摘要
在降雨场景中拍摄的图像会被严重污染,影响视觉效果以及后续图像检测算法的性能,因此,在受降雨污染的图像中恢复清晰图像是非常具有实际意义的.其他去雨网络更多的是关注雨滴条纹的去除,但由于真实雨滴图像中的污染不仅仅包含雨滴条纹,还有水汽在深度方向上堆叠形成的雾化效果,因此,在恢复图像过程中同时去除雾气和雨滴条纹是非常必要的.文章提出一种同时考虑图像上产生的雾气与降雨条纹并将其去除的方法,用于解决降雨场景中的图像恢复问题.该方法提出的卷积神经网络是先在输入的雨雾图上去除雾气,不改变雨滴在图像中的分布,而由于雾气被去除,雨滴在图像中将更加突出和更加容易被检测.接着再在已经去雾的图像上去除雨滴.降雨条纹在图像上纹理特征较为明显,感受野的选择对效果有着较大的影响.文章加入了大小感受野融合的残差结构,在扩大感受野的同时,减少因为大感受野而出现的光晕效应.最后该方法在合成数据集以及真实降雨场景的数据集上进行了测试,测试结果表明该方法在解决此类问题上相较于其他方法有着更好的性能.
The images taken in the rain scene will be seriously polluted,which will affect the visual effect and the performance of the subsequent image detection algorithm.Therefore,it is very practical to restore clear images from the rain-polluted images.Other rain removal networks focus more on the removal of raindrop streaks.However,the pollution in real raindrop images not only includes raindrop streaks,but also the atomization effect caused by the stacking of water vapor in the depth.Therefore,it is very necessary to remove both fog and raindrop streaks in the process of image restoration.This paper proposes a method that considers the fog and rain streaks generated on the image and removes them at the same time to solve the problem of image restoration in the rain scene.The convolutional neural network proposed by this method will first remove the fog on the input rain and fog map,without changing the distribution of raindrops in the image,and because the fog is removed,the raindrops will be more prominent in the image and easier to be detected.Then this approach removes raindrops on the defogged image.Rain streaks have obvious texture characteristics on the image,and the selection of the receptive field has a greater impact on the effect.In this paper,the fusion residual structure of large and small receptive fields is added to expand the receptive field while reducing the halo effect due to the large receptive field.Finally,this method is tested on synthetic datasets and datasets of real rainfall scenarios.The test results prove that our proposed method has better performance than other methods in solving such problems.
作者
余卓权
范俊宇
彭绍湖
刘长红
YU Zhuo-quan;FAN Jun-yu;PENG Shao-hu;LIU Chang-hong(School of Electronics and Communication Engineering,Guangzhou University,Guangzhou 510006,Guangdong;School of Mechanical and Electrical Engineering,Guangzhou University,Guangzhou 510006,Guangdong)
出处
《广州大学学报（自然科学版）》
CAS
2020年第3期76-84,共9页
Journal of Guangzhou University:Natural Science Edition
关键词
去雨
去雾
雾化效果
降雨条纹
图像恢复
残差结构
rain removal
fog removal
atomization effect
rain streaks
image restoration
residual structure
