Interactive image segmentation based on Gaussian Mixture Models with spatial prior

In this paper, an interactive color natural image segmentation method is proposed. The method extends the Gaussian Mixture Model (GMM) by taking into account user markers as useful spatial prior. In this way, a distribution combining with color and spatial location is obtained. The distribution is incorporated in a Bayesian MAP approach. The approach is formalized as an iterative energy minimization problem. A direct global minimization technique based on variational method is employed to get binary solution. After each iteration, the largest connected region from foreground is used to update foreground GMMs and achieve more superior performance than original GrabCut method. Extensive experiments are performed on public benchmark datasets such as GrabCut benchmark, Berkeley segmentation database and Graz benchmark. The results show that the proposed method can extract the object region from the complex background accurately. © 2015 SERSC.