基于自编码器模型的人脸替换算法与视频人脸替换系统  

A Face Replacement Algorithm Based on Autoencoder Model and a Video Face Replacement System

作　　者：刘译键 宁宁 金鑫[1] LIU Yijian;NING Ning;JIN Xin(Beijing Electronic Science and Technology Institute,Beijing 100070,P.R.China)

机构地区：[1]北京电子科技学院,北京市100070

出　　处：《北京电子科技学院学报》2022年第1期50-61,共12页Journal of Beijing Electronic Science And Technology Institute

基　　金：北京高校“高精尖”学科建设项目(项目编号:20210051Z0401)

摘　　要：目前的人脸替换算法生成的图像存在一些问题,比如分辨率低、存在伪影以及棋盘效应等。为了解决这些问题并实现精确的人脸替换,本文提出了基于自编码器模型的人脸替换算法,在原始的Deepfakes算法的基础上对网络结构进行改进,加入了更多的特征提取层和不同的堆叠顺序并在鉴别器中增加对抗损失和感知损失,提升了生成图像的质量,也使人物与背景融合得更自然;针对视频中人脸多样性和背景复杂性的特点,在人脸替换算法的基础上,提出并设计了视频人脸替换系统,通过视频人脸替换系统预处理(视频转图片、人脸检测和对齐)、基于编码器的人脸替换模型的网络和模型训练、人脸转换、后处理(图像融合和光照融合)、视频合成五个步骤,最终实现精准的视频人脸替换。At present,face replacement algorithm generated images are facing some problems,such as low resolution,artifact and chessboard effect,etc.To solve these problems and achieve precise face replacement,in this paper a face replacement algorithm based on autoencoder model is proposed.In the proposed algorithm,the network structure is improved on the basis of the original Deepfakes algorithm.More feature extraction layers and diverse stacking sequences are added,and an antagonism loss and a perception loss are also added in the discriminator,to improve the quality of the generated image and to blend the character and the background more naturally.For the characteristics of face diversity and background complexity in videos,a video face replacement system is proposed and designed on the basis of the face replacement algorithm.With the five steps in the video face replacement system including the preprocessing(such as video to picture,face detection and alignment),the encoder-based face replacement model network and model training,the face conversion,the post-processing(such as image fusion and illumination fusion),and the video synthesis,an accurate video face replacement is ultimately realized.

关 键 词：自编码器模型 人脸替换 人脸检测 人脸对齐 视频人脸替换系统 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]