Modified Gradient Linear Discriminant Analysis for Face Recognition

In this paper, we propose two kinds of orthogonalization gradient linear discriminant analysis (OGLDA) algorithms to improve the performance of traditional gradient LDA (GLDA) for undersampled problems in face recognition tasks. In the OGLDA1, the orthogonalization procedure is performed on the discriminant vectors to reduce the redundancy among the discriminant features obtained by GLDA. Thus, all obtained discriminative features can equally contribute to classification performance, which significantly improves the performance of GLDA algorithm for face recognition. In the OGLDA2, the discriminant vectors are resolved one by one in each iterative procedure which overcomes the drawbacks of high computational cost and numerical instability existing in the OGLDA1 algorithm. Since the orthogonalization procedure is applied in the proposed OGLDA methods, the computational stability is improved greatly. The effectiveness of the proposed methods is verified in the experiments on the standard face image databases.