High-quality Image Dehazing with Diffusion Model

TL;DR

DehazeDDPM tackles dehazing in dense-haze scenarios by embedding physics into a conditional diffusion framework. It splits the task into two stages: a physics-guided stage that estimates $trmap$, $J$, and $A$ under the Atmospheric Scattering Model $I(x)=J(x)t(x)+A(1-t(x))$, followed by a diffusion-based refinement conditioned on $J$ and $trmap$ to recover heavily information-lost regions. Key innovations include Fog-aware and Distribution-closer Conditions, Confidence-guided Dynamic Fusion, and a frequency-prior loss that emphasizes high-frequency restoration, along with EMA-based color stabilization. Extensive experiments on synthetic and real hazy datasets show state-of-the-art perceptual and distortion metrics, with improved generalization to natural hazy images, indicating practical impact for surveillance, autonomous driving, and outdoor imaging in adverse weather.

Abstract

Image dehazing is quite challenging in dense-haze scenarios, where quite less original information remains in the hazy image. Though previous methods have made marvelous progress, they still suffer from information loss in content and color in dense-haze scenarios. The recently emerged Denoising Diffusion Probabilistic Model (DDPM) exhibits strong generation ability, showing potential for solving this problem. However, DDPM fails to consider the physics property of dehazing task, limiting its information completion capacity. In this work, we propose DehazeDDPM: A DDPM-based and physics-aware image dehazing framework that applies to complex hazy scenarios. Specifically, DehazeDDPM works in two stages. The former stage physically models the dehazing task with the Atmospheric Scattering Model (ASM), pulling the distribution closer to the clear data and endowing DehazeDDPM with fog-aware ability. The latter stage exploits the strong generation ability of DDPM to compensate for the haze-induced huge information loss, by working in conjunction with the physical modelling. Extensive experiments demonstrate that our method attains state-of-the-art performance on both synthetic and real-world hazy datasets.

Figures (15)

• Figure 1: The visual examples of dehazing results were sampled from real-world hazy images. The second to fourth columns show the results of Dehamer guo2022image, our first-stage, and our DehazeDDPM, respectively. Our method demonstrates unprecedented perceptual quality on the challenging real-world datasets.
• Figure 2: Statistics illustration of the haze-induced information loss, including t-SNE clustering van2008visualizing, distribution distance, histogram, gradient, entropy, and standard deviation. The dense haze causes massive information loss in content and color.
• Figure 3: Thumbnail of main idea. Most previous image dehazing methods learn the mapping from hazy to clear images. Our method memorizes the data distribution of clear images by introducing conditional DDPM into image dehazing.
• Figure 4: Overview structure of the proposed DehazeDDPM. DehazeDDPM works in two stages. In the former stage, the physical modelling network generates $J$, $trmap$, and $A$, governed by the underlying SAM physics. For the latter stage, the Fog-aware and Distribution-closer Conditions (FDC) endows the DehazeDDPM with fog-aware ability and pulls the distribution closer to the clear data. The Confidence-guided Dynamic Fusion (CDF) leverages the transmission map as confidence map to incorporate the well restored region of the first stage into the second stage, mitigating the learning difficulty of DDPM for image dehazing.
• Figure 5: Different components of the first stage. The three rows are respectively sampled from SOTS li2018benchmarking, Dense-Haze Dense-Haze_2019, and NH-HAZE ancuti2020nh datasets. The estimated $trmap$ perfectly reflects the density of the fog, where the darker region denotes denser fog.