← All posts
GuideBy Filip Molčík · Aug 31, 2026 · 6 min read

How Vision Systems and Machine Vision Work in Manufacturing

A vision system combines industrial cameras, lighting and software to inspect every product on a production line in real time. Below is how the hardware and software work together and where they pay off.

Abstract 3D illustration of a light signal passing through a row of glass panels, like the layers of a neural network

Walk through a bakery, pharma packaging hall or bottling plant and you will see a camera in a metal housing watching products pass at several pieces per second. That camera, together with its lighting, optics and the software behind it, is a vision system. It's one of the least glamorous and most useful pieces of automation on the market, and the industry around it is far from niche: analysts put the global machine vision market at roughly 13–15 billion USD in 2025, growing at 7–9 % a year. This article explains what a vision system actually is, how it works, and where it earns its keep.

Difference between computer vision, machine vision and a vision system

The terms get mixed up constantly, so let's untangle them:

  • Computer vision is the broad research field: teaching computers to extract meaning from images. It covers everything from medical imaging to self-driving cars.
  • Machine vision applies computer vision to factory inspection, measurement, robot guidance and code reading in real time.
  • A vision system is the concrete, deployed package: camera(s), lens, lighting, a processing unit and software, integrated with the machine or line it serves.

In other words, computer vision is the science, machine vision is the discipline, and a vision system is the thing bolted above your conveyor.

Tasks a vision system handles on the line

Most industrial applications fall into four families, often abbreviated GIGI — guidance, identification, gauging, inspection:

  • Inspection checks that the product is complete, the surface is free of cracks or contamination, the seal holds, the cap sits straight and every blister pocket is full, then sorts each piece into OK or NOK.
  • Identification — reading barcodes, DataMatrix codes, printed text (OCR) such as lot numbers and expiry dates, and verifying that the right label is on the right product.
  • Gauging (measurement) — non-contact measurement of dimensions: diameter, gap, fill level, position of a component, all to fractions of a millimetre.
  • Guidance — telling a robot or actuator where something is, so it can pick, place or align it.

Counting is a fifth, humble but common task: how many biscuits in the tray, how many bottles in the crate, how many pieces went past this shift.

The economic argument is simple. Humans are surprisingly poor at repetitive visual inspection: studies of manual inspection tasks report error rates of 20–30 %, and performance degrades further with fatigue, shift changes and line speed. A vision system inspects piece number 100,000 with exactly the same attention as piece number one — and it never looks away.

A vision system checks 100 % of output instead of one piece in a hundred, applying the same criteria at full belt speed on every shift.

How a vision system works, step by step

Every vision system, from a smart sensor to a multi-camera cell, runs the same basic loop:

  1. Trigger. A photoelectric sensor, encoder pulse or PLC signal tells the system "a product is in position now". At typical line speeds, the timing window is a few milliseconds.
  2. Image acquisition. The camera captures a frame. Lighting, lens choice and exposure time decide whether the defect is visible in the frame at all. Backlight for silhouettes and fill levels, low-angle light to make scratches pop, diffuse domes for shiny foil: lighting design is half the job.
  3. Processing. Software analyses the image. Classic rule-based tools measure edges, count blobs, match templates and read codes. Deep-learning models classify products or flag anomalies on variable materials such as bread, meat and natural wood, where no two good pieces look identical.
  4. Decision. The result is condensed to something actionable: OK/NOK, a measured value, a decoded string.
  5. Action. The system tells the line what to do — fire a pneumatic ejector to reject the NOK piece, stop the machine, or simply log the result. This is why integration with the PLC (or using a controller that is itself an AI-capable PLC) matters as much as the AI.

All of this typically happens in tens of milliseconds. At three pieces per second, the full trigger-to-reject loop must finish within 300 ms minus mechanical actuator delay.

Two approaches compared, rule-based tools and AI models

Traditional machine vision is deterministic. You define rules: this edge must be here ±0.2 mm, this region must contain at least N dark pixels, this code must decode. It's fast, explainable and excellent for geometry, measurement and code reading — and it's been running reliably in factories for four decades.

Its weakness is variability. Rule-based systems struggle when good products naturally vary (food is the classic example), when lighting drifts, or when defects can't be exhaustively described in advance.

Deep learning works from examples instead of written rules. You train the model on photos of OK and NOK pieces so it learns normal product variation, while anomaly-detection models need mostly OK images to flag unfamiliar defects. Deep-learning models handle lighting shifts and shape variation better than template matching or hand-tuned rules, in exchange for training data and extra compute.

In practice, modern systems combine both: rules for measurement and codes, neural networks for surface defects and classification. Edge AI hardware runs inference directly at the line without a cloud link, network delay or production images leaving the food or pharma plant.

What deployment actually looks like

Both in-house builds and camera subscriptions such as Robopipe follow the same commissioning steps:

  • Define the defect. Collect real NOK examples and agree with quality what counts as a reject. Vague criteria ("looks bad") are the number-one cause of failed projects.
  • Fix the imaging. Mounting position, lens, lighting, shrouding against ambient light. In wet areas and washdown environments, IP-rated housings are non-negotiable.
  • Collect data and train. For AI-based inspection, a few hundred to a few thousand images are usually a workable starting point; the model then improves as operators confirm or correct borderline cases.
  • Integrate with the line. Trigger input, reject output, and ideally OK/NOK counts flowing into your MES or a simple dashboard so the numbers are visible beyond the line itself.
  • Validate and hand over. Run the system in shadow mode alongside manual inspection, compare results, tune the threshold between false rejects and escapes, then let it take over.

An industrial camera and edge AI controller now cost a fraction of what a vision cell cost a decade ago. The remaining effort is in lighting, data and integration, which is exactly why turnkey and subscription models have emerged: they shift that engineering burden off the plant team.

The bottom line

A vision system is a camera, controlled lighting, software that makes an OK/NOK decision and an output wire that triggers within milliseconds. Deep learning and lower-cost edge hardware now bring 100 % inline inspection to mid-sized plants handling variable products. Add up the cost of customer complaints, rework and manual end-of-line checks to see how fast a line camera pays back.

See what a vision system catches on your own line

Book a demo