SDE-YOLO: A lightweight synergistic dynamic enhancement framework for small object detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

To address the challenges of dense small objects, complex backgrounds, and limited computational power in edge devices for UAV aerial imagery, this paper proposes an efficient synergistic dynamic enhancement detection framework (SDE-YOLO) that builds on YOLOv11n as the baseline. First, an Inverse Gated Bidirectional Feature Pyramid Network (IGB-FPN) is constructed to effectively suppress background noise while preserving spatial boundary details in shallow layers. Second, a Synergistic Dynamic Detail-Enhanced Head (SDDH) is designed, achieving spatial alignment and multi-task collaboration with a low parameter budget via dynamic task decomposition and shared convolutions. Third, to mitigate the phenomenon of over-suppression of dense small objects, an Adaptive Scale-aware Dynamic NMS (ASD) mechanism is proposed, effectively improving the detection accuracy of small objects. Furthermore, Layer Adaptive Magnitude Pruning (LAMP) is utilized for structural model compression, and the optimized model is deployed on the RK3588 edge computing platform. To comprehensively evaluate the generalization and robustness of the proposed model, extensive validations are conducted on three challenging public datasets: VisDrone, TinyPerson, and DOTA 1.0. Experimental results on the core VisDrone dataset show that, compared to baseline models, SDE-YOLO improves the mean accuracy (mAP50) by 16.83 percentage points, while further reducing the number of parameters and computational complexity by 63.98% and 42.70%, respectively, through pruning. Final embedded field deployment tests confirm that SDE-YOLO achieves an effective balance between detection accuracy and lightweight deployment, providing a viable application solution for drone edge vision systems.