Researches On Clustering-aided Sparse Representation Classification Algorithm

Sparse Representation Classification(SRC)is a classification algorithm based on sparse representation and over-complete dictionary construction.The idea of the SRC algorithm is to use as few over-complete dictionary atoms as possible to linearly represent the original signal,and then obtain the classification result by comparing the similarity between the sparse signal and the dictionary atoms of each class.At present,based on the high accuracy and strong robustness of this algorithm,it is widely used in the fields of image recognition,text classification and signal detection.With the development of SRC,an SRC algorithm that combines clustering advantages to optimize dictionary construction is proposed.The method reduces the dimensionality of the dictionary by extracting valid information from the training data to reduce the time complexity.This paper investigates the clustering-aided SRC algorithm,and uses the MNIST dataset to simulate the proposed algorithms.The main work and contributions are as follows:(1)Clustering-aided SRC algorithm based on training set segmentation.Clustering-aided SRC algorithm enjoys high accuracy and robustness in signal classification field,but as the number of training samples increases,the time complexity of clustering will be greatly increased.To solve the problem,this paper proposes two improved clustering SRC algorithms based on training set segmentation:a)Clustering-aided SRC algorithm based on direct training set segmentation.First,we split the total training set into multiple subsets and implement clustering-aided SRC algorithm on each subset;second,we combine the results of each subset to classify the signal.Compared with the traditional clustering-aided SRC algorithm,this algorithm greatly reduces the clustering time and improves the parallelism of the algorithm when the classification performance is similar.b)Clustering-aided SRC algorithm based on training set segmentation and improved residual combination.Based on the previous algorithm,this algorithm compares the reconstruction residuals of each subset,and constructs a new dictionary for SRC.Then,it uses the newly obtained residual to improve the residual combination result of the previous algorithm to classify the signal.Compared with the previous proposed algorithm,the classification accuracy of this algorithm improves with slightly increased time complexity.(2)Clustering-aided SRC algorithm based on prior information.Clustering-aided SRC algorithm does not use the prior information of the test object in the sparse representation process and residual calculation process,resulting in information waste and performance degradation.This paper proposes two improved clustering-aided SRC algorithms based on prior information:a)Clustering-aided SRC algorithm based on prior information assisting support set selection.In the sparse representation process,we use the distance between the test object and the dictionary atoms to calculate the probability of each dictionary atom as a support set atom,and then take these probabilities as prior information to improve support set selection process.Compared with the traditional clustering SRC algorithm,this algorithm greatly improves the classification accuracy.b)Clustering-aided SRC algorithm based on prior information assisting residual calculation.Based on the previous algorithm,we use the distance between the test object and dictionary atoms as priori infomiation to improve residual calculation process.Compared with the previous algorithm,this algorithm further improves the classification accuracy.The improved algorithms proposed in this paper optimize the traditional clustering-aided SRC algorithm from two aspects of clustering time and classification accuracy,and achieve better performance,which is helpful for the practical applications of SRC.