A background recovery based matting approach

Alpha matting aims at extracting foreground object accurately from an image. In alpha matting, all pixels are considered as the composition of its foreground and background color by an opacity value (alpha). Alpha matting is an extremely challenging task for its ill-posed nature. Some existing techniques aim at finding the best foreground/background color pair to every pixel by sampling nearby pixels. Some others rely on color propagation like solving matting affinity matrix. In this paper, we propose a novel method combined the advantages of both sampling and propagation based approaches to get high quality matte. First, we propagate the information of background to unknown region using inpainting technique. Then, we use a novel sampling method to get the estimation of foreground color. Finally, we solve a cost function that minimizes the color distortion combined with alpha matte smoothness error. Experimental results show that high quality matte can be obtained using our method.