Abstract:

In order to solve the problem that high-dimensional input data may have singular value, as well as to improve the operation efficiency and robustness of the algorithm, we propose a new algorithm named block two dimensional locality preserving projections based on L1-norm (B2DLPP-L1). Traditional locality preserving projection (LPP) uses the principal component analysis (PCA) to project input data to PCA subspace to avoid singular value problem, however, input data can lose some effective information in this way. The B2DLPP-L1 algorithm chooses two dimensional data as input data, and it divides original input images into modular images and use the images which are divided into two types as the new input data afterwards. Then we apply the proposed algorithm to the sub-images to reduce the dimensionality. In theory, the B2DLPP-L1 algorithm can better reduce dimensionality, preserve effective information of input data, reduce computation, improve operation efficiency of the algorithm, and overcome the problem of low classification accuracy and improve algorithm robustness. Experimental results on face databases reveal that the B2DLPP-L1 algorithm utilizes less time to accomplish the nearest-neighbor classification and obtain more accurate classification rate.

Key words: locality preserving projections, dimensionality reduction, manifold learning, face recognition