基于多分辨率深层网络的图像去雨方法研究  

Research on Image Rain Removal Method Based on Multi-resolution Deep Network

作　　者：廖建敏 李丽红 义翔 LIAO Jianmin;LI Lihong;YI Xiang(Guangxi Machinery Industry Research Institute Co.,Ltd.,Nanning 530007,China;Department of Computer Applications,Guilin University of Technology at Nanning,Nanning 530001,China;Guangxi Guilu Expressway Co.,Ltd.,Nanning 530022,China)

机构地区：[1]广西机械工业研究院有限责任公司,广西南宁530007 [2]桂林理工大学南宁分校,计算机与应用系,广西南宁530001 [3]广西桂鹿高速公路有限公司,广西南宁5300022

出　　处：《微型电脑应用》2023年第8期201-204,224,共5页Microcomputer Applications

摘　　要：雨天条件下视频监测图像中含有雨滴,使得输变配电设备监测目标的细节信息模糊,降低了目标的辨识度,影响了视频监测的性能。针对雨滴导致的视频监测图像退化问题,提出了一种基于深层神经网络的多分辨率模型。首先,利用先验信息,采用小波滤波提取表征图像高频分量的雨滴模糊特征图,将模型聚焦于雨滴信息,从而有效抑制图像中的背景干扰。其次,借鉴HRNet网络多分辨率提取多阶特征结构,构建多分辨率深层神经网络融合底层和高层特征,有效丰富网络的信息表征能力。最后,采用强化学习方法,调节模型多分辨率分支数,优化模型去雨性能。在合成及真实世界的雨天图像集上,对所提出的方法进行了实验验证,实验结果表明本文方法较其他方法具有更好的去雨效果。Under rainy weather,the video monitoring image contains raindrops,it blurs the detailed information of the monitoring target,reduces the recognition of the target,and affects the performance of video monitoring.Aiming at the problem of video monitoring image degradation caused by raindrops,this paper proposes a multi-resolution model architecture based on deep neural networks.Using the prior information,the wavelet filtering is used to extract the fuzzy feature maps of raindrops that characterizes the high-frequency components of the image,which renders the model to focus on the raindrop information and effectively suppress the background interference in the image.Learning from the multi-branch extraction structure of the multi-resolution feature of HRNet,a multi-resolution deep neural network is built to fuse the bottom and high-level features,and effectively enrich the information representation capabilities of the network.Reinforcement learning method is used to adjust the number of multi-resolution branches of the model to optimize the network performance.On the synthetic and real-world rainy day image sets,the method proposed in this paper has been experimentally verified.The experimental results show that the method has better rain removal effect than other methods.

关 键 词：小波滤波 深层神经网络 强化学习方法 图像处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]