Computer vision fundamentals · CAB107

Computer vision enables machines to interpret images and video. A digital grayscale image is stored as a grid of pixel intensity values, usually from 0 (black) to 255 (white); a colour image has three such channels, red, green and blue.

A convolutional neural network (CNN) is the standard model for images.

Layer	What it does
Convolution	Slides small filters over the image to produce feature maps that detect local patterns such as edges
Activation (ReLU)	Adds non-linearity
Pooling	Reduces the spatial size of feature maps while keeping the strongest responses
Fully connected	Combines features into the final class scores

Worked example: 2 x 2 max pooling with stride 2 on the 4 x 4 map with rows [1,3,2,1], [4,6,5,0], [7,2,9,8], [3,1,4,6] takes the maximum of each window: top-left 6, top-right 5, bottom-left 7, bottom-right 9, giving [[6, 5], [7, 9]]. Average pooling would give [[3.5, 2.0], [3.25, 6.75]].

import numpy as np
m = np.array([[1,3,2,1],[4,6,5,0],[7,2,9,8],[3,1,4,6]])
print(m.reshape(2, 2, 2, 2).max(axis=(1, 3)))

Typical tasks: classification, object detection and segmentation.

Exam trap: pooling shrinks the map but has no learned filters, and max pooling takes the largest value in each window, not the average.

Try two questions

Question 1 · Tier 3, Apply

How many bytes does an uncompressed 100 x 100 RGB image need at 8 bits per channel?

10,000
240,000
30,000
3,000

Show the answer

C. 100 x 100 x 3 channels x 1 byte = 30,000 bytes.

Why not A: 10,000 is the number of pixels, not the bytes.

Why not B: 240,000 counts bits rather than bytes.

Why not D: 3,000 loses a factor of 10.

Question 2 · Tier 2, Understand

What is the purpose of a pooling layer in a CNN?

To increase the number of pixels so that the feature maps become larger and sharper
To reduce the spatial size of the feature maps while keeping the strongest responses
To label each of the input images with a class name before training starts
To change the colour balance of the picture so that the network sees more contrast

Show the answer

B. Max pooling keeps the largest value in each window, reducing computation and adding small shift tolerance.

Why not A: Pooling shrinks the maps instead of enlarging them.

Why not C: Labelling is a human or training step.

Why not D: Pooling does not recolour the image.

Key points

Try two questions