Do we need thousands of defect images?
Not always. Anomaly-detection models learn from good parts and flag anything unusual, which helps when defects are rare. You still need real defect examples to test against.
Guide / AI
Computer vision quality inspection for small manufacturers
Vision inspection works when the defect is visible, the image is repeatable and you can collect labelled examples of good and bad parts. Fix lighting and part position first; the model is the easy part. Start with one defect on one station, run it alongside your inspector, then let it reject.
01
When vision inspection fits
- The defect is visible to a person looking at a photo.
- Parts arrive in a repeatable position and orientation.
- The same check runs on every part, all shift.
- A missed defect costs more than the station.
If an inspector needs to touch, measure or turn the part to decide, a camera alone won't do it.
02
What it takes
| Piece | What it means |
|---|---|
| Defect spec | Written pass/fail rules with example photos. If two inspectors disagree, the model will too. |
| Image capture | Fixed camera, lens and lighting. A trigger from a sensor or PLC so every image shows the part the same way. |
| Labelled data | Good and bad parts from real production, labelled by the people who inspect today. |
| Model | Classification, detection or anomaly detection, picked for the defect and the data you have. |
| Line integration | A reject signal, a divert or an alert, and a log of every decision with its image. |
| Review loop | Borderline parts go to a person. Their calls become new training data. |
Most failed projects fail on capture and spec, not on the model.
03
Where it breaks
| Problem | Fix |
|---|---|
| Glare on shiny parts | Diffuse or polarised lighting, fixed exposure. |
| Parts shift or rotate | Fixture the part, or detect and align before inspecting. |
| Defects too rare to collect | Anomaly detection trained on good parts. |
| New product variant | A labelling pass per variant before it goes live. |
| Lighting drifts over months | A reference image check at shift start. |
04
How to start
- Pick one defect on one station. The one that costs the most in scrap or returns.
- Fix the image. Mount the camera and lights, capture a week of production.
- Label and train. Your inspectors label; hold back a set to test on.
- Run in shadow. The model flags, the inspector decides. Compare both.
- Let it reject. Once misses and false rejects are within your tolerance.
Questions
Direct answers
Can we use the cameras we already have?
Sometimes. Security or phone cameras rarely give consistent images. Most lines need a fixed industrial camera, controlled lighting and a trigger tied to the part.
Will it replace our inspectors?
It takes the repetitive checks. Inspectors review borderline parts, handle new defect types and own the reject decision until the error rate is proven.