Our good faceplate fails when someone wears a white shirt

We tried a temporary dark screen and the long mark disappeared. A finer line near the screw opening remains. I don't want to assume both have the same cause.

Keep that distinction. Can quality find the fine line on the actual piece?

Mark its location on a copy of the image, not on the inspection surface itself.

Yes. A pen mark would give your classifier an impressively easy job.

We had operators move a screen to reach the cleaning kit, then wonder why inspection changed after lunch. Who needs access around this station once it's running? The final enclosure needs to survive ordinary work.

Good point. A test screen proves an effect, not a usable enclosure.

Access is from the rear for lens cleaning. Quality has now looked at the fine line and called it acceptable finish variation, not a scratch. We've kept that part as an accepted reference with the line location recorded.

Accepted variation. Have you got an actual rejected part as well? Easy to make all the good ones pass.

Yes, one with a handling scuff across the centre. It was rejected independently before this trial. We have photographed it in the same static setup and the scuff appears faint compared with the acceptable line.

Then brightness alone sounds a poor judge, at least with that lighting.

Agreed. Don't tune until that scuff is reliably visible.

Is the accepted part and scuffed part comparison being kept together in the handover? Someone arriving later could see only the cleaner background and think the inspection problem was solved. We've managed that particular misunderstanding ourselves.

Label the trial result as reflection control, with detection still unproven.

Try the lighting options on those physical samples before pricing replacement cameras, this still looks like an image formation problem.

Both parts are going into the imaging trial with the vision supplier. I have put the rear cleaning access on the enclosure sketch as well. The temporary screen won't be handed over as the finished solution.

Include repeat presentation too; one flattering image doesn't establish detection.

That's in the request. We're asking them to show accepted variation and the rejected scuff across repeated presentations, with the imaging setup documented. No threshold has been selected yet.

That gives them a concrete problem to solve.

Curious what they come back with. Our shiny parts still take more discussion than the complicated machined ones.