Evaluation of YOLO Key Modules Using Synthetic Aperture Images

ABSTRACT

Target detection and recognition in synthetic aperture images remain challenging. Given the advantages of the YOLO series, four enhancement schemes (Marine, SDS, DINOv3, MoonNet) are evaluated comprehensively. These modules are integrated into YOLOv8, YOLOv11, YOLOv12 and YOLOv13 in this letter. Sixteen comparative models are constructed. Three main conclusions are obtained. First, MoonNet is the most robust method. Performance improvements are achieved for YOLOv8 and YOLOv11. Slight drops in performance are observed for YOLOv12 and YOLOv13. Second, high sensitivity to the base architecture is shown by SDS and marine. A 0.0172 drop is observed in marine‐based YOLOv11. This is the largest drop. Third, the performance of all DINOv3‐based methods is poor. No positive improvements are achieved at all. These conclusions provide practical guidance for users in model selection.