Manifold-Based Supervised Feature Extraction and Face Recognition

Abstract

Unsupervised discriminant projection (UDP) has a good effect on face recognition problem, but it has not made full use of the training samples' class information that is useful for classification. Linear discrimination analysis (LDA) is a classical face recognition method. It is effective for classification, but it can not discover the samples' nonlinear structure. This paper develops a manifold-based supervised feature extraction method, which combines the manifold learning method UDP and the class-label information. It seeks to find a projection that maximizes the nonlocal scatter, while minimizes the local scatter and the within-class scatter. This method not only finds the intrinsic low-dimensional nonlinear representation of original high-dimensional data, but also is effective for classification. The experimental results on Yale face image database show that the proposed method outperforms the current UDP and LDA.