The vision system passes 99.4 percent of parts. The 0.6 percent it rejects get scrapped without review. Then a customer finds a bad part that passed. The vision system has a false negative, and nobody caught it because the rejects went straight to the bin.
False passes cost more than false rejects
A false reject costs you one good part. A false pass costs you a warranty claim, a customer complaint, and a recall. Most vision lines tune the threshold to minimize rejects, which means they’re letting marginal parts through. The false pass rate is invisible because nobody checks the parts that passed.
Sample 1 in 100 accepted parts for manual review. If the sample is clean, you’re fine. If you find a bad part, the threshold needs tightening. This is the only way to know the false pass rate. The vision system can’t tell you; it’s the part the system got wrong.
Rejects need a review loop
Don’t scrap rejected parts immediately. Route them to a review station where an operator looks at the image and decides. The image is already captured. The operator clicks good or bad. Over weeks, you’ll see which defects the system over-flags and which it misses. Use that data to tune the threshold.
If rejects go straight to the bin, you have no feedback. The system runs at whatever threshold it was tuned at during install, and nobody knows if it’s still right a year later.
The lighting drifts
Lamps dim. Camera sensors collect dust. The machine vibrates slightly, shifting the working distance. A system tuned at install drifts over months. The threshold that was correct on day one starts rejecting good parts on month six. This is why ongoing review matters more than the initial tune.
Run a golden part through every shift. If the system reads it differently today than last week, something changed. Check the light, the lens, and the fixture. Don’t wait for the reject rate to climb.
Log the images
Every rejection should save the image to a folder. When a customer complains, you can go back and see what the system actually saw. Without the image, you’re guessing whether it was a real defect, a lighting change, or a fixture shift.
Storage is cheap. Keep rejected images for 90 days. If a defect type recurs, you have a pattern to fix.
Bottom line
A vision system is a tool, not an inspector. Tune the threshold with real reject data, sample accepted parts for false passes, and review rejects instead of scrapping them. The system gets better when you close the loop, not when you install it and walk away.
