Abstract

Locality Preserving Projection (LPP), as a linear manifold learning algorithm, has attracted much interests in recent years. LPP considers an n₁ × n₂ image as a vector in €^n1×n2 space, and thus is limited by the curse of dimensionality. To deal with such problem, Principal Component Analysis (PCA) is traditionally used for dimensionality reduction first. However, some discriminant information would be discarded by PCA. In this paper, a modified version of LPP (MLPP) is proposed which can be applied directly in the original face space without the aid of PCA. Moreover, the mapping matrix derived from MLPP is orthogonal in contrast to that of LPP. Experiments show competitive performance of our method. © 2008 Binary Information Press.

Paper Details