Insight / Blog
Semiconductor Component AI Vision : Pass, Defect, and Rework Decisions
Wire bonding in semiconductor packaging relies on extremely small precision components. Tiny cracks, breakage, or shape differences in these parts can directly affect final semiconductor quality.
Because the parts are so small and inspection criteria are demanding, manual review can be slow and decisions can vary between inspectors.
Background : Slow and Inconsistent Manual Inspection of Precision Parts
Company A, a manufacturer of precision semiconductor packaging components, relied on inspectors to classify good parts, defects, and rework candidates from monitor images. As volume grew, processing slowed and judgment varied when separating recoverable parts from scrap.
The company needed an AI Vision system that could increase inspection speed while keeping decisions consistent across a repetitive quality-control process.
Approach : Region Extraction and Three-Stage AI Classification
SURROMIND built an integrated AI Vision system that identifies each component in an inspection image and classifies its quality status in three stages.
- Automatic region extraction: located target components in source images and organized them into individual inspection units.
- Multi-stage decisions: filtered clear good parts first, then classified the remaining parts as defects or rework candidates.
- AutoML: compared candidate inspection models and selected the strongest-performing option for deployment.
- Continuous improvement: connected field corrections with retraining so model decisions could improve over time.
Results : About 90% Faster Inspection with 97%+ Final Accuracy
| Metric | Previous process | SURRO Inspection |
|---|---|---|
| Inspection processing time | 27 minutes | 2 minutes 40 seconds, approximately 90% faster |
| Component-region extraction | - | 99% accuracy |
| Final decision accuracy | Varied by inspector | Consistent decisions at 97%+ accuracy |
| Operating model | One-time manual classification | Continuous cycle of decisions, label correction, and retraining |
Operational Impact : Connecting Rework Decisions with Field Feedback
The project went beyond image classification. It was designed to reduce inspector fatigue and help recover parts that could be reworked instead of discarded.
AI Vision located component regions with 99% accuracy and filtered clear good and defective parts first. Inspectors could then focus on ambiguous cases and components identified for rework.
Processing time fell from 27 minutes to 2 minutes 40 seconds, reducing a major inspection bottleneck while improving the consistency of rework decisions.
