Human Body Region Extraction from Photos
Yi Yong Hu · 2007
This paper presents an approach to automatically extract human body region from color photos, which introduces trimap shape updating into iterated GrabCut image segmentation technique. It is based on an observation that human torso is relatively stable in appearance compared with various human poses formed by hands and feet, and on a fact that estimation on a small region is more accurate than on a large region if a few cues for estimation are just available. At first a human face is found by scanning a face detector across the whole target unknown image. Then a body trimap, an image showing potential body area, is initialized according to the found face. With this trimap, body torso is estimated with GrabCut image segmentation method. After that, the trimap is updated by dynamically growing its contour according to local image information, and new body region is estimated by applying GrabCut to the target image. With the iterated processing of trimap shape updating and GrabCut applying, human body region is finally extracted. The approach has been tested with 400 photo images, and the results show its usefulness. 1.