PART 3 · MODULE 2Teaching AI to see
CNNs & Computer Vision
Convolutional neural networks for image understanding
Loading module...
Convolutional neural networks for image understanding
Computer vision is the field of teaching machines to understand images and video. For a computer, an image is just a grid of numbers — pixel values from 0 to 255. The challenge is turning those numbers into meaning: "that's a cat" or "that sign says stop."
A small 256x256 image has over 196,000 pixels. If every pixel connects to every neuron in the next layer, you get billions of connections. That is way too many — the network would be impossibly slow and would memorize rather than learn.
Convolutional Neural Networks (CNNs) solve this with a clever trick: instead of looking at the whole image at once, they scan small patches with a sliding filter — like reading a page with a magnifying glass, one small area at a time.
A convolution filter is a small grid of numbers (usually 3x3 or 5x5) that slides across the image. At each position, it multiplies its values with the underlying pixels and sums them up. Different filters detect different features — one might detect vertical edges, another horizontal lines.
A CONVOLUTION FILTER SLIDING OVER AN IMAGE GRID
A CNN processes images through a sequence of steps. Each step extracts and refines features until the network has enough understanding to make a decision.
HOW A CNN PROCESSES AN IMAGE
Detect features like edges, textures, and patterns using sliding filters.
Simplify the data by keeping only the strongest signals. Shrinks the image down.
Combine all detected features to make the final decision: what is this image?
The magic of CNNs is that each layer learns to see more complex things. Early layers detect simple features; deeper layers combine them into higher-level concepts.
CNNs learn to see the world in layers — from simple edges to complex objects. No one programs these features by hand. The network discovers them on its own during training, which is why CNNs can recognize things that would be nearly impossible to describe with manual rules.