AI Quality Control: Visual Inspection and Defect Prediction

By Tara Minh · Updated Oct 11, 2026

AI quality control is the use of computer vision and machine learning to find defects, or to predict them before they happen, so that fewer bad products reach a customer. Vision systems look at each unit and decide pass or fail. Prediction models read process data such as temperatures, test results and timestamps and score how likely a unit or batch is to fail. People still own the quality standard, the cost trade-off and the decision to stop a line.

It's a detection and prediction tool, not a quality system. In the terms of QA vs QC, most of what you'll buy sits on the control side: checking output against a requirement. The best results come when the findings feed back into assurance, meaning the process changes that stop the defect from being made at all.

This article is part of the AI for Operations collection, which starts at AI in operations management. It covers how the two main techniques work, why accuracy is the wrong thing to ask a vendor about, how to set the false-reject trade-off, and how to pilot safely.

Key Facts

• A press-shop camera system at BMW's Dingolfing plant used to flag dust and oil residue as cracks. BMW said in July 2019 that a neural network trained on about 100 real images per feature stopped those false alarms (BMW Group). That's the company's own claim.
• Precision is the share of flagged items that really are defective. Recall is the share of real defects that get flagged. Raising the decision threshold tends to cut false positives and raise false negatives (Google Machine Learning Crash Course).
• In the public Bosch production-line dataset, fewer than 1 percent of parts were in the failure class, across 1,184,687 training samples. The best training score reported in one analysis was a Matthews correlation coefficient of 0.227 (arXiv 1701.00705).
• Defect samples are hard to collect, and labeling them takes substantial labor, which is why a whole branch of research targets inspection with few labeled examples (arXiv 2203.05733, 2022).
• The pattern across the field: the model is rarely the hard part. Lighting, labeled defect examples and a clear rule for what counts as a defect decide the outcome.

The Two Techniques

Managers often hear "AI inspection" and "predictive quality" as one thing. They're different jobs with different data.

Visual inspection answers: is this unit defective, and where? A camera captures an image, a model compares it with what a good part looks like, and the system passes, fails or routes the unit to a person. It works on anything you can photograph: scratches on painted panels, missing components on a circuit board, mislabeled packs, fill levels, weld spatter.

Defect prediction answers: which units or batches are likely to fail, before or without inspecting them? It uses structured data from the line, such as measurements at each station, machine settings, material lots, test results and timestamps. The output is a risk score, not a verdict. A high score can trigger extra testing, a hold or a process check.

In ACE Framework terms, both start with Ingest (images or process data), then Analyze and Predict. Execute is the step that rejects a unit or stops a line, and it's where you set the tightest limits.

How Computer Vision Inspection Works

The pipeline has four parts, and the weakest one sets your ceiling.

1. Capture. Camera, lens, lighting and mounting. Poor lighting makes a defect invisible to any model. Controlled lighting often matters more than the choice of algorithm.

2. Training data. You need images labeled by someone who knows what the standard is. Real defects are rare, so you'll have plenty of good parts and few bad ones.

3. Model. Two broad approaches exist. A supervised model learns from examples of each defect type. An anomaly-detection model learns only what good looks like and flags anything that deviates. The MVTec AD benchmark, a widely used public dataset for industrial anomaly detection, is built this way: it has over 5,000 high-resolution images across fifteen object and texture categories, with defect-free training images in each category (MVTec). Anomaly detection helps when defects are too rare or varied to collect, though it tends to flag harmless variation too, which leads straight to false rejects.

4. Decision and handoff. A score becomes a pass, fail or "send to a person." That last category matters. A system that can say "I'm not sure" is much safer than one forced into two answers.

The research on small defect datasets is useful context. A 2022 survey of surface defect detection notes that defect samples are hard to collect and annotation is labor-heavy, and it reviews methods built for that case, including data augmentation, transfer learning, and semi-supervised, weakly supervised and unsupervised learning (arXiv). The practical takeaway for a buyer is to ask a vendor how many defect images it needs for each defect type, and what it does for defects you've never seen.

A real example: cutting false alarms

BMW's Dingolfing press shop is a clear public case because the story is about false rejects, not about catching more defects. According to BMW, dust particles or oil residue left after forming can resemble very fine cracks, and earlier camera systems sometimes flagged these harmless traces. BMW says its AI application, trained on about 100 real images per feature, including images of perfect parts, parts with dust and parts with oil droplets, removed those pseudo-defects (BMW Group). The same release describes AI comparing order data with a live image of the model designation at final inspection and alerting the team to a mismatch.

Two cautions. These are BMW's statements in a press release, and the release gives no accuracy figure you could compare with your own line. Treat it as an illustration of how a system gets tuned: show it the look-alikes, such as dust and oil, that it must learn to ignore.

How Defect Prediction Works

Prediction models don't look at the product. They look at everything recorded about how it was made.

Take a plant where each part passes through dozens of stations, and each station records measurements and test results. A model learns which combinations of readings, which machines and which times tend to precede a failure, then scores new parts. A good example is the public Bosch production-line dataset, released for a Kaggle competition. It holds 1,184,687 training samples with 968 numerical, 2,140 categorical and 1,156 date features, and the failure class is under 1 percent of parts. In one published analysis the best training score was a Matthews correlation coefficient of 0.227, and the competition-style leaderboard score was about 0.215 (arXiv). That's a modest score, and a fair reminder that predicting rare failures from process data is hard. The authors concluded that flagging likely failures was feasible (arXiv abstract), which is a more realistic ambition than "predict every defect."

Where prediction earns its keep:

• Replacing a late, expensive test. If a final test is slow or destructive, a risk score can decide which units need it.
• Finding the upstream cause. Even a weak model can point to a station, supplier lot or shift where failures cluster. That feeds root-cause work such as Six Sigma.
• Triage. Send high-risk units to extra inspection while low-risk ones flow through.

Accuracy Is the Wrong Question

"How accurate is it?" sounds sensible and hides the problem. If 1 unit in 100 is defective, a system that passes everything is 99 percent accurate and useless. That's why quality teams use the pair of measures defined in Google's machine learning course: precision is the proportion of positive classifications that are actually positive, and recall is the proportion of all actual positives that were classified correctly (Google). In inspection terms:

• Recall is the share of real defects the system catches. Low recall means escapes: bad units reach customers.
• Precision is the share of flagged units that are truly bad. Low precision means false rejects: good units are scrapped or reprocessinged.

The two pull against each other. Google's course notes that raising the classification threshold tends to reduce false positives, which also tends to increase missed positives (Google). Scikit-learn's documentation describes the same mechanism, with precision and recall curves computed by varying a decision threshold (scikit-learn). So the question to put to any vendor is: at what threshold, on whose images, and what do the escapes and false rejects look like in my units?

An illustrative example of the trade-off

This is a made-up scenario, with round numbers chosen to show the arithmetic.

A line makes 100,000 units a month, and 1 percent (1,000 units) are truly defective. Setting A is strict: it catches 98 percent of defects (980) and wrongly rejects 3 percent of good units (2,970). Setting B is lenient: it catches 90 percent (900) and wrongly rejects 0.5 percent of good units (495).

Setting A looks safer, but three out of four things it rejects are fine. If a false reject costs a few dollars in reprocessing, that's affordable. If an escape costs a recall, or someone's safety, it's the right choice anyway. The point isn't which setting wins. It's that the right threshold depends on the cost of each kind of error, which only your business can estimate. Most plants use two thresholds: automatic fail above one score, automatic pass below another, and a human check in the middle.

Function by Function

• Manufacturing: surface defects, assembly completeness, weld and seal checks, label and print verification. See manufacturing quality management for the wider system this plugs into.
• Logistics and warehousing: package damage at receiving, label and barcode checks, count verification.
• Suppliers: incoming inspection photos and lot-level risk scores, which pair with AI supplier performance scorecards.

How to Implement AI Quality Control, Step by Step

1. Pick one defect that costs real money. Choose a defect with a known cost per escape and per false reject. Without both numbers you can't set a threshold.
2. Write the defect standard. Define pass, fail and borderline with photos. The pharmaceutical literature on manual inspection makes a useful point here: inspection conditions and inspector qualification need to be controlled for consistent results (AAPS PharmSciTech). The same holds for the people who label your training images.
3. Measure your current process. Record how many defects your human or rule-based check catches and misses today, using a challenge set of known good and bad units. That's your baseline.
4. Fix capture first. Stabilize lighting, camera position and part presentation before you train anything.
5. Collect and label examples. Include the look-alikes: dust, glare, harmless marks. Ask how many images each defect type needs.
6. Run in shadow mode. The system scores every unit while people keep making the decisions. Compare its calls with theirs, and investigate each disagreement.
7. Set two thresholds and a review lane. Auto-pass, auto-fail, and "send to a person" in between. Review the size of that middle band weekly.
8. Close the loop. Feed every confirmed defect back into training data and into the root-cause process. A detection tool that never changes the process only moves the scrap pile.
9. Monitor for drift. New materials, new suppliers and a changed lamp all shift what images look like. Assign an owner who re-checks recall on a fresh challenge set on a schedule.

For the data work in steps 3 to 5, see operational data for AI. When you're ready to scale beyond one line, implementing AI in operations covers sequencing, and measuring AI in operations covers how to prove the gain.

Pitfalls to Avoid

Buying on a single accuracy number. Ask for escapes and false rejects at a stated threshold, on images from your line.

Training only on perfect parts and obvious defects. The system then rejects every harmless variation. BMW's fix was to show the system dust and oil on purpose.

Ignoring lighting. A model can't find what the camera can't see.

Forcing a binary answer. Without a review lane, uncertainty becomes either scrap or escapes.

Skipping the root cause. Catching defects at the end of the line is the most expensive place to find them.

The Limits of AI Quality Control

Rare defects. The rarer the failure, the less there is to learn from, and the harder prediction gets, as the Bosch data shows.

Defects nobody labeled. A supervised model can't recognize a defect type it has never seen. An anomaly model can flag it, with more false alarms.

Subjective standards. If two experienced inspectors disagree about a cosmetic flaw, a model will inherit the disagreement. Settle the standard first.

Changing conditions. New materials, finishes and suppliers can quietly degrade performance.

Accountability. Safety-critical and regulated decisions need a named human owner. The US National Institute of Standards and Technology describes its AI Risk Management Framework as voluntary guidance, which makes it a reasonable checklist for ownership and monitoring.

What to Measure

• Escapes: defects found after the inspection point, by customer complaint, return or audit.
• False-reject rate and cost: good units rejected, and the reprocessing or scrap they cause.
• Review-lane size: the share of units sent to a person, and how it trends.
• Agreement with experts: how often the system's call matches a verified expert call on the challenge set.
• Cost of quality: scrap, reprocessing, warranty and inspection labor together, not any one of them alone.

Where to Go Next

If you do one thing this quarter, build a challenge set of a few hundred units, with known good and bad examples, and measure how well your current inspection does on it. That single exercise tells you the baseline any AI system has to beat, and it exposes most of the disagreement over what counts as a defect. For the quality concepts behind it, read QA vs QC.