Convolutional Neural Network: A Quick Overview

Photo by Hunter Harritt on Unsplash

In the world of AI, Machine Learning, Deep Learning, and Computer Vision, we have come across and heard about various tasks like Image Classification, Object Detection, Image Pattern Detection, Text Classification, and Face Recognition. In this article, I have written a quick overview of Convolutional Neural Networks.

Convolutional Neural Network (CNN or ConvNets) is a Deep Learning technique which is generally used to perform the tasks mentioned above. Here, input is an image (simply a matrix of pixels) which is fed into a CNN model that assigns some learnable weights and biases to various aspects of an image to analyze input images for recognition and classification. Just like teachers help toddlers recognize different letters and digits through images and diagrams, similarly, computers are also trained on large datasets (millions and thousands of images) for image or text recognition tasks.

A (28 x 28 x 1) pixel grayscale image from the EMNIST dataset for character Z. This is fed as input to a CNN model which then predicts its label based on the learning and features.

The elementary role of a CNN model is to modify the images into a form that is computationally less expensive with no loss of features, which are important for getting good and accurate predictions.

Types of Layers in a CNN model:

1. Convolutional Layer

2. Pooling Layer

3. Fully Connected (Dense Layer)

1. Convolutional Layer: In the Convolutional Layer, the convolution operation is performed to get the position and strength of features of the input image. It takes an input image (say 6 x 6 image) and convolves (dot product) with a filter (say 3 x 3 filter/kernel) to give a convolved image of size 4 x 4(if p=0 i.e. valid padding and s = 1).

({[n + 2p — f]/s}+ 1) x ({[n + 2p — f]/s}+ 1) → dimension of convolved image after convolution operation.

Note: n = Input image dimensions, p = padding, f = size of filter, & s = stride

Padding: Padding is a hyperparameter of a convolutional layer. It is a process of adding zeros to our input image. If we use the convolutional layer without setting the padding parameter, the output image shrinks and there are chances of losing some information/features from edges.

Note: * is a convolution operation

A Quick Note: Filter/kernel is a matrix of size (f x f) and f — a hyperparameter is usually chosen to be odd like 3 x 3, 5 x 5, 7 x 7, etc.

No. of channels of Filter = No of channels of image

Strides: Strides is another hyperparameter in the convolutional layer which tells about the number of pixels to shift the filter over the input matrix. If we set s = 3, shifting the filter over input image by 3 pixels.

2. Pooling Layer: Usually used after a convolutional layer. The pooling layer is responsible for reducing the size of the convolved features. In this layer, there are no parameters and weights to learn for the model. It aims to reduce the size of the representation to speed up the computation and to make some features more robust i.e. extracting dominant features from convolved images.

Max Pooling and Average Pooling of 4 x 4 image (stride = 2, pool size = 2)

A Quick Note: Feature Maps in CNN captures features and that is the output of applying filters to the input image.

3. Fully Connected (Dense): A Fully Connected Layer (Dense Layer) takes the output of the previous layers (usually Pooling Layers) and flattens it to a single vector. The 1st fully connected layer takes in the features and assigns weights to them, and the output Dense Layer gives the probabilities for each class using an appropriate activation function.

A general strategy for creating a CNN Model:-

1. Input an image into the convolutional layer. Select appropriate hyperparameters (filters, strides, and padding). Perform convolution on the image and use an activation function depending on the task to be performed.
2. Add a Pooling Layer and perform pooling to reduce dimensionality. Max or Avg. Pooling can be used based on functionality.
3. Add Convolutional layers and Pooling Layers to the model as per model’s accuracy/loss i.e. the metric chosen and depth of Neural Network required based on image complexity.
4. Flatten the output (Flatten Layer) and feed into a fully connected dense layer (FC Layer).
5. Output the classes using an activation function using Dense Layer. Additionally, we can use Batch Normalization and Dropout Layers to improve the performance of our model.

Generally, Conv — Pool — Conv — Pool — Fully Connected (Flatten Layer)— Fully Connected — Fully Connected (Output Layer)

Let’s take an example to understand the dimensions of images at each layer of a CNN model. Let’s take a simple CNN model as shown below:

Basic CNN Architecture

The input is a (32 x 32 x 3) dimensional image which is fed into a Convolutional Layer.

In 1st Convolutional Layer, f = 5 → a filter/kernel matrix of size (5 x 5) is convolved with (32 x 32 x 3) input image with same padding and stride of s = 1 to output a (32 x 32 x n) dimensional image. Let’s assume there are n filters of size (5 x 5). Hence, dimension of image is (32, 32, n).

Calculations: p = (f-1)/2 = (5–1)/2 = 2

({[n + 2p — f]/s} + 1) = ({[32 + 2(2) — 5]/1} + 1) = 32

Using p = same, we are getting an output image of the same size as the input which is also evident from our calculations.

In Pooling Layer 1, we get a matrix with features having either maximum or average value. For a (32 x 32 x n) convolved image with a stride of s = 2 and pool size = 2, we get a (16 x 16 x n) image.

In 2nd Convolutional Layer, f = 3 → a filter matrix of size (3 x 3) is convolved with (16 x 16 x n) image with valid padding to get a (14 x 14 x n`) dimensional image. Here, s = 1, and the filter moves by 1 pixel over input image. Let’s assume there are n` filters of size (3 x 3). Hence dimension of the convolved image is (14, 14, n`).

Calculations: p = 0

({[n + 2p — f]/s} + 1) = ({[16 + 2(0) — 3]/1} + 1) = 14

In Pooling Layer 2, either a max pool or an average pool can be used. For a (14, 14, n`) convolved image we get a (7 x 7 x n`) image since we have used a stride of s = 2 and pool size = 2.

This image is then fed to a Flatten Layer to convert the image into a 1D vector of size 7 x 7 x n` = 49xn`. This 1D vector is then fed into a Dense Layer having x units/neurons with some normalization techniques/layers (like Batch Normalization or a Dropout Layer) to improve the model’s performance. Another Output Dense Layer with y classes (depends on the problem being considered) outputs labels or probabilities.

How do we calculate no. of parameters for a Convolutional Layer and Fully Connected Layer?

# Parameters in Conv. Layer : [ (f x f x no. of channels in previous layer)+ 1] x no. of filters

# Parameters in FC Layer : [no. of units in current layer x no of units in previous layer] + no. of units in current layer

Note: There are no parameters for Pooling Layer because it’s used to reduce the dimensionality of the image and has no learning of parameters while training the model.

Let’s take an example to calculate no. of parameters for a Convolutional Layer and FC Layer:

No. of parameters in Convolutional Layer 1 of ‘Basic CNN Architecture’ figure :-

Here, f = 5, no. of filters = n = 64(say) and no. of channels in previous layer = 3

Therefore, [(5 x 5 x 3) +1] x 64 = 4864 no. of parameters

No. of parameters in Fully Connected Layer (Dense Layer) of ‘Basic CNN Architecture’ figure :-

Here, no. of units in current layer = 120 (say) and no. of units in previous layer = 400(say)

Therefore, [120 x 400] + 120= 48120 no. of parameters