Research On Sparse And Low Rank Representation Based Classification Algorithms

Feature extraction and classification techniques are the hot topics in the community of pattern recognition.Due to the variations in illumination,expression,pose and occlusion,how to extract robust features from images remains a key problem.Moreover,representa-tion based classification methods(RBCM)have attracted considerable attention in recent years and applied in various image classification tasks,and one representative RBCM is sparse representation based classification(SRC).However,SRC has the following three drawbacks:1)the coefficient vector of test sample contains negative entries;2)SRC needs clean training images;3)SRC directly employs all the training data as the dictionary.The above drawbacks hamper the performance of SRC in practical applications.In this disser-tation,we focus on the aspects of robust feature extraction,nonnegative representation based classification,discriminative low rank representation and discriminative dictionary learning,and we propose several improved approaches.The main research results are as follows:(1)A subspace learning method from the second order image gradient orientation-s(SOIGO)is proposed.Recent researches uncover that,in contrast to the first order gradient information,the second order gradient information can better characterize the geometric properties of images.Inspired by this finding,we perform principal compo-nent analysis(PCA)on the second order image gradient orientations to obtain com-pact features.Furthermore,due to its merit of computational efficiency,collaborative representation based classification(CRC)is employed to classify the extracted features.Experimental results indicate that,under the scenarios of real-world occlusion,artificial occlusion,expression and illumination variations,the proposed method is superior to its competing approaches with few training samples,and even outperforms some prevailing deep neural network based approaches.(2)A class-specific residual constraint non-negative representation based classifier(CRNRC)is developed.The recently proposed NRC ignores the relationship between the coding and classification stages.Moreover,there is no regularization term on the coefficient vector in the formulation of NRC,which may result in unstable solution leading to misclassification.In this dissertation,a class-specific residual constraint is introduced into the formulation of NRC,which encourages training samples from different classes to competitively represent the test sample.The proposed CRNRC can be viewed as an integration of CRC and nearest subspace classifier(NSC)with the non-negative constraint.The superiority of CRNRC is validated on the face databases,handwritten digit datasets and large-scale visual classification datasets.(3)A face recognition method based on low rank matrix recovery with structural incoherence(LRSI)and low rank projection is proposed.Conventional low rank matrix recovery approaches mainly focus on recovering the low rank part of original training samples,while ignoring the corrupted test samples in the test stage.To address this problem,we learn a low rank projection matrix in LRSI,which can correct corrupted test samples by projecting them onto their corresponding underlying subspaces.The low rank part of training samples is employed as the dictionary,through which the corrected test samples are classified.Experimental results show that the proposed method can achieve superior classification performance compared to LRSI.(4)A low rank representation with block-diagonal structure is presented.SRC can-not effectively tackle the situation when both the training and test images are corrupted.To obtain discriminative coefficient matrix,a regularization term that captures the block-diagonal structure of the target representation matrix of the training data is incorporated into the framework of low rank representation.Meanwhile,dictionary learning is intro-duced to learn a compact and discriminative dictionary.The resulting problem is solved by the augmented Lagrange multiplier(ALM)method.Finally,a simple yet effective linear classifier is used for the classification task.Experimental results on benchmark datasets show that the proposed method is robust to appearance variations in illumination,real disguise,random pixel corruption and artificial occlusion.(5)A locality constraint dictionary learning with support vector discriminative term is designed.The locality of atoms is not fully explored in conventional discriminative dictionary learning approaches,and they utilize a single decision rule to classify the test samples.In a departure from traditional methods in which the graph Laplacian matrix is derived from the original training data,we preserve the locality of atoms on the basis of the learned dictionary.Consequently,the graph Laplacian matrix used in this dissertation is updated with the updating of dictionary in the optimization process.Moreover,we exploit both the regularized residual and multi-class SVM for classification.Experimental results on standard databases demonstrate the superiority of our proposed method over previous dictionary learning approaches on both hand-crafted and deep features.