Application Of The Subtensor Algorithm In Face Recognition

The present study of face recognition based on vector pattern has been relatively successful, the related theoretical methods have been relatively mature system. But there is a common problem in the face recognition based on vector pattern, you need to put people face image stretched into a vector, then use the whole or partial blocking of the vector arithmetic operations, and this method will causes three problems. At first, when stretched the images into a vector, the natural structure information of face images will be destroyed, and the relationship between the local pixel adjacent face also will be destroyed. Second, if there are fewer training samples, and the image has been stretched into a vector, respect to the number of training samples, the dimension of vector will be very large, so the "curse of dimensionality" and "small sample size" problems will be happen. Third, after the image has been stretched into a vector, there will be appear slow convergence or no convergence situation. Response to these problems, in recent years, more and more scholars began to focus on the tensor, tensor can represent inside geometry structure information between each image naturally, and effectively retain the relationship between the images in operation, avoid the vector pattern "curse of dimensionality" and "small sample size" problem successfully, and tensor data also can get a better stable convergence in the training phase. This article will be applied the subtensor algorithm to face recognition, expect the face recognition algorithm’s mature theoretical system as a support, join tensor in terms of information protection as well as an image representation of superiority, while referring to the block pattern of picture, application subtensor methods, which can effectively overcome the local image or defaced accessories covering problems in image, in the case of the overall feature of the image is not obvious, fully exploit local information using image features, get better recognition accuracy. In this paper, principal component analysis, linear discriminant analysis and locality preserving projections are applied to the subtensor, and then TLPCA, TLLDA and TLLPP is proposed. Based on Yale, Yale_B and ORL standard face database data, and comparison with principal component analysis, linear discriminant analysis, locality preserving projections in the whole tensor pattern and vector pattern respectively, evaluation algorithm from the merits of recognition accuracy rate and computation time two angles.