Physics-Guided Image Dehazing Diffusion

TL;DR

IDDM addresses the synthetic-to-real generalization gap in image dehazing by embedding the atmospheric scattering model into a diffusion framework. It introduces a time-indexed haze component $\mathbf{h}_t$ and a joint forward diffusion of haze and noise, with a DDIM-like sampler to recover clear images via $J = \frac{\mathbf{x}_0}{1 - \tilde{h}_T}$. The method uses a two-stage training: physics-guided denoising (Stage I) and a haze estimator (Stage II) trained in a bi-directional loop, enabling robust domain generalization. Experiments on RTTS, BEDDE, and other datasets show state-of-the-art performance and strong cross-domain transfer, even when trained on indoor synthetic data.

Abstract

Due to the domain gap between real-world and synthetic hazy images, current data-driven dehazing algorithms trained on synthetic datasets perform well on synthetic data but struggle to generalize to real-world scenarios. To address this challenge, we propose \textbf{I}mage \textbf{D}ehazing \textbf{D}iffusion \textbf{M}odels (IDDM), a novel diffusion process that incorporates the atmospheric scattering model into noise diffusion. IDDM aims to use the gradual haze formation process to help the denoising Unet robustly learn the distribution of clear images from the conditional input hazy images. We design a specialized training strategy centered around IDDM. Diffusion models are leveraged to bridge the domain gap from synthetic to real-world, while the atmospheric scattering model provides physical guidance for haze formation. During the forward process, IDDM simultaneously introduces haze and noise into clear images, and then robustly separates them during the sampling process. By training with physics-guided information, IDDM shows the ability of domain generalization, and effectively restores the real-world hazy images despite being trained on synthetic datasets. Extensive experiments demonstrate the effectiveness of our method through both quantitative and qualitative comparisons with state-of-the-art approaches.

Figures (6)

• Figure 1: Motivation for physics-guided Dehazing diffusion. Haze formation and the diffusion forward process share a common characteristic: both progressively degrade image clarity. IDDM integrates the physical haze formation process (depth-indexed) into diffusion (time-indexed) by simultaneously adding atmospheric scattering and noise.
• Figure 2: Overview of the proposed framework. We introduce a physics-guided diffusion process that explicitly models haze formation through the atmospheric scattering model. (A) During forward diffusion, both physical haze $\mathbf{h}_t$ and noise ${\boldsymbol{\epsilon}_t}$ are progressively added to clear images through shared timesteps, where $\mathbf{h}_t$ accumulates according to depth-dependent scattering attenuation. (B) The reverse process leverages a Haze Estimator (HtNet) to predict timestep-dependent haze intensity, which conditions the denoising network for physics-aware image restoration. (C) A bi-directional training strategy is employed: the frozen denoising U-Net supervises HtNet training via consistency regularization, while predicted haze components guide noise prediction across varying degradation levels.
• Figure 3: Visual comparisons on dense hazy images from RTTS dataset reside and Fattal's dataset fattal. All models are trained on OTS dataset. Please zoom in for more details.
• Figure 4: Visual comparisons on a low-light hazy image from RTTS dataset reside. All models are trained on ITS dataset.
• Figure 5: Visual comparison of methods from Table. \ref{['tab:ridcp']} on a very dense hazy image from RTTS dataset reside.