Abstract: Neighborhood Preserving Projections(NPP) is a method for unsupervised linear dimensionality reduction, which can preserve the local neighborhood structure on the data manifold. However, when NPP is applied to the classification tasks, it has some limitations such as ignorance of the class labels information. This paper proposes a novel subspace method named Supervised Neighborhood Preserving Projections(SNPP) for face recognition, in which local geometric relations are preserved according to prior class-label information. It gains a perfect approximation of face manifold and enhances its discriminant power in a feature space. Experiments on ORL face database demonstrate the effectiveness of the method.

Key words: dimensionality reduction, Neighborhood Preserving Projections(NPP), supervised subspace learning, face recognition

摘要： 保持近邻投影是一种无监督线性降维方法，具有保持数据流形上局部近邻结构特性，但应用到分类任务时具有局限性，如忽略类标签的信息。该文提出一种新的人脸识别子空间学习方法——监督保持近邻投影，根据先验的类标签信息保持局部几何关系，能获得较好的近似人脸流形以及增强特征空间的判别力。在ORL人脸数据库上的实验表明该方法是有效的。

关键词: 降维, 保持近邻投影, 监督子空间学习, 人脸识别

CLC Number: 

• TP391.41