电子测量与仪器学报

郭嘉城,罗晓忠,钟寒.基于双路径重参数化网络的图像去雾方法[J].电子测量与仪器学报,2026,40(6):35-48

基于双路径重参数化网络的图像去雾方法

Dual-path structural reparameterized network for high-quality image dehazing

DOI：

英文关键词:image dehazing structural reparameterization large kernel convolution dense non-uniform haze

基金项目:中国人民公安大学“双一流”创新研究项目（2023SYL07）、高等学校学科创新引智基地项目（B20087）资助

作者	单位
郭嘉城	中国人民公安大学信息网络安全学院北京100038
罗晓忠	北京警察学院科技信息网络处北京102202
钟寒	中国人民公安大学信息网络安全学院北京100038

Author	Institution
Guo Jiacheng	School of Information Network Security, People’s Public Security University of China, Beijing 100038,China
Luo Xiaozhong	Technology Information Network Department, Beijing Police College, Beijing 102202, China
Zhong Han	School of Information Network Security, People’s Public Security University of China, Beijing 100038,China

摘要点击次数: 48

全文下载次数: 15

中文摘要:

针对密集非均匀雾霾场景中雾分布复杂、空间变化剧烈，导致现有深度去雾模型在真实场景下泛化能力不足、细节恢复不稳定且部署开销较高的问题，设计了一种双路径结构重参数化去雾网络（re-parameterized large kernel dehaze，RepLKDehaze）。该网络由大核结构重参数化子网（large-kernel rep subnet，RepLK）与全分辨率注意力子网（full-resolution attention subnet，FRA）构成：前者基于裁剪的UniRepLKNet-s编解码框架，在训练阶段通过扩张重参数化构建多分支表示，并在推理阶段等效折叠为单一大核卷积，实现对全局雾分布的高效建模；后者在原始分辨率下引入残差通道注意力机制，用于补偿下采样带来的局部纹理与边缘信息损失。两路特征经融合后实现清晰图像重建，从而在全局一致性与局部细节恢复之间取得平衡。实验结果表明，尽管RepLKDehaze在RESIDE-SOTS等合成数据集上并未在所有指标上取得最优，但在更贴近真实成像退化的Dense-Haze与NH-Haze（2020/2021）数据集上表现出明显优势：在Dense-Haze上达到17.66 dB/0.613/0.494（PSNR/SSIM/LPIPS），在NH-Haze 2020上达到21.66 dB/0.716/0.254，在NH-Haze 2021上取得22.74 dB/0.857/0.169，在颜色还原、结构保真与残雾抑制方面均优于对比方法。结果表明，所提出的双路径结构重参数化去雾网络能够有效提升真实、密集非均匀雾场景下的去雾鲁棒性，并兼顾模型部署效率，为实际雾天视觉感知任务提供了一种可行解决方案。

英文摘要:

Dense and non-uniform haze in real-world scenes exhibits complex spatial variations, which significantly degrades the generalization ability of existing deep image dehazing models, leading to unstable detail recovery and high deployment costs. To address these challenges, this paper proposes a dual-path structurally re-parameterized dehazing network (re-parameterized large kernel dehaze, RepLKDehaze). The proposed framework consists of a large-kernel structural re-parameterization subnet (large-kernel rep subnet, RepLK) and a full-resolution attention subnet (full-resolution attention subnet, FRA). The former is built upon a trimmed UniRepLKNet-s encoder-decoder backbone, where multi-branch representations are introduced via dilated re-parameterization during training and equivalently merged into a single large-kernel convolution at inference, enabling efficient modeling of global haze distribution. The latter operates at full spatial resolution and incorporates residual channel attention to compensate for local texture and edge information loss caused by downsampling. By fusing features from both paths, RepLKDehaze achieves a balanced reconstruction between global consistency and fine-grained detail preservation. Experimental results demonstrate that although RepLKDehaze does not achieve the best performance across all metrics on synthetic datasets such as RESIDE-SOTS, it consistently outperforms state-of-the-art methods on real-world datasets that better reflect realistic imaging degradation. Specifically, it achieves 17.66 dB/0.613/0.494 (PSNR/SSIM/LPIPS) on Dense-Haze, 21.66 dB/0.716/0.254 on NH-Haze 2020, and 22.74 dB/0.857/0.169 on NH-Haze 2021. The proposed method shows clear advantages in color fidelity, structural preservation, and residual haze suppression. These results indicate that the proposed dual-path structurally re-parameterized dehazing network effectively enhances robustness in real dense non-uniform haze scenarios while maintaining deployment efficiency, providing a practical solution for real-world hazy-scene visual perception tasks.

查看全文查看/发表评论下载PDF阅读器