2D object detection and recognition

Detecting and recognizing two-dimensional objects in grayscale pictures are two central challenges in machine vision, and this book takes them on directly. It explains how to build and train the models involved, how to make the computations efficient, and how to carry out parallel implementations in neural network designs that resemble biological systems. The work rests on statistical modeling and estimation, and it favors models that are simple, easy to follow, and fast to run.

The book examines a family of deformable template models. At one end are coarse, sparse versions built from discrete steps that compute quickly. At the other end are finely detailed versions formulated over continuous space, which demand heavy optimization. Each model is specified by a set of points placed on a reference grid, called the template, together with a collection of permitted deformations of those points, and a statistical description of the image data for any given deformed instance of the object. Moving from coarse to fine solutions recurs throughout as a guiding idea for vision problems.

Detailed explanations of the algorithms accompany the text, along with the code. The software and data sets used are available online.