An efficient two-stage region merging method for interactive image segmentation

Abstract

Interactive image segmentation aims to extract user-specified regions from the background. In this paper, an efficient two-stage region merging based method is proposed for interactive image segmentation. An image is first over-segmented into many super-pixels using a bottom-up method. The color histogram is exploited to represent each super-pixel, and the Bhattacharyya coefficient is computed to measure the similarity of two adjacent super-pixels. Then some strokes, denoting the desired object and background, are manually labeled by the user on the over-segmented image. With the labeled seed super-pixels, a merging strategy is designed to realize adaptive region merging. The whole merging process is divided into two stages, which are repeatedly executed until no new merging occurs. In the first stage, some unlabelled super-pixels are merged into the labeled foreground or background super-pixels if the labeled ones are their nearest neighbors. In the second stage, any two unlabelled super-pixels are merged together if one super-pixel is the nearest neighbor of the other. Extensive experiments are conducted to evaluate the performance of the proposed method. The results show that the proposed method can extract the object reliably and quickly from the background.