Accurate recognition model for tiny objects based on YOLOv10s

To address the challenges of insufficient feature resolution and computational redundancy in tiny object detection within UAV remote sensing scenarios, this study proposes a feature pyramid reconstruction-based model named P-YOLOv10s. By introducing a high-resolution P2 feature layer to construct a refined P2-P4 feature pyramid, the proposed model preserves shallow-layer detailed information while eliminating the redundant P5 layer to reduce model complexity. Experimental results on the VisDrone2019 dataset demonstrate that the enhanced model achieves 25.9% mAP50 at 640×640 resolution, representing a 1.3% improvement over the baseline YOLOv10s while reducing parameter count by 26%. Visualization analysis confirms the effectiveness of the proposed approach, offering a novel solution for real-time lightweight object detection applications.