Research Of Image Segmentation Based On Fuzzy Clustering Algorithm

In computer vision, image segmentation is the process of partitioning image intomultiple segments. The goal of segmentation is to simply or change the representationof image into something that is more meaningful and easier to analyze. The result ofimage segmentation directly influences the accuracy of representation and description,feature extraction, which will further influence the accuracy of pattern recognition andcomputer vision. Although a large number of image segmentation algorithms areproposed, there is no uniform method or strategy. So image segmentation is regardedas the bottleneck of computer vision, which inspires further research.Image segmentation is an ill-conditioned problem. On the one hand, theboundaries between image features may be non-crisp and ill-defined owing to unevenillumination or other external factors. On the other side, the partition criterion isusually non-unique or uncertainty due to the subjectivity of human cognition andvision. This kind of fuzziness and uncertainty in image segmentation cannot beeffectively modeled by the classical set theory, but can be preferably solved by fuzzyset theory. Methods based on fuzzy theory provide a solid theoretical foundation forthe analysis of fuzziness in image segmentation; and the fuzziness of pixel’smembership can be well described by fuzzy membership function.Fuzzy C-Means, abbreviated as FCM, is a kind of fuzzy clustering algorithmsbased on objective function; it not only preferably describes the fuzziness of pixel’smembership, but also meets the need of unsupervised segmentation. In FCM, imagesegmentation can be considered as the optimal solution of nonlinear programming withconstraints; the optimal distribution of membership functions is acquired byminimizing objective function; and the partition of pixels is realized afterdefuzzification of membership functions. FCM is simple in design, easy to implement,and can get effective segmentation, especially for medical images. Therefore, it hasbeen a popular image segmentation method in recent years. This paper studies theimage segmentation algorithms based on FCM, and mainly completes the followingwork:(1) Analyze the problem of clustering center initialization and similarity measure in FCM algorithm, and propose an improved FCM algorithm based on parameterinitializations. Initializing the clustering centers randomly has a great influence on thesegmentation of FCM. It is easy to trap the objective function into a local minimum,which will cause non-uniform segmentation results in different executions of FCM,besides affects the convergence of FCM. To solve problems above, an initializationmeans of clustering centers is proposed on the basis of image histogram. Moreover, toenhance the algorithm robustness to different cluster structures, high dimensiondistance metric is introduced into with the help of Gaussian kernel function, andmeanwhile an estimate of radial width of Gaussian kernel function is given. To acertain extent, the improved FCM algorithm enhances the effectiveness as well as theefficiency of image segmentation.(2) Analyze the anti-noise performance of image segmentation by FCM，andpropose an adaptive FCM algorithm, which improves the neighborhood control factorby combing the relationship of spatial location, intensity and homogeneity in structurebetween pixels. The interaction of neighborhood pixels can ensure the intensityhomogeneity of segmented regions; and with influence of normal pixels the clusteringprocess of noised pixels can be corrected. The anti-noise performance thereby can beenhanced. Besides, to enhance segmentation accuracy of the edge, the localre-segmentation is applied. Therefore, the improved adaptive FCM algorithm has goodnoise immunity as well as edge preserving.