Discriminative feature extraction based on sparse and low-rank representation.

Feature extraction plays an important role in the pattern recognition and computer vision. Existing methods only consider either the global structure or consider the local structure, while cannot capture both of them. To address this problem, in this paper, we propose a novel feature extraction algorithm called discriminative feature extraction based on sparse and low-rank representation (DFE), which considers the global structure and local manifold structure of the sample points simultaneously. Specifically, the global information is preserved by low-rank constraint, and the local structure information is captured by sparse constraint. To exploit the discriminative information, we require a sample from each class to be reconstructed only by the corresponding samples from the same class. Furthermore, the projection matrix can be also obtained in the proposed method and it can be used to solve the out-of-sample problem. The results of extensive experiments show that the proposed DFE can extract robust and discriminative features and has achieved better classification performance compared with many existing methods.