Research On Non-constrained Face Recognition Via Joint Multiple Dictionary Combined With Extended Sparse Representation

Face recognition(FR)has become the representative technology of biological recognition,thanks to its advantages of difficult loss,unforgeability,easy carrying and acquirement in the mainstream of identity authentication.Although in many areas,face recognition has shown strong recognition and security performance,but there are still some difficulties in the practical application,such as the illumination,posture and expression changes,artificial occlusion,and inadequate sample size under-constrained situation.How to solve these problems and extend the robustness of the existing face recognition algorithms has become the main research.direction of this paper.According to the achievement of dictionary learning and sparse representation in the pattern recognition,joint multiple dictionary learning combined with extended sparse representation method is put forward for face recognition.The performance of face recognition algorithms have been improved by establishing multiple dictionaries and combining with joint extended sparse classification method in practical application.The concrete research contents are as follows:(1)To solve the problem of cost sensitive face classification,a coarse to fine facial security authentication method is presented,which is named limited expression and pose based cost sensitive face secure verification(FSV)model.Considering the fusion of the cost sensitive information and sparse coding learning process,the two-fold dictionaries: the discrimination dictionary and the confirm dictionary are constructed,which are respectively used for coarse recognition and fine confirmation.Meanwhile,for enhancing the robustness and reducing the sensitive cost values,adaptive space distance weighted way is carried out.Then by sparse representation,the classification is proceeded.The relationship of cost sensitive penalty value per case is fully estimated to adjust the algorithm for the face re-authentication research.(2)By the methods of occlusion dictionary modification and image intersection clustering,an algorithm of varying occlusion detection and iterative recovery for face recognition is proposed.Image processing and modified surface occlusion dictionary are used to determine occlusion parts of the samples and separate the occlusion area to set up occlusion maps.Then with the help of occlusion maps,the occlusion areas are iteratively recovered by sparse method and combined to form a non-occlusion face image.The global features can be reflected t.These images can improve the performance of face recognition in the classification process.Finally,by comparing with three typical occlusion map detection methods,the good detection and recovery performace of our algorithm is verified.(3)Considering that the dictionary which contains rich information can contribute a lot to sparse coding,this paper put forward the group dictionary learning and joint kernel collaborative model.This model can mine the effective information from prototype and virtual space.The original and virtual training images are used to constitute a sample group and misplaced atom dictionary by misplaced combination.On the one hand,the amount of sample varying information is expanded.On the other hand the encoding and distinctiveness ability of the dictionary are greatly enhanced.And kernel collaboration representation classifier by joint optimization can be used to capture more nonlinear structure and reduce the time consumption.Though simulation experiments,the recognition performace of the algorithm is verified in the case of small training set.(4)Due to the unbalanced contribution of sample structure in manifold space to correct classificationt,kernel collaboration representation based manifold regularized model is proposed by merging face manifold structure and sparse coding in kernel space.The general kernel collaboration algorithm is extended to enhance the robustness and improve the recognition accuracy by using optimal neighbor base.At the same time,the local texture feature extraction combined with the optimized kernel function algorithm is used for image classification can not only reduce the sensitivity of the algorithm to occlusion and illumination image,but also get more nonlinear structure of the face image.By simulation experiments,the recognition performace of the algorithm is verified under different occlusion and noise.