False positive rate in AOI is the metric that quality engineers argue about the most, and the one most often miscalibrated in the initial system setup. A 1% FP rate looks like a rounding error in a performance report. At 200 ppm on a two-shift operation, it's 48 good parts per hour going into the reject bin — 768 parts per shift, 1,536 parts per day. Whether that matters depends entirely on what those parts cost and what the manual re-inspection process looks like. This post is about building the actual cost model rather than accepting that 1% is or isn't acceptable in the abstract.
The direct cost: part value and rework labor
The most straightforward component of false positive cost is the value of the falsely rejected part multiplied by the volume. But this calculation requires knowing what happens to parts that go into the reject bin — and most operations have three possible outcomes:
Re-inspection and return to production. A quality technician reviews the rejected parts, confirms they're actually good, and releases them back to production. This is the most common outcome for high-value parts. The part itself isn't lost, but the labor cost is real: at typical quality tech labor rates, reviewing 768 false rejects per shift consumes 2–4 hours of technician time depending on how thorough the confirmation process is. On a two-shift operation, that's a dedicated resource.
Scrap without re-inspection. For low-value parts where the manual re-inspection cost exceeds the part cost, the reject bin goes to scrap directly. Here the entire part value is lost. For a $0.40 stamped bracket, the scrapping threshold might be reached quickly. For a $3 die-cast housing with a metal insert, it almost certainly isn't.
Accumulation in a pending bin. The most operationally damaging outcome: false rejects pile up in a "maybe" bin waiting for someone to have time to review them. This delays production if the rejected parts are needed for a downstream assembly step. We've seen stamping lines where false positives were backlogging a week's worth of production in the pending queue before anyone addressed the FP rate calibration problem.
The indirect cost: throughput loss and operator trust
False positive rate has a non-linear effect on throughput when it erodes operator trust in the vision system. When operators see that the system is calling out good parts, they start overriding rejects — manually releasing parts from the reject gate or disabling the system during high-priority runs. Once operators develop the habit of second-guessing the AOI, the system's actual false negative performance becomes irrelevant because the escape rate climbs toward whatever it was before deployment.
This trust erosion is one of the main reasons AOI deployments fail in practice — not because the detection performance is poor, but because the false positive rate was high enough to generate constant friction without being high enough to obviously break the economics. Quality engineers who deploy AOI systems often focus on detection rate as the primary success metric and treat FP rate as a secondary concern. In practice, for a production line where operators interact with the reject system daily, FP rate has a larger influence on long-term system adoption than detection rate does.
We're not saying false negatives don't matter — they're obviously the primary quality failure mode. The point is that a system with 99% detection rate and 5% FP rate will often perform worse in practice than a system with 96% detection rate and 0.3% FP rate, because the former will be disabled or bypassed within a month of deployment.
How to calculate your actual FP cost per shift
The model has four inputs:
- FP rate (% of good parts rejected) — measured from your system, not assumed
- Good parts per shift — line speed × run time × (1 − true defect rate)
- Re-inspection cost per false reject — labor time for technician review × labor rate, or part cost if scrapping without review
- Throughput opportunity cost — value of the production time consumed by handling the reject queue
Take a concrete example: a 150 ppm line, two 8-hour shifts, 95% uptime, 0.5% true defect rate. That's approximately 136,800 good parts per day. At 1% FP rate, 1,368 false rejects per day. If each false reject requires 45 seconds of technician time to confirm and release, that's 17 hours of technician time per day — more than two full-time equivalents. At $28/hour blended rate for quality labor, that's $476 per day, roughly $120,000 per year in inspection labor alone.
For the same line, cutting FP rate from 1% to 0.15% — a realistic target for a well-trained Procunit model on a stable line — reduces that labor cost by 85%, to about $18,000 per year. The model performance investment that achieves that improvement pays back in labor cost within months on lines of this volume.
What drives false positive rate on stamping lines specifically
On stamped metal lines, the most common source of high FP rates is illumination-induced variation that the model interprets as defects. The three main culprits:
Specular reflectance variation. Die-formed stampings have variable surface angles across the part face. Depending on part orientation on the conveyor, regions of the part will reflect the ring light at different intensities. If the part rotates slightly between captures — due to conveyor vibration or part-to-part stacking variation — what appears to be a surface anomaly in one frame is a normal reflectance pattern in the next. This is why part fixturing at the inspection station matters significantly for FP rate control on reflective metal parts.
Coolant or lubricant residue. Stamping operations use lubricants that leave residue on the part surface. If the lubricant distribution is uneven — which it typically is — regions with heavier lubricant film reflect light differently than dry regions. This creates apparent surface anomalies that correlate with lubricant distribution rather than actual defects. The fix is either better lubricant consistency upstream or including parts with representative lubricant variation in the training set so the model learns to ignore it.
Die wear-related surface texture variation. As tooling wears, the surface texture of the shear zone and the formed surfaces changes gradually. A model trained early in the die's life will see increasing anomaly scores as the die wears, even if the parts are still within specification. This is the model calibration drift problem — the model's reference for "normal" is now misaligned with the actual current production baseline.
Setting a target FP rate for your application
The right target FP rate is application-specific and is driven by the cost model above, not by a universal benchmark. For a high-value assembly with a 1% true defect rate, a FP rate of 0.2% is reasonable — the ratio of false rejects to true rejects is 1:5, which is manageable for a re-inspection workflow. For a low-value commodity stamping with a 0.1% true defect rate, a 0.2% FP rate means the reject bin contains twice as many good parts as defective parts — which typically doesn't meet the economics threshold for automated rejection without human review.
As a general calibration starting point, FP rate should be below the true defect rate by at least a 3:1 margin before committing to automated hard rejection without a re-inspection gate. Above that ratio, the labor cost and operator trust dynamics start working against you faster than the detection value is being captured.
The calibration lever is the detection threshold. A lower threshold improves detection rate at the cost of higher FP rate; a higher threshold reduces FP rate at the cost of detection rate. The right operating point depends on your escape cost versus your re-inspection cost — and that calculation needs to be done explicitly, not assumed from a vendor's default configuration.