For a person, recognising a mug on a table is instant. For a computer, an image is just a grid of numbers. Computer vision is the branch of artificial intelligence that teaches machines to extract useful information from those grids: working out what is in a photo, where it is, whether a part made in a factory has a defect. It is one of the most mature and widespread applications of AI, often invisible because it is built into everyday objects.
How a computer "sees" an image
A digital image is made up of pixels. In a colour image each pixel is described by three values indicating the intensity of red, green and blue, usually from 0 to 255. A 1000 by 1000 pixel photo therefore corresponds to three million numbers.
The problem is that the same object can produce completely different numbers: a mug photographed in morning light or under a lamp, close up or from a distance, from the side or from above. A vision system must learn to recognise what stays constant beyond these variations.
Convolutional networks
For many years the reference architecture was the convolutional neural network (CNN). Its key element is the filter: a small window, for example 3 by 3 pixels, that slides across the image and responds to a specific local pattern, such as a vertical edge or a change of colour.
It works in levels:
- the first layers pick out simple elements: lines, corners, contrasts;
- the intermediate layers combine them into shapes and textures: curves, grids, rough surfaces;
- the deep layers recognise parts of objects and whole objects: a handle, a circular rim, a mug.
Nobody programs these filters by hand: the network learns them during training by looking at thousands of labelled images. More recently Vision Transformers have also become widespread; they divide the image into small patches and treat them much as language models treat words.
The main tasks
| Task | Question it answers | Example |
|---|---|---|
| Classification | What is in the image? | "This photo contains a cat" |
| Object detection | What is there and where? | Boxes around every car on a street |
| Segmentation | Which object does each pixel belong to? | Precise outline of a tumour in a medical image |
| Text recognition (OCR) | Which characters are there? | Reading a scanned invoice |
| Pose estimation | How is a body positioned? | Analysing movements in physiotherapy |
Concrete applications
Industry and quality control
Cameras above a production line photograph each part and a model flags scratches, dents, irregular welds or crooked labels. The advantage is not only speed but consistency: the system does not tire at the end of the shift.
Logistics
Automatic reading of codes and addresses on parcels, counting goods on shelves, checking that a pallet has been loaded correctly.
Agriculture
Analysing images from drones or from sensors on machinery to identify areas of a field with struggling plants or weeds, so that action can be targeted.
Documents
Modern OCR, combined with models that understand document structure, extracts data from invoices, forms and receipts, even when they are photographed with a smartphone.
Accessibility
Apps that describe the surroundings aloud or read labels and signs for people with visual impairments.
Medicine
Support systems that highlight suspicious areas in radiological images, always as an aid to the doctor's judgement and not as a replacement.
An example from a small business
A craft workshop producing decorated tiles can install a camera at the kiln exit. After collecting and classifying a few hundred images of good and defective pieces, a model learns to flag cracks and glaze blemishes. At first every alert is checked by an operator; over time, clearly sound pieces pass without checks and human attention focuses on doubtful ones.
Limits worth knowing
- Dependence on shooting conditions. Changing lighting, camera or background can greatly worsen performance. It is best to keep conditions constant or include their variety in the data.
- Rare cases. A defect never seen during training may go unnoticed.
- Shortcuts. A model may learn an irrelevant clue, for example recognising defective parts by the colour of the tray they were photographed on rather than by the defect itself.
- Privacy. When images show people, data protection rules apply, and some biometric uses are heavily restricted or prohibited by the AI Act.
In short
- To a computer an image is a grid of numbers; computer vision extracts meaning from it.
- Convolutional networks learn filters that recognise increasingly complex patterns.
- The main tasks are classification, detection, segmentation, OCR and pose estimation.
- Quality control, logistics, documents and accessibility are well-established applications.
- Shooting conditions, rare cases and privacy are the main limits to manage.
