【作者】 陆涛；

【导师】 谭晓阳；

【作者基本信息】 南京航空航天大学 ， 计算机应用技术， 2012， 硕士

【摘要】 随着摄影、照相等相关硬件技术的不断发展，在电影、游戏、医疗等领域，三维技术得到了广泛的应用。相比于二维图像，三维图像具有更多的空间信息且贴近人们的生活。近年来，在计算机视觉和图形学中，三维人脸建模的研究成为该领域的核心内容，在三维人脸动画与识别中起到了关键作用。三维人脸建模一般有四种方法，参数模型，肌肉模型，视觉模型和形变模型。形变模型因具有真实感好、自动化等优点而被认为是三维人脸建模最成功的方法之一。前人提出了ICP、NICP、流光法等方法来建立三维形变模型，但这些方法在人脸对齐精度及准确性上都存在着局限性。因此，在一个三维人脸库中建立形变模型仍是一个难题。本文对三维人脸形变模型的建立及其应用进行研究，主要研究内容和创新包括：1.对标准形变算法的改进。在大规模的三维人脸数据库上，研究多种建立形变模型的方法并分析其优劣。实验证明，标准的三维表面形变方法在人脸对齐的精度及准确性上效果良好。然后根据本实验中三维人脸数据库的特点，改进局部寻找匹配点对时的策略，不仅使得匹配点对准确性提高，而且使得后期的能量函数优化速度加快。最后，运用非刚性形变算法的思想，在标准三维形变变方法的基础上，提出了迭代稠密对齐算法。2.评估三维人脸形变模型的准确性和有效性。针对不同应用领域，评估标准也不同。本文分别从四个方面和标准三维表面形变算法进行对比。实验证明，本文算法在对齐精度和准确性上都要高于标准形变方法。3.在已建立的形变模型上完成基于单幅图像的三维人脸重建。根据单幅正面人脸图像来重建三维人脸，由于信息的不对等，这是不可能实现的。但是三维形变模型具有三维人脸丰富的先验知识，可以近似重构出图像的三维人脸。由于三维形变模型的参数复杂，本文使用三维形变稀疏模型，通过对特征点的拟合，达到对三维人脸形状的重建。然后，通过二维到三维的纹理映射，完成对三维纹理重建。实验表明，重建的三维人脸形状光滑，纹理逼真。更多还原

【Abstract】 With the development of hardware technology of photograph,3D technology has been widelyused in the field of movie, games, and medical care and other fields. Compared with2D image,3Dimage enjoys more space information and has closely lined to people’s daily life. In recent years,research based on3D face model has become a hot issue in computer vision and graphic, which playsan important role in3D face animation and recognition.3D face model generally have four methods, parameter model, muscle model, the vision modeland the morphable model. Morphable model when compared to the other method has the advantagesof good quality and automation, is supposed to be the most successful method in3D face model.However, it is still difficult to establish Morphable model in a3D face database. Many predecessorshad put forward the methods like ICP, NICP, time method, etc. But they have limitations in facealignment accuracy and the accuracy of established correspondence. This paper focuses on theconstruction and application of3D face Morphable model. The main contents of the present paperinclude:1. The construction of3D face morphable model. We study various methods of constructing3Dface morphable model on the basis of a large-scale3D face database and analyze their advantages anddisadvantages. The experimental result shows that the standard3D surface deformation methods havegood effect in face alignment accuracy and the accuracy of correspondences. And then, according tothe characteristics of the database, we modify the strategy in searching local correspondences, whichnot only improve the accuracy of correspondences, but also speed up the process of energy functionoptimization. As the deformation of face belongs to non-rigid transformation, we present our iterativedense3D face alignment method on the basis of standard3D surface deformation method, which referto the frame of non-rigid algorithms.2. The evaluation of3D face morphable model. Measures of alignment quality have to be domainspecific. We compare our method with the traditional surface deformation method from four aspects.The results show that our method is superior to traditional one in the alignment accuracy and theaccuracy of correspondence.3.3D face reconstruction based on a single face image. According to only one face image3Dface reconstruction is an ill-problem. However, since the3D morhpable model has lots of prioriknowledge,3D face shape can be estimated by this face model. Because of the complexity ofmorphable model parameters, the sparse3D face morphable model is choosen to reconstruct the3D face shape by means of fitting the feature points between the3D face model and the face image. Andthen,3D face shape is projected to image to generate the face texture. According to the experimentalresult, the reconstruction3D face shape is smooth and the texture is realistic.更多还原