Multi-scale atmospheric adaptive fusion network for water surface image dehazing
Wang Hongru, Hu Cheng, Jingtao Zhang · Applied Optics · 2026
Image dehazing is a critical task in computer vision, widely applied in environmental monitoring, autonomous driving, and surveillance. However, non-uniform haze distribution with complex water-surface reflections and refractions often suppresses image details and invalidates conventional atmospheric priors. To address these challenges, this paper proposes a novel, to our knowledge, U-Net-based architecture, the multi-scale atmospheric adaptive fusion U-Net (MSAAFU-Net). The framework incorporates three tailored modules: 1) an atmospheric scattering attention module (ASAM) that embeds learnable physical priors to model spatially varying haze and aquatic optical properties; 2) a multi-scale frequency enhancement module (MSFEM) designed to restore both low- and high-frequency structures; and 3) an adaptive bidirectional fusion module (ABFM) to strengthen feature interaction between the encoder and decoder. Extensive experiments on RESIDE-OTS, NH-Haze, and a self-made water-surface dataset demonstrate that MSAAFU-Net consistently outperforms recent state-of-the-art methods, including FFA-Net, AOD-Net, DehazeFormer, and several diffusion models. On average, it achieves a +1.24dB improvement in PSNR, with noticeable SSIM improvements on the self-made and NH-Haze datasets, while performing competitively on RESIDE-OTS. In terms of perceptual quality, our method reduces NIQE by 0.081 and LPIPS by 0.011 over the strongest baseline. These results confirm that MSAAFU-Net provides obvious advantages in both objective fidelity and subjective visual quality, underscoring its effectiveness for challenging water-surface dehazing tasks.