Guides / Computer vision

Guide · Basics

Computer Vision in Plain English

01What it actually is

Computer vision is software that's learned to recognise things in pictures — in the same way the AI behind ChatGPT learned patterns in text. Shown thousands of labelled examples ("this is a scratch", "this isn't"), it learns to recognise the same thing in images it's never seen.

You already use it: your phone unlocking with your face, your banking app reading a cheque, number-plate cameras at a car park. The difference now is that the same technology is affordable for a single workshop, shop or farm.

02What it's good at

• Counting — items on a shelf, vehicles in a yard, people through a door, animals in a pen.
• Spotting — whether something's present or missing: a component, a label, safety gear.
• Reading — text, serial numbers, barcodes, meter dials, number plates, handwritten forms.
• Inspecting — defects, damage, wear, inconsistent colour or finish.
• Measuring — sizes, fill levels, gaps, positions.

03Real small-business examples

• Stock and inventory — a photo of a shelf or storeroom turned into a count, instead of an afternoon with a clipboard.
• Quality control — a camera on a workbench or production line flagging products with defects before they ship.
• Paperwork capture — delivery notes, meter readings and job sheets photographed and turned into spreadsheet rows.
• Site and yard monitoring — vehicles logged in and out by number plate; alerts when someone enters a restricted area.
• Condition records — before-and-after photos of hire equipment or vehicles checked for new damage.
• Agriculture — counting livestock, spotting lameness, checking crop health from drone images.

See it working: our live computer vision demo runs in your browser.

04Two routes: off-the-shelf or custom

Off-the-shelf first. The everyday AI assistants can now look at photos, so for occasional jobs — reading a faded label, describing damage, extracting figures from a photo of a form — just upload an image and ask. Many apps for specific jobs (stock counting, receipt capture, vehicle inspection) already exist too.

Custom when it's repetitive and specific. When you need the same check hundreds of times a day, or to spot something unique to your products, a custom model trained on your own images is more accurate, faster and cheaper to run at volume.

05What it needs to work well

• A narrow job. "Is this weld acceptable?" works. "Tell me everything wrong in the workshop" doesn't.
• Consistent conditions. Same camera angle, decent lighting. Most accuracy problems are really lighting problems.
• Examples. For a custom model, typically a few hundred labelled photos of good and bad examples — often collected over a few weeks of normal work.
• A plan for mistakes. No system is 100% accurate. Decide what happens when it's unsure — usually, a person takes a look.

06Cameras, people and privacy

If your cameras capture people — staff, customers, the public — the footage is personal data and UK data protection law applies. In practice: have a clear reason, tell people (signage, staff notices), keep footage only as long as you need it, and point cameras at the job, not at people, wherever possible. For anything that monitors staff or uses facial recognition, the ICO expects a formal risk assessment (a DPIA) first — and facial recognition is rarely justifiable for a small business.

The best projects sidestep the issue entirely: a camera looking down at a workbench sees products, not people.