Grade 12

Advanced

8 min

Lesson 7 of 12

Computer Vision and CNNs

Not Started

Exploring how Convolutional Neural Networks process visual information and recognize patterns in images.

How does a self-driving car distinguish a pedestrian from a lamppost in a split second? It doesn't 'see' like we do; it performs millions of matrix multiplications to translate pixels into patterns.

Learning Objectives

1.Explain the convolution operation and how kernels extract specific features from raw pixels.
2.Identify how pooling layers reduce computational load while preserving essential information.
3.Analyze the hierarchical nature of CNNs as they progress from detecting simple edges to complex objects.

The Convolution Operation

At its core, a Convolutional Neural Network (CNN) treats an image as a matrix of numbers representing pixel intensities. The 'magic' happens during the convolution step, where a small matrix called a kernel (or filter) slides across the input image. At each position, the network performs an element-wise multiplication and sums the results. This process creates a feature map. For example, a vertical edge-detection kernel might have high values on the left and low values on the right, effectively 'lighting up' whenever it passes over a vertical boundary in the image. Mathematically, for an input

I

and kernel

K

, the output

S

at position

(i, j)

is calculated as:

S(i, j) = \sum_{m} \sum_{n} I(i+m, j+n) K(m, n)

Quick Check

If a kernel is designed to detect horizontal edges, what would happen to the output values when it passes over a perfectly smooth, solid-colored wall?

Think about the difference between the pixel values the kernel is multiplying.

Answer

The output values would be zero (or very low) because there is no change in pixel intensity for the kernel to detect.

Let's apply a $2 \times 2$ kernel to a small section of an image. 1. Input Matrix section: $\begin{bmatrix} 10 & 20 \\ 10 & 20 \end{bmatrix}$ 2. Kernel (Vertical Edge Detector): $\begin{bmatrix} -1 & 1 \\ -1 & 1 \end{bmatrix}$ 3. Calculation: $(10 \times -1) + (20 \times 1) + (10 \times -1) + (20 \times 1)$ 4. Result: $-10 + 20 - 10 + 20 = 20$ . The positive result indicates a feature (an edge) was found!

Downsampling with Pooling

Images contain a massive amount of data, much of which is redundant. Pooling layers are used to reduce the spatial dimensions (width and height) of the feature maps. This process, known as downsampling, decreases the number of parameters the model needs to learn, which prevents overfitting and speeds up computation. The most common method is Max Pooling, where the network slides a window over the feature map and keeps only the maximum value from that window. This ensures that the most prominent features are preserved even as the 'resolution' of the map decreases. By doing this, the network becomes more 'translation invariant,' meaning it can recognize a feature regardless of its exact pixel location.

Quick Check

What is the primary trade-off when using a large pooling window (e.g., $4 \times 4$ instead of $2 \times 2$ )?

Think about how much information is discarded in each step.

Answer

A larger window reduces data more aggressively, which saves memory but risks losing fine-grained spatial details.

Perform

2 \times 2

Max Pooling with a stride of 2 on the following

4 \times 4

feature map:

\begin{bmatrix} 1 & 3 & 2 & 1 \\ 4 & 5 & 8 & 2 \\ 1 & 0 & 3 & 4 \\ 2 & 1 & 0 & 9 \end{bmatrix}

1. Top-left

2 \times 2

block:

\max(1, 3, 4, 5) = 5

2. Top-right

2 \times 2

block:

\max(2, 1, 8, 2) = 8

3. Bottom-left

2 \times 2

block:

\max(1, 0, 2, 1) = 2

4. Bottom-right

2 \times 2

block:

\max(3, 4, 0, 9) = 9

Resulting

2 \times 2

matrix:

\begin{bmatrix} 5 & 8 \\ 2 & 9 \end{bmatrix}

The Hierarchy of Features

CNNs are organized in a hierarchical structure. The early layers (closest to the input) act as 'microscopes,' using kernels to find simple patterns like edges, lines, and colors. As we move deeper into the network, the layers combine these simple patterns to recognize textures and shapes (like circles or honeycombs). Finally, the deepest layers assemble these shapes into complex objects (like eyes, wheels, or faces). This hierarchy mimics the human visual cortex. By the time the data reaches the final 'Fully Connected' layer, the spatial information has been distilled into a high-level vector that represents the probability of the image belonging to a specific class, such as 'Dog' or 'Car'.

In advanced CNN design, you must calculate the size of the next layer. Given an input width

W

, kernel size

K

, padding

P

, and stride

S

, the output width

O

is:

O = \frac{W - K + 2P}{S} + 1

If you have a

32 \times 32

image, a

5 \times 5

kernel, padding of 2, and a stride of 1: 1.

O = \frac{32 - 5 + 2(2)}{1} + 1

O = \frac{32 - 5 + 4}{1} + 1

O = 31 + 1 = 32

. The output remains

32 \times 32

Key Takeaways

1Convolution uses kernels to perform element-wise multiplication, extracting features like edges from raw pixels.
2Pooling layers (like Max Pooling) reduce the spatial size of data, making the model more efficient and robust.
3CNNs are hierarchical: early layers detect simple edges, while deeper layers detect complex objects.
4The output size of a layer is determined by the formula: $O = \frac{W - K + 2P}{S} + 1$ .

Quick Check

Multiple Choice

What is the primary purpose of a 'Kernel' in a CNN?

Multiple Choice

Which layer type is specifically designed to reduce spatial dimensions and prevent overfitting?

True/False

In a CNN hierarchy, the layers closest to the output are responsible for detecting basic edges and lines.

What's Next?

Review Tomorrow

In 24 hours, try to sketch the process of a $2 \times 2$ Max Pooling operation and explain why it helps a computer recognize a cat even if the cat is slightly shifted in the photo.

Practice Activity

Use the output dimension formula to calculate the size of a feature map if you start with a $224 \times 224$ image, use a $7 \times 7$ kernel, a stride of 2, and no padding.

Browse

Browse

Computer Vision and CNNs

Learning Objectives

The Convolution Operation

Downsampling with Pooling

The Hierarchy of Features

Key Takeaways

Quick Check

What's Next?

Computer Vision and CNNs

Learning Objectives

The Convolution Operation

Downsampling with Pooling

The Hierarchy of Features

Key Takeaways

Quick Check

What's Next?