Automatically extracting hairstyles from 2D images

Abstract

Automatic hair extraction from a given 2D image has been a challenging problem for a long time, especially when complex backgrounds and a wide variety of hairstyles are involved. This paper has made its contribution in the following three aspects. First, it proposes a novel framework that successfully combines the techniques of face detection, outlier-aware initial stroke placement and matting to extract the desired hairstyle from an input image. Second, it introduces an alpha space to facilitate the choice of matting parameters. Third, it defines a new comparison metric that is well suited for the alpha matte comparison. Our results show that, compared with the manually drawn trimaps for hair extraction, the proposed automatic algorithm can achieve about 86.2% extraction accuracy. © 2013 Springer-Verlag.