Support Vector Machine, developed on the basis of statistical learning theory, is a new machine learning method with the advantage of perfect theory, adaptability, global optimization, short training time and good generalization performance. Since been proposed, it has been paid much attention and has been widely applied to the problems of pattern recognition and regression, of which face recognition is one typical application. Comparing to other biometric-based identification methods, it has many advantages, such as friendly, easy sampling, without psychological burden, and so on. What is more, the modern image communication requirement is very urgency, it is very important to study the face recognition technology. In this paper, the main work includes:Firstly, Support Vector Machines is one popular pattern recognition method. It has a number of unique advantages in solving the small samples, nonlinear and high dimensional pattern recognition problems .In traditional Support Vector Machines, Inner-product is used to evaluate the simulation between samples. After the kernel function and its parameter of SVM are studied, a new kernel function based on Euclidean distance is proposed. The excremental results on ORL, Yale, Essex face databases show that this kernel functions can complete face recognition better. Comparing to other kernels, it has higher recognizing accuracy and the recognize time doesn't increase.Secondly, for a SVM classifier, kernel function is the key factor. A kernel function is introduced in this paper, known as conditionally positive definite kernel, which does not satisfy the Mercer condition but can be used in the study of kernel methods. This kernel function is proven more effective than any other kernel function based on inner-product by theory and experiment at the same time. The experimental results show that this kernel function has better results than the traditional kernel functions in the aspect of face recognition.Thirdly, under the fact that the identify time is quite long and the efficiency is quite low, several different image processing methods is selected to process the images of these three face databases. The experimental results prove the effectiveness of the two new kernel functions again, these image process methods are compared at the same time.At last, the whole work of this paper is summarized, and a target and hope to the future research is given. |