YOLO-SS: optimizing YOLO for enhanced small object detection in remote sensing imagery

The identification of minuscule objects in remote sensing data presents a formidable challenge in computer vision, where objects may occupy a mere handful of pixels. The lack of unique shape features in such small objects hinders the effectiveness of established object detection algorithms. Remote sensing of small object detection plays an important role in areas such as environmental monitoring and estimating agricultural production. To address this challenge, in this study, we introduce YOLO-SS, an enhanced version of the YOLO algorithm tailored specifically for small object detection in remote sensing imagery. YOLO-SS incorporates an optimized backbone network, a restructured loss function and an asymmetric training sample weighting strategy. These improvements prioritize the model’s attention toward high-quality positive samples of small objects while reducing sensitivity to complex backgrounds. Evaluation on the AI-TOD dataset demonstrates YOLO-SS’s exceptional performance, achieving an AP50 score of 0.535, surpassing YOLOv6L by 13.4% and other popular object detection algorithms. Our findings offer a novel pathway for advancing small object detection capabilities in diverse remote sensing applications.