MsDA: Multi-scale Domain Adaptation Dehazing Network

Single image dehazing is a typical ill-posed problem among the computer vision tasks. Recent years have also witnessed the excellent progress of deep learning-based methods in single image dehazing. However, most existing methods are stuck in the pattern of training models on synthetic datasets due to the lack of real image pairs and hence, cannot generalize well to real hazy photos. To this issue, we propose a multi-scale domain adaptation framework to bridge the domain gap. More specifically, we utilize an end-to-end dehazing model and align the features extracted from its encoder of both the synthetic and real images, to initially mitigate the domain shift. Besides, we close the gap between feature maps in skip connections of two domains to further improve its generalization ability. Moreover, we incorporate the modified dark channel prior and color attenuation prior, conducting the real domain dehazing in an unsupervised manner. Through extensive experiments and ablation study, we show that the proposed architecture is effective and necessary to perform favorably against state-of-the-art.