一种面向图像修复的改进生成对抗网络
An Improved Generative Adversarial Network for Image Inpainting
刘庆俞 1胡莹 2陈磊 3肖强 3刘磊3
作者信息
- 1. 淮南师范学院计算机学院(安徽 淮南 232038);菲律宾National University,College of Computing and Information Technologies
- 2. 淮南市职业教育中心
- 3. 淮南师范学院计算机学院(安徽 淮南 232038)
- 折叠
摘要
为解决图像修复后存在的模糊、边界突出等问题,提出了一种基于生成对抗网络的图像修复模型.首先,使用UNet结构改造Context Encoder(CE)模型的生成器部分;其次,在生成器下采样模块中使用多路分支残差模块,以提高模型的特征提取能力;最后,损失函数中引入了全局一致损失和TV Loss,联合对抗损失和局部损失,从而提高图像的修复效果.生成器网络和判别器网络交替、对抗训练,直至得出稳定的、生成效果较好的生成器模型以完成图像修复.修复模型在CelebA数据集上进行了测试,结果表明修复效果较好,SSIM和PSNR两项指标均有明显提升.
Abstract
An image inpainting model based on generative adversarial network is proposed to solve the problems of poor inpainting effect such as blurring and boundary protrusion after image restoration.First,the generator part of the Context Encoder(CE)model is modified using the UNet structure;second,multi-branch residual modules are used in the generator downsampling modules to improve the feature ex-traction capability of the model;finally,global consistency loss and TV loss are introduced in the loss function to combine the adversarial loss and local loss,thus improving the image inpainting.The generator network and discriminator network are trained alternately and adversarially until a stable generator model with better generation effect is derived to complete the image inpainting.The inpainting model was tested on the CelebA dataset,and the results showed a better inpainting effect with significant improvement in both SSIM and PSNR metrics.
关键词
图像修复/残差网络/生成对抗网络/UNetKey words
image inpainting/residual network/generative adversarial network/UNet引用本文复制引用
基金项目
2023年度安徽省高等学校科研计划项目(2023AH051546)
中国高校产学研创新基金项目(2021ITA07029)
教育部产学研协同育人项目(202102089017)
认知智能全国重点实验室开放课题(COGOS-2023HE02)
出版年
2024
NETL
NSTL
维普
万方数据