北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2024, Vol. 47 ›› Issue (2): 30-37.

• 论文 • 上一篇    下一篇

基于双注意力卷积及Transformer融合的非均匀去雾算法

王科平1,张自娇1,2,杨艺1,费树岷3,韦金阳1   

  1. 1. 河南理工大学
    2. 河南省智能装备直驱技术与控制国际联合实验室
    3. 东南大学
  • 收稿日期:2023-06-25 修回日期:2023-08-01 出版日期:2024-04-28 发布日期:2024-01-24
  • 通讯作者: 张自娇 E-mail:1132410091@qq.com
  • 基金资助:
    国家重点研发计划项目;河南省科技公关项目

Non-Homogeneous Dehazing Algorithm Based on Fusion of Dual Attention and Transformer

  • Received:2023-06-25 Revised:2023-08-01 Online:2024-04-28 Published:2024-01-24
  • Contact: Zi-Jiao ZHANG E-mail:1132410091@qq.com
  • Supported by:
    National Key Research and Development Plan Project;Science and Technology Project of Henan Province

摘要: 针对现有大部分去雾算法中对不同雾霾区域关注不足以及浓雾区域细节信息恢复不理想的问题, 提出了一种结合卷积神经网络和 Transformer 模块的非均匀去雾算法。首先,为了更好地关注浓雾区域,在浅层特征提取阶段构建了并联双注意力卷积网络,分别从像素和通道的角度给图像分配不同的权重;其次,在深层特征提取中,引入了 Transformer 模块进行全局非均匀雾霾区域特征提取,既能有效捕捉特征之间的长距离依赖关系,又避免了普通卷积扩大感受野导致细节信息丢失的问题;最后, 设计了多特征融合重建网络,能够自适应地融合浅层和深层特征,从而重构清晰图像。在公共数据集和自建非均匀雾霾数据集上进行了大量实验,结果表明,所提算法在视觉效果和客观评价指标上均优于其他先进对比算法。

关键词: 非均匀去雾, 双注意力卷积, Transformer模块, 多特征融合重建网络

Abstract: Most of the existing dehazing algorithms lack of attention to different concentrations of hazy images,which will lead to unsatisfactory recovery of the details in dense haze areas. Furthermore, image dehazing is a pixel-level reconstruction process,detail extraction is critical to image restoration. To address this issue, this paper proposes a non-homogeneous dehazing algorithm based on fusion of dual attention convolution and Transformer. Firstly, in shallow feature extraction, in order to improve the attention to areas of the dense haze, a parallel dual attention convolution network is constructed to assign different weights to the images from the perspective of pixels and channels. Secondly, in deep feature extraction, a Transformer block is integrated into the global non-homogeneous hazy region features, which can fully capture the long-range dependence between features and avoid the problem of detail loss in ordinary convolution enlarged receptive fields. Finally, a multi-feature fusion reconstruction network is designed to adaptively fuse shallow and deep features to reconstruct clear images. To verify the effectiveness of the algorithm, experiments are conducted on I-HAZE, O-HAZE, NH-HAZE, self-built non-homogeneous hazy datasets, and SOTS. The experimental results demonstrate that the proposed algorithm outperforms other state-of-the-art comparative algorithms in terms of visual effects and objective evaluation metrics.

Key words: non-homogeneous dehazing, dual attention convolution, Transformer block, multi-feature fusion reconstruction network

中图分类号: