一种基于深度残差网络的图像修复算法设计与实现  

Design and implementation of an image inpainting algorithmbased on deep residual network

作　　者：吴金苗 王育欣 韩江宁 张家亮 张志 魏雨露 Wu Jinmiao;Wang Yuxin;Han Jiangning;Zhang Jialiang;Zhang Zhi;Wei Yulu(College of Computer and Information Engineering,Tianjin Agricultural University,Tianjin 300392,China;Business Management Department,China Unicom Video Technology Co.,Ltd.,Tianjin 300300,China)

机构地区：[1]天津农学院计算机与信息工程学院,天津300392 [2]联通视频科技有限公司业务管理部,天津300300

出　　处：《天津农学院学报》2024年第3期85-91,共7页Journal of Tianjin Agricultural University

基　　金：中国高校产学研创新基金重点项目(2022IT017);教育部产学合作协同育人项目(220900287260151);天津农学院人才资助项目计划(Y0400907);天津农学院研究生教育教学研究与改革重点项目(2021-YA-6);天津农学院教育教学研究与改革项目(2023-A-14);天津农学院教材研究项目(2021-B-01);校企合作技术研发项目(TNHXKJ2022074,TNHXKJ2023009)。

摘　　要：随着数字图像在人们日常生活中使用越来越多,针对破损图像缺失区域的数字图像修复问题,本文提出一种基于深度残差网络的图像修复模型。该模型总体架构基于编码-解码结构,编码器采用不同深度的残差网络,解码器分别使用反卷积网络结构和上采样-卷积结构。通过实验探讨在本模型中不同结构的解码器、编码器以及不同的损失函数对图像修复效果的影响。实验结果表明:本文提出的基于深度残差网络的图像修复模型,采用修改后的Resnet 34-layer作为编码器,反卷积网络作为解码器,L1 Loss作为损失函数,能够达到较好的图像修复效果。With the increasing use of digital images in people's daily life,restoring the missing areas of damaged images has become a problem,for which this paper proposes an image inpainting model based on a deep residual network.The overall architecture of the model is based on the encoder-decoder structure.The encoder uses residual networks of different depths,and the decoder uses a deconvolution network structure and an upsampling-convolution structure.The effects of different structure decoders,encoders and different loss functions on the image inpainting effect in this model were discussed through experiments.The experimental results showed that the image inpainting model based on the deep residual network proposed in this paper could achieve good image inpainting effect by adopting the modified Resnet 34-layer as the encoder,the deconvolution network as the decoder,and the L1 Loss as the loss function.

关 键 词：图像修复 残差网络 反卷积 编码器-解码器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]