摘要： 为更有效地进行数据降维，将核映射思想引入到邻域保持判别嵌入中，提出一种核邻域保持判别嵌入的流形学习算法。以类内相似度矩阵与类间散度矩阵之差作为鉴别准则，使类间散度矩阵不受满秩的约束，从而解决人脸数据的非线性和小样本问题。在ORL和Yale人脸库上的实验结果表明，该算法具有较好的人脸识别性能。

关键词: 核方法, 邻域保持判别嵌入, 数据降维, 流形学习, 人脸识别

Abstract: In order to have a data reduction more effectively, this paper proposes a new manifold learning algorithm named Kernel Neighborhood Preserving Discriminant Embedding(KNPDE) which puts kernel mapping into the Neighborhood Preserving Discriminant Embedding(NPDE). The algorithm adopts the difference of between within-class similarity matrix and between-class scatter matrix as the discriminant criterion. So it can avoid receiving the restraint of full rank of within-class scatter matrix. The algorithm is applied to the face recognition and solves the problem of nonlinear and small sample for face data. The experiment results on the ORL and Yale face database show that this algorithm has a good recognition performance.

Key words: kernel method, Neighborhood Preserving Discriminant Embedding(NPDE), data dimension reduction, manifold learning, face recognition

中图分类号: 

• TP391