Unsupervised Image Dehazing Based on Physical Priors and Contrastive Learning

Abstract

Due to the challenges of acquiring paired foggy and non-foggy images in real-world settings, and the limited applicability of synthesized foggy images to real conditions, this paper proposes an unsupervised image dehazing algorithm that leverages prior features and contrastive learning to mitigate these issues. Specifically, by constructing a multi-scale non-local reconstruction network, the self similarity information of images at different scales is obtained, and deformable convolution is introduced to improve the visual effect of dehazing images. Employing labeled data for supervised training with bright channel priors and contrastive learning losses, in conjunction with unlabeled, unpaired real-world foggy and non-foggy images for unsupervised training within a cyclically consistent adversarial network, can effectively enhance the model's generalization and robustness.