Soft-Constrained Random Walk Propagation for Hierarchical Image Matting

For image matting, the affinity-based algorithms adore the hierarchical propagation to handle the memory problem. But they rely heavily on the accuracy of the pre-segmentation, and the hard input constrains may lead to the detail missing or foreground error. In this paper, we present a random walk based hierarchical algorithm with extended Dirichlet function. Regarding constrains as regularization, the new random walk framework can receive the guidance information from the known or pre-computed regions, while respect the image context as much as possible. Based on this framework, we perform the two levels affinity propagation. First, we conduct the regularized random walk in the superpixel-based graph and obtain the coarse matte which will be served as the soft constraints in the following propagation. Then in the pixel-based graph, the improved random walk is carried out the second time to recover more details for the final matte. Experimental results reveal that the soft constrained approach can propagate more efficiently and preserve more details, which make the proposed method perform better in accuracy and robustness.