A Fuzzy Similarity Based Image Segmentation Scheme Using Self-organizing Map with Iterative Region Merging

This paper presents a new region-based segmentation scheme which considers homogeneous regions as constituted of pixel blocks that are highly similar to their neighborhoods. Based on the postulate that each homogenous region can be represented by an exemplary pixel block, segmentation is done by grouping contiguous pixel blocks whose neighborhoods are highly similar to the exemplary pixel blocks. In our approach, the degree of similarity between one pixel block and its neighborhood is determined via fuzzy similarity, while the exemplary pixel blocks are automatically discovered by Kohonen self-organizing map. The discovered pixel blocks are later used to split the image into its constituent regions. To obtain a more discernible result, a two-stage iterative merging technique based on Region Adjacency Graph (RAG) is applied. The proposed scheme has been evaluated using real images with results that are comparable and in certain cases better than the morphological watershed segmentation.