An Efficient Face Recognition System Using a New Optimized ...
An Efficient Face Recognition System Using a New Optimized Localization Method Hamidreza Rashidy Kanan, Karim Faez and Mehdi Ezoji Image Processing and Pattern Recognition Lab., Electrical Engineering Department, Amirkabir University of Technology (Tehran Polytechnic), Hafez Avenue, Tehran, Iran, 15914 {rashidykanan, kfaez, m_ezoji}@aut.ac.ir Abstract In this paper a system is developed for face recognition processes. Preprocessing and face localization is necessary to obtain a high classification rate in face recognition tasks. In this study after preprocessing of face images, for omitting the redundant information such as background and hair, the oval shape of face is approximated by an ellipse using shape information. Then the parameters (orientation and center coordinates) of this ellipse are optimized using Genetic Algorithm (GA). High order Pseudo Zernike Moment Invariant (PZMI) which has useful properties is utilized to produce feature vectors. Also Radial Basis Function Neural Network (RBFNN) with HLA learning rule has been used as a classifier. Simulation results on ORL database indicate that the error rate of proposed system which uses genetic algorithm for optimizing the face localization step is lower than an older system which described in [2].
1. Introduction Automatic Face recognition has been an active research topic due to their extensive range of applications, such as access control systems, criminal identification and authentication in secure systems [1]. Face region segmentation or face localization is a fundamental step in the process of face recognition. The accuracy of the localized face center coordinates and orientation has a heavy influence on the recognition performance [2]. Genetic algorithms (GA’s) are optimization techniques based on the mechanics of natural selection [3]. Because of their advantages, recently, GA’s have been widely used as a tool in pattern recognition applications. PZMI is a moment that has been frequently utilized for a number of image processing tasks [4]. The nature of rotation invariance makes the PZMI descriptors very valuable.
RBF neural network has been found to be very attractive for many engineering problems due to its simple topological structure, its locally tuned neurons and its ability to have a fast learning algorithm [2]. In this paper, we have proposed a new method for face localization that is utilizes shape information and genetic algorithms approach. After preprocessing, connected components are determined by applying a region growing algorithm (coarse segmentation). Then the best-fit ellipse for the face area is computed. We have used GA to find the best location (including the best orientation and the best position) of face in image. High order PZMI has been used for feature extraction of the preprocessed face images. Also RBF neural network which have been found to be very attractive for many engineering problem is employed in this system. The organization of this paper is as follows: Section 2 presents preprocessing and face localization. Feature extraction and Classifier are described in section 3 and 4 respectively and finally, section 5 and 6 attain the experimental results and conclusion.
2. Preprocessing and Face Localization One of the key problems in building automated systems that perform face recognition task is face localization. Many algorithms have been proposed for face localization and detection, which are based on using shape, color information, motion etc. A critical survey on face localization and detection can be found in [5]. After preprocessing (histogram equalization) of the facial images, we extract connected components by applying a region growing algorithm. Then the ellipse that is a good approximation of connected components is selected by the shape information method [6]. We have used genetic algorithm to find the best location (including the best orientation and the best position) of face in image. Architecture of this method is shown in Fig. 1.
0-7695-2521-0/06/$20.00 (c) 2006 IEEE
Image Enhancement
Coarse Segmentation
Face Localization based on Shape Information
Input Face Image
θ = 0.5 × Arc tan(2µ11 ( µ 20 − µ 02 )) Where µ pq denotes the central moment
Optimization of Face Location using GA’s Localized Face Image
Figure 1. Basic flowchart of the algorithm At the first, we extract the face region from the input image using a region growing method [7] that is described in the following: Let S(x, y) denote the square region with center (x, y) and side length d. then C(x, y) is given by:
C ( x, y ) =
y+d / 2
∑
j = y −d / 2
Vr ( j ) +
x+d / 2
∑ V (i ) c
(1)
i = x− d / 2
number of pixels (i, j), x − d / 2 ≤ i ≤ x + d / 2 , such that one of the I(i-1, j) and I(i, j) is greater than µ plus K and the other is smaller than µ mines K where µ is the average intensity of pixels (i, j) and K is a constant. Vc (i) denote the number of pixels (i,j),
y − d / 2 ≤ j ≤ y + d / 2 , such that one of the I(i, j-1) and I(i, j) is greater than µ plus K and the other is smaller than µ minus K. The oval shape of face can be approximated by an ellipse. Therefore looking for faces in an image means to detect objects with nearly elliptical shape [6]. The advantage of the region-based method is its robustness in the presence of the noise and changes in illumination. In the region-based method, the connected components are determined by applying a region growing algorithm, then for each connected component, the best-fit ellipse having the best minimum size, is computed using the properties of the geometric moments. To find a face region, an ellipse model with five parameters is used: X 0 , Y0 are the centers of the ellipse, θ is the orientation, α and β are the major and minor axes of the ellipse respectively, as shown in Fig. 2. To calculate these parameters, we use the best fit ellipse equations. Therefore, the center of the ellipse is given by the center of gravity of the connected components. The orientation θ of the ellipse can be calculated by determining the least moment of inertia:
of the
connected components. The length of the major and the minor axes of the best-fit ellipse can also be computed by evaluating the moment of inertia. With the least and the greatest moment of inertia of an ellipse defined as:
I Min = ∑∑ [( x − x0 ) cosθ − ( y − y 0 ) sin θ ] (3) 2
x
y
I Max = ∑∑ [( x − x0 ) sin θ − ( y − y0 ) cosθ ] (4) 2
x
y
The lengths of the major and minor axes are calculated as follows:
α = (1 / π ) × [I 3 Max I Min ]
18
(5)
β = (1 / π ) × [I 3 Min I Max ]
18
Where Vr ( j ) and Vc (i) are the mean crossing number of row j and column i which are defined as follows. Let I(i, j) denote the intensity values of pixels (i, j) in the image. Then, for each row j, Vr ( j ) represents the
(2)
Y
(6)
( X 0 , Y0 ) β α θ
X
Figure 2. Face model based on ellipse model Because of the background effect, the ellipse that is obtained in previous part is not the best-fit ellipse. So, we must optimize its orientation and center coordinates. We calculate the major and minor axes of best-fit ellipse by the equation (5) and (6) and its exact orientation and center coordinates which uses genetic algorithm is described in following. In the optimization section, the population was initialized with random points. In order to select the individuals for the next generation, GA’s roulette wheel selection method was used. Further genetic parameters are: Population size: 30 Chromosome length: 23 Probability of crossover: 0.8 Probability of mutation: 0.003 We use two different coding methods (gray and binary) for genotype coding and the symmetric of localized image with respect to major axis of obtained ellipse for fitness function. With this optimization method, the position of ellipse is found precisely while the Tilt and Translation Errors are lower before optimization. For more details, please see [8].
0-7695-2521-0/06/$20.00 (c) 2006 IEEE
3. PZMI Feature Extraction The advantages of considering orthogonal moments are that they are shift, rotation and scale invariant and very robust in the presence of noise [2]. Pseudo Zernike polynomials are orthogonal sets of complex-valued polynomials defined as:
and hidden layers have unit weights and, as a result, do not have to be trained. The RBF units are also fully connected to the output layer. 1
R nm ( x , y ) =
n− m|
∑
2
2
D n ,|m |, s ( x + y )
2
(8)
( 2 n + 1 − s )! (9) s!( n − | m | − s )!( n − | m | − s + 1)! The PZMI of order n and repetition m can be computed using the scale invariant central moments CM p, q and the radial geometric moments RM p, q as D n ,|m|, s = ( −1) S
follow:
PZM
nm
n +1
=
π
n −|m |
∑D
k
m
∑ ∑ ( )( )
n ,| m |, s ( n − m − s ) even , s = 0 a =0 b= 0
k a
m b
( − j ) b CM 2 k + m − 2 a − b , 2 a + b +
n +1
π
n −|m |
∑D
m
d a
m b
b
(10)
2 d + m − 2 a −b , 2 a + b
Where k = (n−s −m) / 2, d = (n − s − m −1) / 2 and also CM p , q and RM p , q are as follow:
CM p ,q =
RM p,q
µ p,q M 00 x
(11)
( p + q +2 ) / 2
∑∑ =
1
y
f ( x, y )( xˆ 2 + yˆ 2 ) 2 xˆ p yˆ q ( p+ q+ 2)
n
s r
Wrs
For designing a classifier based on RBF neural network, we have set the number of input nodes in the input layer of neural network equal to the number of feature vector elements. The number of nodes in the output layer is set to the number of image classes. The RBF units are selected using the clustering procedure [2]. Training of the RBF neural network involves estimating output connection weights, centers and widths of the RBF units. We use the HLA method, which combines the gradient method and the linear least squared method for training RBF neural network [2].
5. Exprimental Results
∑ ∑ ( )( ) d
n , |m |, s a =0 b=0 ( n − m − s ) odd , s = 0
( − j ) RM
2 3
Figure 3. RBF neural network structure
n−s 2
s=0
Where
1 2
y (7) V nm ( x , y ) = R nm ( x , y ) exp( jm tan −1 ( )) x 2 2 Where x + y ≤ 1 , n ≥ 0 , | m |≤ n and Radial polynomials { Rnm } are defined as:
W11
1
(12)
M 00 Where xˆ = x − x0 , yˆ = y − y0 and x0 , y0 , 2
To check the utility of our proposed algorithm, experimental studies are carried out on the ORL database images of Cambridge University. This database contains 400 facial images from 40 individuals in different states. The total number of images for each person is 10. After preprocessing by means of histogram equalization and face localization that is described in part 2, we have face images in which the redundant information such as background and hair are omitted and also the tilt and translation errors of localized images are reduced. Fig. 4 shows the localized faces before and after using the genetic algorithm optimization.
M pq and µ pq have been defined in previous parts.
4. RBF Neural Network Classifier An RBF neural network structure is shown in Fig. (3) The construction of the RBF neural network involves three different layers with feed forward architecture. The input layer of this network is a set of “n” units that are fully connected to the hidden layer with “r” hidden units. Connections between the input
(a)
(b)
(c)
Figure 4. Localized Face: (a) Original image (b) Before optimization (c) After optimization
0-7695-2521-0/06/$20.00 (c) 2006 IEEE
%Error Rate
In the feature extraction step, because of PZMI advantages, we extract PZMI of localized face images with different orders and different element numbers. Simulation results (Fig. 5) show that the classification error rate using PZMI with the order of 9 or higher is lower than the other orders. 45 40 35 30 25 20 15 10 5 0
1
2
3
4
5
6
7
8
9
10
Moment Order
Figure 5. Classification error rate for different order of PZMI In classifier step, we use RBF neural network with HLA learning algorithm. We use 5 images from each class randomly for training and the rest for the test. There is no overlap between the training and test sets. The RBF classifier is trained using the HLA method with the training sets, and finally the classifier error rate is computed using the test images. In this study, the classifier error rate is computed as the number of misclassifications in the test phase over the total number of test images. Simulation results on the ORL database show that the error rate in face recognition system with the face localization step is lower than the same system without the face localization step. Furthermore, our method for localization of faces which uses genetic algorithm improves the performance of classifier more than the method defined in [2]. These results are summarized in Table 1 and 2. Table 1. Experimental results for different order of PZMI Features Vectors Category
Number of Feature Elements
n=1, 2, …, 6 n=4, 5, 6, 7 n=6, 7, 8 n=9, 10
27 26 24 21
Testing Phase Classification Error Rate Proposed Ref Method [2] 6.9% 7.5% 4.4% 4.5% 2.8% 3% 1% 1.3%
Table 2. Experimental results for the best order of PZMI Face Recognition System Without Face Localization With Localization (Ref. [2]) With Localization (Proposed Method)
Tilt Error
Errors Translation Error
Classification Error Rate
**
**
5.2%
17.5
26.7
1.3%
4.2
5.4
1%
6. Conclusion In this research, we have evaluated two kinds of face localization methods for face recognition. After preprocessing, we find the exact location of oval shape of face (best fit ellipse) in database images with genetic algorithm. Then the high orders PZMI are used as a feature vectors which are fed to classifier stage. We have used RBF neural network classifier with HLA learning algorithm. Experimental results on ORL database indicate that the error rate of our system which uses genetic algorithm for optimizing the face localization step is lower than the other systems which have the same feature extraction and classifier stages.
7. Acknowledgment This research was supported by the Telecommunication Research Center (ITRC).
Iran
8. References [1] Chellappa R., Wilson C.L., Sirohry S., “Human and Machine Recognition of faces: A survey”, Proceedings of the IEEE, Vol. 83, No. 5, 1995, P.P. 705-740. [2] H. Haddadnia, K. Faez and M. Ahmadi, “AN EFFICIENT HUMAN FACE RECOGNITION SYSTEM USING PSEUDO ZERNIKE MOMENT INVARIANT AND RADIAL BASIS FUNCTION NEURAL NETWORK”, Int. Journal of Patt. Recogn. and Artificial Intelligence, Vol. 17, No. 1, 41-62, 2003. [3] M. Srinivas, Lalit M. Patnik, “Genetic Algorithms: A Survey”, IEEE, 1994. [4]. C. H. Teh and R. T. Chin, “On image analysis by the methods of moments”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 10, No. 4, 1988, PP. 496-513. [5] J. Daugman, “Face detection: a survey”, Comput. Vis. Imag. Underst. 83, 3, 2001, P.P. 236-274. [6] K. Sobotta and I. Pitas, “Face localization and facial feature extraction based on shape and color information”, IEEE Int. Conf. on Image Processing, Vol. 3, Lausanne, Switzerland, 16-19 September, 1996, P.P. 483-486. [7]. C.H. Lin and J.L. Wu, “Automatic facial feature extraction by genetic algorithms”, IEEE Transaction of Image Processing, Vol. 8, No. 6, 1999, p.p. 834-845. [8] H. R. Kanan, K. Faez, S. Mozaffari and M. Ezoji, “FACE LOCALIZATION USING SHAPE INFORMATION AND GENETIC ALGORITHM”, Proceeding of the First Int. Conference on Modeling, Simulation and Applied Optimization, Sharjah, U.A.E., 2005.
**: Not defined
0-7695-2521-0/06/$20.00 (c) 2006 IEEE
1. Introduction Automatic Face recognition has been an active research topic due to their extensive range of applications, such as access control systems, criminal identification and authentication in secure systems [1]. Face region segmentation or face localization is a fundamental step in the process of face recognition. The accuracy of the localized face center coordinates and orientation has a heavy influence on the recognition performance [2]. Genetic algorithms (GA’s) are optimization techniques based on the mechanics of natural selection [3]. Because of their advantages, recently, GA’s have been widely used as a tool in pattern recognition applications. PZMI is a moment that has been frequently utilized for a number of image processing tasks [4]. The nature of rotation invariance makes the PZMI descriptors very valuable.
RBF neural network has been found to be very attractive for many engineering problems due to its simple topological structure, its locally tuned neurons and its ability to have a fast learning algorithm [2]. In this paper, we have proposed a new method for face localization that is utilizes shape information and genetic algorithms approach. After preprocessing, connected components are determined by applying a region growing algorithm (coarse segmentation). Then the best-fit ellipse for the face area is computed. We have used GA to find the best location (including the best orientation and the best position) of face in image. High order PZMI has been used for feature extraction of the preprocessed face images. Also RBF neural network which have been found to be very attractive for many engineering problem is employed in this system. The organization of this paper is as follows: Section 2 presents preprocessing and face localization. Feature extraction and Classifier are described in section 3 and 4 respectively and finally, section 5 and 6 attain the experimental results and conclusion.
2. Preprocessing and Face Localization One of the key problems in building automated systems that perform face recognition task is face localization. Many algorithms have been proposed for face localization and detection, which are based on using shape, color information, motion etc. A critical survey on face localization and detection can be found in [5]. After preprocessing (histogram equalization) of the facial images, we extract connected components by applying a region growing algorithm. Then the ellipse that is a good approximation of connected components is selected by the shape information method [6]. We have used genetic algorithm to find the best location (including the best orientation and the best position) of face in image. Architecture of this method is shown in Fig. 1.
0-7695-2521-0/06/$20.00 (c) 2006 IEEE
Image Enhancement
Coarse Segmentation
Face Localization based on Shape Information
Input Face Image
θ = 0.5 × Arc tan(2µ11 ( µ 20 − µ 02 )) Where µ pq denotes the central moment
Optimization of Face Location using GA’s Localized Face Image
Figure 1. Basic flowchart of the algorithm At the first, we extract the face region from the input image using a region growing method [7] that is described in the following: Let S(x, y) denote the square region with center (x, y) and side length d. then C(x, y) is given by:
C ( x, y ) =
y+d / 2
∑
j = y −d / 2
Vr ( j ) +
x+d / 2
∑ V (i ) c
(1)
i = x− d / 2
number of pixels (i, j), x − d / 2 ≤ i ≤ x + d / 2 , such that one of the I(i-1, j) and I(i, j) is greater than µ plus K and the other is smaller than µ mines K where µ is the average intensity of pixels (i, j) and K is a constant. Vc (i) denote the number of pixels (i,j),
y − d / 2 ≤ j ≤ y + d / 2 , such that one of the I(i, j-1) and I(i, j) is greater than µ plus K and the other is smaller than µ minus K. The oval shape of face can be approximated by an ellipse. Therefore looking for faces in an image means to detect objects with nearly elliptical shape [6]. The advantage of the region-based method is its robustness in the presence of the noise and changes in illumination. In the region-based method, the connected components are determined by applying a region growing algorithm, then for each connected component, the best-fit ellipse having the best minimum size, is computed using the properties of the geometric moments. To find a face region, an ellipse model with five parameters is used: X 0 , Y0 are the centers of the ellipse, θ is the orientation, α and β are the major and minor axes of the ellipse respectively, as shown in Fig. 2. To calculate these parameters, we use the best fit ellipse equations. Therefore, the center of the ellipse is given by the center of gravity of the connected components. The orientation θ of the ellipse can be calculated by determining the least moment of inertia:
of the
connected components. The length of the major and the minor axes of the best-fit ellipse can also be computed by evaluating the moment of inertia. With the least and the greatest moment of inertia of an ellipse defined as:
I Min = ∑∑ [( x − x0 ) cosθ − ( y − y 0 ) sin θ ] (3) 2
x
y
I Max = ∑∑ [( x − x0 ) sin θ − ( y − y0 ) cosθ ] (4) 2
x
y
The lengths of the major and minor axes are calculated as follows:
α = (1 / π ) × [I 3 Max I Min ]
18
(5)
β = (1 / π ) × [I 3 Min I Max ]
18
Where Vr ( j ) and Vc (i) are the mean crossing number of row j and column i which are defined as follows. Let I(i, j) denote the intensity values of pixels (i, j) in the image. Then, for each row j, Vr ( j ) represents the
(2)
Y
(6)
( X 0 , Y0 ) β α θ
X
Figure 2. Face model based on ellipse model Because of the background effect, the ellipse that is obtained in previous part is not the best-fit ellipse. So, we must optimize its orientation and center coordinates. We calculate the major and minor axes of best-fit ellipse by the equation (5) and (6) and its exact orientation and center coordinates which uses genetic algorithm is described in following. In the optimization section, the population was initialized with random points. In order to select the individuals for the next generation, GA’s roulette wheel selection method was used. Further genetic parameters are: Population size: 30 Chromosome length: 23 Probability of crossover: 0.8 Probability of mutation: 0.003 We use two different coding methods (gray and binary) for genotype coding and the symmetric of localized image with respect to major axis of obtained ellipse for fitness function. With this optimization method, the position of ellipse is found precisely while the Tilt and Translation Errors are lower before optimization. For more details, please see [8].
0-7695-2521-0/06/$20.00 (c) 2006 IEEE
3. PZMI Feature Extraction The advantages of considering orthogonal moments are that they are shift, rotation and scale invariant and very robust in the presence of noise [2]. Pseudo Zernike polynomials are orthogonal sets of complex-valued polynomials defined as:
and hidden layers have unit weights and, as a result, do not have to be trained. The RBF units are also fully connected to the output layer. 1
R nm ( x , y ) =
n− m|
∑
2
2
D n ,|m |, s ( x + y )
2
(8)
( 2 n + 1 − s )! (9) s!( n − | m | − s )!( n − | m | − s + 1)! The PZMI of order n and repetition m can be computed using the scale invariant central moments CM p, q and the radial geometric moments RM p, q as D n ,|m|, s = ( −1) S
follow:
PZM
nm
n +1
=
π
n −|m |
∑D
k
m
∑ ∑ ( )( )
n ,| m |, s ( n − m − s ) even , s = 0 a =0 b= 0
k a
m b
( − j ) b CM 2 k + m − 2 a − b , 2 a + b +
n +1
π
n −|m |
∑D
m
d a
m b
b
(10)
2 d + m − 2 a −b , 2 a + b
Where k = (n−s −m) / 2, d = (n − s − m −1) / 2 and also CM p , q and RM p , q are as follow:
CM p ,q =
RM p,q
µ p,q M 00 x
(11)
( p + q +2 ) / 2
∑∑ =
1
y
f ( x, y )( xˆ 2 + yˆ 2 ) 2 xˆ p yˆ q ( p+ q+ 2)
n
s r
Wrs
For designing a classifier based on RBF neural network, we have set the number of input nodes in the input layer of neural network equal to the number of feature vector elements. The number of nodes in the output layer is set to the number of image classes. The RBF units are selected using the clustering procedure [2]. Training of the RBF neural network involves estimating output connection weights, centers and widths of the RBF units. We use the HLA method, which combines the gradient method and the linear least squared method for training RBF neural network [2].
5. Exprimental Results
∑ ∑ ( )( ) d
n , |m |, s a =0 b=0 ( n − m − s ) odd , s = 0
( − j ) RM
2 3
Figure 3. RBF neural network structure
n−s 2
s=0
Where
1 2
y (7) V nm ( x , y ) = R nm ( x , y ) exp( jm tan −1 ( )) x 2 2 Where x + y ≤ 1 , n ≥ 0 , | m |≤ n and Radial polynomials { Rnm } are defined as:
W11
1
(12)
M 00 Where xˆ = x − x0 , yˆ = y − y0 and x0 , y0 , 2
To check the utility of our proposed algorithm, experimental studies are carried out on the ORL database images of Cambridge University. This database contains 400 facial images from 40 individuals in different states. The total number of images for each person is 10. After preprocessing by means of histogram equalization and face localization that is described in part 2, we have face images in which the redundant information such as background and hair are omitted and also the tilt and translation errors of localized images are reduced. Fig. 4 shows the localized faces before and after using the genetic algorithm optimization.
M pq and µ pq have been defined in previous parts.
4. RBF Neural Network Classifier An RBF neural network structure is shown in Fig. (3) The construction of the RBF neural network involves three different layers with feed forward architecture. The input layer of this network is a set of “n” units that are fully connected to the hidden layer with “r” hidden units. Connections between the input
(a)
(b)
(c)
Figure 4. Localized Face: (a) Original image (b) Before optimization (c) After optimization
0-7695-2521-0/06/$20.00 (c) 2006 IEEE
%Error Rate
In the feature extraction step, because of PZMI advantages, we extract PZMI of localized face images with different orders and different element numbers. Simulation results (Fig. 5) show that the classification error rate using PZMI with the order of 9 or higher is lower than the other orders. 45 40 35 30 25 20 15 10 5 0
1
2
3
4
5
6
7
8
9
10
Moment Order
Figure 5. Classification error rate for different order of PZMI In classifier step, we use RBF neural network with HLA learning algorithm. We use 5 images from each class randomly for training and the rest for the test. There is no overlap between the training and test sets. The RBF classifier is trained using the HLA method with the training sets, and finally the classifier error rate is computed using the test images. In this study, the classifier error rate is computed as the number of misclassifications in the test phase over the total number of test images. Simulation results on the ORL database show that the error rate in face recognition system with the face localization step is lower than the same system without the face localization step. Furthermore, our method for localization of faces which uses genetic algorithm improves the performance of classifier more than the method defined in [2]. These results are summarized in Table 1 and 2. Table 1. Experimental results for different order of PZMI Features Vectors Category
Number of Feature Elements
n=1, 2, …, 6 n=4, 5, 6, 7 n=6, 7, 8 n=9, 10
27 26 24 21
Testing Phase Classification Error Rate Proposed Ref Method [2] 6.9% 7.5% 4.4% 4.5% 2.8% 3% 1% 1.3%
Table 2. Experimental results for the best order of PZMI Face Recognition System Without Face Localization With Localization (Ref. [2]) With Localization (Proposed Method)
Tilt Error
Errors Translation Error
Classification Error Rate
**
**
5.2%
17.5
26.7
1.3%
4.2
5.4
1%
6. Conclusion In this research, we have evaluated two kinds of face localization methods for face recognition. After preprocessing, we find the exact location of oval shape of face (best fit ellipse) in database images with genetic algorithm. Then the high orders PZMI are used as a feature vectors which are fed to classifier stage. We have used RBF neural network classifier with HLA learning algorithm. Experimental results on ORL database indicate that the error rate of our system which uses genetic algorithm for optimizing the face localization step is lower than the other systems which have the same feature extraction and classifier stages.
7. Acknowledgment This research was supported by the Telecommunication Research Center (ITRC).
Iran
8. References [1] Chellappa R., Wilson C.L., Sirohry S., “Human and Machine Recognition of faces: A survey”, Proceedings of the IEEE, Vol. 83, No. 5, 1995, P.P. 705-740. [2] H. Haddadnia, K. Faez and M. Ahmadi, “AN EFFICIENT HUMAN FACE RECOGNITION SYSTEM USING PSEUDO ZERNIKE MOMENT INVARIANT AND RADIAL BASIS FUNCTION NEURAL NETWORK”, Int. Journal of Patt. Recogn. and Artificial Intelligence, Vol. 17, No. 1, 41-62, 2003. [3] M. Srinivas, Lalit M. Patnik, “Genetic Algorithms: A Survey”, IEEE, 1994. [4]. C. H. Teh and R. T. Chin, “On image analysis by the methods of moments”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 10, No. 4, 1988, PP. 496-513. [5] J. Daugman, “Face detection: a survey”, Comput. Vis. Imag. Underst. 83, 3, 2001, P.P. 236-274. [6] K. Sobotta and I. Pitas, “Face localization and facial feature extraction based on shape and color information”, IEEE Int. Conf. on Image Processing, Vol. 3, Lausanne, Switzerland, 16-19 September, 1996, P.P. 483-486. [7]. C.H. Lin and J.L. Wu, “Automatic facial feature extraction by genetic algorithms”, IEEE Transaction of Image Processing, Vol. 8, No. 6, 1999, p.p. 834-845. [8] H. R. Kanan, K. Faez, S. Mozaffari and M. Ezoji, “FACE LOCALIZATION USING SHAPE INFORMATION AND GENETIC ALGORITHM”, Proceeding of the First Int. Conference on Modeling, Simulation and Applied Optimization, Sharjah, U.A.E., 2005.
**: Not defined
0-7695-2521-0/06/$20.00 (c) 2006 IEEE
