Full Space Inverse Representation-based Classification Model And Its Application In Face Recognition

As a biometric feature recognition technology for identity recognition through face visual feature information,face recognition has always been a research hotspot in recognition field.And it has broad application prospect in many fields.Sparse representation-based methods have been praised by researchers because of their good performance since applied to the field of face recognition.There are,however,following problems in the face recognition based on sparse representation: 1.dependence on sufficient training samples of the same class;2.ignorance of complementary between training samples and test samples;3.sensitivity of proportion of training samples and test samples;4.without considering the subspace structure characteristics of the data.In view of these problems,in this paper we have done the following research work:Firstly,a full space inverse representation classification model with sparse and low-rank constraints is constructed.Firstly,in order to extract the information hidden in test samples and make full use of complementary information between test samples and training samples,a new representation named full space inverse representation is proposed.The representation method uses dictionary consisted of training samples and test samples to linearly represent training samples.Furthermore,in order to improve the representation ability and discriminant ability of the model and to study the structural information between samples,we add lowest-rank and sparsest constraints on model.Then we solve the model rapidly by improved split Bregman iterative algorithm and give the convergence theorem and proof of the algorithm.Secondly,we apply the classification model to face recognition.We conduct experiments on three public face image database.For the lack of labeled samples and superfluous unlabeled samples in face recognition,we mainly designs the experiments of four aspects: 1.analyze the influence of classification results with change of parameters in model.2.discuss the complementarity between samples.3.the convergence of model and corresponding algorithm is analyzed.4.discuss the robustness of model and corresponding algorithm.It is worth noting that,we analyze emphatically the influence of recognition results of different models with change of number of training samples and test samples.The experimental results show that the proposed model is robust and has better recognition performance than other classicalmethods when the number of training samples is little and the unknown samples were occluded.