DDSF-UNet: dual-domain spatial–frequency learning for underwater image enhancement

Abstract Underwater image enhancement (UIE) is still difficult. This is mainly due to wavelength-dependent light absorption, scattering, and environmental noise, which lead to low visibility, color distortion, and loss of fine details. Most existing deep learning (DL) based models primarily operate on spatial-domain convolutional operations and do not explicitly transform or manipulate image representations in the frequency domain (FD), where wavelength-related degradation characteristics can be more effectively modeled. In order to overcome these limitations, this study presents DDSF-UNet, a dual-domain spatial-frequency learning framework that jointly models spatial representations and FD information within a unified end-to-end architecture. A multi-residual module (MRM) mechanism is employed throughout the encoder and decoder to enhance local feature representation. At the network bottleneck, a dual-domain fusion module (DDFM) is introduced to integrate spatial features with FD cues via adaptive Fourier magnitude modulation, enabling effective suppression of underwater noise while preserving structural and textural details. Furthermore, a strengthen–operate–subtract (SOS) reconstruction strategy is employed in the decoder to refine features during upsampling and mitigate blur artifacts progressively. To ensure stable and effective optimization, we design a composite loss function that integrates pixel-wise reconstruction, structural similarity, perceptual consistency, total variation regularization, and FD constraints. Comprehensive experiments on benchmark UWI datasets show that DDSF-UNet achieves competitive quantitative performance and improved visual quality compared with recent state-of-the-art UIE methods, while extended ablation studies support the effectiveness of the main proposed components. The code is available on: https://github.com/najm-h/DDSF-UNet .