【作者】 胡永利；

【导师】 高文； 尹宝才；

【作者基本信息】 北京工业大学 ， 计算机应用技术， 2004， 博士

【摘要】 真实感三维人脸建模和人脸动画是计算机图形学、计算机视觉领域最根本、最困难的问题之一,也是近年来国际上最活跃的研究热点。真实感人脸建模的研究在电影、广告、人物动画、计算机游戏、视频会议、人机交互、医学、公安安全等许多领域有广阔的应用前景。形变模型(Morphable Model)是近几年出现的三维人脸建模新方法。虽然该模型同以往的参数模型、肌肉模型以及视觉模型相比,具有真实感好、自动化等特点,但目前该模型在模型建立、优化以及光照等方面还存在诸多问题,在模型应用方面也缺乏探索。针对上述问题,本文在基于形变模型的人脸建模方法及应用方面展开了研究,主要的研究内容包括: (1) 面向中国人特征的大规模三维人脸数据库的建立。三维人脸数据库是建立形变模型的基础。数据库的建立包括数据获取、数据预处理、数据规格化以及数据库结构设计等研究内容。关键是要解决高维三维人脸数据基于特征的像素级稠密对应问题。(2) 基于形变模型的人脸建模方法及改进。首先,基于三维人脸数据库,研究三维人脸形变模型的建立方法,主要包括参数模型的建立和针对输入人脸图像的模型匹配。模型建立涉及模型基础数据的量化处理和人脸参数表示模型的建立,模型匹配涉及光照模型、摄像机标定以及大规模优化问题。其次,对现有的形变模型进行了改进,主要针对非理想光照环境下的模型匹配问题和匹配优化加速问题展开研究。(3) 形变模型的应用研究。主要研究模型在人脸动画、人脸识别和人脸特征提取方面的应用。在人脸动画方面,研究人脸动画的参数表示模型和人脸动画的驱动方法。在人脸识别方面,针对人脸识别中的光照和姿态问题,研究基于三维人脸模型的人脸识别方法。在人脸特征提取方面,研究形变模型在二维人脸图像特征表示方面的应用。更多还原

【Abstract】 Realistic 3D face modeling and animation is the research focus in computer graphics and computer vision, and has many challenging open problems. It can be widely used in many areas, such as film, advertisement, character animation, computer game, teleconference, human computer interface, medicine, and safeguard etc. Recently, a new face modeling method, the Morphable Model is proposed. Comparing with the classic face modeling methods, such as the parametric model, the physical model and the deformable model, the Morphable Model has advantages in realism and automation, while there are many aspects of the model to be improved and explored, for examples, the model construction, the model matching, the lighting model and its application. Our work concentrates on the construction and application of the Morphable Model. The main study includes the following contents: (1) The construction of the Chinese 3D face database with large quantity. The 3D face database is the key step to the Morphable Model that demands many prototypic 3D faces as the base of the face space. The construction of the 3D face database includes prototypes acquiring, face data preprocessing and standardizing, and structure designing of the database. The difficult problem, the pixel-to-pixel alignments among the dense facial mesh of the prototypes must be analyzed and resolved. (2) The research on the Morphable Model and its improvement. The Morphable Model is implemented firstly. The implementation consists of the model construction and the model matching to the given facial image. The model construction includes the vectorization of the prototypes and the construction of the parametric facial models. The illumination, camera calibration and optimization with massive data are investigated in the model matching. On the other hand, the Morphable Model is considered to be improved in the matching with abnormal illumination and finding the faster matching algorithm. (3) The research on the application of the Morphable Model. We consider three applications of the Morphable Model: face animation, face recognition and facial feature extraction. In face animation, we want to find simple method to produce realistic facial animation with the Morphable Model. It is a good idea using the 3D face modeling method to solve the pose and light problems in face recognition. Applying the Morphable Model to facial feature extraction is inspired by its basic conception, the linear combination. As result of the research on the above areas, we have constructed a large scale 3D face database and the Morphable Model is constructed from the prototypic faces. By the Morphable Model we can reconstruct the 3D face from the given 2D facial image. Some difficult problems about the face modeling method are also resolved. The main contributions and innovations of our research are listed below: (1) Proposing a 3D face alignment method base on mesh resampling. Firstly, the proposed meshing resampling method gives a resolution to the difficult problem of the pixel-to-pixel alignments among the prototypic faces. It is important to the construction of the Morphable Model. Secondly, from the coarse to fine resampling strategy, a multi-resolution Morphable Model is constructed and is used in accelerating the model matching. Lastly, the mesh resampling method can be used in extracting the animation vectors between neutral and expressive faces. (2) Improving the original Morphable Model by the proposed multi-resolution and multi-light model. Considering the time-consuming model matching and the disadvantage in matching case with abnormal illumination, the Morphable Model is improved in two aspects. The multi-resolution face model is proposed to accelerate the model matching with more stability. And the Morphable Model is driven to fit the matching problem with abnormal illumination by the designed multi-light models. (3) Proposing a general face animation method based on the Morphable Model.更多还原