Face Recognition Using Modular Locality Preserving Projections

Facial-image data are always distributed in the high-dimensional space, which makes it difficult to use for accurate face recognition. Recently, many manifold learning methods have been proposed to reduce the dimensionality of the image data. In this paper, a novel method, named Modular Locality Preserving Projection (modular LPP), is proposed. This proposed method is derived from the LPP methods, and is designed to handle face images with various illuminations and facial expressions. In the proposed method, the face images are divided into smaller sub-images and the LPP approach is applied to each of these sub-images. As some of the local facial features of an individual do not vary even when the lighting directions and facial expressions vary, the proposed method is expected to cope with these variations. The Modular LPP and its variant are compared with LPP, based on the Yale and YaleB face database. Experimental results show the significant improvement of our proposed algorithm.