Alpha matte estimation of natural images using local and global template correspondence

Natural image matting is an interesting and difficult problem of computer vision because of its under-constrained nature. It often requires a user interaction, a trimap, to aid the algorithm in identifying the initial definite foreground and background regions. Current techniques use local or global image statistics of these definite regions to estimate the alpha matte for the undefined region. In this paper we propose a novel non-parametric template correspondence approach to estimate the alpha matte. This technique alleviates the problem of previous parametric algorithms that rely solely on colour information and hence are unable to exploit the image structure to their advantage. The proposed technique uses global and local template correspondence, to the definite know regions, to construct the background and foreground layers. Once the foreground and background colours are estimated, the final alpha matte is computed. According to the quantitative analysis against the ground truth, the proposed algorithm outperforms the current state-of-the-art parametric matting techniques.