Efficient Interactive Image Segmentation with Local and Global Consistency

Abstract

Interactive image segmentation models aim to classify the image pixels into foreground and background classes given some foreground and background scribbles. In this paper, we propose a novel framework for interactive image segmentation which builds upon the local and global consistency model. The final segmentation results are improved by tackling two disadvantages in graph construction of traditional models: graph structure modeling and graph edge weights formation. The scribbles provided by users are treated as the must-link and must-not-link constraints. Then the graph structure is modeled as an approximately k-regular sparse graph by integrating these constraints and our extended neighboring spatial relationships. Content driven locally adaptive kernel parameter is proposed to tackle the insufficiency of previous models which usually employ a unified kernel parameter. After the graph construction, a novel three-stage strategy is proposed to get the final segmentation results. Experimental results and comparisons with other state-of-the-art methods demonstrate that our framework can efficiently and accurately extract foreground objects from background.