Content uploaded by Fayaz Ali Dharejo
Author content
All content in this area was uploaded by Fayaz Ali Dharejo on Oct 23, 2019
Content may be subject to copyright.
Content uploaded by Awais Ahmed
Author content
All content in this area was uploaded by Awais Ahmed on Aug 24, 2019
Content may be subject to copyright.
LBPH Based Improved Face Recognition At Low Resolution
Aftab Ahmed
School of Information & Software Engineering,
University of Electronic Science & Technology of
China
Chengdu, China
e-mail: aftabahmed@ibacc.edu.pk
Jiandong Guo
School of Information & Software Engineering,
University of Electronic Science & Technology of
China
Chengdu, China
e-mail: jdguo@uestc.edu.cn
Fayaz Ali
School of Information & Software Engineering,
University of Electronic Science & Technology of China
Chengdu, China
e-mail: fayazdharejo40@gmail.com
Farha Deeba
School of Information & Software Engineering,
University of Electronic Science & Technology of China
Chengdu, China
e-mail: farahdeebauestc@hotmail.com
Awais Ahmed
School of Information & Software Engineering,
University of Electronic Science & Technology of China
Chengdu, China
e-mail: engr.awais86@yahoo.com
Abstract—Automatic individual face recognition is the most
challenging query from the past decade in computer vision.
However, the law enforcement agencies are inadequate to
identify and recognize any person through the video
monitoring cameras further efficiently; the blur conditions,
illumination, resolution, and lighting are still the major
problems in face recognition. Our proposed system operates
better at the minimum low resolution of 35px to identify the
human face in various angles, side poses and tracking the face
during human motion. We have designed the dataset (LR500)
for training and classification. This paper employs the Local
Binary Patterns Histogram (LBPH) algorithm architecture to
address the human face recognition in real time at the low level
of resolution.
Keywords-face recognition; LBPH; low resolution; feature
extraction
I. INTRODUCTION
Currently, the Face recognition becomes the more
important topic in computer vision and having much
importance in many applications such as for security,
surveillance, banking and so on. But it becomes more
challengeable because of accuracy and efficiency. Over the
years, many scholars have developed variety kinds of face
recognition algorithms, including Sparse Coding (SC)
algorithm [1], Local Binary Pattern (LBP) algorithm [2],
Histograms of Oriented Gradients (HOG) algorithm [3],
Linear Discriminant Analysis (LDA) algorithm [4], and
Gabor feature algorithm [5].These all algorithms provide
accuracy rate between 50% - 76% [6. Compared with the
above algorithms the LBPH algorithm can not only
recognize the front face, but also recognize the side face,
with 90% accuracy rate [6].
II. WORK FLOW OF FACE RECOGNITION SYSTEM
Figure 1. Face recognition system work flow.
Mostly face recognition system includes four main parts:
information acquisition module, feature extraction module,
classification module and training classifier database module
144
,QWHUQDWLRQDO&RQIHUHQFHRQ$UWLILFLDO,QWHOOLJHQFHDQG%LJ'DWD
978-1-5386-6987-7/18/$31.00 ©2018 IEEE
[7]. The image information collected by the learning
acquisition module which will be used as a test sample for
analysis. In the feature extraction module, which can
represent human identity information is extracted and
examined. In the classification module, the classifier trained
by the database is used to classify test samples to determine
the identification information of individuals.
A. Face Detection
We have used OpenCV which presents a Haar cascade
classifier [8], [12], which is used for face detection. The Haar
cascade classifier uses the AdaBoost algorithm to detect
multiple facial features. First, it reads the image to be
detected and converts it into the gray image, then loads Haar
cascade classifier to decide whether it contains a human face.
If so, it proceeds to examine the face features and draw a
rectangular frame on the detected face. Otherwise, it
continues to test the next picture.
B. Feature Extraction
The LBP operator is applied to describe the contrast
information of a pixel to its neighborhood pixels. The
original LBP operator is defined in the window of 3*3.
Using the median pixel value as the threshold of the window,
it compares with the gray value of the adjacent 8 pixels. If
the neighborhood pixel value is larger or equal compare to
the median pixel value, the value of pixel position is marked
as 1, otherwise marked as (0) [9]. The function is defined as
shown in equation 1. It can be illustrated in Figure 2.
()= 1, ≥0
0, <0
Figure 2. Original LBP Operator.
In this way, 8 points in the 3*3 neighborhood are
compared to generate 8-bit binary numbers. Changing it to
decimal numbers, the LBP values of the middle pixel points
of the window are obtained, which is used to display the
texture features of the region. The current LBPH algorithm
uses an improved circular LBP operator. It can be
represented by Figure 3 and equation 2.
Figure 3. Circular LBP Operator.
= ∑(−)2−1
=0 (2)
The gray value GP of P neighborhoods of the pixel C, the
radius of which is R. GC is the gray value pixel value C
(xc,yc). This algorithm makes the LBP operator no longer
limited to fixed radius and neighborhood and can meet the
needs of more different size and texture features. For each
pixel of an image, it computes its LBP eigenvalues. Then
these eigenvalues can form the LBP feature spectrum. The
LBPH algorithm uses the histogram of the LBP characteristic
spectrum as the feature vector for classification. It divides a
picture into several sub regions, then extracts LBP feature
from each pixel of the sub-region, establishing a statistical
histogram of the LBP characteristic spectrum in each sub
region, so that each sub region can using a statistical
histogram to describe the whole picture through a number of
statistical histogram components. The advantage is to reduce
the error that the image is not fully aligned with a certain
range.
C. Dataset LR500
We have designed our own database named LR500,
which stores 500 images of each person. It is created on the
basis of face detection. Make different facial expressions and
postures to a scene and detect faces. The saved pictures are
stored in the same folder to form the generated face database.
During image acquisition step, the dataset images have been
converted into gray scale images for features extraction; and
then normalized those images for good recognition results.
Normalization technique has been applied on all images to
remove noise and set the alignment position of images.
Figure 4. Test Images of Face Database LR500.
III. FACE RECOGNITON ALGORITHM
To perform the face recognition system here the Local
Binary Pattern Algorithm has been applied. The LBP
operator is used in local features through Local Binary
Pattern acts which shorten the local special arrangement of a
face image [10]. The LBP operator is the number of binary
ratios of pixels intensities within the pixel of center and it’s
around eight pixels. It can be shown in below equation.
(,)=∑
7
=0 (−)2 (3)
Where ic indicates the value of the center pixel and (xc,
yc), shows eight surrounding pixels information. Therefore, it
is very helpful in determining the face features. From the
original matrix Features of the image are extracted then these
values are compared with the center pixel values, the later
binary code is generated.
The Algorithm works as below:
1. First, we need to start with temp=0
2. Where I, is the training for each image
3. H=0,then Initialize the pattern histogram
145
4. Calculate the model label of LBP
5. Keep adding the corresponding bin by 1.
6. Get the greatest LBP feature during each face image and
then merging into the unique vector.
7. It's time to compare the features.
8. Finally, if it resembles with the stored database the image
is recognized.
Figure 5. LBPH algorithm flowchart
A. Feature Vectors
In order to receive the feature vectors, the pattern for
each pixel is obtained [11]. To represent all faces efficiently,
the image has to be subdivided into K2 regions, i-e 82= 64
regions. A histogram with each potential label is composed.
Each bin in a histogram gives the information about a pattern.
While the feature vectors can be obtained from the
histograms. So we can say that each regional histogram hold
of P (P − 1) + 3 bins: P (P −1).
To achieve the area with a distance with the help of the
LBP system from the edges of the image, if it's not then it
means some area on the border of the image is not used.
For the image (NxM), the feature vector is designed with
the help of calculating the LBP code for all pixels (Xc, Yc)
with xc є {R + 1,. .., N − R} and yc {R + 1, . . . , M − R}.
If an image is divided into k × k regions, then the
histogram for region (kx, ky), with kxє {1, . . . , k} and kyє
{1, . . . , k}, Mathematically, :
,= ∑, ,(,)= (),=1,…….(−1)+3 (4)
x ∈
⎩
⎪
⎨
⎪
⎧
R+1,………….,N
KKx= 1
(Kx−1)N
K+1,…….N−R Kx= K
(Kx−1)N
K+1,…….K
xN
K else
(5)
∈
⎩
⎪
⎨
⎪
⎧
R+1,………….,M
K Ky= 1
Ky−1M
K+1,…….M−R Ky= K
Ky−1M
K+1,…….K
yM
K else
(6)
in which L is the label of binary i and
()= 1,
0, (7)
The three distinct levels of locality of the face can be
determined from feature vector: the labels include
information at the little environmental level and design
architecture of histograms provides information about the
face.
IV. RESULTS AND DISCUSSION
In this experiment, each image in the face database has
the distinct ID number. First, prepare the face database, and
then extract the LBP texture features of each test image.
Finally, classify and recognize the face information. For this
test we have collected 2500 face images, those face images
are taken with a TTQ HD 1080px camera.
We compare the input face images with database face
images and work as if the given appearance images, after
extracting features compared with the dataset so finally we
can figure-out the face image is favorably recognized
otherwise the face image would not be recognized. It can be
shown in Figure 6.
Figure 6. unknown person.
Based on the algorithm, this information of face image of
known and an unknown identity is compared with the face
image of known individuals from the available database. In
the research, we have performed major three tasks, capture,
train, and recognize the face images by using the camera.
A. Face Detection
In face detection step, the system detects the face in an
input image via camera and captures the gray scale image.
146
B. Training Face Images
After image acquisition and pre-processing task, we have
to perform dataset training. For training phase, the training
recognizer is applied to store the histogram values of face
images.
Figure 7. Face detection.
TABLE I. TRAINING IMAGES STATISTICS
Total
Recognized
Unrecognized
Training
Images
Images
Images Time
2500
2470
30
35 sec
Figure 8. Dataset training.
C. Recognize Face Image
The final task is to recognize face images. The Haar
cascaded classifier and training recognizer will be used for
face recognition. The classifier will compare the stored face
images with input face images. If the face features of input
images matched with the database images, the recognition
result will be displayed on the camera screen.
@35px @45px
Figure 9. Recognizing face images.
TABLE II. RECOGNITION ACCURACY RATE COMPARISON
Algorithm
45px
35px
LBPH 94% 90%
V. CONCLUSION AND FUTURE IMPROVEMENTS
We used Local Binary Patterns at low resolution for the
face recognition. It essentially contains three major parts, i-e
the representation of the face, feature extraction, and finally
classification. While in Face representation describes the
input of face behaves and moreover, it limits the algorithms
for the detection and recognition. Further, for feature
extraction, this LBPH histogram found a novel result and
finally we classify input detected face compare with the
proposed DATASET (LR500). Then we can analyze our
system either recognized a known person or unknown person.
In future, this proposed approach will be more beneficial
for security agencies to identify criminals, whose have
criminal record in database. It will help to recognize any
unknown or known person in surveillance area at low
resolutions due to long distance of camera and observed
subject.
VI. ACKNOWLEDGEMENT:
This work is supported by Vice Professor Jiandong Guo
and School of Information and Software Engineering,
University of Electronic Science and Technology of China.
REFERENCES
[1] J. Olshausan B A, Field D J.Emergence of simple-cell receptive field
properties by learning a sparse code for natural images. Nature, 1996,
381(6583):607-609
[2] J. CHAO W L, DING J J, LIU J Z. Facial expression recognition
based on improved local binary pattern and class-regularized locality
preserving projection. Signal Processing, 2015, 117:1-10.
[3] J. HU Liqiao, QIU Runhe. Face recognition based on adaptive
weighted HOG. Computer Enigeering and Applications, 2017, 53(3):
164-168.
[4] J. Yu yan JIANG, Ping LI, Qing WANG. Labeled LDA model based
on shared background topic. Acta Electronica Sinica, 2015, 2013, (9):
1794-1799.
[5] J. WU Qi, WANG Tang-hong, LI Zhan-li. Imporved face recognition
algorithm based on Gabor feature and collaborative representation.
Computer Engineering and Design, 2016, 37(10): 2769-2774.
[6] Aruni Singh, Sanjay Kumar Singh, Shrikant Tiwari, Comparison of
face Recognition Algorithms on Dummy Faces, The International
Journal of Multimedia & Its Applications (IJMA) Vol.4, No.4,
August 2012.
[7] XueMei Zhao, ChengBing Wei, A Real-time Face Recognition
System Based on the Improved LBPH Algorithm, 2017 IEEE 2nd
International Conference on Signal and Image Processing.
[8] Varun Garg, Kritika Garg, Face Recognition Using Haar Cascade
Classifier, Journal of Emerging Technologies and Innovative
Research (JETIR) , December 2016, Volume 3, Issue 12.
[9] Hongshuai Zhang, Zhiyi Qu Liping, Yuan GangLi, A Face
Recognition Method Based on LBP Feature for CNN, 2017 IEEE 2nd
Advanced Information Technology, Electronic and Automation
Control Conference (IAEAC).
[10] T. Chen, Y. Wotao, S. Z. Xiang, D. Comaniciu, and T. S. Huang,
“Total variation models for variable lighting face recognition” IEEE
Transactions on Pattern Analysis and Machine Intelligence,
28(9):1519{1524, 2006
[11] Zhao and R. Chellappa “Robust face recognition using symmetric
shape from-shading” Technical Report, Center for Automation
Research, University of Maryland, 1999
[12] Zheng Xiang, Hengliang Tan, Wienling Ye. The excellent properties
of dense gird-based HOG features on face recognition compare to
gabor and LBP, 2018 volume issue 99.
147
