DSFNet: Directional Spatial–Frequency Collaboration Network for Underwater Optical Image Enhancement

Abstract

Underwater optical images frequently suffer from wavelength-dependent attenuation and scattering, resulting in color distortion, contrast reduction, and structural-detail loss. Recent methods have already explored directional modeling and spatial–frequency fusion; coordinating these operations through explicitly specified feature interactions remains a design question. To address this issue, we propose the Directional Spatial–Frequency Collaboration Network (DSFNet) for underwater optical image enhancement. Building on LU2Net’s encoder–decoder and axial depthwise convolution, DSFNet adds separate sparse main- and anti-diagonal branches to form OctaRayDW. A spatially conditioned Fourier module updates amplitude-derived coefficients and phase responses, while two reciprocal cross-domain attention operations exchange spatial and frequency-enhanced features before gated fusion. A parallel convolutional path is adaptively combined with histogram attention for local and contextual modeling. These components are integrated into a three-stage encoder–decoder containing 5.58 M parameters. On the Underwater Image Enhancement Benchmark (UIEB) dataset, DSFNet achieves 25.352 dB PSNR and 0.929 SSIM, exceeding the retrained original two-stage LU2Net by 0.704 dB and 0.004, respectively. Additional evaluations on C60, U45, MABLs, and UFO120 indicate favorable perceived quality under the adopted no-reference metrics across diverse underwater conditions. The results confirm the effectiveness of directional spatial–frequency collaboration for underwater optical image enhancement.