EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Foundation Vision Models for Manufacturing Defect Recognition Under Open Set

Marta Alvarez, Yujin Han, Alessandra Greco

Journal of geospace science · 01.06.2026 · 6 (01) · 180-191

DOI: 10.71465/jgs55 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of geospace science, online erschienen am 01.06.2026. Verfasst von Marta Alvarez und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The transition toward automated quality control in modern manufacturing necessitates robust visual inspection systems capable of identifying both known and unforeseen anomalies. Traditional deep learning approaches typically operate under a closed-set assumption, where the classes of defects encountered during deployment are strictly identical to those present in the training data. This paradigm fails in dynamic industrial environments where novel defect typologies emerge due to variations in raw materials, tool wear, or environmental shifts. Recently, foundation vision models, trained on massive scale datasets using self-supervised or weakly supervised objectives, have demonstrated unprecedented generalization capabilities across various downstream tasks. This paper presents a comprehensive evaluation of state-of-the-art foundation vision models deployed for open-set manufacturing defect recognition. We systematically analyze the zero-shot and few-shot anomaly detection performance of prominent architectures, exploring their internal feature representations when exposed to out-of-distribution industrial anomalies. Furthermore, we propose a rigorous evaluation framework tailored for open-set scenarios, integrating multi-domain industrial datasets to assess model resilience, scalability, and practical applicability. Our extensive empirical analysis reveals that while foundation models significantly outpace traditional supervised baselines in recognizing unseen defects, they exhibit distinct vulnerabilities related to domain shift and fine-grained texture discrimination. The insights derived from this study provide a crucial roadmap for integrating large-scale vision models into robust, adaptive, and scalable Industry 4.0 inspection pipelines.

Zitieren

APA

Alvarez, M., Han, Y., & Greco, A. (2026). Foundation Vision Models for Manufacturing Defect Recognition Under Open Set. Journal of geospace science, 6(01), 180-191. https://doi.org/10.71465/jgs55

Vancouver

Alvarez M, Han Y, Greco A. Foundation Vision Models for Manufacturing Defect Recognition Under Open Set. Journal of geospace science. 2026;6(01):180-191. doi:10.71465/jgs55

BibTeX
@article{Alvarez2026,
  author = {Alvarez, Marta and Han, Yujin and Greco, Alessandra},
  title = {Foundation Vision Models for Manufacturing Defect Recognition Under Open Set},
  journal = {Journal of geospace science},
  year = {2026},
  volume = {6},
  number = {01},
  pages = {180--191},
  doi = {10.71465/jgs55}
}
Kostenlos starten →