Insight / Blog
Semiconductor AI Vision Inspection : Reducing AOI False Positives
Microscopic defects in semiconductor test components can directly affect product quality and reliability. AOI systems can rapidly flag large numbers of potential defects, but a high false-positive rate creates a heavy secondary review workload and drives up the total cost of reaching a final defect decision.
Company A, a manufacturer of semiconductor test components, wanted to keep its existing AOI process in place while using AI Vision to reclassify AOI images and results. The goal was to reduce false positives and improve confidence in defect decisions.
Background : High AOI False Positives and Review Burden
Company A’s AOI process produced too many false positives, creating a significant review burden. When large numbers of candidates must be checked before real defects can be identified, inspection speed and quality-control efficiency both suffer.
The company needed a separate AI Vision layer that could address AOI over-detection without replacing its existing equipment or disrupting established inspection workflows.
Approach : AI Vision as a Secondary Decision Layer
SURROMIND designed a secondary AI Vision layer that reviews the images and defect candidates flagged by AOI. Rather than replacing AOI, the solution narrows the set of candidates that require human review.
- Built defect datasets from inspection images
- Labeled data by defect type and trained classification models
- Configured the inspection server and recipe editor
- Connected labeling, model training, deployment, and inspection log management in one operating workflow
- Designed the integration around existing AOI equipment and production requirements
Results : Good Parts Reduced from 80% to 5% of Defect Candidates
| Item | Result |
|---|---|
| Existing inspection issue | About 80% of AOI defect candidates were good parts |
| False-positive reduction | After AI Vision was added, good parts represented about 5% of the remaining defect candidates |
| Model performance | 98.1% average accuracy on validation data |
| Review productivity | Review time fell from 0.4 hours to 0.14 hours, while inspection-process productivity improved by 10% |
Operational Impact : Production-Ready Inspection Operations Beyond Model Accuracy
AOI excels at quickly flagging potential defects, but high over-detection means more candidates must be checked manually. This project kept the customer’s existing equipment and AOI workflow intact, then added AI Vision as a second decision layer to narrow the candidates requiring review.
Model performance alone was not enough. The project also brought together the inspection server, recipe editor, labeling, model training and deployment, and log management to create a workflow that teams could operate in production.
The model achieved an average accuracy of 98.1% on validation data. With fewer candidates to review, review time decreased from 0.4 hours to 0.14 hours and inspection-process productivity increased by 10%.
