Degradation information-guided Mamba for underwater image enhancement

Abstract

Underwater imaging process is degraded by absorption and scattering, hindering the development of marine science. Underwater Image Enhancement (UIE) techniques have been developed to address this issue. Most of the current best performing UIE methods adopt solutions combining deep learning and physical models. However, mainstream deep learning networks fail to strike a good balance between globality and computational cost. In addition, existing physics-informed deep learning typically relies on a single source of physical information with degradation from a single aspect. The issue of integrating and fusing multiple physical information to UIE is not well addressed. Based on the above analysis, a degradation information-guided Mamba for underwater image enhancement is proposed, named UIEMamba. In the framework, an advanced Mamba-based backbone is designed to achieve the balance between globality and computational cost. Under the Mamba framework, the Two-Stream Swap subnet (TSS-subnet) is designed to extract and fuse information from transmission maps and color difference maps, indicating degradation levels at multiple sources. To fully utilize and fuse this degradation information, a Local to Global Multi-level Fusion subnet (LGMF-subnet) is proposed with traditional local convolution and a newly designed global Degradation Information Guided Mamba (DIG-Mamba). Our method is tested on four real-world underwater datasets, and the experimental results demonstrate that UIEMamba outperforms existing state-of-the-art methods in both quantitative metrics and visual quality.