Machine vision / Article

Reducing AOI False Calls Without Letting Defects Escape

A false call is a good board flagged as bad. On typical SMT lines, 30 to 80 percent of AOI calls are false, and every one of them consumes verification labour without removing a single defect. The fix is not loosening thresholds — that trades false calls for escapes. This article covers where false calls come from, how to tune them down safely, and when tuning stops being enough.

The numbers at a glance

30-80%
Of AOI calls reported as false on typical SMT lines
<10%
Working false call rate target for a tuned program
-60%
Manual verification cut reported with AI screening
2 JOBS
Every threshold does two: block defects, pass good boards

01 / The cost

What a false call actually costs

Industry reporting puts AOI false call ratios anywhere between 30 and 80 percent of total calls, a range confirmed independently by shop-floor data vendors. Read that again: on a badly tuned line, most of what the AOI flags is not a defect. Someone still has to look at every flag.

That someone is the verification operator, and their time is the first cost. The second is throughput: boards queue at the verify station instead of moving down the line. The third is the dangerous one. Fatigued review of endless good boards desensitizes the operator, and desensitized operators start passing real defects. High false call rates do not just waste labour — they manufacture escapes. We covered what those escapes cost in AOI vs manual inspection.

02 / The causes

Where false calls come from

Traditional AOI programs are built from golden-board tolerances: acceptable appearance ranges derived from a handful of known-good samples. Real production varies more than any handful of samples. New component batches look slightly different, board finishes drift run to run, and the moment production steps outside the training window, the flags flood in.

01Tolerances derived from too few golden boards — the program's idea of "normal" is narrower than the process.
02Component batch and finish variation the library has never seen — same part number, different look.
03Lighting and optics limits — low-contrast features that the camera cannot cleanly separate.
04Pre-reflow flags that self-correct — offset components re-align during reflow, so a pre-reflow "defect" passes post-reflow.

The distribution matters as much as the causes: false calls concentrate. Drill-down analyses routinely find that a handful of component types produce over half of a line's false calls. That concentration is good news — it means most of the problem is fixable by working a short list.

03 / The safe sequence

Tuning down false calls without opening escape paths

The tempting fix is to loosen the thresholds that fire most. Resist it. Every widened tolerance is a door a real defect can walk through, and the same components that generate your false calls are often also your top sources of real defects. The safe sequence attacks causes, not symptoms:

01Measure first. Track false call rate and escape rate as a pair, per program and per component. One number without the other is unreadable.
02Pareto the flags. Rank components by false call count and work the top of the list — the concentration does most of your prioritisation for you.
03Fix libraries before thresholds. Re-teach the offending components with samples spanning real batch variation, not one golden board.
04Fix physics before software. Lighting, focus, and 3D measurement settings resolve low-contrast false calls that no threshold change can.
05Only then adjust tolerances — one change at a time, watching the escape rate for the next runs before making the next change.

Systematic work along these lines typically brings a program from crisis levels down to the 10 to 15 percent band. A well-tuned process targets under 500 PPM per inspection opportunity. If you are above 20 percent of flags, you have a programming problem, not a process problem.

04 / Beyond tuning

When tuning stops — AI takes the review

Below roughly 10 percent, rule-based tuning hits diminishing returns: the remaining false calls come from variation the rule system cannot express. The current answer is machine learning at two points in the pipeline. At the review stage, ML classifiers re-judge every AOI flag and screen out high-confidence false calls before an operator sees them — deployments report up to 60 percent of manual verification eliminated. At the detection stage, deep-learning inspection models trained on wide, realistic image sets replace brittle golden-board tolerances altogether.

The training data is the catch: real defect images are scarce precisely because defects are rare. That is the problem CODETRACE attacks with NVIDIA-platform defect image generation — synthesizing realistic defect images so deep-learning inspection models see enough variation to separate real defects from process noise, before the line ever produces the defect.

Never trade escapes for quiet. Fix what the program believes, not what it's allowed to say.

FAQ / False calls

Questions, answered.

01

What is an AOI false call?

A false call is when the AOI flags a good joint or component as defective. An operator then has to re-inspect and pass it manually. Industry-reported false call ratios run anywhere from 30 to 80 percent of total AOI calls depending on program quality.

02

What is an acceptable AOI false call rate?

Common working targets are a false call rate below 10 percent of total flags, or below roughly 500 PPM measured per inspection opportunity on a well-tuned program. Above 20 percent of flags usually indicates a systematic programming problem rather than normal process variation.

03

Why does my AOI have so many false calls?

Most false calls trace to tolerances derived from too few golden boards, component batch and board finish variation the program never saw, lighting and optics limitations, and pre-reflow flags that self-correct during reflow. A small number of component types usually accounts for the majority of false calls.

04

Can I fix false calls by loosening thresholds?

Loosening thresholds trades false calls for escapes. Every widened tolerance opens a path for a real defect to pass. The safe sequence is to fix component libraries, lighting, and program logic first, and to track escape rate alongside false call rate so you can see if a tuning change leaked.

05

How does AI reduce AOI false calls?

Machine learning classifiers re-judge each AOI flag before it reaches the operator, screening out high-confidence false calls. Vendors report up to 60 percent reduction in manual verification work. Deep learning inspection programs trained on realistic defect image sets attack the problem earlier, at the detection stage.

Drowning in flags? Send us your FCR.

Talk to our engineers
Scroll to Top