Computer Vision Using Deep Learning : Neural Network Architectures with Python and Keras

Organizations spend huge resources in developing software that can perform the way a human does. Image classification, object detection and tracking, pose estimation, facial recognition, and sentiment estimation all play a major role in solving computer vision problems. This book will bring into focus these and other deep learning architectures and techniques to help you create solutions using Keras and the TensorFlow library. You'll also review mutliple neural network architectures, including LeNet, AlexNet, VGG, Inception, R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN, YOLO, and SqueezeNet and see how they work alongside Python code via best practices, tips, tricks, shortcuts, and pitfalls. All code snippets will be broken down and discussed thoroughly so you can implement the same principles in your respective environments. __Computer Vision Using Deep Learning__ offers a comprehensive yet succinct guide that stitches DL and CV together to automate operations, reduce human intervention, increase capability, and cut the costs. ## What You'll Learn * Examine deep learning code and concepts to apply guiding principals to your own projects * Classify and evaluate various architectures to better understand your options in various use cases * Go behind the scenes of basic deep learning functions to find out how they work ## Who This Book Is For Professional practitioners working in the fields of software engineering and data science. A working knowledge of Python is strongly recommended. Students and innovators working on advanced degrees in areas related to computer vision and Deep Learning. Table of Contents 5 About the Author 11 About the Technical Reviewer 12 Acknowledgments 13 Introduction 15 Chapter 1: Introduction to Computer Vision and Deep Learning 20 1.1 Technical requirements 21 1.2 Image Processing using OpenCV 22 1.2.1 Color detection using OpenCV 23 1.3 Shape detection using OpenCV 25 1.3.1 Face detection using OpenCV 28 1.4 Fundamentals of Deep Learning 31 1.4.1 The motivation behind Neural Network 33 1.4.2 Layers in a Neural Network 34 1.4.3 Neuron 35 1.4.4 Hyperparameters 36 1.4.5 Connections and weight of ANN 37 1.4.6 Bias term 37 1.4.7 Activation functions 38 1.4.7.1 Sigmoid function 39 1.4.7.2 tanh function 40 1.4.7.3 Rectified Linear Unit or ReLU 41 1.4.7.4 Softmax function 42 1.4.8 Learning rate 44 1.4.9 Backpropagation 45 1.4.10 Overfitting 47 1.4.11 Gradient descent 48 1.4.12 Loss functions 50 1.5 How Deep Learning works? 51 1.5.1 Popular Deep Learning libraries 55 1.6 Summary 57 1.6.1 Further readings 58 Chapter 2: Nuts and Bolts of Deep Learning for Computer Vision 59 2.1 Technical requirements 60 2.2 Deep Learning using TensorFlow and Keras 60 2.3 What is a tensor? 61 2.3.1 What is a Convolutional Neural Network? 63 2.3.2 What is convolution? 64 2.3.3 What is a Pooling Layer? 69 2.3.4 What is a Fully Connected Layer? 70 2.4 Developing a DL solution using CNN 71 2.5 Summary 82 2.5.1 Further readings 84 Chapter 3: Image Classification Using LeNet 85 3.1 Technical requirements 86 3.2 Deep Learning architectures 86 3.3 LeNet architecture 87 3.4 LeNet-1 architecture 88 3.5 LeNet-4 architecture 89 3.6 LeNet-5 architecture 90 3.7 Boosted LeNet-4 architecture 93 3.8 Creating image classification models using LeNet 94 3.9 MNIST classification using LeNet 95 3.10 German traffic sign identification using LeNet 102 3.11 Summary 118 3.11.1 Further readings 119 Chapter 4: VGGNet and AlexNet Networks 120 4.1 Technical requirements 121 4.2 AlexNet and VGG Neural Networks 121 4.3 What is AlexNet Neural Network? 122 4.4 What is VGG Neural Network? 124 4.5 VGG16 architecture 124 4.6 Difference between VGG16 and VGG19 127 4.7 Developing solutions using AlexNet and VGG 128 4.8 Working on CIFAR-10 using AlexNet 130 4.9 Working on CIFAR-10 using VGG 145 4.10 Comparing AlexNet and VGG 153 4.11 Working with CIFAR-100 154 4.12 Summary 155 4.12.1 Further readings 156 Chapter 5: Object Detection Using Deep Learning 157 5.1 Technical requirements 158 5.2 Object Detection 158 5.2.1 Object classification vs. object localization vs. object detection 159 5.2.2 Use cases of Object Detection 160 5.3 Object Detection methods 162 5.4 Deep Learning frameworks for Object Detection 163 5.4.1 Sliding window approach for Object Detection 164 5.5 Bounding box approach 166 5.6 Intersection over Union (IoU) 168 5.7 Non-max suppression 170 5.8 Anchor boxes 171 5.9 Deep Learning architectures 173 5.9.1 Region-based CNN (R-CNN) 173 5.10 Fast R-CNN 176 5.11 Faster R-CNN 178 5.12 You Only Look Once (YOLO) 181 5.12.1 Salient features of YOLO 182 5.12.2 Loss function in YOLO 183 5.12.3 YOLO architecture 185 5.13 Single Shot MultiBox Detector (SSD) 188 5.14 Transfer Learning 193 5.15 Python implementation 195 5.16 Summary 198 5.16.1 Further readings 200 Chapter 6: Face Recognition and Gesture Recognition 202 6.1 Technical toolkit 203 6.2 Face recognition 203 6.2.1 Applications of face recognition 205 6.2.2 Process of face recognition 207 6.2.2.1 Deep Learning modes for face recognition 208 6.2.3 DeepFace solution by Facebook 209 6.2.4 FaceNet for face recognition 214 6.2.5 Python implementation using FaceNet 221 6.2.6 Python solution for gesture recognition 223 6.3 Summary 232 6.3.1 Further readings 234 Chapter 7: Video Analytics Using Deep Learning 235 7.1 Technical toolkit 236 7.2 Video processing 236 7.3 Use cases of video analytics 237 7.4 Vanishing gradient and exploding gradient problem 239 7.5 ResNet architecture 244 7.5.1 ResNet and skip connection 244 7.5.2 Inception network 248 7.5.2.1 1x1 convolutions 249 7.5.3 GoogLeNet architecture 251 7.5.4 Improvements in Inception v2 253 7.6 Video analytics 257 7.7 Python solution using ResNet and Inception v3 258 7.8 Summary 268 7.8.1 Further readings 269 Chapter 8: End-to-End Model Development 270 8.1 Technical requirements 271 8.2 Deep Learning project requirements 271 8.3 Deep Learning project process 275 8.4 Business problem definition 276 8.4.1 Face detection for surveillance 278 8.4.1.1 Defect detection for manufacturing 279 8.4.2 Source data or data discovery phase 281 8.4.2.1 Face detection for identification 281 8.4.2.2 Live environment on a manufacturing line 282 8.5 Data ingestion or data management 283 8.6 Data preparation and augmentation 285 8.6.1 Image augmentation 287 8.7 Deep Learning modeling process 292 8.7.1 Transfer learning 295 8.7.2 Common mistakes/challenges and boosting performance 297 8.8 Model deployment and maintenance 302 8.9 Summary 307 8.9.1 Further readings 309 References 310 Major activation functions and layers used in CNN 310 Google Colab 311 Index 315 Organizations spend huge resources in developing software that can perform the way a human does. Image classification, object detection and tracking, pose estimation, facial recognition, and sentiment estimation all play a major role in solving computer vision problems. This book will bring into focus these and other deep learning architectures and techniques to help you create solutions using Keras and the TensorFlow library. You'll also review mutliple neural network architectures, including LeNet, AlexNet, VGG, Inception, R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN, YOLO, and SqueezeNet and see how they work alongside Python code via best practices, tips, tricks, shortcuts, and pitfalls. All code snippets will be broken down and discussed thoroughly so you can implement the same principles in your respective environments. Computer Vision Using Deep Learning offers a comprehensive yet succinct guide that stitches DL and CV together to automate operations, reduce human intervention, increase capability, and cut the costs. You will: Examine deep learning code and concepts to apply guiding principles to your own projects Classify and evaluate various architectures to better understand your options in various use cases Go behind the scenes of basic deep learning functions to find out how they work