UV空间中的人脸颜色纹理和几何细节协同补全  

Collaborative Completion of Facial Texture and Displacement in UV Space

作　　者：程轩 刘仁帅 郑英林 曾鸣[1] Cheng Xuan;Liu Renshuai;Zheng Yinglin;Zeng Ming(School of Informatics,Xiamen University,Xiamen 361005)

机构地区：[1]厦门大学信息学院,厦门361005

出　　处：《计算机辅助设计与图形学学报》2022年第1期113-120,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61802322,61402387,62072382);中央高校基本科研业务费(20720190003);福建省科技计划引导性项目(2018H0037).

摘　　要：当图像中的人脸存在较大角度的偏转时,由于自身遮挡,单幅图像3D人脸重建方法较难获取整张人脸的纹理和几何细节.考虑人脸纹理特征分布和几何细节的特征分布的双向关联特性,提出一种统一框架下的协同补全模型TDGAN.首先,将颜色纹理和几何细节映射到同一UV空间;然后,通过统一的生成对抗网络协同补全纹理与几何,并对这2部分信息分别设计全局与局部判别器,以实现纹理和几何的全局与局部一致性;最后,为了充分利用颜色纹理和几何细节共有特征,增加了一个纹理-几何一致性约束网络,从而得到高完整度和高一致性的颜色纹理与几何细节UV图.在当前最大3D人脸数据集FaceScape的实验表明,TDGAN比独立的UV空间补全方法能得到更高质量的补全结果.When the input face has large pose,it’s hard for single view 3D face reconstruction methods to estimate complete facial texture and geometric details due to self-occlusion.Considering the correlation between the feature distributions of texture and geometric details,TDGAN is proposed,a collaborative completion model with GAN structure which could complete texture and geometric details collaboratively in a uniform framework.Firstly,the texture and geometric details are mapped into the UV space.Then,the collaborative completion is performed in a generative adversarial network which includes a generator,two global discriminators and two local discriminators.Finally,to exploit the common structures of texture and geometric details,a consistency constraint module is incorporated to the framework.The complete and consistent texture and geometric details UV maps are synthesized.Experiments on the currently largest 3D face dataset demonstrate that the proposed collaborative completion method could produce more high quality results than the independent UV completion methods.

关 键 词：人脸补全 人脸3D重建 生成对抗网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]