Chapter 12 – Computer vision

12.1  Overview

12.2  Introduction

12.2.1  Why Computer Vision Is Difficult

12.2.2  Phases of Computer Vision

12.3  Digitisation and Signal Processing

12.3.1  Digitising Images

12.3.2  Thresholding

12.3.3  Digital Filters

12.3.3.1  Linear Filters

12.3.3.2  Smoothing

12.3.3.3  Gaussian Filters

12.3.3.4  Practical Considerations

12.4  Edge Detection

12.4.1  Identifying Edge Pixels

12.4.1.1  Gradient Operators

12.4.1.2  Robert's Operator

12.4.1.3  Sobel's Operator

12.4.1.4  Laplacian Operator

12.4.1.5  Successive Refinement and Marr's Primal Sketch

12.4.2  Edge Following

12.5  Region Detection

12.5.1  Region Growing

12.5.2  The Problem of Texture

12.5.3  Representing Regions -- Quadtrees

12.5.4  Computational Problems

12.6  Reconstructing Objects

12.6.1  Inferring Three-Dimensional Features

12.6.1.1  Problems with Labelling

12.6.2  Using Properties of Regions

12.7  Identifying Objects

12.7.1  Using Bitmaps

12.7.2  Using Summary Statistics

12.7.3  Using Outlines

12.7.4  Using Paths

12.8  Facial and Body Recognition

12.9  Neural Networks for Images

12.9.1  Convolutional Neural Networks

12.9.2  Autoencoders

12.10  Generative Adversarial Networks

12.10.1  Generated Data

12.10.2  Diffusion Models

12.10.3  Bottom-up and Top-down Processing

12.11  Multiple Images

12.11.1  Stereo Vision

12.11.2  Moving Pictures

12.12  Summary

accuracy, active vision, ambiguous image, aspect ratio, auto-associative memory, autoencoder, backpropagation, Bayes Theorem, binary image, bitmap image, Boltzmann machine, bottom-up reasoning, camera!pan, camera!zoom, clustering, connectionist model, constraint satisfaction, constraints, contour following, convolutional neural network, convolutions, correlation, crowdsourcing, data structure, database, deep fakes, deep neural network, differential (calculus), diffusion models, digital filtering, digital signal processing, digitisation, edge detection, edge following, emotion recognition, facial recognition, false positive, frame of video, game playing, Gaussian filter, generative adversarial network, geographic information system, geometric constraints, gesture recognition, Google, gradient descent, gradient operators, graphics processing unit, grey-scale image, ground truth, handwriting recognition, heuristic evaluation function, Hopfield networks, human perception, hybrid architecture, image thresholding, image understanding, labelling, Laplacian operator, Laplacian-of-Gaussian filter, line labelling, linear filter, machine learning, Marr's primal sketch, moving images, multiple images, neural network, neural-network architecture, Normal distribution, normalisation, object identification, object identification!bitmaps, object identification!outlines, object identification!paths, object identification!summary statistics, object recognition, OCR (optical character recognition), octree, optical flow, parallax, parallel processing, pattern matching, pen-based systems, position independent, pre-processing, privacy, quadtree, reasoning with uncertainty, receptive field, region detection, region growing, restricted Boltzmann machine, ridge, Robert's operator, robotics, segmentation, sensation, sensor fusion, sharpening filters, signal processing, similarity measure, Skynet, smoothing, Sobel's operator, standard deviation, stereo vision, successive refinement, template matching, texture, three-dimensional objects, threshold, time series, unsupervised learning, voxel, Waltz's algorithm, wavelet transform, zero-sum game