Single Sample Face Recognition More Manifolds Discriminant Analysis

Currently, there are four kinds of popular methods to deal with this problem, i.e., local feature representation, general learning framework, virtual sample extended and image partition method. Among those methods, the image partition method is the most popular. Many image partition methods can improve the recognition accuracy to some extent. However, most of them usually ignore geometrical information of the images.A generic learning method is proposed to improve face recognition accuracy with single sample per person. Firstly, a suitable generic training set is selected and they are superposed with each single sample by proportion so as to increasing training samples for each class, and PCA is used to reduce dimensionality. Then, traditional extraction methods are used to extract features and all training and testing samples are projected to the subspace. Finally, Euclidean distances between testing and training samples are computed and the maximum membership principle is used to realize face recognition. Experimental results on ORL and Yale face datasets show that proposed method has improved face recognition accuracy with single training sample per person.Traditional face recognition methods with single training sample per person usually ignore geometrical information of the images. To tackle this problem, the multi-manifold discriminant analysis method based on the fusion of both generic learning framework and Fisher criterion is proposed. Firstly, generic learning framework is used to generate multiple samples for each single sample, and the multiple samples of each class are assumed to distribute in a manifold. Then, an optimization function is proposed based on Fisher criterion, and iterative algorithm is used to get the optimal projection matrix. Finally, distances between testing and training images are computed and sparse coefficient reconstruction is used to accomplete the face recognition. The effectiveness of the propsed method has been verified by the experimental results on ORL and AR face datasets. The experimental results show that proposed method has better recognition accuracy than several advanced algorithms when processing face recognition with single sample per person.