Research On Multiple Kernel Learning And Its Application On Image Classification

Choosing the right kernel function for the given task is the core issue of the kernel method.The support vector machine(SVM)algorithms mostly use single kernel function to deal with non-linear classification problems.However,the generalization capability and robustness of single kernel based SVM are usually limited in practical application as the data are usually heterogeneous,multiple feature and huge sample size,so it does not apply to all specific problems.Compared to the single kernel method,multiple kernel learning(MKL)methods are more flexibility while not only depend on the choice of kernel function by combining several basic kernel functions according to different strategies to transfer the data representation in the feature to the choice of the kernel weights.In recent years,various theories and application researches have improved the MKL model in some degrees,which not only the generalization ability but also the performance of SVM.This paper is based on support vector machine,which describes the basic theory of MKL.What's more,this paper does some researches on several classical MKL algorithms and analyzes the optimization processes and the existing problems of these algorithms.Use this as a foundation,this paper proposed an elastic multiple kernel learning method based on the newton gradient optimization and the multiple kernel learning based on feature selection and kernel target alignment.The main work of this paper is reflected in the following aspects:(1)This paper is based on the two-class SVM and introduces the basic concepts and properties of the linear and non-linear classification of the SVM;The theory and framework of MKL are expounded,and the method of multiple kernel support vector machine is given;Learning current MKL algorithm deeply and doing comparative analysis of their advantages and problems.(2)For the present sparse multiple kernel learning model may lead to discard useful information,and the generalization ability is degenerated when produces the sparse weight of kernel function.Furthermore,the multiple kernel learning method which is based on gradient descent method has a slow convergence speed when close to the optimal solution,this paper establishes the elastic multiple kernel learning model which is based on multiple kernel learning and proposes an elastic multiple kernel learning method based on the Newton gradient optimization.By using the second order Newton gradient descent method,the algorithm has less number of iterations to reach convergence.The validity of the algorithm is verified by experiments.(3)As the pre-selected kernel will greatly affects the performance of multiple kernel learning,this paper proposes a general method to pre-select a set of reasonable base kernels based on feature selection and kernel target alignment before the optimization process of multiple kernel learning.After choosing the base kernels by feature selection,these base kernels will be used in the optimization process of the existing multiple kernel learning to generate better result.The validity of the algorithm is verified by experiments.