Dual-branch Contrastive Learning for Image Enhancement of Underwater Internet of Things

Learning a single underwater image enhancement network from unpaired degraded and clear images is of practical interest. In reality, it is almost infeasible to obtain clear reference images corresponding to captured images in the IoT underwater. As a result, training enhancement networks in a supervised manner is challenging in the absence of such paired data. At the same time, existing methods are often insufficient to learn the semantic and texture knowledge inherent in clear images from limited data due to the significant variability between clear and degraded image domains. In response, we propose a two-branch contrast enhancement framework (DCE-Net) for unpaired underwater image enhancement, which learns mutual information between clear and degraded image domains from a limited amount of unpaired data. The proposed DCE-Net consists of a Cyclic Consistency Module (CCM) and a Contrast Enhancement Module (CEM). Specifically, the CCM is designed to guide feature transformation and latent feature learning between underwater clear images and underwater degraded images. the CEM is designed to constrain the consistency of semantic information between underwater clear images and underwater degraded images, encouraging better enhancement and improving image recovery quality. Extensive experiments are conducted using publicly available underwater datasets. The results demonstrate the effectiveness of the proposed method.