A saliency-based sampling method for image matting

Image matting is the problem of softly and accurately extracting the foreground from a natural digital image. Matting is inherently an ill-posed problem. Most previous methods use user-defined trimap as additional constraints. Given a trimap, a simple way to compute the alpha value is to select some known foreground and background colors for each unknown pixel. Existing sampling-based matting methods often consider samples near the unknown pixels only. They may fail if good samples cannot be found in the sample set. In this paper, we propose a new matting algorithm based on visual saliency. Our saliency-based method use adaptive sample collection method for each unknown pixel, and then propose an improved confidence function for these samples. It can avoid missing true color samples for each unknown pixel. Finally a Random Walk optimizer is employed to solve for the final alpha matte. Experiments show that our new matting algorithm produces both visually and quantitatively high-quality matting results. We demonstrate that our sampling matting approach improves the result of matting on complex scene images.