一种可解释的双分支深度学习去雾算法

An interpretable dual-branch deep learning dehazing algorithm

关键词:  图像去雾  特征提取  残差连接  注意力机制  细节增强

Key words:  image dehazing  feature extraction  residual connections  attention mechanism  detail enhancement

摘要:

为解决目前大多数深度学习去雾算法去雾过程中细节丢失、可解释性差的问题,提出了一种可解释的双分支深度学习去雾算法。该算法设计双分支协同架构解耦去雾任务,上分支雾霾提取子任务通过设计的雾霾提取块（haze removal block,HRB）在频域中提取雾霾特征,并引入通道注意力机制增强浓雾区域的特征捕获；下分支细节修复子任务采用聚合式残差框架修正提取特征时损失的纹理细节；通过计算模糊图像与雾霾特征图像的负残差得到初步去雾图像,并利用下分支修正细节获得最终的去雾图像。在SOTS、NH-HAZE和真实场景数据集上进行相关实验,结果表明,相较于现有的主流去雾算法,本文算法恢复的图像不仅雾霾去除更加彻底、细节保留更加完整,而且客观评价指标也有显著提升。研究成果不仅为深度学习在图像去雾领域开辟了新的研究方向,还为实际应用中的图像清晰化处理提供了切实可行的解决方案。

Abstract:

In order to address the issues of detail loss and poor interpretability in current deep learning-based dehazing algorithms, this paper proposes an interpretable dual-branch deep learning dehazing algorithm. The algorithm employs a dual-branch collaborative architecture to decouple the dehazing task: the upper branch focuses on haze extraction through a designed haze removal block (HRB) that captures haze features in the frequency domain, while incorporating a channel attention mechanism to enhance feature extraction in dense haze regions. The lower branch adopts an aggregated residual framework for detail restoration to correct texture details lost during feature extraction. By computing the negative residual between the hazy image and the haze feature image, a preliminary dehazed image is obtained, which is then refined by the lower branch to produce the final dehazed result. Experiments on the SOTS, NH-HAZE, and real-world datasets demonstrate that compared to existing mainstream dehazing algorithms, the proposed method achieves more thorough haze removal, more complete detail preservation, and significant improvements in objective evaluation metrics. This work not only establishes new research directions for deep learning in image dehazing field but also provides a practical solution for real-world image clarity enhancement.