-
 Algorithms
for Image Processing and Computer Vision, J. R. Parker.
John Wiley & Sons, 1997.
If it were possible to learn to do OCR from one book, this might be the
right one. The algorithms surveyed (with a wealth of "C" source code
included on CD-ROM) cover a wide range of techniques but particularly emphasize
those which arise in an OCR context, such as segmentation, edge detection,
skeletonization, and morphological filtering. Chapter 8 is a tutorial
in which an OCR system is developed, component by component, until a usable
system is obtained for reading machine-print characters from fax
images. Handprint recognition is treated as a separate problem in
Chapter 9, which also contains a well-organized survey of techniques for
combining multiple classifiers. This book covers both classical
image processing heuristics and newer general techniques such as wavelets,
neural networks and genetic algorithms. However, the overall approach
is bottom-up, with the OCR system of Chapter 8 being designed as a succession
of solutions to particular problems. This is a practical approach
(made yet more practical by the inclusion of source code) and reflects
the manner in which much of the OCR field was established. However,
it doesn't do much to place OCR in the context of the larger field of
pattern analysis and recognition. So, if you were to choose two
books, a good complement to Parker would be one that takes a
top-down approach, such as Schürmann.
ISBN 0-471-14056-2 (softcover with CD-ROM).
|
-
 Digital Document Processing, H. S.
Hou. John
Wiley & Sons, 1983. Gives a complete view of computerized document
processing from scanner to output, with compact but informative descriptions
of the important algorithmic techniques at each stage. Covers text recognition
from the image quality, character feature, and language-dependent perspectives,
as well as other techniques used in a complete document processing system
including image halftoning, compression, transmission, and document retrieval.
Because the book concentrates on the major basic algorithms in each area
rather than specific implementations, the material holds up well despite
the age of the book. Age does, however, make the book difficult to
obtain. ISBN 0471862479.
|
This "unified view" has been a long time
in coming and is presented with wonderful lucidity by Schürmann.
Fundamentals of classifier performance are described (with frequent reference
to character recognition, but in a way that doesn't depend on the application.)
Neural approaches including multilayer perceptrons and radial basis functions,
and classical regression approaches such as the polynomial classifier,
are developed as alternative extensions of the same principles. Methods
of selecting features, setting reject criteria, evaluating classifier performance,
and combining classifiers are described in good detail and in a way that
does not depend on the classification approach selected. An accompanying
statistical modeling program can be downloaded by FTP from Wiley's site.
ISBN 0471135348. 
the Bayes Decision Theorem from an almost-trivial formula to the basis
for a real understanding of how to deal with probabilistic or incomplete
data. Despite its basis in perhaps the best-known rule in pattern recognition,
I would still call this book an "alternative" approach since Bayesian networks
have not (yet?) enjoyed the widespread application of some other techniques.
Contains detailed comparisions with related approachs such as Markov fields
and Dempster-Shafer theory. Doesn't skimp on mathematical rigor, but is
also simply a great read. ISBN 0934613737. 
and K-S. Fu, eds. 
J. McLachlan. 
a tutorial as the Kohonen and Kosko books are, but a collection of
IEEE-refereed
papers selected for their importance to the theory or actual application
of fuzzy logic in the design of pattern recognition systems. Includes Zadeh's
original 1965 paper describing fuzzy sets, and many more on cluser analysis,
feature selection and classifier design, image processing, and adaptive
("learning") behavior.
image processing since the publication
of the first edition in 1976. A section on human visual perception provides
a context for the enhancement and restoration techniques described. Besides
the usual linear techniques, Rosenfeld and Kak describe geometry-preserving
nonlinear approaches (e.g. region growing, line thinning) applicable to
image segmentation. All these techniques are important to image recognition,
but the book is not directed at recognition specifically and does not delve
into classifier design. ISBN 0125973012. There is a second volume edited
by Rosenfeld, ISBN 0125973020.