基于样本的人像自动生成.pdfVIP

ACCV2002: The 5th Asian Conference on Computer Vision, 23–25 January 2002, Melbourne, Australia 1 Example-based Automatic Portraiture 1 1 1 1 2 Hong Chen , Lin Liang , Ying-Qing Xu , Heung-Yeung Shum and Nan-Ning Zheng 1 Microsoft Research, China 2 Xi’an Jiaotong University, China Abstract that can generate portraits with varying styles. We adopt an inhomogeneous Markov Random Field model, which can In this paper, we present an example-based approach for model not only the likelihood of a portrait given the origi- automatically generating a life-like portrait from a frontal nal image, but also the prior statistical characteristics of the face image. Based on an inhomogeneous Markov Random portrait. Based on this statistical model, we propose two Field Model, an inhomogeneous non-parametric sampling sampling strategies: iterative sampling which is simple and scheme is used to capture the complex statistical character- efficient, and simulated annealing which is more robust. istics of face image and corresponding portrait. In our ap- The rest of this paper is organized as follows. We present proach, only those pixels corresponding to a portrait point our example-based learning framework in Section 2. The are sampled. Such a strategy is crucial for maintaining fa- statistical model for portraiture is described in Section 3. cial structure and guaranteeing coherence of portrait lines. T