AI vision is not simply a replacement for human eyes. It lowers quality cost only when lighting, speed and reflectivity are controlled and when images connect to process actions.
Why projects fail
Plants often group scratches, stains, delamination and plating-boundary defects as “appearance defects.” Different causes and acceptance limits require different training data and responses. A large sample of acceptable variation is as important as defect images.
| Metric | Question | Practical target |
|---|---|---|
| Detection | Can it find the smallest critical defect? | Zero customer-critical escapes |
| False rejects | Does it flag normal gloss or oil? | Within reinspection cost limit |
| Speed | Does it work at production speed? | Within line takt |
| Traceability | Are image, coil, lot and time linked? | Immediate segment isolation |
| Governance | Who approves model changes? | Versioned validation record |
A 90-day pilot
- Weeks 1–2: classify critical, major and minor defects.
- Weeks 3–4: collect normal and defect images by lot, lighting and speed.
- Weeks 5–8: validate one product and one defect offline.
- Weeks 9–10: run alongside human inspection.
- Weeks 11–12: calculate escapes, false rejects, reinspection time and metal savings.
BIZPECTRA view
The starting point is not a vendor's “99% accuracy” claim. It is the balance between customer escapes and false-reject cost. Start with one or two defect types that create the largest claim or precious-metal loss.
Review AOI placement, precious-metal loss, traceability and customer inspection criteria for R2R and spot plating.
Explore CHEMNEX solutions