Research On Occlusion Face Recognition Based On Low Rank Sparse Decomposition

Face recognition,including face image acquisition and acquisition,face detection,image preprocessing and classifier design,is one of the most widely used techniques for biometric identification.In recent years,with the rapid development of computers,Internet of Things and sensing technology,non-occlusion face recognition technology has made significant progress.However,in the case of occlusion,the captured face image is lost compared with the original image.The original facial features of the face lead to a sharp decline in the recognition rate of existing algorithms.Therefore,for the problem of occlusion interference,this thesis mainly studies and learns the low-rank sparse decomposition of occlusion face images,and discusses how to extract the low-rank structure representing the eigen features of faces from the occlusion face image,thus improving face recognition.Accuracy,the specific work content is as follows:(1)In view of the shortcomings of the traditional Robust Principal Component Analysis(RPCA)to obtain inaccurate facial features,this thesis proposes occlusion face recognition algorithm based on Iterative Weighted Low Rank Decomposition(IWLRD).Firstly,the weighted function is used to weight the sparse interference matrix.The weight of the occlusion part is larger,and the weight of the non-occlusion part is smaller,so that the occlusion and noise interference factors contained in each training sample are accurately extracted.Then,for the test sample The problem of the occlusion of the training sample is different,and then the IWLRD algorithm is used to extract the information of the training samples covered by the occlusion area in the test sample.Finally,inspired by the Extended SRC(ESRC)classifier,a new joint dictionary is constructed by constructing a new joint dictionary for the face low-rank feature matrix,occlusion matrix,noise matrix and test sample for each training sample.It is expressed as a linear combination of linear sparse representations of the new joint dictionary,and the residuals are used for classification and discrimination.The experimental simulations are carried out on the AR library and the Extended Yale B library.The experimental results show that compared with the existing algorithms such as Extended Sparse Representation based Classification and Low-rank Sparse Representation based Classification,the IWLRD algorithm proposed in this thesis has better occlusion robustness and accurate face recognition.The rate has improved significantly.(2)In the traditional RPCA algorithm,the nuclear norm is used instead of the rank function for low rank sparse decomposition.However,the nuclear norm is greatly affected by the singular value,which leads to the bias of the optimization results.In view of this,this thesis proposes an Iterative Weighted Low Rank Decomposition based on Logarithmic Determinant(IWLRDLD)algorithm.Firstly,the logarithmic determinant function is used to replace the nuclear norm for low rank sparse decomposition,and the face low rank feature matrix and the sparse interference matrix are obtained.Then,the face occlusion matrix is extracted by the IWLRD algorithm.Finally,the face low-rank feature matrix,the occlusion matrix,the noise matrix and the occlusion vector of the test sample of each type of training samples are constructed into a new joint dictionary,and the test samples are represented as a sparse linear combination of the new joint dictionary.The residual is calculated by sparse approximation,and the obtained coefficient is used for classification and discrimination.The experimental results show that the recognition rate of the IWLRDLD algorithm is further improved in AR library and Extended Yale B library.At the same time,for the problem that the singular value decomposition of large-scale matrix results in high time complexity,this thesis proposes a Fast Iterative Weighted Low Rank Decomposition based on Logarithmic Determinant(FIWLRDLD)algorithm,the idea of matrix tri-factorization is introduced into the logarithmic determinant function,which can avoid solving the singular value decomposition problem of large-scale matrix and effectively reduce the time complexity of the algorithm.