PA-YOLO: Small Target Detection Algorithm With Enhanced Information Representation for UAV Aerial Photography

Small object detection in drone aerial imagery faces significant challenges due to the small object size, frequent occlusion, and complex background interference. To address these issues, this letter proposes an improved YOLOv8 algorithm—PA-YOLO [poly Kernel inception and contextual anchor attention (PC-C2f) and attention scale sequence fusion feature pyramid network and dysample upsampling operator (ASD-FPN)]. First, we propose a PC-C2f structure to enhance feature extraction capabilities. Second, the ASD-FPN network is designed in the neck section to effectively prevent information loss during downsampling, while an additional small object detection layer significantly improves detection performance. Finally, soft nonmaximum suppression is applied to reduce missed detections in dense scenes. Experimental results show that the proposed algorithm has accuracy of 42.3% in terms of mAP0.5 metric, improving that of 11.3% compared to YOLOv8 baseline model, significantly enhancing the accuracy of small object detection and making it suitable for resource-constrained environments.