Building a defect catalogue
Scratches, missing parts, incorrect placement and colour variation may need different methods. Each defect needs examples around acceptable and unacceptable boundaries. Record inspected surfaces, field of view and unseen regions. Annotated examples help teams share a defect definition and reduce inconsistent labels.
Rule-based method or learned model?
Thresholding or contours may suffice to check a consistently lit part for presence. A learned approach may suit a defect with highly variable appearance. Compare methods on the same independent samples. Before selecting a more complex model, examine whether improving image acquisition reduces the errors.
Connecting results to the production flow
Associate each decision with an item identity and timestamp. Verify that a rejected item was actually separated, and consider a review queue for uncertainty. Track rejection rate, missed defects and camera faults separately. Reporting only accepted samples is misleading; disclose test coverage and defects outside the visible scope too.