Beruflich Dokumente
Kultur Dokumente
liuyg@xaut.edu.cn lujiwen@xaut.edu.cn
Abstract - This paper presents a new approach to face representation of face images. ICA has a number of
recognition using independent component analysis (ICA) and advantages over PCA. First, ICA can decorrelate high-order
fuzzy support vector machine (FSVM). Firstly, 2D wavelet statistics while PCA only decorrelate second-order statistics.
transform is adopted to obtain different level of wavelet Second, ICA basis vectors are more spatially local than PCA
coefficients. Secondly ICA is applied on the low-frequency which
basis vectors and it is usually related to local descriptions such
contains most discrimination information of the original face
image. One criterion that not all ICs are useful for face as edges and facial expression variations. As face is a non-
recognition is demonstrated and a rule for selecting ICs is rigid object, local representation of faces will reduce the
proposed. For reducing the computational cost, a fast ICA sensitivity of the face variations due to different facial
method is proposed. Then, FSVM classifier is designed for expressions and pose variations.
recognition. Lastly, this algorithm is tested on the ORL and Yale Support vector machine (SVM) [3], [9] and [10] is a new
face databases and the experimental result is encouraging, which classifier which has high generalization capabilities in many
achieves comparatively high recognition accuracy and is more tasks especially in the object recognition problem. SVM is
computationally efficient than using general PCA ˉ based based on structural risk minimization, which is the expectation
recognition method. of the test error for the trained machine. As a recent advances
theory in statistical learning principle, SVM has made it
Index Terms – Face recognition, independent component
analysis (ICA), fuzzy support vector machine (FSVM), wavelet
possible to obtain high accuracies for the object recognition.
transform, feature extraction. Early researchers found that variations in facial expressions
mainly affect high-frequency components of the face,
I. INTRODUCTION therefore, we apply wavelet transform to original face image
and extract feature in the low frequency which can overcome
Face recognition by computer has been started since the expression and light variety. Based on this point, this paper
1970s and currently becomes an active and important research put forward an automatic face recognition method using ICA
topic in the domain of image processing and pattern and SVM.
recognition in the past few years because of its wide potential The main advantages of our approach in this paper are as
applications such as biometrics systems, security control, bellow: (1) based on wavelet transform and ICA [8] and [11],
personal identification and human-computer communication we make a meaningful attempt to human face feature
etc. Up till now, many kinds of face recognition methods have extraction. (2) Two public available face databases, namely
been proposed by earlier researchers and among these Yale University and Olivetti Research Laboratory face
methods, they can be mainly divided into two classes, databases are selected to test our method and it proves that this
constituent-based and face-based [1], [2], [3], [4], [5]and[6]. method can perform well in face recognition. (3) Two
In the constituent-based approaches, recognition is based on classifiers namely nearest neighbor (NN) and fuzzy support
the relation between human facial features such as eyes, vector machine (FSVM) are applied to face recognition and
mouth, noise, face silhouettes and boundaries. Face–based experimental results are compared with the two different
approaches attempt to capture and define the face as a whole. classifiers. (4) It is easy to implement and has better
In this kind of method, face is treated as a two-dimension or recognition accuracy. The remainder sections of this paper are
three-dimension matrix and is matched through finding its organized as follows: in the next section, face image feature
statistical regularities. One of the most typical and general extraction using wavelet transform and ICA is proposed while
approaches is principle component analysis (PCA) or the final experimental results and recognition accuracy is
eigenface. Along this direction many PCA-based face given in sections 3. At last, Section 4 concludes this paper and
recognition systems have been developed. PCA method can put forward future research direction of this field.
perform well in scaled and aligned human face, but it will II. FACE IMAGE FEATURE EXTRACTION USING WAVELET
degrade dramatically for not-aligned faces. To overcome the TRANSFORM AND INDEPENDENT COMPONENT ANALYSIS
limit of this approach, a better method named independent
component analysis (ICA) [7] and [8] is presented, whose task A. Perform wavelet transform on face image
is to find basis functions which are local and give good
9890
the within-class scatter and between-class scatter is minimized Bi , j = min{−1,− Di ( x)} i≠ j (7)
[5]. The method is proposed as follows.
If the matrix X contains n individual persons and each Then, we can classify the test sample according to (8)
person has m face images, aij represents the entry at the ith F = ω , ωi = arg max mi ( x) (8)
i =1, 2 ,,n
row and the jth column. W j , which is called as the mean of Where
within-class distance in the jth column, is then given by mi ( x) = min Bi , j ( x) (9)
j =1, 2 ,,n
n m m
1
Wj = ∑∑∑
nm(m −1) i=1 u=1 v=1
(a(i−1)m+u, j − a(i−1)m+v, j )2 (2) III. RECOGNITION AND E XPERIMENTAL RESULTS
Two public available face databases, namely, Yale
B j , which is called as the mean of between-class distance in University and Olivetti Research Laboratory face databases
the jth column: are selected to evaluate the performance of our proposed
n n system. The Yale face database adopted in this paper is just
1
Bj = ∑∑ ρ (a s' , j − at' , j )
n ( m − 1) s =1 t =1
(3) portion of the whole Yale face database, which contains all
frontal view images with different facial expression and here
where we adopted consists of ten persons and each person has nine
1 m different views images. There are forty persons in the whole
a i' , j = ∑ a( i−1) m+u , j
m u =1
(4) Olivetti face image database and we use the whole face image
for training and recognition. There existing ten different facial
In this paper, we employ the ratio of within-class distance and images in the dataset such as various expressions, small
between-class distance to select stable mixing feature occlusions and orientations. Two person face image in Olivetti
from A .The ratio γ j is defined as database and Yale database which contains different views can
be seen in Fig.2 and 3.
Wj
γj = (5)
Bj
From the definition of γ j , the smaller γ j is, the better the
classifier will be. Using (5), we choose the smallest γ j and
select the top k ( k < n ) column features from A and S . If there
is a zero value γ j , one IC is sufficient for recognize all face
images. Generally this is a very special case and seldom
happens in practice. To find the best γ j for recognition, we
Fig. 2 E xamples of various view images in ORL face database
can calculate the recognition accuracy for each value of γ j ,
E
and
9891
Yale database Olivetti database greatly and extract main face information which is sufficient
25 ICs Using all 40 ICs Using all for recognition.
selected ICs selected ICs (3) A criterion that not all ICs are useful for face
Rank 1 90.0% 90.0% 80.0% 75.0%
recognition is demonstrated and a rule for selecting ICs is
proposed.
Rank 2 95.0% 91.5% 85.0% 80.0% Finally, although the experimental results show that ICA
can give good performance on face images with different
Rank 3 96.5% 95.0% 87.5% 85.5% expression and orientations, there is still a limitation that the
computational load of calculating ICs is very high even if we
T
adopt the FastICA algorithm. Therefore, developing a more
accurate and fast ICA algorithm is the main research in our
TABLE 2. THE RECOGNITION RESULTS USING THE NN CLASSIFIER
future work and we believe that the recognition rate will
improve and increase if a better computation of ICs is
Yale database Olivetti database
available.
25 ICs Using all 40 ICs Using all
selected ICs selected ICs ACKNOWLEDGMENT
Rank 1 92.5% 91.5% 82.5% 77.5%
The authors would like to express their thanks to Yale
University and Olivetti Research Laboratory for their standard
Rank 2 95.0% 93.5% 87.5% 85.0%
public face databases to test our algorithm. Portion of the
Rank 3 96.6% 94.7% 92.5% 87.5%
research in this paper uses Yale and Olivetti databases
collected by Yale University and Olivetti Research
Laboratory. The author would also like to thank Dr Hyvarinen
for providing the fixed-point algorithm for independent
TABLE 3. THE RECOGNITION RESULTS USING THE NN CLASSIFIER
component analysis.
Yale database Olivetti database REFERENCES
25 ICs Using all 40 ICs Using all
selected ICs selected ICs [1] J. H. Lai. Pong C. Yuen and Guo Can Feng. “Face recognition using
holistic invariabnt features,” Pattern Recognition, vol. 34, pp. 95-109,
Rank 1 95.0% 94.5% 95.0% 81.9% 2001
[2] Jeff Fortuna and David Capson. “Improved support vector classification
using PCA and ICA feature space modification”, Pattern Recognition,
Rank 2 97.5% 97.5% 97.5% 86.1%
vol. 37, pp. 1127-1129, 2004
[3] O. Deniz. M.Castrillon and M. Hernandez. “Face recognition using
Rank 3 99.5% 97.5% 99.5% 89.6% independent component analysis and support vector machines”, Pattern
Recognition Letters. Vol 24, pp. 2153-2157, 2004
[4] M. S. Bartlett. Javier R. Movellan and Terrence J.Sejnowski. “Face
From the above three tables, we can find that FSVM classifier Recognition by Independent Component Analysis”, IEEE Trans. on
is better than other classifiers used in face recognition but here Neural Network. Vol. 24, pp. 450-1464, 2001
we still use NN classifier as it is very simple and it can save a [5] Pong C. Yuen and J. H. Lai. “Face representation using independent
component”, Pattern Recognition. vol 35, pp. 1247-1257, 2002
large of computational time for our recognition which is very [6] Lu, J.W. and Zhang, E.H. “Face Recognition Using Wavelet Transform
important for gait recognition system. FSVM is a new and Independent Component Analysis”. Proceedings of the Sixth
classifier as it has strong generalization and it is very suit not International Symposium on Test and Measurement, Dalian, China, pp.
only for 2-class classification, but also for multiple-class 7055-7058, 2005
recognition, so if we apply face image to distinguish the [7] Cao, L.J., Chua, K.S., Chong, W.K., Lee H.P., Gu, Q.M. “A Comparison
of PCA, KPCA and ICA for Dimensionality Reduction in Support Vector
gender of the human identification, FSVM classifier is the best Machine”. Neurocomputing, vol 55, pp. 321-336, 2003
choice. [8] Hyvarinen, A., Oja, E.: “Independent Component Analysis: Algorithm
IV. CONCLUSION AND FUTURE WORK and Applications”. Neural Networks, vol 13 , pp. 411-430, 2000
[9] N. Cristianini, J. Shawe-Taylor. “An Introduction to Support Vector
A new face recognition method using wavelet transform Machines and Other Kernel-based Learning Methods”. Publishing House
and independent component analysis has been reported and of Electronics Industry press, 2005, pp.35-45 (in Chinese)
discussed in this paper. Two public standard face databases, [10] Kecman, V.: Learning and Soft Computing, Support Vector machines,
Neural Networks and Fuzzy Logic Models, The MIT Press, Cambridge,
namely Yale University and Olivetti Research Laboratory, MA, 2001, pp.100-115
have been adopted to evaluate the performance of the method
proposed. The recognition accuracy is comparatively high.
Meanwhile, we can draw the following conclusions:
(1) ICs representation gives better performance than PCA
representation and it is less sensitive to pose and expression
variations.
(2) Wavelet transform and multisolution analysis perform
well in face image decomposition. It can reduce image data
9892