Research On Virtual Sample Generation Method For Face Recognition

Face recognition technique has been develop a lot during the past two to three decades with the recognition accuracy rising remarkablely and the computing speed improving incredibly,a lot of problems has been addressed.While many researchers have been devoted to improving the recognition accuracy,they keep ignoring the significant drop in recognition accuracy due to insufficient training samples in small-dataset situations.Especially,when we only have one face image of every individual for training,some recognition algorithms can no longer be applied and some will suffer from obvious decline in recognition accuracy.However,under some circumstance such as face recognition in criminal investigation,single image is all we have and we have no choice but address the problem.Therefore,we conduct research in face recognition of single sample per person,aiming to guarantee the recognition accuracy through augmenting the training data by designing methods to generate virtual samples and applying robust methods of feature extraction.Generating virtual multi-view face images from a single face image has always been a challenge in the area of computer vision.It often suffers from appearance distortions and artifacts if the generation rule is not well-defined.In order to generate virtual samples that can contribute to recognition accuracy,we propose a Three-dimensional(3D)face model reconstruction method based on Generative Adversarial Networks(GANs)to expand the training dataset,transforming the issue of single sample per person(SSPP)face recognition to the general issue of face recognition by augmenting the training dataset.In order to address the inaccuracy recovering of the depth information,we propose a noise filtering and image smoothing method based on manifold learning to synthesize virtual multi-view samples with strong sense of reality.Our synthesized virtual samples have been proved efficient on different dataset using different algorithms.We propose a novel method called multi-feature discriminant and multi-manifold analysis based on heterogeneous graph and block separation.In order to make robust expression of the image blocks,we extract several features of the image blocks.To separate the inner-class information and intra-class information,we take the generated virtual sample that have the same semantics with the orign single training sample as a manifold to conduct discriminant feature extraction.To improve the expression of discriminant of the image blocks in sub-spaces,we put heterogeneous graph into the discriminant criterion to keep thereconstruction relation between image blocks of the same class and to inhibit the similarity relation between image blocks of different classes.Evetually,in the process of recognition,we use KNN classifier in every sub-space to determine the unknow label on the basis of integrating voting.Our proposed algorithm has been tested superior on different datasets.