Skip to content
AI in Operations | Part 7 of 10

Vision Inspection With AI: Light, Labels, Integration

What decides success in an AI vision station: controlled lighting, a defect catalog, a clear reject path and the human check that stays.

Vision inspection is the most common first AI use case in factories, and for good reasons. The loss is visible, the check repeats hundreds of times per shift, and the reaction is clear: a bad part is rejected.

Yet many vision projects disappoint. The model is rarely the reason.

Light decides more than the algorithm

A camera sees what the light shows. Sunlight through a hall window, a lamp that ages, a reflection on a glossy surface: each changes the image more than any defect does.

Shield the station from ambient light, use defined lighting for the defect type, and treat the lamp as a wear part with a replacement interval. In one pilot, false rejects fell from 4.1 to 3.3 percent in a single week after the station was shielded, without any change to the model.

Labels decide what the model learns

A model learns what people tell it is good or bad. If two inspectors judge the same scratch differently, the model learns the disagreement.

Before labeling, write a defect catalog: each defect type with photos, a limit sample and a rule such as “scratches longer than 3 mm on the visible face are rejects”. Label with the catalog in hand. Start with a few hundred good examples per defect type and add more where the model is weak.

Rare defects need special care. Keep every real example and collect them deliberately. Synthetic variants can help, but test the model on real parts.

Integration decides whether it is used

The station must fit the line:

  • inference time well inside the takt, for example 0.8 seconds at a takt of 14 seconds,
  • a reject path that cannot be bypassed by accident,
  • a signal to the line control and a record per part,
  • a check station where rejected parts are verified by a person.

The check station is not a weakness. It catches false rejects, produces new labels and keeps people involved in quality.

The human check that stays

After go-live, keep a sample check of passed parts, for example one part per hour. It is the only way to measure the detection rate in production. Without it, the first sign of a weak model is a customer complaint.

Plan for change

Materials, colors, suppliers and variants change. In one case a new material supplier delivered housings with a slightly different gloss, and detection dropped from 98.7 to 97.2 percent within a month. Monthly monitoring caught it, 300 new images fixed it.

Make every process change a trigger to test the model before it goes live, just like a new gauge needs a check before use.

The package for the next line

When the first station works, write down what made it work: hardware list, mounting drawing, lighting rules, defect catalog, model card, training for operators. That package turns the second and third line from a project into a routine.

Newsletter

One Lean idea.
Every week.