Research On Face Recognition Based On Low-rank Decomposition And Sparse Representation

With the development of scientific research and commercial applications,artificial intelligence technology has been widely used in many fields such as medical care,transportation,security,and Internet finance.As an important branch of artificial intelligence technology,biometric recognition has been researched and paid attention by research organizations and scholars at home and abroad,and has gradually become a research hotspot in computer vision and artificial intelligence.Among them,face recognition technology is most universal.Therefore,it still has great theoretical significance and application value for face recognition research.Based on the above points of view,this paper mainly focuses on the feature extraction and classification recognition process of face recognition algorithms,aiming to propose face recognition algorithms with high accuracy and robustness in a complex and varied training environment.The main research contents of the thesis are as follows:1.In the real application,the result of face recognition is easily affected by the acquisition angle of the training and test data,and multi-view will cause larger identification errors.This paper presents a face recognition algorithm for multi-view data.Firstly,we perform face detection,noise reduction and other image preprocessing operations on the data.Then the low-rank decomposition model based on two different regular terms is designed to complete the feature extraction operation,and the first type low-rank feature and the second type low-rank feature are extracted.Finally,the sparse representation classification is based on the characteristics of minimum residuals for classification and discrimination.A residual ratio comparison model based on the minimum residual and the second least residual is designed,and the final classification results are determined based on the residual ratio comparison.By designing a strategy using a comparison model,it is possible to effectively select the correct recognition result among the two characteristics.2.In order to further deal with the complex sampling situation,this paper considers the multi-view effect and considers other complicated sampling environments such as illumination,occlusion and expression.A face recognition algorithm with high efficiency and robustness is proposed to overcome the complex environmental changes and insufficient training samples.The algorithm uses non-convex rank approximation norm and kernel norm to perform two low-rank matrix decompose methods to achieve the purpose of removing the occlusion interference.Firstly,we use the non-convex robust principal component analysis to acquire low-rank dictionaries that remove light and occlusion.In order to speed up the convergence of the algorithm,the obtained low-rank dictionary is used as the initialization,and the second low-rank decomposition based on the kernel norm is performed to obtain the low-rank dictionary with the relevance between classes removed for classification.Finally,for the problems of undersampled training data and excessive proportion of occlusion samples,the external data that do not participate in training and classification are used as auxiliary data.The auxiliary data are used to learn the dictionary in simulating the possible occlusion,light,and other effects that may occur through the classification.We obtain the minimum reconstruction residual and the final classification result by solving the convex funtion with auxiliary dictionary and the low rank dictionary jointly used.This paper carries out detailed experiments on the two proposed face recognition algorithms.Experiments show that,the proposed algorithms can deal with the data in many classic face databases,and all of the experiments can obtain high-efficiency recognition accuracy.