An Improved DLDA Based Method- Nonparametric DLDA

The Small Sample Size problem always arises when Linear Discriminant Analysis (LDA) is used in face recognition. Direct LDA(D-LDA) is a method to solve this problem by firstly whitening the between-class scatter Sb and then diagonalizing the within-class scatter Sw, but the discriminative information in the null space of Sw is always lost indirectly. To overcome the shortcoming of D-LDA, an improved D-LDA based method called Nonparametric D-LDA (ND-LDA) is proposed in this paper. ND-LDA introduces a new nonparametric definition for Sb, which is full rank and can capture the boundary structural information for different classes. As a result, ND-LDA can exploit the discriminative information in both the principal and the null space of Sw, and can perform well on non-Gaussian cases. The results of the experiments show the improvements of ND-LDA.