Image Classification Research Based On Improved PSO Algorithm

Recently, Support Vector Machine(SVM) has been widely used in image classification, which is a common and well-performed classifier. It is still an aporia to optimize the parameters of SVM.The performace of SVM suffers from whether choosing appropriate parameters or not. As a swarm intelligence algorithm, Particle Swarm Optimization(PSO) algorithm shows good results for SVM’s parameter optimization. This paper researches from three aspects: the speed of SVM, the optimization ability of PSO, the performance on image classification. The main contribution of this paper can be summarized as follows:1. A new version of SVM with fast intersection kernel is proposed in this paper. By transforming the way to calculate the intersection kernel, the time of classification during prediction is reduced with the time complexity from O(m ?n) to O(n ?log m).2. Aiming at the problem of premature convergencein PSO, the reason for falling local optimum is discussed. To solve this problem, the mutation PSO is addressed.The particles mutate at a particular probability during iteration, which enhance PSO’s ability to global searching and prevent prematuring convergencein.3. Inspired by Quantunm Particle Swarm Optimization(QPSO), the PSO algorithm based on normal distribution function is proposed, with the analisis on the reason of the QPSO’s great performance. During the process of updating particles’ position, with normal distribution function added, the distribution of particles is more scientific and uniform. Great result is obtained in the experiment of standard test functions, and the proposed algorithm has much better ability to the speed and robustness of convergence compare to standard PSO and QPSO.4. When dealing with unbalanced data in image classification, using SVM method will lead to low accuracy on minority class. In this paper, weight-SVM(with fast intersection kernel) is used to increase the performance on classification. In addition, because of adding too many parameters by using weight-SVM, the impoved PSO proposed in this paper is applied. This algorithm is experimented in Caltech 101 dataset and Pascal Voc2007 dataset, which has a certain improvement compared with other traditional and classic algorithms.