Every vendor demo looks convincing. Someone holds up a defective part, the screen flashes red, the room nods. Then the system arrives, meets your actual lighting and your actual part variation, and the conversation gets more complicated. Evaluating visual inspection systems well means looking past the demo part and asking what happens on a Tuesday afternoon with a supplier change and a new operator.
This is a buyer’s guide, not a spec sheet. The goal is to help you ask better questions before the purchase order.
What Visual Inspection Systems Are Actually Made Of
Treat any quote as five separate purchases bundled together, and price each one in your head.
- Optics and lighting — lens, working distance, and the illumination that makes the defect visible in the first cybercrime
- Capture hardware — camera resolution and frame rate matched to your line speed
- Compute — where inference runs: cloud, on-premises server, or an edge device beside the line
- The model — what it was trained on, and how you retrain it when your product changes
- Integration — how a result reaches a PLC, a reject gate, an operator screen, or your quality system
Weak links are rarely in the model. They are usually in lighting or integration, and both are cheaper to fix before installation than after.
Rule-based machine vision or AI-based inspection
Both still have a place, and confusing them leads to disappointment.
| Rule-based machine vision | AI-based inspection | |
| Best for | Measurement, presence/absence, reading codes, fixed geometry | Cosmetic and variable defects, “looks wrong” judgments |
| Setup | Configure parameters and tolerances | Collect and label example images, train |
| Handles variation | Poorly — needs tight control | Well, if the variation appears in training data |
| Explaining a decision | Straightforward | Needs review tooling and thresholds |
| Changing the product | Reconfigure | Retrain with new examples |
If a caliper could do the job, you probably want rules. If your current standard is a veteran operator saying “that one’s not right,” you want a learned model.
The questions that separate serious options
Who can retrain it? If every product change means a service ticket and a vendor visit, your real cost is not the licence. Systems where your own team can add images and retrain age far better.
How much labelled data does it need to start? Ask for a realistic number for a defect like yours, not a best case.
Can I see it fail? Ask the vendor to show a false positive and explain how you would fix it. The answer tells you more than any accuracy claim.
Where does inference run, and what happens when the network drops? For anything touching a reject mechanism, that is a hard requirement, not a preference.
What does the audit trail look like? You will eventually need to prove why a specific part was accepted, possibly to a customer.
Be sceptical of a single accuracy number
Accuracy quoted without a dataset description means little. A system that is highly accurate on a heavily-defective test set may behave differently on your line, where the overwhelming majority of parts are fine. Ask separately about missed defects and false alarms, because they cost your business different amounts. Whichever one hurts more should drive how the system is tuned.
Prove it on one line first
Run a parallel trial. Keep your existing inspection in place, let the system watch the same parts, and compare disagreements case by case. Every disagreement is either a defect your people missed or a false alarm you now know how to tune — both useful.
Agree success criteria before the trial starts, in writing, and include an operator-experience measure. A system that hits its accuracy target while generating nuisance alerts every few minutes has not succeeded.
Budget for the second year
Visual inspection systems are not appliances. Products change, suppliers change, cameras get bumped, lighting ages. Plan for periodic retraining, a named owner, and a simple routine for adding fresh images when something new shows up. The projects that quietly fail are rarely the ones that failed at launch — they are the ones nobody maintained.
Conclusion
The best of the visual inspection systems on your shortlist is the one your own team can still operate and improve a year after the vendor’s engineer has gone home. Judge on retrainability, honest failure behaviour, deployment fit, and the audit trail — not on how cleanly the demo part triggered a red box. Start on one station, define success before you begin, and expand only once the people running the line believe the results.
