Face recognition with discriminant locality preserving projections in complete kernel space

To efficiently utilize the discriminant information in the whole space of kernel locality preserving total scatter, this paper proposes a complete kernel discriminant locality preserving projections (CKDLPP) algorithm for face recognition. In our previous research, a kernel locality preserving discriminant analysis (KLPDA) algorithm, which is derived by extending discriminant locality preserving projections (DLPP) method to its kernel form, is presented to address the classification limitation of DLPP. However, KLPDA loses the discriminant information in the null space of kernel locality preserving within-class scatter. To address this issue, in the proposed CKDLPP, discriminant features which are extracted from both the principal and the null subspaces of so-called reduced kernel locality preserving within-class scatter separately are combined to enhance the recognition performance. Experiments of comparing the proposed algorithm with some other popular linear and nonlinear subspace learning methods on UMIST and FERET face databases show that the proposed algorithm consistently outperforms the others.