You are on page 1of 4

Proceedings of the Seventh International Conference on Machine Learning and Cybernetics, Kunming, 12-15 July 2008

COMPARATIVE STUDY OF DIMENSION REDUCTION AND RECOGNITION ALGORITHMS OF DCT AND 2DPCA
BIN JIANG1, GUO-SHENG YANG2, HUAN-LONG ZHANG3
1

Institute of Advanced Control and Intelligent Information Processing, Henan University, Kaifeng 475001, China 2 Information Engineering College,Central University for Nationalities, Beijing, 100081, China 3 Department of Computer, Luoyang Institute of Science and Technology, Luoyang 471023, China E-MAIL: jiangbin19831983@yahoo.com.cn,ygs@henu.edu.cn, zhl888zhl_3@yahoo.com.cn

Abstract:
Based on the application of DCT discrete cosine transformin the image compression, the feasibility of DCT to be used in image feature dimension reduction is analyzed, and the basic principle of the image feature dimension reduction based on DCT is given in this paper. And then, taking the face recognition and the facial expression recognition as the research background, the theoretical analysis that DCT algorithm has the higher recognition than 2DPCA two-dimensional principal component analysisin the face recognition and the facial expression recognition is given under the condition that DCT and 2DPCA algorithms have the approximate dimension reduction effect. At last, the comparative simulation experiment is performed on DCT and 2DPCA algorithms respectively by use of the AT&T face database and JAFFE facial expression database.

Keywords:
DCT; Image Compression; Feature Dimension Reduction; Facial Recognition; Facial Expression Recognition

1.

Introduction

In image pattern recognition, high dimension image feature always has heavy workload on computation and storage, which exerts seriously negative effect on the real time performance of the algorithm. So, image dimension reduction is the key to solve the problem of image pattern recognition. PCA Principal Component Analysis and 2DPCA are two common and useful approaches to dimension reduction. They have obvious advantages in dimension reduction, but it is to be further researched that they have the same advantages in recognition. DCT is a common approach to image compression. The nature of DCT is that much more information can be denoted in less data by using DCT on image. The image compression is implemented in such the way so that the destination of reducing data storage and speeding can be achieved. Viewing from this point, we can see there are 978-1-4244-2096-4/08/$25.00 2008 IEEE

many similarities between the image compression and image feature dimension reduction. It is based on this idea that many researchers try to apply DCT in the image feature dimension reduction. Reference [1] expatiates the approach of face recognition based on DCT, reference [2] uses DCT to work on the local wavelet transform image, and achieve the further feature extraction. The functions of DCT decorrelation and dimension reduction are explained in both references, but the basic principle of the image feature dimension reduction based on DCT is not given. It is in the reference [3] that the basic principle of the image feature dimension reduction based on DCT is given. Based on above-mentioned, the paper explicit the basic principal of feature dimension reduction by use of DCT. And then taking the face recognition and facial expression recognition as the research background, we analyze DCT and 2DPCA algorithm of feature dimension reduction theoretically, and get the reason why the two algorithms have different recognition rate under the approximate level of feature dimension reduction. Finally, the simulation result shows the validation of theoretical analysis. The paper is organized as follows: Section 2 gives the principal of DCT feature dimension reduction. Section 3 explains the reason why DCT and 2DPCA algorithms have different recognition rate under the approximate level of feature dimension reduction. Section 4 gives the comparative simulation result and some analysis. And the conclusion is given in Section 5. 2. 2.1. Feature dimension reduction based on DCT The basic principal of DCT image compression Given that f ( m, n) represent an image with the size of N N , C (u , v ) be the coefficients obtained from the two-dimension DCT of the image, then [4]:

407

Proceedings of the Seventh International Conference on Machine Learning and Cybernetics, Kunming, 12-15 July 2008
N 1 N 1 C ( u , v ) = { f (m, n)(u)(v) m=0 n=0 cos[(2m + 1)u 2N] cos[(2m + 1)v 2N]} 1 N, u=0 (u) = u = 1, 2, ....N 1 2 N, 1 N, v=0 (v) = v = 1, 2, ....N 1 2 N, u, v = 0, 1, 2...N 1

same to the image dimension, which leads to the high dimension of the feature matrix. But there are many zero elements in the DCT coefficient matrix, and the coefficients Di (u , v) are small when coordinates of u and v are big. Moreover, the bigger coefficients are located in the left-top of DCT coefficient matrix where u and v are small [6]. So the left-top corner is taken as the useful information area, and an area template A is used to pick up bigger coefficients by the dot multiply of A and Di (u , v ) .

(1)

From equation (1), it can be seen that the workload of computing C (u , v ) is increased as the image size is increased. So in actual application, we use the approach of image blocking to divide the image matrix with size of

N N into ( N h) 2 (h N ) image blocks M i ( i = 1,2,, ( N h) 2 ) with size of h h . DCT is performed on each M i , which results in:

H i = A Di , 1 1 1 1 1 0 A = 1 0 0 0 0 0 0 0 0 0 0

i = 1,2, , ( N / h) 2 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

Di = TM iT '

(2)

(4) Let p elements in the left-top corner be 1 in the area template A , and the remains are set to be zero. p is determined by the image compression ratio. Rearrange these coefficients in the format of a vector to realize the feature dimension reduction. Then

Where T = T {t ki } is the transform matrix derived from equation (1), and the elements of T are defined as follows:

i = 0, 1 N, t ki = i = 1,2, , N 1 2 N, k = 1,2, N 1

H i = [hi1 , hi 2 , hip , 0, 0, 0]

(5)

(3)

where hiq 0 , ( q = 1,2, p ) . Because the zero elements dont have any meaning for image feature, non-zero elements are selected to form a new vector H i = [ hi1

According to the requirement of image quality and compression rate [5], we can select the quantization matrix which is performed on equation (2). The quantized matrix contains many zero elements which represent no information. Accordingly, we can use a few non-zero elements to represent an image to realize the image compression. 2.2. The principal of feature dimension reduction based on DCT As usual, the coefficient matrix obtained from the two-dimension DCT of the image can be used to describe image features well. If the coefficient matrix is directly used as a feature matrix, the rank of the feature matrix is

h i 2 hip ] . We can align these


2

vectors H i from the top to down to form a feature matrix H with the size is ( N / h) p .

h11 H1 h1p = H = (6) h H h ( N / h)2 p ( N / h)2 ( N / h)2 P ( N / h)2 1


3. Comparative analysis

Face recognition based on PCA is a simple, fast and effective algorithm. But this algorithm needs to transform

408

Proceedings of the Seventh International Conference on Machine Learning and Cybernetics, Kunming, 12-15 July 2008 two-dimensional face image matrix into one-dimensional vector at first, then constructs covariance matrix. As to the face image with size of N * N , the covariance matrix dimension is N * N if PCA is used to pick up feature. It is difficult to compute the covariance matrix because of the high dimension. Although we can use Singular Value Decomposition to solve it, we cant avoid constructing covariance matrix. Reference [7] proposals a method named 2DPCA. 2DPCA is a straightforward image projection technique which does not need to transform the face image into a one-dimension vector. Instead an image covariance matrix is constructed directly using the original image matrices. Compared with the conventional PCA, 2DPCA has two important advantages over PCA. First, it is easier to evaluate the covariance matrix accurately. Second, the sample number of image has less influence on feature extraction. From the analysis mentioned as above, it can be seen that the sample number of image has influence on the feature dimensions of PCA and 2DPCA to some different extents. But the DCT feature dimension has relations only with the image dimension and the compression ratio p , and has no relation with the sample number of the image. And moreover, DCT has the excellent performance of de-correlation. DCT can save the image feature in low frequency, and get rid of the correlation of image feature to achieve the feature dimension reduction. Image feature in low frequency represent the facial features, for example the most of face apparatus, which change slowly. So DCT algorithm can reserve the most information of human face, which results in that DCT algorithm has the higher recognition than 2DPCA in the face recognition under the condition that DCT and 2DPCA algorithms have the approximate dimension reduction effect. Also in facial expression recognition, DCT has a better recognition rate. The reason is that: first, DCT performs a excellent decorrelation [8], so we can distinguish facial expression feature easily; second, the major facial expression feature areas are eyes and mouth. These areas are the major area of DCT feature extraction. 4. Experiment result and analysis
2 2

simulation environment is Matlab 7.0 which runs on the personal computer with ACER P4, 2.6GHz. The results of simulation experiment are as follows, shown from table 1 to table 3. Table 1 Comparison of DCT with 2DPCA in face recognition Method 2DPCA DCT Size of Image 112*92 112*92 Size of Feature 112*15 168*10 matrix Recognition 90% 92.86% Rate In Table 1, AT&T face database is used, 100 images of 10 individuals (each person had 10 different images) are selected. The first 7 images of each person were used for training, and the remains were used for testing. The training images were 70, and the testing images were 30. Table 2 Comparison of DCT with 2DPCA in facial expression recognition Method 2DPCA DCT Size of Image 124*124 124*124 Size of Feature 124*17 156*10 matrix Recognition 75.52% 79.30% Rate In Table 2, taking the facial expression recognition as the research background, the comparative study of DCT and 2DPCA algorithms is done by using all 213 images of 10 individuals of the JAFFE Database. 70 images (include all individuals) were used for training, and the remains were used for testing. In order to remove the noise such as image background or hairs, we follow the criterion of reference [9] to cut the image by hand, and get the 124*124 pure face images. Table 3 Recognition accuracy of each facial expression Expression 2DPCA DCT Method Angry 85% 85% Disgust 73.68% 89.47% Fear 72.73% 81.82% Neutral 80% 95% Sad 57.14% 57.14% Table 3 shows the recognition ratios of the 5 facial expressions including angry, disgust, fear, neutral and sad. From table 1 to table 3, we can see: (1) From table 1 to table 2, we can see that DCT and 2DPCA have the similar effect on feature dimension reduction. The dimension reduction of DCT depends

Taking the face recognition and facial expression recognition as the research background, the comparative simulation study of DCT and 2DPCA algorithms of feature dimension reduction and recognition is done by use of AT&T face database and JAFFE database. In the simulation, the accumulative contributions ratio of 2DPCA is 90%, a nearest neighbor classifier is used for classification. The

409

Proceedings of the Seventh International Conference on Machine Learning and Cybernetics, Kunming, 12-15 July 2008 on the template size and the size of p . When the size of template and p are fixed, the change of DCT dimension is steady. (2) From table 1, it can be seen that the face recognition ratio of 2DPCA is 90%, but the recognition ratio of DCT is 92.86%, which is higher than 2DPCA. From table 2, we find the facial expression recognition ratio of 2DPCA is 75.52%, and the recognition ratio of DCT is 79.30%, which is also higher than 2DPCA. So, as a whole, DCT has higher recognition ratio than 2DPCA in face recognition and facial expression recognition. (3) From table 3, it can be seen that DCT and 2DPCA have the same recognition ratio for angry and sad expression, but for the disgust, fear and neutral expression, DCT has higher recognition ratio than 2DPCA. 5. Conclusion Computer Engineering, Vol 30, No.16, pp.53-54, Aug. 2004. Shufen Liang, Junning Gan, Face Recognition Based on Local Wavelet Transform and Discrete Transform, Control&Automation, Vol 22, No.2, pp.206, 2006. Yankun Zhang, Chongqing Liu, Efficient face recognition method based on DCT and LDA, Journal of Engineering and Electronics, Vol 15, No.2, pp.211-216, 2004. Rafael.C.Gonzalez, Digital Image Processing (second edition), Publishing House of Electronics Industry, Beijing, 2004. K Cabeen, and P Gent, Image Compression and Discrete Cosine Transform, http://online.redwoods.cc.ca.us/instruct/darnold/LAPR OJ/Fall98/PKen/dct.pdf. Jian Lu, and Hui Wang, Comparison between SVD and DCT Feature Extraction Methods in Face Recognition, Geomantics and Information Science of Wuhan University, Vol 30, No.2, pp.119, Feb.2006. Jian YangDavid ZhangAlejandro F Frangi et al. Two- dimensional PCA A new approach to appearancebased face representation and recognition[J].IEEE Trans Patten Analysis and Machine Intelligence2004261 131-137. Syed Ali Khayam, The Discrete Cosine Transform (DCT): Theory and Application, http://www.egr.msu.edu/waves/people/Ali_files/DCT_ TR802.pdf Zhaoqi Bian, and Xuegong Zhang, Pattern Recognition (second edition), Tsinghua University Press, Beijing, 2000.

[2] [3]

[4] [5]

[6]

From theory and experiment, this paper analyzes the feasibility of feature dimension reduction based on DCT. And taking the face recognition and facial expression recognition as the research background, two algorithms of DCT and 2DPCA are studied. Compared to 2DPCA, DCT algorithm has the higher recognition than 2DPCA in the face recognition and the facial expression recognition under the condition that DCT and 2DPCA algorithms have the approximate dimension reduction effect. This has been demonstrated by the simulation experiments. But DCT algorithm of feature dimension reduction can be improved in some aspects, for example, we can find different templates to enhance the compression rate. References [1] Ning Fang, Jingzhi Li, and Guiming He, A Face Recognition Based on DCT and Neural Network,

[7]

[8]

[9]

410

You might also like