Abstract

Locality Preserving Projections (LPP) is an unsupervised and graph-based dimensionality reduction algorithms which has been successfully applied in face recognition. The performance of LPP is mainly depend on its neighbor graph whose construction may suffers from one problem: face images were transformed into their vector patterns by the traditional graph construction algorithm to compute knearest neighbors of each face sample, which leads to the loss of the sample's original matrix structure information. In this paper, we propose a new neighbor graph construction algorithm which is based on image blocks sampled by a sliding window to determining neighbors of each sample, so we name this new neighbor graph construction algorithm as Sliding Window (SW) based Neighbor Graph Construction, and we name the neighbor graph construct by SW as Sliding Window based Neighbor Graph (SWG). The SWG is aiming to preserve image sample's original matrix structure information. We incorporate SWG into LPP, and proposed a new algorithm called SWG-LPP. To evaluate the SWG-LPP, several experiments were performed on three well-known face databases, and experimental results show that SWG-LPP achieves better performance than LPP.

Paper Details