基于双注意力卷积及Transformer融合的非均匀去雾算法
Non-Homogeneous Dehazing Algorithm Based on Fusion of Dual Attention and Transformer
王科平 1张自娇 1杨艺 1费树岷 2韦金阳1
作者信息
- 1. 河南理工大学 电气工程与自动化学院, 焦作454003;河南省智能装备直驱技术与控制国际联合实验室, 焦作454003
- 2. 东南大学 自动化学院, 南京210096
- 折叠
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对比算法.
Abstract
In order to solve the problems of the lack of attention to different haze regions and the unsatisfactory recovery of detailed information within the dense haze areas by existing dehazing algorithms, we propose a non-homogeneous dehazing algorithm combining convolutional neural network and Transformer.Firstly, in order to better focus on the dense haze areas, a parallel dual attention convolutional network is constructed, with different weights assigned to pixels and channels within the image.Secondly, for deep feature extraction, Transformer blocks are introduced to capture the global non-homogeneous hazy region features, enabling the comprehensive capture of long-range dependence between features and mitigating the problem of detail loss associated with conventional convolution approaches with enlarged receptive fields.Finally, a multi-feature fusion reconstruction network is designed to adaptively fuse shallow and deep features to reconstruct clear images.Through extensive experiments conducted on both public datasets and a self-built non-homogeneous haze dataset, the results indicate that the proposed algorithm outperforms other state-of-the-art algorithms in terms of visual quality and objective evaluation metrics.
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Key words
non-homogeneous dehazing/dual attention convolution/Transformer block/multi-feature fusion reconstruction network引用本文复制引用
基金项目
河南省科技攻关计划(232102210040)
出版年
2024
NETL
NSTL
万方数据