This slim volume considers the question of representing and matching textures and similar patterns. This question is closely related to recognition--for example, recognizing that something is a picture of a leafless tree--and segmentation, that is, picking out the distinct (leaved) trees in a view of a forest or extracting a figure from a background. The aim of the book is to introduce readers to current ideas in the field, in particular the use of deep learning. The book has ten chapters and an appendix.
The first chapter is an overview of patterns and textures, particularly: “What are the set of desired requirements for a computational representation for patterns or texture?” Chapter 2 introduces textons, originally a “local configuration of image features”; by recognizing when the density of a texton changes, the region can be segmented much in the way that human pre-attentive vision works. Nice illustrations clarify the ideas. Currently a texton is defined as a collection of filter outputs, for example, a wxh image convolved with nine filters.
Chapter 3 discusses texture recognition and introduces the use of deep learning for texture recognition. Several systems that use deep learning are cited; readers will need to consult them for detailed descriptions of the techniques. Deep learning’s ability to handle large quantities of data allows for the creation of substantial texton dictionaries. Chapter 4 discusses segmentation and gives detailed yet simple worked examples of the ways in which this can be done. Graph-based and Markov random field methods are considered along with deep learning methods. Chapter 5 deals with texture synthesis, first describing traditional methods, for example, image quilting, before outlining the use of deep learning in texture synthesis.
Chapter 6 deals with texture style transfer, where the style of one image is transferred to another while preserving the semantic context of the latter, for example, aging a face. Nice pictures illustrate the effect. The very brief chapter 7 recalls the ideas of pyramid methods in which multi-scale representations are layered. The chapter cites a number of systems that use multi-scale representations. Chapter 8, another brief one, considers some open issues in the understanding of visual patterns. A particular and potentially important example is the detection of subtle change, such as might be relevant for the early detection of disease.
Chapter 9 introduces several application areas: medicine, industry, e-commerce, textured solar panels, and road analysis for automated driving. Chapter 10 discusses tools for mining patterns, specifically software libraries and cloud services.
The appendix is a concise description of deep learning. In keeping with the style of the book, the coverage is quite high level with many citations. Each of the first six chapters and the appendix include valuable exercises. The substantial bibliography would have benefited from being organized in some categorical way.
Altogether the book fulfills its purpose of providing an overview of the topic. Readers will need to consult the various cited sources for details on how existing systems work.
The book will be particularly useful as an introduction for people wishing to work in the area.