You glance at a photo and instantly know it shows a dog on a beach at sunset. A computer sees only millions of numbers. Computer vision is the craft of closing that gap — turning a grid of pixels into a judgement. This course explains how, from the basics of what an image even is, to the systems that unlock your phone and read medical scans.
- Understand what an image is to a computer, and why vision is hard
- Grasp "features" — the building blocks vision is built on
- See how CNNs turned vision from clever tricks into reliable systems
- Know the main vision tasks and where each is used in the real world
What an image is to a computer
To a computer, an image is a grid of pixels, each just numbers for its colour. A modest photo is millions of these numbers — and nowhere in them is it written "this is a dog." The meaning has to be computed from the raw values.
That is why vision is hard. The same dog looks utterly different in different photos: lighting changes the numbers, angle changes them, distance, background, a paw over the face — all change the pixels completely while the dog stays the same. A vision system has to see through all that variation to the thing itself. Humans do this effortlessly; for a computer it is a genuine feat.
- An image is a grid of pixels — just colour numbers, with no built-in meaning
- The same object produces wildly different pixels under different conditions
- Vision is hard because meaning must be computed despite all that variation
Features — the building blocks of seeing
Before deep learning, the key idea was the feature: a small, distinctive pattern like an edge, a corner, or a texture. The insight is that you can recognise and match things by their features, even across different photos.
Think of how you'd recognise a friend in two different pictures — not by comparing every pixel, but by distinctive features: the shape of their nose, a particular smile. Early computer vision did the same: find the standout points in each image, then match them. This is how panorama stitching lines up two photos, and how early apps recognised a landmark.
photo A photo B (same scene, shifted)
•corner •corner
•edge ───► •edge matched points
•texture •texture reveal it's the same thing
- A feature is a small distinctive pattern — an edge, corner, or texture
- Things can be recognised and matched by their features across different images
- This hand-crafted idea became the learned features inside CNNs
- Why do two photos of the same object look completely different to a computer?
- Why is matching by features smarter than comparing images pixel by pixel?
CNNs — when vision became reliable
For decades, vision relied on clever hand-crafted features and worked only so-so. Then came the CNN (from the Deep Learning course), which learns the features itself — and vision suddenly leapt from "sometimes works" to "better than humans" on many tasks.
Recall how a CNN works: it scans the image with small sliding filters, early layers finding edges, later layers combining them into shapes and whole objects. The magic is that nobody specifies those filters — the network discovers the right ones by training on labelled images. Hand-crafted features were a person's best guess; learned features are tuned by millions of examples.
- Hand-crafted features capped how well vision could work
- CNNs learn the right features from labelled examples instead
- This leap made vision reliable enough for real products
The vision tasks, and where they're used
"Computer vision" is really a family of tasks. The main ones, and where you meet them:
| Task | What it answers | Everyday use |
|---|---|---|
| Classification | "What is in this image?" | Photo apps sorting pictures by content |
| Detection | "What's here, and where?" (boxes) | Self-driving spotting cars and pedestrians |
| Segmentation | "Which exact pixels are the object?" | Medical scans outlining a tumour; photo background blur |
| Tracking | "Where did it move across frames?" | Sports analysis; security cameras following motion |
| Recognition | "Which specific one is this?" | Face unlock; identifying a product |
- Vision is a family of tasks: classification, detection, segmentation, tracking, recognition
- Each maps to real products, from face unlock to medical imaging to self-driving
- These systems give probabilities and can fail on unusual inputs — verify, and keep a human in high-stakes loops
- What changed when CNNs let networks learn features instead of hand-crafting them?
- For a self-driving car spotting pedestrians, which vision task is it, and why must it report where, not just what?