期刊文献+

基于渐进多尺度注意力残差网络的单幅图像去雨方法

Single Image Rain Removal Method Based on Progressive Multi-scale Attention Residual Network
下载PDF
导出
摘要 雨水会严重影响场景的能见度,降低成像质量,影响许多计算机视觉系统,如视频监控、自动驾驶等的正常工作。因此从退化的含雨图像中去除雨水是一项迫切的任务。论文提出了一种新的基于渐进式多尺度注意力残差网络模型(PMARnet)用于单幅图像去雨。首先考虑到复杂雨天场景一般包含多个不同特性的雨层,该网络将去雨过程分解为多个阶段,每个阶段使用残差网络预测不同的雨层,避免梯度消失。进一步采用了多尺度注意力残差模块(MAR),以更好地利用多尺度信息提取各层雨带的语义和空间细节特征,有效地表征每个雨层。在Rain100H和Rain100L两个公开数据集中与十一种先进的模型和方法进行了实验对比,我们的模型得到了最好的结果。其中,在Rain100H中,峰值信噪比(PSNR)达到28.06,结构相似度(SSIM)为0.89,较第二好的方法分别提升2.41%和1.14%;在Rain100L中,PSNR达到37.25,SSIM为0.98,较第二好的方法分别提升3.16%和1.03%,证明了该方法的有效性。论文所提出的PMARnet可以有效地在雨条纹层和干净背景图像层之间传播信息。PMARnet网络很好地利用了雨条纹层和背景层,取得了良好的去雨效果。 Rains can exert severely impacts on the visibility of scenes,reducing the quality of imaging and affecting a large number of computer vision tasks and systems,such as video surveillance and self-driving car and the like.Eliminating rain streaks,therefore,is a crucial task.This paper proposes a novel deraining model,coined as progressive multi-scale attention residual net-work(PMARnet),to remove rain streaks from a single frame image.Considering that complex scenes usually consist of multiple rain layers,PMARnet contains several stages.Each of them possess a residual network to alleviate gradient vanishing.In further,a multi-scale fusion attention residual model(MAR)is proposed to better characterize the semantic feature and local spatial feature in detail for each rain layer.Two publicly available benchmark datasets,Rain100H and Rain100L,are used for experimental valida-tion.Compared with eleven existing advanced methods,PMARnet performs the best with an PSNR of 28.06 and an SSIM of 0.89 for Rain100H and accordingly 37.25 and 0.98 for Rain100L.Compared with the second best method,there is an improvement of 2.41%and 1.14%for Rain100H and that of 3.16%and 1.03%for Rain100L.In this study,the proposed PMARnet can effectively propa-gate information between the rain streaks layer and the clean background image layer.PMARnet makes good use of rain streaks layer and background layer,and can achieve good rain removal effect.
作者 顾小豪 王欢 GU Xiaohao;WANG Huan(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094)
出处 《计算机与数字工程》 2024年第3期827-833,879,共8页 Computer & Digital Engineering
基金 国家自然科学基金项目(编号:61703209)资助。
关键词 单幅图像去雨 深度学习 渐进式图像去雨 多尺度融合 注意力网络 single image rain removal deep learning progressive image deraining multi-scale fusion attention network
  • 相关文献

参考文献2

二级参考文献2

共引文献14

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部