A Novel Graph Construction Method Based on Locality Sensitive Histogram for Locality Preserving Projections

Abstract

This paper presents a novel construction method of neighbor graph based on locality sensitive histogram for Locality Preserving Projections (LPP), which is called LSH Graph Construction. Unlike the conventional construction method of neighbor graph that vectorizes the original data set to compute the k-nearest neighbor graph, the LSH graph construction method which we proposed is acting on 2D image matrixes directly. Firstly, we compute each sample’s locality sensitive histogram, so that each sample can be easily divided into multiple overlapping regions. Then we construct the samples’ neighbor graph basing on corresponding region-to-region matching. Our proposed LSH Graph Construction method have the following two advantage: it can preserve samples’ intrinsic structural information and is insensitive to illumination and pose variation in some extent. We apply LSH Graph Construction into the famous dimensionality algorithm: LPP to develope a new algorithm called LSHG-LPP. Experiments on two wellknown face databases (Yale and ORL face databases) demonstrate that the proposed method outperforms LPP.