A method for image segmentation based on rough sets theory was presented. Objects in the clustering process often have similarity relation instead of equivalence relation. Rough sets theory was applied in similarity relation to solve clustering issue. In general, clustering results are easily corrupted by noises. In order to decrease noise disturbance, maximum membership principle of the boundary points and roughness are defined. Clustering evaluation function was presented on this base. The method provides us a new viewpoint on processing image.