Computer vision on the factory floor: a 6-week case study

A regional beverage manufacturer reached out with a familiar story: defect detection on the bottling line was manual, slow, and inconsistent across shifts. Customer complaints had ticked up two quarters in a row. They'd seen the vendor demos. They wanted to know what was actually involved.

The problem

Bottles move past a fixed camera position at roughly 1.4 per second. Defects include cap misalignment, label tears, fill-level deviation, and the occasional foreign particle. Human inspectors catch about 88% of defects with high false-positive rates that slow throughput.

Weeks 1–2: framing

We spent the first two weeks not training a model. We spent them on the line, with the inspectors, watching what they actually caught and what they missed. Three things became clear: the lighting was inconsistent across the shift, the existing camera position was wrong for two of the four defect classes, and the 'defect rate' nobody could agree on was 0.7% — too low to justify a deep model trained from scratch.

Weeks 3–4: modeling

We fine-tuned a pre-trained vision backbone on roughly 12,000 labeled frames, with heavy augmentation for the rarer defect classes. A separate, smaller model handled fill-level via classical CV — no neural net needed. By end of week 4 we were at 94% recall with a 2.1% false-positive rate on the holdout set.

• Train backbone: vision transformer, fine-tuned.
• Auxiliary head for fill-level: classical thresholding, deterministic.
• Inference: edge device on the line, ~28ms per frame.
• Human-in-the-loop UI for ambiguous cases, reviewed at shift change.

Weeks 5–6: production

The last two weeks were the unglamorous part: integrating with the existing PLC, setting up shadow mode so the operators could compare model output to their own calls in real time, drift monitoring, and a one-page runbook for the maintenance team. By end of week 6 the system was running in production, in shadow mode for the first week, then live.

Results and lessons

• Defect detection recall: 88% (human) → 96.4% (system, after 90 days).
• False positives: down 41% versus inspectors.
• Customer complaints related to bottling defects: down 62% in the following quarter.
• Inspector role shifted from rejection to root-cause analysis — a quietly large win.

Six weeks. One model, one classical CV trick, one lighting rig, one redesigned role. The shape of most good engagements looks like that — not like the keynote.