Computer vision for quality control: what is realistic in 2020
A plain assessment of where computer vision actually delivers in factory quality control right now - and what the common misconceptions are about cost and scope.
Visual inspection is one of the clearest application areas for AI in manufacturing. The task is repetitive, the criteria are usually well-defined, the volumes are high, and human fatigue is a real problem. On paper, this is a near-ideal use case for machine learning.
In practice, I regularly see projects that start with enthusiasm and arrive at a result nobody is satisfied with. The technology works - but the problem is usually not where people assumed it was.
What computer vision for QC actually does
The typical system uses a camera - sometimes a standard industrial camera, sometimes something more specialised - and a model trained on images of acceptable and defective parts. The model flags parts that do not match the acceptable pattern. An operator reviews the flagged items or, in more mature setups, the system makes the accept/reject decision directly.
The well-functioning version of this is not futuristic. It has been in automotive supply chains for several years. What has changed recently is that the threshold for setting it up has dropped: you no longer need to build the computer vision stack from scratch, and the hardware costs less.
Where it works well
Three conditions make visual inspection a strong fit for automation:
- The defect is visual and consistent. Surface scratches, missing components, incorrect colour coding, dimensional non-conformance that is visible in an image - these are addressable.
- The parts have low variance. If you are inspecting one type of connector in one orientation, the training data requirement is manageable.
- The lighting and positioning are controlled. This is the engineering constraint people underestimate most. Variable lighting and inconsistent part positioning are the most common reasons a model that looked good in testing fails in production.
Where it struggles
Computer vision QC does not work well for defects that are invisible to cameras (internal cracks, material density issues), for highly varied products inspected in low volumes, or in environments where setup discipline is low.
The failure mode I see most often: a pilot succeeds under controlled lab conditions, then fails when deployed to the actual production line because positioning and lighting were not standardised. The cost of standardising the physical setup is frequently larger than the cost of the model itself.
The cost picture
A realistic entry-level system for a single inspection point - camera, lighting rig, hardware for running the model, integration with the line - costs somewhere between 20,000 and 80,000 EUR depending on the specifics. This is not a precise number, but it gives a sense of the range.
Training and customisation add time, not usually cost at this scale. Off-the-shelf platforms (Cognex ViDi, Landing AI, and several others) have reduced the time to first working model substantially. The remaining effort is mostly on the physical and process side.
The honest starting point
Before engaging with any vendor, have your quality team document: what defects you want to catch, at what rate they currently occur, what the cost of a missed defect is, and what the inspection point physically looks like. That documentation shapes every subsequent decision and prevents the most common failure mode - building for a problem that was never precisely defined.