Why Machine Vision Systems False-Reject Good Product — and How to Diagnose It

An escape is defective product that passed. A false reject is good product that failed. They move in opposite directions when you change a threshold, which is why a single "accuracy" number is useless for diagnosis and why both need their own budget.

Before touching anything, quantify the current state: rate of rejects by reason code, over time, and whether the rejected units are genuinely acceptable on re-inspection. A surprising share of "false reject" complaints turn out to be the system correctly catching a process change nobody had noticed.

The most common cause. Guides get bumped, a worn belt lets parts skew, a fixture is re-installed a few millimetres off, spacing changes when the upstream machine is re-timed. The inspection sees a part in a position it was never configured for.

• Compare current images with images captured at commissioning — position, rotation, scale, framing.
• Check whether the located datum still appears where the recipe expects it.
• Inspect fixturing and guides physically, including anything replaced during recent maintenance.
• Look at whether the reject rate correlates with a specific lane, head, or upstream machine.

Illuminators age, lenses and windows accumulate dust and aerosol film, ambient light changes when a bay door opens or a fixture above the line is replaced with a different lamp. Contrast falls, the measured feature moves toward the threshold, and reject rate rises gradually.

• Measure a reference grey or a known feature's contrast and compare to the commissioning value.
• Clean lens, window and light face; re-measure before changing anything in software.
• Check for a new ambient source, including a neighboring station's strobe leaking into the field.
• Confirm strobe timing and current have not been altered during unrelated work.

A system commissioned with the operating point sitting just beside the population of good product will false-reject as soon as anything shifts. Margin is measurable: for each check, how far is the threshold from the nearest acceptable sample, and how wide is the good-product distribution?

Illustrative example. If good units score 82–94 on a match metric and the reject threshold sits at 80, a two-point contrast loss from a dusty lens starts producing rejects. Widening the imaging contrast — better geometry, cleaner optics, more resolution on the feature — restores margin. Moving the threshold to 70 does not; it just hides the escape risk.

Record per check

Good-product score distribution, defect score distribution, threshold, and the gap between them.

Healthy sign

Two clearly separated distributions with the threshold in the empty space between them.

Warning sign

Overlapping distributions. No threshold fixes this; the imaging or the specification must change.

Parameters edited on the floor to get through a shift, never documented, never reverted. Six months later the active configuration is nobody's design. This is a procedural failure and it is fixed procedurally.

• Diff the active recipe against the released version; investigate every difference.
• Confirm recipe selection is still driven by the work order and not by a manual override left in place.
• Restrict who can change what, with an audit trail of changes and reasons.
• Require a short validation run after any parameter change, and log it.

Late verdicts, a shifted encoder scale, a mistriggering photo-eye, or a unit-tracking offset after a jam recovery all produce a specific signature: rejects that do not correlate with product condition, or a good unit rejected next to a bad one.

• Verify trigger-to-image latency and worst-case decision time against the reject window.
• Check for double triggers from a dirty or misaligned photo-eye.
• Confirm encoder scaling against a measured distance, and unit tracking after a simulated jam.
• Look for reject-mechanism timing that has been adjusted mechanically without updating the window.

Condensation after a chilled step, a new film supplier with different gloss, a seasonal substrate color shift, a promotional artwork variant, a different ink lot. The inspection is behaving exactly as configured; the world changed.

The remedy is to bring the new variation into a labeled panel, decide whether it is acceptable product, and re-establish the operating point deliberately. Absorbing it by relaxing a threshold on the floor is how systems lose credibility.

A labeled sample panel — clean units plus each defect class, including marginal cases — is the only instrument that lets you distinguish a real change from an argument. Run it at speed, report escapes and false rejects separately, and keep the panel physically available for re-runs.

Re-run after any change to fixturing, optics, illumination, recipe or model version. Those are exactly the events that invalidate prior results, and re-running takes less time than the debate that follows not re-running.