FR5 bezel inspection proposal, not commissioned. Bright streaks move between presentations but get scratch labels. I need physically grounded references before tuning.
Have quality assess identified bezel regions and retain those parts. Record the present camera, light and holder arrangement, then repeat presentation of the same examples. Moving image features are evidence of variation, not by themselves proof that the surface is acceptable or defective.
You can still record repeat-presentation behaviour while those decisions are pending. Keep that diagnostic work separate from threshold selection, and include both confirmed defects and acceptable variation once the reference decisions arrive.
Who is keeping the parts and their image links together? I'd put that responsibility in the trial handover now. Otherwise the next person can inherit the same pictures with better filenames but no way to find the physical bezel.