Parameter Optimization And Application Of SVM Based On Improved Particle Swarm Optimization

Support Vector Machine(SVM)is a machine learning method based on statistical learning theory.Because SVM has uni que advantages in solving problems such as sm all samples,high dimensions,and nonlinearity,it is widely used in the fields of face recognition,text classification,and image processing.When using SVM to deal with classification problems,the se lection of SVM parameters has a great impact on the classification results,so optimizing SVM parameters is the main research direction of applying SVM.Particle Swarm Optimization(PSO)has advantages of simple algorithm,easy implementation and fast convergence speed,many scholars have used it to optimize SVM parametersin recent years.However,PSO has disa dvantages of prem ature convergence,easy to fa ll into local optimum and low classification accuracy.Aiming at these problems,this paper focuses on the problem of SVM parameteroptimization based on improved PSO.The specific work of this paper is as follows:(1)An adaptive PSO(WPSO)is used.This paper uses adaptive weight instead of original inertial weight,the weight of each particle is dynamically adjusted according to its fitness value,so as to balance the global and local optimal capabilities of the particle;On the basis of adaptive weight,this paper uses adaptive mutation to optimize PSO,som e particles are given a certain probability of mutation,andthegroup best value of some particlesare mutated,thereby enhancing population diversity.(2)A PSO with simulated annealing mechanism(SAPSO)is u sed.In the speed update formula,the self-cognition part and the social cognition part of particle are dynamically adjusted.Thispapercompares the fitness value of the current particle with the fitness value of the group best.If the current particle is better than the group best,the current particle is accepted as the group best;otherwise,according to the roulette strategy,the current particle is accepted as the group best with a certain probability,so that the particle can jump out of the local optimal solution and achieve the global optimalsolution.(3)SVM classification modelsare established based on improved PSO,namely SVM optimized by adaptive PSO(WPSO-SVM)and SVM optimized by PSO with simulated annealing mechanism(SAPSO-SVM).This paper applies the algorithms WPSO-SVM and SAPSO-SVM to the classification of remote sensing images.The remote sensing image of Jiangsu University of Science and Technology and remote sensing dataset Pavia U are selected as experimental data.The experiments are performed using the algorithms SVM,GA-SVM,ACO-SVM,PSO-SVM and the algorithms WPSO-SVM and SAPSO-SVM proposed in this paper.The classification accuracy of remote sensing image are compared and analyzed using classification confusion matrix,overall classification accuracy and Kappa coefficient.The experimental results show that the algorithms WPSO-SVM and SAPSO-SVM can more easily extract features of ground objects and improve the classification accuracy of remote sensing images.