A Semi-supervised Nighttime Dehazing Baseline with Spatial-Frequency Aware and Realistic Brightness Constraint

Existing research based on deep learning has extensively explored the problemof daytime image dehazing. However, few studies have considered thecharacteristics of nighttime hazy scenes. There are two distinctions betweennighttime and daytime haze. First, there may be multiple active colored lightsources with lower illumination intensity in nighttime scenes, which may causehaze, glow and noise with localized, coupled and frequency inconsistentcharacteristics. Second, due to the domain discrepancy between simulated andreal-world data, unrealistic brightness may occur when applying a dehazingmodel trained on simulated data to real-world data. To address the above twoissues, we propose a semi-supervised model for real-world nighttime dehazing.First, the spatial attention and frequency spectrum filtering are implementedas a spatial-frequency domain information interaction module to handle thefirst issue. Second, a pseudo-label-based retraining strategy and a localwindow-based brightness loss for semi-supervised training process is designedto suppress haze and glow while achieving realistic brightness. Experiments onpublic benchmarks validate the effectiveness of the proposed method and itssuperiority over state-of-the-art methods. The source code and SupplementaryMaterials are placed in the https://github.com/Xiaofeng-life/SFSNiD.