Content uploaded by Mohammad A. N. Al-Azawi
Author content
All content in this area was uploaded by Mohammad A. N. Al-Azawi on Apr 20, 2018
Content may be subject to copyright.
VOL. 13, NO. 6, MARCH 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.
www.arpnjournals.com
2259
FACE RECOGNITION TECHNIQUE BASED ON ADAPTIVE-OPPOSITION
PARTICLE SWARM OPTIMIZATION (AOPSO) AND SUPPORT VECTOR
MACHINE (SVM)
Mohammed Hasan Abdulameer1, DhurghamA.Mohammed1, Saad Ali Mohammed2, Mohammed Al-Azawi3,
Yahya Mahdi Hadi Al-Mayali4 and Ibrahim A. Alameri5
1Department of Computer Science, Faculty of Education for Women, University of Kufa, Iraq
2College of Imam Kadhim for Islamic Sciences University, Iraq
3Oman College of Management and Technology, Oman
4University college of Humanities, Iraq
5Jabir Ibn Hayyan Medical University, Iraq
E-Mail: mohammed.almayali@uokufa.edu.iq
ABSTRACT
OPSO and AAPSO are the most recently developed face recognition techniques, in order to optimize the
parameters of SVM. However, in order to increase the optimization, a combination between OPSO and AAPSO techniques
has been proposed in this paper. The proposed technique is called Adaptive-Opposition particle swarm optimization
(AOPSO). In AOPSO, the random values in the initial generation of the population in PSO is solved by OPSO and the
randomization fixed values in the velocity coefficient is solved using AAPSO in the same time. T hen, the proposed
algorithm is used with support vector machine to find the optimal parameters of SVM. The performance of the proposed
AOPSO method has been validated with two face images datasets, YALE and CASIA datasets. In the proposed method,
we have initially performed feature extraction, followed by the recognition of the extracted features. In the recognition
process, the extracted features have been employed for SVM training and testing. During the training and testing, the SVM
parameters have been optimized with the AOPSO technique. The comparative analysis has demonstrated that, the AOPSO-
SVM proposed in this study has outperformed the existing PSO-SVM technique.
Keywords: face recognition, PSO, AOPSO, SVM.
1. INTRODUCTION
Classification, regression and other learning tasks
can be done using a popular machine learning method
known as Support Vector Machine (SVM) [1]. The SVM
classifier [2, 3] is a supervised learning algorithm, which
depends on statistical learning theory [4, 5]. The SVM
classifier is aimed at determining a hyperplane, which
effectively divides two classes by using training data sets
[6]. It is noteworthy that, SVM is a highly effective
technique, which focuses on the concept of increasing the
margin, or level of separation, in the training data. Several
hyperplanes are available, which facilitates the division
of data for classification. Nevertheless, it is crucial to
select optimal hyperplane, which signifies the largest
separation, or margin, between two classes. SVM looks for
the ideal hyperplane making use of support vectors [7].
The support vectors are the training samples, which
estimate the ideal separating hyperplane; furthermore, they
are the most challenging patterns to classify [8]. This
means that, the support vectors comprise of the data
points, which are nearest to the optimal hyperplane. As
SVM deals with a subset of data points (support vectors)
that are close to the decision boundary, generally the SVM
solution depends on the local features of the data. Over the
last ten years, SVM has triggered the interest of
researchers as a contemporary machine-learning technique
for solving problems in a number of fields, such as, pattern
recognition, bioinformatics and other non-linear problems
with small sample sizes. Fundamentally, SVM has robust
theoretical framework and a superior generalization
capability [9]. In terms of implementation, training an
SVM in classification is similar to that of solving a
linearly constrained quadratic programming (QP) problem,
which utilizes considerable volume of memory and
computation time, when dealing with huge volume of
samples. Even though PSO-SVM [6, 10], AAPSO-SVM
[11] and OPSO-SVM [12] have been widely used to
address the above mentioned problems, still it has a room
of an improvement in term of SVM parameters
optimization. Therefore, for resolve this problem, in this
study we have proposed a AOPSO-SVM method. The rest
of the paper has been organized as follows: the related
work has been discussed in section 2, followed by the
description of our proposed method. The experimental
result and analysis have been described in section 3.
Finally, the conclusions and potential future studies have
been presented in Section 4.
2. RELATED WORK
As discussed above, the PSO technique has been
employed in several fields, such as, video-based face
recognition, face verification and face recognition. In
traditional PSO, the populations are randomly generated,
which results in the random generation of population; this
feature creates an element of doubt about accuracy of
results produced by PSO. Another problem in PSO, the
randomization that is occurring in choosing the velocity
coefficients of PSO which is usually fixed random
number.
VOL. 13, NO. 6, MARCH 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.
www.arpnjournals.com
2260
Consequently, a number of studies have
attempted to resolve the above mentioned problems.
Mohammed et al., [12] have proposed OPSO-SVM for the
purpose of minimizing the constraint of randomization in
PSO; the OPSO has been employed to enhance the
parameters of the SVM. The optimized SVM has been
trained for effective face recognition. In another study, and
to address the problem of random values for the
acceleration coefficients in PSO; [11] have proposed
adaptive acceleration particle swarm optimization
(AAPSO). They utilized AAPSO to optimize the
parameters of SVM. In addition, many researchers have
been used PSO as a part of face recognition technique.
Raghavendra et al., [13] have proposed a two-image
fusion scheme for incorporating visible and near infrared
face images (NIR), to enhance the performance of face
verification; for this purpose, they have employed particle
swarm optimization (PSO) to discover an ideal approach
for performing fusion of the visible and NIR sub-band
coefficients. For addressing the dynamic optimization
problem Jean-Francois Connolly et al. [14] have proposed
a dynamic particle swarm optimization for a video-based
face recognition system. The authors have proposed an
incremental learning strategy based on dynamic particle
swarm optimization (DPSO), to develop heterogeneous
ensembles of classifiers (where each classifier fits to a
particle) in response to new reference samples. This
approach has been employed to resolve video-based face
recognition, using an AMCS that includes a pool of fuzzy
ARTMAP (FAM) neural networks, which are used for
classifying facial regions; and the enhanced version of
DPSO, optimizes all FAM parameters, so that the
classification rate has been maximized.
3. THE PROPOSED FACE RECOGNITION
TECHNIQUE
The proposed face recognition technique is
performed in three phases: feature extraction by PCA,
Adaptive-Opposition particle swarm optimization
(AOPSO), and parameters selection for SVM with
AOPSO. These three phases have been performed
repeatedly on the input database face images, and thus, the
face images are recognized more effectively. The three
phases are discussed in Sections 3.1, 3.2 and 3.3. The
basic structure of our proposed face recognition system is
shown in Figure-1.
Figure-1. Structure of the proposed recognition technique based on AOPSO-SVM.
3.1 Feature extraction using PCA
The purpose of the feature extraction is to extract
the information that represents the face. Principal
component analysis (PCA) is one of the popular feature
extraction methods [15]. Using an information theory
approach, PCA has been used for feature extraction in face
recognition [16]. This approach could be proficiently and
easily coded to extract the appropriate information from a
face image. The subspace of the image space extended by
the training face image data has been identified and de-
correlated, using the pixel values. The conventional
representation of a face image has been obtained, by
projecting the face image on to the coordinate system
described by the principal components. Information
compression, de-correlation and dimensionality reduction
have been employed for decision making in the projection
of face images into the principal component subspace.
Dealing with an image as a vector in a highly dimensional
face space mathematically attempts the principal
components of the distribution of faces or the eigenvectors
of the covariance matrix of the set of face images. We
have applied PCA on the training and testing database face
images and have obtained exclusive one dimensional
feature vectors.
3.2 The proposed adaptive-opposition particle swarm
optimization (AOPSO)
Particle swarm optimization (PSO) is inspired by
the social behavior of biological creatures, such as fishes
and birds, which have the ability to group together to work
as a whole to locate desirable positions in a certain area,
e.g., fish searching for a food source. This type of search
behavior is equivalent to searching for solutions of
equations in a real-valued search space [17].PSO emulates
the swarm behavior of individuals who represent potential
solutions in a D-dimensional search space. In the proposed
AOPSO method, the populations are generated randomly
as in the standard PSO and also based on opposite number
(Jabeen et al. 2009). The opposite number generation
process is described as below:
Let be the real number which is generated between
the intervals , then the opposite number is
defined by
Let be a point, where and
based on these points the opposite points are defined
as
VOL. 13, NO. 6, MARCH 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.
www.arpnjournals.com
2261
where
By utilizing the aforementioned process, the
opposition based populations are initially generated in
OPSO (Jabeen et al. 2009). After the population
generation, each individual particles fitness value is to be
computed. The fitness value of and point
can be replaced by
p
otherwise we will continue
with . To find the more fit ones the opposite point
and its point are evaluated simultaneously.
Particle
i
is often composed of four vectors:
),,,( 21 D
iiii xxxX
, where
d
i
x
is its position in the dth
dimension;
),,,( 21 D
iiii pbestpbestpbestpbest
,
where
d
i
pbest
is the best position in the dth dimension
that particle
i
has found on its own;
),,,( 21 D
iiii vvvV
, where
d
i
v
is the velocity in the dth dimension; and
),,,( 21 D
gbestgbestgbestgbest
, where
d
gbest
is the global best position in the dth dimension
that all particles have found. Particles in a swarm move
through the search space as follows:
(1)
, (2)
1
r
and
2
r
are two independent random numbers
uniformly generated in the range [0.1] at each updating
iteration from d=1 to D,
d
i
V
is the velocity of the ith
particle,
d
i
x
is the current position of the particle
i
,
d
i
pbest
is the position of the best fitness value of the
particle at the current iteration and
d
gbest
is the position
of the particle with the best fitness value in the swarm. In
addition,and are the adaptive velocity coefficients
based on the proposed formula Adaptive PSO
(Abdulameer et al 2014) :
max
2
minmin
f
f
avg
f
f
+, (3)
max
2
minmin
f
f
avg
f
f
+, (4)
where and represent the minimum
and maximum values of,
and
are the
particle minimum, average and maximum fitness values of
the entire population, and and represent the
minimum and maximum values of. The proposed
AOPSO algorithm is described in the following flowchart:
VOL. 13, NO. 6, MARCH 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.
www.arpnjournals.com
2262
The pseudo code of the proposed AOPSO
For every particle
Initialize particle randomly+ opposite number
END
Do
For every particle in both populations
Compute fitness value and choose the fitter ones form both populations are
chosen as particle.
If the Computed fitness value is better than the best fitness value (pbest) in
history
put current value as the new pbest
End
Select the particle that has best fitness value among all the particles as the
gbest
For every particle
Compute the velocity of particle using equation (1)
Update Particle position using equation (2)
End
Continue while maximum iterations or minimum error criteria is not attained
VOL. 13, NO. 6, MARCH 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.
www.arpnjournals.com
2263
3.3 Parameter selection for support vector machine
using adaptive opposition particle swarm
optimization
The SVM parameters optimized using the
proposed AOSPO method as shown in Figure-2.
Figure-2. SVM parameter optimization using the
proposed AOPSO.
The process of optimal parameter selection by
AAPSO in SVM is as follows:
Step 1: Initially, the particles are generated randomly and
based on the opposite number generationwithin
the interval
],[ yx
. The generated particles are
composed of SVM parameters
i
p
. Then, the
parameters of each particle are initiated,
including position and velocity.
Step 2: The fitness value of every particle is calculated
using Equation. (5). The particles that have the
minimum fitness values are selected as the best
particles as follows:
N
i
ii Cp
1
2
2
1
min
, (5)
Such that
Ni
Nib
y
xp
i
i
i
N
i
ii
,2,1,0
,2,1,
1
.
1
(6)
where,
N
is the size of the training dataset, and C is the
cost function.
Step 3: The
i
pbest
of each particle is updated and
i
gbest
for the domain is updated. Based on
these values, the velocity and position of every
particle are updated using equations (1) and (2).
Step 4: Stop if the current optimization solution is good
enough or if the stopping criterion is satisfied.
4. EXPERIMENTAL RESULTS
The proposed recognition technique has been
experimented in the working platform of MATLAB. The
performance of the proposed AOPSO-SVM technique has
been evaluated using the YALE dataset [19] and CASIA
face dataset [20].
The two datasets were divided into training and
testing datasets. In Yale dataset, there are 15 classes, in
each class there are different images in different
conditions. In the experimental tests, 165 images have
been used in the evaluation process75 images for training
and 90 images for testing.From CASIA database, 500
images have been used for experimentation. In CASIA
database, the images are chosen at five different poses in
different environments and illumination variations. In the
evaluation process, the images in the dataset have been
equally divided for training and testing.
The performance of the proposed technique has
been analyzed by conducting n-fold (n = 10) cross
validation over each datasets, and the corresponding
statistical performance measures are determined. To
perform n-fold cross validation, ten folds of training and
testing datasets are generated by folding operation. The
images are obtained from both databases, and the feature
extraction has been computed using PCA, while the
recognition process has been computed using the proposed
AOPSO-SVM technique. Figure-3 illustrates the sample
face images from the YALE and the CASIA databases.
(i)
(ii)
Figure-3. Sample face images from (i) YALE and (ii)
CASA databases.
To analyze the classification performance, we
have conducted 10 experiments on the Yale and CASIA
datasets. Table-1 and Figure-4 illustrate the results of
classification accuracy obtained for the two dataset. In 10
experiments, our proposed AOPSO-SVM method has
attained higher iris image classification accuracy, as
against the standard PSO-SVM. The average classification
accuracy is 86.9% for Yale dataset, and 88.3% for CASIA
dataset.
VOL. 13, NO. 6, MARCH 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.
www.arpnjournals.com
2264
Table-1. The accuracy of PSO and AOPSO based on SVM classification performance results
for the Yale and property CASIA face datasets.
Experiments
Accuracy (%)
PSO-SVM
Accuracy (%)
AOPSO-SVM
Accuracy (%)
PSO-SVM
Accuracy (%)
AOPSO-SVM
Yale dataset
Yale dataset
2.5d dataset
2.5d dataset
1
89
92
89
91
2
81
90
80
85
3
80
85
85
90
4
70
82
87
90
5
82
80
80
85
6
80
87
85
90
7
81
90
82
85
8
82
87
80
85
9
85
86
89
92
10
80
90
85
90
Average
81
86.9
84.2
88.3
Figure-4. Average accuracy values of PSO-SVM and the proposed AOPSO-SVM.
Additionally, the performance of our proposed
AOPSO method has been compared with the performance
of the PSO method, Equation. (5). Based on the results
illustrated in Figure-5, it is evident that, our proposed
AOPSO method has yielded more accurate particles that
have lower fitness values, than those generated by the
conventional PSO method. The high performance result
shows that, our AOPSO method is able to determine the
more accurate SVM parameters.
(i)
VOL. 13, NO. 6, MARCH 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.
www.arpnjournals.com
2265
(ii)
Figure-5. Performance of AOPSO and PSO methods
(i) YALE database (ii) CASIA database.
In Figure5, our proposed technique has obtained
more accurate particles that have minimum fitness values,
smaller than those obtained with the conventional PSO
method. Therefore, our AOPSO technique has yielded
more accurate SVM parameters. Figure-5(i) and (ii) show
the fitness value performance for particles used on the
YALE and CASIA face images databases. For all
iterations, the fitness values of the particles of our
proposed AOPSO method are lower than those of the
conventional PSO method. Furthermore, the
computational times for our proposed AOPSO and the
PSO methods are shown in Table-2 and Figure-6.
Table-2. Computation times of proposed AOPSO and
PSO techniques with SVM.
Computational Time (sec)
Images
PSO
AOPSO
1
0.182345
0.133412
2
0.024352
0.019061
3
0.025007
0.020171
4
0.02178
0.014853
5
0.020261
0.017977
6
0.020363
0.020351
7
0.019325
0.017685
8
0.020612
0.018368
9
0.020936
0.01778
10
0.018513
0.017532
Figure-6. Computation times of proposed AOPSO-SVM and PSO-SVM techniques.
5. CONCLUSIONS
In this paper, we had introduced AOPSO based
on SVM to address the limitations of the standard PSO
method. These limitations of PSO: using random values in
the initialization process and choosing velocity
coefficients, which may lead to performance instability?
The optimized SVM, using the AOPSO technique, shows
effective face recognition performance. Two human face
databases, YALE and CASIA were utilized to analyze the
performance of our proposed AOPSO-SVM face
recognition technique. The performance and comparative
analysis results had showed that our proposed AOPSO-
SVM technique has yielded higher face recognition
performance results as against the PSO-SVM method.
Based on the 10 experiments we have proved that, our
proposed AOPSO method has attained higher face image
average classification accuracy. The average classification
accuracy was 86.9% for Yale dataset and 88.3% for ASIA
dataset. In addition, the computation time of proposed
AOPSO was lesser than the conventional PSO by 21%, in
terms of finding the optimal parameters in SVM.
REFERENCES
[1] Chih-Chung C, Chih-Jen L. LIBSVM: 2011. A
Library for Support Vector Machines. ACM
VOL. 13, NO. 6, MARCH 2018 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2018 Asian Research Publishing Network (ARPN). All rights reserved.
www.arpnjournals.com
2266
Transactions on Intelligent Systems and Technology.
2(3) : 1-27.
[2] Bazzani A, Bevilacqua A, Bollini D, Brancaccio R,
Campanini R, Lanconelli N, Riccardi A, Romani D.
2001. An SVM classifier to separate false signals
from icrocalcifications in digital mammograms.
Physics in Medicine and Biology. 46(6): 1651-63.
[3] Mukundhan S, Sriram R. 2013. Comparative study on
hidden markov model versus support vector machine
a component-based method for better face
recognition. Proceedings - UKSim 15th International
Conference on Computer Modelling and Simulation,
UKSim. 430-436.
[4] Vladimir N. 1998. AT &T Statistical Learning
Theory. New York, USA: John Wiley & Sons.
[5] Vladimir N. 1998. Statistical Learning Theory.
Adaptive and learning Systems for Signal Processing.
Communications and Control. 2: 1-740.
[6] Jin W, Zhang, Zhang X. 2011. Face recognition
method based on support vector machine and particle
swarm optimization. Expert Systems with
Applications. 38: 4390-4393.
[7] Khan N, Ksantini R, Ahmad I, Boufama B. 2012. A
novel SVM+NDA model for classification with an
application to face recognition. Pattern Recognition.
45: 66-79.
[8] Duda RI, Hart PS, Stork DI 2000. 2nd ed. Pattern
Classification. California, USA: Wiley.
[9] Guermeur Y. 2002. Combining discriminant models
with new multi-class SVMs. Pattern Analysis and
Applications. 5(2): 168-179.
[10] Subasi A. 2013. Classification of EMG signals using
PSO optimized SVM for diagnosis of neuromuscular
disorders. Computers in Biology and Medicine. 43(5):
576-586.
[11] Abdulameer M. H., Sheikh Abdullah S. N., Othman
Z. 2014. A Modified Active Appearance Model Based
on an Adaptive Artificial Bee Colony. The Scientific
World Journal. 4: 1-16.
[12] Abdulameer M. H., Sheikh Abdullah S. N., Othman
Z. 2013. A face recognition system based on
opposition particle swarm optimization (OPSO) and
support vector machine (SVM). In: IEEE.
International Conference on Signal and Image
Processing Applications (ICSIPA); October, Malaka,
Malaysia.
[13] Raghavendra R, Bernadette D, Ashok R, Hemantha
K. 2011. Particle swarm optimization based fusion of
near infrared and visible images for improved face
verification. Pattern Recognition. 44: 401-411.
[14] Jean-Francois C, Eric G, Robert S. 2012. Evolution of
heterogeneous ensembles through dynamic particle
swarm optimization for video-based face recognition.
Pattern Recognition. 45: 2460-2477.
[15] Vaseghi S, Jetelova H. 2006. Principal and
Independent Component Analysis in Image
Processing. In: Proceedings ACM International
Conference on Mobile Computing and
Networking14th; San Francisco, Calif, USA: pp. 1-5.
[16] Subasi A. 2013. Classification of EMG signals using
PSO optimized SVM for diagnosis of neuromuscular
disorders. Computers in Biology and Medicine. 43(5):
576-586.
[17] Chen W, Zhang J, Henry S.H, Wen-liang Z, We-gang
W, Yu-Hui S. 2010. A novel set-based particle swarm
optimization method for discrete optimization
problem. IEEE Transactions on Evolutionary
Computation. 14: 278-300.
[18] Jabeen, Ruqaya, Altaf A, Muhammad I. 2009.
Phytoremediation of heavy metals: physiological and
molecular mechanisms. The Botanical Review. 75(4):
339-364.
[19] Bemd H, Purdy H, Tomaso P. 1997. Eigenfaces vs.
Fisherfaces: recognition using class specific linear
projection. IEEE Transactions on Pattern Analysis
and Machine Intelligence. 19(7): 711-720.
[20] Chinese Academy of Sciences. 2007. CASIA face
database. http://biometrics.idealtest.org/.