Uncertainty-Aware Diffusion Model for Real-World Image Dehazing

Diffusion models have recently achieved remarkable success in image dehazing tasks. However, existing methods struggle to handle real-world hazy images due to the neglect of physical properties, i.e., the haze density generally increases with the scene depth and the haze color changes uncertainly. To address this issue, we propose a novel Uncertainty-aware Diffusion model (UnDiff) for real-world image dehazing. Specifically, inspired by the atmospheric scattering model, we elaborate an Uncertainty Degradation Modeling (UDM) pipeline that considers density and color shifts of haze to better suppress the domain gap between synthetic and real images. To adaptively tackle the diverse hazy images, we design a Context-aware Prior Embedding (CPE) module to efficiently learn the contextual information that varies with the scene depth. Benefiting from the above physical guidance mechanism, our UnDiff achieves high-quality restoration on various real-world hazy images. Extensive experiments show that our method exhibits noteworthy superiority over existing methods.