EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Beyond Random Splits: A Temporal Robustness Benchmark for Industrial Food Anomaly Detection

Taewan Kim

IEEE Access · 2026 · 14 · 72331-72345

DOI: 10.1109/access.2026.3691935 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Taewan Kim. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Korea Institute of Planning and Evaluation for Technology in Food, Agriculture and Forestry (IPET) through the High Value-Added Food Technology Development Pro…, Ministry of Agriculture, Food and Rural Affairs.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Industrial visual anomaly detection has achieved near-saturated performance on benchmarks such as MVTec AD, yet deployed systems face challenges these benchmarks do not capture: deformable subjects, non-laboratory imaging, and temporal distribution shift. We introduce RoboCook-TAD, a benchmark of 10,124 images collected over six months from the quality-inspection camera of a commercially deployed udon-cooking robot. The dataset contains 4,512 normal frames, 2,083 foreign-object images with 5,219 bounding-box annotations, 1,847 process anomalies, and 1,682 operational anomalies for extended evaluation. RoboCook-TAD adopts a temporal train/test split as its primary protocol and introduces the Temporal Generalization Gap (TGG), defined as the AUROC difference between random and temporal splits for each method. We benchmark twelve methods spanning unsupervised, semi-supervised, supervised, and zero-shot families under a unified protocol. Key findings: 1) state-of-the-art methods drop from 99% AUROC on MVTec AD to 74–86% on our data; 2) method rankings change between protocols, with 2–10 AUROC points attributable to train/test similarity; 3) fully-supervised detectors achieve the highest random-split AUROC but the largest TGG, while semi-supervised methods offer the best trade-off; 4) temporal degradation correlates strongly with measurable feature-space distributional drift; and 5) on an extended open-set evaluation of operational anomalies, zero-shot vision-language models outperform all other families, revealing that visual detection alone is insufficient for complete quality assurance.

Förderung

  • Korea Institute of Planning and Evaluation for Technology in Food, Agriculture and Forestry (IPET) through the High Value-Added Food Technology Development Pro…
  • Ministry of Agriculture, Food and Rural Affairs

Zitieren

APA

Kim, T. (2026). Beyond Random Splits: A Temporal Robustness Benchmark for Industrial Food Anomaly Detection. IEEE Access, 14, 72331-72345. https://doi.org/10.1109/access.2026.3691935

Vancouver

Kim T. Beyond Random Splits: A Temporal Robustness Benchmark for Industrial Food Anomaly Detection. IEEE Access. 2026;14:72331-72345. doi:10.1109/access.2026.3691935

BibTeX
@article{Kim2026,
  author = {Kim, Taewan},
  title = {Beyond Random Splits: A Temporal Robustness Benchmark for Industrial Food Anomaly Detection},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {72331--72345},
  doi = {10.1109/access.2026.3691935}
}
Kostenlos starten →