Compared with the traditional statistics, Statistical learning theory (SLT) is a theory that specialized in machine learning with finite samples, and SLT provides a firm foundation to support vector machines (SVM). SVM is considered as a candidate to replace neural networks and other traditional classification methods for its good performance and high generalization ability.However, SVM is a novel learning machine and still has many aspects that need further research, such as most of research is limited in theory, and lack of applications. On the other hand, using Quadratic Programming optimization techniques to train SVM is time-consuming, especially when the training data set is very large.This thesis applies the newest theory, SVM, in the application of image classification. It studies the information of positive and negative images which user provides. We compare the process of study using three kind of kernel function, the polynomial kernel function have better generalization ability than Gaussian and sigmoid. At the mean while, we discuss about the effect on SVM's performance which image features cause, multi-features in different classes for image can raise the performance of classification.A SVM-based image classification System is described based on the research mentioned above. A number of classification experiments with real images are performed to validate theories above and some results are presented. |