Cross-Validated Benchmarking of YOLO Architectures for Ceramic Crack Detection with Efficiency Analysis
Journal of Robotics and Control (JRC) · 01.03.2026 · 7 (2) · 3671-3680
DOI: 10.18196/jrc.v7i2.29037 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Robotics and Control (JRC), online erschienen am 01.03.2026. Verfasst von Benchalak Maungmeesri und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
For inspecting a ceramic surface, detecting a crack is a crucial requirement because a small defect might propagate and lead to failure. However, many defect detection methods based on YOLO are still using a single split for training and testing, and their accuracy is often not analyzed in detail in terms of efficiency during deployment. This paper proposes a benchmarking methodology for detecting cracks in ceramics using five-fold cross-validation for six variants of YOLO-based architectures, namely YOLOv5nu, YOLOv8n, YOLOv9t, YOLO11n, YOLO12n, and YOLO26n, using a common dataset of 300 images with 399 instances for damage detection under control with the same hyperparameters for training, namely input size 640x640, batch size 16, and Stochastic Gradient Descent optimization with augmentation. Performance metrics are Precision, Recall, F1-score, mAP@50, and mAP@50-95. Deployment feasibility metrics are parameter count, model size, and GPU/CPU inference latency. The reliability of the statistical evaluation is improved through fold-level aggregation and 95% bootstrap confidence intervals. The results show that all YOLO architectures are differentiated in terms of mean mAP@50-95 under cross-validation: YOLOv9t performs best in terms of mean mAP@50-95 (0.672 ± 0.050, 95% CI [0.638, 0.716]) and has the smallest model size (4.43 MB, 1.97M parameters), while YOLO12n has the most stable performance across all folds (range 0.088) but has the highest CPU latency (26.06 ms/image). Accuracy-Efficiency plots show that there are non-monotonic relationships between model size/latency and localization quality. Thus, it is not the size of the architecture that matters but its design. The final training of the selected architecture on a 90/10 split validates stable convergence and excellent qualitative detection performance for a confidence threshold of 0.25. The evaluation protocol proposed here improves transparency for choosing YOLO architectures for ceramic crack inspection.
Zitieren
APA
Maungmeesri, B., Crisnapati, P. N., Maneetham, D., & Khonthon, S. (2026). Cross-Validated Benchmarking of YOLO Architectures for Ceramic Crack Detection with Efficiency Analysis. Journal of Robotics and Control (JRC), 7(2), 3671-3680. https://doi.org/10.18196/jrc.v7i2.29037
Vancouver
Maungmeesri B, Crisnapati PN, Maneetham D, Khonthon S. Cross-Validated Benchmarking of YOLO Architectures for Ceramic Crack Detection with Efficiency Analysis. Journal of Robotics and Control (JRC). 2026;7(2):3671-3680. doi:10.18196/jrc.v7i2.29037
BibTeX
@article{Maungmeesri2026,
author = {Maungmeesri, Benchalak and Crisnapati, Padma Nyoman and Maneetham, Dechrit and Khonthon, Sasithorn},
title = {Cross-Validated Benchmarking of YOLO Architectures for Ceramic Crack Detection with Efficiency Analysis},
journal = {Journal of Robotics and Control (JRC)},
year = {2026},
volume = {7},
number = {2},
pages = {3671--3680},
doi = {10.18196/jrc.v7i2.29037}
}
EUVIMED