Research On Matting Methods Under Natural Background

Matting and compositing are the two fundamental operations in computer graphics and visual effects. Traditional compositing operation derived from film industry is a compact and intuitionistic model and it has been used in modern compositing methods until now. Traditional matting operation is widely used in the related field, but it still has some limitations. Matting is an under-constraint problem in theory, because it has seven unknown parameters but only three functions. The existing matting methods were associated by some extra constraints to solve the matting problem. Matting methods can be divided into two kinds by the background of the image disposed by matting algorithm, one is called single color background matting and the other is natural matting.In this dissertation, we analyze all the methods from universality, convenience and veracity and point out some of their limitations. Then, this dissertation brings out three new methods enlightened by some models in computer vision. Firstly, based on that the structure between image and graph are similar, a graph based matting algorithm is realized using the max-flow methods to get the min-cut set. Secondly, an iterative optimize method is realized based on Markov Random Field model to improve the graph-based method. In order to solve the MRF model, we construct an energy function and use the belief propagation algorithm to get the minimal energy. Finally, we get the relationship between the defocus degree in a defocus image and the parameter of the camera by studying the image theory of the camera. We design an automatic matting method using two defocus images which we shot the same sense with different camera parameters as the extra restriction. In the method, we calculate the depth of the field and the defocus degree of the image, and then get the absolute foreground and absolute background of the image. And we get the alpha channel image using the Bayesian matting method to refine the unknown area which is the edge area of the foreground.The matting methods in this dissertation pay their attention to the natural matting and adopt the mainstream human and machine interface which has the least labor intensity. The first two methods use the existing model to the matting problem, and they get proper good results as the same labor intensity as the mainstream methods. The last methods is an original automatic natural matting methods, which has the similar matting result with the best matting methods but has an labor intensity equals to zero.