Night-time image dehazing using deep hierarchical network trained on day-time hazy images
Arpad Kis, Codruta Orniana Ancuti · 2022
Haze influences the visibility in the images of the outdoor scenes and reduces the efficiency of different imaging algorithms. The problem becomes more complex for night-time hazy scenes. In this paper we introduce a novel strategy that employ a deep learning model trained with day-time image dehazing dataset that can achieve competitive results also for night time hazy scenes. We employ Deep Multi-Patch Hierarchical Network a recent deep learning approach that is trained on a realistic day-time image dehazing dataset. In our strategy, to generalize this approach also for night time hazy scene, we apply as a pre-processing step an effective color correction strategy [1]. The extensive qualitative evaluation demonstrate that our approach is competitive and in general yields better results for night time hazy scenes compared with several state-of-the-art dehazing techniques.