In this comprehensive text on optical character recognition (OCR), the authors blend the theoretical with the practical, the breadth of survey with the depth of analysis, and the algorithmic approach with the mathematical. As such, the book will be of interest to anyone interested in OCR, regardless of perspective. Since the book involves theorem proving and integral calculus, the reader should understand that it is not a trivial introduction to the subject. Yet, because of its thorough treatment of the topics discussed, the reader is not left alone in the development of one equation from the previous ones. It is certainly a text for someone interested in the cutting edge of the technology. It does, at times, favor a signal processing perspective in presenting the material.
The book consists of 13 chapters, three appendices, and extensive references for each chapter. It begins with an introduction to the problems involved in OCR and presents simplified approaches that one might conceive of to solve them. These approaches do not work well in practice. A main reason for this is that the data are usually noisy, an issue that is dealt with in the next chapter. A related issue is that templates of characters generally assume a certain typography, while documents (especially handwritten ones) may shift to other fonts and point sizes. Chapter 3, therefore, covers normalization, presenting four different models of normalization (Iijima, Amari, linear, and nonlinear).
Assuming that the issues of noise and scale have been dealt with, the OCR program will need to appropriately segment the relevant portions of the image that contain the actual text scanned in. This is discussed in chapter 4, and a series of experiments are presented that test various nonadaptive thresholding methods. Once the characters have been isolated from the background of the document, preprocessing steps are usually employed in order to simplify the subsequent matching phase. The most standard of these steps is thinning the character to its skeleton. Chapter 5 introduces this concept, based on connectivity, and discusses the Hilditch, Pavlidis, and graph approaches. Other suggested preprocessing steps are discussed in chapter 6, where morphological techniques are defined.
Whereas the first half of the book deals more with preparing the image and segmented character data, the second half addresses topological representations of the features in a given character and matching of the structures obtained from the image to those stored in a character database. Chapter 7 presents feature extraction based on linear methods. The mathematics necessary to understand this chapter is involved; it is based on functional analysis, invariant features using Zernike moments, Fourier expansions and descriptors, and Karhunen-Loeve expansions. Chapter 8 continues the discussion of feature extraction, based on structural methods. These methods include convex decomposition, stream-following, polygonal approximations, split and merge, contour following, and corner detection methods.
Once the different features of a given character have been extracted from the image, chapters 9 and 10 deal with assembling all of this information into one shape descriptor. Chapter 9 accomplishes this with an algebraic approach, and chapter 10 integrates the clear background that a character rests on with the general description of the strokes used to form the actual character. Together, these form a more unified description of shape. Chapters 11 through 13 describe numerous matching techniques for comparing the shape obtained with the shapes in a database. The three appendices provide some extra information about some of the details in the main text and discuss data sets for the experiments considered.
The book is extensive and comprehensive, and is worthwhile reading for researchers interested in optical character recognition. It could also serve as a text for a graduate special topics course in OCR.
