7. Computer Vision with Convolutional Neural Networks

Overview

This chapter covers computer vision and how this is accomplished with neural networks. You will learn to build image processing applications and classify models with convolutional neural networks. You will also study the architecture of convolutional neural networks and how to utilize techniques such as max pooling and flattening, feature mapping, and feature detection. By the end of this chapter, you will be able to not only build your own image classifiers but also evaluate them effectively for your own applications.

Introduction

In the previous chapter, we explored model evaluation in detail. We covered accuracy and why it may be misleading for some datasets, especially for classification tasks with highly imbalanced classes. Datasets with imbalanced classes such as the prediction of hurricanes in the Pacific Ocean or the prediction of whether someone will default on their credit card loan have positive instances that are relatively rare compared to negative instances, so accuracy scores are misleading since the null accuracy is so high.

To combat class imbalance, we learned about techniques that we can use to appropriately evaluate our model, including calculating model evaluation metrics such as the sensitivity, specificity, false positive rate, and AUC score, and plotting the ROC curve. In this chapter, we will learn how to classify another type of dataset—namely, images. Image classification is extremely useful and there are many real-world applications of it, as we will discover...

Computer Vision

To understand computer vision, let's discuss human vision. Human vision is the ability of the human eye and brain to see and recognize objects. Computer vision is the process of giving a machine a similar, if not better, understanding of seeing and identifying objects in the real world.

It is fairly simple for the human eye to precisely identify whether an animal is a tiger or a lion, but it takes a lot of training for a computer system to understand such objects distinctly. Computer vision can also be defined as building mathematical models that can mimic the function of a human eye and brain. Basically, it is about training computers to understand and process images and videos.

Computer vision is an integral part of many cutting-edge areas of robotics: health care and medical (X-rays, MRI scans, CT scans, and so on), drones, self-driving cars, sports and recreation, and so on. Almost all businesses need computer vision to run successfully.

Imagine...

Convolutional Neural Networks

When we talk about computer vision, we talk about CNNs in the same breath. CNN is a class of deep neural network that is mostly used in the field of computer vision and imaging. CNNs are used to identify images, cluster them by their similarity, and implement object recognition within scenes. CNN has different layers— namely, the input layer, the output layer, and multiple hidden layers. These hidden layers of a CNN consist of fully connected layers, convolutional layers, a ReLU layer as an activation function, normalization layers, and pooling layers. On a very simple level, CNNs help us identify images and label them appropriately; for example, a tiger image will be identified as a tiger:

Figure 7.1: A generalized CNN

The following is an example of a CNN classifying a tiger:

Figure 7.2: A CNN classifying an image of a tiger into the class "Tiger"

The Architecture of a CNN

The main components of CNN architecture are as follows:

• Input image
• Convolutional layer
• Pooling layer
• Flattening

Input Image

An input image forms the first component of a CNN architecture. An image can be of any type: a human, an animal, scenery, a medical X-ray image, and so on. Each image is converted into a mathematical matrix of zeros and ones. The following figure explains how a computer views an image of the letter T.

All the blocks that have a value of one represent the data, while the zeros represent blank space:

Figure 7.3: Matrix for the letter 'T'

Convolution Layer

The convolution layer is the place where image processing starts. A convolution layer consists of two parts:

• Feature detector or filter
• Feature map

Feature detector or a filter: This is a matrix or pattern that you put on an image to transform it into a feature map:

Figure...

Image Augmentation

The word augmentation means the action or process of making or becoming greater in size or amount. Image or data augmentation works in a similar manner. Image/data augmentation creates many batches of our images. Then, it applies random transformations to random images inside the batches. Data transformation can be rotating images, shifting them, flipping them, and so on. By applying this transformation, we get more diverse images inside the batches, and we also have much more data than we had originally.

A cylinder can be rotated from different angles and seen differently. In the following figure, a single cylinder can be seen from five different angles. So, we have effectively created five different images from a single image:

Figure 7.13: Image augmentation of a cylinder

The following is some example code that we would use for image augmentation; here, the ImageDataGenerator class is used for processing. shear_range, zoom_range, and...

Summary

In this chapter, we studied why we need computer vision and how it works. We learned why computer vision is one of the hottest fields in machine learning. Then, we worked with convolutional neural networks, learned about their architecture, and looked at how we can build CNNs in real-life applications. We also tried to improve our algorithms by adding more ANN and CNN layers and by changing the activation and optimizer functions. Finally, we tried out different activation functions and loss functions.

In the end, we were able to successfully classify new images of cars and flowers through the algorithm. Remember, the images of cars and flowers can be substituted with any other images, such as tigers and deer, or MRI scans of brains with and without a tumor. Any binary classification computer imaging problem can be solved with the same approach.

In the next chapter, we will study an even more efficient technique for working on computer vision, which is less time-consuming...