This book addresses the study of computer systems for accessing images from very large image databases, where retrieval is based directly on image contents, rather than on supporting metadata.
The first 50 pages, which are effectively chapters 1 and 2, address the role of image retrieval (IR) systems: who would use them, and why. These chapters are a little self-indulgent and rambling. This is probably because the material addressed in these chapters is itself somewhat nebulous; this is the weakest part of the book. The next chapter addresses evaluation, and it is strange to find this material introduced so early in the book. In contrast to earlier chapters, this chapter gets quite technical, with detailed coverage of topics like statistical significance testing.
Chapter 4, “Similarity,” is the chapter that I enjoyed the most. It is a good, self-contained essay on similarity in general: how to measure it, and what kind of metrics can be used. This material will provide an excellent grounding for students of many disciplines in computing. In chapter 5, the author presents a review of content-based image retrieval systems. In this area the author is on comfortable ground, addressing retrieval based on color histograms, textures, and shapes--the materials to be expected in a book on image retrieval. The book as a whole, however, tries to go beyond such “level 1” retrieval. Its aim is to explore the deeper role and meaning of content-based retrieval of images, rather than simply to present algorithms and techniques for image matching. This requires a high level of image analysis and computer vision, so chapter 6 has lots of material on these topics. This material is quite comprehensive and very technical, making it quite hard work to get through.
Chapter 7 covers text-based information retrieval, including probabilistic and vector space models, latent semantic indexing, relevance feedback, web structure, and linkage analysis--all of the topics I cover in my graduate course in information access. Chapter 8 moves on to relational algebras, and presents an extension of the classical relational algebra that allows the specification of image queries. Chapter 9, a very technical exposition of indexing techniques for high dimensional feature spaces found in image retrieval, is not an easy chapter to read.
The final chapter of the book is broadly about interface issues, how to specify queries, and how to present results, though this is presented in a general context, and is not specific to image databases. And then, suddenly, the book ends.
Reading the book is more like reading about an odyssey or a journey than reading a presentation of research achievements or an overview of a field. The book is 570 pages, excluding bibliography and appendices, and covers a huge amount of material that is very broad in nature. The reader is taken into areas that appear to have very tenuous connections. In the end, I was left wondering how it all fits together. The book needs a conclusion, and some attempt to bring closure to the odyssey.
The prose is heavy in places, and the book is not an easy read, but neither is it probably intended to be. The author includes frequent references to an Italian background and culture. The theoretical nature of the topics--similarity functions for measuring the distances between two color histograms for two images, for example--is quite mathematical. Although there are some worked examples, the book is not an undergraduate textbook. If anything, it would be suitable for a graduate course.
So, would I recommend this book? For coverage of certain topics, such as similarity, or as an introduction to IR, or as a quick overview of simple approaches to IR, it is fine. If a reader is new to the area, and is looking for a comprehensive overview of image retrieval, then I recommend Lew [1]. If a reader is already reasonably knowledgeable in this field, however, and is looking for something a little deeper and more thought-provoking, then this book is a good choice.
