Abstract: Image dehazing is a prominent topic in the field of computer vision. However, acquiring large-scale, high-quality paired datasets from the real world is costly and challenging. Consequently, most existing methods resort to using synthetic data for fully supervised deep learning model training, which can lead to suboptimal real-world performance. To bridge the domain gap between synthetic and real data, a semi-supervised image dehazing algorithm based on teacher-student learning was introduced. This approach leverages Exponential Moving Average (EMA) strategies to update the teacher model and conducts end-to-end dehazing learning, significantly mitigating domain shift issues between synthetic and real data and enhancing generalization performance in real hazy scenarios. Experimental results demonstrate the method's strong performance on publicly available synthetic and real hazy image datasets, delivering superior dehazing visual results.

Key words: image dehazing, deep learning, semi-supervised learning, teacher-student learning, domain difference

摘要： 图像去雾是计算机视觉领域的热点话题之一，然而，由于真实世界中大规模高质量的配对数据集的获取存在成本昂贵、实施困难等问题。现有的大多数方法往往利用合成数据对深度学习模型进行全监督训练，但此类方法可能会导致模型在真实场景下的泛化性能较差。为了缓解真实域和合成域之间的域差异问题，提出一种基于师生学习的半监督图像去雾算法，该算法采用师生学习半监督框架，利用指数移动平均(EMA)策略更新教师模型，端到端地进行去雾学习。所提半监督去雾模型架构显著缓解了合成数据与真实数据之间的域偏移问题，并提高模型在真实有雾场景下的泛化性能。实验结果表明，所提算法在公开的合成和真实有雾图像数据集上取得了较好性能，并获得更好的去雾视觉效果。

关键词: 图像去雾,, 深度学习, 半监督学习, 师生学习, 域差异