PDN-Net: a unified physics-driven network for image dehazing

Engineering notes

Experiments were conducted on the RESIDE-SOTS indoor and outdoor benchmarks datasets. PDN-Net achieves 27.32 dB PSNR and 0.94 SSIM indoors, and 30.24 dB PSNR and 0.94 SSIM outdoors, outperforming classical methods such as DCP and advanced deep models, including AOD-Net, Dehaze-UNet, Light-DehazeNet, and Fourmer.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Image dehazing caused by the complex interplays of aerosol scattering, atmospheric turbulence, and wavelength-dependent dispersion is still a difficult problem to address. These interplays jointly degrade visibility and distort scene radiance. Traditional physics-based methods and deep learning-based methods have been employed in the past to address these problems. However, the physics-based methods fail under heterogeneous haze, and the deep learning methods often ignore physical constraints. Thus, both techniques limit interpretability and generalization. To address these challenges, this paper proposes PDN-Net, which is a unified Physics-Driven Network that integrates the three dominant haze-formation mechanisms into a single differentiable mathematical model. PDN-Net introduces a novel optimization framework that embeds gradient-descent iterations directly into the network. The entire framework is a composite of a lightweight ResNet module that models the underlying physical degradation, a learnable channel-weighting mechanism that accounts for dispersion, and an enhanced U-Net corrector that stabilizes iterative updates while preserving structural details. Experiments were conducted on the RESIDE-SOTS indoor and outdoor benchmarks datasets. PDN-Net achieves 27.32 dB PSNR and 0.94 SSIM indoors, and 30.24 dB PSNR and 0.94 SSIM outdoors, outperforming classical methods such as DCP and advanced deep models, including AOD-Net, Dehaze-UNet, Light-DehazeNet, and Fourmer. Ablation studies confirm that removing any core module significantly degrades performance. Robustness tests under noise and brightness attenuation show that PDN-Net maintains the highest MS-SSIM and visual consistency. These results highlight the model’s strong generalization and stability in real-world conditions where haze co-occurs with illumination changes and sensor noise. Therefore, PDN-Net provides a physically interpretable, high-performance, and robust framework for single-image dehazing, offering a reliable foundation for safety-critical applications such as autonomous driving, surveillance, and remote sensing.

Links and sources

Comments