Fuzzy Multi-Manifold Classifier for One-sample Face Identification

With the continuous development of image recognition technology, various classification algorithms have been proposed, which makes the accuracy of prediction increasingly improved. However, in many practical applications, there is only ’one sample per person (OSPP)’ that can be obtained. Thus, many popular methods fail to work well. To address this problem, we explore a Fuzzy Multi-Manifold Classifier (FMMC) method which transforms OSPP into the manifold matching problem. First, we construct the sub-manifold by partitioning each enrolled image of several non-overlapping patches. Then, in order to specify the class-membership degree of each patch, we introduce the theory of fuzzy set to define the membership matrix. Such that each patch has a big enough membership degree, even if it is surrounded by the inter-class neighbors. Lastly, the classification is turned into a problem of finding a category with a maximum decision membership degree of manifold. Experimental results on the YALE and AR face databases show that our method outperforms some other classifiers.