Two Ways an Inspection Configuration Can Be Wrong
An inspection configuration is a decision rule, and a decision rule can fail in two directions. It can reject a good part - a false reject, overkill in shop language - the kind of error that eats good production and turns operators against the vision inspection machine. Or it can accept a defective part - a false accept, the escape that reaches your customer and undoes the reason you bought the machine. No false-reject or false-accept rate is published on this site for these machines - and a brochure detection rate quoted for any machine type, including the flying-camera robot's, is not a number measured on your compound, your sizes or your defect families. What this article covers is the structure underneath those numbers: how the two error directions trade against each other, and how the sample test is built to show you both - before any capital decision. The sample test process itself is in the sample test walkthrough (/posts/vision-inspection-sample-test).
How the Sample Test Measures Each Direction
The published package asks for good, bad and borderline parts - and each error direction gets its own evidence:
• The false-accept direction is probed by your worst-case defective parts: parts carrying each defect family you care about, at the severity you cannot tolerate. What the configuration catches - and what it lets through - is the detection half of the result, returned as images and an OK/NG decision log. That log is also the raw data for your own error math: counts of defective parts passed and good parts rejected are sitting in it.
• The false-reject direction is probed by known-good parts - and the caveat in the sampling article (/posts/sampling-plans-vs-full-inspection) is why good parts from more than one production lot belong in your package: good parts that get rejected show you where the thresholds are tighter than your process can live with; borderline parts show you where the decision line actually sits, as opposed to where you assume it is.
One warning that falls out of this structure: a sample test run with only clean, current-lot good parts cannot measure the false-reject direction well - good parts that all look identical tell you nothing about whether next month's lot will pass. The multi-lot recommendation is not padding; it is the measurement. And if you only have one lot, say so plainly - the package still measures the detection side, but the false-reject side will be under-measured.
The False Reject / False Accept Trade-Off
The two errors are not independent. Move the thresholds tighter and you catch more marginal defective parts - and reject more marginal good ones. Loosen them and good production flows - and marginal defects start passing. Every configuration sits somewhere on that line, and where it should sit is a question about your costs, not about the machine: what does a false reject cost you (re-screening labor, scrapped good parts, line stops), and what does a false accept cost you (a claim, a containment run, a customer audit finding)? The containment workflow (/posts/re-screening-rejected-lot-containment) is what a false accept turns into; chronic over-rejection is what turns a floor against the machine.
What the Buyer Controls
Three inputs decide where on the line your configuration lands, and all three are yours:
• The defect specification. Fuzzy families ("looks bad") force loose thresholds; crisp families with severity levels let the configuration be precise. The method is in defining defects for OK/NG inspection (/posts/defining-defects-ai-visual-inspection).
• The limit samples. Boundary/limit samples - your physical definition of "still acceptable" and "must reject" - are what the decision line gets tuned against. Without them, the line is a guess wearing a number.
• The borderline policy. Parts between the lines need a rule - hold for manual review or reject - set before production runs, not improvised mid-shift.
The Boundaries
• No false-reject rate is published on this site, and the balance shifts with your lot variation - multi-lot good parts are recommended precisely because next month's parts will not look identical to this month's. After a changeover (/posts/vision-inspection-machine-changeover), re-confirm the balance on the new part's samples rather than assuming it carried over.
• The trade-off only exists for defects cameras can see. Internal voids, same-color mixing and material-level properties have no threshold to tune - they are outside the trade-off entirely, which is why the purchase-boundary guide (/posts/when-not-to-buy-vision-inspection-machine) says to check how much of your risk is surface-visible first.
• "Zero false rejects and zero false accepts" is not a configuration; it is a description of a perfect decision rule, which does not exist on any machine. The honest goal is a balance measured on your parts and matched to your costs.
Frequently Asked Questions
Can we get zero false rejects?
No configuration guarantees that. Tightening thresholds toward zero false rejects pushes errors into the other direction - more marginal defective parts passing. The measurable goal is a balance demonstrated on your parts at the sample test.
Which direction should we tune toward?
The direction whose errors cost you more. A safety-critical seal family argues for the strict side; a cosmetic family on high-volume low-value parts argues for the loose side. Weigh your costs on each side; the configuration moves accordingly.
My NG logs look clean - doesn't that mean false accepts are under control?
A clean NG log shows the configuration caught what it was tested for - it cannot show what it missed. Escapes surface as customer claims; that is the containment scenario covered in re-screening a rejected lot (/posts/re-screening-rejected-lot-containment).
How does the sample test measure both directions at once?
By package design: worst-case defective parts for the detection side; good parts - ideally from more than one lot - and borderline pieces for the reject side. All of it runs through the same configuration, and the images and OK/NG decision log come back for each part.
Next Step
Assemble the sample-test package that measures both directions: worst-case defective parts per family, good parts - worth including from more than one lot - and your borderline pieces with the limit samples marked. Send them with your defect specification; confidentiality, test cycle and logistics for your samples are confirmed when you first contact the team. The images and OK/NG decision log will show you exactly where the decision line sits on your parts. Contact the team (/contact-us) - Monica (+86 138 5726 5705, WhatsApp) or Arya (+86 173 2430 9735), or email monica@aivisioncheck.com.