Two-dimensional locality sensitive discriminant analysis

Recently, locality sensitive discriminant analysis (LSDA) was proposed for dimensionality reduction. As far as matrix data, such as images, they are often vectorized for LSDA algorithm to find the intrinsic manifold structure. Such a matrix-to-vector transform may cause the loss of some structural information residing in original 2D images. Firstly, this paper proposes an algorithm named two-dimensional locality sensitive discriminant analysis (2DLSDA), which directly extracts the proper features from image matrices based on LSDA algorithm. And the experimental results on the ORL database show the effectiveness of the proposed algorithm. After that, 2DLSDA plus Fisherface, which was presented for the further dimensionality reduction, was compared with other dimention reduction methods, namely Eigenface, LSDA and 2DLSDA plus PCA. Experiments show that conducting Fisherface after 2DLSDA achieves high recognition accuracy.