基于多阶段双残差网络的图像去雨模型  

Image Deraining Model Based on Multi-stage Dual Residual Network

作　　者：杨浩 周冬明[1] 赵倩 李淼 YANG Hao;ZHOU Dongming;ZHAO Qian;LI Miao(School of Information Science&Engineering,Yunnan University,Kunming 650504,China)

机构地区：[1]云南大学信息学院,云南昆明650504

出　　处：《无线电工程》2022年第11期1933-1943,共11页Radio Engineering

基　　金：国家自然科学基金(62066047,61365001,61463052)。

摘　　要：雨天采集的图像往往因为雨线的影响而降低了图像的质量,从而影响后续计算机视觉任务的执行。为了改善雨天图像的成像质量,提出了一种基于多阶段双残差网络的去雨算法来恢复被雨线遮挡的图像。算法的第1个创新点是多阶段的整体架构,其逐步学习从有雨图像到无雨图像的映射过程,从而将整个去雨过程分解为更容易的子过程。算法的第2个创新点是提出的双残差网络,与传统残差网络相比,提出的双残差网络拥有更多恒等映射路径,减少了信息丢失。算法的第3个创新点是特征信息聚合策略,提出的算法不仅注重从前面阶段到后面阶段之间纵向信息交换,而且在各个阶段之间也存在横向信息交换,以简化信息流并避免信息丢失。实验结果和分析表明,提出的算法对图像中的雨条纹去除效果明显,在Rain100H数据集上达到了33.42 dB的PSNR和0.94的SSIM。相比于最近的图像去雨算法,在主观以及客观评价上都是最优的。Images captured on rainy days are often degraded by rain streaks,which affects the performance of subsequent computer vision tasks.In order to improve the imaging quality of rainy images,a rain removal algorithm based on a multi-stage dual residual network is proposed to recover the images obscured by rain streaks.The first innovation of the algorithm is the multi-stage structure,which gradually learns the mapping process from images with rain to images without rain,thus decomposing the whole deraining process into easier sub-processes.The second innovation of the algorithm is the proposed dual residual network,which has more constant mapping paths and reduces information loss compared with the traditional residual network.The third innovation of the algorithm is the feature information aggregation strategy.The proposed algorithm not only focuses on the vertical information exchange from the previous stage to the later stage,but also has horizontal information exchange between the stages to simplify the information flow and avoid information loss.Experimental results and analysis show that the proposed algorithm is effective in removing rain streaks from images,achieving a PSNR of 33.42 dB and SSIM of 0.94 on the Rain100H dataset.Compared to recent image deraining algorithms,it is optimal in subjective as well as objective evaluation.

关 键 词：图像去雨 残差网络 多阶段网络 图像恢复 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]