Kernel Locality Preserving Low-Rank Representation

Classification based on Low-Rank Representation (LRR) has been a hot-topic in the field of pattern classification. However, LRR may not be able to fuse the local and global information of data completely and fail to represent nonlinear samples. In this paper, we propose a kernel locality preserving low-rank representation with Tikhonov regularization (KLP-LRR) for face recognition. KLP-LRR is a nonlinear extension of LRR, and it introduces the local manifold structures of data sets into LRR methods. Since the feature information in kernel space has a very high dimensionality, and to fit the proposed KLP-LRR method well, we introduce locality preserving factor and Tikhonov regularization into dimensionality reduction. It can get more discriminant coding information, especially in the aspect of combining local features with global features, where it is capable of improving the recognition rate obviously. Explicit experimental results on AR, the extended Yale B, FERET face databases show that KLP-LRR out-performs other comparative methods.