基于法线贴图增强隐式函数的单图像三维头部重建  

Single-Image 3D Head Reconstruction Using Normals Enhanced Implicit Function

作　　者：刘乐元 刘旭[1] 孙见弛 陈靓影 Liu Leyuan;Liu Xu;Sun Jianchi;Chen Jingying(National Engineering Research Center for E-Learning,Central China Normal University,Wuhan 430079;National Engineering Laboratory for Educational Big Data,Central China Normal University,Wuhan 430079)

机构地区：[1]华中师范大学国家数字化学习工程技术研究中心,武汉430079 [2]华中师范大学教育大数据应用技术国家工程研究中心,武汉430079

出　　处：《计算机辅助设计与图形学学报》2024年第12期1906-1919,共14页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金面上项目(62077026);中央高校基本科研业务费优秀青年团队项目(CCNU22QN012).

摘　　要：三维头部重建是建构元宇宙的基础技术之一,在影视制作、游戏娱乐、智能教育等领域也具备广阔的应用前景.从单图像重建三维头部模型可以最大程度地节省成本并大幅提高操作的便捷性.然而,单图像三维头部重建是一个非适定问题,现有算法普遍存在重建的三维头部模型保真度低、细节少以及算法泛化能力差等问题.为此,文中提出了一种基于法线贴图增强隐式函数的单图像三维头部重建算法.首先,设计了一个法线贴图估计子网络从单幅图像估计头部法线贴图;其次,将三维头部模型表面看作由隐式函数描述的水平集,建立了端到端的深度神经网络,从输入图像及法线贴图中提取视觉特征并判别三维空间中各点位于该水平集等值面的概率.文中算法在FCH(FaceScape,CoMA,HeadSpace)和FaceVerse公共数据集上所重建三维头部模型的平均倒角距离分别为0.7696 mm和1.3080 mm,大幅优于对比的单图像三维头部重建算法;在2个从互联网采集的数据集上的实验结果也表明,文中算法能从包含不同年龄、人种、性别、面部表情的单幅肖像图像中重建出具备高保真度和丰富细节的三维头部模型,且具备强大泛化能力.3D head reconstruction is one of the fundamental techniques for building the metaverse,and it al-so has broad applications in the fields of film and cartoon creation,game design,and intelligent education.Compared to modeling 3D heads manually by artists or capturing head scans using 3D scanners or stereo imaging systems,reconstructing 3D heads from a single image is much more economical and practical.However,single-image 3D head reconstruction is an ill-posed problem,and existing methods often suffer from low fidelity,fewer details,and poor generalization ability.To this end,this paper proposes a sin-gle-image 3D head reconstruction method using normals enhanced implicit function.First,a deep network for estimating normal maps of heads from a single image is designed.Second,the surface of a 3D head model is defined as a level set described by an implicit function,and an end-to-end deep neural network is built to extract visual features from the input image and the estimated normal map and predict whether a 3D point lies on the head surface or not.Our method achieves Chamfer distances of 0.7696 mm and 1.3080 mm respectively on the FCH(FaceScape,CoMA,HeadSpace)and FaceVerse datasets,which outperform other single-image 3D head reconstruction methods by a large margin;experimental results on images collected from the Internet also show that our method can reconstruct 3D head models with high fidelity and rich de-tails from single images with different ages,human races,genders,and facial expressions,and has a strong generalization capability.

关 键 词：三维头部重建 单图像 深度神经网络 隐式函数 法线贴图 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]