Multi-Level Feature Fusion Based UAV Object Detection Method Under Foggy Weather

Adverse weather conditions can blur aerial images captured by optical cameras on UAVs, complicating the already challenging task of detecting small-sized targets from a UAV's perspective. In this paper, a novel model called DU-YOLO is proposed. Firstly, an FFA-Net-based defog module is introduced, ensuring effective image preprocessing while meeting real-time requirement. Then, a multi-level feature fusion module based on the squeeze-excitation attention mechanism is proposed to enhance information extraction and detail preservation. Additionally, the focaler-MPDIoU loss function is designed to enhance the learning ability for small target samples and accelerate convergence. Finally, experimental results are tested on the synthetic fog dataset VisDrone-Fog, demonstrating that DU-YOLO achieves a 4.9% increase on mAP compared to the baseline model, YOLOvlOs.