SPECTRAL ANALYSIS FOR FACE RECOGNITION

Different eigenspace-based approaches have been pro-posed for the recognition of faces, i.e. eigenface, fisher-face and Laplacianface. For fisherfaces, the original im-age space is reduced to an n-c dimensional subspace in which the standard LDA is carried out, where n is the number of training samples and c is the number of classes. In this paper, we present a spectral analysis of fisherface which shows that the initial PCA dimensionality reduc-tion might be insufficient. The noise might not be com-pletely eliminated. This is due to the fact that fisherface only takes into account the discriminating structure while ignores geometrical structure. Based on the theoretical analysis, we propose a new method, called enhanced fish-erface, which takes into account the discriminating struc-ture as well as the intrinsic geometrical structure. Ex-perimental results show that the proposed approach is effective in improving the performance of face recogni-tion