Corresponding Block Based Graph Construction for Locality Preserving Projection

摘要

Locality Preserving Projection (LPP) is a typical method of neighbor graph based dimensionality reduction algorithm. So, graph construction plays a key role on the performance of LPP. The original samples were transformed into their vectorial form by the traditional graph construction method before calculate k-nearest neighbors of each samples, which will lost Sample's inner structure information. In this paper, we proposed a new graph construction approach which called Corresponding Block (CB) Based Neighbor Graph Construction Method, and we named the so constructed graph as Corresponding Block Based Graph (CBG). Our new method divided each sample matrix into several blocks and base on corresponding blocks to determine neighbors of each sample, which can well preserve samples' intrinsic structural information and has the ability of non-uniform illumination immunity in some extent. Then, we incorporate CBG into the state-of-art dimensionality algorithm: LPP, and developed a new algorithm called CBG-LPP. To evaluate CBG-LPP, several experiments were conducted on three well-known face databases and achieved satisfactory results.