GFREA-Net: A Lightweight Method for Guided Feature Rapid Extraction and Small-Object Detection From Aerial Imagery

Due to the characteristics of drone aerial images—high resolution, variable perspectives, and complex backgrounds—detecting small objects in traffic scenes poses challenges such as minute target sizes, severe occlusions, and dense distributions. Existing deep learning detection methods have made progress in feature extraction and fusion, but limitations remain in ensuring real-time performance and high accuracy. To address these issues, we propose GFREA-Net, a lightweight method for guided feature rapid extraction and small object detection from aerial imagery. This network achieves a balance between detection performance and model complexity through a multi-layer architecture. First, the Multi-Scale Guided Attention (MSGA) mechanism is introduced. It enhances feature representation for small and occluded objects in complex backgrounds through cross-scale semantic interaction and directional guidance. Second, the Rapid Feature Extraction (RFE) module is designed to achieve lightweight feature encoding. It significantly reduces convolutional redundancy and computational overhead while preserving feature expressiveness. Finally, we propose the Fused Lightweight Coupled Head (FLCH) for small object detection. By incorporating Group-Aware Convolution(GAConv), FLCH enhances modeling capabilities for fine-grained edge features, thereby improving detection accuracy for distant, minute objects. Experiments on our self-built dataset for drone traffic intersections, the public Roboflow dataset, and the VisDrone benchmark dataset demonstrate that GFREA-Net achieves an optimal balance between accuracy and efficiency. Compared to baseline models, it achieves 1.8% and 5.3% improvements in mAP50 and mAP50-95, respectively, while reducing parameters by 16.1%. It also demonstrates excellent generalization performance across test datasets. This research provides an efficient, lightweight solution for detecting small targets from UAVs in complex urban environments, offering significant reference value for intelligent traffic perception and UAV-based smart monitoring.