Guide

How to reduce false rejects without missing defects

Why a vision station rejects good products, and a practical order of work for quality managers and automation engineers.

By Frank GuoTechnical review: Frank GuoPublished 24 September 2026Updated 27 October 202611 min read

Applies to:Label & print OCR/OCV & codes Surface defects

Discuss a station with false rejects

// Short answer

The false reject rate, also called the false rejection rate (FRR), is the share of good products a vision station rejects. It rises when the image of a good product looks like a defect: unstable lighting, glare, a part out of position, blur, dirty optics, or normal variation it was not set up on. Fix the imaging first, then widen the good-sample set. Change limits last, with your quality team, because a looser limit lets more defects through, and prove each change on a fixed sample set.

Key points

A false reject is a good product the station rejects, counted against the good products inspected.
Most false rejects come from images that vary. Fix lighting, presentation, trigger and focus first.
A looser limit rejects fewer good products but passes more defects. Change limits with quality, and record why.
Prove every change on the same fixed sample set, measuring false rejects and missed defects together.
False rejects that creep back usually mean drift: ageing light, dirty optics, moved mounts or a new material.

What a false reject is, and why it matters

A false reject is a good product that the station rejects. Its opposite is a missed defect: a defective product that passes. The false reject rate is counted against the good products inspected; the missed defect rate against the defective products inspected. They are reported together, because either can be lowered by raising the other.

False rejects cost more than the product in the bin. Operators stop trusting the station, and sooner or later someone switches it off or loosens a limit, and real defects get through.

False rejection rate formula and how to measure it

False reject rate (false rejection rate, FRR) = good products rejected ÷ good products inspected × 100 %.
Missed defect rate (escape rate) = defective products passed ÷ defective products inspected × 100 %.
Defect detection rate = 100 % − missed defect rate.

To use the false reject rate formula, measure the false rejection rate by running a known number of good products at production speed, and the missed defect rate by feeding a known number of marked defective products at known positions. Count both from the same test, because a change that improves one usually worsens the other.

The same term is used in biometrics, where the false rejection rate is paired with a false acceptance rate; that is where many formula searches lead. In inspection, the false acceptance side is the missed defect, or escape. A “high rejection rate” on a line is not always a false reject problem: first check how many of the rejected products are really defective.

Find the cause before touching the limit

Start with the images. Save the images of rejected products, let your quality team split them into real defects and good products, and group the good ones by what the station saw: glare, a part out of position, blur, normal variation, or nothing visible at all. That last group often points to reject tracking. The PLC and reject integration guide covers tracking and confirmation.

The drawing shows why the limit comes last. When good and defective results overlap, moving the limit only swaps one error for the other. Better imaging and samples pull the groups apart.

[Schematic]Reject limit trade-off · illustrative · no values

Swipe sideways to see the whole drawing →

[1] False rejects: good products whose result falls on the reject side of the limit.

[2] Missed defects: defective products whose result falls on the pass side.

[3] Reject limit: set and recorded with your quality team.

[4] Better imaging: separates the groups; both errors fall.

Imaging first: lighting, presentation, trigger, focus

Most false rejects start with an image of a good product that differs from the setup images. Four things decide how repeatable the image is.

Lighting. The station light should be the only light that matters. Daylight, overhead lights and nearby stations change the image through the day; a shroud, strobing, or a lens filter matched to the light reduce that (NI practical guide to machine vision lighting). The geometry must suit the surface, because glare on shiny or curved parts causes many false calls. The machine vision lighting guide explains which light shows which defect.
Presentation. Each product should reach the camera at the same position, angle and distance. Worn guides, touching products, tilted parts and speed changes all alter the view.
Trigger. The sensor must see every product once, with the whole part in view. Late or double triggers give images of half a product, which fail.
Focus and blur. Parts at different heights, a turned lens ring, vibration or an exposure too long for the speed soften edges. A soft edge can look like a defect, or hide one.

Samples: good variation and a borderline library

A station set up on a few perfect parts will reject normal ones. The good-sample set should cover what passes today: batches, shifts, moulds or cavities, print shades, colours, suppliers and every variant at the inspection point. Teach references from representative parts, not the best part on the bench. The sample preparation guide explains how to build and label the set.

Keep a borderline library: parts that are just acceptable and just rejectable, each with your quality team's decision, a photo and the reason. It settles arguments on the line, trains operators, and keeps deep-learning labels consistent. The rule-based or deep learning guide compares the two.

Limits: the trade-off, and how to set and record them

Every decision compares a result with a limit, measured on the good, defect and borderline sets. It is the trade-off statisticians describe for any test: making one kind of error less likely makes the other more likely, unless the measurement itself improves (NIST/SEMATECH e-Handbook, statistical tests). Where the limit sits is a quality decision, not a software setting.

Set limits per defect type, not one overall sensitivity. A wrong code or missing seal justifies more false rejects than a cosmetic mark.
Consider a review band between clear pass and clear fail, diverting borderline products to a manual check, if the line allows it.
Record every limit: the value, the reason, the samples used, who approved it and when. Changes follow the same approval, and operator access to limits is restricted.

Verification: re-test with a fixed sample set

Keep a fixed verification set of good, defective and borderline samples that is not used for setup or training, and run it before and after every change. The figures match project acceptance:

False reject rate: good products rejected, divided by good products inspected, at production speed.
Missed defect rate: defective products passed, divided by defective products fed at known positions.

Record the sample set, speed and duration with each figure. Also run the same parts several times: a part that passes once and fails the next time shows poor repeatability, the question a gauge study asks of any measuring system (NIST/SEMATECH e-Handbook, gauge studies). A small run with no misses shows detection still works on those samples; it does not prove the station misses nothing in long-term production. See how we deliver a project for how these figures are written into FAT and SAT.

Drift: why false rejects creep back

Lighting ageing: output falls over the light's life; images slowly darken.
Dirty optics: dust, oil mist and splash build a film on the lens or cover glass.
Mechanical movement: loose fixings, a camera knocked during cleaning, guides moved at changeover.
New materials or suppliers: a new label stock, resin or ink changes gloss or colour.

Reference checks catch drift early. At agreed moments, such as shift start, after cleaning and after changeover, run a few known good and known defective reference samples and compare result and image with the stored reference. Plot the share of good products rejected per shift or batch; a proportions chart shows when it moves beyond normal variation (NIST/SEMATECH e-Handbook, proportions control charts). Then fix the cause and re-run the verification set, rather than retuning limits on the line. On stations we deliver, who runs these checks is written into the support plan.

When operators bypass the vision system

The clearest sign of a false reject problem is an operator who bypasses the vision system because the rejected products “look fine”. They are usually right about most of them, and the station loses the trust it needs to stop the one bad product that matters. A bypass should be possible only with authorisation, shown clearly on the HMI and logged with time and user, and every bypass is a reason to review the false rejects behind it. A line that has bypassed the vision system for a shift has run that shift without the check.

Two common cases on packaging lines

The empty bottle inspector rejects good bottles because of scratches. Returnable glass carries scuffs from earlier trips. If limits were taught on new bottles, worn bottles fail. The fix is to teach and verify limits on real returned bottles, adjust lighting so scuffs stay faint while real defects stay strong, and accept that heavily worn stock keeps the false reject rate higher.

The labeller rejects good products because the encoder miscounts. If the encoder wheel slips or the belt stretches, tracking drifts and the reject fires on the wrong product, so a good product goes into the bin and the faulty one continues. Check the encoder mounting, compare trigger counts with results and rejects, and add a sensor at the reject point where products can shift. See the PLC and reject integration guide.

Diagnostic table

Start from what you see on the line. These are common causes, not the only ones.

Diagnostic table: false reject symptoms, likely causes and what to check
Symptom	Likely cause	What to check
Rejects rise slowly over weeks	Light output falling with age; a film on the lens or cover glass.	Run the reference samples; compare the image with the stored reference; clean and re-check.
Rejects jump after cleaning or a knock	Camera, lens or light moved; focus or aperture ring turned.	Compare position and sharpness with the reference image; check fixings and lens locks.
Bursts at certain times of day	Daylight, overhead lights or a nearby station's light.	Look at images from those times; check the shroud and strobing.
Rejects follow a batch, supplier or material	New label stock, resin, ink or colour not covered by the good samples.	Once quality accepts the batch, add it to the good set and update the recipe under change control.
Part off-centre, rotated or cut off in the image	Guides moved, products touching, trigger misplaced, speed changed.	Check guides, spacing and trigger position; confirm the whole part is in view.
Soft or smeared edges	Exposure too long for the speed, focus shifted, vibration.	Check exposure, strobe, focus lock and mounting.
Rejects start after a changeover	Wrong recipe, or references taught on one unrepresentative part.	Confirm the running recipe; re-teach from representative good parts.
Same part passes, then fails	Result sits close to the limit: borderline part or poor repeatability.	Run the part several times and look at the spread; decide borderline cases with quality.
Flagged "defect" is normal variation	Not enough good variation in setup: texture, print shade, moulding marks.	Widen the good set, narrow the inspected region, or consider deep learning.
Image of the rejected product shows a pass	Tracking or reject timing removed the wrong product.	Check PLC tracking, encoder or delay, and the confirmation sensor.

// Production checks

Separate image false alarms from wrong-item rejection

A good item in the reject bin does not tell you whether the optical decision was wrong. Trace the image, cycle reference and physical reject before changing image thresholds.

Decision or event / Check or action / Acceptance boundary
Decision or event	Check or action	Acceptance boundary
Good image classified faulty	Inspect illumination, presentation, recipe and accepted variation	Recheck known faults after any threshold change.
Correct bad decision, wrong item removed	Inspect encoder tracking, spacing and actuation delay	Image training does not fix physical tracking.
No image or no result	Inspect trigger, acquisition and timeout handling	Do not count an uninspected item as a pass.
Unstable machined surface	Compare oil/chips/fixture seating and permitted cleaning	Avoid tuning out a real burr with the nuisance marks.

Report false optical decisions and failed physical isolation separately. Keep missed defects in the same validation report so reducing one error does not conceal growth in the other.

Machined-part diagnosis · Tracking and confirmation

// FAQ

Questions about false rejects

What is a false reject in machine vision?

A good product that the station rejects, also called a false call, overkill or overdetection. The false reject rate is good products rejected divided by good products inspected, reported together with the missed defect rate.

Why does my vision system reject good products?

Usually because the image of a good product sometimes looks like a defect: changing light, glare, a part out of position, blur, dirty optics, or variation the station was not set up on. Sometimes the limit is too tight, or tracking removes the wrong product. Check the images first.

Should we just loosen the limit?

Only once the imaging is stable, and only with your quality team. A looser limit rejects fewer good products but passes more real defects. If good and defective parts give similar results, no limit separates them; better imaging or samples will.

How is the false reject rate measured?

Run a known number of good products at production speed and count how many are rejected. Missed defects are measured separately, by feeding known defective products and counting how many pass. Record the sample set, speed and duration.

Why did false rejects increase after months of stable running?

Something changed in the image or the product: ageing lights, a film on the optics, a camera or guide that moved, or a new material or supplier. A fixed set of reference samples, run at agreed intervals, shows which.

What is an acceptable false reject rate?

There is no universal figure. It is agreed per project from what a wrongly rejected good product costs compared with what a missed defect costs, and it is measured at FAT and SAT on your own products. A line where a missed defect means a recall accepts more false rejects than one where it means a cosmetic complaint.

How to reduce false rejection rates using the right vision system?

Choose the lighting, optics and presentation that make the defect clearly different from normal variation, then set limits on a sample set that includes normal variation and borderline cases. A more expensive camera does not fix an image in which good and bad products look alike.

How to reduce false positives in AI visual inspection?

Add more images of good products with their normal variation to the training set, label borderline cases consistently with your quality team, and verify on a fixed test set that was not used for training. Where the defect is well defined, a rule-based check is often more stable than deep learning.

How accurate are machine vision systems?

Accuracy is not one figure; it is described by the false reject rate and the missed defect rate for each check, measured on your products at line speed. Both depend more on lighting, presentation and how clearly defects differ from good products than on the camera itself.

Can false rejects be removed completely?

Usually not over the long run. Where good and defective parts look alike near the limit, fewer false rejects means more missed defects. A test run on well-separated samples can show no false rejects, but that does not prove it for long-term production. The aim is a false reject rate the line can live with, at a missed defect rate your customers accept, agreed in writing and checked over time.

Sources

NIST/SEMATECH e-Handbook of Statistical Methods: What are statistical tests? (US National Institute of Standards and Technology, checked 2026-09-24; the two kinds of error)
NIST/SEMATECH e-Handbook of Statistical Methods: Proportions control charts (checked 2026-09-24)
NIST/SEMATECH e-Handbook of Statistical Methods: Gauge R & R studies (checked 2026-09-24)
A Practical Guide to Machine Vision Lighting (NI technical guide, checked 2026-09-24; methods against ambient light)

Machine vision lighting guide →Preparing samples →Rule-based vs deep learning →PLC & reject integration →Surface defect detection →How we deliver a project →Support →

Line solutions using this:

Dairy cups →Wine labels →Fasteners and turned parts →

Common in these industries:

Food & beverage →Metal components →

Typically built with:

Area scan cameras with a vision controller →2D measurement (backlight, telecentric optics) →Vision sensor (HIKROBOT SC1000–SC3000) →

// Next step

Is a station rejecting good product?

Send a few images of rejected good products and of real defects, and describe the line. We will tell you where we would look first and what samples a review needs.

Start an assessment