Validation Proportion Matters: Generalization Stability of YOLO-Based Wafer Defect Detection with Limited Data
Publication Date
1-1-2026
Document Type
Conference Proceeding
Publication Title
IEEE Workshop on Microelectronics and Electron Devices Wmed
DOI
10.1109/WMED71264.2026.11509677
First Page
21
Last Page
25
Abstract
Automated wafer surface inspection increasingly relies on deep learning, yet limited annotated datasets challenge generalization stability. This study investigates how validation proportion affects YOLO-based defect detection using a 300image microscopy dataset with three defect classes: small-area undercut (sUdc), large-area undercut (IUdc), and particles (Ptc). YOLOv8m and YOLOv11m were evaluated under 70/15/15, 70/20/10, and 80/10/10 train/validation/test splits. Under identical training conditions, YOLOv11m with a 70/20/10 split achieved the best performance (m A P 50=0.819, Recall =0.810). Training with 80% data did not improve performance and decreased validation stability. Results demonstrate that, in limited semiconductor datasets, maintaining adequate validation sampling is more critical than maximizing training size.
Keywords
AdamW, Defect Detection, Semiconductor, YOLO
Department
Electrical Engineering
Recommended Citation
Chih Kuan Ho and D. W. Parent. "Validation Proportion Matters: Generalization Stability of YOLO-Based Wafer Defect Detection with Limited Data" IEEE Workshop on Microelectronics and Electron Devices Wmed (2026): 21-25. https://doi.org/10.1109/WMED71264.2026.11509677
