Dynamic Feature Distillation for Enhanced UAV Object Detection

Abstract

Uncrewed aerial vehicle (UAV) target detection is highly challenging in mixed urban scenes due to small target scales, sparse feature representations and strong background coupling, all of which can often lead to undetected targets and incorrect predictions. In order to address these issues, we propose a dynamic feature distillation (DFD) framework with cross-granularity regional guidance (CRG) to improve UAV detection quality while maintaining model efficiency. In particular, DFD adaptively adjusts the distillation weights of channel–spatial features according to the consistency between teacher and student predictions. In addition, CRG dynamically partitions feature maps into hierarchical regions to prioritize knowledge transfer for small-scale targets. We evaluate our method using two datasets: the DUT Anti-UAV dataset and our self-built Urban-UAV dataset. The results show clear performance improvements achieved by our approach. On DUT Anti-UAV, our framework achieves an mAP50 of 0.904 and an mAP50:95 of 0.619, outperforming the YOLO11-s baseline by 1.2% and 1.0%, respectively. Performance on small targets also improves by 0.6%. The experimental results indicate that our method is practical for real-time urban UAV detection tasks.