A vision system for quality control is a station that looks at the product with a camera and decides: good or defective. Plants call it different things — visual inspection, vision inspection, machine vision — but the job is always the same: to replace or support the human eye where inspection has to be repeatable, fast and documented.
This guide brings together what is worth knowing before you decide: how such a system works, what it consists of, how to judge its results, what drives the cost and how to start without much risk. We go deeper into each topic in separate articles, linked along the way.
What visual inspection is for
Most often, detecting defects a person can see with the naked eye but cannot reliably catch on every part and every shift: missing components, incorrect assembly, flash, cracks, label errors and contamination. We describe each of these tasks in Applications, and typical industries — plastics, automotive, metal components and packaging — in Industries.
The value of the system is not just that it "sees". It is that the same decision logic works in the first and the last hour of a shift, and every decision is recorded — with the image, the model version and the settings responsible for it.
How the system decides: rules or a model trained on examples
A classic vision system works on rules: brightness thresholds, edge detection, template matching, distance measurement. It works well when the defect can be described geometrically and the product and lighting are stable. Every change in the product or conditions, however, means retuning the rules.
An AI-based system learns from examples of OK and NOK images. It copes where a defect is hard to describe with a rule — irregular shapes, subtle surface damage, high natural product variability. In return it needs representative data and sound validation. In practice the two approaches complement each other. How to choose between the available forms — a smart camera, an all-in-one AI kit, a cloud platform and software on your own hardware — is covered in Smart camera or AI system?.
What a visual inspection station consists of
- Camera and optics — they must show the defect clearly enough and always in a similar way. Resolution, lens and distance are chosen for the smallest defect that has to be detected.
- Lighting — often more important than the camera. The same scratch is clear at one angle of light and disappears at another; we cover this in Lighting in visual inspection.
- Positioning and triggering — the product must reach the camera in a repeatable position, and the station must know when to take the image: on a sensor signal, from the PLC or in continuous mode.
- Computer and software — they analyse the image and make the decision within the line's cycle time.
- Line integration — the result must reach the PLC, which rejects the defective part or stops the line. A few signals are usually enough, for example over Modbus TCP.
- Data and version management — the training image set, the inspection history and control over which model version runs on which station.
Where the model comes from
In an AI-based system the most important ingredient is data. The model learns to tell a defect from acceptable variation, so it has to see that variation — different batches, shifts and defect severities. How many images that means is covered in How many images do you need to train an AI inspection?, and how not to fool yourself when judging the result in A dataset that does not cheat.
Training itself no longer has to require a machine-learning specialist. In AIVQC the quality engineer labels defects, and automatic training chooses the model and parameters. The result is checked on an independent test set before anything reaches the line.
How to judge whether the system works
The question "what percentage does it detect" is not enough. Two kinds of error matter: a false accept, meaning a defective part let through, and a false reject, meaning a good part rejected for no reason. Both cost money, but differently, and the decision threshold is set to balance them according to the process risk. Details in False accept or false reject?.
Just as important is what the station does when it is not sure. A good system does not guess: it flags the part for a person to check or signals a fault, instead of quietly letting products through. In AIVQC that is what the REVIEW and ERROR results are for, alongside OK and NOK.
What drives the cost
The cost of a vision system is not just the camera. The total cost is made up of:
- hardware: camera, optics, lighting, computer and mounts — unless existing equipment can be reused after qualification (retrofit or new station),
- line integration and commissioning,
- people's time: collecting data, annotation, training and validation,
- maintenance: new product variants, new defects, updates,
- the cost of errors after start-up: missed defects and false rejects.
The biggest difference between approaches is usually not the hardware but who does the work at launch and at every change. If an integrator does it, you pay for their hours every time. If your own team does it, the cost becomes work by people you already employ — that is what deploying with your own team looks like. How to work out the return using data from your own line is covered in How to calculate the return on automating quality inspection.
How to start without much risk
The safest start is one station and one well-defined problem — the defect that costs the most today. Before investing in a full deployment, check on your own line whether the defect is detectable under the available imaging conditions. That is what a pilot is for: in AIVQC it is 30 days on one Trainer workstation and one station, with up to three defect classes and a free extension available. We describe it step by step in AI quality-inspection pilot.
Before you start, it is worth preparing a few things:
- Pick one defect and describe what is still OK and what is already NOK.
- Start archiving defective parts — physically, not just as photos.
- Work out what this defect costs today: complaints, sorting, inspectors' time.
- Check whether the product reaches the camera in a repeatable position.
- Decide who on the team will own the system: the quality engineer, IT and the automation engineer.
Have a question this guide did not cover? Look in the FAQ or get in touch — we will help assess whether your case is a good fit for visual inspection.