Image Segmentation And Object Classification Based On Support Vector Machines

Image segmentation and object classification are two important topics of digital image processing. Traditional classification approaches based on statistical theory have been extensively applied in the two research areas. Traditional classification approaches, which are based on the principle of Experiential Risk Minimization instead of Expected Risk Minimization, achieve the best, when the number of training samples is infinite. Because the number of training samples is often limited and data dimension is high, the performance of traditional classification approaches is often unsatisfied in practice. Compared with statistical theory, statistical learning theory focuses on the machine learning of small sample size and can trade off between the complexity of models and generalization performance. Support vector machines, which are based on Vapnik-Chervonenkis (VC) dimension theory and Structural Risk Minimization principle, are considered good candidates because of their high generalization performance without the need to add a priori knowledge, even when the dimension of the input space is very high and the problem is nonlinear. This dissertation studies image segmentation and object classification based on support vector machines. The main contents and contributions are as follows: (1) The effects of kernel function, model parameters and window size on the segmentation performance of support vector machines are investigated, and some valuable conclusions are drawn, which will provide references for using support vector machines to segment images. (2) In order to improve the segmentation performance of images corrupted by impulse noise and Gaussian noise, fuzzy weighted support vector machine is proposed. Experimental results show that fuzzy weighted support vector machine enhances the robustness to impulse noise and Gaussian noise. (3) To resolve the unclassifiable region of one-against-one support vector machine for multiclass problems, an improved one-against-one support vector machine based on distance measure is presented. Compared with other variants, the proposed method reduces complexity while preserving the classification performance. (4) Because of existing nonlinear correlation and various noises among selected texture features, kernel principal component analysis is used to extract features from selected features. Experimental results indicate that support vector machine is a good choice to segment images with blurry edges, intensity non-uniformity and discontinuity (such as medical images). (5) Because shadows cast by moving vehicles are often detected as a part of the moving vehicles since shadows move in accordance with the movement of vehicles, which will reduce accuracy of vehicle detection, an algorithm is proposed to suppress the moving cast shadow for vehicle detection based on four properties of the moving cast shadow. In order to ensure detection system to normally work under these conditions that light changes dramatically and the camera may vibrate due to wind or traffic resulting in geometric displacements between image frames, a robust filter algorithm for binary subtraction image is proposed. (6) In order to resolve the vehicle classification that is a difficult problem, fuzzy weighted support vector machine is selected to be a classifier. Experimental results indicate that the classification accuracy of fuzzy weighted support vector machine is satisfied.