Computer Vision on the Factory Floor: Lessons from 92% Defect Detection

Vision-based quality inspection is one of the highest-ROI AI applications in manufacturing — and one of the most underestimated in delivery complexity. Here's what we learned reaching 92% defect detection accuracy on precision components.

The model is 20% of the work

Modern architectures make defect classification almost easy. The hard parts are physical: consistent lighting rigs, camera positioning that survives vibration, and capturing enough examples of rare defects. We spent more time on lighting engineering than on model architecture.

Labels are your real asset

• Quality engineers, not annotators, must define defect taxonomies
• Borderline cases need adjudication rules, or your labels encode disagreement
• Every disputed prediction becomes a labeled example — the flywheel that compounds accuracy

Run inference at the edge

Cloud round-trips don't survive takt time. We deploy quantized models on industrial edge hardware, inspecting parts in under 80ms, with the cloud handling retraining and fleet management rather than inference.

Plan the human transition

Inspectors don't disappear — they move up the value chain to adjudicating machine-flagged cases and hunting root causes. Plants that involve quality teams from day one hit adoption targets; plants that impose the system from above fight it for a year.