摘要
有雨图像往往丢失大量的特征及细节信息,严重影响视觉效果和目标检测.针对有雨图像,本文采用图像显著性检测来寻找有雨区域,确定待修复区域,设计了多尺度融合生成对抗网络(Multiscale Fusion Generative Adversarial Network,MsF-GAN)进行图像去雨,在生成器的12/14/16网络层采用Concat进行多尺度融合,使生成的去雨图像中更多地保留图像的原始信息,鉴别器采用全局鉴别保证图像去雨的视觉连贯性,局部鉴别提高有雨区域去除的泛化能力,从而保证了去雨的有效性和可靠性.本文提出的MsF-GAN在公开发布的数据集上进行了实验,结果表明:去雨图像的结构相似度(Structural similarity index,SSIM)和峰值信噪比(Peak Signal to Noise Ratio,PSNR)等相关指标相比其他方法都有了大幅提升.
Raindrop images often lose a lot of feature and detail information,which seriously affects visual effect and target detection.For raindrop images,this paper uses image saliency detection to find rain areas and determine the areas to be repaired,and designs a multi-scale fusion generative adversarial network(MsF-GAN)for image raindrop removal,and uses Concat for multi-scale fusion at the 12/14/16 network layers of the generator,so that the original information of the image is more retained in the generated removal raindrop images.The discriminator adopts global discrimination to ensure the visual consistency of image rain removal,and local discrimination improves the generalization ability of raindrop area removal,thus ensuring the effectiveness and reliability of rain removal.The MsF-GAN proposed in this paper has been tested on publicly released data sets.The results show that the structural similarity index(SSIM)and peak signal to noise ratio(PSNR)and other related indexes of the image removal raindrop have been greatly improved compared with other methods.
作者
徐志京
于帅
XU Zhi-jing;YU Shuai(School of Information Engineering,Shanghai Maritime University,Shanghai 201306,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2021年第5期1050-1055,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61673259)资助.
