Machine Vision Inspection KPIs: Benchmarks for Detection, Escapes and False Rejects

Use these benchmarks as reference points for spotting a gap, not as targets. Your own trend on a fixed seeded panel says more about your station than any single published figure, so let the ranges below tell you which KPI to investigate first.

Most figures here have no public source and are marked no cited source: they are kept exactly as originally published because no public source stating them was found, so treat them as illustrative. The figures with a superscript number describe human inspectors and supplier quality ratings, which give useful context for judging a vision station.

At a glance

Detection rate

Detection rate is the headline KPI. Typical deployed systems land at 90 to 96 percent detection on their trained defect classes no cited source; world-class lines with mature datasets and controlled lighting reach 98 to 99.5 percent no cited source.

Below 90 percent, the system is usually catching easy defects and missing the subtle ones that matter no cited source. Measure it monthly against a fixed seeded panel of known defects, not against production drift, so the number stays comparable quarter to quarter and improvement is visible.

The human baseline

Human inspection is the usual comparison, and it is far from perfect. A 2012 literature review by Sandia National Laboratories notes that error rates of 20 to 30 percent are frequently quoted across many types of inspection task¹.

Detection in individual studies varied widely: 45 to 100 percent for soldering defects, 57 to 98 percent for magnetic particle inspection and 68 percent in aircraft visual inspection¹. Even on simple accept or reject tasks, the lowest human error rate appears to be on the order of 1 in 1,000¹.

Pace matters too. In one study the review cites, doubling the normal inspection rate raised misses from 23 to 30 percent, though easy defects were unaffected¹.

For automated inspection, the review describes a glass bottleneck inspection prototype that scored 98 percent hits with 2.4 percent false alarms on 120 test images¹. That is a lab result, and its false alarm rate sits far above the world-class false reject band later in this guide.

How to measure

• Run a fixed seeded panel of known defects through the station each month.
• Divide defects detected by defects present, for each trained defect class.
• Keep the panel unchanged so results stay comparable quarter to quarter.

What moves it

• Mature, well-labeled datasets.
• Controlled, consistent lighting.
• Added examples of the subtle defects the system misses.

Calculate detection rate (recall) from your seeded panel with the Vision Defect Detection Rate calculator.

Escape rate

Escape rate, the defects that slip through, is the KPI your customer feels. Automotive and medical targets sit near 50 to 200 defective parts per million shipped no cited source; a typical general industrial line runs 500 to 3000 PPM after inspection no cited source.

One published rating comes from aerospace, a different industry. RTX rates suppliers shipping at least 100,000 pieces over 12 months on rolling PPM: under 100 PPM is Performing, up to 250 PPM is Progressing, and above 250 PPM is Underperforming². It counts incoming and floor failures per million received, so it measures received quality at the customer.

The lever is threshold tuning combined with adding examples of the specific escaped defect classes, which usually costs a labeling cycle rather than new hardware. Track escapes by defect type, because one or two classes normally drive 80 percent of the total no cited source and deserve targeted data collection.

How to measure

• Count escaped defective parts per million shipped.
• Split escapes by defect type.
• Use a rolling window so a single bad lot does not swing the result.

What moves it

• Decision threshold tuning.
• Examples of the specific escaped defect classes added to the training data.
• Targeted data collection on the classes that dominate escapes.
• Full camera coverage of every required inspection zone.

Estimate what escapes cost each month with the False Accept Cost calculator, and look for blind spots in the layout with the Camera Coverage Rate calculator.

False reject rate

False reject rate is the KPI that quietly destroys margin. World-class stations hold false rejects at 0.2 to 0.8 percent no cited source; typical deployments sit at 1.5 to 4 percent no cited source, and a poorly tuned pilot can hit 8 percent no cited source.

Every point of false reject scraps good product at full value. The primary levers are decision threshold placement and better lighting consistency.

For example, moving from 3 percent to 1 percent false rejects on a 100,000 part line saves 2 percent of 100,000, or 2,000 good parts.

Report false reject and escape together, since they trade against each other. A threshold change that cuts one usually raises the other.

How to measure

• Confirm rejects on a sample basis to separate true defects from false rejects.
• Divide confirmed good parts rejected by all parts inspected.
• Report false reject rate beside escape PPM on the same page.

What moves it

• Decision threshold placement.
• Lighting consistency.

Estimate the monthly cost of good parts thrown away with the False Reject Cost calculator.

Station uptime and cycle time

Station uptime and throughput adherence tell you whether the cell is production-grade. Target 97 to 99 percent uptime no cited source; below 95 percent, false reject jams and lighting drift are usually stealing availability no cited source.

Throughput adherence is measured cycle time versus takt. World-class cells hold cycle time within 5 percent of the design ceiling no cited source, while struggling cells drift 15 to 30 percent slower as inference queues build no cited source.

Watch cycle time at the 95th percentile, because the slow tail is what actually stalls the conveyor.

How to measure

• Log station downtime by cause, including false reject jams and lighting drift.
• Compare measured cycle time with takt and the design ceiling every shift.
• Report the 95th percentile cycle time, which shows the slow tail.

What moves it

• Removing the causes of false reject jams.
• Stable, maintained lighting.
• Enough inference capacity to keep queues from building.

Downtime and slow cycles erode the business case; check it with the Machine Vision ROI calculator.

Model drift

Model drift is the KPI that separates a project from a program. A healthy line sees detection rate degrade less than 1 percent per quarter no cited source; unmanaged models can lose 5 to 10 percent within six months no cited source as materials, tooling wear and lighting age shift the image distribution.

The lever is a scheduled retraining cadence, typically quarterly no cited source, fed by a captured sample of recent false rejects and escapes. Set an alert when weekly detection on the seeded panel drops more than 2 points no cited source, and trigger retraining rather than waiting for a customer complaint.

How to measure

• Run the seeded panel weekly and chart detection rate over time.
• Mark material, tooling and lighting changes on the same chart.

What moves it

• A scheduled retraining cadence.
• A captured sample of recent false rejects and escapes for each retraining round.
• An alert on the weekly seeded-panel result.

Track the weekly panel result with the Vision Defect Detection Rate calculator.

Data and labeling health

Data and labeling health are leading indicators that predict the outcome KPIs. World-class programs keep at least 200 to 500 labeled examples per active defect class and label agreement above 90 percent between annotators no cited source.

Typical programs run thin at 50 to 100 examples on rare classes and see agreement near 70 percent no cited source, which caps detection rate no matter the model.

Audit a 200 image sample quarterly for label consistency. Inconsistent ground truth is the single most common reason a detection rate stalls below target.

How to measure

• Count labeled examples for each active defect class.
• Measure agreement between annotators on a shared sample.
• Audit a sample of labels for consistency every quarter.

What moves it

• Targeted collection for rare defect classes.
• Clear labeling rules and regular annotator calibration.

Measure the recall each labeling round buys with the Vision Defect Detection Rate calculator.

Cost-side KPIs

Cost-side KPIs keep the program honest without re-running the full cost model. Track cost of quality per million parts and the ratio of false reject scrap to escape cost.

A balanced line keeps that ratio between 0.5 and 2 no cited source, and a ratio above 3 means the threshold is set far too aggressively no cited source.

Also track labeling hours per point of detection rate gained. It typically climbs from 20 hours per point early to 200 hours per point near 99 percent no cited source, signaling when further data collection stops paying off.

How to measure

• Cost of quality per million parts inspected.
• False reject scrap cost divided by escape cost, each month.
• Labeling hours spent per point of detection rate gained.

What moves it

• Threshold placement, which trades false rejects against escapes.
• Stopping data collection once labeling hours per point stop paying off.

Price both sides of the ratio with the False Reject Cost calculator and the False Accept Cost calculator, and the whole business case with the Inspection Automation Payback calculator.

How to build your own baseline

Turn the benchmarks into a scorecard and review it monthly. Put detection rate, escape PPM, false reject rate, uptime, cycle time at the 95th percentile and quarterly drift on one page, with the reference bands from this guide beside each actual. The Vision Defect Detection Rate calculator, the False Reject Cost calculator and the False Accept Cost calculator supply the measured inputs.

Any KPI outside its typical band gets a named owner and a lever. The review closes when the trend, not just the single month, moves back toward the world-class column.

• Write down each KPI's definition and the seeded panel you measure against, and keep both fixed.
• Collect several months of data before you set targets.
• Compare against your own trend first and published ranges second.
• Record every model, lighting, material and threshold change so a KPI move can be traced.