PFR-Det: Progressive Feature Refinement Small Object Detection Framework for UAV Images
Abstract
Vision-based uncrewed aerial vehicle (UAV) detection, which encompasses both UAV-perspective observation and anti-UAV surveillance, is crucial for remote sensing, yet challenging due to resource constraints and complex environments. Targets are often tiny and submerged in background clutter, degrading detection accuracy. To bridge this gap, this article proposes a Progressive Feature Refinement small object detection framework (PFR-Det) based on YOLOv12. Specifically, a Pyramid Context C2f module extracts features from different receptive fields through parallel paths to enhance weak target representations. Furthermore, DySample is employed for dynamic upsampling in the neck to reconstruct fine-grained morphological details with high fidelity. In addition, an asymmetric spatial-channel attention module accentuates salient features while attenuating background noise, complemented by a lightweight GhostHead to streamline feature decoding. Comprehensive evaluations across VisDrone2019-DET, UAVDT, and DUT-AntiUAV benchmarks substantiate that PFR-Det strikes a superior balance between precision and latency. Notably, on VisDrone2019-DET, our model achieves an $mAP_{50}$ of 34.6%, surpassing the YOLOv12 baseline by 6.0% with merely 2.04 M parameters.