An Introduction to Image Classification. From Designed Models to End-to-End Learning

Preface
	What the Book Is About
	How Should the Book Be Used
	How the Book Is Structured
Acknowledgments
Contents
Chapter 1: Image Classification: A Computer Vision Task
	1.1 What Is Image Classification and Why Is It Difficult?
	1.2 Image Classification as a Structured Process
	1.3 Python Implementation Details
		1.3.1 Python Development Environment
		1.3.2 Basic Modules
		1.3.3 Some Basic Operations on Images
	1.4 Exercises
		1.4.1 Programming Project P1.1: Reading and Displaying MNIST Data
		1.4.2 Exercise Questions
	References
Chapter 2: Image Features: Extraction and Categories
	2.1 Image Acquisition Artifacts
	2.2 Using Pixel Values as Features
	2.3 Using Texture Features
		2.3.1 Haralick´s Texture Features
		2.3.2 Gabor Filter Banks
		2.3.3 Local Binary Pattern
		2.3.4 Using Texture Measures for Image Classification
	2.4 Using Edges and Corners
		2.4.1 Edge Detection
		2.4.2 Corners
	2.5 HOG Features
	2.6 SIFT Features and the Bag-of-Visual-Words Approach
		2.6.1 SIFT Computation
		2.6.2 From SIFT to Secondary Features: Bag of Visual Words
	2.7 Exercises
		2.7.1 Programming Project P2.1: Orientation Histograms
		2.7.2 Programming Project P2.2: A Cluster Measure in Feature Space
		2.7.3 Exercise Questions
	References
Chapter 3: Feature Reduction
	3.1 Unsupervised Feature Reduction
		3.1.1 Selection Based on Feature Variance
		3.1.2 Principal Component Analysis
	3.2 Supervised Feature Reduction
		3.2.1 Forward and Backward Feature Selection
		3.2.2 Linear Discriminant Analysis
	3.3 Exercises
		3.3.1 Programming Project P3.1: Comparing Variance in Feature Space
		3.3.2 Exercise Questions
	References
Chapter 4: Bayesian Image Classification in Feature Space
	4.1 Bayesian Decision Making
	4.2 Generative Classification Models
		4.2.1 Likelihood Functions from Feature Histograms
		4.2.2 Parametrized Density Functions as Likelihood Functions
	4.3 Practicalities of Classifier Training
		4.3.1 The Use of Benchmark Databases
		4.3.2 Feature Normalization
		4.3.3 Training and Test Data
		4.3.4 Cross-validation
		4.3.5 Hyperparameter
		4.3.6 Measuring the Classifier Performance
		4.3.7 Imbalanced Data Sets
	4.4 Exercises
		4.4.1 Programming Project P4.1: Classifying MNIST Data
		4.4.2 Exercise Questions
	References
Chapter 5: Distance-Based Classifiers
	5.1 Nearest Centroid Classifier
		5.1.1 Using the Euclidean Distance
		5.1.2 Using the Mahalanobis Distance
	5.2 The kNN Classifier
		5.2.1 Why Does the kNN Classifier Estimate a Posteriori Probabilities
		5.2.2 Efficient Estimation by Space Partitioning
	5.3 Exercises
		5.3.1 Programming Project P5.1: Features and Classifiers
		5.3.2 Exercise Questions
	References
Chapter 6: Decision Boundaries in Feature Space
	6.1 Heuristic Linear Decision Boundaries
		6.1.1 Linear Decision Boundary
		6.1.2 Non-linear Decision Boundaries
		6.1.3 Solving a Multiclass Problem
		6.1.4 Interpretation of Sample Distance from the Decision Boundary
	6.2 Support Vector Machines
		6.2.1 Optimization of a Support Vector Machine
		6.2.2 Soft Margins
		6.2.3 Kernel Functions
		6.2.4 Extensions to Multiclass Problems
	6.3 Logistic Regression
		6.3.1 Binomial Logistic Regression
		6.3.2 Multinomial Logistic Regression
		6.3.3 Kernel Logistic Regression
	6.4 Ensemble Models
		6.4.1 Bagging
		6.4.2 Boosting
	6.5 Exercises
		6.5.1 Programming Project P6.1: Support Vector Machines
		6.5.2 Programming Project P6.2: Label the Imagenette DATA I
		6.5.3 Exercise Questions
	References
Chapter 7: 7 Multi-Layer Perceptron for Image Classification
	7.1 The Perceptron
		7.1.1 Feedforward Step
		7.1.2 Logistic Regression by a Perceptron
		7.1.3 Stochastic Gradient Descent, Batches, and Minibatches
	7.2 Multi-Layer Perceptron
		7.2.1 A Universal, Trainable Classifier
		7.2.2 Networks with More Than Two Layers
	7.3 Training a Multi-Layer Perceptron
		7.3.1 The Backpropagation Algorithm
		7.3.2 The Adam Optimizer
	7.4 Exercises
		7.4.1 Programming Project P7.1: MNIST- and CIFAR10-Labeling by MLP
		7.4.2 Exercise Questions
	References
Chapter 8: Feature Extraction by Convolutional Neural Network
	8.1 The Convolution Layer
		8.1.1 Limited Perceptive Field, Shared Weights, and Filters
		8.1.2 Border Treatment
		8.1.3 Multichannel Input
		8.1.4 Stride
	8.2 Convolutional Building Blocks
		8.2.1 Sequences of Convolution Layers in a CBB
		8.2.2 Pooling
		8.2.3 1 x 1 Convolutions
		8.2.4 Stacking Building Blocks
	8.3 End-to-End Learning
		8.3.1 Gradient Descent in a Convolutional Neural Network
		8.3.2 Initial Experiments
	8.4 Exercises
		8.4.1 Programming Project P8.1: Inspection of a Trained Network
		8.4.2 Exercise Questions
	References
Chapter 9: Network Set-Up for Image Classification
	9.1 Network Design
		9.1.1 Convolution Layers in a CBB
		9.1.2 Border Treatment in the Convolution Layer
		9.1.3 Pooling for Classification
		9.1.4 How Many Convolutional Building Blocks?
		9.1.5 Inception Blocks
		9.1.6 Fully Connected Layers
		9.1.7 The Activation Function
	9.2 Data Set-Up
		9.2.1 Preparing the Training Data
		9.2.2 Data Augmentation
	9.3 Exercises
		9.3.1 Programming Project P9.1: End-to-End Learning to Label CIFAR10
		9.3.2 Programming Project P9.2: CIFAR10 Labeling with Data Augmentation
		9.3.3 Exercise Questions
	References
Chapter 10: Basic Network Training for Image Classification
	10.1 Training, Validation, and Test
		10.1.1 Early Stopping of Network Training
		10.1.2 Fixing Further Hyperparameters
	10.2 Basic Decisions for Network Training
		10.2.1 Weight Initialization
		10.2.2 Loss Functions
		10.2.3 Optimizers, Learning Rate, and Minibatches
		10.2.4 Label Smoothing
	10.3 Analyzing Loss Curves
		10.3.1 Loss After Convergence Is Still Too High
		10.3.2 Loss Does not Reach a Minimum
		10.3.3 Training and Validation Loss Deviate
		10.3.4 Random Fluctuation of the Loss Curves
		10.3.5 Discrepancy Between Validation and Test Results
	10.4 Exercises
		10.4.1 Programming Project P10.1: Label the Imagenette Data II
		10.4.2 Exercise Questions
	References
Chapter 11: Dealing with Training Deficiencies
	11.1 Advanced Augmentation Techniques
		11.1.1 Cutout Augmentation
		11.1.2 Adding Noise to the Input
		11.1.3 Adversarial Attacks
		11.1.4 Virtual Adversarial Training
		11.1.5 Data Augmentation by a Generative Model
		11.1.6 Semi-supervised Learning with Unlabeled Samples
	11.2 Improving Training
		11.2.1 Transfer Learning
		11.2.2 Weight Regularization
		11.2.3 Batch Normalization and Weight Normalization
		11.2.4 Ensemble Learning and Dropout
		11.2.5 Residual Neural Networks
	11.3 Exercises
		11.3.1 Programming Project P11.1: Transfer Learning
		11.3.2 Programming Project P11.2: Label the Imagenette Data III
		11.3.3 Programming Project P11.3: Residual Networks
		11.3.4 Exercise Questions
	References
Chapter 12: Learning Effects and Network Decisions
	12.1 Inspection of Trained Filters
		12.1.1 Display of Trained Filter Values
		12.1.2 Deconvolution for Analyzing Filter Influence
		12.1.3 About Linearly Separable Features
	12.2 Optimal Input
	12.3 How Does a Network Decide?
		12.3.1 Occlusion Analysis
		12.3.2 Class Activation Maps
		12.3.3 Grad-CAM
	12.4 Exercises
		12.4.1 Programming Project P12.1: Compute Optimal Input Images
		12.4.2 Programming Project P12.2: Occlusion Analysis
		12.4.3 Exercise Questions
	References
Index