Content uploaded by Hemantha kumar Kalluri
Author content
All content in this area was uploaded by Hemantha kumar Kalluri on Oct 05, 2018
Content may be subject to copyright.
COMPARATIVE STUDY OF FACE
RECOGNITION TECHNIQUES
N. BalaYesu1, Hemantha Kumar Kalluri2,
1Department of Computer Science & Engineering
Vignan’s Foundation for Science Technology
and Research, Vadlamudi, Andhra Pradesh, India
2Vignan’s Lara Institute of Technology
and Sciences, Vadlamudi, Andhra Pradesh, India
hemanth mtech2003@yahoo.com,
nbalayesu@gmail.com,
nbalayesu vlits@vignan.ac.in
June 12, 2018
Abstract
In this paper presented a comparative study of face
recognition techniques. Features are considered holistic
features, Gabor features, phase congruency features.
Experiments are conducted for verification (Equal Error
Rate) and identification (Correct Identification Rate) on
ORL database using Principal Component Analysis,
Linear Discriminant Analysis, Kernel Fisher Analysis, and
Kernel Principal Component Analysis.
Keywords:Biometrics, Face Recognition, PCA, LDA,
KFA, KPCA.
1 INTRODUCTION
Biometric system verifies and identifies the individual based on
their physiological or behavioral characteristics to provide access
1
International Journal of Pure and Applied Mathematics
Volume 120 No. 6 2018, 3527-3536
ISSN: 1314-3395 (on-line version)
url: http://www.acadpubl.eu/hub/
Special Issue http://www.acadpubl.eu/hub/
3527
to physical or virtual domains instead of passwords, PIN numbers,
smart cards, tokens and so forth. Jain et al., [1] proposed
biometrics to overcome the limitations of the passwords; PIN
numbers (may be lost or forgotten). Nowadays, many
applications, like attendance monitoring in an educational
institutions, banking, and business are moving towards biometric
technologies. Muhtahir et al., [2] specified that the physiological
traits of an individual include face, iris, palm print, finger print,
ear, hand geometry. The behavioral traits include gait, key stroke,
signature and speech.
The hand geometry and the iris scan are very expensive [3] but
they provide good accuracy. Whereas face recognition offers the
following advantages such as Non-Intrusive, low equipment cost, do
not require individual physical interaction, no transfer of any health
risk. Good algorithms can handle noise, and slight variation in
orientation, illumination in an image. Kalluri et al., [4, 5] proposed
palm print identification using WPLI and Gabor features.
Since face recognition offers above advantages, it has been
using in the applications like, security, surveillance, general
identity verification, and smart card, video indexing, image
database investigation, and so forth. Face recognition system
consists of three steps, Face detection (Segmentation), Feature
extraction and Face recognition. The block diagram of Face
recognition is shown in Figure 1. The aim of face detection is to
extract the face from the background. The important task in face
recognition is feature extraction.
Figure 1: Block diagram of Face Recognition
Every image consists of four important features, they are 1.
Visual features (includes edges, contours, textures and regions).2.
Statistical features 3. Transform Coefficient features and 4.
Algebraic features. The approaches for extracting features are
Knowledge based, Mathematical Transform, Neural Network or
Fuzzy Extractor and other methods.
2
International Journal of Pure and Applied Mathematics Special Issue
3528
Gabor Wavelet Transform is used as one of the popular feature
extraction techniques, due to its biological relevance and
computational properties. Gabor feature extraction captures
visual properties such as spatial localization, spatial frequency
and orientation selectivity. The dimensionality of the Gabor
feature space is very large because the face images given by it are
combined with the Gabor filter bank. PCA is opted to reduce the
dimensionality. Phase Congruency extracts the face local features
and it can be computed using Gabor Wavelet transform. It is
insensitive to variations in illumination.
Lerato Masupha et al., [6] classified the Face Recognition
techniques based on static images into feature based and holistic
based techniques. Feature based techniques identify distinct
features in face like eyes, nose, and mouth...etc.as well as other
facial points. Then they measure the geometrical relationship
between facial points and reduce the image into geometrical
feature vector. Feature based techniques are classified into
geometric feature based and Elastic bunch Graph. Holistic based
Techniques: These techniques use complete information about
image to identify faces. They are subdivided into statistical and
Artificial Intelligence methods.
In statistical based approach represents the entire image as the
2D array of intensity values, then the probe image is compared
with the all the faces in the data base. Principal Component
Analysis (PCA), LDA is the examples of statistical approach.
Artificial Intelligence techniques use concepts of neural networks
and machine learning to perform face identification or verification.
Local Binary patterns, Hidden Markov Model, Support Vector
Machine, Neural Networks and Radial Basic fuzzy (RBF) are the
examples of Artificial Intelligence approach.
The rest of the paper is organizes as follows: section 2 explained
about the proposed approach. Experimental results are discussed
in section 3. Conclusions are given in section 4.
3
International Journal of Pure and Applied Mathematics Special Issue
3529
2 PROPOSED APPROACH FOR
FACE RECOGNITION
The block diagram of the proposed face recognition and
identification are shown in Figure 2 and Figure 3. The proposed
approach uses Principal Component Analysis (PCA), Linear
Discriminant Analysis (LDA) and etc. the brief explanation of
PCA and LDA as follow. For feature extraction, we used the
holistic features, Gabor features, Phase congruency features.
Principal Component Analysis:
Karim et al [7] stated that the Principal component analysis is
linear projection method. PCA find the optimal linear
least-square representation in (N-1) dimension space, where N is
the total facial images. A set of N-1 eigenvectors and eigenvalues
are used to describe the representation. The orthogonal can be
finding by normalizing the Eigen vectors. The larger portion in
training data can be encoded by lower order Eigen vectors
whereas s the higher-order eigenvectors changes smaller portion.
PCA is very easy and efficient and also the correlation between
training and testing image is high. A. Mir and A. G. Mir et al. [8]
proved that the accuracy of PCA decreases as lighting effects.
LDA:
Sharkas et al. [9] proposed that the Linear Discriminant Analysis
or Fisher analysis maximizes the inter class differences but not the
data representation. The application of LDA encounters the
problem, when the sample size small. Chen et al. [10] proposed a
method that solves the small sample size problem.
Figure 2: Block diagram of Face Registration
4
International Journal of Pure and Applied Mathematics Special Issue
3530
Figure 3: Block diagram of Face Identification
In this approach, the entire face database is divided into two
partitions named as training set and testing set. Training set is
used to construct the PCA/LDA subspace. Feature vector for the
training set is prepared. The test image is projected onto
PCA/LDA subspace to produce test image feature vector. The
classifier performs classification based on the similarity between
train and test image. Euclidean distance is used to measure the
similarity. If the distance is less than the threshold, test image is
recognized otherwise not recognized.
3 FACE DATABASE
ORL [11] database was created by AT & T institute. ORL database
consists of 400 images of 40 subjects where each subject has 10
different styles. Each image is in PGM format and is of size 92x112
pixels. The number of grey levels is 256 per pixel. Some of the
sample images are shown in Figure 4.
4 RESULTS
The performance of various face recognition algorithms on ORL
face database is given below. Out of 10 images of every face, first
3 images are considered as training and remaining 7 images are for
testing. In the second run, we considered first 5 images are train
images and the remaining 5 images are test images.
5
International Journal of Pure and Applied Mathematics Special Issue
3531
Figure 4: Sample image of ORL database
When the number of training images is 3, the identification
rates for PCA is 66.07, for LDA is 86.07, for KPCA is 49.69, for
KFA is 85.71. The performance of PCA, LDA, KFA, and KPCA
for 5 training images is 70.83, 91.5, 88.5, and 52 % respectively.
Gabor based PCA, KFA, KPCA and LDA produced the
identification rates as 84.5, 93.33, 80, and 93.33 respectively for 3
train images. For 5 train images, they produce 62.5, 95, 67.5 AND
100% respectively. The results are given in Table 1 and depicted
in Figure 5.
Table 1: Results obtained for different face recognition algorithms
Algorithm Correct Identification Rate
3 Training Images 5 Training Images
PCA 66.07 70.83
KPCA 49.29 52
KFA 85.71 88.5
LDA 86.07 91.5
Gabor PCA 84.5 62.5
Gabor KPCA 80 67.5
Gabor KFA 93.33 95
Gabor LDA 93.33 100
Phase Congruency PCA 60.83 57.5
Phase Congruency KPCA 71.67 57.5
Phase Congruency KFA 84.17 87.5
Phase Congruency LDA 82.5 92.5
The Equal Error Rate (EER) for various face recognition
6
International Journal of Pure and Applied Mathematics Special Issue
3532
Figure 5: Performance of algorithms
techniques are given in Table 2. The EERs for PCA, LDA,
KPCA, KFA, Gabor based PCA, LDA, KFA, KPCA , Phase
Congruency based PCA, KPCA, LDA,KFA with 3 training
images are 5.03, 9.29, 7.22, 4.28, 5.01, 4.17, 2.51, 2.51, 5.85, 6,
5.94 and 5.19. The EERs for PCA, LDA, KPCA, KFA, Gabor
based PCA, LDA, KFA, KPCA , Phase Congruency based PCA,
KPCA, LDA,KFA with 5 training images are 4.08, 7.03, 6.13, 2.5,
0, 0, 2.28, 0, 2.5, 2.5, 7.37 and 2.5.
Table 2: Results Equal Error Rate
Algorithm Equal Error Rate
3 Training Images 5 Training Images
PCA 5.03 4.08
KPCA 9.29 7.03
KFA 7.22 6.13
LDA 4.28 2.5
Gabor PCA 5.01 0
Gabor KPCA 4.17 0
Gabor KFA 2.51 2.28
Gabor LDA 2.51 0
Phase Congruency PCA 5.85 2.5
Phase Congruency KPCA 6 2.5
Phase Congruency KFA 5.94 87.5
Phase Congruency LDA 5.19 2.5
7
International Journal of Pure and Applied Mathematics Special Issue
3533
5 CONCLUSION
From experimental results, it is observed that Gabor based face
recognition algorithms produce better results compared to phase
congruency based face recognition algorithms. Gabor PCA, Gabor
KPCA and Gabor LDA obtains 100% accuracy when the number
of training images are 5. It is also observed that using five training
images reduces the Equal Error Rate.
References
[1] A. K. Jain, R. Bolle, and S. Pankanti, ”Biometrics: Personal
Identification in Networked Security,” A. K. Jain, R. Bolle,
and S. Pankanti, Eds.: Kluwer Academic Publishers, 1999.
[2] M. O. Oloyede, and G. P. Hancke, Unimodal and Multimodal
Biometric Sensing Systems: A Review, IEEE Access,pp:7532-
7555, Vol.4, 2016.
[3] R. Jafri and H. R. Arabnia : A survey of Face Recognition
techniques, Journal of Information Processing Systems, pp. 41-
68, Vol.5, No.2, 2009.
[4] H. K. Kalluri, M. V. N. K. Prasad, A. Agarwal,Palmprint
Identification and verification based on wide principal lines
through dynamic ROI, International Journal of Biometrics,
pp. 1-30, Vol. 7, No. 1, 2015.
[5] H. K. Kalluri, M V N K Prasad, Palmprint Identification based
on Goabor and WPLI Features, Procedia Computer Science
93, 706 712, 2016.
[6] L. Masupha, T. Zuva, S. Ngwira, O. Esan: Face
Recognition Techniques, their Advantages, Disadvantages
and Performance Evaluation, International Conference on
Computing, Communication and Security (ICCCS), 2015.
[7] T. F. Karim, M. S. H. Lipu, Md. L. Rahman, F. Sultana : Face
Recognition Using PCA-Based Method, IEEE International
Conference on Advanced Management Science, 2010.
8
International Journal of Pure and Applied Mathematics Special Issue
3534
[8] A. Mir and A. G. Mir, ”feature extraction methods(PCA fused
with DCT),” International Journal of advances in engineering
and technology, vol. 6, pp. 2145- 2152, 2013.
[9] M. Sharkas, M. Abou Elenien, Eigenfaces vs. Fisherfaces
vs. ICA for Face Recognition; A Comparative Study,
9thInternational Conference on Signal Processing, 2008.
[10] L. Chen, H. Yuan, M. T. Ko, J. C. Lin, G. J. Yu : A new LDA-
based face recognition system which can solve the small sample
problem, Pattern Recognition, Vol. 33, No. 10, pp. 1713-1726,
2000.
[11] http://www.cl.cam.ac.uk/research/dtg/attarchive/face
database.html
9
International Journal of Pure and Applied Mathematics Special Issue
3535
3536