A Belief-based Pixel Labeling Strategy for Medical and Satellite Image Segmentation

In this paper, a belief-based pixel labelling strategy is introduced and applied to image segmentation. The procedure exploits the results of a K-means clustering algorithm for first quantifying the membership degree of each pixel to a region in the image. Belief functions are then used to quantify the uncertainty about the pixel labelling. This theoretical framework allows to manage imprecise and uncertain information extracted from spatial neighbors. Using this strategy (classification + post-labelling), the segmentation scheme allows to perform a complementary approach combining region segmentation and edge detection. Segmentation results on different kinds of image are presented and allow to highlight the algorithm performances. Finally, the application of this methodology is also presented in case of 3D medical imaging and multi-spectral satellite images.