Salient Object Segmentation Based on Superpixel and Background Connectivity Prior

Salient object segmentation is well known for detecting and segmenting objects using saliency map as input. In this paper, we propose a salient object segmentation method which integrates saliency, superpixel, and background connectivity prior. First, our method extracts the superpixels of an image by the simple linear iterative clustering algorithm. Second, based on superpixel representation, we use background connectivity prior to characterize the spatial layout of each superpixel in a color space with respect to the image boundary. Third, considering both saliency and background connectivity values, we label four kinds of superpixel-level seeds and feed them to superpixel-level GrabCut method. Because superpixel representation generates only a few seeds, the optimization of GrabCut method converges fast. Finally, for further improvement, we crop a rectangular region which contains the segmented object obtained in the third step and apply GrabCut at the pixel level to produce the final segmentation result. Experimental results on eight typical datasets demonstrate that in terms of both performance and computational efficiency, the proposed segmentation method outperforms existing state-of-the-art methods.