MFR-YOLO: advancing UAV object detection with multi-scale feature refinement via deformable convolution and global attention

Unmanned aerial vehicle imagery exhibits extreme scale variance, high small-object density, and arbitrary geometric distortions, which continue to pose challenges to the state-of-the-art detectors like YOLO series. This paper proposes MFR-YOLO, an enhanced model built upon a multi-scale feature refinement network. First, a multi-scale feature extraction module is designed, integrating SPD-Conv to preserve fine-grained details and employing DCNv4 to dynamically adjust receptive fields for adapting to objects of varying scales and orientations. Standard convolutions in the backbone are systematically replaced with DCNv4 layers, enhancing feature propagation from low-level edges to high-level semantics. A lightweight Global Attention Module is incorporated into both the backbone and neck, leveraging dual channel-spatial self-attention to suppress background clutter and amplify critical object features. Additionally, the original C3K2 bottleneck is upgraded with a Pyramid-Pooling Attention mechanism, enriching multi-scale representations by fusing local and global contextual information. Comprehensive evaluations on the VisDrone2021 and UA-DETRAC benchmarks demonstrate that MFR-YOLO achieves superior overall detection accuracy, exhibits significant gains in small object detection, and maintains a balance between accuracy and real-time efficiency. Ablation studies validate the contribution of each proposed component. This research provides a novel technical framework for object detection algorithms in complex scenes.