Spatio-Temporal Adaptive Airlight Estimation with Depth-Aware Transmission for Feature-Preserving Video Dehazing
Ibrahim Salim Sulaiman Al Farsi, Mohd Shafry Mohd Rahim, Devi Willieam Anggara, Falah Y. H. Ahmed, Papiya Giri · Human-Centric Intelligent Systems · 2026
Spatially non-uniform haze, temporal inconsistencies such as frame-to-frame flickering and abrupt airlight variation, and color distortion artifacts including color shift, desaturation, and contrast imbalance make video dehazing a challenging task. While recent deep learning based methods are capable of producing strong spatial restoration, they often suffer adversely from temporal instability in heavy and dynamically varying conditions of haze. In this paper, we present a Spatio-Temporal Adaptive Airlight Estimation (STAAE) framework for video dehazing. The proposed method tackles heterogeneous atmospheric conditions by modeling airlight as a spatio-temporally varying parameter and blends depth-aware transmission estimation with adaptive post-enhancement. Through the use of region based airlight estimation and temporal consistency constraints as well as transmission refinement aimed at preserving features, STAAE can yield stable haze removal results that are visually consistent under the experimental settings evaluated in this work. Experiments on synthetic and real-world video datasets report quantified improvements in spatial quality, structural fidelity, temporal stability, and downstream vision performance compared with the evaluated baselines and the closest rivals. Specifically, our evaluation demonstrates that STAAE can achieve PSNR = 23.67 ± 0.89 dB, SSIM = 0.913 ± 0.021, and reduces the Temporal Flicker Index (TFI) from 0.068 to 0.031 (54.4% reduction) relative to the strongest baseline approach. Furthermore, YOLOv5 [email protected] improves by 42.8%–63.7% across light-to-dense haze levels.