Underwater Image Enhancement via Physical-Feedback Adversarial Transfer Learning

This article proposes a domain adaptive learning framework based on physical model feedback for underwater image enhancement. Underwater image enhancement involves mapping from low-quality underwater images to their dewatered counterparts. Due to the lack of dewatered images as ground truth, most learning-based methods are trained using synthetic datasets. However, they usually ignored the domain gap between synthetic training data and real-world testing data, which seriously reduces the generalization ability of those models when testing on real underwater images. We solve the problem by embedding a domain adaptive mechanism in a learning framework to eliminate the domain gap. However, the basic formulation of a domain adaptive-based learning framework does not generate realistic images in color and details. Motivated by an observation that the estimated results should be consistent with the physical model of underwater imaging, we propose a physics constraint as a feedback controller so that it can guide the estimation of underwater image enhancement. Extensive experiments validate the superiority of the proposed framework.