Research On Optimization Method Of Support Vector Machine Kernel Based On Multi-scale Kernel Weighted-fusionon

Support vector machine(SVM)is an excellent machine learning method.It is widely used in data mining and pattern recognition applications.Support vector machine is based on the statistical learning theory and its structural risk minimization as the principle.The complexity of the learning machine and the learning ability of the learning machine are controlled by the principle of structural risk minimization.So support vector Machine has better generalization ability.SVM has a very good advantage in solving the problem of small sample machine learning at high latitudes.It does not have many problems with traditional machine learning methods,such as local extremum,over-fitting and so on.Support vector machine has been Extensive application and development of the theory since proposed.And the theory has also been a very big improvement.However,there are still some problems,such as how to choose the kernel function and the related parameters,how to use the multi-scale kernel learning method and so on.There is no perfect unified model to follow.Therefore,SVM kernel function selection,SVM parameter selection,multi-kernel learning and other methods to do research is very necessary and meaningful.Therefore,this article carries out detailed and in-depth research on SVM kernel function selection,SVM parameter selection,multi-kernel learning.Based on the kernel polarization,the kernel function is selected to improve the generalization ability of the SVM algorithm without increasing the complexity of the original algorithm.In this dissertation,we have done four aspects of the work as follows.(1)The machine learning theory,the statistical learning theory,and the support vector machine theory are studied.Then,the mathematical model of the SVM algorithm is described in detail by the minimum space between the optimal hyper planes.Then the linear SVM algorithm is deduced from the mathematical model of the algorithm,and the linear SVM is extended to the nonlinear SVM by kernel function.(2)The method of selecting the kernel function,especially the kernel target alignment which is independent of the SVM algorithm,is mainly studied.The commonly used kernel function metric methods such as kernel metric criterion,kernel polarization criterion and generalized multi-class kernel polarization criterion are analyzed.The method is independent of the specific algorithm and does not need training SVM.The algorithm is simple with low complexity,and it is a very effective kernel function selection method.(3)This paper introduces multi-kernel learning method,and you can find multi-kernel learning method performance better than single-kernel learning method.Multi-scale kernel learning method is a kind of multi-kernel learning method,which is flexible and practical.Inspired by the advantages of kernel and multi-scale kernel learning methods,a multi-scale kernel learning method is proposed,and the theoretical analysis and proof are given.(4)The proposed method is tested on six classical image datasets,and compared with the other four classification methods.It is found that the proposed method has higher recognition rate and stability.