【作者】 王进；

【导师】 彭群生； 鲍虎军；

【作者基本信息】 浙江大学 ， 应用数学， 2003， 博士

【摘要】 虚拟人脸建模和表情动画是当前计算机图形学、计算机视觉和图像处理领域中一个热点研究课题，在视频会议、影视制作、通讯等方面有着广泛的应用；静态图像和动态视频中人脸检测技术在生物认证等领域有着特殊的意义。本文提出了一个视频驱动的人脸动画原型系统，详细介绍了构造该系统所涉及的人脸检测、特征提取、特征跟踪、特定人脸建模、表情动画等方面的技术。在对每个技术细节进行认真分析的基础上，本文对各个主要步骤都分别提出了新的见解和改进。 ※ 提出了一个改进的CANDIDE-4参数化人脸模型。它是MPEG-4的子集，为了通应行为驱动人脸动画设计需要，本文对MPEG-4的FAPs表情行为单元、节点、形状单元进行了改造。 ※ 研讨了基于单幅、两幅、多幅图像的三种不同的构造特定人脸的方法。利用Shape from Shading的思想和人脸的约束信息实现了基于单幅正面人脸图像的重建；以CANDIDE-4的参数调整为手段实现基于两幅正交图像的人脸模型重建；通过跟踪视频中的特征点，标定相机外参，进而估计特征点的3D位置，实现了基于一段视频中小特征点集的人脸建模算法。 ※ 提出并实现了基于RBF、Harmonic Model插值的两种基于约束的纹理映射算法，可以实现形式化表示的、满足一阶或二阶连续的纹理坐标插值。 ※ 提出一种结合小波分解和ERI的参数化表情纹理细节迁移算法。所给出的算法可以维持源图像的基本属性，保留源图像的特征，迁移表情图像的纹理细节；可以利用FAPs的函数控制表情夸张程度。 ※ 实现了对表情特征的检测和跟踪。论文采用统计训练的思想，选择包括各种表情变化的人脸图像建立样本库，取所有样本与平均图像的差构造一个矩阵，利用主元分析方法进行降维，然后通过独立元分析降低主元相关性，建立了人脸的特征子空间；算法采取对主元进行扰动优化匹配的方法检测人脸，本文称此方法为全局最优的方法。为了提高算法精度，我们还借鉴LFA的思想，提出了一个基于局部特征的分层学习、匹配模型ML-IDAM，得到了更精确的实验结果。更多还原

【Abstract】 At present, virtual face modeling and expression animation is one of the research hot in computer graphics, image processing and computer vision, which have great application in teleconference, artificial life, wireless presence, and the like; also, face detection of image and video has especial significance in biology certification area. In this thesis, we present one prototype system of video based performance-driven facial expression animation, and describe in detail the relative content of face detection, feature extracting, feature tracking, special face modeling, expression animation and the like. After analyzing every technique in detail, we present some new idea and improvement to every key step: Present an improved parameterized face model of CANDIDE-4, which is a subset of MPEG-4. In order to meet the requirement of performance-driven facial expression animation, we modify the AUs(Act Units), vertexes and SUs(Shape Units) of FAPs in MPEG-4. Discuss three different methods, based on single image, two images and multi-images, to create special face model. Utilizing the idea of SFS (Shape from Shading) and the facial constrained information, we reconstruct face model by single frontal face image. We utilize orthogonal image method to generate individualized face model by adjusting the parameters of the CANDIDE-4. We realize an algorithm based on minimum features for rapid face modeling from video, by tracking feature points, calibrating exterior parameter, estimating 3D location of feature points. Present two constrain-based texture mapping methods using RBF and Harmonic Modelinterpolation respectively, which are expressed in explicit formulation and satisfy C1 or C2. Combining wavelet decomposition and ERI (Expression Ratio Image), we propose a new parameterized algorithm for transforming facial expressional details. The algorithm can hold basic illumination and key characters of source image, transform texture details of target expression image; in addition, we can control the degree of expression exaggeration by the function of FAPs. Realize the expression feature detection and tracking. In this thesis, we adopt the technique of statistical training, create a sample database of every kinds of expression face images,construct a matrix of the difference of each sample and average image, and reduce dimension by PCA, then decrease the relativity of principle components by ICA, and therefore get the character sub-space of face. When detecting a face, we adopt the method of disturbing principle components of model to match special facial image, which is called whole optimization method in this thesis. In order to improve precision, we refer to the idea of LFA (Local Feature Analysis) and give a model of ML-IDAM (ICA-based Multi-Layer Directly Appearance Model), which can give more precise experiment results.更多还原