Catching defects at line speed with computer vision
A vision system that flags defects in under 200 milliseconds on the line — and, just as importantly, learns which flags the inspectors actually agree with.
The problem
A components manufacturer relied on end-of-line human inspection. Inspectors were good, but attention degrades across a shift, and defects that escaped to customers triggered returns, credits, and — twice in two years — a containment exercise that cost more than the line was worth. Sampling caught trends; it did not catch individual escapes.
The hypothesis
Surface and dimensional defects on a controlled line are a well-suited computer-vision problem. The real risk was not detection accuracy — it was false positives. A system that halts good product teaches operators to ignore it within a week.
The build
- Controlled capture first — fixed lighting, fixed geometry, machine-vision cameras. Most of the accuracy in this project came from the rig, not the model.
- Defect model with a small labeled core — a few thousand labeled images per defect class, expanded with augmentation and, for rare classes, synthetic generation of realistic variants.
- Edge inference — an optimized model running on an industrial edge device at the line, under 200 milliseconds per part, with no cloud dependency and no network in the critical path.
- Inspector-in-the-loop feedback — every flag gets a one-tap agree or disagree from the operator. Those dispositions are the retraining set and the live accuracy metric.
Design choice that mattered: the first release only advised — it never stopped the line. We earned the right to divert automatically by demonstrating agreement with inspectors over several weeks, defect class by defect class.
Rollout
One line, one product family, advisory mode for six weeks. We tracked agreement rate per class and found two classes the model handled better than humans and one it handled badly enough to exclude entirely — that class still routes to a person today. Automatic diversion was enabled only for classes that cleared the bar.
Results
Escapes to customers dropped meaningfully, inspection became consistent across shifts, and the feedback loop meant accuracy improved rather than decayed. The operators' relationship with the system changed too: because they trained it, they trusted it.
What we'd tell you
- Spend on the camera rig before the model. Controlled capture beats clever architecture.
- Ship advisory-only first. Trust, once lost to false positives, is very hard to rebuild.
- Make the operator's disposition a first-class input — it is your label pipeline and your monitoring in one.
- It is fine for the system to opt out of a defect class. Partial coverage that works beats full coverage that doesn't.
More from the field.
Consumer lending, decided in hours
Early-warning credit risk
The 6 a.m. advisor brief
Bring us a hypothesis. Leave with a system.
Tell us what's eating your team's time. We'll give you an honest read on whether AI is the right tool — and if it is, a scoped v1 with a timeline and cost.