Enhanced UAV Detection in Foggy Environments: Integrating Receptive Field Attention and Global Context in YOLOv8

Detecting unmanned aerial vehicles (UAVs) in foggy conditions poses significant challenges due to poor visibility and image degradation. To address this, we propose Defog-YOLO, a neural network that integrates receptive field spatial attention convolution (RFAConv) and the global context attention mechanism (GCNet) into YOLOv8 for enhanced UAV detection. The RFAConv, incorporated into the defogging module of the All-in-One dehazing network (AOD-Net), effectively extracts UAV features from foggy images. Meanwhile, the GCNet, integrated into YOLOv8, leverages long-range dependencies to boost the detection of small objects. An end-to-end joint inference framework optimizes the performance of both defogging and detection. We introduce a new dataset, GUET-UAV-FOG, comprising both real and synthesized foggy UAV images to evaluate our approach. Experimental results demonstrate that Defog-YOLO significantly improves mean average precision (mAP) up to $80.2 \%$, outperforming baseline methods, thus showcasing its effectiveness in UAV detection even under adverse foggy conditions.