Over the past few years, learning-based methods have exhibited the convincing performance in removing single image haze. Nevertheless, the current dehazing approaches are largely designed in terms of land scenes and exhibit low performance during their applications for maritime images. As impacted by the difficulty in the collection of paired hazy and clean image in maritime scenes, this paper proposes Maritime Image Dehazing-GAN (named MID-GAN), the novel end-to-end adversarial model generating real haze-free image by merely exploiting unpaired supervision. To be specific, an attentive-recurrent feature extraction module (ARFEM) is developed, employing a memory for extracting the haze components containing two constrained cycle-consistency branches through the highlight of the hazy and haze-free image domain. Furthermore, the new real-world nature unpaired dataset, termed as REMIDE, is contributed for maritime image dehazing research. As revealed from qualitative and quantitative outcomes, the proposed MID-GAN framework outperforms several outstanding dehazing models on both real-world and synthetic datasets.

在过去的几年中，基于学习的方法在去除单幅图像雾霾方面表现出令人信服的表现。然而，目前的去雾方法主要是根据陆地场景设计的，并且在应用于海洋图像时表现出低性能。受海事场景中成对的模糊和清晰图像采集困难的影响，本文提出了海事图像去雾GAN（命名为MID-GAN），一种新颖的端到端对抗模型，仅通过生成真正的无雾图像利用非配对监督。具体来说，开发了一种注意力循环特征提取模块（ARFEM），该模块采用内存来通过有雾和无雾图像域的高光提取包含两个约束循环一致性分支的雾分量。此外，新的真实世界自然非配对数据集，称为 REMIDE，用于海事图像去雾研究。正如定性和定量结果所揭示的那样，所提出的 MID-GAN 框架在现实世界和合成数据集上都优于几个出色的去雾模型。