Visual inspection is tiring, repetitive work. Inspectors get tired, and two people rarely judge a borderline part the same way. Cameras and deep learning can help, but not on every line. Here is how to tell.
When a camera does the job well
- The defect is visible in an image. Scratches, dents, missing parts, wrong labels, colour or print errors.
- You can control the scene. The part arrives in a similar position, under the same light, at a known speed.
- You have examples. Hundreds of images of good parts and enough images of each defect type to teach the model.
- The rule is clear. People agree on what counts as a reject.
- Volume is high. Many parts a minute, every shift, where people cannot look at every one carefully.
When it struggles
- The defect is inside the part, or needs touch, sound or measurement rather than sight.
- Products change every week, so the model never has time to learn them.
- Defects are so rare that there are almost no examples to learn from.
- Inspectors themselves disagree on borderline parts. A model learns the disagreement.
Deep learning or classic machine vision?
Classic machine vision measures with fixed rules: an edge here, a hole of this size there. It is fast and predictable when parts are uniform. Deep learning learns from examples, and copes better with natural variation: textures, organic materials, defects that never look the same twice. Many lines use both: rules for measurements, a model for appearance.
Where the model runs
Images can be analysed in the cloud, on a server in the plant, or on an edge device next to the camera. On a production line the edge is often the right place: decisions arrive in milliseconds, the line keeps working if the network drops, and images do not leave the factory. The price is a device you have to manage: updates, monitoring and a new model when the product changes.
What a first project looks like
- Pick one station with a clear, visible defect and enough volume.
- Collect and label images with your inspectors, including borderline cases.
- Fix the scene: camera position, lighting, triggering.
- Run in shadow mode: the system judges every part, but people still decide. Compare the two.
- Measure missed defects, false rejects and inspection time against the inspectors' baseline.
- Switch over only when the numbers hold, and keep a person on the parts the model is unsure about.
The numbers that matter
- Missed defects (escapes): bad parts that pass. Usually the most expensive error.
- False rejects: good parts thrown away or re-checked. Too many and people stop trusting the system.
- Cycle time: whether inspection keeps up with the line.
Agree the targets for all three before you start, not after the first results.
We build inspection software and the edge devices it runs on. See our case studies, or tell us about your line.
