SFMNet: a small object detection algorithm for UAV perspectives based on a feature-enhanced backbone network

Small-object detection in UAV aerial imagery remains challenging because targets often suffer from low spatial resolution, occlusion, background clutter, and missed detections. To address these problems, this paper proposes SFMNet, an improved small-object detection algorithm for UAV imagery. First, the original backbone is redesigned as SONet, in which the Small-Object Block (SOBlock) enhances multi-scale feature representation through parallel receptive-field modelling and feature fusion. Second, a dual-attention mechanism is introduced to emphasize informative channel and spatial features while suppressing irrelevant background responses. Third, a dedicated small-object detection layer is added to preserve shallow fine-grained features and reduce feature loss caused by repeated downsampling. Experiments on the VisDrone2019 dataset show that SFMNet improves mAP@0.5 by 7.4% points and mAP@0.5:0.95 by 4.7% points compared with the baseline. Comparative experiments with representative detection methods further demonstrate that SFMNet achieves competitive accuracy while maintaining favourable computational efficiency. Additional experiments under different input resolutions and on the DOTA-v1.0 dataset further verify the stability and applicability of the proposed method in UAV and remote-sensing scenarios.