基于单张照片的三维人脸重建算法研究

发布时间：2018-03-04 18:38

本文选题：三维人脸重建　切入点：深度信息恢复　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：伴随经济日渐繁荣和科学技术日新月异,人类在物质需求和精神层面都有了更高的追求。在游戏动漫、影视剧作、医疗美容、视频通讯和信息安全等众多方面,不再满足于二维世界带来的视觉感官体验,3D电影、电视,游戏动漫等应运而生。其中,三维人脸建模由于直观的显示效果和广泛的应用场景,成为目前科学研究领域和实际工程项目中的一个关注热点。相对于传统的基于三视图、多视图以及视频流的三维人脸重建,基于单张照片对人脸进行重建需要的用户输入量最少,因此是目前研究的重点。然而由于人脸面部构造极为复杂,并且不同人脸之间的差异巨大,只通过单张人脸照片很难直接获取人脸的深度数据,其实际执行操作更具挑战性。因此,本课题的开展极具研究价值和实用价值。本文围绕基于单张照片的重塑三维人脸模型进行了如下两项工作:一是提出了一种基于特征融合的面部特征点深度恢复算法。首先,对人脸数据库采取预处理工作,获取三维模型对应的二维面部图像;其次通过显式形状回归的方法获取面部特征点位置信息;之后,提取面部几何特征,依据所获得的关键点对人脸进行Delaunay三角剖分,划分人脸特征区域,通过比对待恢复人脸与数据库人脸特征区域的几何特征距离相似度,完成基于几何特征的深度信息恢复;接着,提取局部纹理特征,以关键点划分局部纹理区域,并进行区域局部二值模式算子直方图统计,通过衡量待恢复人脸与数据库人脸的直方图距离相似度,完成基于局部纹理特征的深度恢复;最后,运用最小二乘法进行特征融合,提高特征点的恢复准确度。本文提出的基于特征融合的人脸特征点深度恢复算法,所使用的特征易于提取并且算法复杂度低,实验结果表明恢复的面部特征点深度信息较为精确。二是提出了两种三维人脸模型的纹理映射算法。基于约束细化Delaunay三角剖分的纹理映射方法是通过关键点计算二维图像和三维模型之间的投影关系,对关键点构建的特征区域进行撒点插值,进行进一步约束细化的Delaunay三角剖分,依据不同区域的映射关系,完成三维人脸的建立以及对应纹理的贴附。基于径向基插值的纹理映射方法是通过已知特征点信息对径向基网络进行拟合训练,形成特定人脸对应的径向基网络,然后对非特征点经构建完成的网络实现插值映射,重建特定的人脸模型。相较于传统的方法,本文提出的方法不需要进行大量的人机交互操作,在使用少量数据的基础上重建出了较为真实的人脸模型,适合应用在实际工程中。
[Abstract]:With the increasing prosperity of economy and the rapid development of science and technology, human beings have a higher pursuit in both material and spiritual aspects. In many aspects, such as game animation, film and television plays, medical beauty, video communication and information security, etc. No longer satisfied with the visual sensory experience brought by the two-dimensional world, 3D film, television, game animation and so on. Among them, 3D face modeling, due to the visual display effect and extensive application scene, Compared with the traditional 3D face reconstruction based on three-view, multi-view and video stream, it has become a hot topic in the field of scientific research and practical engineering. Face reconstruction based on single photo requires the least input from users, so it is the focus of current research. However, due to the complexity of facial structure and the huge differences between different faces, It is difficult to get the depth data of a face directly by using only a single face photograph, so it is more challenging to perform the operation in practice. This thesis is of great research value and practical value. This paper focuses on the reconstruction of 3D face model based on single photo as follows: first, a facial feature point depth restoration algorithm based on feature fusion is proposed. The face database is preprocessed to obtain 2D facial image corresponding to 3D model. Secondly, the location information of facial feature points is obtained by explicit shape regression. After that, facial geometric features are extracted. According to the key points obtained, the face is triangulated by Delaunay, and the facial feature regions are divided. The depth information restoration based on geometric features is completed by comparing the similarity of geometric features between face and database facial features. Then, the local texture features are extracted, the local texture regions are divided by the key points, and the histogram statistics of the local binary pattern operators are carried out to measure the histogram distance similarity between the face to be recovered and the face in the database. The depth restoration based on local texture features is completed. Finally, the least square method is used for feature fusion to improve the accuracy of feature point restoration. A facial feature point depth restoration algorithm based on feature fusion is proposed in this paper. The features used are easy to extract and have low algorithm complexity. The experimental results show that the depth information of facial feature points restored is more accurate. Secondly, two texture mapping algorithms for 3D face models are proposed. The texture mapping method based on constrained thinning Delaunay triangulation is calculated through key points. The projection relationship between 2D images and 3D models, The characteristic regions constructed by key points are interpolated by scatter points, and further constrained Delaunay triangulation is carried out. According to the mapping relationship of different regions, The method of texture mapping based on radial basis function interpolation (RBF) is to fit and train radial basis function network (RBF) through the information of known feature points to form radial basis function network (RBF) corresponding to a particular face. Then the network constructed by non-feature points is interpolated to reconstruct a specific face model. Compared with the traditional method, the method proposed in this paper does not require a lot of man-machine interaction operations. A real face model is reconstructed on the basis of a small amount of data, which is suitable for practical engineering.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】