A novel merging criterion incorporating boundary smoothness and region homogeneity for image segmentation

A novel joint region merging criterion combining region homogeneity and boundary smoothness is proposed. Previous watershed segmentation method which utilizes region homogeneity or edge integrity or both gives good results in some cases. However, for complex scenes such as images of a vehicle with irregular roadside objects reflected on the window panes, it tends to give undesired segmentation results with region boundaries not residing on real physical boundaries. Aiming at improving the segmentation of these complex scenes, we propose the incorporation of an additional measure of boundary smoothness into a new joint criterion. Based on this, an affine transform invariant measure of the smoothness of boundaries is developed, which is the equivalent width of the energy distribution function over frequencies, obtained from Fourier descriptors of the boundary. Experimental results and evaluation are presented in this paper to demonstrate the merits of the proposed method.