|
Support vector machine (Support Vector Machine, SVM) is a new kind of machinelearning methods, it has made very great progress in recent years, and studies haveshown that this method can be used in many areas, and it has broad application potential.SVM integrates a number of technologies, such as the maximum interval hyper-plane,convex quadratic programming, Mercer nuclear, slack variables and so on. Up to now,SVM gets the best performance, and shows its advantages in challenging applicationssuch as signal processing, image recognition and gene pattern recognition.Handwritten numeral recognition is the core technology in dealing with some datainformation, such as statistical statements, postal code, bank notes and so on. Therefore,the research in this field has application significance. The key of SVM in thehandwritten numeral recognition lies in the digital image processing as well as theselection of the SVM kernel function. Vector set of low-dimensional space is usuallydifficult to divide, the kernel function just solve this problem masterly. In other words,an appropriate kernel function is selected if you want to obtain the classificationfunction of high-dimensional space. Under SVM theory, takes a different kernelfunction will lead to a different SVM algorithm. Therefore, the optimization ofparameter and SVM kernel function as well as the sample characteristics selectionshould be paid more attention, the parameter optimization mainly focused on twoarguments, which are C and.This article majorly focused on the research of the selection of the SVM kernelfunction parameters optimization method. First of all, comparing the three mainparameters optimize algorithm: grid search method, GA and PSO algorithm in theidentification accuracy. By comparison, we can find that the results obtained by thePSO algorithm are the best. Therefore, this algorithm is selected to optimize parameters.Then method that handwritten numeral character recognition based on improvedPSO-LSSVM is proposed, which has the advantages of both improved PSO andLSSVM. This article using Libsvm enhance toolbox and improved PSO-LSSVM todiscriminating digit. Above all, parameters are optimized with improved PSO, thenhave handwritten numeral recognition with LSSVM, which not only improved the localsearch capability, but also solved the problems that the training set category label is notvery accurate in the recognition process. The digital recognition results shows that the proposed new model showed better model stability, and effectively improved theaccuracy of handwritten numeral recognition, which has a certain vale, and can bewidely used in handwritten numeral recognition.
|