FUZZY CLUSTERING METHODS FOR IMAGE SEGMENTATION

Abstract

This thesis focuses on the study of the fuzzy c-means clustering algorithm (FCM) and related validity issue. The aim is to exploit the potential of this automatic and unsupervised learning strategy so as to develop new approaches for image segmentation. The FCM and its validity measures are first introduced. The self-organized (optimized) nature of FCM and the supervisory capacity of the validity criteria are highlighted since they together provide a promising approach through which automatic and unsupervised image segmentation can be achieved. An automatic image segmentation method utilizing both spatial and gray-level domain information is presented. In this approach, thresholding is performed dynamically according to local rather than global characteristics. By applying the validity-guided fuzzy c-means clustering algorithm, the spatial partitioning of an image is context-oriented and fully automatic, unlike conventional local techniques that blindly partition the image into overlapping or non-overlapping rectangular subregions prior to local analysis. Experimental results indicate that the algorithm possesses robustness to uneven illumination, noise and presence of shadows. When clustering algorithms are applied to image segmentation, the goal is to solve a classification problem. However, these algorithms, including FCM, do not directly optimize the classification quality. Consequently, a new proposal emphasizing the importance of post-clustering evaluation and modification is presented. The concept of density is introduced as a validity criterion to assess the quality of an algorithmically generated partition and accordingly guide an amelioration process through split-and-merge operations. Our aim is to adjust the classification to the most reasonable pattern from the point of view of cluster density. Sample images corresponding to 1-, 2- and 3-dimensional pattern analysis are tested and the results confirm the effectiveness of the proposed algorithm.