Machine Vision

Vision AI in Inspection: What Works and What Is Hype

Machine learning has real uses in inspection and real limits. A practical view from a company that builds both the measurement and the AI.

November 17, 2026 · 6 min read

Almost every inspection product now claims artificial intelligence. Some of those claims describe a real capability and some describe a threshold with a new name. It helps to know what machine learning does well in inspection, what it does not, and what to ask.

What it does well

Judging appearance. Whether a solder fillet looks acceptable, whether a mark is a scratch or a fiber, whether a surface is contaminated: these are easy to recognize and hard to express as rules. Models trained on examples handle them well.

Coping with normal variation. Components from different suppliers, differing print on markings and natural variation in materials defeat rigid templates. A trained model generalizes across them.

Finding things. Locating components, reading characters and codes, and segmenting parts from complicated backgrounds are well-solved problems with modern networks.

Reducing false calls. Using a learned classifier to review items already flagged by conventional inspection is one of the most dependable applications.

Anomaly detection. Where defects are rare and unpredictable, a model trained only on good parts can flag anything that looks different.

What it does not do

It does not measure. Height, volume and position are physical quantities. They come from calibrated optics and geometry, with a traceable accuracy. A neural network can help find what to measure; the measurement itself should be metrology.

It does not fix a bad image. If the lighting does not reveal the defect, there is nothing for the model to learn.

It does not work without data. A classifier needs examples that represent what it will see in production, including the defects. Rare defects are, by definition, hard to collect.

It does not explain itself readily. A rule fails for a stated reason. A model gives a score. For regulated products this needs to be planned for.

Measurement and AI together

The strongest systems combine the two. Calibrated 2D and 3D measurement provides the numbers that have to be right. Learned models handle the judgements that are hard to specify. Each covers the other’s weakness.

Questions to ask about an AI claim

  • What exactly does the model decide, and what is still done by conventional measurement?
  • What was it trained on, and how similar is that to my product?
  • How many examples of my own does it need, and who labels them?
  • What is the escape rate and false call rate on a test set, and can I see it run on my samples?
  • What happens when my product or a component supplier changes? Who retrains it and how long does that take?
  • Does it run on the machine, or does my image data leave the plant?
  • How are model versions controlled and validated?

The data problem, and ways round it

Shortage of defect examples is the usual obstacle. Practical approaches include training on good parts only and flagging anomalies, generating synthetic defect images, augmenting the examples you have, and continuous learning, where operator review decisions are fed back to improve the model over time.

A sensible way to adopt it

  1. Pick one specific, costly problem, such as a defect type with a high false call rate.
  2. Run the model in parallel with the existing inspection and compare.
  3. Validate against a held-back set of known defects.
  4. Only then let it make decisions, and keep monitoring.

Ask what the AI decides and how you would know if it were wrong. A clear answer to both is a good sign. A vague one is the hype.

Questions About Your Own Line?

Our engineers are glad to talk through your process and what inspection could do for it.