Full-Space LDA With Evolutionary Selection for Face Recognition

Linear discriminant analysis (LDA) is a popular feature extraction technique for face recognition. However, it often suffers from the small sample size problem when dealing with the high dimensional face data. Some approaches have been proposed to overcome this problem, but they usually utilize all eigenvectors of null or range subspaces of within-class scatter matrix (S_w). However, experimental results testified that not all the eigenvectors in the full space of S _w are positive to the classification performance, some of which might be negative. As far as we know, there have been no effective ways to determine which eigenvectors should be adopted. This paper proposes a new method EDA+Full-space LDA, which takes full advantage of the discriminative information of the null and range subspaces of S_w by selecting an optimal subset of eigenvectors. An estimation of distribution algorithm (EDA) is used to pursuit a subset of eigenvectors with significant discriminative information in full space of S_w middot EDA+Full-space LDA is tested on ORL face image database. Experimental results show that our method outperforms other LDA methods