Sunteți pe pagina 1din 5

2016 International Seminar on Application for Technology of Information and Communication

Face Recognition using 3D GLCM and Elman


Levenberg Recurrent Neural Network

Rocky Yefrenes Dillak Sumartini Dana Marthen Beily


Electrical Engineering Deparment Electrical Engineering Department Electrical Engineering Department
Politeknik Negeri Kupang Politeknik Negeri Kupang Politeknik Negeri Kupang
Kupang, Indonesia Kupang, Indonesia Kupang, Indonesia
rocky_dillak @ yahoo.com sumartinidana @ yahoo.com marthen.beily @ yahoo.co.id

AbstractFace recognition has been widely used as biometric Various research have proven that Gray level co-
systems currently. Because of its ability to recognize a person occurrence matrix (GLCM) is powerful basis for use in
identity reliably and accurately based on face. This research pattern recognition. The textural properties calculated from
aims to develop a method that can be used for face recognition the GLCM help understand the details about the image
system. The proposed approach works as follows:(i) preprocess content. Such as conducted in [2] and [3] proved that the use
input images using gray-scaling, contrast stretch and Amoeba of GLCM could increase the accuracy higher than PCA. In
Median Filter, (ii) extract its characteristic features using 3D [4], they expanded GLCM from 2-dimensional to 3-
GLCM which is expanded from 2D GLCM, (iii) train these dimensional. Thus the result is 3D GLCM is better than 2D
characteristic features using The Elman Levenberg Neural
GLCM. The expanded GLCM also was used in [5] to train iris
Network to obtain the recognition rate. Data used in this paper
were taken from The Essex database which contains 7900 face
features and the result indicated that this algorithm is suitable
images taken from 395 individuals (male and female). Based to use in features extraction.
on experiments conducted, this method reached the recognition Backpropagation Neural Network (BPNN) is one of the
rate of 98.86 percent which showed the considerable excellent classifiers in patterns recognition. Despite its good
enhancement compared to previous methods ability in recognize features, BPNN often get stuck in local
minima consequently it cannot reach best error though be
Keywordsface recognition; 3d-GLCM; elman recurrent;
trained coutinuously. Furthermore, BPNN usually consumes
neural network; levenberg
plenty of time in training phase. Elman Recurrent levenberg
Neural Network (ERNN) is a Recurrent neural network that
I. INTRODUCTION using Elman network as architecture and Levenberg
Biometric system such as fingerprints, iris, or face is a Marquardt as learning algorithm. Some previous research
security system created to measure and analyze biological data showed that The Levenberg algorithm is able to learn patterns
of a person based on certain characteristics. Because of its better than BPNN in both time consumption and accuracy as
uniques characteristic which each person has different features was carried by [6], [7], and [8]. Likewise it has not been used
from others, biometric features can be used as an indicator of in face recognition specifically to train features that are
one's identity to fulfill the information security. extracted from 3D GLCM.
Since known as one of the excellent biometric systems, Based on background explained above, in this paper we
face recognition currently has been widely developed and investigated the use of 3D GLCM and ERNN to recognize
become an interesting research topic in the field of pattern face
recognition and artificial intelligence. As one of the important
biometric identification technology of human body, its II. RELATED WORKS
maturation and improvement will bring brand-new
management ways for state security, social security, bank The research in face recognition has been conducted by
many researchers. Usually, they use PCA as feature extraction
financial security, social welfare and electronic currency
algorithm as conducted by [9]. In this paper, PCA was used to
payments.
extract face features subsequently were trained by SOM
However, the performance of this system usually depends classifier to recognize the face. The result from this research
on features extraction and classifier algorithm. Hence, the use indicated a good accuracy.
of suitable algorithm will increase the accuracy of recognition
The research that closely linked to ours as was carried in
rate. The suitable algorithm is the algorithm that able to
[10]. In this paper, they used GLCM and Eucledian Distance
extract the face features so can be used to differ each face [1].

152
2016 International Seminar on Application for Technology of Information and Communication

to recognize face. The results showed that the method could C. Features Extraction
recognized face with an accuracy of 96.25 %. The gray level co-occurrence matrix (GLCM) method first
The use of Levenberg Marquardt Algorithm (LMA) as that was introduced by Haralick et al [13] in 1973 is most
classifier in face recognition carried in [11]. The result from used method for texture analysis. This method is a second
this research showed that levenberg could accelerate the order statistic method which counts the order of cooccurrence
learning time and reached better accuracy. of brightness value of paired pixels at a certain distance and a
direction. However, since the 2D-GLCM is unable to fully
The related research also was conducted in [12] which represent the texture features of the space domain images,
focused on the use of backpropagation neural network 3D-GLCM which is expanded from the original 2DGLCM as
(BPNN) in classifying images. The result from this paper was proposed in [4] can strengthen and demonstrate the
reported that BPNN was able to recognize face with the texture features of the space domain images and powerfull in
accuracy of 98.7 %. extracting the images properties as were used in some
previous methods. It can be described in the Fig 2

III. THE PROPOSED METHOD

A. Data used
We used The Yale database and The Essex database which
consist of 7900 images, each 395 individuals (male and
female) has 20 images per individuals. Contains images of
people of various racial origins, mainly of first year
undergraduate students, so the majority of individuals are
between 18-20 years old but some older individuals are
also present. Some individuals are wearing glasses and
beards. The image format is 24bit color JPEG, and using
camera S-HVS camcorder. The lightings are artificial,
mixture of tungsten and fluorescent overhead. This
research is using Essex database because many of Fig. 2. The Structure of 3D GLCM
researchers already using the database and achieve a
better accuracy than other database. The proposed method as
shown in Fig 1.
From Fig 2 above, R is the referenced pixel, N2 is the 1st
neighbour, and N2 is the 2nd neighbor. In our approach, we
used all possibilities angle formed at distance =1. Finaly we
extracted six elements from 3D GLCM formed namely :

Maximum Probability=max(pijk ) (1)


Q Q Q
Entopy = Pijk ( ln pijk ) (2)
i =1 j =1 k =1
Q Q Q
Energy = pijk
2
(3)
i =1 j =1 k =1
Q Q Q (i - i )( j j )( k k )
Correlation = Pijk (4)
i =1 j =1 k =1 i j k
Q Q Q
Contrast = Pijk (i j ) 2 + (i k ) 2 + ( j k ) 2
i =1 j =1 k =1
Fig. 1. The proposed method (5)
Q Q Q pijk
B. Preprocessing Homogeneity =
1 + ( i j ) + ( i k ) + ( j k )
2 2 2
i =1 j =1 k =1
All images used in this paper were converted into gray
(6)
scale form and filtered using Amoeba Median Filter (AMF).
This filter is a median filter that is combined with
morphological operations of amoeba. In image enhancement, Subsequently these extracted features were trained by The
AMF is a powerful filter to reduce noise without changing the ERNA/LMA to obtain the accuracy of identification.
contour. After was filtered, input image will be enhanced by
applying contrast stretching

153
2016 International Seminar on Application for Technology of Information and Communication

D. Identification TABLE I. PARAMTERS VALUE OF ERNN / LMA

This phase is divided into two steps namely training and No Parameters Value
testing. Network training process is basically a process of 1 No. Input Neurons
adjusting the weights of all nodes in input layer, hidden layer 6 neurons + bias
and output layer. The process continued till the network have 2 Hidden Layer
1 layer (7-12 neurons + bias)
reached the allowable or acceptable error. After had trained,
network would be tested using images which did not use in 3 Contex Layer 1 layer (7-12 neuron)
training phase. In this paper we used ERNN with levenberg as 4 Output Neuron 1
learning algorithm which has proven that this algorithm could
accelerate the training time
The training algorithm flowchart of ERNN / LMA as As described in Table 1, we used six neurons input + biase
despicted in Fig. 3 (total neuron in the context layer always same as input layer),
which each neuron represent one feature extracted. Namely
x1:maximum probability, x2:correlation, x3: contrast, x4:
energy, x5:entropy, x6 :homogeneity, and b1:bias. Meanwhile
on the output layer we used one neuron to recognize all faces.
For instance face with id=0, if correctly recognized will result
0 likewise face with id=1, if correctly recognized will result
as 1.

The network architecture of ERNN / LMA as can be described


in Fig 4

min < max

= 10 y _ desiredOutput( t +1)

SSEnew < SSE = * (t +1) = y _ desiredOutput (t +1) y( t +1)

y(t +1)

= /

min = 0.1

z1( t +1) z2(t +1) z3(t +1) zn ( t +1) b2( t +1)

( SSEnew SSEmin )
p = p +1 or
(p = epochmax )

Fig. 3. Training Algorithm Flowchart xi (t ) b1(t +1) x1(r t ) x2(r t ) x3(r t ) xnr (t )

In order to get better result in recognition, few parameters of


x1(t ) x6( t )
ERNN / LMA will be tested vary on network. The various
parameters and their respective values used in the training
process are shown in Table 1 Fig. 4. Architecture of ERNN / LMA

154
2016 International Seminar on Application for Technology of Information and Communication

IV. RESULTS AND DISCUSION We compared the results from both database using 6-7-1
We conducted this experiment using The Essex database structure as shown in Fig. 5
and The Yale database. Data images in each database were
divided as training set (105 images) and testing set (60
images)
In this research, we trained ERNN / LMA consecutively
using various structures to find the best parameters. Such as
structures 6-7-1 means input layer has six neurons, hidden
layer has seven neurons and output layer has one neuron. We
investigated from 100 epochs to 500 epochs and observed the
SSE obtained from each epoch as shown in Table 2 (using The
Essex database).

TABLE II. COMPARISON OF CANDIDATE ERNN / LMA STRUTURES (THE


ESSEX DATABASE)

Iterations
Structures
Fig. 5. Result from The Essex database and The Yale database using 6-7-1
100 250 350 500
structure
6-7-1 0.002934 0.000617 0.00027 0.0003
Fig 5 shows that the structure reached minimum error
6-8-1 0.06632 0.0093 0.000964 0.0007391 when trained The Essex database. However the result was not
6-9-1 0.04408 0.006 0.00073 0.0051949 significant compared with result obtained from The Yale
database. This difference occurs because of the characteristics
6-10-1 0.005 0.00097 0.00085 0.00075 of both database in taking faces.
6-11-1 0.0012 0.0006 0.0005 0.000788 To evaluate the performance of ERNN / LMA, we
6-12-1 0.0076 0.005 0.0007 0.00093 compared this classifier with Eucledian Distance (ED)
classifier to measure our correct recognition rate (CRR) as
shown in Table 4.

Based on result in Table 2, the combination of 6-7-1 TABLE IV. COMPARISON OF CORRECT RECOGNITION RATE BETWEEN
structure is better to use because it reached the smallest SSE. DIFFERENT CLASSIFIERS
Meanwhile for The Yale database, we obtained results as
The Essex Database The Yale Database
shown in Table 3. Number of
Testing Eucledian Eucledian
Images LMA/ERNN LMA/ERNN
TABLE III. COMPARISON OF CANDIDATE ERNN / LMA STRUTURES Distance Distance
(THE YALE DATABASE)
15 90.12 85.5 88.3 81.2
Iterations
Structures 30 91.3 83.6 88.9 82.1
100 250 350 500 50 92.6 84.5 89.76 83.5
6-7-1 0.014 0.0021 0.001 0.0005 70 93.6 85.4 90.5 84.8

6-8-1 0.02 0.003 0.0007 0.0009 90 94.4 86.6 92.4 85.3

6-9-1 0.4 0.0621 0.00323 0.0051949 110 95.6 87.6 93.5 85.8

6-10-1 0.25 0.0807 0.0085 0.0075 130 95.7 88.1 94.3 86.3
150 96.88 89.87 95.6 87.7
6-11-1 0.112 0.03 0.0013 0.0008
160 98.86 90.5 97.8 89.8
6-12-1 0.04 0.088 0.004 0.0093

Results from Table 4 showed that ERNN/LMA reached


From Table 3, the structure of 6-7-1 reached SSE of CRR of 98.86 % from The Essex database and 97.8 % from
0.0005 means that this combination should be used as the best The Yale database which is better than ED classifier. The
structure for testing phase. ERNN/LMA is also better when tested using images from The
Yale database.

155
2016 International Seminar on Application for Technology of Information and Communication

In order to measure our method performance, we have REFERENCES


compared with some previous methods which also had used
the same database with our research and the results as can be [1] Liu, C. The Development Trend of Evaluating Face-Recognition
shown in Table 5. Technology. 2014 International Conference on Mechatronics and
Control (ICMC) pp. 1540-1544. July 3 - 5, 2014
[2] Eleyan, A., and Demirel, H. Co-Occurrence based Statistical Approach
TABLE V. COMPARISON WITH SOME PREVIOUS METHODS
for Face Recognition. International Symposium on Computer and
Sciences 2009. pp. 611-615 Sept. 2009
Methods CRR (%)
[3] Appati, J., Fosu, G., Gogovi, G. Face Feature Extraction for
Sekhon et. al 86 Recognition Using Radon Transform. American Journal of Engineering
Research (AJER). Vol 03, Issue 03, pp 219-224. 2014
Khalajzadeh et. al 86.06
[4] Chen, W.S., Huang, R.H., dan Hsieh L., IrisRecognition Using 3D Co-
Assadi et. al 88.96 Occurrence Matrix, Springer-Verlag Berlin Heidelberg, Germany,
LNCS, 5558, pp. 1122-1131. 2009
Narang et. al 96.1 [5] Dillak, R.Y., Bintiri, M.G. A Novel Approach for Iris Recognition.
Ch'ng et. al 96.5 The IEEE Region 10 Symposium (TENSYMP), 09-11 May 2016
[6] Nawi, N.H., Khan, A., Rehman, M.Z., Aziz, M.A., Herawan,T.,
Pankaj et. al 98.7 Abawajy, J.H. An Accelerated Particle Swarm Optimization Based
Levenberg Marquardt Back Propagation Algorithm. ICONIP 2014,
Proposed Method 98.86
Part II, LNCS 8835, pp. 245253, 2014
[7] Faisal, T., Nasir, M., Taib., Ibrahim, F. Neural network diagnostic
system for dengue patients risk classification. J Med Syst. Vol. 36, pp.
Based on Table 5, the comparison of some previous 661676. 2012
methods indicated that our proposed method reached CRR of [8] Bascil, M.S., Temurtas, F., A Study on Hepatitis Disease Diagnosis
98.86 % higher than others Using Multilayer Neural Network with Levenberg Marquardt Training
Algorithm. J Med Syst. Vol 35 pp. 433436. 2011
[9] Anggraini D. R, Face Recognition Using Principal Component
Analysis and Self Organizing Maps Third ICT International Student
V. CONCLUSION Project Conference (ICT-ISPC2014), pp. 91-94, 2014
[10] Huang, Y., Chen, J., Liao, J., Kuang, P. Face Recognition Based On
In this paper, we have presented a novel approach for face Texture Features Using Hybrid Projection Function. International
recognition. Based on our experiments, it could be concluded Conference on Wavelet Active Media Technology and Information
that 3D GLCM and ERNN/LMA were suitable to be used in Processing (ICWAMTIP), pp. 82 85, 17-19 Dec. 2012
face recognition. The feature extraction method used was able [11] Narang, G., Singh, S., Narang, A. Robust face recognition method
based on SIFT features using Levenberg-Marquardt Backpropagation
to extract characteristic features of face images. Meanwhile neural networks. 6th International Congress on Image and Signal
ERNN/LMA was also able to recognize such features with Processing (CISP). pp . 1000 1005, 16-18 Dec. 2013
good accuracy. The proposed method reached CRR of 98.86 [12] Pankaj, D.S ., Wilscy, M. Face Recognition Using Fuzzy Neural
% which improved the CRR from some previous methods. Network Classifier. PDCTA 2011, CCIS 203, pp. 53-62, Springer-
However, in next research it should be combined with more Verlag Berlin Heidelberg 2011
than one feature extraction algorithm such as PCA and 3D [13] Haralick, R.M., Shanmugam, K., Dinstein, L.: Textural Features for
Image Classification. IEEE Trans. on Systems, Man, and Cybernetics 3,
GLCM. 610621 1973
[14] The Essex Face Database, cswww . essex.ac . uk/mv/allfaces/index.html
ACKNOWLEDGMENT [15] The Yale Face Database,cvc.yale.edu/projects/yalefaces/yalefaces.html
We expressed our gratitude to The Essex database [14] and
The Yale database [15] that have provided face images used in
this research

156

S-ar putea să vă placă și