多阶段渐进处理的图像去雨方法  

The image rain removal method based on multi-stageprogressive processing

作　　者：廉继红[1] 王平 李英 李云红[1] LIAN Jihong;WANG Ping;LI Ying;LI Yunhong(School of Electronics and Information,Xi’an Polytechnic University,Xi’an 710048,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《西北大学学报(自然科学版)》2025年第2期297-308,共12页Journal of Northwest University(Natural Science Edition)

基　　金：陕西省科技计划项目(2022GY-053);陕西省自然科学基础研究重点项目(2022JZ-35)。

摘　　要：针对现有图像去雨方法中存在雨纹去除不彻底、纹理信息丢失等问题,提出一种多阶段渐进式处理的图像去雨算法,可以同时将上下阶段的特征融合,使去雨算法的性能有很大的提高。该去雨网络模型由3个阶段构成。前2个阶段采用改进后的U-Net编码器解码器结构学习多尺度上下文特征信息,特征提取部分采用有效通道注意力机制(efficient channel attention network,ECANet),使网络模型参数变小,更加轻量级;第3阶段加入并行注意力机制(parallel attention subnetwork,PASNet),在学习上下文信息和空间细节特征的同时还能生成高分辨率特征,更好地保留图像的输出细节。此外,还引入监督注意力模块(supervised attention module,SAM)以加强特征学习。实验结果表明,在数据集Rain100H上PSNR达到29.37 dB,SSIM为0.88;在Test1200上PSNR达到32.50 dB,SSIM为0.93,验证了所提方法在图像去雨任务上的有效性。Aiming at the problems of incomplete rain pattern removal and texture information loss in the existing image rain removal methods,this paper proposes a multi-stage progressive image rain removal algorithm,which can simultaneously fuse the features of the upper and lower stages and greatly improve the performance of the rain removal algorithm.The rain removal network model consists of three stages.In the first two stages,the improved U-Net coder-decoder structure is used to learn multi-scale context information,and the efficient channel attention network(ECANet)is used for feature extraction,which can reduce the parameters of the network model.In the third stage of becoming lighter,parallel attention subnet(PASNet)is added,which can generate high-resolution features while learning contextual information and spatial details,and can better preserve the output details of images.At the same time,supervised attention module(SAM)is introduced to strengthen feature learning.The experimental results show that the PSNR is 29.37 dB and SSIM is 0.88 on the data set Rain100H;The PSNR is 32.50 dB and SSIM is 0.93 on Test1200,which verifies the effectiveness of the proposed method in the task of image rain removal.

关 键 词：图像去雨 特征提取 监督注意力 并行注意力机制 空间细节 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]