MPD-Net: a maritime port dehazing network for structure-preserving and hue-consistent visibility restoration

Coastal fog and haze reduce the structural clarity, color fidelity, and visibility of images taken at coastal ports. This study proposes MPD-Net, a dual-candidate restoration network tailored for port scenes. It sequentially applies a color and luminance hint module (CLH) and a frequency-aware refinement layer (FARL) to the base output; the detail branch embeds a detail-guided feature modulation module (DGF) into the backbone network and applies CLH at the output. The dual-candidate adaptive fusion module (DCAF) fuses the two candidate results and performs color correction. On 315 paired foggy and clear port images, MPD-Net achieves the best overall PSNR, SSIM, MSE, and LPIPS among the compared approaches. Additional analyses cover scene subsets, three fogging coefficients, module contributions, and red-target preservation. For red targets, MPD-Net obtains the lowest hue MAE, whereas DehazeFormer reports lower CIEDE2000, RGB MAE, and saturation MAE. The color-related advantage is therefore specific to hue fidelity in the evaluated red regions. Future work will involve further validation using licensed real-world marine fog images and task-aware port scenarios.