Updated July 2026 · 7 min read · Adente Vision Engineering Team
Why does 99% accuracy hide a bad inspector?
Accuracy is the share of parts an inspector classifies correctly, good or bad. On a line where defects are rare, that single number flatters almost any inspector, because most parts are good and calling a good part good is easy. Take a line running 10,000 parts with 100 real defects, a 1% defect rate. An inspector that passes every part without looking is right about 9,900 times and scores 99% accuracy, yet it catches zero defects and lets all 100 escape.
The headline number looks excellent and the inspector is useless. Rare defects break accuracy because the metric is dominated by the many good parts, not the few bad ones that decide whether the line is safe to ship. The rarer the defect, the more accuracy overstates the inspector, which is exactly the regime most quality lines operate in.
What do precision, recall and F1 actually measure on a line?
Four counts describe any inspector: true positives (defects caught), false negatives (defects missed, the escapes), false positives (good parts wrongly rejected), and true negatives (good parts correctly passed). Precision and recall each read a different pair of those counts, and each answers a question a quality manager cares about.
Recall, also called the catch rate or sensitivity, is the share of real defects the inspector catches: caught defects divided by all real defects. Low recall means escapes reaching the customer. Precision is the share of rejected parts that were actually defective: caught defects divided by everything the inspector rejected. Low precision means false rejects, good parts thrown away.
F1-score is the harmonic mean of precision and recall, a single number from 0 to 100% that is high only when both are high. Because it is a harmonic mean, one weak side pulls it down hard, so you cannot hide a poor recall behind a good precision or the reverse. That property is what makes F1 the honest single number when accuracy is not.
How do you read F1 as the balance of escapes and false rejects?
F1 moves only when both escapes and false rejects are under control, which is why it exposes an inspector that accuracy flatters. Return to the 10,000-part line with 100 defects and compare two inspectors on the same run.
| Metric | Pass-everything inspector | Balanced AI inspector | What it exposes |
|---|---|---|---|
| Accuracy | 99.0% | 99.7% | Almost identical, so accuracy hides the gap |
| Recall (catch rate) | 0% | 98% | Escapes: defects reaching the customer |
| Precision | n/a, rejects nothing | 77% | False rejects: good parts scrapped |
| F1-score | 0 | 86% | The honest single number |
The pass-everything inspector scores 99% accuracy and an F1 of zero: it never rejects, so it never catches anything. A balanced inspector that catches 98 of the 100 defects and wrongly rejects 30 good parts scores 99.7% accuracy, barely higher, but its F1 is about 86%. Less than one accuracy point separates a worthless inspector from a working one; F1 separates them by 86 points. That is the gap accuracy hides and F1 shows.
What does 99.65% F1 mean next to a 0.69% false-negative rate?
A high F1 and a low false-negative rate together are the pair worth asking for, because one proves overall balance and the other pins down the error that reaches your customer. On a live cap-inspection line, Adente Vision runs at 99.65% F1 with a 0.69% false-negative rate, catching broken, unclosed and hinge-damaged caps in about 30 ms per part. The F1 says precision and recall are both high, so the line is neither leaking defects nor scrapping good caps in volume.
The 0.69% false-negative rate says the specific error that matters most, a bad cap reaching a customer, is down to roughly seven escapes per thousand defects. Read together, the two numbers describe an inspector you can check rather than one hiding behind accuracy. Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators.
Which metric should you demand from a vision vendor?
Ask for F1, precision, recall and the false-negative rate, each measured on parts like yours, and treat a bare "99% accurate" as a non-answer. A vendor that quotes only accuracy is either not measuring the errors that cost you money or is relying on the easy math of a rare-defect line. The honest set is small: F1 as the single balanced number, recall and false-negative rate for escapes, precision for false rejects, each measured on a representative run of your own good and defective parts.
The same metrics describe every task a single unit runs, from surface-defect anomaly detection to counting and dimensional measurement, so you can compare tasks on the same terms across the inspection applications it covers. A number a vendor can put next to a named case, the way a 99.65% F1 sits next to a real cap line, is worth far more than an unsourced percentage on a datasheet. For the metric that most directly tracks escapes, see the false-negative rate inspection metric post; for the wider method behind few-shot and anomaly detection, see the AI visual inspection guide.