Collaborative Representation Based Graph Embedding Discriminant Analysis for Face Recognition

Abstract

Considering the collaborative representation projection constructs the adjacent graph based on the collaborative representation relationship of samples and many samples from different classes may be gathered together after projection, a discriminative competitive and collaborative representation projection is proposed. Firstly, each sample is competitively and collaboratively represented by all samples in the dataset to calculate the similarity of samples. Then, an intraclass graph is constructed to characterize the intraclass compactness, and an interclass graph is built to characterize the interclass separability. On this basis, the label propagation algorithm is applied to calculate the soft label information of unlabeled samples to eliminate the influence of unlabeled samples on the recognition results. Additionally, the nonlinear mapping is used to replace the linear inner in the graph embedding framework to solve the linearly inseparable problem of the original samples in the low-dimensional space. Experimental results on ORL, AR, FERET and Yale B face datasets show that compared to the CRP method, the proposed method improves the maximum recognition rate by 1%~4% and improves 2%~6% in noise and blur results, respectively.