What is Vision?
Vision, in full machine vision, is the industrial technology in which cameras, lighting and image-processing software automatically examine products or processes and derive a decision from them, such as pass/fail, a dimension, a position or a decoded code. The difference from ordinary image processing is that the output is not a new image but information that a PLC, robot or MES can act on directly. Machine vision is an applied systems-engineering discipline; computer vision is the broader scientific field that supplies many of its algorithms.
In production environments vision is used mainly for quality inspection, measurement, code reading and robot guidance, delivering both a control signal and valuable production data.
🧠 How does a vision system work?
Every inspection runs through the same chain, often within a few tens of milliseconds:
- Trigger - a sensor or encoder signals that a product is in position
- Illumination and capture - the lighting strobes and the camera captures one image (or a sequence of lines)
- Pre-processing - filters remove noise and increase contrast
- Analysis - algorithms find edges, blobs or patterns, measure distances, read codes or classify with a neural network
- Decision - the result is compared with tolerances: pass or reject
- Output - a digital signal or network message drives a rejector, passes coordinates to robotics or stores the result for tracking and tracing
🔧 What are the components of a vision system?
| Component | Function | Practical choice |
|---|---|---|
| Camera | Converts light into a digital image | Resolution, frame rate, monochrome or colour, global or rolling shutter |
| Lens | Determines field of view, sharpness and working distance | Fixed focal length; telecentric for accurate measurement |
| Lighting | Makes the feature visible with consistent contrast | Ring light, backlight, dome, coaxial, structured light |
| Vision software | Performs image analysis and decision logic | Rule-based tools or deep-learning models |
| Controller or PC | Runs the software and communicates with the line | Smart camera, industrial PC or edge device with GPU |
| Trigger and I/O | Starts capture and passes on the result | Photo-eye, encoder, digital outputs or fieldbus |
Lighting is often underestimated. A well-lit image keeps the software simple; poor lighting cannot be fixed by any algorithm. A backlight, for example, renders the contour of a part sharply, while dome lighting images shiny or curved surfaces without reflections.
📷 What types of vision systems are there?
| Type | How it works | Typical application |
|---|---|---|
| 2D area scan | One complete image per capture | Presence checks, code reading, position detection |
| Line scan | Builds an image line by line as the product moves | Continuous webs: paper, film, steel, textiles |
| 3D laser triangulation | Laser line on the object; its deformation gives the height | Volume, flatness, weld seams, adhesive beads |
| 3D stereo / time-of-flight / structured light | Depth from two cameras, travel time or a projected pattern | Bin picking, depalletising, navigation |
| Smart camera | Sensor, processor and software in one housing | Simple inspections directly on the line |
| PC-based system | Several cameras on one powerful PC | Complex inspections, AI models, high speeds |
Most industrial applications still work in 2D. Laser triangulation is the most widely used 3D technique; in addition, multispectral, infrared and X-ray cameras can reveal features the human eye cannot see, such as moisture in food or inclusions in castings.
🔌 How does a vision system communicate with the control system?
Four interface standards are common between camera and computer:
| Standard | Maintained by | Characteristics |
|---|---|---|
| GigE Vision | A3 (since 2006) | Standard Ethernet, cables up to 100 m, control via GVCP on UDP port 3956 |
| USB3 Vision | A3 (since 2013) | Plug-and-play over USB 3, short cables, low cost |
| CoaXPress | JIIA | Up to 12.5 Gbit/s per coax cable, for very high frame rates |
| Camera Link | A3 | Older frame-grabber interface, still common in existing installations |
GigE Vision, USB3 Vision and CoaXPress build on GenICam, a standard hosted by the EMVA that describes camera features uniformly, so software can address cameras from different vendors in the same way; for Camera Link, GenICam support is optional. To compare cameras, the EMVA publishes the measurement standard EMVA 1288 (Release 4.0 since June 2021), which also forms the basis of the ISO standard ISO 24942.
Results usually reach the PLC via ProfiNET, IP or digital I/O. For integration with MES and data platforms, the companion specification OPC UA for Machine Vision (OPC 40100-1) has existed since August 2019; it defines control, configuration, recipe management and results of vision systems in a standardised OPC UA information model.
🎯 What is vision used for?
- Quality inspection - detecting scratches, cracks, missing parts, wrong labels or fill levels
- Metrology - contactless measurement of distances, angles and diameters, often to a fraction of a pixel
- Pick-and-place and robot guidance - passing position and orientation so a robot or cobot can pick up a product
- Code reading - barcodes, DataMatrix and QR codes, for example to GS1 standards
- Optical character recognition (OCR) - reading and verifying serial numbers, best-before dates and LOT codes
- Sorting and classifying - separating objects by colour, shape or label
Compared with manual checks, vision inspects without contact, objectively and with consistent quality, even at high line speeds, which reduces error rates and waste. Inspection results are also data: reject rates per line or shift feed SPC, process improvement and a digital twin.
🤖 What is the difference between classical vision and AI vision?
| Classical (rule-based) vision | AI-based vision |
|---|---|
| Programmer sets rules, filters and tolerances | Model learns from example images (deep learning) |
| Deterministic and easy to explain | Output is a probability; explanation is harder |
| Strong at measuring and code reading | Strong with natural variation: wood, food, weld seams, cosmetic defects |
| Little computing power needed | Often requires a GPU or AI accelerator |
| Works without training data | Needs hundreds to thousands of labelled images |
In practice an AI model does not keep learning by itself on the line: it is trained offline, validated and then frozen. New defect types require retraining and revalidation. Many modern systems combine both: rule-based tools for dimensions and codes, machine learning for defects that are hard to define. Read more in Vision AI.
🛠️ How do you approach a vision project?
- Define the task precisely - which feature, which smallest defect, which line speed and what acceptable false-reject rate?
- Collect good and bad samples - without real rejects, reliable testing is impossible
- Choose the lighting first - test different illumination techniques before selecting a camera
- Calculate the resolution - a common rule of thumb is at least 3 pixels across the smallest feature you need to see
- Test offline, validate online - measure rejects and false rejects over a representative production period
- Record recipes and versions - every product variant gets a documented inspection recipe
Worked example: with a field of view of 100 mm and a smallest defect of 0.2 mm, you need 100 / 0.2 × 3 = 1,500 pixels in that direction. A camera with 1,920 × 1,200 pixels is then sufficient, provided the product is positioned so that the long side of the image covers the 100 mm.
European context: in pharmaceuticals, inline vision is indispensable in practice. Since 9 February 2019, Commission Delegated Regulation (EU) 2016/161 has required a unique identifier in a 2D Data Matrix (ECC200) on the packs of prescription medicines (with a few exceptions), with a print quality of at least 1.5 according to ISO/IEC 15415. Vision systems read and grade that code on the packaging line. For the integrity of inspection data, the GMP requirements (in the EU including Annex 11) also apply and, for the US market, 21 CFR Part 11.
🔐 What should you consider for security and regulation?
A vision system is a fully fledged computer on the OT network, often running Windows with a web interface and remote access for the supplier. That calls for the same measures as other OT systems:
- Network segmentation - place cameras and vision PCs in their own zone according to IEC 62443; GigE Vision has no authentication or encryption in its basic form
- Patch management - vision PCs run the same software version for years and quickly become legacy
- Remote access - supplier access only through a controlled route, not through a stray modem or 4G router
- USB policy - images and recipes are often moved on USB sticks
On the regulatory side: a vision system that sorts out rejects usually falls into the lowest risk class of the AI Act. When AI vision is used as a safety function, for example to detect people near a robot, heavier requirements apply via the Machinery Regulation (applicable from 20 January 2027), and under the AI Act it is a high-risk system, with obligations from 2 August 2028. Personnel protection also requires safety-rated devices to standards such as IEC 61496; an ordinary inspection camera is not a light curtain. If cameras also film employees, the AVG (GDPR) applies.
❓ Frequently asked questions
What is the difference between machine vision and computer vision?
Computer vision is the broad scientific field that teaches computers to understand images. Machine vision is its industrial application: a complete system of camera, lighting, software and I/O that makes a concrete decision on the production line. Machine vision therefore uses computer vision algorithms but focuses on reliability, speed and integration with the control system.
How fast is a vision inspection?
A simple machine vision inspection with a smart camera usually takes a few to a few tens of milliseconds, allowing hundreds of products per minute to be checked. Line-scan cameras inspect continuous webs at tens of thousands of lines per second. AI models need more computing time, which is often handled with a GPU or edge accelerator.
How much does a machine vision system cost?
A simple code reader or smart camera costs a few thousand euros. A custom inspection system with several cameras, special lighting, integration and validation can easily run into tens of thousands of euros. The payback usually comes from fewer customer complaints, less manual inspection and better process data.
Can vision fully replace a human inspector?
Machine vision is faster and more consistent and does not tire, so it outperforms people on repetitive, well-defined inspections. For rare, new or hard-to-describe defects, human judgement remains valuable. In practice, people often assess the borderline cases that the vision system sorts out.
Why is lighting so important in machine vision?
Lighting determines the contrast between the feature and its background, and without contrast no algorithm can find a defect. A different lighting angle or wavelength can make an invisible scratch suddenly clear. That is why every machine vision project starts with lighting trials.
Which standards apply to machine vision?
For camera interfaces, GigE Vision, USB3 Vision, CoaXPress and GenICam are the main standards, and EMVA 1288 describes how camera sensors are measured. For code quality, ISO/IEC 15415 (2D codes) and ISO/IEC 15416 (linear barcodes) apply. For integration with IT systems there is the OPC UA for Machine Vision companion specification.
📌 In summary
Machine vision lets machines automatically ‘see’ and decide: camera, lighting and software turn an image into a pass/fail signal, a dimension, a position or a decoded code. It is a key technology for quality inspection and traceability within industrial processes, Smart Industry and Industry 4.0, and as a networked device it deserves the same OT security as any other control system.