东北大学学报(自然科学版) ›› 2025, Vol. 46 ›› Issue (1): 26-34.DOI: 10.12068/j.issn.1005-3026.2025.20239041

• 信息与控制 • 上一篇    下一篇

基于全局残差注意力和门控特征融合的CNN-Transformer去雾算法

李海燕1, 乔仁超1, 李海江2, 陈泉3   

  1. 1.云南大学 信息学院,云南 昆明 650050
    2.云南省交通投资建设集团有限公司,云南 昆明 650000
    3.阿尔多瓦大学 人文学院,法国 阿拉斯 62000
  • 收稿日期:2023-08-07 出版日期:2025-01-15 发布日期:2025-03-25
  • 作者简介:李海燕(1976—),女,云南红河人,云南大学教授,博士生导师.
  • 基金资助:
    国家自然科学基金资助项目(62266049);云南省万人计划“云岭教学名师”项目(20190101055);2023年度云南省档案局科技项目(2023016607)

CNN-Transformer Dehazing Algorithm Based on Global Residual Attention and Gated Feature Fusion

Hai-yan LI1, Ren-chao QIAO1, Hai-jiang LI2, Quan CHEN3   

  1. 1.School of Information Science and Engineering,Yunnan University,Kunming 650050,China
    2.Yunnan Communications Investment and Construction Group Co. ,Ltd. ,Kunming 650000,China
    3.Faculty of Humanities,University of Artois,Arras 62000,France. Corresponding author: LI Hai-jiang,E-mail: li_cannie@163. com
  • Received:2023-08-07 Online:2025-01-15 Published:2025-03-25

摘要:

为解决现有图像去雾算法因缺乏全局上下文信息、处理分布不均匀的雾时效果差且复用细节信息时引入噪声的缺陷,提出了基于全局残差注意力和门控特征融合的CNN-Transformer去雾算法.首先,引入全局残差注意力机制编码模块自适应地提取非均匀雾区的细节特征,设计跨维度通道空间注意力优化信息权重.然后,提出全局建模Transformer模块加深编码器的特征提取过程,设计带有并行卷积的Swin Transformer捕捉特征之间的依赖关系.最后,设计门控特征融合解码模块复用图像重建所需的纹理信息,滤除不相关的雾噪声,提高去雾性能.在4个公开数据集上进行定性和定量实验,实验结果表明:所提算法能够有效地处理非均匀雾区域,重建纹理细腻且语义丰富的高保真无雾图像,其峰值信噪比和结构相似性指数都优于经典对比算法.

关键词: 图像去雾, 全局残差注意力机制, CNN-Transformer架构, 门控特征融合, 图像重建

Abstract:

To address the shortcomings of existing image dehazing algorithms, such as the lack of global contextual information, inadequate performance in dealing with non‑uniform haze, and the introduction of noise during the reuse of detailed information, a CNN-Transformer dehazing algorithm based on global residual attention and gated feature fusion is proposed. Firstly, a global residual attention mechanism is introduced to adaptively extract the detailed features from non‑uniform haze regions, and cross‑dimensional channel‑spatial attention is designed to optimize information weights. Thereafter, a global modelling Transformer module is proposed to deepen the feature extraction process of the encoder, and a Swin Transformer with parallel convolutions is constructed to capture the inter‑feature dependencies. Finally, a gated feature fusion decoder module is designed to reuse the texture information required for image reconstruction, to filter out irrelevant haze noise, and thereby to improve dehazing performance. Qualitative and quantitative experiments conducted on four publicly available datasets indicate that the proposed algorithm can effectively handle non‑uniform haze regions, reconstruct high‑fidelity haze‑free images with fine textures and rich semantics, and achieve higher peak signal‑to‑noise ratio and structural similarity index compared to the classic algorithm.

Key words: image dehazing, global residual attention mechanism, CNN-Transformer architecture, gated feature fusion, image reconstruction

中图分类号: