Research On Interactive Color Image Segmentation Methods Based On Multiway Cuts

Image segmentation is the process of dividing images into several non-overlapping regions by adopting supervised or non-supervised ways. It is one of the most basic research topics in computer vision, directly affecting the quality of the image analysis and image understanding. Graph Cuts is one of the current main image segmentation methods. Its significant advantage is the ability to effectively integrate multiple image features and priori information, ensuring the real-time, effectiveness and accuracy of the segmentation results by less interactions. In this thesis, the researches are conducted in the framework of multiway cuts which is the variant of Graph Cuts extended to the multi-label problem.The researches on multiway cuts are mainly along the two directions. One is about the nonlocal extension of smoothness term(corresponding to the n-links) of energy functional, whose purpose is to overcome the inherent “shrinking bias” problem in graph model. Secondly, the building of the data term(corresponding to the t-links) of the energy functional are optimized. An asymmetrical kernel density estimation method is presented to improve the ability of describing image color feature.Standard multiway cuts model(SMCM) is difficult to segment thin elongated regions, concave regions and other small parts, due to “shrinking bias” problem. In this thesis, nonlocal extension of SMCM is conducted, and a strategy to search for nonlocal homogeneous points based on mean shift trajectories is presented. Because of the local and nonlocal information being integrated to multiway cuts model at the same time, “shrinking bias” problem is thus effectively overcome.Traditional kernel density estimation method treated each kernel function equally in the overall estimation, leading to less discernibility to different image regions. So the obtained segmentation result is not satisfactory. Aiming at this problem, an asymmetrical kernel density estimation method is presented. Particularly, the proposed method treats each kernel function differently, and the weight coefficient is introduced in the kernel density estimation function to express the contribution of each kernel function to the overall estimation.