利用生成对抗网络的人脸图像分步补全法  

Face Image Inpainting with Generative Adversarial Network

作　　者：林椹尠 张梦凯 吴成茂 郑兴宁 LIN Zhen-xian;ZHANG Meng-kai;WU Cheng-mao;ZHENG Xing-ning(School of Science,Xi’an University of Posts&Telecommunications,Xi’an 710121,China;School of Communication and Information Engineering,Xi’an University of Posts&Telecommunications,Xi’an 710121,China;School of Electronic Engineering,Xi’an University of Posts&Telecommunications,Xi’an 710121,China)

机构地区：[1]西安邮电大学理学院,西安710121 [2]西安邮电大学通信与信息工程学院,西安710121 [3]西安邮电大学电子工程学院,西安710121

出　　处：《计算机科学》2021年第9期174-180,共7页Computer Science

基　　金：国家自然科学基金(61671377)。

摘　　要：人脸图像修复技术是近年来图像处理领域的研究热点,而人脸图像大面积缺失导致损失语义信息过多,一直是该领域的重点难点问题。针对这一问题,文中提出了一种基于生成对抗网络的图像分步补全算法。将人脸图像修复问题分为两步,设计两个串联的生成对抗网络,首先残缺图像通过预补全网络进行图像的预补全,预补全图像进入增强网络进行特征增强;判别器分别判断预补全图像和增强图像与理想图像的差异性;采用长短时记忆单元连接两部分的信息流,增强信息的传递。然后使用内容损失、对抗损失和全变分损失相结合的损失函数,提高网络的修复效果。最后在CelebA数据集上进行实验,结果显示,所提算法相较于对比算法在峰值信噪比指标上提高了16.84%~22.85%,在结构相似性指标上提高了10%~12.82%。Face image inpainting is a hot topic of image processing research in recent years.Due to the loss of excessive sematic information,it is a difficult problem to inpaint large area missing of face images.Aiming at the problem of inpainting face images,a step-by-step image inpainting algorithm based on generative adversarial network is proposed.Face images inpainting task is divided into two steps.Firstly,face images are completed through the pre-completion network,and pre-completion images is enhanced feature through the enhancement network.The discriminator judges the difference between the pre-completion images,the enhanced images and the ideal image respectively.The long-term memory unit is used to connect the information flow of two parts.Secondly,the adversarial loss,content loss and total variation loss are combined to improve the effectively.Experiments are conducted on Celeb A dataset,and this algorithm has an improvement of 16.84%~22.85%in PSNR and 10%~12.82%in SSIM compared with others typical image inpainting algorithms.

关 键 词：生成对抗网络 人脸图像 图像补全 长短时记忆 深度学习 缺失区域 跳跃连接 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]