The typical situation
Inspectors check parts by eye at the end of the line, thousands an hour. Cracks and thread damage still reach the customer, and each complaint brings a sorting charge and a difficult call.
No camera system is better than a careful inspector on a good day. The aim is to be as good as that inspector on every part, on every shift.
What the data looked like
No useful images existed at the start. The first four weeks collect and label images from the line, with the quality head deciding what counts as a defect. Customer complaint records show which defects matter most.
What we built
- A camera and lighting station at one inspection point, with a rejection gate.
- A defect model trained on the plant's own labelled images, running on a small computer beside the line.
- A screen for the inspector showing every rejected part and the reason, so false rejects are corrected and fed back.
- A weekly report of catch rate and false-reject rate, measured against a manual audit sample.
Typical outcome
| Measure | Before | After | Period |
|---|---|---|---|
| Defective parts reaching customers | baseline | about 60% fewer | typical, first quarter |
| Good parts wrongly rejected | not measured | under 1% | typical outcome |
What usually trips these projects up
Lighting matters more than the model. Two weeks spent getting the lighting right saves more accuracy than any amount of model tuning afterwards.