Extended class-wise sparse representation for face recognition
Minghua Wu, Shiren Li, Jianguo Hu · 2017
Sparse representation classification (SRC) performs well in classification problems, especially in face recognition. When it comes to the undersampled and occlusion problem, the performance of SRC, however, decreases rapidly. In this paper, we propose a robust method called extended class-wise sparse representation (ECSR) to tackle the issues. ECSR extends SRC to applications where there are very few training samples per gallery subject. The proposed ECSR constructs a subsidiary intraclass variant dictionary to represent the intraclass variations between the query image and training samples. Moreover, it minimizes the number of selected classes of training samples to seek an optimum representation of the probe image. For the undersampled problem with variable expressions, disguise and illuminations, experimental results on EYB, AR and CMU Multi-PIE show that ECSR has much better generalization ability and higher recognition accuracy than the competing methods.