PANet: A Physics-guided Parametric Augmentation Net for Image Dehazing by Hazing

TL;DR

PANet addresses the real-world dehazing domain gap by introducing a physics-guided, parametric haze augmentation framework. It learns pixel-wise haze parameters via a Haze-to-Parameter Mapper and reconstructs hazy images with a Parameter-to-Haze Mapper, guided by a depth map and refined with a Data-driven Haze Refiner, enabling realistic, non-homogeneous haze generation. The approach produces physically interpretable haze variants through pixel-wise resampling of haze density and atmospheric light, significantly boosting state-of-the-art dehazing models on real-world datasets. Empirical results across NH-Haze, O-Haze, and I-Haze demonstrate improved PSNR/SSIM and cross-dataset generalization, highlighting PANet’s practical impact for robust real-world image dehazing.

Abstract

Image dehazing faces challenges when dealing with hazy images in real-world scenarios. A huge domain gap between synthetic and real-world haze images degrades dehazing performance in practical settings. However, collecting real-world image datasets for training dehazing models is challenging since both hazy and clean pairs must be captured under the same conditions. In this paper, we propose a Physics-guided Parametric Augmentation Network (PANet) that generates photo-realistic hazy and clean training pairs to effectively enhance real-world dehazing performance. PANet comprises a Haze-to-Parameter Mapper (HPM) to project hazy images into a parameter space and a Parameter-to-Haze Mapper (PHM) to map the resampled haze parameters back to hazy images. In the parameter space, we can pixel-wisely resample individual haze parameter maps to generate diverse hazy images with physically-explainable haze conditions unseen in the training set. Our experimental results demonstrate that PANet can augment diverse realistic hazy images to enrich existing hazy image benchmarks so as to effectively boost the performances of state-of-the-art image dehazing models.

Figures (12)

• Figure 1: Left: Examples of the augmented images by PANet (yellow rectangles) and other augmentation methods (blue rectangles). Existing non-homogeneous hazy datasets NH-Haze_2020NH-Haze_2021 contain a very limited number of training pairs, as shown in green circles. Previous augmentation methods wu2023ridcpyang2022self cannot effectively generate non-homogeneous hazy images. In contrast, PANet can generate realistic non-homogeneous and homogeneous hazy images, as shown in yellow rectangles. Right: Comparison of the dehazing results with and without using data augmented by PANet.
• Figure 2: Overview of the proposed PANet. PANet comprises a Haze-to-Parameter Mapper (HPM) to project real haze images into a parameter space and a Parameter-to-Haze Mapper (PHM) to revert them back to the real space. In addition to generating original hazy images (green rectangle box), PANet can generate real hazy images not provided in the training set (yellow rectangle box).
• Figure 3: Block diagram of PANet. PANet utilizes a cyclic haze-parameter-haze mapping framework consisting of a Haze-to-Parameter Mapper (HPM) followed by a Parameter-to-Haze Mapper (PHM). Besides the hazy images in the original training set (green boxes), PANet can augment additional hazy images with various haze conditions unseen in the training set (yellow boxes).
• Figure 4: Architecture of Haze-to-Parameter Mapper (HPM). HPM consists of a shared encoder followed by two parallel parameter decoders to estimate the haze density map $\beta_\mathrm{est}(z)$ and atmospheric light map $A_\mathrm{est}(z)$, respectively.
• Figure 5: Visuals of hazy images generated by PANet. Given a hazy image, we can decrease or amplify its haze density by 2, as shown in (A) and (B). In addition, we can reverse its haze location or generate a complex hazy image, as shown in (C) and (D)