基于规范化用户输入空间的自然图像抠图

Natural Image Matting Based on Normalized User Input Space

为了降低自然图像抠图中抠图算法对用户输入的敏感度,提出了规范化用户输入空间(Normalized User Input Space,NUIS)的概念及一种基于NUIS空间的抠图算法(NUIS-Matting)。方法首先将原始图像过分割为超像素(superpixels)并引入超像素前景不透明度以提高算法的抗噪能力,再用超像素构造NUIS空间并将原始用户输入映射到NUIS空间。然后使用一种更有效的采样方法在NUIS空间中采样前景及背景颜色对来计算未知区域像素点的前景不透明度及其置信度,并选取对应高置信度的不透明度作为初始结果;最后使用随机游走(random walk)解一个图标记问题(graph labeling problem)得出优化后的结果。实验结果表明,方法大大降低了抠图对用户输入的敏感度,提高了抠图结果的质量。

A concept was proposed： Normalized User Input Space （NUIS）, and a matting algorithm based on NUIS （NUIS-Matting） was proposed for reducing the sensitivity of matting algorithm to user input. The origin image was oversegmented into superpixels and an alpha term for them was introduced to improve the anti-noise ability of the algorithm. Then, NUIS was constructed with superpixels and the user input was mapped to NUIS. Foreground and background colors were sampled in NUIS with a more effective method to compute the alpha values and confidences for unknown pixels. Alpha values which had the high confidences were picked to form the initial alpha matte. Finally, a graph labeling problem was solved as a Random Walk and the optimized matte was got. The experimental results show that the approach has greatly reduced the sensitivity of matting problem to user input and has improved the quality of matting result.