Research On Digital Matting Method

Image matting refers to the problem of accurately extracting the foreground of arbitrary shape from an image, which plays an important role in image and video editing. It is an extremely challenging problem because we must estimate the foreground and the background colors, as well as the foreground opacity for each pixel from a single color measurement. Due to its highly under-constrained property, human intervention is required to solve this ill-posed problem. Commonly used human interventions are trimap and scribble, i.e. labeling some pixels which are definitely foreground or background pixels.By analyzing problems within several existing digital image matting algorithms, we propose a graph model based matting algorithm and a fusion based matting algorithm in this paper.In order to solve the over smoothing problem suffered by closed form matting algorithm, we propose a graph model based image matting algorithm in this paper. Our graph model based algorithm firstly estimates an initial alpha value for each unknown pixel by optimized color sampling method, and then uses the similarity between pixels to propagate the alpha values of which confidence values are high, so as to refine the alpha values whose confidence values are low. Compared with several related matting algorithms, our graph model based method can achieve more accurate alpha matte, and can effectively avoid over smoothing problem.By analyzing different matting algorithms and comparing their matting results, we find that most algorithms are only effective for images of certain types, and no one claims the best among others. Different algorithms may have their different cons and pros. Inspired by this, we propose a fusion based matting algorithm in this paper, which integrates the strengths of different matting algorithms to achieve the best matting results. In order to acquire more continuous and smoother alpha matte, we propose two optimization methods to smooth our initial fusion result. The first method constructs smooth term by introducing a matting Laplacian matrix, and the second one employs dynamic programming based trajectory optimization algorithm to refine the initial fusion result. In our experiments, we fuse the alpha results of 12 different algorithms to obtain our matting result. Compared with state-of-the-art algorithms, our fusion based matting algorithm produces more superior results according to the evaluation on standard benchmark dataset.