Underwater sequential images enhancement via diffusion and physics priors fusion

Abstract

Although learning-based Underwater Image Enhancement (UIE) methods have demonstrated its remarkable performance, several issues remain to be addressed. A critical research gap is that different water effects are not properly removed, including color bias, low contrast, and blur. This is mainly due to the synthetic-real domain gap of the training data. They are either (1) real underwater images but with synthetic pseudo-labels or (2) synthetic underwater images although with accurate labels. However, it is extremely challenging to collect real-world data with true labels, where the water should be removed to obtain true references. Besides, the inter-frame consistency is not preserved because the previous works are designed for single-image enhancement. To address these two issues, a novel UIE framework fusing both diffusion and physics priors is present in this work. The extensive prior knowledge embedded in the pre-trained video diffusion model is leveraged for the first time to achieve zero-shot generalization from synthetic to real-world UIE task, including both single-frame quality and inter-frame consistency. In addition, a synthetic data augmentation strategy based on the physical imaging model is proposed to further alleviate the synthetic-real inconsistency. Qualitative and quantitative experiments on various real-world underwater scenes demonstrate the significance of our approach, producing results superior to existing works in terms of both visual fidelity and quantitative metrics.