Face Recognition Methods Based On Support Vector Machines

As a focus and difficulty in the pattern recognition and image processing fields, face recognition (FR) research has an important significance, either for its widely practical application. Human faces have large variation in shape at different time. Many influences, such as lighting, background, facial expressions and facial details, will easily affect the recognize result. Limited by the practice, we can't collect sufficient face images for every people. Compare with the face image vector dimension, FR is a small sample problem. When solving this small sample problem with high dimension and nonlinear, many traditional pattern recognition methods will tend to occur overfitting phenomenon. Support vector machines (SVM) is specially devised to solve the small sample problem. Based-on the structural risk minimization (SRM) principle in the statistical learning theory, SVM selects the optimal separate hyperplane as the separate function. The optimal separate hyperplane is the hyperplane that either correctly separate the sample set or get the biggest margin between two classes. Thus the separate problem can be formulated as a quadratic optimization problem satisfied simple restriction. The quadratic problem has singular global maximum point. By introducing the kernel function, the nonlinear separate samples are projected into a high dimension space (so call "feature space"). New separate problem is solved in the linear separate feature space. Applied the kernel function, the compute complexity is not enhanced. Using different kernel functions, SVM works as the same as some traditional pattern recognition methods. It has been the preference classifier at present. This article use the K-L transform and wavelet transform to extract face image feature, then classify them based on SVM. Recognition result demonstrates the effective of the system.