Sparse discriminative multi manifold embedding based on graph optimization

Abstract

Sparse discriminative multi manifold embedding (SDMME) algorithm was used for feature extraction, graph construction and projection learning were independent, the quality of the graph directly affects the effect of projection learning. In order to solve the problem, a new algorithm named sparse discriminative multi manifold embedding based on graph optimization (GOSDMME) was proposed in this paper. First, in proposed approach, the image matrix was divided into blocks. The matrix blocks on the same image were located on the same manifold. Then, the sparse graph was used to establish the connection relationship between different blocks. Finally, in the framework of the same objective function, the sparse constraint graphs and projections were studied simultaneously. The graphs and projections were learned at the same time, iterate and update the graph and projection to obtain a projection matrix that satisfies the accuracy requirements. The face recognition experiments conducted on Extended Yale B and CMU PIE datasets show that the new algorithm has better recognition performance than the SDMME algorithm.