Feature extraction using two-dimensional maximum embedding difference

In this paper we propose a novel method combining graph embedding and difference criterion techniques for image feature extraction, namely two-dimensional maximum embedding difference (2DMED). This method directly extracts the optimal projective vectors from 2D image matrices by simultaneously considering characteristic that is the intraclass compactness graph, the margin graph and inter-class separability graph, respectively. In this method, it is not necessary to convert the image matrix into high-dimensional image vector so that much computational time would be saved. In addition, the proposed method preserves the manifold reconstruction relationships in the low-dimensional subspace. Experimental results on the ORL, Yale face and USPS database show the effectiveness of the proposed method.