@CallumChan1056 Some labels came from the pictures alone. We still have the samples, but those labels aren't tied to physical inspection results. That's irritatingly circular.

Mark those labels uncertain and have the samples inspected properly. Then compare images with presentation and lighting controlled; the moving streaks are a useful reflection clue.

@CallumChan1056 That clue needs a limit: presentation can also change how a real scratch appears. Movement in the image isn't enough to reclassify the sample as acceptable.

You're right to qualify that. The streak behaviour motivates an imaging investigation; it doesn't replace physical classification under the agreed inspection procedure.

Should uncertain images be treated as acceptable, or kept somewhere separate while their status is established?

Keep them explicitly undecided, with the sample link. Don't use them as confirmed examples of either class until the physical review settles them.

Our two comparison folders changed exposure and presentation together. I can't tell which change made the images look better. That comparison needs a caveat too.

I had a separate inspection setup where the 'better lighting' folder used easier samples too. Its name implied a lighting conclusion that the comparison didn't support.

What did you keep beside each image? I can manage sample ID and settings, but I don't want a form nobody fills in.

@SarahBell0676 On my setup we kept sample ID, exposure, lighting arrangement and presentation condition. Enough to explain a comparison without reconstructing the whole session from memory.

Use those details to ask your vision specialist for controlled comparisons on representative surfaces. Keep the sample set fixed when assessing an acquisition change.

Would a learned model just cope with the lighting differences?

It still needs trustworthy labels and representative evaluation. A different model doesn't establish whether your reference scratches are real or your acquisition comparison is fair.

@CallumChan1056 Keep physical sample identity in mind when forming evaluation sets. Different images of a sample used for tuning don't provide an independent sample-level evaluation.

I'll keep all images of one sample together when we split tuning and evaluation sets. Our file names alone wouldn't have caught that overlap.

Use the held-out samples to report defect misses and false rejections independently. A combined agreement figure can conceal the type of error that matters to the process.

@CallumChan1056 Should undecided physical samples count as misses in that report? I don't want them quietly disappearing just because they're awkward.

Show them as undecided with their count, outside confirmed-class error rates. That keeps the coverage limit visible without inventing a ground-truth label.

That helped on my setup. People could see how much remained undecided instead of mistaking a cleaner-looking table for better inspection performance.