84‐1: Attention‐Guided Large Vision Models for Data‐Efficient Display Defect Detection
SID Symposium Digest of Technical Papers · 2026-06 · 57 (1) · 1158-1162
DOI: 10.1002/sdtp.70001_313 ↗Kurz gefasst
Fachartikel in der Zeitschrift SID Symposium Digest of Technical Papers, online erschienen am 11.09.2026. Verfasst von Zhihong Pan und 6 weiteren Autor:innen. Die Arbeit nennt 14 Literaturangaben.
Abstract (Originalsprache)
To facilitate defect detection in new product lines with limited labeled data, we propose a self‐supervised Vision Transformer (ViT) framework tailored for defection detection and classification in display manufacturing. Our approach builds on DINOv2 self‐distillation, adapting it to focus on multi‐scale anomalies in domain‐specific optical inspection images. During pretraining, we utilize the ViT's self‐attention maps to bias data augmentations, ensuring that random crops and masked patches consistently include critical defect regions. This prevents the model from ignoring subtle anomalies common in high‐resolution industrial imagery. In the finetuning stage, we further exploit attention maps to calculate a weighted average of patch tokens, capturing localized defect features that outperform the standard class token. Additionally, we employ attention‐based background masking to enhance data augmentation. Evaluated on 11 diverse labeled datasets, our method improves F1 scores by 2.6% over generic transfer learning baselines and achieves comparable performance using only 30% of the training data. These results demonstrate a scalable, cost‐efficient solution for intelligent quality control.
Zitieren
APA
Pan, Z., Shenoy, R., Lee, H., Balakrishnan, K., Jeon, Y., An, S., & Kim, J. (2026). 84‐1: Attention‐Guided Large Vision Models for Data‐Efficient Display Defect Detection. SID Symposium Digest of Technical Papers, 57(1), 1158-1162. https://doi.org/10.1002/sdtp.70001_313
Vancouver
Pan Z, Shenoy R, Lee H, Balakrishnan K, Jeon Y, An S, et al. 84‐1: Attention‐Guided Large Vision Models for Data‐Efficient Display Defect Detection. SID Symposium Digest of Technical Papers. 2026;57(1):1158-1162. doi:10.1002/sdtp.70001_313
BibTeX
@article{Pan2026,
author = {Pan, Zhihong and Shenoy, Rahul and Lee, Hyemin and Balakrishnan, Kaushik and Jeon, Yongmoon and An, Sooho and Kim, Jaewon},
title = {84‐1: Attention‐Guided Large Vision Models for Data‐Efficient Display Defect Detection},
journal = {SID Symposium Digest of Technical Papers},
year = {2026},
volume = {57},
number = {1},
pages = {1158--1162},
doi = {10.1002/sdtp.70001_313}
}
EUVIMED