Researches On Color Image Segmentation SVM Approach Based On Automatic Selection Of Training Samples

Image segmentation is important in the field of image processing and it is the first step in image analysis. The purpose of image segmentation is to extract the target that people care about from the image. Many researchers at home and abroad have been proposed nearly a thousand kinds of different segmentation methods, but most of the traditional image segmentation algorithm is effective for a certain type of image. There is no a segmentation method can be generally applied to different types of images. Because there is no uniform image segmentation standard, traditional segmentation method only applies to the particular application.In recent years, a learning method based statistical learning theory-support vector machine (SVM) is proposed. It can achieve good generalization ability in the small sample space and it is a strong generalization ability classification algorithm. Therefore, the SVM algorithm applied to image segmentation has become a more generally applicative segmentation method with good effect. SVM-based image segmentation method is largely based on classification using some image pixel’s gray and other features as training samples to train SVM, and then use the trained classifier to segment the whole image. However, SVM algorithm is a supervised classification algorithm. It needs manually providing different types of training samples when applied to image segmentation. But manually selecting training samples. is random, time-consuming and affecting the segmentation results. Therefore, how to automatically select the training samples and making training samples wildly representative of the samples’type will be the emphasis of the SVM applied to image segmentation.For the problem of selecting training samples about image segmentation method based on SVM, this thesis proposed two methods about how to automatically select and label the SVM training samples.The main works are concluded as follows:(1) The thesis analyzed and summarized the image segmentation approaches based on support vector machine. The main problems of SVM image segmentation algorithm were proposed and researched.(2) The color image segmentation method based on FCM and SVM is proposed by combining the SVM and Fuzzy C-Means(FCM). Firstly, FCM clustering algorithm is used to originally segment image with automatically and randomly selecting training samples in the tow categories. Then, color and texture features are extracted from the image as attributes of training samples of SVM. Finally, the images are segmented by the trained classifier. The experiment results on Berkeley image database demonstrate that the proposed approach achieved good segmentation performance.(3) The color image segmentation method based on watershed and SVM is proposed by combining the SVM and watershed algorithm. The center points of areas divided by watershed approach are selected as training samples of SVM and labeled by comparing with the segmentation reference image. Moreover, image color and texture features characteristics are the attributes of the training samples. The experiment results demonstrate that this approach compared with previously proposed approach can further improve the segmentation accuracy rate and achieve better segmentation results.