Domain-Constrained Augmentation for Industrial Bounding Box Detection Under Limited Data Conditions
IEEE Access · 2026 · 14 · 163604-163615
DOI: 10.1109/access.2026.3746074 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Matouš Cejnek und 2 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Institutional Resources of CTU in Prague for Research (RVO12000) and by the Center of Advanced Aerospace Technology, ESIF under the EU Operational Programme Research, Development and Education at the Faculty of Mechanical Engineering, Czech Technical University in Prague.
Abstract (Originalsprache)
In the steel processing industry, machine-vision based tracking via bounding box detection presents unique challenges distinct from common defect detection tasks. These challenges include highly variable lighting conditions, small annotated datasets, and the necessity for highly reliable real-time system performance to ensure safety and operational reliability. While numerous advanced data augmentation methods exist in machine-vision, their complexity, sensitivity, and practical deployment issues limit adoption in industrial applications. This work contributes a practical adaptation of TrivialAugment for industrial bounding box detection by introducing box-secure, domain-aware spatial transformations and systematically evaluating augmentation magnitude limits under real deployment constraints. The evaluation uses three publicly released, real-world datasets collected from long-term steel-production deployments and spanning different production processes, image properties, preprocessing procedures, and object distributions. Emphasizing practical utilization and robustness to extreme lighting variations, we provide empirically grounded guidance for practitioners working with limited data and mainly static backgrounds. Compared with training without augmentation, Conservative TrivialAugment improved mean average precision across intersection-over-union thresholds from 0.5 to 0.95 by 176%, 56%, and 64% for the three evaluated datasets, respectively, demonstrating substantial gains with limited implementation complexity.
Förderung
- Institutional Resources of CTU in Prague for Research (RVO12000) and by the Center of Advanced Aerospace Technology
- ESIF under the EU Operational Programme Research, Development and Education at the Faculty of Mechanical Engineering, Czech Technical University in Prague
Zitieren
APA
Cejnek, M., Prchlik, V., & Vrba, J. (2026). Domain-Constrained Augmentation for Industrial Bounding Box Detection Under Limited Data Conditions. IEEE Access, 14, 163604-163615. https://doi.org/10.1109/access.2026.3746074
Vancouver
Cejnek M, Prchlik V, Vrba J. Domain-Constrained Augmentation for Industrial Bounding Box Detection Under Limited Data Conditions. IEEE Access. 2026;14:163604-163615. doi:10.1109/access.2026.3746074
BibTeX
@article{Cejnek2026,
author = {Cejnek, Matouš and Prchlik, Victor and Vrba, Jan},
title = {Domain-Constrained Augmentation for Industrial Bounding Box Detection Under Limited Data Conditions},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {163604--163615},
doi = {10.1109/access.2026.3746074}
}
EUVIMED