基于GANs无监督回归三维参数化人脸模型  

GANs-based unsupervised regression for 3D morphable model

作　　者：张星星 李金龙[2] Zhang Xingxing;Li Jinlong(School of Software Engineering,University of Science and Technology of China,Hefei 230026,China;School of Computer Science and Technology,University of Science and Technology of China,Hefei 230026,China)

机构地区：[1]中国科学技术大学软件学院,安徽合肥230026 [2]中国科学技术大学计算机科学与技术学院,安徽合肥230026

出　　处：《信息技术与网络安全》2020年第11期50-55,共6页Information Technology and Network Security

摘　　要：针对神经网络回归训练过程中三维人脸数据稀缺的问题,提出了基于生成对抗网络(GANs)回归三维参数化人脸模型(3DMM)的无监督学习方式。首先利用GANs的对抗生成训练使生成器回归的3DMM参数符合真实感人脸形状的参数分布。随后将生成的三维人脸网格渲染成二维图片,利用身份编码器对输入人脸及渲染人脸分别提取身份特征向量,通过不断缩小向量之间的距离使得生成的三维人脸网格靠近输入人脸的身份特征。实验结果表明,该方法在重建结果顶点位置准确性上相对于现有的方法有明显的提升,且拥有较好的RMSE值,能够较好应用于三维人脸重建任务。Aming at the problem of scarcity of 3D face ground-truth data in the training process of neural network regressing for 3D facial mesh,an unsupervised learning method based on Generative Adversarial Networks(GANs),which regresses for 3D Morphable Model(3DMM)parameters,is proposed.Firstly,the adversarial training process of GANs is used to make the generated 3DMM parameters conform to the realistic 3DMM parameters distribution.Then,the generated 3D facial mesh is rendered into a 2D image.Later,the identity encoder is used to extract the identity feature embeddings from the input face and the rendered face respectively.By continuously minimizing the distance between the two embeddings,the generated 3D facial mesh is enforced to maintain the identical features of the input face.The experimental results show that the proposed method has a significant improvement in the accuracy of vertex position compared with the existing methods,and has a good RMSE value,which can be well applied to 3D face reconstruction tasks.

关 键 词：三维人脸重建 生成对抗网络 无监督 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]