Rethinking Adaptive Contextual Information and Multi-Scale Feature Fusion for Small-Object Detection in UAV Imagery

Small object detection in unmanned aerial vehicle (UAV) imagery poses significant challenges due to insufficient feature representation, complex background interference, and extremely small target sizes. These factors collectively degrade the performance of conventional detection algorithms, leading to low accuracy, frequent missed detections, and false alarms. To address these issues, we propose <b>YOLO-DMF</b>, which is a novel detection framework specifically designed for drone-based scenarios. Our approach introduces three key innovations from the perspectives of feature extraction and information fusion: (1) a <b>D</b>etail-Semantic Adaptive Fusion (DSAF) module that employs a multi-branch architecture to synergistically enhance shallow detail features and deep semantic information, thereby significantly improving feature representation for small objects; (2) a <b>M</b>ulti-Scale Residual Spatial Attention (MSRSA) mechanism incorporating scale-adaptive spatial attention to improve robustness against background clutter while enabling a more precise localization of critical target regions; and (3) a <b>F</b>eature Pyramid Reuse and Fusion Network (FPRFN) that introduces a dedicated 160×160 detection head and hierarchically combines multi-level shallow features with high-level semantic information through cross-scale fusion, effectively enhancing sensitivity to both small and tiny objects. Comprehensive experiments on the VisDrone2019 dataset demonstrate that YOLO-DMF outperforms state-of-the-art lightweight detection models. Compared to the baseline YOLOv8s, our method achieves improvements of 3.9% in [email protected] and 2.5% in [email protected]:0.95 while reducing model parameters by 66.67% with only a 2.81% increase in computational cost. The model achieves a real-time inference speed of 34.1 FPS on the RK3588 NPU, satisfying the latency requirements for real-time object detection. Additional validation on both the AI-TOD and WAID datasets confirms the method's strong generalization capability and promising potential for practical engineering applications.