Research On Face Recognition Algorithm Based On Sparse Representation

With the development of computer technology,face recognition technology has been gradually applied to people's daily life.At present,many recognition methods have been proposed,among which the recognition based on sparse representation has its unique advantages,such as excellent robustness against interference of occlusion and corrosion,and insensitivity to feature selection,etc.But,the principle of its operation leads to the fact that sufficient samples are necessary to achieve good recognition effect.However,in practical application,many scenes cannot collect enough faces.On the other hand,the computational complexity of sparse representation classification is relatively high.In order to be applied in practice,the speed and accuracy of recognition must be accelerated.In this paper,improvements are made for these problems.The main research contents and contributions are as follows:(1)In order to solve the problem of insufficient samples,this paper proposes a new virtual sample construction method.First,the original sample was enlarged by double cubic interpolation,and then the enlarged sample was divided into non-overlapping blocks with the size of 2×2 pixels.The pixels in the same position of each block were taken out to form a virtual sample of the same size as the original sample,and a total of 4 virtual samples were obtained.Then all the virtual samples and original samples are combined into a new training sample set,and the discriminant sparse representation is used for two-stage classification.Experiments on different face databases show that the algorithm has a higher recognition rate than most sparse representation algorithms.(2)In order to solve the problem that global structure will ignore local information,this paper proposes a sparse representation method based on local and global fusion.First of all,the global structure corresponding to each class of training samples is constructed.For any class of training samples,its global structure does not include all other training samples of this class.Then,this kind of training sample and its corresponding global structure are respectively used for linear fitting of the test sample to obtain the two fitting errors before and after,and then the former error is divided by the latter error to obtain the fused error,and finally the error is classified with this error.Experiments on different face databases have proved that the recognition rate of the fusion algorithm proposed in this paper is higher than that of the algorithms based on local or global only,and the fusion algorithm combined with the virtual samples proposed in this paper is also higher than that of most of the current face recognition algorithms based on sparse representation.(3)In order to speed up the recognition,this paper proposes a fast two-stage recognition algorithm.First,k-means clustering algorithm is used to deal with the training sample,the training samples are closer to each other's sample aggregate into a broad categories,for a new test samples,only need to be calculated and various kinds of clustering center distance,select distance nearly several categories,these categories contains the original training samples included in the category of the all the training sample,a new training sample set,the use of the new training sample set,a second phase of the identification.Experiments on different face databases have proved that the fast twostage recognition algorithm proposed in this paper greatly accelerates the recognition speed on the basis of a small increase in accuracy.(4)Integrating the factors of recognition speed and accuracy,this paper proposes a sparse representation recognition method based on fast two-stage and virtual samples.Firstly,the virtual sample generation algorithm proposed in this paper is used to expand the training samples,and then the expanded samples are classified by the fast two-stage algorithm proposed in this paper.A large number of experiments show that this algorithm can greatly speed up the recognition speed when the accuracy is similar to that of the single algorithm with virtual samples added.