基于生成对抗网络的图像修复  

Image Inpainting Based on Generative Adversarial Networks

作　　者：孙全 曾晓勤 SUN Quan;ZENG Xiao-qin(Institute of Intelligence Science and Technology,College of Computer and Information,Hohai University,Nanjing 211100,China)

机构地区：[1]河海大学计算机与信息学院智能科学与技术研究所,南京211100

出　　处：《计算机科学》2018年第12期229-234,261,共7页Computer Science

基　　金：国家重点研发计划项目(2017YFC0405805)资助

摘　　要：针对现有图像修复算法存在受损区域的形状和大小受限以及修复痕迹明显、修复边缘不连续的问题,文中提出一种基于生成对抗网络的图像修复方法。该方法采用生成对抗网络(Generative Adversarial Networks,GAN)这种新的生成模型作为基本架构,结合Wasserstein距离,同时融入条件对抗网络(CGAN)的思想;以破损图像作为附加条件信息,采用对抗损失与内容损失相结合的方式来训练网络模型,以修复破损区域。此方法能够修复大多数破损情况下的图像。在CelebA和LFW两个数据集上的实验结果表明,所提方法能够取得很好的修复效果。Aiming at the problems of the restricted shape and size of the damaged area,the obvious inpainting tracks and the discontinuous inpainting edge in the existing image impainting algorithms,this paper proposed an image inpainting method based on generative adversarial networks.In this method,a new generative model named generative adversarial networks(GAN)is used as the basic framework with combining Wasserstein distance and the idea of conditional genera-tive adversarial networks(CGAN).The network receives the damaged image as additional conditional information and combines adversarial loss with content loss to train the network model for restoring pixels of missing areas.This method can be used to repair most of the damages in images.The experimental results on two datasets of CelebA and LFW suggest the capability of this method to obtain good performance.

关 键 词：生成对抗网络 图像修复 对抗学习 Wasserstein距离 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]