Face recognition using novel LDA-based algorithms

Facial feature extraction with enhanced discriminatory power plays an important role in face recognition (FR) applications. Linear discriminant analysis (LDA) is a powerful tool used for dimensionality reduction and feature extraction in FR tasks. However, the classification performance of traditional LDA is often degraded, due to two factors: 1) their classification accuracies suffer from the small sample size problem (SSSP), which widely exists in FR; 2) their Fisher discriminant criteria are not directly related to the classification ability. Recently, so-called direct fractional-step LDA (DF-LDA) algorithm has been proposed to solve this problem. In this paper, the limitations of DF-LDA are discussed and a novel DF-LDA has been proposed to solve those problems. The novel DF-LDA has been tested, in terms of classification accuracy, on two face databases. Results reveal that the proposed method outperforms the existing methods, including: the Eigenfaces, Fisherfaces, D-LDA, previous DF-LDA, and EFM methods.