UMIENet: Underwater image enhancement based on multi-degradation knowledge integration

Due to the absorption and scattering effects of water on light, underwater images generally suffer from multiple degradations such as blur, color cast, and non-uniform illumination, which severely affect image quality and visual processing tasks. Therefore, underwater image enhancement (UIE) has gained widespread application in various marine exploration tasks. While supervised learning-based methods currently dominate this field, existing methods have two main problems. The limited availability of real paired images and the incompleteness of the degradation types of synthetic datasets restricts the model training performance, and most UIE models are designed for specific types of degradation, lacking systematic processing of multiple underwater degradations. These problems lead to poor model performance. In this work, we construct an Underwater Multi-Degradation Knowledge Integration dataset, called UMDKI, it models multiple degradation factors including blur, color cast, and non-uniform illumination by incorporating a revised image formation model and point light mathematical modeling. Besides, we propose an Underwater Multi-degradation Image Enhancement Network, called UMIENet, it integrates the advantages of various traditional methods and achieves collaborative enhancement of multiple degradations. Extensive experiments demonstrate that the proposed UMIENet achieves excellent performance on multiple benchmarks and shows good effectiveness in real underwater vision tasks.