Dual Contrastive Learning for Unpaired Image Dehazing

Abstract

In real hazy conditions, the existing unpaired image dehazing by using cycle-consistency loss constraints and adversarial loss constraints effect is not enough for the network to be more effective for the acquisition of potential relationship between hazy images and clear images. In this paper, we propose an effective framework for unpaired image dehazing. Our approach improves the Cycle-Consistent Adversarial Networks (Cyclegan) by incorporating an improved Deep Residual Shrinkage Network in the generator, an attentional mechanism for the deep features to filter out the efficient subset as well as soft thresholding to remove the noise to further mine the potential feature distributions between the domains, introducing the contrast learning in self-supervised learning, double layer contrastive learning constrains deep feature relationships for better image dehazing and restoration. Extensive experiments show that we demonstrate superiority over existing unpaired dehazing methods on Nyu-depth and Reside datasets, and produce dehazing results comparable to several fully supervised dehazing methods.