Document processing is not just optical character recognition (OCR). This book demonstrates this with topics as diverse as the problem of writer analysis in the forensic field to the processing of mathematical expressions. However, OCR is the main subject of this text. This is a handbook of document processing topics and provides broader coverage than the typical compilation of research papers. Most chapters are reviews of a topic, rather than reports on particular methods or experiments.

Chapter 1, “Reading Systems,” contains a broad review of the issues to be addressed in document processing systems, and surveys the system components required for a computer to read a document. The current capabilities and limitations of each component are discussed. A comparison is made between the reading abilities of humans and computers. Chapter 2 provides a broad review of layout-analysis methods, concentrating on scanned, offline image analysis.

The mainstay OCR application of book scanning has come to prominence with the large-scale scanning of libraries by Google and others. While conversion of machine-generated Roman script to ASCII or Unicode is robustly and accurately done with available OCR systems, this book covers more specialized topics where software is not available, or where, if software is available, the accuracy or robustness is still low. OCR of multiscript texts is still a challenge. OCR of Japanese, Tibetan, and Indian scripts is covered in three chapters of this book. Another ongoing OCR challenge is handwriting, especially online with tablets and digital pens. This area is addressed in chapters 6 through 9. The systems view of chapter 1 is useful here, because it allows us to see the quite different emphasis required for digital tablet input compared to OCR input of historical handwritten journals or diaries.

The systems-level nature of this book comes through with the chapters on postal address analysis and check processing (chapters 10 and 20). More technically focused chapters present OCR of math expressions, handling robustness, graphics recognition (the pixel-to-vector transform and shape classification), and super-resolution techniques.

The latter part of the book switches its focus to information extraction and retrieval. Chapter 15 reviews the task of metadata extraction from bibliographic documents, and chapters 18 and 19 review extracting parts of Web pages.

Although I am not an active researcher in document processing, and thus not in the target audience, I found this book to be an accessible, useful review of the field. There is some variability in writing quality across the chapters, but the quality, look, and feel are consistently maintained. The review chapters make fine use of diagrams, charts, and images to ground the topics being reviewed with visual examples.

A theme that arises either explicitly or implicitly across the chapters is the need for benchmark corpora and agreed-upon performance evaluation metrics. The metrics of the National Institute of Standards and Technology’s (NIST’s) TREC, in the overlapping field of information retrieval, are an obvious example of the effect of benchmarking on a field’s progress. In this respect, the document processing field is relatively immature with regard to its mature years and widespread deployment. Perhaps this is explained by the diversity of niche problems as reviewed in this book, leading to a mature dismissal of the dream of a universal reading machine.