Research On Image Segmentation Algorithm Based On Fuzzy Clustering

Image segmentation is a classical issue in the field of image treatment and computer visual, as well as a key technology in image analysis, recognition and identification. It is widely applied to many areas, such as computer vision, image coding, model identification, and medical image analysis. The results of image segmentation directly affect the quality of the analysis, recognition and interpretation in the later phase. In the numerous segmentation algorithms, image segmentation based on clustering analysis is a very important and popular algorithm. And the algorithm of FCM (fuzzy C-means) is the most widely used in clustering analysis.Fuzzy C-means (FCM) clustering segmentation is a kind of datum clustering means on the basis of the optimization of the fuzzy objective function. Its purpose is to partition the special vector data points into c interspaces. And the characteristic of clustering result is that a datum is represented by its degree of membership to the clustering center. However, FCM clustering segmentation has its inherent drawbacks. On one hand, the data amount n of the image sample is so great (an image of 256 x 236, n=65536), that a great deal of time and space resource will be consumed for the sample data of big size by the use of FCM. On the other hand, the segmentationmodel is incomplete due to the weak resistance to the noise or the low level of robustness in the process of image segmentation by means of FCM, and the only use of grey feature without consideration of the special features of pixels. So FCM is only suitable for segmenting images of low noise. For the above mentioned, this thesis focuses on how to make effective use of space information, and offers possible solutions to the drawbacks of FCM in image segmentation. This thesis can be summed up as the following:Firstly, this thesis explores the standard image segmentation of FCM and studies such issues as the selection of the initial number of clustering, and the identification of the initial center and the initial membership matrix.Secondly, this thesis studies the existing FSC image segmentation based on gray information and space information and proposes a modified FCM algorithm: incorporating the space information in standard FCM, projecting pixel space onto feature space of its gray histogram, and thus realizing fast clustering. Then on the basis of FCM fast clustering, of the full use of pixel's neighbor spatial feature, and of the modified membership function, new membership function will be established. With the new membership function, image will be segmented. The experiment of the new segmentation has shown its effectiveness and the robustness to noise.Finally, from the perspective of neighbor membership constraints, a new clustering objective function is proposed, and FCM image segmentation based on neighbor membership constraints is obtained. With the new segmentation putting into experiment, the result of the experiment has shown that with the various value of penalized coefficient the results of image segmentation have been also varied.