Sparse Representation Based Feature Extraction And Classification Methods

Recently,sparse representation has been widely applied in variety of fields such as computer vision,pattern recognition and machine learning.Learning useful and compu-tationally convenient representations from complex redundant,,and highly dimensional visual data is very important for the success of sparse representation technique.However,how to select appropriate features that can best work with the learned sparse represen-tation remains an open question.In addition,dictionary plays an important role in representation based classification methods,how to learn an appropriate dictionary to improve the classification performance also a hot research topic.The main contributions of this paper are summarize as follows:Firstly,we propose two optimal discriminative projection methods for sparse repre-sentation based classification and a multiple kernel sparse representation based discrimi-native projection method.The basic idea of these methods are based on the decision rule of classifier.which aims to learn a projection matrix such that the within-class sparse reconstruction residual is minimized and the between-class is maximized in the projected low dimensional subspace.Thus,sparse representation based classification or multiple sparse representation based classification can achieve better performance in the projected low dimensional space.We use the trace ratio maximization method and the stochas-tic gradient ascent algorithm to optimize the objective function,respectively.Extensive experimental results demonstrate the superiority of the proposed algorithms.Socondly,we propose a sparse representation based joint,kernel dietionary and dis-criminative analysis learning mcthod.The basic idea of the proposed method is to map the data from the input space to a high dimensional feature space.Then learn a projec-tion matrix and a kernel dictionary smultaneously sucli that in the reduced space the sparse representation of the data can be easily obtained and the reconstruction residual can be further reduced.For improving the discrinminant ability of sparsce coding,we also learn a dictionary in the proposed method.Thirdly,we propose a,collaborative representation based joint kernel dictionary learning and discriminative projection learning method.In order to reduce the com-putational complexity,we use the l2 norm to replace the l1 nrom.The proposed method aims at joint learning a projection matrix and a dietionary which can achieve the smallest reconstruction residual of a given data set while the largest between-class reconstruction rcesicdual in the low dimensional subspace.The optimization scheme can be effectively solved based on gradient descent.Extensive experimental results validate the superiority of the proposed approach.Finally,we propose two cost-sensitive dictionary learning methods,i.e.,cost-sensitive dictionary learning and cost-sensitive joint feature and dictionary learning methods.The main idea is to introduce the cost information into dictionary learning process such that the designed dictionary is able to produce cost-sensitive spare coding,resulting in mini-mum cost rather minimum error.In order to further reduce the total cost,we introduce cost information into feature extraction phase and simultaneous learned with the die-tionary.Extensive experimental results on a series of face databases demonstrate the effectiveness of the proposed methods.