A two-stage linear discriminant analysis for face recognition

Abstract

A two-stage linear discriminant analysis technique is proposed that utilizes both the null space and range space information of scatter matrices. The technique regularizes both the between-class scatter and within-class scatter matrices to extract the discriminant information. The regularization is conducted in parallel to give two orientation matrices. These orientation matrices are concatenated to form the final orientation matrix. The proposed technique is shown to provide better classification performance on face recognition datasets than the other techniques.