Comparison among YOLO Series in SAR Ship Detection with Preliminary Results

Deep learning has shown significant advantages in object detection, particularly with the You Only Look Once (YOLO) model. YOLO adopted an end-to-end training and detection method that balances speed and accuracy, making it an effective way for synthetic aperture radar (SAR) ship detection. This paper mainly compared several mainstream YOLO series, focusing on detection accuracy and efficiency in SAR ship detection. The experiment results with SAR Ship Detection Dataset indicated that YOLOv6 outperformed other YOLO series models in terms of accuracy and efficiency, achieving a mAP of 0.742 with FPS being 200.00. However, it has a larger parameter size (of 16.30M). Meanwhile, both YOLOv5 and YOLOv8 showed good overall performance with fewer parameters compared against YOLOv6, while their mAP and FPS were relatively lower. This comprehensive analysis highlighted the advancements of different YOLO series, providing valuable insights for further investigation on improving object detection (e.g., SAR ship detection).