P‐89: Structural Anomaly Detection via Vision Transformer with Dual Attention Mechanism
SID Symposium Digest of Technical Papers · 2026-06 · 57 (1) · 2039-2042
DOI: 10.1002/sdtp.70001_554 ↗Kurz gefasst
Fachartikel in der Zeitschrift SID Symposium Digest of Technical Papers, online erschienen am 11.09.2026. Verfasst von Lei Song und 4 weiteren Autor:innen. Die Arbeit nennt 7 Literaturangaben.
Abstract (Originalsprache)
Anomaly detection, particularly structural anomaly detection, is a critical topic in display manufacturing. This paper presents a novel structural anomaly detection method based on Vision Transformer, with the following key innovations: (1) Meta‐Feature Extractor is designed to learn robust normal sample prototypes, effectively addressing the challenge of modeling normality under high intra‐class variation; (2) Dynamic Reconstruction Mechanism is introduced to capture correlations between input features and normal prototypes, mitigating class imbalance by reconstructing only normal patterns; (3) a unified cost function integrating global anomaly loss, local anomaly consistency loss, and KoLeo regularization is proposed to enhance fine‐grained feature discriminability for hard sample mining, enabling fully unsupervised anomaly detection without pixel‐level annotations. Requiring only normal samples for training, the proposed method achieves accurate anomaly detection, improving the metric I‐AUROC from 0.72 to 0.75 and P‐AUPRO from 0.63 to 0.72. The experimental results demonstrate potential to positively influence industrial defect inspection paradigm.
Zitieren
APA
Song, L., Guo, H., Yin, S., Lv, B., & Jiang, X. (2026). P‐89: Structural Anomaly Detection via Vision Transformer with Dual Attention Mechanism. SID Symposium Digest of Technical Papers, 57(1), 2039-2042. https://doi.org/10.1002/sdtp.70001_554
Vancouver
Song L, Guo H, Yin S, Lv B, Jiang X. P‐89: Structural Anomaly Detection via Vision Transformer with Dual Attention Mechanism. SID Symposium Digest of Technical Papers. 2026;57(1):2039-2042. doi:10.1002/sdtp.70001_554
BibTeX
@article{Song2026,
author = {Song, Lei and Guo, Haiyang and Yin, Shujuan and Lv, Bin and Jiang, Xingqun},
title = {P‐89: Structural Anomaly Detection via Vision Transformer with Dual Attention Mechanism},
journal = {SID Symposium Digest of Technical Papers},
year = {2026},
volume = {57},
number = {1},
pages = {2039--2042},
doi = {10.1002/sdtp.70001_554}
}
EUVIMED