Updated July 2026 · 6 min read · Adente Vision Engineering Team
What is a false positive in inspection?
A false positive in inspection is a good part that the system rejects as defective. It is the opposite error to a false negative, or escape, which is a defective part the system passes as good. Both are inspection mistakes, but they fail in opposite directions: a false positive throws away something that was fine, while a false negative lets something bad through.
The vocabulary matters because the two errors have very different owners. A false negative is paid downstream, by the customer who receives the defect. A false positive, also called a false reject or over-rejection, is paid inside your own plant, quietly, as scrapped good parts and extra handling. Because it never leaves the building, over-rejection is the error that hides in the yield number instead of showing up as a complaint.
Why is over-rejection so costly?
Over-rejection is costly because every false reject is good product destroyed or delayed, plus the work of handling it. A rejected good part gets scrapped, or it gets pulled for manual re-inspection, which adds labor, buffers and sometimes a line stoppage. Do that on a small percentage of a high-volume line and the loss compounds into real yield erosion, all of it invisible to the customer and easy to miss in a headline accuracy figure.
There is a second cost that is harder to measure and just as damaging: trust. When operators see the inspector reject parts they can see are good, they start to override it, wave parts through, or ask for the threshold to be loosened. Once the check is treated as a nuisance rather than a gate, its real defects get waved through too. Over-rejection therefore does not just burn parts; it can quietly disable the inspection it is supposed to provide.
Why does chasing zero escapes spike false rejects?
Chasing zero escapes spikes false rejects because both errors are controlled by one setting: the decision threshold. An inspection model outputs a score for how abnormal a part looks, and the threshold is the line that turns that score into a pass or fail. Move the threshold to catch every possible defect, and you also start rejecting borderline-good parts, so escapes fall while false rejects climb. Loosen it to stop scrapping good parts, and some defects slip through. You cannot push one error to zero without paying in the other.
That is why a single number like accuracy hides the real behaviour, and why F1-score is the honest measure: it balances precision, how many rejects were truly bad, against recall, how many defects were caught. Adente Vision is an edge-AI visual inspection unit built by Adente AI, part of the Aden Group, sold through automation system integrators. On a live cap-inspection line, a delivered installation reached a 99.65% F1-score with a 0.69% false-negative rate: a high F1 means the system caught escapes without over-rejecting, because F1 only stays high when both errors are low at once.
| Inspection outcome | What it means | Who pays |
|---|---|---|
| True positive | Defect correctly rejected | Nobody, the check worked |
| False positive (false reject) | Good part wrongly rejected | You: scrap, rework, re-inspection, lost yield |
| False negative (escape) | Defect wrongly passed | Your customer: field failure, recall, warranty |
| True negative | Good part correctly passed | Nobody, the check worked |
How do you set the threshold to balance both errors?
Set the threshold to your real cost of quality, not to a slogan like zero defects. The right operating point depends on how much an escape costs you versus how much a scrapped good part costs, and those are rarely equal. A safety-critical part may justify some over-rejection to keep escapes near zero; a low-margin, high-volume part may not tolerate the scrap. The decision is a business one, expressed as a threshold.
Two things make that balance easier to hold. First, a good reference set: a model trained on good parts that span the real range of acceptable variation will not flag normal spread as a defect, which keeps false rejects down without loosening the threshold. Second, an honest metric: track F1, precision and the false-negative rate together on your own parts, so you can see both errors move as you tune, rather than optimising one blind. For the escape side of the same trade-off, see the sibling post on the false-negative rate. For the full method, see the pillar guide, and for where this fits real line problems, see the applications page.