Algorithm Research Of Face Recognition Based On Kernel2DLDA

Face recognition is pattern recognition research hot spot subject. In the present face feature extraction algorithm, based on the standards Linear Discriminant Analysis (LDA) algorithm is one of the more successful feature extraction algorithm. The aim of LDA is to find the optimal projection in order to maximize the between class scatter matrices(Sb)and to minimize the within class scatter matrices(Sw), which makes the ratio of the determinants of the between class and the within-class scatter matrices of the projected samples reach its maximum.However, high dimensional face data leading LDA often met with "small sample problem".Two-Dimensional Linear Discriminant Analysis (2DLDA) algorithm is an excellent algorithm for feature extraction and dimensional reduction, in which the "small sample problem" of LDA algorithm is overcome. However, Traditional scatter matrix could hardly tell the difference between two classes when the category mean and the global mean are close to each other. So, this article introduces the each two between-class scatter matrix and within-class overlap coefficient matrix to the scatter matrix theory. Each two between-class scatter matrix is to distinguish the vectors whose category mean and global mean are close. Experimental results demonstrate that the proposed N2DLDA method compared with Fisherfaces,2DLDA, its has certain advantages.In this paper, through a large number of experiments to determine the introducing adjustable parameters (k) on the influence of recognition rate, through the impact factor can indirectly regulate recognition rate and recognition time balance.In many real-world applications the distributions of original samples are usually highly complex and nonlinear. N2DLDA is the same as Fisher Linear Analysis (LDA), it is a kind of supervisory linear space dimension-reduction methods, it overcomes the LDA "Small Sample Problem", but cann’t use data of nonlinear information. The proposed algorithm research of face recognition on improved2DLDA combination kernel-based methods (K-N2DLDA) could make up for the deficiency. In this method, the sample data through the nuclear map is mapped to kernel space, and then in the nuclear space for linear transformation. The process fully solved N2DLDA cannot use images of the nonlinear characteristics. In this paper, a large number of experimental verification of the fractional polynomial kernel function relative to other nuclear functions have certain advantages. Finally, simulation results showed the superiority of K-N2DLDA algorithm.