MF-DETR: an efficient end-to-end framework for small object detection in UAV imagery

Abstract

Unmanned aerial vehicle (UAV) object detection plays an important role in intelligent transportation, environmental monitoring, and public safety. However, UAV images usually contain dense small objects, large scale variations, occlusion, and complex background interference, making accurate and efficient detection challenging. In addition, end-to-end models designed for natural images do not work well in small object detection, while real road environment in the wild makes object detection more difficult. Therefore, this paper proposes MF-DETR, a degradation-oriented end-to-end detection framework for UAV small-object detection. Instead of simply adding independent enhancement blocks to RT-DETR, MF-DETR follows a progressive representation recovery strategy. First, a Statistical Intra-scale Semantic Enhancement module is introduced to suppress background-dominated token responses and strengthen weak intra-scale semantics. Second, a Multi-scale Frequency-spatial Collaboration Fusion module is designed to recover texture and boundary details that are easily lost during multi-scale feature fusion. Finally, a Frequency-guided Feature Alignment and Enhancement module is developed to correct cross-level semantic and spatial misalignment before detection. Through this semantic recalibration–detail recovery–feature alignment pipeline, MF-DETR improves the representation quality of small and occluded UAV objects while maintaining efficient inference. Experiments on VisDrone demonstrate that MF-DETR improves AP and \(\hbox {AP}_{50}\) by 3.5% and 4.9%, respectively, compared with the RT-DETR-R18 baseline. The source code is available at https://github.com/zcjMoxx/MF-DETR .