Extended Marginal Fisher Analysis Based on Difference Criterion

The reasonable construction of neighbor graph plays an important impact to improve the efficiency of face recognition in Marginal Fisher Analysis (MFA) algorithm. In this paper, a new method of constructing neighbor graph is proposed to improve the recognition accuracy of MFA. Specifically, we try to map the raw data of the training samples to a new feature space z, then we select the nearest neighbor points directly based on the new distances, thereby constructing the graph. In addition, we also introduce a new criterion function, which can overcome some shortcomings of MFA. The contrastive experiments on several benchmark face databases show the effectiveness of proposed method.