EUVIMEDEUVIMED entdecken
Crossref

P‐104: Category‐Aware Adaptive Masking for Industrial Anomaly Detection

Wei Gao, Shujuan Yin, Bin Lv, Xingqun Jiang

SID Symposium Digest of Technical Papers · 2026-06 · 57 (1) · 2093-2096

DOI: 10.1002/sdtp.70001_569 ↗

Kurz gefasst

Fachartikel in der Zeitschrift SID Symposium Digest of Technical Papers, online erschienen am 11.09.2026. Verfasst von Wei Gao und 3 weiteren Autor:innen. Die Arbeit nennt 9 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Industrial anomaly detection is critical for quality control. In display manufacturing (e.g. OLED panel production), common defects such as bright spots, dark spots, line defects, and Mura pose significant challenges for automated inspection. Modern manufacturing faces the need to detect multiple defect types and continually adapt to new defect classes as product lines diversify and new models or processes introduce novel anomalies. Traditional methods often struggle to separate the feature distributions of different object categories, leading to ambiguous anomaly judgments, and they typically require a dedicated model per category ‐ an inefficient and inflexible approach. In this paper, we propose Category‐Aware Adaptive Masking (CAAM) for multi‐class industrial anomaly detection. CAAM employs two key innovations: (1) We inject learnable category tokens into a Transformer's self‐attention mechanism, guiding the model to encode each category's features in an independent subspace and thereby enhance inter‐class separability. (2) We use a masked image reconstruction strategy ‐ during training, random patch masking forces the model to learn robust global representations of normal data, and during inference, adaptively masking suspected anomalous regions prevents the model from reconstructing those regions. By focusing reconstruction on normal regions, our approach produces large errors on defective areas, making anomalies easy to detect and localize. We evaluate CAAM on standard industrial defect datasets in a unified multi‐class setting. Experiments show that CAAM achieves 98.50% image‐level AUROC on the MVTec‐AD dataset and image‐level I‐F1 of 68.58% on the challenging MVTec‐AD2 dataset.

Zitieren

APA

Gao, W., Yin, S., Lv, B., & Jiang, X. (2026). P‐104: Category‐Aware Adaptive Masking for Industrial Anomaly Detection. SID Symposium Digest of Technical Papers, 57(1), 2093-2096. https://doi.org/10.1002/sdtp.70001_569

Vancouver

Gao W, Yin S, Lv B, Jiang X. P‐104: Category‐Aware Adaptive Masking for Industrial Anomaly Detection. SID Symposium Digest of Technical Papers. 2026;57(1):2093-2096. doi:10.1002/sdtp.70001_569

BibTeX
@article{Gao2026,
  author = {Gao, Wei and Yin, Shujuan and Lv, Bin and Jiang, Xingqun},
  title = {P‐104: Category‐Aware Adaptive Masking for Industrial Anomaly Detection},
  journal = {SID Symposium Digest of Technical Papers},
  year = {2026},
  volume = {57},
  number = {1},
  pages = {2093--2096},
  doi = {10.1002/sdtp.70001_569}
}
Kostenlos starten →