A Optimal Kernel Fisher Nonlinear Discriminant Analysis Method and Applied on Face Recognition

The kernel fisher nonlinear discriminant analysis (KFDA) has become one of the most effective methods applied to extract the nonlinear discriminant face features. However, the face recognition problem is a typical problem of high dimension with small sample, the KFDA is imperfect because the within-class scatter matrix is irreversible. In this paper, a new method (L-KFDA) is proposed to extract the nonlinear discriminant face features: First to solve the projection vectors which make the fisher criterion function larger than zero; then defines a priority function which is highly co-related with the fisher criterion function; using this priority function, we obtain the priority values of all the projection vectors, several maximum projection vectors are selected as the final discriminant vectors. Experimental results show that the proposed method is efficient and it significantly outperforms the traditional kernal fisher linear discriminant analysis (KFDA) on Yale Face Database.