Computer Vision for Quality Inspection: From Pilot to Production Line
The physical setup determines your ceiling
Model accuracy is capped by image quality, and image quality is decided by lighting, camera placement and vibration — none of which are software problems. A fixed camera with controlled diffuse lighting and a consistent background will outperform a better model shooting under changing daylight from a shaky mount. Spend the first week on the rig, not the network. On several projects, moving a light and adding a shroud improved defect detection more than any modelling change we made afterwards.
You will not have enough defect images
That is the normal situation on any line running well — defects are rare by design. Two approaches work. Anomaly detection trains mostly on good parts and flags deviation, which suits low-defect-rate lines and catches novel defects nobody catalogued. Supervised classification needs examples of each defect class and is better when defect types are known and must be distinguished for routing. Many production systems run anomaly detection as a first pass and classification on flagged items only.
Set the threshold with the cost of each error
A false negative that lets a defective part reach a customer and a false positive that pulls a good part off the line have very different costs, and those costs are specific to the plant. Get finance and quality in a room and put numbers on both, then tune the decision threshold to minimise total cost rather than maximise accuracy. This one conversation prevents the most common post-launch complaint, which is that the system flags too much and operators start ignoring it.
Plan for drift from day one
Suppliers change, materials shift, a lens gets dusty, someone adjusts a light during a night shift. Accuracy decays quietly. Production systems need a sampled human audit — a small percentage of decisions reviewed daily — and an alert when the flag rate moves outside its normal band. Retraining is not an emergency response; it should be a scheduled routine with fresh labelled data collected continuously by the audit process.
Edge or cloud is a bandwidth and latency question
If the system must reject a part within milliseconds while it moves down a conveyor, inference runs on the line. Modern industrial edge hardware handles this comfortably for most inspection models. The cloud's role is aggregation: storing flagged images, tracking defect trends across shifts and sites, and retraining. Sending raw video off-site continuously is usually neither necessary nor affordable, and in some jurisdictions raises its own questions.
Roll out alongside people, not instead of them
The pilots that stick run the vision system in parallel with existing inspection for several weeks, showing operators what it flags without acting on it. That builds trust, exposes edge cases and produces labelled disagreements that improve the model. Switching to automatic rejection on day one, before anyone believes the numbers, is how these systems end up quietly switched off.
Have a project that needs this kind of thinking?
Let's talk through what you're building — free consultation, no commitment.
Get in Touch