After sorting the cover labels in my earlier thread, I've been handed a UR5e trim-plate image set with the same problem: bright marks move between presentations, but every image says scratch. Physical samples retained, at least
Start by having quality inspect the retained plates and connect each result to its images. Then compare the lighting and presentation used for the captures. Are the rejected examples independently identified, or did all those labels originate from looking at the screen?
All screen labels. Quality has now accepted two plates and held one for closer inspection. I have removed the unconfirmed scratch labels from the working reference set
Keep that held sample identifiable while they decide. For the accepted pieces, document the capture conditions and see whether you can repeat the images. The lighting setup needs to survive routine production use, not only one careful photograph.
Plate repositioned, lamp reportedly fixed, but no record of the surrounding setup. I'm arranging another capture session rather than guessing all that from the old thumbnails
Quality accepted a cover I had labelled damaged too. Did they finish examining your held plate? Keep that decision separate from how clear its next photograph looks
Did the new session give you repeatable presentation? That will help determine how much the imaging changes when the robot eventually presents these parts.
Held plate was rejected for a real edge scuff. New static captures are repeatable with the background screened, but the scuff is faint. Supplier is testing lighting options; we're still short of reliable defect detection