How does AI recognize images? A short, plain explainer

What does the computer actually see?

Numbers. A 12-megapixel photo is about 12 million pixels, each stored as three values for red, green and blue. Nothing in those values says dog or beach; meaning has to be learned from examples.

Early systems used hand-written rules, such as edge detectors. Since around 2012, neural networks have learned the rules themselves. A big step was training on ImageNet, a collection of more than 14 million images labelled by people, where a network called AlexNet clearly beat earlier methods in the 2012 competition.

Colour and size vary endlessly, which is why rules alone failed. A cat photographed in shade and the same cat in sunlight produce very different numbers, yet a person sees one animal. Learning from many examples is how networks cope with that variety.

How does a neural network learn to recognise things?

1. The network looks at a labelled image and guesses its label.
2. The guess is compared with the right answer, giving an error.
3. Its internal weights are nudged to make that error smaller.
4. This repeats over millions of images until guesses are reliably right.

Inside, early layers react to simple features like edges and colour blobs. Later layers combine those into textures, parts and eventually whole objects. Vision transformers, introduced in 2020, take a different route: they split the image into 16×16-pixel patches and learn how the patches relate, much as language models relate words.

Training takes huge computing power, but using a trained network is far cheaper. That is why a model trained on servers can later run on a phone, often in a smaller, compressed form.

How is this used to compare photos?

Instead of a label, a network can output an embedding: a list of a few hundred numbers summarising what the image shows. Similar pictures get similar lists, so measuring the distance between them finds near-duplicates even when the files differ in size or format. Models such as CLIP, trained on 400 million image and text pairs, place pictures and words in the same space, which is how a search for beach at sunset can find matching photos.

Photo cleaners combine these signals with simpler ones like a perceptual hash and blur measures. The AI photo cleaner guide shows how they turn into keep-or-delete suggestions.

The same building blocks power photo search in Google Photos, where typing a word such as dog brings up matching pictures.