Research On Face Recognition Algorithms Based On Kernel Method

Face recognition has the unique advantages of non-intrusiveness,convenience and friendliness,which makes it have important application and research value in the field of biometric recognition.Because of the changes of illumination,expression,occlusion and posture,the complex nonlinear changes of face images are produced,which makes face recognition a typical nonlinear problem.Collaborativet Neighbor Representation Classification improves recognition rate by searching the nearest neighbor representations of unknown samples in Euclidean space,but it belongs to linear algorithm,so the nonlinear problem in face recognition will limit its classification effect.When facing the nonlinear problem of face recognition,kernel method and manifold learning are two classical solutions.Most of the existing face recognition algorithms based on kernel method ignore the local manifold structure of samples in high-dimensional feature space,and only consider the nonlinear similarity of features.Traditional face recognition algorithms based on manifold learning construct local manifold structure of samples in input space,which can not make full use of the nonlinear structure of features.In view of the above problems,this paper studies the non-linear problem in face recognition.The main research work is as follows:(1)A new metric is defined in the kernel local projection space of the kernel Fisher discriminant analysis.By using this metric criterion,the nearest neighbor representation base of unknown samples is found in the kernel space.(2)Projecting all samples into the kernel space,constructing the cooperative nearest neighbor representation in the kernel space.(3)Based on the distribution of samples in feature space,we use manifold learning to find the local manifold domain of samples,embedding class labels for samples in the local manifold domain,constructing local manifold structure by neighboring classes of samples,constructing constraints by using local manifold structure in the encoding process.Experiments on several face databases,AR,Extended Yale-B,FERET and CMU Multi-PIE,the results show that our algorithm are compareble to others.