局部自适应输入控制的随机游走抠图

针对传统性抠图算法中，非完全正确用户标注及不精确超像素分割造成的信息误扩散，以随机游走算法为基础，提出带软性约束的抠图算法。通过对扩展Dirichlet问题的推导，指出带软约束的随机游走与部分自吸收随机游走概率的关联性。以吸收概率为指导，在传统相似扩散所构建的图模型上，根据局部窗口内特征矩阵的秩与方差设计了输入控制矩阵，使得信息扩散的过程能够跟随图像的局部特征进行自适应扩散。最后将软约束随机游走应用到单帧双层抠图及视频抠图中。实验表明，所提算法具有信息远距传播能力和良好的容错性能，尤其在用户标注不够充分的情况下能够取得更加优良的抠图结果。

Random-walk matting with local adaptive input control

In traditional image-matting algorithms, incomplete user labeling and inaccurate super-pixel segmentation lead to the incorrect propagation of information. To solve this problem, we propose the use of soft constrained matting based on a random-walk algorithm. Through the derivation of the extended Dirichlet problem, we identify the relationship between a random walk with soft constraint and the probability of a partial self-absorption random walk. Guided by the absorption probability, an input control matrix is designed according to the rank and variance of the feature matrix in the local window. This is performed via a graph model that was constructed using traditional similarity diffusion such that the process of information diffusion could follow the local image features to realize adaptive diffusion. Finally, we applied the soft constrained random walk to single-frame two-layer image and video matting. The experimental results reveal that the proposed algorithm can transmit information over long distances and has good fault tolerance. In addition, it can achieve better image matting results, particularly in cases wherein user labeling is insufficient.