I'm sorting test images for a UR5e faceplate inspection station. A bright line shifts when the faceplate turns, but the folder calls every version scratched. Nobody has shown me a scratch on the actual plate. Where should we start before anyone tunes a threshold?
Get the plate inspected under the product's agreed appearance check. Keep its identity with the pictures. I couldn't approve a reference set from somebody's guesses about moving bright lines, even if the folder names sound certain
The person labelling them thought anything bright was a defect. We've kept the faceplates, so quality can go back to the actual surfaces. At least we haven't thrown the answer away.
Then ask quality to mark which area each decision refers to. A whole-plate good or bad label won't help much if the photograph's bright line and the real mark are in different places
Was anyone checking the part before taking the pictures? We get boxes labelled bad with no explanation, then somebody expects the operator to know whether bad means scratched, bent or just dirty. Makes for a rotten training example.
That will decide how many examples you can rescue. Unmatched ones may still illustrate a lighting problem, but I wouldn't give them a physical defect label on that basis
Most match through the sample sheet. Three don't, so they're out of the labelled set. Quality found a real nick on one plate, nowhere near the bright line everyone argued about.
This sounds familiar from the cover thread I was in recently. We had to sort the physical marks before the lighting discussion made sense. Have you checked whether the plate seats the same way each time?
Not yet. I can rock it slightly in the presentation fixture. I've asked the fixture person to check that before we blame the camera for every changed image.
Useful thing to investigate next. Keep the nicked sample available for the imaging trial too, so improved consistency doesn't accidentally mean a light setting that hides the mark you need to see