An Improved Semi-Supervised Learning-Based U-Net for Underwater Image Enhancement

The attenuation and scattering of light in underwater environments often lead to degradation issues such as image blurring, color cast, and low contrast. Due to the lack of high-quality paired datasets, the performance of deep learning-based methods is limited. To address these challenges, we propose an improved U-Net architecture based on semi-supervised learning for underwater image enhancement. First, a multi-scale input-level fusion module is proposed to allow our network to fully exploit global spatial information from input images, natural light fields, and depth maps. Second, a multi-branch hybrid convolutional attention residual block is designed to improve the network's feature extraction capability and focus on important spatial and channel information in underwater images. Finally, a semi-supervised learning approach is used to fully exploit unpaired datasets, improving the generalization ability of the network model. Through qualitative and quantitative analysis, the proposed network achieves satisfactory image enhancement results in various underwater scenarios.