用于雨线和雨滴去除的像素级核预测网络  

Pixel-level kernel prediction network for rain streaks and raindrops

作　　者：李克文[1] 姚贤哲 蒋衡杰 刘文龙[1] LI Ke-wen;YAO Xian-zhe;JIANG Heng-jie;LIU Wen-long(College of Computer Science and Technology,China University of Petroleum(East China),Qingdao 266580,China)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,山东青岛266580

出　　处：《计算机工程与设计》2025年第1期182-189,共8页Computer Engineering and Design

基　　金：国家自然科学基金重大基金项目(51991365);山东省自然科学基金项目(ZR2021MF082)。

摘　　要：针对现有的图像去雨算法通常需要设计先验知识和多阶段优化框架导致算法适应场景单一的问题,提出一种基于图像过滤统一去除雨线和雨滴的去雨核预测网络(RKPN)。有雨图像通过RKPN估计像素级去雨内核。采用空洞卷积提取多尺度特征,提出一个多分支特征聚合模块。通过数据增强构建一个雨线和雨滴混合数据集(RDRS),提升现实世界复杂场景中图像去雨算法的效果。对4个公开数据集和RDRS数据集的广泛实验结果表明,所提模型取得均高于MPRNet等主流网络的峰值信噪比和结构相似度。To solve the problem that existing deraining algorithms usually need to design prior knowledge and multi-stage optimization framework to adapt to a single scene,an image filtering-based deraining network(RKPN)was proposed to uniformly remove rain streaks and raindrops.The image with rain was estimated by RKPN to gain the pixel-level deraining kernel.Dilated convolution was used to extract multi-scale features,and a multi-branch feature aggregation module was proposed.A rain streaks and raindrops dataset(RDRS)was constructed through data enhancement to improve the effectiveness of image derain-removal algorithms in real-world complex scenes.Extensive experimental results on four existing datasets and RDRS dataset show that the proposed model achieves higher SSIM and PSNR than mainstream networks such as MPRNet.

关 键 词：深度学习 图像去雨 核预测网络 计算机视觉 混合雨模式 图像过滤 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]