知识迁移引导的空频双域联合去雾网络  

Knowledge Transfer Guided Space Frequency Dual Domain Joint Defogging Network

作　　者：杨燕[1] 梁皓博 林雷 YANG Yan;LIANG Haobo;LING Lei(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Yinchuan Power Supply Company,State Grid Ningxia Electric Power Co.,Ltd.,Yinchuan 750000,China)

机构地区：[1]兰州交通大学电子与信息工程学院,甘肃兰州730070 [2]国网宁夏电力有限公司银川供电公司,宁夏银川750000

出　　处：《湖南大学学报(自然科学版)》2025年第4期16-26,共11页Journal of Hunan University:Natural Sciences

基　　金：甘肃省高等学校产业支撑计划项目(2021CYZC-04);兰州交通大学研究生教改项目(JG201928)。

摘　　要：目前一些基于CNN的方法在去雾方面有着不错的性能,但网络鲁棒性欠佳.这主要归因于雾霾分布复杂和数据集难以收集,导致去雾过程中纹理细节丢失严重并且在小规模数据集上存在严重的过拟合问题.为了解决上述问题,提出了空频联合的双分支结构.上分支捕获更多的纹理细节,利用三级小波变换在频域中获取特征;下分支提升网络泛化能力,采用域迁移方法在空域中提供额外的知识先验,以Res2Net作为该分支的核心部分.最后,本文在NH-HAZE数据集上对模型进行训练,在I-HAZE和NTIRE 2023数据集上进行泛化能力测试.此外,为了保证实验的公平性,本文对所有对比算法也采用NH-HAZE数据集进行训练.实验结果表明,本文网络在细节纹理恢复和泛化能力方面获得了显著提升.The current CNN-based methods exhibit satisfactory performance in fog removal,but their network robustness is compromised due to the intricate haze distribution and challenging dataset collection.Consequently,there is a significant loss of texture details during the fog removal process and severe overfitting issues on small-scale data sets.To address these challenges,we propose a two-branch structure incorporating space-frequency joint techniques.The upper branch focuses on capturing finer texture details by utilizing three-level wavelet transform to extract features in the frequency domain.Meanwhile,the lower branch enhances network generalization by employing domain migration method to incorporate additional prior information from airspace and leveraging Res2Net as its core component.Finally,the proposed model is trained on the NH-HAZE dataset and evaluated for generalization ability using the I-HAZE and NTIRE 2023 datasets.Furthermore,to ensure fairness in comparison experiments,all competing algorithms are also trained using the NH-HAZE dataset.Experimental results demonstrate that our proposed network significantly improves both detail texture recovery capability and generalization performance.

关 键 词：图像去雾 域迁移 小波变换 注意力机制 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]