What is computer vision quality inspection?

Computer vision quality inspection uses cameras and trained models to check every unit at line speed against a learned standard, replacing sampling plans and manual visual checks with continuous automated detection.

Sampling was always a compromise. Inspecting one part in five hundred catches systemic defects and misses intermittent ones entirely. Vision systems inspect all of them, and they do it at the same standard on the third shift as on the first.

When does deep learning beat rule-based machine vision?

Rule-based vision wins on rigid geometry with tight tolerances: presence checks, measurement, barcode reading, and alignment. Deep learning wins where the good part varies naturally and the defect does not have a fixed shape.

Welds, castings, textiles, food, wood grain, and painted surfaces all sit in the second category. An engineer cannot write a threshold for “acceptable grain variation” but can show a model two thousand examples. Cognex, Keyence, Landing AI, Instrumental, and Elementary ship tooling built around labeled examples rather than hand-written rules. Many lines run both, with rules handling measurement and a model handling cosmetic judgment.

How many defect images do you need to train a model?

Fewer than teams expect for detection, more than they expect for rare defect classes. A working starting point is 100 to 300 labeled examples per defect class, with the good-part set several times larger.

Class imbalance is the harder constraint. If a defect occurs once in ten thousand units, collecting 200 real examples takes months. Seed the line with known-bad parts, mine quarantine and scrap history, or augment synthetically. Label consistency beats label volume. Two inspectors who disagree on borderline parts produce a model that disagrees with itself.

How do you manage false rejects and false accepts?

Set the operating point from the cost asymmetry. A false accept that reaches a customer usually costs far more than a false reject that costs one part and one re-inspection, so most lines tune toward over-rejecting and add a human adjudication station.

Track both rates weekly against a fixed holdout set. A rising false-reject rate is usually the first visible symptom of drift, and it appears before quality escapes do. Give operators a one-click way to flag a wrong call. That feedback becomes next quarter’s training data.

What breaks a vision model in production?

Lighting changes, fixture wear, camera drift, and new supplier material. All four alter the image without altering the part, and the model has no way to know which happened.

Lock lighting and fixturing as controlled parameters under the quality management system, not as maintenance items. In FDA-regulated production an automated inspection decision is a quality record, which brings 21 CFR Part 11 for electronic records and the design controls referenced through ISO 13485:2016 into scope. Under ISO 9001, IATF 16949, or AS9100, the vision system is a measurement system and needs the same validation and change control as a gauge.

What to do next

Pick the single station with the highest scrap or rework rate and pull three months of defect photos from quarantine records. If you can sort them into three or four consistent defect classes with 100 or more examples each, you have a trainable first use case.

Read next: AI for Manufacturing Operations: Quality, Uptime, and Safety

The post Computer Vision for Automated Quality Inspection appeared first on Scadea Solutions.