EFAF-YOLO: A Novel Enhanced Framework for Small-Object Detection

Remote sensing platforms provide crucial data for applications such as urban planning and traffic monitoring. However, detecting small objects remains challenging due to low resolution, cluttered backgrounds, and scale variations. In this work, we propose efficient feature aggregation framework YOLO(EFAF-YOLO), a lightweight and efficient architecture for real-time detection in remote sensing imagery. The model incorporates two novel modules: the Multi-scale Feature Enhancement (MFE) module, which expands receptive fields via multi-branch and dilated convolutions to improve small-object representation, and the Adaptive Multi-scale Feature Fusion (AMFF) module, which selectively integrates shallow and deep features to recover targets often overlooked in deeper layers. Additionally, an Efficient Multi-scale Attention (EMA) module is embedded in the backbone to emphasize small targets in cluttered scenes. Evaluated on VisDrone2019, AI-TOD, and RS-STOD, EFAF-YOLOs achieves 45.4% mAP50 on VisDrone2019 (+6.8 over YOLO11s), 53.7% mAP50 on AI-TOD (+4.0), and 77.3% mAP on RS-STOD (+5.2), with only 1.2M additional parameters, while maintaining real-time inference at 5.2 ms per image (192 FPS on RTX 5070), well above the 30 FPS requirement. These results demonstrate the method’s effectiveness for efficient and real-time interpretation of remote sensing images.