Research On The Improvement And Expansion Of Linear Regression Classifiers

With the rapid development of information technology and its rapid integration in daily life,there is a growing need for people to identify users accurately and effectively,thereby ensuring the security of information and public sphere.Face recognition has attracted widespread attention because of its non-invasive and concealed characteristics.Although face recognition technology has improved day by day,how to maintain the stability of the algorithm and improve the classification results under the conditions of illumination,facial expression change and mask change still faces great challenges.Regarding the key research problems existing in the face recognition method as the starting point,this dissertation improves and expands the linear regression classifier.The specific work is as follows:(1)An orthogonal discriminative projection for linear regression classification(ODP-LRC)is proposed.Linear regression classification(LRC)is an effective classifier which exhibits good performance in face recognition.However,the original feature space cannot guarantee the classification efficiency of LRC.To solve this problem,this dissertation proposes a new dimensionality reduction method called ODP-LRC,which has the direct connection to LRC.According to the classification rule of LRC,the objective function of ODP-LRC is designed.What's more,the orthogonal constraint is imposed on the projection matrix to remove redundant information.Based on the orthogonal constraint,a new trace ratio optimization method is used to solve the objective function.Experiment results demonstrate that ODP-LRC is better than many competitive methods.(2)An orthogonal discriminative projection for locality-regularized linear regression classification(ODP-LLRC)is proposed.All the linear subspace-based algorithms have a same assumption that samples of the same class are in the same linear subspace,but images in the real world are not always linear.To solve this problem,locality-regularized linear regression classification(LLRC)extends the traditional LLRC with a manifold learning procedure to remove the face image that does not conform to the linear hypothesis.This dissertation proposes a new dimensionality reduction method called ODP-LLRC,which has the direct connection to LLRC.According to the classification rule of LLRC,the objective function for ODP-LLRC is derived by maximizing the inter-class local reconstruction errors and simultaneously minimizing the intra-class local reconstruction errors.To obtain the solutions,the relationship between the linear regression coefficients and the projection matrix is exploited,and an alternative iterative optimization method is used.Experiment results demonstrate that ODP-LLRC is better than many competitive methods.(3)A joint discriminative projection and dictionary learning for domain adaptive linear regression classification(JDPDL-LRC)is proposed.LRC has shown impressive performance in many recognition problems.However,when confronted with situations where the training data has different distribution with the test data,the performance of LRC will be degraded significantly.To address this problem,this dissertation proposes a novel dictionary learning method called JDPDL-LRC.JDPDL-LRC is designed according to the classification rule of LRC,and it aims to learn the projection matrix of each domains and the shared dictionary such that the between-class reconstruction residuals are maximized and the with-class reconstruction residuals are minimized in the low-dimensional subspace.To obtain the optimal solution,the relationship between optimization variables is exploited,and an alternative iterative optimization method is used.Experiment results on single-source domain and multi-source domain adaptation databases show that the proposed method is better than many competitive domain adaptation methods.