A Novel Image Segmentation Algorithm: Region Merging Using Superpixel-based Local CRF Model

摘要

Natural image segmentation is a special application of pattern classification for the discrete random variables in two-dimensional space. Many remarkable segmentation algorithms are proposed in the framework of Bayesian Networks. In this paper, a region merging approach based on the SLIC superpixels and local CRF models is provided. The corresponding interaction potential function of the CRF model involves both color and edge features, which are computed based on the diffusion distance of histograms and the Canny operator respectively. The merging of regions is realized with the propagation through neighboring overlapped local CRF models. By comparing with some competing models, the experimental results show that the proposed model is more precise and robust and it is capable of dealing with uneven illumination, complex background and local textures.