Abstract: Abstract: Although image dehazing algorithms have achieved promising results on multiple datasets under normal lighting conditions, their performance in complex nighttime lighting scenarios remains unsatisfactory. Nighttime image dehazing is a challenging task due to the dual impact of significant haze and low light conditions at night. Existing research has largely overlooked the effects of poor lighting in nighttime scenes, resulting in suboptimal dehazing performance. Observing that the inverted image of a low-light image closely resembles a daytime hazy image, we propose a dual-branch nighttime dehazing network that simultaneously processes both the nighttime hazy image and its inverted images to address the distortions of haze and low light. To effectively integrate information from both the nighttime hazy image and its inverted image, we design a simple but effective fusion strategy by Branch Interaction and Channel Shuffling (BICS) blocks, which enables the information exchange between and within the two branches. We further employ cross-attention mechanisms to capture long-range relations between the nighttime hazy image and its corresponding inverted image. Multi-scale input-output network design further enhances feature representation. Extensive experiments on public benchmarks demonstrate that our method outperforms state-of-the-art nighttime dehazing techniques in both image quality and assessment performances.

Key words: Keywords: nighttime dehazing, deep learning, feature fusion, cross attention, image inversion

摘要： 摘 要: 图像去雾算法在正常光照条件下的多个数据集上取得了良好效果，而在夜间复杂光照场景下的表现却差强人意。夜间存在重度雾霾与低光照的双重影响，使得夜间图像去雾成为一项极具挑战性的任务。现有研究在很大程度上忽略了夜间场景光照不佳带来的影响，导致去雾效果未达最优。研究发现，低光照图像的反转图像与白天雾霾图像高度相似，据此提出一种双分支夜间去雾网络，同时处理夜间雾霾图像及其反转图像，以解决雾霾和低光照造成的图像失真问题。为有效融合夜间雾霾图像与其反转图像的信息，本文设计了一种基于分支交互与通道混洗（BICS）模块的简洁高效融合策略，实现两个分支之间及分支内部的信息交互。此外，引入交叉注意力机制捕捉夜间雾霾图像与其反转图像之间的长距离关联，并采用多尺度输入输出的网络设计进一步增强特征表征能力。在公共基准数据集上的大量实验表明，该方法在图像质量和客观评价指标上均优于当前已有的夜间图像去雾技术。

关键词: 关键词: 夜间图像去雾, 深度学习, 特征融合, 交叉注意力, 图像反转