Journal of Applied Computer Science & Mathematics, no. 12 (6) /2012, Suceava
Enhanced SVD Based Face Recognition
1
Muhammad SHARIF,
2Saad ANIS,
3Mudassar RAZA,
4Sajjad MOHSIN
Department of Computer Science, COMSATS Institute of Information TechnologyWah Cantt., 47040, Pakistan 1
[email protected], [email protected], [email protected],4 [email protected]
Abstract-One of the demanding tasks in face recognition is to handle illumination and expression variations. A lot of research is in progress to overcome such problems. This paper addresses the preprocessing method that is composed of grouping SVD perturbation and DWT. The proposed technique also performs well under one picture per person scenarios. The resulting image of this method is fed in to the simple SVD algorithm for face recognition. This paper performs its accuracy test on ORL, Yale, PIE and AR databases and focuses on the illumination problems.
Keywords: Eigen face, Singular value decomposition (SVD), Discrete wavelet transforms (DWT), SVD perturbation.
I INTRODUCTION
Face identification is a constructive research fields in image processing. Due to its wide range of applications in real (everyday) life, it gains more importance among the researchers. A number of face recognition techniques are available which can be classified into geometric feature and template based techniques [2].
First category makes use of local facial appearance portions such as eyes, nose, mouth and the distance between these features. Elastic bunch graph matching (EBGM) [3] is the most widely used technique in this category whereas; the second category uses the whole face as input for recognition. The most basic technique for this category is Eigen face [4] which projects a high dimensional face space into lower dimensional Eigen face subspace.
SVD [5] is a projection based recognition technique. It requires less space and is also an efficient approach than Eigen face. A comprehensive study on available face recognition techniques can be made at Zhao et al [6]. But the performance of these algorithms decreases under varying illumination conditions and expressions.
One sample per individual problem has also gained much importance in the recent era which is a great challenge for researchers. Zhi-Hua et al [6] defines the one sample problem for one image of a person; the purpose is to classify an individual from the database by using the single pose of any person and identifying the similar person.
According to Zhang et. al. [7], the performance of PCA, 2DPCA [8] and many other face recognition techniques degrades when only one image is available as a training sample. They also provide a detailed survey on the
techniques available that address the one sample problem [17].
This paper focuses on the preprocessing stage for face recognition so that useless information can be removed from face and the feature extraction stage with useful information can be provided to work further. For a very reason, this paper first makes use of SVD perturbation, then applies discrete wavelet transforms on face images and finally, uses these images as input to FR algorithm named as “SVD based projection for face recognition” [18].
II RELATEDWORK
Much of the work has been done to cope with the challenges of small sample size. Most of these techniques are focused on the preprocessing stage.
In (PC)2A, Wu and Zhou et al [9] addressed the one sample problem and enhanced the simple PCA by enriching the important information on face. Their work is to find out the first order projection (horizontal and vertical) of face and combine it with the original face image to get a more enhanced image. After that, they applied PCA on this enhanced image [6]. Zhi-Hua and Zang et al [10] presented an enhanced version of (PC) 2A named E (PC) 2A [19].
“Eigen hills” [11] is another preprocessing technique that is based on applying Eigen face method on the extracted edges. This method cannot perform well under varying expressions because edges are more sensitive to the expression variations [12].
Songcan Chen et al [1] developed SVD perturbation method to address the single sample image for a person. The concept behind this is to derive a new enhanced image by increasing the singular values of original face image. Derived images are also treated as independent images to make the one sample database a normal sample size database. Finally, the PCA is applied on both (original and derived) face images.
Wavelet transforms are now widely used to handle such variations. GangYan et al [13] presented a characteristics selection method that is based upon the combination of discrete wavelet and discrete cosine transform. Y.Z.Goh et al [14] devised a method to remove illumination effect from the face image by using discrete wavelet transforms. The main idea behind this is to set the coefficients of low frequency
Computer Science Section
component to zero which makes the system invariant to illumination.
Dimension reduction is also an area of concern. One purpose of conventional methods such as PCA and 2DPCA is to reduce the dimension of face space. Gang Yan et al. [10] presented a method that is based on the combination of Discrete Wavelet Transform and Discrete Cosine Transform. This combination is used to find features of face images. The use of DCT reduces the dimension of the original face image and retains only the useful information in feature subspace. The SVM [11] is then used to categorize these feature vectors.
III PROPOSED TECHNIQUE
This paper has two main objectives to fulfill. The first one is to cope up with the varying illuminations and expressions. The second intention is to provide an answer for one sample dilemma. This paper presents a preprocessing technique that has three steps. In the first step, the histogram equalization is applied to the face image, so that large intensity variations can be handled to some extent at this step. In the second step, the paper makes use of singular value decomposition (SVD) perturbation, which at first, applies the SVD decomposition on face image (I).
L D R, ,
svd I
(1) Here L and R are left and right odd vectors respectively, D is the diagonal matrix of particular values.SVD perturbation uses these singular values to make the derived image (J).
i
J L D R (2)
where i varies between 1 and 2. Finally the derived image is combined with the original image.
1 a J C a (3) where a is the combination parameter and it varies from 0 to 1.The second step fades out the unimportant information from the face image and the salient features of face become richer. Output of this step is shown in Figure 4. Individually, this step is not able to perform well under varying conditions. So finally, this paper makes use of wavelet transforms to handle those variations. Wavelet transforms decompose a face image into a number of coefficients that represent an image into different frequency sub bands.
In the third step, the technique applies discrete wavelet transforms (DWT) on the processed image of second step. Two dimensional discrete wavelet transforms decompose an image into the approximation and detailed coefficients (horizontal, vertical and diagonal) as shown in figure1.
Fig1. Flow diagram of 2D-DWT
Approximation variables are the low frequency components (LL) and detail variables are high frequency components and are represented as LH, HL, and HH.
LL1= [Lc *[Lr*I(x,y)]2↓1] 1↓2 LH1= [Hc *[Lr*I(x,y)] 2↓1] 1↓2 HL1= [Lc *[Hr*I(x,y)] 2↓1] 1↓2 HH1= [Hc*[Hr*I(x,y)] 2↓1] 1↓2
where L is the low pass filter and H is the high pass filter. I is an original image.
Low frequency component contains the useful information that is helpful in recognition process. Also, unimportant information gets lost in this component, by which recognition stage can fully focus on important features. The (LL) component is ineffective with illumination changes and expression variations. So this technique takes an advantage of this fact and uses these low frequency components as input images to the recognition algorithm. Finally, the approximation coefficient is used as an input to the SVD based projection for face recognition, where the mean image is calculated by using preprocessed images as:
1 1 M M i i
(4) and then feature vectors are computed using SVD. The proposed method computes performance experiment using the orthogonal ‘haar’ wavelet function as a mother wavelet. The system decomposes an image up to level 1 as shown in figure 5.Many researchers used level2 decomposition but the idea behind level1 decomposition is that it restrains enough energy or information that is useful for recognition. As decomposition goes deeper (level2, level3…), valuable information gets lost form face image that really degrades the recognition ratio.
Another benefit of using wavelet transform is the dimension reduction. When DWT level1 decomposition is applied on an image, it returns the coefficients of half size of original image. Therefore, with reduced image vectors, computational complexity also gets reduced. The proposed system shows great success under one sample problem and varying conditions.
The block diagram of proposed algorithm is shown in figure 2.
Journal of Applied Computer Science & Mathematics, no. 12 (6) /2012, Suceava
Fig.2. Working and functionality of the proposed system IV EXPERIMENTS AND RESULTS
ORL, Yale, PIE and AR face datasets are used to test the proposed method. ORL consists of 400 images of 40 personals (10 images of per person of size 112 x 92). The database contains the images with different illuminations, expressions and poses. Sample images of ORL database are shown in figure 3. Here in this paper two experimental ways are used to compute the result; first for varying conditions in which five images per individual were chosen as a training and the remaining five images were chosen for testing as shown in figure 3. See Table I for Results.
The second experimental way is used for one sample problem in which first image per person is selected for training and remaining for testing purpose. This procedure is repeated 50 times. See Table II for Results..
In experimental results proposed method uses the parameters i=3/4 in equation 2 and alpha = 0.25 in equation 3. Processed image is shown in figure 4.
Proposed system includes grey level equalization, removing effect of illuminations and expression variations and also reducing the dimension of original images with the help of using the combination of SVD perturbation and discrete wavelet transforms. The recognition accuracy of the technique is compared with previous proposed systems like PCA and SVD for the first way and with PCA, (PC) 2A, SVD and SVD perturbation for second way of experiment. The proposed system shows better accuracy/recognition rate than the existing techniques as shown in figure 6. In table I top1 matched rate is mentioned with 10 Eigen faces used for PCA. The recognition rates of different techniques for one image per person are shown in table II, where the proposed system shows higher accuracy as shown in figure 7.
Table III computes the accuracy of the proposed method
(a)
(b)
Fig.3. Face images from ORL database. (a)Training Images (b) Testing Images.
(a) (b) (c) Fig.4. SVD perturbation (a) Original image (b) Derived image (c) Combined
image
a) (b) (c) (d) Fig5. (a) LL1 (b) LH1 (c) HL1 (d) HH1
TABLE I Recognition rate on ORL database with 5 images per person
Image No PCA SVD ESVD
2 72 80 83 3 80 82 86 4 81 88 91 5 86 91 95
Fig 6. Comparison chart with varying no. of images TABLE II Recognition rate with one image per person on ORL dataset
Method Recognition ratio
PCA 60
(PC)2A 62
SVD Perturbation 62
Computer Science Section
Fig7. Comparison chart with one sample image where ptb is perturbation
TABLE III The results on the Yale face datasets
Methods Accuracy
PCA 71.56% (119/165)
2DPCA 84.24% (139/165)
Enhanced (SVD) (Proposed) 90.30% (149/165) (PC)2 A 69.70% (115/165)
The table shows the accuracy of different facial recognition methods with the proposed enhanced support vector machine, 2DPCA and PCA.
Table IV shows the accuracy of different facial recognition methods along with the proposed method enhanced (SVD). Similarly figure 9 shows the accuracy result of proposed method in graphical form.
Fig8. The results on the Yale face datasets
The PIE facial database comprises of 68 images and a total of 41,368 images from different angles [15].
TABLE IV The results on the PIE face datasets with Comparison of Eigenfaces, Fisher faces, Laplacian faces and proposed method enhanced
(SVD). Methods Accuracy Eigen Faces 80.66% (121/150) Fisher Faces 94.60% (142/150) Laplacian Faces 95.33% (143/150) Enhanced (SVD) (Proposed) 96.00% (144/150)
Fig9. The results on the PIE face datasets
The images of this database were captured by synchronized cameras from 13 different directions under varying expression, illumination and pose variation. This paper uses 150 images of each individual, 50 for training and 100 for testing the accuracy percentage.
The AR face data base was collected at Purdue University with a total of 4,000 images of 126 persons (70 men, 56 women) [16].
Table V provides recognition accuracy percentage of AR dataset on 1000 frontal face images out of which 100 were used at training side (60 men and 40 women) and 900 on testing side of the algorithm. Images were cropped as frontal faces with variation in facial expressions, pose variation and illumination conditions. Similarly in figure 10 the error percentage per each method and proposed method is shown where the proposed method has very low percentage of error.
V CONCLUSION
Enhanced SVD idea focuses on preprocessing stage of facial recognition. Hence, the paper focused on to remove the challenge of illumination and expression variations by the combination of SVD perturbation and discrete wavelet transforms. This paper identifies the low frequency component as the input image to recognition algorithm that is invariant to these challenges. Finally, the use of wavelet transforms on the input images reduces the calculation complexity. It also reduces the size of original image.
TABLE V The results of AR face datasets with comparison of Eigenfaces, Projection, B-2DPCA, ILLFR (DCT) in LRD FR-using (LBD) and
Proposed method enhanced (SVD).
Methods Accuracy Error
Percentage Eigen Faces 74.44% 25.56% Projection (PC)2A 77.78% 22.22% B-2DPCA 86.11% 13.89% ILLFR(DCT) in LRD 90.55% 09.45% FR-using(LBD) 94.44% 05.56% Enhanced(SVD) (Proposed) 95.00% 05.00%
Journal of Applied Computer Science & Mathematics, no. 12 (6) /2012, Suceava
Fig.10. The results of AR face datasets
VI REFERENCES
[1] D.Q. Zhang, S.C. Chen, Z.-H. Zhou, “A new face recognition method based on SVD perturbation for single example image per person”, Appl. Math. Comp. 163 (2) (2005) 895–907. [2] R. Bmnelli, and T. Poggio, “Face Recognition: Features
versus Templates”, IEEE Trans., Oct. 1993
[3] L. Wiskott, J.-M. Fellous, N. Kruger, and C. von der Malsburg. “Face recognition by elastic bunch graph matching. IEEE Tans. on Pattern Analysis and Machine Intelligence 19 (7) (1997)775-779”
[4] M. Turk and Pentland, “Eigenfaces for recognition”,Journal of cognitive neuro Science, March 1991
[5] Chou-Hao Hsu and Chaur-Chin Chen,“SVD-Based Projection for Face Recognition”, IEEE EIT 2007 Proceedings
[6] W. Zhao,r. Chellappa,P.J.Phillips and A. Rosenfeld, ”Face Recognition: A Literature Survey” ACM Computing Surveys, Vol. 35, No. 4, December 2003, pp. 399–458. [7] Xiaoyang Tana,b, Songcan Chena,c, , Zhi-Hua Zhoub,
Fuyan Zhangb, “Face recognition from a single image per
person: A survey ”, Pattern Recognition 39 (2006) 1725 – 1745
[8] J. Yang and D. Zhang, A.F. Frangi, and J.Y.Yang, ”Two-dimensional PCA: a new approach to appearance-based face representation and recognition”, IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 26, no. 1, pp.131-137, 2004.
[9] J. Wu, Z.-H. Zhou, “Face recognition with one training image per person”, Pattern Recognition Lett. 23 (14) (2002) 1711– 1719.
[10] S.C. Chen, D.Q. Zhang, Z.-H. Zhou, Enhanced (PC)2A for
face recognition with one training image per person, Pattern Recognition Lett. 25 (10) (2004) 1173–1181.
[11] A. Yilmaz, M. Gohen, “Eigenhill vs. eigenface and eigenedge”, Pattern Recognition, vol. 34, pp. 181-184, 2001 [12] Iulian B. Ciocoiu, Brenf Valmar, “a comparison between two
preprocessing techniques in. pca-based face recognition”. [13] Ming yu, Gang yan, Qing-wen zhu, “New face recognition
method based on dwt/dct combined feature selection” Proceedings of the Fifth International Conference on Machine Learning and Cybernetics, Dalian, 13-16 August 2006
[14] vcY. Z. Goh, Andrew B. J. Teoh, Michael K. O. Goh, “Wavelet Based Illumination Invariant Preprocessing in Face Recognition”, 2008 Congress on Image and Signal Processing.
[15] http://cvc.yale.edu/projects/yalefaces/yalefaces.html
[16] A.U. Batur and M.H. Hayes, “Linear Subspace for Illumination.Robust Face Recognition,” Proc. IEEE Int’l Conf. Computer Vision and Pattern Recognition, Dec. 2001. [17] P.N. Belhumeur, J.P. Hespanha, and D.J. Kriegman,
“Eigenfaces vs. Fisherfaces: Recognition Using Class Specific Linear Projection,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19, no. 7, pp. 711-720, July 1997.
[18] M. Belkin and P. Niyogi, “Laplacian Eigenmaps and Spectral Techniques for Embedding and Clustering,” Proc. Conf. Advances in Neural Information Processing System 15, 2001. [19] M. Belkin and P. Niyogi, “Using Manifold Structure for
Partially Labeled Classification,” Proc. Conf. Advances in Neural Information Processing System 15, 2002
Mr. Muhammad Sharif is working as an Assistant Professor in department of computer science, COMSATS Institute of
Information Technology, Wah Campus and having over 17 years of experience of teaching graduate and under graduate classes. His areas of research are Image Processing and Computer Networks.
Mr. Saad Anis is a graduate from COMSATS Institute of Information Technology, Wah Campus. He is working as
Professional Data Warehouse Consultant in Teradata GCC Pakistan
Mr. Mudassar Raza is working as a lecturer in department of computer science, COMSATS Institute of Information
Technology, Wah Campus and having over 5 years of experience of teaching under graduate classes. His area of research is Image Processing.
Dr. Sajjad Mohsin is working as a Professor/Dean, Faculty of Information Sciences and Technology, COMSATS Institute
of Information Technology. He is the member of editorial boards of 3 International Journals including IEEE Journal as well. He is also member of Board of Faculties and Convener Board of Studies of Computer Science.
