Research On Sparse Representation-based Classifier

Sparse representation-based classifier is an important research field of data mining and machine learning.It can learn and establish classification model based on a large of relative sample data.The classification function,which is obtained by the classification model,is able to map the unknown sample to a certain one of given categories,thus forecasting this unknown data.It has been widely used in all kinds of fields,such as finance,security,face recognition and so on.Under the support of the funds of National Natural Science Foundation(61471124),aiming at the problem of low recognition rate,slow computation rate and poor self-adaption of the sparse representation-based classifier and its improved algorithms,this paper proposes kernel and sparse representation-based fast classifier algorithm,two-phase collaborative sparse representation classifier algorithm and kernel and sparse representation-based adaptive classifier,respectively.The details are as follows:Firstly,aiming at the problem of the low recognition rate and slow computation rate of the sparse representation-based classifier,this research topic comes up with an improved local sparse representation based classification algorithm,namely kernel and sparse representation-based fast classifier.And this algorithm blends the idea of KNN algorithm and sparse representation algorithm.First of all,it uses the kernel-induced distance to look for the N nearest training samples.Then,we can find the relative classes of testing sample through these near training samples and use them to compose the dictionary.Finally,we use the learning dictionary to collaboratively represent the testing sample.And the smallest constructive residual deviation is regarded as the standard to classify.The experimental results show that the improved local sparse representation based classification algorithm has excellent recognition rate and calculation speed.Secondly,we explore the performance of the improved two-phase collaborative sparse representation classifier based on the above classifier algorithm.The first stage is to calculate the kernel-induced distance between the testing sample and each training sample.Then the samples,of which the kernel-induced distance of each class of training samples is minimum,are selected to compose a dictionary.And we use it to collaboratively represent the testing sample.Then,we choose N classes training samples by the smaller residual error.In the second stage,we apply the selected N classes of training samples to compose a new dictionary and use it to collaboratively represent testing sample again.And the smallest constructive residual error decides which the testing sample belongs to.The experimental results show that the improved two-phase collaborative sparse representation classifier algorithm has outstanding recognition rate.Finally,directing at the self-adaption problem that how to choose the value of the number N of the nearest samples can achieve the optimal effect in the design of local sparse representation-based classification,this paper presents an improved algorithm which is named kernel and sparse representation-based adaptive classifier.Specially,firstly,it uses the method,which is similar to leave one out cross validation,to obtain a residual curvilinear function in the training samples.Then we exploit the degree of mutation of the curvilinear function to forecast the most appropriate value of the number N of the nearest samples.Finally,we apply this value of N to the above classifier algorithm.The experimental results show that the algorithm can easily reach the best result of the above classifier algorithm.The improved local sparse representation algorithm is proposed on the basis of the technologies of collaborative sparse representation classifier algorithm.Then this issue further studies the problem of recognition rate and adaptability of this improved algorithm and the software development is completed in the Matlab platform.Experiments show that the design of this paper has good performance and high practical value.