Underwater image enhancement via deep modeling with frequency decomposition

Underwater images often suffer from spatially varying illumination, color cast, low contrast, and veiling effects caused by wavelength-dependent attenuation and scattering. To address these degradations, this paper proposes FDM-Net, a frequency-decomposition modeling network for underwater image enhancement. During the downsampling stage, FDM-Net explicitly decouples low-frequency global information and high-frequency local details, and incorporates physics-based underwater imaging cues into feature learning via a frequency-expert-guided bidirectional modulation strategy, thereby enhancing the robustness and interpretability of the model. During the upsampling stage, a multi-scale fusion scheme is adopted to reconstruct fine structures while preserving cross-scale consistency. On the T-90 and LSUI-TE datasets, FDM-Net achieves the best PSNR/SSIM scores of 23.58/0.915 and 28.18/0.925, respectively, and demonstrates superior visual enhancement performance. These results indicate that the proposed method has strong robustness and promising potential for real-world underwater vision applications.