EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Efficient Micro-LED Defect Detection Using a Pruned YOLO With CBAM and Class-Oriented Replacement Augmentation

Jeong-Sang Yu, Jun-Hee Lee, Eung-Tae Kim

IEEE Access · 2026 · 14 · 86257-86272

DOI: 10.1109/access.2026.3700082 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Jeong-Sang Yu und 2 weiteren Autor:innen. Die Arbeit nennt 35 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Collaborative Research and Development between Industry, Academy, and Research Institute funded by the Ministry of Small and Medium-sized Enterprises and Start….

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Recently, deep learning–based object detection methods have been widely adopted in automated visual inspection systems for real-time defect detection. However, micro-LED inspection presents two major challenges: severe class imbalance and strict computational constraints in embedded environments. Severe class imbalance often leads to biased optimization toward majority classes and degraded recall for minority instances, while strict computational constraints impose practical real-time requirements in embedded inspection environments. To address these issues, we propose a unified framework that integrates a pruned YOLO-based architecture with a novel data augmentation method, termed Class-Oriented Replacement Augmentation (CORA). The proposed architecture reduces computational complexity by removing task-irrelevant detection layers and incorporates attention mechanisms to enhance small-defects. In addition, CORA enables controllable minority-class augmentation by replacing majority-class instances at the object level, preserving spatial consistency and improving recall for defect instances, which constitute the minority class in the dataset. Experimental results show that the proposed method significantly improves detection performance and computational efficiency under class-imbalanced conditions. The final model achieves an F1-score of 0.9436, reduces model parameters by 51%, and yields a $2.57\times $ speedup in inference compared to the baseline. These results indicate that the proposed framework enables high detection accuracy with efficient real-time inference for micro-LED defect inspection.

Förderung

  • Collaborative Research and Development between Industry, Academy, and Research Institute funded by the Ministry of Small and Medium-sized Enterprises and Start…

Zitieren

APA

Yu, J., Lee, J., & Kim, E. (2026). Efficient Micro-LED Defect Detection Using a Pruned YOLO With CBAM and Class-Oriented Replacement Augmentation. IEEE Access, 14, 86257-86272. https://doi.org/10.1109/access.2026.3700082

Vancouver

Yu J, Lee J, Kim E. Efficient Micro-LED Defect Detection Using a Pruned YOLO With CBAM and Class-Oriented Replacement Augmentation. IEEE Access. 2026;14:86257-86272. doi:10.1109/access.2026.3700082

BibTeX
@article{Yu2026,
  author = {Yu, Jeong-Sang and Lee, Jun-Hee and Kim, Eung-Tae},
  title = {Efficient Micro-LED Defect Detection Using a Pruned YOLO With CBAM and Class-Oriented Replacement Augmentation},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {86257--86272},
  doi = {10.1109/access.2026.3700082}
}
Kostenlos starten →