Machine vision for quality: the honest buyer's guide

What it does and what it doesn't

A camera (or several) looks at every part, a model compares it with what it learned as "good", and it rejects or alerts when it sees a defect: scratches, porosity, missing material, out-of-tolerance colours, misplaced labels, faulty welds. At line speed, without getting tired and without having a bad day.

What it does NOT do: understand the process, decide whether the defect matters, or fix anything. It detects the visible. If your quality problem can't be seen in a photo, machine vision is not your tool.

When it pays: the maths

A serious inspection point costs 5,000-25,000€ today (camera, lighting, integration and model - lighting, which nobody mentions, is half the battle). It pays when:

• The cost of escapes is high: customer claims, returns, rework, manual sorting. Add it up per year: if it clearly exceeds the system cost, there's a case.
• You already pay for human inspection: a person watching parts 8 hours a day costs ~25,000€/year and catches 70-85% of defects. The machine catches 95-99% and never blinks. If someone does this job, the maths usually does itself.
• The defect is visible and repeatable: always in the same place, good contrast, and you can show hundreds of good and bad examples.
• The pace is stable: same part or part family, same presentation to the camera.

It doesn't pay when: the defect changes shape every week, you make 50 parts a year of each reference, or your real problem is that the process isn't capable (fix the process first, then).

The 8 questions to ask the supplier

1. How many good and defective parts do you need for training, and who collects and labels them? (That work is usually yours.)
2. What false-reject rate does the system have on MY part? A 3% false-positive rate at 1,000 parts/day is 30 good parts thrown away daily.
3. Who re-trains the model when I change reference or material, and how much does that cost?
4. Does the lighting survive my environment (dust, vibration, changing workshop light)?
5. What happens to the doubtful part: rejected, diverted to human review, or just flagged?
6. Can I see it running in a plant similar to mine, with my type of defect?
7. Is the system mine or a subscription? What happens if I stop paying?
8. What guarantee on the detection rate, in writing?

If the supplier squirms at these questions, you already have your answer.

The right pilot

Don't buy the whole line. A serious pilot: ONE inspection point, YOUR most expensive defect, 4-8 weeks in parallel with current inspection, measuring real escapes and false rejects. If the pilot doesn't prove the promised rate on your parts, there is no project. Any serious supplier accepts this; the one who only sells "the full project" is selling risk - yours.

And keep the pilot data: every inspected part is a quality data point you didn't have before. That's where your real defect Pareto starts, and that's worth as much as the detection itself.