GrowMatting: A GPU-based real-time interactive method for image matting

摘要

Natural image matting is one of the major challenges in modern image editing, whose goal is to estimate fractional opacity of foreground layer from an image and recover the colors of the foreground/background layers respectively. Most of existing methods require complicated mathematics computation and are implemented on a single processor. The closed-form matting [1] is relatively faster and more robust than previous methods. However, the sequential matting process is still time consuming and may not produce the best matting results because of the over-smoothing problems. In this paper, we propose a real-time interactive matting method on a graphic processing unit(GPU), with the following advantages: First, it is overall faster to get high quality results. Second, the user can modify constraints during the whole process. Third, the over-smoothing problem is overcome with a novel stopping criterion. Fourth, users can refine matting result inside a local region in a more convenient way.