A Dual-Stage Frequency-Driven Network for Texture and Structure-Aware Underwater Image Enhancement

Underwater images often suffer from color distortion, texture degradation, and structural blurring due to wavelength-dependent absorption and scattering. To address these issues, we propose FD-DMTNet, a novel two-stage enhancement framework that integrates frequency-domain priors with fine-grained structural refinement. In the first stage, a frequency-aware U-Net is built using Frequency-Domain Correction Blocks (FDCB) and Multi-Scale Feature Stream Blocks (MSFS), while a Frequency-Domain Transformer (FETB) with multi-head self-attention enables global context learning. In the second stage, a Fine-Grained Enhancement Module (FGEN) comprising three branches is introduced: A Texture Enhancement Branch (TEB) for multiscale texture recovery, a Color Correction Branch (CCB) for frequency-guided color adjustment, and a Structure Refinement Branch (SRB) using edge-aware attention and FETB to restore structural details. Extensive experiments on multiple benchmark datasets demonstrate that FD-DMTNet significantly outperforms existing methods in terms of color accuracy, texture clarity, and structural consistency. Compared with state-of-the-art approaches, it achieves average improvements of 3.66%, 2.04%, 2.48%, and 1.83% in PSNR, SSIM, UIQM, and NIQE, respectively.