HFG-YOLOv11: a high-frequency-guided feature-enhanced method for small-object detection

Small-object detection in aerial scenes remains challenging because repeated downsampling erodes already limited pixel information, while standard fusion and prediction heads inadequately exploit high-resolution cues and contextual dependencies. To address these limitations, we propose HFG-YOLOv11, a stage-coupled detector based on YOLOv11-n. The high-frequency-guided detail preservation stem combines a fixed channel-wise Laplacian-of-Gaussian branch with a multi-path spatial semantic branch to preserve edge and texture cues before further downsampling. The high-resolution feature fusion neck constructs a seven-node bidirectional topology across P2 to P5 to coordinate semantic propagation and localization feedback. The P2 tiny-object detection branch extends prediction to four scales, while the transformer prediction head applies window attention to P2/P3 and global attention to P4/P5 for scale-dependent contextual modeling. Across five independent runs, HFG-YOLOv11 achieves 29.03±0.46% mAP50:95, 47.34±0.28% mAP50, and 13.80±0.45% APS on VisDrone2019 and 24.05±0.43%, 38.76±0.26%, and 12.19±0.48% on UAVDT, outperforming YOLOv11-n by 5.76/7.20/3.11 and 5.82/7.18/2.38 percentage points, respectively. At 640×640 resolution on an NVIDIA RTX 3090 graphics processing unit (GPU), it reaches 100 and 92 frames per second (FPS), respectively. These results demonstrate that HFG-YOLOv11 improves small-object detection accuracy while maintaining real-time GPU inference.