High-quality prior-guided dehazing with multi-color space fusion and haze-aware prompting
Yunyun Ji, Liang Xiong, Xiaobao Shi · The Imaging Science Journal · 2025
Haze degrades image quality by reducing contrast andobscuring details, adversely affecting computer vision tasks such as objectdetection and segmentation. Current dehazing methods face several mainchallenges: the synthetic-to-real domain gap, inaccurate color restorationresulting from RGB channel coupling, and poor adaptability to varying hazedensities. To address these limitations, a novel framework termed HPG-DM isproposed, which integrates High-Quality Priors with Multi-color space fusionand haze-aware Prompting. The approach is based on a pretrained VQGAN to bridgedomain gaps by mapping degraded inputs into a high-fidelity semantic space. Adual-branch encoder processes RGB and HSV spaces concurrently, whilecross-attention fusion enhances color fidelity. A Residual Transformer equippedwith a Prompt Block dynamically adjusts dehazing intensity based on Value–Saturation differences as a haze density proxy. Extensiveexperiments on the SOTS-outdoor and RTTS datasets demonstrate that the proposedmethod achieves state-of-the-art performance in detail preservation, colorfidelity, and generalization.