate the
impact of different disk scheduling algorithm on the performance. Since main memo-
ry database system, we have also developed another model followed by description of
various components such as system model, network model, cohort execution model,
database model [11]. In our system model, data object are divided into two type
namely temporal data and non temporal data. Temporal data are the sensor data from
physical world. Each temporal data object has a validity interval and is updated by
periodic sensor update transaction. Non –Temporal data object do not have validity
intervals and therefore there are no periodic system updates with them.
Over Load Detection and Admission Control Policy in DRTDBS 47
3 A New Admission Control Policy Architecture
A database system can be overloaded if many user transactions are executed concur-
rently. As a result computational resources such as CPU cycle and memory spaces
can be exhausted. More ever many transactions can be blocked or aborted and res-
tarted due to data contention also. Fig. 1 shows the architecture of a new admission
control policy in DRTDBS. The architecture has 4 layer, remote data access layer,
QoS enforcement layer, real time database management layer and DRTDBS layer. In
DRTDBS layer does exact work as to admitted or reject the transaction All the trans-
action are submitted in arrival queue and overload admission controller check the
incoming transaction as the system resources is already hold by other transaction.
Fig. 1. The System Model for New ACP Protocol
Overload Resolver send current transaction to the ORA (Overload Resolver Array)
table. After getting the response from the cohort the overload admission controller
interact with the transaction handler (TH) and rejected queue.
The real time database layer does typical real time transaction handling; the incom-
ing transaction are dispatched and processed by transaction handler. The transaction
handler consists of a concurrency controller (CC) a freshness manager (FM) and a
scheduler (SC). In the SC, update transaction are scheduled in the low priority queue.
Update transaction are either updates from local sensors to local data objects. Within
each queue, transactions are scheduled with Earliest Deadline First (EDF)[12, 13].
The third layer is guaranteed the desired miss ratio even in the presence of unpre-
dictable workload; QoS enforcement layer exploits two feedback control loops. It has
local controller (LC) which operates on the local transaction and global controller
(GL) which operate on the global transaction.
48 Nuparam and U. Shanker
The remote data access layer enables transparent access to remote data within a
bounded communication time. Remote temporal data are replicated locally to provide
timely access to them.
4 Algorithms for Centralized Admission Control in DRTDBS
In each node, there are a local miss ratio controller and local utilization controller.
The local miss ratio controller takes the miss ratio from latest sampling period, com-
pare them with serial execution time and compute the local miss ratio control signal
δLmisr used to adjust the target utilization at the next sampling period. The equation
used to derive δLmisr is as follows
δLmisr (Mri-Mrs) + (Ltmri-Mrsi) -- ---------------- (1)
Mri is the miss ratio of class i transaction of last period and Ltmri is the long term aver-
age miss ratio of class i transaction. Mrsi is the specified miss ratio requirement by the
serial execution specification. And n is the specified serial execution level.
and are two controller parameter. In order to prevent under-utilization a
utilization feedback loop is added. At each sampling period, the local utilization con-
troller compares the utilization and generates the local utilization control signal δLutil
using equation.
δLutil Iutil × (Lutil - Lutilpset) + Jutil × (LTutil - Lutilpset) (2)
Lutil is the CPU utilization of last sampling period and LTutil is the long term average
CPU utilization of the system. Lutilpset is the preset CPU utilization threshold. Iutil & Jutil
are controller parameter.
4.1 Load Balancing Factor
The load sharing process is guided by the load balancing factor (LBF). The LBF at
each node is an array of real number which denotes the amount of workload the local
node transfer to other node during the next sampling period.
Tremain work (a) Ttrsnsfer (a,i,j) (3)
The left expression stands for remaining executing (predicted) time of transaction ‘a
in the original node and its value is the difference between the transaction ‘average
executing time and executing time of transaction ‘a’. The value of right expression
stands for the time cost of transferring from node i to node j of transaction ‘a’ which
is determined by:
Ttransfer (a, i , j) = Tcode (a, i, j) +Tdata (a, i, j) (4)
There into, Tcode (a, i, j) = sizeof_code(a)/R, Tdata (a, i, j) = sizeof_data(a)/R.
The value of sizeof_code(a) stands for the total size of executing environment pa-
rameter and log of transaction ‘a’. The value of sizeof_data(a) stands for the size of
fetched data and immediate results by transaction ‘a’ in the current database server
node i. R stands for the average network transferring rate.
Over Load Detection and Admission Control Policy in DRTDBS 49
5 Algorithm for Global Load Balancing in Decentralized
DRTDBS
5.1 Work Load Transfer Test
The first step is to test whether there exist load transfer between nodes. To do that we
calculate the mean deviation of Mri from different node
Mean Deviation Mri) - Mean (Mri) -------------- (5)
Where Mri is the Miss ratio of node k. ABS(Mri) returns the absolute value of Mri and
Mean(Mri) returns the mean of Mri., n is the nodes in the system. The mean deviation
of Mri is a measure for workload balance in the system.
5.2 LBF Adjustment
The LBF adjustment is divided into two cases, depending on whether there is a load
transfer among the node.
Load Imbalance -: When there is load transfer in the system i.e the mean deviation
of Mri is larger than the threshold, it is necessary to share the load between nodes. The
load balancing algorithm at the overloaded nodes will shift some workload to the less
loaded nodes. A node i is considered to be overloaded compared to other nodes if and
only if the difference between its MRI and MRI mean is larger than the present mean
deviation threshold. i.e. as follow
a. When the difference of MRI and MRI Mean (T1) ≥Mean Deviation Threshold (T2)
b. When the difference of MRI and MRI Mean < 0
c. 0≤When the difference of MRI and MRI Mean(T1)≤Mean Deviation Threshold
(T2)
5.3 Algorithm Description
In order to realize the algorithm we introduce three data structures to record
the node’s states namely Rqueue, Squeue. and Oqueue. and several other variables such
as the maximum value of probing time to avoid too probing effect to the system
performance.
5.3.1 Knocking State
A node i is considered overloaded if the difference of MRI and MRI Mean ≥ Mean
Deviation Threshold T2. It divides into two parts: sender side and receiver side. The
sender side execute the following operation when a new transaction is created in node
i namely Ti (m+1), which cause the node i to be sender.
I. Then select a node j from LBF Rqueue which satisfies:
Ttransfer(a,i,j) = Min (Ttransfer(a,i,j)),
Therein j=1,2,3,4…………n. n is the recorded receiver number in structure LBF
Rqueue. It is necessary to judge if the node can be receive in terms of the probing
resut.
50 Nuparam and U. Shanker
II. IF ( node j is the receiver)
Puts the transaction Ti(m+1) into the node j LFB queue Tj and transfer the log,
then the algorithm terminate
ELSE
Remove the node I from LBF Rqueue and put it into the queue of LBF head
of sender queue Squeue
III. Select another node from LBF to repeat the process until one of the following
condition satisfied. If then terminate the algorithm
III.I Rqueue is empty
III.II Probing time is beyond the maximum time
III.III Node I is no longer a sender
If receiving probe information then the receiver’s side executes the following
operation.
IV. Remove the node from the current LBF queue to the head of the Squeue
Send a message about node j’s states to node i.
5.3.2 Responding State
A node i is considered less overloaded if the difference of MRI and MRI Mean < 0. It
also divides into two parts: receiver’s side and sender’s side, the first part executes the
following operation if the node i becomes a receiver when a transaction, Tik, finished
or be removed.
I. The node i sends the probing message to node j which is in the head of Squeue .
And judge if node j is a sender.
II. IF (node j is a sender)
IF (Exists a transaction that can be transferred to node j)
Transfer the transaction Tjr which resides in node j before to the transaction
LBF queue of node i then executes the Tjr from the beginning in the node i.
The log will be sent i at the same time. After that, the algorithm terminates.
ELSE ( Remove the node j to the tail of Squeue)
ELSE
Removes the node j from Squeue and put into the queue’s head of Rqueue or
Oqueue Select another new node to repeat I and II until one of the following
conditions satisfied, if then, the algorithm terminates.
a. Squeue is empty
b. The probing time beyond the limit
c. Node i is no longer a receiver.
When a node j receives the probe information from node i, then sender’s side execute
the following action:
III IF (node j is a sender)
Node j sends a notice message about the node j’s states of node i, and
evaluate the value of Ttransfer(a,i,j) by the equation 3 & 4 which is the
transferring cost of the transaction. And judge if it is worthy of the
transferred.
IF (exists a transaction is worthy of the transferred)
Node j sends a message to indicate there is an available transaction to be
transfers the most cost-efficient transaction and corresponding logs. After
that the algorithm terminates.
Over Load Detection and Admission Control Policy in DRTDBS 51
ELSE
Send a message indicating there is no available transaction can be
transferred to node i.
ELSE
Remove the node i from the current queue to the queue’s head of Rqueue, and
send the current states of node j and i.
5.3.3 Balance State
A node i is considered balanced if the Mean Deviation of MRI is less than the speci-
fied threshold, the LBF will reduce the load transferring factors.
LBFqueue(i,j) = LBFqueue(i,j) ×µ (6)
Where 0 < µ < 1. µ is called the LBF Regression Factor which regulates the load
transferring factors regression process. After reducing LBF, if a LBF becomes suffi-
ciently small(less than 0.005), it is reset to 0.
6 Performance Evaluation
In this section, we show the value of admission control policy by overload detection
comparing the performance achievable through workload admission control policy.
The transaction inter-arrival rate, which is drawn from an exponential distribution, is
varied from 50 transactions per second up to 300 transactions per second in incre-
ments of 50, which represents light-to-medium loaded system. Each simulation was
run three times, each time with a different seed, for 20000 ms. the results depicted is
the average over the three runs. The settings for the user transaction workload are
given in table 1. A user transaction consists of operations on both local data object
and global data object to 1000 microseconds. At each node, the transaction workload
consists of many periodic transactions and the average arrival rate and throughput
shown in the simulation result.
Table 1. Baseline Workload Parameter
Parameter Meaning Value
CPUTime CPU time per page access 2.5 ms
DBsize Database size in pages 1,000
ArrivalRate Transaction arrival rate 5-100 TPS
CTComp Time Mean Compensating Task Time 10 ms
CTStdDev St. Dev. of CT Time 0.5 T CompTime
SlackFactor Slack Factor 2
RegFactor Regression Factor 0 to .005
TaskSchd Task scheduling protocol EDF
CTSchd CT scheduling protocol FF, LF, LMF
Thrsh CT computation Threshold 0.125
CCntrl Concurrency Control Protocol OCC-BC
52 Nuparam and U. Shanker
Fig. 2. Mean Deviation throughput Fig. 3. Global Load Balancer with Throughput
As shown in Fig. 2 the system running best effort algorithm keeps unbalanced
throughout the workload burst periods; with admission control policy the system
workload become balanced (mean deviation of MRI becomes less than 0:1) within 5
seconds. The miss ratios at overloaded nodes are shown in Fig. 3 As we can see, for
the best-effort algorithm, QoS requirements are violated and the miss ratio of class i
transactions remains over 90%.
7 Related Work
This work differs from previous research in that our transaction model incorporates
not only primary tasks, with unknown WCET, but also compensating tasks. The new
admission control mechanism used admits transaction into the system with the abso-
lute guarantee that either the primary task will successfully commit or the compensat-
ing task safely terminate. Distributed Real Time Database System (DRTDBS) have
drawn research attention in recent years [14, 15]. Instead of providing strong logical
consistency, DRTDBS focus on data freshness and timeliness of transaction. However
most previous DRTDBS work targeted small scale system, but we extended it to large
scale system in wide area network environment.
A New Admission Control Policy (NACP) and feedback mechanism could employ
in variety of DRTDBS component: Transaction scheduling [15], Memory Allocation for
Query Management [16], Concurrency Admission Control Management in ACCORD
[8], and an Efficient Call Admission Control for Hard Real Time Communication in
Differentiated Service Network [17].The main idea of the admission control policy is to
associate an local update to each submitted transaction in order to favors, when the sys-
tem is overloaded, the executions of the most important transactions according to the
application-transactions set. In overload conditions, each sub-transaction of the global
Over Load Detection and Admission Control Policy in DRTDBS 53
transaction is executed on a site that has the lowest workload among those sites that
have executed the data items needed by the sub transaction.
8 Conclusion
Most previous DRTDBS studies have assumed that the only possible outcome of a
transaction execution is either the commitment or the abortion of transaction. In many
systems a third outcome of an outright rejection may be desirable. A process control
application the outright rejection of a transaction may safer than attempting to execute
that transaction only to miss its deadline. Out system allows the system to reject a
transaction because the admitted transaction holds some resources, thus making it
possible utilization by other transaction, to be taken in timely fashion. Also this flex-
ibility allows the system to relate its resources in the most profitable way, by only
admitting high value transaction when the system is overloaded while being less
choosy when the system is under loaded.
Our current research efforts focus on evaluating the performance of pessimistic as
well as speculative CACM techniques. Moreover, our work to date has concentrated
on uniprocessor systems. We are currently investigating the extension of our admis-
sion control and scheduling protocols to multiprocessor systems. A number of chal-
lenging questions arise. How are transactions, both their primary tasks and compen-
sating tasks allocated to processors? What type of CPU scheduling discipline should
be used? How valuable is the use of the WACM in a multiprocessor system? How the
concurrency will maintained in CACM?.
References
[1] Aldarmi, S.A.: Real Time Database System, Concept and design. Department of computer
science, University of York (April 1998)
[2] Kim, Y., Son, S.: Supporting predictability in real time database system. In: Proc. 2nd
IEEE Real Time Technology and Application Symposium (RTAS 1996), Boston,
pp. 38–48 (1996)
[3] Abbott, R., Garcia-Molina, H.: Scheduling real time transaction: A performance evalua-
tion. In: Proceeding of the 14th International Conference on very large Data Bases, Los
Angeles, CA, pp. 1–12 (1988)
[4] Kang, W., Son, S.H., Stankovic, J.A.: Managing deadline miss ratio and sensor data
freshness in real time databases. IEEE Transactions on Knowledge and Data Engineering
(October 2004)
[5] Stankovic, J.A., He, T., Abdelzaher, T., Marley, M., Tao, G., Son, S., Lu, C.: Feedback
Control Scheduling in Distributed Real Time Systems Symposium (RTSS 2001), Wash-
ington, DC, USA, p. 59 (2001)
[6] Lam, K.W., Lee, V.C.S., Hung, S.L.: Transaction scheduling in distributed real time sys-
tem. Int. J. Time – Crit. Comput. Syst. 19, 169–193 (2000)
[7] Lee, V.C.S., Lam, K.-W., Hung, S.L.: Concurrency control for mixed transactions in real-
time data-bases. IEEE Trans. Comput. 51(7), 821–834 (2002)
[8] Nagy, S., Bestavros, A.: Concurrency Admission Control Management in ACCORD.
Ph.D Thesis at Bostons Univesity (1997)
54 Nuparam and U. Shanker
[9] Bestavros, A., Nagy, S.: Value-congnizant admission control for rtdb systems. In: RTSS
1996 the 17th Real Time System Symposium, Washington DC (December 1996)
[10] Saab Systems Pty Ltd. “Ship Control System” in Saab System Website,
http://www.saabsystems.com.au
[11] Shanker, U.: Some performance issues in Distributed Real Time Database Systems. PhD
Thesis, Department of Electronics & Computer Engineering, IIT Roorkee (December
2005)
[12] Chetto, H., Chetto, M.: Some results of the earliest deadline scheduling algorithm. IEEE
Transaction on Software Engineering (October 1989)
[13] Liu, C.L., Layland, J.: Scheduling algorithms for multiprogramming in hard real time en-
vironments. Journal of the Association of Computing Machinery (January 1973)
[14] Kang, W., Son, S.H., Stonkovic, J.A., Amirijo, M.: I/O aware deadline miss ratio man-
agement in real time embedded database. In: 28th IEEE Real Time System Symposium
(RTSS) (December 2007)
[15] Lee, V.C.S., Lam, K.W., Hang, S.L.: Transaction Scheduling in Distributed Real Time
System. Intr. J. Time Crit. Computer System (2000)
[16] Pang, H., Carey, M.J., Livny, M.: Managing memory for real time queries. In: Proceed-
ings of the 1994 ACM SIGMOD Conference on Management of Data, pp. 221–232
(1994)
[17] Baronia, P., Sahoo, A.: An efficient Call Admission Control for Hard Real Time Commu-
nication in Differentiated Services Network. Proc. IEEE (2003)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 55–60, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Wavelet Transform Based Image Registration
and Image Fusion
Manjusha Deshmukh1 and Sonal Gahankari2
1 Asst. Prof., Electronics and Telecommunication, Saraswati college of Engg.,
Navi Mumbai, India
manju0810@yahoo.com
2 Lecturer, Electronics and Telecommunication, Saraswati college of Engg.,
Navi Mumbai, India
sonalgahankari@rediffmail.com
Abstract. Image Registration is a fundamental task in image processing used to
match two or more pictures taken, for example, at different times, from differ-
ent sensors, or from different viewpoints. Image registration is particularly
difficult when images are obtained through different sensor, (multi-modal regis-
tration). Mutual Information can be used for multimodal image registration. But
this method has its own limitations one of it is of speed, method is very slow
hence when time is an important constraint one cannot use this method. In this
paper an attempt has been made to overcome this limitation.
Keywords: Image Registration, Image Fusion, Wavelet Transform, Mutual In-
formation.
1 Introduction
Image registration is establishment of correspondence between images of the same
scene. Many image processing applications like remote sensing for change detection,
estimation of wind speed and direction for weather forecasting, fusion of medical
images like PET-MRI, CT-PET etc need image registration.
A comprehensive survey of image registration methods is presented by Brown,
Barbara Zitova and Jan Flusser [1, 2, and 3]. Subunku proposed various entropy based
algorithms for multimodal image registration [4]. S. Chaudhari and U. Bhosale pro-
posed methods for multispectral panoramic mosaicing and fast method for image
mosaicing using geometric hashing [5,6]. Flusser classified feature based methods
using special relations, methods using invariant descriptors, relaxation method and
pyramid and wavelets [7, 8].
Image registration is classified into two types that are registration of images that
are in same spectral band and registration of images that are in different spectral band.
T. Sao proposed mutual information based method for image registration [9]. Cahill,
N D Williams, C.M. Shoupu propose an approach to incorporate spatial information
into the estimate of entropy to improve multimodal image registration [10]. A. survey
56 M. Deshmukh and S. Gahankari
of medical image registration based on mutual information is presented by J. P. W.
Pluim [11]. Xiaoxiang Wang and Jie Tian in their paper proposed a mutual informa-
tion based registration method using gradient information [12]. Frederik Maes and
Andre Collignon apply mutual information to measure the statistical dependence
between the image intensities of corresponding voxels in both images [13].
2 Wavelet Domain Image Registration
Registration of image using wavelets is being discussed in this section.
Using discrete wavelet transform (DWT), a function f(t) can be represented by
(1)
Where aj,k are wavelet coefficients , Ψj,k(t) are basis function , j is scale , k is transla-
tion of mother wavelet Ψ(t). Two dimensional DWT can be obtaine by applying DWT
across rows and columns of an image.The two dimensional DWT of image f(x,y)is
(2)
Where C J0 is approximation coefficient, φj,k,l (x,y) is scaling function, DjS is set of
detail coefficients and Ψ S
j ,k, l is set of wavelet function.
The DWT coefficients are computed by using a series of low pass filter h[k], high
pass filters g[k] and down samplers across both rows and columns. The results are the
wavelet coefficient the next scale. The filter bank approach to calculate two dimen-
sional dyadic DWT is shown in figure 3 and dyadic representation of the DWT is
shown in figure 4 . The wavelet coefficients are of smaller spatial resolution as they
go from finer scale to coarser scale. The coefficients are called the approximation (A),
horizontal detail (H), vertical detail (V) and diagonal detail (D) coefficient.
Fig. 1. Two-dimensional orthogonal wavelet decomposition
)()(
,
,
=
∑
tatf jk
kj
kj
ψ
),(],[),(),(),(
,,,
,,,,
,0
0+= ∑∑∑∑=
∞
=
yxlkDyxlkCyxf
DVHSJJlk
S
lkj
S
jlkj
kj
J
ψφ
Wavelet Transform Based Image Registration and Image Fusion 57
2.1 Cross Correlation as Similarity Measure
Cross correlation is a similarity measure or match metric. For template T and image I,
where T is small compared to I, the two dimensional normalized cross-correlation
function measures the similarity for each translation.
)],(2.[
),(),(
),( −−
−−
=∑
∑
∑
∑
vyuxyIx
vyuxIyxyTx
vuC
(3)
If template matches the image, then cross correlation will have it’s peak.
2.2 Mutual Information as Similarity Measure
MI is an entropy-based concept and denotes the amount of information that one varia-
ble can offer to the other. Mutual Information criteria presented here states that, mu-
tual Information of image intensity values of corresponding voxel pairs is maximum
if images are geometrically aligned. Let A and B represent random variables and PA
(a) and PB (b) represents its marginal probability distributions. Let
represents joint probability distribution then are independent if
Mutual Information I (A, B) is given by
(4)
Mutual Information is related to entropy by following equations.
(5)
3 Application of Image Registration for Image Mosaicing
We collected series of images using digital camera on a leveled tripod in front of
Hiranandani complex, Kharghar, Navi Mumbai.
Hir-1 Hir-2 Hir-3 Hir-4 Hir-5
Fig. 2. Mosaic of images from Hir1 to Hir-5
),( baP AB
BA &
)()(),( bPaPbaP BAAB ∗=
)}]().(/{),(log[),(),(
,
=∑bPaPbaPbaPBAI BAAB
ba
AB
),()()(),(
−
−+= BAHBHAHBAI
58 M. Deshmukh and S. Gahankari
Table 1. Location of Maximum Match
S.N. Image
combination
Wavelet
Method
Using MI
Wavelet
Method
Using
Correlation
1 Hir-1-Hir-2 146 149
2 Hir-2-Hir-3 237 240
3 Hir-3-Hir-4 127 129
4 Hir-4-Hir-5 127 128
4 Wavelet Based Image Fusion
Image fusion is useful technique for merging similar sensor and multi-sensor images
to enhance the information.
Fig. 3. Wavelet multi-dimensional fusion
4.1 Wavelet Based Algorithm
• Apply wavelet transformation separately to each source image to establish various
images of
• wavelet tower shaped transformation.
• Fuse images at each transformation level.
• Apply inverse Wavelet transform on fused wavelet pyramid.
In wavelet transformation due to sampling, the image size is halved in both spatial
directions at each level of decomposition process thus leading to a multi-resolution
signal representation. The most important step for fusion is the formation of fusion
PET MRI W-fused Le-1 W-fused Le-2 W-fused Le-3 W-fused Le-4
Fig. 4. Sample images and Wavelet based fused images at different levels
Wavelet Transform Based Image Registration and Image Fusion 59
pyramid. We used mutual information based method for registering source images. In
the process of fusion, we fused images at four different levels. In the next section, we
make a quantitative evaluation of fusion at different levels.
Wavelet based fusion can deal with images of different spectral and spatial resolu-
tions. However, this method cannot handle cases where data is scattered or when
input images differ greatly in either their spectral or spatial resolution.
5 Conclusion
From experimental results it is observed that Mutual information method yields a
more accurate registration. But this method has its own limitations. When images
are of low resolution, when images contain little information, or when the region of
overlap is small then mutual information result in mis-registration. It has one more
limitation of speed, when time is an important constraint one cannot use this method.
Although it has some limitations entropy and mutual information are best approaches
for multimodal image registration. It is observed that combinational approach of
wavelet and mutual information gives better results as compared to wavelet - correla-
tion combination.Even wavelet mutual information combination can be used in case
of multimodal image registration. Wavelet based fusion can deal with images of
different spectral and spatial resolutions.
References
[1] Brown Gottesfeld, L.: Survey of Image Registration techniques. ACM Computing Sur-
veys 24(4), 325–376 (1992)
[2] Zitova, B., Flusser, J.: Image Registration Methods: A survey. Image and Vision Compu-
ting 21, 977–1000 (2003)
[3] Antoine Maintz, J.B., Vierger, M.A.: A Survey of Medical Image Registration. Medical
Image Analysis 2(1), 1–37 (1998)
[4] Sabuncu, M.R.: Spatial Information in Entropy – Based Image Registration. In: Gee, J.C.,
Maintz, J.B.A., Vannier, M.W. (eds.) WBIR 2003. LNCS, vol. 2717, pp. 132–141.
Springer, Heidelberg (2003)
[5] Chaudhari, S., Bhosale, U., Dutta Roy, S.: Multispectrul Panoramic Mosaicing. In: Inter-
national Conference on Advances in Pattern Recognition (ICPR), pp. 188–191. Indian
Statistical Institute, Kolkatta (2003)
[6] Bhosale, U., Chaudhari, S., Dutta Roy, S.: A Fast Method For Image Mosaicing Using
Geometric Hashing. IETE Journal of Research: Special Issue on Multimodal Media
Processing, 317–324 (May-August 2002)
[7] Flusser, J., Suk, T.: Degraded Image Analysis: an Invariant approach. IEEE Transaction
on Pattern Analysis and Machine Intelligence 20, 590–603 (1998)
[8] Flusser, J., Suk, T.: A Moment Based Approach to Registration of Images with Affine
Geometric Distortion. IEEE Transactions on Geoscience and remote Sensing 32, 382–387
(1994)
[9] Tsao, J.: Interpolation Artifacts in Multimodality Image Registration Based on Maximiza-
tion of Mutual Information. IEEE Medical imaging 22(7), 854–864 (2003)
60 M. Deshmukh and S. Gahankari
[10] Cahill, Williams, N.D., Shoupu, C.M.: Biomedical Imaging, Nano to Macro. In: 3rd IEEE
International Symposium, April 6-9, pp. 832–835 (2006)
[11] Pluim, J.P.W., Maintz, J.B.A.: Mutual Information Based registration of medical images,
survey. IEEE Medical Imaging
[12] Wang, X., Tian, J.: Image Registration based on Maximization of Gradient Code Mutual
Information. Image Anal. Streol. 24, 1–7 (2005)
[13] Maes, F., Collignon, A.: Multimodality Image Rregistration by Maximization of Mutual
Information. IEEE Transactions on Medical Imaging 16(2) (April 1997)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 61–66, 2011.
© Springer-Verlag Berlin Heidelberg 2011
60 GHz Radio Channel Characteristics in an Indoor
Environment for Home Entertainment Networks
T. Rama Rao1, S. Ramesh2, and D. Murugesan2
1 Dept. of Telecommunication Engineering,
SRM University. Kattankulathur – 603203, TN, India
ramarao@ieee.org
2 Dept. of ECE, Valliammai Engineering College
Kattankulathur – 603203, Tamil Nadu, India
rameshsvk@gmail.com, mdmurugesan@gmail.com
Abstract. 60 GHz radio channel propagation characteristics in a typical indoor
environment are addressed in this paper using a simple deterministic 2-
Dimensional model utilizing ray-tracing technique based on geometrical optics
(GO) and image principle. Values of the received power, rms delay spread (DS)
and power delay profile (PDP) for horn, omni-directional and isotropic antennas
are presented.
Keywords: Radio channel characterization, Indoor Environments, Millimeter
Waves, Ray-Tracing Technique, Antennas, Received Power, Delay Spread,
Wireless Networks.
1 Introduction
In today’s world, ´wireless´ is the new electricity and wireless technologies are allow-
ing us to enter an age where information becomes contextualized geographically as
well as personally. The last few years have seen a growth in the demand for wireless
broadband access. The reason for this growth can be seen as the emergence of multi-
media applications, demands for ubiquitous high-speed Internet connectivity, the
massive growth in the wireless and mobile communications and the deregulation in
the telecommunications industry. Recently, applications of millimeter (mm) waves for
high-speed broadband WLAN (Wireless Local Area Network) WPANs (Wireless
Personal Area Network) communication systems such as multimedia equipment,
home appliances, video signal transmissions, and personal computers in indoor envi-
ronment are increasingly gaining importance due to the spectrum availability and
wider bandwidth requirements [1, 2]. Within these mm wave bands, various kinds of
wireless data transmissions are expected to develop with/without license.
Propagation characteristics of mm waves, especially at 60 GHz are very different
from the ones at lower frequencies. Strong attenuation over the free space due to the
smaller wavelengths, oxygen absorption and severe attenuation by walls allow fre-
quency reuse and user privacy [3, 4, 5, 6]. This makes 60 GHz an attractive proposi-
tion for high-speed indoor WLAN & WPANs [7, 8]. It is important to know the
62 T. Rama Rao, S. Ramesh, and D. Murugesan
propagation characteristics of the wireless radio channel at 60 GHz for designing effi-
cient WLAN/WPAN systems. Next generation WLAN/WPAN’s at 60 GHz is now a
subject of intensive research topic.
In the present work, in order to describe the radio signal propagation at 60 GHz in
a typical indoor environment, we present a deterministic approach of a simple 2-
Dimensional model utilizing ray-tracing technique based on classical geometrical
optics (GO) and image method [9, 10, 11] to account for the direct ray and reflected
rays from wall, floor and ceiling, respectively. The key measurable parameters char-
acterizing the wideband indoor radio channel at 60 GHz are the received power, rms
delay spread and power delay profile. These are evaluated for horn, omni-directional
and isotropic antennas utilizing Matlab simulations. This paper is organized as fol-
lows; Section 2 deals with 60 GHz propagation channel model describing the geome-
try of the environment under consideration, the proposed ray-tracing model and the
simulation procedure. Section 3 deals with results obtained in our preset work and
discussions. Finally, Section 4 gives conclusions.
2 Propagation Channel Modeling
2.1 Description of the Indoor Environment
The considered simulation environment is a long corridor with dimensions 40.0 x 4.0 x
3.0 m3. The left and right wall surfaces are made of brick and plasterboard (relative
permittivity εr = 5.0 [12, 13]). In order to simplify the simulation procedure we assume
the surface as a uniform wall made of brick and plasterboard. The floor is concrete
covered with marble (εr = 4.0 [12, 13]) and furred ceiling is made of aluminum (εr = 1.0
[12, 13]). The beginning and the end of the corridor are open areas and are not taken
into account in the simulations.
2.2 Modeling
The radio channel propagation modeling at mm wave frequencies can be realized
based on ray-tracing theory. The ray-tracing method is among the available methods
for the relatively accurate estimation of field strengths to deal with the type of
complex layout that is often found in indoor environments [14] and allows fast com-
putation of single & double reflection processes. In 60 GHz region the diffraction
phenomenon can be neglected and the sum of the direct ray and the reflected rays are
enough to describe the behavior of the propagation channel with great accuracy [1,
15].In this present work, the proposed simple 2D model is a general case of the two-
ray model [9, 16]. The reflected components may exhibit single or double reflection
from a plane surface. If we know the geometry of the environment where the signal
propagates and the surface reflection coefficients, one may calculate the propagation
losses [11]. The total received power (RR) of the multi-rays are calculated by the
summation of ‘X’ single reflected and ‘W’ double reflected rays given by
2
3
21
1
2
0
1
1
23
21 )()()(
4d
e
RR
d
e
R
d
e
aa
jkd
W
j
jkd
X
i
jkd
rt
RR TR −
=
−
=
−∑∑ ++
⎟
⎠
⎞
⎜
⎝
⎛
=
θθθ
π
λ
(1)
60 GHz Radio Channel Characteristics in an Indoor Environment 63
where λ is the wave length; k is the wave number; d1 is the distance of the direct path;
d2is the distance of the single reflected path; d3is the distance of the double reflected
path; at, ar are the antenna functions; R(ө0)is the reflection coefficient of the single re-
flected ray on the reflecting surface; R(ө1), R(ө2) are the reflection coefficient of the
double reflected rays on respective reflecting surfaces; and TR is the transmitted power.
For isotropic antennas (at = ar = 1) the total received power (RR) is
2
3
21
1
2
0
1
1
23
21 )()()(
4d
e
RR
d
e
R
d
ejkd
W
j
jkd
X
i
jkd
RR TR −
=
−
=
−∑∑ ++
⎟
⎠
⎞
⎜
⎝
⎛
=
θθθ
π
λ
(2)
2.3 Simulation Procedure
In our simulations, we considered the direct ray, floor reflected ray, wall reflected rays,
wall-floor reflected rays, ceiling reflected ray and ceiling-floor reflected ray, and wall-
ceiling reflected ray. The simulations are conducted with Matlab script. The initial
transmitter position is at the beginning of the corridor and the receiver is moving away
from Tx with 1 m initial separation and we collect a signal sample as a function of dis-
tance every 0.00125 m (λ/4). During the entire simulation procedure vertical polariza-
tion is assumed. Hence, for the rays reflected from vertical walls we use the perpen-
dicular reflection coefficient, whereas for the rays from floor and ceiling surfaces we
use the parallel reflection coefficient. Both reflection coefficients are calculated using
the equations described in [16]. To examine how the antenna radiation patterns affect
the signal propagation in the indoor environment, we assumed three different transmis-
sion systems with different antenna characteristics and transmitted power. The systems
are 1) Horn antenna with 10 dBm output power, at = ar = 20.8 dBi, 2) Omni directional
antennas with 20 dBm output power, at = ar = 8.5 dBi, and 3) Isotropic antennas with
20 dBm output power. Further in our simulations, we are not considered the factors
such as the diffraction loss, atmospheric propagation losses, third/fourth order reflec-
tions and the non-uniformities of the surface materials, as these are almost negligible
and not contribute to the total received power.
3 Results and Discussions
3.1 Received Power
The received power provides an indication of the received signal strength with respect
to receiver position. It can be used by system designers to gauge what transmitted
power is required in order to achieve the desired performance parameters. The received
power is vital to determining cell sizes and co-channel interference effects; it also pro-
vides a preliminary measure of maximum raw data rates prior to modulation. The
antenna radiation patterns, have a significant impact on the received power, the simu-
lated results differ significantly for isotropic, omni-directional and horn antennas. The
uniform radiation characteristics of isotropic and omni-directional antennas, allow sig-
nals to be transmitted and received equally well from all directions. This results in all
the multipath components to reach the receiver, making the multipath effect much
more severe. Whereas for the horn antennas the narrow antenna bandwidths suppress
more of the multipath rays, hence in our simulation we discovered that using horn
64 T. Rama Rao, S. Ramesh, and D. Murugesan
antennas only first order reflected rays had a significant contribution to the received
power, with second order reflections being suppressed.
Table 1. Comparison of total received power values for different antenna configurations
Received Power for
Horn Antenna
(dBm)
Received Power for
Omni-directional Antenna
(dBm)
Received Power for
Isotropic Antenna (dBm)
Min -79.73 -94.33 -109.33
Max -13.44 -28.04 -43.04
Mean -36.47 -51.07 -66.07
Median -36.61 -51.22 -66.22
SD 7.26 7.26 7.26
Table 1 summarizes the received power statistics for isotropic, omni and horn an-
tennas. It is observed that the mean received power is -66.07, -51.07 and -36.47 dBm
for isotropic, omni-directional and horn antenna respectively. An increase of 29.7 dB
in received power between horn and isotropic antennas can be attributed to the ability
of the horn antenna to minimize its response to unwanted signals not in the favored
direction of the antenna. Inspecting the above table, we observe that the received
power is far greater when a horn antenna is employed in comparison to the isotropic
and omni-directional antennas.
3.2 Power Delay Profile (PDP)
The wideband channel is characterized by the time or space-variant channel impulse
response (CIR) [16]. In an indoor propagation environment where the typical time-
varying factors are human movement, it can be assumed that the channel is quasi-
stationary. The phase variations are assumed to be mutually independent random
variables, which have a uniform distribution over[-п,п]. Thus we considered only the
amplitude and delay components in our simulations. The most significant parameter
derived from the wideband channel model is the power delay profile [16] which is a
representation of the individual contributions of each multipath component with re-
spect to excess time delay and provides a visualization of vital delay statistics. The
PDP can be expressed as
)()()(
1
i
N
iRD d
RP
ττδτ
−= ∑
=
(3)
where RR(d) is the received signal power of the ith multipath component,τi is the excess
delay which is the relative delay of the ith component as compared to the first arriving
component and N is the total number of equally spaced multipath components. During
our quantization process we discretize the delay axis to have a time resolution of 1 ns
[16]. In this process we assigned each multipath echo to the nearest value of delay
equal to a multiple of 1 ns. The average received power in each bin is normalized to
the direct ray component. In the simulation process we obtain a PDP for each Tx-Rx
distance traversed, resulting in a set of PDPs. Each PDP is the individual power
contributions of each multipath component at that particular distance. To gain an
60 GHz Radio Channel Characteristics in an Indoor Environment 65
overall characterization of the channel we take the averages of these PDP’s. The PDPs
of horn, isotropic and omni antennas followed respective trends with large delays since
the entire multipath components contribute to the total received signal power.
3.3 rms Delay Spread (DS)
The important parameter derived from the PDP is the rms (root mean square) delay
spread (DS), is defined as square root of the second central moment of the average
PDP [16]. At mm wave frequencies the channel dispersion is smaller when compared
to values encountered at lower frequencies because echo paths are shorter on average.
Simultaneous measurements at 5 and 60 GHz indicate a difference of a factor 1.5 – 2
[17]. The rms DS of the channel may range from a few to 100 ns if linear polarisation
is used. It is expected to be highest if omni-directional antennas are used in large re-
flective indoor environments [18]. When, instead, high gain antennas are used, the rms
DS may be limited to a few ns only [18], but this is only the case when the antennas
are exactly pointed towards each other. In our simulations we discovered that when
isotropic antennas were employed, an average rms DS of approximately 2.2 ns was
obtained, this increase in the rms DS can be attributed to the fact that for isotropic an-
tennas all the multipath components contribute to the total received power. Obtained
rms DS of 1.59 ns with horn antennas due to its narrow antenna beamwidths sup-
presses some of the multipath components, reducing the rms DS, and hence increasing
the channel performance. With Omni directional antennas we obtained rms DS of 1.21
ns. Comparing our results with similar research works [11, 19, 20, 21], we find that our
simulated results are in agreement. The close correlation with our simulated observa-
tions is due to the similarity in configuration methodologies.
4 Conclusions
With the view to analyze 60 GHz wireless indoor scenario for WLAN & WPAN ap-
plications, radio channel characterization has been made in an indoor environment
using a simple deterministic 2D model utilizing ray-tracing technique based on geo-
metrical optics and image principle. From our results we observed that the antenna
radiation patterns have a significant effect on received power and the use of directive
horn antennas greatly improved the power performance in comparison to its iso-
tropic/omni-directional counterpart. In our simulations we observed that the PDP is
dominated by 1st order reflected components, whilst 2nd order reflected components
have a relatively insignificant contribution and the PDP follows an exponentially de-
caying relationship with respect to excess delay. Further, we observed that the rms DS
values decreased with the use of high gain directive antennas, due to suppression of
unwanted signal components by the narrow antenna beam width. It is believed that
the huge demand for bandwidth and higher data rate services will make the 60 GHz
channel an inevitable eventuality. However due to the complex nature of mm wave
propagation, there are still a lot of unknowns that need to be quantified before a work-
ing standard is achieved. With this in mind the work presented in this paper serves as
a contribution to the deployment of mm wave based WLAN & WPANs.
66 T. Rama Rao, S. Ramesh, and D. Murugesan
References
1. Correia, L.M., Prasad, R.: An overview of wireless broadband communications. IEEE
Commun. Mag., 28–33 (January 1997)
2. Smulders, P.F.M.: Exploiting the 60 GHz band for local wireless multimedia access: pros-
pects and future directions. IEEE Communications Magazine 40(1) (January 2002)
3. Andrisano, O., Chiani, M., Tralli.: Millimeter wave short range communications for ad-
vanced transport telematics. European Transactions on Telecommunications (July-August
1993)
4. Smulders, P.F.M., Wagemans, A.G.: Wideband Indoor Radio Propagation Measurements
at 58 GHz. Electron. Lett. 28(13), 1270–1272 (1992)
5. Dardari, D., Minelli, L., Tralli, V., Andrisano, O.: Wideband Indoor Communication
Channels at 60 GHz. In: Proceedings of PIMRC 1996, Taiwan, October 15-18 (1996)
6. Prasad, R.: Overview of Wireless Personal Communications: Microwave Perspective.
IEEE Communications Mag., 104–108 (April 1997)
7. Xiao, S.-Q., Zhang, M.-T.Z.Y.(eds.): Millimeter Wave Technology for Wireless LAN,
PAN and MAN. Auerbach Publications (2008)
8. Yong, S.K., Chong, C.-C.: An Overview of Multi gigabit Wireless through Millimeter
Wave Technology: Potentials and Technical Challenges. EURASIP Journal on Wireless
Communications and Networking, ArticleID 78907 (2007)
9. Betroni, H.L.: Radio Propagation for Modern Wireless Systems. Prentice Hall, Englewood
Cliffs (2000)
10. Hammoudeh, A.M., Graham, A.: Millimetric wavelengths radiowave propagation for LoS
Microcellular mobile communications. IEEE Trans. 44(3) (August 1995)
11. Nektarios, M., Philip, C.: Propagation Modeling at 60 GHz for Indoor Wireless LAN Ap-
plications. IST Mobile & Wireless Telecommunications Summit (2002)
12. Sato, K., et al.: Measurements of the complex refractive index of concrete at 57.5 GHz.
IEEE Trans. Antennas Propa. 44(1), 35–39 (1996)
13. Sato, K.: Measurements of reflection and transmission characteristics of interior structures
of office building in the 60-GHz band. IEEE Trans. Ant. Prop. 45 (December 1997)
14. Imai, T., Fujii, T.: Indoor micro cell area prediction system using ray-tracing for mobile
communication systems. In: Proc. PIMRC 1996, vol. 1, pp. 24–28 (1996)
15. Hübner, J., Zeisberg, S., Koora, S., Finger, A.: Simple Channel model for 60 GHz indoor
wireless LAN Design Based on Complex Wideband Measurements. In: IEEE 47th Vehicu-
lar Technology Conference, pp. 1004–1008 (1997)
16. Rappaport, T.S.: Wireless Communications. Prentice Hall, Englewood Cliffs (2000)
17. Plattner, A., Prediger, N., Herzig, W.: Indoor and outdoor propagation measurements at 5
and 60 GHz for radio LAN application. IEEE MTT-S Digest., 853–856 (1993)
18. Smulders, P.F.M.: Broadband wireless LANs: a feasibility study. Ph.D. Thesis, Eindhoven
University of Technology, The Netherlands (1995); ISBN 90-386-0100-X
19. Dardari, D., Minelli, L., Tralli, V., Andrisano, O.: Wideband indoor, communication chan-
nels at 60 GHz. In: PIMRC 1996, vol. 3, pp. 791–794 (October 15-18, 1996)
20. Ghobadi, C., Shepherd, P.R., Pennock, S.R.: 2D ray-tracing model for indoor radio propa-
gation at millimetre frequencies, and the study of diversity techniques. In: IEE Proc. Mi-
crow. Ant. Prop., vol. 145(4), pp. 349–353 (August 1998)
21. Chiu, C.C., Wang, C.P.: A Comparison of Wideband Communication Characteristics for
Various Corridors at 57.5 GHz. Wireless Personal Communications (12), 71–81 (2000)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 67–73, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Improved Back Propagation Algorithm to Avoid Local
Minima in Multiplicative Neuron Model
Kavita Burse1, Manish Manoria2, and Vishnu Pratap Singh Kirar1
1 Department of Electronics and Communication, Truba Institute of Engineering and
Information Technology, Bhopal, India
kavitaburse14@gmail.com, vishnupskirar@live.com
2 Department of of Computer Science and Engineering, Truba Institute of Engineering and
Information Technology, Bhopal, India
manishmanoria@rediffmail.com
Abstract. The back propagation algorithm calculates the weight changes of
artificial neural networks, and a common approach is to use a training algorithm
consisting of a learning rate and a momentum factor. The major drawbacks of
above learning algorithm are the problems of local minima and slow
convergence speeds. The addition of an extra term, called a proportional factor
reduces the convergence of the back propagation algorithm. We have applied
the three term back propagation to multiplicative neural network learning. The
algorithm is tested on XOR and parity problem and compared with the standard
back propagation training algorithm.
Keywords: Three term back propagation, multiplicative neural network,
proportional factor, local minima.
1 Introduction
Artificial Neural Network (ANN) consists of a number of interconnected processors
known as neurons, which are identical to the biological neural cells of the human
brain. Neural network is defined by its architecture, neuron model and the learning
algorithm. Architecture refers to a set of neurons and the weighted links connecting
the layers of neurons. Neuron model refers to information processing unit of the
neural network. The weights are adjusted during the training process. A learning
algorithm is used to train the NN by modifying the weights in order to model a
particular learning task correctly on the training examples. Learning is a fundamental
and essential characteristic of ANN. ANN training usually updates the weights
iteratively using the negative gradient of a Mean Squared Error (MSE) function. The
error signal is then back propagated to the lower layers.
The back propagation (BP) algorithm was developed by Rumelhart, Hinton and
Williams in 1986. Efficient learning by the BP algorithm is required for many
practical applications. The BP algorithm calculates the weight changes using a two-
term algorithm consisting of a learning rate and a momentum factor. The major
drawbacks of the two-term BP learning algorithm are the problems of local minima
68 K. Burse, M. Manoria, and V. P. S. Kirar
and slow convergence speeds. The addition of an extra term called a proportional
factor (PF) to the two-term BP algorithm was proposed in 2003 by Zweiri and has
outperformed standard two- term BP in terms of low complexity and computational
cost [1]. BP is a method for calculating the first derivatives or gradient of the cost
function required by some optimization methods. It is certainly not the only method
for estimating the gradient. However, it is the most efficient [2]. The major limitations
of this algorithm are the existence of temporary, local minima resulting from the
saturation behavior of the activation function. A number of approaches have been
implemented to avoid the local minima which are based on selection of dynamic
variation of learning rate and momentum, selection of better activation function and
better cost function.
In [3] the learning rate and momentum coefficient are adapted according to the
coefficient of correlation between the downhill gradient and the previous weight
update. In [4] modification is based on the solving of weight matrix for the output
layer using theory of equations and least squares techniques. Drago et al. have
proposed an adaptive momentum BP for fast minimum search [5]. A randomized BP
algorithm is proposed by Chen et al. It is obtained by choosing a sequence of
weighting vectors over the learning phase [6]. A new generalized BP algorithm is
proposed in [7] to change the derivative of the activation function so as to magnify the
backward propagated error signal, thus the convergence rate can be accelerated and
the local minimum can be escaped. An adaptive BP algorithm is proposed in [8]
which can update learning rate and inertia factor automatically based on dynamical
training error rate of change. In [9] an improved BP is proposed where each training
pattern has its own activation function of neurons in hidden layer to avoid local
minima Wang et al. have proposed an individual inference adjusting learning rate
technique to enhance the learning performance of the BP neural network [10]. In [11]
Conjugate Gradient (CG) algorithm which is usually used for solving nonlinear
functions is combined with the modified Back Propagation (BP) algorithm yielding a
new fast training multilayer algorithm. The proposed algorithm improved the training
efficiency of BP-NN algorithms by adaptively modifying the initial search direction.
This paper is organized as follows. In the next section we propose the learning rule
with improved BP algorithm to avoid local minima in multiplicative neuron model. In
section 3, we illustrate the basic results of our paper. Section 4 concludes the paper.
2 Learning with Improved BP Algorithm
The McCulloch-Pitts model initiated the use of summing units as the neuron model,
while neglecting all possible nonlinear capabilities of the single neuron and the role of
dendrites in information processing in the neural system. It is widely agreed that there
is only minor correspondence between these neuron models and the behavior of real
biological neurons. In particular, the interaction of synaptic inputs is known to be
essentially nonlinear. In search for biologically closer models of neural interactions,
neurobiologists have found that multiplicative-like operations play an important role
in single neuron computations. For example, multiplication models nonlinearities of
dendritic processing and shows how complex behavior can emerge in simple
Impro
v
networks. In recent years e
nervous system of several
a
increases the computation
a
known from extensions o
f
[12].
The back-propagation
l
(MNN) has to be more effi
c
learning is to update th
e
difference between the ac
t
vector. The rapid comput
a
since, in general, the numb
convex multimodal objecti
broad flat regions adjoine
d
described as follows. The
b
as depicted in Fig. 1[13].
A node receives a numb
set of weights
are multiplied to form a
p
subjected to a nonlinear fu
n
In the MNN a numbe
multidimensional input is
p
first layer nodes then beco
m
output of the network is
connections exist from a
connections exist between
MNN is increased decisio
n
and have highly nonlinea
r
forward MNN where the s
u
n
www ...,21
v
ed Back Propagation Algorithm to Avoid Local Minima
vidence has accumulated regarding specific neurons in
a
nimals compute in a multiplicative manner. Multiplica
t
a
l power and storage capacity of neural networks is
w
f
ANN where this operation occurs as higher order
u
l
earning algorithm with multiplicative neural netw
o
c
ient as both single-units and also in networks. The go
a
e
network weights iteratively to minimize globally
t
ual output vector of the network and the desired ou
t
a
tion of such a global minimum is a rather difficult
t
er of network variables is large and the corresponding
n
ve function possesses multitudes of local minima and
d
with narrow steep ones. The architecture of the MN
N
b
asic building block of the MNN is a single neuron or
n
Fig. 1. Node structure of MNN
er of real inputs which are then multiplied
b
and bias terms are added. The resultant va
l
p
olynomial structure. This output of the node is fur
t
n
ction defined as .
r of nodes described above are arranged in layers
.
p
assed to each node of the first layer. The outputs of
m
e inputs to the nodes in the second layer and so on.
T
the output of the nodes of the final layer. Weig
h
node
t
o every node in the succeeding node but
nodes of the same layer. As the number of layers in
n
regions are formed which are considerably more com
p
r
boundaries. Fig. 2 shows a general model of a
f
u
mmation at each node is replaced by the product unit.
n
xxx ...,21
n
n
bbb ...,21
fx
x
e
e
f−
−
+
−
=1
1
69
the
t
ion
w
ell
u
nits
o
rks
a
l of
the
t
put
t
ask
n
on
has
N
is
n
ode
b
y a
l
ues
t
her
.
A
the
T
he
h
ted
no
the
p
lex
f
eed
70 K. Burse, M. Manor
i
Fig
The MNN is trained
u
corresponding target vector
network is trained increme
n
input is presented. The M
N
and the corresponding targ
e
bias. Output of the node u
b
The bipolar sigmoidal acti
v
An error back propagation
b
Where, is the number o
f
The weight update equa
t
Where, is the learning r
a
quickly, but if it is too large
The bias is updated as
p
i
w
−
=Δ
−=
η
Δ
i
a, and V. P. S. Kirar
. 2. Architecture of feed forward MNN
u
sing supervised learning where a set of input and
is used to adjust the scalar parameters weight and bias.
T
n
tally so that the weights and biases are updated after
e
N
N is trained using supervised learning where a set of i
n
e
t vector is used to adjust the scalar parameters weight
b
efore applying activation function is given by [14].
v
ation function f is given by
b
ased learning rule is used for training. The MSE is given
f
input patterns.
t
ion for single layer algorithm is given by
a
te and is the desired signal. If is large, learning oc
c
it may lead to instability and errors may even increase.
)(
1
ii
n
i
ibxwu += ∏
=
u
u
e
e
ufy −
−
+
−
== 1
1
)(
2
1
)(
2
1p
N
p
pdy
N
E−= ∑
=
i
w
E
∂
∂
−
η
i
iii
x
bxw
u
yydy )(
)1)(1)((
2
1
+
−+−
η
d
η
i
ib
E
b∂
∂
−=
Δ
η
)(
)1)(1)((
2
1
iii bxw
u
yydy +
−+−−=
η
the
T
he
e
ach
n
put
and
(1)
(2)
by
(3)
(4)
c
urs
(5)
Improved Back Propagation Algorithm to Avoid Local Minima 71
The standard BP algorithm calculates the new weights and biases as
(6)
(7)
The standard algorithm is further modified by adding the momentum term and
proportional factor term. The momentum term is a fraction of the previous weight
change. The momentum term prevents extreme changes in the gradient due to
anomalies and suppresses oscillations due to variations in the slope of the error
surface [15] and prevents the network to fall into shallow local minima. The
convergence still remains relatively slow because of the saturation behavior of the
activation function. In the saturation area of the output activation function, the
corresponding gradient descent takes very small value leading to small changes in
weight adjustments. The problem of slow convergence is solved by adding a term
proportional to the difference between the output and the target. The improved BP
weight update is calculated as
(8)
(9)
is the proportional term
is the previous weight change
is the proportional term
is the difference between the output and the target at each iteration
is the previous bias change
The error function optimization depends on three independent quantities. The three
term back propagation function as a PID controller used in control application [16].
3 Simulation and Results
We have tested the convergence of the three term BP algorithm for the MNN network
on the XOR problem which is the most used nonlinear pattern classification problem
as compared to other logic operations. The architecture of the network is a single
layer MNN with 2 inputs, 3 hidden layer neurons and 1 output neuron. The
convergence curve for the three term BP algorithm is compared with the standard BP
algorithm in Fig. 3. The convergence of the improved BP algorithm with momentum
and PF factor is five times faster as compared to the standard BP algorithm. Table 1
compares the testing performance for the XOR problem.
i
old
i
new
iwww Δ+=
i
old
i
new
ibbb Δ+=
)( dywww old
ii
improved
i−+Δ+Δ=Δ
γβ
)( dybbb old
ii
improved
i−+Δ+Δ=Δ
γβ
β
wold
i
Δ
γ
)( dy −
bold
i
Δ
72 K. Burse, M. Manor
i
Fig.
3
Table
Input Tar
g
0 0 0
0 1 1
1 0 1
1 1 0
Table 2.
T
Input Ta
r
0 0 0
0
0 0 1
1
0 1 0
1
0 1 1
0
1 0 0
1
1 0 1
0
1 1 0
0
1 1 1
1
We have further tested
t
three bit parity problem w
h
bit output is 1 for odd
n
performance for parity pro
b
i
a, and V. P. S. Kirar
3
. Convergence curves for XOR problem
1. Testing performance for XOR problem
g
et Output with MNN
trained with standard
BP algorithm
Output with MNN
trained with three
term BP algorithm
0.0004 0.0001
0.9841 0.9942
0.9835 0.9997
0.0002 0.0001
T
esting performance for 3 bit parity problems
r
get Output with MNN
trained with
standard
BP algorithm
Output with MNN
trained with three
term BP algorithm
0
0.0312 0.0032
1
0.8921 0.9978
1
0.9876 0.9886
0
0.0214 0.0021
1
0.8953 0.9778
0
0.0021 0.0041
0
0.0032 0.0034
1
0.9873 0.9921
t
he three term BP algorithm for the MNN network on
h
ich maps 3 bit binary numbers onto its parity. The p
a
n
umber of 1 else it is 0. Table 2 compares the tes
t
b
le
m
.
the
a
rity
t
ing
Improved Back Propagation Algorithm to Avoid Local Minima 73
4 Conclusion
In this paper we have proposed an improved BP algorithm to avoid local minima and
for faster convergence of multiplicative neural network training algorithm. We have
tested the algorithm for XOR and three bit parity problem and compared the result
with standard BP multiplicative neural network algorithm. The addition of PF term
helps in convergence of the algorithm five times faster.
References
1. Zweiri, Y.H., Whidborne, J.F., Althoefer, K., Seneviratne, L.D.: A three term back
propagation algorithm. Neurocomputing 50, 305–318 (2003)
2. Edward, R.J.: An Introduction to Neural Networks. In: A White paper. Visual Numerics
Inc., United States of America (2004)
3. Yam, Y.F., Chow, T.W.S.: Extended back propagation algorithm. Electronics
Letters 29(19), 1701–1702 (1993)
4. Verma, B.K., Mulawka, J.J.: A modified back propagation algorithm. In: IEEE World
Congress on Computational Intelligence, pp. 840–844 (1994)
5. Drago, G.P., Morando, M., Ridella, S.: An adaptive momentum back propagation. Neural
Computing and Application 3, 213–221 (1995)
6. Chen, Y.Q., Yin, T., Babri, H.A.: A stochastic back propagation algorithm for training
neural networks. In: International Conference on Information, Communications and Signal
Processing, Singapore, pp. 703–707 (1997)
7. Ng, S.C., Leung, S.H., Luk, A.: Fast convergent generalized back propagation algorithm
with constant learning rate. Neural Processing Letters 9, 13–23 (1999)
8. Wen, J.W., Zhao, J.L., Luo, S.W., Han, Z.: The improvements of BP neural network
learning algorithm. In: ICSP 2000, pp. 1647–1649 (2000)
9. Wang, X.G., Tang, Z., Tamura, H., Ishii, M., Sun, W.D.: An improved back propagation
algorithm to avoid the local minima problem. Neuro Computing 56, 455–460 (2004)
10. Wang, C.H., Kao, C.H., Lee, W.H.: A new interactive model for improving the learning
performance of back propagation neural network. Automation in Construction 16(6),
745–758 (2007)
11. Bayati, A.Y., Al, S.N.A., Sadiq, G.W.: A modified conjugate radient formula for back
propagation Neural Network Algorithm. Journal of Computer Science 5(11), 849–856
(2009)
12. Mel, B.: Information processing in dendritic trees. Neural Computing 6, 1031–1085 (1994)
13. Yadav, R.N., Kalra, P.K., John, J.: Time series prediction with single multiplicative neuron
model. Applied Soft Computing 7, 1157–1163 (2007)
14. Yadav, R.N., Singh, V., Kalra, P.K.: Classification using single neuron. In: IEEE Int.
Conf. on Industrial Informatics, Banff, Alberta, Canada, pp. 124–129 (2003)
15. Yu, C.C., Liu, B.D.: A back propagation algorithm with adaptive learning rate and
momentum coefficient. In: The International Joint Conference on Neural Networks,
IJCNN 2002, pp. 1218–1223 (2007)
16. Zweiri, Y.H.: Optimization of a Three-Term Backpropagation Algorithm Used for Neural
Network Learning. International Journal of Engineering and Mathematical Sciences 3(4),
322–327 (2007)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 74–81, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Technical White Paper on
“Time and Frequency Synchronization in OFDM”
Anagha Rathkanthiwar1 and Mridula Korde2
1 Department of Electronics Engineering
Priyadarshini College of Engineering, Nagpur
anagharathkanthiwar@yahoo.co.in
2 Department of Electronics Engineering
Shri Ramdeobaba Kamala Nehru College of Engineering
Nagpur
mridulakorde@yahoo.com
Abstract. Orthogonal frequency division multiple access (OFDMA) is a
emerging standard for broadband wireless access. Synchronization in OFDMA
represents one of the most challenging issues and plays a major role in the
physical layer design. Aim of this paper is to provide a overview of various
frequency and time synchronization errors in OFDM systems. This paper also
discusses the effect of timing and frequency errors on system performance. It
focuses on time and frequency error estimation algorithms for OFDM based
systems in downlink transmission as well.
Keywords: Downlink synchronization, frequency correction, frequency
estimation, orthogonal frequency division multiple access (OFDMA),
orthogonal frequency division multiplexing (OFDM), timing estimation.
1 Introduction
The demand for wireless communications is growing today at an extremely rapid pace
and this trend is expected to continue in the future. The common feature of many
current wireless standards for high-rate multimedia transmission is the adoption of a
multi-carrier air interface based on orthogonal frequency division multiplexing
(OFDM). OFDM is a special case of multi-carrier transmission, where a single data
stream is transmitted over a number of lower rate (thus increased symbol duration)
sub-carriers. The increased symbol duration of OFDM symbol improves the
robustness to channel delay spread and use of orthogonal sub-carriers makes efficient
use of the spectrum by allowing overlap. It eliminates Inter Symbol Interference (ISI)
and Inter Block Interference (IBI) through use of a cyclic prefix to some extent.
Furthermore, it provides larger flexibility by allowing independent selection of the
modulation parameters (like the constellation size and coding scheme) over each sub-
carrier. OFDM Modulation can be realized with Inverse Fast Fourier Transform
(IFFT). Due to all these favorable features, many digital transmission systems have
adopted OFDM as the modulation technique such as digital video broadcasting
Technical White Paper on “Time and Frequency Synchronization in OFDM” 75
terrestrial TV (DVB-T), digital audio broadcasting (DAB), terrestrial integrated
services digital broadcasting (ISDB-T), digital subscriber line (xDSL) etc.
IEEE802.11a standard is the first one to use OFDM in packet-based communications,
while the use of OFDM until now was limited to continuous transmission systems [1].
Now it is being used in other packet based systems like multimedia mobile access
communications (MMAC), and the fixed wireless access (FWA) system in IEEE
802.16.3 standard [2]. It has become fundamental technology in the future 3GPP LTE
and 4G-multimedia mobile communications systems.
Despite its appealing features, the design of an OFDM system poses several
technical challenges. One basic issue is related to the stringent requirement on
frequency and timing synchronization [4]. In OFDM downlink transmission, each
terminal has to perform timing and frequency synchronization by exploiting the
broadcast signal transmitted by the BS. Synchronization is performed in two phases
one is acquisition phase and other is tracking phase. During acquisition step reference
blocks are exploited to get coarse estimates of synchronization parameters i.e.
frequency and timing errors. These estimates are then refined during tracking step.
1.1 Effect of Timing Errors
This section describes the effect of uncompensated timing error on system
performance. Timing error occurs because of multipath dispersion. Due to this timing
error the receiver’s time-domain FFT window spans samples from two consecutive
OFDM symbols. This results in inter-OFDM symbol interference. Additionally, even
small misalignments of the FFT window result in an evolving phase shift in the
Fig. 1. Partial overlapping between received blocks due to multipath dispersion
frequency domain symbols, leading to BER degradation [5]. As shown in Fig. 1 the
tail of each received block extends over the first L - 1 samples of the successive block
as a consequence of multipath dispersion. If cyclic prefix is greater than Channel
Impulse Response (CIR) duration ( 0 to L-1) then there is certain interval which is
not affected by previous block. As long as the DFT window starts anywhere in this
interval, no IBI is present at the DFT output. This situation occurs whenever the
timing error Δθ belongs to interval -Ng + L - 1 ≤ Δθ ≤ 0 and only results in a cyclic
shift of the received OFDM block [6]. Thus, recalling the time-shift property of the
76 A. Rathkanthiwar and M. Korde
Fourier transform and assuming perfect frequency synchronization the DFT output
over the nth subcarrier takes the form
Ri(n) = ej2πnΔθ/N H(n)di(n) + Wi(n)
Here
N
lnj
L
l
elhnH
π
2
1
0
)()( −
−
=
∑
=
(2)
where )(lh is the channel impulse response and di is the ith OFDM block. The
equation of Ri(n) indicates that timing error Δθ appears as a linear phase across
subcarriers and it can be compensated by the channel equalizer. On the other hand, if
the timing error is outside the interval -Ng + L - 1 ≤ Δθ ≤ 0 , samples at the DFT input
will be contributed by two adjacent OFDM blocks. In addition to IBI, this results in a
loss of orthogonality among subcarriers which, in turn, generates ICI. In this case, the
nth DFT output is given by
Ri(n) = ej2πnΔθ/N α(Δθ)H(n)di(n) + Ii(n,Δθ) + Wi(n) (3)
where α(Δθ) is an attenuation factor while Ii(n,Δθ) accounts for IBI and ICI and can
reasonably be modeled as a zero-mean random variable with power σI2(Δθ). The loss
in SNR parameter can be obtained by using
(4)
Where ideal
SNR is the SNR of a perfectly synchronized system and real
SNR is the
SNR in the presence of a timing offset.
For a normalized channel response with unit average power )(
θ
γ
Δ is given by
(5)
1.2 Effect of Frequency Errors
This section describes the effect of uncompensated carrier frequency error on system
performance. Carrier frequency errors occurs due to offset between the incoming
waveform and the local references used for signal demodulation. This error result in a
shift of the received signal in the frequency domain. If the frequency error is an
integer multiple of the sub-carrier spacing, then the received frequency domain
modulated sub-carriers are shifted by sub-carrier positions. The sub-carriers are still
mutually orthogonal but the received data symbols are now in the wrong position in
the demodulated spectrum, causing increase in BER. If the carrier frequency error is
not an integer multiple of the sub-carrier spacing, then it results in loss of mutual
orthogonality between the sub-carriers. ICI is then observed between the sub-carriers,
which deteriorates the BER performance of the system. To better explain this concept,
assume ideal timing synchronization and compute the DFT output corresponding to
real
ideal
SNR
SNR
=Δ )(
θγ
⎥
⎦
⎤
⎢
⎣
⎡Δ
+
Δ
=Δ 2
2
1
2
)(
1
)(
1
)(
w
σ
θσ
θα
θγ
Technical White Paper on “Time and Frequency Synchronization in OFDM” 77
the ith OFDMA block in the presence of a frequency error ε [6]. In case when
frequency error ε is integer multiple of subcarrier spacing, the DFT output is given by
Ri(n) = e jΨiH(|n- ε|N)di (|n- ε|N) + Wi(n) (6)
Here (|n- ε|N) is the value of n- ε reduced to interval [0, N-1]. This equation shows that
even though the received symbols appear in a wrong position at the DFT output, no
ISI is present as orthogonality among subcarriers is preserved. The situation is
drastically different when frequency error ε is not integer valued. In this case, the
subcarriers are no longer orthogonal and DFT output can conveniently be rewritten as
Ri(n) = e jΨiH(n)di (n)fN(ε) + Ii(n,ε) + Wi(n) (7)
Here Ii(n,ε) is a zero mean ICI term. Loss in SNR and )(xfNare given by (8) and
(9) as follows
(8)
(9)
2 Synchronization Algorithms for Downlink Transmission
As discussed OFDMA system is extremely sensitive to timing errors and carrier
frequency offsets between the incoming waveform and the local references used for
signal demodulation. Inaccurate compensation of the frequency offset destroys
orthogonality among subcarriers and produces ICI. Timing errors result in IBI & ISI
and must be counteracted to avoid severe error rate degradations. Therefore,
synchronization is extremely crucial to the OFDM systems.
Synchronization algorithms can be divided into following categories:
1.Non data aided methods :based on use of internal structure of OFDM symbols.
2.Data aided methods : based on training symbols or pilots.
In continuous mode transmission systems, there is no stringent requirement on
acquisition time hence averaging method can be used to improve estimation accuracy.
It is appropriate to apply non data aided methods which makes use of CP for this
mode. However in the burst packet mode, synchronization ought to be established at
any time because when data streams are ready to transmit is unknown hence there is
stringent requirement on synchronization time.
The report [1] described use of CP (non data aided method) for both timing and
frequency synchronization considering that first TG (duration of CP) seconds of each
symbol is identical to last part. This algorithm correlates a TG long part of the signal
with a part that is T (symbol duration) seconds delayed. Frequency offset is then
estimated by averaging correlation output function over interval TG and finding phase
of this output function.
The conventional algorithms for the coarse symbol timing synchronization in time
domain are MLE (Maximum Likelihood Estimation) utilizing the cyclic prefix of the
OFDM symbols. This technique of synchronization was proposed by J. J. Van de
[]
⎭
⎬
⎫
⎩
⎨
⎧−+= 2
2
2
2)(11
)(
1
)(
ε
σ
ε
εγ
N
w
N
f
D
f
NNxj
Ne
NxN
x
xf /)1(
)/sin(
)sin(
)( −
=
π
π
π
78 A. Rathkanthiwar and M. Korde
Beek [2]. However, good performance is achieved only for AWGN channel. For
multipath fading channels data is badly contaminated by ISI and there is significant
influence on carrier frequency offset (CFO).
To improve performance of MLE under multipath fading channels, a novel scheme
was proposed in [3]. This scheme utilizes both redundancy in CP & Pilots symbols for
coarse symbol timing synchronization. This estimator allow a shorter CP and thus a
more spectrally efficient system.
Fig. 2. Training Symbol of S & C algorithm
Fig. 3. Training Symbols as reference blocks placed in the beginning of frame
But because of limited number of pilots used for estimation, non negligible
fluctuation still exists. In order to mitigate the fluctuation T.M. Schmidl and D.C. Cox
[4] (S & C) introduced the method allowing large acquisition range for the carrier
frequency offset by using one unique symbol composed of two identical halves of half
the length of symbol which is transmitted at the beginning of each frame as shown in
Fig. 3. The symbol is shown in Fig.2. The algorithm is based on concept that a
training symbol with two identical halves in the time domain, will remain identical
after passing through the channel, except that there will be a phase difference between
them caused by the carrier frequency offset. Consider the training symbol where the
first half is identical to the second half (in time order), except for a phase shift caused
by the carrier frequency offset. If the conjugate of a sample from the first half is
multiplied by the corresponding sample from the second half ( seconds later), the
effect of the channel should cancel, and the result will have a phase of approximately
φ= π T Δf. At the start of the frame, the products of each of these pairs of samples will
have approximately the same phase, so the magnitude of the sum will be a large
value. As long as the CP is longer than the CIR duration, the two halves of the
reference block will remain identical after passing through the transmission channel
except for a phase shift induced by the CFO and the received samples corresponding
to the first half as
r(k) = s(R)(k)e j2πεk/N + w(k), θ ≤ k ≤ θ + N/2-1 (10)
Received samples corresponding to second half are given by
r(k + N/2) = s(R)(k)e j2πεk/N ejπε + w(k +N/2), θ ≤ k ≤ θ + N/2-1 (11)
Technical White Paper on “Time and Frequency Synchronization in OFDM” 79
Timing estimate is
(12)
Where
The above method is rapid and suitable for continuous transmission or a burst
operation over a frequency-selective channel. Unfortunately metric of this algorithm
exhibits a large plateau that may greatly reduce the estimation accuracy. The start of
the frame and the beginning of the symbol can be found, and carrier frequency offsets
of many subchannels spacing can be corrected. The algorithms operate near the
Cramer–Rao lower bound for the variance of the frequency offset estimate, and the
inherent averaging over many subcarriers allows acquisition at very low signal-to-
noise ratios (SNR’s).
JungJu Kim; Jungho Noh; KyungHi Chang, presented [5] a preamble timing
synchronization method for OFDMA timing estimation as modification to T. M.
Schmidl and D. C. Cox to reduce uncertainty due to timing metric plateau. The
training symbol used in this method consists of four segments in which first & third
segments are identical and second & fourth segments are symmetric conjugate with
first segment as shown in figure 4.Therefore, the proposed efficient timing
synchronization method guarantees the better performance of the initial timing
synchronization of OFDMA systems.
Thomas Keller, Lorenzo Piazzo, Paolo Mandarini, and Lajos Hanzo in their paper
[7] suggested use of cyclic post amble to mitigate effect of timing errors. The
reference symbols used by their algorithm consists of repetitive copies of a
synchronization pattern as shown in figure 5 and proposed algorithm rely on
evaluation of correlation functions as given in equations (14), (15) one corresponding
to OFDM symbol and other corresponding to reference symbol.
Fig. 4. Training symbol used by algorithm in [5]
Fig. 5. Repetitive copies of a synchronization pattern
{
}
)
~
(maxarg
ˆ
θθ
Γ=
∑
∑
−+
=
−+
=
+
+
=Γ 1
2
~
~
2
1
2
~
~
*
)2/(
)()2/(
)
~
(N
q
N
q
Nqr
qrNqr
θ
θ
θ
θ
θ
80 A. Rathkanthiwar and M. Korde
)(JG is used both for symbol synchronization and frequency tracking. Where as
)(JR is used for frequency acquisition. Also they proposed in the same paper joint
frequency and time synchronization acquisition algorithm using periodicity of CP of
OFDM symbol.
Two algorithms for timing synchronization using preamble with special properties
consisting of two part each of one symbol duration were described by M. Gertou, G.
Karachalios, D. Triantis, K. Papantoni and P. I. Dallas [8]. One algorithm utilizes
only first part of preamble and take average of two detected positions obtained in the
defined metric function giving better timing estimation. Second algorithm uses first
part of preamble for coarse and second part for fine symbol synchronization.
Moose, P.H in his paper [9] describes a technique to estimate frequency offset
using repeated data symbols with MLE. It has been shown that for small error in the
estimate, the estimate is conditionally unbiased and is consistent in the sense that the
variance is inversely proportional to the number of carriers in the OFDM signal.
Furthermore, both the signal values and the IC1 contribute coherently to the estimate
so that it is possible to obtain very accurate estimates even when the offset is too
great. Since the estimation error depends only on total symbol energy, the algorithm
works equally well in multipath spread channels and frequency selective fading
channels. However, it is required that the frequency offset as well as the channel
impulse response be constant for a period of two symbols.
The synchronization schemes described in [1], [7], [2], [9] are Non Data Aided
methods and the schemes described in [3], [4], [5], [8] are Data Aided methods of
synchronization.
3 Conclusion
As mentioned in the earlier section, synchronization methods can be divided into two
categories: DA and NDA algorithms. Pilots, training symbols or the combination of
them are generally applied to the DA-type of methods achieving synchronization in
less time but at the expense of the reduced bandwidth efficiency. For the NDA
methods, data used for the estimation may be contaminated by ISI, resulting in the
inaccurate estimation. Throughput and power efficiency are improved but time taken
to establish synchronization is increased. Synchronization time, algorithm
complexity, the required system performance and etc. are all the factors that should
be considered when choosing the synchronization scheme for the particular system.
After studying and analyzing various available OFDM synchronization algorithms
a suitable algorithms can be developed. Simulation of these algorithms can be
performed for various parameters like probability of missed detection, probability of
∑
−
=
−−−=
1
0
*
)().()(
g
N
m
NmjzmjzJG
∑−−+
=
−−−=
1
0
*
)().()(
Sg NNN
m
NmjzmjzJR
Technical White Paper on “Time and Frequency Synchronization in OFDM” 81
false alarm, for different values of threshold. Also simulation of estimation error
probability vs SNR can be done. Complexity of algorithms vs performance kind of
simulation can also be done.
Design and appropriate use of reference symbols to be used for synchronization is
important. So suitable reference symbol can be developed for estimation of time and
frequency errors.
References
1. Intini, A.L.: Orthogonal frequency division multiplexing for wireless networks standard
IEEE 802.11a. University Of California Santa Barbara (December 2000)
2. van de Beek, J.J., Sandell, M., Borjesson, P.O.: ML estimation of time and frequency offset
in OFDM systems. IEEE Transactions on Acoustics, Speech and Signal Processing 45(7),
1800–1805 (1997)
3. Landström, D., Wilson, S.K., Van de Beek, J.J., Odling, P., Börjesson, P.O.: Symbol time
offset estimation in coherent OFDM systems. In: Proc. Int. Conf. On Communications,
Vancouver, BC, Canada, vol. 1, pp. 500–505 (June 1999)
4. Schmidl, T.M., Cox, D.C.: Robust frequency and timing synchronization for OFDM. IEEE
Trans. on Commun. 45(12), 1613–1621 (1997)
5. Kim, J., Noh, J., Chang, K.: An efficient timing synchronization method for OFDMA
system. In: IEEE/ACES International Conference on Wireless Communications and
Applied Computational Electromagnetics, April 3-7, pp. 1018–1021 (2005)
6. Morelli, M., Kuo, C.-C.J., Pun, M.-O.: Synchronization Techniques for Orthogonal
Frequency Division Multiple Access (OFDMA): A Tutorial Review. Proceedings of the
IEEE 95(7), 1394–1427 (2007)
7. Keller, T., Piazzo, L., Mandarini, P., Hanzo, L.: Orthogonal Frequency Division Multiplex
synchronization techniques for frequency-selective fading channels. IEEE Journal on
selected areas in communications 19(6) (June 2001)
8. Gertou, M., Karachalios, G., Triantis, D., Papantoni, K., Dallas, P.I.: Synchronization
Approach for OFDM based Fixed Broadband Wireless Access Systems. INTRACOM S.A
9. Moose, P.H.: A technique for orthogonal frequency division multiplexing frequency offset
correction. IEEE Transactions on Communications 42(10), 2908–2914 (1994)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 82–86, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Cell-ID Based Vehicle Locator and Real-Time
Deactivator Using GSM Network
Nilesh Dubey1, Vandana Dubey1, and Shivangi Bande2
1 Sanghvi Institute of Management and Science, Indore-453332
2 Institute of Engineering and Technology DAVV, Indore-452009
Abstract. The work “Cell-ID Based Vehicle Locator and Real-Time Deactivator
Using GSM Network” (VLRD) uses GSM network for locating and real-time
controlling on the vehicle by using the cell-id sent by guardian software and real
time decoding of the encoded commands which are communicated through the
circuit switched network. The system gets the command from the user and acts
according to a pre-programmed set of instructions in real time to avoid the prob-
lems of delayed or undelivered commands in the existing system and improves
the assurance of command execution. Removing the GPS usage can be reached
through detecting the moments that Cell-ID is providing sufficient accuracy for
the envisaged application.
Keywords: Cell-ID, DTMF, Embedded, GSM, Immobilizer.
1 Introduction
NICB (National Insurance Crime Bureau) of United States had a report which stated
that there were more than one million cars stolen in 2006. The value of the stolen cars
was about ninety seven hundred millions dollars, and every twenty six seconds there
was a car being stolen. Even advanced technologies such as electric locks and elec-
tronic immobilizers have been applied to vehicles; cars are still stolen by thieves to
resell the components or the whole car.
The presently used systems are very costly and need high maintenance cost as well.
Presently used system uses GPS tracking which can only track the vehicle but cannot
send control command to vehicle. Some systems allow immobilizing the vehicle
through sending the control command through SMS.
But the problem with the presently used technique is, they uses the GPS tech-
nology for getting the position of the vehicle and sends the GPS data via GSM/GPRS
modem through SMS.
1. These units are very expensive approximately Rs.16000 to 25000 in Indian
market.
2. Second problem is monthly rental of the service provided which is approximately
Rs.500 per month + SMS charges + GPRS charges.
3. The strength of the GPS signals, which is very weak in the urban areas, under the
tree, basement of building or any type of shadows. The GPS almost in non work-
ing condition in case of weak signals
Cell-ID Based Vehicle Locator and Real-Time Deactivator Using GSM Network 83
4. GPS need a long antenna for receiving the signals from the satellites.
Not all the tracking unit provides the function of vehicle immobilization and which
are providing are facing the problem of non real-time controlling on the vehicle be-
cause of the undelivered or delayed SMS command.
The work suggests solution of the above problems. It uses only GSM network for the
whole working of the system. The mobile phone market lacks a satisfactory location tech-
nique that is accurate, but also economical and easy to deploy. Current technology pro-
vides high accuracy, but requires substantial technological and financial investment.[4]
The GSM signals are much stronger then the GPS signals and no need of any ex-
ternal antenna in the unit. It is very cost effective and it sends the location data only
when we need.
The user needs not to pay any monthly charges. It gives the real time controlling on
the vehicle because of the DTMF commands over the voice channel which has high
priority then the data channel which used for sms service.
VLRD uses GSM network for locating the vehicle. It uses the database of the GSM
service provider for getting the location of the cell in which the vehicle is currently run-
ning. The hidden unit in side the vehicle sends the Cell-ID of the current location when we
need. The same network used to send the DTMF (Duel tone multiple frequency) com-
mands from the remote side to the hidden unit. The DTMF command activates pre-
programmed tasks to De-activate the vehicle and for other actions like alarming, photo
capturing, Video recording, central car locking etc. The VLRD also provides the facility of
full duplex communication between driver and remote user. The VLRD is useful in case
of vehicle jacking and lifting and also used in transportation business for finding the loca-
tion for managing the time and to keep eyes on your vehicles.
DTMF is a generic communication term for touch tone (a Registered Trademark of
AT&T). The tones produced when dialing on the keypad on the phone could be used
to represent the digits, and a separate tone is used for each digit.
DTMF dialing uses a keypad with 12/16 buttons. Each key pressed on the phone
generates two tones of specific frequencies. One tone is generated from a high fre-
quency group of tones and the other from low frequency group[3]. The frequencies
generated on pressing different phone keys are shown in the Table 1.
Table 1. DTMF Frequency Assignment
Button Low freq. (Hz) High Freq. (Hz)
1 697 1209
2 697 1336
3 697 1477
4 770 1209
5 770 1336
6 770 1477
7 852 1209
8 852 1336
9 852 1477
0 941 1209
* 941 1336
# 941 1477
84 N. Dubey, V. Dubey, and S. Bande
2 System Design and Methodology
The concept is, whenever a user wants to immobilize or locate the vehicle which is
installed with the VLRD. He just have to make a call on the mobile attached with the
system, when call received he has to dial a password to enter in the system operate
mode.
Then he has to press the key according to which action he wants to perform on the
vehicle either of immobilization or location. If he wants location than mobile sends a
SMS which contains Cell-ID data.
The Cell-ID data contains LAC(Location Area Code), MCC(Mobile Country
Code), NAC(Network Area Code) and Operator code.
The work is basically divided in two parts one is Hardware which is an Embedded
system shown in fig. 1 and second is Software which is a Web based service.
2.1 Hardware Section
The Embedded system is interfaced with the Symbian 60 series of mobile phone here
we used Nokia N-72. The mobile is on auto answer mode so that when a call make
from remote side the call will automatically received.
After call received any key pressed from remote mobile sends a unique DTMF sig-
nal to the system’s mobile according to the Table-1 which will be decoded in BCD
through a DTMF to BCD Decoder CM8870 [5] this decoded commands are than send
to the Microcontroller AT89c51 [6]. The outputs of the Microcontroller acti-
vate/deactivates the relays using relay driver ULN 2003 [7] according to the command
input and the embedded program.
Fig. 1. Hardware System Block Diagram
These relays then perform different actions.
1. The VLRD’s one of the most important works is to immobilize the vehicle and this
task done by the immobilizer panel. The immobilizer immobilizes the vehicle by
many types. The VLRD activates and deactivates some other systems to immobi-
lize the vehicle like deactivating the ignition system, deactivating fuel pump, lock-
ing gear, locking steering, etc.
Cell-ID Based Vehicle Locator and Real-Time Deactivator Using GSM Network 85
2. A video recording facility is available in the VLRD for legal evidence and recogni-
tion of theft/ unknown person.
3. It provides full duplex voice communication facility between driver and the remote
user using car’s audio system.
4. The Project VLRD has an loud alarm system which can be activated remotely
through a mobile phone. The purpose of the alarm is to alert the surrounding per-
sons about the theft car and also useful to locate the car in a particular cell.
5. The VLRD sends the Cell-ID of the particular GSM cell in which the vehicle is
running using Guardian software which is free available on internet. Every time
the user powers up the phone, Guardian automatically starts and check if the
inserted SIM card is present in the authorized list. If this is not authorized the soft-
ware will send a notification SMS to the number previously stored. .[9]
The restarting of system’s mobile is done through microcontroller. The circuit is
shown in Fig. 2.
Cell-ID: The mobile operators keep the locations of GSM masts a secret. Every
GSM mast sends its Cell ID shown in fig. 3, and it is possible to read this Cell ID of
the nearest-by mast on your GSM mobile. By gathering these Cell IDs with accompa-
nying locations and by putting them into a database, it should be possible to read out
Mobile’s current location. With this information it is possible to make location based
applications using GSM Mobile.
Fig. 2. Complete circuit diagram
2.2 Software Section
The Geolocation API is an abstraction for various location APIs that currently exist on
mobile platforms (GPS-based, network/cellid-based). Geolocation implementations
could be straightforward mappings to native APIs (e.g the S60 Location Acquisition
86 N. Dubey, V. Dubey, and S. Bande
API) or have a more complex design that combines several location providers (e.g. a
GPS-based provider and a cell id-based provider) and returns the location from the most
accurate provider at any given time.[8].
The cell-id based location service is a Web based application which locates the
vehicle on Google Map used in Website according to the Cell-ID data entered. It com-
pares the input data to the cell ID database which is free provided by Google. The
Google has a worldwide cell id database which can be fetched by using Google API
on any website.
This database has the Geo co-ordinates of the Cell specified by Cell-ID.
3 Conclusion
A novel method of designing a low-cost, compact theft control system for a vehicle
was designed demonstrated in this paper and also practically implemented by us. This
work is an ultimate threat for vehicle thieves. By this work which is presented in this
paper, it is very easy to track the vehicle at a higher degree of reliability, since it is
based on GSM Technology, which is very developed now.
References
1. Nagaraja, B.G., Rayappa, R., Mahesh, M., Patil, C.M., Manjunath, T.C.: Design & Devel-
opment of a GSM Based Vehicle Theft Control System. In: IEEE International Conference
on Advanced Computer
2. Li, S., Xiong, Z., Li, T.: Distributed Cooperative Design Method and Environment for Em-
bedded System. In: Priceedings of the 9th International Conference on Computer Supported
Cooperative Work in Design, pp. 956–960.
3. Ladwa, T.M., Ladwa, S.M., Kaarthik, R.S., Ranjan, A., Dhara, N.D.: Control of Remote
Domestic System Using DTMF. In: IEEE ICICI-BME 2009 Bandung, Indonesia (2009)
4. Trevisani, E., Vitaletti, A.: Cell-ID location technique, limits and benefits: an experimental
study. In: Proceedings of the Sixth IEEE Workshop on Mobile Computing Systems and
Applications, WMCSA 2004 (2004)
5. CM8870C DTMF Decoder Datasheet,
http://www.calmicro.com/products/data/pdf/cm8870.pdf
6. Microcontroller 89c51 datasheet,
http://www.atmel.com/atmel/acrobat/doc0265.pdf
7. http://pdf1.alldatasheet.co.kr/datasheetpdf/view/25575/STMICR
OELECTRONICS/ULN2003.html
8. http://code.google.com/p/gears/wiki/GeolocationAPI
9. http://www.guardianmobile.com/usage.pdf
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 87–94, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Novel Design of Reconfigurable Architecture for
Multistandard Communication System
T. Suresh1 and K.L. Shunmuganathan2
1 Research Scholar, R.M.K Engineering College,
Anna University, Chennai, India
fiosuresh@yahoo.co.in
2 Professor & Head, Department of CSE, R.M.K Engineering College,
Anna University, Chennai, India
kls_nathan@yahoo.com
Abstract. The goal of future mobile communication systems will be to incorpo-
rate and integrate different wireless access technologies and mobile network
architectures in a complementary manner so as to achieve a seamless wireless
access infrastructure. To support this seamless user mobility across different
wireless access technologies it is needed to design reconfigurable multistandard
receiver architecture. This paper presents the system-level design of a wireless
receiver’s baseband architecture, which supports two wireless access technolo-
gies: Wideband Code Division Multiple Access (WCDMA) and Orthogonal
Frequency Division Multiplexing (OFDM). In this paper, efficient method of
Fast Fourier Transform (FFT) algorithm for OFDM standard and Rake Receiver
design for WCDMA standard were implemented. This architecture efficiently
shares the resources needed for these two standards while reconfiguring. The
proposed architecture is simulated using ModelSimSE v6.5 and mapped onto a
Spartan 3E FPGA device (3s5000epq208) using the tool Xilinx ISE 9.2. Simu-
lation results show that the proposed architecture can be efficiently reconfi-
gured in run-time and proved as area efficient.
Keywords: FPGA, WCDMA, OFDM, FFT/IFFT, Rake Receiver, Reconfigur-
able.
1 Introduction
The need to support several standards in the same handheld device, associated with
the power consumption and area restrictions, created the necessity to develop a porta-
ble, power efficient, integrated solution [1]. Users carrying an integrated open termin-
al can use a wide range of applications provided by multiple wireless networks, and
access to various air interface standards. The continuous evolution of wireless net-
works and the emerging variety of different heterogeneous, wireless network plat-
forms with different properties require integration into a single platform [2]. This has
lead to an increased interest in the design of reconfigurable architecture. The idea of
the reconfigurable architecture is that it should be possible to alter the functionality
of a mobile device at run-time by simply reusing the same hardware for different
88 T. Suresh and K.L. Shunmuganathan
wireless technologies and ultimately for users to connect to any system that happens
to be available at any given time and place.
2 Reconfigurable Architecture
Multistandard wireless communication applications demand high computing power
[3], flexibility, and scalability. An Application-Specific Integrated Circuit (ASIC)
solution would meet the high computing power requirement, but is inflexible [4] and
the long design cycle of ASICs makes them unsuitable for prototyping. On the other
hand, general purpose microprocessors or Digital Signal Processing (DSP) chips are
flexible, but often fail to provide sufficient computing power. Field Programmable
Gate Arrays (FPGAs) [5] signal processing platforms are now widely being accepted
in base_station designs. However, low power and form factor requirements have pre-
vented their use in handsets. Reconfigurable hardware for Digital Base-Band (DBB)
[6] processing is rapidly gaining acceptance in multi-mode handheld devices that
support multiple standards. In Reconfigurable Hardware tasks that are required initial-
ly can be configured in the beginning. When another task is required, the configura-
tion to load it can then be triggered. In this paper we present the design methodology
for reconfigurable baseband signal processor architecture that supports WCDMA and
OFDM wireless LAN standards.
During recent years, a number of research efforts focused on the design of new re-
configurable architectures. In [7] the flexibility of the MONTIUM architecture was
verified by implementing HiperLAN/2 receiver as well as a Bluetooth receiver on the
same architecture. In [8], a broadband mobile transceiver and a hardware architecture
which can be configured to any cyclic-prefix(CP) based system reconfigurable archi-
tecture for multicarrier based CDMA systems is proposed. Reconfigurable Modem
(RM) Architecture targeting 3G multi-standard wireless communication system was
proposed in [3]. This architecture targeted two 3G wireless standards WCDMA and
CDMA 2000 and the design objectives are scalability, low power dissipation and low
circuit complexity. It is seen that though different functions can be reconfigured on a
reconfigurable hardware, the major challenge is to have an efficient system configura-
tion and management function which will initiate and control the reconfiguration as
per the different application requirements.
3 Reconfigurable Receiver Architecture
Figure 1 shows the Block Diagram of Reconfigurable Receiver System. This Receiver
System is able to reconfigure itself to the WCDMA or OFDM Wireless LAN
(WLAN) standard. The Proposed architecture comprises functional blocks, which is
in the form of reusable, reconfigurable [9-12] functional blocks for use in implement-
ing different algorithms necessary for OFDM and WCDMA standards. One or more
reusable functional blocks as given in Fig. 1, can be configured to implement a
process including multiplication, addition, subtraction and accumulation. By
accommodating the above mentioned capabilities, the architecture should be
configured to support WCDMA and WLAN OFDM Standards. For example Fast
Fourier Transform (FFT) (basic butterfly function) for WLAN OFDM and Rake.
A Novel Design of Reconfigurable Architecture 89
Fig. 1. Block Diagram of Reconfigurable Receiver System
Fig. 2. Block Diagram of Reconfigurable Receiver Architecture
Receiver algorithms (multiply and accumulate select function) for WCDMA are
implemented in the architecture as shown in Fig.2. This architecture allows for trans-
formation of the chip from WCDMA chip to WLAN Wi-Fi chip on-demand wherein
new algorithms can be accommodated on-chip in real time via different control sets.
3.1 Rake Finger Implementation
In WCDMA receivers, the demodulation is performed in the Rake fingers by correlat-
ing the received signal with a spreading code over a period corresponding to the
spreading factor. The output of the ith Rake finger can be expressed as
∑
. (1)
where Cs is the combined spreading and scrambling code and R is the received signal
and both are complex numbers [3]. Since the scrambling and spreading codes are
always of +/-1, the multiplication and addition of each correlation stage are simpli-
fied. So the equation(1) is simplified to
(Rr+jRi)(Csr-jCsi) = (RrCsr+ RiCsi)+j(RiCsr- RrCsi) . (2)
Data in Memory
Delay
Demodulation
Rake
finger
FFT
Code
generators
Twiddle
factors
WCDMA
RF
RECEIVER
WLAN RF
RECEIVER
I/P
RECONFIGURABLE
WCDMA/OFDM
DIGITAL BASEBAND
90 T. Suresh and K.L. Shunmuganathan
If the value +1 is represented as logic ‘0’ and the value -1 is represented as logic ‘1’,
the equation (4) is simplified as follows
Rr+Ri+j(Ri-Rr), when Csr=0, Csi=0
Rr-Ri+j(Ri+Rr), when Csr=0, Csi=1
= -(Rr-Ri)-j(Ri-Rr), when Csr=1, Csi=0
-(Rr+Ri)-j(Ri-Rr), when Csr=1, Csi=1 (3)
Since the code input is binary valued, the complex multiplication in the correlations is
simplified to one real addition/subtraction and one imaginary addition/subtraction.
Selection of addition or subtraction is done with the help of multiplexer. So the total
resources required to implement Rake Receiver using (3) are two adders, two subtrac-
tors and one multiplexer.
3.2 FFT Implementation
In OFDM, the demodulation is performed by applying 64-point FFT. The twiddle
factor is calculated and put in a table in order to make the computation easier and can
run simultaneously. The Twiddle Factor table is depending on the number of points
used. During the computation of FFT, this factor does not need to be recalculated
since it can refer to the Twiddle factor table, and thus it saves time. Figure 3 shows
the 2 point Butterfly structure [13] where multiplication is performed with the twiddle
factor after subtraction.
Fig. 3. 2 Point Butterfly Structure
Multiplication is certainly the most vital operation in Communication processing,
and its implementation in an integrated circuit component requires large hardware
resources and significantly affects the size, performance, and power consumption of
a system [14]. So an efficient way of multiplier reduction in FFT processing is done
as follows. Consider the problem of computing the product of two complex numbers
R and W
X = RW = (Rr+jRi)(Wr+jWi)
= (RrWr-RiWi)+j(RrWi+RiWr) (4)
From equation (4), the direct architectural implementation requires total of four mul-
tiplications and one real subtraction and one imaginary addition to compute the com-
plex product. However, by applying the Strength Reduction Transformation we can
reformulate equation (4) as:
-1 WN
b
a A=a+b
B=(a-b)W
N
A Novel Design of Reconfigurable Architecture 91
Xr=(Rr-Ri)Wi+Rr(Wr-Wi) (5a)
Xi=(Rr-Ri)Wi+Ri(Wr+Wi) (5b)
As can be seen from Equations (5a) and (5b), by using the Strength Reduction Trans-
formation the total number of real multiplications is reduced to only three. This how-
ever is at the expense of having two additional subtractors and one adder.
4 Processing Element
Figure 4 shows the Processing Element(PE) and its resources required for the imple-
mentation of FFT in WLAN OFDM and figure 5 shows the Processing Element(PE)
and its resources required for the implementation of Rake finger in WCDMA. It is
shown that the two adders and subtractors(red coloured) are shared by both the stan-
dards. So the proposed Reconfigurable Architecture consists of processing units ,their
computational elements are shared by both the Rake Receiver operation of WCDMA
and FFT operation of OFDM. The processing units perform the multiply-accumulate
operation in the Rake mode as described in section 3.1 and butterfly operations in the
FFT mode as described in section 3.2. The computational resources required by the
proposed architecture are 5 adders, 4 subtractors, 3 multipliers and multiplexers.
Fig. 4. PE and its Resources of WLAN OFDM
sub
mul
add sub
sub sub add
mul mul
add add
add
Registers
92 T. Suresh and K.L. Shunmuganathan
Fig. 5. PE and its Resources of WCDMA
5 Results and Discussion
The proposed reconfigurable architecture described in section 3 and 4 were simulated
using ModelSimSE v6.5 and mapped onto a Spartan 3E FPGA device
(3s5000epq208) with speed grade (-5) using the tool Xilinx ISE 9.2 and synthesized.
The proposed Reconfigurable Architecture with Resource sharing is compared with
Reconfigurable Architecture without Resource sharing. Table 1 and Figure 6 show
the Resources utilized by the proposed Architecture(Reconfigurable Architecture with
Resource sharing) and the Reconfigurable Architecture without Resource sharing.
From the results presented above it seems that there is a significant reduction in large
number of computational resources which forms the proposed architecture which is
more efficient than the conventional Architecture in terms of area.
Table 1. Resource utilization of Reconfigurable Architecture without and with Resource
sharing
Resources Utilized Reconfigurable Architecture
without Resource Sharing
Reconfigurable
Architecture with
Resource Sharing
Number of Slices 1172 out of 4656 (25%) 1070 out of 4656
(23%)
Number of 4 input
LUTs
2195 out of 9312 (23%)
2048 out of 9312
(22%)
Number of IOBs
140 out of 158 (88%)
140 out of 158 (88%)
Registers
add
add/sub
mux
sub
add/sub
A Novel Design of Reconfigurable Architecture 93
Fig. 6. Comparison of the percentage of resources utilized by Reconfigurable Architecture
without Resource Sharing and with Resource Sharing
6 Conclusion
An architecture which can reconfigure itself to wireless LAN OFDM and WCDMA
standards, was presented in this paper. While configuring these two standards, it was
also presented to implement FFT operation for OFDM and Rake Receiver functioning
for WCDMA efficiently. To lower the number of multipliers in FFT and eliminate the
multipliers in Rake Receiver, we adopted Strength Reduction Transformation
technique and multiplier-less technique. The proposed architecture was simulated
using ModelSimSE v6.5 and mapped onto a Xilinx Spartan 3E FPGA device and
synthesis report was generated. Simulation results demonstrated that the proposed
architecture can reduce hardware overhead, enhance circuit efficiency and
significantly reduce area. Moreover, the proposed architecture can be improved to
reconfigure to various other advanced wireless standards.
References
1. Atallah, J.G., Ismail, M.: Future 4G front-ends enabling smooth vertical handover. IEEE
Circuits and Devices Magazine XXII, 6–15 (2006)
2. Liljana Gavrilovska, M., Vladimir Atanasovski, M.: Interoperability in Future Wireless
Communications systems: A Roadmap to 4G. Microwave Review, 19–28 (2007)
3. Lee, J.-S., Ha, D.S.: FleXilicon: a Reconfigurable Architecture for Multimedia and Wire-
less Communications. In: IEEE International Symposium on Circuits and Systems,
pp. 4375–4378 (2006)
4. Kim, J., Ha, D.S.: A New Reconfigurable Modem Architecture for 3G Multi-Standard
Wireless Communication Systems. In: IEEE International Symposium on Circuits and
Systems, pp. 1051–1054 (2005)
5. David, R., Chillet, D., Pillement, S., Sentieys, O.: A compilation framework for a dynami-
cally reconfigurable architecture. In: Glesner, M., Zipf, P., Renovell, M. (eds.) FPL 2002.
LNCS, vol. 2438, pp. 153–194. Springer, Heidelberg (2002)
0
20
40
60
80
100
No. of
Slices
No.of LUTs No. of IOBs
Reconfigurable
Architecture without
Resource sharing
Reconfigurable
Architecture with
Resource sharing
94 T. Suresh and K.L. Shunmuganathan
6. Harju, L., Nurmi, J.: A Programmable Baseband Receiver Platform for WCDMA/OFDM
Mobile Terminals. In: IEEE Wireless Communications and Networking Conference, USA,
pp. 33–38 (2005)
7. Rauwerda, G.K., Smit, G.J.M., van Hoesel, L.F.W., Heysters, P.M.: Mapping Wireless
Communication Algorithms to a Reconfigurable Architecture. The Journal of Supercom-
puting 30, 263–282 (2004)
8. Liang, Y.-C., Naveen, S., Pilakkat, S.K., Marath, A.K.: Reconfigurable Signal Processing
and Hardware Architecture for Broadband Wireless Communications. EURASIP Journal
on Wireless Communications and Networking 3, 323–332 (2005)
9. Hauck, S., Fry, T.W., Hosler, M.M., Ko, J.P.: The Chimaera Reconfigurable Functional
Unit. IEEE Transactions on Very Large Scale Integration(VLSI) Systems 12(2), 206–217
(2004)
10. Parizi, H., Niktash, A., Kamalizad, A., Bagherzadeh, N.: A Reconfigurable Architecture
for Wireless Communication Systems. In: Third International Conference on Information
Technology: New Generations, pp. 250–255 (2006)
11. Hartenstein, R.: Coarse Grain Reconfigurable Architectures. In: Conference on Asia South
Pacific Design Automation, pp. 564–570 (2001)
12. Qu, Y., Tiensyrj, K., Soininen, J.-P., Nurmi, J.: Design Flow Instantiation for Run-Time
Reconfigurable Systems: A Case Study. EURASIP Journal on Embedded Systems 11
(2008)
13. Heysters, P., Smit, G., Molenkamp, E.: A Flexible and Energy-Efficient Coarse- Grained
Reconfigurable Architecture for Mobile Systems. The Journal of Supercomputing 26,
283–308 (2003)
14. Hinkelmann, H., Zipf, P., Li, J., Liu, G., Glesner, M.: On the design of reconfigurable mul-
tipliers for integer and Galois field multiplication. Microprocessors & Microsystems 33(1),
2–12 (2009)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 95–99, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Two Novel Long-Tail Pair Based Second Generation
Current Conveyors (CCII)
Amisha Naik1 and N.M. Devashrayee2
1 Asst. Prof, EC Department, Institute of Technology, Ahmedabad-382481, Gujarat, India
a_p_niak@yahoo.com
2 Asso. Prof, EC Department, Institute of Technology, Ahmedabad-382481, Gujarat, India
nalin_deepika@yahoo.com
Abstract. Two novel long tail pair based CCII are proposed in this paper. The
first OTA CCII offer 1.98GHZ current transfer bandwidth and 10MHZ voltage
transfer bandwidth. The second proposed design of CCII is independent against
bias current variation. The second proposed design offer voltage transfer
bandwidth of 10MHZ and current transfer bandwidth of 1.2GHZ with a very
accurate voltage and current copy at corresponding x and Z nodes. The none
idealities of first CCII are also measured using Spice simulation with TSMC
180nm model parameters.
Keywords: OTA CCII, Voltage transfer bandwidth, Current transfer band-
width.
1 Introduction
The current-mode approach considers the information flowing on time-varying currents.
Current-mode techniques are characterized by signals as typically processed in the cur-
rent domain. The current-mode approach is also powerful if we consider that all the
analog IC functions, which are traditionally designed in the voltage-mode, can also be
implemented in current-mode. In voltage mode circuits, the main building block used to
add subtract, amplify, attenuate, and filter voltage signals is the operational amplifier. A
current-mode approach is not just restricted to current processing, but also offers certain
important advantages when interfaced to voltage-mode circuits. Since the introduction
of conveyors in early 70’s, lot of research has been carried out to prove usefulness of
this CCII. The CCII is a functionally flexible and versatile, rapidly gaining acceptance
as both a theoretical and practical building block. Internal architecture of CCII is voltage
follower cascaded with current follower as shown in figure:1.
2 OTA CCII
The simple current Mirror based CCII can be improved by replacing Simple diode
connected level shifter in a flipped voltage follower based simple current mirror CCII
96 A. Naik and N.M. Devashrayee
Fig. 1. CCII Block Representation
proposed by A.J.Lopez martin etal by OTA level shifter. The resulting circuit is de-
picted in figure:2. The proposed circuit is a combination of a OTA and FVF cell.In
the proposed circuit(fig:2), transistors M1 to M4 form differential pair. Transistor M7
and M8 forms Flipped voltage follower. The X terminal voltage is connected to the
gate of M2 transistor which is controlled by source voltage of M7.So,transistor M1-
M4 and from terminal a through M7 forms a level shifter, the voltages at X and Y
terminals follow each other. The drain voltage of M7 is also used to control M8 and
M11 transistors. So, the current flowing at X terminal is conveyed to Z terminal via
M8.Here M7 and M8 ensure low resistance at X node. While Y input is at gate of M1
which gives very high input impedance at Y input. The circuits of proposed CCII in
figure 2 was simulated using 0.18 µm CMOS technology with NMOS and PMOS
threshold voltages of approximately 0.4 V and -0.39v. The Transistor aspect ratios are
shown in Table 1. Bias voltage was +/-1 V, and bias current IB was 70 µA. First, its
time response was evaluated by configuring the the circuit as unity-gain voltage am-
plifiers. In order to do so, ports X and Z were loaded with 15 kΩ resistances. The in-
put voltage, a 100KHZ, 100 mVpp, sinusoid. was applied to the Y port Figure2 and
the result is tabulated in Table:2 The AC small-signal frequency response for the cir-
cuit is subsequently obtained, using the same load resistors. The simple structure of
Figure 2 has a unity bandwidth of 100 MHz,as expected. Table 2 also compares simu-
lation results of the proposed circuit with low-voltage current conveyor reported in
the literature[6]. The advantages in terms of power dissipation ,offset and compact-
ness at reduced circuit complexity can be clearly evidenced.
Fig. 2. OTA CCII based on differential pair
Two Novel Long-Tail Pair Based Second Generation Current Conveyors (CCII) 97
Table 1. Aspect Ratio for Fig:2
Transistor W/L
M1,M2,M7 30u/.9u
M3,M4,M9,M10 15u/.9u
M5,M6,M8,M10 13.5u/2.7u
The small signal terminal impedances at X,Y and Z node is as follows.
7
8|| 9
177
ro
Z
xro ro
gm ro
⎡
⎤
⎛⎞
=+
⎢
⎥
⎜⎟
+
⎝⎠
⎣
⎦,***
Z
yWLcox
γ
=,10 || 11
R
zro ro=,
The small signal ratios Vx/Vy and Iz/Ix is given as follows. The non-idealities of
CCII shown in figure is measured using TSMC 180nm model parameters using spice
simulation. The non ideal matrix is shown below.
07 7( / 2) 1
177(1(/2)2)
Vx r gm ro gm
Vy ro gm ro gm
=++ ,( 11* 11* 08)
( 11 10)(1 8 8)
Iz gm ro r
Ix ro ro gm ro
=++
3 Cascode OTA and FVF Based CCII
The problem with the above CCII is the output offset is a function of Ib , technology
parameters and input voltage. The mathematically it is given by,
Table 2. Simulation Results For OTA CCII
Characteristic
Parameters
Proposed
FV mirror based Arch-1 A.J.L.Martin etal
Voltage Supply +/1V 1.5V
Power Cons. 0.58mwatt 0.75mwatt
Iz/Ix transfer BW 1.98Ghz 20Mhz
Vx/Vy transfer Bw 10Mhz 100Mhz
IBIAS 90uA 100uA
Offset -23.72mv 300mv
Iz / Ix 1 1.1
Vx / Vy 1 1
Y Para. Imp. 6GΩ 80k
X Para. imp 1.4kΩ 10KΩ
Z Para. imp. 300kΩ 11KΩ
THD 1.21% @ 100Mhz 1%
-7 -3
-3
0.33x10 0 0.195x10
11.8k0.310
0.09 1 0.005x10
Iy Vy
Vx Ix
Iz Vz
⎡⎤
⎡⎤ ⎡⎤
⎢⎥
⎢⎥ ⎢⎥
=⎢⎥
⎢⎥ ⎢⎥
⎢⎥
⎢⎥ ⎢⎥
±
⎣⎦ ⎣⎦
⎦
⎣
98 A. Naik and N.M. Devashrayee
[]
12
2b
offset DS DS
I
VVV
β
=−
The variation with Ib is plotted infigure:3 shows that
the offset varies with biasing current Ib. From figure it is clear that as biasing current
increases from 10mA to 170mA the output offset varies from -400mv to 0v.The cir-
cuit can be made independent from biasing current by adding one more pair of pMOS
current mirror load on the top of the pMOS current mirror load in OTA based CCII in
figure:2.The resulting CCII is shown figure:4 .
Table 3. Simulation Results For Cascoded OTA CCII
Characteristic
Parameters
Proposed
FV mirror based
Arch-1
Cascoded OTA
CCII A.J.L.Martin
etal
Voltage Supply +/1V +/-1.25V 1.5V
Power Cons. 0.354mwatt 0.9mV 0.75mwatt
current transfer bandwidth 2Ghz
1.2Ghz 100Mhz
IBIAS 70uA 60uA 100uA
offset Function of Ibias Independent of
Ibias 300mv
Iz / Ix 1 1 1.1
Vx / Vy 1 1 1
Fig. 3. Plot of biasing current dependency on output offset of OTA CCII in figure:2
5 Conclusion
Two novel CCII topologies are proposed simulated and compared with the present
state of art design. The topologies are very compact, low power and wideband. The
Two Novel Long-Tail Pair Based Second Generation Current Conveyors (CCII) 99
second topology is a high precision with zero offset and independent of biasing cur-
rent. The results are tabulated in Table-2 and 3.
Fig. 4. OTA CCII based on differential pair
References
[1] Rajput, S.S., Jamuar, S.S.: Low voltage, low power, high performance current conveyors.
In: Proc. ISCAS 2001, Sydney, pp. 1123–1726 (May 2001)
[2] Ramirez-Angulo, J., Carvajal, R.G., Torraiba, A., Galan, A., Vega-Leal, A.P., Tombs, I.:
The flipped voltage follower: a useful cell for low-voltage low-power circuit design. In:
Proc. ISCAS 2002, Phoenix, AZ, pp. Ill 615–Ill 618 (May 2002)
[3] Lopez Martin, A.J., Ramirez-Angulo, J., Carvajal: Low voltage Low power wideband
CMOS current conveyors based on a flipped voltage follower. Proc. IEEE, Ill 801– Ill 804
(2003)
[4] Lopez Martin, A.J., Angulo Sheetal Gupta, J.R., Carrvajal, R.G.: Comparison of conven-
tional and New flipped voltage structure with increased input and output signal swing and
current sourcing /Sinking capacity. IEEE proceedings (2005)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 100–105, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Generating Testcases for Concurrent Systems Using
UML State Chart Diagram
Debashree Patnaik1, Arup Abhinna Acharya1, and Durga P. Mohapatra2
1 School of Computer Engineering
KIIT University
debashree.patnaik@gmail.com, arupacharya.kiit@gmail.com
2 Department of Computer Science and Engineering
National Institute of Technology, Rourkela
durga@nitrkl.ac.in
Abstract. Communication and concurrency are the major factors needed for the
construction of a concurrent system. In concurrent environment systematic test-
ing becomes a complex task. Generating test cases in concurrent environment is
a difficult task because of arbitrary meddling of the concurrent thread. The in-
terference of the concurrent thread may lead to a deadlock. In this paper we
propose a methodology to generate the test cases for the conformance of dead-
lock in concurrent systems using UML State Chart Diagram. For system speci-
fication we have used the UML State Chart Diagram, from which event tree is
generated. Along with a case study an algorithm is proposed to generate the test
suite and confirm whether it is free from deadlock.
Keywords: Concurrency, State Chart Diagram, Deadlock, Event Tree, Test
Sequences.
1 Introduction
The design of the concurrent software system is a complex activity leading to issues
in testing. Environment is distributed means there are many machines distributed over
the network which runs concurrently. In object-oriented system, objects become the
means of communication. In concurrent execution environment we have multiple
objects communicating concurrently [1]. In one class we can have many objects
which can be active at the same time behaving differently from each other (multiple
role playing). The concurrency of the objects has to be properly tested to establish the
desired confidence in the functionality of any object. Testing the concurrent object-
oriented system has to deal with the complexities arising from the physical distribu-
tion and parallel execution of the objects. An object has a definite state at a particular
instant of time [1]. The state of the object changes due to the reaction of particular
event. So event becomes the basic unit to observe the behavior of the object. States do
not qualify to become the basic unit of observation because of the possibility of being
complex and containing complex sub-states.
The main challenges in designing concurrent programs are ensuring the correct se-
quence of interactions between different computational processes, and coordinating
the access of resources that are shared among processes [3].
Generating Testcases for Concurrent Systems Using UML State Chart Diagram 101
The rest of the paper is organized as follows: Section-2 discusses deadlock issue in
concurrent system. Section-3 describes deadlock analysis in UML State Chart. The
proposed model is discussed in Section 4. Section-5 discusses the Future work. And
finally Section-6 concludes the paper.
2 Deadlock: An Issue in Concurrent System
There are many challenges for a concurrent system. We have a number of different
processes running together at a time. To synchronize and coordinate between the
processes is very contending.
A deadlock is a situation where two or more competing actions are waiting for
each other to finish and neither of the processes finishes [5]. It is a lock having no
keys. Deadlock refers to a specific condition when two or more processes are each
waiting for each other to release a resource, or more than two processes are waiting
for resources in a circular chain. Deadlock is a common problem in multiprocessing
where many processes share a specific type of mutually exclusive resource. Comput-
ers intended for the time-sharing and/or real-time markets are often equipped with a
hardware lock guarantying exclusive access to processes, forcing serialized access.
There is no general solution to avoid deadlocks.
3 Deadlock Analysis in UML State Chart Diagram
Deadlock is a common problem in multiprocessing where many processes share a
specific type of mutually exclusive resource known as a software lock or soft lock.
Sometimes, two or more transactions in their course of operation may attempt to
access the same table or resource in a database. This prevents both the transactions
from proceeding forward. This is called a Deadlock.
Fig. 1. Concurrency Leading To Deadlock In A Banking System
3.1 Deadlock in State Chart
In FIG.1 deadlocks is represented in terms of State Chart Diagram. Deposit and
Balance Enquiry are the two states. Customer A, Customer B, Bank account, Bank
database, ATM, Customer Details are the sub-states. When the event is fired from one
state it reaches another state, it is called as transition. In deposit the Customer A is in
initial state, it wants to carry out the deposit action. When certain sum of money is to
be deposited by the customer, first we verify whether the account is a valid account. If
the account is a valid then the bank database is accessed and then the customer details
are searched. On the other hand in balance enquiry customer B wants to find a mini
102 D. Patnaik, A.A. Acharya, and D.P. Mohapatra
statement of his account. He inserts the card into the ATM and types the pin. The pin
number is verified. If the pin code is OK then the database is accessed and the cus-
tomer details are searched. There may be an instant where the Deposit and the Bal-
ance Enquiry are trying to access the same customer details at same instant of time. In
this case deadlock occurs; the two actions try to take place concurrently.
4 Proposed Method
In this paper the authors have proposed a model to generate Test cases, for a concur-
rent system. The main issue for a concurrent system is deadlock. The concurrent sys-
tems are modeled by using UML State Chart Diagram from which event tree are gen-
erated. Finally event trees are traversed to generate the test cases.
4.1 Creation of Event Tree
Using Chow’s Algorithm [1] the authors have generated the event tree for individual
State Charts Diagram. In Fig.2 Deposit and Balance Enquiry are the generated event
trees from the State Chart Diagram. Even though both sub states run concurrently
they are represented in different graphs [6]. To generate the test sequence we combine
both the event tree, and generate a graph for the whole system and then find the test
Suite for the scenario. In the system graph, we have to take all possible scenarios and
generate the test cases.
Fig. 2. Event tree of two sub-states (Deposit and Balance Enquiry) D-DEPOSIT, A-ACCESS, S-
SEARCH, I-C-INSERTCARD
4.2 Traversing Event Graph
With the backtracking algorithm we make an effort to visit every node. All nodes are
initialized with value 0, when a node is discovered the value of the node changes to 1.
If the node is visited then the value of the node is changed to 2. If all the nodes are
traversed then finally we get a graph with value 2. If there is deadlock as shown in
Fig.3, then the transactions are rolled back.
We have applied backtracking algorithm for traversal of the tree and generation of
Test Cases and Test Sequences.
V= {account, withd, aaccess, baccess, search,p_val,n_bal,open,close,found,
n_found}
Node status= {0, 1, 2}
0 – Nodes that are Unexplored, 1 - Nodes that are discovered, 2 – Nodes that are
explored.
Generating Testcases for Concurrent Systems Using UML State Chart Diagram 103
ALGORITHM: BACKTRACKING_ALGORITHM
INPUT: Event Tree
OUTPUT: Test Sequences
Step1: Start
Step2: For each u that belongs to V [G]
Value [u] = 0
Pie[u] = nil // records the event by setting u’s predecessor field pie[u]
time Å 0
// all the vertices are initialized to value 0 and their pie fields to nil. Time global counter) is set to 0.
Step3: For each u
If (value [u] = = 0)
DFS visit (u) // Depth First Search
Value[u] = = 1
time Å time + 1
//When a vertex with value 0 is found it visits using the DFS visit. Vertex is discovered and
set to value 1, global counter is incremented.
Step4: Find each vertex V which is adjacent to u.
D[v] Å u
If (value [u] == 0) || (reaches the last node)
Pie[v] Åu
DFS visit (v)
Value[u] Å 2
time Åtime + 1
//The discovered node are explored and while leaving the vertex the value of the vertex are set
to 2, with the increment of the global counter).
Step5: Stop
Fig. 3. Event Graph representing Deadlock Fig. 4. Event graph free from deadlock
Account(Root node)- as it enhances a situation to enter into the State Chart Dia-
gram, Withd-Withdrawal, insert-inserting the ATM card, access-accessing the bank
database, Search-Searching the required account number,. cust_details- accessing
the same customer details, p_val- permitted value,n_val-no balance, open-account is
active, close-account is inactive, found-customer details is found, n_fou-account is
not found.
4.3 Generation of Test Sequences
The test sequences are generated by the backtracking traversal algorithm. In Fig.4 the
system is free from deadlock and the test sequences generated are given in table 1. In
104 D. Patnaik, A.A. Acharya, and D.P. Mohapatra
Fig. 3 the system shows a deadlock event as at same instant of time the access to cus-
tomer details is required.
Table 1. Test Sequences Generated for Conformance of Deadlock
Test Sequence #1
Test Sequence #2
Test Sequence #3
Test Sequence #4
withd
aaccess
search
p_val
withd
aaccess
search
n_val
insert
baccess
search
open
found
insert
baccess
search
close
n_fou
4.4 Generation of Test Cases
The test cases generated using the proposed methodology is listed in Table 2. As
deadlock is a major concern in concurrent systems, only a subset of test cases are
generated are shown here to address this issue.
Table 2. Generated Test Cases
Step # Action Performed Expected Results
1
Card is inserted in the ATM
machine
The ATM prompts us to enter the 4 digit
PIN/Transaction code
2
Key in the 4 digit
PIN/Transaction code
On entering the PIN code the System starts a
action or transaction for checking the Customer
details and account details
3
System searches for the Customer
details attached to this card Correct Customer is found from the customer table
4
Once the PIN is validated. The
system prompts for the required
action, we select the "Balance
Inquire"
On selecting Balance Inquire the system searches
for the Account details in the Account Details
table
5
The system checks for the status
of the “Account Details” table
If another Banking transaction is trying to
simultaneously access the “Account Details” table,
a Deadlock situation is faced by both of the
transactions and our transaction is ends abruptly
giving a message as “Your transaction cannot be
processed currently. Please try again later”.
If the “Account Status” table is available/free for
the transaction then the “Account Search” action is
initiated.
5 Future Work
The deadlock analysis can be determined with the help of a mathematical tool called
as petrinets. Both static and dynamic analysis can be determined by petrinets concur-
rently. Graphs can be constructed avoiding the interleaving. The authors intend to
Generating Testcases for Concurrent Systems Using UML State Chart Diagram 105
analyze and explore on the idea we have established in a wider range of the distri-
buted systems.
6 Conclusion
Problems that arise from concurrency and deadlock are discussed in context to distri-
buted environment. The paper emphasizes on testing the behavioral aspects of the
distributed objects using events to stimulate the object’s behavior. A method of dead-
lock analysis is presented through UML State chart diagrams. From the State Chart
diagram an event tree is generated with event as node. The event tree is further trans-
formed to an event graph generating the test cases and the test sequences. For the tra-
versal of the event graph the authors have implemented the back-tacking algorithm.
References
[1] Adnan Bader, A.S.M., Sajeev, S.R.: Testing concurrency and communication in distributed
objects
[2] Asaadi, H.R., Khosravi, R., Mousavi, M., Noroozi, N.: Towards Model-Based Testing of
Electronic Funds Transfer Systems
[3] Hessel, A., Larsen, K.G., Mikucionis, M., Nielsen, B., Pettersson, P., Skou, A.: Testing
real-time systems using UPPAAL. In: Hierons, R.M., Bowen, J.P., Harman, M. (eds.)
FORTEST. LNCS, vol. 4949, pp. 77–117. Springer, Heidelberg (2008)
[4] Hierons, R.M., Bogdanov, K., Bowen, J.P., Cleaveland, R., Derrick, J., Dick, J., Gheorghe,
M., Harman, M., Kapoor, K., Krause, P., Lüttgen, G., Simons, A.J.H., Vilkomir, S.A.,
Woodward, M.R., Zedan, H.: Using formal specifications to support testing. ACM Com-
puting Surveys 41(2) (2009)
[5] Koopman, P.W.M., Plasmeijer, R.: Testing reactive systems with GAST. In: Post-
Proceedings of TFP 2003, Intellect, pp. 111–129 (2003)
[6] Mikucionis, M., Nielsen, B., Larsen, K.G.: Real-time system testing on-the-fly. In: Pro-
ceedings of NWPT, pp. 36–38 (2003)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 106–110, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Intelligent Agent Based Resource Sharing in Grid
Computing
V.V. Srinivas1 and V.V. Varadhan2
1 Department of Computer Science and Engineering,
National Institute of Technology – Tiruchirappalli
srinivas15j1988@gmail.com
2 Department of Information Technology,
Madras Institte of Technology - Chennai
varadhan3n90@gmail.com
Abstract. Most of the resource present in grid are underutilized these days.
Therefore one of the most important issue is the best utilization of grid resource
based on users request. The architecture of intelligent agent proposed to handle
this issue consists of four main parts. We discuss the need and functionality of
such an agent and propose a solution for resource sharing which satisfies prob-
lems faced by today’s grid. A J2EE based solution is developed as a proof of
concept for the proposed technique. This paper addresses issues such as re-
source discovery, performance, security and decentralized resource sharing
which are of concern in current grid environment.
Keywords: grid, resource sharing, intelligent agent, decentralization.
1 Introduction
Grid computing is distributed, large-scale cluster computing, as well as a form of
network-distributed parallel processing. Each computer present in grid has computa-
tional power and resources such as memory, printer etc., which are underutilized. In
order to utilize resources and provide service to customers resource sharing was in-
troduced. Resource sharing [1] provides access to a particular resource on a computer
to be accessed by clients on grid. The need for resource sharing arises in case of com-
plex mathematical modeling and simulations like the network simulation or simula-
tion of automatic test pattern algorithms, virtual supercomputing or DNA mapping.
Resource sharing involves three main process namely: resource discovery, resource
management and resource allocation. Resource discovery is finding resources availa-
ble in grid. This problem is solved in tools such as Globus and Condor Matchmaker
[2]. Resource management involves collecting resource. The challenge involved is
finding the right quantity of resource [3].
1.1 Contribution
There are a number of ongoing research in the field of grid computing most of them
trying to address some of the challenges faced in grid environment. This paper
Intelligent Agent Based Resource Sharing in Grid Computing 107
addresses aspects like: resource discovery in grid, security, decentralized resource
sharing, scheduling of resources and threshold based resource allocation.
The rest of this paper is organized as follows: Section 2 discusses ongoing research
followed by section 3 discussing proposed architecture. Section 4 describes experi-
mental results. In section 5, the paper deals with application and advantages followed
by future research and conclusion.
Fig. 1. Resource discovery Fig. 2. Grid model of entire system
2 Ongoing Research
2.1 Globus Toolkit
Globus is an open source toolkit that is used for construction of grids. It provides
access to resource present within the grid located in any geographical area [5]. Globus
uses GSI (Grid Security Infrastructure), GRAM (Grid Resource Allocation and Man-
agement) and MDS (Monitoring and Discovery Services) for resource management
and discovery. Globus provides a remote front end to multiple batch systems. Our
paper tries to preserve all the advantages of Globus and eliminates all complex confi-
guration and installation of number of tools.
2.2 Meta-broker Architecture
The meta-broker architecture focuses on how to allocate a particular resource present in
some other network or grid to a user requesting for resource [6]. Previous works deals
with MESS (Multi Engine Search Services) and ISS (Internet Search Service) based on
CORBA [7] and meta-broker architecture for management of grid resources [8].
2.3 Negotiation Algorithms
This work mainly deals with negotiation protocols between the client and provider.
The key focus is on contracts [9]. One important algorithm used is G-Negotiation
algorithm. Our paper discusses a simple mechanism for secure communication.
108 V.V. Srinivas and V.V. Varadhan
3 Proposed Architecture
The entire work is split up into several modules which are discussed in detail. First is
the server module in which client registers to provide resource. The client is provided
with an address. The details registered include resource type, amount of resource and
time duration when the resource would be available. Resource discovery [4] module
is used to keep check on parameters such as processor time, print queue, system
threads, disk queue length and cpu usage. The values are obtained from performance
logs and alerts. The retrieved values are stored in .csv format. These obtained values
Fig. 3. Four program instances run on 2, 3, 6 and 8 nodes. X axis represents injection rate and
Y axis represents latency. Buffer size are taken as 4 and 6.
Intelligent Agent Based Resource Sharing in Grid Computing 109
Fig. 4. Performance analysis in terms of execution time
Fig. 5. Four program instances run on 2, 3, 6 and 8 nodes with buffer size are taken as 8 and 12
110 V.V. Srinivas and V.V. Varadhan
are compared with threshold values. The third is the intelligent agent module which reads
the files obtained from the resource discovery module running on each machine. The re-
source files are set to refresh after tk time. The resource requester requests the agent for a
particular resource. The intelligent agent searches and retrieves the various resources it has
from various clients. Once the right resource is obtained, a secret key is transmitted to the
resource requester and provider. Along with the secret key, the resource requester receives
the providers address and vice-versa. This ensures security.
4 Experimental Results
The performance was evaluated between the number of process and execution time. We
took the number of resource providers in the grid to be 4 and the number of process was
taken as multiples of 4. A graph was plotted for the process running on single system to
the process split among the four systems and the resultant graph is shown in Fig. 4.
5 Conclusion
Our paper discusses a solution for resource sharing at the same time preserving fea-
tures like security, authentication, resource discovery and decentralization. This paper
defines a simple solution to implement intelligent agent in grid environment.
References
1. Cruz-Perez, F.A., Ortigoza-Guerrero, L.: Equal resource sharing allocation with QoS diffe-
rentiation conversational services in wireless communication networks. IEEE Proceedings
Communications, 150, 391–398 (2003)
2. De Smet, A.: Computer Science Department. University of Wisconsin Madison,
http://www.cs.wisc.edu/condor
3. Li, Y., Wolf, L.: Adaptive Resource Management in Active Nodes. In: 8th IEEE Interna-
tional Symposium on Computer and Communication (2003)
4. Giovanni, A., Massimo, C., Italo, E., Maria, M., Silvia, M.: Resource and Service Discov-
ery in the iGrid Information Service. In: Gervasi, O., Gavrilova, M.L., Kumar, V., Laganá,
A., Lee, H.P., Mun, Y., Taniar, D., Tan, C.J.K. (eds.) ICCSA 2005. LNCS, vol. 3482,
pp. 1–9. Springer, Heidelberg (2005)
5. Ian, F., Carl, F.: The Globus project: A Status Report. In: Heterogeneous Computing
Workshop, pp. 4–18 (1998)
6. Kertesz, A., Kaeushk, P.: Meta-broker for future generation grids: A new approach for
high level inter operable resource management. In: CoreGrid workshop (2007)
7. Yue-Shan, C., Hsin-Chun, H., Shyan-Ming, Y., Lo, W.: An agent based search engine
based on Internet search service on CORBA. In: Proceedings of International Symposium
on Distributed Objects and Applications, pp. 26–33 (1999)
8. Kertesz, A., Kacsuk, P.: Grid Interoperability Solutions in Grid Resource Management.
Systems Journal 3, 131–141 (2009)
9. Antoine, P., Phelipp, W., Oliver, W., Wolfgang, Z., Dynamic, S.L.A.: negotiation based on
WS agreement. In: Core Grid Technical Report TR-0082 (2007)
10. Karl, W.: The Management of Change among Loosely Coupled Elements. In: Making
Sense of the Organization (1982)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 111–116, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Wideband Miniaturized Patch Antenna Design and
Comparative Analysis
Sanket Patel1,*, Yogeshwar Kosta2, Himanshu Soni1, and Shobhit Patel2
1 G.H. Patel College of Engineering and Technology, V.V. Nagar, Gujarat, India
sanket_patel_2020@yahoo.co.in
2 Charotar University of Science and Technology, Changa, Gujarat, India
Abstract. A novel kind of miniaturized wideband patch antenna is designed
and comparative analysis is presented. The antenna is having size of 2.1cm X
2.1cm X 1.25cm and patch is having area of 153mm2. Single or dual bands fre-
quency response can be obtained by varying the feed location, height of the
substrate and the geometric specifications of the antenna. This proposed an-
tenna is very small in size although it provides good impedance behavior, return
loss S11 behavior and VSWR which is very much nearer to 1 at each band.
Bandwidths up to more than 56% can be obtained. Far-field radiation pattern
and field distributions on the coaxial probe feed patch have been analyzed.
Keywords: Dual Bands, Return Loss, VSWR.
1 Introduction
Microstrip antennas have gained extensive applications in recent times due to their
light weight, small size, easy reproduction and integration ability with the circuitry
[1-10]. There is a tremendous growth in demand for wireless RF systems in applica-
tions such as local area networks, point-to-point communications and applications in
medical and industrial sectors. By time patch structure is modified to have application
specific resonating frequencies and to have higher gain and bandwidth response.
[11-13]. Rectangular, circular and triangular patch geometries are the most exten-
sively analyzed antenna geometries in recent years and now these geometries are
modified to improve their performance [14-16]. Lu [15] analyzed a circular patch
antenna and its arrays with a pair of L-shaped slots for broadband dual-frequency
operation. Wong and Hsu [16] applied a V shaped slot in an equilateral triangular
microstrip antenna. In this article, novel patch antenna is designed and comparative
analysis is done with the normal patch antenna. By varying proposed antenna parame-
ters one can have single or dual frequency bands of operation. Antenna provides en-
hanced bandwidth. Comparison is done based on the geometric specifications as well
as in terms of responses.
* Sanket S. Patel is currently pursuing M.E. in Communication Engineering from G.H. Patel
College of Engineering and Technology, Gujarat Technological University.
112 S. Patel et al.
2 Antenna Design
The geometry of the antenna is shown in Fig.1. The antenna parameters are also given
in Fig.1. The antenna is mounted on a duroid (tm) substrate and fed by a coaxial
transmission line.
Simulations were performed using HFSS™[17]. HFSS (High Frequency Struc-
ture Simulator), is the industry-standard simulation tool for 3D full-wave electromag-
netic field simulation. HFSS provides E and H-fields, currents, S-parameters and near
and far radiated field results. It integrates simulation, visualization, solid modeling,
and automation. Ansoft HFSS employs the Finite Element Method (FEM) for EM
simulation by developing/ implementing technologies such as tangential vector finite
elements and adaptive meshing.
In this proposed antenna patch and ground plane are made of copper having rela-
tive permittivity as 1. The substrate material duroid (tm) has relative permittivity as
2.2 and dielectric loss tangent 0.0009.
Convergence was tested for each case separately in terms of evaluating S11 (dB) at
selected frequencies for a number of times. Once convergence was obtained, simula-
tions were conducted in order to obtain swept frequency response extending from 1 to
10 GHz. The swept response gave the S11, which was used to calculate the VSWR.
After that radiation pattern was computed. Here three different designs are presented
and analyzed.
Fig. 1. Antenna Geometrical Configuration. Parameters are: sub_w=21mm, sub_L=21mm,
sub_h=12.5mm, l=3mm.
Wideband Miniaturized Patch Antenna Design and Comparative Analysis 113
3 Comparative Analysis
The tabular analysis is given in the Table 1. Table compares the proposed miniaturized
antenna with the conventional design. Comparison is done considering geometrical
specifications. Table narrates that the dimensions of the proposed antenna are very
much smaller then the conventional antenna. Proposed antenna has the substrate area is
about 20 times smaller then the conventional antenna. It should be noted that patch area
is about 8 times smaller then the conventional design results in low conducting material
requirement and so low cost. Volume of the proposed antenna is 5 times lesser then the
conventional patch antenna so proposed antenna can be encapsulated even in small size
instruments. Substrate dielectric material used for the both the antennas are having same
relative permittivity as 2.2 and same dielectric loss tangent 0.0009.
Considering specific feed locations the return loss S11 behavior of both the anten-
nas are compared. Conventional antenna provides single band of operation and
%bandwidth at 10dB is 2.54% (60MHz). While in second proposed antenna, dual
band of operation is seen respective to feed locations. For feed location
(0.7mm,0.7mm) achieved dual bands have the %BWs of 40.30%(730MHz) and
5.43%(500MHz). For feed location (2,0) achieved dual bands have the %BWs of
40.30%(730MHz) and 7.26%(660MHz). So the antenna can be used for UWB an-
tenna. Corresponding VSWR plots are compared, and VSWR much nearer to 1 is
achieved at each frequency band. Far Field Radiation Pattern for specific angles is
compared. Pattern depicts the directivity of the antennas. Total Gain (3D Polar Plot) is
generated and is matching with the directivity 2D plots as gain is directly proportional
to directivity. It should be noted that here in both antennas the material used for the
patch and for the ground plane is the copper. Mesh refinement and E-field distribution
on the patch for both the antennas are also depicted in Table 1.
Table 1. Comparative analysis of conventional antenna and proposed antenna
Parameter Conventional Design Proposed design
Top View
Front View
Dimensions Sub_L=100mm, sub_w=90mm,
L=40mm, w=30mm, sub_h=3.2mm
Sub_L=21mm, sub_w=21mm,
sub_h=12.5mm, l=3mm
Ground Plane
area 9000mm2 441mm2
114 S. Patel et al.
Table 1 (continued)
Patch Area 1200mm2 153mm2
Volume of
Antenna 28.8cm3 5.5125cm3
Dielectric
Material of
the substrate
Material: Rogers RT/duroid 5880
(tm)
Relative Permittivity 2.2
Loss tangent 0.0009
Material: duroid (tm)
Relative Permittivity 2.2
Loss tangent 0.0009
Feed
Location
(x,y)
(-0.5mm,0mm) Case:1 (0.7mm,0.7mm)
Case:2 (2mm,0mm)
Return Loss
S11
1.00 1.50 2.00 2.50 3.00 3.50
Freq [GHz]
-35.00
-30.00
-25.00
-20.00
-15.00
-10.00
-5.00
0.00
dB(St(coax_pin_T1,coax_pin_T1))
Ansoft Corporation HFSSDes ig n1
S11 Return Loss
m1
m2 m3
Curve Inf o
dB(St(coax_pin_T1,coax_pin_T1))
Setup1 : Sweep1
Name X Y
m1 2.3600 -30.6972
m2 2.3300 -8.7353
m3 2.3900 -8.7632
Name Delta(X) Delta(Y) Slope(Y) InvSlope(Y)
d(m2,m3) 0.0600 -0.0279 -0.4648 - 2.1513
0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 10.00
Freq [G Hz]
-18.00
-16.00
-14.00
-12.00
-10.00
-8.00
-6.00
-4.00
-2.00
0.00
dB(S(WavePort1,WavePort1))
Ansoft Corporation HFSSDesign1
Return loss S11
m1
m2
m3
m4 m5 m6 m7
m8 m9
Curve Inf o
dB(S(WavePort1,WavePort1))
Setup1 : Sweep1
fdx ='0. 7mm' fd y='0 .7mm' h ='12. 5mm' l= '3mm' xx ='2 1mm' yy ='2 1mm'
dB(S(WavePort1,WavePort1))
Setup1 : Sweep1
fdx ='2mm' f dy ='0mm' h= '12. 5mm' l= '3mm' xx ='2 1mm' yy ='2 1mm'
Name X Y
m1 1.8110 -12.2846
m2 9.2110 -12.9463
m3 9.0810 -16.2443
m4 1.4610 -9.9544
m5 2.1910 -10.0007
m6 8.9510 -10.0809
m7 9.4510 -10.0228
m8 8.7510 -9.9104
m9 9.4110 -9.9357
Name Delta(X) Delta(Y) Slope(Y) InvSlope(Y)
d(m4,m5) 0.7300 -0.0463 -0.0634 -15.7808
d(m6,m7) 0.5000 0.0581 0.1163 8.6013
d(m8,m9) 0.6600 -0.0253 -0.0383 -26.1319
VSWR
1.00 1.50 2.00 2.50 3.00 3.50
Freq [GHz]
0.00
50.00
100.00
150.00
200.00
250.00
300.00
VSWRt(coax_pin_T1)
Ansoft Corporation HFSSDesign1
VSWR
m1
Curv e In f o
VSWRt(coax_pin_T1)
Setup1 : Sweep1
Name X Y
m1 2.3600 1.0601
1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 10.00
Freq [GHz]
0.00
5.00
10.00
15.00
20.00
25.00
30.00
VSWR(WavePort1)
Ansoft Corporation HFSSDesign1
VSWR
m1 m2m3
Curve Info
VSWR(WavePort1)
Setup1 : Sweep1
fdx= '0. 7mm' f dy =' 0.7 mm' h ='12 .5 mm' l=' 3mm' x x ='21 mm' y y ='21 mm'
VSWR(WavePort1)
Setup1 : Sweep1
fdx= '2mm' fd y= '0mm' h='1 2. 5mm' l ='3 mm' x x=' 21mm' yy ='2 1mm'
Name X Y
m1 1.8110 1.6423
m2 9.0810 1.3643
m3 9.2210 1.4958
Far Field
Radiation
Pattern
(Directivity)
1.00
2.00
3.00
4.00
90
60
30
0
-30
-60
-90
-120
-150
-180
150
120
Far Field_Radiation Pattern
Curve Info
DirT ot a l
Setup1 : LastAdaptive
Phi='0deg'
DirT ot a l
Setup1 : LastAdaptive
Phi='40deg'
0.08
0.16
0.24
0.32
90
60
30
0
-30
-60
-90
-120
-150
-180
150
120
Far Field_Radiation Pattern
Total Gain
(Polar Plot)
G=kD
Patch
/
Ground
Plane
Material
Copper Copper
Wideband Miniaturized Patch Antenna Design and Comparative Analysis 115
Table 1 (continued)
Mesh Plot
E-field
Distribution
Fig. 2. Return Loss and VSWR plots for proposed wideband antenna
The case when the feed location of the proposed antenna is at (0mm,0mm) position
and the substrate height is 10mm then the return loss behavior and the VSWR plot is
shown in the Fig. 2. Resonant frequency is at 1.90GHz, which can be used for GSM
1900 transmit frequency (uplink) band (1850 MHz –1910MHz). Bandwidth achieved
at 10dB is 56.81% which is equivalent to 1.088GHz of band span as shown in Fig. 2.
4 Conclusion
Proposed antenna provides large bandwidth, good return loss and VSWR behavior.
It is very much small in size compare to the conventional antenna. This proposed
116 S. Patel et al.
antenna is the efficient design in terms of bandwidth, number of bands of operation
and the cost.
References
1. Carver, K.R., Mink, J.W.: Microstrip antenna technology. IEEE Trans. Antennas
Propag. AP-29, 2–24 (1981)
2. Mailloux, R.J., McIlvenna, J.F., Kemweis, N.P.: Microstrip array technology. IEEE Trans.
Antennas Propagat. AP-29, 25–37 (1981)
3. Bahl, I.J., Bhartia, P.: Microstrip Antennas. Artech House, Dedham, MA (1980)
4. James, J.R., Hall, P.S., Wood, C.: Microstrip Antenna Theory and Design
5. James, J.R., Hall, P.S.: Handbook of Microstrip Antennas. Peter Peregrinus, London, U.K
(1989)
6. Richards, W.F., Lo, Y.T., Harrison, D.: An improved theory for microstrip antennas and
applications. IEEE Trans. Antennas Propagat. AP-29, 38–46 (1981)
7. Pozar, D.M.: Considerations for millimeter wave printed antennas. IEEE Trans. Antennas
Propagat. AP-31, 740–747 (1983)
8. Schaubert, D.H., Pozar, D.M., Adrian, A.: Effect of microstrip antenna substrate thickness
and permittivity: Comparison of theories and experiment. IEEE Trans. Antennas Propa-
gat. 37, 677–682 (1989)
9. Pozar, D.M.: Microstrip Antennas. Proceedings of the IEEE 80(1) (January 1992)
10. Garg, R., Bhartia, P., Bahl, I.J., Ittipiboon, A.: Microstrip antenna design handbook.
Artech House, New York (2001)
11. Carver, K.R.: Practical analytical techniques for the microstrip antenna. In: Proc. Work-
shop Printed Circ. Antennas, pp. 7.1–7.20. New Mexico State University (1979)
12. Binu Paul, S., Mridula, C.K., Aanandan, P., Mohanan: A new microstrip patch antenna for
mobile communications and Bluetooth applications. Microwave and Opt. Technol.
Lett. 33(4), 285–286 (2002)
13. Tiwari, V.K., Kimothi, A., Bhatnagar, D., Saini, J.S., Saxena, V.K.: Theoretical and ex-
perimental investigation of circular sector microstrip antenna. Indian J. Radio and Space
Physics 35, 206–211 (2006)
14. Bhardwaj, D., Bhatnagar, D., Sancheti, S., Soni, B.: Design of square patch antenna with a
notch on FR4 substrate. lET Microwaves, Antennas & Propagation 2(8), 880–885 (2008)
15. Lu, J.H.: Broadband Dual-Frequency Operation of Circular Patch Antennas and Arrays
with a Pair of L-Shaped Slots. IEEE Transactions on Antennas and Propagation 51(5),
1018–1023 (2003)
16. Wong, K.L., Hsu, W.S.: Broadband triangular microstrip antenna with V-shaped slot.
Electron. Lett. 33(25), 2085–2087 (1997)
17. Ansoft HFSS, Ansoft Corporation, http://www.ansoft.co.jp/hfss.htm
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 117–122, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Predicting Number of Zombies in a DDoS Attack Using
ANN Based Scheme
B.B. Gupta1,2,*, R.C. Joshi1, M. Misra1, A. Jain2, S. Juyal2,
R. Prabhakar2, and A.K. Singh2
1 Department of Electronics and Computer Engineering, Indian Institute of Technology
Roorkee, Roorkee, India
gupta.brij@gmail.com
2 Department of Computer Science and Engineering, Graphic Era University, Dehradun, India
Abstract. Anomaly based DDoS detection systems construct profile of the traf-
fic normally seen in the network, and identify anomalies whenever traffic devi-
ate from normal profile beyond a threshold. This deviation in traffic beyond
threshold is used in the past for DDoS detection but not for finding zombies. In
this paper, two layer feed forward neural networks of different sizes are used to
estimate number of zombies involved in a DDoS attack. The sample data used
to train the feed forward neural networks is generated using NS-2 network si-
mulator running on Linux platform. The generated sample data is divided into
training data and test data and MSE is used to compare the performance of var-
ious feed forward neural networks. Various sizes of feed forward networks are
compared for their estimation performance. The generalization capacity of the
trained network is promising and the network is able to predict number of zom-
bies involved in a DDoS attack with very less test error.
1 Introduction
Denial of service (DoS) attacks and more particularly the distributed ones (DDoS) are
one of the latest threat and pose a grave danger to users, organizations and infrastruc-
tures of the Internet. A DDoS attacker attempts to disrupt a target, in most cases a
web server, by flooding it with illegitimate packets, usurping its bandwidth and over-
taxing it to prevent legitimate inquiries from getting through [1,2]. Anomaly based
DDoS detection systems construct profile of the traffic normally seen in the network,
and identify anomalies whenever traffic deviate from normal profile beyond a thresh-
old [3]. This extend of deviation is normally not utilized. Therefore, this extends of
deviation from detection threshold and feed forward neural networks [4-6] are used to
predict number of zombies. A real time estimation of the number of zombies in DDoS
scenario is helpful to suppress the effect of attack by choosing predicted number of
most suspicious attack sources for either filtering or rate limiting. We have assumed
that zombies have not spoof header information of out going packets. Moore et. al [7]
have already made a similar kind of attempt, in which they have used backscatter
* Corresponding author.
118 B.B. Gupta et al.
analysis to estimate number of spoofed addresses involved in DDoS attack. This is an
offline analysis based on unsolicited responses.
Our objective is to find the relationship between number of zombies involved in a
flooding DDoS attack and deviation in sample entropy. In order to predict number of
zombies, feed forward neural network is used. To measure the performance of the
proposed approach, we have calculated mean square error (MSE) and test error.
Training and test data are generated using simulation. Internet type topologies used
for simulation are generated using Transit-Stub model of GT-ITM topology generator
[8]. NS-2 network simulator [9] on Linux platform is used as simulation test bed for
launching DDoS attacks with varied number of zombies and the data collected are
used to train the neural network. In our simulation experiments, attack traffic rate is
fixed to 25Mbps in total; therefore, mean attack rate per zombie is varied from
0.25Mbps to 2.5Mbps and total number of zombie machines range between 10 and
100 to generate attack traffic. Varies sizes of feed forward neural networks are com-
pared for their estimation performance. The result obtained is very promising as we
are able to predict number of zombies involved in DDoS attack effectively.
The remainder of the paper is organized as follows. Section 2 contains overview of
artificial neural network (ANN). Intended detection scheme is described in section 3.
Section 4 contains simulation results and discussion. Finally, Section 5 concludes the
paper.
2 Artificial Neural Network (ANN)
An Artificial Neural Network (ANN) [4-6] is an information processing paradigm that
is inspired by the way biological nervous systems, such as the brain, process informa-
tion. The key element of this paradigm is the novel structure of the information proc-
essing system. It is composed of a large number of highly interconnected processing
elements (neurons) working in unison to solve specific problems. ANNs, like people,
learn by example. An ANN is configured for a specific application, such as pattern
recognition or data classification, through a learning process. Learning in biological
systems involves adjustments to the synaptic connections that exist between the neu-
rons. This is true for ANNs as well. Neural networks, with their remarkable ability to
derive meaning from complicated or imprecise data, can be used to extract patterns
and detect trends that are too complex to be noticed by either humans or other com-
puter techniques. A trained neural network can be thought of as an "expert" in the
category of information it has been given to analyze. This expert can then be used to
provide projections given new situations of interest and answer "what if" questions.
3 Detection of Attacks
Here, we will discuss propose detection system that is part of access router or can
belong to separate unit that interact with access router to detect attack traffic. Entropy
based DDoS scheme [10] is used to construct profile of the traffic normally seen in
the network, and identify anomalies whenever traffic goes out of profile. A metric that
Predicting Number of Zombies in a DDoS Attack Using ANN Based Scheme 119
captures the degree of dispersal or concentration of a distribution is sample entropy.
Sample entropy H(X) is
2
1
() log()
N
ii
i
HX p p
=
=−
∑ (1)
where i
p is ni/S. Here ni represent total number of bytes arrivals for a flow i in
{t −∆, t} and
1
, 1, 2....
N
i
i
Sni N
=
==
∑. The value of sample entropy lies in the range
0-log2 N.
To detect the attack, the value of ()
c
HX
is calculated in time window Δ continu-
ously; whenever there is appreciable deviation from ()
n
XX
, various types of DDoS
attacks are detected. ()
c
HX
, and ()
n
XX
gives Entropy at the time of detection of
attack and Entropy value for normal profile respectively.
4 Results and Discussion
4.1 Training Data Generation
Neural network has to be trained by giving sample inputs and corresponding output
values and a training algorithm will adjust the connection weight and bias values until
a minimum error or other stopping criteria is reached. The training data has to be
taken carefully to consider the complete input range. Normalization and other pre-
processing of the data improve the training performance.
In our paper, in order to predict number of zombies ( ˆ
Y) from deviation (HC - Hn)
in entropy value, training data samples are generated using simulation experiments in
NS-2 network simulator. Simulation experiments are done at the same attack strength
25Mbps in total and varying number of zombies from 10-100 with increment of 5
zombies i.e. mean attack rate per zombie from 0.25Mbps-2.5Mbps. The data obtained
is divided into two parts, 78.95% of the data values are used for training. The remain-
ing data values which are selected randomly are used for testing.
4.2 Network Training
For the prediction of the number of zombies in a DDOS attack, three feed forward
neural networks have been tested. The feed forward networks used have different sizes.
The size of a network refers to the number of layers and the number of neurons in each
layer. There is no direct method of deciding the size of a network for a given problem
and one has to use experience or trial error method. In general, when a network is
large, the complexity of the function that it can approximate will also increase. But as
the network size increase, both training time and its implementation cost increase and
hence optimum network size has to be selected for a given problem. For the current
120 B.B. Gupta et al.
Table 1. Training results of various feed forward networks
Network
used
Network
size
Number of
Epochs
MSE
in training
5-1 400 6.86
10-1 400 0.36
2 layer
network
15-1 400 0.0025
problem, two layer feed forward networks with 5, 10 and 15 neurons are selected. The
training algorithm used is the Levenberg-Marquardt back propagation algorithm of
MATLAB’s neural network toolbox. The training results are given in Table 1.
4.3 Network Testing
Table 2 shows the result of the testing of the networks using the test data values.
Table 2. Test results of various feed forward networks
Network
used
Network
size
MSE in
Testing
5-1 2.91
10-1 2.59
2 layer
network
15-1 3.14
From the result of table 1, we can see that the MSE in training decreases linearly
as the network size increase. This is as expected. But in table 2, we can see that in
spite of the smaller MSE in training and the increase in network size, the test result
for the feed forward network having 15 hidden layer neurons is greater than the
networks having 5 and 10 neurons. One reason for this is, for a good network per-
formance the ration of number of tunable parameters to that of training data size has
to be very small and in here network size has increased but training data size is the
same. For the last network, the number of tunable parameters is 31 and ration is
1.63. And because of this over fitting has occurred and the generalization perform-
ance of the last network is poor though it has good training performance. The
training performance is measured using the mean square error (MSE). MSE is the
difference between the target and the neural network's actual output. So, the best
MSE is the closest to 0. If MSE is 0, this indicates neural network's output is equal
to the target which is the best situation. Number of zombies of the individual net-
works can be compared with actual number of zombies for each test data values and
the results are given in figure 1, 2 and 3. The simulation results show that two layer
feed forward networks with 10 neurons performs best. Two layer feed forward net-
works with 10 neurons is able to predict number of zombies involved in a DDoS
attack with very less error.
Predicting Number of Zombies in a DDoS Attack Using ANN Based Scheme 121
0
10
20
30
40
50
60
70
80
90
100
0.048 0.121 0.157 0.189
Deviation in Entropy
Number of Zombies
Observed number of Zombies Pridicted number of zombies using Feed Forword neural network of Size 5-1
Fig. 1. Comparison between actual number of zombies and predicted number of zombies using
feed forward neural network of size 5-1
0
20
40
60
80
100
120
0.048 0.121 0. 157 0.189
Deviation in Entropy
Number of Zombies
Observed number of Zombies Pridicted number of zombies us ing Feed Forword neural network of Size 10-1
Fig. 2. Comparison between actual number of zombies and predicted number of zombies using
Feed forward neural network of size 10-1
0
20
40
60
80
100
120
0.048 0.121 0.157 0.189
Deviation in Entropy
Number of Zombies
Observed number of Zombies Pridicted number of zombies using Feed Forword neural network of Size 15-1
Fig. 3. Comparison between actual number of zombies and predicted number of zombies using
Feed forward neural network of size 15-1
122 B.B. Gupta et al.
5 Conclusion and Future Work
The potential of feed forward neural network for predicting number of zombies in-
volved in a flooding DDoS attack is investigated. The deviation ( ()
c
HX
-()
n
XX
) in
sample entropy is used as an input and MSE is used as the performance measure. Two
layer feed forward networks of size 5, 10 and 15 have shown maximum mean square
error (MSE) of 2.91, 2.59 and 3.14 respectively in predicting the number of zombies.
Therefore, total number of predicted zombies using feed forward neural network is
very close to actual number of zombies. However, simulation results are promising as
we are able to predict number of zombies efficiently, experimental study using a real
time test bed can strongly validate our claim.
References
1. Gupta, B.B., Misra, M., Joshi, R.C.: An ISP level Solution to Combat DDoS attacks using
Combined Statistical Based Approach. International Journal of Information Assurance and
Security (JIAS) 3(2), 102–110 (2008)
2. Gupta, B.B., Joshi, R.C., Misra, M.: Defending against Distributed Denial of Service At-
tacks: Issues and Challenges. Information Security Journal: A Global Perspective 18(5),
224–247 (2009)
3. Gupta, B.B., Joshi, R.C., Misra, M.: Dynamic and Auto Responsive Solution for Distrib-
uted Denial-of-Service Attacks Detection in ISP Network. International Journal of Com-
puter Theory and Engineering (IJCTE) 1(1), 71–80 (2009)
4. Burns, R., Burns, S.: Advanced Control Engineering. Butterworth Heinemann (2001)
5. Dayhoff, U.E., DeLeo, J.M.: Artificial neural networks. Cancer 91(S8), 1615–1635 (2001)
6. Yegnanarayana, B.: Artificial Neural Networks. Prentice-Hall, New Delhi (1999)
7. Moore, D., Shannon, C., Brown, D.J., Voelker, G., Savage, S.: Inferring Internet Denial-
of-Service Activity. ACM Transactions on Computer Systems 24(2), 115–139 (2006)
8. GT-ITM Traffic Generator Documentation and tool,
http://www.cc.gatech.edu/fac/EllenLegura/graphs.html
9. NS Documentation, http://www.isi.edu/nsnam/ns
10. Shannon, C.E.: A mathematical theory of communication. ACM SIGMOBILE Mobile
Computing and Communication Review 5, 3–55 (2001)
11. Gibson, B.: TCP Limitations on File Transfer Performance Hamper the Global Internet.
White paper (2006),
http://www.niwotnetworks.com/gbx/
TCPLimitsFastFileTransfer.htm
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 123–131, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Novel Biometric Watermaking Approach Using
LWT- SVD
Meenakshi Arya1 and Rajesh Siddavatam2
1 Lecturer, 2 Associate Professor
Department of CSE & IT,
Jaypee University of Information Technology, Waknaghat,
Solan, Himachal Pradesh, India
meenakshi.arya@juit.ac.in, srajesh@juit.ac.in
Abstract. The lifting wavelet transform (LWT) is a recent approach to wavelet
transform, and singular value decomposition (SVD) is a valuable transform
technique for robust digital watermarking. While LWT allows generating an
infinite number of discrete biorthogonal wavelets starting from an initial one,
singular values (SV) allow us to make changes in an image without affecting
the image quality much. This paper presents an approach which tries to amal-
gamate the features of these two transforms to achieve a hybrid and robust digi-
tal image watermarking techniques. Certain performance metrics are used to
test the robustness of the method against common image processing attacks.
Keywords: Biorthogonal wavelets, Biometric watermarking, Lifting wavelet
transform, Singular value decomposition.
1 Introduction
Biometrics based authentication systems are becoming increasingly popular as they
offer enhanced security and user convenience as compared to traditional token-based
(I.D. card) and knowledge based (password) systems. Biometric watermarking refers
to embedding a biometric trait like fingerprint [1], face [2], handwritten signature [3]
etc. for the purpose of content authentication.
In [7], LSB and DWT have been synergistically combined to embed face template
in fingerprint image while the same idea was applied in [4] for embedding offline
handwritten signature in a host image. Vatsa et al. [6] developed a v- SVM based
biometric watermarking method which was further revised by Cheng et al. [5] for
biometric watermarking based on offline handwritten signature.
In the current literature, neither lifting wavelet transform nor singular value de-
composition have been used for biometric watermarking. In this paper, a novel LWT-
SVD based biometric watermarking technique for offline handwritten signatures has
been proposed. The lifting scheme has been used firstly for separating the significant
pixels of the host image from the insignificant ones and then the Singular Value De-
composition (SVD) is applied. The watermark is embedded at this level using a gain
factor (k). The watermarked image is then obtained by taking inverse LWT transform.
The proposed algorithm gives excellent results for various attacks on the host image.
The rest of the paper is organized as follows: Section 2 explains the theoretical
124 M. Arya and R. Siddavatam
framework of SVD and LWT while Section 3 presents the proposed method. In
Section 4, the significance measures PSNR and SSIM have been described to assess
the quality of the watermarked image and the recovered signature image. The efficien-
cy of the proposed method along with the results have been presented in Section 5.
Section 6 concludes the work.
2 Theoretical Framework of Lifting Wavelet Transform (LWT)
and Singular Value Transform (SVD)
2.1 Lifting Wavelet Transform (LWT)
The basic idea of wavelet transforms is to exploit the correlation structure present in
most real life signals to build a sparse approximation. The lifting scheme is a tech-
nique for both designing fast wavelets and performing the discrete wavelet transform.
The technique was introduced by Swelden [8, 9]. While the discrete wavelet trans-
form applies several filters separately to the same signal, the signal is divided like
zipper for the lifting scheme. Then a series of convolution- accumulate operations
across the divided signals is applied. Generally speaking, lifting scheme includes
three steps that are splitting, prediction and update. The basic idea of lifting is de-
scribed here briefly:
Split: The original signal is divided into two disjoint subsets. Although any disjoint
split is possible, we will split the original data set into 2 , the even
indexed points and [n] 2 1 , the odd indexed points.
Predict: The wavelet coefficients d[n] is generated as error in predicting xo[n] from
[n] using prediction operator P.
o e . (1)
Update: and are combined to obtain scaling coefficients c[n] that represent
a coarse approximation to the original signal . This is accomplished by applying an
update operator U to the wavelet coefficients and adding the result to e :
e . (2)
These three steps form a lifting stage. Iteration of the lifting stage on the output c[n]
creates the complete set of DWT scaling and wavelet coefficients c j [n] and d j [n]. At
each scale we weight the c j [n] and d j [n] with ke and ko respectively as shown in
Fig. 1. This normalizes the energy of the underlying scaling and wavelet functions.
The lifting steps are easily inverted even if P and U are nonlinear, space-varying, or
noninvertible. Rearranging equation (1) and (2) we have
. (3)
. (4)
The original signal will be perfectly reconstructed as long as the same P and U are
chosen for the forward and the inverse transforms.
A Novel Biometric Watermaking Approach Using LWT- SVD 125
Fig. 1. Lifting Steps
2.2 Singular Value Decomposition (SVD)
The Singular Value Transform (SVD) was explored a few years ago for watermarking
purposes. In recent years, SVD has been used in watermarking as a different trans-
form as it is one of the most powerful tools of linear algebra with several applications
in watermarking[10,11,12,13,14]. Singular values are the luminance values of SVD
image layer, changing these values slightly do not affect the image quality much .The
purpose of singular value decomposition is to reduce a dataset containing a large
number of values to a dataset containing significantly fewer values, but which still
contains a large fraction of the variability present in the original data. SVD analysis
results in a more compact representation of these correlations, especially with multi-
variate datasets and can provide insight into spatial and temporal variations exhibited
in the fields of data being analyzed. SVD is optimal matrix decomposition in a least
square sense packing the maximum signal energy into a few coefficients as possible
[11]. The SVD theorem decomposes a digital image A of size M × N, as:
∑. (5)
T∑. (6)
Since U and V are orthogonal, this becomes the singular value decomposition
∑
T. (7)
The full singular value decomposition of an (M × N) matrix involves an (M × M) U,
an (M × N) ∑, and an (N × N) V. In other words, U and V are both square and ∑ is the
same size as A. The singular value decomposition is the appropriate tool for analyz-
ing a mapping from one vector space into another vector space, possibly with a differ-
ent dimension.
3 Proposed Technique
An image comprises of certain high frequency components (edges) known as the
approximation coefficients and low frequency components (smooth areas) known as
the detailed coefficients. Most of the previous SVD and DWT-based watermarking
Odd even
split
x(n)
x
o
(n)
x
e
(n
)
-P
+
+
U
k
e
k
o
c(n)
d(n)
126 M. Arya and R. Siddavatam
techniques treat different parts of the image in the same way. Therefore, the edges and
the smooth areas of the image, related to different sub-bands, accept similar effects.
The HVS is less sensitive to noise on edges, hence making similar changes to percep-
tually significant and insignificant areas of the image consequently lead to noticeable
alternation in smooth areas, thereby causing a significant degradation to the image
quality.
The paper proposes a novel biometric watermarking technique with imperceptible
image quality alteration. Additional advantages of the presented technique could be
highlighted as high capacity and robustness of the method against different types of
common attacks. Since LWT provides high redundancy in transform domain, the high
capacity of the transformed host could utilize as the beneficial point to scatter the
watermark data.
3.1 Biometric Feature Processing
To employ offline handwritten signature as watermark, the preprocessing algorithm as
depicted in Fig 3 is applied on the signature image. Initially, the signature image is
binarized and resized to an image of 300 pixels x 200 pixels. This is to isolate single
stroke or a cluster of separated strokes of a handwritten signature from the back-
ground. Median filter is applied to this binary image to eliminate noise which might
be present in the form of speckles, smears, scratches etc. that might thwart feature
extraction. Hough transform (HT) is then applied to the signature image for projection
into feature space. The step is followed by applying Principle Components Analysis
(PCA) is to compress the feature space generated by HT without losing the significant
attributes [15]. Lastly, PCA feature is statistically discritized into binary representa-
tion signature code as proposed in [16].
3.2 Watermark Embedding
The following steps explain the embedding phase.
(i) Let Ioriginal be the host image of size N × N.
(ii) The Lifting Wavelet Transform Ilwt(i,j) of the host image is calculated according
to the selected decomposition level (L), sub-bands of size
can be
achieved.
(iii) Let Soriginal be the original offline handwritten signature of m x n where m<=n.
Resize the signature image such that size (Ioriginal) = size (Soriginal)
(iv) Calculate Slwt(i,j),the corresponding wavelet transform of the signature image.
(v) At L= 2, apply SVD to the horizontal detailed sub-band of the cover image as
well as to the signature image.
(vi) The singular values of the cover image sub-band are modified with the singular
values of the signature sub-band obtaining modified LWT coefficient at the 2nd
level.
2, 2,
2,. (8)
A Novel Biometric Watermaking Approach Using LWT- SVD 127
Embedding at this level is described as
2, 2,
2, . (9)
(vii) Using the inverse wavelet transformation the final watermarked image Iwm will
be constructed.
3.3 Watermark Extraction
Since the SVs of the original images are needed in the extraction phase, the proposed
technique is non-blind as it uses the singular vector matrices of the original signature
image as the keys. The extraction phase is explained by the following steps
(i) Compute the Lifting Wavelet Transform of the watermarked image according to
the selected decomposition level (L)
(ii) Locate the embedded coefficients and extract the singular values of the corres-
ponding sub-band of the signature image through Equation 10.
∑ / (10)
(iii) Combine the SVs thus obtained to recover the 2nd level approximation coeffi-
cient.
(iv) Perform 2 –level Inverse LWT to obtain the watermark.
3.4 Template Matching Based Authentication
This extracted watermark is fed as an input to the biometric feature processing algo-
rithm for template matching. The database contains 250 offline handwritten signatures
collected from 50 users at different times to capture the intrapersonal differences in
signing by a single user. Initially all the steps mentioned in biometric feature
processing are applied to the entire signature database to generate a feature vector
comprising the feature vectors corresponding to each signature image. These steps are
applied to the recovered signature image to extract its features. The Euclidean dis-
tance between the feature vector of the recovered signature and the feature vectors of
all the signatures in the database is calculated according to the formula as given by
Equation 11.
,, . (11)
The database image with the least Euclidean distance with the extracted image is the
corresponding template and hence the verification of the signature of the user.
4 Significance Measures
4.1 Peak Signal to Noise Ratio PSNR)
The proposed algorithm has been tested for various signal processing attacks like
median filtering, salt and pepper noise addition, histogram equalization, Gaussian
128 M. Arya and R. Siddavatam
noise and JPEG compression. The experimental results have been gauged using Mean
Square Error (MSE) and Peak Signal to Noise Ratio (PSNR) which have been given
below.
∑∑
,,
. (12)
where I and W are the original and the watermarked images having a resolution of
m*n.
10 log
. (13)
4.2 Structural Similarity Index Measure (SSIM)
SSIM [14] is a new paradigm metric designed to improve on traditional methods like
peak signal-to-noise ratio (PSNR) and mean squared error (MSE) for quality assess-
ment. It is based on the hypothesis that the HVS is highly adapted for extracting struc-
tural information. The measure of structural similarity compares local patterns of
pixel intensities that have been normalized for luminance and contrast. In practice, a
single overall index is sufficient enough to evaluate the overall image quality; hence a
mean SSIM (MSSIM) index is used as the quality measurement metric.
,
. (14)
,
∑,
. (15)
5 Results and Discussions
The proposed algorithm based on Lifting wavelet transform and singular value de-
composition generates results that are superior to the existing methods for offline
handwritten signature watermarking [4,5]. The algorithm embeds the signature as a
whole thus providing better authentication than the previous methods. Table 1 shows
the PSNR between the original and the recovered watermark varies between 53 dB to
60 dB for the signatures of 30 users when the watermarked image is not subjected to
any attack, while in Table 2, it can be seen that the PSNR value of recovered water-
marks (for 3 users) after the watermarked image is subjected to various attacks varies
between 50 dB to 55 dB.
Even after subjecting the watermarked image to a JPEG compression ratio ranging
between 90% to 30%, the watermark recovery is pretty good. For various noise ratios
between 10% to 20%, the PSNR varies between 30 dB to 40 dB as shown in Table 2.
For implementing the algorithm, MATLAB 7 on a 1.73 GHz Pentium M Processor
with minimum 256 MB of RAM has been used. The results have been verified on
various standard images like Lena, Peppers, Baboon and Elaine. Figures 2 and 3 show
the effect of varying the embedding factor while Figures 4 and 7 show the effect of
varying the decomposition level on various host images.
A Novel Biometric Watermaking Approach Using LWT- SVD 129
Table 1. Extracted Watermarks
PSNR=54
MSSIM=0.99886
PSNR=53
MSSIM=0.99826
PSNR=54
MSSIM=0.99909
PSNR=54
MSSIM=0.9983
PSNR=55
MSSIM=0.99995
PSNR=54
MSSIM=0.99922
PSNR=54
MSSIM=0.99863
PSNR=53
MSSIM=0.99929
PSNR=53
MSSIM=0.99853
PSNR=55
MSSIM=0.99949
PSNR=59
MSSIM=0.99949
PSNR=54
MSSIM=0.99969
PSNR=54
MSSIM=0.9994
PSNR=54
MSSIM=0.99914
PSNR=57
MSSIM=0.99998
Table 2. Extracted Watermarks after Simulation of Various Attacks
CROPPING HISTOGRAM MEDIAN
SALT & PEPPER GAUSSIAN SHARPENING
JPEG
PSNR=51
MSSIM=0.997
PSNR=28
MSSIM=0.850
PSNR=31
MSSIM=0.965
PSNR=29
MSSIM=0.790
PSNR=26
MSSIM=0.704
PSNR=24
MSSIM=0.784
PSNR=44
MSSIM=0.993
(a) Signature 1
CROPPING HISTOGRAM MEDIAN GAUSSIAN SHARPENING SALT & PEPPER JPEG
PSNR=50
MSSIM=0.999
PSNR=30
MSSIM=0.979
PSNR=30
MSSIM=0.990
PSNR=28
MSSIM=0.956
PSNR=26
MSSIM=0.963
PSNR=30
MSSIM=0.974
PSNR=44
MSSIM=0.99
7
(b) Signature 21
Fig. 2. PSNR versus embedding factor for various cover images
0
20
40
60
80
0.1 0.2 0.3 0.4 0.5
K
PSNR dB
LENA
CAMERAMAN
MANDRIL
PEPPERS
130 M. Arya and R. Siddavatam
Fig. 3. MSSIM versus embedding factor for various cover images
Fig. 4. PSNR versus decomposition level for various cover images
Fig. 5. MSSIM versus decomposition level for various cover images
Furthermore, the performance of the proposed algorithm has been tested for vari-
ous values of embedding factor. The effect of embedding factor on PSNR and
MSSIM has been presented in Tables 3 and 4 while Tables 5 and 6 show the effect of
varying the decomposition level.
6 Conclusion
In this paper, a novel biometric watermarking scheme using LWT-SVD for offline
handwritten signature has been proposed. The proposed technique shows superior
results as compared to the existing technique. The work can be further expanded by
incorporating the latest signature verification techniques so as to reduce the FAR or
FRR of the proposed system and also amalgamate the two areas of biometric water-
marking and signature authentication/ verification.
0.975
0.98
0.985
0.99
0.995
1
1.005
0.1 0.2 0.3 0.4 0.5
K
MSSIM
LENA
CAMERAMAN
MANDRIL
PEPPERS
0.98
0.985
0.99
0.995
1
1.005
LENA CAM ERAMAN MANDRIL PEPPERS
PSNR dB
Level 1
Level 2
0.98
0.985
0.99
0.995
1
1.005
LENA CAMERAMAN MANDRIL PEPPERS
MSSIM
Level 1
Level 2
A Novel Biometric Watermaking Approach Using LWT- SVD 131
References
1. Jain, A.K., Hong, L., Bolle, R.: On-line Fingerprint Verification. IEEE Trans. PAMI 19(4),
302–314 (1997)
2. Pang, Y.H., Teoh, A.B.J., David Ngo, C.L.: Enhanced Pseudo Zernike Moments in Face
Recognition. IEICE Electron, Express 2(3), 70–75 (2005)
3. Low, C.Y., Teoh, A.B.-J., Tea, C.: A Preliminary Study on Biometric Watermarking for
Offline Handwritten Signature. In: Proc. Of 2007 IEEE International Conference on Tele-
communications, Malaysia (2007)
4. Low, C.Y., Teoh, A.B.-J., Tea, C.: Fusion of LSB and LWT Biometric Watermarking for
Offline Handwritten Signature. In: 2008 Congress on Image and signal processing,
pp. 702–708. IEEE Computer Society, Los Alamitos (2008)
5. Low, C.Y., Teoh, A.B.-J., Tea, C.: Support Vector Machines (SVM) based Biometric Wa-
termarking for Offline Handwritten Signature. In: 2008 Congress on Image and signal
processing, pp. 702–708. IEEE Computer Society, Los Alamitos (2008)
6. Vasta, M., Singh, R., Noore, A.: Improving Biometric recognition accuracy and robustness
using DWT and SVM watermarking. IEICE Electronics Express 2(12), 362–367 (2005)
7. Vasta, M., Singh, R., Noore, A.: Robust Biometric Image Watermarking for Fingerprint
and Face Template Protection. IEICE Tran. On Fundamentals of Electronics (2006)
8. Sweldens, W.: The lifting scheme: A New Philosophy in Biorthogonal Wavelet Construc-
tions. In: Proceedings of SPIE, pp. 68-79 (1995)
9. Daubechies, I., Sweldens, W.: Factoring Wavelet Transforms into Lifting Schemes. The
Journal of Fourier Analysis and Applications 4, 247–269 (1998)
10. Chandra, D.V.S.: Digital Image Watermarking using Singular Value Decomposition. In:
Proc. of 45th IEEE Midwest Symposium on circuits and Systems, Tulsa, OK, pp. 264–267
(2002)
11. Liu, R., Tan, T.: A SVD based Watermarking scheme for protecting rightful ownership.
IEEE Transactions on Multimedia 4(1), 121–128 (2002)
12. Zhou, B., Chen, J.: A Geometric Distortion Resilient Image Watermarking Algorithm
Based on SVD. Chinese Journal of Image and Graphics 9, 506–512 (2004)
13. Bao, P., Ma, X.: Image Adaptive Watermarking Using Wavelet Domain Singular Value
Decomposition. IEEE Transactions on Circuits and Systems for Video Technology 15(1),
96–102 (2005)
14. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image Quality Assessment: From
Error Visibility to Structural Similarity. IEEE Transactions On Image Processing 13(4)
(2004)
15. Ooi, S. Y., Teoh, A.B.J., David Ngo, C.L.: Offline Signature Verification through Discrete
Radon Transform and Principal Component Analysis. In: Proc. of International Conference
on Computer and Communication Engineering (ICCCE), (2006).
16. Tuyls, P., Akkermans, A.H.M., Kevenaar, T.A.M.: Face Recognition with Renewable and
Privacy Preserving Binary Templates. In: 4th IEEE Workshop on Automatic Identification
Advanced Technologies (2005)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 132–137, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Detection and Prevention of Phishing Attack Using
Dynamic Watermarking
Akhilendra Pratap Singh, Vimal Kumar, Sandeep Singh Sengar, and Manoj Wairiya
Department of Computer Science and Engineering
MNNIT, Allahabad
{akhil121282,vimaliitr10,sansen0911,wairiya}@gmail.com
Abstract. Nowadays phishing attacks are increasing with burgeoning rate
which is highly problematic for social and financial websites.Many anti-
phishing mechanisms currently focused to verify whether a web site is genuine
or not. This paper proposes a novel anti-phishing approach based on Dynamic
watermarking technique. According to this approach user will be asked for
some additional information like watermark image, its fixing position and secret
key at the time of user’s registration and these credentials of particular user will
be changed at per login. During each login phase a user will verify the authentic
watermark with its position and decide the legitimacy of website.
Keywords: Phishing Attack, Watermarking, Website, Authentication.
1 Introduction
Today all organizations are using the internet for sharing the message so the commu-
nication must be secure. Few unethical hackers are doing the cyber criminal activity
by committing fraud. Attacker sends the phishing message by the fake website that
looks like a original site. This attack treated as a deceptive phishing attack which
target to the financial organization. Criminals complete their life cycle in very short
period by the login and personal detail of the people. Phishing attack has various
types as deceptive, malware, keyloggers, data theft, search engine, content injection
and web trojan. Many anti phishing tools exist to protect the attack. Juan Chen et.al.
have proposed a very standard approach in [1]. According to their idea they have
propose a new end-host based anti-phishing algorithm, which is called Link Guard,
which works by utilizing the generic characteristics of the hyperlinks in phishing
attacks. These characteristics are derived by analyzing the phishing data archive
provided by the Anti-Phishing Working Group (APWG). Because it is based on the
generic characteristics of phishing attacks, Link Guard can detect not only known but
also unknown phishing attacks. Michael Atighetchi et .al. have given the idea in [2].
They suggested a framework based on attribute based checks for defending against
phishing attacks. According to Amir Herzberg et.al.s approach presented in [3] they
present an improved security and identification indicators, as author implemented, a
browser extension in Trust Bar. Dmytro Iliev et.al.s proposed idea in [7] phishing
prevention approach based on mutual authentication is provided. Mohsen Sharifi et.al.
Detection and Prevention of Phishing Attack Using Dynamic Watermarking 133
have suggest an idea in [4] AntiPhishing Authentication (APA) technique to detect
and prevent real-time phishing attacks.Daisuke Miyamoto et.al.s gave an approach in
[5] on performance of machine learning-based methods such as Regression Trees and
Classification, Naive Bays, Additive Regression Trees , Logistic Regression for de-
tection of phishing sites. Daisuke Miyamoto et.al. Proposed approach in [6] that is
study of users past trust decisions (PTDs) for improving the accuracy of detecting
phishing sites. Mercan Topkara et.al. have suggested an approach using watermark in
[8] which is called ViWiD. ViWiD is an integrity check system based on visible wa-
termarking of logo images. Various webpage related security using watermarking
technique is discussed in [9][10][11][13][14].Section 2 describes the proposed algo-
rithm. Section 3 contains all experimental results and section 4 concludes the paper.
2 Proposed Algorithm
There are many methods proposed earlier to detect and prevent phishing attack. Some
of them use watermark technique and some them uses another approach. There is
some limitation in this approach like many checks and enforcements which are used
by the client-side defense tools can be tricked by attackers after getting a reasonable
knowing of web site construction [4]. For example, using mosaic attack, an attacker
can make fool the image check system of SpoofGaurd by partitioning the logo image
into small parts and show it in such way that it looks like legitimate one. Some time
due to irritation of some protection means like antivirus, user turnoff the protection
mechanism of client side. Hence all client side schemes will live all defensive actions.
Similarly in case of tools which are based on cryptography, they need individual
downloaded software in each client side machine. In some cases in spite of using SSL
secure connection, if client authentication tool will be turned off, it may suffers with
Phishing attack. Few earlier proposed algorithms also use watermarking technique but
visible and stationary nature of their watermark may also suffer with phishing attack.
In this paper we are proposing an approach for prevention of phishing attack based
on dynamic position watermarking technique. This approach is divided in to three
modules viz. Registration process, Login verification process and Web site closing
process. Different position for watermark image can be top left, bottom left, top right,
bottom right, center.
2.1 Registration Process
Registration process is the first phase whenever we open the website and trying to
become a member of the website. Hence first communication between client and
server is done in this phase. Most of the financial, social networking web site gives
the opportunity to make an account in their server by uploading user credentials like
user name and password etc. These credentials play vital role for further communica-
tion with those web site. There are three phase in registration process. You have more
than one assuming, please make sure that the Volume Editor knows how you are to be
listed in the author index.
134 A.P. Singh et al.
Step 1. Client will open the web page which starts with registration phase. There will
be five most essential credentials of user viz. Username, Password, Secret
key, Watermark Image and position of watermark. Here Secret key will be
unique for each user and hence it is called primary key. Since we are assum-
ing a secure channel between client and server, hence all credentials will be
in encrypted form and there will be no attacks like man in the middle.
Step 2. In this phase web site hosting server will store all data related to particular
user. Here Secret key will be primary key for the database.
Step 3. User is acknowledged with a proper web page having desired watermark at
predefined location. Now the user will get login page for entering his user-
name and password.
2.2 Login Verification Phase
Once user has created his account he needs to log in the web site. This is the very
critical phase which mostly suffers with phishing attack because most of the attackers
create a fake website similar to the legitimate one. This phase consist of eight steps.
Step 1. Whenever user will open the website he will be asked for proper Secret key
which is unique for each user. In this phase there will not be login window.
Step 2. A query will be forwarded to database of server for retrieving all credentials
of user associated with that particular secret key.
Step 3. After getting a proper match from database, two important credentials (Wa-
termark image and its position) will be returned to website hosting server.
Step 4. Login page will be displayed to user with proper watermark image and its
location as set by user at the time of registration.
Step 5. After ensuring the correctness of watermark image and its position, a legiti-
mate user will verify the authenticity of the website and then only he will
enter his login id and password.
Step 6. A query will be passed to database for entered username and password.
Step 7. Now according to username and password, all information related to that
particular user will be retrieved from database to the server.
Step 8. A proper account will be shown to user which is of legitimate website. Now
we can trust on that website.
2.3 Website Closing Phase
Since proposed algorithm is based on dynamic watermarking, hence at per login the
position and nature of watermark must be changed which is only known to legitimate
user.
Step 1. User starts the closing process by clicking on the close button.
Step 2. During logout, user will be prompt for reentering the new watermark image
and its location. This step is very essential and can not be ignored.
Step 3. After resetting the watermark image and position, user must have to wait for
acknowledgment from server.
Step 4. These all new information will be stored in database and old information will
be invalidated.
Step 5. After successful updating user will be acknowledged.
Detection and Prevention of Phishing Attack Using Dynamic Watermarking 135
2.4 Determination of Phishing Website
Suppose attacker has created a phishing website which looks similar to the original
one. As soon as user will click on suspicious link, fake website will be open which is
ask for secret key.
Step 1. In this step user will enter the secret key and wait for the desired watermark
image at particular position.
Step 2. A fake website will not have the database related to watermark information
and its position. That’s why it will be very difficult to determine or guess the
correct watermark and position by attacker.
Step 3. Due to absence of proper watermark and its location user can determine that
it is not authorized one and he is going to be suffered with phishing attack.
In this case the user must open the proper website by verifying the URL and then
must change his current secret key by newer one. Because attackers now aware with
the old secret key. Hence we will invalidate that secret key. When user enters correct
username and password, then only he can see his account of website, so here our main
aim is to protect username and password. Suppose if an attacker knows the secret key,
which is not changed by newer one till now. At that condition the watermark image
and its position doesn’t matter for the attacker. Now he will get login page but
still unaware of username and password so he will not be able to see the account
information as well as he can not change the secret key because during changing it,
a attacker must know all credentials of user. Hence the flowchart given in figure 1
will demonstrate the procedure to determine whether the website is phishing one or
original one.
Fig. 1. Flow chart for phishing Identification
3 Experimental Results
Proposed algorithm is implemented and verified over the Local area network. In our
experiment we have created a website for a bank named ABC Bank. First of all users
has created his account and stored his credentials over the server database. Here we
have taken a passport size photograph as a visible watermark and chosen the top left
position as a location for watermark. After creating an account whenever we open the
web site it will be looked like figure and it will only prompt for secret key as set by
136 A.P. Singh et al.
Fig. 2. Webpage for secret key insertion
Fig. 3. Web page for login and password insertion
legitimate user at the time of registration. Once we have entered correct secret key it
will show the web page look like figure 2. by which a user can verify the watermark
and its location and according to this information he can assure that opened website is
authentic one.
Now the user will be prompted for username and password. After entering correct
user name and password user will able to see his actual account information from a
legitimate and desired website. This is the exact way to prevent from Phishing attack
as proposed in this paper.
4 Conclusion
In this paper, we propose a new anti-phishing approach based on Dynamic water-
marking technique. This scheme neither requires online interactions with a Detection
and Prevention of Phishing Attack using Dynamic Watermarking third party, nor
requires any plug in or online tool hence this approach is more user friendly than the
previous approaches. Experimental results show the working of our approach in
which a user requires only different watermark at the time of per login which is more
tolerable than being hacked by attacker. It is clearly revealed that by determining the
main differences between the legitimate website and the phishing, one can reduce the
risk of this type of attack. According to experimental results it is clear that proposed
approach is more applicable for social and financial websites than others.
Detection and Prevention of Phishing Attack Using Dynamic Watermarking 137
References
1. Chen, J., Guo, C.: Online Detection and Prevention of Phishing Attacks. IEEE, Los
Alamitos (2006)
2. Atighetchi, M., Pal, P.: Attribute-based Prevention of Phishing Attacks. In: IEEE Int.
Symposium on Network Computing and Applications (2009)
3. Herzberg, A., Jbara, A.: Security and Identification dicators for Browsers against Spoofing
and Phishing Attacks, ACM Mohsen Sharif, A Zero Knowledge Password Proof Mutual
Authentication Technique Against Real-Time Phishing Attacks. Springer, Heidelberg
(2007)
4. Sharif, M., Saberi, A., Vahidi, M., Zorufi, M.: A Zero Knowledge Password Proof Mutual
Authentication Technique Against Real-Time Phishing Attacks. Springer, Heidelberg
(2007)
5. Miyamoto, D., Hazeyama, H., Kadobayashi, Y.: An Evaluation of Machine Learning-
Based Methods for Detection of Phishing Site. Springer, Heidelberg (2009)
6. Miyamoto, D., Hazeyama, H., Kadobayashi, Y.: Human Boost: Utilization of Users past
Trust Decision for Identifying Fraudulent Websites. Springer, Heidelberg (2009)
7. Iliev, D., Sun, Y.B.: Website forgery prevention IEEE (2010)
8. Topkara, M., Kamra, A., Atallah, M.J., Rotaru, C.N.: ViWiD: Visible Watermarking
Based Defense Against Phishing. Springer, Heidelberg (2005)
9. Frattolillo, F.: Watermarking Protocol for Web Context. IEEE Transactions on Information
Forensics and Security 2(3) (September 2009)
10. Jin, C., Xu, H., Zhang, X.: Web Pages Tamper-Proof Method Using Virus-Based Water-
marking. IEEE, Los Alamitos (2008)
11. Sun, P., Lu, H.: An efficient web page watermarking scheme. IEEE, Los Alamitos (2009)
12. Sun, P., Lu, H.: Two efficient fragile web page watermarking schemes. In: Fifth Interna-
tional Conference on Information Assurance and Security (2009)
13. Long, X., Peng, H., Zhang, C.: A Fragile Watermarking Scheme Based On SVD for Web
Pages. IEEE, Los Alamitos (2009)
14. Okada, M., Okabe, Y., Uehara, T.: A Web-based Privacy-Secure Content Trading System
for Small Content Providers Using Semi-Blind Digital Watermarking. In: IEEE Communi-
cations Society subject matter experts for publication in the IEEE CCNC 2010 (2010)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 138–143, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Search Tool Using Genetic Algorithm
M.K. Thanuja and C. Mala
Department of Computer science and Engineering, National Institute of technology,
Thiruchrapalli, India
thanujaprakash@yahoo.in, mala@nitt.edu
Abstract. In the current business scenario, with explosive growth of the amount
of information resources available over the Internet and Intranets in any organi-
zation, the retrieval of the required information at the spot and time of require-
ment is very essential, for the effective performance of the organization, by re-
ducing the process time and meeting customer satisfaction of meeting delivery
schedule. When the number of document collection and users of the document
go beyond an extent, an efficient search tool for the retrieving the information
from the collection of documents becomes vital. Actual information retrieval
means searching for keyword within documents. This study investigates the
various stages of information retrieval, selection of best method for implemen-
tation at each stage and optimizing the solution using of genetic algorithm with
different parameters. The method is tested with a training data of document col-
lections, where more relevant documents are presented to users in the genetic
modification. In this paper a new fitness function is presented in the genetic al-
gorithm for appropriate information retrieval which is found to be efficient than
other fitness functions.
Keywords: Cosine similarity, Fitness function, Genetic Algorithm, Information
Retrieval.
1 Introduction
With large amount of information resources available over the Internet and Intranets
in organizations, the information overload for the user has become overwhelming.
With the corresponding dramatic increase of the number of users in finding the best
and the newest information has increased exponentially. The absence of suitable al-
ternatives in the functionality of most of current information systems are, the user
looking for some topic and the Information Retrieval System(IRS) retrieves too much
information. The systems provide no qualitative distinction between the relevant and
irrelevant documents [1].Genetic algorithms (GAs) are not new to information re-
trieval. Gordon suggested representing a posting as a chromosome and using genetic
algorithms to select good indexes [2]. Yang et al. suggested using GAs with user
feedback to choose weights for search terms in a query [3].Morgan and Kilgour sug-
gested an intermediary between the user and IR system employing GAs to choose
search terms from a thesaurus and dictionary [4]. Boughanem et al. [5] examine GAs
for information retrieval and they suggested new crossover and mutation operators.
A Search Tool Using Genetic Algorithm 139
2 Motivation
BHEL is a engineering and manufacturing organization in which a large volume of
documents are stored into a vault through different process based applications. Due to
lack of search tools, the past and the present knowledge in the documents are not rea-
dily available. This paper aims at the study the process of information retrieval Sys-
tem and the efficient methods at each stage and optimizing the solution for giving the
best results to the users.
3 Proposed Method
3.1 Information Retrieval System
Information Retrieval System (IRS), a system used to extract and store information
pertaining to a collection, in required form , that need to be processed, searched and
retrieved based on the user’s query. The methods for information retrieval that have
been proposed in the literature are full text scanning, indexing, signature files and
clustering[6].When a comparison is made among the above methods indexing method
is found to be advantageous in terms of access time as well as storage space. So it is
proposed to use indexing method of data retrieval, for the proposed search tool.
3.2 Frame Work
3.3 Indexing
Most IRSs use keywords to retrieve documents. The complete indexing process in-
volves the steps namely markup and format removal of tags, tokenization, filtration
and indexing. Various index techniques are available in the field of IR, but the com-
mon and more efficient indexing technique found in the literature is inverted index. It
is proposed to use inverted index technique for the proposed search tool. An inverted
index is an index data structure storing a mapping from content, such as words, to it’s
location in a document or a set of documents. It is the most popular data structure
used in document retrieval systems [6].A general inverted file index consists of two
parts: a set of inverted file entries, being lists of identifiers of the records containing
each indexed word, and a search structure for identifying the location of the inverted
file entry for each term. The search structure may contain the information like the
document id in which the term is located. A typical inverted index appears as follows
Parsing
the doc-
Indexing the
terms of the doc-
uments (B tree
representa-
Query
processing
and
retrieval
of documents
Genetic
Algorithm
140 M.K. Thanuja and C
.
It is proposed to imple
m
every word(Term) that oc
c
includes a document num
b
search presents the followi
n
W(t
i
,d
j
)
=
where f(t
i
,d
j
) is the no. of t
i
in d
j,
f(t
i
) is the no. of docs.
3.4 Document Retrieval
a
In this paper it is proposed
viewed as a vector in n-di
m
tinguishing terms used to d
e
is also treated in the same
w
in the user request. Docu
m
between the query and the
d
presented with respect to t
h
It is proposed to use Co
s
3.5 Optimization Using
G
In GA, the search space is
a string is termed as a chr
o
fitness. A set of chromoso
m
at a given iteration of the
algorithm to choose weigh
t
for the retrieval of the rel
e
relevant document are not
c
this paper a new fitness fu
n
.
Mala
Fig. 1. An index structure
m
ent the search structure using a B tree. In this method,
c
urs in a document, the inverted list contains an entry
t
b
er and the weight of the term in the document. This
n
g weighting scheme for the weight of the terms[6].
=
f(
t
i
,d
j
)/sqrt(|d
j
|) x ln(D/f(t
i
))
i
mes term t
i
appears in document d
j,
|d
j
| is the no. of te
r
Containing t
i
and D is the no. of docs. in the collection
a
nd Relevancy
to use the vector model of the IRS, in which a docume
n
m
ensional document space (where n is the number of
d
e
scribe contents of the documents in a collection). A q
u
w
ay and constructed from the terms and weights provi
m
ent retrieval is based on the measurement of the simil
a
d
ocuments. In this method, the retrieved documents ca
n
h
eir relevance to the query [7].
s
ine similarity function in this wor
k
G
enetic Algorithm
composed of solutions to the problem each represente
d
o
mosome. Each chromosome has a function value, ca
l
m
es with their fitness is called the population. Populat
i
GA, is called a generation. Pathak et al. used a gen
t
s for such a combination [8].It is found the time requ
i
e
vant documents or the time required for downloading
c
onsidered in any of the fitness functions of GAs of IR
S
n
ction is introduced involving the time factors.
term1
term2
term3
term4
(1,2) (3,6) (3,3)
(3,4) (2,1)
(1,3) (2,3) (3,1)
(2,4)
for
t
hat
re-
(1)
r
ms
.
n
t is
d
is-
u
ery
i
ded
a
rity
n
be
d
by
l
led
i
on,
n
etic
i
red
the
S
. In
A Search Tool Using Genetic Algorithm 141
3.5.1 The Genetic Approach
Once significant keywords are extracted from training data (relevant and irrelevant
documents) including weights are assigned to the keywords. The binary weights of
the keywords are formed as a query vector. The GA have been applied for a fitness
function to get an optimal or near optimal query vector, also the results of the GA
approach have been compared with the IR Systems without using GA.
3.5.1.1 Representation of the Chromosomes. These chromosomes use a binary
representation, based on the relevancy of the documents with the given query. The no.
of genes will be equal to the no. of documents in the collections. In this genetic
approach the documents are represented as follows
Doc1={term1, term2 , term3,…..termn}
Query={qterm1, qterm2 ,q term3,…..qtermm}
The relevancy of the Doc1 with respect to the query is given by the cosine similarity
function mentioned in section 3.4 and is either 1 or 0. The relevancy of the document
with respect to the given query is encoded as chromosome as follows
Chromosome C1={0 1 1 0 1 0 0 0 0 10 0 0 0 0 0 0 0 0 0} means the documents 2,3
& 10 are relevant to the query for a collection of 20 documents. The no. of genes in a
chromosome is the total no. of documents in the collection.
3.5.1.2 Fitness Function. Fitness function is a performance measure or reward
function, which evaluates how each solution, is good. In this work, the GA with
following fitness function is used
F=rel(d) +1/Tr + 1/Td +S (2)
where rel(d) is the total no. of relevant documents retrieved with respect to the given
query i.e ∑x
where xk is kth gene of the ith chromosome, Tr is the time taken to
select the relevant documents, Td is the time taken to download the relevant docu-
ments, S total size of the relevant documents.
3.5.1.3 Selection. As the selection mechanism, the best chromosomes will on
average achieve more copies, and the worst fewer copies. The algorithm stops when
the fitness value of all the chromosomes of a generation is equal and are equal to the
maximum fitness of the previous generations. In the GA approaches, two GA
operators are used to produce offspring chromosomes, namely Crossover which
occurs with crossover probability Pc. GAs construct a better solution by mixture of
good characteristic of chromosome together. Higher fitness chromosome has an
opportunity to be selected more, so good solution always alive to the next generation.
In this work the crossover is done between the alternate chromosomes with maximum
fitness. Mutation involves the modification of the gene values of a solution with
some probability Pm. The mutation in this work is with probability of 0.001 and
occurs at random.
142 M.K. Thanuja and C. Mala
4 Experimental Results
The test databases for the GA approach are documents in a vault. The experiment was
applied on 10 queries. Initial generation and subsequent generation of a sample of 4
chromosomes are shown in Table 1.In the cross over process the chromosomes with
maximum fitness are selected.
Table 1. Chromoses and the fitness for initial generation and subsequent generation
Initial Generation Subsequent generation
Chromosomes Fitness Chromosomes Fitness
01101111101100111010 63.40 11110011111010111000 63.40
11110011111010111000 63.42 01101111101100111010 63.42
01010101101001101000 59.69 0101110000000110100 57.27
01011100000001000000 55.38 01010101101001000000 57.34
The maximum fitness values of 10 queries are listed in Table 2.
Table 2. Queries and the maximum fitness values
Query Maximum fitness value
1-5 58.617 to 62.226
6-10 57.456 to 60.31
Precision= Number of documents retrieved and relevant/Total Retrieved.
From the experimental observation, the best values for this test documents collec-
tions at crossover probability Pc = 0.8 and mutation rate is Pm= 0.0001 for the GA
follows
Table 3. Comparison of the precision with the study GA and without GA for first 10 queries
Precision for 50 queries Precision for 50 queries
Query Precision Query Precision
With GA Without GA With GA Without GA
1 0.44 0.5 6 0.63 0.7
2 0.55 0.55 7 0.625 0.75
3 0.65 0.66 8 0.65 0.66
4 0.78 0.555 9 0.8 0.85
5 0.5 0.5 10 0.56 0.56
The precision of 10 queries listed in the above table is plotted in the Fig 2.
Fig. 2. Comparison of th
e
5 Conclusion
From pervious results, it is
the paper gives more sophi
lections. Also, from the re
s
sion value better than the o
t
References
[1] Martin-Bautista, M.J., Vi
l
an Adaptive Information
R
[2] Gordon, M.: Probabilisti
c
of the ACM 31(10), 1208
–
[3] Yang, J., Korfhage, R.,
R
genetic algorithms–a rep
o
1st text retrieval conferen
[4] Morgan, J., Kilgour, A.:
P
algorithm. In: Moscardi
n
Intelligence, pp. 142–149
[5] Boughanem, M., Chrism
e
relevance optimization i
n
formation Science and T
e
[6] Faloutsos, C., Christodo
u
Performance Evaluation
[7] Salton, G., McGill, M.H
New York (1983)
[8] Pathak, P., Gordon, M.,
F
based matching function
Science (HICS), Hawaii,
U
A Search Tool Using Genetic Algorithm
e
precision with the study GA and without GA for 20 queries
noted that the new fitness function which is represente
d
sticated results than other fitness functions, in the test
c
s
ults, it is noted that our new fitness function has a pr
e
t
her fitness functions.
l
a, M.-A., Larsen, H.L.: A Fuzzy Genetic Algorithm Approa
c
R
etrieval Agent
c
and genetic algorithms in document retrieval. Communicat
i
–1218 (1988)
R
asmussen, E.: Query improvement in information retrieval u
o
rt on the experiments of the TREC project. In: Proceedings o
f
ce (TREC-1), pp. 31–58 (1992)
P
ersonalising on-line information retrieval support with a ge
n
n
i, A., Smith, P. (eds.) PolyModel 16: Applications of Artif
i
(1996)
e
nt, C., Tamine, L.: On using genetic algorithms for multim
o
n
information retrieval. Journal of the American Society fo
r
e
chnology 53(11), 934–942 (2002)
u
lakis, S.: An access Method for Documents and its Analy
t
.: Introduction to Modern Information Retrieval. Mc
G
raw-
H
F
an, W.: Effective information retrieval using genetic algorit
h
s adaption. In: Proc. 33rd Hawaii International Conferenc
e
U
SA (2000)
143
d
in
c
ol-
eci-
c
h to
i
ons
sing
f
the
n
etic
i
cial
o
dal
r
In-
t
ical
H
ill,
h
ms
e
on
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 144–149, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Heterogeneous Data Mining Environment Based on DAM
for Mobile Computing Environments
Ashutosh K. Dubey1, Ganesh Raj Kushwaha2, and Nishant Shrivastava3
1 Dept. of Computer Science & Engineering
Trinity Institute of Technology and Research
Bhopal, India
ashutoshdubey123@gmail.com
2 Dept. of Computer Science & Engineering
Trinity Institute of Technology and Research
Bhopal, India
ganeshrajkushwaha@gmail.com
3 Dept. of Computer Science & Engineering
JNCT, Bhopal, India
nishantuit@gmail.com
Abstract. Today the concept of Data Mining services is not alone sufficient.
Data mining services play an important role in the field of Communication in-
dustry. Data mining is also called knowledge discovery in several database in-
cluding mobile databases and for heterogeneous environment. In this paper, we
discuss and analyze the consumptive behavior based on data mining technol-
ogy. We discuss and analyze different aspects of data mining techniques and
their behavior in mobile devices. We also analyze the better method or rule of
data mining services which is more suitable for mobile devices. In this paper,
we propose a novel DAM (Define Analyze Miner) Based data mining approach
for mobile computing environments. In DAM approach, we first propose about
the environment according to the requirement and need of the user where we
define several different data sets, then DAM analyzer accept and analyze the
data set and finally apply the appropriate mining by the DAM miner on the ac-
cepted dataset. It is achieved by CLDC and MIDP component of J2ME.
Keywords: Data Mining, DAM, CLDC, MIDP.
1 Introduction
The rapidly expanding demand for digital mobile communication services, along with
the wide popularity of mobile devices, has given rise to the development of new
dimensions and requirements for future mobile communication systems. But today
is also the requirement of heterogeneous environment where we accept different data
sets according to the user and apply data mining on them. It is also use in mobile de-
vices with the use of MIDLET and CLDC component of J2ME. In few years back,
mobile extensions to Grid systems have been increasingly proposed in order to support
ubiquitous access and selection to the Grid and to include mobile devices as additional
Heterogeneous Data Mining Environment Based on DAM 145
Grid resources [1, 2]. In today’s scenario mobile devices, such as mobile phones,
PDAs, notebook and others, provide a basic building block [3][4][5][6].
Finding prevalent mobile user patterns and behavior in a heterogeneous environ-
ment has been one of the major problems in the area of mobile data mining. Particu-
larly, the algorithms of discovering frequent user’s behavior patterns in the mobile
agent system have been studied extensively in recent years. The key feature in most
of these algorithms is that they use a dataset and frequent Item-Sets visited by the
customers. In this case, some problems occur because they do not consider that mo-
bile user’s behavior patterns are dynamically variable as time passes. In this paper we
discuss some of the data mining service which are use in different areas and then ap-
ply those services to mobile devices and then apply those DMS services in mobile
computing and exploiting the need of DMS in mobile computing environments using
CLDC and MIDP components. The Connected, Limited Device Configuration
(CLDC) and the Mobile Information Device Profile (MIDP) have emerged as J2ME
standards for mobile phone applications development which are used with DMS ser-
vices. The role of CLDC and MIDP component is to apply Data Mining Services in
mobile. The remaining of this paper is organized as follows. We discuss CLDC and
J2Me in Section 2. In Section 3 we discuss about MIDP.The proposed heterogeneous
data mining algorithm, namely DAM in section 4. In section 5 we discuss about the
Challenges. The conclusions and future directions are given in Section 6. Finally ref-
erences are given.
2 CLDC and J2ME
The J2ME architecture is described in general before the components in the J2ME
technology are introduced.J2ME applications are also discussed in general, and it is
explained how they are made available to end users.J2ME is a highly optimized Java
runtime environment. Fig 1 shows the J2ME architecture.
The fundamental branches of the J2ME platform are configurations. A
configuration is a specification that describes a Java Virtual Machine and some set of
APIs that are targeted at a specific class of device.
Fig. 1. J2ME Architecture
146 A.K. Dubey, G.R. Kushwaha, and N. Shrivastava
The Connected, Limited Device Configuration is one such specification. The CLDC
specifies the APIs for devices with less than 512 KB of RAM available for the Java
system and an intermittent (limited) network connection. It specifies a stripped-down
Java virtual machine, called the KVM, as well as several APIs for fundamental
application services. Three packages are minimalist versions of the J2SE java.lang,
java.io, and java.util packages. A fourth package, javax.microedition.io, implements the
Generic Connection Framework, a generalized API for making network connections.
Many J2ME games already exist and enjoy great popularity especially among
young generation. Java comes with the immense requirement of the object-oriented
programming language for developers to implement new mobile applications [7].
Configurations provide core functionality and a way to provide greater flexibility but
no services for managing the application life-cycle, for driving the user interface, for
maintaining and updating persistent data on the device or for secure access to infor-
mation stored on a network server [8].Fig 1 shows the CLDC position in J2ME Archi-
tecture in configuration part.
Several networks have conducted a survey on users’ watching behavior [9] which
reflects that user behavior pattern recognition is not so easy task, we can achieve this
by CLDC and MIDP component. Instead of replacing existing TV service, mobile
services should be complementary [10], and offer more interactive means for users to
watch their chosen content.
3 MIDP
The Mobile Information Device Profile is a specification for a J2ME profile. It is
layered on top of CLDC and adds APIs for application life cycle, user interface,
networking, and persistent storage [11]. An application written for MIDP is called a
MIDlet. MIDlet applications are subclasses of the javax.microedition.midlet.MIDlet
class that is defined by MIDP. MIDlets are packaged and distributed as MIDlet suites.
A MIDlet suite can contain one or more MIDlets. A MIDlet is a J2ME application
designed to operate on small computing device. A MIDlet is defined with at least a
single class that is derived from the javax.micoedition.midlet.MIDlet abstract class.
The Position of MIDP is shown in Fig 2.
Fig. 2. MIDP in J2ME Architecture
Heterogeneous Data Mining Environment Based on DAM 147
4 Proposed Method: DAM
This method actually deal with the Heterogeneous Data Mining Environment Based
on DAM for mobile computing environments, where we first design a framework
where several data set activity is stored and their patterns are recognize based on the
environment. When the related data set is input from any other external environment
then our DAM environment recognize the particular data set if it is present in the
DAM environment otherwise reject the data set for further processing.
In the second phase we select the data set if the DAM analyzer analyzes the data
set from the database for further operation. Finally apply the DAM miner Strategy on
the data set for finding frequent pattern and further pruning. We also consider those
mining techniques which provide better performance and speed in terms of computa-
tion. In this section, we describe the proposed method. The entire concept is divided
into four phases: 1) Define the data set for our database .2) Analyze and Select the
data set when ever needed. 3) Apply the mining technique which is best suited. 4)
Mobile computing for the data set.
1) Algorithm for Creating the Data Set
For accepting the data set we consider the following things:
a) Table size
b) No of Columns
c) No of rows
Assumptions:
Col: Column, Dt: Data Type , v1, v2….vn: Values
Algorithm Define Datasets (DAM Definer)
1. Define the Data set from different data Source (WWW, XML, Data Ware-
house etc.).
2. Apply the Create table statement when defining the data source Data Source.
Create table tablename as select * / Col1, Col2, Col from data source.
[We use Oracle as a Database Management Software]
3. Apply the Create table statement[Stand alone]
Create table tablename (col1 Dt1, Col2 Dt2………Col n Dtn) with condition
clause.
4. Insert the values considering the domain.
5. For Standalone [Insert into table name values (v1, v2………vn).]
6. For step no2 when using universal false condition insert into table (Select
query).
7. Finish.
2) Algorithm for Analyzing and selecting the Data Set(DAM Analyzer)
Algorithm Select Datasets (SDS)
1. Select the Data Source (WWW, XML, Data Warehouse etc.).
2. check in the database
3. if(value!=NULL) then
4. For evaluating all columns with rows [Select * from datasource.]
5. For selected set of columns [Select col1, col2 ….coln from datasource.]
6. For some specified set of conditions
Select * / col1, col2...coln from datasource where condition.
5. Finish.
148 A.K. Dubey, G.R. Kushwaha, and N. Shrivastava
3) Algorithm for Alter the Data Set and Mining(DAM Miner)
By alter statement we can add a column, delete a column and increase and decrease
the size of the column in the database.
1. Select the Data Source from (WWW, XML, Data Warehouse etc.)
2. Add new column to the data source.
Alter table tablename add (Col1 Dt1, Col2 Dt2……..Coln Dtn).
3. Modify the size of the column
If(table is empty)
{3a. Increase the size
Alter table tablename modify (Col1 Dt1, Col2 Dt2……..Coln Dtn).
3b. Decrease the size
Alter table tablename modify (Col1 Dt1, Col2 Dt2……..Coln Dtn).}
If(table is not empty)
{3c. Increase the size
Alter table tablename modify (Col1 Dt1, Col2 Dt2……..Coln Dtn).}
Else
{Exit (0) ;}
4) Algorithm for Mobile Communication the Data Set (MCDS)
1. Select the Data Source (WWW, XML, Data Warehouse etc.) .
2. Apply the MIDP Profile and using MIDlet establish the connection
3. Apply the packages of J2ME according to the need.
4. Using WTK (Wireless ToolKit) access the result on the simulator.
5. Apply data Mining techniques.
6. Finish.
After analyzing the several aspects of DAM method the picture is clear for any data-
bases it is easy to manage and whenever necessary we can update the repository sys-
tem. We also apply several data mining techniques very smoothly because our data
base is consistent because of limiting redundancy in the database. Finally apply the
J2ME for mobile devices so that we can coherent the entire above scenario for mobile
computing environments.
5 Challenges
A number of constraints and technical difficulties faced by researchers, which are
discussed in this section. These general problems must be considered for further re-
search in this area to propose new technologies for making mobile computing easier.
Some of these are:
• The screen size of the mobile is a big limitation. The screen size can affect the
approximate visualization of complex results representing the discovered
model.
• Mobile navigation facility is also a big task to achieve and implement.
• The overhead due to the communication between MIDLET and Data Mining
service should not affect the execution time.
• The experiments on system performance depend almost entirely on the comput-
ing power of the server on which data mining task is executed.
Heterogeneous Data Mining Environment Based on DAM 149
6 Conclusions and Future Work
In this paper, we propose a novel DAM (Define Analyze Miner) Based data mining
approach for mobile computing environments. In DAM approach, we first propose
about the environment according to the requirement and need of the user where we
define several different data sets, then DAM analyzer accept and analyze the data set
and finally apply the appropriate mining by the DAM miner on the accepted dataset.
In future we also work on the limitations that were faced by the researchers.
References
[1] Migliardi, M., Maheswaran, M., Maniymaran, B., Card, P.: Mobile Interfaces to Compu-
tational, Data, and Service Grid Systems. ACM, New York (2004)
[2] Wesner, S., Dimitrakos, T., Jeffrey, K.: Akogrimo – The Grid goes Mobile. ERCIM (59)
(2004)
[3] Arcelus, A., Jones, M.H., Goubran, R., Knoefel, F.: Workshops, AINAW (2007)
[4] Bahati, R.M., Bauer, M.A.: Adapting to runtime changes in policies driving autonomic
management. In: ICAS 2008 : Proceedings of the Fourth International Conference on
Autonomic and Autonomous Systems (ICAS 2008). IEEE Computer Society, USA
(2008)
[5] Beetz, M., Bandouch, J., Kirsch, A., Maldonado, A., Müller, A., Rusu, R.B.: Proceedings
of the 4th COE Workshop on HAM (2007)
[6] Bergmann, R.: Ambient intelligence for decision making in fire service organizations. In:
AmI, pp. 73–90 (2007)
[7] Isakow, A., Shi, H.: Review of J2ME and J2MEbased Mobile Applications. International
Journal of Communication and Network Security, 189–198
[8] Ortiz, A Survey of J2ME Today, Sun Developer Network (SDN) (2004a)
http://developers.sun.com/mobility/getstart/articles/survey/
(viewed August 13, 2007)
[9] http://www.3g.co.uk/PR/Sept2005/1943.htm
[10] http://www.cellular-news.com/story/18707.php
[11] http://docs.sun.com
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 150–155, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Selection of Views for Materialization Using Size and
Query Frequency
T.V. Vijay Kumar and Mohammad Haider
School of Computer and Systems Sciences,
Jawaharlal Nehru University,
New Delhi-110067, India
Abstract. View selection is concerned with selecting a set of views that im-
proves the query response time while fitting within the available space for ma-
terialization. The most fundamental view selection algorithm HRUA uses the
view size, and ignores the query answering ability of the view, while selecting
views for materialization. As a consequence, the view selected may not account
for large numbers of queries. This problem is addressed by the proposed algo-
rithm, which aims to select views by considering query frequency along with
the size of the view. The proposed algorithm, in each iteration, computes the
profit of each view, using the query frequency and size of views, and then se-
lects from amongst them, the most profitable view for materialization. The
views so selected would be able to answer a greater number of queries resulting
in improvement in the average query response time. Further, experimental
based comparison of the proposed algorithm with HRUA showed that the pro-
posed algorithm was able to select views capable of answering significantly
greater number of queries at the cost of a slight increase in the total cost of
evaluating all the views.
Keywords: Materialized Views, View Selection, Greedy Algorithm.
1 Introduction
Data warehouse stores subject oriented, integrated, time variant and non-volatile data
to support processing of analytical queries [7]. These analytical queries, which are long
and complex, consume a lot of time when processed against a large data warehouse.
Further, the exploratory nature of these analytical queries contributes to high average
query response time. This query response time can be reduced by materializing views
over a data warehouse[9]. Materialized views contain pre-computed and summarized
information, computed from the information stored in the data warehouse. They are
significantly smaller in size, when compared with the data warehouse, and can signifi-
cantly reduce the response time if they contain relevant and required information for
answering analytical queries. The selection of such information and storing them as
materialized view is referred to as the view selection problem[4]. View selection deals
with selecting appropriate set of views that provide answers to most of the future que-
ries. View selection is formally defined in [4] as “Given a database schema D, storage
space B, Resource R and a workload of queries Q, choose a set of views V over D to
Selection of Views for Materialization Using Size and Query Frequency 151
materialize, whose combined size is at most B and resource requirement is at most R”.
The number of possible views is exponential in the number of dimensions and for
higher dimensions it would become infeasible to materialize all views due to space
constraints[6]. Further, the space and resource constraint translates the views selection
problem into an optimization problem that is NP-Complete[6]. Alternatively, views
can be selected empirically, based on past query patterns[12], or heuristically using
algorithms that are greedy, evolutionary etc. This paper focuses on greedy based view
selection.
Greedy based view selection, in each iteration, select the most beneficial view for
materialization. Among the several greedy based algorithms presented in literature
[1, 2, 3, 5, 6, 8, 10, 11, 13, 14], the algorithm in [6] is considered the most fundamen-
tal one. This algorithm, which hereafter in this paper would be referred to as HRUA,
selects the top-T beneficial views, from amongst all possible views, in a multidimen-
sional lattice. HRUA computes the benefit of a view in terms of its cost, which is de-
fined in terms of the size of the view. HRUA computes benefit as given below:
BenefitV = ∑{(Size(SMA(W)) – Size(V)) | V is an ancestor of view W in the lattice
and (Size(SMA(W)) – Size(V)) > 0}
where Size(V) = Size of view V
Size(SMA(V)) = Size of Smallest Materialized Ancestor of view V
HRUA uses size of the view to compute the benefit, It does not consider the number
of queries that can be answered by a view, referred to as its query frequency. As a
consequence, the views selected using HRUA may not be beneficial with respect to
answering most of the future queries. As an example, consider a three dimensional
lattice shown in Fig. 1(a). The size of the view in million (M) rows, and the query
frequency (QF) of each view, is given alongside the view. Selection of Top-3 views
using HRUA is shown in Fig. 1(b).
Fig. 1. Selection of Top-3 views using HRUA
HRUA selects AB, C and A as the Top-3 views. These selected views result in a
Total View Evaluation Cost (TVEC) of 492. Considering the query frequency along
with the size of each view, the Total Queries Answered (TQA) by the selected views
AB, C and A is 3130, from among 6214 queries. An increase in this TQA value
would result in more queries being answered by the selected views. The proposed
algorithm aims to select Top-T profitable views for materialization that improves the
TQA value by considering query frequency along with the size of each view. As a
Benefit
View Size QF 1st Iteration 2nd Iteration 3rd Iteration
AB 56 1032 176 - -
AC 68 1616 128 64 32
BC 62 1468 156 76 38
A 22 440 156 68 48
B 30 810 140 52 32
C 36 848 128 84 -
Selection of To
p
-3 views usin
g
HRUA
3-Dimensional lattice
ABC
AB BC AC
C B A
NONE
(100,
2000)
(56,
1032) (68,
1616
)
(22,
440) (30,
810
)
(62,
1430)
(36,
848)
(1,
0
)
(a) (b)
152 T.V. Vijay Kumar and M. Haider
result, the views selected would be able to answer a greater number of queries thereby
improving the average query response time. The paper is organized as follows: The
proposed algorithm is given in section 2 followed by experimental results in section 3.
Section 4 is the conclusion.
2 Proposed Algorithm
Unlike HRUA, the proposed algorithm aims to select views that are not only profit-
able with respect to size but are also capable of answering greater number of queries.
The proposed algorithm, in each iteration, considers the query frequency, along with
the size of each view, to select the most profitable view for materialization. The query
frequency of each view reflects past trends in querying and is computed as the number
of queries, posed in the past, that can be answered by the view. The proposed algo-
rithm, as given in Fig. 2, takes the lattice of views along with the size and query fre-
quency of each view as input and produces the Top-T views as output.
Fig. 2. Proposed Algorithm
The proposed algorithm, in each iteration, computes the profit of each view P(V)
as given below:
()()
()() ()()()()
∑⎪
⎭
⎪
⎬
⎫
⎪
⎩
⎪
⎨
⎧>−
−
−
=0VSWSMAS and lattice in the W viewofancestor an is V
)(WSMAS
)(WSMAQF
P(V) VS
VQF
Input: lattice of views L along with size and query frequency of each view
Output: Top-T views
Method:
Let
VR be the root view in the lattice, S(V) be the size of view V, QF(V) be the query frequency of V in the lattice,
SMA(V) be the smallest materialized ancestor of V, D(V) be the set of all descendent views of V, MV be the set
of materialized views, P (V) = Profit of view V, PM = Maximum Profit, VP = View with maximum profit
FOR V ∈ L
SMA(V) = RootView
END FOR
REPEAT
PM = 0
FOR each view V∈ (L – VR ∪ MV)
VP = V
P(V) = 0
FOR each view W ∈ D(V) and (S(SMA(W)) – S(V)) > 0
()()
()()
)(WSMAS
)(WSMAQF
P(V)P(V) VS
VQF
−
−
+=
END FOR
IF PM < P(V)
PM = P(V)
VP = V
END IF
END FOR
MV = MV ∪ {VP}
FOR W ∈ D(VP)
IF S(SMA(W)) > S(VP)
SMA(W) = VP
END IF
END FOR
Until |MV| < T
Return MV
Selection of Views for Materialization Using Size and Query Frequency 153
The profit of a view V is computed as the product of the number of dependents of V
and the ratio of frequency difference between V and its smallest materialized ancestor
and the size difference between V and its smallest materialized ancestor. The pro-
posed algorithm (PA), in each iteration, computes profit of the as yet unselected views
and selects, from amongst them, the most profitable view for materialization. The
selection continues in this manner until T views are selected.
Let us consider the selection of the Top-3 views from the multidimensional lattice
in Fig. 1(a) using PA. The selection of Top-3 views is given in Fig. 3.
Fig. 3. Selection of Top-3 views using PA
PA selects AB, A and BC as the Top-3 views. The views selected using PA has a
TVEC of 480, which is less than TVEC of 492 due to views selected using HRUA.
Also, the views selected using PA have a comparatively higher value of TQA of 4598
against the TQA of 3130 due to views selected using HRUA. Thus, it can be said that
PA, in comparison to HRUA, is capable of selecting views that account for a greater
number of queries at a lower total cost of evaluating all the views.
In order to compare the performance of PA with respect to HRUA, both the algo-
rithms were implemented and run on data sets with varying dimensions. The experi-
mental based comparisons of PA and HRUA are given next.
3 Experimental Results
The PA and HRUA algorithms were implemented using JDK 1.6 in Windows-XP
environment. The two algorithms were experimentally compared on an Intel based 2
GHz PC having 1 GB RAM. The comparisons were carried out on parameters like
TQA and TVEC for selecting the Top-10 views for materialization. The experiments
were conducted by varying the number of dimensions of the data set from 5 to 10.
First, graphs were plotted to compare PA and HRUA algorithms on TQA versus
number of dimensions. The graphs are shown in Fig. 4(a). It is observed from the
graph that the increase in TQA, with respect to number of dimensions, is higher for
PA vis-à-vis HRUA.
In order to ascertain the impact of higher TQA on TVEC due to views selected us-
ing PA, graphs for TVEC against number of dimensions were plotted and are shown
in Fig. 4(b). It is evident from the graph that the TVEC of PA is slightly more than
that of HRUA. This small difference shows that the PA selects views which are al-
most similar in quality to those selected by HRUA.
Profit
View Size QF 1st Iteration 2nd Iteration 3rd Iteration
AB 56 1032 88 - -
AC 68 1616 48 24 24
BC 62 1468 56 28 28
A 22 440 40 35 -
B 30 810 34 17 9
C 36 848 36 27 18
154 T.V. Vijay Kumar and M. Haider
It can be reasonably inferred from the above that PA trades significant improve-
ment in TQA with a slight increase in TVEC of views selected for materialization.
4 Conclusion
In this paper, an algorithm is proposed that selects Top-T views from a multidimen-
sional lattice using both the size and the query frequency of each view. The proposed
algorithm, in each iteration, computes the profit of each view using the size and query
frequency of the views and then selects, from amongst them, the most profitable view
for materialization. Unlike HRUA, the proposed algorithm is able to select fairly good
quality views that are able to account for large number of queries. This would result in
improvement in the average query response time.
The experiment based comparison of PA with HRUA on parameters TQA and
TVEC showed that PA was found to achieve higher TQA at the cost of a slight in-
crease in the TVEC in respect of views selected for materialization. That is, PA is
able to select views capable of answering significantly greater number of queries at
the cost of a slight drop in the quality of views selected for materialization.
References
1. Agrawal, S., Chaudhuri, S., Narasayya, V.: Automated Selection of Materialized Views
and Indexes in SQL Databases. In: Proceedings of VLDB 2000, pp. 496–505. Morgan
Kaufmann Publishers, San Francisco (2000)
2. Aouiche, K., Darmont, J.: Data mining-based materialized view and index selection in data
warehouse. Journal of Intelligent Information Systems, Pages, 65–93 (2009)
3. Baralis, E., Paraboschi, S., Teniente, E.: Materialized View Selection in a Multidimen-
sional Database. In: Proceedings of VLDB 1997, pp. 156–165. Morgan Kaufmann Pub-
lishers, San Francisco (1997)
4. Chirkova, R., Halevy, A., Suciu, D.: A Formal Perspective on the View Selection Problem.
The VLDB Journal 11(3), 216–237 (2002)
5. Gupta, H., Mumick, I.: Selection of Views to Materialize in a Data Warehouse. IEEE
Transactions on Knowledge and Data Engineering 17(1), 24–43 (2005)
Fi
g
. 4. PA Vs. HRUA – (TQA, TVEC) Vs. Dimensions
HRU Vs. PA
Top-10 views
0
500000
1000000
1500000
2000000
2500000
3000000
3500000
5678910
Dimension
TQA
PA
HRU
HRU Vs. PA
Top-10 views
0
1000000
2000000
3000000
4000000
5000000
6000000
5678910
Dimension
TVEC
PA
HRU
(b) TVEC - PA Vs. HRUA (a) TQA - PA Vs. HRUA
Selection of Views for Materialization Using Size and Query Frequency 155
6. Harinarayan, V., Rajaraman, A., Ullman, J.: Implementing Data Cubes Efficiently. In:
Proceedings of SIGMOD 1996, pp. 205–216. ACM Press, New York (1996)
7. Inmon, W.H.: Building the Data Warehouse, 3rd edn. Wiley Dreamtech, Chichester (2003)
8. Nadeau, T.P., Teorey, T.J.: Achieving scalability in OLAP materialized view selection. In:
Proceedings of DOLAP 2002, pp. 28–34. ACM, New York (2002)
9. Roussopoulos, N.: Materialized Views and Data Warehouse. In: 4th Workshop KRDB
1997, Athens, Greece (August 1997)
10. Serna-Encinas, M.T., Hoya-Montano, J.A.: Algorithm for selection of materialized views:
based on a costs model. In: Proceeding of Eighth International Conference on Current
Trends in Computer Science, pp. 18–24 (2007)
11. Shah, A., Ramachandran, K., Raghavan, V.: A Hybrid Approach for Data Warehouse
View Selection. Int. Journal of Data Warehousing and Mining 2(2), 1–37 (2006)
12. Teschke, M., Ulbrich, A.: Using Materialized Views to Speed Up Data Warehousing,
Technical Report, IMMD 6, Universität Erlangen-Nümberg (1997)
13. Vijay Kumar, T.V., Ghoshal, A.: A Reduced Lattice Greedy Algorithm for Selecting Mate-
rialized Views. CCIS, vol. 31, pp. 6–18. Springer, Heidelberg
14. Vijay Kumar, T.V., Haider, M., Kumar, S.: Proposing Candidate Views for Materializa-
tion. CCIS, vol. 54, pp. 89–98. Springer, Heidelberg (2010)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 156–161, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Key Validation Using Weighted-Edge Web of Trust
Model
Sumit Kumar1, Nahar Singh2, and Ashok Singh Sairam1
1 Department of Computer Science & Engineering
Indian Institute of Technology Patna
Patna, India
sumit.itech@gmail.com, ashok@iitp.ac.in
2 Department of Image Engineering
Chung-Ang University
Seoul, Korea Republic of
nahar.iitg@gmail.com
Abstract. Public key cryptography is widely used in the establishment of se-
cure communication. An important issue in the use of public key cryptography
is to ensure that a public key actually belongs to its owner. This problem is re-
ferred to as the key validation problem. One solution to this key validation
problem is the centralized public key infrastructure (PKI) model ([1]). A second
alternative is the decentralized Web of Trust model ([2]). In this model a user
depends on digital certificate introduced by other users, who may be trustwor-
thy or otherwise, to establish the binding between a public key and its owner. In
this work we propose to model the trust relationship between users as weighted
edges. We propose an algorithm to find trusted paths and thus validate public
keys. The proposed algorithm is simple and has a low complexity both in terms
of time and space.
Keywords: Web of Trust, Trust Hierarchy, Public key infrastructure, Key Vali-
dation Problem, PGP.
1 Introduction
The need to protect e-mail and data files has become very essential with the wide-
spread availability of attack tools. Public key encryption allows users to securely
communicate with each other over a non-secure channel without having to agree upon
a shared secret key beforehand. In public-key encryption a user has two keys - a pub-
lic key and a private key pair. The public key is made publicly available. Other users
wanting to securely communicate with a user encrypt their data with the public key of
the user. The encrypted data is decrypted by the user by using his private key which is
only known to him. In public key cryptography an important problem is to establish
that a public key actually belongs to its supposed owner. We refer to it as the key
validation problem. A solution to the key validation problem is the use of public-key
infrastructures (PKI) where one or several central authorities are responsible for issu-
ing digital certificates for public keys. Such a certificate is a trusted warranty for the
Key Validation Using Weighted-Edge Web of Trust Model 157
binding between the involved public key and its owner. Another solution to the key
validation problem is a distributed trust model called Web of Trust used in PGP,
GnuPG and other OpenPGP compatible systems.
In the Web of Trust model the users themselves (also called clients) are able to is-
sue public keys as introducers. The decision of whether to follow a trust endorsement
depends on the individual users. Given a user and a key endorsement by a client, it is
often possible for the user to find a chain of clients such that there exist bidirectional
trust relationship between adjacent clients. In this paper we propose to design the trust
relationship between users as weighted edges depending on the trust level of users.
Our proposed algorithm facilitates to find a trusted path. The paper is organized as
follows. Section 2 reviews the web of trust model and essential trust levels to validate
destination. Section 3 is describing related work and their defects while the section 4
provides the details of proposed key validation using weighted edge Web of Trust
model. Section 5 presents the theoretical analysis, while section 6 concludes.
2 The Web of Trust Model
The PGP trust model has some particular characteristics. First of all, (only) four levels
of trust are supported: complete trust, marginal trust, no trust and legitimate. The
owner of the key ring (Someone’s personal collection of certificates is called key
ring); who needs to manually assign these trust values for all other users, automati-
cally receives full trust (also called implicit or Ultimate Trust). When a user places
trust in an introducer, implicitly it means that the user possesses a certain amount of
confidence in the introducer’s capability to issue valid certificates, i.e. correct bind-
ings between users and public keys. Therefore, Trust levels can be one of these:
1. Undefined: we cannot say whether this public key is valid or not.
2. Marginal: This public key may be valid be we cannot be too sure.
3. Complete: we can be wholly confident that this public key is valid.
4. Legitimate: Deemed legitimate by you.
Based on such trust values, the PGP trust model suggests accepting a given public key
in the key ring as completely valid, if either
(a) The public key belongs to the owner of the key ring,
(b) The key ring contains at least C certificates from completely trusted introducers
with valid public keys,
(c) The key ring contains at least M certificates from marginally trusted introducers
with valid public keys.
As we discussed before PGP Trust Models vary in two categories, one of them is Web
of Trust. Web of trust is the encroachment of classical trust hierarchy model (HPKI).
Web of trust supports graph based structure consequently it is a de-centralized ap-
proach. In Web of Trust each client able to issue the authentication certificate as in-
troducer. The issuers of such certificates are called introducers, who can make them
publicly available. All clients placed in the graph as nodes which able to authorize
another node. There may be multiple paths between two entities & bi-directional trust
158 S. Kumar, N. Singh, and A.S. Sairam
relationship flow of trust in a two directions (node to node). Web of trust is not more
scalable & give difficulties in path discovery due to graph based structure.
3 Related Work
• Proposed model solves major part of key validation problem like limited trust
level, limited validity levels, counter-intuitive key validation and hidden key de-
pendencies. Proposed model gives strong assurance to trust someone if it covered
all necessary conditions of approach.
• Probabilistic key validation approach [3, 4] has reliability problems such as how
would you map trust with continuously varying probability values beside of this
there may be complex mathematical computation to compute key validity but
proposed weighted graph model is much reliable to compute trust with simple
calculations.
• Self organized key Based Trust [5, 6] is superior than other existing approaches
but its main concerns is only to have updated relationship status, but proposed
model able to adopt the problem start point to end.
4 The Proposed Weighted Edge Trust Model
In order to resolve the defects of existing methods, we put forward a new proposal as
“Weighted Edge Trust Model for PGP’’.
Fig. 1. Weighted Web of Trust Model Scenario
In general setting, the starting point of scenario as a ‘‘weighted edge graph ’’,
where graph represent the key ring, vertices are introducers, edges are certificates. If
an edge comes from a complete trust node, its weight will be ‘1’ and the key intro-
duced is said to be complete trust. If edge comes from a marginal trust node, its
weight will be ‘0.5’ (or less than 1 greater than 0) and the key introduced is said to be
Key Validation Using Weighted-Edge Web of Trust Model 159
marginal trust. If an edge comes from a un-trusted node, its weight will be‘0’ and the
key introduced is said to be not-trusted, as shown in Figure 1.
In this approach when a user D wants to validate the public key of S it will execute
the following algorithm:
Algorithm:
1. For each existing path S to D (prefer shortest path first).
2. {
3. if (weight= N) then
4. {
5. key trust is complete;
6. CreateNewEdge(S, D);
7. }
8. if(weight< N)
9. {
10. if(CheckEdgeWeight(E)!=0 && CheckWeight(D)>=1)
11. {
12. if ((exist alternate path S and D) && CheckEdgeWeight(E)!=0)
13. {
14. key trust is complete;
15. CreateNewEdge(S, D);
16. }
17. }
18. Else
19. {
20. key trust is incomplete or un-trusted;
21. }
22. }
23. }
- N = no. of edges between source & destination.
- Weight = total weight from source to destination (by adding weight of each
existing edge in path).
- CheckWeight ( ) = a function which calculate the weight of each incoming
edges at the each node in path (also consists destination node).
- CheckEdgeWeight ( ) =checks the weight of individual edges.
CreateNewEdge (S, D) = create new edge between source(S) to destination
(D) with weight = 1: where trust is complete.
- Source always will be complete trust.
- E=set of edges in path
- V=set of nodes in path
- Dotted edges created by CreateNewEdge (S, D) with weight ‘1’.
The input to the algorithm will be the weighted edge graph as shown in fig 1. In order
for a node (referred to as source) wants to validate the public key of another node
(referred to as destination) it will first try to find a path between them. There are three
possibilities. The first case can be if total weight of the edges between the two is equal
to no. of edges then it means all existing node in the path must be completely trusting
160 S. Kumar, N. Singh, and A.S. Sairam
except the last node. Then key introduced by this path will be complete trust. As
shown in fig.1, source A to destination G through D have all edges 1. Hence we can
create a new edge (CreateNewEdge( )) between them. This new edge will decrease
the time complexity for future communication between the nodes.
The second case can be if total weight between sources to destination path is less
than no. of edges & the weight is an integral value then it means there must be at least
one un-trusted node in the path. As shown in fig.1, the path from source A to destina-
tion H through D, G has a total weight of 2 but the number of edges is 3. Such a path
is an incomplete trust path. The third possibility is that the total weight between
sources to destination path is less than no. of edges but none of the edges have a
weight 0. The weight of the path will not be an integer value. Such a path may be
valid or invalid. If we can find another path between the source and destination with
exactly identical conditions than the source can be validated else it will be invalid. For
example in fig.1, the path from A to F through B has a weigh of 1.5 but the total num-
ber of edges is 2 and there are no edges with 0 weight. However there is another path
from A to F through C with weight of 1.5 and no edges with 0 weight. Thus F can
validate A.
5 Analysis of Proposed Algorithm
To analyze the efficiency of proposed algorithm, we apply the web of trust probabilis-
tic key validation method [7] in the scenario shown in figure 1. Assume client F wants
to validate A’s public key. The key validation in the probabilistic model consists of
two phases:
Phase 1: Determine all certificate paths leading from A to F.
Path1:[A, B, E, F], Path2:[A, B, F], Path3:[A, C, F]
Phase 2: Compute the probability that at least one operating path exists.
Let presume probabilities of nodes are as follow:
A=1.0, B=0.4, C=0.5, D=0.9, E=0.1, F=0.2, G=0.1, H=0.2.
The probability of a single path is simply the product of its (stochastically independ-
ent) trust values and for the overall probability of the set minpath (A, F), we can apply
the so-called inclusion-exclusion formula, so probability of Path1, Path2, and Path3
are 0.04, 0.4, and 0.5 respectively. Exact result (hard to compute): key validity of F =
P (path1, path2, path3, path4) = 0.488 (approx).
The result of whether F validates A certificate will depend on F own validation
policy. For example if F specifies the validity threshold as 0.5 then the key of A will
not be accepted.
On the other hand our proposed deterministic proposal will give a uniform result
and will not depend on the client’s validation policy. Moreover, the proposed algo-
rithm has a lower complexity as can be seen below.
• Time Complexity: Probabilistic key validation models calculate the trust at each
intermediary node so if a scenario has n number of nodes than the worst case
complexity would be O (n). In proposed model time complexity will be minimum
after creating the all possible edges it will be O(1).
Key Validation Using Weighted-Edge Web of Trust Model 161
• Space Complexity: Probabilistic web of trust models need to store information
about intermediary nodes but in proposed model, information of them would be
useless after creating direct edge consequently space complexity will decrease.
And formation of direct edges gives better path discovery.
• Ambiguity: Ambiguity is a major problem of probabilistic models. Probabilistic
model has to decide probability ranges to make trust. It is quite ambiguous that
what probabilities make sure complete trust but in proposed model, there is no
ambiguity trust in path due to frequent calculation.
6 Conclusion
Establishment of a hierarchical PKI is costly and lacks dynamicity. Further PKI being
a centralized approach there is the possibility of a single point of failure and the sys-
tem doesn’t allow users to make their own decisions. On the other hand, the decen-
tralized Web of Trust model is more scalable and do not suffer from the limitations of
PKI. In the Web of Trust model any client can act as an introducer and issue digital
certificates. However, this introducers or clients may not be completely trustworthy.
In this paper we propose to model the trust relationship between clients deterministi-
cally using a weighted edge graph. We propose a simple algorithm whereby a user
can validate digital certificates issued by the introducers.
References
1. Abdul-Rahman, A.: The PGP Trust Model: EDI-Forum (April 1997)
2. Zimmermann, P.R.: The Official PGP User’s Guide. MIT Press, Cambridge (1994)
3. Cristina, S., Rafael, P., Jordi, F.: PROSEARCH: A Protocol to Simplify Path Discovery in
Critical Scenarios. In: López, J. (ed.) CRITIS 2006. LNCS, vol. 4347, pp. 151–165.
Springer, Heidelberg (2006)
4. Rolf, H., Jacek, J.: A New Approach to PGP’s Web of Trust: ENISA European eIdentity
conference, Paris (2007)
5. Hideaki, K., Osamu, M., Hiroaki, N., Hiroshi, I.: Self-Organized Key Management based on
Trust Relationship List: Course of Computer and Communications. Graduate School of En-
gineering, Tokai University
6. Ross, D.: Pretty Good Privacy (1998-2008)
7. Jacek, J., Markus, W., Rolf, H.: A Probabilistic Trust Model for GnuPG. In: 23C3, 23rd
Chaos Communication Congress, pp. 61–66 (2006)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 162–169, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Novel Reconfigurable Architecture for Enhancing
Color Image Based on Adaptive Saturation Feedback
M.C. Hanumantharaju1, M. Ravishankar1, D.R. Rameshbabu2, and S. Ramachandran3
1 Department of Information Science & Engineering,
{mchanumantharaju, ravishankarmcn}@gmail.com
2 Department of Computer Science & Engineering,
Dayananda Sagar College of Engineering, Bangalore, India
bobramysore@gmail.com
3 Deptartment of Electrical & Electronics, SJB Institute of Technology, Bangalore, India
ramachandar@gmail.com
Abstract. In this paper, a novel architecture suitable for Field Programmable
Gate Array (FPGA) implementation of an adaptive color image enhancement
based on Hue-Saturation-Value (HSV) color space is presented. The saturation
feedback is used in order to enhance contrast and luminance of the color image.
The saturation component is enhanced by stretching its dynamic range. Hue is
preserved in order to avoid color distortion. The adaptive luminance enhance-
ment is achieved by using a simple arithmetic mean filter. An efficient architec-
ture for histogram equalization is also developed in order to evaluate the
performance of proposed algorithm. The algorithm is implemented on Xilinx
Vertex II XC2V2000-4ff896 FPGA device. The pipelining and parallel process-
ing techniques have been adapted in order to speed up the enhancement process.
The experimental results show that the color images enhanced by the proposed
algorithm are clearer, vivid and efficient.
Keywords: FPGA, Adaptive Color Image Enhancement, Saturation Feedback,
HSV, Arithmetic Mean Filter.
1 Introduction
Digital Image Enhancement [1] refers to accentuation, sharpening of image features
such as edges, boundaries, or contrast to make a graphic display more useful for dis-
play and analysis. The color image enhancement can be classified into two categories
according to the color space: (i) Color Image Enhancement in RGB color Space and
(ii) Color Image Enhancement based on Transformed space [2].
HSV color space discriminates between color and intensity and hence these spaces
reconstruct better images than is possible with the RGB space. In this work, HSV
color space is chosen since it offers good image enhancement. Further, arithmetic
mean filter [3] has been adopted that achieves very good quality reconstructed im-
ages. The algorithm is implemented on Xilinx Vertex II XC2V2000-4ff896 FPGA
device. The pipelining and parallel processing techniques have been adapted in order
A Novel Reconfigurable Architecture for Enhancing Color Image 163
to speed up the enhancement process. It is the first of its kind in the literature, which
FPGA implementation of adaptive color image enhancement [4] based on HSV color
space has been proposed.
2 Proposed Method
This paper proposes a new adaptive color image enhancement based on arithmetic
mean filter [5] in order to improve the quality of an image. The arithmetic means for
luminance and saturation may be expressed by Eqns. (1) and (2):
(, )
(, )w
ij w
VVij
∈
=
∑
(1)
(, )
(, )w
ij w
SSij
∈
=
∑
(2)
where m = 3 and n = 3 for 3×3 window. The local variance for luminance and satura-
tion are now expressed as: (3) and (4)
22
(, )
(, ) [ (, ) ]vw
ij w
xy Vij V
σ
∈
=−
∑
(3)
22
(, )
(, ) [ (, ) ]sw
ij w
xy Sij S
σ
∈
=−
∑
(4)
Further, the local correlation coefficient of luminance and saturation follows the
Eqn. (5).
(, )
22
[(,) (,)][(,) (,)]
(, ) (, ) (, )
ww
ij w
vs
Vxy V xy Sxy S xy
xy xy xy
ρσσ
∈−−
=∑
(5)
The new luminance enhancement with saturation feedback is given by the Eqn. (6).
12(,) (,) [(,) (,)] [(,) (,)] (,)enhV xy Vxy kVxy Vxy k Sxy Sxy xy
ρ
=+ − − −× (6)
The value of k1 and k2 was arrived at a value of 2 after conducting elaborate experi-
ments. In order to improve the whole effect of color image with brighter and richer
color the saturation component is enhanced by stretching its dynamic range. The
mathematical model for saturation component enhancement is given by Eqn. (7).
enhSS
γ
= (7)
where S represents original saturation component and Senh is the enhanced saturation
component. The value of gamma was arrived at a value of 0.77 after experiment
analysis. In this work, hue is preserved in order to avoid color distortion.
164 M.C. Hanumantharaju et al.
3 Hardware Implementation
The Fig. 1 shows the block diagram for FPGA implementation adaptive color image
enhancement system [6]. The proposed system has been realized using Register
Transfer Level (RTL) compliant, Verilog Hardware Description Language (HDL). In
the proposed work, the image size is chosen as 256×256 pixels.
Fig. 1. Block diagram of the Proposed Adaptive Image Enhancement System
The first step in the proposed scheme is RGB to HSV conversion. This module
transforms pixels from RGB space to HSV. RGB to HSV conversion needs a total of
18 clock cycles to complete the conversion process. The digital hardware for RGB to
HSV conversion is developed using Eqns. (8) to (10).
43
0,(,,)
(,,) (,,)
43
85 , ( , , )
(,,) (,,)
43
171 , ( , , )
(,,) (,,)
H
GB
M
ax R G B R
Max R G B Min R G B
BR
M
ax R G B G
Max R G B Min R G B
RG
M
ax R G B B
Max R G B Min R G B
=
⎧×−
+=
⎪−
⎪
⎪×−
⎪+=
⎨−
⎪
⎪×−
⎪+=
−
⎪
⎩
(8)
{
}
(, , ) (, , )
255 (, , )
SMax R G B Min R G B
Max R G B
=−
× (9)
(, , )VMaxRGB= (10)
The architecture for adaptive color image enhancement shown in Fig. 2 is developed
using Eqns. (6) and (7) described in the previous section. The enhanced pixels are
transformed back to RGB space using HSV to RGB converter module. The HSV to
RGB converter module takes 23 clock cycles to complete the conversion process. In
order to speed up the complete system pipelining and parallel processing technique
has been adapted. The Fig. 3 shows the hardware architecture for histogram equaliza-
tion. This module performs contrast enhancement [7] of an image. Histogram equali-
zation [8] module includes special decoder (8 to 256), 256 counters (16-bit), and a
memory 256 words (each word of 8-bits) followed by a combinational logic for pixel
mapping. In order to realize the digital hardware for histogram equalization we use
Eqn. (11)
A Novel Reconfigurable Architecture for Enhancing Color Image 165
0
(1)
k
kj
j
L
Sn
MN =
−
=
∑
(11)
where Sk is the discrete transformation, L represents intensity levels in the image
(L=256 for an 8-bit image), The product MN is the total number of pixels in the
image.
Fig. 2. Proposed Architecture for Adaptive Color Image Enhancement
H = Hue; S = Saturation; V = Value (Intensity); D = Delay; W = Window
Su = Sum; SQRE = Square; SQRT = Square Root; HENH = Hue Enhanced
SENH = Saturation Enhanced; VENH = Value (Intensity) Enhanced
Fig. 3. Proposed Architecture for Histogram Equalization
166 M.C. Hanumantharaju et al.
4 Experimental Results and Comparative Study
The proposed FPGA implementation of adaptive color image enhancement [9] based
on HSV color space has been coded and tested in Matlab (Version 8.1) first in order
to ensure the correct working of the algorithm. Subsequently, the complete system has
been coded in Verilog HDL so that it may be implemented on an FPGA or Applica-
tion Specific Integrated Circuit (ASIC). The system simulation is done using Model-
Sim (Version SE 6.4) and synthesized using Xilinx ISE 9.2i. The algorithm is imple-
mented on Xilinx Vertex II XC2V2000-4ff896 FPGA device. In this work, window
size was chosen as 3×3 since the image looks better and more colorful than the origi-
nal image. The experiment was conducted and presented in Fig. 4 by considering
three poor quality test images.
Fig. 4. Experimental Results of Tree, Tractor and Squares
First Column: Original Image; Second Column: Histogram Equalized Image obtained from
the Proposed Architecture shown in Fig. (3); Third Column: Enhanced Images obtained from
proposed hardware architecture shown in Fig. (2).
The first column of Fig. 4 shows three different original color images. The second
column of Fig. 4 shows the images enhanced by proposed histogram equalization of
R, G and B channels in RGB color space. The result shows that histogram equalized
images suffer from color distortion. The last column of Fig. 4 shows the images en-
hanced by proposed luminance enhancement algorithm based on saturation feedback.
A Novel Reconfigurable Architecture for Enhancing Color Image 167
Table 1. Summary of the FPGA Device Utilization XC2V2000-4ff896
Logic Utilization Used Available Utilization
No. of Slice Flip Flops 3128 21504 14%
No. of 4 input LUTs 1436 21504 6%
Logic Distribution
No. of Occupied Slices 3427 10752 31%
No. of Slices containing
only related logic
3427 4792 71%
No. of Slices containing
unrelated logic
0 4792 0%
Total No. of 4 input LUTs 7483 21504 34%
Total Equivalent Gate
Count for Design
64312
Table 1. shows the device utilization summary of FPGA implementation [10] adap-
tive color image enhancement based on adaptive saturation feedback. Experimental
results show that the color image enhancement based on the proposed adaptive lumi-
nance and saturation feedback offers much better enhanced images than that of other
methods. The reconstructed images offer richer color, clearer details and higher con-
trast. In order to evaluate the performance of the proposed method, we present con-
trast enhancement performance, luminance enhancement performance Peak Signal to
Noise Ratio (PSNR) and histogram plots. The contrast enhancement performance C,
Luminance enhancement performance L and PSNR is evaluated using the following
Equations.
out in
in
C
σσ
σ
−
=
(12)
out in
in
II
LI
−
=
(13)
2
10 255
10logPSNR
M
SE
⎡
⎤
=
⎢
⎥
⎣
⎦
(14)
The Mean Square Error (MSE) is given by
()
2
11
(, ) (, )
pq
xy
Exy Ixy
MSE pq
==
−
=∑∑ (15)
where σout and Iout are luminance variance and luminance mean values of the output
image, σin and Iin are luminance variance and luminance mean values the input image,
E(x, y) and I(x, y) are the enhanced and original gray pixel at position (x, y), p and q
denote the size of the gray image.
168 M.C. Hanumantharaju et al.
Table 2. Performance Comparison
Parameter Histogram Equalization Method
Tree Tractor Squares
Contrast 0.45 0.48 0.63
Luminance 0.53 0.61 0.77
PSNR 34.7 36.7 39.1
Parameter Proposed Method
Tree Tractor Squares
Contrast 0.41 0.44 0.64
Luminance 0.44 0.65 0.72
PSNR 32.5 33.6 34.2
Table 2. shows the performance comparison. It is clear that, the contrast enhance-
ment performance of the proposed method is good as compared to histogram equali-
zation. The luminance of the histogram equalization has increased; hence the image
looks over enhanced leading to color distortion. The proposed method has better
PSNR as compared to other methods.
5 Conclusion
This paper presented the adaptive color image enhancement and its FPGA implemen-
tation based on arithmetic mean filter. Arithmetic mean filter provides better visual
effects as compared with other filters. The proposed scheme uses HSV color space
since it discriminates between color and intensity. In addition, it offers good image
enhancement. In this work, efficient hardware architectures for RGB to HSV, HSV to
RGB, Adaptive image enhancement and histogram equalization has been developed.
In order to increase the processing speed the techniques such as pipeline and parallel
processing schemes has been adapted. The Verilog code developed for the complete
system is RTL compliant and works for ASIC design. The picture resolution of
256×256 pixels is chosen for this scheme. The implementation presented in this paper
has been realized on an FPGA. The experimental results show that the enhanced im-
ages are better in visual quality compared to other methods.
References
1. Gonzalez, R.C., Woods, R.E.: Digital Image Processing. Addison- Wesley, Reading
(1992)
2. Jain, A.K.: Fundamentals of Digital Image Processing. Prentice-Hall, Englewood Cliffs
(1989)
3. Thomas, B.A., Strickland, R.N., Heffrey, J.: Color Image Enhancement using Spatially
Adaptive Saturation Feedback. In: IEEE International Conference on Image Processing,
vol. 3, pp. 30–33 (1997)
4. Song, G., Qiao, X.-L.: Adaptive Color Image Enhancement based on Human Visual Prop-
erties. In: Proceedings of International Congress on Image and Signal Processing (2008)
A Novel Reconfigurable Architecture for Enhancing Color Image 169
5. Song, G., Qiao, X.-L.: Color Image Enhancement based on Luminance and Saturation
Components. In: Proceedings of International Congress on Image and Signal Processing
(2008)
6. Kokufuta, K., Maruyama, T.: Real-time Processing of Contrast Limited Adaptive Histo-
gram Equalization on FPGA. In: Proceedings of IEEE International Conference on Field
Programmable Logic and Applications, pp. 155–158 (2010)
7. He, K.-J., Chen, C.-C., Lu, C.-H., Wang, L.: Implementation of a New Contrast Enhance-
ment Method for Video Images, pp. 1982–1987 (2010)
8. Xie, X., Shi, Z., Guo, W., Yao, S.: An Adaptive Image Enhancement Technique based on
Image Characteristics. In: Proceedings of IEEE International Conference on Image and
Signal Processing, pp. 1–5 (2009)
9. Zhang, M.Z., Seow, M.-J., Asari, V.K.: A Hardware Architecture for Color Image
Enhancement Using a Machine Learning Approach with Adaptive Parameterization. In:
Proceedings of IEEE International Joint Conference on Neural Network, pp. 35–40 (2006)
10. Ngo, H.T., Zhang, M.Z., Tao, L., Asari, V.K.: Design of a Digital Architecture for Real-
Time Video Enhancement Based on Illuminance- Reflectance Model. In: Proceedings of
49th IEEE International Midwest Symposium on Circuits and Systems, pp. 286–290
(2006)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 170–177, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Signal Processing Approach for Prediction of Kink in
Transmembrane α-Helices
Jayakishan K. Meher1, Nibedita Mishra2, Pranab Kishor Mohapatra3,
Mukesh Kumar Raval4, Pramod Kumar Meher5, and Gananath Dash6
1 Department of Electronics & Telecom Engineering, SITE, Balangir, Odisha,
India -767002
jk_meher@yahoo.co.in
2 Department of Chemistry, Rajendra College, Balangir, Odisha, India -767002
nibedita1976@yahoo.com
3 Department of Chemistry, CV Raman College of Engineering, Bhubaneswar,
Odisha, India – 752054
pkmohapatra@yahoo.co.in
4 Department of Chemistry, Gangadhar Meher College, Sambalpur, Odisha, India – 768004
mraval@yahoo.com
5 Institute for Infocomm Research, Singapore – 138632
pkmeher@i2r.astar.edu.sg
6 School of Physical Sciences, Sambalpur University, Jyoti Vihar, Odisha, India – 768019
gndash@ieee.org
Abstract. The functions of transmembrane proteins are attributed by kinks
(bends) in helices. Kinked helices are believed to be required for appropriate
helix-helix and protein-protein interaction in membrane protein complexes.
Therefore, knowledge of kink and its prediction from amino acid sequences is
of great help in understanding the function of proteins. However, determination
of kink in transmembrane α-helices is a computationally intensive task. In this
paper we have developed signal processing algorithms based on discrete Fou-
rier transform and wavelet transform for prediction of kink in the helices with a
prediction efficiency of ~80%. The numerical representation of the protein in
terms of probability of occurrence of amino acids constituted in kinked helices
contains most of the necessary information in determining the kink location,
and the signal processing methods capture this information more effectively
than existing statistical and machine learning methods.
Keywords: DFT, Kink, Transmembrane α-Helices, Wavelet.
1 Introduction
Knowledge of segments of transmembrane proteins and the bends in helices help in
the study of tertiary structure and hence understanding the role played by that protein.
20-30% of all the proteins in any organism are membrane proteins. These are of par-
ticular importance because they form targets for over 60% of drugs on the market.
Transmembrane α-helix bundle is a common structural feature of membrane proteins
Signal Processing Approach for Prediction of Kink in Transmembrane α-Helices 171
except porins, which contains β-barrels. Membrane spanning α-helices differ from
their globular counterpart by the presence of helix breakers, Pro and Gly, in the mid-
dle of helices. Pro is known to induce a kink in the helix [1, 2]. A hypothesis suggests
that Pro is introduced by natural mutation to have a bend and later further mutated
leaving the bend intact for required function during the course of evolution [3]. The
role of Pro and kinks in transmembrane helices were extensively investigated both
experimentally and theoretically to unravel the nature's architectural principles [2,4].
Another observation suggest induction of kink at the juncture of α-helical and 310 heli-
cal structure in a transmembrane helix [2-6]. Mismatch of hydrophobicity of lipid
bilayer and peptide may also result in distortion of α-helical structure [7]. Sequences
of straight and kinked helices were further subjected to machine learning to develop a
classifier for prediction of kink in a helix from amino acid sequences. Support vector
machine (SVM) method [8] projects that helix breaking propensity of amino acid se-
quence determines kink in a helix. Kinked and straight helix of protein Type-4 Pilin
and Chlorophyll A-B binding protein respectively are shown in Fig.1.
(a) (b)
Fig. 1. Backbone representation of (a) Kinked helix of Type-4 Pilin protein (2pil) (b) Straight
helix (second helix) of Chlorophyll A-B binding protein (1rwt)
A kink in a helix may be formed by helix-helix interaction. In such cases the in-
trinsic kink forming or helix breaking tendency may not be required. Even a helix
forming tendency may be overridden. This possibility clamps a theoretical limit to
predict a kink with high accuracy. Hence there is a need to develop advanced algo-
rithm for faster and accurate prediction of kink in transmembrane helices. This moti-
vates to develop novel approach based on signal processing methods such as DFT and
wavelet transform to effectively predict kink in transmembrane α-helices.
2 Materials and Methods
Database. List of transmembrane proteins and their coordinate files were obtained
from the Orientation of Proteins in Membranes (OPM) database at College of Phar-
macy, University of Michigan (http://www.phar.umich.edu).
Determination of α-helical regions. Dihedral angles were computed using MAP-
MAK from coordinate files and listed for each residue along with assignment of
conformational status of the residue namely right or left helical, β-strand. Molecular
172 J.K. Meher et al.
visual tools RasMol were used to visually confirm the transmembrane α-helical
regions.
Computation of helix axis. Helix axis was computed from the approximate local
centroids θi’(xi
0,yi
0,zi
0) of the helix by taking a frame of tetrapeptide unit [9].
(1)
where xi, yi, and zi are the coordinates of Cα atoms of the tetrapeptide frame. Unit vec-
tor in the direction of resultant of vectors θ'iθ'i+1 yields direction cosines (l, m, n) of
axis of helix (A). The axis pass through the centroid of the helix θ0 = (X0, Y0, Z0).
(2)
where n is the number of residues in a helix. Refined local centers θi of helix are then
calculated for each Cα by computing the foot of perpendicular drawn from Cαi to A.
Location of hinges. Hinges were located in a helix by a distance parameter d(CiNi+4),
where Ci is the backbone carbonyl carbon of ith residue and Ni+4 is backbone peptide
nitrogen of i+4th residue [9]. Value of d(CiNi+4) beyond the range 4.227±0.35 Å
reflects a hinge at the ith residue in the helix. Hinge was quantified by two parameters
kink and swivel [3].
3 Proposed Signal Processing Methods for Kink Prediction
Signal processing methods such as Fourier transform and wavelet transform can iden-
tify periodicities and variations in signals from a background noise. This property of
signal processing approach plays a major role in prediction of helix kink in amino
acid sequence. In this paper the presence of kink in amino acid sequence is deter-
mined effectively in transform domain analysis.
In our approach, helix forming propensity of amino acid residues in a sequence
window of nine residues is taken as an input vector represented as xi є R9, where
i =1,2,....,N and N represents the number of samples.. The fifth residue of the
sequence exhibits kink in the sequence of nine in case of kinked helix. A sequence of
i-4 to i+4 is selected because to determine a kink at ith residue axes of i-4 to i and i to
i+4 is necessary and minimum five residues are required to determine axis of a helix.
Hence, a sequence of nine residues is minimum requirement for determination of
kink. The sequence xi is converted to numerical represention. The window of 9 resi-
dues slides across the sequence and at each position an average value in the transform
domain is calculated and assigned to the middle residue. The plot of the average value
of each window against the relative residue location is found to exhibit peaks at kink
locations which indicate the presence of a kink region in the amino acid sequence and
no such peak at straight helices. Based on this method two transformed domains such
as DFT and wavelet transform are demonstrated in this paper.
∑∑∑ +++ ===
3
0
3
0
3
0
4
1
,
4
1
,
4
1i
i
ii
i
i
ii
i
i
ii zzyyxx
∑∑∑ ===
===
n
i
i
n
i
i
n
i
iz
n
Zy
n
Yx
n
X
1
0
1
0
1
01
,
1
,
1
Signal Processing Approach for Prediction of Kink in Transmembrane α-Helices 173
DFT Based Approach. Discrete Fourier transform plays a role in the implementation
of digital signal processing algorithm for kink prediction. It is found that DFT of a
given input amino acid sequence having kink helices exhibits peaks at the kink
locations thus detecting periodicity in the sequence. The amino acid sequence of 9
characters is converted into numerical sequence by substituting its probability of oc-
currence of residues (Table 1) from a known 500 dataset of OPM database. Alter-
nately physicochemical property such as polarizability as shown in Table 1 that has
been computed from Hyperchempro 8.0 software play also similar role.
Table 1. Physico-chemical properties of amino acids
Residues
Probability of
occurrence
Polarizability Residues Probability of
occurrence
Polarizabil-
ity
A 0.8492 4.44 L 1.0908 9.95
R 0.6979 14.16 K 1.4422 10.72
N 1.1960 7.72 M 1.0261 11.11
D 0.8988 6.55 F 1.2654 14.10
C 1.2513 7.44 P 10.000 8.79
Q 0.8762 11.39 S 0.8676 5.08
E 1.6414 8.38 T 0.7235 6.92
G 0.9859 2.61 W 1.0967 19.37
H 0.8137 11.84 Y 0.898 14.74
I 0.6388 9.95 V 0.8624 8.11
For example, the aminoacid sequence x(n) = [SWWNFGSLL], the corresponding
numerical sequence xi(n) substituting probability of occurrence of residues is xi(n) =
[0.8676 1.0967 1.0967 1.1960 1.2654 0.9859 0.8676 1.0908 1.0908]. The DFT of
an N-point of sequence xi(n) is defined as
(3)
where N is the length of the segment of amino acid sequence. The window slides
across the sequence and at each position an average DFT value for the nine residues is
calculated and assigned to the middle residue. The frequency spectrum of X(k) is
found to exhibit peaks at kink locations which indicate the presence of a kink region
in the amino acid sequence. The plot shows that there are remarkable peaks in kinks
(bend) helices and no such peak in straight (nonbend) helices as shown in Fig.2 and
Fig.3. This unique property is used to predict the kink in amino acid sequence.
Wavelet transform based approach. Recently, the use of wavelet transform, both
continuous and discrete in the Bioinformatics field is promising. Continuous Wavelet
Transform (CWT) allows one-dimensional signal to be viewed in a more discrimina-
tive two-dimensional time-scale representation. CWT is calculated by the continuous
shifting of the continuously scalable wavelet over the signal. In discrete wavelet trans-
form (DWT) a subset of scales and positions are chosen, in which the correlation be-
tween the signal and the shifted and dilated waveforms are calculated. Consequently,
the signal is decomposed into several groups of coefficients, each containing signal
features corresponding to a group of frequencies. Small scales refer to compressed
∑
−
=
−−≤≤⋅=
1
0
/2 10 ,)()(
N
n
Nknj
iNkenxkX
π
174 J.K. Meher et al.
wavelets, depicted by rapid variations appropriate for extracting high frequency
features of the signal. An important attribute of wavelet methods is that, due to the
limited duration of every wavelet, local variations of the signal are better extracted
and information on the location of these local features is retained in the constituent
waveforms.
DWT has been applied on hydrophobicity signals in order to predict hydrophobic
cores in proteins [10]. Protein sequence similarity has also been studied using DWT
of a signal associated with the average energy states of all valence electrons of each
amino acid [11]. Wavelet transform has been applied for transmembrane structure
prediction [12]. In this work, the wavelet transform is used to determine kink in seg-
ments of amino acid sequences of α-helical membrane proteins.
A wavelet is a waveform that is localised in both time and frequency domains. This
wavelet is dilated and translated along the signal to perform the analyses. The com-
monly used wavelets in practice are Haar, Daubechies, Gaussian wave, Mexican hat
and Morlet wavelets. The selection of particular wavelet for any analysis depends on
the kind of signal being studied and kind of signal variation to be captured. In case of
analysis of protein sequence signal the Mexican hat wavelet seemed to be choice. The
mother wavelet of Mexican hat wavelet is defined as
(4)
The wavelet transform gives rise to patterns that are distinct between the kink regions
from straight helix regions. To analyze the protein sequence for prediction of kink, it
is first transformed into a numerical signal based on probability of occurrence of resi-
dues along a protein sequence.
A sliding window of 9 residues has been used. Wavelet coefficients are computed
by translating the wavelet along the signal. The window slides across the sequence
shifted by one residue with the window size of nine residues. At each position an
average coefficients for the nine residues is calculated and assigned to the middle resi-
due. The length of the signal produced across the protein sequence is equal to the
number of residues of the protein. The plot of average value against relative residues
is found to exhibit peaks at kink locations that indicate the presence of kink regions in
the amino acid sequence of transmembrane α-helices as shown in Fig. 2 and Fig.3.
(a) (b)
Fig. 2. Cytochrome BC1 Complex Protein (pdb id: 1BGY) showing four kinked helices includ-
ing two overlapped kinks by (a) DFT, (b) Wavelet transform
2/2 2
)1()( t
ett −
−=
ψ
0 20 40 60 80
0
0.2
0.4
0.6
0.8
1
Relative Residue Locations
Magnitude
0 20 40 60 80
0.2
0.4
0.6
0.8
1
Relative Residue Locations
Magnitude
Signal Processing Approach for Prediction of Kink in Transmembrane α-Helices 175
.
(a)
(b)
0 50 100
0
0.2
0.4
0.6
0.8
1
Relative Residue Locations
Magnitude
0 50 100
0.2
0.4
0.6
0.8
1
Relative Residue Locations
Magnitude
Fig. 3. Patassium Channel protein (pdb id: 1BL8) showing three kink helices including two
overlapped kinks by (a) DFT, (b) Wavelet transform.
4 Result and Discussion
We have used the DFT and wavelet transform techniques to detect the kink locations
using numerical representation based on probability of occurrence of amino acid resi-
dues and polarizability, physico-chemical properties obtained from Hyperchempro 8.0
software of HyperCubeInc, USA. 200 proteins data sets used as bench mark for this
purpose. In a good number of cases all the proposed methods performed well. Helix
forming propensity of amino acid residues in a sequence window of nine residues is
taken as an input vector of dimension 9. The fifth residue of the sequence exhibits
kink in the sequence of nine in case of kinked helix. For example in protein Cyto-
chrome BC1 Complex with accession no: 1JBY, proposed method using DFT and
wavelet transform have detected all four kinked helices including two overlapped
kinks as shown in Fig. 2. Again in Potassium Channel protein with Accession
No: 1BL8, three kink helices including two overlapped kinks are identified by the
proposed methods as shown in Fig. 3.
The list of proteins under study and the corresponding results obtained using
proposed methods are shown in Table 2. It is found that the proposed methods show
better accuracy in predicting kink helices. The performance analysis of various meth-
ods can be made by prediction measures such as accuracy (A), precision (P) and recall
(R) which are defined in terms of four parameters true positive (tp), false positive (fp),
true negative (tn) and false negative (fn). tp denotes the number of actual kinks and are
also predicted as kinks, fp denotes the number of actually straight helices but are pre-
dicted to be kinks, tn is the number of actually straight helices and also predicted to be
straight helices, and fn is the number of actually kinks and predicted to be straight
helices.
Accuracy. The accuracy of prediction kink in amino acid sequence is defined as the
percentage of helices correctly predicted of the total helices present. It is computed as:
; (5)
nnpp
np
ftft
tt
A+++
+
=
helicesofnumberTotal
spredictioncorrectofNumber
A
=
176 J.K. Meher et al.
Precision. Precision is computed for kink and straight helix classes separately. It is
defined as the percentage of kinks correctly predicted to be one class of the total kink
predicted to be of that class. Precision is computed as:
; (6)
Recall. Recall is also computed separately for kink and straight helix classes. It is
defined as the percentage of the kinks that belong to a class that are predicted to be
that class. Recall is computed as:
; (7)
Table 2. Accuracy comparisons of various techniques of OPM protein data set
Prediction measures
PDB Id., Protein Name Methods
A P R
SVM 0.75 1 0.75
DFT .8 1 0.8
1BL8, Potassium
channel protein
DWT .8 1 0.8
SVM 0.75 0..75 1
DFT 0.8 0.8 1
2PIL, Type-4 Pilin,
DWT 0.8 0.8 1
SVM 0.8 1 0.8
DFT 0.88 1 0.88
1BGY, Cytochrome
BC1 complex
DWT 0.88 1 0.88
SVM 0.66 1 0.66
DFT 0.75 1 0.75
1JBO, C-Phycocyanin
Alpha chain
DWT 0.75 1 0.75
SVM 0.75 1 0.75
DFT 0.75 0.75 1
1NEK , Succinate De-
hydrogenase Flavopro-
tein DWT 0.75 0.75 1
SVM 0.8 1 0.8
DFT 88 1 0.88
2BL2, V-Type Sodium
ATP synthase SUB-
UNIT K DWT 88 1 0.88
The result suggests that non-helix former amino acids may induce kink. But no
such correlation with single amino acid residues at kink is observed, rather helix
formers namely Ala, Leu, Ile, Phe, more frequently occur at kink position. However,
at the i+4th position Pro and Gly occur with higher frequency. Again no correlation
is observed between propensity for helix formation of i+4th and frequency of its
occurrence in kinks at i+4th position. Combinations of residues in a definite sequence
may be responsible for non-Pro kinks in helices [13]. It appears that there may be a
non-linear correlation between helix forming tendency of amino acid residues and
formation of kink.
pp
p
ft
t
P+
=
predictedkinksofnumberTotal
kinkspredictedcorrectlyofNumber
P
=
kinksactualofnumberTotal
kinkspredictedcorrectlyofNumber
R
=np
p
ft
t
A+
=
Signal Processing Approach for Prediction of Kink in Transmembrane α-Helices 177
5 Conclusion
Signal processing approach plays a vital role in the prediction of kink in transmem-
brane α-helix. The proposed method is not only fast but also has improved accuracy
(more than 80%) as compared to SVM learning system reported by us earlier [8].
However prediction of kink in a helix depends on the features of amino acid se-
quence. Feature vector with probability of occurrence of residues and polarizability
are only used for numerical representation in the present study. Although kink predic-
tion has its own limitations, the present work is the first report in the area of helix
kink prediction from amino acid sequence based on signal processing algorithms.
References
1. Ramachandran, G., Ramakrishnan, C., Sasisekharan, V.: Stereochemistry of Polypeptide
Chain Configuration. J. Mol. Biol. 7, 95–97 (1963)
2. Sankararamakrishnan, R., Vishveshwara, S.: Conformational Studies on Peptides with
Proline in the Right-Handed-Helical Region. Biopolymers 30, 287–298 (1990)
3. Cordes, F., Bright, J., Sansom, M.P.: Proline Induced Distortions of Transmembrane Heli-
ces. J. Mol. Biol. 323, 951–960 (2002)
4. von Heijne, G.: Proline Kinks in Transmembrane-Helices. J. Mol. Biol. 218, 499–503
(1991)
5. Yohannan, S., Faham, S., Whitelegge, J., Bowie, J.: The Evolution of Transmembrane He-
lix Kinks and the Structural Diversity of G-protein Coupled Receptors. Proc. Natl. Acad.
Sci. U.S.A. 101, 959–963 (2004)
6. Pal, L., Dasgupta, B., Chakrabarti, P.: 3(10)-Helix Adjoining Alpha-helix and Beta-strand:
Sequence and Structural Features and Their Conservation. Bioploymers 78, 147–162
(2005)
7. Daily, A., Greathouse, D., van der Wel, P., Koeppe, R.: Helical Distortion in Tryptophan-
and Lysine-Anchored Membrane-Spanning Alpha-Helices as a Function of Hydrophobic
Mismatch: A Solid-State Deuterium NMR Investigation using the Geometric Analysis of
Labeled Alanines Method. Biophys. J. 94, 480–491 (2008)
8. Mishra, N., Khamari, A., Mohapatra, P.K., Meher, J.K., Raval, M.K.: Support Vector Ma-
chine Method to Predict Kinks in Transmembrane Helices, pp. 399–404. Excel India Pub-
lishers, India (2010)
9. Mohapatra, P.K., Khamari, A., Raval, M.K.: A Method for Structural Analysis of Helices
of Membrane Proteins. J. Mol. Model. 10, 393–398 (2004)
10. Hirakawa, H., Muta, S., Kuhara, S.: The Hydrophobic Cores of Proteins Predicted by
Wavelet Analysis. Bioinformatics 15, 141–148 (1999)
11. de Trad, C., Fang, Q., Cosic, I.: Protein Sequence Comparison Based on the Wavelet
Transform Approach. Protein Eng. 15, 193–203 (2002)
12. Murray, K.B., Gorse, D., Thornton, J.: Wavelet Transforms for the Characterization and
Detection of Repeating Motifs. J. Mol. Biol. 316, 341–363 (2002)
13. Chou, K., Nemethy, G., Scheraga, H.: Energetic Approach to the Packing of Helices. 2.
General treatment of Nonequivalent and Nonregular Helices. J. Am. Chem. Soc. 106,
3161–3170 (1984)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 178–183, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Cascaded H-Bridge Multilevel Boost Inverter without
Inductors for Electric/Hybrid Electric Vehicle
Applications
S. Dhayanandh1, A.P. Ramya Sri1, S. Rajkumar1, and N. Lavanya2
1 Assistant Professor, Department of ECE,
Kathir College of Engineering
Coimbatore, India
dhais@sify.com, ramyaashree123@gmail.com, rajped22@gmail.com
2 Assistant Professor,Department of ECE,
Maharaja Engineering College
Coimbatore, India
lavanya.nataraj87@gmail.com
Abstract. This paper presents a cascaded H-bridge multilevel boost inverter for
electric vehicle (EV) and hybrid EV (HEV) applications implemented without
the use of inductors. Currently available power inverter systems for HEVs use a
dc–dc boost converter to boost the battery voltage for a traditional three-phase
inverter. A cascaded H-bridge multilevel boost inverter design for EV and HEV
applications implemented without the use of inductors is proposed in this paper.
The proposed design uses a standard three-leg inverter (one leg for each phase)
and an H-bridge in series with each inverter leg which uses a capacitor as the dc
power source. Experiments show that the proposed dc–ac cascaded H-bridge
multilevel boost inverter can output a boosted ac voltage without the use of
inductors.
Keywords: Cascaded H-bridge multilevel boost inverter, Three-leg inverter,
electric vehicle (EV)/hybrid electric vehicle (HEV).
1 Introduction
Now a day’s increasing of oil prices and environmental issues, hybrid electric ve-
hicles (HEVs) and electric vehicles (EVs) are gaining increased attention due to their
higher efficiencies and lower emissions associated with the development of improved
power electronics and motor Technologies.HEV typically combines a smaller internal
combustion engine of a conventional vehicle with a battery pack and an electric motor
to drive the vehicle. This combination having advantage of lower emissions. And
convenient fueling of conventional (gasoline and diesel) vehicles. An EV typically
uses rechargeable batteries and an electric motor. Here batteries want to be charged
regularly. Both HEVs and EVs need a traction motor and a power inverter to drive the
traction motor. The requirements for the power inverter include high peak power and
low continuous Power rating. Currently available power inverter systems for HEVs
Cascaded H-Bridge Multilevel Boost Inverter without Inductors 179
use a dc–dc boost converter to boost the battery voltage for a traditional three-phase
inverter. If the motor is running at low to medium power, in this case dc–dc boost
converter is not necessary; here battery voltage will be directly applied to the inverter
to drive the traction motor. If the motor is running in a high power mode, the dc–dc
boost converter will boost the battery voltage to a higher voltage, so that the inverter
can provide higher power to the motor. Present HEV traction drive inverters have low
power density, are expensive, and have low efficiency because they need bulky induc-
tors for the dc–dc boost converters. To achieve a boosted output ac voltage from the
traditional inverters for HEV and EV applications, the Z-source inverter is proposed,
which also requires an inductor. A cascaded H-bridge multilevel boost inverter shown
in Fig. 1 for EV and HEV applications is described in this paper. Traditionally, each
H-bridge of a cascaded multilevel inverter needs a dc power supply. The proposed
cascaded H-bridge multilevel boost inverter uses a standard three-leg inverter (one leg
for each phase) and an H-bridge in series with each inverter leg which uses a capaci-
tor as the dc power source. In this topology, the need for large inductors is eliminated.
A fundamental switching scheme is used to do modulation control and to output five-
level phase voltages. Experiments show that the proposed dc–ac cascaded H-bridge
multilevel boost inverter without inductors can output a boosted ac voltage.
2 Working Principle of Cascaded H-Bridge Multilevel Boost
Inverter without Inductors
The topology of the proposed dc–ac cascaded H-bridge multilevel boost inverter is
shown in Fig. 1. The inverter uses a standard three-leg inverter and an H-bridge with a
capacitor as its dc source in series with each phase leg.
Fig. 1. Topology of the proposed dc-ac cascaded H-bridge multilevel boost inverter
2.1 Switching Technique of Cascaded H-Bridge Multilevel Boost Inverter
without Inductors
There are several kinds of modulation control methods such as traditional sinusoidal
pulse width modulation (SPWM),space vector PWM, harmonic optimization or selec-
tive harmonic elimination, and active harmonic elimination, and they all can be
180 S. Dhayanandh et al.
used for inverter modulation control. For the proposed dc–ac boost inverter control, a
practical modulation control method is the fundamental frequency switching control
for high output voltage and SPWMcontrol for low output voltage, which only uses the
bottom inverter. In this paper, the fundamental frequency switching control is used.
(1)
The key issue of fundamental frequency modulation control is choice of the two
switching angles θ1 and θ2. In this paper, the goal is to output the desired fundamen-
tal frequency voltage and to eliminate the fifth harmonic. Mathematically, this can be
formulated as the solution to the following:
cos(θ1) + cos(θ2) =ma
cos(5θ1) + cos(5θ2) =0. (2)
This is a system of two transcendental equations with two unknowns θ1 and θ2, and
ma is the output voltage index. Traditionally, the modulation index is defined as
(3)
Fig. 2. Switching angle solutions for proposed dc–ac cascaded H-bridge multilevel boost inver-
ter control
Therefore, the relationship between the modulation index m and the output voltage
index ma is
(4)
There are many ways one can solve (2) for the angles. Here, the resultant method is
used to find the switching angles. A practical solution set is shown in Fig. 2, which is
continuous from modulation index 0.75 to 2.42. Although it can be seen from Fig. 2
that the modulation index range for the five-level fundamental frequency switching
control method can reach 2.42, which is double that of the traditional power inverter,
it requires the capacitors’ voltage to be kept constant at Vdc/2. Traditionally, the max-
imum modulation index for the linear operation of a traditional full-bridge bi-level
inverter using SPWM control method is 1 (without third harmonic compensation) and
Cascaded H-Bridge Multilevel Boost Inverter without Inductors 181
1.15 (with third harmonic compensation, and the inverter output voltage waveform is
an SPWM waveform, not a square waveform). With the cascaded H-bridge multilevel
inverter, the maximum modulation index for linear operation can be as high as 2.42;
however, the maximum modulation index depends on the displacement power factor,
as will be shown in the next section.
3 Output Voltage Boost
As previously mentioned, the cascaded H-bridge multilevel inverter can output a
boosted ac voltage to increase the output power, and the output ac voltage depends on
the displacement power factor of the load. Here, the relationship of the boosted ac
voltage and the displacement power factor is discussed. It is assumed that the load
current displacement angle is
φ
. To balance the capacitor voltage, the net capacitor
charging amount needs to be greater than the pure discharging amount. The inverter
can regulate the capacitor’s voltage with a displacement power factor of one if the
modulation index is below 1.27; if the modulation index is above 1.27, the displace-
ment power factor must be less than a specified amount. For practical applications,
the highest output voltage is determined when the load is determined. As mentioned
previously, there are many methods to do modulation control for the proposed dc–ac
cascaded H-bridge multilevel boost inverter without inductors. The fundamental fre-
quency method with regulated Vdc/2 capacitor voltage is only one of the possible me-
thods to output continuous power. The traditional SPWM method can also be applied
to this inverter to boost the output voltage with a lower maximum continuous output
power and high switching loss but better THD for a lower output frequency range. It
is also possible to use SPWM for low output frequency low output voltage conditions
and staircase waveform for high output frequency high output voltage range to
achieve optimal performances with maximum continuous output power, lower switch-
ing loss, and lower THD. It can also be seen that accurate load inductance is not re-
quired for controller design, and the controller is robust independent of the leakage
inductance of stator windings. For HEV and EV applications, sometimes, only short
period peak Power is required. The modulation control can store energy to the capaci-
tors by boosting the capacitor voltage to a higher voltage, which could be higher than
Vdc when the vehicle is working in a low power mode. When the vehicle is working
in high power modes, the capacitors will deliver much higher power than the conti-
nuous power to the motor load combined with the battery, fuel cell, or generator. This
feature will greatly improve the vehicle’s dynamic (acceleration) performance.
4 Experimental Implementation and Validation
A real-time variable output-voltage variable-frequency three-phase motor drive con-
troller based on an Altera FLEX 10 K field programmable gate array (FPGA) is used
to implement the control algorithm. For convenience of operation, the FPGA controller
is designed as a card to be plugged into a personal computer, which uses a peripheral
component interconnect bus to communicate with the microcomputer. To maintain the
capacitors ‘voltage balance, a voltage sensor is used to detect the capacitors voltage
182 S. Dhayanandh et al.
and feed the voltage signal into the FPGA controller. The FPGA controller will output
the corresponding switching signals according to the capacitor’s voltage. To further
test the cascaded multilevel boost inverter, experiments with load current versus mod-
ulation indexes with different fundamental frequencies were performed to achieve the
highest output voltages. These were implemented by using an R–L load bank and com-
pared to a traditional inverter. For these experiments, the R–L load was fixed, the mod-
ulation index was changed with different fundamental frequencies, and the load cur-
rents were recorded. In this experiment, to achieve the highest output voltages for the
cascaded multilevel boost inverter without inductors and the traditional inverter, two
steps were involved. First, the load was connected to the bottom traditional
inverter to output its highest voltage; second, the load was connected to the cascaded
H-bridge multilevel inverter with the same dc power supply voltage. The output vol-
tages for the two cases are shown in Table 1. Table 1 shows that the highest output
voltage of the cascaded H-bridge multilevel inverter is much higher than that of the
Traditional inverter. The voltage boost ratio is higher than 1.4 for the whole testing
frequency range.
Table 1. Highest Output Voltage For Traditional Inverter And Cascaded H-Bridge Multilevel
Inverter (Dc Bus Is 40 V)
Table 1 also shows that the highest output voltage of the inverter is decreasing
when the frequency is decreasing; this is because the impedance of the inductor is
decreasing. Another issue is that the boost voltage ratio is decreasing when the
frequency is decreasing; this is because the power factor is increasing for the fixed
R–L load.
5 Conclusion
The proposed cascaded H-bridge multilevel boost inverter without inductors uses a
standard three-leg inverter and an H-bridge in series with each inverter leg.The appli-
cation of this dc–ac boost inverter on HEV and EV can result in the elimination of the
bulky inductor of present dc–dc boost converters, thereby increasing the power density.
References
1. Du, Z., Tolbert, L.M., Chiasson, J.N.: DC–AC Cascaded H-Bridge Multilevel Boost Inver-
ter With No Inductors for Electric/Hybrid Electric Vehicle Applications. IEEE Transac-
tions On Industry Applications 45(3) (May/June 2009)
Cascaded H-Bridge Multilevel Boost Inverter without Inductors 183
2. Rahman, K.M., Patel, N.R., Ward, T.G., Nagashima, J.M., Caricchi, F., Crescimbini, F.:
Application of direct-drive wheel motor for fuel cell electric and hybrid electric vehicle
propulsion system. IEEE Trans. Ind. Appl. 42(5), 1185–1192 (2006)
3. Hinkkanen, M., Luomi, J.: Braking scheme for vector-controlled induction motor drives
equipped with diode rectifier without braking resistor. IEEE Trans. Ind. Appl. 42(5),
1257–1263 (2006)
4. Rivetta, C.H., Emadi, A., Williamson, G.A., Jayabalan, R., Fahimi, B.: Analysis and con-
trol of a buck dc–dc converter operating with constant power load in sea and undersea ve-
hicles. IEEE Trans. Ind. Appl. 42(2), 559–572 (2006)
5. Lai, J.S., Peng, F.Z.: Multilevel converters—A new breed of power converters. IEEE
Trans. Ind. Appl. 32(3), 36–44 (1996)
6. Lai, J.S., Rodriguez, J., Lai, J., Peng, F.: Multilevel inverters: A survey of topologies, con-
trols and applications. IEEE Trans. Ind. Appl. 49(4), 724–738 (2002)
7. Tolbert, L.M., Peng, F.Z., Habetler, T.G.: Multilevel converters for large electric drives.
IEEE Trans. Ind. Appl. 35(1), 36–44 (1999)
8. Jacobina, C.B., dos Santos, E.C., de Rossiter Correa, M.B., da Silva, E.R.C.: AC motor
drives with a reduced number of switches and boost inductors. IEEE Trans. Ind.
Appl. 43(1), 30–39 (2007)
9. Ben-Brahim, L., Tadakuma, S.: A novel multilevel carrier-based PWM-control method for
GTO inverter in low index modulation region. IEEE Trans. Ind. Appl. 42(1), 121–127
(2006)
10. Schuch, L., Rech, C., Hey, H.L., Grundling, H.A., Pinheiro, H., Pinheiro, J.R.: Analysis
and design of a new high-efficiency bidirectional integrated ZVT PWM converter for dc-
bus and battery-bank interface. IEEE Trans. Ind. Appl. 42(5), 1321–1332 (2006)
11. Shen, M., Wang, J., Joseph, A., Peng, F.Z., Tolbert, L.M., Adams, D.J.: Constant boost
control of the Z-source inverter to minimize current ripple and voltage stress. IEEE Trans.
Ind. Appl. 42(3), 770–778 (2006)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 184–189, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Design of Microstrip Meandered Patch Antenna for
Mobile Communication
Shobhit Patel1,*, Jaymin Bhalani2, Yogesh Kosta3, and Sanket Patel4
1,4 Assistant Professor, 2 Associate Professor, 3 Professor
Electronics & Communication Department, Charotar University of Science and Technology,
Changa-388 221, Gujarat, India
shobhit_65@yahoo.com
Abstract. The enhancing bandwidth and size reduction mechanism that im-
proves the performance of a conventional micro strip patch antenna on a rela-
tively thin substrate (about 0.006 λ ), is presented in this research. The design
adopts meandered patch structure. Introducing the novel meandered square
patch, offer a low profile, broadband, high gain, and compact antenna element.
The proposed patch has a compact dimension of 0.384 λ × 0.384 λ (where λ is
the guided wavelength of the centre operating frequency). The design is suitable
for applications with respect to a given frequency of 750-1100 MHz. The simu-
lated bandwidth of the proposed antenna is about 39%.
Keywords: Microstrip Patch Antenna, Meandered Patch.
1 Introduction
With the ever-increasing need for mobile communication and the emergence of many
systems, it is important to design broadband antennas to cover a wide frequency range.
The design of an efficient wide band small size antenna, for recent wireless applications,
is a major challenge. Microstrip patch antennas have found extensive application in
wireless communication system owing to their advantages such as low-profile, conform-
ability, low-cost fabrication and ease of integration with feednetworks [1]. However,
conventional microstrip patch antenna suffers from very narrow bandwidth, typically
about 5% bandwidth with respect to the center frequency. This poses a design challenge
for the microstrip antenna designer to meet the broadband techniques [2][3].
There are numerous and well-known methods to increase the bandwidth of anten-
nas, including increase of the substrate thickness, the use of a low dielectric substrate,
the use of various impedance matching and feeding techniques, the use of multiple
resonators, and the use of slot antenna geometry [4][5][6][7]. However, the bandwidth
and the size of an antenna are generally mutually conflicting properties, that is, im-
provement of one of the characteristics normally results in degradation of the other.
Recently, several techniques have been proposed to enhance the bandwidth. A
novel meandered patch antenna with achievable impedance bandwidth of greater than
39% has been demonstrated [9].
* Corresponding author.
Design of Microstrip Meandered Patch Antenna for Mobile Communication 185
In this paper, a novel meandered shape patch is investigated for enhancing the im-
pedance bandwidth on a thin substrate (about 0.006 λ ). In this paper, the design and
simulations results of the novel wideband micro strip patch antenna, is described.
Many techniques have been reported to reduce the size of microstrip antennas at
a fixed operating frequency. In general, microstrip antennas are half-wavelength
structures and are operated at the fundamental resonant mode TM01 or TM10, with a
resonant frequency given by (valid for a rectangular microstrip antenna with a thin
microwave substrate)
r
L
c
f
ε
2
≅ (1)
where c is the speed of light, L is the patch length of the rectangular micro strip an-
tenna, and εr is the relative permittivity of the grounded microwave substrate.
Meandering the excited patch surface current paths in the antenna’s radiating patch
is also an effective method for achieving a lowered fundamental resonant frequency
for the micro strip antenna [9][10][11]. For the case of a rectangular radiating patch,
the meandering can be achieved by inserting several narrow slits at the patch’s nonra-
diating edges. It can be seen in Fig.1 that the excited patch’s surface currents are ef-
fectively meandered, leading to a greatly lengthened current path for a fixed patch
linear dimension. This behavior results in a greatly lowered antenna fundamental
resonant frequency, and thus a large antenna size reduction at a fixed operating fre-
quency can be obtained.
Fig. 1. Surface current distributions for meandered rectangular microstrip Patch with meander-
ing slits
2 Designing and Modeling
It is known that increasing the thickness of the patch antenna will increase the imped-
ance bandwidth. However, the thicker the substrate of the antenna, the longer the
coaxial probe will be used and, thus, more probe inductance will be introduced [8],
which limits the impedance bandwidth. Consequently, a patch antenna design that can
counteract or reduce the probe inductance will enlarge the impedance bandwidth.
Fig.2 depicts the geometry of the proposed patch antenna. The square patch,
with dimensions as given in Table 1. is supported by a low dielectric substrate with
186 S. Patel et al.
dielectric permittivity ε (2.2). it is in between patch and ground plane. Here the patch
contains two slits on both the sides of centerline of the x-axis as shown in Fig.2 mak-
ing total four slits. The dimensions of slits are given in Table 1.
Fig. 2(a). Actual HFSS Model (top view) of meandered microstrip patch
Fig. 2(b). Actual HFSS Model (side view) of meandered microstrip patch
Table 1. Dimentions of Patch Antenna
Size in m.m
Patch 128×128×0.5
Substrate 135×135×2
Slit1 6×126×0.5
Slit2 4×126×0.5
Slit3 6×126×0.5
Slit4 4×126×0.5
3 Simulation Results and Discussions
For simulation we used High Frequency Structure Simulator (HFSS) version 11 of
ansoft, which is very good simulator for RF antennas. After simulating the design the
results we got is as follows. Fig. 3 shows the Return Loss(S11) plot of the design.
Feed
Slit 3
Slit 1 Slit 4
Slit 2
Design of Microstrip Meandered Patch Antenna for Mobile Communication 187
0.70 0.75 0.80 0.85 0.90 0.95 1.00 1.05 1.10 1.15
Freq [GHz]
-40.00
-35.00
-30.00
-25.00
-20.00
-15.00
-10.00
-5.00
dB(S(WavePort1,WavePort1))
Ansoft Corporation HF SSDesign1
XY Plot 1
Curv e Inf o
dB(S(WavePort1,WavePort1))
Setup1 : Sw eep1
Fig. 3. S11 paramater of the antenna
0.70 0.75 0.80 0.85 0.90 0.95 1.00 1.05 1.10 1.15
Freq [GHz]
1.00
1.50
2.00
2.50
3.00
3.50
VSWR(WavePort1)
Ansoft Corporation HFSSDesign1
XY Plot 2
Curv e Inf o
VSWR(WavePort1)
Setup1 : Sw eep1
Fig. 4. VSWR of the antenna
Table 2. S11 & VSWR values
Frequency
MHz
VSWR
Return loss(S11)
dB
Frequency
MHz
VSWR
Return loss(S11)
dB
750 1.86 -10 930 1.09 -27
760 1.84 -11 940 1.15 -23
770 1.85 -11 950 1.23 -20
780 1.87 -10 960 1.26 -18
790 1.88 -10 970 1.32 -17
800 1.87 -10 980 1.36 -16
810 1.83 -11 990 1.40 -16
820 1.78 -11 1000 1.43 -15
830 1.71 -12 1010 1.46 -15
840 1.63 -12 1020 1.48 -14
850 1.55 -13 1030 1.49 -14
860 1.46 -15 1040 1.53 -13
870 1.37 -16 1050 1.56 -13
880 1.28 -18 1060 1.62 -12
890 1.20 -21 1070 1.68 -12
900 1.10 -25 1080 1.72 -11
910 1.05 -31 1090 1.81 -11
916 1.03 -36 1100 1.87 -10
188 S. Patel et al.
Fig. 4 shows the VSWR plot of the design and Table 2 shows values of S11 &
VSWR for different frequencies. For the whole range VSWR less than 2 and S11 is
less than -10 dB and at frequency 916 MHz both are minimum
-9.20
-6.40
-3.60
-0.80
90
60
30
0
-30
-60
-90
-120
-150
-180
150
120
Ansoft Corporation HFSSDesign1
Radiation Pattern 1
Curv e Info
dB(rETheta)
Setup1 : Sw eep1
Fig. 5. Radiation Pattern of the antenna
Fig.5 shows the radiation pattern for frequency 916 MHz. Same way patterns for
other frequencies can be generated.
The driven model simulations are carried out for frequency sweep of 700 MHz
to 1150 MHz as shown in Fig. 3-5 under first order fast mode time domain analysis-
set-up. The solution frequency is 900 MHz with linear step size of 0.0004 GHz,
Maximum number of passes taken was 6, maximum delta S to be 0.02 and error toler-
ance was assumed to be 0.5%.
4 Conclusion and Future Scopes/Perspectives
The theoretical-geometric investigations and analysis, design, modeling and iterative
simulations are carried out for center frequency of 900 MHz. The result indicates the
total band of 750 MHz to 1100 MHz. So the antenna can be used for GSM 900,
CDMA 800 and CDMA 850 bands. Further the band can be changed by modifying
the design and further it can be designed for GSM 1800. The results are in very good
agreement with the industry and standard published antenna-requirements with re-
spect to ease of fabrication, compactness and volume miniaturization compared to
other antennas so far designed for similar applications.
References
1. He, W., Jin, R., Geng, J.: E-Shape patch with wideband and circular polarization for mil-
limeter- wave communication. IEEE Transactions on Antennas and Propagation 56(3),
893–895 (2008)
2. Lau, K.L., Luk, K.M., Lee, K.L.: Design of a circularly-polarized vertical patch antenna.
IEEE Transactions on Antennas and Propagation 54(4), 1332–1335 (2006)
3. Zhang, Y.P., Wang, J.J.: Theory and analysis of differentially-driven microstrip antennas.
IEEE Transactions on Antennas and Propagation 54(4), 1092–1099 (2006)
Design of Microstrip Meandered Patch Antenna for Mobile Communication 189
4. Matin, M.M., Sharif, B.S., Tsimenidis, C.C.: Probe fed stacked patch antenna for wide-
band applications. IEEE Transactions on Antennas and Propagation 55(8), 2385–2388
(2007)
5. Pozar, D.M., Schaubert, D.H.: Microstrip Antennas. IEEE press, New York (1995)
6. Wi, S.H., Sun, Y.B., Song, I.S., Choa, S.H., Koh, I.S., Lee, Y.S., Yook, J.G.: Package-
Level integrated antennas based on LTCC technology. IEEE Transactions on Antennas and
Propagation 54(8), 2190–2197 (2006)
7. Wi, S.H., Kim, J.M., Yoo, T.H., Lee, H.J., Park, J.Y., Yook, J.G., Park, H.K.: Bow-
tieshaped meander slot antenna for 5 GHz application. In: Proc. IEEE Int. Symp. Antenna
and Propagation, vol. 2, pp. 456–459 (2002)
8. Yang, F., Zhang, X., Rahmat-Samii, Y.: Wide-band E-shaped patch antennas for wireless
communications. IEEE Transactions on Antennas and Propagation 49(7), 1094–1100
(2001)
9. Dey, S., Mittra, R.: Compact microstrip patch antenna. Microwave Opt. Technol. Lett. 13,
12–14 (1996)
10. Wong, K.L., Tang, C.L., Chen, H.T.: Acompact meandered circular microstrip antenna
with a shorting pin. Microwave Opt. Technol. Lett. 15, 147–149 (1997)
11. George, J., Deepukumar, M., Aanandan, C.K., Mohanan, P., Nair, K.G.: New compact mi-
crostrip antenna. Electron. Lett. 32, 508–509 (1996)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 190–195, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Building Gaussian Mixture Shadow Model for Removing
Shadows in Surveillance Videos
Archana Chougule and Pratap Halkarnikar
Shivaji University, Kolhapur
Maharashtra, India
chouguleab@gmail.com, pp_halkarnikar@rediffmail.com
Abstract. Moving cast shadows are a major concern for foreground detection
algorithms. Processing of foreground images in surveillance applications typi-
cally requires that such shadows have been identified and removed from the de-
tected foreground.. It is critical for accurate object detection in video streams
since shadow points are often misclassified as object points, causing errors in
segmentation and tracking. Processing of foreground images in surveillance ap-
plications typically requires that such shadows have been identified and re-
moved from the detected foreground. We have implemented a novel pixel based
statistical approach to model moving cast shadows of non-uniform and varying
intensity by modelling each pixel as a mixture of Gaussians and using an on-
line approximation to update the model.
Keywords: Gaussian mixture model, cast shadow, chromaticity.
1 Introduction
Shadows result from the obstruction of light from the light source. Shadows cause
serious problems while segmenting and extracting moving objects due to the misclas-
sification of shadow points as foreground. To obtain a better segmentation quality,
detection algorithms must correctly separate foreground objects from the shadows
they cast. Model-based approaches have shown less robustness than property-based
algorithms [1]. A review of shadow detection techniques can be found in [2]. In [3],
Salvador et al. use the fact that a shadow darkens the surfaces on which it is cast to
identify an initial set of shadowed pixels that is then pruned by using color invariance
and geometric properties of shadows. In [4], Cucchiara et al. use the hypothesis that
shadows reduce surface brightness and saturation while maintaining hue properties in
the HSV color space. Schreer et al. [5] adopt the YUV color space to avoid using the
time consuming HSV color transformation and segment shadows from foreground
objects based on the observation that shadows reduce the YUV pixel value linearly. In
[6], Horprasert et al. build a model in the RGB color space to express normalized
luminance variation and chromaticity distortions. [7] uses edge width information to
differentiate penumbra regions from the background. The algorithm in [8] combines
luminance, chrominance, and gradient density scores in a shadow confidence score
function for segmentation.
Building Gaussian Mixture Shadow Model for Removing Shadows 191
In this paper, we present an approach for detecting and modelling moving cast
shadows using Gaussian mixture model (GMM) in surveillance video.
2 Gaussian Mixture Model
The foreground object is detected as described in [9] and modified for online imple-
mentation in [10]. For each pixel, a fixed number of states K, typically between 3 and
5, is defined. Some of these states will model the YUV values of background sur-
faces, and the others, foreground surfaces. Each pixel value Xt is a sample in a YUV
color space of a random variable X. The K states are ordered decreasingly by their
ωk/
ǁσ
k
ǁ
ratio and the first B states whose combined probability of appearing is greater
than the threshold T , i.e.,
)min(arg
1
TWkB
b
k
>= ∑
=
(1)
are labelled as background states, and the other states are labelled as foreground
states.
2.1 Shadow Modelling
Our approach is based on the hypothesis that, for a given pixel, the shadow cast by
different moving foreground objects is relatively similar. The states that capture cast
shadows are identified, and their pdfs are used to build a second GMM, which we call
the Gaussian mixture shadow model (GMSM).
When using a GMM to build a background model for foreground detection, it is
possible to observe that some of the states model the shadows cast by persons moving
across the scene. A first person crosses the scene, and on pixels where its shadow is
cast, a new state representing the value of this shadow will be created with a small a
priori probability ωinit. When the next person crosses the scene and casts a shadow
on the same pixels, the algorithm will associate the pixel values to the same new
state describing the shadow value. The a priori probability of this state will then
increase as
ωk,t+1 = ωk,t + αs,t Mk,t (2)
where αs,t is the learning parameter, Mk,t is equal to 1 for the state that is associated to
the pixel value, and zero for the other states. When the pixel value could describe a
shadow over the surface background, we increase the learning rate of the pixel associ-
ated state parameter ωk. This modification allows a state representing a shadow on a
background surface to rapidly become a stable foreground state and this without pre-
venting the creation of states describing the background. To do so, we define
αs,t = αt S (3)
where the parameter S > 1 when Xt correspond to a shadowed background surface,
and S = 1 in the other cases. When there are no persons or objects crossing the scene
for a long time, the a priori probability wk of the foreground states modelling the cast
shadows will tend toward zero. Since wk will become smaller than winit , any future
192 A. Chougule and P. Halkarnikar
detection of a foreground event will result in a new state and the destruction of fore-
ground states capturing the values of shadow surfaces.
2.2 Properties of Shadowed Surfaces
As in [5], we adopted the YUV color space and our model is also based on the obser-
vation that a shadow cast on a surface will equally attenuate the value of its three
components. We first estimate this attenuation ratio using the luminance component
Y, and we then verify that both U and V components are also reduced by a similar
ratio. More specifically, if color vector X represents the shadow cast on a surface
whose average color value is μ with variance σμ, we have
αmin < αY < 1 with αY = XY /μY (4)
(1/σμU )|XU − αY μU| < ΛU (5)
(1/σμV )|XV − αY μV | < ΛV (6)
where αmin is a threshold on maximum luminance reduction. ΛU,V represent the toler-
able chromaticity fluctuation around the surface value μU, V. If we match a pixel
value Xt to a non-background state of the GMM, we then verify if that pixel value
could be the shadowed surface of the most frequent background state k = 1:
αmin < αY < 1 with αY = Xt,Y /μ1,t,Y (7)
(1/σ1,μ,t,U ) |Xt,U − αY μ1,t,U | < ΛU (8)
(1/σ1,μ,t,V ) |Xt,V − αY μ1,t,V | < ΛV (9)
If these three conditions are met, we increase the learning rate αs,t of the a priori prob-
ability ωk of that state.
2.3 Gaussian Mixture Shadow Models
Updating the Gaussian mixture shadow model is the next step. Unlike the GMM [9],
the GMSM filters the input values, and these input values are Gaussian probability
density functions fX|k with parameters θk ={μk, σk}. These Gaussian pdf are those of the
states which have been identified as describing cast shadows on background surfaces.
We use the following conditions on fX|B+1 to determine if this state describes a cast
shadow on the background surface of state k = 1:
αmin < αY < 1 with αY = μB+1,t,Y /μ1,t,Y (10)
(1/σ1,μ,t,U ) |μB+1,t,U − αY μ1,t,U | < ΛU (11)
(1/σ1,μ,t,V ) |μB+1,t,V − αY μ1,t,V | < ΛV (12)
Building Gaussian Mixture Shadow Model for Removing Shadows 193
If these conditions are met, pdf fX|B+1 is transfered to the GMSM. It is then compared
to the existing GMSM pdfs:
dTk,t dk,t < λ2s,a
where
dk,t = (σsk,t I)−1(μB+1,t − μsk,t) (13)
and where we use the superscript s when referring to the Gaussian mixture shadow
model. If there is a match, the parameters θsk are updated:
μsk,t+1 = (1 − α) μsk,t + α μB+1,t (14)
σsk,t+1 = (1 − α) σsk,t + α σB+1,t (15)
where α is a constant. Also, we set Msk,t = 1 in the following equation :
ωsk,t+1 = ωsk,t + αMsk,t (16)
If there is no match, a new state is added in the GMSM, upto a maximum of Ks states.
For this new state,
ωsk,t+1 = ωsinit (17)
μsk,t+1 = μB+1,t (18)
σsk,t+1 = σB+1,t (19)
The a priori probabilities ωsk,t are then normalized, and the states sorted in decreasing
order of ωsk,t. We do not use the ratio ωk/||σk|| since the variance is relatively constant
for all states in the GMSM. Within the GMSM, the first Bs states, where
)(minarg
1
,
∑
=
>= b
k
s
tk
s
b
sTwB (20)
are used to model moving cast shadows on a background surface.
2.4 Removing Shadows
In a GMM, a pixel is labelled as foreground if its value Xt is matched to a foreground
state. This detected pixel is then compared to the shadow states of the GMSM
dTk,t dk,t < λ2s,b
where
dk,t = (σsk,t I)−1(Xt − μsk,t) (21)
If this condition is met for a state ωsk with k ≤ Bs, the pixel is labeled as representing
a moving cast shadow. It is then simply removed from the detected pixels of the
194 A. Chougule and P. Halkarnikar
GMM. We are then left with detected pixels representing foreground objects without
their cast shadows.
3 Experimental Results
The method explained above has been tested on a number of image sequences. The
experiments were performed on Intel Core2Duo CPU and the algorithm was imple-
mented using Matlab 7.6. Results shown here are raw results, without any post treat-
ment. For each environment, parameters were set once. Fig. 1 shows results for the
human object scene when GMM is applied. Here moving object along with the
shadow is extracted. Fig. 2 shows the result after applying GMSM.
Fig. 1. Moving object detection
Fig. 2. Result of GMSM
4 Conclusion
In this paper, we have described a system for removing shadows from a video, which
is pixel-based statistical approach to model and detect moving cast shadows.
The proposed approach uses a GMM to learn from repetition the properties of the
Building Gaussian Mixture Shadow Model for Removing Shadows 195
shadowed background surfaces. The algorithm identifies distributions that could rep-
resent shadowed surfaces, modify their learning rates to allow them to converge
within the GMM, and then uses them to build a GMM for moving shadows on back-
ground surfaces. This approach can be used with various shadow models in different
color spaces.
References
1. Prati, A., Mikic, I., Trivedi, M.M., Cucchiara, R.: Detecting Moving Shadows: Algorithms
and Evaluation. IEEE Trans. Pattern Analysis and Machine Intelligence 25(7), 918–923
(2003)
2. Nadimi, S., Bhanu, B.: Physical Models for Moving Shadow and Object Detection in
Video. IEEE Trans. Pattern Analysis and Machine Intelligence 26(8) (August 2004)
3. Salvador, E., Cavallaro, A., Ebrahimi, T.: Cast Shadow Segmentation Using Invariant
Color Features. Computer Vision and Image Understanding, 238–259 (2004)
4. Cucchiara, R., Grana, C., Piccardi, M., Prati, A., Sirotti, S.: Improving Shadow Suppres-
sion in Moving Object Detection with HSV Color Information. In: Proc. Intelligent Trans-
portation Systems Conf., pp. 334–339 (2001)
5. Schreer, O., Feldmann, I., Goelz, U., Kauff, P.: Fast and Robust Shadow Detection in Vid-
eoconference Applications. In: Proc. Fourth IEEE Int’l Symp. Video Proces. and Multime-
dia Comm., pp. 371–375 (2002)
6. Horprasert, T., Hardwood, D., Davis, L.S.: A Statistical Approach for Real-Time Robust
Background Subtraction and Shadow Detection. In: Proc. Int’l Conf. Computer Vision
FRAMERATE Workshop (1999)
7. Stauder, J., Mech, R., Ostermann, J.: Detection of Moving Cast Shadows for Object Seg-
mentation. IEEE Trans. Multimedia 1(1), 65–76 (1999)
8. Fung, G.S.K., Yung, N.H.C., Pang, G.K.H., Lai, A.H.S.: Effective Moving Cast Shadows
Detection for Monocular Image Sequences. In: Proc. 11th Int’l Conf. Image Analysis and
Processing (ICIAP), pp. 404–409 (2001)
9. Stauffer, C., Grimson, W.E.L.: Learning Patterns of Activity Using Real-Time Tracking.
IEEE Trans. Pattern Analysis and Machine Intelligence 22(8), 747–757 (2000)
10. Power, P.W., Schoonees, J.A.: Understanding Background Mixture Models for Fore-
ground Segmentation. Proc. Image and Vision Computing, 267–271 (2002)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 196–202, 2011.
© Springer-Verlag Berlin Heidelberg 2011
FAutoREDWithRED: To Increase the Overall
Performance of Internet Routers
K. Chitra1 and G. Padmavathi2
1 Associate Professor, Dept. of Computer Science,
D.J. Academy for Managerial Excellence, Coimbatore,
Tamil Nadu, India – 641032
chitrakandaswamy@yahoo.com
2 Department of Computer Science,
Avinashilingam university for Women, Tamil Nadu, India
Abstract. Active Queue Management is a solution to the problem of congestion
control in Internet routers. The problem of congestion degrades the overall per-
formance of Internet routers. As data traffic is bursty in routers with varying
flows, burstiness must be handled intelligently without comprising the overall
performance. A congested link leads to many problems such as large delay, un-
fairness among flows, underutilization of the link and packet drops in burst. In
this paper, we propose an AQM scheme that considers the characteristics of
queue length based AQMs and uses the flow information to increase the overall
performance and to satisfy the QOS requirements of the network.
Keywords: Congestion, Drop Probability, Fairness, Queue length, Misbehav-
ing flows.
1 Introduction
A router in the Internet may receive heavy traffic at any given time due to high use of
real-time applications. A Internet receives varying flows that should receive its fair share
while sharing queue in the router. Another possibility is that the heavy load tends to vary
in an Internet router resulting in a queue oscillation. Router must take care of the above
problem. The robustness of today’s Internet depends heavily on the congestion control
mechanism. So the buffer in the routers is to be used effectively by using an efficient
Active Queue Management Mechanisms. Active Queue Management prevents conges-
tion and provides quality of service to all users. A router implementing RED AQM [1]
maintains a single queue that drops an arriving packet at random during periods of con-
gestion. RED suffers from lockout and global synchronization problems when parame-
ters are not tuned properly. RED allows unfair bandwidth allocation when a mixture of
traffic types traverses a link. Adpative RED [2], AutoRED with RED [3] tries to im-
prove the parameter tuning problem in RED. While AQMS like PD-RED [4], MRED
[5], DS-RED [6] tries to improve the performance compared to RED.
Some AQMs [7] [8] techniques were introduced that used input rate as congestion indi-
cators besides using average queue size. AQM [9] was proposed that used both congestion
indicators queue size and input rate to detect congestion where as BLUE [10] uses link
history and packet loss as congestion indicator to compute the packet drop probability.
FAutoREDWithRED: To Increase the Overall Performance of Internet Routers 197
The objective of this paper is to propose an algorithm that improves the overall per-
formance of Internet routers. Because of the simplicity and easy implementation that
exists in AutoREDwithRED, this paper takes this Queue based AQM as the base algo-
rithm with certain modifications. The proposed AQM is implemented in this to bring in
the advantages of Queue based algorithm and to remove the problems like parameter
tuning problem. This algorithm is simple to implement, improves the overall perfor-
mance of the routers. The rest of the paper is organized as follows: Section 2 explains
the background study that includes the various AQM algorithms. In Section 3, the con-
cepts regarding the proposed algorithm are discussed. In section 4, the simulation is
carried out with discussion of results. Our conclusions are presented in section 5.
2 Background
The Internet routers face the problem of congestion from the birth of real-time appli-
cations in Internet. So research in this field of congestion has become a continuous
process to bring out the best Active Queue Management algorithm to improve the
performance of Internet even in case of heavy traffic. The various existing AQMs
detect congestion based on different factors and calculate the packet dropping proba-
bility. Floyd et al proposed the first RED AQM in 1993 with the objective of
preventing congestion with reduced packet loss. This AQM alleviates congestion by
detecting incipient congestion early and delivering congestion notification to the
source to reduce its transmission rates avoiding overflow from occurring. But the ap-
propriate selection of the RED parameters defines the success of RED. Incase of
heavy traffic, RED AQM also leads to global synchronization, lock-out problem and
unstable queue size if parameters not properly tuned. Further in improving RED and
RED based AQMs, AutoRED technique was implemented in them that used the con-
cept of dynamic wq.
The YELLOW AQM proves that the packet drop probability just does not depend
only on the queue length rather on input rate also that helps in identifying the real
congestion in the queue. In case of the rate-based AQM AVQ, it maintains a virtual
queue whose capacity is less than the actual capacity of the link. In improving the
method for setting the value for γ, SAVQ [11] is proposed. SAVQ stabilizes the dy-
namics of queue maintaining high link utilization. In REM, both queue length and
load is used as congestion indicators. The BLUE algorithm resolves some of the prob-
lems of RED by employing two factors: packet loss from queue congestion and link
utilization. So BLUE performs queue management based on packet loss and link
utilization. It maintains a single probability pm to mark or drop packets. SRED in [12]
pre-emptively discards packets with a load-dependent probability when a buffer in a
router is congested. It stabilizes its buffer occupancy at a level independent of the
number of the active connections. GREEN [13] algorithm uses flow parameters and
the knowledge of TCP end-host behavior to intelligently mark packets to prevent
queue build up, and prevent congestion from occurring.
198 K. Chitra and G. Pa
d
3 Proposed Algorith
The proposed algorithm is
queue size and fair bandw
i
gestion characteristics of t
h
tuning problem results in
h
tivated to identify a sche
m
queue size. The proposed
queue-based and uses the
f
the Queue based AutoRE
D
This algorithm calculates t
h
al. As in Fig. 1, average q
u
every arriving packet. If a
v
the average queue size is g
r
randomly selected packet
flow id, both are dropped.
with a probability depen
d
AutoRED technique wher
e
RED. This algorithm will
and parameterized.
Fig
.
The dynamic value of
w
traffic. The w
q
is redefin
e
oscillation. The definition
o
characteristics as follows:
T
t
= p
t
(1-p
t
)
J
w
q,t
= T
t
* J
t
* K
t
For every packet arr
i
Calculate Q
ave
if (Q
ave
< min
th
)
Forward the
Else {
Select randomly
a
Compare arrivin
g
If they have the s
a
Else
if (Q
ave
m
Calculate
Drop the
p
Else
}
}
d
mavathi
m
motivated by the need for a stable operating point for
i
dth allocation irrespective of the dynamic traffic and
c
h
e n flows. An unstable queue size due to the param
e
h
igh queue oscillation in queue based AQMs. We are
m
m
e that penalizes the unresponsive flows with the st
a
algorithm - FAutoREDWithRED enforces the concep
t
f
low information. We propose an algorithm that modi
D
withRED algorithm to penalize the unresponsive fl
o
h
e average queue size of the buffer for every packet ar
r
u
eue size is compared with the thresholds min
th
, max
h
,,
v
erage queue size is less than min
th
, every arriving W
r
eater than min
th
, every arriving packet is compared wi
t
from the queue for their flow id. If they have the s
a
In this proposed algorithm, the arriving packet is drop
p
d
ing on the average queue size. This algorithm
u
e
w
q
is dynamic in nature compared to a constant w
q
work fine as the parameters are well tuned automatic
a
.
1. Pseudocode of FAutoREDwithRED
w
q
adapts itself to the varying nature of the congestion
e
d as in [4].and results in reduced instantaneous q
u
o
f weighting parameter w
q
is written as a product of t
h
J
t
= K
t
=
i
val {
new packet
a
packet from the queue for their flow id
g
packet with a randomly selected packet.
a
me flow id
Drop both the packets
m
ax
th
)
the dropping probability p
a
p
acket with probability p
a
Drop the new packet
the
c
on-
e
ter
m
o-
a
ble
t
of
fies
o
ws.
r
iv-
for
hen
t
h a
a
me
p
ed
u
ses
q
in
a
lly
and
u
eue
h
ree
FAutoREDWithRED: To Increase the Overall Performance of Internet Routers 199
The first characteristic Tt represents the dynamic status of the congestion in the
network. It signifies the probability with which the system can lead to congestion with
the information available at time t. The second characteristic Jt projects the current
status of traffic in the network at time t. Tt and Jt is a time dependent function. The
third characteristic Kt is time independent parameter and it allows normalization of
instantaneous queue size changes with respect to the buffer size (bs). Therefore these
three characteristics are used to incorporate the dynamic changes in the congestion
and traffic in the calculation of average queue size.
Though the dynamic varying nature of wq takes care of the network characteristics
it keeps the average queue length high and in a unstable point in case of heavy traffic.
This algorithm overcomes this problem with the help of the flow based information.
So both wq and the flow information take care of the unresponsive flows and brings in
stable fair queuing. Packets of unresponsive flows are dropped more often than the
adaptive flows and well behaved flows to bring in fairness.
4 Experimentation
In this section, we will use the packet-simulator ns-2 to simulate the FAutoRED-
WithRED algorithm. In this simulation the network topology is as in Fig. 2. The con-
gestion link is in between the two routers R1 and R2. The TCP flows are derived from
FTP sessions which transmit large size files. The UDP hosts send packets at a con-
stant bit rate of 2 Mbps. In the simulation setup we consider 32 TCP flows and 1 UDP
flow in the network. The minimum threshold minth in the FAutoREDWithRED
scheme is set to 100 and the maximum threshold maxth to be twice the minth and the
physical queue size is fixed at 300 packets.
Fig. 2. Network Topology
In a dynamic varying mixture of traffic, the control parameter wq alone does not
help in achieving the stable operating point for the queue size. As shown in Fig. 3
dynamic varying parameter wq maintains the average queue size and instability at a
higher level in case of AutoREDwithRED for dynamic varying traffic. But this algo-
rithm shows a stable and a low queue size compared to other AQMs as in Fig 3.
RED and other AQMs are unable to penalize unresponsive flows. The misbeha-
vingtraffic like UDP can take up a large % of the link bandwidth and starve out TCP
200 K. Chitra and G. Pa
d
Fig. 3. Compari
s
friendly flows in existing
A
nalizes misbehaving flows
existing RED based AQM
s
uses only 2% of the link c
a
throughput by limiting the
U
link capacity. The total T
C
FAutoREDwithRED.
Fig. 4. CBR and TC
Fig. 5. Comparison of CBR
WithRED
To enforce the degree t
o
tion, CBR throughput of U
D
the Fig 5. A misbehaving
incurs high packet droppin
g
the entire bandwidth for
U
d
mavathi
s
on of Average queue size/Queue Stability of AQMs
A
QMs as in Fig 4. FAutoREDWithRED identifies and
effectively compared to the existing AQMs as in Fig
4
s
, the UDP flow uses 98% of the link capacity while
T
a
pacity. This algorithm indicates the improvement of
T
U
DP throughput to 212 Kbps which is around 21 % o
f
C
P throughput is increased from 2 Kbps to 750 Kbp
s
P Utilisation of other AQMs with FAutoREDWithRED
Fair share and Throughput of other AQMs with FAutoR
E
o
which FAutoREDwithRED gains fair bandwidth all
o
D
P connection along with their ideal fair share is show
n
flow with a high arrival rate and high buffer occupa
g
mostly due to matches. The other AQMs almost tak
e
U
DP flow showing unfairness though its actual CBR
pe-
4
. In
T
CP
T
CP
f
the
s
in
E
D-
o
ca-
n
in
a
ncy
e
up
fair
FAutoREDWithRED:
T
Fig. 6. Comparison of
P
share is very minimum. B
u
is only 21% of the bandwi
d
REDwithRED tries to giv
e
UDP utilises the link to the
5 Conclusions
This paper proposes an A
Q
crease the overall perform
a
case of mixture of traffic a
t
prising high utilisation, lo
w
with well dynamically tun
e
scheme discriminates agai
n
cation. This proposed AQ
M
based AQM and uses fl
o
network.
References
[1] Floyd, S., Jacobson, V
IEEE/ACM Trans. Net
w
[2] Floyd, S., Gummadi, R
.
the robustness of RED’
s
http://www.icir.
[3] Suthaharan, S.: Reduct
i
Science Direct, Comput
e
[4] Sun, J., Ko., K.-T.,
C
Performance of Red. IE
E
[5] Koo, J., Song, B., Chun
Early Detection. In: 15t
h
2001)
[6] Zheng, B., Atiquzzama
n
Generation Networks.
I
Networks, LCN 2000 (
N
T
o Increase the Overall Performance of Internet Routers
P
acket Drop Rate of other AQMs with FAutoREDWithRED
u
t in this algorithm, UDP is penalized and CBR throug
h
d
th. In case of packet drop rate shown in Fig 6, the FA
u
e
a fair share by dropping the UDP packets otherwise
maximum without allowing the TCP packets.
Q
M scheme called FAutoREDwithRED which aims to
a
nce of Internet routers. It reduces the queue oscillatio
n
t
the link. It penalises the non-adaptive flows without c
o
w
queuing delay and controlled packet loss. It is achie
e
d parameters and flow information. This packet drop
p
n
st the unresponsive flows resulting in fair congestion i
n
M
scheme inherits the advantages of these queue le
n
o
w information to satisfy the QOS requirements of
.: Random early detection gateways for congestion avoida
n
w
orking 1, 397–413 (1993)
.
, Shenkar, S., ICSI: Adaptive RED: An algorithm for Increa
s
s
active Queue Management, Berkely, CA
org/floyd/red.html (online)
i
on of queue oscillation in the next generation Internet rou
t
e
r Communication (2007)
C
hen, G., Chan, S., sukerman, M.: PD –Red: To Imp
r
E
E Communications Letter (August 2003)
g, K., Lee, H., Kahng, H.: MRED: A New Approach To Ran
d
h
International Conference on Information Networking (Febr
u
n
, M.: DSRED: An Active Queue Management Scheme for
N
I
n: Proceedings of 25th IEEE conference on Local Com
p
N
ovember 2000)
201
h
put
ut
o-
the
in-
n
in
o
m-
ved
p
ing
n
di-
n
gth
the
a
nce.
s
ing
t
ers.
r
ove
d
om
u
ary
N
ext
p
uter
202 K. Chitra and G. Padmavathi
[7] Kunniyur, S., Srikant, R.: Analysis and design of an adaptive virtual queue (AVQ)
algorithm for active queue management. In: Proceedings of ACM SIGCOMM, San
Diego (2001)
[8] Long, C., Zhao, B., Guan, X., Yang, J.: The Yellow active queue management algorithm.
Computer Networks (November 2004)
[9] Athuraliya, S., Li, V.H., Low, S.H., Yin, Q.: REM: Active queue management. IEEE
Network Mag 15, 48–53 (2001)
[10] Feng, W., Kandlur, D.D., Saha, D., Saha, D.: The Blue active queue management
algorithms. IEEE Transactions on Networking (2002)
[11] C.N., long, C.N., Zhao, B., Guan, X.-P.: SAVQ: Stabilized Adaptive Virtual Queue
Management Algorithm. IEEE Communications Letters (January 2005)
[12] Ott, T.J., Lakshman, T.V., Wong, L.: SRED: Stablised RED. IEEE Infocomm (March
1999)
[13] Feng, W.-c., Kapadia, A., Thulasidasan, S.: GREEN: Proactive Queue Management over
a Best-Effort Network. In: IEEE GlobeCom, Taipei, Taiwan (November 2002)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 203–206, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Scan Based Sequential Circuit Testing Using DFT
Advisor
P. Reshma
M. Tech, VLSI design, Amrita Vishwa Vidyapeetham, Coimbatore
reshma_p2004@rediffmail.com
Abstract. This paper shows that not every scan cell contributes equally to the
power consumption during scan based test. The transitions at some scan cells
cause more toggles at the internal signal lines of a circuit, which have a larger
impact on the power consumption during test application than the transitions at
other scan cells. They are called power sensitive scan cells. A verilog based
approach is proposed to identify a set of power sensitive scan cells. Additional
hardware is added to freeze the outputs of power sensitive scan cells during
scan shifting in order to reduce the shift power consumption.
Keywords: Scan, Scan DFlip-Flop, DFTAdvisor.
1 Introduction
Scan-based tests might cause excessive circuit switching activity compared to a cir-
cuit’s normal operation Higher switching activity causes higher peak supply currents
and higher power dissipation. High power dissipation during test can cause many
problems, which are generally addressed in terms of average power and peak power.
Average power is the total distribution of power over a time period. Peak power is the
highest power value at any given instant. When peak power is beyond the design
limit, a chip cannot be guaranteed to function properly due to additional gate delays
caused by the supply voltage drop. The power consumption within one clock cycle
may not be large enough to elevate the temperature over the chip’s thermal capacity
limit. To damage the chip, high power consumption must last for an enough number
of clock cycles. The test power consumed during scan shifting and capture cycles is
referred to as shift power and capture power, respectively. Average power consump-
tion is determined by the shift power. This paper, focus on reducing the average shift
power in scan-based tests. An effective method to identify a set of power sensitive
scan cells is used. By inserting additional hardware to freeze power sensitive scan cell
outputs to pre-selected values, the average shift power consumption can be reduced
dramatically.
2 The Proposed Method
A mentorgraphics tool called DFTAdvisor is used for Scan insertion. An S27 ISCAS
benchmark circuit is taken and scan is inserted in that circuit using DFTAdvisor. In
204 P. Reshma
that scan inserted circuit, number of toggling is calculated using verilog. The toggling
between the first and second ScanD-FF is higher. By introducing logic gates at the
output of first ScanD-FF we can block the complete toggling occurring at the combi-
national part, when only the Flip-Flops are selected. Implementation of a frozen scan
cell is shown in Fig. 1. An additional AND gate with the second input inverted is
inserted between the scan cell output and the function logics it drives. During scan
shifting, the scan enable signal se is asserted to be 1. It makes additional gate output
become constant. During normal operations, se is de-asserted to be 0 .Vector re-
ordering is also done, so as to reduce the toggling activity, there by reducing the
power dissipiation to a larger extend.
Fig. 1. Hardware implementation of frozen scan cell
3 Experimental Results
For testing a sequential circuit, both combinational cloud testing as well as memory ele-
ment testing is needed. DFTAdvisor accepts gate level net list format and generates a
new net list with scan cells inserted. The transitions at some scan cells cause more tog-
gles at the internal signal lines of a circuit than the transitions at other scan cells. These
scan cells are called power sensitive scan cells. Using verilog, power sensitive scan cells
are identified. Additional hardware is added to freeze the outputs of power sensitive scan
cells during scan shifting in order to reduce the shift power consumption. The number of
togglings is higher in between first and second flipflop. By introducing the circuit in
Fig.1b the combinational cloud in between d_out_0 and d_out_1 can be isolated,when
scan enable becomes 1(only flip-flops are selected).Thereby power dissipiation of the
whole circuit can be reduced to a larger extend.
Fig. 2. ISCAS S27 Benchmark Circuit
Scan Based Sequential Circuit Testing Using DFT Advisor 205
Fig. 3. Scan inserted circuit obtained from DFTAdvisor
Fig. 4. Scan inserted module generated by DFTAdvisor
Fig. 5. Freezing Power sensitive scancells by logic gates
206 P. Reshma
4 Conclusion
This paper proposed an effective and efficient method to reduce switching activity
during scan shifting. The method identifies a set of power sensitive scan cells and an
optimal input control vector through a fast signal probability analysis. By inserting
additional hardware to freeze the outputs of the identified scan cells and applying the
input control vector during scan shifting, reduction in scan shift power by more than
45% on average for the industrial circuits that are experimented, even if only 5% of
scan cells are frozen, can be achieved. Compared with the previous work that freezes
all the scan cells, this method achieves a good tradeoff between shift power reduction
and area overhead.
References
[1] Girard, P.: Survey of low-power testing of VLSI circuits. IEEE Design Test Com-
put. 19(3), 80–90 (2002)
[2] Ravi, S.: Power-aware test: Challenges and solutions. In: Proc. Int. Test Conf., pp. 1–10
(October 2007)
[3] Butler, K.M., Saxena, J., Jain, A., Fryars, T., Lewis, J., Hetherington, G.: Minimizing
power consumption in scan testing: Pattern generation and DFT techniques. In: Proc. Int.
Test Conf., pp. 355–364 (October 2004)
[4] Wang, S., Gupta, S.: An automated test pattern generator for minimizing switching activ-
ity during scan testing activity. IEEE Trans. Comput.-Aided Design Integr. Circuits
Syst. 21(8), 954–968 (2002)
[5] Sankaralingam, R., Oruganti, R.R., Touba, N.: Static compaction techniques to control
scan vector power dissipation. In: Proc. VLSI Test Symp., pp. 35–40 (2000); A Study of
Wavelet Thresholding Denoising, Proceedings of ICSP 2000
[6] Wen, X., Yamashita, Y., Kajihara, S., Wang, L.-T., Saluja, K., Kinoshita, K.: On low-
capture power test generation for scan testing. In: Proc. VLSI Test Symp., pp. 265–270
(2005)
[7] Wen, X., Kajihara, S., Miyase, K., Suzuki, T., Saluja, K., Wang, L.-T., Abdel-Hafez, K.,
Kinoshita, K.: A new ATPG method for efficient capture power reduction during scan
testing. In: Proc. VLSI Test Symp., pp. 60–65 (2006)
[8] Lin, Y.-T., Wu, M.-F., Huang, J.-L.: PHS-fill: A low power supply noise test pattern gen-
eration technique for at-speed scan testing
[9] Kajihara, S., Ishida, K., Miyase, K.: Test vector modification for power reduction during
scan testing. In: Proc. VLSI Test Symp, pp. 160–165 (2002)
[10] Saxena, J., Butler, K.M., Jayaram, V.B., Kundu, S., Arvind, N.V., Spreeprakash, P.,
Hachinger, M.: A case study of IR-drop in structured at-speed testing. In: Proc. Int. Test
Conf., pp. 1098–1104 (2003)
[11] Li, W., Reddy, S.M., Pomeranz, I.: ―On reducing peak current and power during test. In:
Proc. IEEE Comp. Soc. Annu. Symp. VLSI, May 2005, pp. 156–161 (2005)
[12] Remersaro, S., Lin, X., Zhang, Z., Reddy, S.M., Pomeranz, I., Rajski, J.: Preferred fill: A
scalable method to reduce capture power for scan based designs. In: Proc. Int. Test Conf.,
October 2006, p. 110 (2006)
[13] Badereddine, N., Girard, P., Pravossoudovitch, S., Landrault, C., Virazel, A., Wunderlich,
H.-J.: Structural-based power-aware assignment of don’t cares for peak power reduction
during scan testing. In: Proc. IFIP Int. Conf. VLSI, October 2006, pp. 403–408 (2006)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 207–213, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Rate Adaptive Distributed Source-Channel Coding Using
IRA Codes for Wireless Sensor Networks
Saikat Majumder and Shrish Verma
Department of Electronics and Telecommunication,
National Institute of Technology, Raipur, India
{smajumder.etc, shrishverma}@nitrr.ac.in
Abstract. In this paper we propose a scheme for rate adaptive lossless distri-
buted source coding scheme for wireless sensor network. We investigate the
distributed source-channel coding of correlated sources when correlation para-
meter is not fixed or may change during sensor network operation. For achiev-
ing rate adaptability we propose the puncturing and extension of IRA code de-
pending on the value of correlation between two sources and the quality of
channel. In our scheme we need to transmit only incremental redundancy for
decreased correlation or fall in channel quality to meet energy constraints and
reducing computation cost.
Keywords: Distributed source-channel coding, IRA code, sensor networks.
1 Introduction
Distributed lossless compression of information is required in many applications like
wireless sensor networks. To achieve efficient transmission the sources may be com-
pressed at the individual nodes independently and sent to the fusion center through
wireless channel. Compression is required before transmission because wireless sen-
sor nodes are characterized by low power constraint and limited computation and
communication capabilities. If there is correlation between the data sent by sensors,
then, though they do not communicate with each other, exploiting the correlation
would lead to further reduction in number of transmitted bits. Slepian-Wolf theorem
[1] shows that lossless compression of two separate correlated sources can be as effi-
cient as if they were compressed together, as long as decoding is done jointly at the
receiver. Such a scheme is known as distributed source coding (DSC). Practical
schemes for exploiting the potential of Slepian-Wolf theorem were introduced based
on channel codes [2], [3], [4], [5] and some of them used modern error correcting
codes (e.g. LDPC, Turbo codes) to achieve performance very close to theoretical
Slepian-Wolf limit. Schemes for distributed joint source-channel coding has been
proposed in literature [6], [7], [8] and the references therein which use a ‘good’ chan-
nel code for jointly performing the operation of compression (source coding) and
adding error correction capability (channel coding).
The basic idea of distributed source coding with side information at the decoder is
shown in the Fig. 1. It is a communication system of two binary sources X and Y with
208 S. Majumder and S.
V
conditional probability ma
s
tions, this statistical depen
d
ment. Instead of designing
the code should adapt to th
e
tion requires larger generat
Punctured low density par
i
required rate. Though punc
t
any way ease the computat
i
er. It is because puncturi
n
reinserts null bits before d
e
achieving rate-adaptive co
d
Fig. 1. Distributed so
u
The primary contributi
o
adaptive distributed source
-
requirements than the meth
Though puncturing has be
e
extension has not been ap
p
We apply efficient punct
u
desired code rate for give
n
of applying code extension
memory and computation
a
besides conserving battery
p
The paper is outlined a
s
plains the relation and tra
d
desired BER. Section 3 dis
c
IRA codes. Finally in secti
o
2 System Model
Consider the system of Fi
g
There are two correlated s
o
length k. The dependency
b
ity mass function Px
|y
before deployment, optima
l
is available lossless and er
r
possible; encoding operat
i
apply systematic irregular
encoder using the method i
n
V
erma
s
s function P(X|Y) = p. For many sensor network appl
i
d
ency between X and Y may not be known before dep
l
the code for lowest correlation probability p
min
, the rat
e
e
correlation parameter. Designing code for lowest corr
e
or and parity check matrices and it results in longer co
d
i
ty check (LDPC) codes were used [10] for achieving
t
uring reduces the number of transmitted bits, it does n
o
i
onal and memory requirements of the encoder and de
c
n
g channel code removes bits after encoding process
e
coding. Other variations of LDPC code has been used
d
es for distributed source coding [11].
u
rce-channel coding with side information at the decoder
o
n of this paper is the development of a scheme for r
-
channel coding that has lesser computational and me
m
ods using puncturing only for achieving rate-compatibi
l
e
n the accepted method for achieving rate adaptation, c
p
lied in the context of distributed source-channel cod
i
u
ring [12] and extending [13] methods for obtaining
n
value of source correlation and channel quality. Our
i
methods for distributed source-channel coding reduces
a
l requirements at both sensor nodes and fusion ce
n
p
ower.
s
follows; Section 2 describes the overall system and
d
eoffs between correlation parameter, channel quality
c
usses the methods employed for puncturing and exten
d
o
n 4 and 5 we present simulation results and conclusion
.
g
. 1 with the following notations used for rest of the pa
p
o
urce vectors X = [x
1
, x
2, ... ,
x
k
] and Y = [y
1
, y
2
, ... , y
k
b
etween these two sources is given by conditional prob
a
. Since the correlation parameter p may not be kn
o
l
rate of transmission has to be decided during run tim
e
r
or free at the decoder. We try to encode X as efficientl
y
i
on being joint source-channel coding. In this paper
repeat accumulate (IRA) codes as joint source-cha
n
n
[6].
i
ca-
l
oy-
e
of
ela-
d
es.
the
o
t in
c
od-
and
for
ate-
m
ory
l
ity.
ode
i
ng.
the
i
dea
the
n
ter,
ex-
and
d
ing
.
p
er.
] of
a
bil-
o
wn
e
. Y
y
as
we
n
nel
Rate Adaptive Distributed Source-Channel Coding Using IRA Codes 209
Fig. 2. Rate adaptive source-channel IRA encoder. Hu is the mother code matrix, whereas
matrices El are obtained by code extension. All the generated parity bits are concatenated to
produce the final code word Z. Puncture is applied when increase in code rate is required.
IRA codes, a special class of LDPC codes, are used here in the context of sensor
network because they enjoy extremely simple encoding and low complexity decoding.
Parity check matrix for systematic IRA code has the form H = [Hu Hp], where Hu is a
sparse matrix and Hp is an dual diagonal matrix. The output of the systematic
encoder is Z = [X P], where P is the parity vector. We design the IRA encoder in Fig.
2 which provides incremental redundancy and is based on such an LDPC encoder in
[14]. In the encoder matrix HuT followed by accumulator block provides the parity
check bits P0 of mother code. This part is same as any conventional IRA encoder. The
matrices E1, E2... Elmax take into account extra parity bits (for decreasing code rate)
provided by code extension. The puncturing block is for increasing the code rate more
than the rate of mother code. For example, when rate index l is specified, the concate-
nated code is given by Z = [X P0 P
1 P
2 . . . Pl]. Parity bits are obtained as P
XH
H
and PXE for i1,…, [13], [15].
When sources are highly correlated or if the channel is good, only Z = [X P0] is
produced and code is punctured to get the required rate. Code design for given rate
index (or rate) using extension and puncturing of IRA code is discussed in section 3.
The generated code Z and lossless source Y can be jointly decoded using a message
passing decoder [15]. The decoding process is same as any IRA decoder, only differ-
ence being the initialization of the log-likelihood ratios (LLR) [6]. For unit code word
energy and code rate r, LLR of the parity bit nodes is 2P14r N
⁄ and that of
information bit nodes is 2Y1ln1
⁄2X14r N
⁄ where, P = [P0
P1 P2 . . . Pl] and function f(x) indicates effect of channel on signal x.
Next we derive the relation for code rate under channel constraints and varying
amount of correlation between the sources. Since Y is a vector of equiprobable binary
random variable, its rate is its entropy HYkHyk. From Slepian-Wolf theo-
rem, the theoretical limit on rate after lossless compression of X is HZ
kHx|ykH. For nonideal channel we have to account for the limited channel
capacity. For error free communication of a binary source of per bit entropy H(p),
using a channel code of rate r, the bit energy Eb required is related as EN
⁄
21/2r where, N0/2 is the noise power spectral density. Appropriate value
210 S. Majumder and S. Verma
of rate r for given channel condition and conditional entropy H(p) can be found by
solving this equation. For finite length codes and continuous channel, the SNR
(Eb/N0) entered has to be few dB higher than actual value. This shift depends on code
length and minimum BER required and can be obtained empirically. For example, in
our case of IRA code of 1024 information bits and tolerable BER of 10-4, shift re-
quired would be of 1.45 dB.
3 Rate Adaptive Code Design
We now elaborate the code design method which uses extended and punctured IRA
code for a obtaining a specific rate r. Rate-compatible IRA codes through determinis-
tic extending based on congruential extension sequences [13] is used in our study.
Compared to other code extension methods, this method uses low-complexity, alge-
braic operations without any post-construction girth conditioning. As already men-
tioned, puncturing of code though reduces the transmitted energy; it does not reduce
the memory and computational requirements of encoding and decoding operation. So,
starting out with a channel code of lowest anticipated rate and puncturing saves
transmission power but not computational resource. Therefore, we consider a mother
code of moderate rate rk/n and length n0. Corresponding parity check matrix H
= [Hu Hp] is of size mn. Let r,r,…,r with rrr to
be set of desired target rates lower than r0 for corresponding parity check matrices
H,H,…,H. For a particular rate index l, starting from H0, we construct
HHH0
E0I
(1)
where, E is obtained by concatenation of sub-matrices E1, E2, ... , El as ET = [E1T, E2T,
..., ElT]. Each of the submatrices El is of size εkεsε, 0 are all zero matric-
es, I is a εε identity matrix. sk/q, where q = ε if ε is odd, otherwise q = ε –
1. ε is the number of rows in El and is chosen to have the desired code rates. Two
random sequences, :0,…,1 and :0,…,1, with elements from
GF(q) is generated. Care must be taken so that no two element in same sequence are
same. A matrix A = [aij : 0 ≤ i ≤ l - 1, 0 ≤ j ≤ s - 1] of dimension s is formed with
its elements q, where we have taken d = 1. Using the
elements of A, a new s matrix L is formed with its each element I(aij) being
εε identity matrices with rows cyclically shifted to the right by aij positions. Fi-
nally, our extended matrix E is obtained from L by method of masking [9].
To obtain code rates 1,…,r— ,…,r,r with 1r r r
we use the puncturing method in [12]. In this method only parity bits are punctured
and punctured node is chosen such that it is at equal distance (in terms of number of
nodes) from neighboring unpunctured nodes. The punctured codes of different rates
obtained from this type of deliberate puncturing are not rate compatible. To obtain
rate compatibility, they have proposed this simple algorithm which performs punctur-
ing in reverse way. In their algorithm, mother code is deliberately punctured to the
highest code rate, then unpuncturing those punctured nodes to obtain lower rates.
Rate Adapti
v
4 Simulation Result
s
For simulation we have
u
0.6
0.4
, rat
e
done for a binary phase shi
f
sum product algorithm (S
P
decoding iterations is 100
source-channel coding of
parameter. We then compa
r
In our case mother code
ing and extension, respecti
v
{1/3, 0.8} and different val
for given p, any suitable ra
As can be seen from the pl
Shannon’s limit. This gap
c
may not be always possibl
e
Fig. 3. Simulation results for
j
correlation parameter p. The v
Table 1 compares the
m
using systematic IRA code
tor values only and are obt
a
seen that if we start out w
i
nificantly less than the cas
e
only.
v
e Distributed Source-Channel Coding Using IRA Codes
s
u
sed mother code with node-wise degree-distributio
n
e
r
0
= 0.5 and k = 1024 as in [13]. All the simulations
f
t keying (BPSK) modulated AWGN channel. The itera
t
P
A) was used for decoding, and the maximum numbe
r
. We simulate the error performance of distributed j
o
two correlated sources for various rates and correla
t
r
e the memory requirements for obtaining various rates.
rate is 0.5. Code of rate 0.8 and 1/3 is obtained by punc
v
ely. Fig. 3 gives the BER plot for two different rates
ues of correlation parameter. These plots clearly show
t
te r
l
can be chosen in accordance to the channel condit
i
ots the code performance is the about 2 dB away from
c
an be further reduced by increasing the code length, w
h
e
for low cost sensor nodes.
j
oint source-channel decoding of source X for different valu
alues in parenthesis beside correlation values are code rates.
m
emory requirements with an algorithm based on [10]
instead of systematic LDPC code. The numbers are ind
i
a
ined for simulation program running on Matlab. It ca
n
i
th a medium rate code, our memory requirements are
e
where one starts with lower rate code and uses punctu
r
211
n
of
are
t
ive
r
of
o
int
t
ion
tur-
r =
t
hat
i
on.
the
h
ich
u
e of
but
i
ca-
n
be
sig-
r
ing
212 S. Majumder and S. Verma
Table 1. Memory utilization (in kilobytes) for encoding and decoding processes at different
code rates. Two different rate adaptation methods are compared, our method utilizing both
puncturing and extension, and other one uses code puncturing only.
Code
rate
Rate adaptation by puncturing and
extension (our method)
Rate adaptation by puncturing only
(mother code rate is 1/3)
Encoder Decoder Encoder Decoder
8/24 256 1351 256 1351
8/20 192 1110 256 1351
8/18 160 989 256 1351
0.5 128 869 256 1351
0.6 128 869 256 1351
0.7 128 869 256 1351
0.8 128 869 256 1351
5 Conclusion
We proposed a scheme for lossless distributed source-channel coding of correlated
sources with side information using extended and punctured IRA codes. Extension
and puncturing is used for achieving the required rate, which in turn depends on cor-
relation between sources and channel condition. The simulation results confirm that
proposed method can achieve any required rate for given correlation and channel
condition. We have shown that the use of code extension for rate adaptation for joint
source-channel coding requires lesser memory than the methods using puncturing
only. Thus, besides saving transmission power by adapting to most suitable rate, our
method also minimizes the memory requirement for same code rate.
References
1. Cover, T.M., Thomas, J.A.: Elements of Information Theory. Wiley-India, New Delhi
(2006)
2. Pradhan, S.S., Ramchandran, K.: Distributed Source Coding Using Syndromes (DISCUS):
Design and Construction. In: Proceedings of IEEE DCC, pp. 158–167 (1999)
3. Stankovic, V., Liveris, A.D., Xiong, X., Georghiades, C.N.: On Code Design for Slepian-
Wolf Problem and Lossless Multiterminal Network. IEEE transactions on Information
Theory 52(4) (2006)
4. Garcia-Frias, J., Zhao, Y.: Compression of Correlated Binary Sources Using Turbo Codes.
IEEE Communication Letters 5(10) (2001)
5. Fresia, M., Vandendorpe, L., Poor, H.V.: Distributed Source Coding Using Raptor Codes
for Hidden Markov Cources. IEEE Transactions on Signal Processing 57(7) (2009)
6. Liveris, A.D., Xiong, Z., Georghiades, C.N.: Joint Source-Channel Coding of Binary
Sources with Side information at the Decoder Using IRA Codes. In: Proceedings of IEEE
Workshop on Multimedia Signal Processing, pp. 53–56 (2002)
7. Garcia-Frias, J.: Joint Source-Channel Decoding of Correlated Sources Over Noisy Chan-
nels. In: Proceedings of IEEE DCC (2001)
Rate Adaptive Distributed Source-Channel Coding Using IRA Codes 213
8. Zhao, Y., Garcia-Frias, J.: Turbo Compression/Joint Source-Channel Coding of Correlated
Sources With Hidden Markov Correlation. Elsevier Signal Processing 86, 3115–3122
(2006)
9. Xu, J., Chen, L., Djurdjevic, I., Lin, S., Abdel-Ghaffar, K.: Construction of Regular and Ir-
regular LDPC Codes: Geometry Decomposition and Masking. IEEE Transactions on In-
formation Theory 53(1) (2007)
10. Sartipi, M., Fekri, F.: Distributed Source Coding Using LDPC Codes: Lossy and Lossless
Cases with Unknown Correlation Parameter. In: Proceedings of Allerton Conference on
Communication, Control and Computing (2005)
11. Varodayan, D., Aaron, A., Girod, B.: Rate-adaptive Codes for Distributed Source Coding.
Elsevier Signal Processing 86, 3123–3130 (2006)
12. Yue, G., Wang, X., Madihian, M.: Design of Rate-Compatible Irregular Repeat Accumu-
late Codes. IEEE Transactions on Communications 55(6) (2007)
13. Benmayor, D., Mathiopoulos, T., Constantinou, P.: Rate-Compatible IRA Codes Using
Quadratic Congruential Extension Sequences and Puncturing. IEEE Communications Let-
ters 14(5) (2010)
14. Li, J., Narayanan, K.: Rate-Compatible Low Density Parity Check Codes for Capacity-
Approaching ARQ Schemes in Packet Data Communications. In: Proceedings of IASTED
CIIT, pp. 201–206 (2002)
15. Ryan, W.E., Lin, S.: Channel Codes: Classical and Modern. Cambridge University Press,
New York (2009)
16. Hou, J., Siegel, P.H., Milstein, L.B.: Performance Analysis and Code optimization of Low
Density Parity-Check Codes on Rayleigh Fading Channels. IEEE Journal on Selected
Areas in Communication 19(5) (2001)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 214–221, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Web Cam Motion Detection Surveillance System Using
Temporal Difference and Optical Flow Detection with
Multi Alerts
V.D. Ambeth Kumar1 and M. Ramakrishan2
1 Research Scholar, Sathyabama University, Chennai, India
ambeth_20in@yahoo.co.in
2 Professor and Head, Velemmal Engineering College, Chennai, India
ramkrishod@gmail.com
Abstract. This paper proposes a method for detecting the motion of a particular
object being observed. The Motion tracking Surveillance has gained a lot of
interests over past few years. The Motion tracking surveillance system is brought
into effect providing relief to the Normal video surveillance system which offers
time-consuming reviewing process. Through the study and Evaluation of
products and methods, we propose a Motion Tracking Surveillance system
consisting of its method for motion detection and its own Graphic User Interface.
Various methods are used in Motion detection of a particular interest. Each
algorithm is found efficient in one way. But there exits some limitation in each
of them. In our proposed system those disadvantages are omitted and combining
the usage of best method we are creating a new motion detection algorithm for
our proposed Motion Tracking Surveillance system. The proposed system in this
paper does not have its effect usage in office alone. It also offers more
convenient, effective and efficient usage in home.
Keywords: Motion detection, Surveillance, GUI, Real time environment.
1 Introduction
The task of a Motion Tracking Surveillance system is to detect a motion present in
“region of awareness”, where the region of awareness, or the field of view, is defined
as the “Portion of Environment (or) surrounding being monitored” [1]. The motion of
moving objects is the activity of the portion of environment may be represented as
Region of interest. The system captures images only when the motions exceed a
certain threshold that is present in system. It thus reduces the volume of data that
needs to be reviewed and is therefore a more convenient way of monitoring the
environment, especially with the increasing demand for Multi-camera. Also, it helps
to save Data space by not capturing static images which usually do not contain the
object of interest.
There are two Main components that concern basic Motion Tracking Surveillance
software, (i.e) GUI and method for Motion Detection. As part of the literature review,
we evaluated 4 popular motion detection surveillance products that are currently
Web Cam Motion Detection Surveillance System 215
available in market. We considered about the terms of format and their features which
are based on features that are required of a Surveillance system and the additional
features that are required for purpose of motion detection. The four products, namely
‘Active Webcam’ by py software [2], ‘Watcher’ by Digi-Watcher [3], ‘FG Engineering
Surveillance 4 cam Basic’ by FG Engineering [4], and ‘Supervision Cam’ by Peter
Kirst [5] are chosen by considering users feedback.
Also in the literature review, the existing methods for motion detection are
discussed. They include some of popular methods, such as Temporal Difference[1][6]
and background modelling [1][6][7][8]; as well as methods that are not so widely
used due to certain constraints eg. Optical flow [9][10] and Spatio-Temporal
entropy[11].
2 Motion Detection
Methods for motion detection can be categorized into 2 Main classes, i.e Region-
based and pixel-based algorithm [8]. The former, based on Spatial dependencies of
neighboring pixel colours to provide results more robust to false alarms. The later
based on binary difference by employing local or pixel-based model of intensity, is a
simple model often used in real-time application.
2.1 Optical Flow
It is the 2-D velocity field induced in an image due to projection of moving objects
onto the image plane. An optical flow shows velocity of each pixel in image. Most
optical flow techniques assume that uniform illumination is present [12]. However
only small movements can be accurately detected in gradient technique due to
Taylor’s approximation of the gradient constraint equation.
2.2 Spatio-Temporal Entropy
It works on assumption that pixel state change brought about by noises would be in
small range while those brought about by motion would be large. However, it is
impossible to predict all types of noises, thus accuracy of detection is most of time not
to satisfied level.
2.3 Temporal Difference
The consecutive frames are compared on basis of pixel by pixel basis for calculating
motion sequence and a threshold is applied for classification of objects in stationary
or in motion. But when there is motion in objects, the image intensities are not varied
in a short time interval [13]; so that motion is detected only in boundaries. Also it
makes more false alarms as they don’t show relationship of pixel with neighbourhood.
2.4 Background Modelling
Background modelling methods are classified into two types namely pixel-based and
region-based models .Background subtraction is often used in case of pixel-based. An
216 V.D. Ambeth Kumar and M. Ramakrishan
image of the stationary background is generated by averaging the image sequence
over a period of time of time on mixture of Gaussian distribution. With the help of
Gaussian density the likelihood of each incident pixel colour is computed .On the
deviation between the two pixels, they will be labelled as foreground pixels and
detected as motion. Krumm et al[14] has done an excellent analysis on this method.
There are two popular methods for region-based background modelling, the three
tiered algorithm which process image at the pixel, region and frame level[14] , and
the eigen-space decomposition method[8].
3 Proposed Method
We propose a method that uses both temporal difference and optical flow methods
together with morphological filter for the purpose of motion detection. Figure 1 gives
an flowchart of the proposed motion detection process.
3.1 Reducing the Number of Pixel Using Temporal Difference Method
The initial image frame is provided as the input for our proposed method. The image
frame with reduced number of pixel is obtained as output.
First of all, the temporal difference method is used to obtain an initial coarse image
so as to reduce the number of pixels that a downstream tracking algorithms have to
process. It is used despite its shortcomings as it is a cheap and simple method. No
knowledge of the background is required as in background subtraction where training
periods in the absence of foreground are required for bootstrapping [14]. Also, the
method’s inability to detect the entire shape of objects of interest does not act as a
serious obstacle considering the fact that the system is for simple office and home uses.
3.2 Image Denoising Using Optical Flow Method
The image frame with reduced number of pixels is given as the input. Noise pixels are
removed in the output.
The optical flow method is then used to further analyze the detected motion area to
reduce noises such as movements of trees and the capturing device. Tian et al[ 15]
mentioned that objects of interest tend to have consistent moving directions over time.
Mittal et al [16] also stated that persistent motion characteristics are shown by most
objects of interest. Therefore, optical flow is a suitable method in this aspect as it is
able to estimate the direction and speed of moving objects to reduce false alarms
incurred by undesirable external factors like oscillating fans and trees.
3.3 Morphological Filtration
The morphological filter is then applied which is used to suppress noises while
preserving the main object characteristics [17]. It consists of ways for digital image
processing based on mathematical morphology which is a nonlinear approach
developed based on set theory and geometry. It is able to decompose complex shapes
into meaningful parts and separate them from the background. In addition, the
mathematical calculation involves only addition, subtraction and maximum and
minimum operations with no multiplication and division.
Web Cam Motion Detection Surveillance System 217
Fig. 1. Flowchart of motion detection process
The two fundamental morphological operations are dilation and erosion on which
many morphological algorithms are based on.Experiments done by Lu et al [18]
proved that the method is effective in preserving moving object areas and eliminating
noises. Finally, a binary image is generated with the motion region coloured white
and regions with no motion detected coloured black.
3.4 GUI
One guiding principle in designing the GUI of commercial software is user-
friendliness. This is most of the times determined by two factors, format and the
Start the Motion tracking surveillance system
Capture the image frame
Obtain the threshold value
Set the image input frame
Reducing the pixels by temporal difference method
Ima
g
e denoisin
g
b
y
o
p
tical flow detection
Morphological filteration
Motion
detected
Storing the motion and
creating Multi alarm
The consecutive frame is
set as input and the process is
continued
Yes
No
Compare with
defined threshold
value
Motion>Threshold
218 V.D. Ambeth Kumar and M. Ramakrishan
features. Clear and neat formatting makes it easier for users to access necessary
features in the system gives rise to neat formatting.
Two basic formats of GUI are used to design the format of GUI. Type 1 is the GUI
in which users can easily switch from capture panel to log panel for monitoring and
reviewing. Type 2 is the GUI is the more conventional GUI design whereby users
have to click on the menu bar every time for another window to appear.
4 Results and Discussion
Our aim is to present a new technology with unique features. We are going to create
three types of alerts: 1.Audio alert 2. Video alert 3. E-mail alert. This will be
explained in detail in the forth coming paragraphs. This model is based on GUI which
makes the user to familiarize the concept. The function of the settings, camera and log
is shown in the picture. These functions give a easy access to the users.
Fig. 2. Login panel Fig. 3. Camera panel
Fig. 4. Video alert Fig. 5. E-mail alert
Web Cam Motion Detection Surveillance System 219
The “Camera” panel (shown Fig 3) displays detected motions that automatically
pop up on the main window for close monitoring. The “Log” panel (shown Fig 2)
displays the images saved by the system.
Now we discuss on the three types of alerts first Audio alert will produce the
audible sound when any motion is detected across the prohibited region. Next the
visual alert (shown Fig 4) indicates the occurrences of the motion by changing the
color displayed in the monitor. When the normal video is playing a grey strip will be
seen at the top of the video.
But when the motion is detected the grey color strip changes to red and the play
symbol is also displayed along with that strip. Finally E-mail alert (shown in Fig 5)
generates the alert as mail with frame attachments.
5 Comparison with Other Methods
The detection accuracy of motion detection using proposed algorithm is compared
with other detection methods .The following table shows the comparison of number
of alerts and detection accuracy of motion detection for images.
Table 1. Comparison of Alert (Performance) of the Proposed Method with Other Motion
Detection Methods
Motion
Detection
Methods
No.of Alerts Detection
Accuracy (%)
Optical Flow
[Ref.14]
01 83.34
Temporal
Difference
[Ref.12]
01 87.43
Proposed
Method
03 94.375
The corresponding graph for the Alert for motion detection using the above
mentioned methods is given as under.
0
0.5
1
1.5
2
2.5
3
3.5
Optical Flow Temporal Difference Proposed Method
Methods
No. of Ale rts
Fig. 6. Comparison of Alert of Other Motion Detection Methods with Proposed Method
220 V.D. Ambeth Kumar and M. Ramakrishan
The corresponding graph for the Detection Accuracy for motion detection using
the above mentioned methods is given below.
76
78
80
82
84
86
88
90
92
94
96
Optical Flow Temporal Difference P roposed Method
Detection Methods
Accurasy(%
)
Fig. 7. Comparison of Accuracy in % of Existing Detection Methods with (OF, TD) with
Proposed Method
6 Conclusion
The software system is capable of working at any circumstances. This is simple to
install and access it. This system provides an enhanced feature of email alert, run in
surveillance mode, live view about the happenings of that particular area etc. This
system is very powerful since it uses motion detection algorithm to compare the
frame. These frames are compared by pixels when there is any change even in the
single pixels they indicate the alerts to the user. We are using J2SE for our
implementation. This will be a more convenient way of monitoring the real time
environment, especially with the increasing demand for multi-camera.
References
1. Stefan Gachter, “Motion Detection as an Application for the Omnidirectional Camera”,
Research Reports of CMP, Czech Technical University in Prague, Omnidirectional Visual
System (7) , 5-13 (2001); FPS RTD – FET, Project No: IST-1999-29017
2. About Active Web Cam,
http://www.pysoft.com/ActiveWebCamMainpage.htm
3. Digi-Watcher Features of our webcam software for video surveillance,
http://www.digi-watcher.com/surveillance_features.htm
4. http://www.fgeng.com-vsmfeatures.htm
5. SupervisionCam Homepage English, http://www.supervisioncam.com/
6. Abdelkader, M.F., Chellappa, R., Zheng, Q.: Integrated Motion Detection and Tracking for
Visual Surveillance. In: Proc. Of the Fourth IEEE International Conference on Vision
Systems (ICVS 2006), pp. 1–3 (2006)
7. Miezianko, R.: Motion Detection and Object Tracking in Grayscale Videos Based on
Spatio Temporal Texture Changes. In: Partial Fulfillment of the Requirements for the
Degree of Doctor of Philosophy. Temple University Graduate Board, pp. 8–14 (2006)
Web Cam Motion Detection Surveillance System 221
8. Sheikh, Y., Shah, M.: Bayesian Modeling of Dynamic Scenes for Object Detection. IEEE
Transaction on Pattern Analysis and Machine Intelligence 27(11), 1778–1780 (2005)
9. Rivo, J.-E.S., Cajote, E.R.: Object Motion Detection Using Optical Flow. In: Digital
Signal Processing Laboratory, Department of Electrical and Electronics Laboratory,
University of the Philippines, pp. 1–2
10. Sarker, M.H., Bechkoum, K., Islam, K.K.: Optical Flow for Large Motion Using Gradient
Technique. Serbian Journal of Electrical Engineering 3(1), 103–113 (2006)
11. Ma, Y.-F., Zhang, H.-J.: Detecting Motion Object by Spatio-temporal Entropy. In: Proc.
IEEE International Conference on Multimedia and Expo., pp. 265–268 (2001)
12. Zelek, J.S.: Bayesian Real-time Optical Flow. In: School of Engineering, University of
Guelph
13. Davis, L.: Time varying image analysis,
http://www.umiacs.umd.edu/~1sd/426/motion.pdf
14. Krumm, J., Toyama, K., Brumitt, B., Meyers, B.: The Ultimate Futility of Background
Subtraction, Int. Journal of Computer Vision, 3–8 (Submitted)
15. Tian, Y.-L., Hampapur, A.: Robust Salient Motion Detection with Complex Background
for Real-time Video Surveillance. IBM T.J.Watson Research Center, 1–6
16. Gonzalez, R.C., Woods, R.E.: Digital Image Processing, 2nd edn. Prentice-Hal,
Englewood Cliffs (2002); ISBN: 0130946508
17. Mittal, A., Paragios, N.: Motion-Based Background Subtraction Using Adaptive Kernel
Density Estimation. In: Real-Time Vision and Modeling Siemens Corporate Research
Princeton and C.E.R.T.I.S. Ecole Nationale de Ponts etc Chaussees Champs sur Marne,
France, pp. 1–8
18. Lu, N., Wang, J., Wu, Q.H., Yang, L.: An Improved Motion Detection Method for Real-
Time Surveillance. IAENG International Journal of Computer Science 35, 1,
IJCS_35_1_16, 1–10
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 222–226, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Rts-Mirror: Real Time Synchronized Automated Rear
Vision Mirror System
Kuldeep Verma1, Ankita Agarkar1, and Apoorv Joshi2
1 Malwa Institute of Technology, Nipaniya, Bypass Road, Indore 452 016, India
{lucky_kd_007,ankita_agarkar}@yahoo.com
2 Royal College of Technology, Opp. Pitra Parvat, Jamburdi Hapsi, Indore 452 003, India
apoorv76652@gmail.com
Abstract. In this paper, we present ‘RTS-Mirror’, an attempt to transform the
overall usage and functioning of rear vision mirror to an entire stage of compli-
ance in accordance to the automated technology of rear mirror that will leverage
the present technology of movement of the rear vision mirror by a joystick or
rather by a human hold. ‘RTS-Mirror’ is an intelligent system running on
RTOS that tracks the face of driver in real-time and will apparently adjusts the
rear mirror in a contiguous synchronized manner. This project explores the use
face tracking algorithm like Haar-Classifier cascade algorithm [3], [5], Lucas-
Kanade algorithm [14], [15], and joystick mechanism of rear mirror can de-
velop a fully automated digital Rear vision mirror.
Keywords: Joystick controlled rear mirror, Facial features recognition algo-
rithm, Synchronized movement of Mirrors, Image Tracking.
1 Introduction
The affixing of a rear vision mirror in Marmon race car at the inaugural Indianapolis
500 race [1] in 1911 ascended the importance of the newly invented automobile part.
The main area of development occurred so far concerns with the optical area counting
such as HCM, Electro-chromic Mirror [2], [4], [7]. Despite this, human palm interfer-
ence still exists in adjusting the rear mirror.
‘RTS-Mirror’ is an endeavour to integrate a camera into the central dashboard rear
mirror that will track the facial position of the driver guided by Real Time Operating
System (RTOS) for computing the critical angle at real time. RTS-Mirror primarily
functions with the aid of two different algorithms, namely Haar-Classifier cascade
algorithm [3], [5] and Lucas-Kanade algorithm [14], [15]. Lastly, the joystick mecha-
nism of rear vision mirror will be embedded into RTS-Mirror technology with input
provided by the resultant of the algorithm.
2 Related Works
Several projects and products have tried to meld the simple rear vision mirror into a
more compound form that claims to automate the rear vision mirror using joystick
Rts-Mirror: Real Time Synchronized Automated Rear Vision Mirror System 223
interface [2], etc. But all in all, no such project has been developed that claims to have
facial gestures as the primary input for adjustment of the rear vision mirror.
This project is a blend of two different projects; the facial position detection
and the joystick oriented automated approach. But still it is solely a genuine project
with extreme possibilities, which is still unlisted in any sort of currently developing
projects.
3 Statistical Data
According to the National Highway Traffic Safety Administration (NHTSA) in
Washington D.C., there are approximately 840,000 side-to-side blind spot collisions
with 300 fatalities every year in the United States [8]. These accidents are due to the
incorrect adjustment of the rear vision mirror. By carefully examining the accurate
angle that gives the best viewing angles, three cases has arrived.
3.1 Accurate Angle Adjustment
Accurate/Critical viewing angles, showed in fig. 1, is calculated by the derivations
below. These angle measurements will be patched in RTOS environment to maintain
a rough idea about the actual critical angles [9], [10], [11].
Fig. 1. Approximately calculated adjusting angles for all three rear view mirrors of car deliver-
ing the best area covered in vision
4 RTS-Mirror
4.1 What Is RTS-Mirror?
RTS-Mirror is an integrated project that delivers an intelligent environment that acts
at the change in posture of rider (though slight). RTS-Mirror tracks the facial specific
parts such as eyes as main concern which is represented as dots in algorithm. After-
wards, the system adjusts the mirror using Joystick controlled mechanism. The input
is given by the algorithm and thus eliminating any interference of human palm.
224 K. Verma, A. Agarkar, and A. Joshi
4.2 How Does RTS-Mirror Work?
A high-level overview of how the RTS-Mirror works is shown in Figure 2 by flow
chart. Firstly, the embedded camera in the rear vision mirror primarily tracks the mo-
tion in its perspectives. Consecutively, it uses Haar-Classifier Cascade Algorithm to
filter the human face from the tracked motion. Successively, the next algorithm Lucas-
Kanade Algorithm will tracks the features of the face. Later, it calculates the critical
angle [6] which results the calculation of accurate angle, resulting in synchronization
phase of all the mirrors in correspondence with the central rear view mirror. This
mechanism will be handles by RTOS environment. The mirror’s assembly includes an
actuator, capable of transforming electrical signals into a mechanical force, most com-
monly a torque applied to an output shaft. The mirrors are mechanically attached to the
output of the actuator through a mechanical means, such as a threaded shaft.
4.3 Calculating Critical Angle
Critical angle is an angle that provides the accurate viewing angle perfectly appropri-
ate for the driver [6], [11]. Primarily, we consider a horizontal plane angle from the
operator's eyes to the central rear view mirror, θIR, where δIR represents the horizontal
plane adjustment angle for the manually adjusted internal rear view mirror. θIR can be
estimated as, follows,
θIR=2δIR (1)
Now, given θIR and applying a simple trigonometric function based upon a right trian-
gle having perpendicular side dIR and IIR, the longitudinal distance from the operator's
eyes to the rear view mirror.
IIR = dIR tan ( 2δIR) (2)
Thus, the horizontal plane angle from the operator's eyes to the left side mirror, θLS,
utilizing a right triangle with perpendicular sides dLS and e+IIR, can be expressed as
follows, and similarly, the horizontal plane angle from the operator's eyes to the right
side mirror, θRS, can be expressed as follows:
θLS = tan-1(dLS * e+IIR) = tan-1(dLS * e+dIR tan(2δIR) ) (3)
θRS = tan-1(dRS * e+IIR) = tan-1(dRS * e + dIRtan(2δIR) ) (4)
Calculated thusly, input angles from a manual adjustment of a rear view mirror can be
used to generate side mirror adjustment angles [12], [13].
4.4 Implementation
The implementation phase of this project is still in its coding part. At this instance, the
camera is programmed to track any motion. This part has been designed and coded
with Aforge.NET framework that provides efficient classes of imaging and visual
related forms, etc. The next phase will be consisting of using specified algorithms to
monitor only the facial parts of human (driver). Afterwards, the synchronization of
the side mirrors in accordance with the central rear mirror will be managed. Lastly,
Rts-Mirror: Real Time Synchronized Automated Rear Vision Mirror System 225
Fig. 2. Detailed System design of ‘RTS-Mirror’
the project will consist of embedding the above proposed software into the present
joystick technology.
The overall system design is implemented in accordance with the Flow chart
(fig.2). All in all, the fully automated system will be developed with RTOS control
delivering accurate angles of the mirror at real time providing efficient removal of
blind spots as well.
5 Results and Conclusion
The starting result of developing RTS-Mirror has been substantial with high grade
statistical data. Also, the implementation of the algorithms has provided high per-
formance characters to the proposed system. ‘RTS-Mirror’ system bridges the gap
between the mechanical and digital world by providing an interface, linking the exist-
ing technology of joystick mechanism with the advanced motion tracking technology.
By augmenting the familiar and ubiquitous rear vision system, the system leverages
existing patterns of behaviour, merging fully mechanical mechanism with a user’s
digital world.
226 K. Verma, A. Agarkar, and A. Joshi
Acknowledgements
We thank our Professor Divya Gautam and Lokendra Songare for their valuable feed-
backs. We would also like to thank Sapan Jain, post graduate student of Texas A&M
University who had contributed ideas and time to this research.
References
1. Ward’s AutoWorld magazine, Ward’s Auto World: Rearview Mirror webpage on Ward-
sauto (2002), http://waw.wardsauto.com/ar/auto_rearview_mirror/
2. Murakami Corporation,
http://www.murakami-kaimeido.co.jp/english/product/
mirror/pro04.html
3. Adolf, F.: How-to build cascade of boosted classifiers based on Haar - like features,
http://robotik.inflomatik.info/other/opencv/
OpenCV_ObjectDetection_HowTo.pdf
4. Murakami Corporation, Hydrophilic Clear Mirror (HCM), http://www.murakami-
kaimeido.co.jp/english/product/mirror/pro01.html
5. Wilson, P., Fernandez, J.: Facial Feature Detection using Haar Classifiers. Texas A&M
University – Corpus Christi, JCSC 21(4) (April 2006)
6. Ryu, J., Shin, K., Litkouhi, B., Lee, J.: Automatic Rearview Mirror adjustment system for
vehicle. USPC class 70149, Patent application number 20100017071
7. Murakami Corporation Electro-Chromic Mirror,
http://www.murakami-kaimeido.co.jp/english/operation/mirror/
product/pro02.html
8. AccidentMI.com, 5674 Lancaster, Commerce, MI 48382,
http://www.accidentmi.com/accidentstatistics.html
9. Hannah, J., Palin, B., Kratsch, C., Gaynor, B.: http://www.wikihow.com/
Set-Rearview-Mirrors-to-Eliminate-Blind-Spots
10. The Web Page of Kristopher Linquist,
http://www.linquist.net/motorsports/tech/mirrors/
11. PhysicsForums,
http://www.physicsforums.com/showthread.php?s=63a4705cd010d3
b75fca0df6d3ff6423&t=442701
12. Adjusting Your Mirrors Correctly, http://www.smartmotorist.com/car-accessories-fuel-
and-maintenance/adjusting-your-mirrors-correctly.html
13. Part571.111: Standard No. 111; Rearview Mirrors, http://www.fmcsa.dot.gov/rules-
regulations/administration/fmcsr/fmcsrruletext.aspx?reg=r49CFR571.111
14. Lucas, B.D., Kanade, T.: Lucas–Kanade Optical Flow Method,
http://en.wikipedia.org/wiki/Lucas%E2%80%93Kanade_Optical_Fl
ow_Method
15. Lucas, B.D., Kanade, T.: Computer Science Department, Carnegie-Mellon University,
http://cseweb.ucsd.edu/classes/sp02/cse252/lucaskanade81.pdf
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 227–232, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Fuzzy Based PSO for Software Effort Estimation
P.V.G.D. Prasad Reddy1 and CH.V.M.K. Hari2
1 Department of Computer Science & Systems Engineering, Andhra University
prasadreddy.vizag@gmail.com
2 Department of IT, Gitam Institute of Technology, GITAM University
kurmahari@gmail.com
Abstract. Software Effort Estimation is the most important activity in project
planning for Project Management. This Effort estimation is required for estima-
tion of resources, time to complete the project successfully. Many models have
been proposed, but because of differences in the data collected, type of projects
and project attributes, no model has been proven successful at effectively and
consistently predicting software development effort due to the uncertainty fac-
tors. The Uncertainty in effort estimation controlled by using fuzzy logic and
the parameters of the Effort estimation are tuned by the Particle Swarm Optimi-
zation with Inertia Weight. We proposed three models for software effort esti-
mation using fuzzy logic and PSO with Inertia Weight. The valuated effort is
optimized using the incumbent archetypal and tested and tried on NASA soft-
ware projects on the basis of three touchstones for assessment of software cost
estimation models. A comparison of the all models is done and it is found that
the incumbent archetypal cater better values.
Keywords: Person Months, Thousands of Delivered Lines of Code, Fuzzy
Logic System, Particle Swarm Optimization, Software Effort Estimation.
1 Introduction
Software project management activities are mainly classified into three categories:
project planning, project monitoring and control, project termination. In project plan-
ning, cost estimation is one of the most important activity. Software cost estimation is
the process of predicting how much amount of effort is required to build software.
The effort is measured in terms of person – months, thus later on it is converted into
actual dollar – cost. The basic input parameters for software cost estimation is size,
measured in KDLOC (Kilo Delivered Lines of Code). A number of models have been
evolved to establish the relation between Size and Effort [9]. The parameters of the
algorithms are tuned using Genetic Algorithms [3], Fuzzy models [4], Soft-
Computing Techniques [5][7], Computational Intelligence Techniques[5], Heuristic
Algorithms, Neural Networks, Radial Basis and Regression [8][9] .
Fuzzy logic [11][12] has rapidly become one of the booming technologies in to-
day’s world for developing convoluted control systems. Fuzzy logic is so efficacious
that it can address any such applications perfectly as it resembles human decision
making with an ability to generate precise solutions from expurgated or unexpurgated
information. It is an interlude in engineering design methods, left vacant by purely
228 P.V.G.D. Prasad Reddy and CH.V.M.K. Hari
mathematical approaches and purely logical approaches in system design. While an
alternative approaches require accurate equations to statuette real -world behaviours,
fuzzy design can inveigle the ambiguities of real-world human language and logic. It
provides both an intuitive method for describing systems in human terms and auto-
mates the conversion of those system specifications into effective models.
A paradigmatic fuzzy system comprises of three crucial components [11][12], are
shown in figure 1, fuzzifier, fuzzy inference engine (fuzzy rules), and defuzzifier. The
fuzzifier metamorphose the input into linguistic monikers using membership function
that delineate how much a given numerical value of a particular variable fits the lin-
guistic term mapping between the input membership functions and the output mem-
bership function using fuzzy rules that can be obtained from proficient knowledge
over relationships being modelled. The greater the input membership degree, the
stronger the pull towards the output membership functions can be contained in the
consequents of the rules triggered.
Fig. 1. Fuzzy Logic System
A defuzzifier carries out the defuzzification process to amalgamate the output into
a single expression or numerical value as required.
2 Fuzzy Logic and PSO for Software Effort Estimation
The accuracy of the estimate will depend on the amount of accurate information, we
have about the final product. The project is being initiated and estimated all alone,
during the feasibility study. Specifications with bewilderment represent a range of
possible final products and not one precisely defined product. Hence, the cost estima-
tion cannot be accurate, the uncertainties are reduced, and more accurate cost esti-
mates can be made. In order to reduce the skeptism, we use fuzzy logic. The cost for a
project in a function of many parameters. Foremost amongst them in size of project in
order to reduce the skeptism at the input level i.e. size, we use triangular membership
function, this process is known as fuzzification. The parameters of the cost model
equation are tuned by using PSO algorithm. By applying fuzzy inference, we get the
suitable equation, finally defuzzication is done through weighted average method
which is actually translates fuzzy values into output.
Particle Swarm has two primary operators: Velocity update and Position update.
During each generation each particle is accelerated toward the particles previous best
position (Pbest) and the global best position (Gbest). The inertia weight is multiplied
Fuzzy Based PSO for Software Effort Estimation 229
by the previous velocity in the standard velocity equation and is linearly decreased
throughout the run. This process is then iterated a set number of times or until a
minimum error is achieved.
The basic concept of PSO lies in accelerating each particle towards its Pbest and
Gbest locations with regard to a random weighted acceleration at each time [12].
3 Model Description
In this model, we considered input sizes as fuzzy values, the parameters a, b, c, d are
tuned by using PSO with inertia weight algorithm. The following in the methodology
is used to estimate the effort.
3.1 Methodology (Algorithm)
Input: Size of the software project, measured effort, methodology.
Output: Optimized parameters and Estimated Efforts.
STEP 1: FUZZIFICATION: The input size in fuzzified by using triangular member-
ship Function shown in Figure 2. The triangular membership function is defined as (α,
m, β), where α, β are left and right side of boundaries and m is the model value. It is
defined as follows:
0, x
x- , x
m-
(; , , )
, x
0,
Y Triangle x m xm
m
x
α
ααβ
α
αβ ββ
β
β
≤
≤≤
==
−≤≤
−
≤
⎧
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎩
(1)
STEP 2:- FUZZY INFERENCE: In this step we apply the following fuzzy rules to
determine the Effort equation.
Fig. 2. Triangular Member Function
a. If input is only size then apply the equation
Effort = a*(size)b (2)
b. Instead of having resources estimates as a function of one variable, resources
estimates can depend on many different factors, giving rise to multivariable mod-
els hence create a non linearity. The cost factors are required software reliability,
230 P.V.G.D. Prasad Reddy and CH.V.M.K. Hari
Execution time constraint, Application experience, Software tools…etc. The
products of all these factors are called Methodology. If input in size and Method-
ology (ME) then
Effort = a*(size)b + c*(ME) (3)
c. As per the above two models size and effort are directly proportional. But such a
condition is not always satisfied giving rise to eccentric inputs. This can be ac-
counted for by introducing a biasing factor (d). If input is size, methodology with
bias then
Effort = a*(size)b + c*(ME)+d (4)
STEP 3:- PARAMETER TUNING: The parameters “a, b, c, d” of the most model
equations are termed by using Particle Swarm Optimization with inertia weight. The
Update of velocity and positions of Parameter “a” is
Vai
k+1 = w * Vai
k + c1 * rand()1 * (Pbest – Sai
k) + c2 * rand()2 * (Gbest – Sai
k) (5)
Sai
k+1 = Sai
k + Vai
k+1 (6)
Similarly for b, c and d.
Step 4:- DEFUZZIFICATION: Finally fuzzy values are translates into actual output
by using weighted average method by using Optimal Parameters obtained in Step2,
Effort equation obtained in step3.
*
ii i
Output w w
μ
=
∑
∑ (7)
4 Model Analysis
4.1 Implementation
We have implemented the above methodology for tuning parameters a, b, c and d in
“C” language. For the parameter’ a ‘the velocities and positions of the particles are
updated by applying the (5) & (6) equations with w=0.5, c1=c2=2.0. And similarly
for the parameters b, c and d.
4.2 Triangular Membership Function
Fuzziness of TFN (α, m, β) is where m= model value, α= left boundary, β=right
boundary. Fuzziness of TFN (F) = β-α/2m, 0<f<1, the higher the value of fuzziness,
the more fuzzy in TFN. The value of fuzziness to be taken depends upon the confi-
dence of the estimator. A confident estimator can take smaller values of f. Let (m, 0)
divides internally, the base of the triangular in ratio k=1, where k in the real positive
number. If we consider F=0.1 and k=1 then α=0.9m and β=1.1m.
4.3 Weighted Average Defuzzification
Defuzzification is a mapping process from fuzzy logic control action area to a non
fuzzy (crisp) control action area. It is done through weighted average method, by
using Equation (7). The Values of a, b, c, d are obtained by using PSO with inertia
Fuzzy Based PSO for Software Effort Estimation 231
weight implemented in C language are as, a = 2.646251 and b = 0.857612 for Model
1, a=2.771722, b=0.84795 and c =-0.007171 for Model 2 and a=3.131606,
b=0.820175, c = 0.045208 and d=-2.020790 for Model 3.
4.4 Performance Measures
We consider three performance criterions [13], Variance accounted – For (VAF),
Mean Absolute Relative Error (MARE), Variance Absolute Relative Error (VARE).
5 Model Experiment
For the study of these models we have taken data of 10 NASA [10] .The following
table1 shows estimated effort of our proposed model and Bailey –Basili, Alaa F.
Sheta and Harish estimation models ,shows using fuzzy based PSO a better software
effort estimation is achieved.
Table 1. Estimated efforts of Various Models
Size
Meas
ured
effort
METH
ODOL
OGY
Bailey –
Basili
Estimate
Alaa F.
Sheta
G.E.
Model
Estimate
Alaa F.
Sheta
Model 2
Estimate
Harish
model1 Harish
model2 MODEL
I MODEL
II MODEL
III
2.1 5 28 7.226 8.44 11.271 6.357 4.257 5.18221 5.18613 5.20013
3.1 7 26 8.212 11.22 14.457 8.664 7.664 7.23726 7.30845 7.35087
4.2 9 19 9.357 14.01 19.976 11.03 13.88 9.39037 9.55991 9.35261
12.5 23.9 27 19.16 31.098 31.686 26.252 24.702 23.92765 24.25429 24.91985
46.5 79 19 68.243 81.257 85.007 74.602 77.452 73.8252 74.34293 74.38504
54.5 90.8 20 80.929 91.257 94.977 84.638 86.938 84.59242 85.06759 84.93563
67.5 98.4 29 102.175 106.707 107.254 100.329 97.679 101.6272 101.9511 101.8466
78.6 98.7 35 120.848 119.27 118.03 113.237 107.288 115.8015 115.9854 115.7576
90.2 115.8 30 140.82 131.898 134.011 126.334 123.134 130.3124 130.4128 129.4197
100.8 138.3 34 159.434 143.0604 144.448 138.001 132.601 143.3405 143.2897 142.0119
6 Performance Analysis
It can be seen that Fuzzy based PSO models out perform the Bailey –Basili, Alaa F.
Sheta and Harish Estimation models. The computed MARE, VAF and VARE for all
models are given in Table 2.
Table 2. Performance Analysis of various models
Model VAF (%) MARE (%) VARE (%)
Bailey –Basili Estimate 93.147 17.325 1.21
Alaa F. Sheta G.E. Model I Estimate 98.41 26.488 6.079
Alaa F. Sheta Model II Estimate 98.929 44.745 23.804
Harish model1 98.5 12.17 80.859
Harish model2 99.15 10.803 2.25
MODEL -I 98.92 6.1657 0.26
MODEL-II 98.545 6.539 0.237
MODEL-III 98.656 6.4731 0.209
232 P.V.G.D. Prasad Reddy and CH.V.M.K. Hari
7 Conclusion
Accurate software development cost estimation is very important in the budgeting,
project planning and control, trade off and rink analysis of effective project manage-
ment. The uncertainty in the input sizes will be reduced by using fuzzy logic. The
parameters of the cost model are lined by using PSO with inertia weight will give the
optimal result. The developed models are tested in NASA software project and proved
to be best on the basis of VARE, MARE and VAF.
References
1. Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learning. Addi-
son-Wesley, Reading (1989)
2. Deb, K.: Multi-Objective Optimization Using Evolutionary Algorithms. John Wiley and
Sons, Chichester (2002)
3. Sheta, A.F.: Estimation of the COCOMO Model Parameters Using Genetic Algorithms for
NASA Software Projects. Journal of Computer Science 2(2), 118–123 (2006); ISSN 1549-
3636
4. Sheta, A., Rine, D., Ayesh, A.: Development of Software Effort and Schedule Estimation
Models Using Soft Computing Techniques. In: 2008 IEEE Congress on Evolutionary
Computation, CEC 2008 (2008); 978-1-4244-1823-7/
5. Gonsalves, T., Ito, A., Kawabata, R., Itoh, K.: Swarm Intelligence in the Optimization of
Software Development Project Schedule. IEEE, Los Alamitos (2008); 0730-3157/08
6. Pahariya, J.S., Ravi, V., Carr, M.: Software Cost Estimation using Computational Intelli-
gence Techniques. 2009 World Congress on Nature & Biologically Inspired Computing
(2009)
7. Attarzadeh, I., Ow, S.H.: Soft Computing Approach for Software Cost Estimation. Interna-
tional Journal of Software Engineering (IJSE) 3(1) (January 2010)
8. Huang, X., Ho, D., Ren, J., Capretz, L.F.: Improving the COCOMO model using a neuro-
fuzzy approach. Elsevier, Amsterdam (2005), doi:10.1016/j.asoc.2005.06.007
9. Sheta, A., Rine, D., Ayesh, A.: Development of Software Effort and Schedule Estimation
Models Using Soft Computing Techniques. IEEE, Los Alamitos (2008) 978-1-4244-1823-
7/08
10. Bailey, J.w., Basili, v.R.: A meta model for software development resource expenditures.
In: Fifth International conference on software Engineering, pp. 107–129. IEEE, Los
Alamitos (1981), CH-1627-9/81/0000/0107500.75@ 1981
11. Ying, H.: The Takagi-Sugeno fuzzy controllers using the simplified linear control rules are
nonlinear variable gain controllers. Automatica 34(2), 157–167 (1998)
12. Pang, W., Wang, K.-P., Zhou, C.-G., Dong, L.-J.: Fuzzy Discrete Particle Swarm Optimi-
zation for Solving Traveling Salesman Problem. In: Proceedings of the Fourth Interna-
tional Conference on Computer and Information Technology (CIT 2004) (2004) 0-7695-
2216-5/04
13. Hari, C.V.M.K., Prasad Reddy, P.V.G.D., Jagadeesh, M.: Interval Type 2 Fuzzy Logic for
Software Cost Estimation Using Takagi-Sugeno Fuzzy Controller. In: Proceedings of 2010
International Conference on Advances in Communication, Network, and Computing.
IEEE, Los Alamitos (2010), doi:10.1109/CNC.2010.14, 978-0-7695-4209-6/10
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 233–239, 2011.
© Springer-Verlag Berlin Heidelberg 2011
SLV: Sweep Line Voronoi Ad Hoc Routing Algorithm
E. Rama Krishna1, A. Venkat Reddy2, N. Rambabu3, and G. Rajesh Kumar4
1 VITS, Karimnagar, Andhra Pradesh, India
erroju.ramakrishna@gmail.com
2 JITS, Karimnagar, Andhra Pradesh, India
venkat_scce@yahoo.co.in
3 SCCE, Karimnagar, Andhra Pradesh, India
neelabhi_2004@rediffmail.com
4 VITS, Karimnagar, Andhra Pradesh, India
gunupudirajesh@gmail.com
Abstract. Mobile Ad hoc routing is challenging issue due to dynamic and dis-
tributed nature of mobile nodes required frequent observed updations of node
information to route the packets appropriately, there are several proactive and
reactive routing methods and geographical hierarchical routing techniques to
route the packets from source to destination effectively. Based on the character-
istics of Mobile Ad Hoc Network (MANET) routing protocols the routing tech-
niques should address the distributed, dynamic topology, periodic updates and
physical security. We proposed a novel mechanism concentrating to minimize
the existing DSDV routing limitations; it is originated from computational
geometric concepts. Fortune’s Algorithm (Sweep line method) which address
the characteristic complexities of MANET routing and representing the route
information in binary search method which minimizes the searching for optimal
route.
Keywords: MANET, DSDV, routing, Fortune’s Algorithm, Geographical hier-
archical routing, Sweepline method.
1 Introduction
Mobile Ad hoc networks (MANETs) are those that instinctively form without the
need of an infrastructure or centralized controller this type of peer-to-peer system
infers that each node in the network can act as a data endpoint or intermediate re-
peater. A node in a wireless ad hoc network is connected with unrestricted nomadicity
(or mobility), the topology of the network is dynamic. The random walk mobility
model which is a stochastic process that models random continuous motion a mobile
node moves from its current location with a randomly selected speed in a randomly
selected direction. Routing evaluate packet delivery ratio, packet drop, end to end
delay etc and determining performance parameters of any ad hoc communication
network.
Computational geometry is the algorithmic study of geometric problems. Its emer-
gence coincided with application areas such as computer graphics, computer-aided
design/manufacturing, and scientific computing, which together provide much of the
234 E.R. Krishna et al.
motivation for geometric computing [1].The basic idea of the design is to operate
each Mobile Host as a specific router, which frequently advertises its view of the in-
terconnection topology with other Mobile Hosts within the network and the idea
is based on computational geometry, a sweepline algorithm or plane sweep algorithm
is a type of algorithm that uses a conceptual sweep line or sweep surface to solve
various problems in Euclidean space, a set of points satisfying certain relationships,
expressible in terms of distance and angle. It is one of the key techniques in computa-
tional geometry.
In order to make this complete mathematically precise it must clearly define the
notions of distance, angle, translation, and rotation is to define the Euclidean plane as
a two-dimensional real vector space equipped with an inner product [2]
• the vectors in the vector space maps to the points of the Euclidean plane,
• the addition operation in the vector space maps to conversion, and
• the inner product implies notions of angle and distance, which can be used to
define rotation
The design of network protocols for Mobile Ad hoc Networks (MANETs) is a com-
plex issue. These networks need efficient distributed algorithms to determine network
organization (connectivity), link scheduling, and routing. The shortest path (based on
a given cost function) from a source to a destination in a static network is usually the
optimal route, this idea is not easily extended to MANETs. Factors such as power
expended, variable wireless link quality, propagation path loss, fading, multiuser in-
terference, and topological changes, become relevant issues [3].
Any protocol must competently handle several inherent characteristics of
MANETs:
• Dynamic topologies, Power constrained operation: Nodes are free to move ran-
domly thus, the network topology which is typically multi hop may transform
randomly and rapidly at erratic times, power conservation is crucial in mobile
wireless systems since these networks typically operate off power-limited sources
and may consist of both bidirectional and unidirectional links.
• Bandwidth-constrained, variable capacity links: Wireless links will continue to
have significantly lower competence than their hardwired counterparts. In addi-
tion, the realized throughput of wireless communications after accounting for the
effects of multiple access, fading, noise, and interference conditions, etc.
• Physical security: Mobile networks are more vulnerable to physical security
threats such as eavesdropping and congestion attacks.
2 Dynamic Destination-Sequenced Distance-Vector Routing
Protocol (DSDV)
The Destination-Sequenced Distance-Vector (DSDV) Routing Algorithm is based on
the idea of the classical Bellman-Ford Routing Algorithm with certain improvements
[4]. An n-node ad hoc network, maintains n rooted trees, one for each destination. The
main role of the algorithm is to solve the routing loop problem
SLV: Sweep Line Voronoi Ad Hoc Routing Algorithm 235
Advantages of DSDV
DSDV is an efficient protocol for route discovery. Whenever a route to a new destina-
tion is required, it already exists at the source. Hence, latency for route discovery is
very low, DSDV also guarantees loop-free paths.
Disadvantages of DSDV
DSDV needs to send a lot of control messages. These messages are important for
maintaining the network topology at each node. This may generate high volume of
traffic for high-density and highly mobile networks. Special care should be taken to
reduce the number of control messages.
In this paper a novel approach to minimize the limitations of the DSDV is pro-
posed, where a systematic arrangement of mobile nodes performs periodic updates of
the routing information about its clustered neighbor nodes with systematic arrange-
ment of all possible routing paths from source node to destination node. This ap-
proach two phase process, in which Nearest neighbor search followed by Fortune’s
algorithm which generates Voronoi diagram from a set of points (nodes) in a plane.
3 Nearest Neighbor Search
Nearest neighbor search (NNS), also known as proximity search, similarity search or
closest point search, is an optimization problem for finding closest points in metric
spaces. The problem is: given a set S of points in a metric space M and a query point
q ε M, find the closest point in S to q. In many cases, M is taken to be d-dimensional
Euclidean space and distance is measured by Euclidean distance [5]. The Nearest
Neighbor Search Algorithm has different solution like linear search, locality sensitive
hashing, vector approximation file, Space partitioning methods.
3.1 Space Partitioning Method
Space-partitioning systems are often hierarchical, meaning that a space (or a region of
space) is divided into several regions, and then the same space-partitioning system is
recursively applied to each of the regions thus created. Most space-partitioning sys-
tems use planes (or, in higher dimensions, hyper planes) to divide space: points on
one side of the plane form one region, and points on the other side form another.
Points exactly on the plane are usually arbitrarily assigned to one or the other side.
Fig. 1. A set S of n points in d dimensions
query point q Fig. 2. Nearest neighbor graph on a set S of
n points links each vertex
236 E.R. Krishna et al.
Recursively partitioning space using planes in this way produces a BSP tree, one of
the most common forms of space partitioning [5].
4 Fortune’s Algorithm
The nearest neighbor queries as the Voronoi diagram is being computed, instead of
storing it in an auxiliary data structure, using Fortune's algorithm which is a plane
sweep algorithm for generating a Voronoi diagram from a set of points in a plane us-
ing O(n log n) time and O(n) space. This algorithm is based on the following clever
idea: rather than considering distances between the various sites, we will introduce a
line that moves through the plane and use this line to facilitate a more efficient com-
parison of distances. We call this line the sweep line and think of it as uncovering the
Voronoi diagram as it sweeps through the plane.
Let's first remember that if we are given a point p and a line l (not containing p),
then the set of points whose distance to p equals its distance to l forms a parabola. We
will use Pp,l to denote this parabola. As before, it is helpful to think of the parabola as
dividing the plane into two regions: one consisting of points closer to p and other con-
sisting of points closer to l. This is an important point so, let's make it a little more
clear. Consider a point q whose coordinates are q=(qx,qy) and whose distance to p is
denoted by d(p,q) . In what follows, the sweep line will always be a horizontal line;
we will call its vertical coordinate ly , so that the distance between q and l is qy-ly. The
condition that q lies on the parabola Pp,l is therefore d(p,q) = qy-ly
Fig. 3. Parabola dividing the plane in to two regions
More generally
d(q,p)< qy-ly if q lies above Pp,l
d(q,p)= qy-ly if q lies on Pp,l
d(q,p)< qy-ly if q lies below Pp,l
now the horizontal sweep line l will move down through the plane. At any time, we
will only consider the sites that lie above the sweep line and the parabolas Pp,l define
by those.
Fortune's algorithm: the beach line is the curve formed by the lowest parabolic
arcs. That is, any vertical line will pass through several parabolas; the point at which
the vertical line passes through the beach line is the lowest such point. Notice that
each of the arcs that compose the beach line is associated to one of the sites above the
sweep line.
SLV: Sweep Line Voronoi Ad Hoc Routing Algorithm 237
Fig. 4. Collection of sites and the position o
f
the sweep line
Fig. 5. The parabola
Fig. 6. Collection of sites and the position of the sweep line with parabola
The beach line is well suited for constructing the Voronoi diagram.
Let's now determine when the beach line passes through some arbitrary point q.
Suppose that q is as close to site p1 as any other; that is, d(q,p1) d(q,pi)for all other
sites pi . The condition that q lies on Pp1,l may be expressed as
d(q,p1)= qy-ly
Therefore
d(q,pi) d(q,p1) = qy-ly
This means that when q appears on Pp1,l, it cannot be above another parabola Ppi.l,
therefore, when q appears on the parabola Pp1,l, it is on the beach line When a point
appears on the beach line, it is on a parabolic arc associated to its nearest site. The
points on the beach line that lie on two parabolic arcs are called breakpoints. The
breakpoints lie on the edges of the Voronoi diagram. This means that the breakpoints
will sweep out the edges of the Voronoi diagram as the sweep line moves down the
plane. So to construct the Voronoi diagram, we simply need to keep track of the
breakpoints [6]. The Fortune’s algorithm generates Voronoi diagram as an arrange-
ment of nodes maintaining the information about the neighbor nodes these updates
occur periodically which concludes the shortest and optimal path from source to des-
tination is estimated.
238 E.R. Krishna et al.
5 Delaunay Triangles
The subdivision of the space determined by a set of distinct points so that each point
has associated with it the region of the space nearer to that point than to any other is
called Dirichlet tessellation. This process applied to a closed domain generates a set
of convex distinct polygons called Voronoi regions which cover the entire domain.
This definition can be extended to higher dimension where, for example in three di-
mensions, the Voronoi regions are convex polyhedrons [8]. If we connect all the pairs
of points sharing a border of a Voronoi region we obtain a triangulation of the convex
space containing those points. This triangulation is known as Delaunay triangulation.
The Delaunay triangle considered as a conceptual path to cover the nodes in par-
ticular region (Voronoi region), each node gets the information about its neighbor
regions and the Delaunay property measure the existence of node.
Fig. 7. Voronoi diagram (in dotted lines) Fig. 8. Delauna
y
trian
g
ulation, on to
p
of the
Voronoi diagram
6 Routing Information Representation
Fortune’s algorithm or beach line method can be used to represented routing informa-
tion as Beach Line Data Structure (T), derive a balanced binary search tree
• Internal nodes represent break points between two arcs
• Leaf nodes represent arcs, each arc in turn is represented by the site that has
generated it
This tree structure constructed from the Sweepline (or Beach line) method of for-
tune’s algorithm.
The operations performed on the data structure representation are
Search: Given the current y-coordinate of the sweep line and a new site pi, deter-
mine the arc of the beach line lies immediately above pi. Let pj denote the site that
contributes this arc. Return a reference to this beach line entry.
Insert and split: Insert a new entry for pi within a given arc pj of the beach line
(thus effectively replacing the single arc <…, pj,….> with the triple <….,pj , pi,
pj ,….>.Return a reference to the newly added beach line entry (for future use).
Delete: Given a reference to an entry pj on the beach line, delete this entry. This re-
places a triple <…., pi, pj, pk,….> with the pair <….,pi, pk,…..>
It is not difficult to modify a standard dictionary data structure to perform these
operations in O(log n) time each [7].
SLV: Sweep Line Voronoi Ad Hoc Routing Algorithm 239
Fig. 9. Balanced binary search tree Fig. 10. Beach line method
7 Conclusion
Sweep Line Voronoi (SLV)method is follow the geographic routing technique which
exhibit the better performance with highly dynamic and distributed natured mobile
nodes which play significant role in estimating the node location with periodic and
uniform updating the node information with a geometric relative attributes of the
nodes enables to process the connectivity related parameters and routing information
processing without flood the packets to identify the nodes to establish the route.
Sweep line or beach line is systematic process which process the node information
to establish the route in a step by step approach, which creates a Voronoi based
structure indicating the node range and Delaunay structure to possible route path
these mathematical related information can predict the route alternates when the ex-
isting route is disturbed. The scan line creates a balanced binary search tree based
route information which enables the optimal processing of route discovery and node
estimation.
References
1. Skiena, S.S.: The algorithm design manual
2. http://en.wikipedia.org/wiki/Euclidean_space
3. Subbarao, M.W.: Performance of Routing Protocols for Mobile Ad-Hoc Networks
4. Mahdipour, E., et al.: Performance Evaluation of DSDV Routing Protocol
5. http://en.wikipedia.org/wiki/Nearest_neighbor_search
6. Voronoi Diagrams and a Day at the Beach, David Austin Grand Valley State University
7. http://ww.cs.umd.edu/class/spring2010/cmsc754/Lects/lect11.pdf
8. Gorbach, M.: Image Stained Glass using Voronoi Diagrams
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 240–244, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Hybrid Routing for Ad Hoc Wireless Networks
Ravilla Dilli1, R.S. Murali Nath2, and P. Chandra Shekar Reddy3
1 Associate Professor, Dept. of ECE, St. Martin’s Engineering College,
Secunderabad A.P., India
dilli.ravilla@gmail.com
2 Professor, Dept. of CSE,
St. Martin’s Engineering College
Secunderabad, A.P., India
muralinath.r.s@gmail.com
3 Professor Coordinator, Dept.of ECE,
JNT University,
Hyderabad, A.P, India
drpcsreddy@gmail.com
Abstract. The Zone Routing Protocol (ZRP) is a hybrid routing protocol that
proactively maintains routes within a local region of the network (which we re-
fer to as the routing zone). Here, we describe the motivation of ZRP and its ar-
chitecture also the query control mechanisms, which are used to reduce the traf-
fic amount in the route discovery procedure. In this paper, we address the issue
of configuring the ZRP to provide the best performance for a particular net-
work, at any time. Through NS2 simulation, we draw conclusions about the per-
formance of the protocol.
Keywords: Zone Routing Protocol, Routing zone, Query control mechanisms.
1 The ZRP Architecture
As proactive routing uses excess bandwidth to maintain routing information, while
reactive routing involves long route request delays. The Zone Routing Protocol (ZRP)
aims to address the problems by combining the best properties of both approaches.
ZRP divides its network into different zones. A routing zone is defined for each
node separately, and the zones of neighboring nodes overlap. ZRP refers to the locally
proactive routing component as the Intra-zone Routing Protocol (IARP) [5] which
maintains routing information for nodes that are within the routing zone of the node
and the globally reactive routing component as the Inter-zone Routing Protocol (IERP)
[2] which offer enhanced route discovery and route maintenance services based on
local connectivity monitored by IARP. The protocol identifies multiple loop-free
routes to the destination, increasing reliability and performance.
ZRP [3] uses a concept called bordercasting using the Bordercast Resolution Proto-
col (BRP) [6] is used in the ZRP to direct the route requests initiated by the global
reactive IERP [2] to the peripheral nodes, thus removing redundant queries and
maximizing efficiency. By employing query control mechanisms, route requests can
be directed away from areas of the network that already have been covered. In order to
Hybrid Routing for Ad Hoc Wireless Networks 241
detect new neighbor nodes and link failures, the ZRP relies on a Neighbor Discovery
Protocol (NDP) provided by the Media Access Control (MAC) layer. NDP transmits
“HELLO” beacons at regular intervals. IERP uses the routing table of IARP to respond
to route queries. IERP forwards queries with BRP.
1.1 Routing
A node that has a packet to send first checks whether the destination is within its local
zone using information provided by IARP. In that case, the packet can be routed proac-
tively. Reactive routing is used if the destination is outside the zone.
Fig. 1. ZRP architecture
The reactive routing process is divided into two phases: the route request phase and
route reply phase. In the first, the source sends a route request packet to its peripheral
nodes using BRP [6]. If the receiver of a route request packet knows the destination, it
responds by sending a route reply back to the source. Otherwise, it continues the proc-
ess by bordercasting the packet. The reply is sent by any node that can provide a route
to the destination. To be able to send the reply back to the source node, routing infor-
mation is accumulated when the request is sent through the network. In ZRP, the
knowledge of the local topology can be used for route maintenance. Link failures and
sub-optimal route segments within one zone can be bypassed. [4]
2 Query-Control Mechanisms
ZRP uses 3 types of query-control mechanisms: Query Detection, Early Termination,
and Random Query- Processing Delay. To be able to prevent queries from reappearing
in covered regions, the nodes must detect local query relaying activity. BRP provides
two query detection methods: QD1 and QD2. QD1 is used to detect the nodes that re-
lay the query. QD2 is used in single-channel networks to listen to the traffic by other
Table 1. Variable Simulation Parameters
Parameter Symbol Values Defaults
Zone Radius (hops) ρ 1-10 ---
Node Density(neighbors/node) δ 2-10 5
Rel. node speed(neighbors/s) V 0.2-2.2 1.0
# of Nodes(nodes) N 200-1500 400
242 R. Dilli, R.S.M. Nath, and P.C. Shekar Reddy
nodes within the radio coverage (QD2). With Early Termination (ET), a node can pre-
vent a route request from entering already covered regions. To reduce the probability
of receiving the same request from several nodes, a Random Query-Processing Delay
(RQPD) can be employed.
3 ZRP Performance Results
We expect that the amount of IARP [5] traffic should be proportional to the node den-
sity (δ) and the routing zone area (ρ2). The IERP traffic/query decreases with the zone
radius. For a given zone radius, we observe that the amount of received traffic/query
increases with the zone density.
Keeping the zone radii and node density fixed, an increase in network population
(N) is reflected by an increase in the network span. This has the effect of increasing
the length of an IERP source route, thereby reducing the route’s reliability. Fig. 2(a)
and (b) illustrates the effect of node density (δ) on the production of ZRP traffic.
(a) (b)
Fig. 2. ZRP traffic node (N=1500 nodes, v=1.0 neighbors/s) (a) Rroute -usage << Rroute-failure. (b) Rroute-usage
>> Rroute-failure.
We have seen that an increase in node density results in more IARP [5] route up-
dates and more IERP packets per query. Fig. 2(b) demonstrates that the reduction in
route query rate is more than offset by the increase in overlapping query packet
transmission. In general, it appears that the ZRP traffic is an increasing function of
node density. Also, the average node velocity (v) is a measure of the rate of network
reconfiguration. Higher node velocities result in a linear increase in the IARP routing
zone updates and IERP route failures.
(c) (d)
Fig. 2. (Continued): ZRP traffic node (N=1500 nodes, d=5.0 neighbors/s) (c) Rroute-usage <<Rroute-
failure. (d) R route-usage >> Rroute-failure
Hybrid Routing for Ad Hoc Wireless Networks 243
The overall ZRP traffic increases linearly with v, and ρopt remains constant. The
node population influences the ZRP through its effect on the rate of received route
queries. The addition of a new network node places additional load on the network
through the extra route queries that it initiates. All other factors remaining constant,
an increase in N results in an increase in network span. When the rate of route queries
is driven by route failure, we have shown that larger network span reduces route
reliability, further increasing the query load by all other nodes. Fig.2 (e) and (f) dem-
onstrates this behavior.
(e) (f)
Fig. 2. (Continued) ZRP traffic node (5.0 neighbors, v=1.0 neighbors/s) (e) Rroute-usage <<Rroute-fa ilure.
(f)Rroute-usage >> Rroute-failure
As we expected, the amount of ZRP traffic increases with N and larger N favors a
larger routing zone radius.
4 Conclusions
The ZRP combines two radically different methods of routing into one protocol. The
tradeoff between the costs of proactive and reactive components of the ZRP deter-
mines the optimal zone radius for a given network configuration. The span of the
network does not affect the amount of intrazone traffic, but the amount of reactive
route query traffic increases with network span, thereby favoring larger routing zones.
Relative node velocity has the potential to increase both intrazone updates and inter-
zone route queries. We have seen that the total ZRP traffic increases with node
density. We have proposed and evaluated zone radius estimation algorithm ‘min
searching”, which attempts to minimize the amount of ZRP traffic based on direct
measurements of the traffic. Our results demonstrate that the proposed route estima-
tion techniques, applied in conjunction with a simple radius update protocol, allow the
ZRP to perform more efficiently than traditional routing protocols without the need
for centralized control or knowledge of the network operating conditions.
References
[1] Park, V.D., Corson, M.S.: A highly adaptive distributed routing algorithm for mobile wire-
less networks. In: Proc. IEEE INFOCOM 1997, Kobe, Japan, pp. 1405–1413 (1997)
[2] Haas, Z.J., Pearlman, M.R., Samar, P.: Interzone Routing Protocol (IERP). IETF Internet
Draft, draft-ietf-manet-ierp-02.txt (July 2002)
244 R. Dilli, R.S.M. Nath, and P.C. Shekar Reddy
[3] Haas, Z.J., Pearlman, M.R., Samar, P.: The Zone Routing Protocol (ZRP) for Ad Hoc
Networks. IETF Internet Draft, draft-ietf-manet-zone-zrp-04.txt (July 2002)
[4] Beijar, N.: Zone Routing Protocol,
http://www.tct.hut.fi/opetus/s38030/k02/Papers/
08-Nicklas.pdf
[5] Haas, Z.J., Pearlman, M.R., Samar, P.: Intrazone Routing Protocol (IARP). IETF Internet
Draft, draft-ietf-manet-iarp-02.txt (July 2002)
[6] Haas, Z.J., Pearlman, M.R., Samar, P.: Bordercasting Resolution Protocol (BRP). IETF
Internet Draft, draft-ietf-manet-brp-02.txt (July 2002)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 245–247, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Implementation of ARINC 429 16 Channel Transmitter
Controller on FPGA
Debasis Mukherjee1, Niti Kumar2, Kalyan Singh3, Hemanta Mondal4,
and B.V.R. Reddy5
1 SPSU, E&CE, Udaipur & GGSIPU, USIT, Delhi, India
debasismukherjee1@gmail.com
2 GTBIT, E&CE, Delhi, India
niti221@rediffmail.com
3 AIT, RF & Microwave, New Delhi, India
kalyan_mail@yahoo.com
4 Freescale Semiconductor, Noida, India
hemanta.mandal@rediffmail.com
5 GGSIPU, USIT, Delhi, India
bvrreddy64@rediffmail.com
Abstract. Multi-Channel transmitter Controller is designed on ARINC 429 Pro-
tocol. Its new architecture shows better performance and saves high percentage
of total power i.e. static power and dynamic power. It is specially designed for
providing a dedicated transmission link between processor and various equip-
ments placed in aircraft by selecting the particular transmitter for transmitting
the data to many equipments like Microwave Landing system, Tire Pressure
Measuring Systems etc. Also, an efficient VLSI implementation of proposed
system on PROASIC 3 plus ACTEL FPGA device is described.
Keywords: Total power, dynamic power and static power.
1 Introduction
Multi-Channel transmitter Controller is designed using the specifications of
ARINC429. It supports up to 16 transmitter channels. It sets up unidirectional data
transmission between micro-processor and ARNIC429 supported equipments & has
externally programmable control unit for selecting the transmission rate.
ARINC 429 supported equipments are installed in various commercial transport
aircrafts. Many electronic equipments are placed in aircraft and each equipment has
its unique identification number. Multi-channel receiver controller receives data from
these equipments.
A VLSI implementation of the proposed architecture on FPGA device is described.
The whole design was captured in VHDL language and was synthesized.
2 ARINC 429 Specifications
ARINC 429 specifications defines how the avionic equipments and systems should
communicate with each other. These specifications includes electrical and data
246 D. Mukherjee et al.
characteristics. The transmitter is used to transmit information from the defined port
to 16 receivers over a single twisted and shielded wire pairs. Data is transmitted at a
bit rate of 12.5 Kbps for low speed or 100 Kbps for high speed. ARNIC429 supported
electrical signal has three states, is shown in fig-1. ARINC 429 supports three states
of a signal which are high state driven by +5 V, Null state driven by 0V and low state
driven by -5V. Bipolar return to zero encoding scheme is used to solve the synchroni-
zation problem. Implementation of ARNIC429 specified signal is not possible
through single pin because it has three states so desired signal is separated into two
digital signals called Signal A and signal B as shown in the fig– 2. These signals are
transmitted via two pins. The bit value of each state of signals A and B specifies the
three states of desired signal. ARNIC429 supports 32 bit data word. The data word
format consists of parity, label, sign/status matrix, data and source/destination identi-
fier. MSB is the parity bit which is an odd parity. The label is used to specify the type
of data. Source/destination identifier is used for multiple receivers to identify the re-
ceiver for which the data is destined.
Fig. 1. ARINC Standard Fig. 2. Signal A and B
3 Architecture and Functional Description of Single Channel
Transmitter
Single channel transmitter has two 16 bits latches .These latches are used to store the
16-16 bits data at positive and negative edge trigger of load signal. It uses signal
Txsel to select the transmission rate. If Txsel is high then high transmission speed is
passed and if it is low then low transmission speed is passed. This speed is given to
the counter and modulator. Just as the 16 bits MSB of data is loaded, FF1 is set to
enable the counter. This counter is of 6 bits but it counts only up to 36 counts where
32 counts are for transmitting the 32 bits and 4 counts are used to generate the four bit
gap to distinguish between two words .The counter gives its count value to the modu-
lator and based on this value of it check the bits of 32 bits data word and performs
parallel to serial conversion. Its architecture is shown in fig.3.
4 Architecture and Functional Description of 16-Channel
Transmitter Controller
16 channel transmitter controller operates on 16 MHz clock. It has four bit address
bus to select the particular transmitter channel. It has a 16 bit control register and each
Implementation of ARINC 429 16 Channel Transmitter Controller on FPGA 247
bit of which is assigned to each transmitter. The value of bits will decide at which
transmission rate the transmitter will transmit the data serially. If bit value of control
register is ‘0’ then low frequencies such as 100 KHz and 12.5 kHz is selected and if
the bit is ‘1’ then high frequency such as 200 KHz and 100 KHz are selected. The
decoder is used to select the particular channel for transmission of data: fig-4.
Fig. 3. Single Channel Transmitter Fig. 4. 16 Channel Transmitter
5 Results and Conclusions
Family- ProASIC3, Die-A3P250, Package-208 PQFP is used.
Table 1. The comparison of Battery life and power consumption
Architecture Transmitter Controller with 16
Control registers
Transmitter Control-
ler with 1 Control
Register
Battery Capacity 1000.000 mA*Hours 1000.000 mA*Hours
Battery Life 2.026 Hours 2.118 Hours
Total Power 1576.309 mW 1510.567 mW
Static Power 5.220 mW 5.220 mW
Dynamic Power 1571.089 mW 1505.347 mW
New architecture of ARNIC429 Transmitter and it’s multi-channel transmitter con-
troller is introduced which provides an easy interfacing with 16 bit micro-processor
and controlled by micro-processor. Proposed architecture use less hardware of FPGA
chip, consumes less power and increases battery life also.
References
1. Pschierer, C., Kasten, J., Schiefele, J., Lepori, H., Bruneaux, P., Bourdais, A., Andreae, R.:
ARINC 424A A next generation navigation database specification. In: IEEE/AIAA 26th
Digital Avionics Systems Conference, DASC 2007, pp. 2.B.6-1 – 2.B.6-8 (2007)
2. Arnic, http://www.ARINC.com
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 248–254, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Segmentation of Image Using Watershed and Fast Level
Set Methods
Minal M. Puranik and Shobha Krishnan
S.A.K.E.C, Mumbai
+91-9323106641
Minalpuranik@gmail.com
VESIT, Mumbai
25227638(404)
shobha krishnan@hotmail.com
Abstract. Technology is proliferating. Many methods are used for medical im-
aging. The important methods used here are fast marching and fast level set in
comparison with the watershed transform. Since watershed algorithm was ap-
plied to an image has over clusters in segmentation. Both methods are applied
to segment the medical images. First, fast marching method is used to extract
the rough contours. Then fast level set method is utilized to finely tune the ini-
tial boundary. Moreover, Traditional fast marching method was modified by the
use of watershed transform. The method is feasible in medical imaging and de-
serves further research. In the future, we will integrate level set method with
statistical shape analysis to make it applicable to more kinds of medical images
and have better robustness to noise.
Categories and Subject Descriptors
I.4 Image Processing and Computer Vision.
General Terms
Design, Experimentation.
Keywords: Segmentation, Watershed transform, fast marching method and fast
level set method.
1 Introduction
In computer vision, segmentation refers to the process of partitioning a digital image
into multiple segments (sets of pixels) (Also known as super pixels). The goal of
segmentation is to simplify and/or change the representation of an image into some-
thing that is more meaningful and easier to analyze. Image segmentation is typically
used to locate objects and boundaries (lines, curves, etc.) in images. More precisely,
image segmentation is the process of assigning a label to every pixel in an image such
that pixels with the same label share certain visual characteristics.
Segmentatio
n
1.1 Fast Marching Meth
o
The Fast Marching Metho
red), stand on the lowest s
p
little bit of the surface that
the process over and over
,
building that little bit of t
h
been built.
Fig. 1.
C
Gr
e
The speed from this m
e
scaffolding; fortunately, th
e
quickly.Equation of the ev
o
zero level set is then given
b
The major advantages of
u
clude the following [10]. F
tion, but the propagating
f
sharp corners as Φ evolv
e
may be easily determined
f
vector is given by n =
Φ
interfaces in three dimensi
o
called fast marching metho
speed > 0 (the case where
F
monotonically advancing f
r
This simply says that the g
r
of the surface. There are t
w
face: iteration towards the
Eq. (1) or explicit construc
t
method depends on the latt
e
tions. Sethian proved that
i
in order to get the arrival ti
m
n
of Image Using Watershed and Fast Level Set Methods
o
d
d imitates this process. Given the initial curve (show
n
p
ot (which would be any point on the curve), and bui
l
corresponds to the front moving with the speed F. Re
p
,
always standing on the lowest spot of the scaffold,
h
e surface. When this process ends, the entire surface
C
onstruction of stationary level set solution
e
en squares show next level to be built
e
thod comes from figuring out in what order to build
e
re are lots of fast sorting algorithms that can do this
o
lution of Φ, inside which our surface is embedded as
b
y the following equation:
Φt+F| Φ|=0
u
sing this method over other active contour strategies
irst, the evolving level set function Φ(x, t) remains a f
u
f
ront Γ (t) may change topology, break, merge and f
o
e
s. Second, the intrinsic geometric properties of the f
r
f
rom Φ. For example, at any point of the front, the nor
m
Φ
. Finally, the formulation is unchanged for propaga
t
o
ns. One of the most popular level set algorithms is the
d. Now consider the special case of a surface moving
w
F
is everywhere negative is also allowed). We then ha
v
r
ont whose level set equation is of the followingForm:
|T|F=1
r
adient of arrival time is inversely proportional to the s
p
w
o ways of approximating the position of the moving
s
solution by numerically approximating the derivative
s
t
ion of the solution function T from Eq. (2). Fast marc
h
e
r approach. Equation (2) is one form of the Eikonal e
q
i
t is equivalence to solve the Following quadratic equa
t
m
e T of the Eq. (2).
249
n
in
l
d a
p
eat
and
has
the
job
the
(1)
in-
u
nc-
o
rm
r
ont
mal
t
ing
so-
w
ith
v
e a
(2)
p
eed
s
ur-
s
in
h
ing
q
ua-
t
ion
250 M.M. Puranik and S. Krishnan
(3)
Where D+ and D− are backward and forward difference operators:
(4)
The steps of the traditional fast marching method are as follows:
1. Initializing step:
1) Alive points: Let A be the set of all grid points {iA, jA} which represents the
initial curve. In our algorithm, Alive points are the seeded points users assign
to. See Fig. 2;
2) Narrowband points: Let Narrowband points be the set of all grid point neigh-
bors of A. In our Algorithm, those are the 4-nearest points of the seeded
points. Set T(x, y) = 1/F(x, y).
3) Faraway points: Let Faraway points be the set of all others grid points {x, y}.
Set T(x, y) =TIME MAX, ;
1.2 Marching Forwards
1) Begin loop: Let (imin, jmin) be the point in Narrowband with the smallest value
for T;
2) Add the point (imin, jmin) to A; remove it from Narrowband;
3) Tag as neighbors any points (imin − 1, jmin), (imin + 1, jmin), (imin, jmin − 1),
(imin, jmin + 1)
that is either in Narrowband or Faraway. If the neighbor is in Faraway, remove it
from that list
and add it to the set Narrowband;
4) Recomputed the values of T at all neighbors according to equation 3, selecting the
largest possible
solution to the quadratic equation;
5) Return to top of Loop.
1.3 Level Set Method
In implementing the traditional level set methods, it is numerically necessary to keep
the evolving level set function close to a signed distance function. Reinitialization, a
technique for periodically re-initializing the level set function to a signed distance
function during the evolution, has been extensively used as a numerical remedy for
maintaining stable curve evolution and ensuring usable results. However, as pointed
out by Gomes, re-initializing the level set function is obviously a disagreement
Segmentation of Image Using Watershed and Fast Level Set Methods 251
between the theory of the level set method and its implementation. Moreover, many
proposed re-initialization schemes have an undesirable side.
In this paper, we present a new variational formulation that forces the level set
function to be close to a signed distance function, and therefore completely eliminates
the need of the costly re-initialization procedure. Our variational energy functional
consists of an internal energy term and an external energy term, respectively. The
internal energy term penalizes the deviation of the level set function from a signed
distance function, whereas the external energy term drives the motion of the zero lev-
el set to the desired Formulation has three main advantages over the traditional level
set formulations. First, a significantly larger time step can be used for numerically
solving the evolution PDE, and therefore speeds up the curve evolution. Second, the
level set function could be initialized as functions that are computationally more effi-
cient to generate than the signed distance function. Third, the proposed level set evo-
lution can be implemented using simple finite difference scheme, instead of complex
upwind scheme as in traditional level set formulations. The proposed algorithm has
been applied to both simulated and real images with promising results. In particular it
appears to perform robustly in the presence of weak boundaries.
1.4 Drawbacks Associated with Re-initialization
Re-initialization has been extensively used as a numerical remedy in traditional level
set methods. The standard re-initialization method is to solve the following reinitiali-
zation equation
where φ0 is the function to be re-initialized, and sign(φ) is the sign function. There
has been copious literature on re-initialization methods , and most of them are the
variants of the above PDE-based method. Unfortunately, if φ0 is not smooth or φ0 is
much steeper on one side of the interface than the other, the zero level set of the re-
sulting function φ can be moved incorrectly from that of the original function. More-
over, when the level set function is far away from a signed distance function, these
methods may not be able to re-initialize the level set function to a signed distance
function. In practice, the evolving level set function can deviate greatly from its value
as signed distance in a small number of iteration steps, especially when he time step is
not chosen small enough. So far, re-initialization has been extensively used as a nu-
merical remedy for maintaining stable curve evolution and ensuring desirable results.
From the practical viewpoints, the re-initialization process can be quite complicated,
expensive, and have subtle side effects. Moreover, most of the level set methods are
fraught with their own problems, such as when and how to re-initialize the level set
function to a signed distance function. There is no simple answer that applies general-
ly to date. The variational level set formulation proposed in this paper can be easily
implemented by simple finite difference scheme, without the need of re-initialization.
252 M.M. Puranik and S. Krishnan
2 Variational Fast Level Set Formulation of Curve Evolution
without Re-initialization
2.1 General Variational Fast Level Set Formulation with Penalizing Energy
As discussed before, it is crucial to keep the evolving level set function as an approx-
imate signed distance function during the evolution, especially in a neighborhood
around the zero level set. It is well known that a signed distance function must satisfy
a desirable property of φ = 1. Conversely, any function φ satisfying |φ| = 1 is the
signed distance function plus a constant Naturally, we propose the following integral
as a metric to characterize how close a function φ is to a signed distance function in
. This metric will play a key role in our variational level set formulation.
With the above defined functional P(φ), we propose the following variational formu-
lation E(φ) = μP(φ) + Em(φ).
where μ > 0 is a parameter controlling the effect of penalizingthe deviation of φ
from a signed distance function, and Em(φ) is a certain energy that would drive the
motion of the zero level curve of φ. For a particular functional E(φ) defined explicitly
in terms of φ, the Gateaux derivative can be computed and expressed in terms of the
function φ and its derivatives.
2.2 Variational Level Set Formulation of Active Contours without
Re-initialization
In image segmentation, active contours are dynamic curves that moves toward the
object boundaries. To achieve this goal, we explicitly define an external energy that
can move the zero level curve toward the object boundaries. Let I be an image, and g
be the edge indicator function defined by
,
Where Gσ is the Gaussian kernel with standard deviation σ. We define an external
energy for a function φ(x, y) as below
where λ > 0 and ν are constants, the Heaviside function.Now, we define the following
total energy functional
The external energy Eg,λ,ν drives the zero level set toward the object boundaries,
while the internal energy μP(φ) penalizes the deviation of φ from a signed distance
function during its evolution. The gradient flow is the evolution equation of the level
Segmentation of Image Using Watershed and Fast Level Set Methods 253
set function in the proposed method. The second and the third term in the right hand
side of correspond to the gradient flows of the energy functional, respectively, and are
responsible of driving the zero level curves towards the object boundaries.To explain
the effect of the first term, which is associated to the internal energy μP (φ).
2.3 Selection of Time Step
In implementing the proposed level set method, the time step τ can be chosen signifi-
cantly larger than the time step used in the traditional level set methods. We have
tried a large range of the time step τ in our experiments, from 0.1 to 100.0. For exam-
ple, we have used τ = 50.0 and μ = 0.004 for the image in and the curve evolution only
takes 40 iterations, while the curve converge to the object boundary precisely.
3 Watershed Transform
Watershed transformation is apowerful tool for image segmentation. In this paper, the
different morphological tool used in segmentation is reviewed together with an abun-
dant illustration of the methodology through examples of image segmentation coming
from various areas of image analysis. The gradient image is often used in the wa-
tershed transformation, because the main criterion of the segmentation is the homo-
geneity of the grey values of the object present in the image .But, when other
criteria are relevant, other functions can be used. In particular, when the segmentation
is based on the shape of the objects, the distance functions is very helpful. The
watershed transform finds the catchment basins and ridge lines in such a grayscale
image.
4 Experimental Details (Results)
Input image watershed
fast marching Fast Level set
Fig. 2. Result for microscopic cell image
254 M.M. Puranik and S. Krishnan
Fig. 3. Result for microscopic cell image
5 Conclusion
In this paper, we present a new variational level set formulation that completely eli-
minates the need of he reinitialization, results demonstrates desirable performance of
our method in extracting weak object boundaries, whcich is very difficult for the tra-
ditional level set method, watershed and fast marching method.
References
1. Qiu, P.: Some recent developments on edge detection and image reconstruction based on lo-
cal smoothing and nonparametric regression. In Book Recent Research Developments in
Pattern Recognition 1, 41–49 (2000)
2. McInerney, T., Terzopoulos, D.: Deformable models in medical image analysis: a survey.
Medical Image Analysis 1(2), 91–108 (1996)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 255–258, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Tree Structured, Multi-hop Time Synchronization
Approach in Wireless Sensor Networks
Surendra Rahamatkar1, Ajay Agarwal2, Praveen Sen1, and Arun Yadav3
1 Dept. of Computer Science, Nagpur Institute of Technology, Nagpur, India
rahamatkar_s@rediffmail.com
2 Dept. of MCA, Krishna Institute of Engg. & Technology, Ghaziabad, India
3 Dept. of Computer Sc., BMAS Engineering College, Agra,(U.P.), India
Abstract. Time synchronization for wireless sensor networks (WSNs) has been
studied in recent years as a fundamental and significant research issue. Time
synchronization in a WSN is a critical for accurate time stamping of events and
fine-tuned coordination among the sensor nodes to reduce power consumption.
This paper proposes a reference based, tree structured multi-hop time synchro-
nization service for WSNs in which sensor nodes synchronize by collecting ref-
erence points with reference to Root node of the logically constructed tree
structure of the network. Proposed approach is lightweight as the number of re-
quired broadcasting messages is restricted to one broadcasting domain.
Keywords: Ad Hoc Tree Structure, Clock synchronization, Wireless sensor
networks, Hierarchical sensor network.
1 Introduction
Time synchronization for wireless sensor networks (WSNs) has been studied in recent
years as a fundamental and significant research issue. Many applications based on
these WSNs assume local clocks at each sensor node that need to be synchronized to a
common notion of time. Time synchronization, which aims at creating a common
time scale in the network, is a necessity for many applications.
Two of the most prominent examples of existing time synchronization protocols
developed for the WSN domain are the Reference Broadcast Synchronization (RBS)
algorithm [4] and the Timing-sync Protocol for Sensor Networks (TPSN) [5]. Tiny-
SeRSync [7] protocol works with the ad hoc deployments of sensor networks. This
protocol proposed two asynchronous phases: Phase I –secure single-hop pair wise
synchronization, and Phase II–secure andresilient global synchronization to achieve
global time synchronization in a sensor network.
2 Tree Structured Multi-hop Time Synchronization Approach
In this Section we proposed Tree Structured Referencing multi-hop Time Synchroni-
zation scheme, which is based on the protocol, proposed by [1], [2] that aims to
256 S. Rahamatkar et al.
minimize the complexity of the synchronization. Thus the needed synchronization
accuracy is assumed to be given as a constraint, and the target is to devise a synchro-
nization algorithm with minimal complexity to achieve given precision. The proposed
scheme works on two phases. First phase used to construct an ad hoc tree structure
and second phase used to synchronize the local clocks of sensor nodes followed by
network evaluation phase.
The goal of the TSRT is to achieve a network wide synchronization of the local
clocks of the participating nodes. We assume that each node has a local clock exhibit-
ing the typical timing errors of crystals and can communicate over an unreliable but
error corrected wireless link to its neighbors as mentioned in [6]. This scheme syn-
chronizes the time of a sender to possibly multiple receivers utilizing a single radio
message time-stamped at both the sender and the receiver sides. Linear regression is
used in this scheme to compensate for clock drift as suggested in [4].
2.1 Tree Structured Multi-hop Time Synchronization Phase
The proposed approach synchronizes the time of a sender with respect to multiple
receivers by utilizing a single synchronization message time-stamped at both sides
(sender and receiver). We assume that the root of the network is dynamically decided
through the election algorithm, which maintains the global time of the network and all
other nodes synchronize their clocks with reference to the root node. A logical tree
structure from a designated source point (root) will be formed using the algorithm
suggested at [2]. The nodes form an adhoc tree structure to transfer the global time
from the root to all the nodes.
2.2.1 Time Synchronization
Practically the network radius of WSN network is generally greater than one hop. In
the case of network-wide synchronization, the proposed multi-hop approach can be
applied; the detailed description is described in this section. We assume that every
node contains a unique identification.
Nodes in multi-hop synchronization, utilize reference points to achieve synchroni-
zation. A reference point contains a pair of global and local time stamps where both
of them refer to the same time instant. Periodically by sending and receiving the
synchronization message, reference points are generated. The transmissions of syn-
chronization message are either initiated by the root node of the logically constructed
adhoc tree of the network or any synchronized node in the network. The root node of
the network is dynamically decided through the election algorithm and with reference
to the root node; the whole network is being synchronized.
A node within the broadcast range of the root can collect reference points directly
from the root. Nodes outside the broadcast range of the root can gather reference
points indirectly through other synchronized nodes, are synchronized with reference
to the root. When a node collects enough consistent reference points, by estimating
the clock offset and skews of its own local clock and becomes synchronized. The
newly synchronized node can then broadcast synchronization messages to other nodes
in the network.
The proposed scheme is explained in more detail as follows: Periodically a Root
Node namely the reference node broadcasts the Synchronization message to all other
Tree Structured, Multi-hop Time Synchronization Approach 257
nodes of its transmission range. After receiving the consistent number of Synchroni-
zation messages, the node (say node-1 and node-2 in fig. 1(A)) generates the
reference points and by estimating the clock offset and skews of own local clock can
become synchronized. The nodes from beyond the range of Root Node may synchron-
ize by collecting the reference points from the nearest synchronized node from broad-
cast range. As per the fig. 3(B), Nodes 5,6,7,8 and 9 may collect the reference points
from the nodes 3 and 4. Nodes 3 and 4 are synchronized with reference to Root
Node and periodically broadcast the Synchronization message to all other nodes of
their transmission range. The nodes 5 & 6 may be synchronized with reference to the
node 3; similarly nodes 7 and 9 may be synchronized with reference to synchronized
node 4.
Fig. 1. (A) Reference node broadcasts & (B) Other layer nodes are synchronized
neighbours are synchronized
3 Comparison and Conclusion
The performance comparison of proposed scheme and TPSN in terms of the average
number of message exchanges M with respect to the number of beacons N when net-
work wide sync error probability Ps is assumed as 0.01% and 1% respectively. This
comparison is based on the linear network model where the depth of the network B =
5, εmax =10ms, σεo = 16.67µ, d = 10ms, t = 400ms, and σεs =1.58µ have assumed.
It can be observed that proposed scheme requires a less number of timing messages
than TPSN when there multiple numbers of beacon transmissions are required. More-
over, the gap of the average number of required timing messages between proposed
scheme and TPSN significantly increases as N increases, and thus proposed scheme is
by far more efficient than TPSN for large value of N. It can be also seen that a few
number of beacons is enough to minimize M. Besides, as expected, a larger number of
beacons required to meet a more strict constraint of the network-wide error probabili-
ty Ps. In practice, a lower number of N is highly preferable since it is proportional to
the synchronization time, i.e., a lower N induces better latency performance, although,
it may not be optimal in terms of energy consumption.
While the proposed approach is especially useful in WSN which are typically, ex-
tremely constrained on the available computational power, bandwidth and have some
of the most exotic needs for high precision synchronization [3]. The proposed syn-
chronization approach was designed to switch between TPSN and RBS. These two
algorithms allow all the sensors in a network to synchronize themselves within a few
microseconds of each other, while at the same time using the least amount of
258 S. Rahamatkar et al.
resources possible. In this work two varieties of the algorithm are presented and their
performance is verified theoretically and compared with existing protocols. The
comparison with RBS and TPSN shows that the proposed synchronization approach is
lightweight since the number of required broadcasting messages is constant in one
broadcasting domain.
References
1. Rahamatkar, S., Agarwal, A.: An Approach towards Lightweight, Reference Based, Tree
Structured Time Synchronization Scheme in WSN. In: COSIT, Part I. CCIS, vol. 131,
pp. 189–198. Springer, Berlin (2011)
2. Rahamatkar, S., Agarwal, A., Sharma, V.: Tree Structured Time Synchronization Protocol
in Wireless Sensor Network. J. Ubi. Comp. & Comm. 4, 712–717 (2009)
3. Rahamatkar, S., Agarwal, A., Kumar, N.: Analysis and Comparative Study of Clock Syn-
chronization Schemes in Wireless Sensor Networks. Int. J. Comp. Sc. & Engg. 2(3),
523–528 (2010)
4. Elson, J.E., Girod, L., Estrin, D.: Fine-Grained Network Time Synchronization using Reference
Broadcasts. In: 5th Symp. on Operating Systems Design and Implementation,
pp. 147–163 (2002)
5. Ganeriwal, S., Kumar, R., Srivastava, M.B.: Timing-Sync Protocol for Sensor Networks. In:
First ACM Conference on Embedded Networked Sensor System (SenSys), pp. 138–149
(2003)
6. Dai, H., Han, R.: TSync: a lightweight bidirectional time synchronization service for wire-
less sensor networks. SIGMOBILE Mob. Comp. Comm. Rev. 8(1), 125–139 (2004)
7. Sun, K., Ning, P., Wang, C.: TinySeRSync: Secure and Resilient time synchronization in
wireless sensor networks. In: 13 ACM Conf. on Comp. Comm. Security, pp. 264–277
(2006)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 259–262, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A New Markov Chain Based Cost Evaluation Metric for
Routing in MANETs
Abhinav Tiwari, Nisha Wadhawan, and Neeraj Kumar
School of Computer Science and Engineering
Shri Mata Vaishno Devi University, Katra (J&K), India
abhinavtiwari.1003@gmail.com, nishadhamija74@yahoomail.com,
nehra04@yahoo.co.in
Abstract. In recent time, routing remains a major concern in Mobile adhoc
networks (MANETs) due to the mobility of nodes. In particular, the link stabil-
ity remains a major concern in existing protocols. In this paper, we propose a
Markov Chain based Cost Evaluation Metric (MCCE) which is an extension of
existing WCETT metric with an enhancement of link stability. The proposed
routing metric is capable of providing better link stability and higher packet de-
livery fraction than the other proposed metrics in this category.
Keywords: Cost Evaluation function (CE), MCCE (Markov Chain Cost
Evaluation), Mobile Adhoc Networks, ETX, ETT, WCETT.
1 Introduction
A mobile Adhoc Network (MANET) is that network which does not have any central-
ized control. Each node in MANET can send data to any node within its transmission
range through a series of intermediate nodes, i.e., that every node can act as a sender
or a receiver or a router. These networks are used in areas where a fast and efficient
network deployment is required such as in military operations and in floods & earth-
quakes affected areas where the deployment of any wired or wireless network quickly
is not feasible[1,2]. Since in MANETs all the nodes are mobile, network topology
keeps on changing with respect to mobility of nodes in the network [3]. It is very im-
portant to route data to the destination through the most optimal route in order to im-
prove the performance of the network [4] [5].
Routing in MANETs is a very active area of research. Lot of efforts have been
made to improve the process of transferring data on the basis of various routing met-
rics which are used to select the best route for the data transfer between any
two nodes. Various metrics have been proposed in order to find high quality paths
taking into consideration topology changes which can be chosen depending upon the
type of environment which is required to be built. [1]
The remainder of this paper is organized as follows. Section 2 discusses the related
work, Section 3 describes the system architecture and proposed approach, Section 4 de-
scribes the simulation results and comparison, and finally section 5 concludes the article.
260 A. Tiwari, N. Wadhawan, and N. Kumar
2 Related Works
A number of MANETs routing protocols have been proposed [6, 7, 8] which are
based on the minimum hop-count as their routing metric. But hop-count does not pro-
vide a path with maximum throughput in many scenarios [9]. To remove the problems
associated with hop-count metric, a new metric called Expected Transmission Count
(ETX) was proposed [9]. WCETT [10] was an enhancement over the ETT metric
which contained all the features of ETT metric and it also included channel diversity
for the first time. WCETT along a path with n hops and k different channels is given
by the following formula
∑
=
∑
=≤≤
×+×−= j channel on is i hop i
ETT
j
XWhere
j
X
n
ikj
i
ETTWCET
T
,
1max
1
)1(
ββ
For 1 ≤ j ≤ k and 0 ≤ β ≤ 1. Finally, ETT is estimated as
B
S
ETXETT ×=
Where S is the packet size and B is the bandwidth of the link. This metric did not ac-
count for link stability in its calculation. Therefore, we propose an improvement over
WCETT metric called Markov Chain Based Cost Evaluation (MCCE) metric.
3 Proposed Approach
3.1 System Architecture
The system architecture in the proposed approach is shown in figure 1. It consists of
mobile devices and wireless connection. There is no centralized control as shown in
figure1.
Fig. 1. Scenario in mobile adhoc network
We propose a new metric called MCCE which introduces a new feature of
link stability in the selection of the path. The link stability in a network can be
defined as-
Link Stability= t
e
λ
−
−1 (1)
Where is the job arrival rate at any instant of time t. The Cost Evaluation (CE) func-
tion for a link between nodes i and j can be calculated as follows-
A New Markov Chain Based Cost Evaluation Metric for Routing in MANETs 261
StabilityLink Ji =j) , (i CE × (2)
Where Ji is the jitter value which is equal to some threshold value initially and then
later it is calculated as
|)
2-ttime
j) , (CE(i - )
1-ttime
j) , (CE(i| =Ji == (3)
Therefore the metric value for a link between nodes i and j in the path can be calcu-
lated as-
j) , (i CE WCETT= j) , (i MCCE × (4)
The probability of successful transmission of a link between any two nodes i and j in
a network of n number of nodes can be represented using a Markov Chain Probability
Matrix (MCPM) as:
(5)
Pseudo Code
Input: - , WCETT value of the link (i, j), N. Output: - MCCE (i,j)
Repeat
Calculate the link stability in the network using equation (1).
Calculate jitter in the link (i,j) using last two values of CE from equation (3).
Calculate the value of CE (i,j) at time t using the equation (2).
The metric value MCCE (i,j) for the link (i,j) is calculated from equation (4).
While )(
φ
≠N
The Markov Probability Matrix is calculated using the equations (5) and (6).
4 Simulation Results and Comparison
The performance of proposed algorithms is evaluated using ns-2 [11] with respect to
packet delivery fraction and average throughput. Figure 2 shows the impact of the
proposed scheme on existing WCETT based scheme. The results obtained show
Number of nodes
0 20 40 60 80 100 120
Throughput (Mbps)
4
6
8
10
12
14
16
18
20
WCETT based
Proposed
Number of nodes
0 20 406080100
Packet delivery fraction
0.0
0.2
0.4
0.6
0.8
1.0
1.2
Proposed
WCETT
(a) (b)
Fig. 2(a)-(b). Throughput and packet delivery fraction in proposed and existing schemes
][ j
i
aMCPM =
∑
=
=n
j
jiMCCE
ji
M
CC
E
ij
awhere
1
),(
),(
,,
262 A. Tiwari, N. Wadhawan, and N. Kumar
Table 1. Relative comparison of routing metrics
Routing Metrics State Inde-
pendent Link
stability
Load
Balancing Routing Decision
Hop Count [1999] No No Per hop basis
ETX [2003] No No Number of retransmission
ETT [2004] No No Transmission time
WCETT [2004] No No Transmission time
MCCE [2010] Yes Yes Link stability
that the proposed scheme has higher throughput and packet delivery fraction than
WCETT scheme with an increase in number of nodes.
5 Conclusions
MCCE metric can enhance the functionality of WCETT metric by introducing link sta-
bility in the calculation of optimal paths. The metric is very useful for applications in real
time applications where the nodes are highly mobile and it is difficult to estimate the
exact position of a node. Moreover, the proposed metric determines more reliable paths
with less packet loss and delivery time. The results obtained also show that the proposed
scheme is quite effective with respect to the existing scheme in terms of packet delivery
fraction and throughput.
References
1. Safwat, A.M., Hassanein, H.S.: Infrastructure-based routing in wireless mobile ad hoc net-
works. Computer Communications 25(3), 210–224 (2002)
2. IETF MANET Working Group Charter,
http://www.ietf.org/html.charters/manetcharter.html
3. Raimundo, J., Macêdo, A., Assis Silva, F.M.: The mobile groups approach for the coordi-
nation of mobile agents. J. Parallel Distributed Computing 65, 275–288 (2005)
4. Das, S.R., Perkins, C.E., Royer, E.M.: Performance Comparison of Two On demand Rout-
ing Protocols for Ad Hoc Networks. In: IEEE INFOCOM (March 2000)
5. Spohn, M., Garcia-Luna-Aceves, J.J.: Neighborhood aware source routing. In: Proceedings
of ACM MobiHoc, 6th IEEE/ ACIS International Conference on Computer and
Information Science, LongBeach, CA (2001)
6. Johnson, D.B., Maltz, D.A.: Dynamic source routing in ad hoc Wireless networks. Mobile
Computing, 153–181 (1996)
7. Perkins, C., Belding-Royer, E., Das, S.: Ad hoc On-Demand Distance Vector (AODV)
Routing. In: IETF.RFC, vol. 3561 (2003), http://tools.ietf.org/html/rfc3561
8. Tønnesen, A.: Implementing and extending the Optimized Link State Routing Protocol.
UniK University Graduate Center, University of Oslo (2004)
9. Couto, D.S.J.D., Aguayo, D., Bicket, J., Morris, R.: A high throughput path metric for
multi-hop wireless routing. In: Proc. MOBICOM (2003)
10. Draves, R., Padhye, J., Zill, B.: Routing in multi-radio, multi-hop wireless mesh network.
In: Proc. MOBICOM (2004)
11. Fall, K., Varadhan, K. (eds.): NS notes and documentation. The VINT project, LBL
(February 2000), http://www.isi.edu/nsnam/ns/S
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 263–266, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Texture Image Classification Using Gray Level Weight
Matrix (GLWM)
R.S. Sabeenian1 and P.M. Dinesh2
1 Professor and Centre Head Sona SIPRO, 2 Research Associate, Sona SIPRO,
ECE – Department, Sona College of Technology, Salem, India
sabeenian@sonatech.ac.in, pmdineshece@live.com
Abstract. The texture analysis plays an important role in image processing and
image classification field. Texture is an important spatial feature useful for
identifying objects in an image. The local binary pattern and entropy are the
most popular statistical methods used in practice to measure the textural infor-
mation of images. Here, we proposed new statistical approach for the classifica-
tion of texture images. In this method, the local texture information for a given
pixel and its neighborhood is characterized by the corresponding texture unit
and the global textural aspect of an image is revealed by its texture spectrum.
The proposed method extracts the textural information of an image with a more
complete respect of texture characteristics.
Keywords: Textures Spectrum, Local binary pattern, Images, Statistical me-
thods, Spatial Features, Gray Level Weight Matrix.
1 Introduction
In remote sensing data with a high spatial resolution, some of the image elements are
represented by a group of pixels, not by only one pixel. This means that image classifi-
cation and interpretation based on the analysis of individual pixels will result in a rela-
tively high rate of classification and will no longer be sufficient to satisfy the needs. A
good understanding or a more satisfactory interpretation of remotely sensed imagery
should include descriptions of both the spectral and textural aspects.
The purpose of this study is to present a new statistical method of texture analysis
[2], which is focused on texture characterization and discrimination. The concept of
texture unit is proposed first. It may be considered, as the smallest complete unit,
which best characterizes the local texture aspect of a given pixel and its neighborhood
in all eight directions of a square raster. Then a texture image is characterized by its
features like local binary pattern, texture spectrum and entropy, which describe the
distribution of all the texture units within the image. Some natural images have been
used to evaluate the discriminating performance of the texture spectrum.
264 R.S. Sabeenian and P.M. Dinesh
2 Previous Method
2.1 Local Binary Pattern (LBP)
LBP is invariant against any monotonic gray scale transformation. The method is rota-
tion variant like most existing texture measures. LBP does not address the contrast of
texture, which is important in the discrimination of some textures. For this purpose, we
can combine LBP with a simple contrast measure, and consider joint occurrences of
LBP. LBP and LBP/Contrast perform well also for small image regions (e.g., 16x16
pixels), which is very important. A LBP is called uniform if the binary pattern contains
at most two bitwise transitions from 0 to 1 or vice-versa.
2.2 Entropy
If an image is interpreted as a sample of a “gray level source” that emitted it, we can
model that source’s symbol probabilities using the gray-level histogram of the ob-
served image and generate an estimate, called the first-order estimate, H, of the source
entropy
∑ log
where k =1, 2 ,..L
2.3 Texture Spectrum
The previously defined set of 6561 texture units describes the local-texture aspect of a
given pixel. Thus, the statistics of the frequency of occurrence of all the texture units
over a large region of an image should reveal texture information. We termed the
texture spectrum the frequency distribution of all the texture units, with the abscissa
indicating the texture unit number NTU and the ordinate representing its occurrence
frequency. Where is the texture unit number, is the ith element of texture
unit set
3
3 Proposed Method
In LBP the signs of the eight differences are recorded into an 8-bit number. The origi-
nal 3x3 neighborhood is thresholded by the value of the center pixel [4].
In proposed method instead of thresholding the image transforming is made with
neighborhood to a texture unit with the texture unit number under the ordering way as
shown in figure1.
Texture Image C
The transforming condit
i
E
i
= 0 if V
i
<
V
1 if V
i
=
V
2 if V
i
>
V
Neighborhood (
V
62 85 9
2
29 40 3
6
67 36 6
6
The values of the pixels
by the weights given to the
are summed to obtain a nu
m
ter pixel value till end of t
h
the calculation of GLWM i
s
4 Results and Discu
s
In order to evaluate the per
f
and classification, several e
images [1]. These images
w
er and also that they resem
b
Table
1
Method LBP TSO En
t
Image1 94.41 91.25 7
8
Image2 96.6 86.15 8
1
Image3 85.82 87.82 8
2
Image4 93.25 93.5 8
8
Image5 95.37 92.5 9
6
lassification Using Gray Level Weight Matrix (GLWM)
i
ons
V
o
V
o
V
o
Where:
V
i
= The center pixel value
V
o
= The neighboring pixel value
V
i
) Texture unit (E
i
)
2
6
6
2 2 2
0 1 0
2 0 2
Fig. 1.
Texture unit transformation
in the transformed texture unit neighborhood are multip
corresponding pixels. Finally, the values of the eight pi
x
m
ber for this neighborhood. This method considers the
c
h
e process for each 3 x 3 matrix. The equation involve
d
s
shown below.
∑
2
s
sion
f
ormance of the texture spectrum in texture characteriza
t
xperimental studies have been carried out on 9 of Brod
a
w
ere selected because they are broadly similar to one an
o
b
le parts of remotely sensed images.
Fig. 2.
Nine Brodatz's text images
1
.
Analysis of Texture Image Classification
t
ropy GLWM Method LBP TSO Entropy GL
W
8
.09 95.63 Image6 84.24 82.5 87.5 89.
2
1
.57 94.32 Image7 82.67 93.5 93.6 94.
3
2
.68 91.35 Image8 93.51 96.8 76.55 94.
6
8
.55 93.64 Image9 80.46 87.8 86.5 88.
6
6
.52 95.37
265
lied
x
els
c
en-
d
in
t
ion
a
tz's
ot
h-
W
M
2
5
3
2
6
4
6
5
266 R.S. Sabeenian and P.M. Dinesh
We evaluated local binary pattern, texture spectrum, entropy and GLWM. The re-
sults obtained are tabulated. According to tabulation results the GLWM technique has
the best one. The GLWM technique is calculated with overlapping of each 3 x 3 of the
image hence the result of this technique is good when compared to other.
We extended our database to 60 image of the Brodatz’s texture database [1]. The
classification rate of image at different categories are tabulated below
Table 2. Classification percentage for 60 images of database
Image Size No of Sample Classification
Rate
512 X 512 60 X 1 100%
256 X 256 60 X 4 98%
128 X 128 60 X 16 95%
64 X 64 60 X 64 94%
Total Classification Rate 96.75 %
5 Conclusion
Based on the concept of GLWM, Local Binary Pattern, texture spectrum operator and
entropy measure, texture analysis has been presented. We used 60 images from the
Brodatz’s database. From them we took 5100 samples 512 X 512 size of 60 samples,
256 X 256 size of 240 samples, 128 X 128 size of 960 samples, 64 X 64 size of 3840
samples. Totally 5100 samples among them GLWM able to classify 4935 samples
successfully. Evaluations show that the GLWM is able to reveal texture information
in digital images.
References
[1] Lin, H., Wang, J.-y., Liu, S.: The Classification Study of Texture Image Based on the
Rough Set Theory. In: 2010 IEEE International Conference on Proceeding of the IEEE
xplore Granular Computing (GrC), August 14-16 (2010)
[2] Khelifi, R., Adel, M., Bourennane, S.: Texture classification for multi-spectral images us-
ing spatial and spectral Gray Level Differences. In: 2010 2nd International Conference on
Proceeding of the IEEE xplore Image Processing Theory Tools and Applications (IPTA),
July 7-10 (2010)
[3] Sabeenian, R.S., Palanisamy, V.: Texture Based Medical Image Classification of Com-
puted Tomography images using MRCSF. Published in International Journal of Medical
Engineering and Informatics (IJMEI) Published by Inderscience Publications 1(4),
459–472 (2009)
[4] Sabeenian, R.S., Palanisamy, V.: Crop and Weed Discrimination in Agricultural Field us-
ing MRCSF Published. In the International Journal of Signal and Imaging Systems Engi-
neering (IJSISE) Published by Inderscience Publishers 3(1), 61–69
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 267–272, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Formal Verification of IEEE802.11i WPA-GPG
Authentication Protocol
K.V. Krishnam Raju1 and V. Valli Kumari2
1 Dept. of Computer Science Engineering, SRKR Engineering College, Bhimavaram,
Andhra Pradesh, India
kvkraju.srkr@gmail.com
2 Dept. of Computer Science & Systems Engineering, Andhra University,
Visakhapatnam, Andhra Pradesh, India, 5300 03
vallikumari@gmail.com
Abstract. IEEE802.11i is the standard designed to provide secured communi-
cation of wireless LAN. The IEEE802.11i specification contains WPA-GPG au-
thentication protocol. It allows a wireless station to gain access to a protected
wireless network managed by access point. This paper models the WPA-GPG
authentication protocol by formal verification using CasperFDR and analyzes
the output. A few attacks are found in this protocol. The specifications through
which these attacks are found are presented.
Keywords: Formal Verification, WPA-GPG, CasperFDR.
1 Introduction
WPA protocol is a subset of IEEE 802.11i standard. The IEEE 802.11i standard is
designed to provide secured communication of wireless LAN as defined by all the
IEEE 802.11 specifications. IEEE 802.11i enhances the WEP (Wire line Equivalent
Privacy), a technology used for many years for the WLAN security, in the areas of
encryption, authentication and key management. IEEE 802.11i is based on the Wi-Fi
Protected Access (WPA), which is a quick fix of the WEP weaknesses [1].
The IEEE 802.11i standard includes authentication, encryption and data integrity
check of three components, is a complete security program. The core is 802.1x (Port
Access Control) and TKIP (Temporal Key Integrity Protocol). The IEEE 802.1X
offers an effective framework for authenticating and controlling user traffic to a
protected network, as well as dynamically varying encryption keys. 802.1X ties a
protocol called EAP (Extensible Authentication Protocol) to both the wired and wire-
less LAN media and supports multiple authentication methods, such as token cards,
Kerberos, one-time passwords, certificates, and public key authentication.
WPA-GPG is a modification on WPA-PSK authentication protocol that allows a
wireless station (also known as supplicant or STA) to gain access to a protected wire-
less network, managed by an Access Point (also known as AP or authenticator), by
means of its personal GPG key [2]. WPA-GPG verifies that the authenticating STA
is the owner of the key that it is providing to the AP, this represents the core of the
268 K.V. Krishnam Raju and V. Valli Kumari
authentication process. WPA-GPG does not require specific hardware nor firmware
since, its modifications to the original protocol are important but not radical.
WPA-GPG ensures the same level of security of WPA-PSK but it also guarantees in-
network user privacy and non-repudiation of network traffic. With the term in-
network user privacy we mean that network traffic is encrypted in a way that even
STAs which have the right to authenticate are not able to decrypt it. As we will show
this is not possible in WPA-PSK. As a direct consequence WPA-GPG also allows the
AP to know exactly which STA has generated what traffic, ensuring non-repudiation.
The goal of WPA-PSK four-way handshake protocol is to create a PTK known
both to supplicant and authenticator while not revealing the PMK. WPA-GPG has
exactly the same goal, with the difference that no PSK is shared between STA and AP
but STA uses its GPG key to authenticate. In WPA-GPG the PMK is randomly gener-
ated by the AP, encrypted and sent to the authenticating STA. Both AP and STA will
therefore be able to derive a PTK from the PMK, exactly as it happens with WPA-
PSK, with the main difference that each PTK cannot be derived by other STAs.
Modeling and analysis of security protocols with Communicating Sequential
Process (CSP) and Failure Divergence Refinement (FDR) have been proven to be
effective and have helped the research community find attacks in several protocols.
However, modeling directly in CSP is time-consuming and error-prone. Lowe thus
designed Casper [3], which takes more abstract descriptions of protocols as input and
translates them into CSP. CSP was first described by Hoare in [4] [5], and has been
applied in many fields.
First, we formally model and analyze the WPA-PSK protocol with CasperFDR.
Next, we use CasperFDR to show that there are no other known attacks on WPA-PSK
protocol.
The rest of the paper is organized as follows. Section 2 deals with related works. In
Section 3, WPA-GPG protocol is modeled with CasperFDR and is analyzed and fi-
nally we conclude in Section 4.
2 Related Works
Johnson and Walker are among the first researchers who discussed the security issues
in IEEE 802.16 [6]. Sen Xu, Chin-Tser Huang, Manton M. Matthews [7], have
analyzed security issues on the PKMv1, PKMv2 protocols using Casper FDR and
proposed solutions. K.V.Krishnam Raju, V.Valli Kumari, N.Sandeep Varma,
KVSVN Raju [8], also have analyzed security issues on the PKMv3 protocol using
Casper FDR.
According to [9] Wi-Fi Protected access (WPA) is really less secure. Generally
wireless protection is enabled using MAC filtering. According to survey there is no
clear insight into this protocol by the research community. But much work has been
done on other wireless protocols like PKMV1, PKMV2, PKMV3 that belong to
802.16 standards [10]. On the other hand in this paper we have done verification
process for WPA-GPG authentication protocol using some standard verification tool
like CasperFDR.
Formal Verification of IEEE802.11i WPA-GPG Authentication Protocol 269
3 Modeling and Analysis of WPA-GPG Authentication Protocol
3.1 WPA-GPG Authentication Protocol Structure
Message 1: The first message is sent by the AP and contains the AP random nonce
(ANonce). This message is not encrypted and no MIC is attached (the PTK cannot be
derived yet).
Message 2: After receiving the first message, the STA will send its nonce (SNonce)
and attaches the STA public key (GPG Key) for authentication, and attaches the sig-
nature of ANonce and received RSN IE(Information elements) .The signed RSN IE is
added to allow AP to verify that the STA has received a correct element and to pre-
vent version rollback attack.
Message 3: After receiving the second message, the AP will verify the validity of
STA signature on ANonce, using the key attached to message 2. The AP will gener-
ates a random 32 byte PMK which is kept secret and derives the PTK. Attaches IE
and GTK to the message and computes message MIC. Encrypt the generated PMK
using STA public key.
Message 4: After receiving the third message, the STA will decrypt the PMK with its
public key and derives the PTK and verifies the MIC value of message 3. If MIC code
verification is successful STA can send the acknowledgment which contains nothing
but a MIC and AP can authorize it to access the network.
Fig. 1. WPA-GPG Authentication Protocol [2]
3.2 Modeling WPA-GPG Authentication Protocol in CasperFDR
The modeled WPA-GPG authentication protocol in CasperFDR is shown in Fig.2.
In the specification the initiator A and responder S represents Authenticator and
Supplicant.
3.3 Analysis of WAP-GPG Authentication Protocol with CasperFDR
After compiling and checking the above model in CasperFDR tool, attacks were
found for two of the four properties declared in specification part in Fig.2. Out of four
property1 and property2 are related to secret specifications next property3 and prop-
erty4 are related to authentication specifications. CasperFDR tool found attacks on
every specification in the specification part.
Messa
g
e1 AP ÆSTA : N
a
Message2 STA Æ AP : Ns | GPG KEY | Sign(Na + IE)
Mesgesa3 AP Æ STA : GTK | MIC | Cry(PMK)
Message4 STA Æ AP : ACK | MIC
270 K.V. Krishnam Raju and V. Valli Kumari
Fig. 2. WPA-GPG key agreement Protocol Specification
Fig. 3. Attack generated by CasperFDR for property1
#Free variables
A, S : Agent
na, ns: Nonce
kpm, kpt, gtk: SessionKey
gpg : Agent ÆPublicKey
sk : AgentÆ SecretKey
f: HashFunction
InverseKeys = (gpg,sk), (kpm,kpm), (kpt,kpt), (gtk,gtk)
#Processes
INITIATOR(A,na,kpt,kpm,gtk) knows gpg, sk(A)
RESPONDER(S,ns,kpt,kpm,gtk) knows gpg, sk(S)
#Protocol description
0. Æ A : S
[A!=S]
1. AÆS : na
2. SÆA : ns, gpg(S), {na}{sk(S)}
3. A ÆS: gtk, {kpm}{gpg(S)}
4. SÆA : f(kpm)
#Specification
Secret(A, na, [S])
Secret(S, ns, [A])
Agreement(A, S, [na,ns])
Agreement(S, A, [na,ns])
#Actual variables
Authenticator, Supplicant, Mallory : Agent
Na, Ns, Nm : Nonce
Kpm, Kpt, Gtk : SessionKey
InverseKeys = (Kpm, Kpm), (Kpt, Kpt), (Gtk, Gtk)
#Inline functions
symbolic gpg,sk
#System
INITIATOR(Authenticator, Na, Kpt, Kpm, Gtk)
RESPONDER(Supplicant, Ns, Kpt, Kpm, Gtk)
#Intruder Information
Intruder = Mallory
IntruderKnowledge={Authenticator,Supplicant,Mallory,Nm,gpg,sk(Mallory)}
0. Æ Authenticator : Mallory
1. Authenticator Æ I_Mallory : Na
1. I_Supplicant Æ Supplicant : Nm
2. Supplicant Æ I_Supplicant : Ns, gpg(Supplicant),{Nm}{sk(Supplicant)}
2. I_Mallory Æ Authenticator : Nm, gpg(Mallory), {Na}{sk(Mallory)}
3. Authenticator Æ I_Mallory : Gtk, {Kpm}{gpg(Mallory)}
3. I_Supplicant Æ Supplicant : Gtk, {Kpm}{gpg(Supplicant)}
4. Supplicant Æ I_Supplicant : f(Kpm)
Formal Verification of IEEE802.11i WPA-GPG Authentication Protocol 271
An attack on the property1 Secret(A, na, [S]) as shown in Fig.3. It represents the
Supplicant believes Ns is secret but the intruder also knows Ns value.
An attack on the property2 Secret(S, na, [A]) as shown in Fig.4.It also represents
the Supplicant believes Ns is secret but the intruder also knows Ns value.
Fig. 4. Attack generated by CasperFDR for property2
An attack on the property3 Agreement(A, S, [na,ns]) as shown in Fig.5. Top level
trace generated by CasperFDR is Supplicant believes that it has completed a run of
the protocol, taking role RESPONDER, with Supplicant, using data items Nm, Ns.
Authenticator also believes that it is running the protocol.
Fig. 5. Attack generated by CasperFDR for property3
Fig. 6. Attack generated by CasperFDR for property4
An attack on the property4 Agreement(S, A, [na,ns]) as shown in Fig.6. Top level
trace generated by CasperFDR is Authenticator believes (s)he has completed a run of
0. Æ Authenticator : Mallory
1. Authenticator Æ I_Mallory : Na
1. I_Supplicant Æ Supplicant : Nm
2. Supplicant Æ I_Supplicant : Ns, gpg(Supplicant), {Nm}{sk(Supplicant)}
2. I_Mallory Æ Authenticator : Nm, gpg(Mallory), {Na}{sk(Mallory)}
3. Authenticator Æ I_Mallory : Gtk, {Kpm}{gpg(Mallory)}
3. I_Supplicant Æ Supplicant : Gtk, {Kpm}{gpg(Supplicant)}
4. Supplicant Æ I_Supplicant : f(Kpm)
0. Æ Authenticator : Supplicant
1. AuthenticatorÆ I_Supplicant : Na
1. I_Mallory Æ Supplicant : Na
2. Supplicant Æ I_Mallory : Ns, gpg(Supplicant), {Na}{sk(Supplicant)}
2. I_Supplicant Æ Authenticator : Nm, gpg(Supplicant), {Na}{sk(Supplicant)}
3. AuthenticatorÆ I_Supplicant : Gtk, {Kpm}{gpg(Supplicant)}
3. I_Mallory Æ Supplicant : Gtk, {Kpm}{gpg(Supplicant)}
4. Supplicant Æ I_Mallory : f(Kpm)
4. I_Supplicant Æ Authenticator : f(Kpm)
0. Æ Authenticator : Mallory
1. I_Supplicant Æ Supplicant : Nm
1. Authenticator Æ I_Mallory : Na
2. Supplicant Æ I_Supplicant : Ns, gpg(Supplicant),{Nm}{sk(Supplicant)}
2. I_Mallory Æ Authenticator : Nm, gpg(Mallory), {Na}{sk(Mallory)}
3. Authenticator Æ I_Mallory : Gtk, {Kpm}{gpg(Mallory)}
3. I_Supplicant Æ Supplicant : Gtk, {Kpm}{gpg(Supplicant)}
4. Supplicant Æ I_Supplicant : f(Kpm)
272 K.V. Krishnam Raju and V. Valli Kumari
the protocol, taking role INITIATOR, with Supplicant, using data items Na, Nm.
Supplicant also believes that it is running the protocol.
From the above attacks we can find that not only the Authenticator and the Suppli-
cant but the Intruder can also derive GTK key from PMK that is the key used between
Authenticator and Supplicant for message encryption.
4 Conclusions and Future Work
In this paper, the WPA-GPG protocol is modeled using CasperFDR. The compilation
was done with CasperFDR. Attacks were found in this protocol. The attacks that are
interpreted by CasperFDR and the message sequence results are reported. In future we
will fix the attacks found in the WPA-GPG protocol.
References
[1] IEEE 802.11i. WLAN Security Standards,
http://www.javvin.com/protocol80211i.html
[2] WPA-GPG. Wireless authentication using GPG Key, Gabriele Monti, December 9 (2009)
[3] Lowe, G.: Casper: A compiler for the analysis of security protocols. Journal of Computer
Security 6, 53–84 (1998)
[4] Hoare, C.A.R.: Communicating sequential processes. Commun. ACM 21(8), 666–677
(1978)
[5] Hoare, C.A.R. (ed.): Communicating Sequential Processes. Prentice Hall International
(1985)
[6] Johnston, D., Walker, J.: Overview of IEEE 802.16 security. IEEE Security & Privacy
(2004)
[7] Xu, S., Matthews, M.M., Huang, C.-T.: Modeling and Analysis of IEEE 802.16 PKM
Protocols using CasperFDR. In: IEEE ISWCS 2008 (2008)
[8] Krishnam Raju, K.V., Valli kumari, V., Sandeep varma, N., Raju, K.V.S.V.N: Formal
Verification of IEEE802.16m PKMv3 Protocol Using CasperFDR. In: International Con-
ference on advances in information and communication technologies, ICT 2010, Kerala,
India (September 2010)
[9] Wpa-Wpa2,
http://www.lylebackemort.com/blog/2008/5/10/
wpa-wpa2-as-insecure-as-i-expected (accessed on March 14, 2010)
[10] Xu, S., Huang, C.-T., Matthews, M.M.: Modeling and Analysis of IEEE 802.16 PKM
Protocol using CasperFDR. University of South California, Columbia SC, 29208
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 273–276, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Adaptive Steganography Based on Covariance and Dct
N. Sathisha1, Swetha Sreedharan2, R. Ujwal2, Kiran D’sa2, Aneeshwar R. Danda2,
K. Suresh Babu3, K.B. Raja3, K.R.Venugopal3, and L.M. Patnaik4
1 Department of Electronics and Communication Engineering.
R L Jalappa Institute of Technology, Doddaballapura, Bangalore Rural Dist. 561 203, India
nsathisha@gmail.com
2 Department of Electronics and Communication Engineering, BMSIT, Bangalore, India
swethasree2@yahoo.com
3 Department of Computer Science and Engineering,
University Visvesvaraya College of Engineering, Bangalore University, Bangalore 560 001,
4 Defence Institute of Advanced Technologies, Pune, India
Abstract. The steganography is a covert communication to transfer confidential
information over an internet. In this paper we propose Adaptive Steganography
based on Covariance and Discrete Cosine Transform (ASCDCT) algorithm.
The Average Covariance of the Cover Image (ACCI) is computed. The ACCI
of 0.15 is considered as the threshold value. The cover image is segmented into
8*8 cells and DCT is applied to derive coefficients. The payload Most Signifi-
cant Bits (MSBs) are embedded into the cover image based on ACCI and DCT
coefficients. It is observed that the capacity, Peak Signal to Noise Ratio (PSNR)
and security is better compared to the existing algorithm.
Keywords: Covariance, Cover Image, DCT, Payload, Steganography.
1 Introduction
Steganography is an art and science of hiding confidential information into a cover
media to ensure the security of information over the communication channel. The
cover media can be text, audio, video and image. Weiqi Luo et al., [1] embeds the
secret message into sharper edge regions of cover image adaptively according to size
of the message and the gradients of the content edges of cover image. Cheng-Hsing
Yang et al., [2] developed a technique to embed the secret information by Pixel Value
Differing (PVD) method. The number of secret bits embedded depends on the differ-
ence between two consecutive pixels. Bo-Luen Lai and Long-Wen Chang [3] pro-
posed a transform domain based adaptive data hiding method using haar discrete
wavelet transform. The image was divided into sub-bands (LL1, HL1, LH1 and HH1)
and most of the data is hidden in the edge region as it is insensitive to the Human eye.
If these sub-bands were complex, then further division of the bands were performed
so that more number of data bits could be embedded. Raja et al., [4] proposed a high
capacity, secure steganographic algorithm in which the payload bits are encrypted and
embedded in the wavelet coefficients of the cover image. This method utilizes the
274 N. Sathisha et al.
approximation band of the wavelet domain to improve robustness. Carlos velsco
et al., [5] presented an adaptive data hiding method using convolutional codes and
synchronization bits in DCT domain. The cover image is divided into suitable and
ineligible blocks based on the DCT energy features from the horizontal, vertical and
diagonal frequency information. The suitable blocks are used for embedding data us-
ing Quantization Index Modulation (QIM). The two synchronization bits are used for
desynchronization problem and convolution codes are used for decoding errors.
2 Algorithm
Problem definition: The algorithm is given in the Table 1 and objectives are (i) The
payload is to be embedded into the cover image to derive stegoimage (ii) The high
capacity and security with reasonable PSNR.
Table 1. Algorithm of ASCDCT
Input: Cover Image and Payload; Output: Stego Image.
Step 1) A cover image of any size and format is considered and if it is color ima
g
e convert it
into grayscale image.
Step 2) Applying
p
ixel mana
g
ement to the cover ima
g
e to avoid overflow and underflow of
the pixel values 0 and 255.
Step 3) Covariance of cover ima
g
e is determined and avera
g
e is com
p
uted to
g
et avera
g
e
covariance.
Step 4) The average covariance of cover image value is fixed as 0.15, if ACCI
> 0.15 go to step 5 else step 6
Step 5)
(i) The cover image is segmented into 8*8 matrix and DCT is applied on each matrix.
(ii) Payload bit length L to be embedded based on DCT coefficients:
L=4, if Co ≥ 25; L=3, if 24 ≤ Co ≤ 25; L=2, if 23 ≤ Co ≤ 24;
else L=1;
(iii) The stego image obtained in the DCT domain is converted back to the s
p
atial
domain using IDCT.
Step 6)
(i) The cover image is segmented into 8*8 matrix and DCT is applied on each matrix.
(ii) Payload bit length L to be embedded based on DCT coefficients:
L=5, if Co ≥ 25; L=4, if 24 ≤ Co ≤ 25; L=3, if 23 ≤ Co ≤ 24;
else L=2;
(iii) The stego image obtained in the DCT domain is converted back into the spatial do-
main using IDCT.
3 Performance Analysis
The payload is embedded into the DCT coefficients of cover image based on ACCI
and length L. The performance parameter such as PSNR between cover image and
stegoimage is computed and given in the Table 2. It is observed that the PSNR de-
pends on the ACCI of the cover image and also the PSNR decreases as the Hiding
Adaptive Steganography Based on Covariance and Dct 275
Table 2. ACCI and PSNR for different cover images with payload boat
H C
12.5%
H C
20.0%
H C
25.0%
HC
34.0%
Cover Image ACCI
PSNR PSNR PSNR PSNR
Lena 0.064 46.31 41.51 40.51 39.41
Old image 0.122 45.41 41.95 40.85 39.09
Baboon 0.154 47.25 42.64 41.88 40.71
Barbara 0.203 47.08 43.48 42.54 41.17
Ranch House 0.280 46.06 41.50 40.46 39.20
Bridge 0.358 46.13 41.78 40.96 39.69
Casa 0.504 47.56 43.33 42.57 41.35
Capacity (HC) increases. The PSNR value is maintained around 42 dB for the
capacity of 34%.
The Maximum Hiding Capacity (MHC) and the PSNR between the Cover Image
(CI) and stego image for different Payload (PL) is tabulated for existing algorithm An
Adaptive Steganographic Technique Based on Integer Wavelet Transform (ASIWT)
[6] and the proposed algorithm ASCDCT is given in the Table 3. It is observed that
the PSNR is improved in the proposed algorithm compared to the existing algorithm
is due to ACCI.
Table 3. PSNR of existing and proposed techniques for a MHC of 47%
Existing Method
(ASIWT)
Proposed Method
(ASCDCT)
Image
PSNR PSNR
CI: Lena; PL: Barbara 31.80 39.35
CI:Baboon; PL: Cameraman 30.89 37.96
4 Conclusions
The steganography is used to transfer secret message over open channel. In this paper
ASCDCT is proposed. The cover image covariance is computed to consider number
of MSBs of payload to be embedded based on DCT coefficients. The cover image is
divided into 8*8 cells and converted into DCT coefficients to determine the length of
the payload bits to be embedded into the cover image. It is observed that the capacity,
security and the PSNR values are improved compared to the existing algorithm. In
future the robustness of algorithm can be verified.
References
1. Luo, W., Huang, F., Huang, J.: Edge Adaptive Image Steganography Based on LSB Matching
Revisited. IEEE Transactions on Information Forensics and Security 5, 173–178 (2010)
2. Yang, C.-H., Weng, C.-Y., Wang, S.-J., Sun, H.-M.: Adaptive Data Hiding in Edge Areas of
Images with Spatial LSB Domain Sysytems. IEEE Transactions on Information Forensics and
Security 3, 488–497 (2008)
276 N. Sathisha et al.
3. Lai, B.-L., Chang, L.-W.: Adaptive Data Hiding for Images Based on Harr Discrete Wave-
let Transform. In: Chang, L.-W., Lie, W.-N. (eds.) PSIVT 2006. LNCS, vol. 4319,
pp. 1085–1093. Springer, Heidelberg (2006)
4. Raja, K.B., Vikas, Venugopal, K.R., Patnaik, L.M.: High Capacity Lossless Secure Image
Steganography using Wavelets. In: International Conference on Advances Computing and
Communications, pp. 230–235 (2006)
5. Velasco, C., Nano, M., Perez, H., Martinez, R., Yamaguchik: Adaptive JPEG Steganogra-
phy using Convolutional Codes and Synchronization Bits in DCT Domain. In: 52nd IEEE
International Midwest Symposium on Circuits and Systems, pp. 842–847 (2010)
6. El Safy, R.O., Zayed, H.H., El Dessouki, A.: An Adaptive Steganographic Technique Based
on Integer Wavelet Transform. In: IEEE Proceedings on International Conference on Net-
works and Media, pp. 111–117 (2009)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 277–280, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Image Segmentation Using Grey Scale Weighted Average
Method and Type-2 Fuzzy Logic Systems
Saikat Maity1 and Jaya Sil2
1 Dr. B.C. Roy Engineering College
Tel.: +919932252306
saikat.maity@bcrec.org
2 Bengal Engineering & Science University (B.E.S.U)
Tel.: +919433283641
js@cs.becs.ac.in
Abstract. In the paper, the difficulty in image segmentation based on the popular
level set framework to handle an arbitrary number of regions has been addressed.
There is very few work reported on optimized segmentation with respect to the
number of regions. In the proposed model, first the image is classified using
type-2 fuzzy logic to handle uncertainty in determining pixels in different color
regions. Grey scale average (GSA) method has been applied for finding accurate
edge map to segment the image that produces variable number of regions.
Keywords: Segmentation, Grey Scale Average (GSA), Type-2 fuzzy logic,
Fuzzy weighted average (FWA).
1 Introduction
Edge of an image represents the sudden change in the pixel intensity value creating
two separate regions of different intensities. However, other factors like poor focus or
refraction may be the cause of formation of an edge in the image [1]. Relatively early,
the problem of image segmentation has been formalized by Health [5] as the minimi-
zation of an energy function that penalizes deviations from smoothness within regions
and the length of their boundaries [4]. In the level set approach [2], N regions are
represented using exactly log2N level set functions. Grey scale average (GSA) method
has been proposed here for finding accurate edge map to segment the image that pro-
duces variable number of regions [3], described in figure-1. The paper has been di-
vided into five sections. Section 2 describes type-2 fuzzy logic to classify the color
regions while in section 3, fuzzificaion and defuzzification method in image segmen-
tation and how to determine the grey scale average has been discussed. The proposed
algorithm with masking and GSA method is presented in section 4 followed by ex-
perimental results and discussions in section 5.
Fig. 1. Complete Block Diagram of the System
278 S. Maity and J. Sil
2 Classification of Image
Fuzzy Gaussian filter [6] is used to remove noise and sharpen the image, a preproc-
essing step of edge detection with the objective of enhancing the edges [7]. After
removal of noise, the image is classified using type-2 fuzzy rule based classifier.
Let the pixel intensity of an input image f(x,y) is represented by a vector X = (x1, x2,
…..xm) where m is the number of pixels in the image. Assume r is the total number of
fuzzy rules and one of such rule is represented as: If x1 is Rj1 and x2 is Rj2 and …..and
xm is Rjm then X =(x1,x2,x3, …xm) belongs to class k with CF=CFj where j=1,2,…,r and
k=1,2,…,M is the number of color classes classified using equation (1).
ji
jCFxGc ).(
max
arg= (1)
j
CF ∈[0,1] is the certainty factor of the j-th rule such that
)()( 1iR
m
i
ixxG
μ
=
∏= where )( iR x
μ
is the degree of membership value of pixel i in
type-2 fuzzy set R.
3 Grayscale Weighted Average Method
The classified regions are now segmented by detecting edges using GSA method. All
the edge points of the image constitute a set, called an edge map. Edge map is the
specific region bounded by neighborhood pixels within the same object, shown in
figure-2 where the color lines are the edges separated by different b/w color intensity.
Fig. 2. Edge Map in B/W Image
Pixels with membership value ‘1’/’0’ definitely belong/don’t belong to the edge
map set. However, pixels with intermediate membership values may or may not be-
long to the edge map, which are determined with certainty depending upon a pre-
scribed threshold value. After thresholding, a binary image is obtained, which is the
edge map representation of the original image.
The generalized version of the segmentation problem with an arbitrary number of
regions N is considered by providing the energy function of the model in (2).
∑∫++
∫
=Ω =ΓΩ
N
ii
i
i
ii ds
v
dxpNpE
12
log),,(
λ
(2)
Image Segmentation Using Grey Scale Weighted Average Method 279
Where i
Ω= i-th region, i
p= a priori probability of ith pixel, i
Γ = ith region boundary,
v = weighted parameter of boundary
Γ
, ds is the deviation of distance between two
region and the additional term of this energy functional penalizes the number of
regions with the parameter λ. Starting with the entire image domain Ω as a single
region, the two-region segmentation is applied in order to determine the best splitting
of the domain. If energy decreases by splitting, two regions are formed, which are
again divided and so on, until the energy does not decrease by further splits and thus
the optimum number of regions is determined.
GSA deals with the grey scale value having the same range between 0-1 like FWA
method. For the grey scale, the weighted average GSavg is computed as given in (3)
),...,,...(/ 25512551
255
1
255
1
ppwwfwpwGS
iii
iiavg == ∑∑ ==
(3)
Where wi is the weighted intensity and pi is the pixel intensity.
4 Proposed Algorithm with GSA
Begin
Read input image f(x, y) of size M×N
Create the mask W (m×n) with mask coefficients, using
Sparse matrix so that sum of all coefficients of each
Let mask=0 ;
Mask weighted average, a=(m–1)/2 // small size
Mask weighted average, b=(n–1)/2 // large size.
g(x, y) = 0 // output image
For y = b to (N – b – 1) do
For x = a to (M – a – 1) do
Calculate the (largest value) among all the maximum column values;
Calculate the (smallest value) among all the minimum column values;
GSA_range = (largest value) – (smallest value) / n
For x = 1 to N do
for y = 1 to M do
g(x, y) = (f(x, y) – smallest pixel value) × 255 / GSA_range;
End for
Select a color image and convert it into grey scale input image;
Store pixel values of the image along x and y coordinates in matrix form;
Generate the Convolution mask for different gradient operators and store it ;
SUMcoeffmatrix = 0; //Set all the points as black //
Each mask along the horizontal and vertical direction is convolved
with the input image; Calculate magnitude of the gradient vector;
If Ethresold < 0.55 // Threshold is required to determine whether the point
belongs to a specific region or not //
Truncate unwanted edges from edge map information;
Else include edge in the edge map set;
End_for
End.
280 S. Maity and J. Sil
5 Conclusion and Result Discussion
The proposed algorithm is reasonably fast where 169 × 250 size image took 22.5 sec-
onds on an Intel Atom 1.6 Ghz Processor having 1 GB of RAM size. By keeping the
advantages of the level set framework, its main problem has been solved in the paper.
The customized Gaussian filter [8] has a good contrast and sharpness characteristics,
which is required to sharpen an image. This paper shows that the proposed segmenta-
tion model exhibits better performance in edge detection compare to the conventional
method of segmentation [9].
References
[1] Argyle, E.: Techniques for edge detection. Proc. IEEE 59, 285–286 (1971)
[2] Chidiac, H., Ziou, D.: Classification of Image Edges. In: Vision Interface 1999, pp. 17–24.
Troise-Rivieres, Canada (1999)
[3] Hueckel, M.: A local visual operator which recognizes edges and line. J. ACM 20(4),
634–647 (1973)
[4] Malik, J., Belongie, S., Shi, J., Leung, T.K.: Textons, contours and regions: cue integration
in image segmentation. In: Proc. IEEE International Conference on Computer Vision,
Corfu, Greece, pp. 918–925 (September 1999)
[5] Heath, M., Sarkar, S., Sanocki, T., Bowyer, K.W.: Comparison of Edge Detectors: A
Methodology and Initial Study. Computer Vision and Image Understanding 69(1), 38–54
(1998)
[6] Shin, M.C., Goldgof, D., Bowyer, K.W.: Comparison of Edge Detector Performance
through Use in an Object Recognition Task. Computer Vision and Image Understand-
ing 84(1), 160–178 (2001)
[7] Peli, T., Malah, D.: A Study of Edge Detection Algorithms. Computer Graphics and Image
Processing 20, 1–21 (1982)
[8] Gonzalez, R.C., Woods, R.E.: Digital Image Processing, 2nd edn. Prentice-Hall, Inc.,
Upper Saddle River (2002)
[9] Pratt, W.K.: Digital Image Processing, 4th edn. John Wiley & Sons, Inc., Hoboken (2007)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 281–286, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Cluster Analysis and Pso for Software Cost Estimation
Tegjyot Singh Sethi1, CH.V.M.K. Hari2, B.S.S. Kaushal1, and Abhishek Sharma1
1 Dept. of CSE, Gitam University, Visakhapatnam, India
{tjss401, kaushalb09, abhisheksharma9129}@gmail.com
2 Dept. of IT, Gitam University, Visakhapatnam, India
kurmahari@gmail.com
Abstract. The modern day software industry has seen an increase in the number
of software projects .With the increase in the size and the scale of such projects
it has become necessary to perform an accurate requirement analysis early in
the project development phase in order to perform a cost benefit analysis. Soft-
ware cost estimation is the process of gauging the amount of effort required to
build a software project. In this paper we have proposed a Particle Swarm Op-
timization (PSO) technique which operates on data sets which are clustered us-
ing the K-means clustering algorithm. The PSO generates the parameter values
of the COCOMO model for each of the clusters of data values. As clustering
encompasses similar objects under each group PSO tuning is more efficient and
hence it generates better results and can be used for large data sets to give accu-
rate results. Here we have tested the model on the COCOMO81 dataset and also
compared the obtained values with standard COCOMO model. It is found that
the developed model provides better estimation of the effort.
Keywords: Particle Swarm Optimization (PSO), K-Means, Software Cost
Estimation, Constructive Cost Model (COCOMO).
1 Introduction
The software industry today is all about efficiency. The provident allocation of the
available resources and the judicious estimation of the requisites form the basis of any
planning and scheduling activity. With the increase in the expanse and impact of mod-
ern day software projects, the need for accurate requirement analysis early in the
software development phase has become pivotal. For a given set of requirements, it is
desirable to cognize the amount of time and money required to deliver the project pro-
lifically. The chief aim of software cost estimation is to enable the client and the devel-
oper to perform a cost – benefit analysis. The cost / effort estimates are determined in
terms of person-months(pm) which can be easily commuted to actual currency cost.
The cost of the software varies depending on both complexity and lines of code and to
estimate the cost we make use of Particle Swarm Optimization on clustered data. A
common approach to the estimation of the software effort is by expressing it is as a
function of the project size[1] and the Effort Adjustment Factor (EAF). The equation
of effort in terms of size and methodology is considered as follows:
282 T.S. Sethi et al.
Effort = a*(size)b * (EAF)+c (1)
Here a, b, c are constants. The constants are usually determined by regression analysis
applied to historical data[11]. There are a number of models proposed for tuning
parameters using Neural Networks, Machine learning techniques, fuzzy techniques
and Genetic algorithms[2][3][5][6][8].
PSO is a robust stochastic optimization technique [4][9][10] based on the
movement of intelligent swarms . PSO applies the concept of social interaction to
problem solving. It uses a number of agents (particles) that constitutes a swarm
moving around in the search space looking for the best solution. Each particle is
treated as a point in an N- dimensional space which adjusts its movement. According
to its own flying experience (Pbest- personal best) as well as the flying experience of
other particles (Gbest –global best). The basic concept of PSO lies in accelerating each
particle towards its Pbest and Gbest locations with regard to a random weighted
acceleration at each time. The modifications of the particle’s positions can be
mathematically modeled by making use of the following equations:
Vk+1 = wVik + c1 rand1 (Pbest – Sik) + c2 rand2 (Gbest – Sik) (2)
Sik+1 = Sik + Vik+1 (3)
Where, Sik is current search point; Sik+1 is modified search point; Vik is the current
velocity; Vk+1 is the modified velocity; Vpbest is the velocity based on Pbest;Vgbest =
velocity based on Gbest; w is the weighting function; cj is the weighting factors; randj
are uniformly distributed random numbers between 0 and 1. In the particle swarm
optimization technique, the particles searches the solutions in the solution space with-
in the range [-s,s]. The PSO works better when operated on datasets having similar
valued objects. Hence to enhance the PSO computations we have clustered the given
dataset into groups of similar projects using the K Means clustering algorithm. K-
Means clustering[12] is a method of cluster analysis which aims to partition n
observations into k clusters in which each observation belongs to the cluster with
the nearest mean. Here an iterative refinement method is employed to find the
means(centroids) of the cluster. Any new value entered is first ascertained to be of a
particular cluster and then the PSO estimation is performed on it. The parameter thus
estimated gives lesser error and closer results than those without using clustering.
2 Model Descriptions
In this model we have considered “The standard PSO with inertia weights” which
works on data clustered by using the K Means clustering algorithm. The effort is giv-
en by Equation 1, in which a, b and c are parameters to be tuned by PSO.
In PSO, swarm behavior is used for tuning the parameters of the Cost/Effort estima-
tion. As the PSO is a random weighted probabilistic model the previous benchmark
data is required to tune the parameters. Based on that data, swarms develop their intel-
ligence and empower themselves to move towards the solution .We initially develop
clusters having similar values and then apply the PSO to each cluster individually to
Cluster Analysis and Pso for Software Cost Estimation 283
obtain the parameter value. The following is the methodology employed to tune the
parameters in the proposed models following it.
3 Methodology
Here we describe the methodology for our model which uses K means clustering
algorithm and implements PSO on these clusters.
Input: k the number if clusters, d-data set containing size of software projects,
measured efforts, EAF (complexity factor).
Output: Set of k clusters, Optimized parameters for the clusters.
Step 1: Choose k objects from d as initial centroids.
Step 2: Assign each object to a cluster based on the minimum distance between the
centroid and the value.
Step 3: Update the cluster mean by assigning centroid as the mean of all data values
as the new centroid.
Step 4: Repeat the above steps until we get the stable clusters. These clusters
obtained are tuned using PSO.
Step 5: Initialization: Initialize particles with random positions and velocity vectors
of tuning parameters .We also need the range of velocity between [- Vmax,
Vmax].
Step 6: Evaluation of Fitness Function: For each particle position with values of
tuning parameters, evaluate the fitness function. The fitness function here is
Mean Absolute Relative Error (MARE). The objective in this method is to
minimize the MARE by selecting appropriate values from the ranges
specified in step 1.
Step 7: Finding the Pbest – Personal best: If fitness (p) better than fitness (Pbest) then:
Pbest = p. Here the P
best is determined for each particle by evaluating and
comparing measured and estimated effort values of the current and previous
parameters values.
Step 8: Finding the Gbest (global best): Set the best of ‘Pbest’ as global best – Gbest.
The particle value for which the variation between the estimated and
measured effort is the least is chosen as the Gbest particle.
Step 9: Update values: Update the velocity and positions of the tuning parameters
with equations (2) & (3).
Step 10: Repeat steps 2 to step 5 until “particles exhaust”.
Step 11: Give the Gbest values as the optimal solution.
Step 12: Stop.
4 Model Analysis
We have implemented the above methodology, for tuning the parameters a, b and c in
“C” language. For the parameter’ a,b and c ‘the velocity and position of the particles
are updated by applying the equations (2) & (3), with parameters w=0.5, c1=c2=2.0.
The data is clustered by using K Means. The final allocation matrix defines sets of
values in each cluster. We apply PSO individually on each cluster and output a set of
parameter values for each cluster.
284 T.S. Sethi et al.
5 Performance Criterion
We consider three performance criterions which are Variance Accounted –For (VAF),
Mean Absolute Relative Error (MARE) and Variance Absolute Relative Error
(VARE) [7].
6 Experimental Study
For the study of this model we have taken 45 data values from COCOMO81 dataset, for
both training and testing. We have considered 3 clusters (0, 1, 2) the clusters are indi-
cated in Table 1. By running the ‘C’ implementation of the above methodology we
have obtained the following parameters for the proposed model.
Cluster 0: a=0.314800; b=1.862067; c= -5.062736
Cluster 1: a=0.079505; b=2.137003; c= -3.657698
Cluster 2: a=4.182602, b=0.963337; c= -1.449817
The Measured and Estimated Efforts Corresponding to Cluster PSO & COCOMO are
given in Table 1 and the corresponding graph is depicted in Fig 1. The Performance
Criterion is given in Table 2.
Table 1. Measured (ME) and Estimated (EE) Efforts Corresponding to Cluster PSO &
COCOMO
Cluster
No. Size EAF ME COCOMO EE
Cluster
No. Size EAF ME COCOMO EE
0 16 0.66 33 39 31.22 1 28 0.96 83 102 90.80
0 18 2.38 321 214 157.87 1 30 1.14 87 130 126.33
0 20 2.38 218 243 193.19 1 32 0.82 106 100 103.67
0 24 0.85 79 108 94.36 1 28 0.45 50 47 40.62
0 13 2.81 98 133 99.89 2 4 2.22 43 30 33.85
0 22 1.76 230 201 170.01 2 6.9 0.4 8 9.8 9.30
0 13 2.63 82 161 93.16 2 3 5.86 73 60 69.18
0 12 0.68 55 33 16.82 2 3.9 3.63 61 52 54.88
0 15 0.35 12 20 12.00 2 3.7 2.81 40 38 40.00
0 19.5 0.63 45 46 45.00 2 1.9 1.78 9 10.7 12.37
0 23 0.38 36 33 36.00 2 9.4 2.04 88 89 72.43
0 24 1.52 176 193 172.73 2 2.14 1 7.3 7 7.25
0 15 3.32 237 239 156.79 2 1.98 0.91 5.9 5.9 5.90
0 25 1.09 130 145 132.51 2 6.2 0.39 8 8.4 8.01
0 21 0.87 70 68 74.30 2 2.5 0.96 8 8.1 8.26
1 46 1.17 240 212 328.92 2 5.3 0.25 6 4.7 3.76
1 30 2.39 423 327 268.86 2 10 3.18 122 114 120.79
1 37 1.12 201 238 196.26 2 8.2 1.9 41 55 58.88
1 48 1.16 387 239 357.48 2 5.3 1.15 14 22 22.53
1 50 3.14 1063 962 1063.00 2 4.4 0.93 20 14 14.76
1 40 2.26 605 529 472.89 2 6.3 0.34 18 7.5 6.92
1 34 0.34 47 44 47.00 2 6.7 2.53 57 60 64.67
2 9.1 1.15 38 42 38.92
Cluster Analysis and Pso for Software Cost Estimation 285
Fig. 1. Measured Effort Vs Estimated Effort
Table 2. Performance Criterion
Model VAF (%) MARE (%) VARE (%)
Cluster-PSO Model 94.6618 17.3564 3.7287
COCOMO Model 95.4722 20.7606 4.1224
7 Conclusion
Software cost estimation is based on a probabilistic model and hence it does not gen-
erate exact values. However if good historical data is provided and a systematic tech-
nique is employed we can generate better results. In this study we have proposed a
new model to estimate the software effort. In order to cluster the values into similar
groups we have used K-means clustering algorithm and in order to tune the parame-
ters we use particle swarm optimization methodology algorithm. It is observed that
the clustered-PSO gives more accurate results when juxtaposed with the standard
COCOMO. On testing the performance of the model in terms of the error rate the
results were found to be useful. This method can be applied to large datasets to gener-
ate efficient values. These techniques can be applied to other software effort models.
References
1. Bailey, J.w., Basili, V.R.: A meta model for software development resource expenditures.
In: Fifth International conference on software Engineering, pp. 107–129. IEEE, Los
Alamitos (1981), CH-1627-9/81/0000/0107500.75@
2. Briand, L.C., El Emam, K., Bomarius, F.: COBRA: A Hybrid Method for Software Cost
Estimation, Benchmarking, and Risk Assessment, International Software Engineering Re-
search Network Technical Report ISERN-97-24 (Revision 2), PP 1-24 (1997)
3. Gruschke, T.: Empirical Studies of Software Cost Estimation: Training of Effort Estima-
tion Uncertainty Assessment Skills. In: 11th IEEE International Software Metrics Sympo-
sium (METRICS 2005),1530-1435/05. IEEE, Los Alamitos (2005)
286 T.S. Sethi et al.
4. Jørgensen, M.: Evidence-Based Guidelines for Assessment of Software Development Cost
Uncertainty. IEEE Transactions on Software Engineering 31(11), 942–954 (2005)
5. Sheta, A.F.: Estimation of the COCOMO Model Parameters Using Genetic Algorithms for
NASA Software Projects. Journal of Computer Science 2(2), 118–123 (2006)
6. Auer, M., Trendowicz, A., Graser, B., Haunschmid, E., Biffl, S.: Optimal Project Feature
Weights in Analogy-Based Cost Estimation: Improvement and Limitations. IEEE Transac-
tions on Software Engineering 32(2), 83–92 (2006)
7. Hari, C.V.M.K., Prasad Reddy, P.V.G.D., Jagadeesh, M.: Interval Type 2 Fuzzy Logic for
Software Cost Estimation Using Takagi-Sugeno Fuzzy Controller. In: Proceedings of 2010
International Conference on Advances in Communication, Network, and Computing.
IEEE, Los Alamitos (2010), doi:10.1109/CNC.2010.14, 978-0-7695-4209-6/10
8. Jørgensen, M., Shepperd, M.: A Systematic Review of Software Development Cost Esti-
mation Studies. IEEE Transactions on Software Engineering 33(1), 33–53 (2007)
9. Poli, R., Kennedy, J., Blackwell, T.: Particle swarm optimization An overview. In: Swarm
Intell., pp. 33–57. Springer, Heidelberg (2007), doi:10.1007/s11721-007-0002-0
10. Felix, T.S.C., Tiwari, M.K.: Swarm Intelligence: Focus on Ant and Particle Swarm Opti-
mization, pp. 1–548. I-TECH Education and Publishing (2007), ISBN 978-3-902613-09-7
11. Keung, J.W., Kitchenham, B.A., Jeffery, D.R.: Analogy-X: Providing Statistical Inference
to Analogy-Based Software Cost Estimation. IEEE Transactions on Software Engineer-
ing 34(4), 471–484 (2008)
12. Bin, W., Yi, Z., Shaohui, L., Zhonghi, S.: CSIM: A Document Clustering Algorithm Based
on Swarm Intelligence, pp. 477–482. IEEE, Los Alamitos (2002), 0-7803-7282-
4/02@2002
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 287–290, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Controlling Crossover Probability in Case of a Genetic
Algorithm
Parama Bagchi1 and Shantanu Pal2
1 Department of Computer Science and Engineering,
Dream Institute of Technology,
Kolkata, India
2 AKC School of Information Technology,
University of Calcutta,
Kolkata, India
{paramabagchi,shantanu.smit}@gmail.com
Abstract. Genetic Algorithms (GAs) are commonly used today worldwide.
Various observations have been theorized about genetic algorithms regarding
the mutation probability and the population size. Basically these are the search
heuristics that mimic the process of natural evolution. This heuristic is routinely
used to generate useful solutions for optimization and search problems. GAs
belong to the larger class of evolutionary algorithms (EAs), which generate so-
lutions to maximize problem solving by using techniques inspired by natural
evolution, such as inheritance, mutation, selection, and crossover. In this paper
we study of a simple heuristic in order to control the crossover probability of a
GA. We will also explain how stress factors in on the crossover probability and
why it is an important phenomenon in case of a GA and how it can be con-
trolled effectively. Experimental results show that, for reaching lower probabili-
ty from higher probability, we can get faster optimal solutions for any problem.
These experimental values are derived by taking the values at the high probabil-
ity and then slowly yet steadily decreasing them.
Keywords: genetic algorithm, crossover probability, heuristic, evolutionary
algorithms.
1 Introduction
GAs is generally portrayed as a search procedure which can optimize functions based
on a limited sample of function values. A function based on a minimal spanning tree
of data points can be used for clustering and GAs in an attempt to optimize the speci-
fied objective function in order to detect natural grouping in a given data set [1]. The
method has also found to provide good results for various real life data sets [2].
Unlike conventional search methods GAs deals with multiple solutions simultaneous-
ly and computes the fitness function values for these solutions [2]. GAs was inspired
by the way living things evolved into more successful organisms in nature [3]. In
this paper we work on the general ideas of the GAs and the experimental values
gained from the input values regarding the proposed field. We do not try to define any
288 P. Bagchi and S. Pal
re modification of traditional GAs, rather we shows that taking values from higher
probability to lower probabilities will make a successive change to reach the optimal
solution of any problems. This is intern the basic motivation if this paper. The rest of
the paper is organized as follows: Section 2 provides an understanding of the simula-
tion and natural selection. Section 3 recalls the enlisted crossover techniques. In sec-
tion 4 we extend our experimental results which is the motivate aim of our work. Fi-
nally section 5 draws our conclusions.
2 Simulation and Natural Selection
To simulate the process of natural selection in a computer, representation of an indi-
vidual and fitness function is essential to determine. They are as follows:
2.1 Representation of an Individual
At each point during the search process in [4], Filho et all maintains a "generation" of
"individuals." Each individual is a data structure representing the "genetic structure"
of a possible solution. In GAs, the alphabet {0, 1} is usually used for 2 clusters. This
string is interpreted as a solution to the problem in [4] trying to solve, if we want to
find the optimal quantity of the three major components, we can use the alphabet
{1, 2, 3 ..., 9} denoting the number of each ingredient.
2.2 Fitness Function
Given an individual, we must assess how good a solution it is so that we can rank
individuals. A possible function is Fitness (i) = (sum of edge weight of MSTs)/ (Total
no of points in each cluster). This function has a value between 0 and 1 and is mono-
tonically increasing [5].
3 The Crossover Technique Enlisted
The crossover technique [6] is one of the most fundamental operations in the evalua-
tion of GA. Generally if there is a medium - large population a great responsibility is
own to this crossover operator. This is primarily because of the fact that in compari-
son to the mutation operator, the crossover operator owes and brings about a great
change in the structure of the chromosome which is finally responsible for the con-
vergence of the GA. Normally in case of a very small population; crossover might not
have a very significant importance as compared to mutation. The test so as to ensure
that crossover works best when it is kept at a high range of probability values was
done.
4 Experimental Results
As the main aim was to show the result generated by the experiment which reach the
optimal solution quicker. It is also cleared that we have started the experiment with a
Controlling Crossover Probability in Case of a Genetic Algorithm 289
high probability and gradually decreased it to a lower probability region, which in-
deed reflect the quicker optimal solution for a problem. The following observations
were seen by conducting the underneath experiment. Here we have taken clusters
consisting of a number of points. For the points correct clustering is found when slow-
ly the crossover probability is decreased from high to low. Excellent results are found
if the change between a higher probability and a lower probability is decreased by an
amount of 0.1. Decreasing the crossover probability, slowly and step by step would
help us to actually see the increasing change in the nature of chromosomes, which
finally help us to achieve convergence. By decreasing the crossover probability in this
fashion, the sample string changed slowly, till we got the final convergence. In con-
trast to this when the crossover probability was kept at a low range say [0.5], and
slowly increased, we did not time taken to converge was longer. The following table
gives a snapshot of the discussions made above.
Table 1. Crossover Probability by the Experiment
Size of
Data
Number of
Iterations
Number of Points
Correctly Clustered
Percentage (%) of Correct
Classification
20 Points 700
800
900
5000
6000
7000
12
11
11
11
11
20
40%(Crossover Æ0.6)
30%
30%
45% (Crossover Æ0.8)
45%
100%
30 Points 900
1000
2000
9000
10000
15000
21
18
30
21
21
21
30% (Crossover Æ0.6)
40%
100%
30% (Crossover Æ0.8)
30%
30%
Fig. 1. Percentage of Correct Classification for 20 and 30 Points Data
0
30
60
90
120
123456
Percentage (%) of Correct
Classification
Number of Iterations
20 Points
data
30 Points
data
290 P. Bagchi and S. Pal
The above experimental values were actually derived by taking the values at a high
probability and then slowly made to decrease steadily. The final experimental results
show that the probabilities have been taken starting from a high value to steadily low
values. The graph below shows the results.
5 Conclusion
If the conception of a computer algorithms being based on the evolution of organism
is surprising, the extensiveness with which this algorithms is applied in so many areas
is no less than astonishing. These applications, be they commercial, educational and
scientific, are increasingly dependent on this algorithms, the GAs. Its usefulness and
gracefulness of solving problems has made it a more favorite choice among the tradi-
tional methods, namely gradient search, random search and others. In conclusion, it
can as well be said that by increasing the no of iterations it is possible to get a better
set of results. The size of the population should be increased to see the optimal solu-
tion and to be able to deduce the fittest chromosome for a large population. The stop-
ping criterion for genetic algorithms is still not clear. A significant work can also be
done in that region.
References
1. Togelius, J., Preuss, M., Yannakakis, G.N.: Towards Multiobjective Procedural Map Gener-
ation. In: PCGames Monterey, CA, USA, June 18 (2010)
2. Chowdhury, N., Jana, P.: Finding the Natural Groupings in a Data Set Using Genetic
Algorithms. In: Manandhar, S., Austin, J., Desai, U., Oyanagi, Y., Talukder, A.K. (eds.)
AACC 2004. LNCS, vol. 3285, pp. 26–33. Springer, Heidelberg (2004)
3. Mathew, T V.: Genetic Algorithm,
http://www.civil.iitb.ac.in/tvm/2701_dga/
2701-ga-notes/gadoc/gadoc.html (last access November 2009)
4. Filho, J.R., Alippi, C., Treleaven, P.: Genetic Algorithm Programming Environment’s.
ACM Journal Computer 27(6) (June 1994), doi:10.1109/2.294850
5. Rokach, L., Maimon, O.: Clustering Methods. In: Data Mining and Knowledge Discovery
Handbook, vol. 3, pp. 321–325 (2005), doi:10.1007/0-387-25465-X_15
6. Genetic Algorithms: Principles of Natural Selection Applied to Computation. Science, 261,
872-878
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 291–296, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Qualitative Survey on Unicast Routing Algorithms in
Delay Tolerant Networks
Sushovan Patra1, Anerudh Balaji1, Sujoy Saha2,
Amartya Mukherjee1, and Subrata Nandi1
1 Department of Computer Science and Engg., 2 Department of Computer Application National
Institute of Technology, Durgapur, 713209, India
{bubususpatra, anerudhbalaji, sujoy.ju, mamartyacse1,
subrata.nandi}@gmail.com
Abstract. DTN is an emerging research area that takes a different approach to
(inter)networking and allows working in stressed as well as in highly heteroge-
neous environments.DTN features a number of unique properties which make
this concept applicable to challenging environments in which traditional com-
munication paradigms would fail or perform poorly. In DTN intermediate nodes
takes the responsibility to store-and-forward delivery as well as physical data
carriage using deterministic and/or probabilistic routing. In this paper we pre-
sent a comprehensive up to date survey of different routing protocols as well as
perform qualitative comparison of different routing strategies with respect to
important issues in DTN. Further we highlight some of the upcoming issues re-
lated to design of DTN routing strategy.
1 Introduction
In wired and wireless network a large amount of wire-less bandwidth capacity remains
unused because the current communications paradigm (i.e. the Internet) has not been
designed to make use of local and intermittent connectivity. DTN was an attempt to
extend the reach of the network which enables communication between ‘Challenged’
networks which is a technical by-product of absence of end to end connection and
unpredictable node mobility. DTN has greater application in disaster management [11]
but as the cost of the handle devices like laptops, mobiles phones, pager has decreased,
it can be introduced in wide applications of modern technologies.
Routing Protocols in Delay Tolerant Networks classifies the routing family in two
categories - Forwarding based and Flooding based and compares the protocols pro-
posed in the flooding families and the forwarding families on the basis of resource
usage, effectiveness amongst other factors [8][15][1]. Previous survey papers do not
consider the latest works on Delay tolerant network routing and are updated till the
year 2006 only this has motivated us to come up with this survey that encompasses
the most recent work. Our paper also provides a different perspective on sub classifi-
cation of routing techniques in Delay Tolerant Networks. The paper is divided into
three sections. In the following section we present the Taxonomy and produce the
summary in tabular form (Table 1) and section 3 presents the Conclusion and future
directions. The figure below represents our proposed taxonomy.
292 S. Patra et al.
Routing algorithms for DTN
Flooding Forwarding
Fig. 1. Classification of DTN routing algorithms based on a new taxonomy
2 Proposed Taxonomy of Dtn Routing Strategies
The routing strategy for Delay Tolerant networks can be categorized based on two
properties namely Replication (flooding) and Knowledge (forwarding) [1].
2.1 Flooding Based
The relays [1] in DTN store the messages until they connect with the destination, at
which point the message is delivered.
2.1.1 Direct Contact
Waits until source comes into contact with the destination before forwarding the data
but does not necessarily consider shortest path as it does not use knowledge of the
network [1][8].It is assumed to be degenerated case of forwarding as it does only use
single copy of the message.
2.1.2 Tree Based Flooding
Tree Based Flooding distributes the task of making copies to the relayed node; with
an indication for the number of copies required [1]. The concept is quite similar to
diffusion computation algorithm. Two hop relay is a special case of Tree based flood-
ing since this strategy is a case of tree based flooding with depth 1.
2.1.3 Exchange Based
This category includes those routing strategies in which exchange of information
plays a pivotal part in the procedure. Three algorithms – PRoPHET [2], Epidemic [1]
and Practical Routing [10] fall under this category. Epidemic routing [8] maintains a
Summary Vector that will keep track of the entire message Ids .Two nodes send each
other the list of all the messages IDs they have in their buffers on contact and ex-
change the messages they do not have using IDs. Extreme flooding leads to huge
amount of resources consumption [1], [8]. PRoPHET[2] is extension of epidemic
routing which along with Summary Vector also contain the delivery predictability
information stored at the nodes to update the internal delivery predictability vector [2]
[8]. Extreme flooding of previous one is here controlled by threshold value. Practical
Routing [10] uses only observed information about the network. A metric is designed
that estimates how long a message will have to wait before it can be transferred to the
Direct Tree Exchange Utility
Contact Based Based Based Location Metric Social Structure
Based Based Based
A Qualitative Survey on Unicast Routing Algorithms in Delay Tolerant Networks 293
next hop. Messages are exchanged if the topology suggests that a connected node is
“closer” (has better metric) than the current node.
2.1.4 Utility Based Routing
Flooding leads to network congestion and drastic fall in network throughput, which
forms the basis for utility based routing schemes. In strategies that fall under utility
based routing, the utility of the message is considered while making routing decisions.
ORWAR [13] proposes a multi copy routing scheme, using controlled replication and
a fixed number of copies distributed over network. At each contact the node tries to
forward half of the message copies, keeping the rest for itself, messages with best
utility per bit ratio selected first and forwarded. RAPID [14] optimizes a specific rout-
ing metric such as worst case delivery delay or the fraction of packets that are deliv-
ered within a deadline. It models DTN routing as a utility driven resource allocation
problem.
In Direct contact [1] [8] delivery ratio is dependent on the proximity of the destina-
tion with respect to the source. Tree based flooding [1] supports multi path routing of
the messages. Epidemic routing promises a high delivery ratio and exhibits average
effectiveness as it consumes much more buffer space with compare to any other rout-
ing. Utility based routing have the highest delivery ratio while creating a far lesser
overhead. A qualitative comparison as per the above mentioned category is illustrated
in the table 1.
2.2 Forwarding Based
It uses the network topology information to select the best path, and the message is
then forwarded from node to node along this path.
2.2.1 Location Based Routing
This forwarding approach requires the least information about the network and as-
signs coordinates (can be GPS) to each of the nodes. A distance function is used to
estimate the cost of delivering messages from one node to another. In general a
message is forwarded to a potential next hop if that node is closer in than the current
custodian with respect to destination node [1].DTN routing in mobility space [15]
defines a generic routing scheme for DTNs using a high-dimensional Euclidean space
constructed upon nodes’ mobility patterns, and assigns coordinates to the nodes in this
space based upon their probability of being found in each possible location.
2.2.2 Metric Based Routing
In this approach, a weight is assigned to each link based on its suitability in delivering
messages to a given destination as in Gradient Routing[9].Each node must store a
metric for all potential destinations. When the custodian of a message contacts an-
other node that has a better metric for the message’s destination, it passes the message
to it [1], [9]. Friendship based routing [7] finds a metric that reflects the node rela-
tions more accurately by considering the following three behavioral features of close
friendship: high frequency, longevity, regularity. These properties addressed in new
metric called Social Pressure Metric, SPM which gives the information regarding
delay for message passing. Then information in SPM is use to choose best forwarder.
294 S. Patra et al.
2.2.3 Social Structure Based Routing
This class of routing algorithms discovers the heterogeneity of human interactions
such as community formation from real world human mobility traces.
Predictability of the node mobility motivates the idea of BUBBLE Rap [3] .Each
node has a global ranking across the system and a local ranking for its local commu-
nity and may belong to multiple communities with multiple local rankings. The algo-
rithm detects the community using K-clique by Palla et al. and weighted network
analysis (WNA) by Newman. The forwarding is done based on popularity of the node.
Table 1. Summary of flooding and forwarding based routing schemes. Sub classification has
been done according to the propose taxonomy (fig 1).
Routing
Family
Routing Categories Motivation Remarks
Direct Contact
[1]
The need for a simple
and somewhat
effective model of
flooding.
1. Simple and does not consume mu
c
resource.
2. Works only if source contacts
destination.
Tree based
[1]
The need to reduce the
number of copies in
the network and
reduce network
congestion
1. Can deliver messages to
destinations that are multiple hops
away.
2. Tuning of parameters is a major
hindrance.
Utility based
[13], [14]
The need to optimize
specific routing metric
such as average
delivery delay.
1. It does minimize retransmissions.
2. Effective use of resources at
system level.
Flooding
Exchange based
[1],[2],[10]
It was originally
proposed so that
synchronization of the
databases is possible
1. Huge amount of resources-buffer
space, bandwidth are used.
2. No knowledge of network
required.
3. High delivery ratio.
Location based
[1],[15]
Run time
determination of
destination location to
avoid router
maintenance cost.
1. Less knowledge requirement
2. Location does not necessarily
correspond to network topology.
3. Change in network topology
complicates routing.
Metric based
[1], [7], [9]
Using information’s
like last contact with
destination, battery
energy, mobility in
from of metrics.
1. Requires more knowledge
compare to Location based.
2. Achieve higher Delivery
Probability than Direct contact
3. Fail to balance load.
Forwarding
Social structure
based
[3], [4], [5], [6]
Mobility traces of the
real world have been
used and heterogonous
social relationship in
term of group and
individual has been
identified for first
time.
1. Delivery ratio achieved is better
than PRoPHET [3].
2. Requires a setup period.
3.Latency is reduced by 20%
compared to PRoPHET
A Qualitative Survey on Unicast Routing Algorithms in Delay Tolerant Networks 295
Incompetent assumption of the network graph is fully connected in a social network
motivates the idea of SimBet which is based on the concept of Ego network. A node
forwards messages to a node with higher Social Similarity [4] or Higher Between’s
[4]. The major disadvantage of SimBet is that it models social relationship in a binary
form, which is not realistic and leads to inconsistencies which lead to SimBetAge [5].
It improves on similarity estimation and calculation of between’s, by using weighted
time dependant graph and exponential decay function [5]. Fair routing [6] attempts to
overcome the existing problems of security and load balancing. Interaction Strength is
an indicator of the likelihood of a contact to be sustained over time Nodes will only
accept forward request from those nodes of equal or higher status (popularity) i.e.
higher Queue length which may lead to starvation for new node in the network.
Location based routing offers very little flexibility and have scalability issues but
consume minimal resources. Metric based routing offers comparatively lesser delay
and a better delivery ratio. Social Structure based routing, achieves a better delivery
ratio than the other routing categories and also addresses load balancing and security
issues as in Fair Routing [6]. A qualitative illustration is presented in the Table 1.
3 Conclusions and Future Directions
In spite of the continuous research in DTN in last few years, none of the works pre-
sented here pursue dynamic allocation of buffering spaced based on applications.
Other than Fair routing [6], none of the papers surveyed here considers the issue of
security, which is of paramount importance while using Delay tolerant networks for
applications such as Social Networks. The techniques proposed for load balancing
such as Message splitting [12] and Fair Routing [6] are not convincing. Even though
they achieve significant improvement in terms of balancing, the overall performance
of the routing protocol is comparable to that of PRoPHET [2]. The recent improve-
ment in research arena has shown the glimpse that DTN is quite sustainable in large
network structure. So we have identified the issues security, flexibility, scalability,
and mobility support, load balancing require to be addressed so the DTN can be used
as a platform to host upcoming application like Social Network, bulk data transfer,
multimedia data transfer and streaming in Sparse Network.
References
1. Jones Paul, E.P.C., Ward, A.S.: Routing Strategies for Delay Tolerant Networks. Submit-
ted to Computer Communication Review (2008)
2. Lindgren, A., Doria, A., Schelen, O.: Probabilistic Routing in Intermittently Connected
Networks, vol. 3126, pp. 239–254 (2004)
3. Hui, P., Crowcroft, J., Yoneki, E.: BUBBLE Rap: Social-based Forwarding in Delay Tol-
erant Networks. In: MobiHoc, Hong Kong SAR, China, May 26–30 (2008)
4. Daly, E., Haahr, M.: Social Network Analysis for Routing in Disconnected Delay-Tolerant
MANETs. In: MobiHoc, Montreal, Canada, September 9-14 (2007)
5. Link, J.Á.B., Viol, N., Goliath, A., Wehrle, K.: SimBetAge: Utilizing Temporal Changes
in Social Networks for Pocket Switched Networks, U-NET, Rome, Italy, December 1
(2009)
296 S. Patra et al.
6. Pujol, J.M., Toledo, A.L., Rodriguez, P.: Fair Routing in Delay Tolerant Networks. In:
Proc. of IEEE Infocom Rio de Janeiro, Brazil, April 19-25 (2009)
7. Bulut, E., Szymanski, B.K.: Friendship Based Routing in Delay Tolerant Mobile Social
Networks. In: IEEE GLOBCOM Miami Florida USA (December 2010)
8. Shen, J., Moh, S., Chung, I.: Routing Protocol in Delay Tolerant Network: A Comparative
Survey. In: The 23rd International Technical Conference on Circuits/System Computer
and Communication (2008)
9. Poor, R. D.: Gradient routing in ad hoc networks. MIT Media Laboratory (2000),
http://www.media.mit.edu/pia/Research/ESP/texts/
poorieeepaper.pdf (unpublished manuscript)
10. Jones, E.P.C., Li, L., Ward, P.A.S.: Practical routing in delay-tolerant networks. In: Proc of
the ACM SIGCOMM workshop on Delay-tolerant networking (2005)
11. Uddin, Y.S., Nicol, D.M., Abdelzaher, T.F., Kravets, R.H.: A Post-Disaster Mobility
Model for Delay Tolerant Networking. In: Proc. Winter Simulation Conference, Austin,
Texas (December 2009) (invited)
12. Jain, S., Fall, K., Patra, R.: Routing in a Delay Tolerant Network. In: SIGCOMM,
August 30-September 3 (2004)
13. Sandulescu, G., Tehrani, S.: Opportunistic DTN Routing with window aware adaptive rep-
lication. In: Proceedings of the ACM 4th Asian Conference on Internet Engineering (AIN-
TEC), Bangkok, Thailand (2008)
14. Balasubramanian, A., Levine, B., Venkatramani, A.: Replication Routing in DTN: A Re-
source allocation approach. In: SIGCOMM, Kyoto, Japan, August 27-31 (2007)
15. Leguay, J., Friedman, T., Conan, V.: DTN Routing in a mobility pattern space. In:
ACM WDTN, pp. 276–283 (2005)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 297–300, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Designing and Modeling of CMOS Low Noise Amplifier
Using a Composite MOSFET Model Working at
Millimeter-Wave Band
Adhira Raj1, Balamurugan Karthigha2, and M. Jayakumar3
1 M. Tech, VLSI Design, Amrita Vishwa Vidyapeetham, Coimbatore, India
adhira.raj@gmail.com
2 Assistant Professor, Amrita Vishwa Vidyapeetham, Coimbatore, India
b_karthigha@cb.amrita.edu
3 Associate Professor, Amrita Vishwa Vidyapeetham, Coimbatore, India
m_jayakumar@cb.amrita.edu
Abstract. In this paper, MOSFET modeling for millimeter wave integrated cir-
cuits is discussed. High frequency MOSFET is built using BSIM3v3 as intrinsic
core and the parasitics due to HF are designed as extrinsic subcircuit. The pro-
posed methodology is then used in designing a low power, mm-wave CMOS
low noise amplifier. The operation of the circuit is simulated using a circuit si-
mulator. The wideband characteristics are verified by implementing the LNA
circuit with and without composite model.
Keywords: CMOS millimeter-wave integrated circuits, LNA-low noise am-
plifier, high frequency (HF) behaviour, composite model, Bsim3v3.
1 Introduction
Millimeter wave WLAN is getting popular due to the enhancement band of 7GHz in
the frequency of 53-67 GHz. The major advantage of this band is its high oxygen at-
tenuation making it interface free for short range indoor WLAN application. This re-
quires realization of semiconductor technology band devices operating at millimeter
waves. Also CMOS implementation promises higher level of integration, low cost and
low power consumption. The advanced performance of MOSFET could be used for
high frequency circuit design in view of a system-on-a-chip realization. To acquire an
efficient design environment, accurate MOSFET model working at millimeter wave
frequencies is essential.
An approach to model a MOSFET at HF is to build a subcircuit based on the in-
trinsic MOSFET that has been modelled for low-frequency analog applications. With
the added parasitic components at the gate, at the drain, at the source and at the sub-
strate, these models could be made to work at millimeter wave frequencies. Once the
MOSFET model is designed, then the working could be examined by using the model
to construct LNA circuit.
In this paper the section 2 describes about the concepts of MOSFET modeling
at HF and designing of composite MOSFET model. Section 3 addresses the design
298 Adhira Raj, B. Karthigha, and M. Jayakumar
parameters of LNA and issues in selecting a multistage low noise amplifier. To dem-
onstrate the effectiveness of the proposed MOSFET model, Section 4 describes the
implementation of the composite model in LNA and finally section 5 illustrates the
simulation results obtained from MICRCAP 10 circuit simulator.
2 MOSFET Modelling
2.1 AC Small-Signal Modeling
In this, a four-terminal MOSFET can be divided into two portions: intrinsic part and
extrinsic part. The concepts of equivalent circuits representing both intrinsic and ex-
trinsic components in a MOSFET are analyzed to obtain a physics-based RF model.
2.2 Modeling of the Intrinsic MOSFET
A BSIM3v3 model as the intrinsic component is selected. BSIM3v3 has been widely
accepted as a standard CMOS model. This model is mainly selected as it is found to
be more advantageous, as the Drain current could easily be modified by current equa-
tion and this model mainly comprises of an accurate capacitance model.
2.3 Subcircuit Model
As we discussed above, a MOSFET contains many extrinsic components such as gate
resistance, source/drain series resistance, substrate resistance and capacitance, and
gate overlap capacitance.
Fig. 1. Composite MOS Model
3 LNA Architecture
3.1 Multistage LNA Architecture
For many applications such as the automatic gain control in WLAN receivers with
single signal path, variable gain of amplifiers is needed. In this paper, we propose a
Designing and Modeling of CMOS Low Noise Amplifier 299
millimeter wave CMOS multistage LNA. A multistage LNA is selected over a single
stage LNA. LNA is a critical block in radar systems. The resistive loss is to be mini-
mized thus using interconnection lines and the inductors. The LNA consists of four
cascaded common-source stages. In the resistive-feedback second stage, the peaking
inductor and the gate-source capacitance of transistor M2 were series resonance at the
geometric mean frequency. LNA based on inductive-source cascade. This proposed
architecture provides gate-to-source and source-to-body parasitic capacitances in two
stage gain cascaded architecture.
4 Experimental Results
The composite MOSFET model suitable for high frequency IC is done. Using the
composite MOSFET model, a multistage LNA is constructed and simulated using
circuit simulator MICROCAP-10. The simulated results depicts that the gain from 4.5
dB increases to 12 dB with the inclusion of the composite MOS model. Similarly the
composite MOS model is included in a common source amplifier whose gain from 20
dB increases to 40 dB. The amplifier without subcircuit is shown in figure 2 and am-
plifier with subcircuit is shown in figure 3.
Fig. 2. Amplifier gain without subcircuit Fig. 3. Amplifier gain with subcircuit
The comparison of multistage and single stage low noise amplifier is done. It is
found that a increase in gain is found when using a multistage circuitry. The Table 1
shows the comparative study results of different low noise amplifier design with sub-
circuit and without subcircuit. The comparison considering single stage and multi-
stage is also shown and thus showing that the gain with multistage is more efficient
than single stage low noise amplifier.
Table 1. Comparitive Study on Lna
Parameters Multi -stage LNA Multi -stage LNA Single -stage LNA
CMOS technology .18μm .13μm .13μm
GAIN(dB)without
composite MOS model
11.9
(at 53GHz)
6.03
(at100 MHz)
6.04
(at15 GHz)
GAIN(dB)with
composite MOS model
19.9
(at60 GHz)
12.9
(at60 GHz)
6.5
(at60 GHz)
300 Adhira Raj, B. Karthigha, and M. Jayakumar
5 Conclusion
The modelling of both intrinsic and parasitic components in MOSFETs is crucial to
describe the HF behaviour of MOS devices operated at GHz frequencies. The brief
discussion of the MOSFET modeling at millimetre wave frequency is done. The LNA
design consisting of multistage is found to have better gain and with the inclusion of
the designed composite MOS model it is found that the circuit gives better perform-
ance in the aspect of gain. Thus a high frequency MOSFET model for low noise am-
plifier is designed and discussed. This work can be extended to study the distortion
behaviour of MOSFET working at millimetre wave frequencies.
References
1. Rezaul Hasan, S.M.: Analysis and Design of a Multistage CMOS Band-Pass Low-Noise
Preamplifier for Ultrawideband RF Receiver. IEEE Transactions on Very Large Scale
Integration (vlsi) Systems, TVLSI.2009.2014166 (2009)
2. Choi, W., Jung, G., Kim, J., Kwon, Y.: Scalable Small-Signal Modeling of RF CMOS FET
based on 3-D EM-Based Extraction of Parasitic Effects and its application to Millimeter-
Wave Amplifier Design. IEEE Transaction on Microwave Theory and Techniques 57(12)
(2009)
3. Cheng, Y., Jamal Deen, M., Chen, C.-H.: MOSFET Modeling for RF IC Design. IEEE
Transaction Electronic Devices, TED.2005.850656 (2005)
4. Shigematsu, H., Hirose, T., Brewer, F., Rodwell, M.: Millimeter-wave CMOS circuit De-
sign. IEEE Tran. Microwave Theory Tech. 53(2) (2005)
5. Yang, M.T., Ho, P.P.C., Wang, Y.J., Yeh, T.J., Chi, Y.T.: Broadband Small-Signal Model
and Parameter Extraction For deep sub-micron MOSFETs valid up to 110GHz. In: IEEE
Radio Frequency Integrated Circuits Symposium (2003)
6. Ou, J.J., et al.: CMOS RF modelling for GHz communication ICs. In: Symposium on VLSI
Technology Digest of Technical Papers, pp. 94–95 (1998)
7. Zirath, H., Ferndahl, M., Motlagh, B.M., Masud, A., Angelov, I., Vickes, H.O.: CMOS de-
vices and circuits for microwave and millimeter wave applications. In: Proceedings of Asia-
Pacific Asia Microwave conference (2006)
8. Shahriar Rashid, S.M., Ali, S.N., Roy, A., Rashid, A.B.M.H.: A 36.1GHz Single stage Low
Noise Amplifier using 0.13μm CMOS Process. World Congress on Computer and Informa-
tion (2009)
9. Yu, Y.-H., Yang, Y.-S., Chen, Y.-J.E.: A Compact Wideband CMOS Low Noise Amplifier
With Gain Flatness Enhancement. IEEE journal of solid-state circuits 45(3) (2010)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 301–304, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Designing Dependable Business Intelligence Solutions
Using Agile Web Services Mining Architectures
A.V. Krishna Prasad1, S. Ramakrishna2, B. Padmaja Rani3,
M. Upendra Kumar4, and D. Shravani5
1 Research Scholar Computer Science S.V. University Tirupathi A.P. India
kpvambati@gmail.com
2 Professor Computer Science S.V.University Campus Tirupathi A.P. India
drsramakrishna@yahoo.com
3 Associate Professor CSE JNTU CEH Hyderabad A.P. India
padmaja_jntuh@yahoo.co.in
4 Research Scholar JNTU Hyderabad A.P. India
uppi_shravani@rediffmail.com
5 Research Scholar Rayalaseema University Kurnool A.P. India
sravani.mummadi@yahoo.co.in
Abstract. Next generation Business Intelligence web application development
uses integrated and intensified technologies like Web 2.0 architectures, Agile
Modeling, and Service-orientation (or Web Services). Applying Web Services
Mining strategies to Agile Modeled Web architectures will eventually provide
valuable insights to Business Intelligence users, Operational Business Decision
makers, and more importantly Web application architects. These insights are
important in maintenance of these developed applications and also in their scal-
ability purposes. Our research focuses on applying Mining for Software (or
Web) Engineering for designing dependable solutions for these integrated tech-
nologies, which will eventually improve the Web Engineering process in terms
of architecture, its security, requirements etc. In this paper, we discuss about
our Mining approach for Business Intelligence to improve insights of Web En-
gineering applications. We validate our approach with a suitable exemplar.
Keywords: Web Mining, Business Intelligence, Agile Modeling, Web 2.0
architectures, Web Services Mining.
1 Introduction to Next Generation Business Intelligence Web
Applications
Now a days, most of the Business Intelligence applications are developed as Web
based applications with little Web Engineering principles used in developing them.
Next generation Business Intelligence applications development are using an integra-
tion of specific technologies like Web 2.0 architectures, Agile modeling, and Service-
orientation (or Web Services).[1] Our research is based on the premise that, Applying
Web Services Mining approaches to Agile Modeled Web 2.0 architectures will even-
tually provide valuable insights to Business Intelligence users, Decision makers, and
302 A.V. Krishna Prasad et al.
importantly Web application architects. Our research focuses on designing technical
solutions for these integrated technologies, which will eventually improve the Web
Engineering process in terms of architectures security requirements. When all these
different technologies are integrated together for the development of Business Intelli-
gence web based applications, it creates many research challenges pertaining to in-
sights of decision making regarding architectures of the developed application, its
inherent security of that architecture or its requirements etc. These insights are re-
quired for maintenance of this application or in its future scalability issues. Moreover,
Business Intelligence has to be shifted from enhancing the data warehousing and data
mining techniques such as OLAP (Online Analytical Processing), OLAM (Online
Analytical Mining), multi dimensional modeling, design methodologies, optimization,
indexing and clustering techniques, to how to securely protect these knowledge capi-
tals from being tampered with by unauthorized use. [9]
Agile Modeling. Agile modeling embraces change as a part of the software develop-
ment process.[2] In most approaches, change is usually considered a bad word. Agile
developers work in pairs, create many prototypes of their solutions, and incorporate
user’s feedback throughout the entire process. Agile software development has en-
couraged developers to tailor their methods to meet their specific needs. Agile model-
ing using Unified Modeling Language is geared towards small development projects
with tight deadlines, like building Web front ends.
Web 2.0 Architectures. The relationship between Web 2.0 design patterns, models,
and architecture artifacts are based on Web 2.0 technologies like search engine
optimization, web services, wikis etc. to name a few for our consideration. Models
guides Reference architectures and finally specialized architectures refines reference
architectures, accounts for domain specific requirements, and also enables solution
patterns. [3]
Web Services Architectures. Service Oriented Architectures (and their implementa-
tions Web Services) uses a series of independent services that can communicate busi-
ness logic with one another. These services can be used independently or together to
form business platforms that come together to form business platforms that come
together and provides value.
2 Mining Approaches for These Business Intelligence Applications
Mining Agile Architectures. Agile software development methods are used to build
secure systems. There are different methods defined in agile development as extreme
programming (XP), scrum, feature driven development (FDD), test driven develop-
ment (TDD), etc. Agile processing includes the phases as agile analysis, agile design
and agile testing. These phases are defined in layers of Model Driven Architecture
(MDA) to provide security at the modeling level which ensures that “security at the
system architecture stage will improve the requirements for that system”.
Mining Web 2.0 architectures. Traditionally, to mine web 2.0 architectures in gen-
eral, we use the methodology for mining patterns from examples, hence capture the
knowledge, and then construct models and architecture based on the commonalities in
Designing Dependable Business Intelligence Solutions 303
the patterns. Design patterns are micro architectures that have proved to be reliable,
easy to implement and robust. Three types of Design Patterns i.e. Creational Design
Patterns, Structural Design Patterns and Behavioral Design Patterns. Design Patterns
are described by listing the intents, motivations, applicability, structure (UML dia-
grams), participants, collaborations, consequences, implementation details are known
as related patterns. The structure of the patterns is represented in Graphs or Matrices
(Abstract Class Matrix, Generalization Matrix, Association Matrix etc). The pattern
descriptions are easy to modify to suit the needs of users by using Design Pattern
Markup Language. Reality Mining is a mashup pattern in terms of MIT which states
that, it is the collection of machine-sensed environmental data pertaining to human
social behavior.
Web Services Mining Architectures. Web Service mining is a search process aim-
ing at the discovery of interesting and useful compositions of existing web services.
Recall is the fraction of relevant services in the collection that were returned by the
system and precision is the fraction of the returned results that are relevant. The Busi-
ness Process Execution Language (BPEL) deals with Web service composition and
attempts to solve the problem of composing a number of web services into a business
process. In Business Process Query, there is a need to interact with web services. Web
services interaction mining provides three levels of abstraction that represent three
complementary web services. The levels are Web service Operations level, Web ser-
vice Interactions level and web service work flow level.
Designing Dependable Solutions for Business Intelligence. Software Engineering
problems must be treated by both theoretical and empirical methodologies. The for-
mer is characterized by abstract, inductive, mathematics-based, and formal-inference
centered studies; while the latter is characterized by concrete, deductive, data-based,
and experimental-validation-centered studies. We propose to build a qualitative or
descriptive model along with appropriate notation or tool for providing specific solu-
tions with validations of a case study. Dependability attributes can be seen from dif-
ferent perspectives, depending on the application. Eventually dependability is in-
tended to prevent errors from becoming failures.
Insights of Web Engineered Business Intelligence Applications. General Business
objectives and their functionality for Rich Security Model that users can administer
are: Provide more effective mechanisms to move work between business entities,
such as self-service for customers or partners or enabling outsourcing by providing
business partners a collaborative environment or business data on an extranet. The
valuable insights from this approach include ease of use, scalability, disconnected
from processes for these developed applications, improved customer satisfaction,
increased business agility, reduced time to market, increased revenue and operational
efficiency and improvements.
Implementations and Validations. Software Engineering for Web (Web Engineer-
ing) covers the definition of processes, techniques and models suitable for its envi-
ronment to guarantee quality of results. An important design artifact in any software
development project is the Software Architecture. Software Architecture’s important
part is the set of architectural design rules. A primary goal of the architecture is to
capture the architecture design decisions. An important part of these design decisions
304 A.V. Krishna Prasad et al.
consists of architectural design rules. In an MDA (Model-Driven Architecture) con-
text, the design of the system architecture is captured in the models of the
system. MDA is known to be layered approach for modeling the architectural design
rules and uses design patterns to improve the quality of software system. And to
include the security to the software system, security patterns are introduced that offer
security at the architectural level. Moreover, agile software development methods are
used to build secure systems. We had implemented various case studies like Web
Services Mining Dashboard application CRM application with spatial capabilities as
primary work. Later on we worked on a case study of Design of Agile Modeled
Web Services design for secure stock exchange, with focus on mined security archi-
tectures insights. For details of these implementations, please refer to the website
http://sites.google.com/site/upendramgitcse
3 Conclusions
In this paper we discussed mining approach for Business Intelligence to improve
insights of Web Engineering applications. Future work includes developing a formal
security plan for every connected computer. Performing security inspections of
requirements and specifications. Utilizing high-security programming language such
as “E”.
References
1. Czernicki, B.: Silverlight 4 Business Intelligence Software. Apress, USA (2010)
2. Mordinyi, R., Kuhn, E., Schatten, A.: Towards an Architectural Framework for Agile Soft-
ware Development. In: IEEE 17th International Conference and Workshop on Engineering
of Computer Based Systems (ECBS), pp. 276–280 (2010)
3. Governor, J., Hinchcliffe, D., Nickull, D.: Web 2.0 Architectures. Oreilly Publishers, USA
(2009)
4. Arsanjani, A., Zhang, L.-J., Ellis, M., Allam, A., Chennabasavaiah, K.: S3 A Service-
Oriented Reference Architecture. In: IEEE IT Pro., pp. 10–17 (May/June 2007)
5. Chivers, H., Paige, R.F., Ge, X.: Agile Security Using an Incremental Security Architecture.
In: Baumeister, H., Marchesi, M., Holcombe, M. (eds.) XP 2005. LNCS, vol. 3556,
pp. 57–65. Springer, Heidelberg (2005)
6. Yee, C.G., Radha Krishna Rao, G.S.V.: An Introductory study on Business Intelligence
Security, pp. 204–217. Idea Group Inc., USA (2007)
7. Fernandez, E.B., Yoshika, N., Washizaki, H., Jurjens, J., VanHilst, M., Pernul, G.: Using
Security Patterns to Develop Secure Systems. IGI Global, pp. 16–31 (2011),
doi:10.4018/978-1-61520-837-1.ch002
8. Chung, W.: Designing Web-based Business Intelligence Systems: A Framework and Case
Studies. In: DESRIST, California, CA, USA, February 24-25, pp. 147–171 (2006)
9. Chivers, H., Paige, R.F., Ge, X.: Agile Security Using an Incremental Security Architecture.
In: Baumeister, H., Marchesi, M., Holcombe, M. (eds.) XP 2005. LNCS, vol. 3556,
pp. 57–65. Springer, Heidelberg (2005)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 305–311, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Modified Continuous Particle Swarm Optimization
Algorithm for Uncapacitated Facility Location Problem
Sujay Saha1, Arnab Kole2, and Kashinath Dey3
1 Assistant Professor, CSE Department, Heritage Institute Of Technology, Kolkata, India
sujay.saha@heritageit.edu
2 M. Tech Student, CSE Department, Heritage Institute Of Technology, Kolkata, India
arnab.kole@heritageit.edu
3 Associate Professor, CSE Department, University Of Calcutta, Kolkata, India
kndey55@rediffmail.com
Abstract. A continuous version of particle swarm optimization (CPSO) is em-
ployed to solve uncapacitated facility location (UFL) problem which is one of
the most widely studied in combinatorial optimization. The basic algorithm had
already been published in the Research Article “A Discrete Particle Swarm Op-
timization Algorithm for Uncapacitated Facility Location Problem” [1]. But in
addition to that, the algorithm is slightly modified here to get better result in a
lesser time. To make a reasonable comparison, the same benchmark suites that
are collected from OR-library [6] are applied here. In conclusion, the results
showed that this modified CPSO algorithm is slightly better than the published
CPSO algorithm.
Keywords: Swarm Intelligence, Continuous Particle Swarm Optimization,
UFL, Inertia Factor, Acceleration Coefficient.
1 Introduction
Efficient supply chain management has led to increased profit, increased market
share, reduced operating cost and improved customer satisfaction for many busi-
nesses. One strategic decision is related to physical distribution structure in supply
chain management including locating facilities and allocating customers to them is
facility location. Facility Location, also known as Location Analysis, is a branch of
Operation Research [4] concerning itself with mathematical modeling and solution
of problems concerning optimal placement of facilities in order to minimize the
transportation costs, avoid placing hazardous materials near housing, outperform
competitors’ facilities, etc. We are trying to solve a ‘Uncapacitated Facility Location
Problem’ which finds the number of facilities to be established and specify those
facilities such that the total cost will be minimized considering the total cost means a
fixed cost of setting up a facility in a given site and a transportation cost of satisfying
the customer requirements from a facility. We maintain the constraints that there is
no limit of capacity for any candidate facility and the whole demand of each cus-
tomer has to be assigned to one of the facility.
306 S. Saha, A. Kole, and K. Dey
The organization of the paper is as follows: in Section 2, a basic idea is given for
Particle Swarm Optimization Technique. Section 3 reports about the definition of the
Uncapacitated Facility Location Problem along with the constraints. Section 4
describes the published CPSO algorithm. Section 5 describes the modified CPSO
algorithm for solving the Uncapacitated Facility Location problem. Section 6 provides
the comparison results of these two algorithms. Finally, Section 7 presents the conclu-
sion driven.
2 Particle Swarm Optimization (PSO)
The class of complex systems sometimes referred to as swarm systems is a rich
source of novel computational methods that can solve difficult problems efficiently
and reliably. When swarms solve problems in nature, their abilities are usually attrib-
uted to swarm intelligence [3]; perhaps the best-known examples are colonies of so-
cial insects such as termites, bees, and ants. One of the best-developed techniques of
this type is particle swarm optimization (PSO). In PSOs, which are inspired by flocks
of birds and shoals of fish, a number of simple entities, the particles, are placed in the
parameter space of some problem or function, and each evaluates the fitness at its
current location. Each particle then determines its movement through the parameter
space by combining some aspect of the history of its own fitness values with those of
one or more members of the swarm, and then moving through the parameter space
with a velocity determined by the locations and processed fitness values of those other
members, along with some random perturbations. The members of the swarm that a
particle can interact with are called its social neighborhoods. Together the social
neighborhoods of all particles form a PSOs social network. More precisely, in the
canonical version of PSO, each particle is moved by two elastic forces, one attracting
it with random magnitude to the fittest location so far encountered by the particle, and
one attracting it with random magnitude to the best location encountered by any of the
particle’s social neighbors in the swarm. If the problem is N dimensional, each parti-
cle’s position and velocity can be represented as a vector with N components (one for
each dimension). Starting with the velocity vector, v = (v1. . . vN), each component, vi,
is given by:
))(())(()()1( 2211 txxRtxxRtvtv ipisii ii −+−+=+
ψψω
(1)
where i
s
x is the ith component of the best point visited by the neighbors of the
particle, )( tx i is the ith component of the particle’s current location, i
p
xis the ith
component of its personal best, R1 and R2 are two independent random variable
uniformly distributed over [0, 1] is a constant known as the inertia weight, and
21
ψ
ψ
and are two constants, known as the acceleration coefficients, which
control the relative proportion of cognition and social interaction in the swarm. The
same formula is used independently for each dimension of the problem, and synchro-
nously for all particles. The position of a particle is updated every time step using the
equation:
)1()()1( ++=+ tvtxtx iii . (2)
A Modified Continuous Particle Swarm Optimization Algorithm 307
The next iteration takes place after all particles have been moved. Eventually, the
swarm as a whole, like a flock of birds collectively foraging for food, is likely to
move close to the best location.
3 UFL Definition
In a UFL problem, there are a number of customers, m, to be satisfied by a number of
facilities, n. Each facility has a fixed cost, fcj. A transport cost, cij , is accrued for serv-
ing customer, i, from facility, j. There is no limit of capacity for any candidate facility
and the whole demand of each customer has to be assigned to one of the facilities. We
are asked to find the number of facilities to be established and specify those facilities
such that the total cost will be minimized. The mathematical formulation of the prob-
lem can be stated as follows:
(3)
(3
subject to:
(4)
(5)
where i = 1. . . m; j = 1, . . . , n; xij represents the quantity supplied from facility i to
customer j; yj indicates whether facility j is established (yj = 1) or not (yj = 0). Con-
straint (4) makes sure that all customers demands have been met by an open facility
and (5) is to keep integrity. Since it is assumed that there is no capacity limit for any
facility, the demand size of each customer is ignored and therefore (3) established
without considering demand variable.
4 Earlier Related Work for UFL Problem
According to Sevkli and Guner [1], CPSO considers each particle has three key vectors:
position (Xi), velocity (Vi), and open facility (Yi). Xi = ],.......,,[ 21 inii XXX de-
notes the ith position vector in the swarm, where ik
X is the position of the ith particle
with respect to the kth dimension. Similarly Vi = ],.....,,[ 21 inii VVV is the ith ve-
locity in the swarm, where ik
Vis the velocity of the ith particle with respect to the kth
dimension. Yi represents the opening or closing facilities based on the position vector
(Xi), Yi = ].....,,[ 21 inii YYY where ik
Y represents opening or closing kth facility
of the ith particle. For an n-facility problem, each particle contains n number of dimen-
sions. Initially the positions and velocities are generated as continuous uniform random
variables using the following rules:
xij = xmin + (xmax − xmin) × r1. (6)
vij = vmin + (vmax − vmin) × r2.
∑∑ ∑
== =
+= m
i
n
j
n
j
jjijij yfcxcZ
11 1
.)..min(
.1
1
minix
n
j
ij ∀=
∑
=
jij yx ≤≤0
{}
.1;0∈
j
y
308 S. Saha, A. Kole, and K. Dey
where xmin = -10.0, xmax = 10.0, vmin = -4.0, xmax = 4.0, ]1,0[, 21 ∈rr . The
position vectors do not represent a candidate solution to calculate the total cost (fit-
ness value). In order to create a candidate solution the position vector is converted to
a binary variable ii XY ←, which is also a key element of a particle. This conver-
sion is done using the following formula:
⎣
⎦
)2(mod
ii xy =. (7)
For example, if the position value of one particle with respect to one dimension is -
7.47, then the corresponding open facility vector will be
⎣
⎦
⎣⎦⎣⎦
147.1)2(mod47.7)2(mod47.7 ===− (8)
After generating the position vectors and open facility vectors, the total cost of each
particle is calculated as per equation (3) and initially set these cost as local best or
personal best (Pti) for each particle and the minimum of these personal bests is set as
global best (Gt) of the swarm. Next, the velocity of each particle is updated as per
equation (1) and the position is updated as per equation (2) where
ω
is the inertia
factor used to control the impact of the previous velocities on the current one. After
updating the position value for all particles, the corresponding open facility vector can
be determined to start a new iteration if the predetermined stopping criterion is not
yet met. The stopping criteria can be either getting an optimal solution or reaching
the maximum number of iterations chosen for obtaining the result in a reasonable
CPU time.
5 Modifications
To reduce the execution time & the cost, the above algorithm is modified in the
following manners: first, to generate the position and the velocity in equation (6), if
r1 = 0.5 then xij = 0 and similarly if r2 = 0.5 then vij = 0. So instead of doing the addi-
tion and multiplication which take some time, the values of both xij and vij can be set
to zero for the above said value of r1 and r2. Secondly, instead of keeping the fixed
inertia factor, the value of
ω
can be varied from 0.9 to 0.4 in various iterations to get
the lesser optimal cost. So the modified CPSO algorithm is as follows:
Modified CPSO Algorithm for UFL
Begin
Initialize positions (population) randomly For each particle as:
Generate r1 randomly in [0, 1]
If r1 = 0.5 then xij = 0, else
x
ij = xmin + (xmax − xmin) × r1
End
Initialize velocities randomly For each particle as:
Generate r2 randomly in [0, 1]
If r2 = 0.5 then vij = 0, else
vij = vmin + (vmax − vmin) × r2
End
A Modified Continuous Particle Swarm Optimization Algorithm 309
Calculate open facility vector as per equation (7)
Calculate fitness value using open facility vector as per equation (3)
Set to position vector and fitness value as personal best (Pti )
Select the best particle and its position vector as global best (Gt)
End
Do {
For each particle
Update velocity as per equation (1)
Update position as per equation (2)
Find open facility vectors & calculate the fitness value using open facility
vector
Update personal best (Pti )
Update the global best (Gt) value with position vector
End
Decrement the value of
ω
by: (0.9 – 0.4) / total no. of iterations
} While (Maximum Iteration is not reached)
Table 1. An illustration of deriving open facility vector from position vector for a 4-customer
3-facility problem
Table 2. An example of 3-facility to 4-customer
Facility Locations 1 2 3
Fixed Cost 15 8 3
1 12 3 7
2 2 8 5
3 4 6 14
Customers
4 9 1 10
Considering the 3-facility to 4-customer problem shown in Table 1, the total
cost of open facility vectors can be calculated as follows:
Total Cost = {open facilities fixed cost (fcj) + min (cost of supply from open
facilities to customers i[cij] = {(15+3) + min (12, 7) + min( 2, 5)
+ min (4, 14) + min (9, 10)} = {18+2+7+4+9} = {40}.
6 Comparison of Results
The modified CPSO algorithm is coded with C language in Linux Platform and run
on an Intel Core 2 Duo 2.4 GHz Laptop with 1GB memory. The results are shown in
the following table:
Particle dimension (k)
ith particle vectors 1 2 3
Position Vector (Xi) 3.8 -0.93 5.42
Open Facility Vector (Yi) 1 0 1
310 S. Saha, A. Kole, and K. Dey
Table 3. Experimental Results gained with modified CPSO algorithm
Benchmark
As Per Earlier CPSO
Algorithm
As Per Modified CPSO
Algorithm
Problems Size (m x n) Optimal
Cost ACPU Time Optimal Cost ACPU
Time
Cap 71 16 x 50 932615.75 0.1218 932597.00 0.9700
Cap 72 16 x 50 977799.40 0.1318 977779.00 0.7500
Cap 73 16 x 50 1010641.45 0.1865 1010619.00 0.5900
Cap 74 16 x 50 1034976.98 0.1781 1034956.00 0.5300
Cap 101 25 x 50 796648.44 0.8818 796626.00 1.3800
Cap 102 25 x 50 854704.20 0.7667 854682.00 1.2400
Cap 103 25 x 50 893782.11 0.9938 893758.00 0.9900
Cap 104 25 x 50 928941.75 0.6026 928918.00 0.7700
Cap 131 50 x 50 793439.56 3.6156 794137.00 2.1300
Cap 132 50 x 50 851495.33 3.5599 853149.00 1.8900
Cap 133 50 x 50 893076.71 3.7792 894072.00 1.8900
Cap 134 50 x 50 928941.75 3.3333 928918.00 1.2400
7 Conclusion
In this paper, CPSO algorithm is applied to solve UFL problems. We have slightly
modified this algorithm for initializing the particle’s positions and velocities for a
particular value of the random numbers r1 and r2 to reduce the average CPU time.
Secondly, instead of keeping a fixed ω value, we have varied the value in the range of
0.9 to 0.4 over the total number of iterations. This modified algorithm has been tested
on 12 files of OR-Library and the better results either in terms of time or in terms of
optimal cost are obtained.
References
1. Guner Ali, R., Mehmet, S.: A Discrete Particle Swarm Optimization Algorithm For Unca-
pacitated Facility Location Problem. Hindawi Publishing Corporation Journal of Artificial
Evolution And Applications 2008, Article ID 861512, 9 (2008), doi:10.1155/2008/861512
2. Kennedy, J., Eberhart, R.C., Shi, Y.: Swarm Intelligence. Morgan Kaufmann, San Francisco
(2001)
3. Ghosh, D.: Neighborhood search heuristics for the uncapacitated facility location problem.
European Journal of Operational Research 150(1), 150–162 (2003)
A Modified Continuous Particle Swarm Optimization Algorithm 311
4. Eberhart, R.C., Kennedy, J.: New optimizer using particle swarm theory. In: Proceedings of
the 6th International Symposium on Micro Machine and Human Science (MHS 1995),
Nagoya, Japan, October 1995, pp. 39–43 (1995)
5. Beasley, J.E.: OR-Library (2005),
http://people.brunel.ac.uk/mastjjb/jeb/info.html
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 312–316, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Design of Hybrid Genetic Algorithm with Preferential
Local Search for Multiobjective Optimization Problems
J. Bhuvana and C. Aravindan
Department of Computer science and Engineering,
SSN College of Engineering, Chennai, India
{bhuvanaj,aravindanc}@ssn.edu.in
Abstract. Evolutionary algorithms are used to obtain the optimal solutions for
varieties of engineering problems. The performance of evolutionary algorithms
can be enhanced by integrating them with local search methods. The idea is to
combine the best of both global and local optimization approaches to perform a
better exploration of search space. This paper presents a preferential hybrid
evolutionary algorithm, where the gradient descent method is integrated into the
NSGA-II. This new algorithm has been verified on a set of multiobjective
benchmark problems using four different performance metrics. The results
show that the proposed algorithm brings out the optimal solutions with better
diversity and closeness to the known optimal solutions than NSGA-II and also
consumes less time than traditional hybrid algorithm.
Keywords: Hybrid GA, Gradient Descent, Preferential local search.
1 Introduction
Memetic or hybrid genetic algorithms combine the best of both global and local
search techniques to improve the quality of the solutions. The ability of the local
search to produce a local optimal solution when integrated with global search method
using genetic algorithms will generate diverse set of pareto optimal solutions and also
prevent premature convergence [5]. A hybrid GA with preferential local search has
been proposed in this paper, which addresses the issues such as where to integrate
local optimization and how solutions are favored to undergo the local search to reduce
the computational time of any hybrid approach. Experimental results show that this
approach out-performs the traditional hybrid algorithm in terms of metrics and
computational time.
The proposed Hybrid GA with Preferential Local search is based on NSGA-II as
the global search and Gradient Descent as the local search method. NSGA-II [8] is
one of the most popular elitist genetic algorithms [2] uses ranking method based on
non dominated sorting. Gradient Descent local search minimizes a function f(x)
where x ε R, by computing a new solution using xk+1 = xk +αkdk; k = 0, 1,…, where
the k is the step size[12] and dk is the descent direction given by -Δf(xk).
Varieties of Hybrid algorithms have been reported in the literature [3]. Review
about most commonly used hybrid architectures is available in [6]. Adaptive Local
Design of Hybrid Genetic Algorithm with Preferential Local Search 313
Search (ALS) is applied by combining weighted fitness and restricted mutation in [1].
New weighted fitness is assigned to Ns superior individuals and applies restricted
crossover and mutation. Local search based EMO in [11] is a hybrid algorithm with
Achievement Scalarizing Functions (ASF) and local search method. Local search
begins from a reference point, minimizes the associated ASF. Two hybridization
schemes are discussed in [7] such as, GA with LS, and GA then LS. First hybrid
scheme applies LS to solutions of each generation of GA where the second one
applies LS to solutions obtained from GA. And also recommends the second approach
as the preferable scheme.
2 Algorithm Description
The idea is to propose a preferential hybrid approach that identifies the optimal
solutions by combining local search and global search methods. Whenever new
individuals are generated, they undergo a limited local search. As generation
progresses, when the individuals survive across generation further local search is
applied on them. Thus the local search iteratively deepens on potential candidates
thereby introducing elitism in local search.
In designing a preferential hybrid approach, this paper proposes a new method of
combining the NSGA-II and the Gradient Descent local search method. The basic
assumption here is the objective functions considered are differentiable. The random
values populate the initial population and mating pool is created using binary
tournament selection. SBX cross over with polynomial mutation are used to generate
new offsprings. A limited local search is applied on the offsprings and will exploit the
neighborhood space and prepare them to compete with their parents for survival. To
emphasize the potential solutions, preferential local search is applied on the surviving
parents. This approach is entirely different from traditional memetic or hybrid
algorithms and addresses the issue of termination condition in a way. The pseudo
code for the proposed hybrid algorithm with preferential Local search is given in
Table 1.For local search, the weighted sum method is employed to convert the
multiple objectives into a single objective function, where the weights are uniformly
distributed to all the objectives.
Table 1. Pseudo code for Hybrid GA with Preferential Local search (HPLS)
1. Initialize population with random solutions.
2. Repeat until termination condition is met.
3. Evaluate the fitness.
4. Sort the population according to non domination. Assign ranks accordingly.
5. Perform the selection.
6. Apply preferential local search on elitist parent solutions
7. Apply crossover and Mutation to generate new offsprings.
8. Apply limited local search on all offsprings.
9. Update the population with values obtained from Local search and best parents
314 J. Bhuvana and C. Aravindan
3 Implementation and Experiments
The Hybrid NSGA-II with preferential local search (HPLS) was implemented in C,
with a population size of 40 for 7000 generations. The crossover and mutation
probability were set to 1 and 1/(number of variables) respectively. Same initial
population was fed for three approaches, HPLS, Hybrid NSGA-II with local search
(HLS) and NSGA-II for a single run and 10 such runs were accomplished.
The average of performance metrics were computed and used for further statistical
testing. The gradient for one descent step with a step size of 0.01 was computed. Six
unconstrained bi-objective test problems such as SCH, BINH, ZDT1, ZDT2, ZDT3
and ZDT6 [4, 9] have been used for establishing the performance of the HPLS. The
performances of HPLS, HLS and NSGA-II over the benchmark problems were
compared on the basis of two major criteria such as, Closeness of obtained solutions
to known Pareto optimal solutions and the diversity of solutions along the Pareto
optimal front. Four different metrics such as, Error Ratio(ER), Generational Distance
(GD), Spread and Hyper volume ratio (HVR) have been used to compare the
performances of three approaches. Time taken by the algorithms have been observed
and used for comparison.
Table 2. Performance Metrics
Proble
m
Algo. ER GD Spread HVR Avg. Time(ms)
ZDT1 HPLS 0.0144 0.1511 0.8960 0.2399 117.898
NSGA-II 0.0250 0.1555 0.9264 0.1822 206.915(HLS)
ZDT2 HPLS 0.0225 0.1503 0.9437 0.2309 138.455
NSGA-II 0.1250 0.1544 0.9966 0.1747 287.227(HLS)
ZDT3 HPLS 0.0150 0.1566 0.0561 0.5736 137.1282
NSGA-II 0.0250 0.1708 0.2681 0.4671 359.080(HLS)
ZDT6 HPLS 0.0175 0.1550 0.6260 0.4173 28.918
NSGA-II 0.0250 0.1650 0.6480 0.3155 46.836(HLS)
SCH1 HPLS 0.0250 0.1548 0.3161 0.0644 14.659
NSGA-II 0.2825 0.1631 0.5513 0.0338 18.327(HLS)
BINH HPLS 0.0755 0.1497 0.0649 0.0100 15.1496
NSGA-II 0.3875 0.1655 0.3068 0.0062 38.3964(HLS)
4 Results and Discussions
The performance metrics in Table 2 state that HPLS outperformed NSGA-II in all the
metrics. When HPLS compared with HLS following observations were made, 51.5%
improvement in the ER, about 2.2% enhancement in GD, 14.5% improvement in
Spread and for HVR there was 18.2% increase in performances. Apart from
performance metrics, HPLS has consumed less than 50% of time consumed by HLS
for 10 fixed steps shown in Table 2. HPLS has yielded more number of non
dominating optimal solutions than other two was inferred from Figure 1. The global
search method when accompanied by preferential local search made a thorough
exploration of the search space than global search alone or the traditional hybrid
method. Paired t test [10] was applied over the four metrics computed across
Design of Hybrid Genetic Algorithm with Preferential Local Search 315
generations. The results confirmed that there was a significant difference between
these algorithms. This inference when associated with the obtained performance
metrics revealed the fact that HPLS and HLS were different and the HPLS behaved
better than the HLS and NSGA-II in terms of closeness to the known optimal
solutions and also generated comparatively good spread of solutions. And the
objective of the new approach, exploring the search space globally and also locally
was henceforth achieved.
Fig. 1. Pareto Fronts of Binh, SCH, ZDT2 and ZDT6
5 Conclusion
This paper has introduced a new elitist approach with reduced computation time by
applying preferential local search integrated with a multiobjective GA. Two local
searches have been used, where the first one was preferentially applied over elitist
parents. Parents surviving in the next generation reveal the fact that they are
prominently good solutions and require to be preserved. A limited second local search
applied on offsprings enable the optimization to exploit and explore the unexplored
areas of search space. This work can be extended further by deciding the termination
conditions for the local search applied. The experimental results show that HPLS not
only taken evolution more closely towards the optimal solutions than the traditional
hybrid approach but also yielded more number of additional non dominant solutions
with less than 50% of computational time.
References
1. Ahn, C.W., et al.: A hybrid evolutionary algorithm for multiobjective optimization. In: Fourth
International Conference on Bio-Inspired Computing (BICTA 2009), pp. 1–5 (2009)
2. Coello, C.A.C.: Twenty years of evolutionary multi-objective optimization: what has been
done and what remains to be done. In: Computational Intelligence: Principles and Practice,
ch. 4, pp. 73–88. IEEE Computational Intelligence Society, Los Alamitos (2006)
316 J. Bhuvana and C. Aravindan
3. Coello, C.A.C.: Evolutionary Multiobjective Optimization: some current research trends
and topics that remain to be explored, vol. 3(1), pp. 18–30. Higher Education Press and
Springer Verlag (2009)
4. van Veldhuizen, D.A., et al.: Multiobjective evolutionary algorithm test suites. In:
Proceedings of the 1999 ACM symposium on Applied computing, Texas, United States,
pp. 351–357 (1999)
5. Elmihoub, T., et al.: Hybrid genetic algorithms - a review. Engineering Letters 13,
124–137 (2006)
6. Grosan, C., Abraham, A.: Hybrid evolutionary algorithms: Methodologies, architectures,
and reviews. In: Grosan, C., Abraham, A. (eds.) Hybrid Evolutionary Algorithms. SCI,
vol. 75, pp. 1–17. Springer, Heidelberg (2007)
7. Ken, H., et al.: Hybridization of genetic algorithm and local search in multiobjective
function optimization: recommendation of GA then LS. In: GECCO 2006: Proceedings of
the 8th Annual conference on Genetic and Evolutionary Computation, pp. 667–674. ACM,
USA (2006)
8. Deb, K., et al.: A Fast Elitist Non-dominated Sorting Genetic Algorithm for Multi-
objective Optimization: NSGA-II. In: Deb, K., Rudolph, G., Lutton, E., Merelo, J.J.,
Schoenauer, M., Schwefel, H.-P., Yao, X. (eds.) PPSN 2000. LNCS, vol. 1917,
pp. 849–858. Springer, Heidelberg (2000)
9. Deb, K.: Multi-Objective Optimization Using Evolutionary Algorithms. John Wiley &
Sons(Asia) Pte Ltd., Singapore (2001)
10. Ross, S.M.: Introduction to Probability and Statistics for Engineers and Scientists.
Academic Press, Singapore (2004)
11. Karthik, S., et al.: Local search based evolutionary multiobjective optimization algorithm
for constrained and unconstrained problems. In: CEC 2009: Proceedings of the Eleventh
conference on Congress on Evolutionary Computation, Norway, pp. 2919–2926 (2009)
12. Salomon, R.: Evolutionary algorithms and gradient search: similarities and differences.
IEEE Transactions on Evolutionary Computation 2(2), 45–55 (1998)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 317–322, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Synergy of Multi-agent Coordination Technique and
Optimization Techniques for Patient Scheduling
E. Grace Mary Kanaga1 and M.L. Valarmathi2
1 Department of Computer science and Engineering,
Karunya University, Karunya Nagar, Coimbatore-641 114,
Tamil Nadu, India
grace@karunya.edu
2 Department of Computer science and Engineering,
Government College of Technology, Coimbatore-641 013,
Tamil Nadu, India
ml_valarmathi@rediffmail.com
Abstract. This paper discusses about the synergy of multi-agent coordination
technique with different optimization techniques viz. integer programming
optimization with Lagrangian Relaxation and a simple heuristic approach with
experience based learning effect in case of patient scheduling problem. The ob-
jective is to achieve good reduction in waiting time of the patients in hospital
while achieving better resource utilization. The proposed methods have been
implemented in JADE and the results are compared with the traditional schedul-
ing techniques. The performance of the proposed methods based on total
weighted waiting time of the patient increases by 15% to 52% when compared
to traditional scheduling techniques. The waiting time of the patient is further
reduced by 5%, when experience based learning effect is incorporated. Optimi-
zation based on simple heuristic method shows that it gives near optimal solu-
tion with drastic reduction in the execution time ie. 90.84% when compared to
integer programming approach.
Keywords: Multi-agents, Coordination technique, Optimization techniques, Pa-
tient scheduling.
1 Introduction
A high amount of complexity is faced by the patient scheduling system in hospitals
and coordinating patients for having an optimum schedule is a difficult task. Depend-
ing on the available resources, the schedules may be more or less congested. In coun-
tries like India demand regularly exceeds the resource availability. According to the
statistics, WHO (World Health Organization) says that India has one doctor per 2500
people. It is well below international standards. However, even in this situation, there
is no effective scheduling is used to reduce the waiting time of the patients and to
improve the resource utilization in the hospital. An efficient agent-based patient sche-
duling system with various optimization techniques is discussed in this paper.
318 E. Grace Mary Kanaga and M.L. Valarmathi
2 Related Work
In many current scheduling approaches focus is only on single units which are centra-
lized. These systems do not account for the distribution and dynamics of the patient
scheduling problem and this has been discussed by Paulussen et al. [1]. Emergencies
or complications lead to disturbances in the schedule and result in variable waiting
times for other patients according to Bartelt et al. [2]. The main challenge of the
patient scheduling problem is that it is distributed and it is a safety critical real time
system.
The hospital scheduling problem is diverse and many papers have addressed differ-
ent aspects of the hospital scheduling problem. Some of the major problems addressed
in literature are, patient/medical appointment scheduling as seen in Vermuelen et al.
[3], and Kaandorp and Koole [4]. Another type is the specialized case of Surgical
Case Scheduling (SCS) found in Pham and Klinkert [5] / Operating Room (OR) sche-
duling as given by Becker and Navarro [6]. Marinagi et al. [7] and Kapamara et al. [8]
discussed about patient scheduling in laboratories and scheduling for treatment of
radiotherapy respectively.
3 Agent-Based Patient Scheduling
In the proposed Agent-based Patient Scheduling (APS) framework, each patient and
resource is represented by an agent and a combinatorial auction is used to produce a
schedule. This framework consists of the Patient Agent (PA), the Resource Agent
(RA) and the Common Agent (CA). CA represents a general physician who assigns
the initial treatment plan. Each patient agent, PAi has a set of tasks T1i,T2i,…,TMi.
These tasks represent the treatment plan prescribed for each patient. Each resource
agent, RAm conducts an auction at a decision point, tc. The patient agents available at
the time of auction participate in the auction. Each patient agent PAi participates in
auctions conducted by the resource agents for the tasks to be performed. The time
slot is allocated to the PA by RA, such that the total weighted waiting time is mini-
mum. Different optimization techniques such as Integer Programming Optimization
with Lagrangean Relaxation(IPLR), a heuristic optimization and an Experience Based
Learning (EBL) effect is incorporated in this APS frame work is discussed in the
following sections.
3.1 Agent-Based Patient Scheduling with IP-LR
The Integer Programming (IP) formulation for agent-based patient scheduling prob-
lem is modeled as follows.
The objective function which minimizes total weighted waiting time is given by,
(1)
is subject to the following constraints:
∑∑
=
+
=
cc
c
i
N
i
Lt
tt
tiOii XTw
1
Synergy of Multi-agent Coordination Technique and Optimization Techniques 319
Constraint 1: To ensure that each task of each patient can finish within an allotted
time slot.
(2)
Constraint 2: Task Precedence constraints of two tasks.
(3)
Constraint 3: Resource capacity constraints to ensure that the capacity of the re-
source is not violated in each timeslot.
(4)
Constraint 4: Patient release date constraints to ensure that the first task cannot be
completed before the patient has been in the hospital for at least the time equal to the
processing of the first task.
(5
)
where wi is the weight based on priority assigned to patient i, Ti the tardiness (waiting
time) of patient i, Xijt is the integer variable equals 1 if task (i,j) is completed at time t,
Yijm is the integer variable equals 1 if task (i,j) is processed on resource m, Pij is the
processing time of task j of patient i and i, j ,m are the indices of the patient, task and
resource respectively.
The above problem formulation is for a centralized approach [9]. Lagrangean
Relaxation (LR) of the resource constraints is done to decompose the problem into
patient level and resource level sub problems in order to solve the problem in a distri-
buted environment.
3.2 Agent-Based Patient Scheduling with Experience-Based Learning
The phenomenon of “learning effect” is investigated in the context of scheduling [10]
and its shows that the Experience Based Learning (EBL) gives better solution. Hence
in this paper this EBL effect is incorporated in the previously explained IPLR method.
In this approach each resource has a Learning Agent (LA) and this learning agent
gains its experience whenever a patient is allocated to that particular resource. The
time required to process a particular task is reduced as the experience gained by it
increases. Based on EBL effect, the completion time of patient i scheduled in the vth
slot in a sequence using experience based learning is given by:
(6)
∑
+
=
∀=
Lt
tt
ijt
c
c
jiX ,,1
∑∑
+
=
+
=++ ∀≤+
Lt
tt
Lt
tt
tjiijijt
c
c
c
c
jitXptX ,,
,1,1
∑∑ ∑∑∑ ==
−+
+===
∀≤×+ ci
i
ij
ci
i
N
i
O
Kj
ijmijt
ptL
Lt
N
i
O
Kj
ijmijt tmYXYX
1
'
}1,min{
1'1
,,1
∑
+
=
∀+≥
Lt
tt
iiijt
c
c
irptX ,
1
320 E. Grace Mary Kanaga and M.L. Valarmathi
where the processing time of patient i scheduled in the vth position in a sequence of
resource j is given as follows:
, (7)
where Pij is the normal processing time without experience and ,1,…,,
1,…, and the experience of the resource at the start of execution of the vth
patient is,
1 e
(8)
where ∑
denotes the experience already possessed by the resource and
0,1 is the amount of experience (percentage of ) provided to the resource
by patient in the th position at the start of its execution. The other variables, bi > 0
denote the linear learning ratio of the patient i, and gi is its learning threshold.
3.3 Agent-Based Patient Scheduling with EBL and Simple Heuristic Approach
Heuristic algorithms called suboptimal procedures or simply, heuristics are designed
to find the best possible solutions with small computational efforts.
Heuristic algorithm
4 Results and Discussion
The proposed methods are implemented in the JADE platform. In order to evaluate
the performance of the proposed methods with other traditional scheduling techniques
Step 1: Obtain the list of patients for each of the resources.
Step 2: Set j=1, y=2 and k=2. Pick the first two patients from the list of patients
of jth resource and schedule them in order to minimize the weighted
sum of total completion time and total tardiness.
Step 3: Check whether the schedule is satisfied for the precedence constraint.
If so,
a)
Update the selected partial solution as the new current solution
otherwise
b)
Repeat the process with the next lowest value of weighted sum of total
completion time.
where the completion time of a patient i processed in vth time slot is
calculated using equation(6).
Step 4: Check whether y exceeds nj
If yes go to Step 6
Otherwise increment k and y by one.
Step 5: Generate k candidate sequences by inserting the first patient in the
remaining patient list into each slot of the current solution. Among these
candidates select the best one with the least partial minimization of the
weighted sum of total completion time and total tardiness and go to Step3.
Step 6: Update the values of i and k and repeat until i=m.
Step 7: Stop.
Synergy of Multi-agent Coordination Technique and Optimization Techniques 321
such as FCFS, MS, LPT and SPT, data instances are created with 10 resources and
varying number of patients. The arrival time and the weight of the patients are gener-
ated randomly. The weight of the patients varies from 1 to 5. Higher weight values are
assigned to the patient in poor health state. The task sequence for each patient is allot-
ted randomly.
4.1 Performance Metrics
There are several performance metrics used for scheduling problems. The objective of
patient scheduling problem is to minimize the patient waiting time and improve the
resource utilization in the hospitals. A good scheduling technique will effectively
minimize the values of all these metrics.
▪Total weighted Waiting Time (∑wiTi) ▪Total Waiting Time ( ∑Ti )
▪ Maximum Waiting Time ( Tmax ) ▪ Maximum Completion Time ( Cmax )
▪Total Completion time ( ∑Ci ) ▪ Total Weighted Completion time (∑wiCi )
4.2 Comparison of Results
The proposed methods have been tested with the data set for 50x10 patient scheduling
problem to study their performance. The chief metric, weighted waiting time of the
APS-IPLR method is compared with the traditional techniques and it is seen that
APS-IPLR performance is similar to FCFS and better than traditional techniques MS,
SPT, LPT. For 50 patients, the total weighted waiting time APS-IPLR is 1542 whe-
reas in other algorithms it is 1600 to 3200 which varies from 5% - 52% of increase in
performance.
(a) (b)
Fig. 1. (a) Comparison of performance metrics (b) Comparison of execution time
Perusal of Fig. 1(a) shows that the waiting time as well as the completion time of
the APS with EBL effect is considerably reduced when compare to IP-LR without
EBL. The weighted waiting time of the patients has reduced up to 4.47% in APS-
EBL. The waiting time as well as the completion time of the heuristics approach is
slightly high when compare to IP-LR. Fig. 1 (b) shows the comparison of the execu-
tion time of the proposed methods. The average reduction of execution time is 90.84
% when compared to IP-LR. The above discussions show that the simple heuristic
322 E. Grace Mary Kanaga and M.L. Valarmathi
based patient scheduling gives near optimal solutions in terms of weighted waiting
time and completion time with an emphatic decrease in the execution time.
5 Conclusion
The agent-based solution for the resource and task allocation problem in the healthcare
domain, namely patient scheduling problem is discussed in this paper. The results
shows that agent-based patient scheduling with various optimization techniques gives
better solution than traditional approaches and also handles dynamic and distributed
situations in a robust manner. A schedule that minimizes the patient waiting time may
even aid in saving the person’s life.
This agent based approaches with optimization techniques can be extended for
any complex resource allocation problems. Further the system may be tested with
other optimization techniques such as simulated annealing, bee colony optimization
approaches.
References
1. Paulussen, T.O., Jennings, N.R., Decker, K.S., Heinzl, A.: Distributed patient scheduling
in hospital. In: International Joint Conference on Artificial Intelligence, vol. 18,
pp. 1224–1232 (2003)
2. Bartelt, A., Lamersdorf, W., Paulussen, T.O., Heinzl, A.: Agent Oriented Specification for
Patient-Scheduling Systems in Hospitals. Dagstuhl Article (2002)
3. Vermeulen, I., Bohte, S., Somefun, K., La Poutré, H.: Multi-agent Pareto appointment ex-
changing in hospital patient scheduling. Service Oriented Computing and Applica-
tions 1(3), 185–196 (2007)
4. Kaandorp, G., Koole, G.: Optimal outpatient appointment scheduling. Health Care Man-
agement Science 10(3), 217–229 (2007)
5. Pham, D.N., Klinkert, A.: Surgical case scheduling as a generalized job shop scheduling
problem. European Journal of Operational Research 185(3), 1011–1025 (2008)
6. Becker, M., Navarro, M., Krempels, K.H.A., Panchenko: Agent based scheduling of opera-
tion theatres. In: EU-LAT eHealth Workshop on Agent Based Scheduling of Operation
Theatres (2003)
7. Marinagi, C.C., Spyropoulos, C.D., Papatheodorou, C., Kokkotos, S.: Continual planning
and scheduling for managing patient tests in hospital laboratories. Artificial Intelligence in
Medicine 20(2), 139–154 (2000)
8. Kapamara, T., Sheibani, K., Hass, O.C.L., Reeves, C.R., Petrovic, D.: A review of sche-
duling problems in radiotherapy. In: Proceedings of 18th International Conference on
Systems Engineering, pp. 207–211 (2006)
9. Liu, N., Abdelrahman, M.A., Ramaswamy, S.: A Complete Multi-agent Framework for
Robust and adaptable dynamic job Shop Scheduling. IEEE Trans. On Systems Man And
Cybernetics Part C Applications And Reviews 37(5), 904–916 (2007)
10. Janiak, A., Rudek, R.: Experience-Based Approach to Scheduling Problems With the
Learning Effect. IEEE Transactions on Systems, man, and Cybernetics—part a: systems
and human 39(2), 344–357 (2009)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 323–326, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Naive Bayes Approach for Website Classification
R. Rajalakshmi and C. Aravindan
Department of Computer Science and Engineering
SSN College of Engineering, Chennai, India
{rajalaxmi,aravindanc}@ssn.edu.in
Abstract. World Wide Web has become the largest repository of information
because of its connectivity and scalability. With the increase in number of web
users and the websites, the need for website classification gains attraction. The
website classification based on URLs alone plays an important role, since the
contents of web pages need not be fetched for classification. In this paper, a soft
computing approach is proposed for classification of websites based on features
extracted from URLs alone. The Open Directory Project dataset was considered
and the proposed system classified the websites into various categories using
Naive Bayes approach. The performance of the system was evaluated and Pre-
cision, Recall and F-measure values of 0.7, 0.88 and 0.76 were achieved by this
approach.
Keywords: Website classification, URL, Bayes classifier.
1 Introduction
Web page classification is the process of assigning a web page to one or more prede-
fined category labels. The classification of web pages is necessary and it is normally
performed based on the content of the website. If the web sites are classified based on
URLs alone, the web pages need not be fetched and analyzed. The URL based web-
site classification can be used to identify the abnormal traffic generated from an or-
ganization by observing packets. So systems are designed to automate the classifica-
tion process based on URLs.
In the proposed system, the classification of websites was done based on their
URLs alone using Naïve Bayes approach. The websites were classified into one of the
categories viz., Arts, Business, Computers, Games, Health, Home, News, Recreation
and Reference. The performance of the system was evaluated using ODP dataset and
Precision, Recall and F-measure of 0.7, 0.88 and 0.76 were achieved.
2 Related Works
Qi, X. and Davison [1] surveyed the existing automated classification systems and
compared the approaches in terms of features and classifiers used. But, all these sys-
tems require the content be downloaded before classification. This method of classifi-
cation after downloading the web page content may not be useful if the objective is to
324 R. Rajalakshmi and C. Aravindan
block objectionable content or when speed is of crucial importance. Min Yen Khan
[2] proposed an approach for categorizing web pages without web content using the
ILP98 WebKB dataset. For this SVM based classification, the features viz., URL text,
anchor text, title text and page text were considered and they reported an average F-
measure of 0.43 for the following 4 categories: Course, Faculty, Project and Student.
In the approach proposed by Min-Yen [3], the URLs were segmented into meaningful
chunks and features such as URL component length, content, orthography, token
sequence and precedence were considered. Then Maximum-entropy learning was
applied to classify the web sites and an F-measure of 0.62 was achieved. Lim Wern
Han et al.[4] proposed a framework for classifying the websites considering the fea-
tures from URL, web page title and metadata information from web pages. They re-
ported an average F-measure of 0.78 for 6 categories viz., Business, Economy, Enter-
tainment, Government, Health, News and Sports. But they used additional features
that are not part of the URL, and considered only 6 categories among which 3 are
subcategories. In the proposed system, Naïve Bayes approach for classification of
websites was explored using the features extracted only from the URLs.
3 URL Classification
The system was designed to classify the given URL into one of the 9 categories, viz.,
Arts, Business, Computers, Games, Health, Home, News, Recreation and Reference.
A total of 8,55,939 URLs of 9 categories were extracted from the ODP dataset [5].
Among this 7,70,309 URLs were chosen randomly for forming the dictionary. The URLs
from each category were parsed into tokens. The common words such as "www",
"http", "html" etc. were removed from the list of tokens and only the unique tokens
were used for constructing the dictionary. For 9 categories, a total of 9 dictionaries
were used for training the Bayes classifier.
By applying Bayes theorem, the probability of given URL to belong to Category Ci
can be computed and the URLs can be classified using Naïve Bayes Classifier. To
obtain the most probable hypothesis, given the training data, Maximum a posteriori
hypothesis hMAP was used.
hMAP=arg max P(D|h) * P(h) (1)
For classification, an URL was separated into tokens and each token was checked
with all the dictionaries in the training set. In each dictionary, the occurrence of these
tokens was checked and the number of equal matches and the partial matches were
counted. The value of each token was computed as shown in Equation (2). The value
of the URL depends on its individual tokens, and the dictionary size of the corre-
sponding category.
valueTokeni = eqmatch +(0.5 * partmatch) (2)
The likelihood probability of an URL, given the Category was estimated by Eqn. (3),
in which DictSizeCi is the total number of words in the dictionary for the category
Naive Bayes Approach for Website Classification 325
Ci , and i represents category number that varies from 1 to 9. The token count denoted
by t, varies from 0 to n, where n represents the maximum number of tokens in a URL.
i
n
t
t
iDictSizeC
valueToken
CURLP
∑
=
=0
)|( (3)
To estimate the probability of the given URL to belong to Category Ci, prior probabil-
ity and likelihood probability were used. The prior probability was assumed for every
category according to ODP dataset. To obtain the prior probability of each category,
total number of URLs in each category was divided by total number of URLs in the
dataset. The probability of given URL to be classified under the Category Ci was
computed using equation (4).
P(Ci | URL) = P(URL|Ci) * P(Ci) (4)
Here, P(Ci | URL) is the posterior probability of given URL to be in the Category Ci,
and P(URL|Ci) is the likelihood probability that is calculated using Equation (3) and
P(Ci) is the prior probability of Category Ci. By Bayes theorem, final decision of
classification was made. The category for which the URL has the maximum value
was adjudged as the class of the URL. The experiments showed acceptable results for
this approach of classification.
4 Experiments and Performance Evaluation
For experimental purposes, 9 categories of URLs from ODP dataset were used as
shown in Table 1. A set of 9 experiments were performed by taking 80% for training
and remaining for testing and cross validation was done. Finally 90% of training set
was used and tested with remaining 10% URLs in each category. The system was
implemented in Java. To evaluate the performance of the classifier, Precision, Recall
and F-measure values were calculated and shown in Table 2. The average value of
Precision, Recall and F-measure of 0.7, 0.88, and 0.76 were achieved.
Table 1. Number of URLs in Training and Testing Dataset
Category Number of URLs Training Set Testing Set
Arts 227337 204597 22740
Business 227000 204300 22700
Computers 109201 98390 10921
Games 51207 46080 5127
Health 57515 51750 5755
Home 25368 22824 2544
News 8235 7407 828
Recreation 94565 85104 9461
Reference 55521 49967 5554
Total 8,55,939 7,70,309 85,630
326 R. Rajalakshmi and C. Aravindan
Table 2. Performance measures
Arts Business Comp Games Health Home News Recn Refer
Precision 0.80 0.97 0.65 0.69 0.76 0.80 0.53 0.66 0.44
Recall 0.82 0.57 0.92 0.90 0.97 0.97 0.93 0.93 0.88
F-measure 0.81 0.72 0.76 0.79 0.86 0.88 0.67 0.77 0.59
5 Conclusion
The proposed system classifies websites purely based on the URLs. With this ap-
proach of classification the Precision, Recall and F-measure values of 0.7, 0.88 and
0.76 were achieved. This system helps in monitoring the browsing behavior of users
inside an organization and can assist the administrators to block unnecessary sites.
This classifier can be further extended to identify the abnormal traffic generated from
an organization and the classification accuracy can still be improved by using more
features.
References
1. Qi, X., Davison, B.D.: Web page classification: Features and algorithms. ACM Comput.
Surv. 41(2), 1–31 (2009)
2. Kan, M.-Y.: Web Page Categorization without the Web Page. In: Proceedings of the 13th
International World Wide Web Conference, pp. 262–263. ACM, New York (2004)
3. Kan, M.-Y., Thi, H.O.N.: Fast Webpage Classification using URL Features. In: Proceedings
of CIKM 2005, Germany (2005)
4. Han, L.W., Alhashmi, S.M.: Joint Web-Feature (JFEAT): A Novel Web Page Classification
framework. Communications of IBMA, Article ID 73408 (2010)
5. Open Directory Project, http://www.dmoz.org
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 327–331, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Method to Improve the Efficiency of the Software by the
Effective Selection of the Test Cases from Test Suite
Using Data Mining Techniques
Lilly Raamesh1 and G.V. Uma2
1 Research Scholar, Anna University, Chennai 25
lillyraamesh@yahoo.co.in
2 Professor & Head/IST, CSE, Anna University, Chennai-25
Abstract. This paper proposes a new technique for the efficient selection of test
cases from the large suite. The main aim of selection of test cases is that the
fewer test cases may reduce the testing time and project cost. And also test case
selection may increase a test suite’s rate of fault detection. An increased rate of
fault detection during testing provides earlier feedback on the system under test,
allowing debugging to begin earlier, and supporting faster strategic decisions
about release schedules. Further, an improved rate of fault detection can in-
crease the likelihood that if the testing period is cut short, test cases that offer
the greatest fault detection ability in the available testing time will have been
executed.
Keywords: software testing, test case, test suite, data mining, clustering.
1 Introduction
In test automation, software is used to control the execution of tests, the comparison
of actual outcomes to predicted outcomes, the setting up of test preconditions, and
other test control and test reporting functions. Test automation involves automating a
manual process already in place that uses a formalized testing process. Ideally, testing
is completely automatic after the specification has been given. The basic objective of
test cases is to validate the testing coverage of the application. An optimized set of
test cases brings standardization and minimizes the ad-hoc approach in testing.
2 Software Testing
Software testing is the process of validating and verifying a software program. Testing
can never identify all the defects within software. Instead, it provides a comparison that
compares the state and behavior of the product against principles or mechanisms by
which someone might recognize a problem.
A test case has components that describe an input, action or event and an expected
response, to determine an application is working correctly.
328 L. Raamesh and G.V. Uma
A test suite otherwise known as a validation suite is a collection of test cases that
are intended to be used as input to a software program to show that it has some speci-
fied set of behaviors.
3 Data Mining
Data mining, the extraction of hidden predictive information from large databases, is
a powerful new technology with great potential. Data mining tools predict future
trends and behaviors, allowing businesses to make proactive, knowledge-driven deci-
sions. The automated, prospective analyses offered by data mining move beyond the
analyses of past events provided by retrospective tools typical of decision support
systems. Data mining tools can answer business questions that traditionally were too
time consuming to resolve. They scour databases for hidden patterns, finding predic-
tive information that experts may miss because it lies outside their expectations.
4 System Architecture
The diagram explains the system architecture
• Converting the UML state diagrams into FSM (Finite State Machine)
• Generate test cases from the mined FSM diagrams using available techniques.
• Reduce the test suite size.
5 Converting the UML State Diagrams into FSM
The Unified Modeling Language (UML) is a standard in industry for modelling with
several diagrams that express static and dynamic aspects of a system. As one of
the behavioural models, state diagram is often used to model the life cycle of certain
object.
A FSM(Finite State Machines) A is a quintuple (Q, L, δ, q0, q), where Q is a finite
set of states of A, L is a finite set of transition labels of A, δ : Q×L→ Q is the transi-
tion function relating two states by the transition going between them.
Xholon tool is used for transformation mechanism from state diagrams to FSM
models. Xholon is a flexible open source tool for multi-paradigm (UML 2,
ABM, SBML, NN, GP, PSys, CA, ...) modeling, simulation, design, execution, and
transformation.
UML State
Diagram FSM Test Cases Generation
Minin
g
techni
q
ue
O
p
timized
test suite
Method to Improve the Efficiency of the Software by the Effective Selection 329
6 Test Case Generation from State Machine Diagrams
The method given in [3] is used to create test cases from state machine diagrams
which is as follows:
There are three main steps in test case generation, in the first step a predicate is
selected on a transition from a UML state machine diagram. In the next step, the se-
lected predicate is transformed into a predicate function. In the third step, test data are
generated corresponding to the transformed predicate function.
6.1 Predicate Selection and Transformation
For selecting a predicate, a traversal of the state diagram is performed using depth
first (DFS) traversal or breadth first (BFS) traversal to see that every transition is
considered for predicate selection. DFS traversal is used here. During traversal, condi-
tional predicates on each of the transitions are looked. Corresponding to each condi-
tional predicate, test data are generated.
Let I0 consists of all variables that affect a predicate q in the path P of a state ma-
chine diagram, then two points named ON and OFF for a given border satisfying the
boundary-testing criterion are created. The relational expressions of the predicates are
transformed into a function F called predicate function.
6.2 Test Data Generation
The basic search Procedure used for finding the minimum of a predicate function F is
the alternating variable method. This method is based on minimising F with respect to
each input variable in turn. An initial set of inputs can be randomly generated by in-
stantiating the data variables. Each input data variable xi is increased/ decreased in
steps of Sxi, while keeping all other data variables unchanged. Here, Sxi refers to a
unit step of the variable xi. The exact value of unit step can be defined conveniently.
For example, unit step of 1 is normally used for integer values. Unit step can easily
be defined for many other types of data such as float, double, array, and pointer and
so on.
The refinement is done by reducing the size of the step and comparing the value of
F with the previous value. Also, the distance between the data points is minimised by
reducing the step size. For each Conditional predicate in the state machine diagram,
the test data is generated. The generated test data are stored in a file. A test executor
can use these test cases later for automatic testing.
The above said procedure produces a test suite that is of somewhat smaller size.
But the size is further reduced by using mining techniques.
7 Mining Techniques for Test Suite Optimisation
This paper proposes a clustering Approach. Clustering is the process of grouping the
data into classes or clusters so that object within a cluster has high similarity in com-
parison to another, but is dissimilar to object in other clusters. It doesn’t require the
class label information about the data set because it is inherently a data driven ap-
proach. It is the process of grouping or abstract object into classes of similar object.
330 L. Raamesh and G.V. Uma
Among all the mining techniques, clustering is the most effective technique, which
can be used for test case mining. Clustering analysis helps constant meaningful parti-
tioning of a large set of object based on a “divide and conquer” methodology, which
decomposes a large scale system into smaller components to simplify design and im-
plementation. As a data mining task, data clustering identifies cluster or densely
populated regions, according to some distance measurement, in a large, multidimen-
sional data. Given a large set of multidimensional data points, the data space is
usually not uniformly occupied by the data points. Data clustering identifies the
sparse and the crowded places, and hence discovers the overall distributions patterns
of the data set.
algorithm TestSuiteOptimisation
begin
currentClusterIndex := 0; Sort(clusters,ascending);
while there exists ri such that marked[i] == FALSE do
if currentClusterIndex == k then currentClusterIndex := 0;
list := all tj €
€clusters[currentClusterIndex];
if Card(list) == 1 then
test := t€
€list;
if there exists rj €
€requirements where cv[test,rj] == TRUE and
marked[j] == FALSE then nextTest := test;
else
currentClusterIndex := currentClusterIndex + 1;
continue;
endif
else nextTest := SelectTest(list);
endif
if nextTest ≠ 0 then
RS := RS €
€ {nextTest};
foreach rj €
€requirements where cv[nextTest,rj] == TRUE do
marked[i] := TRUE;
endfor
endif
currentClusterIndex := currentClusterIndex + 1;
endwhile
return RS;
end TestSuiteOptimisation
8 Conclusion and Future Work
The proposed approach helps to generate optimised number of test cases which
results in manageable size test suite. So that the test suite can be run many times as
software is modified and evolved. The automation of the process can be concentrated
in the future.
Method to Improve the Efficiency of the Software by the Effective Selection 331
References
1. Wang, J., Mong, W.H., Lee, L., Sheng, C.: A Partition-Based Approach to Graph Mining
National University of Singapore, Singapore 117543 (1989),
wangjunm, whsu, leeml shengcha @comp.nus.edu.sg
2. Zhang, S., Yang, J., Cheedella, Monkey, V.: Approximate Graph Mining Based on Span-
ning Trees EECS Dept. Case Western Reserve University (1995),
jiong.yang@case.edu, venumadhav@gmail.com
3. Samuel, P., Mall, R., Bothra, A.K.: Automatic test case generation using unified Modeling
language (UML) state diagrams Department of Computer Science and Engineering, Indian
Institute of Technology, Kharagpur 721302, West Bengal, India (1999),
philips@cusat.ac.in
4. Wang, X., Guo, L., Miao, H.: An Approach to Transforming UML Model to FSM Model
for Automatic Testing School of Computer Engineering and Science Shanghai University
China _whitecn, glory (1994), hkmiao@sh
5. Lin, J.-W., Huang, C.-Y.: Huang Analysis of test suite reduction with enhanced tie-
breaking techniques *Department of Computer Science, National Tsing Hua University,
No. 101, Section 2, Kuang Fu Road, Hsinchu 300(1996), Taiwanu.edu.cn
6. Bader, J., Chaudhuri, A., Rothberg, J., Chant, J.: Gaining confidence in high-throughput
protein. Interaction Networks, Nature Biotechnology 22(1), 78–85 (2004)
7. Koyuturk, M., Grama, A., Szpankowski, W.: An efficient algorithm for detecting frequent
subgraphs in bioloicalnetworks. Bionformatics 20, 200–207 (2004)
8. Kuramochi, M., Karypis, G.: Frequent subgraph discovery. In: Proc. of ICDE (2001)
9. Nijssen, S., Kok, J.: A quickstart in frequent structure mining can make a difference. In:
Proc of KDD (2004)
10. Yan, X., Han, J.: CloseGraph: Mining Closed Frequent Graph Patterns. In: Proc. of
SIGKDD (2003)
11. Yan, X., Han, J.: gSpan: graph-based substructure pattern mining. In: Proc. of ICDM
(2002)
12. Bertolino, A.: Software Testing: Guide to the software engineering body of knowledge.
IEEE Software 16, 35–44 (1999)
13. Quan, Z., Bernhard, Gioranni: Automated Software Testing and Analysis: Techniques,
Practices and Tools. In: Proc. of Intl Conf. on System Sciences, HICSS 2007, p. 260
(2007)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 332–335, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Hybrid Intelligent Path Planning Approach to
Cooperative Robots
K. Prasadh1, Vinodh P. Vijayan2, and Biju Paul2
1 Professor, CSE, Valia Koonambaikulathamma college of Engg & Tech, Parippally
ksprasadh@yahoo.com
2 Assistant Professor, Rajagiri School of Engineering and Technology, Cochin-39
vinodhpvijayan@yahoo.com, biju_paul@rajagiritech.ac.in
Abstract. The main problem in a distributed cooperative mobile robots when
they navigate to a target is that fixing the degree of freedom in movement for an
individual robot without breaking the communication link with other robots. In
this paper an agent based approach to intelligently coordinate the group mem-
bers in a cooperative robots are done. Initial steps of optimized link state rout-
ing algorithm is used to get the information of other members. And a fuzzy
based decision can be taken regarding the movement without breaking the
communication link. So this technique ensure , each and every time all the ro-
bots are in communication with at least one robot in the group, at the same time
system provides maximum degree of freedom for movement.
Keywords: Type2 Fuzzy, distributed cooperative robot, Multi Agent, OLSR.
1 Introduction
Cooperation among the robots are the major design issues in a distributed cooperative
mobile robots system. The distributed cooperative means, there is no master slave
relation all the robots are autonomous and capable of making decision. But the indi-
vidual decision of a robot is depend on the other robot in the group, hence it is called
cooperative. So they travel to destination as team. Here the communication among the
robot is need to be well maintained. And communication is established using some
kind of wireless technique in physical layer, hence there will be a specific range for
the communication. The problem that may occur is when any of the robot move in
different direction as a result of obstacle avoidance etc may go out of wireless range
and the word cooperative become meaning less. Hence it is need to be calculate a the
degree of freedom for robot movement, means the range in meters a robot can move
with out breaking the communication with other robot in the group or by maintaining
the communication with at least one robot in the group. Once this is calculated indi-
vidual robot can take turning decision by keeping this distance in consideration. The
use of Optimized Link State Routing Protocol (OLSR)[5,6] in this context will
provide the next hop or neighbor robot information to the current robot. So once the
distance to different neighbor robots are known, a robot can calculate how much it
can travel in each direction with out breaking the wireless link. Here we specifically
use a communication agent in each robot, which maintain the communication
333 K. Prasadh, V.P. Vijayan, and B. Paul
between robots. Other functions like sensing , collecting link state information, main-
tain wireless communication, final actuator to robots legs are also designed as agents.
Hence a multi agent approach to the system provide a higher level of intelligence.
2 System Diagram
Figure 1 shows the general functional block diagram that map with in a robot. The
sensor ring with fuzzy based collision avoidance is used for dynamic obstacle avoid-
ance in target path. Sensor ring is used to identify the obstacle direction, size etc. The
wireless LAN maintain necessary connectivity between the robots. The Link informa-
tion block runs a optimized link state algorithm and finds neighbors using hello
packet, an echo packets can be used to find the transmission time. Once the transmis-
sion time is obtained, the distance to particular neighbor can be calculated and update
in the table. Based on the distance calculated to each neighbor and the prior knowl-
edge of range of wireless system, the degree of freedom for movement of a particular
robot with respect to each neighbors can be calculated.
Fig. 1. System block diagram
3 Tools and Techniques Used
3.1 Optimized Link State Routing Protocol (OLSR)
The OLSR[5,6,7,8] is an IP routing protocol optimized for mobile ad-hoc networks.
Here OLSR is executed and the neighbor robot information is collected as link state
information. Assume let robot R1 and robot R2 is L meter apart and maximum range
of available wireless is R meter, then the distance for which R1 is free to move,
Dr = (R/2-L/2) meters.
We consider only half of available wireless range because other robot itself may move
away. Next issue is how frequent the link information need to be updated. Simple
solution is the time taken to travel Dr, and if the speed or all robot are similar then it
is very straight to calculate.
A Hybrid Intelligent Path Planning Approach to Cooperative Robots 334
3.2 Fuzzy Logic Controller as Decision Making System
Decision making can be done through Fuzzy [3] logic. we can use either type-1 fuzzy
or type-2 fuzzy according to the requirement.
Fig. 2. Membership function for ' Direction and Range'
Figure 2 shows Type-2 fuzzy membership[1,2] function for direction and range A
range has the subsets: Far (F), Medium (M), and Near (N). A direction has the
subsets: Large Negative (LN), Medium Negative (MN), Zero Negative (ZN), Zero
Positive (ZP), Medium Positive (MP), and Large Positive (LP).
4 System Architecture: Implementation Details
Figure 3 show the multi agent architecture of the system, the simplified diagram repre-
senting the multi-agent system using a interface agent. The agents basically interact
with the other components of the system by manipulating information on the interface
agent. The information on the interface agent may represent facts, assumptions, and
deductions made by the system during the course of solving problem.
Fig. 3. Multi-agent System
5 Experimentation and Result
Figure 4 shows a working Environment and the doted Circle with radius Dr1 and Dr2
shows the range in which the robot is free to move, The Dr1 and Dr2 is calculated for
a particular robot from its two different neighbor. Figure 5 show the performance
measures for calculation of link updating interval. Which shows that the link interval
calculation need to be more frequent as speed of the robot increases.
335 K. Prasadh, V.P. Vijayan, and B. Paul
Fig. 4. Working Environment Fig. 5. Link state interval for various speed
6 Conclusion
Cooperative robots can be well organized using optimized link state routing technique
and type 2 fuzzy based decision making. This paper basically addressed the problem
of degree of freedom for movement of a robot when they are in a group. So using the
proposed technique an individual robot will get maximum freedom for moving with
in the group without breaking communication link.
References
1. Vijayan, V.P., John, D., Thomas, M., Maliackal, N.V., Vargheese, S.S.: Multi Agent Path
Planning Approach to Dynamic Free Flight Environment. International Journal of Recent
Trends in Engineering 1(1) (May 2009)
2. Poornaselvan, K.J., Kumar, G., Vinodh, T., Vijayan, P.: Agent Based Ground Flight Control
Using Type-2 Fuzzy Logic And Hybrid Ant Colony Optimization To A Dynamic Environ-
ment. In: First International Conference on Emerging Trends in Engineering and Technol-
ogy 2008 (2008)
3. Sotel, M.A., Narnjo, J., Gonzalez, C., Garcia, R., de Pedro, T.: Using Fuzzy Logic in Auto-
mated Vehicle Control. In: Intelligent Transportation Systems. IEEE, Los Alamitos (2007)
4. Coupland, S.: Type-2 fuzzy control of a mobile Robot. Centre for Computational Intelli-
gence De Montfort University, U.K
5. RFC_3626
6. Abolhasan, M., Hagelstein, B., Wang, J.C.-P.: Real-world performance of current proactive
multi-hop mesh protocols (2009), http://ro.uow.edu.au/infopapers/736/
7. Chandra, M., Roy, A.: Extensions to OSPF to Support Mobile Ad Hoc Networking. RFC
5820 (March 2010)
8. Ogier, R., Spagnolo, P.: MANET Extension of OSPF using CDS Flooding. RFC 5614
(August 2009)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 336–339, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Towards Evaluating Resilience of SIP Server under Low
Rate DoS Attack
Abhishek Kumar1, P. Shanthi Thilagam1, Alwyn R. Pais1, Vishwas Sharma2,
and Kunal M. Sadalkar1
1 Dept. of Computer Engineering, NITK-Surathkal, India-575025
{aksinghabsk, santhisocrates, alwyn.pais,
officialmails.kunal}@gmail.com
2 Dept. of SERC, IISc. Bangalore, India-560012
vishwas@mmsl.iisc.ernet.in
Abstract. Low rate Denial-of Service, DoS, attack recently emerged as the
greatest threat to enterprise VoIP systems. Such attacks are difficult to detect
and capable of discovering vulnerabilities in protocols with low rate traffic and
it noticeably affects the performance of Session Initiation Protocol, SIP, com-
munication. In this paper, we deeply analysis the resilience of SIP server
against certain low rate DoS attacks. For this purpose we define performance
metrics of SIP server under attack and non-attack scenarios. The performance
degradation under attacks gives a measure of resilience of the SIP server. In or-
der to generate normal SIP traffic and the attacks, we defined our own XML
scenarios and implemented them using a popular open source tool known as
SIPp. The system under evaluation was an open source SIP server.
1 Introduction
Voice over Internet Protocol (VoIP) is a technology that is reshaping the future of
telephony. While enterprise VoIP offers low cost and various functionality it’s also
opens the door for external threats [1]. Most VoIP services uses the SIP infrastructure,
because its simplicity and wide range of features, which makes its service vulnerable.
To provide a better VoIP services we need to understand the behavior of VoIP server
under different attack rates [2] and its countermeasure. The attack rate is comparable
with that of normal traffic and hence it is not flooding the target to cause saturation.
This qualifies the attacks to be termed “low rate”. These attack strategies are novel
approach to launch flooding DoS attack without sending high rate traffic to the vic-
tim. These attacks are mixed approaches between flooding and vulnerability attacks
by which attackers get advantages after reducing traffic rate.
We use a freely available open source Asterisk SIP server version 1.6.20 [3] & [4].
We populated 8000 unique username and passwords in user account and directory
data of server. We automate a registration flooding low rate DoS attack scenario and
populate a CSV file of 8000 valid and invalid username and passwords to generate
legitimate and illegitimate traffic to SIP application server. Fig. 1 (A) shows the no-
tional diagram of DoS attack target (SIP Registrar) of SIP application server.
Towards Evaluating Resilience of SIP Server under Low Rate DoS Attack 337
Fig. 1. A. DoS Attack Target SIP Registrar, B. Normal Registration and Registration Flooding
Attack
2 Performance Metrics and Experimental Configuration
2.1 Performance Metrics
Our all metrics are SIP based to evaluate the resilience of VoIP SIP server. If any
metrics are degraded, it will result to un-resilient SIP server. In our study we define
following metrics:
• Database Lookup Efficiency (R12). The ratio of total number of first REGISTER
request for database lookup by SIP server to total number of responded second
“401-Unauthorized or 404- Not found” messages. This will measure the searching
efficiency of SIP server for large number of user database in attack and no attack
scenario.
• SIP URI Binding Efficiency (R34). The ratio of total number of third REGISTER
request with MD5 digest to total number of responded fourth “200 OK” messages.
This measure defines the digest computation efficiency of SIP server in attack and
no attack scenario.
• Successful Registration Acceptance Rate (SRAR). Total number of successful
URI binding per second. This measure defines the successful acceptance rate in at-
tack and non attack mode. This metric determine the utilization of resources of SIP
server.
• Registration Drop Rate (RDR). Total number of rejected URI binding per sec-
ond. This measure defines the registration drop rate in attack and non attack sce-
nario. This metric determine the loss of potential resources of SIP server under
flooding attack.
2.2 Configuration Module
In our experiment in Fig. 2, we used SIPp tool [5] with Transport Layer Security
(TLS) support as an attack generation and analysis module and an open source
version of ASTERISK 1.6 as a SIP server. For attack generation module and client
configuration module, we added 8000 illegitimate and legitimate user’s credentials
in our CSV file respectively, which is an input of created XML attack scenario.
We compiled and run a C code to add 8000 users in SIP database to configure
SIP.conf file.
338 A. Kumar et al.
Fig. 2. Experimental Testbed
3 Resilience Result of SIP Server
According to our SIP server specification and experiment, SIP server is capable to
process 105 legitimate REGISTER packet/sec for 8000 configured phones in no at-
tack scenario. The results of our experiment show that REGISTER flooding attack
significantly degrades the input statistics and derived statistics R12, R34, SRAR and
RDR as shown in Fig.3 (A), (B), (C), (D), (E), (F), (G) and (H).
From Fig. 3 (B) & (C), the differences of successful call are almost more than 600
in attack to no attack scenario for each legitimate traffic rate. An attack rate of zero
and for legitimate traffic rate 100 REGISTER packet/sec on X-axis shows expected
efficiency. In Fig. 3 (E) and (F), R12 and R34 is 100%, for 100 REGISTER pack-
ets/sec in no attack scenario. But those metrics significantly decreases for different
legitimate traffic rate as well as the attack rate increases. For attack rate 200
packet/sec, the database lookup efficiency and SIP URI binding efficiency of SIP
server decreases to more than 94% and 98% respectively, that causes to occur 480 and
160 clients denied by registrar per second respectively. For higher traffic rate, the
Fig. 3. A. Elapsed Time vs Attack Rate, B. Successful Reg. vs Attack Rate, C. Failed Call vs
Attack Rate, D. No. of Reg. Req. with Credential vs Attack Rate, E. R12 vs Attack Rate, F.
R34 vs Attack Rate, G. SRAR vs Attack Rate, H. RDR vs Attack Rate.
Towards Evaluating Resilience of SIP Server under Low Rate DoS Attack 339
difference in value of R12 and R34 at attack rate 200 packets/sec changes signifi-
cantly. Note that the absolute value of 1% loss cause 80 clients (out of 8000 clients)
denied to service per second.
In Fig. 3(G), it is clear that, in no attack scenario the SRAR rate is 100 registration
per second for 100 REGISTER packet/sec traffic, because for 8000 successful SIP
URI binding the server takes 80 seconds. But it decreases to 28 registrations per sec-
ond, when attack rate is 200 for same benign traffic rate. In Fig. 3(G), we can easily
see that the differences in SRAR from no attack scenario to highest attack rate sce-
nario are almost constant (75 registrations/sec) for each traffic rate. From Fig. 4(H),
we see that the RDR value increases 0 to 2.5 in no attack scenario to highest attack
rate scenario. That means, in each second 2.5 REGISTER requests are discarded by
SIP server when attack rate is 200. This drop rate significantly degrades the perform-
ance of SIP server, when requests come to server for long time.
4 Conclusion and Future Work
SIP server is not capable of defending itself against low rate DoS attack also. All
performance metrics change significantly for low attack rates and server becomes less
resilient and more vulnerable to low rate DoS attack. Our study analyzes the perform-
ance of different protocol supporting server of VoIP infrastructure under low rate DoS
attack. In future we proposed to build a detection tool to detect low rate DoS attack on
SIP server and define a protocol state machine to mitigate it.
Acknowledgments. We would like to express our appreciation to Prof. N. Balakrish-
nan, Dept. of Supercomputer Education and Research Center, Indian Institute of Sci-
ence Bangalore for valuable support of this research.
References
1. Collier, M.D.: Enterprise Telecom Security Threats,
http://download.securelogix.com/library/
Enterprise_Telecom_Security_Threats_Draft_10-12-04.pdf
2. Sengar, H.: Overloading Vulnerability of VoIP Networks. In: DSN 2009, pp. 419–428
(2009)
3. Asterisk, The Open Source Telephony Projects, http://www.asterisk.org
4. VoIP Wiki- A Reference Guide to all things VoIP,
http://www.voip-info.org/wiki/view
5. Gayraud, R., et al.: SIPp, http://sipp.sourceforge.net
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 340–343, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Performance Analysis of a Multi Window Stereo
Algorithm on Small Scale Distributed Systems:
A Message Passing Environment
Vijay S. Rajpurohit1 and M.M. Manohara Pai2
1 Gogte Institute of Technology, Belgaum, India
vijaysr2k@yahoo.com
2 Manipal Institute of Technology, Manipal, India
mm.pai@rediffmail.com
Abstract. Stereo vision systems determine the depth from two or more images
which are taken at the same time, but from slightly different viewpoints. A
novel approach for depth map generation is proposed for a multi-window stereo
algorithm on a cluster computing setup using Message Passing Instructions
(MPI) to overcome the speed limitations.
Keywords: Stereo vision, Depth map, Symmetric stereo with Multiple Win-
dowing algorithm, Parallel Depth map Generation, Cluster computing setup,
Message Passing Instructions.
1 Introduction
Stereo vision refers to the process, which transforms the information of two plane
images (2-D) into a description of the 3D scene and recovers depth information in
terms of exact distance (Depth map). The Stereo vision in mobile Robot is attained by
equipping the Robot with two stereo cameras similar to Human Visual System (HVS).
To overcome the speed limitations of vision-based algorithms several hardware
software implementations were presented[2][3][4][5][6]. The existing hardware archi-
tectures are difficult to implement in real-life environment and not suitable low cost
commercial applications due to their high equipment cost and complex configura-
tions. In this work, a novel method for depth map generation is presented for an exist-
ing state of the art algorithm SMW[1] to overcome its speed limitations.
2 Depth Map Generation: A Parallel Implementation
The sequential SMW(Symmetric Stereo With Multiple Windowing) algorithm is
adaptive and uses multiple windowing approach. For each pixel a correlation is per-
formed with nine different windows. The SMW algorithm has the computational
complexity of O(n2 ).
Performance Analysis of a Multi Window Stereo Algorithm 341
2.1 Parallel SMW Algorithm
The parallel algorithm is designed to work with a cluster of computers connected in
parallel. Each node in the cluster processes a part of the left-right image pair to gener-
ate partial depth map. The server receives the sensor data of right and left images. The
received data is divided into equal number of segments which is equal to the number
of processors in the cluster (Fig. 1) including the server. Each slave processor has the
copy of SMW algorithm implementation. The server sends each pair of image seg-
ments to a unique processor in the cluster. Each processor executes SMW algorithm
and calculates the partial depth map to each image segment. Partial depth maps are
sent to the server and are combined to form the complete depth map of the image pair.
Fig. 1. Cluster computing setup for depth map generation
3 Analysis of the Algorithm
Based on the Amdahl’s law analysis, the speedup achievable by parallel SMW algo-
rithm for a fixed size problem increases initially with the increase in the number of
processors and becomes stable ( Fig. 2). As the number of processors increases, the
parallel overhead increases which reduces the overall performance of the algorithm.
The expected execution time is given by
[m2 /p + n[log p]λ + [log p](k/β)] . (1)
342 V.S. Rajpurohit and M.M. Manohara Pai
Where λ (latency) represent the time needed to initiate a message. β (bandwidth)
represent the number of data items that can be sent down a channel in one unit of
time. k indicates the data items. n indicates the message length.
Fig. 2. Predicted speedup of depth map generation
The algorithm is tested for various synthetic and real life stereo images (Fig 3) and
a speedup in the band from 1.48 to 1.68 and efficiency from 0.74 to 0.84 on two node
cluster. By adding one more node i.e. on three node cluster, speedup varied in the
band from 2.28 to 2.58 and efficiency from 0.77 to 0.86. Based on the results it can
concluded that the SMW algorithm can be implemented in parallel to improve
speedup, yet the practical speedup is less than the speedup predicted by Amdahl’s
law.
Fig. 3. Depth map creation using stereo vision
4 Conclusion
A parallel multi-window stereo algorithm (SMW) for the depth map generation,
on cluster computing setup is developed to overcome the slow response time of the
existing multi-window algorithms. Based on the results it can be concluded that the
Performance Analysis of a Multi Window Stereo Algorithm 343
performance of the algorithm initially improves with increase in number of processors
but the observed improvement is less than the improvement expected by the Amdahl's
law in the cluster computing environment.
References
1. Fusiello, A., Roberto, V., Trucco, E.: Symmetric stereo with multiple windowing. Interna-
tional Journal of Pattern Recognition and Artificial Intelligence 14(8), 1053–1066 (2000)
2. Koschan, A., Rodehorst, V.: Towards Real-Time Stereo Employing Parallel Algorithms For
Edge-Based And Dense Stereo Matching. In: Proceedings of the IEEE Workshop on Com-
puter Architectures for Machine Perception (CAMP 1995), Como, Italy, pp. 234–241
(1995)
3. Rosselot, D., Hall, E.L.: Processing real-time stereo video for an autonomous Robot using
disparity maps and sensor fusion. In: Proceedings of SPIE Intelligent Robots and Computer
Vision XXII: Algorithms, Techniques and Active,Vision, vol. 5608, pp. 70–78 (2004)
4. Laine, A.F., Roman, G.C.: A Parallel Algorithm for Incremental Stereo Matching on SIMD
Machines. In: Proceedings of the. 10th ICPR, Atlantic City, New Jersey, USA, vol. II,
pp. 484–490 (1990)
5. Szeliski, R., Zabih, R.: An experimental comparison of stereo algorithms. International
Journal of Computer Vision 32(1), 45–61 (1999)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 344–347, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Ant Colony Algorithm in MANET-Review and Alternate
Approach for Further Modification
Jyoti Jain, Roopam Gupta, and T.K. Bandhopadhyay
Rajeev Gandhi Technical University, Bhopal, India
jyotijain.phd@gmail.com
Abstract. Mobile ad hoc network(MANET) is the latest application of tele-
communication. This is one of the most innovative and challenging area of
wireless networking. Ant Colony Algorithm has been used in Mobile Network
since long because of isomorphism between them. Pheromone graph and stig-
mergic architecture of ant colony algorithm are comparable with structure &
constraints of communication network. In this paper, A literature survey about
the application of ACO is given. In this paper we review ANT algorithm and
different approaches proposed by researchers for the improvement of routing
performance. In this proposed work, ACO will be used in case of link failure.
Path will be discovered by reactive routing, and maintained by periodically
generating HELLO messages by all the nodes in the link. All nodes in the link
will also find an alternate route for next to next node proactively. By using this
method, throughput, and end to end delay parameters can be improved probably
the on the cost of increase in the overhead. Overhead will increases in proactive
route finding at the same time number of route failure reduces so the bits re-
quired in alternate route finding will reduce.
Keywords: MANET, Routing, ACO, AODV, Pheromone.
1 Introduction
Mobile ad hoc network (MANET) is the latest application of telecommunication. This
is one of the most innovative and challenging area of wireless networking. MANET is
an autonomous collection of mobile users that communicate over relatively band
width constrained wireless links. Due to movement of the nodes, network topology
may change rapidly and unpredictably over time. In this decentralized network, dis-
covering the route and delivering of data becomes complicated. Various protocols[1]
for routing are developed for MANET over past few years. Characteristics of mobile
ad hoc network and ant colony algorithm are common. Both of them are self config-
ured, self built and distributed. The concept of ant colony can be utilized to enhance
the performance of Routing protocol[2]. In this paper part I give the details of ant
colony routing algorithm. part II will discuss about the work done by different re-
searcher. In part III, new concept of application of ant algorithm in MANET routing
protocol is discussed. In part IV conclusion is given. Some of the desirable character-
istics such as scalability and robustness are exhibited in simple biological systems of
insects like ant colony.
Ant Colony Algorithm in MANET-Review and Alternate Approach 345
Part I
The Ant Colony Optimization Metaheuristic (ACO)
The ACO metaheuristic is a family of multi-agent algorithms to solve combinatorial
optimization problems. The representation of the combinatorial problem exploited by
the ant-like agents is split in two parts, forward ant and backward ant as shown in
figure 1 and figure 2. Each agent is an autonomous construction process making use
of a stochastic policy and aimed at building a single solution to the problem at hand,
possibly in computationally light.
Fig. 1. Working of forward ant Fig. 2. Working of backward ant
These ant agents are the instrument used to repeatedly sample the solution set ac-
cording to the bias implemented by pheromone variables. The colony’s management
tasks involve the generation of ant agents, the scheduling of optimization procedures
other than the ants (e.g. local optimizers) and whose results can be used in turn by the
ants, and the management of the pheromone, in the sense of modifying the pheromone
values.
Part II
Recently, some ACO-based routing algorithms have emerged for the employment in
MANET. Some of the work done in this field by the researchers is discussed here.
Schoonderwoerd et al.[3] designed ABC (Ant-Based Control) for circuit-switched
telephone networks. In this method ACO were able to optimize performance in the
network by balancing the load in the network. The AntNet algorithm introduced by Di
Caro and Dorigo [4] in 1998 for adaptive routing in packet switching wired networks.
AntNet, was highly adaptive to network and traffic changes, uses lightweight mobile
agents (called ants) for active path sampling, is robust to agent failures, provides mul-
tipath routing, and automatically takes care of data load spreading. Mobile Ants
Based Routing (MABR)[5] by M. Heissenbüttel, & Braun is introduced routing algo-
rithm for MANET’s inspired by social insects. a short hello message was used to
announce its presence and position to update routing table from the corresponding
node in highly dynamic networks. The advantages include the ability to react and deal
quickly with local and global changes. Ant Colony Based Routing (ARA) [6] by
Mesut G¨unes, gives route discovery mechanism similar to other algorithms such
as AODV and DSR. Liu Zhenyu developed EARA-CG (Emergent Ad hoc Routing
Algorithm) and EARA-QoS(Emergent Ad hoc Routing Algorithm)[7], In these
346 J. Jain, R. Gupta, and T.K. Bandhopadhyay
algorithm, the principle of swarm intelligence is used to evolutionally maintain rout-
ing information. This algorithm provides Congestion control. In ADRA by X. Zheng
et al.[8] 2004, The ants deposit simulated pheromones as a function of their distance
from their source node, the quality of the link, the congestion encountered on their
journey. The ADRA system is shown to result in fewer call failures.
Part III
Researchers have done lot of work in improving routing. Most of the work is done at
the level of Monitoring and Admission control. This algorithm is of flexible nature
and can be used for further modification to improve routing protocol. Most of the
researchers developed a technique to find a new path from the source node in case of
route failure. This increases end to end delay during the transmission. Work on alter-
nate route finding from the nearer of the faulty node is still untouched. In this pro-
posed work, In place of selecting a new path from source node, ACO can be used in
finding new path from a precursor of the failed node. In this proposed work path will
be discovered by reactive routing, and maintained by periodically generating HELLO
messages by nodes in the link. All nodes in the link will also find an alternate route
Fig. 3. Algorithm for proposed work
Generate forward ant to find route for the alternate destination in the link
HELLO message to all its neighbors
Update neighbors list
Link break at particular
node
Find alternate route available in the
memory of the predecessors node
Transmit data by
previously available link
Transmit data by new link
End transmission
Initialize Data transmission
End transmission
NO YES
Ant Colony Algorithm in MANET-Review and Alternate Approach 347
for next to next node proactively. In case of route failure alternate route in the mem-
ory of these nodes can be used. Following flow chart presenting proposed algorithm.
Ant colony algorithm can be used to modified maintenance procedure of AODV
protocol. End to end delay will be reduced and hence throughput of the network will
increased. By local repairing the path in case of link failure the overhead during the
communication will also reduce which helps in maintaining the scalability. In place of
sending error message to the source & destination during route failure, fresh route
request can be generated by node just previous to the lost node. By reducing over
head better efficiency and scalability can be achieved. Over head can be reduced by
reducing number of bits in control packets like forward ant backward ant and Hello
packets. Some speeding techniques can be applied for the algorithm from the closer
point to the optimum solution.
2 Conclusion
In this paper we analyzed basic ant colony algorithm and its application in the field of
Mobile ad-hoc network to improve the qualitative characteristic of routing process.
Flexibility of the ant colony algorithm provides numerous approaches to modify mo-
bile ad-hoc network. This flexibility can also be used at the maintenance level. Alter-
nate route finding in case of route failure can be done from node nearer (just before)
to the faulty node. The use of location information at the nodes is used as a heuristic
parameter. This resulted a significant reduction of time needed to establish a route
from a source to a destination which is important for a reactive routing algorithm.
Application of ant colony algorithm at the maintenance level can provide a new scope
in the optimization of routing protocol.
References
1. Perkins, C.E., Royer, E.M.: Ad-hoc on-demand distance vector routing. In: Proc. Of the 2nd
IEEE Workshop on Mobile Computing Systems and Applications (1999)
2. Dorigo, M., Gambardella, L.: Ant colony system: a cooperative learn approach to the travel-
ing salesman problem. IEEE Transactions on Evolutionary Computation 1(1), 53–66 (1997)
3. Schoonderwoerd, R., Holland, O., Bruten, J., Rothkrantz, L.: Ant-Based Load Balancing In
Telecommunications Networks. Adaptive Behavior (1996)
4. Di Caro, G., Dorigo, M.: AntNet: distributed stigmergetic control for comm. networks.
Journal of Artificial Intelligence Research 9, 317–365 (1998)
5. Heissenbuttle, M., Braun, T.: Ants-Based Routing in Large Scale Mobile Ad-Hoc Net-
works. In: Kommunikation in Verteilten Systemen, KiVS (2003)
6. Gunes, M., Sorges, U., Bouazizi, I.: ARA-The Ant-Colony Based Routing Algorithm
for MANETs. In: International workshop, WAHN 2002, British Columbia, Canada,
August18-21 (2002)
7. Liu, Z., Kwaitkowska, M.Z.: Costas Constantinou:A Biologically Inspired QoS Routing
Algorithm for MANET. In: 19th International Conference AINA 2005, pp. 426–431 (2005)
8. Zheng, X., Guo, W., Liu, R.: An ant-based distributed routing algorithm for ad-hoc
networks. In: ICCCAS, pp. 412–417. IEEE, Washington, DC (2004)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 348–351, 2011.
© Springer-Verlag Berlin Heidelberg 2011
RDCLRP-Route Discovery by Cross Layer Routing
Protocol for Manet Using Fuzzy Logic
Mehajabeen Fatima, Roopam Gupta, and T.K. Bandhopadhyay
R.G.P.V., Bhopal, M.P., India
mehajabeen.fatima@gmail.com, roopam_1710@yahoo.co.in,
bandotushar_bando@hotmail.com
Abstract. MANET links have dependency on node mobility and node battery
power causes degradation of network’s performance and it could be alleviated
by allowing cross layer interaction in the protocol stack. In this paper Route
Discovery by Cross Layer Routing Protocol (RDCLRP) is proposed for route
discovery and route maintenance and AODV is modified. Instead of periodic
hello message of AODV, non periodic hello warning message is used for route
maintenance and route discovery in RDCLRP. HELLO warning message is
used to give warning when link is going to break before link breakage when it
touches the threshold and it is broadcasted whenever it is required and not peri-
odically. With the help of Hello warning message new route can also be discov-
ered before link breakage. In this paper only HELLO warning message interval
is optimized with the help of fuzzy logic. This reduces overhead and power
consumption, in turn it reduces delay in search of new route. The result illus-
trates the improvement of optimized RDCLRP over the basic AODV. Network
performance such as throughput, delay, jitter, overhead and battery power con-
sumption can be improved significantly with this method.
Keywords: MANET, AODV, RDCLRP, Cross layer approach, Hello warning
message, Hello warning message interval ( HI).
1 Introduction
A Mobile Ad-Hoc Network (MANET) is a collection of wireless mobile nodes form-
ing a temporary network without using any centralized access point, infrastructure, or
centralized administration. Some or possibly all of these nodes are mobile. This net-
work can be deployed rapidly and flexibly. The person-to-person, person-to-machine
or machine-to-person can communicate instantaneously, immediately and easily with
MANET.
Cross-layer design allows exchange of information of a layer with any other possi-
bly non-adjacent layer in the protocol stack to improve the network performance. This
helps in reduction of delays, overhead, congestion, it saves bandwidth e.t.c.
In Part 2 RDCLRP-new proposal for route maintenance and route discovery is
given. In this proposal AODV is modified using Cross layer approach and Fuzzy
Logic. In Part 3 results and graphs are given. Then conclusion is drawn.
RDCLRP-Route Discovery by Cross Layer Routing Protocol 349
2 RDCLRP
In basic AODV the path is searched when it is required to send the packets. The path
is searched by RREQ and RREP messages, after searching of path, route has to be
maintained. Route is maintained by periodic hello messages. Every path which par-
ticipate in transmission of packets acts as active path or active route and every node
which participate in transmission of packets acts as active node send hello message
shows its presence in the transmission range and if a node does not send the hello
message for long time then it is assumed that it is die. Then a RERR message is sent
back to source. Then source will reinitiate the route discovery procedure. This in-
creases delay and it is not good for real time traffic. Hello message sends after every
one second in AODV it increases overhead and traffic inturn increases possibility of
congestion, consumes bandwidth unnecessarily and most important consumes battery
power[2]. So in [3], it is proposed that instead of this periodic hello message, non
periodic hello message can be send as hello warning message. If speed, transmission
range varies then a node will go out of transmission range in different times, so Hello
warning message(HWM) should be given adaptively. So a RDCLRP is proposed in
which adaptive HWM is used. This HWM in RDCLRP is send only once whenever it
is required and its interval is calculated on the basis of speed, transmission rage, bat-
tery power using fuzzy Inference system. This saves bandwidth and battery power
since transmission of packet consumes maximum of energy[5]. In RDCLRP, Hello
Warning Message Interval( HI ) is made adaptive w.r.t. speed, transmission range,
battery power. When a node is going out of transmission range or if its battery power
is going to zero called critical node. Then critical node will send a HWM to warn the
neighbour nodes that the link is going to break after sometime. So warning has to be
given before link breakage. Neighbour nodes listen the warning message, so
neighbour nodes reply to critical node. Critical node search the route locally and try to
find out the route before link breakage. So instead of sending hello message periodi-
cally, hello sends as HWM only once whenever it is required. This HWM broadcast
before link breakage. The node which reply first will participate in new route. The
critical node will reply service replicate message consist of new node id to its previ-
ous node and old route will replace by new route before link breakage. So a new route
is discovered through HWM locally. This reduces battery power consumption, delay
and overhead simultaneously.
HWM depends on prediction of position of node and remaining battery power.
Position of a node depends on transmitted power, this transmitted power is present at
Physical layer[4]. The value of transmitted power and corresponding transmission
range will call on network layer from physical layer. The position of a node also de-
pends on speed of a node and this speed is available at physical layer and will be call
on network layer - cross layer approach.
The proposed protocol is based on sharing of MAC & physical layer information at
Network layer[3]. These parameters are transmit power, transmission range, full
charge battery capacity, remaining battery power of node at time t, speed of node at
time t, direction of node at time t.
In this paper Fuzzy inference system is used to make Hello warning message inter-
val ( HI ) adaptive. Low, medium and high values are taken for speed, transmission
350 M. Fatima, R. Gupta, and T.K. Bandhopadhyay
range, battery power for making of fuzzification rules. Membership functions
for different parameters consider different ranges. Rules for adaptive HI are given
in Table 1.
Table 1.
p
B
Low(L)
p
B
Medium(M)
p
B
High(H)
r
T r
T r
T
S
L M H L M H L M H
L L L L H M H H H H
M L L L M M H M M H
H L L L L L M L L M
3 Results and Graphs
To analyze RDCLRP performance, fifty times simulation was run on Qualnet network
simulator. Scenario area is 1500x1500 sqm, number of nodes is 30. The communica-
tion distance of nodes is 300m. Nodes move every second. So, the speeds of nodes
from 0mps to 18mps are taken. The traffic model has 4000 packets. The packet length
is 512bytes. The transporting protocol is UDP. The MAC protocol is 802.11 and the
channel rate is 11Mbps. Mobility model is random way point and the work frequency
is 2.4GHz. The simulation period is 400s. HI is calculated by fuzzy logic. In this
paper calculations are made for HI for high transmission range and high battery
power. Here, basic AODV and RDCLRP are compared for throughput, delay, over-
head and number of hello messages. Fig. 1 to 4 drawn for r
T=H, p
B = H.
THROUGHPUT
-20000
0
20000
40000
60000
80000
100000
0 5 10 15 20
Spe ed (m ps )
Throughput(kbps
)
basic AODV
RDCLRP
END TO END DELAY
0
0.5
1
1.5
05101520
speed(mps)
basic AODV
RDCLRP
Fig. 1. Throughput Vs. speed Fig. 2. End to end delay Vs. speed
JITTER
0
0.5
1
1.5
2
2.5
3
3.5
0 5 10 15 20
Speed(mps)
basic AODV
RDCLRP
NO. OF HELLO MESSAGES
0
50
100
150
200
250
300
05101520
speed(mps)
basic AODV
RDCLRP
Fig. 3. Jitter Vs. speed Fig. 4. No. of Hello messages Vs. Speed
RDCLRP-Route Discovery by Cross Layer Routing Protocol 351
Fig. 1 shows that throughput has remarkable improvement specially at lower
speeds. This shows that if a node is moving slowly then there is no need to send Hello
packet again and again. This saves battery power. Fig. 2, 3 shows reduction in delay
and jitter which is best suited to real time traffic. The hello messages or control pack-
ets reduces in RDCLRP inturn reduces overhead, power consumption, undesirable
traffic again in turn it reduces congestion. Less overhead is favourable for non real
time traffic. This saves battery power consumption.
4 Conclusion
From above graphs it is clear that delay, jitter and overhead are reduced in RDCLRP
in comparison of basic AODV. The no. of hello messages reduced considerably in
RDCLRP comparative to basic AODV. This reduces power consumption and conges-
tion indirectly. And this improves throughput as shown in graphs. RDCLRP reduces
control overhead considerably as very less number of hello messages are sent and
they are sent as warning message which discovers new route without any increase in
overhead i.e. without addition of new control packets. This reduces delay also. Con-
cluded that RDCLRP can be used for both real time and non real time traffic.
References
1. Stojmenovic, I., Lin, X.: Power-aware Localized Routing in Wireless Networks. IEEE, Los
Alamitos (2000), ISBN: 0-7695-0574-0/2000
2. Zhang, X., Gao, X., Shi, D.: Lifetime-aware Leisure Degree Adaptive Routing Protocol for
Mobile Ad hoc Networks. In: Proceedings of the Third International Conference on Wire-
less and Mobile Communications (ICWMC 2007). IEEE, Los Alamitos (2007), ISBN:
0-7695-2796-5/07
3. Fatima, M., Gupta, R., Bandhopadhyay, T.K.: Route maintenance improvement By Warn-
ing Hello In AODV Of MANET Using Cross Layer Approach. In: IEEE International Con-
ference on advance Computing and communication techniques, Panipat, Haryana, India,
October 30 (2010)
4. Cano, J.-C., Kim, D.: Investigating Performance of Power-aware Routing Protocols for Mo-
bile Ad Hoc Networks. In: Proceedings of the International Mobility and Wireless Access
Workshop (MobiWac 2002). IEEE, Los Alamitos (2002), ISBN: 0-7695-1843-5/02
5. Marie feeney, L., Nilson, M.: Investigating the energy consumption of a wireless network
interface in an adhoc networking environment. In: Proceedings of IEEE INFOCOM,
Anchorage, AK (2001)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 352–357, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A New Classification Algorithm with GLCCM for the
Altered Fingerprints
R. Josphineleela1 and M. Ramakrishnan2
1 Research scholar, Sathyabama University, Chennai
ilanleela@yahoo.com
2 Department of Information Technology, velammal Engineering College, Chennai
ramkrishod@gmail.com
Abstract. Fingerprints have always been the most practical and positive means
of Identification. Offenders, being well aware of this, have been coming up with
ways to escape identification by that means. Erasing left over fingerprints, using
gloves, fingerprint forgery; are certain examples of methods tried by them, over
the years. Failing to prevent themselves, they moved to an extent of mutilating
their finger skin pattern, to remain unidentified. This article is based upon oblit-
eration of finger ridge patterns .In this article, we propose a new classification
algorithm GLCCM (Gray Level Correlation Coefficient Co-occurrence Ma-
trix) algorithm for altered fingerprints classification. It is based on the fact that
altered fingerprint image is composed of regular texture regions that can be
successfully represents by co-occurrence matrix. So, we first extract the fea-
tures based on certain characteristics and then we use these features for classify-
ing altered fingerprints.
Keywords: altered fingerprints, enhancement, classification and co-occurrence
matrix.
1 Introduction
Fingerprints are used to Identify Objects, Compare Objects Remotely and Test an
Object for Changes. Since fingerprints are smaller, they are very useful as stand-ins
for remote objects. The primary purpose of fingerprint alteration is to evade identifi-
cation using techniques that vary from abrading, cutting, accidents etc. In this paper,
we propose new classification algorithm GLCM(Gray level co-occurrence matrix)
for identifying criminals by purposely altering their fingerprints .Gray level co-
occurrence matrix for altered fingerprint classification consists of Pre-processing,
Feature Extraction, Classification.
Motivation: The motivation behind the work is growing need to identify a person for
security. The fingerprint is one of the popular biometric methods used to authenticate
human being. The proposed enhancement method provides reliable and better per-
formance than the existing technique.
A New Classification Algorithm with GLCCM for the Altered Fingerprints 353
Contribution: In this paper we used altered and natural Fingerprint enhancement
using adaptive wiener 2 filter coefficient with the help of MATLAB codes. Organiza-
tion: This paper is organized into the following. Pre-processing, Feature Extraction,
Co-occurrence matrix (CM), classification, Result finally the Conclusions.
Input
Altered/Natural
Fig. 1. Block diagram of Altered Fingerprints Classification
2 Preprocessing
Preprocessing consists of histogram equalization, Image enhancement, binarization
Fig. 2. Block diagram of Pre-processing
Feature Extraction
Co-occurrence matrix
creation(CM)
Classification
Preprocessing
Histogram Equalization
Binarization
Image Enhancement
354 R. Josphineleela and M. Ramakrishnan
2.1 Histogram Equalization
Because there are many noises in original fingerprint image, an image enhancement
algorithm such as histogram equalization is usually applied to reduce the influence of
the noise in fingerprint image and to emphasize the ridges structures of damaged
fingerprint patterns.
Fig. 3. Altered fingerprint noise reduction experiment samples
2.2 Image Enhancement
To decreases the noise and improve the definition of ridges against the valleys. For
this enhancement the adaptive filter is applied separately to each block of the image
Fig. 4. Altered fingerprint enhancement experiment samples
2.3 Binarization
Converts the enhanced image into binary image adaptive threshold is used, pixel val-
ues above the threshold are assigned to 1 and pixel values below the threshold are
assigned to 0.
3 Feature Extraction
In the feature extraction phase, the feature region is divided into 8*8 blocks and
their directional values are estimated. The directional field is calculated using Gabor
A New Classification Algorithm with GLCCM for the Altered Fingerprints 355
filter and the location of the pixels; orientation and ridge type are stored in the co-
occurrence matrix.
4 Co-occurrence Matrix Creation
For Co-occurrence matrix creation the following algorithm is used GLCMA (Gray
level co-occurrence matrix algorithm) is used.
GLCMA(ip,jp)
{
if (ridge type==2)or(ridge type==4)
{
//termination minutiae has been found
store termination minutiae(it,jt);
}
else if(ridge type==1) or (ridge type==3)
{
//bifurcation minutiae may exist
if( point is valid(it,jt))
store bifurcation minutiae(it,jt);
}
else
{
delete the false pixel position
}
//Creation of co-occurrence matrix
}//end of algorithm
5 Classification
For the classification task, the GLCCM (Gray level Correlation coefficient Co-
occurrence matrix) algorithm is used. The co-occurrence matrix of the input image is
compared with the template image matrix,
x is the template gray level image,
_
x is the average gray level in the template image
y is the source image section
_
y is the average gray level in the source image
N is the number of pixels in the section image
(N= template image size = columns * rows)
The value cor is between –1 and +1, with larger values representing a stronger rela-
tionship between the two images.
356 R. Josphineleela and M. Ramakrishnan
()
()()
∑∑
∑
−
=
−
=
−
=
−⋅−
−⋅−
=1
0
1
0
22
1
0)(
N
i
N
i
ii
i
N
ii
yyxx
yyxx
cor (1)
6 Result and Discussion
In this paper, we have proposed altered fingerprint classification based on the co-
occurrence matrix approach. The above analysis, with experiment result support,
shows that the proposed design is of better performance and security. The features
extracted from the matrices can well characterize the regular texture of altered finger-
print images and Natural fingerprint images.
(a)input image (b)Output image –Natural
Fig. 5. Sample Output of fingerprint image
7 Conclusion
In this paper we have prompted GLCCM algorithm for altered fingerprints classifica-
tion. Experiments were conducted on both real world altered fingerprints. The pro-
posed method GLCCM gives better classification.
A New Classification Algorithm with GLCCM for the Altered Fingerprints 357
References
1. Jain, A.K., Feng, J.: Latent fingerprint Matching. IEEE Trans. on Pattern Analysis and
Machine Intelligence (2010)
2. Jain, A.K., Ross, A.: Detecting Altered Fingerprints. In: ICPR 2010, Istanbul, Turkey
(2010)
3. Bhuyan, M.H., et al.: An effective method for fingerprint classification. International Arab
Journal of e-technology 1(3) (January 2010)
4. Ne’ma, B., Ali, H.: Multipurpose code generation using fingerprint images. International
Arab Journal of Information Technology 6(4) (October 2009)
5. Ravi, J., et al.: Fingerprint Recognition Using Minutia Score Matching. International Arab
Journal of e-Technology 2(35-42), 89 (2009)
6. Yazdi, M., Gheysari, K.: A new approach for the fingerprint classification based on gray-
level co-occurrence matrix. World Academy of Science Engineering and Technology 30
(July 2008)
7. Wang, X., Long, H., Su, X.: Method of Image enhancement based on differential evolution
Algorithm. In: International Conference on Measuring Technology and Mechatronics
Automation (2010)
8. Sihalath, K., Choomchuay, S.: Fingerprint image enhancement with second derivative
Gaussian filter and directional wavelet transform. In: 2010 Second International Confer-
ence on Computer Engineering and Applications (2010)
9. Kukula, E.P., Blomeke, C.R., Modi, S.K., Elliott, T.J.: Effect of Human Interaction on Fin-
gerprint Matching Performance, Image Quality, and Minutiae Count. In: International
Conference on Information Technology and Applications, pp. 771–776 (2008)
10. Ji, L., Yi, Z.: Fingerprint Orientation field Estimation using Ridge Protection. The Jour-
nalof the Pattern Recognition 41, 1491–1503 (2008)
11. http://bias.csr.unibo.it/fvc2000/download.asp
12. http://bias.csr.unibo.it/fvc2002/download.asp
13. http://bias.csr.unibo.it/fvc2004/download.asp
14. http://www.neurotechnologija.com/download.html
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 358–367, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Footprint Based Recognition System
V.D. Ambeth Kumar1 and M. Ramakrishan2
1 Research Scholar, Sathyabama University, Chennai, India
ambeth_20in@yahoo.co.in
2 Professor and Head, Velemmal Engineering College, Chennai, India
ramkrishod@gmail.com
Abstract. This paper proposes a new method of personal recognition based on
footprints. In this basis method, an input pair of raw footprints is normalized,
both in direction and in position. The foot image is segmented and its key points
are located. The foot image is aligned and cropped according to the key points.
The footprint image is enhanced and resized. Sequential Modified Haar trans-
form is applied to the resized footprint image to obtain Modified Haar Energy
(MHE) feature. The sequential modified Haar wavelet can map integer-valued
signals onto integer-valued signals abandoning the property of perfect recon-
struction. The MHE feature is compared with the feature vectors stored in
database using Euclidean Distance. The accuracy of the MHE feature and Haar
energy feature under different decomposition levels and combinations are com-
pared. More than 88.37% accuracy achieved from the proposed MHE feature.
Keywords: Footprint, Sequential Haar Transform network, Heel Shape.
1 Introduction
Innumerable automated biometrics based identification and verification systems have
been developed [1]. The biometrics features derived from fingerprints [2][3] faces [2],
palm print[4], irises, retinas, a speaker’s voice, and perhaps a variety of other charac-
teristics. The systems are now used in a wide range of environments, such as law
enforcement, social welfare, banking, and various security applications. Practical meth-
ods for automated biometrics based identification have not yet been developed for use
with unconstrained subjects. The main Problem in automatic personal identification is
how to verify the sampled feature against the registered feature with high reliability.
Some systems for personal identification use fundamental biometric features derived
from fingerprints and irises. To acquire the feature, subjects must input their biometrics
to a sensor. Signatures [5] and speakers’ voices [6] are also features useful for verifica-
tion; however, obtaining these features requires the subjects’ cooperation. Automatic
face recognition based on vision technique can be a promising method since it can
work without any help user [7] but changing elimination, occlusion and hair style
change are still very difficult problem. Footprint based recognition is another emerging
method which does not need any operation from user. A high recognition rate by veri-
fying raw footprints directly is difficult to obtain, because people stand in various posi-
tions with different distances and angles between the two feet. To achieve robustness
Footprint Based Recognition System 359
in matching an input pair of footprints with those of registered footprints, the input pair
of footprints must be normalized in position and direction. Such normalization might
remove useful information for recognition, so geometric information of the footprint
prior to normalization into an evaluation function for personal recognition decision is
included. In this paper, we propose a footprint-based personal recognition method and
test its reliability. Footprint texture features are usually extracted using transform-
based method such as Fourier Transform [8] and Discrete Cosine Transform [9]. Be-
sides that, Wavelet Transform [10] is also used to extract the texture features of the
footprint. In this work, a Sequential modified Haar Wavelet is proposed to find the
Modified Haar Energy (MHE) feature. Fig 1,2 shows the proposed footprint identifica-
tion using Sequential modified Haar Transform.
Fig. 1. Footprint identification using Sequential Modified Haar Wavelet
Fig. 2. Footprint identification using Feature Extraction
RGB to Gray
Scale conver-
sion
Image
acquisition
Image
Cropping
LPCC and
LPC
Transformation
Techni
q
ues
Modified
Haar Energy
(MHE)
Foot
p
rint
Database
storage
Pressure Sensin
g
Mat
Foot
p
rint Ima
g
e
Feature Extraction
Feature Extraction
Acce
p
t or Re
j
ect
Database
Matchin
g
360 V.D. Ambeth Kumar and M. Ramakrishan
In this work, hundred images of left leg of 30 people are acquired using a pressure
sensor. The foot image is aligned according to the keypoints and it is cropped. The
energy features of the footprint are extracted using Sequential modified Haar Wave-
let. The Modified Haar Energy (MHE) is represented using feature vector and com-
pared using Euclidean Distance with the threshold values stored in the database. Fig 2
shows the process of footprint identification using feature extraction. The footprint
image captured using pressure sensor is extracted for its features. The already existing
features of footprint images stored in database are compared. Then, the footprint im-
age is accepted or rejected based on the comparison.
2 Proposed Method
To obtain a footprint image, we use a pressure-sensing mat [11]. The pressure distri-
bution of a footprint is represented as a monochrome image f(x, y): -X/2< x <= X/2, -
Y/2<y <=Y/2 (Where X and Y are widths of the mat). Details of the sensor and
method for acquiring the footprint are described in the next section. In this section, we
describe a normalization procedure and a recognition method.
2.1 Normalization
2.1.1 Image Acquisition
Footprints are normalized both in direction and in position, so that the subject does
not need to always stand in the same position as when first recorded. The normaliza-
tion procedure is as follows:
Firstly, the position of the center of mass of the whole footprint image is calculated
and moved to the center of the frame. The image of the footprint remains otherwise
unaltered.
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎝
⎛
∑
∑
∑
∑
∑∑
yx
xy
yx
yx
yxf
yxfy
yxf
yxfx
,,
),(
),(
,
),(
),(
Secondly, a scanning line method is employed which gives a number for each seg-
ment of the footprint. When the scanning line meets a segment, a number is given to
the segment. Each segment is copied into two new frames, the left footprint image
L(x,y) or the right footprint image R(x,y), by considering the sign of the x-coordinate
of the center of mass of each segment. For example, if the x-coordinate of the
center of mass of a segment has a negative sign, the segment is assumed to belong to
the left foot.
Thirdly, the center of mass of the left footprint, including all segments of the left
foot, is calculated and moved to the center of the frame. The left footprint remains
otherwise unaltered. The center of mass of the left footprint is indicated by “+ ” and
that of the right footprint by “x” , which is shown in the diagram Fig.3.
Footprint Based Recognition System 361
Fig. 3. Calculation of the center of foot
Fourthly, by using an orthogonal transformation (x, y) is normalized in direction.
First, we obtain the covariance matrix summation of L(x, y)
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎝
⎛
=∑∑
∑
∑
yxyx
yxyx
LyyxLxyyxL
xyyxLxyxL
H
,
2
,
,,
2
),(),(
),(),(
The principal components (the eigenvectors ordered by highest to lowest correspond-
ing eigen value), eL1and eL2 , are extracted. Assuming that the footprint is an oval
shape, eL1and eL1 show the major and minor axes, respectively. The diagonal matrix
uL=(eL1eL2) is defined from two eigenvectors eL1 and eL2. Image L(x,y) is rotated
so that the first eigenvector eL1 fits the vertical direction by the orthogonal transfor-
mation using uL. The same procedures, from Step 3 to Step 4 (Shown in Fig. 3,
Fig. 4), are carried out for the right footprint image R(x,y).
Fig. 4. Normalization in direction
362 V.D. Ambeth Kumar and M. Ramakrishan
Finally, the footprints of both feet are normalized in position. Both left and right
footprints are integrated into a reconstructed image, I(x,y), with a fixed distance, T
(Shown in Fig. 5). The center of mass of the two feet is the center of the frame.
Fig. 5. Reconstructed whole footprint image
2.1.2 Image Cropping
After acquiring the foot image, the next step is to crop the image. Cropping can be
done by keypoints determination and extracting the image. The foot image is cropped
to extract the heel shape as the intensity is highest in the heel portion. Fig. 6 shows
the extracted footprint image using the proposed method.
Fig. 6. Footprint image
2.1.3 Conversion to Grayscale Format
The footprint image is acquired in RGB format. It is converted into grayscale inten-
sity format before image enhancement. Fig. 7 shows the footprint image in grayscale
format.
Footprint texture features are usually extracted using transform-based method such
as Fourier Transform [8] and Discrete Cosine Transform [9]. While using Discrete
Cosine Transform, some of the points are missed leading to an inaccurate inference.
Also Fourier Transform involves floating-valued signals to integer-valued signals,
thus less accuracy. Besides that, Wavelet Transform [10] is also used to extract the
texture features of the footprint. In this work, a Sequential modified Haar Wavelet is
proposed to find the Modified Haar Energy (MHE) feature [12].
Footprint Based Recognition System 363
Fig. 7. Footprint image in Grayscale Intensity format
The Haar wavelet coefficients are represented using decimal numbers. It required
eight bytes for storing each of the Haar coefficients. The division of the subtraction
results in y-axis operation will also reduce the difference between two adjacent pixels.
Due to these reasons, sequential Haar wavelet that maps an integer-valued pixel onto
another Integer-valued pixel is suggested. Sequential Haar coefficient requires only
two bytes to store each of the extracted coefficients. The cancellation of the division
in subtraction results avoids the usage of decimal numbers while preserving the dif-
ference between two adjacent pixels. Fig.8. shows the decomposition using sequential
Haar wavelet.
Fig. 8. Sequential Haar Transform
where X or Y-axis represents the direction of addition and subtraction. “/2” is the
rounding of the “divide by 2” results towards the negative infinity. Fig 9 shows the
location of horizontal details, vertical details and diagonal details for every decompo-
sition levels.
For every image of the detail coefficients, the image is further divided into 4 x 4
blocks. The Modified Haar Energy for each of the block is calculated using (1).
()
∑∑
==
=P
p
Q
q
qpkji CMHE
11
2
,,,
(1)
where i is the level of decomposition, j is Horizontal, Vertical or Diagonal details, k is
the block number from 1 to 16, P x Q is the size of the block. The MHE energy fea-
ture for every detail coefficients are arranged as in (2).
MHEi.j=[MHEi,j,1,MHEi,j,2,…..,HEi,j,16] (2)
364 V.D. Ambeth Kumar and M. Ramakrishan
Fig. 9. Location of details in Every Decomposition Levels
2.1.4 Storing
Lastly, the Modified Haar Energy (MHE) or the threshold value of the footprint im-
age is stored in the database for future references.
B. Recognition Method
2.2 Recognition Method
Recognition Method is also called as Testing. The normalization procedure removes
the geometric information of input footprint (raw image). In Testing, normalization
is done for the footprint image and then it is compared with the one stored in the
database.
2.3 Experiments and Results
For experiment we took image samples of 30 different persons in 10 different angles
using a pressure sensor. We used the block to get the conclusion.
2.3.1 Training
We took image samples of 30 different people and cropped heel portion to find the,
MHE of all blocks. We took a number of MHE out of which we select the minimum.
Let MHE1,MHE2, MHE3 …………….etc be the Modified Haar Energy for n blocks.
Using the formula,
MHE=Minimum (MHE1,………….)
For each person, the MHE is found in this way.
These are stored in the database.
VPerson1= Minimum(MHE1……………..)
VPerson2= Minimum(MHE2……………..)
…..
VPerson30= Minimum(MHE30…………)
Footprint Based Recognition System 365
2.3.2 Testing
The footprint to be tested was taken in 10 different angles and heel part was cropped
so as to get the MHE of the heel shape using:
MHE=minimum (MHE1,………)
This test MHE was compared with the MHE of different persons stored in the data-
base.
Result=TestMHE-VPersoni MHE
Where i=1 to 30
If result is 0, the person with same footprint is found. That person is the master of the
test footprint sample.
Comparison with other methods
The accuracy of Recognition of Footprint images using Sequential Modified Haar
Transform is compared with the other methods for recognition. The following table
shows the comparison of accuracy of recognition for Sequential Modified Haar Trans-
form with Discrete Cosine Transform and Fourier Transform.
Table 1. Comparison of accuracy (performance) of the proposed method with other wavelet
based method
Transform Types Recognition
Accuracy (%)
Computation time in ms
(Recognition)
DCT [Ref.6] 81.64 152
FT [Ref.5] 84.43 138
SHT [Proposed] 88.37 128
The corresponding graph for the accuracy rate of recognition footprints using
above mentioned techniques in as follows
78
80
82
84
86
88
90
Accurancy (%)
DCTFTSHT
Transfor m types
Fig. 10. Comparison of existing transforms (DCT, FT) Accuracy in % with proposed one
(SMHET) The corresponding graph for the computation time (ms) of recognition footprints
using above mentioned techniques in as follows
366 V.D. Ambeth Kumar and M. Ramakrishan
115
120
125
130
135
140
145
150
155
DCTFTSHT
Transform types
Computati on ti me(ms)
Fig. 11. Comparison of existing transforms (DCT, FT) Computation time (ms) with proposed
one (SMHET)
3 Conclusion
The heel portion of the leg is cropped as it is having more intensity at this portion.
This cropping is done using built-in function. The heel portion is divided into blocks
using Sequential Modified Haar Transform. Minimum MHE is selected from all the
calculated MHEs. Comparing the MHE of test image with all person’s MHEs. If both
MHEs match then the master of the footprint is found. This is an efficient method as it
is more accurate. A high accuracy rate of 88.37% is achieved using Sequential Modi-
fied Haar Transform. Future enhancement of this project can be finding the age,
weight and BP a person.
References
1. Special issue on automated biometrics. Proc. IEEE 85, 1348–1491 (1997)
2. Jain, L.C., et al. (eds.): Intelligent Biometric Techniquesin Fingerprint and Face Recogni-
tion. CRC Press, Boca Raton (1999)
3. Masood, H., Mumtaz, M., Khan, S.A.: Wavelet based plamprint authentication system. In:
IEEE Trans. Biometric and Security Tech., pp. 1–7 (2008)
4. Li, W., Zhang, D., Xu, Z.: Plamprint identification by recognition and artificial intelligence
16(4), 417–432
5. Plamondon, R., Lorette, G.: Automatic signature verification and writer identification–The
state of the art. Pattern Recog. 22, 107–131 (1989)
6. Bimbot, F., Hutter, H.P., Jaboulet, C., Koolwaaij Lindberg, J., Pierrot, J.B.: Speaker Veri-
fication in the Telephone network: Research activities in the CAVE project. In: Proc. 5th
Int. Conf. Speech Communication and Technology, Rhodes, Greece (September 1997)
7. Sukthar, R., Stockenton, R.: Argus the digital doorman. IEEE, Intelligient System 2 (2001)
8. Li, W., Zhang, D., Xu, Z.: Palmprint Identification By Fourier transform. Intl. Journal of
Pattern Recognition and Artifical Intelligence 16(4), 417–432 (2002)
9. Jing, X.-Y., Zhang, D.: Face and Palmprint Recognition Approach Based on Discriminant
DCT Feature Extraction. In: IEEE Trans.on Systems, Man. and Cybernetics –Part B: Cy-
bernetics, December 2004, vol. 34(6), pp. 2405–2415 (2004)
Footprint Based Recognition System 367
10. Wu, X.-Q., Wang, K.-Q., Zhang, D.: Wavelet Based Palm print Recognition. In: Proceed-
ings of the First Intl. Conference on Machine Learning and Cybernetics, November 2002,
pp. 1253–1257 (2002)
11. Nakajima, K., Mizukami, Y., Tanaka, K., Tamura, T.: Foot-Based Personal Recognition.
IEEE Tr. On Biomedical Engineering 47(11) (2000)
12. Ambeth Kumar, V.D., Ramakrishan, M.: Footprint Recognition using Modified Sequential
Haar Energy Transform (MSHET). Intl. Journal of Computer science Issuse 7(3/5) (2010)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 368–374, 2011.
© Springer-Verlag Berlin Heidelberg 2011
An Efficient Approach for Data Replication in
Distributed Database Systems
Arun Kumar Yadav1, Ajay Agarwal2, and S. Rahmatkar3
1 Dept. of Computer Science, BMAS Engineering College, Agra,(U.P.), India
arun26977@rediffmail.com
2 Dept. of MCA, Krishna Institute of Engg. & Tech., Ghaziabad (U.P.), India
3 Dept. of Computer Science, Nagpur Institute of Technology, Nagpur, India
Abstract. To increase the availability of data and fault tolerance in distributed
database, it is better to add a backup server for each primary server in the sys-
tem. So, the primary server and backup server must be connected to each other.
To connect these computers to each other when they are at a long distance, it is
necessary to use a leased line which needs to be charged as data is transferred.
When packets are transferred between primary and backup server, more money
need to be paid for charging this line. So if number of transferred packets be-
tween these computers can be reduced, the company can economize in its ex-
penditures. Moreover, the number of transactions which should be performed in
the backup server reduces.
In order to achieve this, a new method is introduced which will reduce the
number of transferred packets between primary and backup server. In this
method, the replicated data of primary server on other computers is used as its
backup.
Keywords: Distributed Database, Remote backup, Data replication, Load bal-
ancing, filtering.
1 Introduction
One of the non-functional requirements in distributed databases is performance [1]. A
technique to improve this requirement is data replication [2]. Data replication maintains
multiple copies of data, called replicas, on separate servers. So the requests to these
servers are answered locally [3]. Data Replication improves performance by the follow-
ing: i) reducing latency, since users can access replicated data, so it avoids remote net-
work access; and ii) increasing throughput, since data are available on multiple com-
puters and can be accessed simultaneously. When data is replicated on more than one
computer, it is necessary to ensure its consistence. There are some protocols such as
single lock, distributed lock, primary copy, majority protocol, biased protocol, and quo-
rum consensus protocol [4] which are responsible to ensure the data consistence.
Another non-functional requirement in distributed system is fault tolerance. So it
ensures that the system can work accurately even in case of occurrence of faults
[5, 6]. We can achieve this ability by performing transaction processing at one server,
called the primary server, and having a remote backup server where all data of the
An Efficient Approach for Data Replication in Distributed Database Systems 369
primary server are replicated. When a primary server fails, its remote backup server is
responsible to answer the requests until the primary server come back to stable state
[4]. Figure 1 shows the architecture of a remote backup system.
Fig. 1. Remote Backup System
In this paper we are introducing a technique which uses backup server data to re-
cover the primary server from failure. The rest of the sections are organized as fol-
lows: Section 2 expresses related work. Section 3 describes how to use replicated data
in backup process. Section 4 includes implementation remarks. In section 5, we have
evaluated our technique and At last in section 6, some constrains in our technique and
conclusion of the paper is presented.
2 Related Works
One of the advantages with distributed databases is high availability of data; that is,
the database must function almost all the time. In particular, since failures are more
likely in large distributed systems, a distributed database must continue functioning
even when there are various types of failures. The ability to continue functioning even
during failures is referred to as robustness. Whenever a transaction writes an object it
also updates the version number in this way:
If data object A is replicated on n different sites, then a lock-request message must
be sent to more than one-half of the n sites in which a is stored. The transaction
does not operate on A until it has successfully obtained a lock on a majority of the
replicas of a.
Read operations look at all replicas on which a lock has been obtained, and read the
value from the replica that has the highest version number.
Writes read all the replicas just like reads to find the highest version number. The new
version number is one more than the highest version number. The write operation
writes all the replicas on which it has obtained locks, and sets the version number at
all the replicas to the new version number. The read one, write all available scheme
can be used if there is never any network partitioning, but it can result in inconsisten-
cies in the event of network partitions [4].
The Database Mirroring method is one of the best methods for backup data. In the
simplest operation of database mirroring, there are two major server-side components,
the principal server instance (Principal) and the mirror server instance (Mirror) [7].
The basic idea behind database mirroring is that synchronized versions of the
database are maintained on the principal and mirror. If the principal database becomes
unavailable, then the client application will smoothly switch over to the mirror
database, and operation will continue as normal. The mirroring architecture of data-
base is shown in figure 2 [7].
370 A.K. Yadav, A. Agarwal, and S. Rahmatkar
Fig. 2. Mirroring Architecture of Database
3 Data Replication to Reduce Backup Cost
To apply our method in the distributed database, some operations in the primary
server, backup server, and guest computer must be changed or added. In the following
sections we describe these necessary changes.
3.1 Changes Required in the Primary Server
As we mentioned, the primary server transfer the transactions on data which is not
replicated in the guest computers. So the primary server and backup server should
know which data item is replicated on which guest computer. Hence when the pri-
mary server replicates a data item to a guest computer, assigns a name to that data and
send this name along data item to the guest computer. After that, the primary server
sends a packet to backup server which includes name of replicated data and the guest
on which is replicated.
If a guest computer requests a data item which is named already, the primary
server does not name it again. The primary server sends data item with its name to the
guest. It also resends data item name and new guest computer which requests that data
to the backup server.
3.2 Changes Required in the Backup Server
As we mentioned, some data on backup server in not updated. Instead, it receives
from the primary server name of this data and the guest computer which has this data.
So when the primary server fails, backup server has to get these data from appropriate
guest and updates its old data by new data. If a data item is replicated to more than
one guest, backup server should receive that data item form a guest which has the last
version of that data.
An Efficient Approach for Data Replication in Distributed Database Systems 371
3.3 Changes Required in the Guest Computers
When a guest computer requests a data form the primary server to replicate, it re-
ceives name of that data along with data. The guest computer should save that name
along data. So in the failure time, if the backup server requests a data item by its
name, the guest computer know which data corresponds with that name, and transfers
the requested data to the backup server.
4 Implementation Remarks
As we have mentioned, the primary sever should name data item which is being repli-
cated on the guest computers. To name data items, primary server can use a counter; a
data item is named according to the counter. After naming a data item, the counter is
increased by one. To assign this name to data items, for example in a relational data-
base, it is better to add a column named data_name to each table in database. So name
of each record is inserted in this column. All records in a data item have the same
name and the value of data_name attribute of these records is equal to each other.
There is a database named Student in the primary server with some records in it is
shown in Table 1.
Table 1. Student Table
Row No. Name of Student Stud_Number Data_Name
1 Nikhil 1001
2 Bharat 1002
3 Jayesh 1003
4 Ankur 1004
5 Nititn 1005
To clarify the issue, suppose a primary server, its backup server, and two guest
computers are connected to each other. Suppose guest 1 requests records number 1, 3,
and 5 to replicate from the primary server. The primary server images these records as
a data item and names it Q1. So the value of data_name attribute for these records is
equal to Q1 and this name is sent to the guest 1 and backup server. This process is
shown in figure 3.a, suppose guest 2 request records number 3, 2, 4. Records number
3 is member of data item Q1, so the primary server send all records in data item Q1 to
the guest 2. After that it assign a name Q2 to records number 2 and 4, and send them
to the guest 2. This process is shown in figure 3.b.
So if a guest requests from the primary server some records which are subset of a
data item Q1, all records in that data item are sent to the guest.
Now suppose the primary server fails. The backup server knows that data item
Q2 is replicated on guest 2, and requests this data item from guest 2. Data item Q1 is
replicated in both guest 1 and 2. The primary requests this data item from the guest
which has the last version of that data item. So the backup server requests this data
item from guest 2 and updates its data. This process is shown in figure 4.
372 A.K. Yadav, A. Agarwal, and S. Rahmatkar
Fig. 3. Replicating Data to Client 1 and Client 2 from Primary Server
Fig. 4. Transferring Data to the Backup Server
5 Evaluations
In this section, we are evaluating our technique. As mentioned in the previous section,
when the primary server data is replicated on some guest computers, this replicated
data can be used to recover system from a failure state.
Firstly we are measuring the network traffic between primary and backup server in
mirroring method. In this method primary server sends the transactions on data to the
backup server in specific interval time. After that we will apply our method to the
system and will measure the network traffic which is imposed to the system. Finally
we can compare the network traffic in our method and mirroring method.
To evaluate our method we are using SQL Server 2005 database system. Our pri-
mary server has 200000 records. We replicate 500 records to a guest computer and
perform 50 update transactions to these records. The network traffic to update backup
server in the mirroring method is equal to 250 Kbyte. If we apply our method to the
system, this value is equal to 55 Kbyte. We continue to increase replicated data and
perform the same 50 update transaction to them and monitor network traffic in our
An Efficient Approach for Data Replication in Distributed Database Systems 373
method and mirroring method. The results are shown in table 2 and figure 5 shows
this result as a diagram. We find that the amount of replicated data effects in perform-
ance in our method, and improves it.
Table 2. Network Traffic and replicated data comparison in mirroring and our method
Network Traffic Between Primary
Server and Backup Server (Mbyte)
Proposed
Method
Mirroring
Method
Number of
Records
Replicated
0.055 0.25 500
0.102 0.25 1000
0.508 3.125 5000
1.016 5.75 10000
3.555 16.812 30000
5.078 30.687 50000
10.148 55.312 100000
Fig. 5. Diagram of relationship between network traffic and replicated data in the mirroring and
our method
6 Method Constrains and Conclusion
There are some constraints in our method. First, if the primary server and one of the
guest computers which has data which is replicated only on that guest computer get
fail concurrently, the backup computer cannot recover all data, and as a result data
may be inconsistent. As you know this condition takes place very seldom. Second,
our method imposes some over load in the primary server. So the primary server
should have a suitable hardware to continue its work. Third, the recovering system
from failure needs more time than previous methods. Because the backup computer
should gets updated information from the guest computers, if the number of guest
computers is high, this process may needs very much time.
374 A.K. Yadav, A. Agarwal, and S. Rahmatkar
This paper presents a new method to decrease network traffic between primary
server and backup server. In this method replicated data is used for backup process.
Transactions on data which is replicated on the guest computers are not transfer from
primary server to backup server. Instead, at the time of primary server fails, this repli-
cated data are transferred from the guest computers to the backup server.
Reduction of network traffic between primary server and backup server causes
fewer transactions execute in the backup server. In addition, the company has to pay
less money for charging the link.
References
1. Sleit, A., AlMobaideen, W., Al-Areqi, S., Yahya, A.: A Dynamic Object Fragmentation
and Replication Algorithm in Distributed Database Systems. American Journal of Applied
Sciences 4(8), 613–618 (2008)
2. Loukopoulo, T., Ahmad, I.: Static and adaptive distributed data replication using genetic
algorithms. Journal of Parallel and Distributed Computing 64(11), 1270–1285 (2004)
3. Abdul-Wahid, S., Andonie, R., Lemley, J., Schwing, J., Widger, J.: Adaptive Distributed
Database Replication Through Colonies of Pogo Ants. In: IEEE International Parallel and
Distributed Processing Symposium, IPDPS 2007, pp. 358–365 (2008)
4. Silberschats, A., Korth, H.F., Sudarshan, S.: Database System Concepts. McGraw-Hill,
New York (2006)
5. Gashi, I., Popov, P., Strigini, L.: Fault tolerance via diversity for off-the-shelf products: A
study with SQL database servers. IEEE Transactions on Dependable and Secure Comput-
ing 4(4), 280–294 (2007)
6. Wang, Mueller, F., Engelmann, C., Scott, S.: A job pause service under lam for transparent
fault tolerance. In: IEEE International Parallel and Distributed Processing Symposium,
pp. 1–10 (2007)
7. Rizzo, T., Machanic, A., Skinner, J., Davidson, L., Dewson, R., Narkiewicz, J., Sack, J.,
Walters, R.: Pro SQL Server 2005, ch. 15. A Press (2006)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 375–378, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Unicast Quality of Service Routing in Mobile Ad Hoc
Networks Based on Neuro-fuzzy Agents
V.R. Budyal1, S.S. Manvi2, and S.G. Hiremath3
1 Basaveshwar Engineering College, Bagalkot, India
vrbudyal@yahoo.co.in
2 Reva Institute of Technology and Management, Bangalore, India
agentsun2002@yahoo.com
3 G. M. Institute of Technology, Davanagere, India
Department of Electronics and Communication Engineering
Abstract. This position paper presents Quality of Service (QoS) routing model
in Mobile Ad hoc Networks (MANETs) by using software agents that employ
fuzzy logic and neural networks for intelligent routing. It uses Dynamic Source
Routing (DSR) in MANETs to find various paths and attributes. Fuzzy static
agents decide whether each node on the path satisfies QoS requirement for mul-
timedia application. The static neuro-fuzzy agents are used for training and
learning to optimize the input and output fuzzy membership functions accord-
ing to user requirement, and Q-learning (reinforcement learning) static agent is
employed for fuzzy inference instead of experts experience. Mobile agents are
used to maintain and repair the paths.
Keywords: mobile ad hoc networks, routing, neuro-fuzzy, software agents.
1 Introduction
A Mobile Ad hoc network (MANET) is a self configuring network consisting of mo-
bile devices connected by wireless links. Each device in a MANET is free to move
independently in any direction. Each node acts as host and a router. The QoS routing
protocol is also needed in a stand-alone multi-hop mobile network for real-time appli-
cations (like voice, video, etc.). QoS routing requires not only finding a route from a
source to destination, but a route that satisfies the end-to-end QoS requirement, often
given in terms of bandwidth or delay [1].
Recently software agents are considered to be one of the important paradigms for
providing flexible and adaptable services in intelligent communication networks.
Agents are autonomous programs activated on an agent platform of host. The agents
use their own knowledge base to achieve the specified goals without disturbing the
activities of the host [2]-[4].
The problem addressed in the paper is to route the packets between source and des-
tination by path discovered using DSR (based on DSR [5]) to satisfy QoS parameters.
Our contributions in the paper are as follows: building the neuro-fuzzy agents for op-
timizing the input and output fuzzy membership functions using feed forward back
376 V.R. Budyal, S.S. Manvi, and S.G. Hiremath
propagation learning, designing fuzzy inference based on Q- learning agent, finding
the QoS satisfied path from source to destination with multi-metric uncertain input
parameters using fuzzy agents and maintaining the QoS path whenever there is failure
either in link or node using mobile agents.
2 Proposed Work
An agent based model comprising of User Agency, DSR Agency and QoS Agency for
QoS routing in MANETs at each node is shown in Fig.1. The knowledge base facili-
tates inter-agent communication.
Fig. 1. Components of Agency
2.1 User Agency
User Agency consists of User Manager Agent and User Agent. User Manager Agent
is a static agent, which is directly accessible to an application, and creates a User
Agent whenever there is request from the user for multimedia data transmission with
specific QoS requirement and it also, triggers Qos Agency and DSR Agency. User
Agent holds the requirement of the user for particular application.
2.2 DSR Agency
DSR Agency comprises of static DSR Manager Agent, DSR Agent, Node Dis-joint
Multipath Routing and Stability agent and Maintenance mobile Agent. DSR Manager
Agent is a static agent. It accepts destination node address from User Manager Agent
and creates and co-ordinates with DSR Agent, NDMR and Stability Agent and Main-
tenance Agent, to obtain stable NDMR path. DSR Agent is static agent. It sends
Route REQuest (RREQ) to its neighbours till destination is found. Route REPly
(RREP) carries information of available bandwidth and packet loss rate at each inter-
mediate node which is saved in DSR Manager Agent.
Unicast Quality of Service Routing in Mobile Ad Hoc Networks 377
Node Dis-joint Multipath Routing and Stability agent is a static agent that performs
the following tasks at the node it is residing. It gets all multipaths from DSR Manager
Agent. NDMR and Stability Agent compute to find node dis-joint paths. The other
task performed by NDMR and Stability Agent is to compute stable path from number
of NDMR paths. DSR Manager Agent decides higher stability path and sends mes-
sage to QoS Agency. Maintenance Agent is a mobile agent that migrates from node to
node periodically on the QoS satisfied route. In case of either mobility or violation of
QoS satisfaction of intermediate node, application may use either new path from DSR
Manager Agent, which satisfies QoS, or local recovery patch up paths are used to,
further continue the communication.
2.3 QoS Agency
QoS Agency comprises of QoS Manager Agent, Neuro-fuzzy Agent and Fuzzy QoS
agent. QoS Manager Agent is a static agent creates Neuro-fuzzy Agent, Fuzzy QoS
agent and Q-learning Agent. Neuro-fuzzy Agent updates optimized fuzzy member-
ship functions in QoS Manager Agent. Q-learning Agent infers the rules of the fuzzy
if-then rules. Fuzzy QoS Agent maps available bandwidth and packet loss on to fuzzy
logic and obtains the crisp value. Neuro-fuzzy Agent is a static Agent that performs the
following tasks at the node it is residing. It generates the training data set and learns
by feed forward backward propagation, which consists of five layers. Neuro-fuzzy
Agent computes output at each layer till the last layer which is compared with desired
value and obtains error. To distribute the error in the hidden layer, Neuro-fuzzy Agent
back propagates the error at each layer and updates the weights.
The same procedure is repeated till error is within the acceptable limits. QoS
Manager Agent holds the optimized input and output fuzzy membership functions.
Q-Learning Agent is a static agent that performs the tasks like evaluating the conse-
quences for the rule, computing the value of Q with number of iterations until the
value Q stabilizes. Consequence computed by Q- learning Agent is updated in Fuzzy
QoS Agent. Fuzzy QoS Agent is static agent which performs the following tasks: (1) It
fuzzifies the input variables on optimized membership functions. (2) Applies if-then
rules to obtain fuzzy inference for each rule where fuzzy inference is obtained from
Q-learning Agent. (3) Fuzzy QoS agent computes defuzzified value using centroid
method to decide QoS satisfaction node.
An algorithm to depict the functioning of the model is as follows.
Algorithm. Functioning of QoS path discovery
1. : Begin
2. : Accept application route request and QoS requirement;
3. : Compute Stable NDMR paths from DSR multipaths;
4. : Optimize input and output fuzzy membership functions based on QoS
Requirement using neuro-fuzzy agents;
5. : Learn fuzzy inference by Q- learning agent;
6. : Check QoS satisfaction using fuzzy logic at each node for the complete stable
NDMR path selected;
7. : IF {node satisfies QoS}
8. : THEN {check QoS for next intermediate node};
378 V.R. Budyal, S.S. Manvi, and S.G. Hiremath
9. : ELSE {accept next higher stable NDMR path, go to step 6};
10. : Start sending the multimedia data packets on QoS satisfied path;
11. : End.
3 Conclusion
Proposed model optimizes input and output fuzzy membership functions and infer-
ence based on neuro-fuzzy agents and Q-learning agents respectively. For QoS mul-
timedia applications requirement, proposed scheme takes appropriate fuzzy decisions
and maintains the path when there is a path failure using mobile agents. Using this
approach, we can guarantee the QoS routing for MANETs. We are planning to simu-
late the model using ns-2 and evaluate the performance of model in terms of applica-
tion rejection ratio, agents overhead and route discovery time.
References
1. Kannan, T., Pushpavalli, M., Natarajan, A.M.: Fortification of QoS routing in MANETs us-
ing Proactive Protocols. In: IEEE International Conference on Wireless Communication and
Sensor Computing (ICWCSC 2010), India (2010)
2. Budyal, V., Manvi, S.S., Hiremath, S.G., Kala, K.M.: Intelligent Agent based QoS Enabled
Dis-joint Multipath Routing in Manets. In: 17th IEEE International Conference on Ad-
vanced Computing and Communication, India (2009)
3. Wang, Z., Seitz, J.: An Agent based Service Discovery Architecture for mobile environ-
ments. In: 1st EuroAsian conference on Information and Communication Technology, UK,
pp. 350–357. Springer, Heidelberg (2002)
4. Manvi, S.S., Vekataram, P.: International Journal on Computer Communications 27(15),
1493–1508 (2004)
5. Zhong, Y., Yuan, D.: Dynamic Source routing Protocol for Wireless Ad hoc Networks in
special scenario using Location Information. In: International Conference on Communica-
tion Technology ICCT, vol. 2, pp. 1287–1290 (2003)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 379–383, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Performance Comparison of Routing Protocols in
Wireless Sensor Networks
Geetika Ganda1, Prachi2, and Shaily Mittal3
1 M-tech Student, 2 Assistant Professor, 3 Assistant Professor
Dept. of Computer Science, Dept. of Information Technology
Itm University, Gurgaon
geetikaganda@gmail.com, prachi@itmindia.edu, shally@itmindia.edu
Abstract. This paper aims to compare performance of some routing protocols
for Wireless Sensor Networks(WSNs). A Wireless Sensor Networks (WSN) is a
set of hundreds or thousands of micro sensor nodes that have capabilities of
sensing, establishing wireless communication between each other and doing
computational and processing operations. The efficiency of sensor networks
strongly depends on the routing protocol used. Routing protocols are needed to
send data between sensor nodes and the base station. In this paper, we analyzed
three different types of routing protocols: Fisheye, LANMAR, LAR1. Sensor
networks were simulated using Qualnet simulator. Several simulations were
conducted to analyze the performance of these protocols on the basis of per-
formance metrices such as hop count, throughput, end-to-end delay.
Keywords: WSNs, LANMAR, Fisheye, LAR1.
1 Introduction
Routing is a function in the network layer which determines the path from a source to
destination for the traffic flow. WSNs routing protocols are broadly divided into two
categories [2] i.e reactive(on-demand) and proactive(table-driven). In Table-driven rout-
ing protocols, each node maintains one or more tables containing routing information to
every other node in the network. All the nodes update these tables so that a consistent and
up-to-date network is maintained. In contrast to table-driven routing protocols, all up-to-
date routes are not maintained at every node; instead the routes are created as and when
required. Various reactive protocols are DSR, AODV, ABR, TORA etc. LSR, OLSR,
DSDV, LAR1, Fisheye, LANMAR, are proactive protocols. In this paper we have
used Fisheye, LANMAR and LAR1 i.e. unicast routing protocols for their performance
comparison. We have taken these protocols as these are not evaluated earlier for such
comparison.
An earlier protocol performance comparison was carried out by Guangyu Pei et all
in [10], who conducted experiments with Ad hoc On-Demand Vector routing (AODV),
Fisheye, Dynamic MANET On-demand (DYMO), Source Tree Adaptive Routing
(STAR) protocol, Routing Information Protocol (RIP), Bellman Ford, LandMark
Ad hoc Routing protocol (LANMAR) and Location Aided Routing protocol (LAR).
This simulation experiment showed that AODV, Dymo and Bellman ford protocols are
having higher end to end delays than others, indicating that the speed of simulation in
380 G. Ganda, Prachi, and S. Mittal
large scale networks will be affected, whereas LANMAR and RIP shows the consider-
able amount of delay in scaled up environment.
Performance comparison of AODV, DSR, FSR and LANMAR is presented by M.
Gerla et all in [11]. According to their simulation results LANMAR outperforms FSR
under all delay and throughput measures.In the last few years, there are several re-
searches have evaluated the performance of routing protocols for mobile Ad- Hoc
network as a function of mobility rate and pause time using ns2(network simulator
2)[9] .There are lesser evaluations available using Qualnet simulator [1] which is
commercially available and faster than ns2 [3]. We are using Qualnet simulator for
comparison evaluation of LANMAR, LAR1 and Fisheye.
DSR and DSDV were simulated and compared to a newly developed Cluster-based
Routing Protocol (CBRP) by Mingliang, Tay and Long [12]. The simulations were
performed with pause times from 0 to 600 seconds and with 25 to 150 mobile nodes.
Their results shows CBRP performed much better with a delivery ratio always greater
then 90 percent and a lower routing overhead than DSR in larger networks.
An earlier protocol performance comparison was carried out by authors in [12],
who conducted experiments with Destination Sequence Distance Vector (DSDV),
Temporally-ordered routing algorithm (TORA) along with DSR and AODV. The
simulations were quite different for they used a constant network size of 50 nodes, 10
to 30 traffic sources, seven different pause times and various movement patterns on
ns2 simulator.The rest of the paper is organized as follows: Section 2 describes three
concerned protocols in detail i.e. Fisheye, LANMAR and LAR1. Section 3 describes
the simulation environment, parameters evaluated and simulation results. Lastly work
is concluded in section 4.
2 Preliminaries
2.1 Fisheye
Fisheye technique proposed by Kleinrock and Stevens [4] to reduce the size of infor-
mation required to represent graphical data. The eye of a fish captures with high detail
the pixels near the focal point. The detail decreases as the distance from the focal
point increases Fisheye State Routing (FSR) [4] generates accurate routing decisions
by taking advantage of the global network information. Fisheye Routing determines
routing decisions using a table-driven routing mechanism similar to link state.
2.2 LANMAR
The Landmark Ad-hoc Routing Protocol (LANMAR) [5] combines the features of
FSR and landmark routing. LANMAR assumes that the large scale ad hoc network is
grouped into logical subnets in which the members have a commonality of interests
and are likely to move as a “group”. LANMAR uses the notion of landmarks to keep
track of such logical subnets[6].The route to a landmark is propagated throughout the
network using a Distance Vector mechanism [11] The routing update exchange of
LANMAR routing can be explained as follows. Each node periodically exchanges
topology information with its immediate neighbors. In each update, the node sends
entries within its Fisheye scope [4]. Updates from each source are sequentially num-
bered. To the update, the source also piggybacks a distance vector of all landmarks.
Performance Comparison of Routing Protocols in Wireless Sensor Networks 381
As a result, each node has detailed topology information about nodes within its Fish-
eye scope and has a distance and routing vector to all landmarks.
2.3 LAR1
The goal of Location-Aided Routing (LAR)[7] is to reduce the routing overhead by
the use of location information. LAR protocol uses the GPS (Global Positioning Sys-
tem) to get location information of mobile hosts. In the LAR routing technique,[8]
route request and route reply packets similar to DSR and AODV are being proposed.
3 Performance Evaluation
We carried out simulations on Qualnet simulator.We designed the network using
Random waypoint model with different number of nodes. We compiled the results
using 5 simulations and the application traffic between the randomly chosen source
and destination is CBR traffic[7]. The metrics used to measure the performance of
protocols are average end to end delay, average TTL based hop count and throughput.
3.1 Simulation Results
3.1.1 Average End to End Delay
End-to-end delay indicates duration for a packet to travel from the CBR source to the
application layer of the destination. According to results obtained in figure 1 LAN-
MAR shows minimum end to end delay of 0.015 s and almost remains constant irre-
spective of increase in no. of nodes. Similar is the case with LAR1 protocol which
shows a slight higher delay in comparison with LANMAR. FSR shows worst per-
formance with highest end to end delay of 0.023s.
3.1.2 TTL Based Average Hop Count
Hop count is the number of hops a packet took to reach its destination. The results for
TTL based hop count in figure 1 shows three protocols have a constant hop count
irrespective of the no. of nodes. With the increase of no. of nodes it remains constant.
LANMAR have highest hop count of 64 hops while FSR and LAR1 require less
number of hops of 19 hops and 1 hop respectively. The plotted graph shows that hop
count is independent of no. of nodes. Hence, according to the results dawn, it is clear
that these three protocols are independent of scalability.
3.1.3 Throughput
Throughput is the average rate of successful message delivery over a communication
channel. This data may be delivered over a physical or logical link, or pass through a
certain network node. It is usually measured in bits per second. The results of
throughput in figure 1 shows that LAR1 performs the best with the highest throughput
in spite of increased no. of nodes while FSR drops down to approx. near zero.
382 G. Ganda, Prachi, and S. Mittal
a) Hop Count b) Throughput c) End to end Delay
Fig. 1. Simulation results of FSR, LANMAR AND LAR1
4 Conclusion
In this paper, a performance comparison of three different routing protocols i.e. FSR,
LANMAR, and LAR1 for wireless sensor network is presented. Three performance
metrics used to compare protocols are average end to end delay; average TTL based
hop count and throughput. LANMAR performs best in measuring end to end delay
and FSR performs best in TTL based hop count. LAR1 performs best in case of
throughput. In future, this work may be extended for analyzing the behavior of these
protocols in heterogeneous networks with many more metrics for evaluation.
References
1. The Qualnet simulator, http://www.Scalable-Networks.com
2. Acs, G., Buttyabv, L.: A taxonomy of routing protocols for wireless sensor networks.
BUTE Telecommunication Department (January 2007)
3. Jorg, D.O.: Performance comparison of MANET routing Protocols in different network
sizes. Institute of computer Science and applied mathematics computer Networks and dis-
tributed systems University of Berne, switzerland
4. Sun, A.C.: Design and Implementation of Fisheye Routing protocol for Ad Hoc Networks.
Dept of Electrical and CSE, Massachusetts institute of Technology (May 2000)
5. Lee, Y.Z., Chen, J., Hong, X., Xu, K., Breyer, T., Gerla, M.: Experimental Evaluation of
LANMAR, a scalable Ad- Hoc routing protocol. In: IEEE Communications Soci-
ety/WCNC 2005. University of California, Los Angles (2005)
6. Pei, G., Gerla, M., Hong, X.: LANMAR: Landmark Routing for Large Scale Wireless Ad-
Hoc Networks with Group Mobility. In: Proceedings of IEEE/ACM MobiHOC 2000, Bos-
ton, MA, August 2000, pp. 11–18 (2000)
7. Arshad, J., Azad, M.A.: Performance Evaluation of Secure on-Demand Routing Protocols
for Mobile Ad-hoc Networks. In: 15th IEEE International Conference on Network Proto-
cols, Beijing, China (2007)
8. Kurkowski, S., Navidi, W., Camp, T.: Discovering Variables that Affect MANET Protocol
Performance. In: Proceedings of IEEE Global Telecommunications Conference, GLOBE-
COM 2007, Air Force Inst. of Technol., Wright Patterson, November 26-30. IEEE, Los
Alamitos (2007)
Performance Comparison of Routing Protocols in Wireless Sensor Networks 383
9. Boomarani Malany, A., Sarma Dhulipala, V.R., Chandrasekaran, M.: Throughput and De-
lay Comparison of MANET Routing Protocols. Int. J. Open Problems Compt. Math. 2(3)
(September 2009), ISSN 1998-6262
10. Pei, G., Gerla, M., Hong, X.: Landmark Routing for Large Scale Wireless Ad Hoc Net-
works with Group Mobility. CSE Department University of California
11. Gerla, M., Hong, X., Pei, G.: Landmark Routing for Large Ad Hoc Wireless Networks. In:
Proceeding of IEEE GLOBECOM 2000, San Francisco, CA (November 2000)
12. Mingliang, J., Tay, Y., Long, P.: A cluster-based routing protocol for mobile ad hocnet-
works (1999/2002), http://www.comp.nus.edu.sg/~tayyc/cbrp/hon
13. Broch, J., Maltz, D.A., Johnson, D.B., Hu, Y.-C., Jetcheva, J.: A performance comparison
of multi-hop wireless ad hoc network routing protocols. In: Proceeedings of the 4th Annual
ACM/IEEE International Conference on Mobile Computing and Networking (MOBICOM
1998), October 1998, pp. 85–97 (1998)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 384–387, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Security and Trust Management in MANET
Akash Singh, Manish Maheshwari, Nikhil, and Neeraj Kumar
School of Computer Science Engineering
Shri Mata Vaishno Devi University (J&K), India
akash72@gmail.com, manish5@live.in, nnikhil.roy@gmail.com,
nehra04@yahoo.co.in
Abstract. A mobile adhoc network (MANET) is a network which does not
have any centralized control. Security and trust management are paramount
concern for these networks for efficient data transfer among the participating
nodes. In this paper, we propose an efficient security and trust management
based algorithm for MANET. The proposed algorithm consists of three steps:
initialization, data transmission, and detection. The time based nonce is gener-
ated at different time interval which gives effectiveness to the proposed ap-
proach in the sense that it is not easy to detect the generated nonce. The
proposed approach is quite effective with the earlier approaches to detect the
security threat in MANET.
Keywords: MANET, Security, Trust management.
1 Introduction
A mobile ad hoc network (MANET) is a collection of wireless mobile hosts forming a
temporary network without the aid of any centralized administration or standard
support services regularly available in wide-area networks to which the hosts may
normally be connected [1].MANET are vulnerable to a powerful attack known as
wormhole attack. In a wormhole attack [2], an attacker introduces two transceivers
into a wireless network and connects them with a high quality, low-latency link. Rout-
ing messages received by one wormhole endpoint are retransmitted at the other end-
point. Attackers can exploit wormholes to build bogus route information, selectively
drop packets, and create routing loops to waste the energy of network.
Several cryptographic approaches have been proposed in literature to mitigate se-
curity attacks but many of these proposals may not be suitable for real-time applica-
tions where time is a real challenge/constraint since encryption and decryption of
message requires time which introduces the factor of latency. Moreover, the devices
used in encryption and decryption have limited resources in terms of resources such
as computation power and memory. Keeping in the mind of above challenges and
constraint, we propose a latency aware real time cryptographic communication algo-
rithm for MANET. Rest of the paper is organized as follows: Section 2 discusses the
related work, Section 3 describes the proposed approach (algorithm), and Section 4
provides conclusions.
Security and Trust Management in MANET 385
2 Related Work
Many solutions for security in MANET have been proposed over years [3]. These
solutions try to minimize the overhead and maximize the security of the system.
Generalize architectures for intrusion detection is given in [4] in which all nodes par-
ticipate in the monitoring of the data transmission. To prevent resource consumption
attacks, LHAP [5] implements lightweight hop-by-hop authentication. Using LHAP, a
node joining MANET only need to perform some inexpensive authentication opera-
tions to bootstrap a trust relationship with its neighbors. It then switches to a very
lightweight protocol for subsequent traffic authentications.
Many Group Key Agreement protocols [6] have been proposed in the literature,
most being derived from the two-party Diffie–Hellman (DH) [7] key agreement pro-
tocol. Some have no formal proofs while some are secure against passive adversaries
only. Boyd et al. [8] proposed an efficient constant-round protocol where the bulk of
the computation is done by one participant (the current group leader), thus making it
highly efficient for heterogeneous ad hoc networks. It is provably secure in the Ran-
dom oracle model [9], but lacks forward secrecy.
3 Proposed Approach
The proposed scheme consists of following three steps: Initialization, data transmis-
sion, and detection. We assume that there are no malicious nodes in the system. After
time synchronization, each node sends packet to its neighbour to know its identity.
This identity is multiplied with identity of node modulus P and stored in table. Here P
is a large prime number. Let n is node id and n1 is id of neighbour then table[]=
(n*n1) mod P. Nonce generation algorithm is started in each node. In this algorithm
according to clock pulse, i is generated. The value of i and i-1 is stored in a hash table
along with the time stamp. This hash table will store the value only for a particular
window of time say n seconds. On every node, for a particular time value, same
nonce value is generated. Nonce corresponding to the latest time stamp present in
hash table is fetched say i. Let a1 be the identity of given node. Then b1 ≡ a1i mod P
where P is large prime number. The value of b1 and the time stamp at which i is cal-
culated is appended with the data encrypted in DES and transmitted as
D=b1.time_stamp.E (K, data).
When this packet is received by the neighbor, the neighbor extracts the value of
time stamp and b1 from the packet. The value of i-1 corresponding to the extracted
time stamp is fetched from the hash table. If the time stamp is not present in hash ta-
ble, it means that the duration of time stamp has expired and the packet is dropped. If
time stamp is present, packet is valid. The calculation (a2^i’*b1) mod P is performed.
This is equivalent to (a2^i’*a1^i) mod P=a2*a1 mod P as i and i’ are inverse of each
other and a2 is the node id. This value is checked with value present in table[]. If
value is found, the data is from authorized node. The proposed algorithm is described
in the following section:
386 A. Singh et al.
clock ( ) : this is the clock of the node which keeps the time of the node and maintains
the node in synchronization.
EXTENDED EUCLID (P, i ) : This is extended Euclidean algorithm used to find in-
verse of i modulo P.
Step 1: Initialization
id;neighbour =n
repeat
0=neigh 0,=i number, prime large=P ),clock(time tsync,=time
)hbour until(neig
++iP; modn * node_id=table[i]
φ
≠
forever
i);inv_i,p,(time_stamhash_table
i);EUCLID(P, EXTENDED =inv_i
clock();=time_stampnonce;=i
repeat
Step 2: Data Transmission
forever
data);amp.E(K,b1.time_st = enc_data
P; mod iia1^=b1
.t.i;hash_table=ii
hashin present stamp elatest tim=t
repeat
node_id; = a1
Step 3: Detection
)hash_table t ble.time_sif(hash_ta
time_stamp>enc_data-= time_stb1;>enc_data-=bb1
repeat
node_id;= a2
∈
discarded is data
else
forwarded is data and called is routineion transmissdata
table[])if(check
P mod b1)*inv_i(a2^=check
∈
discarded is data
else
forever
4 Comparisons and Discussion
Table 1. Relative comparison of existent schemes
Approach Time
Sync
Random
Nonce
Centralized
Authority
Key
Compromise
GPS
Packet Leashes [10] Yes No No N.A. Maybe
Directional Antenna[11] No No No N.A Maybe
Group Key [6] Yes Yes Yes Possible No
Proposed Yes Yes No Very Difficult No
Security and Trust Management in MANET 387
5 Conclusions
The proposed algorithm is nonce based and is resilient to various known attacks in
MANET such as man in the middle attack, and passive eavesdropping, active inter-
ference. By applying nonce concept in proposed algorithm it enhances its capability
for addressing real time application. Source consumption get reduced by calculating
the identity of each neighbor and is stored in the hash table for a particular period of
time. Authentication process get faster as only time stamping factor has to be matched
with the hash table and the encrypted data is just forwarded to destination node which
increases the packet delivery fraction in MANET.
References
1. Raimundo, J., Macêdo, A., Assis Silva, F.M.: The mobile groups approach for the coordi-
nation of mobile agents. J.Parallel Distributed Computing 65, 275–288 (2005)
2. Eriksson, J., Krishnamurthy, S., Faloutsos, M.: Truelink: A practical countermeasure to
the wormhole attack. In: ICNP (2006)
3. Buttyan, L., Hubaux, J.-P.: Stimulating cooperation in self-organising mobile ad hoc net-
works. In: Mobile Networks and Applications (2003)
4. Kumar, N., Patel, R.B.: MASLKE: mobile agent based secure location aware key estab-
lishment in sensor networks. In: Proc. ICON 2008, New Delhi, India (2008)
5. Zhang, Y., Lee, W.: Instruction Detection in wireless adhoc networks. In: Proceedings of
ACM Mobicom, Boston, USA (2000)
6. Kim, Y., Perrig, A., Tsudik, G.: Group key agreement efficient in communication. IEEE
Transactions on Computers 53(7), 905–921 (2004)
7. Rescorla, E.: Diffie-Hellman Key Agreement Method, RFC 2631, IETF Network Working
Group, http://www.ietf.org/rfc/rfc2631.txt
8. Boyd, C., Nieto, J.M.G.: Round-Optimal Contributory Conference Key Agreement. In:
Desmedt, Y.G. (ed.) PKC 2003. LNCS, vol. 2567, pp. 161–174. Springer, Heidelberg
(2003)
9. Bennett, C.H., Gill, J.: Relative to a Random Oracle A, P^A != NP^A != co-NP^A with
Probability 1. SIAM Journal on Computing 10(1), 96–113
10. Hu, Y., Perrig, A., Johnson, D.: Packet Leashes: A Defense against Wormhole Attacks in
Wireless Ad Hoc Networks. In: Proceedings of INFOCOM (2004)
11. Hu, L., Evans, D.: Using Directional Antennas to Prevent Wormhole Attacks. In: Pro-
ceedings of the 11th Network and Distributed System Security Symposium (2003)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 388–391, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Analytical Parametric Evaluation of Dynamic Load
Balancing Algorithms in Distributed Systems
Mayuri A. Mehta1 and Devesh C. Jinwala2
1 Department of Computer Engineering, Sarvajanik College of Engineering and Technology,
Surat, India
mayuri133@yahoo.com
2 Department of Computer Engineering, S. V. National Institute of Technology, Surat, India
dcj@svnit.ac.in
Abstract. With ever increasing network traffic, distributed systems can provide
higher performance using a typical dynamic load balancing (DLB) algorithm.
Dynamic algorithm employs up to date load information of the nodes to make
load distribution decisions and therefore, they have potential to outperform stat-
ic strategies. In this paper, we illustrate the analytical comparative study of ex-
isting dynamic algorithms and result gives a thorough overview of various dy-
namic algorithms, helping designers in choosing the most appropriate approach
for a variety of distributed systems. Moreover, researchers can use it as a cata-
log of available DLB schemes to come up with new design.
Keywords: Load balancing, dynamic load balancing, distributed system.
1 Introduction
In a typical distributed system setting, nodes are of heterogeneous nature and also
jobs arrive at the different nodes in a random fashion. This causes imbalance in work-
load that is harmful to the system performance in terms of mean response time of jobs
and resource utilization. A load balancing mechanism can be used to remove such
load imbalance. Load balancing (LB) is a technique to spread work between two or
more computers, network links, CPUs, hard drives, or other resources, in order to
obtain optimal resource utilization, throughput, and/or response time [1].
The two major categories for load-balancing algorithm are: static and dynamic. In
static load balancing (SLB), distribution of workload is predetermined and carried out
using priory knowledge of the system [2]. Dynamic load balancing (DLB) makes use
of current system state information to make more effective load distribution decisions.
It has been shown in literature that DLB algorithm performs better than SLB algo-
rithm in a variety of system conditions. Several studies have been conducted on a
variety of DLB algorithms. In [3], a novel neural network approach that realized the
selection and location policies was described. Fuzzy based DLB algorithms were
presented to improve the response time in [1][4]. The approach that considered the
node heterogeneity and random communication delays was discussed in [5]. In [6],
Analytical Parametric Evaluation of DLB Algorithms in Distributed Systems 389
three different DLB approaches were discussed. The DLB algorithms suitable to grid
environment were illustrated in [7]. DLB schemes for multi-class jobs were presented
in [8]. As per our observation, none of these existing attempts exhaustively and com-
paratively evaluate the majority of the DLB algorithms. Hence, our survey of the
dynamic algorithms discussed further is more complete and has a wider coverage.
The rest of the paper is organized as follows: In section 2, we list the various DLB
algorithms with brief descriptions. Section 3 covers parameterized comparison of
these algorithms. Finally conclusion and some future work are specified in section 4.
2 Various Dynamic Load Balancing Algorithms
A New Fuzzy Approach (ANFA) [1] took into consideration the crisp inputs and effec-
tively tolerates uncertainty and inconsistency in state information using fuzzy logic to
reduce response time. Fuzzy-Based Algorithm (FBA) [4] that consisted of a fuzzy rule
base, a fuzzy inference engine, fuzzification, and defuzzification improved overall
system performance. A Novel Neural Network Approach (NNNA) [3] used Winner-
Take-All (WTA) model to realize the selection and location policies of a typical DLB
algorithm and considered all delays due to network communication.
Centralized One-Shot LB Policy (COSLBP) [5] considered only one LB instant that
was time at which load balancing was executed [9]. Real-time experiments conducted
by authors had shown that for a given initial load and average processing rates, there
existed an optimal LB gain and an optimal balancing instant associated with the one-
shot LB policy, that together minimized the average overall completion time (AOCT).
A Distributed, Adaptive Sender-Initiated Policy (DASIP) [5] which was adapted to
varying system parameters such as load variability, delay, and variable runtime
processing speed, minimized the average completion time per task.
In Primary Approach (PA) [6], overloaded primary node first tried to find out light
weighted node in the same cluster and if suitable node was not found then nearby
cluster was searched for the same. In Centralized Approach (CA) [6], centralized node
in each cluster accommodated the overload of a primary node. Whenever an over-
loaded node did not find lightly loaded primary node, overload was transferred to
centralized node. To overcome the limitations of centralized approach, there was a
Modified Approach with Supporting Nodes (MASNs) [6]. Centralized node was split
into several Supporting Nodes (SNs). Now overloaded primary node would interrupt
the SN. This SN used interrupt service routine (ISR) to schedule the process.
In Perfect Information Algorithm (PIA) [7], when a job arrived, a processor com-
puted the job’s finish time on all buddy processors. A job was relocated on the buddy
processor if it could finish the job earlier than this processor. In Estimated Load In-
formation Scheduling Algorithm (ELISA) [7], the processor would migrate the jobs to
lightly loaded processor, if its queue length was greater than the average queue length
in its buddy set. In Modified ELISA (MELISA) [7], at each estimation instant, proces-
sor calculated the average load in its buddy set. Now processor would make decision
of job migration taking into consideration the node’s heterogeneity if its load was
greater than the average load in its buddy set. Load Balancing on Arrival Algorithm
(LBAA) [7] balanced load by transferring a job on its arrival time rather than waiting
for the next transfer instant as in case of ELISA and MELISA.
390 M.A. Mehta and D.C. Jinwala
Dynamic Global Optimal Scheme (DGOS) and Dynamic Noncooperative Scheme
with Communication (DNCOOPC) [8] were extended from static schemes and in-
tended for multi-user (multi-class) jobs. The objective of DGOS was to minimize the
expected response time of all jobs over the entire system. The goal of DNCOOPC was
to minimize the expected response time of the individual users.
DLB Algorithm based on Distributed Database System (DLBDDS) [10] was aimed
to provide a reasonable request-response time and transaction throughput. Due to
considering the communication cost and employing the collected information, this
algorithm was much more efficient and had better performance.
3 Analysis
Here we present comparison of these algorithms based on certain imperative issues
such as handling resource and/or network heterogeneity, associated overhead, scala-
bility and delays due to the underlying network or processing delays at the processors.
A proper DLB algorithm tolerates heterogeneity in multitude of resources, incurs
minimum overhead, integrates scalability and considers delays imposed by communi-
cation networks. Table 1 demonstrates the comparison of above mentioned DLB
algorithms based on these issues.
Table 1. Comparison of dynamic load balancing algorithms
Algo-
rithm/Appr
oach
Tolerates
Hetero-
geneity?
Associated
Overhead
Considers
Delays?
Scalability
Incorporated?
Performance
Metric(s) Used
ANFA Yes More No Yes RT
FBA Yes Less Yes Yes RT, T, TT
COSLBP Yes Less Yes No AOCT
DASIP Yes More Yes Yes AOCT, ACTT
PA Yes Less No Yes ET
MASNs Yes More No Moderately ET
PIA Yes More No Yes RT
ELISA Yes Less No Yes RT
MELISA
and LBA
Yes More Yes Yes RT,JMC,ET,RU
DGOS and
DNCOOPC
Yes Less to more No Yes RT
DLBDDS Yes Average No Yes RT, T
NNNA No More Yes Yes Mean RT
The various metrics viz. Response Time (RT), Throughput (T), Turnaround Time
(TT), Average Overall Completion Time (AOCT), Average Completion Time per
Task (ACTT), Execution Time (ET), Resource Utilization (RU), and Job Migration
Cost (JMC) that are used to measure performance of DLB approaches are also differ-
ent for various DLB algorithms and are shown in Table 1.
Analytical Parametric Evaluation of DLB Algorithms in Distributed Systems 391
4 Conclusion and Future Work
Various DLB algorithms have been proposed in literature. We provide an overview of
the existing DLB algorithms in distributed systems. It reinforces the belief that dy-
namic algorithms potentially do better for performance improvements. The principal
objective of our study on DLB algorithms is to travel around the existing DLB ap-
proaches and to be familiar with the existing work to design a new competent dynam-
ic algorithm for a heterogeneous distributed system. As part of our future work we
intend to design a new efficient DLB and compare it with the existing ones.
References
1. Karimi, A., Zarafshan, F., Jantan, A.B., Ramli, A.R., Saripan, M.B.: A New Fuzzy Ap-
proach for Dynamic Load Balancing Algorithm. Int. J. Computer Science and Information
Security 6, 1–5 (2009)
2. Waraich, S.S.: Classification of Dynamic Load Balancing Strategies in the Network of
Workstations. In: 5th Proceedings of the International Conference on Information Tech-
nology: New Generations, pp. 1263–1265. IEEE Computer Society, Washington (2008)
3. El-Abd, A.E., El-Bendary, M.I.: A Neural Network Appraoch for Dynamic Load Balanc-
ing in Homogeneous Distributed System. In: 30th Proceedings of IEEE International Con-
ference on System Sciences, pp. 628–634. IEEE Computer Society, Washington (1997)
4. A Fuzzy-Based Dynamic Load-Balancing Algorithm,
http://jitas.im.cpu.edu.tw/2004-2/4.pdf
5. Dhakal, S., Hayat, M.M., Pezoa, J.E., Yang, C., Bader, D.A.: Dynamic Load Balancing in
Distributed Systems in the Presence of Delays: A Regeneration – Theory Approach. IEEE
Transactions on Parallel and Distributed Systems 18, 485–497 (2007)
6. Jain, P., Gupta, D.: An Algorithm for Dynamic Load Balancing in Distributed Systems
with Multiple Supporting Nodes by Exploiting the Interrupt Service. Int. J. Recent Trends
in Engineering 1, 232–236 (2009)
7. Shah, R., Veeravalli, B., Misra, M.: On the Design of Adaptiveand Decentralized Load Ba-
lancing Algorithms with Load Estimation for Computational Grid Environment. IEEE
Transactions on Parallel and Distributed Systems 18, 1675–1686 (2007)
8. Penmasta, S., Chronopoulos, A.T.: Dynamic Multi-User Load Balancing in Distributed
Systems. In: 21st Proceedings of IEEE International Parallel and Distributed Processing
Symposium, pp. 1–10. IEEE Press, California (2007)
9. Dhakal, S., Paskaleva, B.S., Hayat, M.M., Schamiloglu, E., Abdalla, C.T.: Dynamical Dis-
crete-Time Load Balancing in Distributed Systems in the Presence of Time Delays. In:
42nd Proceedings of IEEE Conference on Decision and Control, pp. 5128–5134. IEEE
Press, Piskataway (2003)
10. Feng, Y., Li, E., Wu, H., Zhang, Y.: A Dynamic Load Balancing Algorithm Based on Dis-
tributed Database System. In: 4th Proceedings of International Conference and Exhibition
on High-Performance Computing in the Asia-Pacific Region, pp. 949–951. IEEE Comput-
er Society, Biejing (2000)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 392–398, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Wavelet Based Electrocardiogram Compression at
Different Quantization Levels
A. Kumar* and Ranjeet
Indian Institute of Information Tehncology Design and Manufacturing,
Jabalpur, MP-482005, India
anilkdee@gmail.com, ranjeet281@gmail.com
Abstract. In this paper, a wavelet based electrocardiogram (ECG) data com-
pression technique is reviewed. The method employs the discrete wavelet trans-
form (DWT), thresholding, Huffman encoding followed by different quantiza-
tion levels. A comparative study of performance at the different quantization
levels and thresholding is made in terms of Signal-to-noise ratio (SNR), Percent
root mean square difference (PRD) and Mean square error (MSE). The simula-
tion results illustrates that good compression ratio can be achieved at lower
quantization levels, while at higher quantization levels, all fidelity measuring
parameters are enhanced.
Keywords: DWT, Thresholding, Quantization and Huffman Encoding.
1 Introduction
An electrocardiogram (ECG) is the graphical representation of electrical impulses due
to ionic activity in the cardiac muscles of human heart. It is an important physiologi-
cal signal which is exploited to diagnose heart diseases because every arrhythmia in
ECG signals can be relevant to a heart disease [1]. ECG signals are recorded from
patients for both monitoring and diagnostic purposes. Therefore, the storage of com-
puterized is become necessary. However, the storage has limitation which has made
ECG data compression as an important issue of research in biomedical signal process-
ing. In addition to these, there are many advantages of ECG compression such as
transmission speed of real-time ECG signal is enhanced and is also economical.
Several efficient methods [2-9] are available in literature which involve in com-
pression schemes of ECG signal without losing and preserving the relevant clinical
information for the accurate detection and classification. These schemes were classi-
fied into three categories [3]: dedicated techniques such as AZTEC, FAN, CORTES,
and turning point. These techniques were based on the detection and elimination of
redundancies on direct analysis of the original signal, and gives minimum distortion.
In second category, all transform based techniques come and here, compression is
achieved based on spectral and energy distribution of the signal. Other hand, the last
technique is based on feature and parameter extraction in which some parameters
* Corresponding author.
Wavelet Based Electrocardiogram Compression at Different Quantization Levels 393
such as measurement of the probability distribution of the original signal is extracted.
During the last two decades, several efficient methods have reported in literature,
which involve compression of ECG signal without losing and preserving the relevant
clinical information for the accurate detection and classification. Multi-resolution
decomposition of signal is efficient for extracting the content information [10]. In this
technique, wavelet transform has been exploited for the ECG processing and extract-
ing the information. Recently, several other methods [6-9] have been developed based
on wavelet or wavelet packets promise that it is an efficient power tool for compress-
ing and analysis of ECG signal.
In this paper, wavelet based ECG compression technique is reviewed and the effect
of different quantization levels on compression is explored.
2 Discrete Wavelet Transform
Wavelets transform is a method to analyze a signal in time and frequency domain, it
is an effective tool for the analysis of time-varying non stationary signal like ECG [7].
Wavelet transform gives the multiresolution decomposition of the signal. There are
three basic concepts of multiresolution: subband coding, vector space and pyramid
structure coding [10]. DWT decomposes a signal at several n levels in different fre-
quency bands. Each level decomposes a signal into the approximation coefficients
(low frequency band of processing signal) and the detail coefficients (high frequency
band of processing signal) [10] as show in Fig. 1.
Fig. 1. Filter bank representation of DWT decomposition
At each step of DWT decomposition, there are two outputs: the scaling coefficients
xj+1(n) and the wavelet coefficients yj+1(n). These coefficients are given:
2
1
1
() (2 ) ()
n
jj
i
xn hnixn
+
=
=−
∑ (1)
x0 (n)
2
2
h(n)
g(n) h(n)
g(n)
2
2
y1 Level-1 DWT coefficients
y2 Level-2
y3 Level-2
x0 (n)
394 A. Kumar and Ranjeet
and
2
1
1
() (2 ) ()
n
jj
i
yn gnixn
+
=
=−
∑ (2)
where, the original signal is represented by x0(n) and j show the scaling number. Here
g(n) and h(n) represent the low pass and high pass filter, respectively. The output of
scaling function is input of next level of decomposition, known as approximation
coefficients. The approximation coefficients are low-pass filter coefficients and high-
pass filter coefficient are detail coefficients of any decomposed signal.
3 Quantization
Quantization is a process of representation of a set of continues value to finite discrete
set of values. A signal divides into a number of interval, each interval having own
codeword in the quantized value. In wavelet based compression, after thresholding the
wavelet coefficients vector xj (n) is quantized [11-13]. Due to quantization process,
the perfect reconstruction of the original signal is not possible at the reconstruction
side. The quantization process depends on these parameters: maximum value (Mmax),
minimum value (Mmin) in the signal and number of quantization level L= 2m (An m-bit
uniform quantizer in used). Once these parameters are found, then step size ( Δ) is
computed by
()
max min
MMLΔ= − (3)
In the uniform quantization, step-size Δ is depend upon the number of quantization
levels, its associated with the value of m. which captains the information in form of
bit/symbol of quantized signal. A detailed discussion on the quantization process is
given in [11-13].
4 Methodology for ECG Compression
A wavelet based methodology of ECG compression is shown in Fig. 2. This technique
involves three steps for the ECG signal compression: DWT decomposition, threshold
and quantization, and entropy encoding. After DWT decomposition of ECG signal, its
wavelet coefficients are selected on the basis of energy packing efficiency of each
subband. After decomposition of the ECG signal, a thresholding is applied to the
wavelet coefficients, which makes a fixed percentage of wavelet coefficients equal to
zero. There are two types of the thresholding: global and level thresholding. In level
thresholding, the threshold value is calculated using Birge-Massart strategy [14, 15].
While, in global thresholding, the threshold value is set manually, this value is chosen
from wavelet coefficient (0….xmaxj) where xmaxj is maximum coefficient in the decom-
position. Detailed discussion on thresholding is given in [12-16].
Wavelet Based Electrocardiogram Compression at Different Quantization Levels 395
Further, uniform quantization is performed on these coefficients. The actual com-
pression is achieved at this stage and this compression can be further enhanced with
the help of entropy encoding technique (Huffman) [12, 13, 16]. In Huffman encoding,
the probabilities of occurrence of the symbols in the signal are computed. These
symbols indices in the quantization table, these symbols arranged according to the
probabilities of occurrence in descending order and a binary tree and codeword table
is created. Finally, the compressed ECG signal is obtained at the output of entropy
encoder.
Fig. 2. Compression methodology for ECG signals
5 Methodology for ECG Compression
In this paper, ECG signal compression is achieved using the methodology discussed
in Section IV and the effect of different quantization level is seen on the compression.
The performance is evaluated by considering the fidelity of the reconstructed signal to
the original signal. For this, many fidelity assessment parameters are considered such
as Compression ratio (CR), Percent root mean square difference (PRD), Mean square
error (MSE) and Signal to noise ratio (SNR) given in [11-16]: ECG records have
been obtained from MIT-BIH Arrhythmia Database [17]. Here, different wavelet
filters, and global thresholding are exploited for signal compression. The simulation
results obtained in each case are included in Table 1. A comparative analysis of dif-
ferent m-bit quantizer at different thresholds is depicted in Fig. 3.
ECG
Signal
Transform
Methods Thresholding &
Quantization
Entropy Encoding
De-quantization
Entropy Decoding
Inverse
Transform
Reconstructed ECG
396 A. Kumar and Ranjeet
0.05 0.10 0.15 0.20 0.25 0.30
6
6.5
7
7.5
Threshold value
CR
Compression Ratio
0.05 0.10 0.15 0.20 0.25 0.30
4
6
8
10
12
Threshold value
PRD
PRD
0.05 0.10 0.15 0.20 0.25 0.30
18
20
22
24
26
Threshold value
SNR
Signal-to-noise ratio
m=5
m=6
m=7
m=8
0.05 0.10 0.15 0.20 0.25 0.30
0
0.2
0.4
0.6
0.8
1x 10
-3
Threshold value
MSE
Mean square error
(a) (b)
(c) (d)
Fig. 3. Variation of performance measuring parameters of wavelet based compression with
different threshold value and quantization level. (a) CR (b) PRD (c) SNR (d) MSE
It is evident from Table I that at low level quantization, compression ratio is more
as compared to high level quantization. While, at higher quantization level, all fidelity
measuring parameters are more improved and gives good compression as well as
preserving more.
6 Conclusions
In this paper, a wavelet based methodology is presented for the ECG signal compres-
sion. A comparative study of performance of different uniform quantization levels for
the ECG signal compression is explored. The simulation results obtained illustrate
that good compression ratio can be achieved at low level quantization and good re-
construction of the original signal can be achieved at higher quantization levels.
Wavelet Based Electrocardiogram Compression at Different Quantization Levels 397
Table 1. Variation of fidelity measuring parameter at different quantization levels
Wavelet Filters Quantization Levels CR PRD MSE SNR
Haar 5 8.53 11.33 8.51×10-4 18.51
db10 5 8.37 10.53 7.32×10-4 19.16
coif5 5 8.08 9.33 5.84×10-4 20.14
sym8 5 8.53 9.96 6.62×10-4 19.59
Haar 6 7.55 9.51 5.66×10-4 20.27
db10 6 7.45 8.65 4.70×10-4 21.09
coif5 6 7.14 7.01 3.12×10-4 22.86
sym8 6 7.60 7.79 3.84×10-4 21.96
Haar 7 6.82 9.04 5.00×10-4 20.81
db10 7 6.74 8.00 3.91×10-4 21.88
coif5 7 6.49 6.25 2.41×10-4 23.99
sym8 7 6.82 7.15 3.14×10-4 22.83
Haar 8 6.26 8.94 4.83×10-4 20.97
db10 8 6.16 7.80 3.66×10-4 22.17
coif5 8 5.95 6.06 2.23×10-4 24.31
sym8 8 6.24 6.98 2.96×10-4 23.09
References
1. Martis, R.J., Chakraborty, C., Ray, A.K.: An integrated ECG feature extraction scheme us-
ing PCA and wavelet transform. In: Proceedings of IEEE India Council Conference (IN-
DICON 2009), pp. 88–93 (2009)
2. Sastry, R.V.S., Rajgopal, K.: ECG Compression using Wavelet Transform. In: Proceed-
ings of RC IEEE-EMBS, BMESI, vol. 14, pp. 499–500 (1995)
3. Nielsen, M., Kamavuako, E.N., Andersen, M.M., Lucas, M.-F.L., Farina, D.: Optimal
wavelets for biomedical signal compression. Med. Bio. Eng. Comput. 44, 561–568 (2006)
4. Khorrami, H., Moavenian, M.: A comparative study of DWT, CWT and DCT transforma-
tions in ECG arrhythmias classification. Expert Systems with Applications 37(8),
5751–5757 (2010)
5. Manikandan, M.S., Dandapat, S.: Wavelet threshold based TDL and TDR algorithms for
real-time ECG signal compression. Biomedical Signal Processing and Control 3(1), 44–66
(2008)
6. Hunga, K.-C., Tsaia, C.-F., Ku, C.-T., Wanga, H.S.: A linear quality control design for
high efficient wavelet-based ECG data compression. Computer Methods and Programs in
Biomedicine 9(4), 109–117 (2009)
7. Saritha, C., Sukanya, V., Murthy, Y.N.: ECG Signal Analysis Using Wavelet Transforms.
Bulg. J. Physics 35, 68–77 (2008)
8. Ahmed, S.M., Al-Shrouf, A., Abo-Zahhad, M.: ECG data compression using optimal non-
orthogonal wavelet transform. Medical Engineering & Physics 22(1), 39–46 (2000)
398 A. Kumar and Ranjeet
9. Rajoub, B.A.: An Efficient Coding Algorithm for the Compression of ECG Signals Using
the Wavelet Transform. IEEE Transactions on Biomedical Engineering 49(4), 355–362
(2002)
10. Mallat, S.G.: A Theory for Multiresolution Signal Decomposition: The Wavelet Represen-
tation. IEEE Transactions on Pattern Analysis and Machine Intelligence 11(7) (1989)
11. Manikandan, M.S., Dandapat, S.: Wavelet threshold based ECG compression using
USZZQ and Huffman coding of DSM. Biomedical Signal Processing and Control 1, 261–
270 (2006)
12. Chen, J., Wang, F., Zhang, Y., Shi, X.: ECG compression using uniform scalar dead-zone
quantization and conditional entropy coding. Medical Engineering & Physics 30, 523–530
(2008)
13. Ebrahimzadeh, A., Azarbad, M.: ECG Compression using Wavelet transform and Three-
Level Quantization. In: IEEE Conference proceedings IDC, vol. 6, pp. 250–254 (2010)
14. Misiti, M., Misiti, Y., Oppenheim, G., Poggi, J.: Matlab tool box. The Math Works, Inc.
(2000)
15. Najih, A.M.M.A., Ramli, A.R.B., Prakash, V., Syed, A.R.: Speech Compression using
Discrete Wavelet Transform. In: 4th National Conference on Telecommunication Tech-
nology Proceedings, Shah Alam, Malaysia, pp. 1–4 (2003)
16. Tohumoglu, G., Sezgin, K.E.: ECG signal compression by multi-iteration EZW coding for
different wavelets and thresholds. Computers in Biology and Medicine 37(2), 173–182
(2007)
17. http://www.physionet.org
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 399–402, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Content Based Image Retrieval by Using an Integrated
Matching Technique Based on Most Similar Highest
Priority Principle on the Color and Texture Features of
the Image Sub-blocks
Ch. Kavitha1, M. Babu Rao2, B. Prabhakara Rao3, and A. Govardhan4
1 Associate Professor, IT department, Gudlavalleru Engineering College,
Gudlavalleru, Krishna district, Andhra Pradesh, India
kavithachaduvula@yahoo.com
2 Associate Professor, CSE department, Gudlavalleru Engineering College,
Gudlavalleru, Krishna district, Andhra Pradesh, India
baburaompd@yahoo.co.in
3 Director of Evaluation, JNTUK, Kakinada, A.P, India
4 Principal, JNTUH college of Engineering, Jagtial, A.P, India
Abstract. In this paper, we propose an efficient technique for content based im-
age retrieval which uses the local color and texture features of the image. Firstly
the image is divided into sub blocks of equal size. The color and texture fea-
tures of each sub-block are computed. Color of each sub-block is extracted by
quantifying the HSV color space into non-equal intervals and the color feature
is represented by cumulative histogram. Texture of each sub-block is obtained
by using gray level co-occurrence matrix. An integrated matching scheme based
on Most Similar Highest Priority principle is used to compare the query and
target image. The adjacency matrix of a bipartite graph is formed using the sub-
blocks of query and target image. This matrix is used for matching the images.
Euclidean distance is used in retrieving the similar images. The efficiency of the
method is demonstrated with the results.
Keywords: Image retrieval, color, texture, GLCM, integrated matching.
1 Introduction
Content based image retrieval is a technique for extracting similar images from an
image database using low level features of an image. The need for efficient image
retrieval is increased tremendously[1]. There are various CBIR systems which used
global features [2], [3], [4] and local features [4]. From these systems it is clear that
local features play a significant role in determining similarity of images.
Color and texture are the most important visual features. Because of the advantage
of HSV color space is its ability to separate chromatic and achromatic components.
we selected the HSV color space to extract the color features. Texture feature is a
kind of visual characteristics that does not rely on color or intensity and reflects the
400 Ch. Kavitha et al.
intrinsic phenomenon of images. So we developed a technique which captures color
and texture features of sub-blocks of the image.
2 Proposed Method
2.1 Partitioning the Image into Sub-blocks
Firstly the image is partitioned into 6 (2X3) equal sized sub-blocks. The size of the
sub-block in an image of size 256X384 is 128X128. The images with other than
256X384 size are resized to 256X384.
2.2 Extraction of Color of an Image Sub-block
Unequal interval quantization HSV color space according the human color perception
has been applied on H, S, and V components. So, we divided color into eight parts.
Saturation and intensity is divided into three parts separately. The quantified hue(H),
saturation(S) and value(V) are showed as equation 1.
In accordance with the quantization level above, three-dimensional feature vector
for different values of H,S,V with different weight to form one-dimensional feature
vector named G:
G =9H+ 3S +V (2)
This paper represents the one-dimensional vector G by constructing a cumulative his-
togram of the color characteristics of image after using non-interval HSV quantization
for G.
2.3 Extraction of Texture of an Image Sub-block
GLCM creates a matrix with the directions and distances between pixels, and then
extracts meaningful statistics from the matrix as texture features. GLCM texture fea-
tures commonly used are shown in the following:
Energy =
()
∑
∑
=
xy
yxPE ,2 (3)
Contrast I =
()()
∑
∑
−,
2yxPyx (4)
Entropy
∑
∑
−=
xy
yxPyxPS ),(log),( (5)
[
)
[
)
[
)
⎪
⎩
⎪
⎨
⎧
∈
∈
∈
=
1,7.02
7.0,2.01
2.0,0
sif
sif
sifo
S
[
)
[
)
[
)
⎪
⎩
⎪
⎨
⎧
∈
∈
∈
=
1,7.02
7.0,2.01
2.0,0
vif
vif
vifo
V
[]
[]
[]
[]
[]
[]
[]
[]
⎪
⎪
⎪
⎪
⎪
⎩
⎪
⎪
⎪
⎪
⎪
⎨
⎧
∈
∈
∈
∈
∈
∈
∈
∈
=
315,2967
295,2716
270,1915
190,1564
155,763
75,412
40,211
20,316
hif
hif
hif
hif
hif
hif
hif
hifo
H(1)
Content Based Image Retrieval by Using an Integrated Matching Technique 401
Inverse difference ),(
)(1
1
2yxP
yx
H
xy
∑∑ −+
= (6)
2.4 Integrated Image Matching
We have designed an algorithm for finding the minimum cost matching based on
most similar highest priority (MSHP) principle using the adjacency matrix of the bi-
partite graph. The minimum distance dij of this matrix is found between sub-blocks i
of query and j of target. The complexity of the matching procedure is reduced from
O(n2) to O(n), where n is the number of sub-blocks involved. The integrated mini-
mum cost match distance between images is defined as: Dqt=∑ ∑ dij, Where i=1,2,--n
j=1,2---n. And dij is the best-match distance between sub-block i of query image q and
sub-block j of target image t and Dqt is the distance between images q and t.
3 Experimental Setup
3.1 Data Set and Feature Set
Wang’s dataset comprising of 1000 Corel images with ground truth. The image set
comprises 100 images in each of 10 categories. The images are of the size 256 x 384.
The feature set comprises color and texture descriptors computed for each sub-block
of an image as we discussed in section 2.
3.2 Computation of Similarity
Matching of the sub-blocks is done based on the most similar highest principle. We
construct the Euclidean calculation model as follows:
D(A, B) =ω1D(FCA , FCB ) + ω2D(FTA , FTB) (7)
Here ω1 is the weight of color features, ω2 is the weight of texture features, FCA and FCB
represents the color features for image A and B. For a method based on GLCM, FTA
and FTB on behalf of texture features correspond to image A and B. Here, we combine
color features and texture features. The value of ω through experiments shows that at
the time ω1=ω2=0.5 has better retrieval performance.
4 Experimental Results
The experiments were carried out as explained in sections 2 and 3. The results are
benchmarked with some of the existing systems using the same database[15].The
query images are shown in Fig 1. The static precision of the images (a),(b),(c) and (d)
in various techniques of 20 random images is represented in the form of Table 1.
From the results, it is observed that our proposed method improves the precision over
the other methods.
402 Ch. Kavitha et al.
(a) (b) (c) (d)
Fig. 1. Query images
Table 1. The Static Precision
Different image retrieval techniques P(20) HSVcolo
r
+GLCM
Of image
HSVcolor+
GLCM of
image
sub-blocks
HSVcolor+GLCM
of image sub-blocks
with matching based
on MSHP principle
A 0.45 0.5 0.6
B 0.3 0.6 0.75
C 0.8 0.85 1.0
D 0.35 0.55 0.8
Average
precision
0.475 0.625 0.7875
5 Conclusions
In this paper a new image retrieval method based on color and texture features of im-
age sub-blocks and an integrated matching scheme based on Most Similar Highest
priority (MSHP) to match the images is proposed. Our experiment results demonstrate
that the proposed method has better retrieval performance compared other retrieval
techniques.
References
[1] Datta, R., Joshi, D., Li, J., Wang, J.: Image Retrieval:Ideas, Influences and trends of the
New Age. In: Proceedings of the 7th ACM SIGMM international workshop on multimedia
information retrieval, Hilton, Singapore, November 10-11 (2005)
[2] Niblack, W., et al.: The QBIC Project: Querying Images by Content Using Color, Texture,
and Shape. In: Proc. SPIE, San Jose, CA, February 1993, vol. 1908, pp. 173–187 (1993)
[3] Pentland, A., Picard, R., Sclaroff, S.: Photobook: Content-based Manipulation of Image
Databases. In: Proc. SPIE Storage and Retrieval for Image and Video Databases II, San
Jose, CA, February 1994, pp. 34–47 (1994)
[4] Stricker, M., Orengo, M.: Similarity of Color Images. In: Proc. SPIE Storage and Retrieval
for Image and Video Databases, February 1995, pp. 381–392 (1995)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 403–406, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Understanding the Impact of Cache Performance on
Multi-core Architectures
N. Ramasubramaniam, V.V. Srinivas, and P. Pavan Kumar
Department of Computer Science and Engineering,
National Institute of Technology - Tiruchirappalli
{nrs, 206110022, 206109026}@nitt.edu
Abstract. Research, recent and old, payed much attention on reducing the cycle
time and increasing the speed of execution of the processor. This lead to the de-
velopment of multiple core processors that distribute and share load among the
many processors. The important question to answer would be, will the shared
cache technology for multi-core work as efficiently as it did for uni-core proc-
essors? Our analysis in this paper, takes into consideration the impact that
caches have on a uni-core environment and in a shared multi-core environment.
In order to demonstrate this we use: DINERO IV for analysing the performance
of uni-core environment and CACTI (Cache Access Cycle Time Indicator) for
analysing the performance of multi-core environment.
Keywords: CACTI, Cache memory, performance, multi-core, shared cache.
1 Introduction
Memory hierarchy has seen a lot of improvement in recent years [8]. Cache technol-
ogy serves to bridge the gap between the high speed processor and the low speed de-
vices [5]. Recent trends show that, in order to increase the data transfer rate of the
secondary memories, cache memories are placed on the secondary storage to prere-
cord data from secondary storage for faster access. Of-late, not much work has been
done about the impact of cache memory on multi-core architectures [7]. Though the
speed at which main memories or the secondary memories operate cannot be com-
pared with the speed of processors, the access time of memories have come down and
the execution time in terms of clock cycle per second has been reduced, due to the
increase in the number of cores inside the processor. Moreover the cost per gigabyte
of the secondary memories is also very much lower. [1] suggests the impact of cache
on access time and cycle time in a uni-core environment.
1.1 Contribution
The key focus of this paper is to analyze the impact that a cache would produce when
used in a uniprocessor environment and a multi-core environment. The entire paper is
organized as follows: Section 2 describes the description of various tools used fol-
lowed by Section 3, dealing with the experimental setup and the parameters used.
Section 4 evaluates the results obtained. Section 5 concludes with the future research.
404 N. Ramasubramaniam, V.V. Srinivas, and P.P. Kumar
2 Tools Used
DINERO IV [2] is a cache simulator for uni-processor environment. This simulator is
capable of simulating multi-level caches. It takes in a configuration file for simulating
the performance of cache. The access time for DINERO is calculated based on the
formula given below. Here hit time is taken as 1 clock cycle. Miss rate is taken from
the simulation result and the miss penalty is taken as 8 clock cycles for L1 cache. Cy-
cle time for L1 cache is 7.5 nanoseconds [9]. CACTI [3] is an analytical tool devel-
oped to evaluate cache access and cycle time. This tool was developed based on Wada
et. als, model considers both direct and set associative cache. The computation for
cache access time is done as a function of cache size, associativity and cache block
size [4]. In this paper, we have used version 6.5 for all the experiments. CACTI can
be accessed as a web-interfaced version [10] as well as source code version.
Access time = Hit time +Misrate * Misspenalty
3 Experimental Setup
CACTI 6.5 takes a configuration file as input. The parameters used in experiments
are: cache size of 2097152,134217728 bytes with each cache consisting of 64
blocks. We have considered 45 nm chip technology in our experiments. For DINERO
IV we have taken the values closest to the parameters mentioned since it does not
support some of the parameters used by CACTI, such as chip technology. In this pa-
per we have purposefully chosen these cache sizes because these are some of the
cache sizes supported by CACTI-6.5. Much attention is given in this paper to CACTI-
6.5 because, recent technology has evolved from uni-core to multi-core technology.
(a) Direct Associative
(b) Set Associative - Associativity 4
Fig. 1. Access time (ns) and Cycle time (ns) for varying manufacturing technology
4 Results Obtained and Avenues for Future Research
From Fig. 1.a and 1.c we can visualize that as the number of cores increases the cache
access time increases, whereas the cycle time decreases; see Fig. 1.b and 1.d. This
behaviour can be attributed to the access of shared cache by multiple processors. The
Understanding the Impact of Cache Performance on Multi-core Architectures 405
(a) Access time for cache size of 2097152
bytes in ns.
(b) Cycle time for cache size of 2097152 bytes
in ns.
(c) Access time for cache size of 134217728
bytes in ns.
(d) Cycle time for cache size of 134217728
bytes in ns.
Fig. 2. Here x axis represents the number of cores. Y axis represents the access/cycle time in
nanoseconds. Z axis represents Associativity (Here we have taken associativity of 1 and 4. The
plane on front represents associativity of 1 and denoted by black dots, the plane in the rea
r
represents associativity of 4 and denoted by red dots. The values for number of cores =1 is
computed from the results obtained from DINERO IV).
Table 1. CACTI results
Technology 32 32 45 45 65 65 90 90
Number of cache banks 1 2 1 2 1 2 1 2
Access time (ns) 5.3939 5.3417 7.0117 6.879 10.59 10.27 13.1722 12.7284
Cycle time (ns) 11.01 10.95 13.72 13.575 20.14 19.79 24.1068 23.5985
only difference between Fig. 1.a and Fig. 1.c is that the range of the scale differs as
the cache size increases from 2097152 and 134217728 bytes. From Fig. 2.a and 2.b
we can infer that the access time for cache decreases as the technology moves from
90nm to 32 nm. This is due to the small structure, distance between the cache and the
processor and high speed buses connecting cache and the processor. Further research
would be to improve the models of CACTI and DINERO IV tools and to compare the
results obtained from the tools with that of the real time system for the given set of
406 N. Ramasubramaniam, V.V. Srinivas, and P.P. Kumar
cache parameters. As mentioned in our previous paper [6], this serves as one more
step in understanding the performance of multi-core architectures.
5 Conclusion
In this paper the authors have made an attempt in understanding the performance of
multi-core architectures with cache memories. Thus in this paper, the time for access-
ing the cache memory and execution time of the processor have been understood us-
ing tools such as DINERO IV for uni-core architecture and CACTI-6.5 for multi-core
architecture.
References
1. Muralimanohart, N., Balasubramonian, R., Jouppi, N.P.: Architecting Efficient
Interconnections for Large Caches with CACTI-6.0. Proceedings of IEEE Micro. 28,
69–79 (2008)
2. Dinero IV, Trace Driven Uniprocessor Cache Simulator,
http://pages.cs.wisc.edu/~markhill/DineroIV/
3. CACTI-6.5 (Cache Access Cycle Time Indicator),
http://www.hpl.hp.com/research/cacti/
4. Smith, A.J.: Line (block) size choice for CPU cache memories. Proceedings of IEEE
Transactions on Computers 100, 1063–1075 (2009)
5. Goodman, J.R.: Using cache memory to reduce processor-memory traffic. In: Proceedings
of International Symposia on Computer Architecture, pp. 255–262 (1998)
6. Srinivas, V.V., Ramasubramaniam, N.: Understanding the performance of multicore
architecture. Accepted in International Conference in Communication, Network and
Computing (2011)
7. Banakar, R., Steinke, S., Lee, B.S., Balakrishnan, M., Marwedel, P.: Scratchpad memory:
design alternatives for cache on-chip memory in embedded systems. In: Proceedings of the
Tenth International symposium on Hardware/Software Co design, pp. 73–78 (2002)
8. Hennessy, J.L., Patterson, D.A.: Computer Architecture: A Quantitative Approach, 4th
edn. Elsevier Inc., Amsterdam (2007)
9. Liu, Z., Zheng, K., Liu, B.: Hybrid cache architecture for high speed packet processing.
Proceedings of Computers and Digital Techniques 1, 105–112 (2007)
10. CACTI web interface version, http://quid.hpl.hp.com:9081/cacti/
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 407–410, 2011.
© Springer-Verlag Berlin Heidelberg 2011
An Overview of Solution Approaches for Assignment
Problem in Wireless Telecommunication Network
K. Rajalakshmi and M. Hima Bindu
Department of Computer Science,
Jaypee Institute of Information Technology, Noida
rajalakshmi_krishna@yahoo.com, hima.bindu@jiit.ac.in
Abstract. Currently, there is great urgency to reduce the cost of wireless tele-
communication networks in order to minimize the number and cost of required
facilities, and thus adds to search for optimal network designs. One such solu-
tion is optimal assignment of cells to switch in wireless telecommunication
network. In this paper, we explored three basic wireless telecommunication
networks, which share analogous network architecture namely Personal Com-
munication Services (PCS), Cellular Mobile Network and Universal Mobile
Telecommunication Services (UMTS). The optimal assignment problem is NP
hard Complex Integer Programming problem. In this paper, we explored vari-
ous issues like network architecture, formulation of objective function, con-
straints of objective function, algorithm used, optimization strategy, and intensi-
fication and diversification methods in algorithms.
Keywords: wireless telecommunication networks, assignment problem, NP
hard, complex integer programming problem, objective function, constraints.
1 Introduction
Drastic increase in the use of mobile phone systems and the need for various adaptive
demands has attracted recent research attention towards wireless telecommunication
networks. In this paper, we consider three basic network architectures namely
Personal Communication Services (PCS), Cellular Mobile Network, and Universal
Mobile Telecommunication Services (UMTS). Broadly, the above three wireless
networks share analogous network architecture. The wireless telecommunication
network consists of geographically distributed hexagonal structure called cells. These
cells are interconnected in a hierarchical order. Each cell has a Base Transmission
Station (BTS). For communication, any registered cellular mobile device has to
transmit through BTS, in turn connected through Base Station controllers (BSCs). The
core network consists of Public switched Network (PSTN) and IP core network. Con-
ventionally, cells and switches are stationary and their locations are already known.
As given by Houeto et al. [2], in the real world, large proportion of budget has been
allocated to the costs of network facilities that carry traffic from cell sites to switches
and thus there exists pressure to reduce costs. One of the key ways to achieve cost
reduction is optimal design of wireless telecommunication networks. These networks
408 K. Rajalakshmi and M. Hima Bindu
involve the assignment of each cell to a switch, while taking into account a certain
number of constraints including capacity constraints, routing redistribution and hand-
offs frequency.
2 Wireless Telecommunication Networks
Basically wireless telecommunication networks like PCN, Cellular Mobile Network,
and UMTS share analogous network architecture; their difference lies in the values of
the four parameters, namely, frequency band, distance between cell sites, rate of data
transfer and technology used by the communication network. As given by Grillo et al.
[4], the frequency band of PCS is 1.7 GHz to 1.88 GHz, which is comparatively
higher than the frequency band width of Cellular Mobile network with 800 MHz and
UMTS frequency with 1885-2025 MHz. Chamberland et al. in [5] detail the design
problem of Cellular Mobile Network. In terms of, the number of cell sites required,
PCS needs cells sites within few feet apart, impacting need for more number of net-
work devices, which as per Houeto et al. [2] is the major factor for increase in net-
work operating cost. UMTS is a third generation telecommunication technology with
a data transfer rate of 42Mbs, higher than PCS and Cellular Mobile Network. Most
common technology used by UMTS is W-CDMA.
3 Objective Functions
Various parameters to formulate the preliminary objective function, for single homed
signal handoff assignment problem are listed below:
N: Number of cells in the Network
M: Number of Switches in the Network
Cik : Cable cost for existing link between cell i and switch k
Hij : Handoff Cost between cell i and cell j per unit time
Xik: 1 if cell i and switch k are connected, otherwise zero.
As given in [1], along with assigning cells to switches in an optimal manner using
link cost, handoff cost is also considered. The handoff cost consists of both complex
handoff and the simple handoff cost. Let Hij be the cost per unit of time for a simple
handoff between cells i and j, providing cell i and j are connected to the same switch.
Let H’ij be the cost per unit time for a complex handoff between cells i and j, con-
nected to different switches. Thus the objective function is given as
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
∑
=∑
=
∑
=∑
=++
∑
=∑
=
N
i
N
jij
H
N
i
N
jij
H
N
i
M
kik
X
ik
C
111111
min '
(1)
In [2] [6], Hij and H’ij are proportional to the handoff frequency between cells i and j
and thus deduces the overall handoff cost as,
hij = H’ij – Hij (2)
An Overview of Solution Approaches for Assignment Problem 409
As per Quintero et al. [3], the total operating cost of a cellular network includes
the monthly amortization cost of installed switches. Let INSk express the monthly
amortization cost for each installed switch k (k = 1 . . . m). Thus objective function
given in equation (2) becomes,
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
∑
=∑
=
∑
=∑
=+++
∑
=∑
=∑
=
N
i
N
jij
H
N
i
N
jij
HINS
N
i
M
kik
X
ik
C
M
k
k111111
min '
1
(3)
4 Constraints
In general the objective function of minimization of total network operating cost is
constrained by factors like (1) single homed constraint (2) switch constraint and (3)
handoff constraint.
Single homed constraint:
The single homed cell indicates that each cell is connected to only one single switch,
that is, cell i is linked to only one switch at any time.
Switch capacity constraint:
The number of connectivity requested through BTS of each cell for a particular
switch, must be less than or equal to the remaining call handling capacity Pk currently
available at switch k.
Handoff constraints:
To formulate the handoff cost per unit, we introduce
Zijk = Xik Xjk for i, j= 1,2,3 … N and k = 1,2 ,…M (4)
When cell i and cell j are connected to same switch k, then Zijk equals one and other-
wise equals zero. Equation (4) is a nonlinear binary product. Converting this nonlin-
ear binary constraint into linear binary constraint we have
Zijk ≤ Xik ; Zijk ≤ Xjk ; Zijk ≥ Xik + Xjk - 1 ; Zijk ≥ 0 ; (5)
The generalized handoff constraint between two cells is given by
Nji andj i for
M
kijk
Z
ij
Y,...2,1,
1=≠
∑
=
=
(6)
5 Algorithmes
In 1995, Merchant et al. [1] introduced heuristic methods for solving NP-hard As-
signment Problem. Comparative analysis of Integer Linear Programming with Heuris-
tics confirmed the excellent performance of heuristic algorithm. However, the heuris-
tic algorithm exhibits high probability of getting struck at local optima. The Taboo
410 K. Rajalakshmi and M. Hima Bindu
Search method is an improvement of the general descending algorithm, because it
attempts to avoid the trap of local minimum as in [2]. The taboo list is used to avoid
the solution already explored during the generation of the set of neighbor candidates.
Quintero et al. [3] proposes Genetic Algorithms, which are robust search techniques
based on natural selection and genetic production mechanisms perform a search by
evolving a population of candidate solutions through nondeterministic operators and
by incrementally improving the individual solutions through selection, crossover, and
mutation. Memetic algorithms (MAs) [6] are population-based heuristic search, in-
spired by Dawkins’ notion. A meme is defined as a unit of information that repro-
duces itself while people exchange ideas. A meme is usually modified by the person
before passing it to the next generation. In memetic algorithm global search is made
on entire population, like genetic algorithms. As described by Menon et al. [7], the
simulated annealing (SA) is a simple and efficient heuristic. It controls slow conver-
gence of local search. The effectual annealing technique avoids unnecessary exchange
of assignments between cells and switches. The simulated annealing algorithm per-
forms search for feasible solution on those neighborhood solutions neglected by local
search algorithm. The acceptance probability allows the current assignment for further
local search if the objective cost is better than the current best solution.
6 Conclusion
In this paper overview of assignment problem in three different wireless networks has
been discussed. The overview is based on various issues like (i) network architecture
(ii) objective function (iii) constraints (iv) algorithms (v) optimization strategy. Algo-
rithms like memetic algorithm, genetic algorithm, tabu search algorithm, simulated
annealing have been discussed.
References
1. Merchant, A., Sengupta, B.: Assignment of cells to switches in PCS networks. IEEE/ACM
Transaction on Networking 3(5), 521–526 (1995)
2. Houeto, F., Pierre, S.: Assigning cells to switches in cellular mobile networks using taboo
search. IEEE Transactions on Systems, Man and Cybernetic - Part B: Cybernetic 32(3),
351–356 (2002)
3. Quintero, Pierre, S.: On the Design of Large-Scale UMTS Mobile Networks Using Hybrid
Genetic Algorithms. IEEE Transactions on Vehicular Technology 57(4), 2498–2508 (2008)
4. Grillo, D., Sasaki, A., Skoog, R.A.: Personal Communications—Services, Architecture, and
Performance Issues. IEEE Journal on Selected Areas in Communications 15(8), 1385–1389
(1997)
5. Chamberland, S., Pierre, S.: On the Design Problem of Cellular Wireless Networks. Wire-
less Networks 11, 489–496 (2005)
6. Quintero, Pierre, S.: A memetic algorithm for assigning cells to switches in cellular mobile
networks. IEEE Communications Letters 6(11), 484–486 (2002)
7. Menon, S., Gupta, R.: Assigning cells to switches in cellular networks by incorporating
a pricing mechanism into simulated annealing. IEEE Transaction on Systems, Man and
Cybernetics, Part B: Cybernetics 34(1), 558–565 (2004)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 411–416, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Design of Domain Specific Language for Web Services
QoS Constraints Definition
Monika Sikri
Cisco Systems India Pvt Ltd,
SEZ Unit, Cesssna Business Park,
Bangalore Karnataka 560103
msikri@cisco.com
Abstract. Semantic Webservices (SWS) has raised interest in mechanisms for
Ontological representation of Web Services. A number of mechanisms most
notably WSMO and OWL-S are being developed to represent the same. An
important area in description of Web Services is the QoS characterization and
discovery which is the focus of research for this paper. A Domain Specific lan-
guage is being proposed for definition of observable QoS characteristics and
conditions. The syntax of this proposed language is being kept closer to WSML
considering it the standard modeling language.
Keywords: Web Services, QoS, SOA, WSML.
1 Introduction
Semantic Web Services are slowly moving from research papers to actual deploy-
ment. Development in this domain is accelerated with the availability of a number of
open source projects like WSMX [2] to support fast adoption. An area of interest in
this domain is development of mechanism for QoS model representation. Two com-
mon ontologies exist in the form of WSMO [15] and OWL-S [11] for definition of
SWS. This work focuses on defining a Domain Specific Language (DSL) for repre-
sentation of QoS Parameters. Emphasis in this paper is primarily on DSL design
which can be both used for definition of QoS capability and also specifying QOS
constraints as required by Web Service client. A domain specific language is devel-
oped as an embedded language for Groovy and hence inherits some groovy con-
structs. The use of standard language allows for fast development and early adoption
of DSL.
2 Domain Specific Language
The proposed language is created as an embedded DSL for Groovy. Groovy [1] is a
dynamic language completely written on top of Java. It provides an easy learning
curve for Java developer [16]. Its dynamic nature supports an easy development of
embedded DSLs [5].The use of standard language also ensures that conditions on QoS
412 M. Sikri
constraints can be easily expressed without the need of an elaborate grammar as in
WSML Rule Syntax [3]. The user can use standard Java or Groovy syntax inter-
changeably.
As a Domain language for definition of QoS characteristics and constraints in Web
Services, the syntax to specify basic QoS characteristics and constraints needs to be
first specified and determined. At the same time, care is being taken to keep it closer
to ontological hierarchy of WSML structure.
3 Related Work
Specifying non functional properties in Web Services has been approached in differ-
ent ways .Extending WSDL to specify QoS is one of the approaches as has been pro-
posed by authors in [4] .This would require definition of concrete QoS properties.
Model-Driven specification of QoS parameters is another approach as also being
specified in [13]. UML system model is used to capture QoS in Web service function-
ality .This model driven approach is based on UML and is specific to it.QoS based
ontology model have also been proposed by [20].It uses OWL-S to define ontology
whereas the study of this paper focuses on WSML. Specifying QoS in UDDI and
using broker based architecture [14] to access service is another approach. This ap-
proach is limited by UDDI capabilities. QoS ontologies to semantically specify QoS
information in Web Services is another approach towards semantic Web Services
[9].There are existing ontologies and language which addresses different aspects of
QoS. A comparative study has been conducted of various discovery mechanisms [12]
and ontologies [18] to understand the pros and cons of various approaches. Frame-
works have also been proposed for QoS ontology based Web Service selection
[10].Authors in [7] have proposed QoS specification using OWL to cater to clients
based on contribution level. QoS ontology based on WSMO [19] as part of its defini-
tion specifies WSML which is a RDF/XML mechanism for definition of Semantic
Webservices. This is a standard ontological model for definition of QoS model .The
work conducted [17] has focused into modeling QoS characteristics using WSMO
though the disadvantage is it has further extended WSMO. The WSML framework
provides a proposed mechanism to specify QoS parameters both functional and non-
functional. Qos is not central to WSML and its definition using nonfunctional proper-
ties and rule syntax is quite complex. The work proposed in this paper is intended to
provide a DSL for better representation of QoS parameters and is in line with WSML
terminology .The DSL is also being implemented in Groovy. It is based on constraint
programming [6] ideas and provide a verbose syntactical mechanism for defining QoS
constraints.
4 Syntax of Proposed DSL
Listing 1.1 provides a sample of the Domain Specific Language. Various important
parts of DSL syntax and elements of QoS characteristics are then defined.
Design of Domain Specific Language for Web Services QoS Constraints Definition 413
DSL syntax is an important aspect of this design since it would play an important role
in its adoption. The salient features that this language exhibits are the separation of
multiple statements using newline and use of equal sign between property name and
its value to separate them. Various keywords are followed by curly brackets.
4.1 Key Elements of DSL
QosLanguage: This keyword forms the external most qualifier which is used to assert
the presence of QoS constraint language. Line 1 in Listings 1.1 defines this tag.
QosConcept: QosConcept is used to define QoS parameter/characteristics. It is an
important element since it defines the parameters on which QoS selection is based.
1 QosLanguage{
2
3 QosConcept {
4 name "Availability "
5 description "Describes the availability capability of Web Service"
6 unit Integer
7 lowerBound "0"
8 upperBound "100"
9 }
10
11 QosWebService{
12 name "webserviceQos"
13 capability "StockWS"
14 qosinterface{
15 Availability {
16 description " availability description"
17 value "8"
18 }
19 }
20 }
21
22 QosAxiom{
23 name "myQosCondition"
24 concept "Availability "
25 description "Describe the mechanism of comparing the availability "
26 conditionf{$valueÆ
27 return ( value > 50)
28 }
29 }
30
31 QosGoal{
32 description "Goals we wanted to achieve"
33 name "goalWeb"
34 axiomRequired [ "myQosCondition" ]
35 }
36 }
Listing 1.1. Domain Specific Language for Webservices Qos Parameters
414 M. Sikri
The concept is used to define both functional and non functional parameters like
availability and security. Lines 3 to 9 in Listings 1.1 define the QosConcept. Various
parameters that are defined as part of QosConcept are as follows
description A verbal comment on the content of this tag.
Name Unique name to identify this particular QoSConcept.
dataType Defines the datatype used to represent this QoSConcept.
This Datatype should be a well defined Groovy Datatype (or Java
datatype).
lowerBound Defines the lower expectable bound on the value of this QoS parame-
ter. The value is defined as a String. And it is converted into dataType
unit.
upperBound defined the upper threshold value on this parameter
QosWebService: This tag is used to define expected/observed QoS values for Web
Service provider and associates value for each QoS concept. This parameter is used
for binding QoS language to actual webservices provider. Elaborate mechanism need
to be developed for mapping the data to WSDL or UDDI based systems. Lines 11 to
20 in Listings 1.1 define the QosWebService. Its various parameters are as follows
description A verbal comment on the content of this tag.
name unique name to identify this particular WebService.
capability Use to associate webservice interface with this tag
qosinterface Embedded tag inside webservice which is used to specify values for
actual QOS concepts. A nested tag, which contains one tag for each
of QoSconcept. The rough syntax for Qosinterface is provided in
listing 1.2.
QosAxiom: This tag forms the center piece of QoS Language. It is used to define
actual conditions on QoS parameters. Condition on QoSAxiom is defined as closure.
It contains a groovy syntax for defining a condition statement on passed value of the
concept say "Availability". The term axiom is chosen from WSML ontology to pro-
vide for easy interoperability. Lines 22 to 29 in Listings 1.1 define the QosAxiom.
The parameters for QosAxiom are as follows
description A verbal comment on the content of this tag.
Name Specifies a unique name to identify this axiom.
concept name of the Qosconcept on which conditions are defined.
condition Closure structure to define actual condition for verification.
Design of Domain Specific Language for Web Services QoS Constraints Definition 415
Listing 1.2 defines syntax for qosinterface
Condition tag contains a closure which must return a boolean value true or false. This
implies, that the last statement in this Groovy block must evaluate to true or false.
Since this is a groovy block, user can write a valid groovy code to define the condi-
tion. One explicit value, which is current QoS concept value of a known Web Service,
is passed to this closure.
QosGoal: It is used to define requirements of Web Service consumer. It defines con-
ditions that web service consumer requires to be satisfied. The actual set of conditions
on individual QoS parameters is defined as part of QosAxioms. Lines 31 to 35 in
Listings 1.1 define QosGoal. The goals are used to specify set of axioms that need to
be satisfied as part of this webservice consumer. The syntax of QosGoal is as follows
QoSGoal is also a binding point for Qos Language to the actual webservice client (or
consumer). The success evaluation of goal ensures the selection of a QoS webservice
endpoint.
description A verbal comment on the content of this tag.
name unique name to identify this goal this axiom.
axiomRequired A groovy string array containing names of QoSAxiom (condi-
tions) that needs to be satisfied.
5 Conclusions
The intent of this work is to provide an easy mechanism for webservice QoS parame-
ters definition. QoS support for Web Services adds a competitive behavior to Web
Service and client interaction. It adds a lot of value in essential for an enterprise envi-
ronment [8] in usage of Web Services. This intends to provide at the same time some
level of inter op with WSML standard. Besides it also helps service consumer to make
the right choice based on differential QoS provided among similar Web Services. The
resultant embedded DSL can be easily be used to define QoS conditions in Semantic
Web Services environment.
Acknowledgments. I would like to thank my organization Cisco Systems Inc., and all
anonymous referees for the valuable contribution and suggestions.
1 QoSWebService{
2 name….
3 qosinterface {
4 <<qosconceptname>>{
5 description = " "
6 value = " "
7 }
8 }
9
}
416 M. Sikri
References
1. Groovy: An agile dynamic language for the java platform,
http://groovy.codehaus.org/
2. Web service execution environment, http://www.wsmx.org/
3. de Bruijn, J., Lausen, H., Polleres, A., Fensel, D.: The wsml rule languages for the seman-
tic web, http://www.w3.org/2004/12/rules-ws/paper/128/
4. Dai, C., Wang, Z.: A flexible extension of WSDL to describe non-functional attributes. In:
2010 2nd International Conference on e-Business and Information System Security
(EBISS), May 2010, pp. 1–4 (2010)
5. Devijver, S.: A groovy DSL from scratch in 2 hours,
http://groovy.dzone.com/news/groovy-dsl-scratch-2-hours
6. Garcia, J.M., Ruiz, D., Ruiz-Cortes, A., Martin-Diaz, O., Resinas, M.: An Hybrid, QoS-
Aware Discovery of Semantic Web Services Using Constraint Programming. In: Krämer,
B.J., Lin, K.-J., Narasimhan, P. (eds.) ICSOC 2007. LNCS, vol. 4749, pp. 69–80. Springer,
Heidelberg (2007), http://dx.doi.org/10.1007/978-3-540-74974-5_6
7. Ha, Y., Park, H.S.: Qos Based on Client information for Semantic Web Service. In: Ad-
vanced Software Engineering and Its Applications, pp. 237–240 (2008)
8. Kritikos, K., Plexousakis, D.: Requirements for Qos-based Web service Description and
Discovery. IEEE Transactions on Services Computing 2, 320–337 (2009)
9. Li, H., Du, X., Tian, X.: Towards Semantic Web Services Discovery with QoS Support
Using Specific Ontologies. In: International Conference on Semantics, Knowledge and
Grid, pp. 358–361 (2007)
10. Li-li, Q., Yan, C.: QoS Ontology based efficient web services selection. In: International
Conference on Management Science and Engineering, pp. 45–50 (2009)
11. Martin, D., Burstein, M., Hobbs, E., Lassila, O., Mcdermott, D., Mcilraith, S., Narayanan,
S., Parsia, B., Payne, T., Sirin, E., Srinivasan, N., Sycara, K.: OWL-S: Semantic Markup
for Web Services (November 2004), http://www.w3.org/Submission/OWL-S/
12. Nair, M.K., Gopalakrishna, V.: Article:Look Before You Leap: A Survey of Web Service
Discovery. International Journal of Computer Applications 7, 22–30 (2010); published By
Foundation of Computer Science
13. Ortiz, G., Bordbar, B.: Model-Driven Quality of Service for Web Services: An Aspect-
Oriented approach. In: IEEE International Conference Web Services (ICWS 2008), pp.
751–748 (2008)
14. Rajendran, T., Balasubramanie, P.: An Optimal Agent-Based Architecture for Dynamic
Web Service Discovery with QoS. In: International Conference on Computing Communi-
cation and Networking Technologies (ICCCNT), pp. 1–7 (2010)
15. Roman, D., Keller, U., Lausen, H., de Bruijn, J., Lara, R., Stollberg, M., Polleres, A.,
Feier, C., Bussler, C., Fensel, D.: Web Service Modeling Ontology. Appl. Ontol. 1,
77–106 (2005),
http://portal.acm.org/citation.cfm?id=1412350.1412357
16. Subramaniam, V.: Programming Groovy (2008)
17. Toma, I., Foxvog, D., Jaeger, M.C.: Modeling QoS Characteristics in WSMO. In: Proceed-
ings of the 1st workshop on Middleware for Service Oriented Computing (MW4SOC
2006), pp. 42–47. ACM, New York (2006),
http://doi.acm.org/10.1145/1169091.1169098
18. Tran, V.X., Tsuji, H.: A Survey and Analysis on Semantics in QoS for Web Services. In:
International Conference on Advanced Information Networking and Applications, pp.
379–385 (2009)
19. Wang, X., Vitvar, T., Kerrigan, M., Toma, I.: A QoS-Aware Selection Model for Semantic
Web Services, pp. 390–401 (2006),
http://dx.doi.org/10.1007/11948148_32
20. Zhang, H., Gao, W.: A Research on QoS-based Ontology Model for Web Services Discov-
ery. In: International Workshop on Knowledge Discovery and Data Mining, pp. 786–789
(2009)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 417–421, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Modified Auxiliary Channel Diffie Hellman Encrypted
Key Exchange Authentication Protocol
Nitya Ramachandran1 and P. Yogesh2
1 Department of Computer Science & Engineering
Anna University, Chennai
nityaram86@gmail.com
2 Department of Information Science & Technology
Anna University, Chennai
yogesh@annauniv.edu
Abstract. In the wireless network, security is a major concern. Bluetooth net-
work is also highly vulnerable to attacks. This paper proposes a protocol called,
Modified Auxiliary Channel Diffie Hellman Encrypted Key Exchange authenti-
cation Protocol (MACDHEKE), which authenticates the two previously un-
known Bluetooth enabled devices. The user inputs a low security, low entropy
pins and converts it into high security, high entropy shared key. It also authenti-
cates the public key as being sent by the claimed sources or devices. It also ana-
lyzes how secure this protocol is against man in the middle attack as well as
passive eavesdropping.
Keywords: Auxiliary channel, pin, shared key, elliptic curve, Diffie Hellman
key exchange protocol.
1 Introduction
A wireless network is a network where the devices are connected via radio signals
rather than wires. One of the important wireless networks is Bluetooth also called
Wireless Personal Area Network (WPANs). Bluetooth network connects devices in
close range say 1m - 100m approximately. It uses 2.4 GHz short range radio fre-
quency unlicensed Industrial Scientific Medical (ISM) band.
Since Bluetooth is a wireless technology, it is highly vulnerable to attacks. One
major attack is man in the middle attack, abbreviated as MITM attack. In this attack, a
third device gets in between two devices and establishes a connection impersonating
as a valid user. He can then intercept all messages send between the devices. This is a
form of active eavesdropping where the attacker can modify the messages. Another
form of attack is passive eavesdropping where the attacker only listens to the mes-
sages without modifying them. These attacks occur when two Bluetooth devices try to
pair with each other.
Rest of the paper is organized as follows. The second section deals with related
work. The MACDHEKE protocol, MITM attacks are given in the third and fourth
section respectively.
418 N. Ramachandran and P. Yogesh
2 Related Work
The different combinations of connectability and discoverability capabilities of a
Bluetooth device can be divided into three categories, or security levels namely silent,
private and public [5]. Since Bluetooth is a wireless communication system, it is
highly vulnerable to attacks such as jamming or intercepting and fabrication of infor-
mation being passed to the piconet devices [1]. Therefore, Bluetooth security is a
major concern now days.
There is a lot of research going on in Bluetooth pairing. Even with longer alpha-
numeric PIN’s, full protection against active eavesdropping cannot be achieved. It is
found that MITM attacks on Bluetooth versions up to 2.0+EDR can be still performed
[6]. Bluetooth versions 2.1+EDR and 3.0+HS (High Speed) add a new specification
for the pairing procedure, namely SSP [7, 10]. The new Simple Pairing procedure in
Bluetooth version 2.1 [8] tries to remedy a number of the weaknesses of previous
versions. Paper [11] suggests a method for authentication and key agreement achieved
through short string comparison and distance bounds.
The above version is enhanced by a key agreement procedure during pairing using
an authenticated challenge-response procedure over an auxiliary channel [9]. Com-
pared with [9] where only unidirectional auxiliary channel is required, [12] specifies
verification of the symmetric key using commitment value, pin and nonce and thus
require bi directional auxiliary channel.
3 Design of MACDHEKE Authentication Protocol
MACDHEKE is designed with aim of preventing man in the middle attack during
pairing between two devices.
The proposed protocol is as shown in figure 1. First, public keys, PKa and PKb of
device A and B are exchanged. Then device A and device B generates its own tempo-
rary elliptic curve public private key pair. That is for device A, the pair generated is
(TPKa, TSka) and for device B, pair is (TPKb, TSKb). The PIN is then sent over
auxiliary channel. Device A then encrypts TPKa with pin displayed in device B and
inputted in device A and send to device B. B then calculates temporary DHKey, Kba
= P192 ( TSKb, TPKa). That is, device B uses its temporary private key and device
A’s temporary public key. It then calculates commitment value, Cb = f’ (PKa, PKb,
Kba, TPKa), Where, f’ (W, X, Y, Z) = HMAC-SHA256Y (W||X||Z). This commitment
value is again encrypted using Vernam cipher. Temporary DHKey, Kba can be used
to encrypt the commitment value. But since there is possibility of the intruder to com-
pute this DHKey, instead using Kba directly, simple secret cryptographic function is
used by two devices to compute a symmetric shared key using DHKey, Kba as func-
tion argument. This encrypted commitment value is send across to device A along
with its temporary public key encrypted with pin A.
Device A then calculates its temporary DHKey, Kab = P192 (TSKa, TPKb) and
using this calculates the commitment of device B. This is verified against the com-
mitment value received from B after decryption. If the device B get verified, it means,
A is sure that it’s to device B it is communicating. Device A calculates its commit-
ment value Ca = f’(PKb,PKa,Kab,TPKb), encrypts it using symmetric shared key and
MACDHEKE 419
sends it to B. B decrypts it, calculates commitment value of device A and verifies. If
verification succeeds, then B is sure that it is communicating with the device A. Now
original DHKey is calculated by device A using P192 (SKa, PKb) and by device B
using P192 (SKb, PKb). After the last step, both devices are assured that they have
exchanged public key values and that a DH key is established between them (DHKe-
yab== DHKeyba). The temporary public and private keys are discarded after the final
DH key is established. It is important that TPKa is not revealed as it would allow an
attacker to brute force the PIN.
Analysis of the this cryptographic protocol is done using ProVerif which is a soft-
ware tool for automated reasoning about the security properties found in crypto-
graphic protocols. Secrecy of a key k can be modeled as a query for the attacker’s
knowledge k. If ProVerif’s algorithm terminates and the set of attacker knowledge
does not include k, then k has been proved to be secret [8].
MACDHEKE protocol can be written in pi calculus. The device A is specified as
initiator and device as responder. The query is checking whether pin A, pin B and
symmetric key is present in the knowledge base of the attacker or not after execution
of the protocol. If the keys are part of knowledge base of attacker, it means pins and
key is no longer secret. In this protocol, that attacker obtaining the values of pin A,
pin B and symmetric key is false.
A B
PKa
PKb
Input Pin in A and B
PINB PINA
Generate
TPKa/TSKa Generate
TPKb/TSKb
EPINB(TPKa) Calculate
Kba, Cb
EPINA(TPKb), ESKey(Cb)
Decrypt, Calculate
Kab, Cb, verify.
Calculate Ca ESKey(Ca)
Calculate
DHKeyab
Calculate
DHKeyba
Decrypt,
Calculate
Ca and verify
Fig. 1. MACDHEKE protocol
420 N. Ramachandran and P. Yogesh
4 MITM Attack Scenario
Figure 2 shows the attack scenario. The attacker first injects its own public key into
the network. Device A and B as well as the attacker will now generate temporary
public private key pair. Attacker then receives the encrypted temporary public key,
EpinB(TPKa) from device A which is encrypted with pinB send over the auxiliary
channel.
Attacker will not know this pin. So it guesses a pin, pinB’ and encrypts its tempo-
rary public key with pinB’ and send it over to device B. Attacker also decrypts
EpinB(TPKa) to obtain TPKa’. Device B decrypts it and calculates temporary DHKey,
Kbe as P192 ( TSKb, DpinB(EpinB’(TPKe))). Then B calculates commitment value,
Ceb= f’(Pke,Pkb,Kbe, DpinB(EpinB’(TPKe)). This value is encrypted as in previous
case. Attacker obtains this encrypted value and temporary public key of B encrypted
with pin A. The attacker again makes a guess on the pin of A as well as the key and
encrypts Cea calculated and also encrypts his temporary public key with guessed pin
of A.. At this point, attacker is mounting attack on device A. He then sends
EpinA’(TPKbe) and ESKey’(Cea) to device A.
Device A decrypts the received message to get TPKbe’ and Cea’. TPKbe’ will be
equal to TPKbe only if pinA = pinA’ and Cea will be equal to Cea’ only if Skey =
Skey’. Device A calculates DHKey, Kae and the commitment value, Cae. Here Cae
will be equal to Cea only if Kae = Kea. It then verifies with received Cea’. This will
A
PKa
PKe
Input Pin in A and B
PINB PINA
Generate
TPKa/TSKa Generate TPKeb
TSKeb,TPKea,
TSKea
EPINB(TPKa) Calculate
Kbe, Ceb
encrypt
EPINA’(TPKeb),
ESKey’(Cea)
Decrypt,
Calculate
Kae, Cb,Ca
and verify
FAIL Calculate
Ca and
verify
PKe
PKb
Guess PinA’
and Pin B’ EPINB’(TPKea)
Calculate TPKa’
,Kea,Keb EPINA(TPKb)
ESKey(Cbe)
B
ESKey’(Ceb)
FAIL
Generate
TPKb/TSKb
Fi
g
. 2. Attack Scenario
Attacker
MACDHEKE 421
fail since Kea is not equal to Kae because, TPKbe’ is not equal to TPKbe. The device
A aborts the connection.
Attacker then launches attack on device B by calculating the commitment value,
Ceb. It then encrypts Ceb with key’ and sends it to device B. Device B then calculates
Cbe. This verification also fails since attacker does not know pinB and key because of
which Kbe and Keb will not equal and therefore Cbe and Ceb also will not be equal.
Thus attacker fails in establishes a connection with either device A or B. So MITM
attack is prevented since both A or B detects the attack.
5 Conclusion
Modified Auxiliary Channel Diffie Hellman Encrypted Key Exchange authentication
protocol is proposed to establish a connection between two Bluetooth enabled de-
vices. It also authenticates the two previously unknown Bluetooth enabled devices.
The scheme protects against MITM attack and passive eavesdropping.
References
1. André, N., Klingsheim: J2ME Bluetooth Programming. Department of Informatics Uni-
versity of Bergen (June 2004)
2. Anoop, M.S.: Elliptic Curve Cryptography. An Implementation Guide
3. Hwang, M.-S.: A Secure Protocol for Bluetooth Piconets Using Elliptic Curve Cryptogra-
phy. Telecommunication Systems 29(3), 165–180 (2006)
4. Chris, K., Cockrum: Implementation of an Elliptic Curve Cryptosystemon an 8-bit Micro-
controller. Springer, Heidelberg (2009)
5. Haataja, K., Toivanen, P.: Two practical man-in-the-middle attacks on Bluetooth secure
simple pairing and countermeasures. IEEE Transactions on Wireless Communica-
tions 9(1), 384–392 (2010)
6. Haataja, K.: Security threats and countermeasures in Bluetooth-enabled Systems. Ph.D.
dissertation, University of Kuopio, Department of Computer Science (February 2009)
7. Bluetooth SIG, Bluetooth Specifications 1.0–3.0+HS,
http://www.bluetooth.com/Bluetooth/Technology/Building/
Specifications
8. Nilsson, D.K., Larson, U.E., Jonsson, E.: Auxiliary Channel Diffe-Hellman Encrypted
Key-Exchange Authentication. In: ACM QShine 2008, Hong Kong, China (July 2008)
9. Nilsson, D.K., Larson, U.E., Jonsson, E.: Unidirectional Auxiliary Channel Challenge-
Response Authentication. In: Proceedings of the 7th Annual IEEE Wireless Telecommuni-
cations Symposium (WTS), Pomona, CA, USA (2008)
10. Bluetooth Special Interest Group. Simple pairing white paper (2006)
11. Cagalj, M., Capkun, S., Hubaux, J.-P.: Key Agreement in Peer-to-Peer Wireless Networks.
In: Proceedings of the IEEE Special Issue on Cryptography and Security, pp. 467–478
(2006)
12. Jakobsson, M.: Lecture Notes in 1400/I590: Issues in Security and Privacy,
http://www.informatics.indiana.edu/markus/i400/
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 422–426, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Bilateral Partitioning Based Character Recognition for
Vehicle License Plate
Siddhartha Choubey, G.R. Sinha, and Abha Choubey
Shri Shankracharya College Of Engg and Technology, Junwani Bhilai
sidd25876@gmail.com, ganeshsinha2003@gmail.com,
abha.is.shukla@gmail.com
Abstract. This paper presents a new methodology for the image segmentation
and character recognition from standard Indian License number plates. This
method first, gets input of the segmented characters that is partitioned by bilat-
eral partitioning method, in which we eliminate similar part from the character
and match it by judging template and return identified character. This partition-
ing may be performed horizontally or vertically. The characters are to be parti-
tioned horizontally or vertically depend on their subgroup, and before this sub-
group the characters are grouped on the basis of the number of holes in it. The
subgroup is made on the basis of some similar features like | , / , \ , _ , ( , - , etc.
Suppose we have alphabet T and I here similar portion is I than both will go to
same subgroup and we partition it horizontally. This method eliminates the
problem of confusion between similar looking elements like C, G and T, I, 1, J
etc by exploiting the small but significant differences among them.
Keywords: Bilateral, Euler, Partitioning, Segmentation, Pixel Dentsity Distri-
bution.
1 Introduction
At present, a vital part in transportation is played by vehicles. In recent years, the
usage of vehicles has been increasing on account of population growth and human
needs. Hence control of vehicles is becoming a huge problem that is hard to solve [1].
An image processing technology called License Plate Recognition (LPR) that is used
to identify vehicles by their license plates is a kind of automatic vehicle identification
[2]. Image acquisition, license plate detection, character segmentation and character
recognition are four major phases of License Plate Recognition [3-4]. Several areas
including traffic volume control, unsupervised park monitoring, traffic law enforce-
ment and auto toll collections on highways extensively use license plate recognition
applications [5-6]. In this paper we have introduced a new algorithm for character
extraction and recognition based on based on fuzzy logic and analyzing some pattern
recognition based works.
Bilateral Partitioning Based Character Recognition for Vehicle License Plate 423
2 Related Works
Mi-Ae Ko et al. [7] proposed a algorithm for recognizing optimal solution and good
optimization effect in fast speed, by using filled function method and BP neural
network.Experimental results show that the performance of the proposed method is
simple and robust. Fei Lu Mei Xie et al.[8] focussed on extra dot appear in the license
plate , this extra “dot” can completely affect the “shape descriptors”, with an XOR-
based kernel, is used . The XOR-based kernel can improve the performance of the
system drastically. J1an-xia Wang et al. [9] emphasised on the time complexity for
template matching procedure. In order to enhance recognition speed and recognition
rate, an improved template matching method is presented.
3 System Overview
We proposed a novel method to recognize vehicle number plate elements which are
captured by any camera. The system has two stages viz. pre-processing and post-
processing. Pre-processing begins with taking input image then vehicle plate is ex-
tracted. Characters of the number plate are segmented. Post-processing stage includes
feature extraction and storing then in the form of templates for matching purpose
followed by character recognition process.
4 Methodology
The first step is vehicle number plate detection which is followed by character and
number segmentation, feature extraction & recognition of the extracted feature and
character recognition.
4.1 Preprocessing (Vehicle Number Plate Extraction)
In preprocessing we capture the image (Fig. 1) and convert it to gray image shown in
Fig. 2.
4.1.1 Vehicle Number Plate Elements Segmentation
Various preprocessing steps used are: (a) Converting image to gray scale image, (b)
Removing all objects smaller than 100 pixels. Fig. 3 shows the result of canny edge
filter. Fig. 4(a) shows vehicle plate extraction and converting into balck and white
gives Fig. 4(b). Finally, Fig. 4(c) shows the result with noise removed.
Fig. 1. Captured Image Fig. 2. Gray scale image Fig. 3. Result of canny filter
424 S. Choubey, G.R. Sinha, and A. Choubey
(a) Vehicle plate extraction (b) Coversion into black and white (c) Noise removal
Fig. 4. Results of extraction and noise removal
4.1.2 Removing Extra Area from Left and Right by Detection of First and Last
Columns for Each Character and Number
The first step in segmentation process is to cutoff the background from each character
and number from the license plate. Vertical scanning is used to detect first and last
columns for each character and number is conducted before horizontal scanning as
explained in Algorithm.
4.1.3 Remove Extra Area from Upper and Lower of Each Component
Horizontal scanning will be done to detect the first and last rows from the result of the
previous step. The results of this operation will be images that contain only the char-
acters or numbers without any extra area. This result will help to extract image fea-
tures easily. Fig. 5 shows the result of vertical segmentation and Fig. 6 depicts the
result horizontal segmentation.
Fig. 5. Vertical Segmentation Fig. 6. Horizontal Segmentation
4.2 Post Processing: Feature Extraction
The Features extraction step contains the following steps:
4.2.1 Algorithm (Post Processing) Detection of (Character and Numbers):
Grouping Process
STEP 1: Calculate the number of holes, h= number of holes in character
STEP 2: Calculate Euler number based on number of holes: e=c-h, where, c=1 //c is
the number of connected components (always equal to one)
STEP 3: Calculate Euler group based on Euler number: If e = 1, then eg=1; else if
e=0 then eg=2; else if e=-1 then eg=3; where eg is the Euler group.
Sub Grouping Process
STEP 4: For each Euler group match with subgroup deciding template for subgroup
deciding templates find their indexes.
STEP 5: Maximum match determines the subgroup. Return the respective subgroup.
STEP 6: Now once subgroup is decided compare with differentiating template: For
the subgroup templates find their indexes.
STEP 7: Maximum match determines the element: For maximum match return the
respective element (character and number).
It is clear that this grouping process will increase system accuracy of recognizing
characters and numbers and it will also reduce the time of recognition process.
Bilateral Partitioning Based Character Recognition for Vehicle License Plate 425
4.2.2 Character and Number Recognition
Here, we use a new methodology in which the characters are separated into subgroups
on the basis of their similarity or different physical appearance. Euler group 1 is di-
vided into the following subgroups: Subgroup C: consists of C and G. Subgroup E:
Consists of E and F, Subgroup I: Consists of I, T, J & 1 and Subgroup S: Consists of
5and S. Euler group 2 is divided into the following subgroups: Subgroup O: consists
of O and Q; Subgroup P : consists of P and R. All rest elements are not categorized
into any subgroup and have a separate template for each of them. Each group has a
template; the template that matches with the input the most decides the group of the
character. The subgroup is decided on the basis of indexes calculated by formulae:
Index of a particular template = (Number of pixels that matched with that template)
/ (Total number of pixels in that template). The subgroup whose template has highest
index with respect to the character is chosen as the subgroup.
4.3 Recognition of Elements belonging to Same Subgroup
After sub-grouping the physical characteristics of each element that differentiate it
from other elements are used to identify the element. In Subgroup E: E and F are
members of this subgroup. The template shown in Fig. 7 is used to check, if the input
belongs to this group then the template shown in Fig. 8 is used to detect if the input is
E or F as this template is a part of E but not F.
Fig. 7. Template to judge group Fig. 8. Template to judge whether E or F
5 Result and Conclusion
In this paper, we have proposed a new method for character recognition which very
effectively eliminates the problem of confusion between characters of same type like
C & G which the previous algorithms were not able to do with high precision. Initial
experiments have shown success in differentiating between similar looking charac-
ters. Inspite of being very effective on standard number plate suffers with the
disadvantage that it is not usefull on those plate which use non-standard characters to
represent the vehicle’s registration number. Fig. 9 shows results for some characters.
Fig. 9. Results for some characters
426 S. Choubey, G.R. Sinha, and A. Choubey
References
[1] Shan, B.: License Plate Character Segmantation and Recognition Based on RBF Neural
Network. In: 2nd International Workshop on Education Technology and Computer Sci-
ence, vol. 2, pp. 86–89 (2010)
[2] Broumandnia, Fathy: Application of pattern recognition for Farsi license plate recognition.
ICGST International Journal on Graphics, Vision and Image Processing 5(2), 25–31
(2005)
[3] Cancer, H., Gecim, H.S., Alkar, A.Z.: Efficient Embedded Neural – Network Based Li-
cense Plate Recognition System. International Journal of Information and Security 57(5),
2675–2683 (2008)
[4] Feng, J., Li, Y., Chen, M.: The Research of vehicle License Plate Character Recognition
Method Based on Artificial Neural Network. In: 2nd International Asia Conference on In-
formatics In Control, Automation And Robotics, vol. 8(1), pp. 317–320 (2010)
[5] Tsai, I.-C., Wu, J.-C., Hsieh, J.-W., Chen, Y.-S.: Recognition of Vehicle License Plates
from a Video Sequence. IAENG International Journal of Computer Science 36(1), 26–33
(2009)
[6] Deb, K., Chae, H.-U., Jo, K.-H.: Vehicle License Plate Detection Method Based on Sliding
Concentric Windows and Histogram. Journal of Computers 4(8), 771–777 (2009)
[7] Zhang, Y., Xu, Y., Ding, G.: License plate character recognition based on fill function
Method Training BP Neural Network. In: Control and Decision Conference (CCDC 2008),
Chinese, pp. 3886–3891 (2008)
[8] Lu, F., Xie, M.: An Efficient Method of License Plate Location in Complex Scene. In:
Second International Conference on Computer Modeling and Simulation, ICCMS 2010,
Sanya, Hainan, January 22-24, vol. 2 (2010)
[9] Wang, J.-X., Zhou, W.-Z., Xue, J.-F., Liu, X.-X.: The research and realization of vehicle
license plate character segmentation and recognition technology. In: Proceedings of the
2010 International Conference on Wavelet Analysis and Pattern Recognition, Qingdao,
July 11-14 (2010)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 427–430, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Strategies for Parallelizing KMeans Data Clustering
Algorithm
S. Mohanavalli, S.M. Jaisakthi, and C. Aravindan
SSN College of Engineering, Chennai, India
{mohanas,jaisakthism,aravindanc}@ssn.edu.in
Abstract. Data Clustering is a descriptive data mining task of finding groups of
objects such that the objects in a group will be similar (or related) to one an-
other and different from (or unrelated to) the objects in other groups [5]. The
motivation behind this research paper is to explore KMeans partitioning algo-
rithm in the currently available parallel architecture using parallel programming
models. Parallel KMeans algorithms have been implemented for a shared mem-
ory model using OpenMP programming and distributed memory model using
MPI programming. A hybrid version of OpenMP in MPI programming also has
been experimented. The performance of the parallel algorithms were analysed
to compare the speedup obtained and to study the Amdhals effect. The compu-
tational time of hybrid method was reduced by 50% compared to MPI and was
also more efficient with balanced load.
Keywords: Data Mining, Clustering, Parallel KMeans, OpenMP, MPI.
1 Introduction
Data Mining is an art of extracting interesting (non-trivial, implicit, previously
unknown and potentially useful) information or patterns from large data sets [5].
Clustering is an unsupervised data mining technique applied on a given data set to
partition it into groups of similar data objects. The KMeans algorithm is one of the
simplest & popularly practised partitioning algorithm based on similarity distance [4].
Sequential design of such a computationally intensive algorithm may not be very
efficient and hence can be parallelized to improve performance. The success of such
straight forward parallelization of a sequential algorithm relies on a good data organi-
zation and decomposition strategy [2]. Many researchers have worked in the area of
parallel data clustering. Zhang et al. have proposed an enhanced parallel algorithm
using Master/Slave model and dynamic load balancing technique [12]. Rao et al. have
explored the potential of Multi-Core hardware under OpenMP API and POSIX
threads for both static and dynamic threads [11]. Chin et al. have addressed the
problem of higher communication overhead, by adopting hybrid MPI and OpenMP
programming model [1]. Zhou et. al have tried a distributed KMeans algorithm to
improve the local clustering and decrease network load [13]. Jin et al. have focused in
reducing the number of iterations performed to cluster the data by carefully choosing
the initial centriods by sampling [9]. Datta et al. have proposed two solutions to
428 S. Mohanavalli, S.M. Jaisakthi, and C. Aravindan
address the distributed clustering problem in a P2P network using local synchroniza-
tion and uniform sampling of peers [3]. Most of the above research works aim to
speed up the convergence and increase the efficiency of the algorithm. In this research
paper the parallel approaches to perform KMeans clustering has been experimented
and observed that the hybrid method outperforms the other methods.
2 Proposed System Design
The sequential KMeans algorithm design has an iterative structure where the centroid
updations are performed in the outer loop and the cluster labels are assigned to each
data point in the inner loop. The latter task was performed parallely using OpenMP
fork-join model. In the MPI model, data points were distributed evenly among the
processors, data labeling and updation of new centroids were done parallely by means
of message passing between the MPI processes. In the hybrid model a combination of
MPI with OpenMP was tried where the data points were evenly distributed among
the MPI processes and the task of data labeling for each local data within the process
was done using OpenMP directives. The performance of the algorithms were com-
pared in terms of speedup and efficiency for varying problem sizes and the hybrid
method was observed as the best parallel algorithm for clustering the input data set.
3 Experimental Results and Performance Analysis
A parallel algorithm is analysed for its cost effectiveness in terms of speedup and
efficiency (processor utilization). A cost optimal parallel algorithm has speedup p
(number of processors) and efficiency 1 in an ideal situation [10].
Table 1. Observed Execution Time for Parallel K-Means using MPI and Hybrid
Clustering Time in seconds
k 5 7 9 11 13
N MPI Hybrid MPI Hybrid MPI Hybrid MPI Hybrid MPI Hybrid
2 17.47 9.01 17.9 9.22 18.39 9.58 18.78 9.61 19.2 9.83
3 12.05 6.49 12.29 6.59 12.57 6.63 12.89 6.76 13.17 6.93
4 9.31 5.11 9.49 5.14 9.73 5.3 9.86 5.65 10.15 5.67
5 7.61 7.73 7.92 8.02 8.21 7.17 8.46 7.36 8.34 7.74
6 6.69 6.71 6.77 6.88 7.1 7.09 7.01 7.17 7.42 7.35
7 6.42 5.89 5.84 5.98 6.08 6.09 6.16 6.22 6.26 6.4
8 7.34 5.33 5.28 5.32 5.42 5.47 5.69 5.68 5.76 5.68
9 6.66 6.26 6.48 6.41 6.53 6.49 7.59 6.71 6.86 6.86
10 7.57 6.33 6.34 6.48 6.42 6.55 8.43 6.59 6.78 6.84
The sequential and all the versions of parallel KMeans algorithm were imple-
mented and initially tested with benchmark data sets from the UCI repository [6] to
ensure the correctness of the algorithms. Their parallel performance was tested with a
forest data set with 581012 data points and 54 attributes which has ideally 7 clusters
[6]. The algorithms were run for various values of k and n and the observed clustering
Strategies for Parallelizing KMeans Data Clustering Algorithm 429
time is tabulated in Table 1 and Table 2. The speedup and efficiency obtained by
running the algorithms is tabulated in Table 3 and Table 4. The algorithms show lin-
ear speedup and the Amdhals effect [10] was also studied to find the saturation point
in execution time with increase in number of processors. Both the MPI as well as
Hybrid MPI/OpenMP versions, show much improvement in speedup and efficiency
compared to the other versions as summarized in Table 2.
Table 2. Performance Comparison for Sequential, OpenMP, MPI, Hybrid
Performance Analysis
Best Clustering Time in sec Speedup* Efficiency*
K SEQ OMP MPI Hybrid OMP MPI Hybrid OMP MPI Hybrid
5 18.8 16.9 6.42 5.11 1.11 2.93 3.68 0.56 0.41 0.46
7 32.7 24.47 5.28 5.14 1.34 6.19 6.15 0.67 0.77 0.77
9 37.9 30.23 5.42 5.3 1.25 6.99 7.15 0.63 0.87 0.9
11 76.2 39.25 5.69 5.65 1.94 13.39 13.49 0.97 1.67 1.68
13 51.73 26.62 5.76 5.67 1.94 8.98 9.12 0.97 1.12 1.14
Avg 43.47 27.49 5.77 5.45 1.52 7.67 7.89 0.76 0.98 0.99
* - N is the minimum no of process required as per Amdhal’s effect
Fig. 1. Speedup for Parallel KMeans using MPI Fig. 2. Speedup for Parallel KMeans using
Hybrid MPI/OpenMP
The Amdhal’s effect studied for various runs of the MPI and hybrid versions are
depicted in Figure 1 & 2. The efficiency obtained using the hybrid method was almost
double that of MPI algorithm, as the number of processors required to achieve best
execution time using hybrid was half that required with MPI. The experiments were
run in a cluster with 4 dual core systems. For varying values of k and n, the hybrid
algorithm showed its best speedup with 4 nodes as it utilized both the cores to run the
MPI code embedded with OpenMP threads. For higher n values the effect of load
imbalance was seen as a drop in speedup and improved with load balance as shown in
in Figure 2. For MPI the speedup gradually increased with increase in n and saturated
at n = 8, beyond which the performance dropped as in Figure 1.
430 S. Mohanavalli, S.M. Jaisakthi, and C. Aravindan
4 Conclusion and Future Work
In this research work, parallel k means clustering algorithms suitable for shared as
well as distributed memory systems were implemented and studied for feasibility of
parallelism and scalability of the algorithm. It was evidently shown that the hybrid
version of the parallel algorithms was more efficient than the other approaches con-
sidering only the clustering time and is thus more suitable for distributed data mining.
It is proposed to further improve the design to reduce the effect of I/O time by using
appropriate file systems and parallel MPI-I/O routines.
References
1. Wu, C.-C., Lai, L.-F., Yang, C.-T., Chiu, P.-H.: Using hybrid MPI and OpenMP pro-
gramming to optimize communications in parallel loop selfscheduling schemes for multi-
core PC clusters. Journal of Supercomputing (2009), doi: 10.1007
2. Skillicorn, D.B.: Strategies for parallel data mining. IEEE Concurrency 7, 26–35 (1999)
3. Datta, S., Gianella, C.R., Kargupta, H.: Approximate distributed k-means clustering over a
peer-to-peer network. IEEE Transactions on Knowledge and Data Engineering 21(10),
1372–1388 (2009)
4. Dhillon, I., Modha, D.: A Data Clustering algorithm on distributed memory multiproces-
sors. IEEE Transactions on Knowledge and Data Engineering (KDD 1999), 47–56 (1999)
5. Han, J., Kamber, M.: Data Mining:Concepts and Techniques, 2nd edn. Morgan Kauf-
manm, San Francisco (2006)
6. http://archive.ics.uci.edu/mll UC Irvine Machine Learning Repository
7. http://www.openmp.org
8. http://www-unix.mcs.anl.gov/mpi
9. Jin, R., Goswami, A., Agrawal, G.: Fast and Exact out of core and distributed KMeans
clustering. Knowledge and Information Systems, 17–40 (2006)
10. Quinn, M.J.: Parallel Programming in C with MPI and OpenMP. Tata Mc- Graw Hill
(2003)
11. Rao, S.N.T., Prasad, E.V., Venkatehwarulu, A.: Critical Performance Study of Memory
Mapping on Multi-Core Processors: An Experiment with k-means Algorithm with Large
Data Mining Data Sets . International Journal of Computer Applications (0975 8887) 1(9)
(2010)
12. Zhang, X.Z., Mao, J., Ling Ou, L.: The Study of Parallel KMeans Algorithm. In: Proceed-
ings of the 6th WCAIAC, pp. 5868–5871. IEEE, Los Alamitos (2006)
13. Zhou, J., Liu, Z.: Distributed Clustering Based on K-means and CPGA. In: Proceedings of
FSKD(2), pp. 444–447. IEEE, Los Alamitos (2008), doi:10.1109
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 431–434, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Comparative Study of Different Queuing Models Used
in Network Routers for Congestion Avoidance
Narendran Rajagopalan and C. Mala
Department of Computer Science and Engineering,
National Institute of Technology, Tiruchirappalli,
Tamil Nadu, India-620015
narenraj1@gmail.com, mala@nitt.edu
Abstract. As the network traffic is increasing exponentially, the queuing model
used in a network decides the degree of congestion that is possible in the net-
work infrastructure. Hence using a suitable queuing model based upon the net-
work infrastructure like buffer size and the type of data traffic flowing through
the network will help in better utilization of system resources by minimizing
congestion to the best. In this paper, we study the different queuing models that
have evolved over time and how they suit the application.
Keywords: Queue Management, Temporal Fairness, Wireless LAN, RED,
QoS, ECN.
1 Introduction
As the network has evolved, the traffic has increased exponentially and also the com-
puting power of the intermediate nodes like routers and gateways, leading to more
packet drop and making queuing models an active area of research. Queuing disci-
plines[1][2] are algorithms which control how packets are buffered and transferred
through the network. The most basic queuing model proposed earlier was First In
First Out (FIFO) or Drop Tail[4][6]. In this model, all the packets are treated equally
by servicing them in the same order that they were placed into the queue. FIFO is the
widely implemented queuing algorithm in the Internet. All the incoming packets are
buffered until the buffer is full, then the incoming packets are dropped until there is
memory to accept the incoming packets.
The two issues of FIFO queue are lockout and fullqueues. Lockout is a phenome-
non in which more flows are denied of service as few flows occupy the queue to
capacity. Fullqueue can occur due to timing issues, leading to large oscillations in
the network utilization as more packets are dropped. It forces many flows to reduce
their load due to congestion notification and falls below the network capacity. This
oscillation between high and low nullifies the very purpose of using buffers for
smoother flow.
Random Early Detection(RED)[3][4] queue model was proposed to overcome the
lockout and fullqueue issues. The design goals of RED include congestion avoidance
by controlling the average queue size, avoidance of global synchronization and bias
against bursty traffic. In RED model, the average queue size is marked low and the
432 N. Rajagopalan and C. Mala
actual queue size is allowed to fluctuate to accommodate bursty traffic. It calculates
the average queue size using a low pass filter, with an exponential weighted moving
average. If the average queue size is below the minimum threshold, none of the in-
coming packets are marked called as no drop mode. Marking is a process of sending a
notification to the process about the congestion, which is possible only if Explicit
Congestion Notification(ECN) support is available, else the packet is dropped. If the
average queue size is above the maximum threshold, all the incoming packets are
marked, called as forced drop mode. When the average queue size is between the
minimum threshold and maximum threshold, the incoming packet is marked with the
probability Pa, which is calculated as a function of the average queue size called prob-
abilistic drop mode. The probability of a packet from a particular flow getting marked
is proportional to the share of the bandwidth used by the flow.
In the BLUE[5] model, instead of using the average queue size, the rate of packet
loss and link utilization history is used for congestion avoidance. Practically, network
traffic is non-poisson, making average queue length as an indicator to congestion
dubious. In the implementation, ECN uses two bits of the type of service (ToS) field
in the IP header. When BLUE decides that a packet must be dropped or marked, it
examines one of the two bits to determine if the flow is ECN capable. If it is not
ECN-capable, the packet is simply dropped. If the flow is ECN-capable, the other bit
is set and used as a signal to the TCP receiver that congestion has occurred. The TCP
receiver, upon receiving this signal, modifies the TCP header of the return acknowl-
edgment using a currently unused bit in the TCP flags field. Upon receipt of a TCP
segment with this bit set, the TCP sender invokes congestion-control mechanisms as
if it had detected a packet loss.
The Stochastic Fair BLUE(SFB)[5] queuing model is based on bloom filters for
protecting TCP flows against UDP flows with very little overhead and small amount
of buffer space. In a WLAN, Access Point is a potential bottleneck because the most
widely used TCP services such as web browsing or file retrieval has a very high data
traffic on the downlink compared to the uplink traffic.
2 Analysis of Different Queuing Algorithms
The advantages of FIFO or Drop tail queuing model are, it is simple and easy to im-
plement. It does not reorder the packets and has deterministic maximum delay. FIFO
is best suited for networks with no bursty traffic hence the queue depth being low.
FIFO also has the following drawbacks. It is not suitable for Quality of Service(QoS)
as it impacts all flows equally and does not differentiate priority classes of traffic.
When the traffic flow is bursty, it results in increased delay and jitter. During conges-
tion, unfairly benefits UDP traffic over TCP.
RED algorithm is the most well known AQM algorithm. It is costly to implement
and requires special supports like Explicit Congestion Notification(ECN). The average
queue size is used as the indication for congestion measurement. In BLUE algorithm,
the congestion measurement is achieved by rate of packet loss and link utilization his-
tory. This algorithm is best suited for bursty non-poisson traffic. The above mentioned
AQM algorithms do not work well for multi-rate WLANs, and squeeze TCP traffic and
increase UDP traffic flow.
It can be observed from the above studies that none of the algorithms can be gener-
alized to be best for all kinds of traffic and network topologies. Hence proper analytical
A Comparative Study of Different Queuing Models Used in Network Routers 433
study is to be done to understand the advantages and disadvantages of different queu-
ing algorithms and choose one best suiting the required case.
3 Simulation of Different Queuing Models Using NS-2.33
To observe the behavior of different queuing models on a network model, Network
Simulator 2.33[7][8] is used. The number of nodes used is 10, with 5 senders and 5
Fig. 1. Rate of Bytes dropped for Total Transmission for Different Queuing Models
Fig. 2. Graph representing Bytes in Queue during transmission for Different Queuing Models
434 N. Rajagopalan and C. Mala
receivers, each with a bandwidth of 10Mbps. The senders and the receivers are sepa-
rated by two routers linked together with a bandwidth of 1 mbps link. This link be-
tween the routers is buffered with a maximum queue limit of 50. The simulation time
is 100 seconds. Different queuing models are applied between the routers and ana-
lyzed. It can be observed from the graphs that FIFO model performs well until the
queue gets filled up. After the queue is full, the incoming packets are dropped con-
tinuously, hence the packet drop rate is exponential. The contents of the queue remain
continuously full until some packets are delivered. But in the case of RED, BLUE and
SFB queue models, the total bytes in the queue are kept in check with ECN ability.
Hence perform better than FIFO in a congested heavy traffic environment.
References
1. Seok, Y., Park., J., Choi, Y.: Queue Management Algorithm for Multirate Wireless Local
Area Networks. In: The 14th IEEE International Symposium on Personal, lndoor and Mo-
bile Radio Communication Proceedings (2003)
2. Huang, J., Wang., J., Jia, W.: Downlink Temporal Fairness in 802.11 WLAN Adopting the
Virtual Queue Management. In: IEEE WCNC 2007 Proceedings (2007)
3. Barrera, I.D., Bohacek, S., Arce, G.R.: Statistical Detection of Congestion in Routers. IEEE
Transactions on Signal Processing (2009)
4. Almomani, O., Ghazali, O., Hassan, S., Nor, S.A., Madi, M.: Impact of Large Block FEC
with different Queue sizes of Drop Tail and RED Queue Policy on Video streaming Quality
over Internet. In: Second International Conference on Network Applications Protocols and
Services (NETAPPS). IEEE, Los Alamitos (2010)
5. Feng, W., Shin, K.G., Kandlur, D.: The BLUE Active Queue Management Algorithms.
IEEE, Los Alamitos (2001)
6. Stanojevic, R., Shorten, R.N., Kellett, C.M.: Adaptive tuning of drop-tail buffers for reduc-
ing queuing delay. IEEE Communications Letters (2006)
7. Gao, W., Wang, J., Chen, J., Chen, S.: A Prediction-based Fair Active Queue Management
Algorithm. In: Proceedings of the 2005 International Conference on Parallel Processing
(2005)
8. Network Simulator, http://www.isi.edu/nsnam/ns
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 435–439, 2011.
© Springer-Verlag Berlin Heidelberg 2011
SAR Image Classification Using PCA and Texture
Analysis
Mandeep Singh1 and Gunjit Kaur2
1 Asst. Prof., Electrical and Instrumentation Engineering Dept.
Thapar University, Patiala-147004 India
2 Associate System Engineer, IBM India Pvt Ltd
Pune, Maharastra- 411057 India
Abstract. In Synthetic Aperture Radar (SAR) images, texture and intensity are
two important parameters on the basis of which classification can be done. In
this paper, 20 texture features are analyzed for SAR image classification into
two classes like water and urban areas. Texture measures are extracted and then
these textural features are further shortlisted using statistical approach, discri-
minative power distance and principal component analysis (PCA). In this study
30 SAR images are studied for 20 texture features. Finally, most effective 5 tex-
ture features are shortlisted for the classification of SAR images and accuracy is
calculated by Specificity and Sensitivity test. The results obtained from test im-
ages give an accuracy of 95% for image classification.
Keywords: Texture analysis, Classification SAR images, PCA analysis.
1 Introduction
SAR (Synthetic Aperture Radar) image classification has been used in numerous ap-
plications like map updating, urban area classification and classification of extremely
randomized clustering forests, Automatic Target Recognition (ATR) and a list to
mention. Texture holds very useful information in synthetic aperture radar (SAR)
image classification. There are different approaches which can be used to identify
texture patterns in a given image. For efficient classification of water cover or land
cover or urban area coverage, textural measures have to be chosen suitably [1]. Tex-
ture is an intrinsic property of virtually all surfaces. Classification, integrating both
intensity and textural measures, will be effective in segmenting SAR images. The use
of sum and difference histograms is presented for texture-based classification [2].
Clausi has compared and fused co-occurrence features, Gabor, and MRF features for
sea SAR imagery classification [3]. The use of wavelet analysis for identifying specif-
ic targets is performed [4]. Dekker has investigated texture measures like histogram
measures, lacunarity, wavelet energy, and fractal dimension for map updating capabil-
ity [5]. In this work, the images are taken from Image database of Essex University
(UK). The images are analysed to find texture measures using GLCM (Gray Level Co
occurrence Matrix) and GLRLM (Gray Level Run Length Matrix). The Discrimina-
tive power and PCA (Principal component analysis) are used for dimension reduction.
436 M. Singh and G. Kaur
2 Methodology
In this paper, the various Haralick’s measures of texture [6] are computed and ana-
lysed to find the water and building areas in the SAR images. Some of the texture
measures are also found using GLRLM. A gray level run is a set of consecutive, colli-
near picture points having the same gray level value [8]. The length of the run is the
number of picture pixels in the run. The main reason for the use of the GLRLM's fea-
tures is that it reflects the size of the texture elements. In the present work, total seven
run length features are used to find the texture features. Chu et al [10] proposed two
new features to extract gray levels information in the GLRLM and they are also used
in this study; Low Gray Level Run Emphasis (LGRE), High Gray Level Run Empha-
sis (HGRE). The above 20 texture features are calculated for 30 SAR images taken
from Image database of Essex University (UK). 30 sub images each of water and ur-
ban area are used to form two separate training sets, and for each image 20 texture
parameters are extracted. Some of the SAR test images are shown in figure 1.
Fig. 1. Test image of Water (left) and test image of Urban area (right)
The images similar to these are analysed. From each image, a square ROI window
(128×128) for Water and Urban land is selected and called them as Test image. After
finding these training sets, box plots of these training sets are plotted taking two sets
of data at a time. On the basis of box plots, it can be found that which texture feature
has more value of class seprability. To further reduce the dimension, discriminative
power distance is used to discriminate between two sets on the basis of Fisher’s dis-
criminative distance.
where subscript 1 is for water and subscript 2 is for land. Therefore, µ1 is the mean of
water and µ2 is mean of urban area coverage and σ12 and σ22 are the variances of water
and urban area coverage respectively. After calculating the ‘d’, top 10 features having
d > 1.5. To further reduce the dimension of the required number of features, we have
found out Pearson’s correlation coefficient of all the texture features for both water
and urban area coverage. In the last, Principal Component Analysis is used to verify if
the reduced feature sets are the same as obtained.
SAR Image Classification Using PCA and Texture Analysis 437
3 Results and Discussion
The training sets of water and urban area coverage are plotted in box plots to find out
if a given feature is discriminative. From the box plots, we have reduced the 20 fea-
tures to 14 discriminative features. Figure 2 shows the box plot of ASM and Contrast.
Fig. 2. Box Plots of ASM and CONTRAST for two classes
For example, in figure 2, the ASM1 (ASM of water training data) is non-
overlapping with ASM2 (ASM of urban area training data). So ASM is the feature
which can be shortlisted for the analysis. After analysing 14 features are selected and
shown in table 1.
Table 1. Selected 14 features after using Box-Whisker plot
WATER Urban d
Mean S.D. Mean S.D.
ASM 0.871 0.160 0.340 0.190 2.138
CONTRAST 0.048 0.054 0.373 0.223 1.416
DIFF_ENTROPY 0.072 0.062 0.280 0.084 1.992
DIFF_VARIANCE 0.045 0.044 0.250 0.125 1.547
ENTROPY 0.123 0.032 0.684 0.225 2.468
IDM 0.975 0.027 0.824 0.066 2.118
INFO_MEAS_CORR2 0.465 0.228 0.917 0.075 1.883
SUM_ENTROPY 0.116 0.075 0.563 0.164 2.479
SRE (Short Run Emphasis) 0.892 0.042 0.970 0.011 1.797
LRE (Long Run Emphasis) 1.593 0.313 1.125 0.054 1.473
GLN (Gray Level Non uniformity) 1119 337 262 80 2.474
RLN(Run Length Non Uniformity) 10917 1840 14830 680 1.995
RP(Run Percentage) 0.858 0.056 0.961 0.015 1.777
HRGE 285690 278950 1490376 1001152 1.159
ASM1 ASM2 CON1 CON2
0
0.2
0.4
0.6
0.8
1
Values
438 M. Singh and G. Kaur
Then on the basis of Fisher’s discriminative distance ‘d’, we further reduced the
number of features to 10. The correlation between these features is then calculated and
ultimately we have reduced the feature set to just 5 features. We have analysed that the
selected features like GLN, Sum Entropy, RLN, ASM and Info_meas_corr2 are
having highest discriminative distance and are least correlated. Then, we have calcu-
lated the sensitivity, specificity and accuracy to evaluate the proposed set of features
for classification as tabulated in table 2.
Table 2. Specificity and Sensitivity test result
True
Positive
False
Positive
False
Negative
True
Negative SENSITIVITY SPECIFICITY ACCURACY
9 0 1 10 90% 100% 95%
Thus, the proposed system has a sensitivity of 90% which means the system recog-
nizes 90% of the images containing water. The specificity of 100% means that system
is very specific in identifying water images and always identified Urban area as Urban
area. Finally, the proposed system is 95% accurate. Furthermore PCA is applied to the
feature set of final 5 texture features obtained after using discriminative power dis-
tance and correlation coefficient. The following plots are obtained as shown:
Fig. 3. PCA applied Water images Fig. 4. PCA applied to Urban area images
From the above shown PCA plots, it is clear that all the 5 texture features are dis-
criminating between the two training sets.
4 Conclusion
We have studied 30 images for 20 texture features. We have analysed these features
and found the 5 highly discriminative features by using box plots, discriminative
power distance and Pearson’s coefficient of correlation. The final five features are
• GLN (Gray Level Non-Uniformity) from GLRLM
• Sum of Entropy from GLCM
• ASM (Angular Second Moment) from GLCM
SAR Image Classification Using PCA and Texture Analysis 439
• RLN (Run Length Non-Uniformity) from GLRLM
• Information measure of correlation 2 from GLCM
Finally, we have evaluated the performance of these features on the test images and
found that sensitivity is 90%, specificity is 100% and the accuracy for the classifica-
tion has been 95%. So, we proposed a classification system for SAR images based on
highly discriminative texture features. Furthermore, we have also applied PCA on
these five features and concluded that these features are discriminative.
References
1. Chamundeeswari, V.V., Singh, D., Singh, K.: An Analysis of Texture Measures in PCA-
Based Unsupervised Classification of SAR Images. IEEE Geoscience and Remote Sensing
Letters 6, 214–218 (2009)
2. Unser, H.: Sum and difference histograms for texture classification. IEEE Trans. Pattern
Anal. Mach. Intell. PAMI-8(1), 118–125 (1986)
3. Clausi, D.A.: Comparison and fusion of co-occurrence, Gabor and MRF texture features
for classification of SAR sea ice imagery. Atmos. Oceans 39(4), 183–194 (2001)
4. Espinal, F., Jaweth, B.D., Kubota, T.: Wavelet based fractal signature analysis for auto-
matic target recognition. Opt. Eng. 37(1), 166–174 (1988)
5. Dekker, R.J.: Texture analysis and classification of ERS SAR images for map updating of
urban areas in The Netherlands. IEEE Trans. Geosci. Remote Sens. 1(9), 1950–1958
(2003)
6. Haralick, R.M., Shanmugam, K., Dinstein, I.: Textural Features for Image Classification.
IEEE Transactions on Systems, Man and Cybernetics (3), 610–621 (1973)
7. Rangayyan, R.M.: Biomedical Image Analysis. CRC Press, Washington DC (2005)
8. Loh, H.H., Leu, J.G., Luo, R.C.: The Analysis of Natural Textures Using Run Length Fea-
tures. IEEE Transactions on Industrial Electronics 2, 323–328 (1988)
9. Galloway, M.M.: Texture Analysis Using Gray Level Run Lengths. Computer Graphics
Image Processing 4, 172–179 (1975)
10. Chu, A., Sehgal, C.M., Greenleaf, J.F.: Use of Gray Value Distribution of Run Lengths for
Texture Analysis. Pattern Recognition Letter 11, 415–420 (1990)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 440–444, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Performance of WiMAX/ IEEE 802.16 with Different
Modulation and Coding
Shubhangi R. Chaudhary
Assistant Professor in E&TC dept.
Cummins College of Engineering, for women
Pune, India
shubhangirc@yahoo.com
Abstract. WiMAX (Worldwide Interoperability for Microwave Access) is the
IEEE 802.16 Standards based wireless technology that provides MAN (Metro-
politan Area Network) broadband and IP connectivity. WiMAX is promising
technology which offers high speed voice, video and data services based on
OFDM. It offers both line of sight and non line of sight wireless communica-
tion. In this paper performance of WiMAX with different modulation (BPSK,
QPSK and QAM) and coding is studied on the basis of Bit Error Rate, Signal to
Noise Ratio and error probability.
Keywords: WiMAX, OFDM, AMC, BER, SNR.
1 Introduction
In wireless communication, the demand of all types of services is not only voice and
data but also multimedia services. A aims for the design of more intelligent commu-
nication systems, capable of providing spectrally efficient and flexible data rate ac-
cess.WiMAX (Worldwide Interoperability for Microwave Access) is a new wireless
OFDM-based technology that provides high throughput broadband connection over
long distances based on IEEE802.16 wireless (Metropolitan Area Network) MAN air
interface standard [1]. WiMAX supports a variety of modulation and coding schemes
to adapt and adjust the transmission parameters based on the link quality, improving
the spectrum efficiency of the system, and reaching, in this way, the capacity limits of
the underlying wireless channel [4].
The WiMAX physical layer is based on OFDM. OFDM is the transmission scheme
of choice to enable high-speed data, video, and multimedia communications and is
used by a variety of commercial broadband systems [6].Adaptive modulation (AMC)
can effectively improve the bit error rate (BER) performance on radio channels,
which had suffered from shadowing and multipath fading. Adaptive modulation en-
ables a WiMAX system to optimize the throughput based on the propagation condi-
tions. Using adaptive modulation scheme, WiMAX system can choose the highest
order modulation provided the channel conditions are good. As the signal-to-noise
ratio (SNR) is very good near the base station (BS), so higher order modulation
scheme is used in this area to increase the throughput. However, in areas close to the
Performance of WiMAX/ IEEE 802.16 with Different Modulation and Coding 441
cell boundary, the SNR is normally poor. So, the system steps down to a lower order
modulation scheme to maintain the connection quality and link stability. The sup-
ported modulations are BPSK, QPSK, 16-QAM and 64-QAM [6], [8].
Fig. 1. Scheme for the utilization of AMC
In AMC, not only the modulation order but also the forward error correction (FEC)
schemes are varied by adjusting their code rate to the variations in the communication
channel. An example of utilization of the cited AMC scheme is illustrated in Fig. 1. It
shows that as the range increases, the system steps down to a lower modulation, but as
closer to the base station, higher order modulations can be used for increased
throughput.
2 System Model
The WiMAX system model is as shown in fig. 2.
Fig. 2. WiMAX System Model
2.1 Transmitter
The data is generated from a random source, consists of a series of ones and zeros.
The generated data is passed on to the next stage, either to the FEC block or directly
to the symbol mapping if FEC is not used.
Channel coding
There are various combinations of modulations and code rates available in the
OFDM. Channel coding includes the randomization of data, forward error correction
442 S.R. Chaudhary
(FEC) encoding and interleaving. The generated random bits are coded by a concate-
nated Reed-Solomon (RS) and Convolutional encoder.
Modulation
There are different modulation types available for modulating the data onto the sub-
carriers: BPSK, QPSK, 16QAM, and 64QAM.
IFFT: To convert mapped data, which is assigned to all allocated data subcarriers of
the OFDM symbol, from frequency domain into time domain, the IFFT is used. We
can compute time duration of the IFFT time signal by multiply the number of FFT
bins by the sample period. Zeros are added at the end and beginning of OFDM sym-
bol. These zero carriers are used as guard band to prevent inter channel interference
(ICI).
Cyclic Prefix insertion (CP)
To avoid inter symbol interference (ISI) a cyclic prefix is inserted before each trans-
mitted symbol. That is achieved by copying the last part of an OFDM symbol to the
beginning. WiMAX supports four different duration of cyclic prefix (i.e. assuming is
the ratio of guard time to OFDM symbol time; this ratio is equal to 1/32, 1/6, 1/8 and
1/4). This data is fed to the channel which represents ‘Rayleigh fading channel model’
and also implements multipath as shown in system model.
2.2 Receiver
The WiMAX receiver is the reverse function of WiMAX transmitter. The received
signal is then passed through the serial to parallel converter. Then to next block.
Removal of CP: The first step after the arrival of data is to remove CP as shown in
figure 2. We know that CP has no effect in case of using AWGN channel. It is useful
when the multipath channel is used. If CP larger than the delay multipath ,the ISI is
completely removed.
Fast Fourier Transform (FFT)
To convert received data from time domain to frequency domain, the FFT is used.
Afterward, the zeros, which were added at the end and beginning of OFDM symbol
(guard bands) at the transmitter are removed from the assigned places.
Demodulation
Demodulator converts the waveforms created at the modulation to the original trans-
formed bits. The demodulator is used for decision rules with the goal of making a
decision about which bit "zero" or "one", was sent.
Channel Decoder: The channel decoder consists of Viterbi decoder and RS decoder
the sequence of bits coming from demodulator pass to channel decoder. The channel
decoder tries to recover the original bits.
Performance of WiMAX/ IEEE 802.16 with Different Modulation and Coding 443
3 Results
The simulation result based on the adaptive modulation technique for BER calculation
was observed with MATLAB7. The adaptive modulation techniques used in the Wi-
MAX are BPSK, QPSK, 16-QAM, 64-QAM and 256-QAM respectively. Binary
Phase Shift Keying (BPSK) is more power efficient and needs less bandwidth. On the
other hand 64-Qadrature Amplitude Modulation (64-QAM) has higher bandwidth
with very good output. During all simulations we got, BPSK has the lowest BER and
256-QAM has the highest BER than other modulation techniques.
010 20 30 40 50 60 70
0
0.5
1
1.5
2
2.5
3
Data Points
Amplitude
Transmitted Data "O"
010 20 30 40 50 60 70
0
0.5
1
1.5
2
2.5
3
Data Points
Ampli tude
Received Data "X"
Fig. 3. Transmitted Data Fig. 4. Received Data Fig. 5. OFDM with Adaptive
Modulation Techniques In PURE
AWGN
4 Conclusion
Thus depending on the channel condition modulation and coding can be employed.
For larger range lower modulation ie BPSK is more power efficient and need less
bandwidth, lowest BER but for near Base station higher modulation ie 64QAM is
used which has higher bandwidth and highest BER. While QPSK and 16 QAM are for
middle range.
References
1. Eklund, C.: IEEE Standard 802.16: A Technical overview of wirelessMANTM Air Inter-
face for Broadband Wireless Access. IEEE Communication Mag. 40(6), 98–107 (2002)
2. Ghosh, A., Wolter, D.R., Andrews, J.G., Chen, R.: Broadband Wireless Access with Wi-
Max/8O2.16: Current Performance Benchmarks and Future Potential. IEEE Communication
Mag. 43(2), 129–136 (2005)
3. Arafat, O., Dimyati, K.: Performance Parameter of Mobile WiMAX: A Study on the Physi-
cal Layer of Mobile WiMAX under Different Communication Channels & Modulation
Technique. In: Second International Conference on Computer Engineering and Applications
(ICCEA), pp. 533–537 (2010)
4. Andews, J.G., Ghosh, A., Muhamed, R.: Fundamentals of WiMAX Understanding Broad-
band Wireless Networks. Prentice-Hall, Englewood Cliffs (2007)
444 S.R. Chaudhary
5. Ashraful Islam, M., Zahid Hasan, M.: Performance Evaluation of WiMAX Physical Layer
under Adaptive Modulation Techniques and Communication Channels. International Jour-
nal of Computer Science and Information Security 5(1) (2009)
6. Fazel, K., Kaiser, S.: Multi-Carrier and Spread Spectrum System, 2nd edn. John Wiley and
Sons Ltd., New York (2003)
7. Proakis: Digital Communications, 4th edn.
8. Zerrouki, H.: High Throughput of WiMAX MIMO-OFDM Including Adaptive Modulation
and Coding. International Journal of Computer Science and Information Security 7(1)
(2010)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 445–448, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Novel Stair-Case Replication (SCR) Based Fault
Tolerance for MPI Applications
Sanjay Bansal1, Sanjeev Sharma2, and Ishita Trivedi3
1 Research Scholar, Rajiv Gandhi Prodhyogiki Vishwavidya
Bhopal, India
sanjaybansalrgpv@gmail.com
2 Head, School of Information Technology
Rajiv Gandhi Prodhyogiki Vishwavidya
Bhopal, India
sanjeev@rgtu.net
Abstract. When computational clusters increase in size, their mean time to
failure reduces drastically. We generally use checkpoint to minimize the loss of
computation. Most check pointing techniques, however, require central storage
for storing checkpoints. This results in a bottleneck and severely limits the scal-
ability of check pointing, while also proving to be too expensive for dedicated
check pointing networks and storage systems. We propose a Stair-Case Repli-
cation (SCR) Based MPI check pointing facility. Our reference implementation
is based on LAM/MPI; however, it is directly applicable to any MPI implemen-
tation. We use the staircase method of fault-tolerant MPI with asynchronous
replication, eliminating the need for central or network storage. We evaluate
centralized storage, a Sun-X4500-based solution, an EMC storage area network
(SAN), and the Ibrix commercial parallel file system and show that they are not
scalable, particularly after 64 CPUs. We use the staircase MPI method which
allows the access point in a lower complexity level to the higher complexity
level which improves the efficiency of the previous method.
Keywords: check pointing, Fault tolerance, MPI, SAN.
1 Introduction
Over the last years, e-Learning, and in particular Computer-Supported Collaborative
Learning (CSCL) [1], [2] applications have been evolving accordingly with more and
more demanding pedagogical and technological requirements. The aim is to enable
the collaborative learning experience in open, dynamic, large-scale and heterogeneous
environments [3], [4], [5]. In contemporary application servers, dynamic reconfigura-
tion capability is well addressed to provide necessary flexibility– a component in an
application can be dynamically loaded, unloaded, or upgraded at runtime on platforms
like JEE (Java Enterprise Edition) [6] and CORBA (Common Object Request Broker
Architecture) [7],and platforms for Fractal [8] and OpenCOM [9] component model.
Despite the benefits of dynamic reconfiguration, it should not impair other properties,
such as availability and reliability. Software-implemented fault tolerance (FT)
446 S. Bansal, S. Sharma, and I. Trivedi
mechanisms have been provided in application servers to achieve high availability
and reliability. As a result, the FT mechanism should be reconfigured either [7].
In current practices, replication-based FT is widely used in application servers to
improve availability and reliability of components, such as EJBs (Enterprise Java
Beans) and CORBA components [10]. Moreover, adaptive FT is considered in current
practices, but most of the studies concentrate on changing FT policies in a mechanism
to cope with either varied timing constraints [11], or other non-functional require-
ments [12]. This fine grain adaption is also called parameter adaptation [13], which
changes parameters in a mechanism to modify its behaviors.
Since MPI [14] provides no mechanism to recover from such failures, a single
node failure will halt the execution of the entire computation. Thus, there exists great
interest in the research community for a truly fault-tolerant MPI implementation. Sev-
eral groups have included checkpointing within various MPI implementations.
MVAPICH2 now includes support for kernel-level checkpointing of InfiniBand MPI
processes [15]. Sankaran et al. also describe a kernel-level checkpointing strategy
within LAM/MPI [16], [17]. Fig1 shows fault tolerant arrangement and fault toler-
ance support.
Fig. 1. Fault Tolerant Arrangement
The remaining of this paper is organized as follows. We discuss LAM/MPI in Sec-
tion 2. In Section 3 we discuss about Checkpointing. The Recent Scenario in section
4.In section 5 we discuss about Algorithm. The conclusions and future directions are
given in Section 6. Finally references are given.
2 LAM/MPI
MPI is suitable for parallel machines such as the IBM SP, SGI Origin, etc., but it also
works well in clusters of workstations. Taking advantage of the availability of the
clusters of workstations at Dalhousie, we are interested in using MPI as a single paral-
lel virtual machine with multiple nodes.LAM is a daemon based implementation of
MPI. Initially, the program lamboot spawns LAM daemons based on the list of host
machines provided by the user.
LAM/MPI [16] is a research implementation of the MPI-1.2 standard [14] with
portions of the MPI-2 standard. The most commonly used module, however, is
the TCP module, which provides basic TCP communication between LAM processes.
A modification of this module, CRTCP, provides a bookmark mechanism for check-
pointing libraries to ensure that a message channel is clear.
A Novel Stair-Case Replication (SCR) Based Fault Tolerance for MPI Applications 447
3 Chekpointing
Checkpointing is commonly performed at one of three levels: the kernel-level, the
ser-level, or the application level including language level, hardware level, and virtual
machine level, also exist). In kernel-level checkpointing, the checkpointer is imple-
mented as a kernel module, making checkpointing fairly straightforward. However,
the checkpoint itself is heavily reliant on the operating system (kernel version, proc-
ess IDs, etc.). Upon failure, all processes are rolled back to the most recent check-
point/consistent state. Message logging requires distributed systems to keep track of
interprocess messages in order to bring a checkpoint up-to-date.
4 Recent Scenario
One first step in the characterization of any computer system is the use of benchmark-
ing, which allows for the analysis of the performance behavior of a system when dif-
ferent workloads representing the whole spectrum of possible loads are applied. Other
different works have appeared for modeling a grid. Bratosin et al. in provide a formal
description of grids by using Colored Petri Nets (CPN).
5 Stair-Case Replication (Proposed Method)
First compute the random replica method by the below algorithm which is given by
John Paul.
Algorithm 1. Compute random replica
placements.
Input-The number of replicas (Int r)
Input-The number of nodes
Output-Replica a [0….n-1]
Step1: for all I from 0 to n do
Step 2: Preload node with replica i
Step 3: for all i=0 to n do
Step 4: for j=0 to r do
z=select random node
v=Replicas[i][j] until z!=i
Step 5: for all k such tha k>=0 to r do
Step 6: if replicas[i][k]!==v
Valid replica=1
Step 7: for all k such that 0<=k<r do
Ifreplica[i][k]==v||replicas[i][j]==repl
icas[z][k]
Valid-replica=0
Step 8: Finish
Algorithm 2 Stair-Case (Replica Re-
sult Array) Input: Input the Array
Output: enhanced replica model for
fault tolerance
1. For all I such that 0< I <n do
2. Apply asynchronous replication
3. Eliminate central storage
4. Calculate the index by replica
method
5. for all I such that 0 <=i < n
6. for all J such that 0 <=j <n do
7. apply replication on the lower node
to the higher node
8. Finish
After that apply this algorithm to stair-case replication method on the computed rep-
lica placements to the better efficiency and the performance.
448 S. Bansal, S. Sharma, and I. Trivedi
6 Conclusions and Future Work
We have shown that it is possible to effectively checkpoint MPI applications using
the LAM/MPI implementation with low overhead by using staircase technique to go
to the lower index to the higher index. Our stair-case replication implementation has
proven to be highly effective and resilient to node failures. Because of lower to higher
index concept.
References
[1] Koschmann, T.: Paradigm shifts and instructional technology. CSCL (1996)
[2] Dillenbourg, P.: Introduction. Elsevier Science, Amsterdam (1999)
[3] Pankatrius, V., Vossen, G.: Towards E-Learning Grids. In: IEEE Workshop on Knowl-
edge Grid and Grid Intelligence, Halifax, New Scotia, Canada (2003)
[4] Caballé, S., Xhafa, F., Daradoumis, T.: A service-oriented platform for the enhancement
and effectiveness of the collaborative learning process in distributed environments. In:
Chung, S. (ed.) OTM 2007, Part II. LNCS, vol. 4804, pp. 1280–1287. Springer, Heidel-
berg (2007)
[5] Bahrami, K., Abedi, M., Daemi, B.: AICT 2007, pp. 29–35. IEEE Computer Society, Los
Alamitos (2008)
[6] Wang, Q., Huang, G., Shen, J., Mei, H., Yang, F.: COMPSAC 2003, November 3-6,
pp. 230–235 (2003)
[7] Blair, G.S., Blair, L., Issarny, V., et al.: Proc. of Middleware, pp. 164–184 (2000)
[8] Bruneton, E., Coupaye, T., Leclercq, M., Quema, V., Sterain, J.-B.: An open component
model and its support in java. In: Crnković, I., Stafford, J.A., Schmidt, H.W., Wallnau, K.
(eds.) CBSE 2004. LNCS, vol. 3054, pp. 7–22. Springer, Heidelberg (2004)
[9] http://www.martinfowler.com/articles/injection.html
[10] Narasimhan, P.: Transparent fault tolerance for CORBA (1999)
[11] Kim, K., Lawrence, T.: Adaptive fault tolerance in complex real-time distributed com-
puter applications. Computer Communications 15(4) (May 1992)
[12] Froihofer, L., Goeschka, K.M., Osrael, J.: Middleware support for adaptive dependability. In:
Cerqueira, R., Pasquale, F. (eds.) Middleware 2007. LNCS, vol. 4834, pp. 308–327.
Springer, Heidelberg (2007)
[13] McKinley, P., Sadjadi, S., Kasten, E., Cheng, B.: Composing adaptive software. IEEE
Computer 37(07), 56–64 (2004)
[14] The MPI Forum. MPI: A Message Passing Interface. In: Proc. Ann. Supercomputing
Conf. (SC 1993) (ICPP 2006),pp. 471-478 (2006)
[15] Burns, G., Daoud, R., Vaigl, J.: LAM: An Open Cluster Environment for MPI. In: Proc.
Supercomputing Symp., pp. 379–386 (1994)
[16] Sankaran, S., Squyres, J.M., Barrett, B., Lumsdaine, A., Duell, J., Hargrove, P., Roman,
E. (2005)
[17] Squyres, J.M., Lumsdaine, A.: A Component Architecture for LAM/MPI (2003)
[18] InfiniBand Trade Assoc., InfiniBand (2007),
http://www.infinibandta.org/home
[19] Myricom, Myrinet (2007), http://www.myricom.com
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 449–453, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Evaluating Cloud Platforms- An Application Perspective
Pankaj Deep Kaur and Inderveer Chana
Computer Science and Engineering Department, Thapar University, Patiala, India
{pankajdeep.kaur, inderveer}@thapar.edu
Abstract. Cloud computing has gained tremendous consideration recently. In
the ‘Era of Tera’ when data sizes are continuously escalating and traffic pat-
terns have become unpredictable; Cloud is a viable alternative for enterprises to
serve their consumers with quicker response times. The businesses can rely on
service provider to host their applications and can thus focus on their core com-
petencies. However, a large number of Cloud providers have spurred up with
their offerings. As the number of Cloud Computing players grows, the run time
services provided to the cloud consumers act as key differentiators. In this pa-
per, traditional applications are compared with Cloud applications and metrics
relevant to the Cloud services have been identified. Further, the services of five
major Cloud providers are compared for fundamental differences that exist in
their offerings on the basis of these metrics.
Keywords: Cloud computing, Web services, Utility Computing, Virtualization,
IaaS, PaaS, SaaS.
1 Introduction
Cloud Computing is an emerging approach focused to decouple the delivery of
computing services from the underlying technology. This new technology intends to
provide Everything as a Service (XaaS). The cloud service model is categorized into
three layers eminently Software as a Service (SaaS), Platform as a Service (PaaS) and
Infrastructure as a Service (IaaS) [2][12]. Cloud Consumers (CC) provision computing
capabilities provided by the Cloud Service Provider (CSP) and in turn pay whatever
they use and as per the duration of usage [3]. The resources in the cloud are abstracted
and can be accessed using easy web interfaces. The interfaces reduce complexities by
exposing minimum set of capabilities based on target use cases [4]. Cloud services
shifts businesses cost from Capital Expenditure (CAPEX) to Operational Expenditure
(OPEX) and are often associated with Utility Computing [6].
2 Traditional Applications versus Cloud Applications
Cloud applications share the design objectives of Distributed applications. These
objectives abbreviated as IDEAS refer to the Interoperability, Distributed scale
450 P.D. Kaur and I. Chana
out, Extensibility, Adaptivity and Simplicity aspects of distributed applications [1].
Acting as a hybrid between the desktop applications and traditional web applications,
a variation in the traditional application life cycle phases is required to architect cloud
applications.
2.1 Application Design
Traditional applications are designed to meet a stable demand rather than fluctuating
loads prevalent in cloud environments. Applications should be designed in a way to
use the underlying resources only when needed and achieve instant scalability as per
demand [7].
2.2 Application Deployment
Traditional applications are deployed manually or using automated scripts. Auto-
mated scripts are faster than manual deployment but suitable for smaller workloads
[5]. Image Based deployment can be used for Cloud environment in which software
stack image can be instantly copied onto the target system.
2.3 Application Execution
Application execution usually requires the exposure of jobs to the scheduling system
and is assigned to be executed at a later stage. In contrast, cloud applications are not
exposed to the scheduling system. Application execution consists of requesting an
instantiation of Virtual Machine (VM) which is assigned by the user or the middle-
ware [1].
3 Evaluation Metrics for the Cloud Market Place
A large number of Cloud platforms are available in the global market. The runtime
management services act as a key differentiator for various cloud platforms. In this
section, some metrics are used to evaluate cloud platforms and a comparative study of
the cloud offerings provided by the five major industry giants eminently Amazon
Web Services (AWS) [7], Google App Engine (GAE) [8], Microsoft Azure [9],
Rackspace [10] and GoGrid [11] are summarized in Table 1 and Table 2.
Table 1. Comparative Table distinguishing the Cloud PaaS Industry Giants
Metrics Microsoft Azure Google AppEngine
Service Launched Feb 2010 Beta Version April 2008
Application Environment Offers a .NET based
framework that scales
transparently. Server sizes vary
with number of CPU cores,
memory and local disk space.
Offers a component based
framework. Applications run
within Java or Python runtime
environment. No specification of
server sizes.
Virtualization Technology Modified Hyper-V hypervisor Undisclosed
Data Storage Blob,Tables,Queues,SQL
Azure Google BigTable, GQL,
MemCache
Computation Web Role,Worker Role Undisclosed
Evaluating Cloud Platforms––An Application Perspective 451
Table 1. (continued)
Communication Azure Storage Queues Task Queues
Prog. Lang support .NET, PHP, Java Java, Python
Underlying Operating system 64-bit Windows Server 2008 Linux, Windows, Mac OS
Pricing
Resources Unit
Bandwidth Out GB
Bandwidth In GB
CPU Time
(small instance) Hours
Stored Data GB/Month
Recipients
emailed Recipients
Compute time is charged based
on the amount of time an
instance is processing
transactions.
$0.15
Free during off peak times
through Oct 31, 2010 and later
$0.10.
$0.12
$0.15
N/A
Compute time is charged based on
the amount of time an instance is
deployed.
$0.12
$0.10
$0.10
$0.15
$0.0001
Free Usage Yes, Introductory offers base
level of monthly usage till Oct
31, 2010.
Yes, Free default quota
Service Availability 99.95% external connectivity
of role instances. Currently does not implement
SLA,
Atleast 99.9% availability
proposed.
Data Replication Data replicated thrice
regardless of storage option
used.
Defines data location (primary,
alternate)and Read policy (strong
or eventual consistent)
Security:User Authentication
And Authorization Windows Live ID
AppFabric Access Control
Services
Google Accounts and associated
URL paths
Table 2. Comparative Table distinguishing the Cloud IaaS Industry Giants
Metrics Amazon Web Services Rackspace GoGrid
Service
Launched
2002 June 2009 March 2008
Application
Environment
Delivers empty virtual
machine. Server sizes
vary with the amount of
compute capacity.
Allows choosing
Operating system and
server size based on
physical memory.
Allows choosing
operating system and
server size based on
RAM allocations.
Virtualization Xen Xen Xen
Data Storage Amazon S3, EBS,
Simple DB, RDS
Uses network attached
storage devices
Cloud Storage
Computation Elastic Compute Cloud
Elastic Map Reduce
Rackspace Cloud
Servers
GoGrid Cloud servers
Dedicated servers.
Communication Simple Queuing
Service
Undisclosed Undisclosed
Programming
Language
support
PHP, Java, .NET,
Python, Ruby .NET, Perl, PHP,
Python, Ruby on Rails Java, .NET, Perl, PHP,
Python, Ruby on Rails
452 P.D. Kaur and I. Chana
Table 2. (continued)
Underl
y
in
g
O
p
eratin
g
system
Red Hat Linux,
Windows Server
2003/2008, O
p
en
Solaris, openSUSE,
Fedora, Gentoo
Linux.
64-bit Linux
Distributions or
Windows Server
2008 , Windows
Server 2003
Linux, Microsoft
Windows, CentOS and
Red Hat Enterprise
Pricing
Resource Unit
Bandwidth
Out GB
Bandwidth
In
GB
CPU Time
(small
instance)
Hours
Stored Data GB/Month
Recipients
emailed
Recipients
Prices var
y
with
instance t
yp
es and
re
g
ions. Char
g
es are
calculated from the
time an instance is
launched until it is
terminated.
$0.08(over 150 TB)
~$0.15(first 10 TB)
Free until Nov 1,
2010. After that
$0.10.
$0.08 Linux/Unix
~$0.12 Windows
$0.055(over 5000 TB)
~$0.15(first 50 TB)
N/A
Prices var
y
with the
amount of
p
h
y
sical
memor
y
reserved.
Compute time
calculated based on
de
p
lo
y
ed time and
not utilization.
$0.22
$0.08
$0.06 Linux ~$0.08
Windows
$0.15 (unlimited no.
of files each u
p
to 5
GB)
N/A
Server usage calculated
b
y
RAM hour i.e. total
amount of RAM
de
p
lo
y
ed multi
p
lied b
y
the total number of
hours it has been
deployed.
$0.07 (57 TB)~$0.20
(500 GB)
Free
$0.175
$0.15 after stora
g
e
exceeds 10 GB
N/A
Free Usage Incomin
g
Bandwidth
Free until Nov 1,
2010.
-Free bandwidth
between different
cloud servers
- F5 Hardware Load
balancing
-10 GB Cloud
Storage/month
-Up to 16 IP addresses
Service Availability 99.95% availability,
99.9% uptime
100% availability
100% uptime
100% uptime, 10,000%
service credits for SLA
violations
Data Replication Multi
p
le Availabilit
y
Zones Three data co
p
ies
across logical zones Content deliver
y
Network (CDN)
User Authentication
and Authorization
AWS Account ID
Access Control List
(ACL)
User ID, API access
ke
y
, session
authentication token
GoGrid
p
artner GSI
(GoGrid Server Images)
4 Conclusion
As the Cloud Computing technology is gaining wider acceptance, providers are
offering varied feature sets for its consumers. It can be concluded that AWS is offer-
ing the largest array of cloud services and ranks best in terms of its Global presence
(datacenters in four regions) and pricing. However, it lacks behind Rackspace and
GoGrid in terms of 24/7 support and freebies offered. Also, the SLA of Rackspace
and GoGrid ensure 100% uptime yet it needs to be considered as a business tactic that
trades off the outages with the service credits. In terms of credits, both Rackspace and
GoGrid provide 100% credit for its warranties while Amazon caps its credit at 10%
Evaluating Cloud Platforms––An Application Perspective 453
per period. Furthermore, Rackspace’s clear specification of one hour time to resolve
failures may lure consumers, as compared to other providers, which fail to specify any
such time bound warranties. As PaaS provider’s, both Google and Microsoft are
tough competitors. However, Azure’s flexibility offered in terms of running a data-
base and web server of consumer’s choice gains advantage over Google’s proprietary
solutions. Nevertheless, in terms of scalability, GAE’s autoscale capability outshines
Azure’s need based configurations. For hosting purposes, GAE is the best choice as
users are provided free support until an application becomes large enough where a
developer can easily pay while earning profits. The cloud market being extremely
dynamic, with varied feature sets, requires application needs and customer’s require-
ment to be meticulously considered for choosing the best provider.
References
[1] Jha, S., Katz, D.S., Luckow, A., Merzky, A., Stamou, K.: Cloud Book Chapter. In: Un-
derstanding Scientific Applications for Cloud Environments. John Wiley & Sons, Chich-
ester (2010)
[2] Youseff, L., Butrico, M., Da Silva, D.: Toward a Unified Ontology of Cloud Computing.
In: Grid Computing Environments Workshop (GCE 2008), Austin, Texas, USA, Novem-
ber 2008, pp. 1–10 (2008)
[3] Armbrust, M., et al.: Above the Clouds: A Berkeley View of Cloud Computing.
UCB/EECS-2009-28 (February 10, 2009)
[4] Jha, S., Merzky, A., Fox, G.: Using Clouds to Provide Grids Higher Levels of Abstraction
and Explicit Support for Usage Modes. Concurrency and computation: Practice and Ex-
perience 21(8), 1087–1108 (2009)
[5] Server provisioning Methods holding back cloud computing initiatives, Ed Scannell,
http://searchcio.techtarget.com/news/article/0,289142,sid182
_gci1517504,00.html
[6] Kaur, P.D., Chana, I.: Unfolding the Distributed Computing Paradigms. In: International
Conference on Advances in Computer Engineering, ACE 2010, pp. 339–342 (2010)
[7] Amazon Web Services, http://aws.amazon.com/what-is-aws/
[8] Google App Engine-Google Code, http://code.google.com/appengine/
[9] Windows Azure Platform-, http://www.microsoft.com/windowsazure/
[10] Rackspace Hosting, http://www.rackspace.com/index.php
[11] GoGrid, http://www.gogrid.com/cloud-hosting/
[12] Pastaki Rad, M., et al.: A Survey of Cloud Platforms and Their Future. In: Computational
Science and Its Applications (ICCSA 2009), pp. 788–796 (2009)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 454–460, 2011.
© Springer-Verlag Berlin Heidelberg 2011
An Intelligent Agent Based Temporal Action Status
Access Control Model for XML Information
Management
N. Jaisankar and A. Kannan
Department of Computer Science and Engineering Anna University, Chennai-25
jaisasi_win@yahoo.com, kannan@annauniv.edu
Abstract. The enormous amount of XML data growing on the Web raises
several security issues that current XML standards do not address. The most
important security issue in such a distributed environment is the lack of effi-
cient Access Control and authorization for XML data currently. In this paper,
we propose a generalized Role Based Access Control (RBAC) model called An
Intelligent Agent Based Temporal Action Status Access Control (IATASAC)
model which uses temporal and action status constraints for efficient access
control. This model addresses certain shortcomings of RBAC model by the use
of semi structured data; hence it is more suitable for distributed web environ-
ment. Since the proposed model has been developed for XML data and needs
smart decision, it heavily relies on XPath and intelligent agents for effective
querying and answering. The Experimental results show that the proposed
model performs well.
Keywords: Intelligent agent, Action status, XML, Ascribed status, Temporal
constraints, Role Based Access Control.
1 Introduction
As a large quantity of information is presented in XML format on the web, there are
increasing demands for XML security which is not addressed by the current XML
standards. Until now, research on XML security has focused only on the security of
data communication, using either digital signatures or encryption technologies. How-
ever, XML security involves not only with communication security but also with
managerial security. In particular, access control and authorization issues are crucial
in this huge distributed environment but have seldom been addressed. Moreover, most
works in the literature did not give importance to the temporal aspects. However, the
most important issue in access control and authorization for XML data is how to con-
trol user accessibility to each datum based on temporal events and roles where the
traditional methods are inefficient. Hence, it is necessary to identify suitable tech-
niques to address temporal access control issues.
In this paper to increase the access to resources, we propose a generalized
RBAC model. A key feature of this model is that decisions on requests from intelli-
gent agents to access resources are determined by considering the intelligent agent’s
An Intelligent Agent Based Temporal Action Status Access Control Model 455
ascribed status, action status and temporal constraints. Additional conditions of rele-
vance are also considered in answering the access request. An agent’s ascribed status
together with action status gives a measure of the agent’s overall status level. The
agent’s status level is used as the basis for determining authorized actions and thus is
used in rendering a decision on the agent’s access request varies from time to time.
An ascribed status may be associated with, for instance: a particular role, a classifica-
tion of trustworthiness, membership of an organization, time etc. An action status is
determined from a history of the deliberative actions performed by the agent. More-
over, it supports various types of temporal constraints called instant as well as interval
constraints on the enabling/disabling of roles, user-role assignment and permission
assignment. Using temporal constraints and role effective decisions are made even
when incomplete information is present. The next section provides a survey of related
works.
2 Related Work
There are many works in the literature that deal with access control and data security.
Sandhu et al. [1] presented a RBAC model based on users, roles and operations. Re-
cent interest in RBAC has been motivated by the use of roles at the application level
to control access to application data. Hao He et al. [2] proposed a RBAC model for
XML information management the access control scheme is represented in XML
itself, with XPath to specify the linkage between the access information and actual
data. Sriram Mohan et.al. [3] Proposed an infrastructure for access control on XML
document by specifying access constraints in the form of virtual security and enforc-
ing the access constraints via query rewrite. Ninghui Li et.al. [4] Proposed the analy-
sis techniques to maintain desirable security admin previlages, more specifically the
authors defined a family of security analysis problems in RBAC model. Steve Barker
[5] has introduced a generalized RBAC model called Action Status Access Control
(ASAC) model based on the key aspect of autonomous changing of access control
policies in response to events that involve agent action and notion of status. Steve
Barker [6] described implementation of ASAC model and performance measures.
Fenghua Li [7] compared the action based access control model with the other exist-
ing access control models and finally concluded that this model is best for web ser-
vices. Elisa Bertino et.al. [8] proposed a Temporal-RBAC (TRBAC) model, a tempo-
ral extension of the RBAC the main feature of this model is it support for periodic
enabling /disabling of roles and actions expressed by role triggers. James B.D Joshi
et.al. [9] Proposed a generalized TRBAC model and specified various temporal con-
straints to use in user-role and role-permission assignment. Comparing with existing
work proposed in this paper is different and new because it uses temporal and status
level constraints for efficient access control using intelligent agents.
3 Preliminaries
Before formal description of proposed model, we described the basics of status level
of the user agent and the use of temporal constraints as follows.
456 N. Jaisankar and A. Kannan
3.1 Status Level
Here, the term authorization is defined as in [6],
Authorized (A, U, O, S, L) → Check (member ((A, level (U, L), privileges (O, S)))
Check (true) → grant
Check (false) → deny
where U is a user, A is an action, O is an object, S is site, L is history of events,
the operator member is the standard membership test operator, the function level
computes the status level of the user and the function privileges returns a list of pairs
(actions, status level of users allowed to perform the action) for a given object at a
given site.
A user agent’s status level is determined from the user agent’s ascribed and action
status. An ascribed status is a status that is associated with a particular role, a
categorization of agents and agent’s actions status relates to a status an agent may
achieve by doing some action. This idea has been described in Table 1 for a university
database consisting of staff and students considered as an example in this work.
Table 1. Status Level Description
Ascribed
Status Actions performed by agent Status Level
PG Student Doing project and publishing papers Teaching Assistant(TA)
PhD Scholar
Papers published in International
journals with High impact factor and
received Best Paper Award
Young Scientist/
Asst. Professor
Professor Completed Best National Projects Principal Scientific
Advisor/Chairman
In table 1, the status of TA may be used as basis for determining what actions the
student with TA status can and cannot do. He may subject to handle the classes. Simi-
larly in the proposed model the ascribed status and action status of user agent ‘u’, that
requests to access a resource ‘r’ , is used to determine what action ‘u’ can do and
cannot do.
3.2 Temporal Constraints
The proposed model permits to specify various temporal constraints called periodicity
and duration constraints as in [9]. In this model, role assumes one of the three states:
enable, disable and active. The temporal constraints periodicity is used to specify the
exact intervals during which a role can be enabled or disabled and duration is used to
specify duration for which role enabling or assignment is valid.
An Intelligent Agent Based Temporal Action Status Access Control Model 457
4 An Intelligent Agent Based Temporal Action Status Access
Control Model
The different component of the proposed model described as follow
4.1 User Agent’s Session
It receives inputs from a user agent and sends request to the Temporal Access Control
manager (TACM) on behalf of the user agent. It is also responsible for sending re-
sponse received from TACM to the user agent.
4.2 XML Repository
XML repository is a key component of the model and it loads and stores large amount
of XML files and access control files. The XML format of access control files should
have the list of users with the same status level. The status level of the user is
achieved by considering both ascribed status and action status which is determined
from the history of deliberate actions performed by the user as in table 1. Access files
also include history of actions performed by the user, operation tree, role tree and
class membership attributes.
4.3 Temporal Access Control Manager
TACM can be regarded as an application of the XML database.TACM intercepts all
user requests and responds for each user session by querying the XML repository and
checks permission for the user according to the role of a user based on status level and
given temporal constraints. When a request is received by TACM it checks temporal
constraints and status level of the user first and then obtains an access set Sa by com-
bining the operation and target. Now it queries the XML database and obtains the list
of role membership M. Finally it obtains global access set of the user Ga from M. It
also obtains T = Sa U Ga after checking validity of T. The TACM executes T and
returns the result to the user session.
4.4 Administrative Manager
As it interacts directly with the XML database, it can modify the access control file.
The main duty of System admin is as follows. Modification of user roles, temporal
constraints and history of actions of user’s membership attributes and any association
between users and roles, or roles and XML objects. These duties include creating new
roles, granting access permission to roles based on temporal constraints, removing
access permission from roles, assigning new users to roles based on temporal con-
straints updating deliberative actions of user, assigning status level for the user’s
agent, deleting users who have no membership and deleting roles, a role which does
not have children or user.
458 N. Jaisankar and A. Kannan
Fig. 1. System Architecture
Moreover, no direct interaction between user agent’s session and XML database al-
lowed at all time. All interactions must be performed through the TACM. ACM as a
self evolving system admin described in the above for various admin tasks. Some
tasks like the total role number and role relationship and also updating the action
history of users and assigning status level to users based on referring history of ac-
tions performed by the user cannot be changed without a system admin.
5 IATASAC in XML Format
The controlling information for the proposed model is stored in a configuration file in
XML format. When the proposed model starts, the file is read and accordingly, user,
role, and XML object associations are established.
We use the simple Extended Backus-Naur Form (EBNF) which is used for the
XML specifications to design configuration file.
symbol ::= expression .
The configuration file consists of description role-tree, users, SSD, and DSD:
IATASAC -xml ::= description* role-tree user SSD* DSD where descriptions are
Optional.
role-tree ::= role
The role hierarchy is represented by a role tree which has a role as root node. In the
University example, University People is the root node.
role::=role-id nun-limit? role* Temporal constraints acx-f unction* admin-f n *
create? (private acc-fn admin-fn* ownership-link* exception-link*) create?
The role has a unique role id, num_limit. a limitation of memberships it also has col-
lection of job functions which may including XML access functions and administra-
tive functions. All functions after the keyword private are not inheritable. Since own-
ership is unique, both ownership link and exception link are private.
acc-function ::= acc-operation* XMLPointer
acc-operation ::= readlwritelcreateldelete/update
An access function consists of access operations allowed and an XMLPointer pointing
to the XML node object.
admin-func:tion ::= |AssignRoleDepriveRole|AddAccess|RmAccess |MvOwner
users ::= user+, user ::= user-id user-info* p* RolePointer*
User Agent
Temporal Access
Control Manager XML
Repository
Admin Manager
User Agent’s Session
An Intelligent Agent Based Temporal Action Status Access Control Model 459
A user must have a unique user_id and may contain a short description of the user
with his status level. A user must have a password. If a user has a Role Pointer point-
ing to a role, then the user has the membership of that role.
The XML format of Indian University role hierarchy configuration file presented
in table 2.
Table 2. TASAC Model’s Configuration File in XML format
<?xml version=' 1.0' ?>
<!-- XML access control -->
<ATASAC_xml>
<users>
<user id="jai"pwd=”mat"></user>
<user id="raj"pwd="123"></user>
<user status=“TAforPG”></users>
<role_tree>
<role id="Uviversity People" >
<role id="Student" > </role>
<role id="UG" > </role>
<role id="PG> </role>
<role id="Ph.D >
<role id="TA" > </role></role>
<role id ="Staff">
<role id="admin”>
<role id="PA”>
</role> </role>
<role id ="Manager">
<role id="Academic">
<role id="Prof.”>
<role id ="Dean ">
<role id ="VC">
</role> </role>
<role id ="Lect.">
</role> </role>
</role></role></role_tree>
</ATASAC_xml >
For example, if Access(student, read, attendance) is true then the role student can
read attendance, which may be owned by, say, the role professor. Since an XML file
forms a tree, if all child nodes share the same access permissions, access control in-
formation can be stored only at their parent node.
6 Results and Discussion
The Fig.2 shows the number of access violations prevented by intelligent temporal
access status access control model is higher than the traditional TRBAC and RBAC
and Fig.3 shows that the document retrieval time while performing the proposed sys-
tem is considerably less while comparing other traditional access control models.
Fig. 2. Analysis of Access Violation
Prevented Fig. 3. Performance Analysis
460 N. Jaisankar and A. Kannan
7 Conclusion
In this paper, An Intelligent Agent based Temporal Action Status Access Control
Model for XML information management has been proposed. The Access Control
(AC) Scheme is represented in terms of XML format, with XPath to specify the link-
age between the access information the ease of administration and control over the
data and separation of duty. The experimental results show that the proposed model
works well when compared with traditional TRBAC and RBAC. Further work in this
direction is to improve the model with spatio-temporal constraints for providing im-
proved security.
References
1. Sandhu, R.S., Coyne, E.J., Feinstein, H.L., Youman, C.E.: Role-Based Access Control
Models. IEEE Computer 29(2), 38–47 (1996)
2. He, H., Wong, R.K.: A role-based access control model for XML repositories. In: proceed-
ing of the first international conferance on web information system engineering, vol. 1, pp.
138–145. IEEE explorer, Los Alamitos (2000)
3. Mohan, S., Sengupta, A., Wu, Y.: A Framework for Access Control for XML. ACM Trans-
actions on System and Information Security 5, 1–38 (2006)
4. Li, N., Mahesh, V., Tripunitara: Security Analysis in Role-Based Access Control. ACM
Transactions on Information and System Security 9(4), 139–420 (2006)
5. Barker, S.: Action Status Access Control. In: Proceedings of the 12th ACM symposium on
Access control models and technologies (SAGMAT), pp. 195–204 (2007)
6. Barker, S.: Access control by Action Control. In: Proceedings of the 13th ACM symposium
on Access control models and technologies (SAGMAT), pp. 143–152 (2008)
7. Li, F., Wang, W., Jianfengna, Su, H.: Action-Based Access Control for Web Services. Jour-
nal of Information Assurance and Security 5, 162–170 (2010)
8. Bertino, E., Bonatti, P.A., Ferrari, E.: TRBAC: A Temporal Role Based Access Control
Model. ACM Transactions on Information and System Security 4(3), 191–223 (2001)
9. Joshi, J.B.D., Bertino, E., Latif, U.: A Generalized Temporal Role Based Access Control
Model. IEEE Transactions on Knowledge and Data Engineering 17(1), 4–23 (2005)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 461–467, 2011.
© Springer-Verlag Berlin Heidelberg 2011
An Economic Auction-Based Mechanism for
Multi-service Overlay Multicast Networks
Mohammad Hossein Rezvani and Morteza Analoui
School of Computer Engineering, Iran University of Science and Technology (IUST)
16846-13114, Hengam Street, Resalat Square, Narmak, Tehran, Iran
{rezvani, analoui}@iust.ac.ir
Abstract. Recently, strategic behavior modeling has attracted much attention of
the researchers focusing on designing protocols in the area of social networks.
The motivation lies in the fact that the overlay peers of the social networks are
selfish in nature and they typically belong to different administrative domains.
In this paper, we model the strategic behavior of the selfish peers by leveraging
the rich theory of mechanism design using the concept of economic auctions.
By considering the bandwidth of the service offered by the origin server as the
commodity, we design dynamic auctions in which downstream peers submit
their value of bids for each commodity at the upstream peers.
Keywords: Social Networks, Multicasting, Resource Allocation, Strategic Be-
havior, Mechanism Design, Auction Games.
1 Introduction
There exists a significant body of researches towards "social networks" in the litera-
ture. Although online video streaming existed long before social-network-assisted
sites such as YouTube, the establishment of social networks has become an equally or
even more important factor toward their success. On the other hand, "overlay multi-
casting" solution has recently been accepted as a paradigm shift in order to dissemi-
nate the digital real-time contents such as teleconferencing, IPTV, and online video
broadcasting in large peer-to-peer (P2P) networks.
With respect to this fact that the selfish behavior of the overlay peers is inevitable
in real P2P networks, it turns out that the most important question towards designing
the overlay multicast protocols is the following; "How should the selfishness of the
peers is exploited, so that the aggregate throughput of the overlay network still is
maximized?" There have recently been a significant body of research towards design-
ing self-organizing overlay multicast networks by exploiting the inherent selfishness
of the peers. These works can be categorized into two significant strands: "strategic
behavior modeling approaches" such as [1, 2, 3] and "non-strategic behavior model-
ing approaches" such as [4, 5]. In the former, each peer is treated as a potential game
player that seeks to maximize its utility regard to the actions that the other peers do,
whereas in the latter each peer seeks to maximize its utility without taking into ac-
count the actions of the other peers. We present a revenue-maximizing monopoly auc-
tion framework by leveraging the mathematical tools from the theory of "mechanism
462 M.H. Rezvani and M. Analoui
design" of microeconomics. In this framework, each offered service in a typical social
network, plays the role of the commodity in the overlay network economy. The buy-
ers in this economy are all the peers either who relay the services to their downstream
peers or who are leaf nodes in the overlay multicast tree. Also, the sellers are either
the origin servers or the peers who forward the media services to the other peers of the
social network. The remainder of this paper is organized as follows: We discuss the
related works in Section 2. Section 3 is devoted to formal description of the proposed
auction mechanism for the overlay multicast network as well as its associated theo-
ries. Section 4 specifies the performance evaluation of the proposed mechanism.
Finally, we conclude in Section 5.
2 Related Work
There are simultaneous works investigating social networks in popular Web 2.0 sites
[6, 7]. While YouTube is also one of the targeted sites in their studies, exploring the
social network for accelerating content distribution has yet to be addressed. To ad-
dress the selfishness of the peers based on the non-strategic behavior modeling, much
of the literature has applied pricing approaches [4, 5]. A few other works have also
been proposed to allocate the bandwidth based on strategic auctions [1, 2, 3]. The
authors of [3] have presented an auction-based model to improve the performance of
BitTorrent. A key difference between our designed auction and that of [3] is that the
bids in our auction are issued based on the price, whereas in [3], the bids are submit-
ted based on bandwidth. The major difference between our work and the work of
[1, 2] lies in the fact that their approach is in actual a bargaining approach rather than
a monopoly auction in the sense that the mechanism design theory of microeconomics
provides.
3 The Proposed Overlay Auction Mechanism
We consider an overlay network consisting of Vend hosts denoted as
{1, 2, . .., }V=V. Let us suppose that the overlay network consists of N media ser-
vices, denoted as }...,,2,1{ N=N. So, there are Norigin servers among V hosts
)( VN <, each serve a distinct type of media service. Suppose the network is shared
by a set of Nmulticast groups. Any multicast group (multicast session) consists of a
media server, a set of receivers, and a set of links which the multicast group uses. Let
us suppose that the overlay network consists of
L
physical links, denoted as
}...,,2,1{ L=L. The capacity of each link, that is the bandwidth of each physical link
L∈l is denoted as l
c. All the nodes, except the origin servers and leaf nodes, for-
ward the multicast stream via unicast in a peer-to-peer fashion. Fig. 1 shows an over-
lay network consisting of two multicast groups. In this example, we can represent the
set },{ 21 ss=S in which 1
s (node 0) indicates one group and 2
s (node 3) indicates
the other group. Here, the solid lines indicate one group and dashed lines indicate the
second group. Also, the physical network consists of eight links ( 8L=) and two
An Economic Auction-Based Mechanism for Multi-service Overlay Multicast Networks 463
routers. Each multicast session
N
∈nconsists of some unicast end-to-end flows,
denoted as the set n
F:
{| , : 1}
nn n
ij ij
fij=∃∈ =ΜFV
. (1)
Where n
Μ denotes “adjacency matrix” of the multicast group n. Each flow n
ij
f of
the multicast group n passes a subset of physical links, denoted as
()
n
ij
f⊆LL
. (2)
For each link l, we have
)}(|{)( n
ij
nn
ij
nflfl LFF ∈∈= . (3)
Where ()
nlF is the set of the flows belonging to the multicast group n and passing
through the link l. Each flow nn
ij
f∈F in the multicast group n has a rate n
ij
x. We
show the set of all downstream nodes of each overlay node i in the multicast group
n by ),( niChd . Also, the set )(iBuy specifies all the multicast groups in the overlay
network from which the node i receives (buys) the services. Similarly, )(iSell speci-
fies all the multicast groups in the overlay network for which the node i provides
(sells) the services.
Fig. 1. Overlay network consisting of two multicast groups
There are a number of standard auction forms that the seller might use to sell the
commodity: first-price, second-price, Dutch, English. Our designed mechanism,
named overlay monopoly auction mechanism (OMAM), lends itself to explain the
equivalence of revenue in all four auction forms. Let us suppose OMAM in n-th mul-
ticast group in which the monopolist seller s wants to sell the bandwidth of service
n to one of ,sn
K bidders. OMAM is a collection of ,sn
K probability assignment
functions, ,, ,
(, )
in in in
pr v v −, ( , )iChdsn∈,n∈N and ns
K, cost functions
,, ,
(, )
in in in
cvv
−. For each bidder i and every vector of values ,,
(, )
in in
vv
−,
,, ,
(, )[0,1]
in in in
pr v v −∈ denotes the probability that bidder i receives the bandwidth
of service n (in actually the commodity in multicast group n) and
,, ,
(, )
in in in
cvv
−∈\ denotes the payment that bidder i must make to seller s. Also,
464 M.H. Rezvani and M. Analoui
note that ni,− denotes all other bidders apart from bidder i in multicast group n.
Consequently, the sum of the probabilities ,, ,
(,)
(, )
in in in
iChdsn
pr v v −
∈
∑
, must never
exceeds unity. On the other hand, we allow this sum to fall short of unity because we
want to allow the seller to keep the commodity. OMAM works as follows. Because
the seller does not know the bidders' values, it asks them to report the values to it si-
multaneously. It then takes those reports ,in
r, which need not be truthful, and assigns
the commodity to one of the bidders according to the probabilities ,, ,
(, )
in in in
pr v v −,
keeping the commodity with the residual probability, and secures the payment
,, ,
(, )
in in in
cvv
− from each bidder ),( nsChdi∈. It is assumed that the entire direct
selling mechanism, i.e., the probability assignment functions and the cost functions,
are public information, and that seller s must carry out the terms of the mechanism
given the vector of the reported values. Clearly, the seller's revenue will depend on
the reports submitted by the bidders. Now, the main question is that: "will the bidders
be induced to report truthfully?"
Definition 1 (Incentive-Compatible OMAM). OMAM is incentive compatible (IC) if
it is an equilibrium for the bidders to report their values truthfully.
Understanding IC OMAM will not only be the key to understanding the connection
among the four standard auctions, but also it will be central to understanding revenue-
maximizing auctions as well. As it is the case in microeconomics, beginning with the
equilibrium of any of the four standard auctions, we can similarly construct our IC
OMAM that yields the same ex-post assignment of the commodity to the bidders and
the same ex-post payments by them. Interested readers can refer to chapter 9 of [8] to
find an in-depth discussion on this topic as well as the associated theories.
By incentive compatibility, each bidder must find it optimal to report its true value
that all other bidders do so. We now proceed to state the necessary conditions for op-
timal OMAM in such a way that leads to maximization of the revenue for seller s as
well. At first, let us define the concept of individual rationality. In order for OMAM
to be optimal for the seller, the so called property must hold.
Definition 2 (Individual Rationality). In microeconomics, a mechanism is said indi-
vidually rational (IR), if it yields each bidder, regardless of its value, a non-negative
expected payoff in the truth-telling equilibrium.
Clearly, if the expected payoff of a bidder is negative, it will simply not participate in
the selling mechanism. The following theorem completely characterizes the optimal
OMAM:
Theorem 1 (Optimal OMAM). In IC IR OMAM, suppose that the private value of
each bidder i in multicast group n is drawn from the continuous positive density
function ,in
f satisfying (4) as follows
,,
,,
,,
1()
is strictly increasing in
()
in in
in in
in in
Fv
vv
fv
−
−. (4)
Then the probability assignments and the cost functions defined in (5) and (6) yield
the monopolist seller s the largest possible expected payoff.
An Economic Auction-Based Mechanism for Multi-service Overlay Multicast Networks 465
,,
,,
,,
*,, , ,
,, ,
1()
1()
1, if max ( 0, )
() ()
(, )
0, otherwise
jn jn
in in
in jn
in in jn jn
in in in
Fv
Fv
vv
fv fv
pr v v −
−
−
−>−
⎧
⎪
=⎨
⎪
⎩
(5)
,
** *
,, , ,, ,, , ,
0
(, ) (, ) (, )
in
in in in in in in in in in
v
cvv prvv v prwv dw
−− −
=−
∫. (6)
4 Performance Evaluation
In order to handle the operations of the OMAM, we consider a dedicated server,
named Overlay Market Control Server (OMCS). The OMCS bears the characteristics
of "rendezvous point" in the former well-known research projects such as [5, 9]. The
OMCS contains the information of all markets including the free uploading and
downloading capacities of the peers, the rates of the services that are allocated to each
peer, the distribution of the values concerning to each service, the information of each
physical link, and so on. Due to space limitation, we do not mention the details of the
algorithms here. We use BRITE topology generator [10] to set up our experimental
network. Each overlay peer is an end host attached to a single router. The backbone
includes 512 routers and 1024 backbone edges (physical links). The bandwidths of all
physical links have Heavy-tailed distributions in the interval [10 Mbps, 100 Mbps].
The propagation delay of each individual underlying link is uniformly distributed in
the interval [1 ms, 2 ms]. The overlay peers are randomly connected to backbone
routers through access links, whose capacities are exponentially distributed with an
average of 15 Mbps. We also assume 2
ii
CD CU=× , that is downloading capacity of
each access link is two times bigger than its uploading capacity. The maximum toler-
able delay and the maximum tolerable loss rate of each flow are 1 Second and 5%
respectively. The maximum allowed bandwidths of each service, namely n
B, is 3
Mbps. The peers join the network following a Poisson process. The inter-arrival times
follow an exponential distribution with an expected length of 1 second. Upon arrival,
each peer randomly selects some services; then the peer stays in the network for a
certain period of time, following an exponential lifetime distribution with an expected
length of 30 minutes.
Figure 2 shows the average throughput of the peers (the average social welfare) re-
sulting from the OMAM in comparison with the case in which no priced-based
mechanism is used. Also, for the sake of completeness, we have compared the
OMAM with the average upper bound throughput. By "average upper bound
throughput", we mean the average upload capacity of non-leaf peers in all multicast
trees. Clearly, the aggregate demands of the overlay peers cannot exceed the sum of
the uploading capacities of the non-leaf peers of the multicast trees. So, we can gain
further insights into the OMAM by evaluating it with the average upper bound
throughput as the best-case metric. It is evident from Fig. 2 that the resultant average
social welfare of the OMAM is so much better than the case in which no price-based
mechanism is used. The reason of decreasing the average throughput is the data con-
straints and the network constraints of the multicast solution. According to data con-
straint, a peer cannot forward the stream to its downstream peers at a rate higher than
466 M.H. Rezvani and M. Analoui
its own receiving rate. The network constraint arises mainly due to inadequate
downloading or uploading capacities of the peers. As the number of peers increases,
the tree becomes deeper. Thus, the degradation of throughput in lower levels of the
tree results in degradation of the average throughput.
Fig. 2. Average throughput per user (average social welfare)
Fig. 3. Percentage of welfare improvements resulting from OMAM
Figure 3 illustrates the percentage of welfare improvements by the OMAM for dif-
ferent peers' population sizes. To this end, for each population of peers, we have
logged the perceived utility of each peer for both cases of using the OMAM and the
non-priced mechanism and then have compared these two values with each other.
Next, we have normalized the number of improved peers by the total number of peers
An Economic Auction-Based Mechanism for Multi-service Overlay Multicast Networks 467
and have represented the result in the form of percent of improvements. It is clear
from the figure that using the OMAM enhances the perceived quality of the services
in the multicast sessions.
5 Conclusion
In this paper, we provided a social-network-assisted framework for multi-service
multi-rate overlay multicasting based on mechanism design theory of microeconomics
by taking into account the inherent selfishness of the peers. The design of the strategic
behavior model has been done by leveraging the concept of auction games. We
proved the convergence of the auction-based strategies. Also, we showed that the
mechanism is incentive compatible and individually rational and also leads to per-
formance improvements in terms of the social welfare.
References
1. Wu, C., Li, B.: Strategies of Conflict in Coexisting Streaming Overlays. In: INFOCOM,
pp. 481–489 (2007)
2. Wu, C., Li, B., Li, Z.: Dynamic Bandwidth Auctions in Multioverlay P2P Streaming with
Network Coding. IEEE Trans. Parallel Distrib. Syst. 19(6), 806–820 (2008)
3. Levin, D., LaCurts, K., Spring, N., Bhattacharjee, B.: BitTorrent is an auction: analyzing
and improving BitTorrent’s incentives. In: SIGCOMM, pp. 243–254 (2008)
4. Analoui, M., Rezvani, M.H.: An Economic Case for End System Multicast. In: Berre, A.J.,
Gómez-Pérez, A., Tutschku, K., Fensel, D. (eds.) FIS 2010. LNCS, vol. 6369, pp. 40–48.
Springer, Heidelberg (2010)
5. Analoui, M., Rezvani, M.H.: Microeconomics-based Resource Allocation in Overlay Net-
works by Using Non-strategic Behavior Modeling. Elsevier J. Commun. Nonlinear Sci.
Numer. Simulat. 16(1), 493–508 (2011)
6. Cha, M., Kwak, H., Rodriguez, P., Ahn, Y.-Y., Moon, S.: I Tube, You Tube, Everybody
Tubes: Analyzing the World’s Largest User Generated Content Video System. In: Proc. of
ACM IMC (2007)
7. Mislove, A., Marcon, M., Gummadi, K., Dreschel, P., Bhattacharjee, B.: Measurement and
Analysis of Online Social Networks. In: Proc. of ACM IMC (2007)
8. Jehle, G.A., Reny, P.J.: Advanced Microeconomic Theory. Addison-Wesley, Reading
(2001)
9. Pendarakis, D., Shi, S.Y., Verma, D., Waldvogel, M.: ALMI: an application layer multi-
cast. In: 3rd USENIX Symp. on Internet Technologies and Systems (2001)
10. Medina, A., Lakhina, A., Matta, I., Byers, J.: BRITE: An Approach to Universal Topology
Generation. In: Proc. IEEE Int’l Symp. Modeling, Analysis and Simulation of Computer
and Telecomm. Systems, MASCOTS (2001)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 468–472, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Server Virtualization: To Optimizing Messaging Services
by Configuring Front-End and Back-End Topology Using
Exchange Server in Virtual Environments
R. Anand and T. Deenadayalan
Computer Science And Engineering,
Dhaanish Ahmed College Of Engineering,
Anna University, Chennai
Tamil Nadu
nowhereanand@yahoo.com, deenadayalan1982@gmail.com
Abstract. Microsoft Exchange Server supports the deployment of Exchange in
a manner that distributes server tasks among front-end and back-end servers.
This work includes messaging services such as “E-Mail”, “DiscussionForum”,
“SearchEngine”, “NewsLetters”, “GuestBook” and “ClassifiedAdverts”.
The E-Mail service that embeds “Exchange Server” as Back-End acts as the
primary messaging service. Microsoft Exchange Server supports the deploy-
ment of Information exchange that distributes server’s tasks among Front-End
and Back-End servers. Front-End server accepts requests from clients and prox-
ies them to the appropriate Back-End server for processing, to optimize the
Front-End and Back-End servers using virtualization technology. The Author-
ized Users are allowed to utilize the messaging services and special authoriza-
tion is made when the user wishes to join the E-Mail service. After receiving a
request, the front-end server uses LDAP (Lightweight Directory Access Proto-
col) to query the Windows 2003 server Active-Directory-service and determine
which back-end server holds the requested resource. This topology reduces
complexity.
Keywords: Server Virtualization, Virtualization, Exchange Server Virtualiza-
tion.
1 Introduction
Microsoft Exchange 2003 Server supports the deployment of Exchange in a manner
that distributes server tasks among front-end and back-end servers. A front-end server
accepts requests from clients and proxies them to the appropriate back-end server for
processing. The general functionality of the front-end server is to proxy requests to
the correct back-end servers on behalf of the client computers; the exact functionality
of the front-end server depends on the protocol and the action being performed.
After receiving a request, the front-end server uses Lightweight Directory Access
Protocol (LDAP) to query the Microsoft Windows 2003 server Active Directory ser-
vice and determine which back-end server holds the requested resource. The back-end
Server Virtualization: To Optimizing Messaging Services 469
server then sends the results of the logon operation back to the front-end server, which
returns the results of the operation back to the client.The front-end and back-end
server topology is recommended for multiple-server organizations that use Microsoft
Outlook Web Access (HTTP), POP, or IMAP and for organizations that want to pro-
vide HTTP, POP, or IMAP access to their users over the Internet.
The proposed network will allow us to centralize many, if not all. This reduction of
servers, combined with Exchange 2003 integration with Active Directory, will
provide a messaging environment that is significantly easier to manage. By putting
Exchange 2003 on the servers, we can expect a significantly more reliable operating
system due to the inherent reliability improvements incorporated within Exchange
Server 2003.
The Existing system usually used IIS (Internet Information Services). This IIS
stores the configuration information in the “metabase”, whereas Exchange stores
configuration information in Active Directory. The metabase is a local configuration
database shared by the protocols that IIS supports. In Exchange 2000 SP1 and earlier
versions, DS Access used RPCs to connect to directory servers and discover the to-
pology. In Exchange SP2 and later versions (Exchange 2003), DSAccess uses LDAP
for most of the operations.
This research work has a vision to provide its user community with the best tools
in a reliable, manageable, and secure environment. This provides a competitive ad-
vantage both internally and externally. In this earlier research work virtualize the
server and also virtualize the Active directory tools.In this research work virtualize
the server ,active directory and mail server and also optimize the topology of the
network. Both the topology can be optimized using Xen Hypervisor. In VMware to
optimize the topology easy to one.In Xen based hypervisor to optimize the topology
using open source environment. This research work is targeting the following five
primary objectives such as
¾ Increase Manageability of the Server and Workstation Environment.
¾ Increase the Security of the LAN Environment.
¾ Increase the Reliability of the Environment.
¾ Provide User Community with Updated Technology.
¾ Reduce the cost and System utility
2 Detailed Description
2.1 Front-End and Back-End Topology without a Firewall
Scenario
A network willing to maintain a single namespace for their e-mail servers but cannot
fit all of their users on a single server.
Setup Instructions
1. Set up a standard collection of servers running Exchange.
2. Set up a single server running Exchange configured as a front-end server.
3. Direct HTTP, POP, and IMAP users to this server, not to their back-end servers.
4. Ensure that all virtual directories and servers are configured identically on all front-
end and back-end servers.
470 R. Anand and T. Deenadayalan
Fig. 1. Front-End and Back-End Topology Fig. 2. Front-End and Back-End Topology
behind a firewall
Issues
This is the default configuration. You do not need to perform any steps other than the
standard front-end and back-end configuration steps.
If the network permits connections between the client and the back-end servers,
there is nothing to prevent users from circumventing the front-end server and connect-
ing directly to the back-end server.
If this is undesirable, you must change the network routing configuration or the
back-end server configuration to prevent direct connections between a client and a
back-end server.
2.2 Front-End Server behind a Firewall
Scenario
To achieve security and still provide access to Outlook Web Access, POP, or IMAP
from the Internet, a corporation wants to place the Exchange system behind the corpo-
rate firewall.
Setup Instructions
1. Set up a standard Exchange front-end and back-end environment in the corpora-
tion.
2. Configure a firewall between the front-end server and the Internet. For more in-
formation about how to configure an Internet firewall for use with a front-end
server running Exchange
Issues
Because the entire configuration is inside the firewall, Exchange does not require any
special configuration. After a request comes through the firewall to the front-end
server, the front-end server returns a response without any configuration changes.
IP address filtering is highly recommended to limit requests through the firewall to
only those going to the front-end server (or servers) running Exchange and block re-
quests through the firewall to other servers in the organization.
2.3 Configuring a Front-End Server
A front-end server is an ordinary Exchange server until it is configured as a front-end
server. A front-end server must not host any users or public folders.
Server Virtualization: To Optimizing Messaging Services 471
A front-end server must be a member of the same Exchange organization as the
back-end servers (therefore, a member of the same Windows 2000 forest).
2.4 Configuring a Back-End Server
Exchange configuration is stored in Active Directory on a per-forest basis, which
means that all front-end and back-end servers must be in the same forest. Back-end
servers can be accessed directly if required, with no effect on the behavior of the
front-end and back-end configuration. If you did not configure any extra virtual serv-
ers or directories on any front-end servers, then you do not need to configure any on
the back-end server. If you created additional virtual servers or directories on any
front-end servers, however, you must add matching virtual servers and directories on
the back-end servers.
3 Experimental Setup
OS virtualization is achieved by inserting a layer of software between the OS and the
underlying server hardware. This layer is responsible for allowing multiple OS
images (and their running applications) to share the resources of a single server. Each
OS believes that it has there sources of the entire machine under its control, but
beneath its feet, the virtualization layer transparently ensures that resources are
properly shared between different OS images and their applications. Using Xen
hypervisor to configure front end server is windows 2003 server and backend server is
windows 2003 server more than 2 or 3 servers can be running simultaneously. In this
situation more than 2 servers takes less time to utilize the hardware. All the hardware
and virtual machines information are available in hypervisor. Xen based hypervisor
Domain-0 contains all the information. To avoid physical partitioning. Instead, the
virtualization platform generally traps instructions issued by virtual machines and
either passes the instruction through to the physical processor or emulates the
instruction by issuing one or more different instructions to the physical processor and
returning the expected result to the virtual processor. Depending on the virtualization
platform and its configuration, it is possible for instructions from a single virtual
processor to be executed sequentially across one or more physical processors in the
host server. This is not in any way multiprocessing, as the instructions are not
executed in parallel, but it can be performed in order to optimize the processor
resource scheduler in the VMM and to help increase virtual machine performance.
Using python script to configure the virtual machines. This script contains all the
virtual machine configurations can use the system requirements processor: Intel Core
i3M3302.13GHz,MotherBoard:Intel5SeriesBoard3400SH,RAM:2GBDDR2,HardDis
k:320GB,OS:CentOS,Hypervisor:Xen,VirtualMachine:Windows2003Server.
(a) Host Cent OS (b) Xen Virtual Manager (c) Windows2003Server
(Guest OS)
472 R. Anand and T. Deenadayalan
(d) Windows Xp(Guest OS) (e) To access Exchange
Server Using
Guest OS (Windows Xp)
(f) Performance Analysis of all
Virtual Machines
4 Conclusion and Future Work
Server Virtualization is to optimize the front end and back end topology using
exchange server. Exchange server provides secure intranet (or) internet messaging
applications. The topology can be optimized using Xen hypervisor. It increases the
speed and utilizes the hardware fully. Further try to optimize the topology using dif-
ferent virtual machine monitors then the performance is analyzed.
References
[1] Vallée, G., Scott, S.L.: Xen-OSCAR for Cluster Virtualization. In: Min, G., Di Martino,
B., Yang, L.T., Guo, M., Rünger, G. (eds.) ISPA Workshops 2006. LNCS, vol. 4331, pp.
487–498. Springer, Heidelberg (2006)
[2] Ueno, H., Hasegawa, S., Hasegawa, T.: Virtage: Server virtualization with hardware trans-
parency. In: Lin, H.-X., Alexander, M., Forsell, M., Knüpfer, A., Prodan, R., Sousa, L.,
Streit, A. (eds.) Euro-Par 2009. LNCS, vol. 6043, pp. 404–413. Springer, Heidelberg
(2010)
[3] Cafaro, M., Aloisio, G.: Grids, Clouds, and Virtualization. Computer Communications and
Networks, 1–21 (2011)
[4] Van Do, T., Krieger, U.R.: A performance model for maintenance tasks in an environment
of virtualized servers. In: Fratta, L., Schulzrinne, H., Takahashi, Y., Spaniol, O. (eds.)
NETWORKING 2009. LNCS, vol. 5550, pp. 931–942. Springer, Heidelberg (2009)
[5] Zhang, B., Wang, X., Lai, R., Yang, L., Wang, Z., Luo, Y., Li, X.: Evaluating and optimiz-
ing I/O virtualization in kernel-based virtual machine (KVM). In: Ding, C., Shao, Z.,
Zheng, R. (eds.) NPC 2010. LNCS, vol. 6289, pp. 220–231. Springer, Heidelberg (2010)
[6] Xie, W., Navathe, S., Prasad, S.K., Fisher, D., Yang, Y.: Optimizing peer virtualization
and load balancing. In: Li Lee, M., Tan, K.-L., Wuwongse, V. (eds.) DASFAA 2006.
LNCS, vol. 3882, pp. 357–373. Springer, Heidelberg (2006)
[7] Hagen, W.V.: Professional Xen® Virtualization.
[8] Marshall, D., Reynolds, W.A., McCrory, D.: Advanced Server Virtualization
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 473–476, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Blind Source Separation for Convolutive Audio Mixing
V. Jerine Rini Rosebell, D. Sugumar, Shindu, and Sherin
Department of ECE, Karunya University, Coimbatore, Tamil Nadu
jerinebell87@gmail.com, sugumar.ssd@gmail.com
Abstract. This paper describes an efficient Blind Source Separation of speech
and music, speech and music which are considered as convolutive mixtures.
The convolutive mixed signals consist of source signals and same amount of
delay or echo of the same source signal. Convolutive BSS of stereo mixtures is
the challenging task in the audio signal processing application.BSS is a tech-
nique for estimating original source signal from their mixtures of signals. The
mixed signals were decomposed by 1D multilevel discrete wavelet decomposi-
tion. Decomposition levels are changed and the signals to noise ratio (SNR) are
calulated.After that ICA has been performed and the sources are separated.
Keywords: Independent Component Analysis, Blind Source Separation, Con-
volutive mixtures, Multilevel Wavelet decomposition.
1 Introduction
Blind Source Separation is the separation of a set of signals from a set of mixed sig-
nals without the aid of information (or with very little information) about the source
signals or the mixing process [1]. BSS thus separates a set of signals into a set of orig-
inal signals, such that the independency of each resulting signal is maximized, and the
dependency between the signals is minimized [3]. Independent component analysis in
time domain holds good for speech. It will produce better results for speech signals.
Independent component analysis (ICA) is a computational method for separating a
multivariate signal into additive subcomponents supposing the mutual statistical inde-
pendence of the non-gaussian source signals [5]. It is a special case of blind source
separation. Blind source separation refers to two class of multichannel signal
processing tasks in which the goal is to extract multiple useful signals from the mul-
tiple convolutive mixtures of these signals without specific knowledge of the source
properties or the mixing characteristics [6].Convolutive BSS assumes a general multi-
path channel thus requiring multichannel filtering.
2 Convolutive Mixing Model
The source signals are called convolved mixtures since acoustic signals recorded si-
multaneously in a reverberant environment can be described as the sums of differently
convolved sources. In the testing process Mixture2 is taken and it is plotted bellow.
474 V.J. Rini Rosebell et al.
x(n)=∑
(1)
Fig. 1. Convolutive Mixing Model
3 Proposed Algorithm
Wavelet decomposition is applied on the mixed signal (speech+music).Multilevel 1D
wavelet decomposition is used to decompose the mixed signal. It will produce the
approximated & detailed signals. After that independent component analysis is ap-
plied on the decomposed signals.ICA effectively separates the mixed signals. Finally
by applying wavelet reconstruction on the ICA Separated signals the mixed sources
are separated as speech, music.
Fig. 2. Flow diagram
4 Independent Component Analysis
The independent component analysis allows two source signals to be separated from
two mixed signals using statistical principles of independent and non-gaussianity [4].
Wavelet decom
p
ositio n Wavelet decom
p
osition
ICA ICA
Wavelet reconstruction
Wavelet reconstruction
Speech Music
Speech+Music S
p
eech+Music
Pre em
p
hasis Pre em
p
hasis
Blind Source Separation for Convolutive Audio Mixing 475
The algorithm requires that there be as many sensors as input signals. For example,
with three independent sources and three mixtures being recorded, the problem could
be modelled as: (t)=a(t)+b(t) (2)
(t)=c(t)+d(t) (3)
5 Results and Discussions
Table 1. Snr Measurements For Mixture1
Wavelet
Decomposition Levels
SNR SNR1
4 21.7932 7.1472
5 22.5731 7.0732
6 22.700 7.1951
7 26.9025 11.2425
Table 2. Snr Measurements For Mixture 2
Wavelet
Decomposition Levels
SNR SNR1
4 21.1932 3.5927
5 21.2475 3.5923
6 21.3349 4.3686
7 19.2031 7.9077
Fig. 3. ICA(approximation &detailed signals )
Fig. 4. Reconstructed signal (Speech, Music)
6 Conclusion
The mixture1 and mixture2 are separated using independent component analysis
using wavelets decomposition method. Input signals are considered as convolutive
mixing signal. The signal strength mainly depends on the levels of decomposition, As
476 V.J. Rini Rosebell et al.
the number of levels increases SNR also increases. The experimental results show that
mixtures2 produce better results than the mixture1.For mixture1 (speech) the SNR
value increases up to 6 levels after that it start decreasing. The best SNR obtained for
mixture1 in the sixth level is 21.3349dB. The maximum SNR obtained in the seventh
level is 26.9025dB.
References
1. Ozerov, A., Févotte, C.: Multichannel Nonnegative Matrix Factorization in Convolutive
Mixtures for Audio Source Separation. IEEE Transactions on Audio, Speech, and Lan-
guage Processing 18(3) (March 2010)
2. Cobos, M., Lopez, J.J.: Stereo audio source separation based on time–frequency masking
and multilevel thresholding. Elsevier -Digital Signal Processing 18, 960–976 (2008)
3. Jafari, M.G., Vincent, E., Abdallah, S.A., Plumbley, M.D., Davies, M.E.: An adaptive ste-
reo basis method for convolutive blind audio source separation. Elsevier Neurocomput-
ing 71, 2087–2097 (2008)
4. Nesbita, A., Plumbleya, M.D., Daviesb, M.E.: Audio Source Separation with A Signal-
Adaptive Local Cosine Transform. Elsevier, Signal Processing 87, 1848–1858 (2007)
5. Douglas, S.C., Gupta, M., Sawada Sr., H., Makino, S.: Spatio–Temporal FastICA Algo-
rithms for the Blind Separation of Convolutive Mixtures. IEEE Transactions on Audio,
Speech, and Language Processing 15(5) (July 2007)
6. Mei, T., Xi, J., Yin, F., Mertins, A., Chicharo, J.F.: Blind Source Separation Based on
Time-Domain Optimization of a Frequency-Domain Independence Criterion. IEEE Trans-
actions on Audio, Speech, and Language Processing 14(6) ( November 2006)
7. Vincent, E., Gribonval, R., Févotte, C.: Performance Measurement in Blind Audio Source
Separation. IEEE Transactions on Audio, Speech, and Language Processing 14(4) (July
2006)
8. Ding, S., Cichocki, A., Huang, J., Wei, D.: Blind Source Separation of Acoustic Signals in
Realistic Environments Based on ICA in the Time-Frequency Domain. Journal of Perva-
sive Computing and Communications 1(2) (June 2005)
9. Das, N., Routray, A., Dash, P.K.: ICA Methods for Blind Source Separation of Instantane-
ous Mixtures: A Case Study. Neural Information Processing – Letters and Reviews 11(11)
(2007)
10. Addison, W., Roberts, S.: Blind Source Separation with Non-Stationary Mixing Using
Wavelets. The University of Liverpool (2006)
11. Araki, S., Makino, S., Mukai, R., Nishikawa, T., Saruwatari, H.: Fundamental Limitation
of Frequency Domain Blind Source Separation for Convolved Mixture of Speech. NTT
Communication Science Laboratories and Nara Institute of Science and Technology, Japan
(2006)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 477–480, 2011.
© Springer-Verlag Berlin Heidelberg 2011
ICA Based Informed Source Separation for Digitally
Watermarked Audio Signals
R. Sharanya, D. Sugumar, T.L. Sujithra, Susan Mary Bose, and Divya Mary Koshy
Department of ECE, Karunya University, Coimbatore, Tamil Nadu, India
sariyu.ece@gmail.com, sugumar.ssd@gmail.com,
sujithraleon@gmail.com, susanmarybose@gmail.com,
dvy.koshy@gmail.com
Abstract. This paper presents an efficient, digital audio watermarking method
based on one dimensional multilevel discrete wavelet transform(MDWT) and
discrete cosine transform (DCT) for the application of copyright protection. The
proposed scheme also employs informed source separation for the watermark
extracted audio signal based on independent component analysis (ICA).In addi-
tion, the performance of the proposed algorithm in terms of Signal to Noise Ra-
tio (SNR), Peak Signal to Noise Ratio (PSNR) and Normalized Root Mean
Square Error (NRMSE) are also evaluated. The proposed scheme achieved
good robustness against most of the attacks such as requantization, filtering, ad-
dition and multiplication of noise. The experimental result shows that the SNR
value is about 36.12dB for the proposed algorithm, whereas by using only
MDWT technique method gives only 30.2dB.
Keywords: Independent Component Analysis, Informed Source Separation,
Audio Watermarking, Copyright protection.
1 Introduction
Digital watermarking techniques are considered to be effective solution to the prob-
lems of copyrights. Currently, watermarking techniques based on transform domain
are more popular than those based on time domain since they provide higher audio
quality and more robust. Also in terms of the performance of watermarks against at-
tacks, the transform domain methods are commonly considered better than that of the
time and frequency domain methods [1].Additional known audio watermarking
classes include spread spectrum [3] and compressed audio techniques, in which the
spread spectrum requirements of hiding a signal against an unintended listener and
ensuring information privacy, are very similar to those in watermark applications
However, computational complexity and synchronization overhead may be unaccept-
ably high.
Blind Source Separation (BSS) method was proposed in [6] to separate a large
number of speech sources. When the number of available audio channels (mixtures)
equals or exceeds the number of individual sources Independent component analysis
(ICA) is used. In this paper, an audio watermarking algorithm that satisfies the re-
quirements of effective audio watermarking inaudibility and watermark robustness has
478 R. Sharanya et al.
been proposed. The requirements were met by the proposed algorithm by exploiting
the attractive properties of two powerful mathematical transforms; one dimensional
Multilevel Discrete Wavelet Transform (MDWT) and Discrete Cosine Transform
(DCT).In the proposed algorithm, watermark bits are embedded directly on coeffi-
cients of DCT, MDWT is taken and it is reconstructed. The watermark embedding and
extraction procedure, source separation is outlined in section two, and experimental
results, in section three. Conclusion is given in section four.
2 Proposed Algorithm
The proposed algorithm employs a cascade of two transforms; the discrete wavelet
transform and the discrete cosine transform. The algorithm is described in this section
by outlining the major steps, the watermark embedding procedure, and watermark
extraction procedure [7].
2.1 Watermark Embedding/Extraction and Source Separation Procedure
The procedure is illustrated in the block diagram shown in Fig 1, and described in
details in the steps which follow below,
Image
Estimated
audio signal
MDWT
&DCT
IMDWT
&IDCT
MDWT
&DCT
IMDWT
&IDCT
Mixed audio
(
S1 &S2
)
Watermarked audio
Secret
key
Embed
ICA
S1 S2
Fig. 1. Watermark Embedding/Extraction, Source Separation Procedure
Two different audio signals (music + speech) of sampling frequency 44100, bit
resolution of 16 and PCM uncompressed signals are mixed linearly and image is tak-
en. Perform a MDWT and DCT transformation. Perform inverse operation of the em-
bedded signal by taking IDCT and IMDWT, the obtained signal is the watermarked
signal. The watermarked audio signal obtained from the embedding process is taken
and MDWT and DCT for the signal are taken. The binary image is then extracted by
using a secret key. Assemble the extracted bits from the inverse operation of MDWT
and IDCT is taken and the estimated audio signal is obtained. estimated audio signal
is obtained by applying ICA and sources are separated [9].
ICA Based Informed Source Separation for Digitally Watermarked Audio Signals 479
3 Experimental Results and Discussion
In this section, the results obtained using the MDWT-DCT algorithm has been pre-
sented. Pop music and speech audio clips were used to evaluate performance of the
proposed algorithm. The watermark used in our experiments is the binary image
shown in Fig 2. Then performance measures such as SNR, PSNR, and NRMSE are
tabulated.
Fig. 2. Detected Watermark Image
Fig. 3. Watermarked Audio Signal
Fig. 4. Separated Music Signal after applying ICA
Fig. 5. Separated Speech Signal after applying ICA
480 R. Sharanya et al.
Table 1. Performance Measures after Watermark Insertion &watermark Extraction for MDWT-
DCT
Performance
measures After Watermark
insertion
After Watermark
extraction
SNR 36.1 21.6
PSNR 78.2 42.1
NRMSE 12.7 3.4
4 Conclusion
In this paper, proposed an imperceptible (inaudible) and robust audio watermarking
technique based on cascading two powerful mathematical transforms; MDWT and the
DCT. By virtue of cascading the two transforms, inaudibility and different levels of
robustness is achieved. This paper presents that preprocessing is important where
SNR measure without pre-processing process gives 0.0014dB which is very low. The
simulation results obtained verify the effectiveness of audio watermarking as a relia-
ble solution to the copyright protection. Contrary to the BSS framework, in the In-
formed Source Separation, source signals are available before the mix is processed.
The proposed ICA based detector can be used for wavelet based watermarking for all
types of multimedia data, e.g., audio, video, images, etc.
References
1. Cox, I.J., Kilian, J., Shamoon, T.: Secure Spread Spectrum Watermarking for Image, Audio
and Video. IEEE Trans. on Image Processing 6, 1673–1687 (1997)
2. Khademi, N., Akhaee, M.A., Ahadi: Audio watermarking based on quantization index
modulation in frequency domain. Signal Processing and Communication, 1127 (2007)
3. Kirosvki, D., Malwar, H.: Robust spread spectrum watermarking. In: Proceedings of Acous-
tics, speech and signal processing (ICASSP 2001). , vol. 3, pp. 1345–1348 (2001)
4. Kaengin, S., Airphaiboon, S., Patthoumvanh, S.: New technique for embedding watermark
image into an audio signal. Communication and Information Technology (ISCIT) (2009)
5. Yilmaz, O., Rickard, S.: Blind separation of speech mixtures time-frequency masking. IEEE
Trans. Signal Processing 52, 1830–1847 (2004)
6. Parvaix, M., Girin, L., Brossier, J.M.: A Watermarking based method for informed source
separation of audio signals with a single sensor. IEEE Trans.on Audio, Speech and Lan-
guage Processing 18, 1464–1475 (2010)
7. Haj, A.A., Mohammad, A.: Digital audio watermarking based on discrete wavelet transform
and singular value decomposition. European Journal of Scientific Research 39, 6–21 (2010)
ISSN 1450-216X
8. Das, N., Routray, A., Dash, P.K.: ICA methods for blind source separation of Instantaneous
Mixtures. Letters and Reviews 11 (November 2007)
9. Douglas, S.C., Gupta, M., Sawada, H.: SpatioTemporal fast ICA algorithms for the blind
separation of convolutive mixtures. IEEE Tran. on Audio, Speech and Language
Processing 15, 1511–1520 (2007)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 481–484, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Evaluation of Retrieval System Using Textural Features
Based on Wavelet Transform
Lidiya Xavier and I. Thusnavis Bella Mary
School of Electrical Sciences, Karunya University, Coimbatore-641114, India
lidiyaxavier@gmail.com, bellamary@karunya.edu
Abstract. A content based image retrieval system allows the user to present a
query image in order to retrieve images stored in the database according to their
similarity to the query image. Content based image retrieval method is used as
diagnosis aid in medical fields. In this paper content based image retrieval is
used for medical applications. The main objective of this paper is to evaluate
the retrieval system using Textural features. The texture features are extracted
by using wavelet transform. The method is evaluated on Diabetic Retinopathy
Database (DRD). Here the precision rate obtained is about 60% for DRD
images.
Keywords: Image Retrieval, Content based image retrieval, Pyramid-structure
wavelet transform, Texture and Medical image.
1 Introduction
Image Retrieval aims to provide an effective and efficient tool for managing large
image databases. Image retrieval (IR) is one of the most exciting and fastest growing
research areas in the field of medical imaging [2]. A CBIR method typically converts
an image into a feature vector representation and matches with the images in the data-
base to find out the most similar images. From reference paper Mathieu Lamard [1],
Aliaa.A.A Yousif [2] and Cazuguel, G. [3] different databases have been used and a com-
parative study is carried out. From Mathieu Lamard [1] and Cazuguel, G. [3], it is con-
cluded that performance of DRD image is less compared to other database images. In
order to improve efficiency of DRD database images pyramid-structured wavelet
transform is used.In this paper evaluation of retrieval system based on Textural fea-
tures is carried out. The major advantage of this approach is that little human inter-
vention is required. The databases used here is Diabetic Retinopathy Database [1] [3].
2 Block Diagram
Figure 1 shows the basic block diagram used in this work. Textural features are ex-
tracted for both query image and images in the database. The distance (ie., similari-
ties) between the feature vectors of the query image and database are then computed
482 L. Xavier and I.T. Bella Mary
and ranked. The database images that have highest similarity to the query image are
retrieved. Then the performance analysis is carried out using precision and recall.
Image
Database
Feature extraction
(Texture)
Result
Query Image
Performance
Measurement
Similarity
Measurement
Feature extraction
(Texture)
Fig. 1. Basic block diagram
3 Texture Extraction
The property of all surfaces that describes visual patterns, each having properties of
homogeneity is termed as texture. Three types of texture features are used in this pa-
per. They are energy, contrast and entropy. Energy is a measure of textural uniform-
ity. Energy is low when all elements are equal and is useful for highlighting geometry
and continuity. Here energy is calculated using pyramid-structure wavelet transform,
and it will explain in the following sections. Contrast is a measure of the contrast or
amount of local variation present in an image or surface. A texture of high contrast
has large difference in intensity among neighboring pixels, while a texture of low
contrast has small difference. Here contrast is calculated for every image in each de-
composition level. Entropy is a measure of disorder or complexity. It is large for sur-
faces that are texturally not uniform.
3.1 Pyramid-Structured Wavelet Transform
The pyramid-structure wavelet transform indicate that it recursively decomposes sub
signals in the low frequency channels [2].Using the pyramid-structure wavelet trans-
form, the texture image is decomposed into four sub images, as low-low, low-high,
high-low and high-high sub-bands. The energy level of each sub-band is calculated.
This is first level decomposition. Using the low-low sub-band for further decomposi-
tion is done. Decomposition is done up to third level in this paper. The reason for this
type of decomposition is the assumption that the energy of an image is concentrated
in the low-low band.
Evaluation of Retrieval System Using Textural Features Based on Wavelet Transform 483
3.2 Energy Level and Euclidean Distance
Calculate the energy of all decomposed images at the same scale, using:
()
EMN Xij
j
n
i
m
=
==
∑∑
1
11
,
(1)
where M and N are the dimensions of the image, and X is the intensity of the pixel
located at row i and column j in the image map. Repeat from step 1 for the low-low
sub-band image, until it becomes third level. Using the above algorithm, the energy
levels of the sub-bands is calculated [2]. These energy level values are stored to be
used in the Euclidean distance algorithm. And the equation is given below. The
Euclidean distance D between two vectors X and Y is
)2^)((
,
∑
−=
ji
YXD (2)
Using the above algorithm, the query image is searched for in the image database.
The Euclidean distance is calculated between the query image and every image in the
database. This process is repeated until all the images in the database have been com-
pared with the query image.
4 Graphical Analysis
Figure 2(a) shows the graphical representation of precision versus number of retrieved
images. From the graph it is understood that to retrieve all the relevant images in the
database almost the maximum number is has to be retrieved for all the class. Almost
50% of the relevant image is retrieved for precision at P(3). The precision rate is
about 60% when first three images are retrieved from the database. Graphical repre-
sentation of recall and number of retrieved images is shown in the figure 2(b). The
recall rate is high when first three images are retrieved in case of background and pro-
filiative retinopathy compared to non profiliative. Figure 2(c) shows the graphical
representation of precision versus recall. Precision value is inversely proportional to
recall value. The maintains of high precision value at various levels indicates that
majority of relevant images are retrieved at the early stage.
Fig. 2. (a) Precision Vs Number of Retrieved Images,(b) Recall Vs Number of Retrieved Im-
ages, (c) Precision Vs Recall
484 L. Xavier and I.T. Bella Mary
5 Conclusion
In this paper pyramid-structure wavelet transform based method is considered for
content based diabetic retinopathy image retrieval. From Fig. 2 it can be concluded
that the best precision rate of the 60% and recall rate of 60% is achieved using this
method. In paper [1] DRD image is evaluated using adaptive nonseparable lifting
scheme which results in poor retrieval rate. Hence by using this technique the re-
trieval rate for DRD image is improved.
References
1. Quellec, G., Lamard, M., Cazuguel, G., Cochener, B.: Adaptive Non- Separable Wavelet
Transform via Lifting and its Application to Content-Based ImageRetrieval. IEEE Transac-
tions on Image Processing 19(1), 25–35 (2010)
2. Yousif, A.A.A., Darwish, A.A., Mohammad, R.A.: Content based medical image retrieval
based on pyramid structure wavelet. IJCSNS International Journal of Computer Science and
Network Security 3 (March 2010)
3. Lamard, M., Cazuguel, G., Quellec, G., Bekri, L., Roux, C., Cochener, B.: Content based
image retrieval based on wavelet transform coefficients distribution. In: Proc. of the 29th
annual Int. Conf. of the IEEE Engineering in Medicine and Biology Society (August 2007)
4. Antani, S., Lee, D.J., Long, L.R., Thoma, G.R.: Evaluation of shape similarity measurement
methods for spine X-ray images. J. Vis.Commun. Image R 15(3), 285–302 (2004)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 485–488, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Behavioural Level Watermarking Techniques for IP
Identification Based on Testing in SOC Design
Newton david Raj, Josprakash, AntopremKumar, Daniel, and Joshua Thomas
School of Electrical Science, Karunya University, Coimbatore-641114, India
{newton.davidraj1, jos.prakash.av, anto.donns, danielpon10,
joshuaalackal}@gmail.com
Abstract. This paper proposes a watermarking scheme for intellectual property
(IP) identification based on testing in soc design. The core concept is embed-
ding the watermarking generating circuit (WGC) and test circuit (TC) in to the
soft IP core at the behavioural design level. Therefore this scheme can success-
fully survive synthesis, placement and routing and can identify the IP core at
various design levels and IP core does not change at any levels. This method
adopts current main system-on-chip (SOC). The identity of the IP is proven
during the general test process without implementing any extra extraction flow.
After the chip has been manufactured and packaged, it is still easy to detect the
identification of the IP provider without the need of microphotograph. This ap-
proaches entail low hardware overhead, tracking costs, and processing-time
costs. The proposed method solves the IP-identification problem.
Keywords: intellectual-property (IP), system-on-a-chip (SOC), very large scale
integration (VLSI) design, watermarking generating circuit (WGC), test circuit
(TC).
1 Introduction
Advances in semiconductor processing technology have led to the rapid increases in
integrated–circuit (IC) design complexity [1],[2].The shift toward very deep sub
micrometer processing technology has encouraged IC designers to design an entire
system implemented on a single chip. This new paradigm, called the system-on-a-chip
(SOC), has changed design methodologies. In order to reduce time to market and to
increase productivity, the reuse of previously designed modules is becoming a com-
mon practice.
Design rules lead to the development of IP-identification techniques. Each IP
should have identification that represents the design information, including designer
identity, version, ownership rights, and provider. The identification can also provide
designer information, IP tracking, ownership proof, and IP management. The ability
to prove the identity of virtual components is increasing in importance. After the IP
has been integrated into a whole chip and packaged, designers can still check the
identity of the IP. This paper deals the design for SOC using watermark techniques
for IP identification.
486 N. david Raj et al.
2 IP-Based Design Flow with Watermarking
In this section, develop an IP-identification approach using the watermarking tech-
nique. This method is developed depending on the current IP-based design flow [1].
The explanation will describe the design process of headed watermarking-sequence
method.
2.1 Watermark Design
First of all, the watermark, which can intuitively represent one’s identity, it is gener-
ated as a binary sequence and inserted in to each IP core. Propose a coding technique
for the design of the digital watermark. The watermark is a symbol that stands for the
organization’s title, a laboratory’s mark, or a personal name, and it is comprised of a
sequence of bits. For example, the symbol “CEDECEKUC” to represent College of
Engineering, Department of Electronics and Communication Engineering, Karunya
University Coimbatore. It describes the symbol according to the coded table that con-
structed beforehand. 00010 to represent letter C, 00100 to represent E, 00011 to
represent D, 00100 to represent E, 00010 to represent C, 00100 to represent E, 01010
to represent K, 10100 to represent U, 00010 to represent C. With this method of en-
coding, just 45 bits are needed to describe “CEDECEKUC”.
2.2 WGC Design
The WGC is composed of several parallel-input–serial output (PISO) registers and
inverter gates. When the test-mode signal is active (test mode = 1), the WGC will be
turned on. The parallel watermark data are generated by the inverters. If the water-
mark value is one, the circuit directly generates the value. If the watermark value is
zero, there is an inverter that translates the test-mode signal into zero. The watermark
data are generated via the test-mode signal and inverters. The PISO translates the pa-
rallel watermark data into a sequence.
3 Combining TC with WGC
After the WGC has been designed, combine the TC with the WGC. How the TC is
combined with the WGC is very important. two methods for combining the TC with
the WGC, and we analyse the characteristics of each method of combining test circuit
with wgc circuit.
3.1 Headed Watermark-Sequence Method
When the chip is in the test mode, the chip sends out first the watermark sequence.
After sending out the entire watermark sequence, the chip sends the output test
patterns. The watermark sequence is like the header of a bit stream.This method
enables the watermark to simply be extracted. Drawback is that the watermark is easy
to remove
Behavioural Level Watermarking Techniques for IP Identification 487
OUTPUT_PIN1 00011……00101011000110…….
OUTPUT_PIN2 10011……10101011010110…….
OUTPUT_PIN3 01011……00101011111110…….
OUTPUT_PIN4 11011……11101011000110…….
Watermark test patterns
sequence
Fig. 1. Headed watermark-sequence method
(a) (b)
Fig. 2. (a) Architecture of soft IP, (WGC) and test circuit. (b) WM-IP Testing By BIST
488 N. david Raj et al.
4 Experimental Results
The IP cores were designed using Verilog HDL and were verified. The watermark
function is not changed after logic synthesis because embed the watermark into the
TC at the behavioural design level. After placement and routing, still detect the identi-
ty, according to the watermark sequence, without error. According to the results, the
proposed method can identify the soft IP core at the behavioural, gate, and physical
design levels.
(a) (b)
Fig. 3. (a) WGCresponse(when clock=1, reset=0 test mode=1 and ready signal =1). (b) Water-
mark sequence and test circuit response
5 Conclusion
The watermark is a general-purpose design methodology that does not need to be de-
signed case by case according to various IPs. The watermark function is not changed
after logic synthesis, placement, and routing because the watermark is embedded into
the TC at the behavioural design level. The approaches have the ability to detect the
presence of the watermark and to identify the soft IP core at various design levels.
References
[1] Fan, Y.-C.: Testing-Based Watermarking Techniques for Intellectual Property Identifica-
tion in SOC Design. IEEE Trans. Instrumentation and measurement 57(3) (March 2008)
[2] Martin, G., Chang, H.: Winning the SoC Revolution: Experiences in Real Design. Kluwer,
Norwell (2003)
[3] Cox, I.J., Miller, M.L., Bloom, J.A.: Digital Watermarking. Morgan Kaufmann, San Mateo
(2002)
[4] Narayan, N., Newbould, R.D., Carothers, J.D., Rodriguez, J.J., Holman, W.T.: IP protec-
tion for VLSI designs via watermarking of routes. In: Proc. IEEE Int. Conf. ASIC/SOC,
September 2001, pp. 406–410 (2001)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 489–492, 2011.
© Springer-Verlag Berlin Heidelberg 2011
Rough Set Approach for Distributed Decision Tree and
Attribute Reduction in the Disseminated Environment
E. Chandra1 and P. Ajitha2
1 Research Supervisor & Director, Department Of Computer Science, D J Academy for
Managerial Excellence, Coimbatore, Tamilnadu, India
2 Research Scholar & Assistant Professor, Department Of Computer Science, D J Academy for
Managerial Excellence, Coimbatore, Tamilnadu, India
ajitha@y7mail.com
Abstract. Attribute reduction is a necessitated step for the disseminated envi-
ronment in regard to classification and prediction of the data. Traditional ap-
proaches were not efficient for optimal attribute reducts. Current techniques are
quiet time consuming and less accuracy. Rough set approach is a mathematical
technique to handle attribute reducts through data dependencies and structural
methods. This paper discusses a novel algorithm for optimal attribute deduct
and also increases the accuracy in predicted results and the distributed decision
tree classification techniques was made use of to implement the same in the dis-
seminated environment. Proposed algorithm for Construction of distributed de-
cision trees with rough sets increases the accuracy and also reduces the attrib-
utes on the time of the massive data sets handling.
Keywords: Rough Sets, Distributed Decision Trees, Distributed data mining.
1 Introduction
Distributed Data Mining mines massive data sets and also heterogeneous type of data.
An appropriate mathematical model is necessary to deal with the data especially of
heterogeneous types where there was large number of attributes. Individual analysis
of each site is different and inadequate for some special treatments [1] and also when
there was uncertain, vague and imprecision data handling of them was very difficult.
Rough sets, one of the mathematical model is dealt with these category of data.
Rough set was proposed by Pawalak [4], one of the most popular techniques ap-
plied in machine learning, data mining[3], pattern recognition etc. It provided the
strategies to discover the data dependences and used structural methods. This paper
proposes a rough sets distributed decision tree algorithm which achieves simultaneous
prototype selection and feature selection, and a scheme to use the rough sets for deci-
sion tree for efficient classification in a disseminated environment.
2 Proposed Computation of Attribute Reduction
To compute aij the attribute reductions, construct the attribute subset enumeration
tree by merely using the non-core attributes. Already Distributed Decision trees[9]
490 E. Chandra and P. Ajitha
constructions with roughs sets was proposed in the literature survey so it was not
discussed here again.
2.1 Computation Method
Sequential Attribute Reduction Algorithm SARA, similar to the algorithm put for-
ward by Zhang [6], and the distributed attribute reduction algorithm DARA based on
peer-to-peer technique both for the client and server was discussed here. The
maximum length of queue Q or T should be (())/2
()
CCoreC
CUCoreC
⊥−
− [9]. when the
computation is amortized on computers in a peer-to-peer network to apportion Q and
T, the reduction of larger data set with more attributes may be carried out efficiently
as the T computation was calculated. Newly proposed Distributed Attribute Reduction
Algorithm DARA is composed of the server DARA-S and the client DARA-C.
Because of page constraint DARA-S is specified here with the small elimination of
integration DARA-S is taken as such
Input: U , C, and D of consistent decision system
Output: all attribute reduction set R
(1) compute Core(C), send M('C', Core(C)) to all N client computers;
(2) for all a ∈ C -Core(C), send M('Q', {a}) toclient computer C[i] in round robin way to the
distributed evt and NQ =|A|,NT =0, R=
φ
; (3) send M ('S ') to all clients to start checking and
computing the attribute reductions;(4) when receive M('R', q), put q → R;(5) when receive
M('T', q), set NT = NT + 1 and q→R with NQ and NT
.(6) when receive M ('N') ,s et NQ =NQ -1; (6.1) if NQ = 0 and NT = 0, stop; if NQ =0 ,s end M('P')
to all clients;(7) when receive M ('Q', t) , set NQ = N Q + 1 and= NT + 1 and q→R with NQ and
NT and (8) when receive M ('F') , set NT = NT -1 ; then if NT =0 and NQ =O,stop;check if NT =0
,goto (3). (9) send these NQ and NT to all sites
Algorithm 1. DARA-S: Distributed Attribute Reduction on Server
Proposed algorithm DARA-S reduces the attributes when dealing in the distributed
architecture. When the computational complexity of the algorithm is defined O(n+m)
and M(log2(n2))O(n). Here M and n represents the attributes that is sent over across
the sites. Computational memory that is taken is very less in the proposed algorithms
of DARA-S and DARA-C.
3 Classification Based on Rough Sets
When the classification technique of Distributed Decision Tree is considered a
mathematical model is necessary to produce an optimize results. For this greater
purpose an algorithm was proposed so that heterogeneous data sets can be classified
with greater accuracy and in minimised time in compared with the other existing
methods[10].
Rough Set Approach for Distributed Decision Tree and Attribute Reduction 491
3.1 Proposed Algorithm for Rough Sets in Distributed Decision Tree
The following is the proposed algorithm for classification based on rough sets. CT-
tree[8][10] proposes algorithm for attribute reduction here both for the attribute reduc-
tion, time and accuracy was calculated using rough set theory (rCTl) –Classification
Tree based on rough sets
Inputs: TS-Test patterns,TR-Training patterns, s–User defined threshold.
Outputs: Ctime––Classification time, CA––Classification accuracy.
Steps: (1) Generate CT-tree using TR and initialize start-time =time( ).
(2) For each branch bi of the DDT-tree, find lba, uba, cba(3) For each sj ∈ TS (a) Find nj, set of
positions of non-zero values corresponding to sj.(b) For each branch dk in rCT-tree,if dk corre-
sponds to ubk then nj’= nj-Cbk else nj’= nj-Ubk(c) Find the nearest neighbour branch, ep in rPC-tree
depending on maximum number of features which are common to both ep and nji. (d) Attach
the label l associated with ep to ni’ (e) If (l==label of sj) then Correct= correct + 1.(4) end-time
= time( ). (5) CA = (correct/ |TS| x 100 | TS| is the number of test patterns. (6)Output Ctime= end-
time - start-time and Output CA.(7) check CA and Ct ime as scheduled (7) DARA –S and
DARA-C recursive call
Algorithm 2. Classification based on rough sets
lba =set of features from root to leaf along ba,uba = set of features from root to leaf
along ba where ‘Count’ field values of the nodes is greater than or equalto s. assign
Cba=lba - uba. (i) uba–if ba is shared by the pattern of same class; (ii) cba––otherwise.
Using all these attributes are selected based on some threshold and time factor is
checked by taking into account the value of nodes on that specified criteria’s. The
proposed algorithm deals with these aspects paved way to the following results.
Table 1. Comparision of CT-DDT and rCT1
Algorithm Classification accuracy (%) Classification time (in s)Storage space(in Bytes)
CT 92.5 1713 1,302,454
rCTl 93.22 1207 1,291,347
From this table1 that rCTl gives the best classification accuracy and computational
space. On comparing with the other methods the proposed algorithm specifies the
87% of accuracy in compared with the others as specified in the table and as in
simulated results.
Fig. 1. Proposed Simulation results of rCT1 and attribute reduction by using proposed algorithm
492 E. Chandra and P. Ajitha
4 Conclusions and Future Work
When the data sets are large and need to be transmitted over the different sites
chances for less accuracy and more time because of large selection of attributes. As
Rough sets and indiscernibility functions was made use of the classification accuracy
is maintained and also attribute reductions which lead to the computational storage is
reduced as it is also maintained by reducts on decision trees. As decision trees are
quite efficient for classification, this paper deals with the decision trees in distributed
environment. rCT1 tree scans the databases only once. Segmentation of data can be
discussed further and also association rules can be integrated for the pruning of trees,
when the outliers or over fitting occurs in the distributed decision trees.
References
1. Chitcharoen, D., Pattaraintakorn, P.: Novel matrix forms of rough set flow graphs with ap-
plications to data integration. Computers and Mathematics with Applications 60,
2880–2897 (2010)
2. Guo, Q.L., Zhang, M.: Implement web learning environment based on data mining.
Knowledge-Based Systems 22(6), 439–442 (2009)
3. Cios, K., Pedrycz, W., Swiniarski, R.: Data Mining Methods for Knowledge Discovery.
Kluwer, Norwell
4. Pawlak, Z., Skowron, A.: Rudiments of rough sets. Information Sciences
5. Zhang, E.M., Song, G.Z., Ma, W., Zhang, W.: Attribute Reduction Algorithm Research
Based on Rough Core and Back Elimination. In: Proceedings of the 9th International Con-
ference for Young Computer Scientists (ICYCS 2008), pp. 1624–1628. Central South Uni-
versity, China (2008)
6. Stanczyk, U.: On Construction of Optimised Rough Set-based Classifier. International
Journal of Mathematical Models and Methods in Applied Sciences 2(4) (2008)
7. Sikder, I.U., Munakata, T.: Application of rough set and decision tree for characterization
of premonitory factors of low seismic activity. Expert Systems with Applications 36, 102–
110 (2009)
8. Chen, Y., et al.: A rough set approach to feature selection based on power set tree. Knowl.
Based Syst. (2010)
9. Ma, G., Lu, Y., Wen, P., Song, E.: A novel attribute reduction algorithm based on peer-to-
peer technique and rough set theory. In: IEEE/ICME International Conference on Complex
Medical Engineering (2010)
10. Mi, J.S., Wu, W.Z., Zhang, W.X.: Approaches to knowledge reduction based on variable
precision rough set model. Information Sciences 159(3-4), 255–272 (2004)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 493–498, 2011.
© Springer-Verlag Berlin Heidelberg 2011
MIMO and Smart Antenna Technologies for 3G and 4G
Vanitha Rani Rentapalli1 and Zafer Jawed Khan2
1 MISTE, Associate Member of IEEE (No. 90410990) Assistant Professor,
ECE Department, Vivekananda Institute of Technology and Science, Karimnagar,
Andhra Pradesh
vanitharani@aol.in
2 Vivekananda Institute of Technology and Science, Karimnagar, Andhra Pradesh
khanzj1@rediffmail.com
Abstract. Evolution of wireless access technologies is about to the reach its
fourth generation (4G). The adaptation of smart antenna techniques in future
wireless systems is expected to have a significant impact on the efficient use of
the spectrum and transparent operation across the multi technology wireless
networks. With the rapid growth of wireless data traffic operators are anxious to
quickly expand the capacity of their wireless networks. To address these 3GPP
standards have incorporated powerful techniques for using so-called smart an-
tennas. This paper focuses on the practical aspects of deploying smart antenna
systems in the existing Radio Access Systems (RAS). Smart antenna tech-
niques, such as multiple-input multiple-output (MIMO) systems, can extend the
capabilities of 3G and 4G systems to provide customers with increased data
throughput for mobile high-speed data applications.
Keywords: Adaptive array beam forming, MIMO, smart antenna, 3G, 4G.
1 Introduction
Mobile radio communications are evolving from pure telephony systems to multime-
dia platforms offering a variety of services ranging from simple file transfers, audio
and video streaming to interactive applications and positioning tasks. These services
have different constraints concerning data rate, delay, and reliability. Hence future
mobile radio systems have to provide a large flexibility, and scalability to match these
heterogeneous requirements. The basic concept of MIMO (Multi-Input Multi-Output)
is that the transmitted signals from all transmit antennas are combined at each receiv-
ing antenna element in such a way to improve the Bit Error Rate (BER) performance
or the data rate of the transmission. Smart antennas, here are referred to adaptive
antennas with electrical tilt, beam width and azimuth control, which can follow
relatively slow-varying traffic patterns that can form beams aimed at particular users
or steer nulls to reduce interference.
Smart antenna technology is being considered for mobile platforms such as auto-
mobiles, cellular telephones and laptops. Operational experience with advanced net-
work systems that have utilized smart antennas for several years provides practical
experience applicable to the upgrade to HSPA (High Speed Packet Access) and LTE
(Long Term Evolution) wireless systems. Most 3G systems are arranged to operate
in 2 GHz frequency band. The 4G represents the next development stage of cellular
494 V.R. Rentapalli and Z.J. Khan
evolution beyond 3G, and offers an ideal basis and bandwidth to provide more
efficient cellular multicast services. To increase capacity, operators need to add cell
sites. Doubling the number of cell sites approximately doubles the network capacity
and the throughput per user and greatly improves the peak user and the aggregate
throughput. [1-4]
In this paper an outline of various smart antenna schemes for improving the capac-
ity and coverage of the emerging generations of wireless networks is given in addition
to downlink transmission modes in relation to practical antenna configurations. Smart
antennas with MIMO systems which improve Quality of Service (QoS), capacity &
link reliability, and spatial multiplexing are described in section 2. Sections 3 & 4
give the concept of minimizing interfering signals and enhancing signal quality by
using adaptive arrays and reconfigurable techniques. Performance evaluation is dis-
cussed in section 5, simulation results have been shown in 6, and finally section 7
provides conclusion for this paper.
2 Antennas for MIMO
MIMO can offer more capacity by adding more antennas and more sectors. To meet
the data rate and QoS requirements of future broadband cellular systems, their spec-
tral efficiency and link reliability should be considerably improved, which cannot be
realized by using traditional single-antenna communication techniques. To achieve
these goals, MIMO systems deploy multiple antennas at both ends of the wireless
link, exploit the extra spatial dimension, besides the time, frequency, and code dimen-
sions, which allows to significantly increase the spectral efficiency and improve the
link reliability.
Array processing techniques are expected to offer the capability of providing
significantly reduced costs per transmitted bit in multi-service, multi technology net-
works. Two main categories of antenna array processing techniques are, beam form-
ing for high element correlation and spatial multiplexing for low element correlation
environments. Beam forming techniques can be used to reduce the total transmitted
power while preserving the data rate. This in turn reduces the overall system interfer-
ence . With these advanced technologies, the capabilities of 3G and 4G systems can
be extended to provide customers with increased data throughput for mobile high
speed data applications.
Fig. 1. Multi column planar Array Architecture Fig. 2. Reconfigurable beam antenna
MIMO and Smart Antenna Technologies for 3G and 4G 495
2.1 MIMO Spatial Multiplexing
Spatial multiplexing allows transmitting different streams of data simultaneously on
the same resource block(s) by taking maximum use of the spatial dimension of the
radio channel. These data streams can belong to one single user (single user MIMO /
SU-MIMO) or to different users (multi user MIMO / MU-MIMO). Each receiving
antenna may receive the data streams from all transmit antennas, the number of data
streams that can be transmitted in parallel over the MIMO channel is given by min
{Nt, Nr} is limited by the rank of the matrix H. The capacity of MIMO system scales
linearly with receiving antennas at low SNR (Signal to Noise Ratio) and at high SNR.
Hence spatial multiplexing is used in LTE to achieve the peak data rates and it also
provides improvement in cell capacity and throughput. [5]
2.2 Transmit Diversity in LTE
Transmit diversity provides a source of diversity for averaging out the channel varia-
tion either for operation at higher (User Equipment) UE speeds (300Kmph) or for
delay sensitive services at both low (15 Kmph) and medium (15-120Kmph) UE
speeds. The increase in the number of parallel channels translates into an increase in
the achievable data rate within the same bandwidth. This increases the signal to noise
ratio at the receiver side. Instead of increasing data rate or capacity, MIMO can be
used to exploit diversity and increase the robustness of data transmission.
3 Adaptive Array Beam Forming
Reliable and high performance transmission continues to be a major goal of wireless
communication systems, which is significantly enhanced by arrays employing beam
forming and diversity techniques. An adaptive beam forming multi-column array an-
tenna can be considered as an advanced multiple antenna technique that will provide
an improvement in the overall communication link between the base station and mo-
bile. The basic architecture of an adaptive beam-forming antenna consists of multiple
columns of radiating elements that are driven by separate transceiver networks. [7]
The adaptive systems are really intelligent in the true sense and can actually be re-
ferred to as smart antennas. The smartness in these systems comes from the intelligent
digital processor that is incorporated in the system which can adjust or adapt its own
beam pattern in order to emphasize signals of interest and to minimize interfering
signals.
The beam forming array for the downlink case is to increase the signal strength in
the desired direction while reducing interference to the undesired directions. Whereas
for the uplink case, this array is to improve the receiver sensitivity in the direction of
the desired signal while reducing interference from the undesired directions if possi-
ble. For the beam forming antenna shown in figure 1 with 0.5 wavelength column
separation, the signals between adjacent columns would be highly correlated in low
angle spread environments. The two-column antenna with column separation on the
order of 1.2 wavelengths would be suitable for a 4X4 MIMO scheme. In this the sig-
nals between any two sets of cross polarized ports would not be highly correlated
even in low angle spread environments. [6]
496 V.R. Rentapalli and Z.J. Khan
Beam forming quality depends on the accuracy of the amplitude and phase values
of each MIMO transreceiver. Due to undesirable variations between each transmit and
receive path, some degree of errors will be formed in multi-column beam forming
antenna systems, resulting in significant beam forming degradation. This degraded
pattern gives undesirable side lobe levels shown in figure 5, squinting of the main
beam, degradation in gains as well as losing the ability to accurately position nulls.
By using some type of calibration networks, beam forming antennas can minimize
these errors. Typical beam forming systems deployed today require the amplitude
variations be limited to +/-0.5dB, while phase variations are limited to more than +/-5
degrees.
4 Reconfigurable Antennas
A modern telecommunication system which uses re-configurable antennas has the
ability to radiate more than one pattern at different frequencies. In a continuously
changing environment, re-configurable and adaptive techniques are used for adjusting
the structure and parameters of the transceivers to allow them to demonstrate the best
performance in a variety of the particular situations. MIMO receivers are capable of
reconfiguring themselves by switching automatically between a beam forming and a
spatial multiplexing. Reconfigurable beam antennas extend the range of remote beam
changes from a single dimension for elevation beam steering (Remote Electrical Tilt,
RET) to multiple dimensions. These antennas include the possibility to change the
bore sight or azimuth direction (panning), as well as the beam width of the antenna
(fanning) remotely. Reconfigurable beam antennas with tilting, panning and fanning
as shown in figure 2, can help balance the load between different cells, leading to a
combination of coverage, interference, and capacity improvements. Reconfigurable
beam antennas can thus significantly increase the basic network coverage.
5 Performance Evaluation
Spectral efficiency measures the ability of a wireless system to deliver a given amount
of billable services in a given amount of radio spectrum. The performance evaluation
of LTE (E-UTRA) and HSPA (UTRA) are being discussed. Simulation parameters
and assumptions follow guidelines provided in [8]. Case 1 represents interference-
limited small urban macro cell environment having carrier frequency 2GHz, the inter-
site distance 500 m, the bandwidth 10 MHz and UE speed 3Km/h. Case 2 represents
an inter-site distance of 1732 m. The cell radii for cases 1 and 2 are 288.7 and 1000m
respectively. Uplink performance for case 2 is slightly worse than for case 1 due to
the larger cell radius and hence coverage is limited for users towards the cell edge.
Uplink spectral efficiency in the LTE system provides more than two times improve-
ment related to HSPA baseline system. In the down link, LTE uses 2×2 SU-MIMO
which provides more than three times improvement in spectral efficiency over 1×2
baseline HSPA system. LTE downlink, 4×2, 4×4 MIMOs provides 10% and 58%
spectral efficiency gain related to 2×2 MIMO. It is also observed from the results that
the MIMO technologies introduced in LTE improves the cell-edge throughput and
spectrum efficiency.
MIMO and Smart Antenna Technologies for 3G and 4G 497
0
0.2
0.4
0.6
0.8
1
1.2
case 1 ca se 2
1×2 Baseline
1×2 LTE
1×4 LTE
Spectrum eff iciency
bps/Hz/cell
Upli nk
0
0.5
1
1.5
2
2.5
3
case 1 ca se 2
1×2 Baseline
2×2 LTE
2×4 LTE
4×4 LTE
Spectrum eff iciency
bps/Hz/cell
Down link
Fig. 3. Uplink and Downlink spectral efficiency
6 Results and Discussion
Simulation for eight element antenna array is performed on Matlab 7.0, using input
user signal at 0 degrees and three interferers at -60, -30 and 60 degrees. Calculating an
array factor for the array from -180 to 180 degrees, the response of every input user
signal and interfering signal is shown in figure 4. The number of elements in the beam
Fig. 4. Beam Pattern, User angle , and interference at angles 60,30,60
Fig. 5. Amplitude Responses, User angle at 0, Interferences at -60, 30, 60
498 V.R. Rentapalli and Z.J. Khan
forming array affect the complexity of the beam forming patterns. The eight column
antenna array demonstrates the ability to generate sharper high resolution beam with
more nulls due to selection of larger array of antenna elements. The response of major
lobe formation is obtained by using Least Mean Square and Recursive Mean Square
algorithms. This response has maximum signal strength in user direction as shown in
figure 5. An eight column array would result in a larger overall antenna, require eight
transreceivers and resulting in a significant increase in cost and complexity.
7 Conclusion
In this paper an overview of the benefits of most recent advances in smart antenna
transceiver architectures is given. The array processing techniques have been taken
into account to provide the basis for scalable, high data rate, high capacity system
solutions and spatial multiplexing gain on the receiving as well as the transmit sides.
A variety of smart antennas with remote controlled bore sight to beam-steering arrays
and beam forming MIMO have been reviewed. In this context the major trends in the
area of smart antennas, such as reconfigurability to varying channel propagation, an-
tennas for MIMO with diversity techniques have been discussed. The multi antennas
are used in a beam-forming transmission scheme to improve overall reception quality,
increase system capacity and extended coverage.
References
1. Tse, D., Viswanath, P.: Fundamentals of wireless communication. Cambridge University
Press, Cambridge (2005)
2. Khan, F.: LTE for 4G mobile broadband. Cambridge University Press, Cambridge (2009)
3. Peng, M., Wang, W.: Technologies and Standards for TD-SDMA Evolutions to IMT- Ad-
vanced. IEEE Communications Magazine 47(12), 50–57 (2009)
4. Greenspan, A., Klerer, M., Tomcik, J., Canchi, R., Wilson, J.: IEEE 802.20: Moblile Broad-
band Wireless Access for the 21st century. IEEE Communications Magazine 46(7), 58–59
(2008)
5. 3GPP, TS 36.211, Evolved Universal Terrestrial Radio Access (E-UTRA); Physical Chan-
nels and Modulation (Release 8)
6. 3GPP, TS 36.212, Evolved Universal Terrestrial Radio Access (E-UTRA); Multiplexing
and channel coding (Release 8)
7. Parkvall, S., Astely, D.: The Evolution of LTE towards IMT-Advanced Journal of Commu-
nications 4(3) (April 2009)
8. 3GPP RAN WG1 TR 25.814 V7.1.0, Physical Layer Aspects for Evolved UTRA (Release 7)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 499–506, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A GA-Artificial Neural Network Hybrid System for
Financial Time Series Forecasting
Binoy B. Nair, S. Gnana Sai, A.N. Naveen, A. Lakshmi,
G.S. Venkatesh, and V.P. Mohandas
Amrita Vishwa Vidyapeetham, P.O Ettimadai, Coimbatore, Tamilnadu, India, PIN-641105
b_binoy@cb.amrita.edu, sai.gnana@gmail.com,
naveen4053@gmail.com, lareekath@yahoo.in, gsvenky89@gmail.com,
vp_mohandas@amrita.edu
Abstract. Accurate prediction of financial time series, such as those generated
by stock markets, is a highly challenging task due to the highly nonlinear nature
of such series. A novel method of predicting the next day’s closing value of a
stock market is proposed and empirically validated in the present study. The
system uses an adaptive artificial neural network based system to predict the
next day’s closing value of a stock market index. The proposed system adapts
itself to the changing market dynamics with the help of genetic algorithm which
tunes the parameters of the neural network at the end of each trading session so
that best possible accuracy is obtained. The effectiveness of the proposed sys-
tem is established by testing on five international stock indices using ten differ-
ent performance measures.
Keywords: Genetic algorithm, artificial neural networks, financial, time series.
1 Introduction
Prediction of financial time series is an extremely challenging problem mainly due to
the fact that these series are inherently nonlinear in nature. Especially, in case of stock
markets, the task becomes still more difficult as the dynamics that govern the market
behavior are very hard to determine and quantify. The behavior of stock markets
depend on many factors such as political (eg. general elections, government policies),
economic (eg. Economic growth rate, unemployment data), natural factors (eg. natu-
ral disasters, rainfall), among many others. However, there have been many stud-
ies,for example,in [1] which indicate that stock markets do not follow a random walk,
as was suggested by [2] and that, it is in fact, possible to make profits in the stock
market. Two commonly used methods of analysis and prediction of stock markets are:
(a) fundamental analysis and (b) technical analysis. Fundamental analysis tried to
predict the behavior of a stock market based on the analysis of data such as macro-
economic indicators, national/international events etc. Technical analysis, on the other
hand, considers only the time-series generated by the stock-price/ market movement
to arrive at conclusions on the likely future trends.
500 B.B. Nair et al.
Soft computing based techniques for predicting stock markets are also now gaining
prominence [2]. Comparison of the effectiveness of time delay, recurrent and prob-
abilistic neural networks for prediction of stock trends based on historical data of the
daily closing price is done in [3].Combinations of technical indicators and soft com-
puting techniques have been used in [4] and [5] for predicting of stock exchanges.
Artificial neural networks in combination with other soft computing techniques have
also been used [6], [7]. In [8], an evaluation method for evolutionary computation
based stock trading systems is proposed. [9] presents the application of a novel multi-
ple-kernel learning algorithm for improved stock market forecasting using multiple-
kernel support vector regression, while in [10], a multiple objective particle swarm
optimization was used to simultaneously optimize two objective functions, namely,
the percentage profit and the Sharpe ratio. In [11] and [12], adaptive neuro-fuzzy
systems were used to predict the stock markets. However, it has been observed that
though soft computing techniques improve the prediction accuracy, selection of ap-
propriate parameters for the prediction algorithm is of critical importance. In the pre-
sent study, a GA-artificial neural network hybrid system for predicting one-day-ahead
closing values in stock markets is proposed. The proposed system can select the opti-
mum algorithm parameters by itself to ensure a high degree of prediction accuracy.
The effectiveness of the proposed system is validated on five stock markets, namely,
the BSE-Sensex, NSE-Nifty, FTSE-100, DJIA and Nikkei-225.
The rest of the paper is organized as follows: Section 2 presents the design of the
proposed hybrid system and section 3 presents the results and the conclusions.
2 Design of the Hybrid System
The stock market prediction system proposed in the present study has three major
parts. The first part involves the determination of the optimal lag from the given time
series and once the lag is determined, the determination of the optimal embedding
dimension. The vectors created from the time series data using the embedding dimen-
sion are then fed into the Elman neural network which generates the prediction for the
next day’s closing. Some of the major issues that confront the design of an artificial
neural network (ANN) based system are: determining the number of neurons in the
input layer, determining the number of neurons in the hidden layer and the selection
of ANN training parameters such as momentum. In the proposed system, the number
of input neurons is determined using the embedding dimension of the time series and
the number of hidden neurons, the learning parameters of the ANN are optimized
using genetic algorithm. The design of the proposed system is presented in the form
of a flowchart in Fig.1.
2.1 Determination of Embedding Dimension
The first step in the prediction process, as per the proposed hybrid system, involves
the identification of the embedding dimension. For this purpose, the optimum delay
value needs to be computed, which is accomplished in the present study using mutual
information. The value of delay for which the mutual information first reaches the
minimum value is taken as the delay[13]. Once the delay has been calculated, vectors
are created from the stock market data with each dimension being represented by the
A GA-Artificial Neural Network Hybrid System for Financial Time Series Forecasting 501
value at the obtained delay. For example, a three dimensional vector created in this
way will be made up of values at the instants: t, t- (delay) and t- (2*delay). Now the
optimum embedding dimension is calculated using the concept of false nearest
neighbors [14].
Fig. 1. Proposed hybrid system
2.2 Elman Neural Network Based Prediction
The Elman neural network (ENN) is one kind of global feedforward locally recurrent
network model proposed by [15]. An ENN can be considered to be an extension of
multilayer perceptron (MLP) with an additional input layer (state layer) that receives
a feedback copy of the activations from the hidden layer at the previous time step.
These context units in the state layer of the Elman network make it sensitive to
the history of input data. In the present study, the ENN is trained using the gradient
N
Y
Historical stock market data
Determine the delay
Identify the embedding dimension using
the delay
Initialize the ANN parameters
Predict next day’s closing value
Calculate training error
Train the ANN
Is training error
minimum?
Use GA to find new set
of ANN parameters
502 B.B. Nair et al.
descent with momentum backpropagation learning algorithm[16]. In the present
study, Elman neural network takes in the input, with the number of input neurons
being the same as the embedding dimension. The output of the system is the predicted
value for next day’s closing. The number of neurons in the hidden layer is optimised
using genetic algorithm.
2.3 Genetic Algorithms (GA)
GA has been very widely used for solving unconstrained optimization problems [17].
These are parallel search algorithms which can be used for optimizing nonlinear func-
tions, for example, see [4],[7]. Usually, the function which needs to be optimized,
called the objective function, takes the shape of a minimization problem. In the pre-
sent study, GA is used to minimize the prediction error (which is another way of say-
ing-maximizing accuracy) by optimizing the number of neurons in the hidden layer of
the ENN and the momentum parameter. Since the dynamics of stock markets change
over time, the number of neurons in the hidden layer is optimized at the end of each
trading session.
3 Experimental Results
3.1 Datasets
The proposed system was validated on the following five major stock markets,
namely, the Bombay stock exchange sensitive index-India (BSE – Sensex) , The na-
tional stock exchange- India(NSE-Nifty), FTSE 100- London , Nikkei 225- Japan and
the Dow Jones Industrial average-United States of America (DJIA). The period under
consideration was the stock market closing data from January 2005 to September
2010 for all the markets under consideration. First 90% of the data was used for train-
ing and the system was validated on the remaining 10%.
3.2 Optimized System Parameters
The hybrid system was trained using the proposed method. The optimum system
parameters obtained are presented in Table 1. The embedding dimension gives the
number of input neurons needed for the network. The optimal number of hidden neu-
rons and the momentum are obtained using GA.
Table 1. Hybrid system parameters obtained using the proposed system
Training
Parameters
BSE-Sensex NSE-Nifty FTSE 100 DJIA Nikkei 225
Hidden neurons 12 18 13 6 7
Embedding
dimension 10 38 32 22 19
Delay 26 15 22 13 13
Momentum 0.93 0.98 0.82 0.92 0.90
A GA-Artificial Neural Network Hybrid System for Financial Time Series Forecasting 503
3.3 Performance Evaluation Measures[18]
The most common measures used to evaluate the performance of forecasting methods
are presented in this section. Let Yt be the observation at time t and Ft denote the fore-
cast of Yt. Then, the forecast error is et =Yt-Ft and the percentage error is pt =100et/Yt.
Another method is to divide each error, by the error obtained with another standard
method of forecasting (in this paper, the standard method or the base method consid-
ered is the random walk). Let rt =et/et* denote the relative error, where et* is the
forecast error obtained from the base method. Usually, the base method is the “naïve”
method or the random walk model. The notation mean (xt) is used to denote the sam-
ple mean of {xt} over the period of interest (or over the series of interest). Similarly,
median (xt) is used for the sample median and gmean(xt) for the geometric mean. The
most commonly used methods are given in Table 2 [18], where the subscript b refers
to measures obtained from the base (random walk) method.
The results obtained from the Makridakis competitions (also known as M Competi-
tions, M2-Competitions and M3- competitions) held to test forecasting accuracy of
various forecasting methods form the basis of accuracy measures listed in Table 2.In
the first M-competition [19], measures used included the MAPE, Mean squared error
(MSE), MdAPE, and percentage better (PB). However, the MSE is not appropriate for
comparisons between series as it is scale dependent. The MAPE also has problems
when the series has values close to (or equal to) zero [20]. MAPE, MdAPE, PB,
GMRAE, and MdRAE have also been used as performance measures [21]. The M3-
competition [22] used three different measures of accuracy: MdRAE, sMAPE, and
sMdAPE. The “symmetric” measures [23] were proposed in response to the observa-
tion that the MAPE and MdAPE have the disadvantage that they put a heavier penalty
on positive errors than on negative errors. Hence, in the present study, the perform-
ance of the proposed system is evaluated using ten different performance measures to
ensure the validity of results.
Table 2. Forecast performance measures
Measure Description Expression
MAPE Mean absolute percentage error mean(|pt|)
MdAPE Median absolute percentage error median(|pt|)
sMAPE Symmetric mean absolute percentage
error
mean(2|Yt-Ft|/( Yt+Ft))
sMdAPE Symmetric median absolute percent-
age error
median(2|Yt -Ft|/(Yt +Ft))
MRAE Mean relative absolute error mean(|rt|)
MdRAE Median relative absolute error median(|rt|)
GMRAE Geometric mean relative absolute
error gmean(|rt|)
RelMAE Relative mean absolute error MAE/MAEb
RelRMSE Relative root mean squared error RMSE/RMSEb
LMR Log mean squared error ratio log(RelRMSE)
504 B.B. Nair et al.
The results obtained using the proposed hybrid system are presented in Table 3.
Table 3. Results for the proposed hybrid system
Performance measure BSE-
Sensex
NSE-Nifty FTSE 100 DJIA Nikkei 225
MAPE (%) 8.22 6.17 5.26 3.98 3.86
MdAPE(%) 5.72 4.28 3.92 3.83 3.19
sMAPE (%) 0.08 0.06 0.05 0.04 0.04
sMdAPE (%) 0.08 0.06 0.05 0.04 0.04
MRAE 26.31 24.98 27.35 23.35 10.19
GMRAE 11.25 7.24 7.57 8.14 3.82
Md RAE 10.46 6.28 10.69 7.57 3.80
Rel. MAE 10.22 7.21 7.67 5.16 3.39
Rel. RMSE 9.55 7.32 7.47 4.28 3.30
LMR 4.51 3.98 4.05 2.91 2.39
The results obtained using the hybrid system are compared to those obtained using
a stand-alone ENN with four hidden neurons and trained using a gradient descent with
momentum backpropagation algorithm with constant momentum of 0.9. An adaptive
learning rate is used, as in the case of the hybrid system,the initial value being
0.1.These parameters are randomly selected.The results obtained are given in Table 4.
Table 4. Results for stand-alone Elman neural network
Performance
measure
BSE-Sensex NSE-Nifty FTSE
100
DJIA Nikkei 225
MAPE (%) 12.32 16.93 14.24 10.11 6.90
Md APE(%) 11.77 17.02 15.05 10.79 6.49
sMAPE(%) 0.14 0.19 0.14 0.11 0.07
sMdAPE(%) 0.14 0.19 0.14 0.11 0.07
MRAE 49.63 65.48 65.02 76.67 19.53
GMRAE 21.07 25.41 24.42 20.53 8.36
Md RAE 20.88 28.41 20.30 25.03 8.56
Rel. MAE 16.67 20.36 20.71 14.40 6.06
Rel. RMSE 14.63 16.93 18.71 10.92 5.20
LMR 5.37 5.66 5.86 4.78 3.30
From the tables 3 and 4 it is clear that the proposed hybrid system shows a signifi-
cant improvement in performance when compare to a standalone ENN. Hence, it can
be said that the proposed hybrid system is well capable of predicting stock markets
with a high degree of accuracy.
References
1. Atsalakis, G.S., Valavanis, K.P.: Surveying Stock Market Forecasting Techniques – Part
II: Soft Computing Methods. Expert Systems with Applications 36, 5932–5941 (2009)
2. Fama, E.F.: Efficient Capital Markets: A Review of Theory and Empirical Work. Journal
of Finance 25, 383–417 (1970)
A GA-Artificial Neural Network Hybrid System for Financial Time Series Forecasting 505
3. Saad, E.W., Prokhorov, D.V., Wunsch, D.C.: Comparative Study of Stock Trend Predic-
tion using Time Delay, Recurrent and Probabilistic Neural Networks. IEEE Transactions
on Neural Networks 9(6), 1456–1470 (1998)
4. Nair, B.B., Mohandas, V.P., Sakthivel, N.R.: A Genetic Algorithm Optimized Decision
Tree-SVM based Stock Market Trend Prediction System. International Journal on Com-
puter Science and Engineering 2(9), 2981–2988 (2010)
5. Nair, B.B., Mohandas, V.P., Sakthivel, N.R.: A Decision Tree- Rough Set Hybrid System
for Stock Market Trend Prediction. International Journal of Computer Applications 6(9),
1–6 (2010)
6. de Faria, E.L., Albuquerque, M.P., Gonzalez, J.L., Cavalcante, J.T.P., Albuquerque, M.P.:
Predicting the Brazilian Stock Market Through Neural Networks and Adaptive, exponen-
tial smoothing methods. Expert Systems with Applications 36(10), 12506–12509 (2009)
7. Kuo, R.J., Chen, C.H., Hwang, Y.C.: An Intelligent Stock Trading Decision Support Sys-
tem through Integration of Genetic Algorithm based Fuzzy Neural Network and Artificial
Neural Network. Fuzzy Sets and Systems 118, 21–45 (2001)
8. Yeh, I.-C., Lien, C.-H., Tsai, Y.-C.: Evaluation Approach to Stock Trading System using
Evolutionary Computation. Expert Systems with Applications 38(1), 794–803 (2011)
9. Yeh, C.-Y., Huang, C.-W., Lee, S.-J.: A Multiple-Kernel Support Vector Regression Ap-
proach for Stock Market Price Forecasting. Expert Systems with Applications 38(3),
2177–2186 (2011)
10. Briza, A.C., Naval Jr., P.C.: Stock Trading System based on the Multi-Objective Particle
Swarm Optimization of Technical Indicators on End-of-Day Market Data. Applied Soft
Computing 11(1), 1191–1201 (2011)
11. Gholamreza, J., Tehrani, R., Hosseinpour, D., Gholipour, R., Shadkam, S.A.S.: Applica-
tion of Fuzzy-Neural Networks in Multi-Ahead Forecast of Stock Price. African Journal of
Business Management 4(6), 903–914 (2010)
12. Atsalakis, G.S., Valavanis, K.P.: Forecasting Stock Market Short-Term Trends using a
Neuro-Fuzzy based Methodology. Expert Systems with Applications 36, 10696–10707
(2009)
13. Fraser, A.M., Swinney, H.L.: Independent Coordinates for Strange Attractors from Mutual
Information. Phys. Rev. A 33, 1134–1140 (1986)
14. Kennel, M., Brown, R., Abarbanel, H.: Determining Embedding Dimension for Phase-
Space Reconstruction using a Geometrical Construction. Phys. Rev. A 45, 3403–3411
(1992)
15. Elman, J.L.: Finding Structure in Time. Cognitive Science 14, 179–211 (1990)
16. Han, J., Kamber, M.: Data Mining:Concepts and Techniques. Morgan Kaufmann, San
Mateo (2006)
17. Goldberg, D.E.: Genetic Algorithms in Search, Optimization & Machine Learning. Addi-
son-Wesley, Reading (1989)
18. De Gooijer, J.G., Hyndman, R.J.: 25 Years of Time Series Forecasting. International Jour-
nal of Forecasting 22, 443–473 (2006)
19. Makridakis, S., Andersen, A., Carbone, R., Fildes, R., Hibon, M., Lewandowski, R.: The
Accuracy of Extrapolation (Time Series) Methods: Results of a Forecasting Competition.
Journal of Forecasting 1, 111–153 (1982)
20. Makridakis, S., Wheelwright, S.C., Hyndman, R.J.: Forecasting: Methods and Applica-
tions. John Wiley and Sons, New York (1998)
506 B.B. Nair et al.
21. Fildes, R., Hibon, M., Makridakis, S., Meade, N.: Generalising about Univariate Forecast-
ing Methods: Further Empirical Evidence. International Journal of Forecasting 14,
339–358 (1998)
22. Makridakis, S., Hibon, M.: The M3-competition: Results, Conclusions and Implications.
International Journal of Forecasting 16, 451–476 (2000)
23. Makridakis, S.: Accuracy measures: Theoretical and Practical Concerns. International
Journal of Forecasting 9, 527–529 (1993)
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 507–512, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Preemptive View Change for Fault Tolerant
Agreement Using Single Message Propagation
Poonam Saini and Awadhesh Kumar Singh
Department of Computer Engineering, National Institute of Technology,
Kurukshetra, 136119, India
nit.sainipoonam@gmail.com, aksinreck@rediffmail.com
Abstract. The paper presents a proactive approach for failure detection. In our
previous work [5], we assumed a trustworthy Transaction Manager, TM, ame-
nable to the job of view creation, detection of faulty primary as well as backup
replicas and to evacuate them from the transaction processing system. In the
end, TM initiates a view in case it detects faulty primary or faulty replica. The
TM provides an efficient failure-resiliency in the protocol; however, it also
introduces the possibility of single-point failure. To eliminate the reliance on
single TM, we propose a protocol that distributes the responsibilities of a trans-
action manager among 3f+1(f are faulty) replicas and results in a distributed
Transaction Manager (DTM). The article attempts to limit the failure detection
time to an optimum value, i.e., single message propagation time between any
two nodes.
Keywords: Proactive View Change, Byzantine Agreement, Distributed Trans-
action Manager.
1 Introduction
The transaction processing poses many challenges in the area of distributed comput-
ing. In comparison with other applications, fault tolerance is a more serious concern
in transaction systems where the behavior of the interacting nodes may be arbitrary.
Moreover, with reference to agreement, the transaction handling protocol should
maintain atomicity, i.e., either the operation is to be committed or aborted. Traditional
Byzantine fault tolerant protocols e.g., BFTDC [1, 2] and Zyzzyva [3], deal with
failures in a reactive manner, i.e., they rely on the specification of the faults to initiate
view change. In a particular view, one of the replicas is chosen as primary and other
replicas as backups. In the middle of agreement, if time out occurs for current view,
due to delay in message propagation or the primary is found faulty, the view change
occurs. The proactive/preemptive approach, on contrary, is designed to minimize the
transaction discontinuity and latency while ensuring stability as well as availability of
replicas through failure notifications in advance. Towards this goal, we build a system
model to analyze the failure resiliency of our protocol under both reactive and proac-
tive approaches. An optimized agreement is run among the replicas to reach the final
decision about the transaction.
508 P. Saini and A.K. Singh
Fig. 1. Schematic view of DTM
Motivation. In most of the contemporary works, the protocols, as described in [4],
replace a replica from the system when it is diagnosed as faulty. Although, the proto-
cols produce desired result, they incur latency in order to initiate the view change
which results in short-lived (i.e., transient) halts. We have devised a technique which
is able to detect, in advance, the tentative fault in the system. The protocol fulfils all
the requirements that are agreement, validity, and termination. Our previous protocol
[5] relies on a trusted and fault-free Transaction Manager, TM. Although, it served as
an efficient oracle for the proactive view change, the dependency is very high on TM.
The proposed DTM (Fig 1) model intends to eliminate it and distributes the jobs of
TM among the 3f+1 replica’s where f is the number of faulty one.
2 System Model
We assume a 2-tier architecture where the clients are not replicated as the Byzantine
faults are considered only at the coordinator site. There are 3f +1 coordinator replicas,
among which at most f can be faulty during a transaction. . The protocol is started for
a transaction whenever a commit request is received from the client. Each coordinator
replica is assigned a unique id i, where i varies from 0 to 3f. The id is required firstly,
to identify the primary (P), and deputy primary (DP), from the replicas and secondly,
to detect the fault in P, DP and replicas, if any. The deputy primary is chosen along
with primary to avoid immediate transaction discontinuity, in case, the primary be-
haves arbitrarily. DP takes over the charge as soon as primary is diagnosed faulty.
A Preemptive View Change for Fault Tolerant Agreement 509
3 Data Structures and Message Types
The protocol uses the following data structures and messages types:
Data Structure
i.
At all nodes
a. n_id: Node ID.
b. n_wt: Node weight randomly assigned, 0 < n_wt < 1.
c. n_tb: A toggle bit {0, 1} i.e., a priority associated with each node to nominate
itself as primary and deputy primary.
d. p: A replica which acts as primary/coordinator for the transaction.
e. dp: A replica nominated as deputy primary to set back as primary in the event
of primary being declared faulty.
f. r: Backup replicas in the system.
ii.
At primary, p node
a. stp.r: A status token, initially, false (0), to verify the current state of replicas.
b. stp.dp: A status token, initially set to true (1), to validate the existence of dep-
uty primary, dp.
c. rec_statusp [r]: An array of tuples declaring the status of all replicas.
iii.
At deputy primary, dp node
a. stdp.p: A status token to validate the existence of primary, p.
b. tp: A tentative primary status field, initially passive, to be set to active as soon
as the value of stp turns true (1) and primary is declared faulty.
Message types
i.
At all nodes
a. View Message (VM): It is a view broadcast message to inform all replicas and
participants about the primary, p, deputy primary, dp and backup replicas, r.
b. Agreement Message (AM): It is the agreement message between the replicas
to decide on a common value for a particular query.
c. Decision Message (DM): It is the decision message to broadcast the final out-
come of the transaction to the intended clients.
4 The Approach
4.1 The View Formation
In a transaction system, replicas move through a sequence of configurations which is
termed as view (Fig 2). All the replicas enter into the transaction processing system
with weighted information, n_wt. A toggle bit n_tb with value 0 or 1 is attached with
all the replicas. In the next step, a primary, deputy primary and backup replicas are
chosen as follows:
a.
For primary p: {n_wt (p) ≥ 0.5 and n_tb(p) = 1}.
b.
For deputy primary dp: {n_wt (dp) ≥ 0.5, n_tb (dp) = 1, and n_id(p) ≠ n_id(dp)}.
c.
For replicas r: {n_wt (r) < 0.5 and n_tb(r) = 0}.
510 P. Saini and A.K. Singh
Fig. 2. The Semantic View Architecture
A view formation ends with the selection of all the above said entities. The current view
message, which contains the current view number v, primary p, deputy primary dp and
transaction id i, is then broadcast to each participating replica. Finally, if 3f replicas
respond with an acknowledgement of current view, the primary begins the agreement
round. In a particular view, the selection of a deputy primary ensures the pause-free
processing among the replicas. However, in case if both, p and dp are fail-silent, a view
change is carried out in order to provide liveness to the transaction system.
4.2 Proactive Fault Detection
Both, primary and deputy primary check each other for their correctness through a
single-bit status token passing mechanism. The formal description of working is given
as follows:
a.
A status token, initialized as 1 by primary, is rotated among both the entities.
b.
The value of status token (stp.dp) is stored in the array of primary and passed to
the deputy primary.
c.
Now, the deputy primary will complement the received value of the status token
and stores the complemented value in status token (stdp.p) and returns it back to
primary.
d.
Afterwards, the same steps are repeated at primary. In this way, the status token
circulates between primary and deputy primary.
e.
The process continues and the protocol for proactive fault detection runs simulta-
neously with the agreement protocol.
f.
If the value of status token stored at primary (deputy primary) and the value of
status token received from deputy primary (primary) are same, then it signifies
that deputy primary (primary) is declared faulty.
In the similar manner, the backup replicas are also continuously audited by the pri-
mary as follows:
a.
A different status token generated by primary with value 0 is circulated among all
the backup replicas in the system.
A Preemptive View Change for Fault Tolerant Agreement 511
b.
Primary declares a backup replica Byzantine faulty, whenever the backup replica
returns a value of status token other than 0 e.g., it returns token value 1.
c.
In case, if no value is returned from the backup replica, the primary would de-
clare it fail-silent or crashed.
This mechanism results in the preemptive detection of faulty primary, deputy primary,
or replica, if exist, in the transaction processing system. Following the proactive de-
tection mechanism of tentative failures into the system, an optimized Byzantine
agreement protocol is executed simultaneously.
5 Correctness Proof
Lemma 1. The protocol maintains liveness. Formally,
(stp.dp = 1 ∧ stdp.p = 1) ∨ (stp.dp = -1∧ stdp.p = -1) tp.active
Argument: Assume that at some instant t, during transaction processing, the value of
the status token (stdp.p) in the array of deputy primary, dp complies with the last re-
turned value of the status token by the primary p, i.e.,
[dp (stdp.p) ∧ p(stp.dp) = 1 ∨ -1]
Then p is declared faulty by dp. Formally,
(stp.dp = 1 ∧ stdp.p = 1) ∨ (stp.dp = -1∧ stdp.p = -1) (p)
Note: (p) denotes the faulty primary.
Now, according to the operational semantics of the protocol, the deputy primary takes
over as primary, p and broadcast a view change message to all active replicas. The
tentative primary field tp is set to active and status of dp is updated as primary p.
Lemma 2. The faulty replicas are exposed by non-faulty primary, eventually.
Formally,
p(rec_statusp [ri]) = (⊥ ∨ 1) (ri)
Note: (ri) denotes the faulty replicas, where the faulty replicas involves both Byzan-
tine and fail-silent.
Argument: Assume that a replica r receives a status token with value 0 from primary
p. Now, there can be two possibilities:
Case 1: The modified value of token is returned, i.e.,
r sets (stp.r) = {1} and broadcast to p.
Case 2: No value is returned, i.e.,
r sets (stp.r) = { } (a null value)
Now, after the first round of transaction, the primary verifies for the status token array
received from the replicas. The modified value in the array or an empty field in the
array would confirm the presence of Byzantine faulty and fail-silent replicas in the
system. Formally,
rec_statusp ϵ {(⊥ ∨ 1} (ri)
512 P. Saini and A.K. Singh
Lemma 3. All the non-faulty replicas will eventually reach the same view in which the
replica to be chosen as primary, initially, is not faulty. Formally,
∀
r ϵ { p, dp ,r}, p ≠ (p)
Argument: Assume the contrary. Let x and t to be primary for view v+1 and v+2 re-
spectively. Now, if x is primary for view v+1, it will send new view message if it has
received 2f+1 view change messages for view v+1. Let, this set is R1. Similarly, t will
send new view message if it has received view change message for view v+2 from
2f+1 replicas. Let, this set is R2. Now, between R1 and R2 there are at least f+1 com-
mon replica’s because there can be at most f faults in the system at a time.
Note: In the worst case, out of f+1 replicas, at most, f replicas could be faulty.
Thus, at least one correct replica is still present. Hence, the correct replica would
not send new view change message for both the replicas to be selected as primary in
their respective views. Hence the lemma holds.
6 Conclusion
The paper proposed an optimized and novel proactive mechanism towards the proac-
tive detection of faulty replicas. The distribution of responsibilities of the Transaction
Manager (TM) among the replicas has removed the dependency on it. The protocol
prompts to evacuate both, the Byzantine as well as fail-silent i.e., a crashed replica.
The static analysis of the protocol verifies the significance of protocol under both the
faults. For future work, the extension of the boundaries for arbitrary behavior of the
nodes is under revision.
References
1. Castro, M., Liskov, B.: Practical Byzantine Fault Tolerance and Proactive Recovery. ACM
Transactions on Computing Systems 20, 398–461 (2002), doi:10.1145/571637.571640
2. Zhao, W.: A Byzantine Fault Tolerant Distributed Commit Protocol. In: IEEE International
Symposium on Dependable, Autonomic and Secure Computing, pp. 37–44 (Septemper
2007)
3. Kotla, R., Alvisi, L., Dahlin, M., Clement, A., Wong, E.: Zyzzyva: Speculative Byzantine
Fault Tolerance. ACM Proceedings of twenty-first Symposium on Operating Systems and
Principles 41(6), 45–48 (2007)
4. Fisman, D., Kupferman, O., Lustig, Y.: On verifying fault tolerance of distributed protocols.
In: Ramakrishnan, C.R., Rehof, J. (eds.) TACAS 2008. LNCS, vol. 4963, pp. 315–331.
Springer, Heidelberg (2008)
5. Saini, P., Singh, A.K.: An Efficient Byzantine Fault Tolerant Agreement. In: ICM2ST
2010: Proceedings of the International Conference on Methods and Models in Science and
Technology, American Institute of Physics (AIP), December 2010, vol. 1324, pp. 162–165
(2010), doi:10.1063/1.3526183, ISBN: 978-0-7354-0879-1
V.V. Das, G. Thomas, and F. Lumban Gaol (Eds.): AIM 2011, CCIS 147, pp. 513–515, 2011.
© Springer-Verlag Berlin Heidelberg 2011
A Model for Detection, Classification and Identification
of Spam Mails Using Decision Tree Algorithm
Hemant Pandey1, Bhasker Pant2, and Kumud Pant3,*
1 Senior Infrastructure Engineer, EDS, Pune, Maharashtra, India
helpmenu@gmail.com
2 Assistant Professor, Department of IT, Graphic Era University, Dehradun, India
pantbhaskar2@gmail.com
3 Senior Research fellow, Department of Bioinformatics, MANIT, Bhopal, India
pant.kumud@gmail.com
Abstract. Spam mails are unsolicited bulk mails which are meant to fulfill
some malicious purpose of the sender. They may cause economical, emotional
and time losses to the recipients. Hence there is a need to understand their char-
acteristics and distinguish them from normal in box mails. Decision tree classi-
fier has been trained with the major characteristics of spam mails and results
obtained with more then 86.7437% accuracy. This classifier can be a valuable
strategy for software developers who are trying to combat this ever growing
problem.
Keywords: spam mails, header characteristics, message body, classifier/
predictor.
1 Introduction
E-mail spam, known as unsolicited bulk Email (UBE), junk mail, or unsolicited
commercial email (UCE), is the practice of sending unwanted e-mail messages, fre-
quently with commercial content, in large quantities to an indiscriminate set of recipi-
ents. Everyday 1000’s and lack’s of spam mails are send by unknown senders through
out the world. At the first instance spam mails may seem harmless unwanted mails
but some of the innocent users may fall prey to the ill intentions encrypted within it.
The infiltration of ones mail box with these mails is not only annoying but it is also
dangerous. It has the potential to cause great financial losses to the administration
which deals with it and to the common men who may fall in this spam trap that has
been laid for them. They are usually loaded with viruses, spy programs and soft-
ware’s. Realizing the importance of spam mails and problems caused by it, here we
present a comparative account of random forest and decision tree algorithm for classi-
fication, filtering and prediction of spam mails. Previously the header characteristics
of these mails have been used for their classification and identification [4]. The same
have been used here with a comparative account of decision tree and random forest
* Corresponding author.
514 H. Pandey, B. Pant, and K. Pant
method. Hence this is a novel step to make general people aware of these silent mon-
sters and also for software developers to device new hybrid models and approaches
using decision tree classifier and random forest algorithm for increasing efficiency of
their spam trapping ability.
2 Materials and Methods
2.1 Data Set
We collected around 61 spam mails and 37 normal mails from our gmail accounts.
They were checked for various attributes and a data sheet was made for its tabulation.
2.2 The Algorithm
For analyzing our data we used two algorithms from weak.
1) Decision tree and its boosting algorithm J48 from Weka and 2) Random forest
algorithm [1-3]. Here using both the above classifiers a model is developed to classify
spam and normal in-box mails by using their header characteristics. After data filter-
ing, each classifier is trained and cross-validated for 10-times with a 10-fold random
sampling. The ten resulting values for each performance parameter are averaged to
obtain the final figures, and Receiver Operating Characteristic (ROC) curves and TP
rates vs. FP rates are plotted and analyzed [4-7]. In the past decision tree algorithm
has been used for classification of plant and animal micro RNA using their various
attributes and Random Forest for classification of MMP’s hence are quiet powerful in
classification of both supervised and unsupervised data [8, 9].
3 Results
The comparative results of both classifiers are depicted in table 1. It was seen that
characteristic A was the most important criteria in distinguishing spam and normal
mails. On removing which the accuracy was reduced.
The decision tree is shown in figure1.
Fig. 1. J 48 Deision tree with A (recipients address not in to: or cc:field) as the principal field
A Model for Detection, Classification and Identification of Spam Mails 515
Table 1. Comparative results of both the classifiers
With Decision tree With Random Forest
Al characteristics included Characteristic A not
included
Al l characteristics included Characteristic A not included
Correctly
classified
instances
85
(86.7347%)
Correctly
classified
instances
79
(80.6122%)
Correctly
classified
instances
82
(83.6735%)
Correctly
classified
instances
78
(79.5918%)
Incorrectly
classified
instances
13
(13.2653%)
Incorrectly
classified
instances
19
(19.3878%)
Incorrectly
Classified
Instances
16
(16.3265%)
Incorrectly
Classified
Instances
20
(20.4082%)
Kappa statistics 0.7132 Kappa
statistics
0.5897 Kappa statistic 0.6489 Kappa statistic 0.5658
Mean absolute
error
0.2313 Mean absolute
error
0.2956 Mean absolute
error
0.2133 Mean absolute
error
0.2618
Root mean
squared error
0.3438 Root mean
squared error
0.4092 Root mean
squared error
0.3598 Root mean
squared error
0.4254
Relative absolute
error
49.1191% Relative
absolute error
62.7718% Relative
absolute error
45.3011% Relative
absolute error
55.6005%
Root relative
squared error
70.8651% Root relative
squared error
84.3524% Root relative
squared error
74.1698% Root relative
squared error
87.6859%
Total number of
instances
98 Total number
of instances
98 Total Number
of Instances
98 Total Nu mber
of Instances
98
4 Conclusion
Besides header files, other criteria can be very important in distinguishing spam from
normal mails like subject line, sender domain, message content etc. The author also
wishes to extend the above classifier by incorporating the above characteristic so as to
further improve the classifier.
References
1. Witten, I.H., Frank, E.: Data Mining – Practical machine learning tools and techniques with
Java implementations. Morgan Kaufmann, San Francisco (2005)
2. Langley, P., Sage, S.: Elements of machine learning. Morgan Kaufmann, San Fracisco
(1994)
3. Weka Data Mining Java Software, http://www.cs.waikato.ac.nz/~ml/weka/
4. Mike Spykerman – CEO Red Earth Software,Typical spam characteristics How to effec-
tively block spam and junk mail
5. Langley, P., Sage, S.: Elements of Machine Learning. Morgan Kaufmann, San Fracisco
(1994)
6. Han, J., Kamber, M.: Data Mining:Concepts and Techniques. Morgan Kaufmann, San Fran-
cisco (2001)
7. Heckerman, D., Geiger, D., Chickering, D.M.: Learning Bayesian network: The combina-
tion of knowledge and statistical data. Machine Learning 20(3), 197–243 (1995)
8. Pant, B., Pant, K., Pardasani, K.R.: ‘Decision tree cassifier for classiification of plant and
animal micro RNA’s. Communications in Computer and Information Science, Part 9 51,
443–451 (2009), doi:10.1007/978-3-642-04962-0_51.
9. Pant, B., Pant, K., Pardasani, K.R.: Machine Learning Model for Domain Based Classifica-
tion of MMP’s. The Internet Journal of Genomics and Proteomics 5(2) (2010)
Author Index
Acharya, Arup Abhinna 100
Agarkar, Ankita 222
Agarwal, Ajay 255, 368
Ajitha, P. 489
Ambeth Kumar, V.D. 214, 358
Analoui, Morteza 461
Anand, R. 468
AntopremKumar 485
Aravindan, C. 312, 323, 427
Arya, Meenakshi 123
Babu Rao, M. 399
Bagchi, Parama 287
Balaji, Anerudh 291
Balamurugan, Karthigha 297
Bande, Shivangi 82
Bandhopadhyay, T.K. 344, 348
Banerjee, Soumik 39
Bansal, Sanjay 445
Bella Mary, I. Thusnavis 481
Bhalani, Jaymin 184
Bhuvana, J. 312
Bose, Susan Mary 477
Budyal, V.R. 375
Burse, Kavita 67
Chakroborty, Debashish 39
Chana, Inderveer 449
Chandra, E. 489
Chaudhary, Shubhangi R. 440
Chitra, K. 196
Choubey, Abha 422
Choubey, Siddhartha 422
Chougule, Archana 190
Danda, Aneeshwar R. 273
Daniel 485
Dash, Gananath 170
david Raj, Newton 485
Deenadayalan, T. 468
Deshmukh, Manjusha 55
Devashrayee, N.M. 33, 95
Dey, Kashinath 305
Dhayanandh, S. 178
Dilli, Ravilla 240
Dinesh, P.M. 263
D’sa, Kiran 273
Dubey, Ashutosh K. 144
Dubey, Nilesh 82
Dubey, Vandana 82
Fatima, Mehajabeen 348
Gahankari, Sonal 55
Ganda, Geetika 379
Govardhan, A. 19, 399
Grace Mary Kanaga, E. 317
Gupta, B.B. 117
Gupta, Roopam 344, 348
Haider, Mohammad 150
Halkarnikar, Pratap 190
Hanumantharaju, M.C. 162
Hari, CH.V.M.K. 227, 281
Hima Bindu, M. 407
Hiremath, S.G. 375
Jain, A. 117
Jain, Jyoti 344
Jaisakthi, S.M. 427
Jaisankar, N. 454
Jayakumar, M. 297
Jinwala, Devesh C. 388
Joshi, Apoorv 222
Joshi, R.C. 117
Josphineleela, R. 352
Josprakash 485
Juyal, S. 117
Kannan, A. 454
Kaur, Gunjit 435
Kaur, Pankaj Deep 449
Kaushal, B.S.S. 281
Kavitha, Ch. 399
Khan, Zafer Jawed 493
Kirar, Vishnu Pratap Singh 67
Kole, Arnab 305
Korde, Mridula 74
Koshy, Divya Mary 477
Kosta, Yogesh 184
518 Author Index
Kosta, Yogeshwar 111
Krishnam Raju, K.V. 267
Krishnan, Shobha 248
Krishna Prasad, A.V. 301
Kumar, A. 392
Kumar, Abhishek 336
Kumar, Neeraj 259, 384
Kumar, Niti 245
Kumar, P. Pavan 403
Kumar, Sumit 156
Kumar, Vimal 132
Kushwaha, Ganesh Raj 144
Lakshmi, A. 499
Lavanya, N. 178
Maheshwari, Manish 384
Maity, Saikat 277
Majumder, Saikat 207
Makhijani, Jagdish 6
Mala, C. 138, 431
Manohara Pai, M.M. 1, 340
Manoria, Manish 67
Manvi, S.S. 375
Meher, Jayakishan K. 170
Meher, Pramod Kumar 170
Mehta, Mayuri A. 388
Mishra, Nibedita 170
Misra, M. 117
Mittal, Shaily 379
Mohanavalli, S. 427
Mohandas, V.P. 499
Mohapatra, Durga P. 100
Mohapatra, Pranab Kishor 170
Mondal, Hemanta 245
Motwani, Mahesh 6
Mukherjee, Amartya 291
Mukherjee, Debasis 245
Murali Nath, R.S. 240
Murugesan, D. 61
Naik, Amisha 95
Nair, Binoy B. 499
Nandi, Subrata 291
Naveen, A.N. 499
Nayak, Rakesh 27
Nikhil 384
Niranjan, Manoj Kumar 6
Nuparam 45
Oza, Shruti 33
Padmaja Rani, B. 301
Padmavathi, G. 196
Pais, Alwyn R. 336
Pal, Shantanu 287
Pandey, Hemant 513
Pant, Bhasker 513
Pant, Kumud 513
Patel, Sanket 111, 184
Patel, Shobhit 111, 184
Patnaik, Debashree 100
Patnaik, L.M. 273
Patra, Sushovan 291
Paul, Biju 332
Prabhakar, R. 117
Prabhakara Rao, B. 399
Prachi 379
Pradhan, Jayaram 27
Prasadh, K. 332
Prasad Reddy, P.V.G.D. 227
Puranik, Minal M. 248
Raamesh, Lilly 327
Rahamatkar, Surendra 255
Rahmatkar, S. 368
Raj, Adhira 297
Raja, K.B. 13, 273
Rajagopalan, Narendran 431
Rajalakshmi, K. 407
Rajalakshmi, R. 323
Rajesh Kumar, G. 233
Rajkumar, S. 178
Rajpurohit, Vijay S. 1, 340
Rajput, Anil 6
Ramachandran, Nitya 417
Ramachandran, S. 162
Ramakrishan, M. 214, 358
Rama Krishna, E. 233
Ramakrishna, S. 301
Ramakrishnan, M. 352
Ramamohanreddy, A. 19
Rama Rao, T. 61
Ramasubbareddy, B. 19
Ramasubramaniam, N. 403
Rambabu, N. 233
Ramesh, S. 61
Ramesha, K. 13
Rameshbabu, D.R. 162
Ramya Sri, A.P. 178
Author Index 519
Ranjeet 392
Rathkanthiwar, Anagha 74
Raval, Mukesh Kumar 170
Ravishankar, M. 162
Reddy, B.V.R. 245
Rentapalli, Vanitha Rani 493
Reshma, P. 203
Rezvani, Mohammad Hossein 461
Rini Rosebell, V. Jerine 473
Sabeenian, R.S. 263
Sachan, A.K. 6
Sadalkar, Kunal M. 336
Saha, Sujay 305
Saha, Sujoy 291
Sai, S. Gnana 499
Saini, Poonam 507
Sairam, Ashok Singh 156
Sastry, C.V. 27
Sathisha, N. 273
Sen, Praveen 255
Sengar, Sandeep Singh 132
Sethi, Tegjyot Singh 281
Shanker, Udai 45
Sharanya, R. 477
Sharma, Abhishek 281
Sharma, Sanjeev 445
Sharma, Vishwas 336
Shekar Reddy, P. Chandra 240
Sherin 473
Shindu 473
Shravani, D. 301
Shrivastava, Nishant 144
Shunmuganathan, K.L. 87
Siddavatam, Rajesh 123
Sikri, Monika 411
Sil, Jaya 277
Singh, A.K. 117
Singh, Akash 384
Singh, Akhilendra Pratap 132
Singh, Awadhesh Kumar 507
Singh, Kalyan 245
Singh, Mandeep 435
Singh, Nahar 156
Sinha, G.R. 422
Soni, Himanshu 111
Sreedharan, Swetha 273
Srinivas, V.V. 106, 403
Sugumar, D. 473, 477
Sujithra, T.L. 477
Suresh, T. 87
Suresh Babu, K. 273
Thanuja, M.K. 138
Thilagam, P. Shanthi 336
Thomas, Joshua 485
Tiwari, Abhinav 259
Trivedi, Ishita 445
Ujwal, R. 273
Uma, G.V. 327
Upendra Kumar, M. 301
Valarmathi, M.L. 317
Valli Kumari, V. 267
Varadhan, V.V. 106
Venkatesh, G.S. 499
Venkat Reddy, A. 233
Venugopal, K.R. 273
Verma, Kuldeep 222
Verma, Shrish 207
Vijayan, Vinodh P. 332
Vijay Kumar, T.V. 150
Wadhawan, Nisha 259
Wairiya, Manoj 132
Xavier, Lidiya 481
Yadav, Arun Kumar 255, 368
Yogesh, P. 417
