Face Recognition Based On Kernel Sparse Representation

Face recognition as one of the most natural and not easy to be aware of biometric identification technology can be widely used in national security, public safety, network information security, home entertainment and other fields. Face image is easily affected by expression, illumination, pose, occlusion, and other factors. The human eye can distinguish face changes, is difficult for a computer to recognize the complicated changes of human face with high accuracy and speed. Face recognition is one of the research hotpots in the field of pattern recognition.In recent years, sparse representation theory is one of the most concerned methods in face recognition. The theory is robust to illumination, noise and occlusion, and requires more quantity of samples. In this paper, to improve the robustness of the theory and better handle the problem of collecting fewer samples of each class, we propose face recognition algorithms based on improved sparse representation, the main research work is as follows:We start from the feature extraction of face image and use PCA(principal component analysis), LDA(linear discriminate analysis) and KPCA(kernel principal component analysis)algorithm to extract the features of face images and reduce the image dimension. It also reduces the computational complexity of sparse representation. Then we introduce the existing sparse representation and its extension addressing less training sample problem.Finally, the classifier of the corresponding algorithm is introduced.The theory of sparse representation uses the information of training samples to reconstruct the test samples in the sample space. In order to make full use of the nonlinear information within the sample data, we use the Gaussian kernel function to determine the nonlinear mapping. Samples of the original input space are mapped to high dimensional feature space, and consequently the distribution of the samples is changed. In the feature space, we use the sample feature information to reconstruct the test sample feature, and propose kernel sparse representation based classification algorithm and kernel extended sparse representation classification algorithm. These algorithms can significantly improve the typical sparse representation based classification algorithms.To improve the effectiveness and efficiency of face recognition, this thesis presents fast sparse and fast kernel sparse representation for face recognition algorithm. In these two fast algorithms, we use a coordinate descent algorithm to improve the calculation speed of solving1 l norm minimization problem. In this process, the corresponding sparse representation coefficients are obtained by iterative optimization, and then the recognition time is shortened.In summary, the experiments prove our methods can obtain satisfactory recognition rate.Especially the improved fast sparse representation and kernel sparse representation algorithm reduce the time complexity.