AI print inspection system

Context

High-speed print lines produce more surface than a person can watch. Inspection was manual and sampled: an operator pulled sheets, examined them, and made a call about the run they represented. Defects that appeared between samples were found by the customer.

The brief was to detect defects on the line itself, across several print surfaces, without slowing the line down and without flooding operators with false alarms.

What was hard

The surfaces disagree. A defect that is high-contrast on one substrate is nearly invisible on another. A single detector tuned on one surface degrades badly on the rest, and averaging across them produces a model that is mediocre everywhere.

Speed sets the budget. Inspection has to keep pace with the line, which caps how much preprocessing and how large a model is affordable per frame.

False positives are expensive twice. They cost operator attention now, and they cost trust later — an alerting system operators learn to ignore is worse than no alerting system, because it looks like coverage.

Approach

Preprocessing normalises what varies between substrates before the detector sees it: illumination correction and contrast handling in OpenCV, so the model is learning defect shape rather than surface brightness.

Detection is YOLOv8, trained across the print surfaces in scope rather than one at a time, so a single deployed model generalises instead of a family of models that each need their own maintenance.

Beyond detection, the fault patterns matter more than the individual hits. More than 700 print sheets were analysed with defect clustering, which turns a stream of isolated alerts into recurring pattern groups — the same recurring fault, recognised as one problem rather than four hundred.

Outcome

• Detection performance improved from 69 to 95 mAP across the print surfaces in scope — 38 points.
• Manual review time fell by 40%.
• Fault-pattern identification accuracy rose by 25% through defect clustering across 700+ sheets.
• Deployed into live inspection with real-time alerts, improving traceability by 35%.

What I'd change

The clustering step earns its keep and deserves to be first-class rather than analysis run alongside the detector. Folding pattern assignment into the serving path would let the line surface this is the same fault as an hour ago instead of a fresh alert each time, which is the difference between a detector and a diagnostic.

The other gap is drift. The model was validated against the surfaces in scope at the time; a production vision system needs a standing check that today's distribution still resembles the one it was trained on, and an alert when it does not.