Semi-PhysHDNet: Orthogonal Content Haze Disentanglement and Uncertainty-Calibrated Learning for Dense Dehazing
Hira Khan, Sung Won Kim · IEEE Access · 2026
Single-image dehazing plays an important role in adverse-weather vision, where visibility degradation affects both human perception and downstream scene understanding. Dense and non-uniform haze remains a stimulating degradation for image restoration since it simultaneously obfuscates scene structure, distorts color, and weakens the reliability of purely supervised models under real-world domain shift. This paper presents Semi-PhysHDNet, a dehazing framework built on orthogonal content-haze disentanglement and uncertainty-calibrated semi-supervised learning. The network integrates a haze-aware hierarchical Swin encoder with subband-selective frequency refinement, a latent disentanglement stage that separates scene content from haze-related degradation, and a physics-constrained atmospheric reconstruction branch for transmission and atmospheric light estimation. To advance adaptation to real hazy data, a heteroscedastic uncertainty mechanism is presented to down-weight unreliable regions during semi-supervised optimization. This joint design enables more stable restoration under dense and spatially varying haze conditions. Experimental results on synthetic and real-world benchmarks demonstrate strong and consistent performance, including 35.25/0.988 on RESIDE Outdoor, 23.98/0.867 on O-HAZE, and 20.87/0.789 on NH-HAZE. The proposed method also improves downstream YOLOv8s detection on Foggy Cityscapes from 8.77 to 11.01 detections per image, indicating practical value beyond pixel-level restoration.