Current progress and boundaries
Machine-vision inspection is mature at some fixed workstations, but often needs retuning when products, cameras, or materials change.
Why it remains difficult
Defect samples are scarce and definitions depend on the process; glare, occlusion, and equipment drift can cause false alarms. The model must also integrate with existing traceability systems.
Paths worth exploring
Start with one defect type and a fixed camera position. Keep original images, decision rationales, and human labels, and provide manual review for low-confidence results and a quick rollback mechanism.
How to evaluate success
Report missed-defect and false-alarm rates by defect type, review time, and performance degradation after a line change. Do not report overall accuracy alone.