Shrinking the unknown region for image matting

摘要

Natural image matting is a technique to extract the foreground object from an image. To process a given image, it requires a trimap, which partitions the original image into three regions: definitely foreground, definitely background, and unknown regions. The unknown region in the trimap requires complicated computations by a matting algorithm. In this paper, a novel approach to shrinking the unknown region is proposed as a preprocessing. To shrink the unknown region, in addition to a single global threshold of gradient magnitude, the local threshold for every pixel in the unknown region is presented. Experimental results show that the local threshold can reduce the unknown region more accurately and effectively than the global threshold. The proposed approach can improve not only the quality but the efficiency of matting.