摘要
Underwater images often suffer from severe quality degradation caused by light absorption, scattering, and color distortion, which significantly hinder their applications in marine exploration, autonomous underwater vehicles, and underwater robotics. To address these challenges, we propose a physics-guided deep learning framework for underwater image enhancement. In the first stage, a Jaffe–McGlamery physical model produces coarse restoration and extracts priors such as transmission maps and background illumination, which, guided by classification (color cast, low light, blur), inform subsequent color restoration, illumination normalization, and deblurring. In the second stage, a deep neural network with multi-scale convolutional feature extraction and Convolutional Block Attention Module (CBAM) is designed to capture residual degradations across different scales. A decoder is then used for high-quality reconstruction. To optimize both fidelity and perceptual quality, we adopt a hybrid loss function consisting of pixel-wise mean squared error, VGG perceptual loss, and a physics-consistency term. Experimental results on both synthetic and real underwater datasets demonstrate that the proposed method achieves superior performance compared to traditional image processing and existing learning-based approaches, yielding higher scores in PSNR, SSIM, UCIQE, and UIQM. Furthermore, the proposed framework is flexible, providing a scalable solution for diverse underwater environments.