Computer Vision Projects with PyTorch: Design and Develop Production-Grade Models

✍ Scribed by Akshay Kulkarni, Adarsha Shivananda, Nitin Ranjan Sharma

• Publisher: Apress
• Year: 2022
• Tongue: English
• Leaves: 355
• Category: Library

✦ Synopsis

Design and develop end-to-end, production-grade computer vision projects for real-world industry problems. This book discusses computer vision algorithms and their applications using PyTorch.
The book begins with the fundamentals of computer vision: convolutional neural nets, RESNET, YOLO, data augmentation, and other regularization techniques used in the industry. And then it gives you a quick overview of the PyTorch libraries used in the book. After that, it takes you through the implementation of image classification problems, object detection techniques, and transfer learning while training and running inference. The book covers image segmentation and an anomaly detection model. And it discusses the fundamentals of video processing for computer vision tasks putting images into videos. The book concludes with an explanation of the complete model building process for deep learning frameworks using optimized techniques with highlights on model AI explainability.
After reading this book, you will be able to build your own computer vision projects using transfer learning and PyTorch.

What You Will Learn

• Solve problems in computer vision with PyTorch.
• Implement transfer learning and perform image classification, object detection, image segmentation, and other computer vision applications
• Design and develop production-grade computer vision projects for real-world industry problems
• Interpret computer vision models and solve business problems

Who This Book Is For
Data scientists and machine learning engineers interested in building computer vision projects and solving business problems

✦ Table of Contents

Table of Contents
About the Authors
About the Technical Reviewer
Introduction
Chapter 1: The Building Blocks of Computer Vision
What Is Computer Vision
Applications
Classification
Object Detection and Localization
Image Segmentation
Anomaly Detection
Video Analysis
Channels
Convolutional Neural Networks
Receptive Field
Local Receptive Field
Global Receptive Field
Pooling
Max Pooling
Average Pooling
Global Average Pooling
Calculation: Feature Map and Receptive Fields
Kernel
Stride
Pooling
Padding
Input and Output
Calculation of Receptive Field
Understanding the CNN Architecture Type
Understanding Types of Architecture
AlexNet
VGG
ResNet
Inception Architectures
Working with Deep Learning Model Techniques
Batch Normalization
Dropouts
Data Augmentation Techniques
Introduction to PyTorch
Installation
Basic Start
Summary
Chapter 2: Image Classification
Topics to Cover
Defining the Problem
Overview of the Approach
Creating an Image Classification Pipeline
First Basic Model
Data
Data Exploration
Data Loader
Define the Model
The Training Process
The Second Variation of Model
The Third Variation of the Model
The Fourth Variation of the Model
Summary
Chapter 3: Building an Object Detection Model
Object Detection Using Boosted Cascade
R-CNN
The Region Proposal Network
Fast Region-Based Convolutional Neural Network
How the Region Proposal Network Works
The Anchor Generation Layer
The Region Proposal Layer
Mask R-CNN
Prerequisites
YOLO
YOLO V2/V3
Project Code Snippets
Step 1: Getting Annotated Data
Step 2: Fixing the Configuration File and Training
The Model File
Summary
Chapter 4: Building an Image Segmentation Model
Image Segmentation
Pretrained Support from PyTorch
Semantic Segmentation
Instance Segmentation
Fine-Tuning the Model
Summary
Chapter 5: Image-Based Search and Recommendation System
Problem Statement
Approach and Methodology
Implementation
The Dataset
Installing and Importing Libraries
Importing and Understanding the Data
Feature Engineering
ResNet18
Calculating Similarity and Ranking
Visualizing the Recommendations
Taking Image Input from Users and Recommending Similar Products
Summary
Chapter 6: Pose Estimation
Top-Down Approach
Bottom-Up Approach
OpenPose
Branch-1
Branch-2
HRNet (High-Resolution Net)
Higher HRNet
PoseNet
How Does PoseNet Work?
Single Person Pose Estimation
Multi-Person Pose Estimation
Pros and Cons of PoseNet
Applications of Pose Estimation
Test Cases Performed Retail Store Videos
Implementation
Step 1: Identify the List of Human Keypoints to Track
Step 2: Identify the Possible Connections Between the Keypoints
Step 3: Load the Pretrained Model from the PyTorch Library
Step 4: Input Image Preprocessing and Modeling
Step 5: Build Custom Functions to Plot the Output
Step 6: Plot the Output on the Input Image
Summary
Chapter 7: Image Anomaly Detection
Anomaly Detection
Approach 1: Using a Pretrained Classification Model
Step 1: Import the Required Libraries
Step 2: Create the Seed and Deterministic Functions
Step 3: Set the Hyperparameter
Step 4: Import the Dataset
Step 5: Image Preprocessing Stage
Step 6: Load the Pretrained Model
Step 7: Freeze the Model
Step 8: Train the Model
Step 9: Evaluate the Model
Approach 2: Using Autoencoder
Step 1: Prepare the Dataset Object
Step 2: Build the Autoencoder Network
Step 3: Train the Autoencoder Network
Step 4: Calculate the Reconstruction Loss Based on the Original Data
Step 5: Select the Most Anomalous Digit Based on the Error Metric Score
Output
Summary
Chapter 8: Image Super-Resolution
Up-Scaling Using the Nearest Neighbor Concept
Understanding Bilinear Up-Scaling
Variational Autoencoders
Generative Adversarial Networks
The Model Code
Model Development
Imports
Running the Application
Summary
Chapter 9: Video Analytics
Problem Statement
Approach
Implementation
Data
Uploading the Required Videos to Google Colab
Convert the Video to a Series of Images
Image Extraction
Data Preparation
Identify the Hotspots in a Retail Store
Importing Images
Getting Crowd Counts
Security and Surveillance
Identify the Demographics (Age and Gender)
Summary
Chapter 10: Explainable AI for Computer Vision
Grad-CAM
Grad-CAM++
NBDT
Step 1
Step 2
Steps 3 and 4
Grad-CAM and Grad-CAM++ Implementation
Grad-CAM and Grad-CAM++ Implementation on a Single Image
NBDT Implementation on a Single Image
Summary
Index