基于UV位置图的视频三维人脸表情克隆  

Video 3D Facial Expression Cloning Based on UV Position Map

作　　者：崔时雨 申冲 齐畅 刘川 张满囤 CUI Shiyu;SHEN Chong;QI Chang;LIU Chuan;ZHANG Mandun(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China;Hebei Data Driven Industrial Intelligent Engineering Research Center,Tianjin 300401,China;Tianjin International Joint Center for Virtual Reality and Visual Computing,Tianjin 300401,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300401 [2]河北省数据驱动工业智能工程研究中心,天津300401 [3]天津市虚拟现实与可视计算国际联合中心,天津300401

出　　处：《郑州大学学报（理学版）》2023年第1期49-56,共8页Journal of Zhengzhou University:Natural Science Edition

基　　金：河北省自然科学基金项目(F2019202054)。

摘　　要：为了解决三维人脸表情克隆时的准确性与实时性问题,提出了一种基于UV位置图的视频三维人脸表情克隆方法。首先逐帧读取视频进行人脸检测,然后裁剪出人脸区域,输入到位置图回归网络中得到UV位置图,最后根据UV位置图获得人脸关键点三维信息,驱动三维人脸模型产生变形,完成人脸表情克隆。为了提高克隆效果,构建了基于Ghost瓶颈的轻量级网络,并根据人脸肌肉运动规律与表情变化特征设计了一种新的人脸分区方式及权重,结合Wing Loss损失函数,对卷积神经网络模型进行训练。使用所提算法与其他算法在AFLW2000-3D、COFW-68数据集上进行对比实验,结果表明,所提算法提高了人脸关键点检测的准确率,能够有效解决视频三维人脸表情克隆的准确性与实时性问题。In order to solve the accuracy and real-time problem of 3 D facial expression cloning, a 3 D facial expression cloning method based on UV position map was proposed. Firstly, the video was processed frame by frame for face detection, then the face region was clipped and input into the position map regression network to get the UV position map. Finally, the 3 D information of the facial landmarks was obtained according to the UV position map, which drove the 3 D face model to deform and complete the facial expression cloning. In order to improve the cloning effect, a lightweight network based on Ghost bottleneck was constructed, and a new face partition method and weight were designed according to the facial muscle movement law and expression change characteristics. Combined with Wing Loss function, the convolutional neural network model was trained. Compared with other algorithms on AFLW2000-3 D and COFW-68 datasets, it increased the accuracy of facial landmark detection, and the accuracy and real-time problem of video facial expression cloning were improved effectively.

关 键 词：表情克隆 UV位置图 人脸关键点检测 卷积神经网络 三维人脸模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]