Kernel inverse Fisher discriminant analysis for face recognition

In this paper, we present a new nonlinear feature extraction method for face recognition. The proposed method incorporates the kernel trick with inverse Fisher discriminant analysis and develops a two-phase kernel inverse Fisher discriminant analysis criterion - KPCA plus IFDA. In the proposed method, we first apply the nonlinear kernel trick to map the original face samples into an implicit feature space and then perform inverse Fisher discriminant analysis in the feature space to produce nonlinear discriminating features. In implementation, kernel IFDA seeks nonlinear discriminating features by minimizing the inverse Fisher discriminant quotient and overcome the singularity problem by projective transformation of scatter matrices. Experimental results on ORL, FERET and AR face databases demonstrate the effectiveness of the proposed method.