【作者】 黄万军；

【导师】 尹宝才；

【作者基本信息】 北京工业大学 ， 计算机应用， 2001， 硕士

【摘要】 脸部特征提取在人脸识别、人脸表情分析与合成、人脸变形和三维虚拟人脸合成等方面具有非常重要的应用，因而成为国内外的一个研究热点。但由于脸部特征的复杂性，以及人脸图像中环境和光照的变化使得它成为一个比较困难的研究问题，特别是当人脸图像具有较大旋转角度时更增加了这一问题的复杂性。 本文在总结前人各种特征提取方法的基础上，实现了一个由粗到细、渐进的人脸特征提取系统，能够提取人脸图像中的眼睛和嘴巴精确特征。本系统主要包括两个新提出的算法，即人脸姿态估计策略和三维可变形模板匹配技术。 一般的人脸姿态估计方法要通过提取脸部特征点之间的关系或根据已知的三维人脸模型来估测人脸的方向。本文提出了一种基于头发模板的人脸姿态估计策略，能够直接从人脸图像中估算出人脸的大致方向，并且经过进一步矫正，得到更加精确的结果。 为了能够提取多方向人脸图像中的器官精确位置，本文提出了一个参数化的三维可变形模板匹配技术，并以眼睛和嘴巴为例子定义了三维眼睛可变形模板和三维嘴巴可变形模板。这些三维可变形模板与人脸图像进行匹配时，还将分别映射成二维映射模板，并分别定义了眼睛映射模板和嘴巴映射模板的能量函数，最后通过贪心算法求取能量函数的最小值，从而得到眼睛和嘴巴的精确特征。更多还原

【Abstract】 Human facial features extraction has become one of the important issues in lots of significant applications, such as face recognition, facial expression analysis and synthesis, face morphing and virtual 3D face synthesis. To extract the exact features of human face is quiet a hard nun to break, mainly for the complicated structure of thce, as well as thc unknown light condition, and it would become more difficult for the face with lager vertical rotation.Here a gradual feature extraction method has been proposed to extract the exact features of eyes and mouth, in which there are two new techniques. namely pose estimation based on hair templates and 3D deformable template matching.A new pose estimation strategy based on seven hair templates has been presented to estimate the orientation of face. The face orientation can be obtained directly from the face image using the new pose estimation, which is different from the traditional method according to the relations between feature points or the reconstructed 3D face model.A new 3D deforrnable template matching technique is also presented here to extract the exact features of eyes and mouth. When the 3D deformable template is applied to match the face image, it should be mapped into 2D template and the energy functions of 2D eye template and 2D mouth template should also be defined respectively. At last the Greedy Algorithm should be employed to minimize the value of energy functions, and the facial features could then be obtained.更多还原