单幅人脸的三维可视化方法研究

发布时间：2018-03-22 08:37

本文选题：三维可视化　切入点：统计模型　出处：《电子科技大学》2013年硕士论文　论文类型：学位论文

【摘要】：目前，三维人脸可视化技术广泛应用于三维游戏、三维动漫以及安防等领域中。因此，该技术在众多研究领域中一直受到专家学者们的重视，尤其是计算机图形学、模式识别等领域。其中基于单幅人脸照片的三维可视化方法，能够在有效避免三维可视化需要耗费大量人力物力的问题的同时，实现较好的三维可视化效果。因此该项技术的研究具有重大意义。单幅人脸的三维可视化方法众多，目的在于建立与照片中人脸一致的三维人脸模型。其中基于统计学习的方法，主要是通过PCA建立可以形变的统计模型获得人脸的三维形状，并采用正交投影的方法来进行纹理映射。本文在此基础上进行研究，，完成了以下几个方面的工作：一、三维人脸样本库统一过程中三角网格的统一统一三维人脸库的人脸结构是整个基于统计学习方法的基础，通常情况下，人脸样本点之间的对应备受关注，而三角网格的对应却很少被提及，本文提出一种三维人脸样本之间三角网格的统一方法，利用原始人脸样本库中的三角网格，通过跟踪其在点对应时的各种变换，进而得出点对应后的三角网格连接。经过处理，整个三维人脸样本间具有统一的连接方式，方便后续环节进一步的处理；二、形状模型构建过程中能量函数的求解方法。在构建模型的测试阶段中，平均模型与人脸照片之间的匹配问题可转化为最小化两者之间距离的能量函数。为避免能量函数的解出现过拟合现象，本文利用l1范数正则化项对能量函数进行约束，使解出的人脸模型在绝大多数情况下匹配照片中的人脸。但是相比利用先验知识作为约束的求解方法，本文方法仍然存在许多不足之处。三、提出两种特征点深度估计方法，并利用估计特征点提高模型的准确性。从人脸模型的侧面观察角度出发，本文提出了两种特征点深度估计的方法：近邻加权法和稀疏表示法，并利用估计的三维特征点进行建模，提高建模精度，使得构建的人脸模型更符合真实照片中人脸的形态；四、基于分割模板的纹理映射方法。在纹理映射中，基于约束点控制的纹理映射往往存在背景残留等现象，本文通过对测试人脸照片进行分割，对所得的分割结果进行腐蚀处理得到一个安全区域的掩膜，利用该掩膜对映射有误的地方进行修正，改善纹理映射的效果。
[Abstract]:At present, 3D face visualization technology is widely used in the fields of 3D game, 3D animation and security. Therefore, the technology has been paid attention to by experts and scholars in many research fields, especially computer graphics. In the field of pattern recognition, 3D visualization method based on single face image can effectively avoid the problem that 3D visualization needs a lot of manpower and material resources. Therefore, the research of this technology is of great significance. There are many 3D visualization methods for single face, the purpose of which is to build a 3D face model consistent with the face in the photograph, in which, based on the statistical learning method, the statistical model can be constructed by PCA to obtain the 3D shape of the face. The method of orthogonal projection is used to carry out texture mapping. The first is the unification of triangular mesh in the process of 3D face sample database unification. The face structure of the unified 3D face database is the basis of the whole statistical learning method. Usually, the correspondence between face sample points is paid close attention to, but the correspondence of triangular mesh is seldom mentioned. In this paper, a unified method of triangulation mesh between 3D face samples is proposed. By tracking the various transformations of the triangulated mesh in the primitive human face sample database, the triangular mesh connection after the point correspondence can be obtained. After processing, the triangulation mesh after the point correspondence can be obtained by using the triangulation mesh in the primitive human face sample database. The whole 3D face sample has a unified connection mode, which is convenient for further processing. Second, the method of solving the energy function in the process of shape model construction. In the test stage of constructing the model, the matching problem between the average model and the face photograph can be transformed into an energy function that minimizes the distance between the two models. In this paper, the energy function is constrained by the L 1 norm regularization term, so that the solved face model matches the face in the photo in most cases, but compared with the prior knowledge as the constraint method, There are still many shortcomings in this method. Thirdly, two methods of depth estimation of feature points are proposed, and the accuracy of the model is improved by using the estimation of feature points. From the side view of face model, this paper presents two methods of depth estimation of feature points: nearest neighbor weighted method and sparse representation method, and use the estimated 3D feature points to model to improve the modeling accuracy. Make the constructed face model more consistent with the face shape in the real photos; Fourth, texture mapping method based on segmentation template. In texture mapping based on constraint point control texture mapping often exists some phenomena such as background residue. In this paper a mask of a safe area is obtained by segmentation of test face images and corrosion of the resulting segmentation results. The mask is used to correct the mapping errors to improve the effect of texture mapping.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】