基于完全残差的双分支去雨网络  

A Dual-branch De-raining Network Based on Fully Residual

作　　者：宋浩 张鸿[1,2] SONG Hao;ZHANG Hong(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System(Wuhan University of Science and Technology),Wuhan 430065,China)

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]智能信息处理与实时工业系统湖北省重点实验室(武汉科技大学),湖北武汉430065

出　　处：《计算机技术与发展》2023年第11期57-63,共7页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61373109)。

摘　　要：针对使用深度学习的单幅图像去雨算法会导致细节信息丢失的问题,提出了一个双分支去雨网络,包括雨痕提取分支和细节恢复分支,通过补全细节使去雨图像更接近真实图像。雨痕提取分支的目的是完全提取出雨纹,通过构造特征金字塔来多尺度地学习雨纹信息,并在其中引入执行了全部身份映射的完全残差块来加强特征的重用和传播。为充分利用上下文信息,采用可变形卷积在动态扩大感受野的同时避免了网格伪影的产生,最后输入雨图去除雨痕便得到了初步去雨图。细节恢复分支需要产生细节特征图反馈给初步去雨图像来找回丢失的细节,使用轻量级的完全残差块捕捉特征信息,并用跳跃连接来连接完全残差块提供长距离的信息补偿。实验结果表明,该网络在合成数据集Rain100H中比较RESCAN、SPANet和JDNet等主流去雨方法,在PSNR和SSIM指标上分别至少提高了0.09 dB和0.02,在真实数据集和自制数据集中的去雨效果和细节保留程度均优于对比方法。Aiming at the problem that the single image de-raining algorithm using deep learning will lead to the loss of detailed information,a dual-branch de-raining network was proposed,including a rain streaks extraction branch and a detail recovery branch,which makes the de-raining image closer to the real image by completing the details.The purpose of the rain streaks extraction branch is to completely extract the rain streaks.A feature pyramid was constructed to learn the rain streaks information at multi-scale,and introduce a fully residual block that performs all identity mapping in it to enhance feature re-usage and propagation.In order to take full advantage of context information,deformable convolution was used to dynamically expand the receptive field while avoiding the generation of grid artifacts,at the end input rain image take out rain streaks to obtain a preliminary de-raining image.The detail recovery branch needs to generate a detail feature map to feed back to the preliminary de-raining image to retrieve the lost details,so a lightweight fully residual block was used to capture feature information,and a skip connections was used to connect the fully residual blocks to provide long-distance information compensation.The experimental results show that the network improves the PSNR and SSIM indicators at least 0.09 dB and 0.02 respectively compared with the populous de-raining methods such as RESCAN,SPANet and JDNet in the synthetic dataset Rain100H,and the de-raining effect and the degree of detail retention are better than the comparison methods in both the real dataset and the self-made dataset.

关 键 词：卷积神经网络 单幅图像去雨 多尺度学习 完全残差 可变形卷积 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]