Packaging Defect Detection in Lean Manufacturing: A Comparative Study of YOLOv8, YOLOv9, and YOLOv10

This paper presents a comprehensive evaluation of various YOLO models for packaging defect detection within a lean manufacturing context. We utilized a dataset comprising images of boxes on a packaging line labeled with seven classes, including six classes of defects and one class without any defects. A comparative study of YOLO models including YOLOv8n, YOLOv8s, YOLOv8m, YOLOv9t, YOLOv9s, YOLOv9m, YOLOv10n, YOLOv10s, and YOLOv10m were exploited to create models for detecting defects on the packaged boxes. Based on the experiments, the YOLOv10 models especially the YOLOv10s and YOLOv10m models, perform better in term of precision and time efficiency, compared to previous versions like YOLOv8 and YOLOv9. The recommended model depends on the trade-off between mAP and processing time. YOLOv10m achieves the highest mAP at 0.989 within 33.5 ms, YOLOv10s achieves the highest mAP at 0.987 within 25.60 ms, and YOLOv10n achieves the highest mAP at 0.976 within 15.40 ms. This paper provides valuable insights into the effectiveness of YOLO for defect detection in packaging within industrial settings, contributing to improved quality control processes and operational efficiency.