OEM

Automated visual inspection in manufacturing: what AI catches that people miss

Oct 09, 2026

7 min read

Automated visual inspection uses cameras, controlled lighting and software, increasingly AI, to check every part on a production line for defects without relying on a person’s eyes. It matters because human inspection is less reliable than it looks: a literature review by Sandia National Laboratories found that error rates of 20% to 30% are frequently quoted for visual inspection, and that even 100% inspection does not catch every defect. An automated system applies the same standard to every part on every shift, finds defects too small or too brief to see at line speed, and records where each one is, so the cause can be fixed upstream.

Why people miss defects

The Sandia review, by human factors researcher Judi See, pulls together decades of inspection studies. Its findings explain most of what goes wrong on a line:

Misses outnumber false alarms. Inspection errors are more likely to be a missed defect than a good part wrongly flagged.
Speed costs accuracy. In one study, when the rate of inspection was doubled, misses rose from 23% to 30%.
Attention fades fast. Defect detection can fall by up to 40% within 30 minutes. In a study of rubber seals for automotive use, experienced inspectors’ hits fell 27% from the first 15 minutes of a session to the second.
Inspectors disagree, even with themselves. Individual inspectors found between 43% and 100% of solder defects, and no defect was found by all of them. When piston rings were submitted twice, 23% of decisions were reversed.

A later Sandia study of precision-manufactured parts found that inspectors correctly rejected 85% of defective items but also rejected 35% of acceptable parts, against an industry average hit rate of 80% for visual inspection. In other words, even a good human inspection lets roughly one defect in five through, or scraps good parts to avoid it.

None of this is a criticism of inspectors. It is what happens when people do a repetitive search task, at speed, for hours.

How automated visual inspection works

Imaging. Cameras are positioned so every surface that matters is seen, with lighting designed for the defect: diffuse light for colour and contamination, structured patterns for reflective surfaces, UV for leaks and some coatings, lasers for dimensions.
Detection and classification. Software finds anomalies and says what each one is and where.
Decision. Pass or fail goes to the line control or manufacturing execution system (MES) within seconds, and failed parts are routed to rework with the defect already documented.
Record and root cause. Every defect is logged against the part or vehicle ID, so trends point back to the station, tool or material that caused them.

Rule-based machine vision and AI

Classic machine vision is rule-based: measure this edge, check that hole, compare against a reference image. It works well for features you can define precisely under stable conditions, such as whether a part is present or a dimension is within tolerance.

Deep learning is better at defects that vary every time: scratches, dust, dents, texture differences, contamination. Instead of rules, the model learns from labelled examples, and it copes better with variable lighting and reflective surfaces. Many systems combine both, using rules where the answer is a measurement and AI where it is a judgement about appearance.

Choosing the right sensor

The camera is not always the answer. Two cases from car making show why:

Reflective surfaces. On painted panels, a plain photo shows the reflection of the surroundings rather than the surface. Deflectometry projects a known pattern and measures how the paint distorts it, which Fraunhofer IOSB describes as automating what an inspector does when holding a surface in the right light. We cover the defects it finds in automotive paint defects.
Dimensions. 2D cameras are sensitive to lighting, reflections and viewing angle, so gap and flush readings between panels vary. Laser profilers capture a cross-section at each measurement point, which gives simpler data and more repeatable results.

What automated inspection catches that people miss

Small defects. On painted car bodies, CamCom’s deflectometry-based inspection finds surface defects down to 50 microns, far smaller than an inspector can reliably spot on a moving line.
Defects that only show at certain angles. Dents and craters on reflective paint that a camera, or an eye, misses in plain light.
Geometry. Gap, flush and alignment between panels, measured in millimetres rather than judged by eye or a gauge.
Consistency. The same threshold at 7am and at the end of a night shift.
Patterns. A defect type that keeps appearing on the same panel or station, which no single inspector sees across shifts.

In a two-phase proof of value with Volvo Trucks, CamCom replaced variable 2D camera readings of truck cabin gap and flush with robotic laser dimensioning. Robots on linear tracks either side of the cabin take laser cross-sections at each panel interface, and a full multi-point inspection of a commercial truck cabin takes about 5 minutes (Volvo Trucks case study).

Where automated inspection is hard

Rare defects. A model needs examples. New defect types, or ones that occur once in thousands of parts, take time to learn, and the system should flag what it hasn’t seen rather than pass it.
Ground truth. Accuracy can only be measured against expert labels, and experts disagree, as the Sandia studies show. Agree the defect standard before measuring the system against it.
Integration. A system that detects defects but can’t stop the part or write to the MES creates another screen to watch, not an inspection.
Changeovers. New models, colours and materials need to be planned into training and validation.

How to evaluate an automated visual inspection system

Test on your parts and your defects, including the rare ones, not on a vendor’s sample set.
Measure misses and false calls separately. A missed defect reaches the customer; a false call costs rework time. One overall accuracy figure hides the trade-off.
Check it keeps up with takt time at full line speed, not in a lab.
Confirm how it writes to your MES and what traceability it keeps against each part or VIN.
Ask how new defect classes are added, how long it takes, and who labels the data.
Start with one station where escapes cost the most, then expand.

For vehicle production lines, CamCom offers this as Modulus by CamCom: a retrofit inspection gantry that fits into an existing line and inspects a vehicle in as little as 20 seconds, with pass or fail written to MES platforms such as Siemens Opcenter, SAP Digital Manufacturing and Rockwell FactoryTalk. See AI paint and body defect inspection.

Common questions

What is automated visual inspection?
The use of cameras, controlled lighting and software, increasingly AI, to check products for defects on a production line without relying on human eyes. Each part gets the same check, and every defect is recorded with its type and location.

How accurate is human visual inspection?
Less than it looks. A Sandia National Laboratories literature review found error rates of 20% to 30% are frequently quoted, and a later Sandia study cites an industry average hit rate of 80%. Accuracy also falls with speed and time on task.

What is the difference between machine vision and AI inspection?
Classic machine vision follows rules, such as measuring an edge or comparing against a reference image, and suits precise, stable checks. AI inspection learns from labelled examples and handles defects that vary in appearance, such as scratches, dust and dents. Many systems use both.

Can automated visual inspection keep up with line speed?
Yes, if it is designed for the line. Systems synchronise with takt time and send pass or fail decisions to the line within seconds. CamCom’s Modulus system, for example, can inspect a vehicle in as little as 20 seconds.

Sources: Judi E. See, Visual Inspection: A Review of the Literature, Sandia National Laboratories, SAND2012-8590, October 2012. Judi E. See, Visual inspection reliability for precision manufactured parts, Human Factors, vol. 57, no. 8, December 2015. Fraunhofer IOSB, Deflectometry for the inspection of specular surfaces. CamCom, Volvo Trucks case study and Modulus by CamCom.