文献类型：期刊文献

中文题名：基于RDSRCNN的单幅图像去雨模型研究

英文题名：Research on De-rain Model for Single Image Based on RDSRCNN

作者：傅继彬[1];李春辉[1]

第一作者：傅继彬

机构：[1]河南财经政法大学计算机与信息工程学院,河南郑州450046

第一机构：河南财经政法大学计算机与信息工程学院

年份：2023

卷号：22

期号：10

起止页码：198-204

中文期刊名：软件导刊

外文期刊名：Software Guide

语种：中文

中文关键词：图像复原;单幅图像去雨;深度学习;混合式损失函数

外文关键词：image inpainting;single image de-rain;deep learning;mixed loss function

摘要：下雨天气在图像上造成雨痕,不仅严重影响图像观感,更会干扰后续图像分析与处理。图像去雨始终是图像复原研究的热点,于是提出RDSRCNN单幅图像去雨模型。为提升特征提取能力,以增强型特征提取方法ESIFEM为特征提取手段,利用其远距离像素关联能力及低局部特征提取代价实现高效特征提取,同时利用l1与MSS-SIM构造的复合损失函数优化训练效率并保证输出图像对视觉友好,将以上方法与增强型DSRCNN去雨网络相结合形成单幅图像去雨模型。在Rain100H数据集上的实验结果表明,该方法在视觉上能将浓密雨分布情况下的图像恢复至细节丰富的干净场景图,并且相较于大部分对比方法,虚影与物体边缘变形的情况减少90%以上,背景的雨痕清除率高于95%。该方法在量化评估中,峰值信噪比和结构相似性参数在较大部分较对比方法更优,且在空间复杂度等方面优于Restormer方法。
Rainy weather will cause rain marks on the image,which will not only seriously affect the look and feel of the image,but also inter?fere with subsequent image analysis and processing.Image rain removal has always been the focus of image restoration research.In view of this,a single image rain removal model of RDSRCNN is proposed.In order to improve the feature extraction capability,ESIFEM,an en?hanced feature extraction method,was used as the feature extraction means to achieve efficient feature extraction by utilizing its remote pixel correlation capability and low local feature extraction cost.Meanwhile,the loss function constructed with the combination of the l1 and MSS-SIM was used to optimize the training efficiency and ensure the visual friendliness of the output image.The above method is combined with the enhanced DSRCNN rain removal network to form a single image rain removal model.The experimental results on the Rain100H dataset show that this method can visually restore the image with dense rain distribution to a clean scene map with rich details,and compared with most comparison methods,the cases of ghost and object edge deformation are reduced by more than 90%,and the clearance rate of background rain marks is higher than 95%.In quantitative evaluation,the proposed method is superior to the comparison method in most aspects of peak SNR and structural similarity parameters and is superior to Restormer method in space complexity.