Frequency-based and Physics-guiding Diffusion Model for Single Image Dehazing

The diverse degradation of real-world hazy images inevitably reduces the performance of deep learning-based methods for image dehazing. However, most approaches primarily operate in the raw pixel space of images, which limits the exploration of the frequency characteristics of haze images, leading to inadequate utilization of deep models’ representational capabilities in producing high-quality images. Moreover, the interpretability of deep learning methods regarding the diffusion models of hazing remains largely unexplored. In this paper, we consider a dehazing framework, called FP-Diff, that fully utilizes the physics-guiding features and frequency domain information based on conditional diffusion models. Motivated by the realization that deep networks impede the learning of local details, captured by high frequencies, we introduce a Spatial Frequency Block (SFB) to focus on high-and mid-frequency features that are missed in the frequency domain to achieve better reconstruction of fine details in images. To improve the dehazing explainability of the diffusion model, we introduce a Physics-guiding Block(PGB). With the above techniques, our method can show state-of-the-art(SOTA) performance on synthetic datasets and real-world datasets, achieving competitive performance in visual quality.