Enhanced small target detection in UAV imagery: a new training paradigm and optimized YOLOv8n architecture

Unmanned aerial vehicles (UAV) aerial imagery presents significant challenges for object detection due to small, densely packed targets and complex backgrounds. This study introduces an enhanced model, MGT-YOLOv8n(Freeze), extending the YOLOv8n architecture to address these challenges. We propose a novel training paradigm with a classification-focused pre-training phase and a detection phase that leverages inherited classification weights to enhance feature extraction capabilities. The pre-training phase involves cropping instances from the detection dataset for category-specific classification, with weights frozen and inherited during detection.We enhance the YOLOv8n architecture through modifications: integration of the VSS module in the backbone network for improved detail feature capture, employing the GDFPN structure in the neck for optimized multi-scale feature fusion and lightweight dynamic up-sampling, and incorporating the TADH module in the head for improved task alignment. Additionally, we introduce the WSIoU loss function to reduce low-quality frame impacts. Experimental results demonstrate the superiority of our training paradigm, On the HIT-UAV dataset, MGT-YOLOv8n(Freeze) achieves substantial gains of 8.7% and 9.1% in mAP@0.5 and mAP@0.95, respectively, while improving mAP@0.5 and mAP@0.95 on VisDrone2019 by 2.7% and 2.2%, significantly outperforming traditional direct training approaches. These comprehensive results validate the effectiveness of our improved model for small target detection in UAV aerial imagery and the viability of the proposed training paradigm.