Why computer vision for QA

Manual visual inspection on a production line catches defects inconsistently — fatigue and line speed both affect a human inspector's accuracy in ways that are hard to measure, let alone improve.

Data collection first

We resisted buying a model before collecting data: the first six weeks were spent mounting cameras at the existing inspection station and labeling defect and non-defect frames from real production runs, not stock defect photos, since lighting and camera angle on this specific line turned out to matter more than the defect type itself.

The implementation pipeline

  1. Camera and lighting rig matched to the existing station, so the model never has to generalize across setups
  2. A defect-classification model trained on the line's own labeled frames, re-trained monthly as new defect types appear
  3. A confidence threshold that routes low-confidence frames to a human reviewer instead of forcing an automatic pass/fail
  4. A dashboard that tracks the model's agreement rate with human reviewers over time, not just raw defect counts

Rolling out to the line

The system ran in shadow mode — flagging but never stopping the line — for three weeks before it was given the authority to halt production on a high-confidence defect.

Lessons learned

The single biggest accuracy gain came from fixing the camera mount, not from a better model — a lesson worth remembering before reaching for a bigger model to fix what is often a data or hardware problem.