Computer vision in manufacturing: where it fits on a line, and where it doesn’t.
For plant, quality and EHS managers weighing cameras for inspection or safety. Which jobs suit vision, what it cannot do, and how to test it on your own line before anyone relies on it.
What is computer vision in manufacturing?
Computer vision in manufacturing is the use of cameras and image-analysis software to check products, processes and safety conditions on a production site. Common uses are visible defect checks, label and packaging checks, and safety events such as a blocked exit. The software flags what it sees; a quality operator or supervisor decides what happens.
Where does vision fit on a line?
Pick a job. Each runs through six steps, and a named person decides at the step that matters.
- 12 Label position and print read
- 23 Compared with the product recipe
- 34 Skewed label flagged as a hold
- 45 Operator releases or rejectsA person
- 56 Result added to the lot record
Step 4 is where a person decides.
- 12 Seal and flaps checked
- 23 Open flap flagged
- 34 Unit held, line keeps running
- 45 Operator checks the cartonA person
- 56 Reason code recorded
Step 4 is where a person decides.
- 12 Image taken under fixed light
- 23 Missing clip suggested
- 34 Region marked on the image
- 45 Operator confirms or overrulesA person
- 56 Overrules fed back for review
Step 4 is where a person decides.
- 12 Object in the exit route
- 23 Still there after the set time
- 34 Clip sent to the review queue
- 45 Supervisor confirms or dismissesA person
- 56 Confirmed: action with a due date
Step 4 is where a person decides.
- 12 Person enters the defined zone
- 23 Hard hat not seen
- 34 Clip and confidence sent
- 45 Supervisor reviews the clipA person
- 56 Logged by zone, not by person
Step 4 is where a person decides.
Illustrative workflows for a fictional plant. Station flows follow /guides/visual-quality-inspection/; safety flows follow /guides/ppe-detection/.
Which plant jobs suit computer vision?
Visible, repeated checks in a fixed place suit it; judgment calls and safety functions don’t.
General guidance, not a test result. Whether vision works on your line is known only after a feasibility review and a test on your own parts or footage.
What can go wrong with vision on a plant floor?
- Light and dirtSunlight through a door, a dusty lens or a new lamp changes what the camera sees. Stations need controlled lighting and a start-of-shift check.
- New variantsA new label or part version the system never saw can be passed or held wrongly. Treat every change as a change request, tested before use.
- Not inspected is not a passA blurred or missing image must send the unit to a person, never through.
- Not a safety functionVision suggests. It can miss events and raise false alarms, and it never replaces guarding, emergency stops, safety professionals or mandated inspections.
- Watching peoplePoint cameras at product and places. Do not score, track or identify workers; report safety events by zone and shift.
How do you know it works on your line?
- EscapesDefective units the system passed. Plant known defects in a test run and count how many get through.
- False rejectsGood units it held. Too many and operators stop trusting it, so track them by shift and product.
- Review loadHow many holds or alerts a person reviews per shift, and how many were real.
- Not inspected countUnits or minutes with no usable image. A rising count usually means a lighting or camera problem.
What should you agree before cameras go in?
- Who decidesA quality operator decides every hold and reject; a supervisor decides every safety alert. Write the names down.
- Tell your workforceExplain what each camera checks, where and why, and consult worker representatives where required.
- Keep littleKeep only the images and clips review needs, for as long as review needs them, and mask faces where possible.
- Release stays yoursLot release and quality records stay in your own quality system and processes.
Which standards and rules are worth reading?
- ISO 2859-1, Sampling procedures for inspection by attributeshow sampling inspection plans are set and read
- EMVA 1288, Standard for characterization of image sensors and camerascomparing machine-vision camera performance on a common basis
- Regulation (EU) 2016/679 (GDPR), Article 35when a data protection impact assessment is needed before monitoring
How do you start with vision in a plant?
- 1Pick one visible problemOne defect family or one safety event that people check today, with a clear rule for what counts.
- 2Check the station or cameraA fixed, lit station for product checks, or an existing camera with a clear view for safety.
- 3Collect labelled imagesGood and bad examples from your line, borderline ones included, checked by whoever decides today.
- 4Run it in shadow firstThe system records while people inspect as before; compare results daily before it advises anyone.
Which Agentra products use computer vision?
None is available yet. One is in discovery and two are planned.
Last updated
Questions, answered.
The use of cameras and image-analysis software to check products, processes and safety conditions on a production site, such as labels, seals, missing parts or a blocked exit. The software flags what it sees, and a person decides what happens.
Machine vision usually means industrial cameras running fixed rules, such as measuring or reading codes. Computer vision with machine learning learns from labelled images, which helps where defects vary in look. In practice the words overlap, and many stations use both.
Sometimes, for safety events such as a blocked exit, after a suitability test on your own footage. For product inspection, usually not reliably: defect checks need a fixed camera, lens and controlled lighting at the station.
It depends on the defect, the parts, the lighting and the station, so any general figure is meaningless. Measure escapes and false rejects on your own line in a test run. We will not quote accuracy before it is measured there.
It shouldn’t at the start. Run it in shadow mode, then as advice to an operator who decides every hold and reject. It is never part of a machine’s safety function.
It shouldn’t be. Point cameras at product and places, report safety events by zone and shift, and never identify, score or rank individual workers.
No. It takes repetitive visual checks off their shift. People still decide holds, rejects and lot release, and handle anything the system was not tested for.
Not yet. Visual Quality Inspection and AI Safety Monitoring are planned, and AI Inspections, which reviews inspection photos, is in discovery. None is available to buy.
Bring us one process. We’ll show you how it would run.
A 30-minute call with a specialist: what it connects to, what a pilot involves, and an honest view on fit, including when it isn’t one.
- 30 minutesOne call, one process
- No commitmentNo sales script
- 4 business hoursTypical reply time · Mon to Fri, 7:00 to 22:00 IST
- 1You get a short agenda by email.
- 2We look at one process with you and how it would run.
- 3You get an honest view on fit, and what a pilot would involve.