EUVIMEDEUVIMED entdecken
Crossref

Training-Free Vision-Language Model for Substrate Glass Defect Detection

Jiaxuan Liu, Hui Zhang, Hai Wang, Bingrui Zhao, Yang Zhao, Yingtian Yu

IEEE Transactions on Industrial Informatics · 2026-08 · 22 (8) · 7359-7370

DOI: 10.1109/tii.2026.3685293 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Industrial Informatics, 2026-08. Verfasst von Jiaxuan Liu und 5 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Gefördert von: National Science and Technology Major Project, Intelligent Manufacturing Systems and Robotics, Major Program of the National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Substrate glass is essential for advanced applications, such as displays and semiconductors, where microscopic defects can critically degrade yield and reliability. Automated optical inspection for this task faces a tripartite challenge: 1) severe data scarcity, stemming from industrial confidentiality and evolving defect morphologies; 2) multiillumination complexity, which requires synthesizing information from complementary bright-field and dark-field imaging; and 3) multifocal sequential dependency, arising from defects located at varying depths within the glass. Conventional supervised methods are infeasible under such data constraints, while existing zero-shot vision-language models (VLMs) are ill-equipped to handle these domain-specific imaging characteristics. To address this gap, we propose Training-Free framework for Substrate Glass Defect Detection (TFSGD). Without requiring any weight updates, TFSGD introduces a novel pipeline that first leverages illumination-specific prompt engineering and clustering-based text embedding selection to generate robust classifiers from a frozen VLM. It then fuses multiillumination information through a novel structure-sparse low-rank fusion module. Finally, it resolves focal dependency using a clarity-guided sequence refinement strategy. Extensive experiments on a real-world industrial dataset demonstrate that TFSGD establishes a new state-of-the-art in zero-shot defect segmentation and localization, offering a practical, readily deployable solution that circumvents the prohibitive costs of data collection and annotation in industrial settings.

Förderung

  • National Science and Technology Major Project
  • Intelligent Manufacturing Systems and Robotics
  • Major Program of the National Natural Science Foundation of China
  • Key Program of the National Natural Science Foundation of China
  • National Key Research and Development Program of China
  • Equipment Industry Department I Project of the Ministry of Industry and Information Technology of China

Zitieren

APA

Liu, J., Zhang, H., Wang, H., Zhao, B., Zhao, Y., & Yu, Y. (2026). Training-Free Vision-Language Model for Substrate Glass Defect Detection. IEEE Transactions on Industrial Informatics, 22(8), 7359-7370. https://doi.org/10.1109/tii.2026.3685293

Vancouver

Liu J, Zhang H, Wang H, Zhao B, Zhao Y, Yu Y. Training-Free Vision-Language Model for Substrate Glass Defect Detection. IEEE Transactions on Industrial Informatics. 2026;22(8):7359-7370. doi:10.1109/tii.2026.3685293

BibTeX
@article{Liu2026,
  author = {Liu, Jiaxuan and Zhang, Hui and Wang, Hai and Zhao, Bingrui and Zhao, Yang and Yu, Yingtian},
  title = {Training-Free Vision-Language Model for Substrate Glass Defect Detection},
  journal = {IEEE Transactions on Industrial Informatics},
  year = {2026},
  volume = {22},
  number = {8},
  pages = {7359--7370},
  doi = {10.1109/tii.2026.3685293}
}
Kostenlos starten →