Graph Embedding Discriminant Analysis and Semi-Supervised Extension for Face Recognition

Abstract

Considering graph embedding based dimension reduction methods are easily affected by outliers and the fact that there exist insufficient labeled samples in practical applications and thus degrades the discriminant performance, a novel graph embedding discriminant analysis method and its semi-supervised extension for face recognition is proposed in this paper. It uses the intraclass samples and their within-class mean to construct the intrinsic graph for describing the intraclass compactness, which can effectively avoid the influence of outliers to the within-class scatter. Meanwhile, we build the penalty graph through the local information of interclass samples to characterize the interclass separability, which considers the different contributions of interclass samples within the neighborhood to the between-class scatter. On the other hand, we apply the low-rank representation with sparsity constraint for semi-supervised learning, aiming to explore the global low-rank relationships of unlabeled samples. Experimental results on ORL, AR and FERET face datasets demonstrate the effectiveness and robustness of our method.