An Adaptive Algorithm Based On Fuzzy C-means And Fuzzy SVM For Image Segmentaion

Image segmentation is a key step in image processing since accuracy of segmentation has a great effect on the quality of image processing. Fuzzy C-means clustering (FCM) and fuzzy support vector machine (FSVM) are two commonly used algorithms in image segmentation. Actually, FCM algorithm has achieved a huge success in medical image processing. While, image collection, storage and the other early stage of the image processing may distort some image data, i.e., images usually contain noises. These noises usually lead to poor segmentation results.In order to reduce the influence of noises, some researchers have attempted to utilize the local information of image to enhance the denoising ability of FCM and FSVM algorithms. However, choosing appropriate parameters in the enhanced algorithms is a big problem since they are very sensitive to noises. For resolving this problem, this thesis is devoted to modifying these algorithms by providing more adaptive parameters. The main results are summarized as follows:To overcome the problem of noise sensitive of neighborhood radius in Fast Generalized FCM, fuzzy memberships obtained by FGFCM algorithm are utilized to measure the noise in-tensity around a pixel, according to which, one can adjust the radius of the corresponding local window. Due to this automatically adjusting, the influence of noise in image segmentation can be reduced and more image information can be reserved. A new adaptive algorithm, N-FGFCM al-gorithm is proposed based on these ideas. Many numerical experiments have been implemented on both artificial and real images for comparing FCM, FGFCM and N-FGFCM algorithms. The experiment result shows that the new algorithm can get better segmentation accuracy, stronger adaptation and is less time consuming.A new algorithm, Auto-NFSVM is proposed for resolving the noise sensitive and heavy computation problem. In this new algorithm, sample labels and fuzzy memberships are obtained by N-FGFCM algorithm automatically. FSVM classifier model trained by these samples is demonstrated to be more robust to noises though the numerical experiments. The results of experiment also illustrate that compared with FCM-FSVM and KNN-FSVM, the new algorithm has stronger generalization ability and can maintain more local image information.