The field manual for machine learning in manufacturing · updated September 1, 2026
ManufacturingML
Home / Encyclopedia / Computer-vision quality inspection
Applications · Entry · updated September 1, 2026

Computer-vision quality inspection

Key takeaways
  • Vision-based inspection classifies defects at line speed and leads industrial AI adoption with roughly 11% of use cases (IoT Analytics).
  • Deep learning replaced rule-based machine vision for cosmetic and variable defects; classic rules still win for gauging and measurement.
  • It's supervised ML: performance tracks the size and quality of your labeled defect library.
  • Vision tells you a defect exists; process ML tells you why it keeps happening — they're complementary layers.

Automated optical inspection is industrial AI's beachhead: the use case with the clearest before/after, the most mature tooling, and the largest share of deployments — about 11% of the industrial AI market by use case, per IoT Analytics. A camera and a trained classifier inspect every unit at line speed, where human inspectors sample and tire.

Deep learning vs. classic machine vision

Rule-based vision (edges, blobs, gauges) still owns dimensional measurement — it's deterministic and auditable. Deep learning owns everything that varies: scratches, textures, welds, surface cosmetics, assemblies photographed at awkward angles. The practical systems mix both, and the vendors' no-code training tools have made retraining for a new defect a technician task rather than an integrator project.

The data appetite

This is supervised learning: it needs labeled defect examples, and rare defects are — by definition — rare. Programs that succeed treat the defect library as an asset: every escape and every false reject gets labeled and fed back. Synthetic augmentation helps stretch small libraries; it doesn't replace a feedback discipline.

Vision finds it; process ML explains it

An inspection system counting the same defect all week is a detection success and a prevention failure. The mature pattern wires vision output into process-optimization ML as another variable, so the platform can correlate defect rates with upstream process behavior — closing the loop from “caught it” to “stopped making it.”

FAQ

How many defect images do we need to start?
Less than the folklore says: modern tooling starts usefully at tens-to-hundreds of examples per defect class and improves with feedback. The real requirement is the discipline of labeling escapes and false rejects continuously.
Does vision inspection remove inspectors?
It moves them: from staring at parts to adjudicating edge cases, tuning the system and running the feedback loop. Plants typically redeploy inspection labor toward prevention rather than cutting it.
Related entries
Comparing platforms that do this? The adoption data pack carries the market numbers behind this entry.