自动从人像中去除眼镜的方法.pdf

ACCV2002: The 5th Asian Conference on Computer Vision, 23–25 January 2002, Melbourne, Australia 1 Automatic Eyeglasses Removal from Face Images Chenyu Wu1,2 Ce Liu1,2 Heung-Yeung Shum2 Ying-Qing Xu2 Zhengyou Zhang3 (a) (b) (c) (d) (e) Automatic Eyeglasses Removal System Figure 1. Automatic eyeglasses removal system overview. (a) An input face image. (b) Face localization by ASM. (c) Glasses recognition by a classi?er. (d) Glasses localization by MCMC. (e) Glasses removal based on a set of training examples. Abstract and support vector machine (SVM) [9] are some of the typ- In this paper, we present a system that can automatically ical approaches. Deformable models such as [7] and Ac- remove eyeglasses from an input face image. Our system tive Shape Model (ASM) [2] have been demonstrated to consists of three modules: eyeglasses recognition, local- be effective to localize faces. Recently, face hallucination ization and removal. Given a face image, we ?rst use an [8] and facial sketch generation [1] can synthesize high- eyeglasses classi?er, to determine if a pair of eyeglasses resolution face images and good-quality face sketches, re- is present. Then we apply a Morkov chain Monte Carlo spectively. Most of above methods need to extract facial method to accurately locate the glasses by searching for the features such as eye, eyebrows and nose with a non-glasses global optimum of the posteriori. Finally, a novel example- assumption on human faces, despite the fact that many peo- based approach is developed to synthesize an image with ple wear glasses. A person wearing glasses is likely to be eyeglasses removed from the detected and localized face missed by a face detector training on faces without glasses, image. Rather than applying conventiona