Vision system false rejects: why your camera fails good parts, and how to stop it.

Vision system false rejects almost always come from three places. The image drifted (light, focus, position or blur), the good parts vary more than the ones you set up on, or the threshold was placed with too little data. Find which one moved and fix that. Loosening the threshold's the last thing to try.

We build robotic cells with machine vision applications inside them, and we inherit plenty of stations someone else commissioned. The usual story is a lens wiped with a shop rag on second shift, a ring light bumped during a changeover, and a threshold loosened at 2 a.m. to keep the line moving. This is the order we work through those stations in. It's the same on a Cognex In-Sight, a Keyence CV-X or a Basler camera on a PC, because the optics don't care whose logo is on the housing.

What is a false reject in machine vision?

A false reject is a good part that the vision system calls bad. It's the opposite of an escape (also called a false accept), where a bad part gets through. Every inspection trades one against the other, and where that trade lands depends on the threshold and on how far apart good and bad parts score.

The AIAG measurement systems manual calls the false reject rate the false alarm rate, and the escape rate the miss rate.6 You'll hear every one of these names on the same floor.

The direct cost is the part you scrap, rework or check by hand. The indirect cost's bigger. A station that cries wolf a few times an hour teaches operators to fish parts out of the reject bin, and then to bypass it. Once that happens, the station bought to stop escapes is producing them.

So the most expensive false reject is the one that teaches people the camera's not worth listening to.

Fig. 01Why a threshold set on 30 parts rejects good ones

Why does my vision system reject good parts?

Because the image, the part or the decision changed after setup. Lighting drifts, lenses get dirty, fixtures wear, suppliers change, and a threshold set on 30 tidy commissioning parts meets the full spread of production. The camera's usually reporting exactly what it sees. It's what it sees that moved.

We sort the causes into six families on a fishbone (Ishikawa) diagram. It's there to make you check every bone before you fall in love with one theory, because a station rejecting good parts usually has two or three small causes stacked together.

Fig. 02Root causes of false rejects · Ishikawa diagram

What are the most common causes of machine vision false rejects?

The eight we see most are ambient light and lighting drift, focus drift and dirty optics, part presentation, motion blur and trigger timing, thresholds set on too few samples, cosmetic variation in good parts, supplier changes, and model drift in deep-learning tools. Each one's got its own telltale pattern in the reject log.

1. Ambient light and lighting drift

A vision station sees its own light plus a skylight at 3 p.m., the dock door and the brighter fixture maintenance put in last month. If the station's light doesn't dominate, your results depend on the time of day. The standard defenses are a shroud, strobing hard enough to swamp the room light, or narrow-band light with a matching bandpass filter.1 Our machine vision lighting guide covers how to pick between them.

The station's own light drifts too. LED output falls with operating hours, and faster when the LED runs hot, which is why the lighting industry projects it with the IES LM-80 test and TM-21 extrapolation at several temperatures.2 A light that's overdriven or cooked in an enclosure ages faster than its datasheet says. And lights warm up, so Monday's first images aren't quite Monday's noon images.

2. Focus drift and dirty optics

A focus ring that wasn't locked will walk with vibration. So will one a technician brushed while cleaning. Coolant mist, oil, dust and fingerprints on the lens or window cut contrast and soften edges, and edge tools suffer first. If false rejects creep up over weeks and a cleaning knocks them back down, this is your bone. Locking screws are cheap, and our guide on how to choose a machine vision lens covers why some lenses hold focus better than others.

3. Part presentation and fixturing variation

A conventional lens images the part at a scale that depends on working distance. At 200 mm working distance, a part sitting 2 mm higher changes the image scale by about 1% (2 ÷ 200). On a 40 mm feature that's 0.4 mm of apparent size change, which is easily bigger than the tolerance a dimensional gauging with machine vision station is checking. Tilt, a worn nest or a gripper that lets the part slip a millimeter do the same thing. The part's fine. It just isn't where the tool expects it.

The telltale sign? Rejects that cluster by fixture nest, pallet or gripper rather than by lot.

4. Motion blur and trigger timing

A moving part smears across the sensor during the exposure. A handy way to size it is blur in pixels = part speed × exposure time × pixels across the field of view ÷ field of view, and blur usually starts to hurt beyond about one pixel.3 Say a part moves at 200 mm/s, the exposure is 1 ms and you have 2,448 pixels across a 120 mm field. That's 200 × 0.001 × 2,448 ÷ 120, or about 4.1 pixels of blur. Cutting the exposure to 0.2 ms brings it to about 0.8 pixels. But you'll need five times the light to keep the same brightness, which is why strobes exist.

Trigger timing is the quieter cousin. With a photo-eye, a PLC scan and a software trigger between the part and the shutter, the delay varies. At 200 mm/s, 5 ms of jitter moves the part 1 mm, enough to push a feature out of a tight region of interest. A hardware or encoder trigger takes most of that out, and it's usually a day's work.

5. Thresholds set on too few samples

This one's the cause we find most often on stations we inherit, and Fig. 01 shows why. At commissioning someone runs 20 or 30 good parts from one lot, cleaned and placed with care. Their scores cluster tightly, and the threshold goes just past the worst of them. Then production shows the station every lot, shift and nest, and the tail of that wider spread crosses the line.

Volume makes it hurt. If a station sees 2,000 parts a shift and the threshold sits at the 99th percentile of production good-part scores, that's about 20 good parts rejected every shift (2,000 × 1%), before a single real defect shows up.

6. Cosmetic variation in good parts

Good parts aren't identical. Machining marks change with tool wear, castings vary in texture, and paint shifts from batch to batch. A surface defect inspection taught only on pristine parts will treat acceptable cosmetics as defects. The cure is a set of boundary samples, signed off by quality, showing the worst acceptable condition of each feature, and a tool that's tuned against them.

7. Supplier, material and process changes

The supplier who switched resin color without telling anyone is a classic. So are a new stamping lubricant and label stock with a different gloss. Any of them can change the image overnight while parts stay in spec. If false rejects jump on one date and line up with a lot, look at incoming material before the camera. And ask purchasing to tell the vision owner when a supplier changes. It sounds obvious. It almost never happens.

8. Model drift in deep-learning inspection

A deep-learning tool only knows the images it was trained on. When the incoming data drifts away from that training set, which the machine-learning literature calls data or concept drift, the model's confidence on perfectly good parts drops and they land in the reject class.5 A new part revision or slow lighting drift can do it. The fix is a loop. Log the false rejects, have quality label them, retrain, and check the new model on a fixed test set before release. Our comparison of deep learning vs rule-based inspection covers when that extra upkeep is worth it, and the AI visual inspection post shows what the loop looks like on a real line.

How do you troubleshoot false rejects on a vision system?

Diagnose before you adjust. Look at the logged reject images first, then run a golden sample to see whether the station drifted, then split rejects by nest, shift and lot. Only then look at the threshold. None of it needs new hardware, and most of it's done in minutes.

Image logging comes first, because without it every false-reject argument's just opinion. Save each rejected image with its score, timestamp, recipe, nest, lot and shift. Set it up at commissioning, not after the first angry phone call.

A golden sample is a few known-good and known-bad parts kept in a drawer at the station and run each shift. If the golden good part scores differently from its baseline, the station moved. If it doesn't, the parts probably changed.

A/B runs settle the "is it the light or the parts?" argument, and they're cheap. Run the same parts twice and change one thing, like a clean lens against a dirty one, or nest 1 against nest 3. Then rerun the rejects. A part that fails once and passes the second time is telling you about repeatability.

Score histograms are the most useful tool of the four. Plot raw scores, not just pass and fail, split by shift, nest, lot and date. You'll see one tight cluster, a wide smear, or two humps that point to two populations.7 Two humps split by nest is a fixture problem. A hump that slides over weeks is drift.

Fig. 03Troubleshooting false rejects · test the image before the decision

Don't skip the first question in Fig. 03. Plenty of "false rejects" turn out to be real defects operators had been quietly passing.

How do you reduce false rejects without letting defects escape?

Fix the cause the diagnosis pointed at, cheapest first, and don't touch the threshold until you've got score data for both good and bad parts. Cleaning optics and restoring a saved job take minutes. Re-fixturing or switching to 3D takes weeks, so you'll want the cheap causes ruled out first.

We'd push back on one common reflex. People reach for a higher-resolution camera when a station rejects good parts, but more pixels rarely help with light, blur or presentation, and they'll make a marginal exposure worse. Check the machine vision camera resolution math or our machine vision resolution calculator before you spend the money. If you really are heading for the last row, read 2D vs 3D machine vision first.

How do you set a vision inspection threshold?

Collect scores for a few hundred good parts across real variation and for every defect type you must catch, plot both histograms on one axis, and put the threshold in the gap between them. Lean it toward whichever side carries the more expensive error. If they overlap, stop. You've got an image problem.

Treat the threshold as a statistical decision, because that's what it is. Good parts make one distribution of scores and bad parts make another. Moving the line between them trades false rejects against escapes, the same trade-off ROC analysis describes for any classifier. You can't lower one error rate without raising the other unless the distributions move apart.4

1. Score good parts from several lots, shifts and nests. A few hundred is a sensible floor, since you're trying to see the tail.
2. Score every defect type, including the faintest one quality calls a reject. Boundary rejects matter more than gross ones.
3. Plot both histograms on one axis. If there's a clear gap, put the threshold in it, closer to the good parts for a safety or customer-critical defect.
4. If they overlap, no threshold fixes it. Go back to the lighting, optics or presentation. A wider gap is what those buy you.
5. Think about a gray zone with two thresholds. Pass below the first, reject above the second, and re-image or route to manual review in between.
6. Write down the data set, threshold, date and approver. On a regulated line that record's part of your vision system validation.

Don't widen the threshold until you know why parts are failing. If the cause was a dirty window, a looser threshold means the station now accepts parts it can't see properly, and it'll keep doing so after the window's cleaned.

How often should you check a vision system for drift?

Run a golden good and golden bad sample every shift and record the scores, clean and check focus weekly, and review score histograms monthly. Drift's slow, so a station checked only when someone complains has usually been rejecting good parts for weeks.

Fig. 04 shows the shape we mean. A golden sample's brightness falls as the window collects film and the LED ages. Cleaning gets the film back but not the LED, and eventually the image leaves the window the threshold was set in. Nothing changes shift to shift, which is why nobody notices.

Fig. 04Lighting drift over a year · illustrative

Here's the routine we leave behind on most stations.

• Every shift, run the golden good and bad samples and log the scores. Out of band means you stop and check before running production.
• Weekly, clean the window, check focus against a printed target, and skim the reject images. Ten minutes of looking catches problems before the numbers do.
• Monthly, plot score histograms by nest and shift, and look at fixture nests for wear.
• After any change (a new supplier, part revision, light, lens, firmware or model), run the golden samples and a short A/B run before the station goes back to production.
• Yearly, plan light replacement, archive the job files, and redo the threshold study with a year of logged scores.

Our machine vision maintenance checklist turns that into a printable sheet. Put a named owner on it, though. The stations we see drifting usually aren't anybody's job, and the PLC tech who doesn't want to touch the vision job isn't going to volunteer.

When should you change the vision hardware instead of tuning it?

When the image is clean, the station is stable, the threshold came from good data and the good and bad distributions still overlap. That means you've hit the limit of the current setup. More tuning will only move false rejects into escapes.

The usual next steps are a different lighting geometry, more resolution on the defect, a telecentric lens, 3D for height, or deep learning for cosmetic calls that rules can't express. Honestly, which one is a judgment call, and we'd test on your parts before buying. The Cognex vs Keyence comparison and the smart camera vs PC-based vision guide help if the answer turns out to be a different platform.

In a robotic cell there's also a mechanical answer. The robot presents the part the same way every time and re-presents it when the score lands in the gray zone. That's how we build our machine vision workcells, and it's often the cheapest way to take presentation out of the problem. If you're deciding whether a fix is worth paying for, the machine vision ROI post shows how to put a dollar figure on a false reject rate.

Frequently asked questions

What is an acceptable false reject rate for a vision system?

There's no universal number. It depends on what a false reject costs you against what an escape costs your customer. The AIAG MSA manual's attribute criteria are commonly quoted with a false alarm rate of 5% or less as acceptable for a study, but production lines usually want far lower. Agree a target with quality first.

How many parts do you need to set a vision inspection threshold?

More than the 20 or 30 most stations get commissioned with. You need to see the tail of the good-part distribution, so a few hundred good parts from several lots and shifts is a floor, plus labeled examples of every defect. A one-in-a-thousand false reject is invisible in a small sample.

Should parts rejected by a vision system be re-inspected?

Yes, at least until you trust the false reject rate. Re-inspecting rejects tells you how many were really good and gives you labeled images for tuning. In a robotic cell the robot can re-present a borderline part at a second angle, which recovers good parts without a manual station.

Can you use different thresholds for different defect types?

Yes, and you usually should. A cosmetic scratch and a missing O-ring don't carry the same risk, so they shouldn't share a threshold. Most smart cameras and deep-learning tools let you set limits per tool or class. Keep zero tolerance where an escape is safety-critical and allow margin where a false reject only costs rework.

Do AI vision systems have fewer false rejects?

They can on cosmetic inspections, where acceptable variation is hard to describe with rules, because a model learns what acceptable looks like from labeled images. But it only knows its training images. When parts or lighting drift it'll start rejecting good parts, so plan a retraining loop and a fixed test set.

Why did my vision system's false rejects suddenly increase?

A sudden jump usually means something changed on a specific date. Check the maintenance log for a cleaning, light swap, firmware update or job edit, then check incoming lots. Run a golden sample. If it still scores normally, the parts changed and the station didn't.