A small target detection algorithm for unmanned aerial vehicles incorporating global information modeling and multi-scale feature interaction

Abstract

Aerial image object detection faces critical challenges due to target scale variation, high background complexity, and the vulnerability of small objects to noise. Existing methods remain limited in global context modeling and cross-scale feature interactions. To address these issues, we propose GM-YOLO, a novel small object detection framework that integrates global semantic modeling with dynamic multi-scale feature fusion. First, the CFC3K2 module synergizes convolutional neural networks (CNNs) and Transformers, leveraging depthwise separable convolutions and multilayer perceptrons (MLPs) to enhance local detail retention and mitigate feature dilution in small objects. Second, the SPPF-LSKA module employs large-kernel separable convolutions and dilated convolutions to optimize multi-scale feature fusion and global response capability. Third, the BiFPN-SDI architecture improves cross-level feature interaction efficiency through nonlinear multiplicative fusion and dynamic scale alignment. Additionally, the Shared Detail Enhancement Detection Head (SDEDH) reduces parameter redundancy via group normalization and parameter sharing while strengthening edge feature extraction. Finally, the SlideLoss function dynamically modulates gradients to alleviate sample imbalance. Experiments demonstrate that GM-YOLO achieves mAP@50 and mAP@50-95 scores of 43.8% and 27.0% on VisDrone2019, outperforming YOLOv11s by 4.7% and 3.6% respectively, with a 16% parameter reduction (7.9 million). Generalization tests on DOTAv2 further validate its robustness, achieving a 5.3% improvement in mAP@50. GM-YOLO surpasses mainstream detectors in both accuracy and efficiency for complex aerial scenarios.