2 Introduction to transfer learning for computer vision

The previous chapter gave you a flavor of the essentials of transfer learning, where you looked at various aspects of transfer learning including types, methodologies, tools, and examples. While we leveraged and saw the power of pre-trained models first-hand in classifying images, the focus was more on inference and not training. In this chapter, we will touch upon the true power of transfer learning – adapting pre-trained models to solve new and novel problems.

You will build upon the fundamentals you learnt in Chapter 1 by expanding upon convolutional neural networks, also known as CNNs or convnets – the working powerhouse of any deep learning-based computer vision system. Key neural network layers and tensor operations will be covered in-depth for CNNs considering a typical image classification scenario. We will also briefly touch upon other avenues of computer vision like recognition, segmentation and generation which will be covered in detail in subsequent chapters.

2.1 The Deep Learning Landscape for Computer Vision

2.2 Convnet Essentials

2.2.1 Architecture

2.2.2 Layers and Layer Operations

2.3 Transfer Learning Strategies

2.3.1 Frozen Pre-trained Models as Feature Extractors

2.3.2 Fine-Tuning Pre-Trained Models

2.3.3 Fine-tuning Tips and Tricks

2.4 Image Classification on Small Datasets

2.4.1 Problem Objective and Workflow

2.4.2 Processing and Creating Datasets

2.4.3 Simple CNN from Scratch

2.4.4 Simple CNN with Image Augmentation

2.4.5 ResNet-50 from Scratch

2.4.6 Frozen pre-trained ResNet-50

2.4.7 Fine-tuning pre-trained ResNet-50

2.4.8 Fine-tuning pre-trained ResNet-50 with one-cycle learning rate

2.4.9 Performance analysis of training strategies

2.5 Summary