A Deep Learning Approach for Automated Re-Inspection of False Calls in Mixed-Production Surface Mount Lines Under Extreme Class Imbalance
- 1College of Management and Economics, Tianjin University, Tianjin 300072, China
Abstract
Automated Optical Inspection equipment on Surface Mount Technology lines is tuned conservatively so that genuine defects are not missed, producing a persistent stream of false calls that a human operator must reconfirm in a secondary re-inspection step. This study reports a deep learning classifier for that decision, developed and evaluated on 34,466 images from one operating mixed-model line with an acceptable-to-defective ratio of 196 to 1. A three-stage procedure screens five architectures, compares three augmentation intensities, and sweeps twenty class weight settings under five-fold cross-validation. Two operating points are reported rather than one best model: a balanced configuration at 84.85% precision and 78.29% recall, and a high-capture configuration at 89.14% recall. Restricting evaluation to boards absent from the training partition gives 73.69% and 87.71% recall, indicating that the image-level split inflates recall by one to five percentage points. Because the classifier filters a population currently reviewed in full by hand, both configurations reduce human review volume by more than 98% on this line. An illustrative cost comparison, conditioned on an assumed accuracy of the manual re-inspection it replaces, indicates which operating point is preferred at a given ratio of released-defect cost to review cost. A secondary observation, offered as a hypothesis rather than a demonstrated result, is that per-model learning rates fixed before an augmentation study may produce apparent architecture-specific failures that are optimization artifacts.
Related
- FDUNet-LRSA: efficient frequency-adaptive segmentation for automated crop disease inspection
- Mitigating Class Imbalance in Graph-Structured Data via Hierarchical Learning: Insights from Protein Binding Site Prediction
- Metric-driven Pre-acquisition Diagnostics for Automated Microscopy Ultraviolet Surface Excitation (MUSE)
- Closed, Automated CAR-T Cell Manufacturing: from Research to Point of Care production.
- Comparative evaluation of machine learning algorithms for Maize Lethal Necrosis Disease detection under class imbalance
BibTeX
@article{tian2026deep,
title = {{A Deep Learning Approach for Automated Re-Inspection of False Calls in Mixed-Production Surface Mount Lines Under Extreme Class Imbalance}},
author = {Kun Tian and Zhanwen Niu},
journal = {Applied Sciences},
year = {2026},
month = oct,
doi = {10.3390/app162010068},
url = {https://doi.org/10.3390/app162010068},
note = {CC BY},
}
RIS
TY - JOUR
TI - A Deep Learning Approach for Automated Re-Inspection of False Calls in Mixed-Production Surface Mount Lines Under Extreme Class Imbalance
AU - Kun Tian
AU - Zhanwen Niu
PY - 2026
DA - 2026/10/26
JO - Applied Sciences
DO - 10.3390/app162010068
UR - https://doi.org/10.3390/app162010068
AB - Automated Optical Inspection equipment on Surface Mount Technology lines is tuned conservatively so that genuine defects are not missed, producing a persistent stream of false calls that a human operator must reconfirm in a secondary re-inspection step. This study reports a deep learning classifier for that decision, developed and evaluated on 34,466 images from one operating mixed-model line with an acceptable-to-defective ratio of 196 to 1. A three-stage procedure screens five architectures, compares three augmentation intensities, and sweeps twenty class weight settings under five-fold cross-validation. Two operating points are reported rather than one best model: a balanced configuration at 84.85% precision and 78.29% recall, and a high-capture configuration at 89.14% recall. Restricting evaluation to boards absent from the training partition gives 73.69% and 87.71% recall, indicating that the image-level split inflates recall by one to five percentage points. Because the classifier filters a population currently reviewed in full by hand, both configurations reduce human review volume by more than 98% on this line. An illustrative cost comparison, conditioned on an assumed accuracy of the manual re-inspection it replaces, indicates which operating point is preferred at a given ratio of released-defect cost to review cost. A secondary observation, offered as a hypothesis rather than a demonstrated result, is that per-model learning rates fixed before an augmentation study may produce apparent architecture-specific failures that are optimization artifacts.
SN - 2076-3417
ER -