Underwater Image Enhancement via Dual-Path Frequency-Domain Fusion Network

Underwater images typically suffer from color distortion and reduced clarity due to wavelength-dependent light absorption and scattering. Many existing learning-based methods learn a direct pixel-domain mapping from degraded images to enhanced results, while seldom exploiting explicit frequency-domain priors that are important for high-quality restoration. To address this, we propose DPF-Net, a Dual-Path Frequency-domain Fusion Network. Specifically, we first decompose the input into low- and high-frequency components via the discrete wavelet transform (DWT). We then design a low-frequency branch integrating Low-Frequency Guided Attention (LFGA) and the Color Feature Extraction Module (CFEM), and a high-frequency branch integrating High-Frequency Aware Modulation (HFAM) and the Detail Feature Extraction Module (DFEM), to capture frequency-specific characteristics. During upsampling, we introduce a WaveUp module consisting of the Dual-path Up-Interaction Module (DUIM) and a Multi-kernel Convolution-based Fusion Network (MCFN) to enable effective cross-frequency feature interaction and fusion, producing high-quality enhanced images. Extensive experiments on benchmark datasets demonstrate that DPF-Net achieves superior qualitative and quantitative performance compared with state-of-the-art methods, and real-world tests further verify its practical potential. Code is available at https://github.com/YuMa-star-lab/DPF-Net.