PSF-YOLO11: an enhanced YOLO11 framework for small object detection in UAV imagery

Abstract Images captured by unmanned aerial vehicles are characterized by complex backgrounds, significant scale variations, and a large proportion of small and densely distributed objects, which pose substantial challenges to object detection algorithms. To address these issues, this paper proposes PSF YOLO11, a detection framework for complex aerial scenes that aims to improve small object detection performance. First, to enhance high level semantic representation under complex backgrounds, a parallel multi scale convolution module is introduced into the high level feature extraction stages of YOLO11. Second, a small object feature pyramid is proposed to address the insufficient representation of small objects in shallow layers by redesigning the small object detection branch, thereby improving fine grained spatial feature extraction and multi scale information interaction. Third, to alleviate feature inconsistency during multi scale aggregation, a shared weight feature pyramid structure is designed to enhance feature fusion effectiveness. Experimental results on the VisDrone2019 dataset demonstrate that, compared with the baseline YOLO11 model, the proposed method improves mAP@0.5 and mAP@0.5:0.95 by 4.6% and 3.0%, respectively, while achieving a 3.5% improvement in Recall over the baseline. On the drones in optics recognition dataset, which contains a broader range of aerial object categories and diverse imaging conditions, the proposed method achieves improvements of 2.0% in mAP@0.5 and 2.3% in mAP@0.5:0.95, and 1.9% in Recall over the baseline. These results indicate that the proposed method exhibits superior performance in detecting small and densely distributed objects in complex aerial environments, with strong generalization capability across diverse remote sensing scenarios.