Computer vision is the broad field of AI that interprets images; machine vision is its industrial use for inspection, measurement and guidance, traditionally with fixed rules, dedicated cameras and controlled lighting. Many Australian plants now run both: AI vision on a production line for faults that vary, and rule-based machine vision for precise, repeatable measurement.
Both terms come up when manufacturers look at automated inspection, and they're often used loosely. This guide covers what each one means, how they compare on a production line and across a site, and why many Australian plants now run both.
What each term means
Computer vision is the broad field of AI that lets software interpret images and video. It covers everything from reading number plates to spotting a missing cap on a bottle.
Machine vision is the industrial use of cameras for inspection, measurement, sorting and robot guidance. Traditionally it meant dedicated cameras, controlled lighting and hand-configured rules: measure this edge, look for that colour, count these pixels.
AI vision is where the two meet. On a production line it uses the same disciplined hardware as machine vision, a camera and lighting chosen for the job, but the checking is done by a model trained on photos of your own product instead of fixed rules. How AI vision inspection works on a production line.
Comparison table
| Dimension | AI vision (computer vision, learned) | Rule-based machine vision (traditional) |
|---|---|---|
| Scope | Broad AI field: any system interpreting images or video, on a line or across a site | Industrial automation subset: inspection, measurement, sorting, robot guidance |
| Hardware on a production line | Industrial camera and lighting chosen for the job, triggered by a sensor, with a PC on site running the model | Industrial smart cameras or vision controllers, controlled lighting and optics |
| Hardware across a site | Existing CCTV or IP cameras are often a workable start for counting, flow and monitoring | Rarely used outside controlled inspection stations |
| Environment | Handles natural variation in colour, position and lighting; also works outdoors for site monitoring | Best in tightly controlled conditions: fixed position, controlled lighting |
| Software approach | Model trained on photos of your product and your real faults | Hand-configured rules and vendor toolkits |
| Adaptability | Retrained on new products and new fault types; each new version is tested before it goes live | Rule changes usually need a vision engineer for each product or fault change |
| Best at | Faults that vary: contamination, damage, cosmetic defects, labels that move, mixed products | Precise, repeatable measurement: gauging, alignment, presence checks on tightly controlled products |
| Speed and precision | Keeps the line's normal rhythm for most inspection tasks; very fine measurement at extreme speed is assessed case by case | The standard choice for high-precision measurement at very high line speeds |
| Capital outlay | Camera, lighting and an on-site PC per inspection point; lower where existing site cameras suit monitoring | Purpose-built hardware and vendor software per station |
| Vendor lock-in | Built on open frameworks such as YOLO; models can be retrained and moved | Often tied to one vision vendor's ecosystem |
| Typical fits | Food and beverage, packaging and FMCG lines, general manufacturing, warehouses, quarries and worksites | Precision assembly, semiconductor and electronics inspection, high-speed gauging |
| Data sovereignty | Runs on site, so images stay in the plant by default | On-premises hardware, with a similar profile |
Choosing an approach
When AI vision fits
- The faults vary from item to item: contamination, damage, cosmetic defects, labels that move
- You run several products or frequent changeovers
- You need the system to adapt as products, conditions or quality rules change
- You want every item checked and recorded with its photo, with rejects handled by the line's own equipment
- The job is counting, monitoring, tracking or safety across a site, where existing cameras give a usable view
- You work in food and beverage, packaging and FMCG, general manufacturing, warehousing, quarrying or construction
When rule-based machine vision fits
- The task is precise measurement: sub-millimetre gauging, alignment, precision assembly or semiconductor inspection
- The line runs at extreme speed and the check is fixed and repeatable
- The station is already tightly controlled (an enclosed cell, fixed position, controlled lighting)
- The product and process are standardised and unlikely to change
- You already have it working well, in which case AI vision can be added alongside it for the faults rules struggle with
Running both side by side
Many plants run both. Rule-based machine vision stays on the stations that need precise measurement, and AI vision is added for the faults that rules struggle with, plus monitoring across the wider site such as safety zones, vehicles and equipment. The decision is made per check, not per plant.
In both cases, results go to the line's own controls, which reject, divert, stop or flag the item, and the AI never takes control of the line. A new AI check is tested offline on your own footage and then runs in watch-only mode before it acts on anything. How plant integration works.
Why the terms now overlap
Three things changed over the last several years:
- AI models got much better at recognising objects and faults. YOLO and similar architectures now run in real time on modest hardware, so checks that once needed hand-tuned rules can be done by a model trained on your product.
- On-site computing got cheaper. A PC with a capable graphics card, or a compact edge device, can run inspection beside the line. Decisions happen locally and keep working if the internet drops.
- Custom training became accessible. Where rule-based systems needed an engineer to reprogram rules for each product change, an AI model can be retrained on new examples and tested on your own photos before it goes live.
As a result, buyers often use “computer vision” and “machine vision” interchangeably, although the terms remain distinct. For a production line, the practical question is which approach suits each check: learned AI vision, rule-based measurement, or both, always with a camera and lighting chosen for the job. For site and vehicle monitoring, the question is whether the cameras already there give a usable view. The AI glossary defines related terms.