Artificial Intelligence for Quality Control: Optimising Inspection Workflows

Artificial intelligence for quality control: which checks suit AI vision, station readiness, slip and false-call metrics, and when costs outweigh savings.

Artificial Intelligence for Quality Control: Optimising Inspection Workflows
Written by TechnoLynx Published on 24 Mar 2025

Artificial intelligence for quality control is worth putting on an inspection station where rule-based vision breaks down: cosmetic defects, textured or variable surfaces, and assembly verification. Gauging, dimensional measurement and fixed presence or label checks are usually better left to conventional machine vision. Three questions to settle before buying anything: Does the station present parts consistently under controlled light? Have you set the slip rate you will accept and the inspection volume you expect to remove? Does the model hold those numbers on data collected after it was trained?

Which quality checks suit computer vision, and which should stay with human inspectors?

Where a check belongs depends on what makes it hard. If the difficulty is precision, conventional machine vision is the better option. If the difficulty is variation, a learned model earns its keep. Cognex’s guidance gives rule-based tools the location, measurement and orientation of very consistent targets, and gives deep learning complex defects and images with a lot of variation. Landing AI’s 2020 whitepaper draws much the same line: conventional machine vision for gauging and precision work, AI for cosmetic inspection, part location and assembly verification. A trade report on inspection costs makes the same point from the buyer’s side: treating every visual quality problem as an AI problem adds cost without improving the decision.

Decision table: who should own each check

Check Usually best served by Why
Gauging, measurement, location or orientation Conventional rule-based vision Rule-based tools are effective on very consistent targets
Barcode reads, presence checks, fixed label verification Conventional rule-based vision Deterministic tools are enough here, and AI adds cost without a better decision
Cosmetic defects, irregular textures, inconsistent material appearance AI vision Normal variation makes hand-written rules brittle
Assembly verification and completeness checks AI vision, usually detection models Detectors such as YOLO and region-based CNNs handle completeness checks
Defect types not seen in training Not reliably caught by a model trained on line data Models trained on repetitive line data can under-detect new defects
Ambiguous defects, modest volume, frequently changing product Human review Human review may be cheaper under these conditions
Low-throughput, high-consequence operations such as nuclear weapons assembly Human inspectors A Sandia report finds automation not yet practical there

If you want the mechanics behind these rows, our explainer on how a machine vision inspection system works covers them. The equipment side of the same split is covered in our piece on where AI adds value in visual inspection equipment and where rules still win.

Human-level performance is the reference value when previously manual inspection is automated. See, Drury and colleagues, writing for Sandia National Laboratories, report a minimum error rate of about one in a thousand only for very simple accept/reject tasks, and 20% to 30% for the more complex tasks that make up most inspection. In human inspection, misses happen much more often than false alarms. Drury and Sinclair found that neither unaided humans nor full automation worked on one tested task, and Hou, Lin and Drury found combined human, sensor and computer systems usually beat any single component.

New defects are where models fall short. In a 2024 study of continuous training for glass bottle inspection, a newly introduced defect left the model’s predictions close to random. Research on open-set detection aims to recognise novel anomalies without extensive defect annotation. We would still not design a station today on the assumption that the model will flag what it has never seen.

How to use AI in quality control without handing it every decision

The most useful framing is a decision gate, not a replacement. In printed circuit board assembly, research approaches put a machine-learning gate between the automated optical inspection (AOI) machine and the manual inspection station. The gate filters out false calls to free operator capacity at the manual inspection station. AI narrows the work that people do. It does not remove the people.

The next choice is what happens when the model says no. A system that only alerts an operator is usually simpler to integrate than one that stops the line, diverts material and triggers downstream containment. Automatic rejection makes sense where volume, speed and defect risk justify it. Whichever you choose, decide in advance whether uncertain items go to human review.

The model’s task follows from the decision the station has to make. The 2024 survey of 196 open-access papers by Hütten et al. sets out four inspection tasks:

  • binary classification (accept or reject);
  • multi-class classification (scrap versus reworkable, for example);
  • localisation (where the problem is);
  • multi-class localisation (where, and what kind).

In manufacturing use cases, the survey found classification and localisation used about equally. Localising has an advantage beyond the extra information. A 2023 arXiv study found that requiring a model to locate the defect made it learn more robust features than a plain OK/NG classifier: a detector trained on uniform parts still reached an AUC of 0.85 on a diverse dataset.

Automation also changes what the remaining inspectors do. An active search task can become passive supervision, with a risk of more pronounced vigilance effects. Research on human-machine trust suggests that machines in an advisory role can rebuild trust after errors in a way that autonomous ones struggle to. Anomaly maps, which highlight the image regions the model used, can be handed to the inspector to help locate errors. For a broader view of computer vision for quality control in manufacturing across sectors, see our overview.

Is this inspection station ready for AI-based quality control?

Use this checklist before anyone selects a camera or a model. A station that fails an item needs the fix first, because a model will not make up for it.

  1. Parts arrive the same way every time. If parts rotate, shift, overlap, arrive with variable spacing or carry uncontrolled reflections, the project will need mechanical correction later. Finding that out before selecting cameras and software is cheaper.
  2. Lighting is designed for the defect. Daryl Martin’s lighting method for machine vision has three steps: identify the features that separate bad parts from good, design lighting that gives consistent contrast between those features and the background, and test it on many good and bad parts. Shielding the inspection area from ambient light is part of the same job.
  3. Acceptance criteria are written down. The quality, line, engineering and finance teams agree which defects matter and which differences are acceptable variation. They also agree whether the system must identify, classify, locate or measure, and what passing a bad part costs compared with rejecting a good one.
  4. Your inspectors agree with each other. The AIAG suggests a kappa of at least 0.75 for good agreement and prefers values such as 0.90. Consistent ratings are not necessarily correct ones, so check agreement against a reference standard as well.
  5. There are enough defect examples, and they vary enough. Defect samples are scarce because mature processes produce few bad parts, and labelling them is expensive. A 2023 arXiv study on defect detection reports that transfer learning on diverse defect data can produce a performant model from a few hundred instances.
  6. For PCB assembly, the AOI process meets its standard. IPC-9716 sets requirements for lighting, calibration, detectability, threshold limits, measurement system analysis, maintenance and verification. One of its stated aims is to reduce false calls.

Inspector disagreement turns into label disagreement, and the Hütten et al. survey notes that consistent labels need a shared understanding of where one class ends and the next begins. If your inspectors score below the AIAG threshold, any comparison of the AI with “current QC” is a comparison with noise.

Measuring AI inspection against your current QC process: slip rate and volume reduction

The AOI false-call study judges business impact on two numbers, measured on data the model has never seen: slip rate and volume reduction. Overall accuracy will not tell you. A 2023 case study on AOI false-call reduction defines a slip as a truly defective board that the model passes on as a false call. It defines volume reduction as the share of non-defective boards the model keeps away from manual inspection. Slips are much worse, because they let defects through to later processes. The authors set targets of at least 40% volume reduction at no more than 1% slip and describe these as realistic shop-floor requirements.

The other side of the ledger is the rate of good parts sent to manual inspection. A 2023 arXiv study on defect detection argues that the rate of good parts sent for manual inspection usually decides whether automating inspection is economically viable, because it sets how much manual work you avoid. A shipped defect is still usually the more serious error. Set both targets before comparing anything. The AOI study found that three related works defined no success criteria and a fourth set only a slip goal, leaving their results vague.

Accuracy fails here for a structural reason. In the AOI dataset the label classes had an imbalance of 99%, and a dummy classifier scored better accuracy than properly trained models.

A model that passes a random test split has not yet been tested against production. The AOI study trained on the first half of its data in time order and evaluated on five later slices. Models that met both targets on average on the test set exceeded the slip target strongly from the first later slice. The authors concluded that the models would have failed from the first moment in production. The 2023 defect detection study saw the same thing with images. A classifier trained on repetitive mending-plate images, almost perfect on its validation set, failed on a holdout set acquired on another day with the same preparation. In the 2024 continuous-training study, one inspection model dropped from a validation F1 of 93.2% to 45.7% on production data.

A fair comparison therefore follows four steps:

  1. Baseline inspector agreement with kappa.
  2. Fix slip and volume-reduction targets.
  3. Hold out the most recent production period as the test set.
  4. Ask any supplier to separate performance on sample images from performance on the live line.

Monitoring after go-live has its own catch. Measuring performance needs ground-truth labels, but the business case is to stop producing them. The AOI authors note that random sampling of passed parts is the naive answer. We would budget for that sampling from day one rather than discover the need later.

What does rolling out AI quality control look like, from pilot station to full production line?

A sound rollout releases funding in stages that reduce uncertainty in the right order. The trade report on inspection costs describes three stages:

  1. Confirm that the defect, the product’s appearance, the imaging approach and the available data support a viable inspection decision.
  2. Validate line integration, throughput, reject handling and operator workflow.
  3. Move to full deployment only once those production conditions are understood.

Each stage gets a practical acceptance condition. Examples include a stable image under normal operating conditions, an agreed treatment of uncertain results, validated communication with the line, and a named owner for model and recipe changes. The same report lists the integration work involved, from PLC communication to reject mechanisms, and the questions a pilot must answer so that it does not become an isolated demonstration. Landing AI’s whitepaper puts a proof of concept at under 10% of the software needed for the first deployment, though that is a vendor’s estimate.

Deployment quality is strongly shaped by what sits below the model. A 2026 arXiv paper by an Advantech engineer on industrial edge AI treats silent fallback as a defect: a runtime that quietly sends some operators to the CPU can look fine in low-volume testing and still break latency limits in the field. The paper also warns that a colour-space mismatch, a hidden resize or a change in exposure policy can alter what the model sees even when the application code is unchanged. It cites Cidon et al., who found that 14% to 17% of images produced divergent classifications across edge devices. That work used mobile devices rather than industrial cameras. The paper’s practical rule is to characterise the camera path and freeze the preprocessing contract before finalising decision thresholds.

Throughput depends on image resolution, so balance detection precision against speed. A 2024 YOLO-based inspection study measured average detection times of 25 ms and 32 ms on simpler scenes (31 to 40 frames per second). On high-resolution images, detection slowed to 50 ms and 20 frames per second. Model compression is a common deployment step for meeting real-time demands. Our guide to automated visual inspection hardware, model selection and false-reject rates goes further into that trade.

Production keeps changing after go-live. Lighting, vibration and machine wear shift the input data, and new defects can appear that the training data never covered. Retraining has to be careful. In the 2024 continuous-training study, naive retraining on self-labelled production data caused catastrophic forgetting on the glass bottle line, with validation F1 falling by more than 13%.

Who can take an AI inspection pilot through to a production line?

TechnoLynx engineers the model, edge deployment and monitoring side of taking an AI quality control pilot to the line:

  • computer vision models evaluated against your slip and volume targets on chronologically later data;
  • porting models to edge boards, with acceleration and preprocessing checked against the validated model;
  • the drift and regression monitoring harness after go-live.

A sensible first step with TechnoLynx’s AI engineering for manufacturing and automotive is testing the model you have, or a candidate, on your most recent production data against your targets.

That engagement does not cover lighting design, mechanical handling or PLC integration. It also does not certify a station against your customer’s quality standards. If the check is a barcode read, a gauge or a fixed presence check, a conventional machine vision supplier is the better route. A team with its own ML engineers and a stable, well-lit station can reasonably build in-house. Whoever does the work, ask them for results on a chronological holdout set and for slip and volume figures, not accuracy on a sample set.

When artificial intelligence for quality control costs more than it saves

AI quality control costs more than it saves when the work of keeping its results stable outweighs the manual inspection it removes. The camera alone rarely determines the cost. The trade report on inspection costs points out that custom lighting, mechanical handling, labelling, plant-network integration and continuous model maintenance can turn a cheap camera into an expensive project. The false-call cost described under measurement applies here too: frequent false alarms eat value through manual checks and line stoppages. At the other extreme, a system built for extremely low miss rates may need more cameras, tighter lighting control, more training data and longer validation.

Volume alone does not settle the case. A high-volume line with frequent changeovers may need more recipes and more validation than a stable single-product line. Lower-volume production can still justify AI where defects are expensive or hard for people to judge consistently. The same report argues that the return should count lower scrap, earlier detection of process drift, less rework and fewer escaped defects, not just reduced direct labour. Running costs are real as well. A 2025 survey on defect detection names specialised hardware, software, trained staff and ongoing upgrades as a particular burden for small and medium-sized manufacturers.

Siemens and Procter & Gamble report scrap reductions of 10% to 20%, depending on the product, and commissioning five to ten times faster than traditional bespoke vision systems. That comes from a vendor and its customer, not an independent measurement.

So the case rests on your own numbers. If you can say which defects matter, what slip rate you will accept, how much manual inspection volume you need removed, and who owns the model once the line changes, you can price the station honestly. If you cannot, the money is better spent on fixing part presentation and inspector agreement first.

Back See Blogs
arrow icon