Svm-based Image Classification

Automatic classification of images is a critical task in many areas, including information retrieval, visual scene detection, Internet data filtering, medical applications, etc. When operated directly in the image, the conventional methods always difficult to obtain good results because they are poor in high dimension performance. However, Support Vector Machine (SVM) can overcome the defect of high-dimensional representation, it has been widely used in image classification.Since the classification capability of SVM greatly dependent on the selection of kernel parameters, therefore, this study focused on the kernel parameter selection methods, then the different color, texture features were employed in image classification experiment.The main work of this paper is as follows:1. This paper analyzes the kernel parameters of SVM classify model, and use the Grid Search method, Genetic Algorithms and Particle Swarm Optimization Algorithm to optimize the SVM kernel parameters. Then compare their capability in kernel parameter optimization by experiments.2. A new kernel parameter optimization method is proposed. Because control parameters selection of PSO have no corresponding theoretical guidance, most choices are based on experience, this paper proposed a Genetic Algorithm and PSO hybrid algorithm. In the new algorithm, Genetic Algorithm is employed to select the PSO control parameters, improvements were also made in the inertia coefficient. Experimental results demonstrate that the proposed PSO algorithm outperforms the traditional approaches while selecting optimal parameters for SVM, it has better search capability and avoid the algorithm fall into local optimal.3. Using different quantitative methods for color features extracting in HSV color space, based on the traditional histogram loss color space information, a regional weighted and color moment combined color feature extraction method is proposed. The co-occurrence matrix, Tamura method, Gabor-wavelate were employed to extract texture features. The dissertation investigated the dimension reduction of LBP pattern, to overcome the limitation of the uniform LBP pattern, a new LBP hybrid pattern combined with PCA was developed for texture feature extraction.4. The GA-PSO based SVM has been applied in image classification, the recognition performance about image's color, texture and other characteristics were tested, analyzed and compared. This paper also analyzes the influence of different parameters optimization methods on image classification accuracy.