How Computers See Pictures

Pictures are made of tiny dots

Look very closely at a screen. A picture is not one solid thing. It is made of thousands of tiny dots called pixels.

Each pixel is one small square of colour. Put millions of them together and your eyes see a photo of a dog or a beach.

The more pixels a picture has, the sharper it looks.

Colours are really numbers

A computer cannot see colour the way you do. So it stores each pixel as numbers.

Most screens mix three colours of light: red, green and blue. Each one gets a number from 0 (none) to 255 (lots).

• Bright red = a big red number, little green, little blue.
• White = all three turned up high.
• Black = all three turned down to 0.

So to a computer, a picture is just a huge grid of numbers. That is all it really has to work with.

How does AI know what is in a photo?

A computer does not understand "dog" or "tree". So how can AI tell what is in a picture?

It learns from examples. People show the AI thousands of photos that already have labels: "this is a cat", "this is a car".

The AI looks for patterns in the pixel numbers. Cats often have pointy ears and whiskers. Cars often have wheels and windows. (Patterns matter a lot here, just like in What Is a Pattern?.)

Then, when it sees a new photo, it checks: "Do these patterns match a cat?" If they do, it guesses "cat".

This is the same idea as Machine Learning: learning from lots of examples instead of fixed rules.

AI can still get it wrong

Image AI is clever, but it is not perfect.

If a photo is blurry, dark, or unusual, the AI can guess wrong. It might call a muffin a puppy, or miss something obvious.

It also does not truly understand what it sees. It only matches patterns of numbers. So people should always check important results, like a doctor checking a medical scan, not just trusting the computer.

Build it yourself one day

All of this is made with code. People write programs that turn pixels into patterns.

If you would like to make computers do clever things, start with Coding. It is the first step to building image AI yourself.