Research On Face Recognition Based On Sparse Representation

In recent years,face recognition technology has been closely concerned by various scientific research institutions and university scholars because of its convenience,concealment and stability,and gradually developed into a key research direction in the field of pattern recognition.It is also widely used in public security systems,financial security systems,media data retrieval and other fields.With the in-depth study and application of face recognition theory,this technology has achieved an ideal recognition effect under controllable conditions.Among the various identification methods,the sparse representation method based on compressed sensing using the principle that the data signal can be compressed,breaks through the limitation of the Shannon theorem,which has good robustness.However,the recognition effects of these methods are not unsatisfactory in an uncontrollable environment.The face recognition algorithm based on sparse representation is widely used in the recognition and classification of face images because it has the advantages of being insensitive to feature extraction and easy to construction algorithm model.In this paper,an improved algorithm is proposed for the problem of poor recognition and instability under the conditions of illumination variation and facial occlusion: we propose the block face recognition algorithm based on the residual fusion,optimal segment strategy,the improved sparse coefficient-norm value detection method for distinguishing non-face images which 2?aims to improve the robustness of the face recognition system.In this paper,simulation experiments are carried out in the standard face database,the experimental results show that the proposed algorithm enhances the robustness of face recognition system.The main innovations are described as follows:(1)In this thesis,we propose the block face recognition algorithm based on the residual fusion.This method segment the test image and the training image into blocks based on the fusion of the sparse coefficient and residual,and performs non-superimposed training in the overcomplete dictionary composed by each sub-block to obtain the classification residual of each sub-block.We use the new residuals as the classification criterion which is obtained by fusing the residuals of each sub-block,this method completely preserves the sub-block information without interference factors,and reduces the negative influence of the noise-containing sub-block on the entire face recognition process.(2)The optimal face segmentation strategy is proposed,by adopting an appropriate face image segmentation method,the integrity of the important organ information of the face can be preserved,and the feature block and the noise-containing block can be effectively segmented.The robustness simulation experiments are carried out in the AR face database and the Extended Yale B face database,the face samples are processed in multiple ways by using the block recognition algorithm,and finally acquired the optimal face segmentation strategy by their best recognition rate.The optimal face segmentation strategy optimizes the block face recognition algorithm,so its still has a good recognition rate in an uncontrollable environment.(3)In order to identify non-human faces,this paper proposes an improved sparse coefficient-norm value detection method for discriminating non-human faces.By training 2?non-face and face samples,the characteristics of the sparse coefficients are analyzed and the classification threshold is obtained.Comparing the sparse coefficient-norm detection 2?value with the classification threshold,when the detection value is greater than or equal to the threshold rule,it is a valid human face,then the residual face fusion algorithm is processed;when the detection value is smaller than the threshold rule,directly output non-human faces.This paper performs multiple verifications in the standard face database,the experimental results show that the improved residual fusion classification method is robust to noise images which contains occlusion and illumination.