TSID-Net: a two-stage single image dehazing framework with style transfer and contrastive knowledge transfer: TSID-Net: a two-stage single image dehazing framework...

Haze-free images have become a prerequisite for many computer vision tasks; therefore, single image dehazing is particularly important in the field of computer vision. However, existing deep learning dehazing methods face two main problems. First, existing dehazing methods are mostly trained based on paired images, but obtaining paired data of the same scene in the real world is challenging, which limits their dehazing performance. Second, most existing dehazing methods are primarily result-driven, which disregards the intermediate process of dehazing, and the rich prior knowledge present in clear and hazy images is not fully utilized, resulting in significant deviations between the dehazed results and the ground truth. Therefore, we propose a novel two-stage single image dehazing network, TSID-Net, to address the above two issues. In the first stage, we consider hazy images as a form of hazy artistic style, while clear images serve as the content information of the artwork. By combining style transfer, we generate high-quality and diverse paired images. This approach significantly mitigates the challenge of acquiring paired data and provides an ample training sample for the second stage. In the second stage, we utilize abundant clear and hazy images to train positive and negative teacher networks with strong robust prior learning capabilities. By combining knowledge transfer, contrastive learning and process-oriented mechanism, we achieve effective knowledge transfer and contrastive knowledge transfer of the intermediate features in the student network. Additionally, we propose a style version bank and incorporate curricular contrastive regularization to achieve dual contrastive learning of both the process and results for student network. Extensive experimental results demonstrate that TSID-Net effectively removes haze and produces visually pleasing results. Code is available at: https://github.com/wsl666/TSID-Net.git.