RS-YOLO: A YOLO-Based Method for Small Object Detection in Remote Sensing

This paper introduces RS-YOLO, a novel model for small object detection in remote sensing, designed to address the challenges of missed and false detections in scenarios with densely distributed, small objects. By integrating the Biformer, a dynamic sparse attention mechanism with low computational cost, the model can better suppress background noise and enhance the ability to detect small objects. We also utilize an efficient neck structure, Gold-YOLO, based on the Gather-and-Distribute (GD) mechanism to enhance feature fusion capabilities. Furthermore, we combine Inner-IoU and CloU loss functions as a new bounding box regression (BBR) loss function, enhancing localization performance and model robustness while accelerating training. On AI-TOD and DIOR datasets, RS-YOLO achieves mAPO.5 scores of 56.3 % and 74.2 %, outperforming baseline models and other mainstream algorithms.