Physics-inspired RGB-depth feature modulation and fusion for hazy aerial object detection
Xiaorong Zhang, Han Liao, Xuting Hu · Journal of King Saud University - Computer and Information Sciences · 2026
Unmanned Aerial Vehicles (UAVs) suffer from severely degraded visual perception in hazy environments due to atmospheric scattering, which reduces contrast and blurs geometric features, making aerial object detection challenging. Existing cascaded dehazing methods often introduce artifacts owing to task misalignment, while purely data-driven approaches lack interpretable mechanisms for handling low-visibility conditions. To address these challenges, we propose a physics-inspired haze modulation detector (PHMDet), a physics-inspired multimodal framework for hazy aerial object detection that incorporates atmospheric-scattering-inspired cues into neural feature fusion. The core Physics-Inspired Haze Modulation Fusion (PHMFusion) module leverages depth priors and learned ASM-inspired degradation descriptors to dynamically modulate RGB features, enabling feature-space suppression of haze-induced degradation. To handle drastic scale variations in aerial imagery, the Scale-Adaptive Selection Module (SASM) promotes interaction between shallow textures and deep semantics via bidirectional feature diffusion. Furthermore, the Context-Guided Refinement Module (CGR) exploits long-range dependencies for global calibration, effectively suppressing background false alarms. Experiments on HazyDet, together with quantitative and qualitative evaluation on Real-world Task-driven Testing Set (RTTS) and additional analyses on Haze-DIOR, support the effectiveness and robustness of PHMDet in hazy scenes. Notably, it achieves 53.5% Average Precision (AP) on HazyDet, surpassing the external YOLOv11s RGB-only baseline by 3.6 AP points. When evaluated as an RGB-depth detector with precomputed RGB-depth inputs, PHMDet has 5.22 million detector parameters and a detector-side speed of 53.2 frames per second (FPS), indicating a favorable detector-side accuracy-efficiency trade-off excluding the offline pseudo-depth generation stage.