What is computer vision?

Computer vision is the field concerned with extracting information from images and video: what is in the picture, where it is, how it moves, and whether it matches something seen before. A computer receives an image as a grid of numbers with no notion of objects or edges, so every vision system has to reconstruct structure from raw intensity values.

How does image recognition actually work?

A network applies small learned filters across the image, each responding to a local pattern such as an edge or a texture. Stacking these layers builds progressively larger patterns: edges combine into shapes, shapes into parts, parts into objects. The final layer turns those detected patterns into a probability for each possible label.

What is the difference between object detection and segmentation?

Detection returns a rectangle around each object plus a label, which is enough to count things or trigger an alert. Segmentation assigns a label to every individual pixel, which is what you need when the exact shape matters, as in medical imaging or measuring surface area. Segmentation costs far more to annotate, which is usually the deciding factor.

Do I need millions of images to train a vision model?

No. You start from a model already trained on a large public dataset and adapt it to your task, which typically needs a few hundred to a few thousand labelled images per class rather than millions. The quality and representativeness of those images matters far more than the count.

Why does a vision model that scores well in testing fail in production?

Almost always domain shift: production images differ from training images in lighting, camera, angle, resolution or background, and a model that learned incidental correlations rather than the object itself breaks immediately. The fix is not a bigger model, it is training data collected under the same conditions as deployment.

Are vision transformers better than convolutional networks?

With enough data, transformers generally win, because they can relate distant parts of an image directly instead of building up context layer by layer. With modest data, convolutional networks often remain the better choice, since their built-in assumption that nearby pixels are related is exactly the prior that scarce data cannot supply.

Computer Vision: how machines turn pixels into decisions

A photograph means nothing to a computer. It receives a grid of numbers — millions of them — with no indication of where one object ends and another begins. That vision works at all is the result of one good idea applied at scale, and this course explains the idea, the tasks it enables, and the ways it still breaks.

What this course sets out to do: make you able to look at a vision project and say what is realistic, what data it needs, and what will go wrong first.

What this course does not do: train models. The premium catalogue covers that with notebooks, datasets and deployment.

What you are about to discover

Course contents

#	Lesson	Main goal	Time
1	What a computer sees	Pixels, channels, and why this is harder than it looks	8 min
2	Convolution: the idea that worked	Learned filters and hierarchies of features	9 min
3	The four tasks	Classification, detection, segmentation, tracking	9 min
4	Modern architectures	From AlexNet to vision transformers and foundation models	8 min
5	Where it still fails	Domain shift, adversarial inputs, bias, honest evaluation	9 min
6	Recap and FAQ	Synthesis, a decision guide, and 12 common questions	6 min
7	Quiz and attestation	Validate what you learned with 5 corrected questions	3 min

Is this course for you?

You want to automate a visual inspection — defects, counting, sorting — and need to know what it takes.
You work with cameras or medical images and want to judge vendor claims.
You keep hearing "CNN", "YOLO" and "segmentation" and want them precise.
You are deciding whether a vision project is feasible before committing budget.

Recommended first: Deep Learning.

What you will be able to do at the end

Explain what an image is to a computer and why that framing creates the difficulty.
Describe convolution in plain language and say why it beat hand-designed features.
Choose between classification, detection, segmentation and tracking for a given problem.
Say what transfer learning buys you and how much data you actually need.
Explain why domain shift is the most common cause of production failure.
Recognise adversarial examples and dataset bias as real engineering risks.

Estimated time

Around 45 to 55 minutes of reading.

Prerequisites and next steps

Prerequisites: Introduction to AI and Deep Learning.

Natural continuation:

Generative AI — producing images rather than reading them
MLOps — keeping a vision model alive in production
Ethics of AI — facial recognition and surveillance, in detail

Frequent questions, answered in one line

Is computer vision solved?

For narrow, well-specified tasks with representative data, largely yes: reading a meter, spotting a missing component, counting vehicles. For open-ended visual understanding — "is anything unusual happening here?" — no, and not close. The gap between those two sentences is where most failed projects live.

Can a model work on video, or only still images?

Both. Video is usually processed as a sequence of frames, which makes single-frame models immediately applicable, and tasks that depend on motion — tracking an individual across frames, recognising an action — need models that carry information between frames. Video also multiplies compute cost by the frame rate, which drives most architecture decisions in practice.

Do I need a GPU?

To train, effectively yes, though renting one for a few hours is cheap. To run a trained model, often no: a small model can classify images on a phone or a Raspberry Pi, which is the whole point of the tooling covered in the premium track.

How accurate does a vision model need to be?

It depends entirely on the cost of each kind of mistake, and stating a single accuracy target is usually a sign the question has not been thought through. A model that flags possible defects for human review can be quite imprecise and still save money; a model that rejects parts automatically cannot.

Want to build vision systems rather than read about them?

The premium catalogue covers convolutional networks, detection, segmentation and edge deployment, with a verifiable certificate after a 40-question examination. Included in every paid plan.

Ready? Start with lesson 1 →

What you are about to discover​

Course contents​

Is this course for you?​

What you will be able to do at the end​

Estimated time​

Prerequisites and next steps​

Frequent questions, answered in one line​