融合面部深度感知的音频驱动人脸重现方法  

Audio Driven Face Reenactment Method Integrating Facial Deep-perception

作　　者：彭雪康 孙国庆 邵长乐 练智超 PENG Xuekang;SUN Guoqing;SHAO Changle;LIAN Zhichao(School of Cyber and Security,Chinese University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学网络空间安全学院,南京210094

出　　处：《指挥与控制学报》2024年第3期365-371,共7页Journal of Command and Control

基　　金：国家重点研发计划(2021YFF0602104-2)资助。

摘　　要：人脸重现是一项条件面部生成任务,现有的基于音频驱动的人脸重现方法难以生成完整且高质量的人脸。针对这一问题,提出一种融合面部深度信息的音频驱动下的人脸重现方法。该方法采用了轻量级的模型框架以降低模型尺寸和提高运行速度。实验在AnnVI数据集上与3种最新的音频驱动人脸重现方法进行了比较。结果表明,所提出的融合面部深度感知的人脸重现方法,极大地提高了音频驱动下生成人脸图像的质量。Face reenactment is a conditional facial generation mission.The current face reenactment method based on audio driven is difficult to generate a complete and high quality face.As for such a problem,the audio driven face reenactment method integrating facial deep information is proposed.Meanwhile,the method adopts a light-weight model framework to reduce the model size and to improve the running speed.The experiment is compared with three kinds of the latest audio driving face reenactment methods on the AnnVI dataset.The results show that the proposed face reenactment method,integrating facial deep perception can greatly improve the quality of generated facial images driven by the audio.

关 键 词：人脸伪造 人脸重现 深度估计 多模态驱动 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]