AI Visual Defect Inspection: Feasibility, Training Data and Acceptance Tests

Agree the defect catalogue

Collect examples of acceptable variation, definite defects and borderline cases. Define each class with the quality team, including minimum size and inspection region. Separate cosmetic criteria from functional tests. Where inspectors disagree, resolve the acceptance rule before treating their labels as ground truth for model training.

Capture representative evidence

Include real materials, suppliers, shifts, lighting variation and production speeds. Keep related images of the same part or batch together when splitting training and evaluation data to avoid overly optimistic results. Rare defects need a deliberate sample strategy; synthetic examples should not be the sole proof of production performance.

Design the inspection decision

Define pass, reject and uncertain routes. If a reject actuator is used, verify timing, item tracking and feedback that the correct item was removed. Loss of image quality must have an agreed response. Keep records linking the decision to product identity and model version where traceability is required.

Validate by defect and consequence

Report missed defects and false rejects separately for each important class. Test an unseen production sample and review difficult results with quality staff. Agree acceptance criteria before tuning the final model. Monitor drift after changes to materials, tooling or lighting and define when revalidation is required.

FAQ