An Affinity Based Matting Method Based on Multi-Scale Space Fusion

At present, affinity based image matting methods are mainly divided into Matting Laplacian methods and Nonlocal methods. However, their search ranges are fixed and cannot be changed according to the input trimap of the image. This paper combines the characteristics of the search of the Nonlocal method and Matting Laplacian method, and proposes an affinity based matting method based on the fusion of multi-scale spaces. This method firstly distinguishes some opaque foreground and opaque background in the image through preprocessing steps. Secondly, this paper uses a nonlocal method to perform an initial global search on the image, and uses a two-layer space search method, so that for different trimap input, more search ranges can be covered. Thirdly, this paper adopts the Matting Laplacian based on the local smoothness assumption of foreground and background colors to perform a further matting step for the local details of the image. Finally, a sparse linear equation is solved to calculate the final alpha result. Experiments show that, compared to other single-layer and single-search affinity based methods, the multi-layer fusion method proposed in this paper achieves a large final accuracy rate in calculating opaque pixels, mixed pixels, and overall results. At the same time, in terms of visual effects, the method in this paper is also superior to various methods that have appeared.