Can AI Do Quality Inspection With Cameras? Line Speed, Honestly

Can AI do quality control with cameras?

AI can do quality inspection with cameras for any defect you can photograph and label, at line speed, with a false-reject rate you agree in advance.

Yes, at line speed and around the clock, for defects you can show it examples of: cracks, chips, missing parts, wrong labels, colour and print faults. It does not catch what it has never seen, and it lives or dies on lighting. This guide says which inspections work, what a line pays, and how a project is scoped so the false-reject rate is agreed before the camera is bought.

By Tom Joseph · Last updated: 14 September 2026

Which inspections camera AI does well

Inspection	Works when	Watch out for
Surface defects (cracks, chips, scratches)	Consistent lighting and a fixed camera position	Reflective or wet surfaces; defects smaller than a few pixels
Presence and completeness (missing screw, seal, part)	The reference state is stable	Product variants that change the reference
Labels, print and codes	OCR on a fixed field	Curved surfaces, glare
Dimensions and fit	Calibrated camera, known distance	Where a sensor or a gauge measures better than a camera
Colour and coating	Controlled light, calibrated reference	Daylight and lamp ageing drift the colours

What camera inspection does not catch

It does not catch the defect you never labelled: a model trained on cracks will pass a foreign object it has never seen. It does not measure what a camera cannot see: internal voids need ultrasound or X-ray, and a tolerance of a tenth of a millimetre needs a gauge. And a model that rejects 3% of good product to catch 99% of bad product costs more than it saves on some lines, which is why the false-reject rate is an agreed number, not a surprise. Lighting is the whole project: the audit is mostly about the light, the lens and where the camera sits, not the algorithm.

What it costs and how it is scoped

Three fixed steps: a €2,500 audit on the line, which produces the defect catalogue, the lighting and camera plan and the acceptance criterion (detection rate and false rejects on your own product); a €20,000 Production-ready MVP that runs on one product and one camera position next to the inspector for two weeks; and production from €50,000, integrated with the PLC or MES so a reject is handled, not just flagged, and monitored for drift. Excl. VAT, all code and models handed over. See computer vision, AI for manufacturing and the crack detection case.

Frequently asked questions

Can AI replace the visual inspector on the line?

For the defects it has been trained on, at line speed and around the clock, yes; the inspector moves to the edge cases the system flags and to the defects nobody has labelled yet. On most lines that is a reallocation, not a removal.

How many example images does it need?

Fewer than people expect for presence and completeness checks, more for rare surface defects. The audit counts what you have and, where a defect is rare, plans how to collect or generate examples before the MVP.

What accuracy can we expect?

The number to agree is a pair: detection rate and false-reject rate, on your own product, measured over a held-out set the model never saw. The MVP has to meet the agreed pair before anything is built into the line.

Does it work with our PLC or MES?

Yes. A reject has to do something: stop the belt, divert, log to the MES with the image. That integration is scoped in the audit and built in production.

What does camera inspection cost?

At Crux Digits: a €2,500 audit, a €20,000 Production-ready MVP on one product and camera position, and production from €50,000 with PLC or MES integration, all excl. VAT, with code and models handed over. Cameras and lighting are quoted separately at cost.