Start with optics and criteria
Stabilize defect definitions, lighting, optics, triggering and normal product variation before optimizing the AI model.
- Defect taxonomy
- Normal variation and process conditions
- Reproducible imaging conditions
Separate the roles of rules and AI
Use rules for explicit dimensional or geometric conditions and AI for complex surface and pattern variation.
- Decision evidence
- Exception handling
- Cost of false rejects and false accepts
Operate a feedback loop
Collect failure samples after deployment and update them through an approved change process.
- Dataset and model versions
- Approval and rollback
- Performance-drift review
