| 摘要: |
| 深度学习的发展，促进了视频换脸技术的提高，产生了深度视频换脸技术，也叫做深度伪造技术。然而，由于人脸特征、背景环境等条件的复杂性，当前产生的换脸视频逼真度有待提高。本文提出了一种基于光照感知处理的深度视频换脸方法，将光照感知模块加入深度视频换脸过程中，促进生成人脸与目标人脸保持光照一致性，使其更好地与背景融合。在FaceForensics++数据集上的实验表明，相较当前主流的深度视频换脸方法，加入光照感知处理能够有效地生成更加真实的人脸，保持目标视频中的光照条件分布，量化实验表明基于光照感知处理的视频换脸方法有效地提高了换脸视频的质量。 |
| 关键词: 深度伪造 光照感知 视频篡改 换脸视频 |
| DOI： |
| 投稿时间：2020-12-31修订日期：2021-02-26 |
| 基金项目:国家重点基础研究发展计划（973计划） |
|
| A novel method to improve fidelity of face-swapping video based on lighting perception process |
|
Liu Xuesong, He Xiaolei, Liu Changjun
|
| (Institute of Information Engineering, Chinese Academy of Sciences) |
| Abstract: |
| The development of deep learning technology boosts face-swapping video technology and brings deep video tampering technology, namely, deepfake. However, due to the complexity of human face feature and background environment, the fidelity of tampered video needs to be improved. In this paper, we propose a method of deepfake based on lighting per-ception process. By using lighting perception module in the process of deepfake, it is effective to keep the consistency of lighting in the fake video. The experiments on the faceforensics++ database show that adding lighting perception process in deepfake can generate more real faces than not. Quantity evaluation shows that the novel face-swapping method ef-fectively improves the quality of forged face. |
| Key words: deepfake lighting perception video tampering face-swapping video |
