Catching defects at the line, not at the customer

Visual inspection depends on who is on shift. Here is how a camera pilot proves, on your own parts, whether software can catch defects reliably before anyone buys hardware.

The problem

Parts are checked by eye. Clear display cases show every scratch, flow line and speck, and what one inspector rejects another passes. Some defects only show up when the customer opens the box.

What was really causing it

Inspection is a human skill done under time pressure, with no record of what was checked. Without records, nobody can tell whether a defect was missed or never seen.

Inspection is a human skill done under time pressure, with no record of what was checked.

What the research says

• 85% of defective parts correctly rejected by trained inspectors in a controlled study (J. E. See, Human Factors (Sandia National Laboratories), Visual Inspection Reliability for Precision Manufactured Parts)
• 35% of good parts wrongly rejected in the same study (J. E. See, Human Factors (Sandia National Laboratories), Visual Inspection Reliability for Precision Manufactured Parts)
• 99.6% image-level AUROC for a published defect-detection method on the MVTec AD benchmark (K. Roth et al., CVPR 2022, Towards Total Recall in Industrial Anomaly Detection (PatchCore))

The Sandia result is the uncomfortable one: careful, trained people missed one defect in seven and threw away a third of the good parts. The benchmark result says software can be very good on a well-defined dataset. The gap between the two is exactly what a pilot on your own parts, in your own lighting, is for.

What we built

A pilot first, not a production line. We collect and label images of good and bad parts, train a model, and measure how often it catches real defects and how often it rejects good parts, on your parts, in your lighting. Clear parts are hard to photograph well, which is exactly why the pilot comes first.

1. Collected Photos of good and defective parts from real production.
2. Labelled Your quality team marks the defects. Their judgment is the standard.
3. Tested The model is measured on parts it has never seen: caught defects and false alarms.
4. Decided Only if the numbers are good do we plan cameras, lighting and the line.

What stays as it is

Your inspectors. In the pilot the software assists them; it does not replace their sign-off.

Where AI fits, and where it does not

This is AI's job: recognising defects in images. The decision to ship stays with quality. Cameras, lighting and safety on the line may need a hardware partner, and we say so at the start.

What it is worth

What comes next

A review screen for flagged parts, defect trends by press, mould and shift, and links into your quality records.

Questions people ask

Do we have to buy cameras first?

No. The pilot starts with photos of real good and bad parts, taken carefully. Cameras and lighting are planned only if the numbers justify them.

Will it replace our inspectors?

Not in the pilot. The software flags parts and the inspector decides; the quality team's judgment is the standard it is measured against.

How many defect samples do we need?

Fewer than most expect, because some methods learn mainly from good parts. The pilot tells you how many your defects need.

1. Visual Inspection Reliability for Precision Manufactured Parts, J. E. See, Human Factors (Sandia National Laboratories) (2015)
2. Towards Total Recall in Industrial Anomaly Detection (PatchCore), K. Roth et al., CVPR 2022 (2022)
3. Annual Report on the U.S. Manufacturing Economy: 2025 (NIST AMS 100-76), D. S. Thomas, NIST (2026)

Canonical URL: https://ugibo.com/case-studies/catching-defects-at-the-line