Region Fusion and Grab-Cut Based Salient Object Segmentation
Hailuo Wang, Bo Wang, Zhou Zhi-qiang, Song Lu, Sun Li, Shujie Wu · 2014
Accurate object segmentation remains a significant procedure in computer vision tasks. In this paper we propose a novel object segmentation method which based on region fusion and grab-cut. In the preprocessing stage, we segment the input image into superpixels as processing units. Then, we use a graph structure to model the superpixels and their correlations. To achieve the goal of region fusion, we transfer graph model into Minimum Spanning Tree (MST) model and fuse similar regions according to a threshold. Big superpixels are used to represent fused regions. By extracting color features and distant features of big superpixels and computing their saliency scores, we can get the high quality saliency map. Finally, we segment the salient object completely by using Grab-cut with the help of saliency map. Experiments show that our method outperforms state-of-the-art methods by achieving better segmentation results when evaluated using publicly available datasets.