Research On Cooperative Representation Subspace And Low-rank Regression Model

Machine learning and pattern recognition as hot research directions,have been paid attention to by many researchers all the time.The study of pattern recognition started early and was developed into a special research field in the 60 s of last century.The research of pattern recognition can be generalized into four parts,namely,data acquisition,data preprocessing,data feature extraction and data classification.Feature selection and extraction is one of the most important steps.The fast development of science and technology has led to the rapid growth of data.However,the high dimensional data is not conducive to the analysis of the structure of data for us.We can understand pattern recognition as the classify for unknown samples and label the data.There are many kinds of classification methods at present.The representative classification methods include k nearest neighbor classifier,bias classifier and random forest method.Besides,subspace learning method,neural network method and popular deep learning method in recent years,etc.,are also associated with pattern recognition.It is found that sparse and low-rank methods can effectively improve the robustness of learning models even if the picture is polluted by something such as noise,illumination or occlusion.Sparse and low-rank methods can also extract features effectively and improve learning efficiency.Sparse representation classifiers,subspace learning methods based sparse representation and low-rank regression methods occupy a high proportion in pattern recognition field,such as face recognition,action recognition and so on.With the development of researches on sparse and low-rank in pattern recognition field,many novel models based on sparse and low-rank have been put forward,which have achieved satisfactory results in practice,and at the same time,these have been attracted the attention of researchers.The main purpose of this master's thesis is to enhance the robustness and improve the efficiency of the algorithm,and then propose a new low-rank regression model based on cooperative representation local discriminant analysis for low-rank regression model.The main contents and innovations of this master's thesis can be summarized from the following aspects:(1)A brief analysis of the study background of this master's thesis and the present situation of the research.Two algorithms,SRC and CRC,are introduced as typical representatives of sparse representation for pattern classification.(2)Several classical subspaces learning methods,subspace learning based on sparse representation and low rank regression model are introduced,and the related methods are briefly analyzed in this master's thesis.Besides,the master's thesis points out corresponding shortcomings of these methods.(3)The objective function of sparsity preserving projection(SPP)is analyzed from two aspects.Based on the analysis,a variable space cooperative representation discriminant analysis(VSCRDA)algorithm is proposed.Firstly,the shortcomings of the related methods are pointed out.Then,a new objective function is proposed,and the correlation coefficient between data and data is represented more accurately by solving the reconstructed coefficients two times.The feasibility of the proposed method is verified through many experiments.(4)A low-rank regression model is proposed.Based on the analysis of the related low-rank models,it is found that these low-rank models do not relate to the associations among different categories,which will increase the possibility of the error approximation of data in mapping space.To overcome the shortcomings,a low-rank regression model based on local discriminant analysis is proposed.By introducing the association among different classes in the objective function,the discriminant and robustness of the model have been improved,and the possibility of error approximation is also reduced.