Download the full version of the ebook at

https://ebookultra.com

Rough Fuzzy Image Analysis Foundations and

Methodologies 1st Edition Sankar K. Pal

https://ebookultra.com/download/rough-fuzzy-image-
analysis-foundations-and-methodologies-1st-
edition-sankar-k-pal/

Explore and download more ebook at https://ebookultra.com

 Recommended digital products (PDF, EPUB, MOBI) that
you can download immediately if you are interested.

Image Analysis and Modeling in Ophthalmology 1st Edition

Eddie Y. K. Ng

https://ebookultra.com/download/image-analysis-and-modeling-in-
ophthalmology-1st-edition-eddie-y-k-ng/

ebookultra.com

Foundations of Fuzzy Logic and Semantic Web Languages 1st

Edition Umberto Straccia

https://ebookultra.com/download/foundations-of-fuzzy-logic-and-
semantic-web-languages-1st-edition-umberto-straccia/

ebookultra.com

Introduction to Fuzzy Logic 1st Edition James K. Peckol

https://ebookultra.com/download/introduction-to-fuzzy-logic-1st-
edition-james-k-peckol/

ebookultra.com

Fuzzy Intelligent Systems Methodologies Techniques and

Applications Artificial Intelligence and Soft Computing
for Industrial Transformation 1st Edition R. Anandan
https://ebookultra.com/download/fuzzy-intelligent-systems-
methodologies-techniques-and-applications-artificial-intelligence-and-
soft-computing-for-industrial-transformation-1st-edition-r-anandan/
ebookultra.com
Dermoscopy image analysis 1st Edition Celebi

https://ebookultra.com/download/dermoscopy-image-analysis-1st-edition-
celebi/

ebookultra.com

The Jazz Image K. Heather Pinson

https://ebookultra.com/download/the-jazz-image-k-heather-pinson/

ebookultra.com

Statistical Modeling Analysis and Management of Fuzzy Data

1st Edition Irwin R. Goodman

https://ebookultra.com/download/statistical-modeling-analysis-and-
management-of-fuzzy-data-1st-edition-irwin-r-goodman/

ebookultra.com

Still Image and Video Compression with MATLAB 1st Edition

K. S. Thyagarajan

https://ebookultra.com/download/still-image-and-video-compression-
with-matlab-1st-edition-k-s-thyagarajan/

ebookultra.com

Pal Benko My Life Games and Compositions 1st Edition Pal

Benko & Jeremy Silman & John L. Watson

https://ebookultra.com/download/pal-benko-my-life-games-and-
compositions-1st-edition-pal-benko-jeremy-silman-john-l-watson/

ebookultra.com
Rough Fuzzy Image Analysis Foundations and
Methodologies 1st Edition Sankar K. Pal Digital Instant
Download
Author(s): Sankar K. Pal, James F. Peters
ISBN(s): 1439803293
Edition: 1
File Details: PDF, 8.65 MB
Year: 2010
Language: english
 Rough Fuzzy
Image Analysis
Foundations and Methodologies

K10185_FM.indd 1 3/29/10 1:38:14 PM

 Chapman & Hall/CRC
Mathematical and Computational
Imaging Sciences

Series Editors

Chandrajit Bajaj Guillermo Sapiro

Center for Computational Visualization Department of Electrical
The University of Texas at Austin and Computer Engineering
University of Minnesota

Aims and Scope

This series aims to capture new developments and summarize what is known over the whole
spectrum of mathematical and computational imaging sciences. It seeks to encourage the
integration of mathematical, statistical and computational methods in image acquisition and
processing by publishing a broad range of textbooks, reference works and handbooks. The
titles included in the series are meant to appeal to students, researchers and professionals
in the mathematical, statistical and computational sciences, application areas, as well as
interdisciplinary researchers involved in the field. The inclusion of concrete examples and
applications, and programming code and examples, is highly encouraged.

Proposals for the series should be submitted to the series editors above or directly to:
CRC Press, Taylor & Francis Group
4th, Floor, Albert House
1-4 Singer Street
London EC2A 4BQ
UK

K10185_FM.indd 2 3/29/10 1:38:14 PM

 Chapman & Hall/CRC
Mathematical and Computational Imaging Sciences

Rough Fuzzy
Image Analysis
Foundations and Methodologies

Edited by
Sankar K. Pal
James F. Peters

K10185_FM.indd 3 3/29/10 1:38:14 PM

 CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2010 by Taylor and Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed in the United States of America on acid-free paper

10 9 8 7 6 5 4 3 2 1

International Standard Book Number: 978-1-4398-0329-5 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been
made to publish reliable data and information, but the author and publisher cannot assume responsibility for the valid-
ity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright
holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this
form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may
rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or uti-
lized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopy-
ing, microfilming, and recording, or in any information storage or retrieval system, without written permission from the
publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.com (http://
www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923,
978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For
organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.

Library of Congress Cataloging‑in‑Publication Data

Rough fuzzy image analysis : foundations and methodologies / editors, Sankar K. Pal, James F. Peters.
p. cm.
“A CRC title.”
Includes bibliographical references and index.
ISBN 978-1-4398-0329-5 (hardcover : alk. paper)
1. Image analysis. 2. Fuzzy sets. I. Pal, Sankar K. II. Peters, James F. III. Title.

TA1637.R68 2010
621.36’7--dc22 2009053741

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com

K10185_FM.indd 4 3/29/10 1:38:14 PM

 Preface

This book introduces the foundations and applications in the state-of-art of rough-
fuzzy image analysis. Fuzzy sets* and rough sets** as well as a generalization of
rough sets called near sets*** provide important as well as useful stepping stones
in various approaches to image analysis that are given in the chapters of this book.
These three types of sets and various hybridizations provide powerful frameworks
for image analysis.
Image analysis focuses on the extraction of meaningful information from digital
images. This subject has its roots in studies of space and the senses by J.H. Poincaré
during the early 1900s, studies of visual perception and the topology of the brain
by E.C. Zeeman and picture processing by A.P. Rosenfeld**** . The basic picture
processing approach pioneered by A.P. Rosenfeld was to extract meaningful patterns
in given digital images representing real scenes as opposed to images synthesized by
the computer. Underlying picture processing is an interest in filtering a picture to
detect given patterns embedded in digital images and approximating a given image
with simpler, similar images with lower information content (this, of course, is at the
heart of the near set-based approach to image analysis). This book calls attention
to the utility that fuzzy sets, near sets and rough sets have in image analysis. One
of the earliest fuzzy set-based image analysis studies was published in 1982 by S.K.
Pal***** . The spectrum of fuzzy set-oriented image analysis studies includes edge
ambiguity, scene analysis, image enhancement using smoothing, image description,
motion frame analysis, medical imaging, remote sensing, thresholding and image
frame analysis.
The application of rough sets in image analysis was launched in a seminal paper
published in 1993 by A. Mrózek and L. Plonka****** . Near sets are a recent gener-
alization of rough sets that have proven to be useful in image analysis and pattern

* See, e.g., Zadeh, L.A., Fuzzy sets. Information and Control (1965), 8 (3) 338-353; Zadeh, L.A.,

Toward a theory of fuzzy granulation and its centrality in human reasoning and fuzzy logic, Fuzzy
Sets and Systems 90 (1997), 111-127. See, also, Rosenfeld, A., Fuzzy digital topology, in Bezdek,
J.C., Pal, S.K., Eds., Fuzzy Models for Pattern Recognition, IEEE Press, 1991, 331-339; Banerjee,
M., Kundu, M.K., Maji, P., Content-based image retrieval using visually significant point features,
Fuzzy Sets and Systems 160, 1 (2009), 3323-3341; http://en.wikipedia.org/wiki/Fuzzy set
** See, e.g., Peters, J.F., Skowron, A.: Zdzislaw Pawlak: Life and Work, Transactions on Rough

Sets V, (2006), 1-24; Pawlak, Z., Skowron, A.: Rudiments of rough sets, Information Sciences 177
(2007) 3-27; Pawlak, Z., Skowron, A.: Rough sets: Some extensions, Information Sciences 177
(2007) 28-40; Pawlak, Z., Skowron, A.: Rough sets and Boolean reasoning, Information Sciences
177 (2007) 41-73.; http://en.wikipedia.org/wiki/Rough set
*** See, e.g., Peters, J.F., Puzio, L., Image analysis with anisotropic wavelet-based near-

ness measures, Int. J. of Computational Intelligence Systems 79, 3-4 (2009), 1-17; Peters,
J.F., Wasilewski, P., Foundations of near sets, Information Sciences 179, 2009, 3091-3109;
http://en.wikipedia.org/wiki/Near sets. See, also, http://wren.ee.umanitoba.ca
**** See, e.g., Rosenfeld, A.P., Picture processing by computer, ACM Computing Surveys 1, 3

(1969), 147-176
***** Pal, S.K., A note on the quantitative measure of image enhancement through fuzziness, IEEE

Trans. on Pat. Anal. & Machine Intelligence 4, 2 (1982), 204-208.

****** Mrózek, A., Plonka, L., Rough sets in image analysis, Foundations of Computing and Decision

Sciences 18, 3-4 (1993), 268-273.

 0-2

recognition******* .
This volume fully reflects the diversity and richness of rough fuzzy image analysis
both in terms of its underlying set theories as well as its diverse methods and appli-
cations. From the lead chapter by J.F. Peters and S.K. Pal, it can be observed that
fuzzy sets, near sets and rough sets are, in fact, instances of different incarnations
of Cantor sets. These three types of Cantor sets provide a foundation for what
A. Rosenfeld points to as the stages in pictorial pattern recognition, i.e., image
transformation, feature extraction and classification. The chapters by P. Maji and
S.K. Pal on rough-fuzzy clustering, D. Malyszko and J. Stepaniuk on rough-fuzzy
measures, and by A.E. Hassanien, H. Al-Qaheri, A. Abraham on rough-fuzzy clus-
tering for segmentation point to the utility of hybrid approaches that combine fuzzy
sets and rough sets in image analysis. The chapters by D. Sen, S.K. Pal on rough
set-based image thresholding, H. Fashandi, J.F. Peters on rough set-based mathe-
matical morphology as well as an image partition topology and M.M. Mushrif, A.K.
Ray on image segmentation illustrate how image analysis can be carried out with
rough sets by themselves. Tolerance spaces and a perceptual approach in image
analysis can be found in the chapters by C. Henry, A.H. Meghdadi, J.F. Peters, S.
Shahfar, and S. Ramanna (these papers carry forward the work on visual perception
by J.H. Poincaré and E.C. Zeeman). A rich harvest of applications of rough fuzzy
image analysis can be found in the chapters by A.E. Hassanien, H. Al-Qaheri, A.
Abraham, W. Tarnawski, G. Schaefer, T. Nakashima, L. Miroslaw, C. Henry, S.
Shahfar, A.H. Meghdadi and S. Ramanna. Finally, a complete, downloadable im-
plementation of near sets in image analysis called NEAR is presented by C. Henry.
The Editors of this volume extend their profound gratitude to the many reviewers
for their generosity and many helpful comments concerning the chapters in this
volume. Every chapter was extensively reviewed and revised before final acceptance.
We also received many helpful suggestions from the reveiwers of the original proposal
for this CRC Press book. In addition, we are very grateful for the help that we
have received from S. Kumar, A. Rodriguez, R.B. Stern, S.K. White, J. Vakili and
others at CRC Press during the preparation of this volume.
The editors of this volume have been supported by the Natural Sciences and En-
gineering Research Council of Canada (NSERC) research grant 185986, Manitoba
Centre of Excellence Fund (MCEF) grant, Canadian Network of Excellence (NCE)
and Canadian Arthritis Network (CAN) grant SRI-BIO-05, and the J.C. Bose Fel-
lowship of the Government of India.

March 2010 Sankar K. Pal

James F. Peters

******* See,e.g.,
Gupta, S., Patnik, S., Enhancing performance of face recognition by using the near
set approach for selecting facial features, J. Theor. Appl. Inform. Technol. 4, 5 (2008), 433-441;
Henry, C., Peters, J.F., Perception-based image analysis, Int. J. Bio-Inspired Comp. 2, 2 (2009),
in press; Peters, J.F., Tolerance near sets and image correspondence, Int. J. of Bio-Inspired
Computation 1(4) (2009), 239-245; Peters, J.F., Corrigenda and addenda: Tolerance near sets and
image correspondence, Int. J. of Bio-Inspired Computation 2(5) (2010), in press; Ramanna, S.,
Perceptually near Pawlak partitions, Transactions on Rough Sets XII, 2010, in press, Ramanna,
S., Meghdadi, A., Measuring resemblances between swarm behaviours: A perceptual tolerance
near set approach, Fundamenta Informaticae 95(4), 2009, 533-552.
 0-3

Table of Contents
1 Cantor, Fuzzy, Near, and Rough Sets in Image Analysis
James F. Peters and Sankar K. Pal . . . . . . . . . . . . . . . . . . . . . 1-1

2 Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images

Pradipta Maji and Sankar K. Pal . . . . . . . . . . . . . . . . . . . . . . . 2-1

3 Image Thresholding using Generalized Rough Sets

Debashis Sen and Sankar K. Pal . . . . . . . . . . . . . . . . . . . . . . . 3-1

4 Mathematical Morphology and Rough Sets

Homa Fashandi and James F. Peters . . . . . . . . . . . . . . . . . . . . 4-1

5 Rough Hybrid Scheme: An application of breast cancer imaging

Aboul Ella Hassanien, Hameed Al-Qaheri, Ajith Abraham . . . . . . . . . 5-1

6 Applications of Fuzzy Rule-based Systems in Medical Image Understanding

Wojciech Tarnawski, Gerald Schaefer, Tomoharu Nakashima and Lukasz
Miroslaw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-1

7 Near Set Evaluation And Recognition (NEAR) System

Christopher Henry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-1

8 Perceptual Systems Approach to Measuring Image Resemblance

Amir H. Meghdadi and James F. Peters . . . . . . . . . . . . . . . . . . . 8-1

9 From Tolerance Near Sets to Perceptual Image Analysis

Shabnam Shahfar, Amir H. Meghdadi and James F. Peters . . . . . . . . 9-1

10 Image Segmentation: A Rough-set Theoretic Approach

Milind M. Mushrif and Ajoy K. Ray . . . . . . . . . . . . . . . . . . . . 10-1

11 Rough Fuzzy Measures in Image Segmentation and Analysis

Dariusz Malyszko and Jaroslaw Stepaniuk . . . . . . . . . . . . . . . . . 11-1

12 Discovering Image Similarities. Tolerance Near Set Approach

Sheela Ramanna . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-1

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I-1
 1
Cantor, Fuzzy, Near, and Rough Sets
in Image Analysis

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–1

1.2 Cantor Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–2
1.3 Near Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–2
Near Sets and Rough Sets • Basic Near Set Approach •

Near Sets, Psychophysics and Merleau-Ponty • Visual

Acuity Tolerance • Sets of Similar Images • Tolerance
Near Sets • Near Sets in Image Analysis
James F. Peters
Computational Intelligence Laboratory,
1.4 Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–8
Notion of a Fuzzy Set • Near Fuzzy Sets • Fuzzy Sets in
Electrical & Computer Engineering, Rm.
Image Analysis
E2-390 EITC Bldg., 75A Chancellor’s Circle,
University of Manitoba, Winnipeg R3T 5V6
1.5 Rough Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–9
Manitoba Canada
Sample Non-Rough Set • Sample Rough Set • Rough
Sets in Image Analysis
Sankar K. Pal 1.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–11
Machine Intelligence Unit, Indian Statistical Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–11
Institute,Kolkata, 700 108, India Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–12

1.1 Introduction
The chapters in this book consider how one might utilize fuzzy sets, near sets, and rough sets, taken
separately or taken together in hybridizations, in solving a variety of problems in image analysis. A
brief consideration of Cantor sets (Cantor, 1883, 1932) provides a backdrop for an understanding
of several recent types of sets useful in image analysis. Fuzzy, near and rough sets provide a wide
spectrum of practical solutions to solving image analysis problems such as image understanding,
image pattern recognition, image retrieval and image correspondence, mathematical morphology,
perceptual tolerance relations in image analysis and segmentation evaluation. Fuzzy sets result from
the introduction of a membership function that generalizes the traditional characteristic function.
The notion of a fuzzy set was introduced by L. Zadeh in 1965 (Zadeh, 1965). Sixteen years later,
rough sets were introduced by Z. Pawlak in 1981 (Pawlak, 1981a). A set is considered rough
whenever the boundary between its lower and upper approximation is non-empty. Of the three forms
of sets, near sets are newest, introduced in 2007 by J.F. Peters in a perception-based approach to the
study of the nearness of observable objects in a physical continuum (Peters and Henry, 2006; Peters,
2007c,a; Peters, Skowron, and Stepaniuk, 2007; Henry and Peters, 2009b; Peters and Wasilewski,
2009; Peters, 2010).
This chapter highlights a context for three forms of sets that are now part of the computational
intelligence spectrum of tools useful in image analysis and pattern recognition. The principal con-

1–1
1–2 Rough Fuzzy Image Analysis

tribution of this chapter is an overview of the high utility of fuzzy sets, near sets and rough sets
with the emphasis on how these sets can be used in image analysis, especially in classifying parts
of digital images presented in this book.

1.2 Cantor Set

To establish a context for the various sets utilized in this book, this section briefly presents the notion
of a Cantor set. From the definition of a Cantor set, it is pointed out that fuzzy sets, near sets and
rough sets are special forms of Cantor sets. In addition, this chapter points to links between the three
types of sets that are part of the computational intelligence spectrum. Probe functions in near set
theory provide a link between fuzzy sets and near sets, since every fuzzy membership function is a
particular form of probe function. Probe functions are real-valued functions introduced by M. Pavel
in 1993 as part of a study of image registration and a topology of images (Pavel, 1993). Z. Pawlak
originally thought of a rough set as a new form of fuzzy set (Pawlak, 1981a). It has been shown
that every rough set is a near set (this is Theorem 4.8 in (Peters, 2007b)) but not every near set is a
rough set. For this reason, near sets are considered a generalization of rough sets. The contribution
of this chapter is an overview of the links between fuzzy sets, near sets and rough sets as well as the
relation between these sets and the original notion of a set introduced by Cantor in 1883 (Cantor,
1883).
By a ‘manifold’ or ‘set’ I understand any multiplicity,
which can be thought of as one, i.e., any aggregate
[inbegri f f ] of determinate elements which,
can be united into a whole by some law.
–Foundations of a General Theory of Manifolds,
–G. Cantor, 1883.

. . . A set is formed by the grouping together

of single objects into a whole.
–Set Theory
–F. Hausdorff, 1914.

In this mature interpretation of the notion of a set, G. Cantor points to a property or law that de-
termines elementhood in a set and “unites [the elements] into a whole” (Cantor, 1883), elaborated
in (Cantor, 1932), and commented on in Lavine (1994). In 1851, Bolzano (Bolzano, 1959) writes
that “an aggregate so conceived that is indifferent to the arrangement of its members I call a set”.
At that time, the idea that a set could contain just one element or no elements (null set) was not con-
templated. This is important in the current conception of a near set, since such a set must contain
pairs of perceptual objects with similar descriptions and such a set is never null. That is, a set is a
perceptual near set if, and only if it is never empty and it contains pairs of perceived objects that
have descriptions that are within some tolerance of each other (see Def. 2).

1.3 Near Sets

How Near
How near to the bark of a tree are drifting snowflakes,
swirling gently round, down from winter skies?
How near to the ground are icicles,
Cantor, Fuzzy, Near, and Rough Sets in Image Analysis 1–3

slowly forming on window ledges?

–Fragment of a Philosophical Poem.
–Z. Pawlak & J.F. Peters, 2002.

The basic idea in the near set approach to object recognition

is to compare object descriptions. Sets of objects X,Y
are considered near each other if the sets contain objects
with at least partial matching descriptions.
–Near sets. General theory about nearness of objects,
–J.F. Peters, 2007.

TABLE 1.1 Nomenclature

Symbol Interpretation
O, X,Y Set of perceptual objects, X,Y ⊆ O, A ⊂ X, x ∈ X, y ∈ Y ,
F, B Sets of probe functions, B ⊆ F, φi ∈ B,
φi (x) φi : X → ℜ, ith probe function representing feature of x,
φ B (x) (φ1 (x), φ2 (x), . . . , φi (x), . . . , φk (x)),description of x of length k,
ε ε ∈ ℜ (reals) such that ε ≥ 0,
1
 · 2 = (∑ki=1 (·i )2 ) 2 , L2 (Euclidean) norm,
∼
=B,ε {(x, y) ∈ O × O :  φ (x) − φ (y) 2 ≤ ε }, tolerance relation,
∼
=B shorthand for ∼ =B,ε ,
A⊂ ∼ =B,ε ∀x, y ∈ A, x ∼ =B,ε y (i.e., A∼ ∼
=B,ε is a preclass in =B,ε ),
C∼
=B,ε tolerance class, maximal preclass of =B,ε , ∼
X B,ε Y X resembles (is near) Y ⇐⇒ X ∼ =B,ε Y .

Set Theory Law 1 Near Sets

Near sets contain elements with similar descriptions.

Near sets are disjoint sets that resemble each other (Henry and Peters, 2010). Resemblance
between disjoint sets occurs whenever there are observable similarities between the objects in the
sets. Similarity is determined by comparing lists of object feature values. Each list of feature values
defines an object’s description. Comparison of object descriptions provides a basis for determining
the extent that disjoint sets resemble each other. Objects that are perceived as similar based on
their descriptions are grouped together. These groups of similar objects can provide information
and reveal patterns about objects of interest in the disjoint sets. For example, collections of digital
images viewed as disjoint sets of points provide a rich hunting ground for near sets. For example,
near sets can be found in the favite pentagona coral fragment in Fig. 1.1a from coral reef near Japan.
If we consider the greyscale level, the sets X,Y in Fig. 1.1b are near sets, since there are many pixels
in X with grey levels that are very similar to pixels in Y .

1.3.1 Near Sets and Rough Sets

Near sets are a generalization of rough sets. It has been shown that every rough set is, in fact, a
near set but not every near set is a rough set Peters (2007b). Near set theory originated from an
1–4 Rough Fuzzy Image Analysis

(1.1a) favite coral (1.1b) near sets

FIGURE 1.1: Sample Near Sets

interest in comparing similarities between digital images. Unlike rough sets, the near set approach
does not require set approximation Peters and Wasilewski (2009). Simple examples of near sets can
sometimes be found in tolerance classes in pairs of image coverings, if, for instance, a subimage of
a class in one image has a description that is similar to the description of a subimage in a class in
the second image. In general, near sets are discovered by discerning objects–either within a single
set or across sets–with descriptions that are similar.
From the beginning, the near set approach to perception has had direct links to rough sets in its
approach to the perception of objects (Pawlak, 1981a; Orłowska, 1982) and the classification of ob-
jects (Pawlak, 1981a; Pawlak and Skowron, 2007c,b,a). This is evident in the early work on nearness
of objects and the extension of the approximation space model (see, e.g., (Peters and Henry, 2006;
Peters et al., 2007)). Unlike the focus on the approximation boundary of a set, the study of near sets
focuses on the discovery of affinities between perceptual granules such as digital images viewed as
sets of points. In the context of near sets, the term affinity means close relationship between per-
ceptual granules (particularly images) based on common description. Affinities are discovered by
comparing the descriptions of perceptual granules, e.g., descriptions of objects contained in classes
found in coverings defined by the tolerance relation ∼ =F,ε .

1.3.2 Basic Near Set Approach

Near set theory provides methods that can be used to extract resemblance information from objects
contained in disjoint sets, i.e., it provides a formal basis for the observation, comparison, and clas-
sification of objects. The discovery of near sets begins with choosing the appropriate method to
describe observed objects. This is accomplished by the selection of probe functions representing
observable object features. A basic model for a probe function was introduced by M. Pavel (Pavel,
1993) in the text of image registration and image classification. In near set theory, a probe function is
a mapping from an object to a real number representing an observable feature value (Peters, 2007a).
For example, when comparing fruit such as apples, the redness of an apple (observed object) can be
described by a probe function representing colour, and the output of the probe function is a number
representing the degree of redness. Probe functions provide a basis for describing and discerning
affinities between objects as well as between groups of similar objects (Peters and Ramanna, 2009).
Objects that have, in some degree, affinities are considered near each other. Similarly, groups of
objects (i.e. sets) that have, in some degree, affinities are also considered near each other.

1.3.3 Near Sets, Psychophysics and Merleau-Ponty

Cantor, Fuzzy, Near, and Rough Sets in Image Analysis 1–5

Near sets offer an ideal framework for solving problems based on human perception that arise in
areas such as image processing, computer vision as well as engineering and science problems. In
near set theory, perception is a combination of the view of perception in psychophysics (Hoogs,
Collins, Kaucic, and Mundy, 2003; Bourbakis, 2002) with a view of perception found in Merleau-
Ponty’s work (Merleau-Ponty, 1945, 1965). In the context of psychophysics, perception of an object
(i.e., in effect, our knowledge about an object) depends on sense inputs that are the source of signal
values (stimularions) in the cortex of the brain. In this view of perception, the transmissions of
sensory inputs to cortex cells senses are likened to probe functions defined in terms of mappings
of sets of sensed objects to sets of real-values representing signal values (the magnitude of each
cortex signal value represents a sensation) that are a source of object feature values assimilated by
the mind.
Perception in animals is modelled as a mapping from sensory cells to brain cells. For example,
visual perception is modelled as a mapping from stimulated retina sensory cells to visual cortex cells
(see Fig. 1.2). Such mappings are called probe functions. A probe measures observable physical
characteristics of objects in our environment. In other words, a probe function provides a basis
for what is commonly called feature extraction (Guyon, Gunn, Nikravesh, and Zadeh, 2006). The
sensed physical characteristics of an object are identified with object features. The term feature
is used in S. Watanabe’s sense of the word (Watanabe, 1985), i.e., a feature corresponds to an
observable property of physical objects. Each feature has a 1-to-many relationship to real-valued
functions called probe functions representing the feature. For each feature (such as colour) one or
more probe functions can be introduced to represent the feature (such as grayscale, or RGB values).
Objects and sets of probe functions form the basis of near set theory and are sometimes referred to
as perceptual objects due to the focus on assigning values to perceived object features.
Axiom 1 An object is perceivable if, and only if the object is describable.
In Merleau-Ponty’s view (Merleau-Ponty, 1945, 1965), an object is perceived to the extent that it
can be described. In other words, object description goes hand-in-hand with object perception. It
is our mind that identifies relationships between object descriptions to form perceptions of sensed
objects. It is also the case that near set theory has been proven to be quite successful in finding
solutions to perceptual problems such as measuring image correspondence and segmentation eval-
uation. The notion of a sensation in Poincaré (Poincaré, 1902) and a physical model for a probe

FIGURE 1.2: Sample Visual Perception

function from near set theory (Peters and Wasilewski, 2009; Peters, 2010) is implicitly explained by
Zeeman (Zeeman, 1962) in terms of visual perception. That is, ‘seeing’ consists of mappings from
1–6 Rough Fuzzy Image Analysis

sense inputs from sensory units in the retina of the eye to cortex cells of the brain stimulated by sense
inputs. A sense input can be represented by a number representing the intensity of the light from the
visual field (i.e., everything in the physical world that causes light to fall on the retina.) impacting
on the retina. The intensity of light from the visual field will determine the level of stimulation of a
cortex cell from retina sensory input. Over time, varying cortex cell stimulation has the appearance
of an electrical signal. The magnitude of cortex cell stimulation is a real-value. The combination
of an activated sensory cell in the retina and resulting retina-originated impulses sent to cortex cells
(visual stimulation) is likened to what Poincaré calls a sensation in his essay on separate sets of
similar sensations leading to a perception of a physical continuum (Poincaré, 1902). This model for
a sensation underlies what is known as a probe function in near set theory (Peters, 2007b; Peters
and Wasilewski, 2009).

DEFINITION 1.1 Visual Probe Function

Let O = {perceptual objects}. A perceptual object is something in the visual field that is a source
of reflected light. Let ℜ denote the set of reals. Then a probe φ is a mapping φ : X → ℜ. For
x ∈ X, φ (x) denotes an amplitude in a visual perception (see, e.g., Fig. 1.2).

In effect, a probe function value φ (x) measures the strength of a feature value extracted from
each sensation. In Poincaré, sets of sensations are grouped together because they are, in some sense,
similar within a specified distance, i.e., tolerance. Implicit in this idea in Poincaré is the perceived
feature value of a particular sensation that makes it possible for us to measure the closeness of an
individual senation to other sensations.
A human sensation modelled as a probe measures observable physical characteristics of objects in
our environment. The sensed physical characteristics of an object are identified with object features.
In Merleau-Ponty’s view, an object is perceived to the extent that it can be described (Merleau-Ponty,
1945, 1965). In other words, object description goes hand-in-hand with object perception. It is our
mind that identifies relationships between object descriptions to form perceptions of sensed objects.
It is also the case that near set theory has been proven to be quite successful in finding solutions to
perceptual problems such as measuring image correspondence and segmentation evaluation.
Axiom 2 Formulate object description to achieve object perception.
In a more recent interpretation of the notion of a near set, the nearness of sets is considered in
the context of perceptual systems (Peters and Wasilewski, 2009). Poincaré’s idea of perception of
objects such as digital images in a physical continuum can be represented by means of perceptual
systems, which is akin to but not the same as what has been called a perceptual information sys-
tem (Peters and Wasilewski, 2009; Peters, 2010). A perceptual system is a pair O, F where O is
a non-empty set of perceptual objects and F is a non-empty, countable set of probe functions (see
Def. 1).
Definition 1 Perceptual System (Peters, 2010)
A perceptual system O, F consists of a sample space O containing a finite, non-empty set of
sensed sample objects and a non-empty, countable set F containing probe functions representing
object features.
The perception of physical objects and their description within a perceptual system facilitates pattern
recognition and the discovery of sets of similar objects. In the near set approach to image analysis,
one starts by identifying a perceptual system and the defining a cover on the sample space with an
appropriate perceptual tolerance relation.
Method 1 Perceptual Tolerance
Cantor, Fuzzy, Near, and Rough Sets in Image Analysis 1–7

1. identify a sample space O and a set F to formulate a perceptual system O, F, and then
2. introduce a tolerance relation τε that defines a cover on O.

1.3.4 Visual Acuity Tolerance

Zeeman (Zeeman, 1962) introduces a tolerance space (X, τε ), where X is the visual field of the right
eye and ε is the least angular distance so that all points indistinguishable from x ∈ X are within ε of
x. In this case, there is an implicit perceptual system O, F, where O := X consists of points that
are sources of reflected light in the visual field and F contains probes used to extract feature values
from each x ∈ O.

1.3.5 Sets of Similar Images

Consider O, F, where O consists of points representing image pixels and F contains probes used
to extract feature values from each x ∈ O. Let B ⊆ F. Then introduce tolerance relation ∼ =B,ε to
define a covers on X,Y ⊂ O. Then, in the case where X,Y resemble each other, i.e., X B,ε Y ,
then measure the degree of similarity (nearness) of X,Y (a publicly available toolset that makes it
possible to complete this example for any set of digital images is available at (Henry and Peters,
2010, 2009a)). See Table 1.1 (also, (Peters and Wasilewski, 2009; Peters, 2009b, 2010)) for details
about the bowtie notation B,ε used to denote resemblance between X and Y , i.e., X B,ε Y

(1.3a) Lena (1.3b) Lena TNS

FIGURE 1.3: Lena Tolerance Near Sets (TNS)

1.3.6 Tolerance Near Sets

In near set theory, the trivial case is excluded. That is, an element x ∈ X is not considered near itself.
In addition, the empty set is excluded from near sets, since the empty set is never something that we
perceive, i.e., a set of perceived objects is never empty. In the case where one set X is near another
set Y , this leads to the realization that there is a third set containing pairs of elements x, y ∈ X × Y
with similar descriptions. The key to an understanding of near sets is the notion of a description.
The description of each perceived object is specified a vector of feature values and each feature is
1–8 Rough Fuzzy Image Analysis

(1.4a) Photographer (1.4b) Photographer TNS

FIGURE 1.4: Photographer Tolerance Near Sets

represented by what is known as a probe function that maps an object to a real value. Since our
main interest is in detecting similarities between seemingly quite disjoint sets such as subimages in
an image or pairs of classes in coverings on a pair of images, a near set is defined in context of a
tolerance space.
Definition 2 Tolerance Near Sets (Peters, 2010)
Let O, F be a perceptual system. Put ε ∈ ℜ, B ⊂ F. Let X,Y ⊂ O denote disjoint sets with
coverings determined by a tolerance relation ∼
=B,ε . Sets X,Y are tolerance near sets if, and only if
there are preclasses A ⊂ X, B ⊂ Y such that A B,ε B.

1.3.7 Near Sets in Image Analysis

The subimages in Fig. 1.3b and Fig. 1.4b delineate tolerance classes (each with its own grey level)
subregions of the original images in Fig. 1.3a and Fig. 1.4a. The tolerance classes in these images
are dominated by (light grey), (medium grey) and (dark grey) subimages along with a few
(very dark) subimages in Fig. 1.3b and many very dark subimages in Fig. 1.4b. From Def. 2,
it can be observed that the images in Fig. 1.3a and Fig. 1.4a are examples of tolerance near sets,
i.e., ImageFig. 1.4a F,ε ImageFig. 1.3a ). Examples of the near set approach to image analysis can
be found in, e.g., (Henry and Peters, 2007, 475-482, 2008, 1-6, 2009a; Gupta and Patnaik, 2008;
Peters, 2009a,b, 2010; Peters and Wasilewski, 2009; Peters and Puzio, 2009; Hassanien, Abraham,
Peters, Schaefer, and Henry, 2009; Meghdadi, Peters, and Ramanna, 2009; Fashandi, Peters, and
Ramanna, 2009) and in a number of chapters of this book.
From set composition Law 1, near sets are Cantor sets containing one or more pairs of objects
(e.g., image patches, one from each digital image) that resemble each other as enunciated in Def. 2,
i.e., X, T ⊂ O are near sets if, and only if X F,ε Y ).

1.4 Fuzzy Sets

A fuzzy set is a class of objects with a continuum
of grades of membership.
–Fuzzy sets, Information and Control 8
–L.A. Zadeh, 1965.
Cantor, Fuzzy, Near, and Rough Sets in Image Analysis 1–9

. . . A fuzzy set is characterized by a membership

function which assigns to each object its grade of membership
(a number lying between 0 and 1) in the fuzzy set.
–A new view of system theory
–L.A. Zadeh, 20-21 April 1965.

Set Theory Law 2 Fuzzy Sets

Every element in a fuzzy set has a graded membership.

1.4.1 Notion of a Fuzzy Set

The notion of a fuzzy set was introduced by L.A. Zadeh in 1965 (Zadeh, 1965). In effect, a Cantor
set is a fuzzy set if, and only if every element of the set has a grade of membership assigned to it by
a specified membership function. Notice that a membership function φ : X → [0, 1] is a special case
of what is known as a probe function in near set theory.

1.4.2 Near Fuzzy Sets

A fuzzy set X is a near set relative to a set Y if the grade of membership of the objects in sets X,Y is
assigned to each object by the same membership function φ and there is a least one pair of objects
x, y ∈ X ×Y such that  φ (x) − φ (y) 2 ≤ ε }, i.e., the description of x is similar to the description y
within some ε .

1.4.3 Fuzzy Sets in Image Analysis

Fuzzy sets have widely used in image analysis (see, e.g., (Rosenfeld, 1979; Pal and King, 1980,
1981; Pal, 1982; Pal, King, and Hashim, 1983; Pal, 1986, 1992; Pal and Leigh, 1995; Pal and Mitra,
1996; Nachtegael and Kerre, 2001; Deng and Heijmans, 2002; Martino, Sessa, and Nobuhara, 2008;
Sussner and Valle, 2008; Hassanien et al., 2009)). In the notion of fuzzy sets, (Pal and King, 1980,
1981) defined an image of M × N dimension and L levels as an array of fuzzy singletons, each with
a value of membership function denoting the degree of having brightness or some property relative
to some brightness level l, where l = 0, 1, 2, . . . , L − 1. The literature on fuzzy image analysis is
based on the realization that the basic concepts of edge, boundary, region, relation in an image do
not lend themselves to precise definition.
From set composition Law 2, it can be observed that fuzzy sets are Cantor sets.

1.5 Rough Sets

A new approach to classification,based on information systems theory,
given in this paper. . . . This approach leads to a new formulation
of the notion of fuzzy sets (called here the rough sets).
The axioms for such sets are given, which are the same
as the axioms of topological closure and interior.
–Classification of objects by means of attributes.
–Z. Pawlak, 1981.
1–10 Rough Fuzzy Image Analysis
TABLE 1.2 Pawlak Indiscernibility Relation and Partition Symbols
Symbol Interpretation
∼B = {(x, y) ∈ X × X | f (x) = f (y) ∀ f ∈ B}, indiscernibility, cf. (Pawlak, 1981a),
x/∼B x/∼B = {y ∈ X | y ∼B x}, elementary set (class),
U/∼B U/∼B = {x/∼B | x ∈ U}, quotient set.
B∗ (X) B∗ (X) = x/∼ (lower approximation of X),
B
x/∼ ⊆X
B
B ∗ (X) B ∗ (X) = x/∼ (upper approximation of X).
B
x/∼ ∩X=0/
B

Set Theory Law 3 Rough Sets

Any non-empty set X is a rough set if, and only if the approximation boundary of X is not empty.

Rough sets were introduced by Z. Pawlak in (Pawlak, 1981a) and elaborated in (Pawlak, 1981b;
Pawlak and Skowron, 2007c,b,a). In a rough set approach to classifying sets of objects X, one
considers the size of the boundary region in the approximation of X. By contrast, in a near set
approach to classification, one does not consider the boundary region of a set. In particular, assume
that X is a non-empty set belonging to a universe U and that F is a set of features defined either by
total or partial functions. The lower approximation of X relative to B ⊆ F is denoted by B∗ (X) and
the upper approximation of X is denoted by B ∗ (X), where

B∗ (X) = x/∼ ,
B
x/∼ ⊆X
B

B ∗ (X) = x/∼ .
B
x/∼ ∩X=0/
B

The B-boundary region of an approximation of a set X is denoted by BndB (X), where

BndB (X) = B ∗ (X) \ B∗ (X) = {x | x ∈ B ∗ (X) and x ∈

/ B∗ (X)}.

Definition 3 Rough Set (Pawlak, 1981a)

A non-empty, finite set X is a rough set if, and only if |B ∗ (X) − B∗ (X)| = 0.

A set X is roughly classified whenever BndB (X) is not empty. In other words, X is a rough set
whenever the boundary region BndB (X) = 0. / In sum, a rough set is a Cantor set if, and only if its
approximation boundary is non-empty. It should also be noted that rough sets differ from near sets,
since near sets are defined without reference to an approximation boundary region. This means, for
example, with near sets the image correspondence problem can be solved without resorting to set
approximation.

Method 2 Rough Set Approach

1. Let (U, B) denote a sample space (universe) U and set of object features B,
2. Using relation ∼B , partition the universe U,
3. Determine the size of the boundary of a set X.

1.5.1 Sample Non-Rough Set

Let x ∈ U. x/∼B (any elementary set) is a non-rough set.
Cantor, Fuzzy, Near, and Rough Sets in Image Analysis 1–11

1.5.2 Sample Rough Set

Any set X ⊂ U where

x/∼ = X.
B
x/∼ ⊆U/∼
B B

In other words, if a set X does not equal its lower approximation, then the set X is rough, i.e.,
roughly approximated by the equivalence classes in the quotient set U/∼ .
B

1.5.3 Rough Sets in Image Analysis

The essence of our approach consists in viewing
a digitized image as a universe of a certain
information system and synthesizing an indiscernibility
relation to identify objects and measure
some of their parameters.
– Adam Mrozek and Leszek Plonka, 1993.

In terms of rough sets and image analysis, it can be observed that A. Mrózek and L. Plonka were
pioneers (Mrózek and Plonka, 1993). For example, he was one of the first to introduce a rough
set approach to image analysis and to view a digital image as a universe viewed as a set of points.
The features of pixels (points) in a digital image are a source of knowledge discovery. Using Z.
Pawlak’s indiscernibility relation, it is then a straightforward task to partition an image and to con-
sider set approximation relative to interesting objects contained in subsets of an image. This work
on digital images by A. Mrózek and L. Plonka appeared six or more years before the publication
of papers on approximate mathematical morphology by Lech Polkowski (Polkowski, 1999) (see,
also, (Polkowski, 1993; Polkowski and Skowron, 1994)) and connections between mathematical
morphology and rough sets pointed to by Isabelle Bloch (Bloch, 2000). The early work on the use
of rough sets in image analysis has been followed by a number of articles by S.K. Pal and others
(see, e.g., (Pal and Mitra, 2002; Pal, UmaShankar, and Mitra, 2005; Peters and Borkowski, 2004;
Borkowski and Peters, 2006; Borkowski, 2007; Maji and Pal, 2008; Mushrif and Ray, 2008; Sen
and Pal, 2009)).
From set composition Law 3, it can be observed that rough sets are Cantor sets.

1.6 Conclusion

In sum, fuzzy sets, near sets and rough sets are particular forms of Cantor sets. In addition, each of
these sets in the computational intelligence spectrum offer very useful approaches in image analysis,
especially in classifying objects.

Acknowledgements

This research by James Peters has been supported by the Natural Sciences and Engineering Re-
search Council of Canada (NSERC) grant 185986, Manitoba Centre of Excellence Fund (MCEF)
grant, Canadian Centre of Excellence (NCE) and Canadian Arthritis Network grant SRI-BIO-05,
and Manitoba Hydro grant T277 and that of Sankar Pal has been supported by the J.C. Bose Fel-
lowship of the Govt. of India.
1–12 Rough Fuzzy Image Analysis

Bibliography
Bloch, L. 2000. On links between mathematical morphology and rough sets. Pattern Recog-
nition 33(9):1487–1496.

Bolzano, B. 1959. Paradoxien des unendlichen (paradoxes of the infinite), trans. by d.a. steele.
London: Routledge and Kegan Paul.

Borkowski, M. 2007. 2d to 3d conversion with direct geometrical search and approximation

spaces. Ph.D. thesis, Dept. Elec. Comp. Engg. http://wren.ee.umanitoba.ca/.

Borkowski, M., and J.F. Peters. 2006. Matching 2d image segments with genetic algorithms
and approximation spaces. Transactions on Rough Sets V(LNAI 4100):63–101.

Bourbakis, N. G. 2002. Emulating human visual perception for measuring difference in images
using an spn graph approach. IEEE Transactions on Systems, Man, and Cybernetics, Part B
32(2):191–201.

Cantor, G. 1883. Über unendliche, lineare punktmannigfaltigkeiten. Mathematische Annalen

201:72–81.

———. 1932. Gesammelte abhandlungen mathematischen und philosophischen inhalts, ed.

e. zermelo. Berlin: Springer.

Deng, T.Q., and H.J.A.M. Heijmans. 2002. Grey-scale morphology based on fuzzy logic. J.
Math. Imag. Vis. 16:155–171.

Fashandi, H., J.F. Peters, and S. Ramanna. 2009. L2 norm length-based image similarity mea-
sures: Concrescence of image feature histogram distances. In Signal and image processing,
int. assoc. of science & technology for development, 178–185. Honolulu, Hawaii.

Gupta, S., and K.S. Patnaik. 2008. Enhancing performance of face recognition systems by
using near set approach for selecting facial features. J. Theoretical and Applied Information
Technology 4(5):433–441.

Guyon, I., S. Gunn, M. Nikravesh, and L.A. Zadeh. 2006. Feature extraction. foundations and
applications. Berlin: Springer.

Hassanien, A.E., A. Abraham, J.F. Peters, G. Schaefer, and C. Henry. 2009. Rough sets and
near sets in medical imaging: A review. IEEE Trans. Info. Tech. in Biomedicine 13(6):
955–968. Digital object identifier: 10.1109/TITB.2009.2017017.

Henry, C., and J.F. Peters. 2007, 475-482. Image pattern recognition using approximation
spaces and near sets. In Proc. 11th int. conf. on rough sets, fuzzy sets, data mining and
granular computing (rsfdgrc 2007), joint rough set symposium (jrs 2007), lecture notes in
artificial intelligence 4482. Heidelberg, Germany.

———. 2008, 1-6. Near set image segmentation quality index. In Geobia 2008 pixels, objects,
intelligence. geographic object based image analysis for the 21st century. University of
Calgary, Alberta.

———. 2009a. Near set evaluation and recognition (near) system. Tech. Rep., Computationa
Intelligence Laboratory, University of Manitoba. UM CI Laboratory Technical Report No.
TR-2009-015.
Cantor, Fuzzy, Near, and Rough Sets in Image Analysis 1–13

———. 2009b. Perception-based image analysis. Int. J. of Bio-Inspired Computation 2(2). in

press.

———. 2010. Near sets. Wikipedia. http://en.wikipedia.org/wiki/Near sets.

Hoogs, A., R. Collins, R. Kaucic, and J. Mundy. 2003. A common set of perceptual observ-
ables for grouping, figure-ground discrimination, and texture classification. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence 25(4):458–474.

Lavine, S. 1994. Understanding the infinite. Cambridge, MA: Harward University Press.

Maji, P., and S.K. Pal. 2008. Maximum class separability for rough-fuzzy c-means based brain
mr image segmentation. Transactions on Rough Sets IX, LNCS-5390:114134.

Martino, F.D., S. Sessa, and H. Nobuhara. 2008. Eigen fuzzy sets and image informa-
tion retrieval. In Handbook of granular computing, ed. W. Pedrycz, A. Skowron, and
V. Kreinovich, 863–872. West Sussex, England: John Wiley & Sons, Ltd.

Meghdadi, A.H., J.F. Peters, and S. Ramanna. 2009. Tolerance classes in measuring image
resemblance. Intelligent Analysis of Images & Videos, KES 2009, Part II, Knowledge-Based
and Intelligent Information and Engineering Systems, LNAI 5712 127–134. ISBN 978-3-
64-04591-2, doi 10.1007/978-3-642-04592-9 16.

Merleau-Ponty, Maurice. 1945, 1965. Phenomenology of perception. Paris and New York:
Smith, Gallimard, Paris and Routledge & Kegan Paul. Trans. by Colin Smith.

Mrózek, A., and L. Plonka. 1993. Rough sets in image analysis. Foundations of Computing
and Decision Sciences 18(3-4):268–273.

Mushrif, M., and A.K. Ray. 2008. Color image segmentation: Rough-set theoretic approach.
Pattern Recognition Letters 29(4):483493.

Nachtegael, M., and E.E. Kerre. 2001. Connections between binary, grayscale and fuzzy
mathematical morphologies. Fuzzy Sets and Systems 124:73–85.

Orłowska, E. 1982. Semantics of vague concepts. applications of rough sets. Polish Academy
of Sciences 469. In G.Dorn, P. Weingartner (Eds.), Foundations of Logic and Linguistics.
Problems and Solutions, Plenum Press, London/NY, 1985, 465-482.

Pal, S.K. 1982. A note on the quantitative measure of image enhancement through fuzziness.
IEEE Trans. Pattern Anal. Machine Intell. PAMI-4(2):204–208.

———. 1986. A measure of edge ambiguity using fuzzy sets. Pattern Recognition Letters
4(1):51–56.

———. 1992. Fuzziness, image information and scene analysis. In An introduction to fuzzy
logic applications in intelligent systems, ed. R.R. Yager and L.A. Zadeh, 147–183. Dor-
drecht: Kluwer Academic Publishers.

Pal, S.K., and R.A. King. 1980. Image enhancement with fuzzy set. Electronics Letters 16(10):
376–378.

———. 1981. Image enhancement using smoothing with fuzzy set. IEEE Trans. Syst. Man
and Cyberns. SMC-11(7):495–501.
1–14 Rough Fuzzy Image Analysis

Pal, S.K., R.A. King, and A.A. Hashim. 1983. Image description and primitive extraction
using fuzzy sets. IEEE Trans. Syst. Man and Cyberns. SMC-13(1):94–100.

Pal, S.K., and A.B. Leigh. 1995. Motion frame analysis and scene abstraction: Discrimination
ability of fuzziness measures. J. Intelligent & Fuzzy Systems 3:247–256.

Pal, S.K., and P. Mitra. 2002. Multispectral image segmentation using rough set initialized em
algorithm. IEEE Transactions on Geoscience and Remote Sensing 11:24952501.

Pal, S.K., and S. Mitra. 1996. Noisy fingerprint classification using multi layered perceptron
with fuzzy geometrical and textual features. Fuzzy Sets and Systems 80(2):121–132.

Pal, S.K., B. UmaShankar, and P. Mitra. 2005. Granular computing, rough entropy and object
extraction. Pattern Recognition Letters 26(16):401–416.

Pavel, M. 1993. Fundamentals of pattern recognition. 2nd ed. N.Y., U.S.A.: Marcel Dekker,
Inc.

Pawlak, Z. 1981a. Classification of objects by means of attributes. Polish Academy of Sciences

429.

———. 1981b. Rough sets. International J. Comp. Inform. Science 11:341–356.

Pawlak, Z., and A. Skowron. 2007a. Rough sets and boolean reasoning. Information Sciences
177:41–73.

———. 2007b. Rough sets: Some extensions. Information Sciences 177:28–40.

———. 2007c. Rudiments of rough sets. Information Sciences 177:3–27.

Peters, J.F. 2007a. Near sets. general theory about nearness of objects. Applied Mathematical
Sciences 1(53):2609–2029.

———. 2007b. Near sets. general theory about nearness of objects. Applied Mathematical
Sciences 1(53):2609–2029.

———. 2007c. Near sets. special theory about nearness of objects. Fundamenta Informaticae
75(1-4):407–433.

———. 2009a. Discovering affinities between perceptual granules: L2 norm-based tolerance

near preclass approach. In Man-machine interactions, advances in intelligent & soft com-
puting 59, 43–55. The Beskids, Kocierz Pass, Poland.

———. 2009b. Tolerance near sets and image correspondence. Int. J. of Bio-Inspired Com-
putation 1(4):239–445.

———. 2010. Corrigenda and addenda: Tolerance near sets and image correspondence. Int.
J. Bio-Inspired Computation 2(5). in press.

Peters, J.F., and M. Borkowski. 2004. K-means indiscernibility relation over pixels. In Lecture
notes in computer science 3066, ed. S. Tsumoto, R. Slowinski, K. Komorowski, and J.W.
Gryzmala-Busse, 580–585. Berlin: Springer. Doi 10.1007/b97961.

Peters, J.F., and C. Henry. 2006. Reinforcement learning with approximation spaces. Funda-
menta Informaticae 71:323–349.
Cantor, Fuzzy, Near, and Rough Sets in Image Analysis 1–15

Peters, J.F., and L. Puzio. 2009. Image analysis with anisotropic wavelet-based nearness mea-
sures. International Journal of Computational Intelligence Systems 3(2):1–17.

Peters, J.F., and S. Ramanna. 2009. Affinities between perceptual granules: Foundations and
perspectives. In Human-centric information processing through granular modelling sci 182,
ed. A. Bargiela and W. Pedrycz, 49–66. Berlin: Springer-Verlag.

Peters, J.F., A. Skowron, and J. Stepaniuk. 2007. Nearness of objects: Extension of approxi-
mation space model. Fundamenta Informaticae 79(3-4):497–512.

Peters, J.F., and P. Wasilewski. 2009. Foundations of near sets. Information Sciences. An Inter-
national Journal 179:3091–3109. Digital object identifier: doi:10.1016/j.ins.2009.04.018.
Poincaré, H. 1902. La science et l’hypothèse. Paris: Ernerst Flammarion. Later ed,,̇ Champs
sciences, Flammarion, 1968 & Science and Hypothesis, trans. by J. Larmor, Walter Scott
Publishing, London, 1905.

Polkowski, L. 1993. Mathematical morphology of rough sets. Bull. Polish Acad. Ser. Sci.Math,
Warsaw: Polish Academy of Sciences.

———. 1999. Approximate mathematical morphology. rough set approach. Rough and Fuzzy
Sets in Soft Computing, Berlin: Springer - Verlag.

Polkowski, L., and A. Skowron. 1994. Analytical morphology: Mathematical morphology of

decision tables. Fundamenta Informaticae 27:255–271.

Rosenfeld, A. 1979. Fuzzy digital topology. Inform. Contrl 40(1):76–87.

Sen, D., and S. K. Pal. 2009. Histogram thrsholding using fuzzy and rough means of associa-
tion error. IEEE Trans. Image Processing 18(4):879–888.

Sussner, P., and M.E. Valle. 2008. Fuzzy associative memories and their relationship to math-
ematical morphology. In Handbook of granular computing, ed. W. Pedrycz, A. Skowron,
and V. Kreinovich, 733–753. West Sussex, England: John Wiley & Sons, Ltd.

Watanabe, S. 1985. Pattern recognition: Human and mechanical. John Wiley & Sons: Chich-
ester.

Zadeh, L.A. 1965. Fuzzy sets. Information and Control 201:72–81.

Zeeman, E.C. 1962. The topology of the brain and the visual perception. New Jersey: Prentice
Hall. In K.M. Fort, Ed., Topology of 3-manifolds and Selected Topics, 240-256.
 2
Rough-Fuzzy Clustering Algorithm
for Segmentation of Brain MR
Images

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–1

2.2 Fuzzy C-Means and Rough Sets . . . . . . . . . . . . . . . . . . . . 2–3
Fuzzy C-Means • Rough Sets
2.3 Rough-Fuzzy C-Means Algorithm . . . . . . . . . . . . . . . . . . 2–5
Objective Function • Cluster Prototypes • Details of
the Algorithm
2.4 Pixel Classification of Brain MR Images . . . . . . . . . . . 2–7
2.5 Segmentation of Brain MR Images . . . . . . . . . . . . . . . . . 2–9
Feature Extraction • Selection of Initial Centroids
Pradipta Maji 2.6 Experimental Results and Discussion . . . . . . . . . . . . . . 2–13
Machine Intelligence Unit, Indian Statistical Haralick’s Features Versus Proposed Features •
Institute, Kolkata, 700 108, India
Random Versus Discriminant Analysis Based
Initialization • Comparative Performance Analysis
Sankar K. Pal 2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–18
Machine Intelligence Unit, Indian Statistical Acknowledgments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–18
Institute, Kolkata, 700 108, India Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2–19

2.1 Introduction
Segmentation is a process of partitioning an image space into some non-overlapping mean-
ingful homogeneous regions. The success of an image analysis system depends on the quality
of segmentation (Rosenfeld and Kak, 1982). A segmentation method is supposed to find
those sets that correspond to distinct anatomical structures or regions of interest in the
image. In the analysis of medical images for computer-aided diagnosis and therapy, seg-
mentation is often required as a preliminary stage. However, medical image segmentation is
a complex and challenging task due to intrinsic nature of the images. The brain has a par-
ticularly complicated structure and its precise segmentation is very important for detecting
tumors, edema, and necrotic tissues, in order to prescribe appropriate therapy (Suetens,
2002).
In medical imaging technology, a number of complementary diagnostic tools such as x-
ray computer tomography (CT), magnetic resonance imaging (MRI), and position emission
tomography (PET) are available. MRI is an important diagnostic imaging technique for
the early detection of abnormal changes in tissues and organs. Its unique advantage over
other modalities is that it can provide multispectral images of tissues with a variety of

2–1
2–2 Rough Fuzzy Image Analysis

contrasts based on the three MR parameters ρ, T1, and T2. Therefore, majority of research
in medical image segmentation concerns MR images (Suetens, 2002).
Conventionally, the brain MR images are interpreted visually and qualitatively by ra-
diologists. Advanced research requires quantitative information, such as the size of the
brain ventricles after a traumatic brain injury or the relative volume of ventricles to brain.
Fully automatic methods sometimes fail, producing incorrect results and requiring the in-
tervention of a human operator. This is often true due to restrictions imposed by image
acquisition, pathology and biological variation. So, it is important to have a faithful method
to measure various structures in the brain. One of such methods is the segmentation of
images to isolate objects and regions of interest.
Many image processing techniques have been proposed for MR image segmentation, most
notably thresholding (Lee, Hun, Ketter, and Unser, 1998; Maji, Kundu, and Chanda, 2008),
region-growing (Manousakes, Undrill, and Cameron, 1998), edge detection (Singleton and
Pohost, 1997), pixel classification (Pal and Pal, 1993; Rajapakse, Giedd, and Rapoport,
1997) and clustering (Bezdek, 1981; Leemput, Maes, Vandermeulen, and Suetens, 1999;
Wells III, Grimson, Kikinis, and Jolesz, 1996). Some algorithms using the neural network
approach have also been investigated in the MR image segmentation problems (Cagnoni,
Coppini, Rucci, Caramella, and Valli, 1993; Hall, Bensaid, Clarke, Velthuizen, Silbiger, and
Bezdek, 1992). One of the main problems in medical image segmentation is uncertainty.
Some of the sources of this uncertainty include imprecision in computations and vagueness
in class definitions. In this background, the possibility concept introduced by the fuzzy
set theory (Zadeh, 1965) and rough set theory (Pawlak, 1991) have gained popularity in
modeling and propagating uncertainty. Both fuzzy set and rough set provide a mathematical
framework to capture uncertainties associated with human cognition process (Dubois and
H.Prade, 1990; Maji and Pal, 2007b; Pal, Mitra, and Mitra, 2003). The segmentation of MR
images using fuzzy c-means has been reported in (Bezdek, 1981; Brandt, Bohan, Kramer,
and Fletcher, 1994; Hall et al., 1992; Li, Goldgof, and Hall, 1993; Xiao, Ho, and Hassanien,
2008). Image segmentation using rough sets has also been done (Mushrif and Ray, 2008;
Pal and Mitra, 2002; Widz, Revett, and Slezak, 2005a,b; Widz and Slezak, 2007; Hassanien,
2007).
In this chapter, a hybrid algorithm called rough-fuzzy c-means (RFCM) algorithm is pre-
sented for segmentation of brain MR images. Details of this algorithm have been reported
in (Maji and Pal, 2007a,c). The RFCM algorithm is based on both rough sets and fuzzy
sets. While the membership function of fuzzy sets enables efficient handling of overlapping
partitions, the concept of lower and upper approximations of rough sets deals with uncer-
tainty, vagueness, and incompleteness in class definition. Each partition is represented by
a cluster prototype (centroid), a crisp lower approximation, and a fuzzy boundary. The
lower approximation influences the fuzziness of the final partition. The cluster prototype
(centroid) depends on the weighting average of the crisp lower approximation and fuzzy
boundary. However, an important issue of the RFCM based brain MR image segmentation
method is how to select initial prototypes of different classes or categories. The concept of
discriminant analysis, based on the maximization of class separability, is used to circumvent
the initialization and local minima problems of the RFCM, and enables efficient segmenta-
tion of brain MR images (Maji and Pal, 2008). The effectiveness of the RFCM algorithm,
along with a comparison with other c-means algorithms, is demonstrated on a set of brain
MR images using some standard validity indices.
The chapter is organized as follows: Section 2.2 briefly introduces the necessary notions
of fuzzy c-means and rough sets. In Section 2.3, the RFCM algorithm is described based on
the theory of rough sets and fuzzy c-means. While Section 2.4 deals with pixel classification
problem, Section 2.5 gives an overview of the feature extraction techniques employed in seg-
Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images 2–3

mentation of brain MR images along with the initialization method of c-means algorithm
based on the maximization of class separability. Implementation details, experimental re-
sults, and a comparison among different c-means are presented in Section 2.6. Concluding
remarks are given in Section 2.7.

2.2 Fuzzy C-Means and Rough Sets

This section presents the basic notions of fuzzy c-means and rough sets. The rough-fuzzy
c-means (RFCM) algorithm is developed based on these algorithms.

2.2.1 Fuzzy C-Means

Let X = {x1 , · · · , xj , · · · , xn } be the set of n objects and V = {v1 , · · · , vi , · · · , vc } be the
set of c centroids, where xj ∈ ℜm , vi ∈ ℜm , and vi ∈ X. The fuzzy c-means provides a
fuzzification of the hard c-means (Bezdek, 1981; Dunn, 1974). It partitions X into c clusters
by minimizing the objective function
X
n X
c
J= (µij )ḿ ||xj − vi ||2 (2.1)
j=1 i=1

where 1 ≤ ḿ < ∞ is the fuzzification factor, vi is the ith centroid corresponding to cluster
βi , µij ∈ [0, 1] is the fuzzy membership of the pattern xj to cluster βi , and ||.|| is the distance
norm, such that
1X X
n n
vi = (µij )ḿ xj ; where ni = (µij )ḿ (2.2)
ni j=1 j=1

and
X
c
dij ḿ−1
)−1 ; where d2ij = ||xj − vi ||2
2
µij = ( ( ) (2.3)
dkj
k=1

subject to
X
c X
n
µij = 1, ∀j, and 0 < µij < n, ∀i.
i=1 j=1

The process begins by randomly choosing c objects as the centroids (means) of the c
clusters. The memberships are calculated based on the relative distance of the object xj to
the centroids by Equation 2.3. After computing memberships of all the objects, the new
centroids of the clusters are calculated as per Equation 2.2. The process stops when the
centroids stabilize. That is, the centroids from the previous iteration are identical to those
generated in the current iteration. The basic steps are outlined as follows:
1. Assign initial means vi , i = 1, 2, · · · , c. Choose values for ḿ and threshold ǫ. Set
iteration counter t = 1.
2. Compute memberships µij by Equation 2.3 for c clusters and n objects.
3. Update mean (centroid) vi by Equation 2.2.
4. Repeat steps 2 to 4, by incrementing t, until |µij (t) − µij (t − 1)| > ǫ.
Although fuzzy c-means is a very useful clustering method, the resulting memberships
values do not always correspond well to the degrees of belonging of the data, and it may
2–4 Rough Fuzzy Image Analysis

be inaccurate in a noisy environment (Krishnapuram and Keller, 1993, 1996). In real

data analysis, noise and outliers are unavoidable. Hence, to reduce this weakness of fuzzy
c-means, and to produce memberships that have a good explanation of the degrees of
belonging for the data, Krishnapuram and Keller (Krishnapuram and Keller, 1993, 1996)
proposed a possibilistic approach to clustering which used a possibilistic type of membership
function to describe the degree of belonging. However, the possibilistic c-means sometimes
generates coincident clusters (Barni, Cappellini, and Mecocci, 1996). Recently, the use of
both fuzzy (probabilistic) and possibilistic memberships in a clustering has been proposed
in (Pal, Pal, Keller, and Bezdek, 2005).

2.2.2 Rough Sets

The theory of rough sets begins with the notion of an approximation space, which is a pair
< U, R >, where U be a non-empty set (the universe of discourse) and R an equivalence
relation on U , i.e., R is reflexive, symmetric, and transitive. The relation R decomposes
the set U into disjoint classes in such a way that two elements x, y are in the same class iff
(x, y) ∈ R. Let denote by U/R the quotient set of U by the relation R, and

U/R = {X1 , X2 , · · · , Xm }

where Xi is an equivalence class of R, i = 1, 2, · · · , m. If two elements x, y ∈ U belong

to the same equivalence class Xi ∈ U/R, then x and y are called indistinguishable. The
equivalence classes of R and the empty set ∅ are the elementary sets in the approximation
space < U, R >. Given an arbitrary set X ∈ 2U , in general it may not be possible to
describe X precisely in < U, R >. One may characterize X by a pair of lower and upper
approximations defined as follows (Pawlak, 1991):
[ [
R(X) = Xi ; R(X) = Xi
Xi ⊆X Xi ∩X6=∅

That is, the lower approximation R(X) is the union of all the elementary sets which are
subsets of X, and the upper approximation R(X) is the union of all the elementary sets
which have a non-empty intersection with X. The interval [R(X), R(X)] is the representa-
tion of an ordinary set X in the approximation space < U, R > or simply called the rough
set of X. The lower (resp., upper) approximation R(X) (resp., R(X)) is interpreted as the
collection of those elements of U that definitely (resp., possibly) belong to X. Further,

• a set X ∈ 2U is said to be definable (or exact) in < U, R > iff R(X) = R(X).
• for any X, Y ∈ 2U , X is said to be roughly included in Y , denoted by X ⊂Y
˜ , iff
R(X) ⊆ R(Y ) and R(X) ⊆ R(Y ).
• X and Y is said to be roughly equal, denoted by X ≃R Y , in < U, R > iff
R(X) = R(Y ) and R(X) = R(Y ).

In (Pawlak, 1991), Pawlak discusses two numerical characterizations of imprecision of a

subset X in the approximation space < U, R >: accuracy and roughness. Accuracy of X,
denoted by αR (X), is simply the ratio of the number of objects in its lower approximation
to that in its upper approximation; namely

|R(X)|
αR (X) =
|R(X)|
Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images 2–5

The roughness of X, denoted by ρR (X), is defined by subtracting the accuracy from 1:

|R(X)|
ρR (X) = 1 − αR (X) = 1 −
|R(X)|
Note that the lower the roughness of a subset, the better is its approximation. Further,
the following observations are easily obtained:
1. As R(X) ⊆ X ⊆ R(X), 0 ≤ ρR (X) ≤ 1.
2. By convention, when X = ∅, R(X) = R(X) = ∅ and ρR (X) = 0.
3. ρR (X) = 0 if and only if X is definable in < U, R >.

2.3 Rough-Fuzzy C-Means Algorithm

Incorporating both fuzzy and rough sets, next a newly introduced c-means algorithm,
termed as rough-fuzzy c-means (RFCM) (Maji and Pal, 2007a,c), is described. The RFCM
algorithm adds the concept of fuzzy membership of fuzzy sets, and lower and upper approx-
imations of rough sets into c-means algorithm. While the membership of fuzzy sets enables
efficient handling of overlapping partitions, the rough sets deal with uncertainty, vagueness,
and incompleteness in class definition.

2.3.1 Objective Function

Let A(βi ) and A(βi ) be the lower and upper approximations of cluster βi , and B(βi ) =
{A(βi ) − A(βi )} denote the boundary region of cluster βi . The RFCM partitions a set of n
objects into c clusters by minimizing the objective function

 w × A1 + w̃ × B1 if A(βi ) 6= ∅, B(βi ) 6= ∅
JRF = A1 if A(βi ) 6= ∅, B(βi ) = ∅ (2.4)

B1 if A(βi ) = ∅, B(βi ) 6= ∅

X
c X X
c X
A1 = ||xj − vi ||2 B1 = (µij )ḿ ||xj − vi ||2
i=1 xj ∈A(βi ) i=1 xj ∈B(βi )

vi represents the centroid of the ith cluster βi , the parameter w and w̃ correspond to the
relative importance of lower bound and boundary region, and w + w̃ = 1. Note that, µij
has the same meaning of membership as that in fuzzy c-means.
In the RFCM, each cluster is represented by a centroid, a crisp lower approximation,
and a fuzzy boundary (Fig. 2.1). The lower approximation influences the fuzziness of final
partition. According to the definitions of lower approximations and boundary of rough sets,
if an object xj ∈ A(βi ), then xj ∈/ A(βk ), ∀k 6= i, and xj ∈
/ B(βi ), ∀i. That is, the object
xj is contained in βi definitely. Thus, the weights of the objects in lower approximation of
a cluster should be independent of other centroids and clusters, and should not be coupled
with their similarity with respect to other centroids. Also, the objects in lower approxima-
tion of a cluster should have similar influence on the corresponding centroid and cluster.
Whereas, if xj ∈ B(βi ), then the object xj possibly belongs to βi and potentially belongs
to another cluster. Hence, the objects in boundary regions should have different influence
on the centroids and clusters. So, in the RFCM, the membership values of objects in lower
approximation are µij = 1, while those in boundary region are the same as fuzzy c-means
(Equation 2.3). In other word, the RFCM algorithm first partitions the data into two classes
- lower approximation and boundary. Only the objects in boundary are fuzzified.
2–6 Rough Fuzzy Image Analysis

Cluster βi

Crisp Lower Approximation

A( βi ) with µ ij = 1

Fuzzy Boundary B( βi )
with µ ij [0, 1]

FIGURE 2.1 RFCM: cluster βi is represented by crisp lower bound and fuzzy boundary

2.3.2 Cluster Prototypes

The new centroid is calculated based on the weighting average of the crisp lower approxi-
mation and fuzzy boundary. Computation of the centroid is modified to include the effects
of both fuzzy memberships and lower and upper bounds. The modified centroid calculation
for the RFCM is obtained by solving Equation 2.4 with respect to vi :

 w × C1 + w̃ × D1 if A(βi ) 6= ∅, B(βi ) 6= ∅
viRF = C1 if A(βi ) 6= ∅, B(βi ) = ∅ (2.5)

D1 if A(βi ) = ∅, B(βi ) 6= ∅

1 X
C1 = xj ; where |A(βi )| represents the cardinality of A(βi )
|A(βi )|
xj ∈A(βi )

1 X X
and D1 = (µij )ḿ xj ; where ni = (µij )ḿ
ni
xj ∈B(βi ) xj ∈B(βi )

Thus, the cluster prototypes (centroids) depend on the parameters w and w̃, and fuzzifica-
tion factor ḿ rule their relative influence. The correlated influence of these parameters and
fuzzification factor, makes it somewhat difficult to determine their optimal values. Since
the objects lying in lower approximation definitely belong to a cluster, they are assigned
a higher weight w compared to w̃ of the objects lying in boundary region. Hence, for the
RFCM, the values are given by 0 < w̃ < w < 1.
From the above discussions, the following properties of the RFCM algorithm can be
derived.
S
1. A(βi ) = U , U be the set of objects of concern.
2. A(βi ) ∩ A(βk ) = ∅, ∀i 6= k.
3. A(βi ) ∩ B(βi ) = ∅, ∀i.
4. ∃i, k, B(βi ) ∩ B(βk ) 6= ∅.
5. µij = 1, ∀xj ∈ A(βi ).
6. µij ∈ [0, 1], ∀xj ∈ B(βi ).
Let us briefly comment on some properties of the RFCM. The property 2 says that if an
object xj ∈ A(βi ) ⇒ xj ∈
/ A(βk ), ∀k 6= i. That is, the object xj is contained in βi definitely.
The property 3 establishes the fact that if xj ∈ A(βi ) ⇒ xj ∈ / B(βi ), - that is, an object
may not be in both lower and boundary region of a cluster βi . The property 4 says that
Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images 2–7

if xj ∈ B(βi ) ⇒ ∃k, xj ∈ B(βk ). It means an object xj ∈ B(βi ) possibly belongs to βi

and potentially belongs to other cluster. The properties 5 and 6 are of great importance
in computing the objective function JRF and the cluster prototype v RF . They say that
the membership values of the objects in lower approximation are µij = 1, while those in
boundary region are the same as fuzzy c-means. That is, each cluster βi consists of a crisp
lower approximation A(βi ) and a fuzzy boundary B(βi ).

2.3.3 Details of the Algorithm

Approximate optimization of JRF (Equation 2.4) by the RFCM is based on Picard iteration
through Equations 2.3 and 2.5. This type of iteration is called alternating optimization.
The process starts by randomly choosing c objects as the centroids of the c clusters. The
fuzzy memberships of all objects are calculated using Equation 2.3.
Let µi = (µi1 , · · · , µij , · · · , µin ) represent the fuzzy cluster βi associated with the centroid
vi . After computing µij for c clusters and n objects, the values of µij for each object
xj are sorted and the difference of two highest memberships of xj is compared with a
threshold value δ. Let µij and µkj be the highest and second highest memberships of xj .
If (µij − µkj ) > δ, then xj ∈ A(βi ) as well as xj ∈ A(βi ), otherwise xj ∈ A(βi ) and
xj ∈ A(βk ). After assigning each object in lower approximations or boundary regions of
different clusters based on δ, memberships µij of the objects are modified. The values of
µij are set to 1 for the objects in lower approximations, while those in boundary regions are
remain unchanged. The new centroids of the clusters are calculated as per Equation 2.5.
The main steps of the RFCM algorithm proceed as follows:

1. Assign initial centroids vi , i = 1, 2, · · · , c. Choose values for fuzzification factor

ḿ, and thresholds ǫ and δ. Set iteration counter t = 1.
2. Compute µij by Equation 2.3 for c clusters and n objects.
3. If µij and µkj be the two highest memberships of xj and (µij − µkj ) ≤ δ, then
xj ∈ A(βi ) and xj ∈ A(βk ). Furthermore, xj is not part of any lower bound.
4. Otherwise, xj ∈ A(βi ). In addition, by properties of rough sets, xj ∈ A(βi ).
5. Modify µij considering lower and boundary regions for c clusters and n objects.
6. Compute new centroid as per Equation 2.5.
7. Repeat steps 2 to 7, by incrementing t, until |µij (t) − µij (t − 1)| > ǫ.

The performance of the RFCM depends on the value of δ, which determines the class
labels of all the objects. In other word, the RFCM partitions the data set into two classes
- lower approximation and boundary, based on the value of δ. In the present work, the
following definition is used:
1X
n
δ= (µij − µkj ) (2.6)
n j=1

where n is the total number of objects, µij and µkj are the highest and second highest
memberships of xj . That is, the value of δ represents the average difference of two highest
memberships of all the objects in the data set. A good clustering procedure should make
the value of δ as high as possible. The value of δ is, therefore, data dependent.

2.4 Pixel Classification of Brain MR Images

2–8 Rough Fuzzy Image Analysis

In this section, we present the results of different c-means algorithms on pixel classification
of brain MR images, that is, the results of clustering based on only gray value of pixels.
Above 100 MR images with different sizes and 16 bit gray levels are tested with different
c-means algorithms. All the brain MR images are collected from Advanced Medicare and
Research Institute, Salt Lake, Kolkata, India. The comparative performance of different c-
means is reported with respect to DB, and Dunn index, as well as the β index (Pal, Ghosh,
and Sankar, 2000), which are reported next.

Davies-Bouldin (DB) Index:

The Davies-Bouldin (DB) index (Bezdek and Pal, 1988) is a function of the ratio of sum of
within-cluster distance to between-cluster separation and is given by
 
1X
c
S(vi ) + S(vk )
DB = maxi6=k
c i=1 d(vi , vk )

for 1 ≤ i, k ≤ c. The DB index minimizes the within-cluster distance S(vi ) and maximizes
the between-cluster separation d(vi , vk ). Therefore, for a given data set and c value, the
higher the similarity values within the clusters and the between-cluster separation, the lower
would be the DB index value. A good clustering procedure should make the value of DB
index as low as possible.

Dunn Index:
Dunn index (Bezdek and Pal, 1988) is also designed to identify sets of clusters that are
compact and well separated. Dunn index maximizes
  
d(vi , vk )
Dunn = mini mini6=k for 1 ≤ i, k, l ≤ c.
maxl S(vl )
A good clustering procedure should make the value of Dunn index as high as possible.

β Index:
The β-index of Pal et al. (Pal et al., 2000) is defined as the ratio of the total variation and
within-cluster variation, and is given by

N Xc X ni Xc X ni Xc
β= ; where N = ||xij − v||2 ; M = ||xij − vi ||2 ; ni = n;
M i=1 j=1 i=1 j=1 i=1

ni is the number of objects in the ith cluster (i = 1, 2, · · · , c), n is the total number of
objects, xij is the jth object in cluster i, vi is the mean or centroid of ith cluster, and v is
the mean of n objects. For a given image and c value, the higher the homogeneity within
the segmented regions, the higher would be the β value. The value of β increases with c.
Consider the image of Fig. 2.3 as an example, which represents an MR image (I-20497774)
of size 256×180 with 16 bit gray levels. So, the number of objects in the data set of IMAGE-
20497774 is 46080. Table 2.1 depicts the values of DB index, Dunn index, and β index of
FCM and RFCM for different values of c on the data set of I-20497774 considering only
gray value of pixel. The results reported here with respect to DB and Dunn index confirm
that both FCM and RFCM achieve their best results for c = 4 (background, gray matter,
white matter, and cerebro-spinal fluid). Also, the value of β index, as expected, increases
Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images 2–9
TABLE 2.1 Performance of FCM and RFCM on
I-20497774
Value DB Index Dunn Index β Index
of c FCM RFCM FCM RFCM FCM RFCM
2 0.51 0.21 2.30 6.17 2.15 2.19
3 0.25 0.17 1.11 1.62 3.55 3.74
4 0.16 0.15 1.50 1.64 9.08 9.68
5 0.39 0.17 0.10 0.64 10.45 10.82
6 0.20 0.19 0.66 1.10 16.93 17.14
7 0.23 0.27 0.98 0.12 21.63 22.73
8 0.34 0.27 0.09 0.31 25.82 26.38
9 0.32 0.28 0.12 0.13 31.75 32.65
10 0.30 0.24 0.08 0.12 38.04 39.31

with increase in the value of c. For a particular value of c, the performance of RFCM is
better than that of FCM.
Fig. 2.2 shows the scatter plots of the highest and second highest memberships of all the
objects in the data set of I-20497774 at first and final iterations respectively, considering
w = 0.95, ḿ1 = 2.0, and c = 4. The diagonal line represents the zone where two highest
memberships of objects are equal. From Fig. 2.2, it is observed that though the average
difference between two highest memberships of the objects are very low at first iteration
(δ = 0.145), they become ultimately very high at the final iteration (δ = 0.652).

At 1st Iteration After 20th Iteration

1 1
Second Highest Membership Value

Second Highest Membership Value

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Highest Membership Value Highest Membership Value

FIGURE 2.2 Scatter plots of two highest membership values of all objects in data set I-20497774

Table 2.2 compares the performance of different c-means algorithms on some brain MR
images with respect to DB, Dunn, and β index considering c = 4 (back-ground, gray
matter, white matter, and CSF). All the results reported in Table 2.2 confirm that the
RFCM algorithm produces pixel clusters more promising than do the conventional methods.
Some of the existing algorithms like PCM and FPCM have failed to produce multiple
clusters as they generate coincident clusters even when they have been initialized with the
final prototypes of FCM. Also, the values of DB, Dunn, and β index of RFCM are better
compared to other c-means algorithms.

2.5 Segmentation of Brain MR Images

2–10 Rough Fuzzy Image Analysis
TABLE 2.2 Performance of Different C-Means Algorithms
Data Set Algorithms DB Index Dunn Index β Index
HCM 0.16 2.13 12.07
I-20497761 FCM 0.14 2.26 12.92
RCM 0.15 2.31 11.68
RFCM 0.13 2.39 13.06
HCM 0.18 1.88 12.02
I-20497763 FCM 0.16 2.02 12.63
RCM 0.15 2.14 12.59
RFCM 0.11 2.12 13.30
HCM 0.18 1.17 8.11
I-20497774 FCM 0.16 1.50 9.08
RCM 0.17 1.51 9.10
RFCM 0.15 1.64 9.68
HCM 0.17 2.01 8.68
I-20497777 FCM 0.16 2.16 9.12
RCM 0.15 2.34 9.28
RFCM 0.14 2.39 9.81

In this section, the feature extraction methodology for segmentation of brain MR images
is first described. Next, the methodology to select initial centroids for different c-means
algorithms is provided based on the concept of maximization of class separability (Maji and
Pal, 2008).

2.5.1 Feature Extraction

Statistical texture analysis derives a set of statistics from the distribution of pixel val-
ues or blocks of pixel values. There are different types of statistical texture, first-order,
second-order, and higher order statistics, based on the number of pixel combinations used
to compute the textures. The first-order statistics, like mean, standard deviation, range,
entropy, and the qth moment about the mean, are calculated using the histogram formed by
the gray scale value of each pixel. These statistics consider the properties of the gray scale
values, but not their spatial distribution. The second-order statistics are based on pairs of
pixels. This takes into account the spatial distribution of the gray scale distribution. In the
present work, only first- and second-order statistical textures are considered.
A set of 13 input features is used for clustering the brain MR images. These include gray
value of the pixel, two recently introduced features (first order statistics) - homogeneity and
edge value of the pixel (Maji and Pal, 2008), and 10 Haralick’s textural features (Haral-
ick, Shanmugam, and Dinstein, 1973) (second order statistics) - angular second moment,
contrast, correlation, inverse difference moment, sum average, sum variance, sum entropy,
second order entropy, difference variance, and difference entropy. They are useful in charac-
terizing images, and can be used as features of a pixel. Hence these features have promising
application in clustering based brain MRI segmentation.

Homogeneity
If H is the homogeneity of a pixel Im,n within 3 × 3 neighborhood, then
1
H=1− {|Im−1,n−1 + Im+1,n+1 − Im−1,n+1 − Im+1,n−1 | +
6(Imax − Imin )
|Im−1,n−1 + 2Im,n−1 + Im+1,n−1 − Im−1,n+1 − 2Im,n+1 − Im+1,n+1 |}
where Imax and Imin represent the maximum and minimum gray values of the image. The
region that is entirely within an organ will have a high H value. On the other hand, the
regions that contain more than one organ will have lower H values (Maji and Pal, 2008).
Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images 2–11

Edge Value
In MR imaging, the histogram of the given image is in general unimodal. One side of the
peak may display a shoulder or slope change, or one side may be less steep than the other,
reflecting the presence of two peaks that are close together or that differ greatly in height.
The histogram may also contain a third, usually smaller, population corresponding to points
on the object-background border. These points have gray levels intermediate between those
of the object and background; their presence raises the level of the valley floor between the
two peaks, or if the peaks are already close together, makes it harder to detect the fact that
they are not a single peak.
As the histogram peaks are close together and very unequal in size, it may be difficult
to detect the valley between them. In determining how each point of the image should
contribute to the segmentation method, the current method takes into account the rate of
change of gray level at the point, as well as the point’s gray level (edge value); that is, the
maximum of differences of average gray levels in pairs of horizontally and vertically adjacent
2 × 2 neighborhoods (Maji et al., 2008; Weszka and Rosenfeld, 1979). If ∆ is the edge value
at a given point Im,n , then

1
∆= max{|Im−1,n + Im−1,n+1 + Im,n + Im,n+1 − Im+1,n − Im+1,n+1 − Im+2,n − Im+2,n+1 |,
4
|Im,n−1 + Im,n + Im+1,n−1 + Im+1,n − Im,n+1 − Im,n+2 − Im+1,n+1 − Im+1,n+2 |}

According to the image model, points interior to the object and background should gen-
erally have low edge values, since they are highly correlated with their neighbors, while
those on the object-background border should have high edge values (Maji et al., 2008).

Haralick’s Textural Feature

Texture is one of the important features used in identifying objects or regions of interest in
an image. It is often described as a set of statistical measures of the spatial distribution of
gray levels in an image. This scheme has been found to provide a powerful input feature
representation for various recognition problems. Haralick et al. (Haralick et al., 1973)
proposed different textural properties for image classification. Haralick’s textural measures
are based upon the moments of a joint probability density function that is estimated as
the joint co-occurrence matrix or gray level co-occurrence matrix (Haralick et al., 1973;
Rangayyan, 2004). It reflects the distribution of the probability of occurrence of a pair
of gray levels separated by a given distance d at angle θ. Based upon normalized gray
level co-occurrence matrix, Haralick proposed several quantities as measure of texture like
energy, contrast, correlation, sum of squares, inverse difference moments, sum average,
sum variance, sum entropy, entropy, difference variance, difference entropy, information
measure of correlation 1, and correlation 2. In (Haralick et al., 1973), these properties were
calculated for large blocks in aerial photographs. Every pixel within these each large block
was then assigned the same texture values. This leads to a significant loss of resolution that
is unacceptable in medical imaging.
In the present work, the texture values are assigned to a pixel by using a 3 × 3 sliding
window centered about that pixel. The gray level co-occurrence matrix is constructed by
mapping the gray level co-occurrence probabilities based on spatial relations of pixels in
different angular directions (θ = 0◦ , 45◦ , 90◦ , 135◦ ) with unit pixel distance, while scanning
the window (centered about a pixel) from left-to-right and top-to-bottom (Haralick et al.,
1973; Rangayyan, 2004). Ten texture measures - angular second moment, contrast, corre-
lation, inverse difference moment, sum average, sum variance, sum entropy, second order
2–12 Rough Fuzzy Image Analysis

entropy, difference variance, and difference entropy, are computed for each window. For
four angular directions, a set of four values is obtained for each of ten measures. The mean
of each of the ten measures, averaged over four values, along with gray value, homogeneity,
and edge value of the pixel, comprise the set of 13 features which is used as feature vector
of the corresponding pixel.

2.5.2 Selection of Initial Centroids

A limitation of the c-means algorithm is that it can only achieve a local optimum solution
that depends on the initial choice of the centroids. Consequently, computing resources may
be wasted in that some initial centroids get stuck in regions of the input space with a scarcity
of data points and may therefore never have the chance to move to new locations where
they are needed. To overcome this limitation of the c-means algorithm, next a method
is described to select initial centroids, which is based on discriminant analysis maximizing
some measures of class separability (Otsu, 1979). It enables the algorithm to converge to
an optimum or near optimum solutions (Maji and Pal, 2008).
Prior to describe the new method for selecting initial centroids, next a quantitative mea-
sure of class separability (Otsu, 1979) is provided that is given by

P1 (T)P2 (T)[m1 (T) − m2 (T)]2

J(T) = (2.7)
P1 (T)σ12 (T) + P2 (T)σ22 (T)
where
T
X L−1
X
P1 (T) = h(z); P2 (T) = h(z) = 1 − P1 (T)
z=0 z=T+1
T L−1
1 X 1 X
m1 (T) = zh(z); m2 (T) = zh(z)
P1 (T) z=0 P2 (T)
z=T+1
T
X L−1
X
1 1
σ12 (T) = [z − m1 (T)]2 h(z); σ22 (T) = [z − m2 (T)]2 h(z)
P1 (T) z=0
P2 (T)
z=T+1

Here, L is the total number of discrete values ranging between [0, L − 1], T is the threshold
value, which maximizes J(T), and h(z) represents the percentage of data having feature
value z over the total number of discrete values of the corresponding feature. To maximize
J(T), the means of the two classes should be as well separated as possible and the variances
in both classes should be as small as possible.
Based on the concept of maximization of class separability, the method for selecting initial
centroids is described next. The main steps of this method proceeds as follows.
1. The data set X = {x1 , · · · , xj , · · · , xn } with xj ∈ ℜm are first discretized to
facilitate class separation method. Suppose, the possible value range of a feature
fm in the data set is (fm,min , fm,max ), and the real value that the data element xj
takes at fm is fmj , then the discretized value of fmj is
 
fmj − fm,min
Discretized(fmj ) = (L − 1) × (2.8)
fm,max − fm,min

where L is the total number of discrete values ranging between [0, L − 1].
2. For each feature fm , calculate h(z) for 0 ≤ z < L.
Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images 2–13

3. Calculate the threshold value Tm for the feature fm , which maximizes class sep-
arability along that feature.
4. Based on the threshold Tm , discretize the corresponding feature fm of the data
element xj as follows


1, if Discretized(fmj ) ≥ Tm
f mj =
0, Otherwise

5. Repeat steps 2 to 4 for all the features and generate the set of discretized objects
X = {x1 , · · · , xj , · · · , xn }.
6. Calculate total number of similar discretized objects N(xi ) and mean of similar
objects v(xi ) of xi as

X
n
1 X
n
N(xi ) = δj and v(xi ) = δj × xj
j=1
N(xi ) j=1


1 if xj = xi
where δj =
0 Otherwise

7. Sort n objects according to their values of N(xi ) such that N(x1 ) > N(x2 ) >
· · · > N(xn ).
8. If xi = xj , then N(xi ) = N(xj ) and v(xj ) should not be considered as a centroid
(mean), resulting in a reduced set of objects to be considered for initial centroids.
9. Let there be ń objects in the reduced set having N(xi ) values such that N(x1 ) >
N(x2 ) > · · · > N(xń ). A heuristic threshold function can be defined as follows
(Banerjee, Mitra, and Pal, 1998):

R X ń
1
Tr = ; where R =
ǫ̃ i=1
N(xi ) − N(xi+1 )

where ǫ̃ is a constant (= 0.5, say), so that all the means v(xi ) of the objects in
reduced set having N(xi ) value higher than it are regarded as the candidates for
initial centroids (means).
The value of Tr is high if most of the N(xi )’s are large and close to each other. The above
condition occurs when a small number of large clusters are present. On the other hand, if
the N(xi )’s have wide variation among them, then the number of clusters with smaller size
increases. Accordingly, Tr attains a lower value automatically. Note that the main motive
of introducing this threshold function lies in reducing the number of centroids. Actually,
it attempts to eliminate noisy centroids (data representatives having lower values of N(xi ))
from the whole data set. The whole approach is, therefore, data dependent.

2.6 Experimental Results and Discussion

2–14 Rough Fuzzy Image Analysis

In this section, the performance of different c-means algorithms on segmentation of brain

MR images is presented. Details of the experimental set up, data collection, and objective
of the experiments are same as those of Section 2.4.
Consider Fig. 2.3 as an example that represents an MR image (I-20497774) along with
the segmented images obtained using different c-means algorithms. Each image is of size
256 × 180 with 16 bit gray levels. So, the number of objects in the data set of I-20497774 is
46080. The parameters generated in the discriminant analysis based initialization method
are shown in Table 2.3 only for I-20497774 data set along with the values of input param-
eters. The threshold values for 13 features of the given data set are also reported in this
table. Table 2.4 depicts the values of DB index, Dunn index, and β index of FCM and

FIGURE 2.3 I-20497774: original and segmented images of HCM, FCM, RCM, and RFCM

RFCM for different values of c on the data set of I-20497774, considering w = 0.95 and
ḿ = 2.0. The results reported here with respect to DB and Dunn index confirm that both
FCM and RFCM achieve their best results for c = 4. Also, the value of β index, as expected,
increases with increase in the value of c. For a particular value of c, the performance of
RFCM is better than that of FCM.

TABLE 2.3 Values of Different Parameters

Size of image = 256 × 180
Minimum gray value = 1606, Maximum gray value = 2246
Samples per pixel = 1, Bits allocated = 16, Bits stored = 12

Number of objects = 46080

Number of features = 13, Value of L = 101

Threshold Values:
Gray value = 1959, Homogeneity = 0.17, Edge value = 0.37
Angular second moment = 0.06, Contrast = 0.12
Correlation = 0.57, Inverse difference moment = 0.18
Sum average = 0.17, Sum variance = 0.14, Sum entropy = 0.87
Entropy = 0.88, Difference variance = 0.07, Difference entropy = 0.79

Finally, Table 2.5 provides the comparative results of different c-means algorithms on
I-20497774 with respect to the values of DB index, Dunn index, and β index. The cor-
responding segmented images along with the original one are presented in Fig. 2.3. The
results reported in Fig. 2.3 and Table 2.5 confirm that the RFCM algorithm produces seg-
mented image more promising than do the conventional c-means algorithms. Some of the
existing algorithms like PCM and FPCM fail to produce multiple segments as they generate
coincident clusters even when they are initialized with final prototypes of the FCM.
Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images 2–15
TABLE 2.4 Performance of FCM and RFCM on
I-20497774 data set
Value DB Index Dunn Index β Index
of c FCM RFCM FCM RFCM FCM RFCM
2 0.38 0.19 2.17 3.43 3.62 4.23
3 0.22 0.16 1.20 1.78 7.04 7.64
4 0.15 0.13 1.54 1.80 11.16 13.01
5 0.29 0.19 0.95 1.04 11.88 14.83
6 0.24 0.23 0.98 1.11 19.15 19.59
7 0.23 0.21 1.07 0.86 24.07 27.80
8 0.31 0.21 0.46 0.95 29.00 33.02
9 0.30 0.24 0.73 0.74 35.06 40.07
10 0.30 0.22 0.81 0.29 41.12 44.27

TABLE 2.5 Performance of Different C-Means

on I-20497774 data set
Algorithms DB Index Dunn Index β Index
HCM 0.17 1.28 10.57
FCM 0.15 1.54 11.16
RCM 0.16 1.56 11.19
RFCM 0.13 1.80 13.01

TABLE 2.6 Haralick’s and Proposed Features on I-20497774 data set

Algorithms Features DB Index Dunn Index β Index Time (ms)
HCM H-13 0.19 1.28 10.57 4308
H-10 0.19 1.28 10.57 3845
P-2 0.18 1.28 10.57 1867
H-10 ∪ P-2 0.17 1.28 10.57 3882
FCM H-13 0.15 1.51 10.84 36711
H-10 0.15 1.51 10.84 34251
P-2 0.15 1.51 11.03 14622
H-10 ∪ P-2 0.15 1.54 11.16 43109
RCM H-13 0.19 1.52 11.12 5204
H-10 0.19 1.52 11.12 5012
P-2 0.17 1.51 11.02 1497
H-10 ∪ P-2 0.16 1.56 11.19 7618
RFCM H-13 0.13 1.76 12.57 15705
H-10 0.13 1.76 12.57 15414
P-2 0.13 1.77 12.88 6866
H-10 ∪ P-2 0.13 1.80 13.01 17084

2.6.1 Haralick’s Features Versus Proposed Features

Table 2.6 presents the comparative results of different c-means for Haralick’s features and
features proposed in (Maji and Pal, 2008) on I-20497774 data set. While P-2 and H-13 stand
for the set of two proposed features (Maji and Pal, 2008) and thirteen Haralick’s features,
H-10 represents that of ten Haralick’s features which are used in the current study. The
proposed features are found as important as Haralick’s ten features for clustering based
segmentation of brain MR images. The set of 13 features, comprising of gray value, two
proposed features, and ten Haralick’s features, improves the performance of all c-means
with respect to DB, Dunn, and β. It is also observed that the Haralick’s three features -
sum of squares, information measure of correlation 1, and correlation 2, do not contribute
any extra information for segmentation of brain MR images.

2.6.2 Random Versus Discriminant Analysis Based Initialization

Table 2.7 provides comparative results of different c-means algorithms with random initial-
ization of centroids and the discriminant analysis based initialization method described in
2–16 Rough Fuzzy Image Analysis

Section 2.5.2 for the data sets I-20497761, I-20497763, and I-20497777 (Fig. 2.4). The

TABLE 2.7 Performance of Random and Discriminant Analysis Based Initialization

Method
Data Set Algorithms Initialization DB Index Dunn Index β Index Time (ms)
I-204 HCM Random 0.23 1.58 9.86 8297
97761 Proposed 0.15 2.64 12.44 4080
FCM Random 0.19 1.63 12.73 40943
Proposed 0.12 2.69 13.35 38625
RCM Random 0.19 1.66 10.90 9074
Proposed 0.14 2.79 12.13 6670
RFCM Random 0.15 2.07 11.89 19679
Proposed 0.11 2.98 13.57 16532
I-204 HCM Random 0.26 1.37 10.16 3287
97763 Proposed 0.16 2.03 13.18 3262
FCM Random 0.21 1.54 10.57 46157
Proposed 0.15 2.24 13.79 45966
RCM Random 0.21 1.60 10.84 10166
Proposed 0.14 2.39 13.80 6770
RFCM Random 0.17 1.89 11.49 19448
Proposed 0.10 2.38 14.27 15457
I-204 HCM Random 0.33 1.52 6.79 4322
97777 Proposed 0.16 2.38 8.94 3825
FCM Random 0.28 1.67 7.33 42284
Proposed 0.15 2.54 10.02 40827
RCM Random 0.27 1.71 7.47 8353
Proposed 0.13 2.79 9.89 7512
RFCM Random 0.19 1.98 8.13 18968
Proposed 0.11 2.83 11.04 16930

FIGURE 2.4 Examples of some brain MR images: I-20497761, I-20497763, I-20497777

discriminant analysis based initialization method is found to improve the performance in

terms of DB index, Dunn index, and β index as well as reduce the time requirement of all
c-means algorithms. It is also observed that HCM with this initialization method performs
similar to RFCM with random initialization, although it is expected that RFCM is superior
to HCM in partitioning the objects. While in random initialization, the c-means algorithms
get stuck in local optimums, the discriminant analysis based initialization method enables
the algorithms to converge to an optimum or near optimum solutions. In effect, the exe-
cution time required for different c-means algorithms is lesser in this scheme compared to
random initialization.

2.6.3 Comparative Performance Analysis

Table 2.8 compares the performance of different c-means algorithms on some brain MR
images with respect to DB, Dunn, and β index. The segmented versions of different c-
Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images 2–17

means are shown in Figs. 2.5-2.7. All the results reported in Table 2.8 and Figs. 2.5-2.7

TABLE 2.8 Performance of Different C-Means Algorithms

Data Set Algorithms DB Index Dunn Index β Index Time (ms)
I-204 HCM 0.15 2.64 12.44 4080
97761 FCM 0.12 2.69 13.35 38625
RCM 0.14 2.79 12.13 6670
RFCM 0.11 2.98 13.57 16532
I-204 HCM 0.16 2.03 13.18 3262
97763 FCM 0.15 2.24 13.79 45966
RCM 0.14 2.39 13.80 6770
RFCM 0.10 2.38 14.27 15457
I-204 HCM 0.16 2.38 8.94 3825
97777 FCM 0.15 2.54 10.02 40827
RCM 0.13 2.79 9.89 7512
RFCM 0.11 2.83 11.04 16930

confirm that although each c-means algorithm, except PCM and FPCM, generates good
segmented images, the values of DB, Dunn, and β index of the RFCM are better compared
to other c-means algorithms. Both PCM and FPCM fail to produce multiple segments of
the brain MR images as they generate coincident clusters even when they are initialized
with the final prototypes of other c-means algorithms.
Table 2.8 also provides execution time (in milli sec.) of different c-means. The execution
time required for the RFCM is significantly lesser compared to FCM. For the HCM and
RCM, although the execution time is less, the performance is considerably poorer than that
of RFCM. Following conclusions can be drawn from the results reported in this chapter:

FIGURE 2.5 I-20497761: segmented versions of HCM, FCM, RCM, and RFCM

FIGURE 2.6 I-20497763: segmented versions of HCM, FCM, RCM, and RFCM

1. It is observed that RFCM is superior to other c-means algorithms. However,

RFCM requires higher time compared to HCM/RCM and lesser time compared
2–18 Rough Fuzzy Image Analysis

FIGURE 2.7 I-20497777: segmented versions of HCM, FCM, RCM, and RFCM

to FCM. But, the performance of RFCM with respect to DB, Dunn, and β is
significantly better than all other c-means. The performance of FCM and RCM
is intermediate between RFCM and HCM.
2. The discriminant analysis based initialization is found to improve the values of
DB, Dunn, and β as well as reduce the time requirement substantially for all
c-means algorithms.
3. Two features proposed in (Maji and Pal, 2008) are as important as Haralick’s
ten features for clustering based segmentation of brain MR images.
4. Use of rough sets and fuzzy memberships adds a small computational load to
HCM algorithm; however the corresponding integrated method (RFCM) shows
a definite increase in Dunn index and decrease in DB index.
The best performance of the segmentation method in terms of DB, Dunn, and β is
achieved due to the following reasons:
1. the discriminant analysis based initialization of centroids enables the algorithm
to converge to an optimum or near optimum solutions;
2. membership of the RFCM handles efficiently overlapping partitions; and
3. the concept of crisp lower bound and fuzzy boundary of the RFCM algorithm
deals with uncertainty, vagueness, and incompleteness in class definition.
In effect, promising segmented brain MR images are obtained using the RFCM algorithm.

2.7 Conclusion
A robust segmentation technique is presented in this chapter, integrating the merits of
rough sets, fuzzy sets, and c-means algorithm, for brain MR images. Some new measures
are reported, based on the local properties of MR images, for accurate segmentation. The
method, based on the concept of maximization of class separability, is found to be suc-
cessful in effectively circumventing the initialization and local minima problems of iterative
refinement clustering algorithms like c-means. The effectiveness of the algorithm, along
with a comparison with other algorithms, is demonstrated on a set of brain MR images.
The extensive experimental results show that the rough-fuzzy c-means algorithm produces
a segmented image more promising than do the conventional algorithms.

Acknowledgments.
The authors thank Advanced Medicare and Research Institute, Kolkata, India, for providing
brain MR images. This work was done when S. K. Pal was a Govt. of India J.C. Bose Fellow.
Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images 2–19

Bibliography
Banerjee, Mohua, Sushmita Mitra, and Sankar K Pal. 1998. Rough Fuzzy MLP: Knowl-
edge Encoding and Classification. IEEE Transactions on Neural Networks 9(6):
1203–1216.

Barni, M., V. Cappellini, and A. Mecocci. 1996. Comments on A Possibilistic Approach

to Clustering. IEEE Transactions on Fuzzy Systems 4(3):393–396.

Bezdek, J. C. 1981. Pattern Recognition with Fuzzy Objective Function Algorithm.

New York: Plenum.

Bezdek, J. C., and N. R. Pal. 1988. Some New Indexes for Cluster Validity. IEEE
Transactions on System, Man, and Cybernetics, Part B 28:301–315.

Brandt, M. E., T. P. Bohan, L. A. Kramer, and J. M. Fletcher. 1994. Estimation

of CSF, White and Gray Matter Volumes in Hydrocephalic Children Using Fuzzy
Clustering of MR Images. Computerized Medical Imaging and Graphics 18:25–34.

Cagnoni, S., G. Coppini, M. Rucci, D. Caramella, and G. Valli. 1993. Neural Network
Segmentation of Magnetic Resonance Spin Echo Images of the Brain. Journal of
Biomedical Engineering 15(5):355–362.

Dubois, D., and H.Prade. 1990. Rough Fuzzy Sets and Fuzzy Rough Sets. International
Journal of General Systems 17:191–209.

Dunn, J. C. 1974. A Fuzzy Relative of the ISODATA Process and its Use in Detecting
Compact, Well-Separated Clusters. Journal of Cybernetics 3:32–57.

Hall, L. O., A. M. Bensaid, L. P. Clarke, R. P. Velthuizen, M. S. Silbiger, and J. C.

Bezdek. 1992. A Comparison of Neural Network and Fuzzy Clustering Techniques in
Segmenting Magnetic Resonance Images of the Brain. IEEE Transactions on Neural
Networks 3(5):672–682.

Haralick, R. M., K. Shanmugam, and I. Dinstein. 1973. Textural Features for Image
Classification. IEEE Transactions on Systems, Man and Cybernetics SMC-3(6):
610–621.

Hassanien, Aboul Ella. 2007. Fuzzy Rough Sets Hybrid Scheme for Breast Cancer
Detection. Image Vision Computing 25(2):172–183.

Krishnapuram, R., and J. M. Keller. 1993. A Possibilistic Approach to Clustering.

IEEE Transactions on Fuzzy Systems 1(2):98–110.

———. 1996. The Possibilistic C-Means Algorithm: Insights and Recommendations.

IEEE Transactions on Fuzzy Systems 4(3):385–393.

Lee, C., S. Hun, T. A. Ketter, and M. Unser. 1998. Unsupervised Connectivity Based
Thresholding Segmentation of Midsaggital Brain MR Images. Computers in Biology
and Medicine 28:309–338.

Leemput, K. V., F. Maes, D. Vandermeulen, and P. Suetens. 1999. Automated Model-

Based Tissue Classification of MR Images of the Brain. IEEE Transactions on
Medical Imaging 18(10):897–908.
2–20 Rough Fuzzy Image Analysis

Li, C. L., D. B. Goldgof, and L. O. Hall. 1993. Knowledge-Based Classification and

Tissue Labeling of MR Images of Human Brain. IEEE Transactions on Medical
Imaging 12(4):740–750.
Maji, Pradipta, Malay K. Kundu, and Bhabatosh Chanda. 2008. Second Order Fuzzy
Measure and Weighted Co-Occurrence Matrix for Segmentation of Brain MR Images.
Fundamenta Informaticae 88(1-2):161–176.
Maji, Pradipta, and Sankar K. Pal. 2007a. RFCM: A Hybrid Clustering Algorithm
Using Rough and Fuzzy Sets. Fundamenta Informaticae 80(4):475–496.
———. 2007b. Rough-Fuzzy C-Medoids Algorithm and Selection of Bio-Basis for
Amino Acid Sequence Analysis. IEEE Transactions on Knowledge and Data Engi-
neering 19(6):859–872.
———. 2007c. Rough Set Based Generalized Fuzzy C-Means Algorithm and Quanti-
tative Indices. IEEE Transactions on System, Man and Cybernetics, Part B, Cyber-
netics 37(6):1529–1540.
———. 2008. Maximum Class Separability for Rough-Fuzzy C-Means Based Brain
MR Image Segmentation. LNCS Transactions on Rough Sets IX(5390):114–134.
Manousakes, I. N., P. E. Undrill, and G. G. Cameron. 1998. Split and Merge Segmenta-
tion of Magnetic Resonance Medical Images: Performance Evaluation and Extension
to Three Dimensions. Computers and Biomedical Research 31(6):393–412.
Mushrif, Milind M., and Ajoy K. Ray. 2008. Color Image Segmentation: Rough-Set
Theoretic Approach. Pattern Recognition Letters 29(4):483–493.
Otsu, N. 1979. A Threshold Selection Method from Gray Level Histogram. IEEE
Transactions on System, Man, and Cybernetics 9(1):62–66.
Pal, N. R., K. Pal, J. M. Keller, and J. C. Bezdek. 2005. A Possibilistic Fuzzy C-Means
Clustering Algorithm. IEEE Transactions on Fuzzy Systems 13(4):517–530.
Pal, N. R., and S. K. Pal. 1993. A Review on Image Segmentation Techniques. Pattern
Recognition 26(9):1277–1294.
Pal, S. K., A. Ghosh, and B. Uma Sankar. 2000. Segmentation of Remotely Sensed Im-
ages with Fuzzy Thresholding, and Quantitative Evaluation. International Journal
of Remote Sensing 21(11):2269–2300.
Pal, S. K., and P. Mitra. 2002. Multispectral Image Segmentation Using Rough Set
Initiatized EM Algorithm. IEEE Transactions on Geoscience and Remote Sensing
40(11):2495–2501.
Pal, Sankar K, Sushmita Mitra, and Pabitra Mitra. 2003. Rough-Fuzzy MLP: Modular
Evolution, Rule Generation, and Evaluation. IEEE Transactions on Knowledge and
Data Engineering 15(1):14–25.
Pawlak, Z. 1991. Rough Sets, Theoretical Aspects of Resoning About Data. Dordrecht,
The Netherlands: Kluwer.
Rajapakse, J. C., J. N. Giedd, and J. L. Rapoport. 1997. Statistical Approach to
Segmentation of Single Channel Cerebral MR Images. IEEE Transactions on Medical
Imaging 16:176–186.
Rough-Fuzzy Clustering Algorithm for Segmentation of Brain MR Images 2–21

Rangayyan, Rangaraj M. 2004. Biomedical Image Analysis. CRC Press.

Rosenfeld, A., and A. C. Kak. 1982. Digital Picture Processing. Academic Press, Inc.
Singleton, H. R., and G. M. Pohost. 1997. Automatic Cardiac MR Image Segmentation
Using Edge Detection by Tissue Classification in Pixel Neighborhoods. Magnetic
Resonance in Medicine 37(3):418–424.
Suetens, Paul. 2002. Fundamentals of Medical Imaging. Cambridge University Press.
Wells III, W. M., W. E. L. Grimson, R. Kikinis, and F. A. Jolesz. 1996. Adaptive
Segmentation of MRI Data. IEEE Transactions on Medical Imaging 15(4):429–442.
Weszka, J. S., and A. Rosenfeld. 1979. Histogram Modification for Threshold Selection.
IEEE Transactions on System, Man, and Cybernetics SMC-9(1):62–66.
Widz, Sebastian, Kenneth Revett, and Dominik Slezak. 2005a. A Hybrid Approach
to MR Imaging Segmentation Using Unsupervised Clustering and Approximate
Reducts. Proceedings of the 10th International Conference on Rough Sets, Fuzzy
Sets, Data Mining, and Granular Computing 372–382.
———. 2005b. A Rough Set-Based Magnetic Resonance Imaging Partial Volume De-
tection System. Proceedings of the First International Conference on Pattern Recog-
nition and Machine Intelligence 756–761.
Widz, Sebastian, and Dominik Slezak. 2007. Approximation Degrees in Decision
Reduct-Based MRI Segmentation. Proceedings of the Frontiers in the Convergence
of Bioscience and Information Technologies 431–436.
Xiao, Kai, Sooi Hock Ho, and Aboul Ella Hassanien. 2008. Automatic Unsupervised
Segmentation Methods for MRI Based on Modified Fuzzy C-Means. Fundamenta
Informaticae 87(3-4):465–481.
Zadeh, L. A. 1965. Fuzzy Sets. Information and Control 8:338–353.
 3
Image Thresholding using
Generalized Rough Sets

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3–1

3.2 Generalized Rough Set based Entropy Measures with
respect to the Definability of a Set of Elements . . 3–3
Roughness of a Set in a Universe • The Lower and
Upper Approximations of a Set • The Entropy
Measures • Relation between ρR (X) and ρR (X { ) •
Properties of the Proposed Classes of Entropy
Measures
3.3 Measuring Grayness Ambiguity in Images . . . . . . . . . 3–11
Debashis Sen
Center for Soft Computing Research, Indian 3.4 Image Thresholding based on Association Error . . 3–15
Statistical Institute
Bilevel Thresholding • Multilevel Thresholding
3.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3–19
Sankar K. Pal Qualitative analysis • Quantitative analysis
Center for Soft Computing Research, Indian 3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3–26
Statistical Institute Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3–27

3.1 Introduction
Real-life images are inherently embedded with various ambiguities. In order to perceive
the nature of ambiguities in images, let us consider a 1001 × 1001 grayscale image (see
Figure 3.1(a)) that has sinusoidal gray value gradations in horizontal direction. When
an attempt is made to mark the boundary of an arbitrary region in the image, an exact
boundary can not be defined as a consequence of the presence of steadily changing gray
values (gray value gradation). This is evident from Figure 3.1(b) that shows a portion of
the image, where it is known that the pixels in the ‘white’ shaded area uniquely belong
to a region. However, the boundary (on the left and right sides) of this region is vague as
it can lie anywhere in the gray value gradations present in the portion. Value gradation
is a common phenomenon in real-life images and hence it is widely accepted (Pal, 1982;
Pal, King, and Hashim, 1983; Udupa and Saha, 2003) that regions in an image have fuzzy
boundaries.
Moreover, the gray levels at various pixels in grayscale images are considered to be impre-
cise, which means that a gray level resembles other nearby gray levels to certain extents. It
is also true that pixels in a neighborhood with nearby gray levels have limited discernibility
due to the inadequacy of contrast. For example, Figure 3.1(c) shows a 6 × 6 portion cut
from the image in Figure 3.1(a). Although the portion contains gray values separated by
6 gray levels, it appears to be almost homogeneous. The aforementioned ambiguities in

3–1
3–2 Rough Fuzzy Image Analysis

(a) A grayscale image (b) Fuzzy boundary (c) Rough

resemblance

FIGURE 3.1: Ambiguities in a grayscale image with sinusoidal gray value gradations in horizontal
direction

images due to fuzzy boundaries of various regions and rough resemblance of nearby gray
levels is studied and modeled in this chapter. Note that, the aforementioned ambiguities
are related to the indefiniteness in deciding an image pixel as white or black and hence they
can be collectively referred to as the grayness ambiguity (Pal, 1999).
Fuzzy set theory of Lofti Zadeh, is based on the concept of vague boundaries of sets in
the universe of discourse (Klir and Yuan, 2005). Rough set theory of Zdzislaw Pawlak, on
the otherhand, focuses on ambiguity in terms of limited discernibility of sets in the domain
of discourse (Pawlak, 1991). Therefore, fuzzy sets can be used to represent the grayness
ambiguity in images due to the vague definition of region boundaries (fuzzy boundaries)
and rough sets can be used to represent the grayness ambiguity due to the indiscernibility
between individual or groups of gray levels (rough resemblance).
Rough set theory, which was initially developed considering crisp equivalence approx-
imation spaces (Pawlak, 1991), has been generalized by considering fuzzy (Dubois and
Prade, 1990, 1992; Thiele, 1998) and tolerance (Skowron and Stepaniuk, 1996) approxima-
tion spaces. Furthermore, rough set theory, which was initially developed to approximate
crisp sets, has also been generalized to approximate fuzzy sets (Dubois and Prade, 1990,
1992; Thiele, 1998).
In this chapter, we propose the use of the rough set theory and its certain generalizations
to quantify grayness ambiguity in images. Here the generalizations to rough set theory based
on the approximation of crisp and fuzzy sets considering crisp equivalence, fuzzy equiva-
lence, crisp tolerance and fuzzy tolerance approximation spaces in different combinations
are studied. All these combinations give rise to different concepts for modeling vagueness,
which can be quantified using the roughness measure (Pawlak, 1991).
We propose classes of entropy measures which use the roughness measures obtained con-
sidering the aforementioned various concepts for modeling vagueness. We perform rigorous
theoretical analysis of the proposed entropy measures and provide some properties which
they satisfy. We then use the proposed entropy measures to quantify grayness ambiguity in
images, giving an account of the manner in which the grayness ambiguity is captured. We
show that the aforesaid generalizations to rough set theory regarding the approximation of
fuzzy sets can be used to quantify grayness ambiguity due to both fuzzy boundaries and
rough resemblance.
We then propose an image thresholding methodology that employs grayness ambiguity
Exploring the Variety of Random
Documents with Different Content
to serve as the mouthpiece of its author). “Thackeray,” so W. C.
Brownell has asserted in his sympathetic study,

enwraps and embroiders his story with his personal philosophy, charges it with his
personal feeling, draws out with inexhaustible personal zest its typical
suggestiveness, and deals with his material directly instead of dispassionately and
disinterestedly.

This is a privilege implacably denied to the playwright, even if he

has abundant compensation in other ways. As Brownell also
reminded us, the novel is
a picture of life, but a picture that not only portrays but shows the significance of its
subject; its form is particularly, uniquely elastic, and it possesses epic advantages
which it would fruitlessly forego in conforming itself to purely dramatic canons.

III

Dickens’s novels were both theatrical and dramatic; they were

influenced by the melodramas and farces of his youth, as has
already been noted; and it was natural that they should tempt
adapters to dramatize them. They abounded in robustly drawn
character, often verging into caricature; and therefore they appealed
to the actor. They had episodes of violence certain to prove attractive
to the public which liked to be powerfully moved and which had little
delicacy as to the passions portrayed. Dickens’s sprawling serials
were too straggling in story ever to make it possible to compress
them into a solidly built framework of plot; but it was not difficult to
disentangle a succession of situations sufficient to make an effective
panorama of action, peopled with familiar figures. And of these there
have been an unnumbered host.
If Thackeray’s novels lend themselves less temptingly to the
paste-and-scissors method of the dramatizer, they had an immediate
vogue and an enduring reputation, which have allured a host of
playwrights, most of whom have confined their exertion to the
singling out of a salient character and to the presentation in a play of
the more important situations in which this personality is involved,
utilizing the other figures and the other episodes only in so far as
these might be necessary to set off the chosen hero or heroine.
Naturally enough it is upon ‘Vanity Fair’ that they have laid hands
most frequently. The final monthly part of the original publication had
scarcely been issued when John Brougham ventured upon a stage-
version of it, which he produced at Burton’s theater in New York in
1849.
This was an attempt to dramatize the novel as a whole, although
necessarily Becky Sharp held the center of the stage. There was a
revival of Brougham’s adaptation a few years later; there was
another attempt by George Fawcett Rowe; and then in 1893 Sir
James Barrie made a one-act playlet out of the last glimpse of Becky
that Thackeray affords us, when she and Jos. Sedley, Amelia and
Dobbin find themselves together in the little German watering-place
and when Amelia learns the truth about her dead husband’s
advances to Becky. Sir James has kindly informed me that he thinks
that every word spoken in his little piece was Thackeray’s, “but some
of them were probably taken from different chapters.”
A few years later two other Becky Sharp pieces were produced,
one on either side of the Atlantic. The American play was adroitly
prepared by Langdon Mitchell; it was called ‘Becky Sharp’; it was
produced in 1899 and it has been revived at least once since; Mrs.
Fiske was the Becky. The British play was by Robert Hichens and
Cosmo Gordon Lennox; it was originally performed in London, with
Marie Tempest as Becky; and she came over to the United States to
present it a few times at the New Theater in New York in 1910.
A similar method—the method of focussing the attention of the
audience on a single dominating personality and of excluding all the
episodes in which this personality was not supreme—was followed in
more recent plays cut out of the ‘Newcomes’ and ‘Pendennis.’ No
doubt this was the only possible way of dramatizing novels of such
complexity of episode. Brownell has declared that the range of the
‘Newcomes’ is extraordinary for the thread of a single story to follow:

Yet all its parts are as interdependent as they are numerous and varied. It is
Thackeray’s largest canvas, and it is filled with the greatest ease and to the
borders.... It illustrates manners with an unexampled crowd of characters, the
handling of which, without repetition or confusion, without digression or discord,
exhibits the control of the artist equally with the imaginative and creative faculty of
the poet.

A story as vast as the ‘Newcomes’ simply defies the dramatizer;

and all he can do is to build his play about a single group or, better
still, around a single character, relentlessly excluding all the other
allied groups of personages not less interesting in themselves. This
has been the method, it may be recorded, chosen by the several
French playwrights who have been moved to make dramas out of
one or another of the almost equally complex novels of Balzac.
So it was that Michael Morton made a ‘Colonel Newcome’ piece
for Beerbohm Tree in 1906 and that Langdon Mitchell made a ‘Major
Pendennis’ piece for John Drew in 1916. So it was that Francis
Burnand made a ‘Jeames’ piece for Edward Terry in 1878 out of the
‘Diary of C. Jeames de la Pluche.’ Altho Edward Terry was an
amusing Jeames and altho Nelly Farren was an amusing Mary Ann
Hoggins, the “New and Original Comedy” (as its adapter styled it) did
not strike me as amusing in itself; it was three-quarters Burnand and
barely one quarter Thackeray—and the blending was not to my
taste. As I sat through the performance patiently I came to
understand the provocation which had led a gallery boy to shout
down to Burnand as he took the author’s curtain call on the first
night,—“I say, Frank, it’s a good thing Thackeray is dead, isn’t it?”
As the author had provided the ‘History of Henry Esmond’ with a
unifying figure, the dramatizers have only too abundant material for a
chronicle-play showing him at different periods in his long and
honorable career. To make a compact play, a true drama, out of the
protracted story, would be plainly impossible, yet it might not be so
difficult to select salient episodes which would serve as a succinct
summary of the story. But altho the attempt has been made several
times—once for Henry Irving—no one of the versions has ever been
put up for a run in any of the principal playhouses of either New York
or London. In any dramatization one scene would impose itself, the
scene in which Esmond breaks his sword before the prince whom he
has loyally served, the scene in which Thackeray is most truly
dramatic in the noblest sense of the word. If this had been put on the
stage it would have been only a rendering unto the theater of a thing
that belonged to the theater, since perhaps Thackeray had it
suggested to him by the corresponding scene in the opera of ‘The
Favorite’—altho the suggestion may also have come from the
‘Vicomte de Bragelonne’ or from the later play which Dumas made
out of his own story.
There remains to be mentioned only one other dramatization, that
of the ‘Rose and the Ring,’ made by H. Savile Clark in 1890. From all
accounts the performance of this little play, with its music by Walter
Slaughter, provided a charming spectacle for children, one to which
we may be sure that Thackeray would have had no objection and
which indeed might have delighted his heart.

IV

It is testimony to Thackeray’s own liking for the theater that he is

continually telling us that this or that character went to the play. He
also informs us that Henry Esmond was the author of the ‘Faithful
Fool,’ a comedy performed by Her Majesty’s Servants and published
anonymously, attaining a sale of nine copies, whereupon Esmond
had the whole impression destroyed. And the first of the George
Warringtons wrote two plays, ‘Carpezan’ and Pocahontas,’ both of
them tragedies, the first of which caught the public taste, whereas
the second failed to prove attractive. We are all aware that Becky
Sharp took part in the private theatricals at Gaunt House, making a
most impressive Clytemnestra; but we are less likely to recall the
hesitating suggestion that she may have been the Madame
Rebecque who failed to please when she appeared in the ‘Dame
Blanche’ at Strasburg in 1830. It was natural enough that Becky
should go on the stage, since her mother had been a ballet-dancer.
Altho neither Thackeray nor Dickens ever attempted to write a
novel of theatrical life, each of them gave us an inside view of a
provincial stock-company in the earlier years of the nineteenth
century. In ‘Nicholas Nickleby’ we are introduced to the actors and
actresses under the management of Mr. Crummles; and in
Pendennis’ we have a less elaborate study of the actors and
actresses under the management of Mr. Bingley. The group that
Dickens portrays is more boldly drawn and more richly colored than
the group that Thackeray sketches in with a few illuminating strokes.
“What a light of benevolence it is that plays round Crummles and the
Phenomenon and all those poor theatrical people in that charming
book,” said Thackeray in his lecture on ‘Charity and Humor.’ “What a
humor! And what a good humor!”
Altho in these episodes neither Dickens nor Thackeray aimed at
the penetrating inquisition into the histrionic temperament that we
find in Henry James’s ‘Tragic Muse’ and in Howells’s ‘Story of a
Play,’ there is both validity and originality in Thackeray’s portrait of
Miss Fotheringay. In all the dozens and scores of theatrical novels
that I have read, I do not recall any other attempt to show the actress
who is only an instrument in the hands of a superior intelligence, a
woman who has the divine gift and who can display it only when she
is taught, perhaps by one himself deficient in the mimetic faculty but
possessed of interpretative imagination. Possibly Thackeray bestows
overmuch stupidity on the Fotheringay; but she was not too stupid to
profit by the instruction of the devoted Bows. She had beauty, voice,
manner, the command of emotion, without which the tragic actor is
naught; and all she lacked was the intelligence which would enable
her to make the most of her native endowment.
Except when she was on the stage Mrs. Siddons was an
eminently uninspired woman; and not a little of her inspiration in the
theater has been credited to the superior intellect of her brother,
John Philip Kemble. Rachel was intelligent, so intelligent that she
was always eager to be aided by the intelligence of others. Legouvé
recorded that if he gave her a suggestion, she seized on it and
transmuted his copper into silver. She used to confess the immensity
of her debt to Samson, a little dried up actor of “old men”; and she
said once that she did not play a part half as well as she could play
it, unless she had had the counsel of Samson. Even if she was a
genius, she was rather a marvellous executant than a great
composer; and there has been many another actress, even in our
own time, who has owed a large part of her talent to the
unsuspected guidance given by some one unknown to the public
which pressed to applaud her.
Miss Fotheringay was not intelligent like Rachel and she was far
duller than Mrs. Siddons, but she had in her the essential quality.
She was teachable and Little Bows taught her.

He shrieked out in his cracked voice the parts, and his pupil learned them from
him by rote. He indicated the attitudes, and set and moved those beautiful arms of
hers.... With what indomitable patience and dulness she followed him. She knew
that he made her; and she let herself be made.... She was not grateful, or
ungrateful, or unkind, or ill-humored.

She might not be grateful, but she knew very well who had made
her; she said so simply enough, explaining why she had not earlier
played the more important parts, “I didn’t take the leading business
then; I wasn’t fit for it till Bows taught me.”
So it was that Adrienne Lecouvreur, in the play which Scribe and
Legouvé wrote for Rachel, thanked the little old prompter, Michonnet,
who had taught her, “I was ungrateful in saying I had never had a
teacher. There is a kind-hearted man, a sincere friend, whose
counsels have always sustained me.” And Legouvé has told us that
at one of the rehearsals Rachel suddenly turned from Regnier, who
was the Michonnet, and knelt before Samson, who was the Duc de
Bouillon, and addressed this speech directly to him.
It would be interesting to know whether Thackeray ever saw
‘Adrienne Lecouvreur,’ which was produced in Paris in April, 1849,
six months before ‘Pendennis’ began to appear in monthly parts.
(1920)
 IX
MARK TWAIN AND THE THEATER
 IX
MARK TWAIN AND THE THEATER

Mark Twain was a born story-teller; he was a born actor; he was not
affrighted by the idea of facing an audience; he was fond of the
theater; he lived in a time when the drama was regaining its proud
position in our literature and when men of letters who had begun as
novelists were turning dramatists. Why is it that he did not leave us
even one play worthy to be set by the side of the ‘Adventures of
Huckleberry Finn’? Why is it that the only piece of his which was
successful on the stage, is a poor thing, not wholly his own? Why is
it that he did not persevere in playwriting as did his fellow humorists,
George Bernard Shaw and George Ade, and his fellow story-tellers,
Barrie and Tarkington?
These are questions which must have occurred to not a few of his
admirers; and they are questions to which it is not easy to find an
immediate answer. Yet there must be an explanation of some sort for
this puzzling fact; and there may be profit in trying to discover it.
Even if the answer shall prove to be incomplete and unsatisfactory,
the inquiry is worth while for its own sake.

II

That Mark Twain was a born story-teller needs no argument; and

that he was a born actor was equally evident not only to his few
intimates but to all the many who heard him talk on his feet. If any
witness must be called, the best would be Howells, his friend for forty
years; and Howells’s testimony is emphatic and decisive. He tells us
Mark
held that an actor doubled the value of the author’s words; and he was a great
actor as well as a great author. He was a most consummate actor, with this
difference from other actors, that he was the first to know the thoughts and invent
the fancies to which his voice and action gave the color of life. Representation is
the art of other actors; his art was creative as well as representative.

This quotation is from Howells’s introduction to the collection of

Mark’s speeches; and I take another from ‘My Mark Twain’:

He was the most consummate public performer I ever saw, and it was an
incomparable pleasure to hear him lecture; on the platform he was the great and
finished actor he probably would not have been on the stage.... When he read his
manuscript to you, it was with a thorough, however involuntary, recognition of its
dramatic qualities.... He was realistic, but he was essentially histrionic; and rightly
so. What we have strongly conceived we ought to make others strongly imagine,
and we ought to use every art to that end.

As a born actor, he understood the necessity of preparation and

rehearsal. He left nothing to chance. He knew how his effects ought
to be made; and he knew how to make them. Even his seemingly
spontaneous after-dinner speeches were thought out and worked
out, in every minutest detail of inflection and hesitation. In his ‘How
to Tell a Story’ he insisted that the total impression of his hair-raising
ghost-story, the ‘Golden Arm,’ depended upon the exact calculation
of a certain pause; and I can testify that on the only occasion I had
the pleasure of hearing him tell the gruesome tale—one summer
evening in 1890 at Onteora, in a cabin dimly lit by a flickering wood
fire—the pause was long enough to be almost unbearable.
He stood in no fear of an audience, because he had an
imperturbable self-confidence, rooted in a knowledge of his certain
power of impressing all who came within sound of his voice.
Moreover, he possessed to the end of his life the boyish delight in
being conspicuous that he ascribed to Tom Sawyer. It is true that he
was diffident before he had proved himself as a lecturer; and in a
little speech he made after a musical recital given by his daughter in
1906, he described his trepidation when he was about to make his
first appearance before an audience:

I had stage-fright then for the first and last time.... After the first agonizing five
minutes, my stage-fright left me, never to return. I know if I was going to be
hanged I could get up and make a good showing—and I intend to.

When he was living in Hartford he often took part in private

theatricals, the other performers being members of his own
household. After a performance of a dramatization of the ‘Prince and
the Pauper’ by the children of the Educational Alliance in 1907, he
was called upon for a speech and he told the thousand little
spectators that he had himself acted the part of Miles Hendon
twenty-two years earlier. One of his daughters had been the Prince
and the daughter of a neighbor was the Pauper. Mrs. Clemens was
the dramatist and stage-manager. “Our coachman was the assistant
stage-manager, second in command.”
He had many friends among stage-folk, authors, actors and
managers. He accepted the invitation to make the opening address
at the Actors’ Fund Fair in 1907. He lent William Gillette the money
which enabled that veracious actor to start his career. He once gave
a characteristically amusing account of his success in passing
through the sternly defended stage-entrance to Daly’s Theater. At a
dinner to Henry Irving in London in June, 1900, he declared that

the greatest of all arts is to write a drama. It is a most difficult thing. It requires the
highest talents possible and the rarest gifts. No, there is another talent that ranks
with it—for anybody can write a drama—I have written about four hundred—but to
get one accepted requires real ability. And I have never had that felicity yet.

He was a persistent playgoer, altho his visits to the theater were

less frequent in later life than they had been earlier. He took the
drama seriously, as he took the other facts of life; and he thought
that the American theater was not doing its duty by the American
people. In an illuminating article “About Play-Acting,” published in a
magazine in 1898 (and most unaccountably not included in any of
the volumes of his complete works) he described a tragedy which he
had seen at the Burg Theater in Vienna. Then he listed the shows on
exhibition in New York in a single week; and he drew a moral from
the contrast:
It is right and wholesome to have these light comedies and entertaining shows;
and I shouldn’t wish to see them diminished. But none of us is always in the
comedy spirit; we have our graver moods; they come to us all; the lightest of us
cannot escape them. These moods have their appetites,—healthy and legitimate
appetites—and there ought to be some way of satisfying them. It seems to me that
New York ought to have one theater devoted to tragedy. With her three millions of
population, and seventy outside millions to draw upon, she can afford it, she can
support it. America devotes more time, labor, money and attention to distributing
literary and musical culture among the general public than does any other nation,
perhaps; yet here you find her neglecting what is possibly the most effective of all
the breeders and nurses and disseminators of high literary taste and lofty emotion
—the tragic stage. To leave that powerful agency out is to haul the culture-wagon
with a crippled team. Nowadays when a mood comes which only Shakspere can
set to music, what must we do? Read Shakspere ourselves? Isn’t it pitiful? It is
playing an organ solo on a jews-harp. We can’t read. None but the Booths can do
it....
Comedy keeps the heart sweet; but we all know that there is wholesome
refreshment for both mind and heart in an occasional climb among the solemn
pomps of the intellectual snow-summits built upon by Shakspere. Do I seem to be
preaching? It is out of my line; I only do it because the rest of the clergy seem to
be on a vacation.

III

Altho I have quoted Mark’s assertion that he had never had the
felicity of having a play accepted, he did have two pieces produced
by managers; and a third, written in collaboration with Howells, had a
brief and inglorious career at the expense of its authors. His first
play, made out of one of his novels, drew delighted audiences for
several seasons; the second, written in partnership with Bret Harte,
and the third, written in partnership with Howells, met with so little
success that they sank at once beneath the wave of oblivion, being
almost unknown except in the hazy memories of the few surviving
spectators who chanced to see one or the other during its brief stay
on the stage. No one of the three has ever been published.
 After Mark had settled in Hartford he formed a close friendship
with his near neighbor Charles Dudley Warner; and in 1873 they
joined forces in a novel, the ‘Gilded Age.’ They wrote it not so much
in collaboration as in conjunction,—that is to say, each of the writers
was responsible for the chapters he prepared himself; and there was
no integrating co-ordination of their respective contributions. Mark
was the author of more than half of the chapters; and he was the
creator of the one outstanding character, Colonel Mulberry Sellers,
an imaginative reproduction of a man he had known since boyhood,
James Lampton. Mark began by writing the first eleven chapters,
then Warner wrote two, Mark followed with two more; and thus they
worked alternately. They labored, so Mark declared, “in the
superstition that we were writing one coherent yarn, when I suppose,
as a matter of fact, we were writing two incoherent yarns.”
It was not long after the publication of their conjoint work that they
were informed of the performance in San Francisco of a
dramatization by one Gilbert S. Densmore, otherwise unknown to
fame, the character of Colonel Sellers being impersonated by John
T. Raymond. Action was at once taken to put a stop to this
infringement on the copyright of the story. In the end a satisfactory
arrangement was arrived at. Densmore was bought out; Warner,
discovering that his share in the story had been but little drawn upon,
relinquished any claim he might have; Mark made the piece over;
and Raymond continued to play Colonel Sellers, under a contract
which divided the profits between the author and the actor. For a
season or two Mark’s agent travelled with the company and reported
on a postal card every night the author’s share; and Howells has
related how these welcome missives would come about dinnertime
and how Mark would read them aloud in triumph. “One hundred and
fifty dollars—two hundred dollars—three hundred dollars, were the
gay figures which they bore and which he flaunted in the air before
he sat down at table.”
It is difficult now to determine how much of the dramatic skeleton
Densmore had put together to enable Colonel Sellers to exhibit the
facets of his lovable character, survived in the play which drew
crowded houses one long winter in New York. Here Mark himself is
the best witness in his own behalf; and his biographer has quoted
from an unpublished letter a clear-cut statement:

I entirely rewrote the play three separate and distinct times. I had expected to
use little of [Densmore’s] language and but little of his plot. I do not think that there
are now twenty sentences of Mr. Densmore’s in the play, but I used so much of his
plot that I wrote and told him I should pay him about as much more as I had
already paid him, in case the play proved a success.

Paine has printed Densmore’s acknowledgment for this second

payment, thanking Mark “for the very handsome manner in which
you have acted in this matter.”
During the run of the play in New York in the winter of 1874-5 I
saw it twice, the second time on the hundredth performance, when
Mark appeared before the curtain to tell the audience the tale of the
man who tried to ride the Mexican plug and to explain that he was
like this man after his fiery steed had thrown him, in that he was
“speechless.” I recall the play as a rickety contrivance; it creaked in
its joints; its plot was arbitrary and violent and unconvincing. Perhaps
it was no worse than the earlier ‘Solon Shingle’ or the later ‘Mighty
Dollar’; but it was little, if any, better. Yet it served its purpose, which
was to be a frame for the humorously veracious character of Colonel
Sellers, the imperturbable visionary admirably acted by John T.
Raymond. Mark himself liked Raymond’s impersonation,—at least
he did at first. Later he and Raymond fell out; and he put into his
autobiography the assertion that Raymond was lacking in the ability
to express the finer qualities of Sellers.
But playgoers could see in the part only what Raymond has
expressed with the keenest appreciation of its histrionic possibilities;
and they were satisfied, even if the author was not. To us Americans
the character had a special appeal, because he represented at once
our ingenious inventiveness and our incurable optimism. We had
never met James Lampton, but we were all ready to accept Colonel
Sellers as an old friend. Raymond told me once that in town after
town he would be accosted by some man, who would say to him, “I
saw you to-night—and I recognized myself. Didn’t Mark ever tell
you? Well, he took Sellers from me! Why, all my friends knew me the
first time they saw you!”
The plot of the play was melodramatic on the verge of burlesque; it
called for the wholly unnecessary explosion of a steamboat; it
culminated in the trial of the injured heroine for the murder of the
villain who had wronged her and insulted her. For the most part
Colonel Sellers had little to do with the main story; and it was only
when the sympathetic heroine was on trial for her life that Colonel
Sellers was integrally related to the main action. I have revived my
own fading memory of the bubbling humor of this final act by reading
again what Howells wrote about it at the time:

But the greatest scenes are in the last act, where Colonel Sellers appears as a
witness for the defence of Laura Hawkins. As he mounts the stand he affably
recognizes and shakes hands with several acquaintances among the jury; he
delivers his testimony in the form of a stump speech; he helplessly overrides all
the protests, exceptions, and interruptions of the prosecution; from time to time he
irresistibly turns and addresses the jury and can scarcely be silenced; while the
attorneys are wrangling together he has seized a juryman by the coat-lapel and is
earnestly exhorting him in whisper. The effect is irresistibly ludicrous. It is farce and
not farce, for, however extravagantly impossible the situation is, the man in it is
deliciously true to himself. There is one bit of pathos, where Sellers tells how he
knew Laura as a little girl, and implies that, though she might have killed a man,
she could not have done murder.

The extravagantly impossible situation may have been taken over

from the Densmore perversion; but the handling of it, the expressing
out of it of all the humor it might be made to contain, that, we may be
sure, was the doing of Mark himself. No one else could have done it.
Forty years ago and more I pointed out, in an article on the
‘American on the Stage’ that in so far as Colonel Sellers was a
schemer, with an incessant activity in devising new methods for
making money, he had been anticipated by a character in Ben
Jonson’s the ‘Devil is an Ass’—added evidence of the kinship of the
descendants of the Puritans with the daring Elizabethan adventurers.
Where the American proposed a liniment for the sore eyes so
multitudinous in the Orient and saw “millions in it!” the Elizabethan
had advocated a device for making wine of raisins:
 What hast thou there?
O, “Making wine of Raisins”; this is in hand now.
Yes, and as true a wine as the wines of France,
Or Spain or Italy: look of what grape
My raisin is, that wine I’ll render perfect,
As of the Muscatel grape, I’ll render Muscatel;
Of the Canary, his; the claret, his;
So of all kinds; and bate you of the prices
Of wine throughout the kingdom half in half.
When it is objected that this enterprise may put up the price of
raisins, the answer comes pat:
Why then I’ll make it out of blackberries,
And it shall do the same. ’Tis but more art,
And the charge less.
There is a significant kinship between Ben Jonson and Mark Twain
in the superb impossibility of their towering fantasies. But there is no
true likeness between Meercraft, whose very name libels him as an
unscrupulous exploiter of the eternal gullibility of mankind, and
Colonel Sellers, who may have deceived others but who did so only
because he had first deceived himself. Colonel Sellers was a man
without guile; he was as sincere as he was frank; and he made no
more profit out of his swift succession of vain imaginings than did
those who were carried away by his magnificent self-confidence. The
similarity between Ben Jonson’s crook and Mark’s enthusiast is only
superficial; yet it may be worth noting that frenzied speculation was
as characteristic of the golden age of England after the dispersal of
the Armada as it was in the gilded age of America which was the
aftermath of the Civil War. Moreover Ben Jonson and Mark Twain
have this in common also, that they were both of them humorists of
soaring exuberance and both of them realists of immitigable veracity.

IV

In the dramatization of the ‘Gilded Age’ Mark had a silent partner,

the otherwise unknown Densmore. In the two other plays of his he
was working in collaboration with associates of an assured fame,
Howells and Bret Harte. In neither case was he fortunate in the
alliance, for they were not experts in stage-craft, altho each of them
had already ventured himself in the drama. What Mark needed, if he
was to trot in double harness with a running mate, was an
experienced playwright with an instinctive knowledge of the theater.
When Mark yoked himself with Howells or with Harte, it was the blind
leading the blind. The author of ‘Out of the Question’ and the author
of ‘Two Men of Sandy Bar’ lacked just what the author of the ‘Gilded
Age’ lacked,—practice in the application of the principles of
playmaking.
The play written in collaboration with Bret Harte was called ‘Ah
Sin,’ the name of the Heathen Chinee in ‘Plain Language from
Truthful Jones.’ It was undertaken to enable Charles Parsloe, an
actor now forgotten, to profit by the skill he had displayed in the
small part of a Chinaman in Bret Harte’s earlier play, ‘Two Men of
Sandy Bar,’ written for Stuart Robson, brought out in 1876 and
withdrawn after a brief and inglorious career on the stage. Bret Harte
did not know enough about playmaking to perceive that its failure
had been due to its deficiency in that supporting skeleton of plot
which is as necessary to a drama as the equally invisible steel-frame
is to a skyscraper. But he was eager to try again, and he persuaded
Mark to join him. Probably he had no need to be persuasive, since
Mark had found his experience with the ‘Gilded Age’ exhilarating and
profitable. Mark invited Harte to Hartford and they set to work. As I
have always been curious about the secrets of collaboration, I asked
Mark many years afterward, how they had gone about it. “Well,” he
said, with his customary drawl, “Bret came to me at Hartford and we
talked the whole thing out. Then Bret wrote the piece while I played
billiards. Of course, I had to go over it and get the dialect right. Bret
never did know anything about dialect.”
Mr. Paine, to whom I transmitted this information, thinks that it is
“scarcely a fair statement of the case,” since “both authors worked
on the play and worked hard.” But while what Mark said to me may
have been an over-statement, I doubt if it was a misstatement. The
original suggestion had come from Harte; and the probability is that
the major part of the story was his also. The two partners may have
worked hard but I doubt if they worked as seriously at their
playmaking as they were wont to do at their story-telling. The man of
letters who is not primarily a man of theater, is prone to be somewhat
contemptuous in his condescending to the drama.
The play was produced in Washington in May, 1877, with Parsloe
as Ah Sin. I saw it when it was brought to New York in the fall of
1877. From two of the foremost writers in America much was
expected; and the result of their combined efforts was lamentably
disappointing. It was unworthy of either of them, still more unworthy
of both. All I can replevin from my dim recollections is a trial before
Judge Lynch, which lit up the last act, and which I now recall as
having more than a little of the energy and the vigor which I found
afterward in the episode of the attempted lynching in ‘Huckleberry
Finn.’ Mr. Paine tells me that the manuscript is still extant. Sooner or
later it ought to be published, since nothing written by either Mark
Twain or Bret Harte is negligible.
Yet this flat failure of ‘Ah Sin’ did not quench Mark’s dramatic
ardor. Even before the ‘Gilded Age’ had been dramatized he had
begun on ‘Tom Sawyer’; and his first intention was to write it as a
play. Fortunately for us he soon perceived that Tom would have
more freedom if his adventures were narrated. After Mark had
published ‘Tom Sawyer’ he was fired with another dramatic idea; and
he wrote Howells in the first flush of his enthusiasm, that he was
deep in a comedy with an old detective as the principal character:
I skeletoned the first act, and wrote the second to-day, and am dog-tired now.
Fifty-four pages of ms. in seven hours.

A few days later he wrote again, telling his friend that he had

piled up one hundred and fifty-one pages. The first, second and fourth acts are
done, and done to my satisfaction too. Never had so much fun over anything in my
life—such consuming interest and delight.

This piece was intended for Sol Smith Russell. But the theatrical
experts to whom it was submitted did not share its author’s
consuming interest. Dion Boucicault said that it was better than ‘Ah
Sin’; but to say this was saying little. John Brougham wrote that it
was “altogether too diffuse for dramatic representation.” In time
Mark’s own opinion of his play seems to have cooled, and he put his
manuscript aside. Possibly he utilized it more or less many years
later when he wrote ‘Tom Sawyer, Detective’; but this is mere
conjecture.
Then, after a longer interval he asked Howells to collaborate with
him in a sequel to Colonel Sellers; and in ‘My Mark Twain’ Howells
has given a detailed account of their conjoint misadventure. Mark
had a host of suggestions but no story, so Howells supplied one as
best he could; and the two friends spent a hilarious fortnight in
writing the play. Mark had quarrelled with Raymond and did not want
to let him reincarnate Sellers; and yet he had ultimately to recognize
that Raymond was the only actor the public would accept in the
character. So the piece was sent to Raymond, who accepted it,
asking for certain alterations; and then most unexpectedly he
returned the manuscript, refusing to have anything to do with it. After
hawking their play about, the authors arranged to produce it
themselves with Burbank (who was not an actor but an elocutionist-
entertainer) as Sellers,—Burbank playing the part in imitation of
Raymond. At last they had lost confidence in it so completely that
they paid a forfeit rather than undertake the risk of a production in
New York. So it was that the ‘American Claimant, or Mulberry Sellers
Ten Years Later’ was made visible in New York only at a special
matinee in the fall of 1887. It had a few performances in unimportant
out of town theaters; and then it disappeared from the stage. Still, it
had not lived in vain since it supplied material for several chapters in
Mark’s later novel, to which he gave the same title, without the
subtitle.
After this play had been withdrawn from the boards Mark’s
ambition to establish himself as a dramatist did not again manifest
itself. However, it is pleasant to believe that the pain of his own
failure may have been more or less assuaged by the better fortune
of dramatizations of two of his novels.
I have already noted that not long after the publication of the
‘Prince and the Pauper’ Mrs. Clemens had arranged scenes from it
to be acted by members of the family and by their young friends, and
that Mark himself had undertaken the part of Miles Hendon. A little
later a dramatization of the whole story was made by Abby Sage
Richardson; and this was produced in New York in January, 1890. It
achieved instant popularity, as well it might, since the story is
indisputably dramatic and since it has a more direct action than any
other of Mark’s novels. This version, revised by Amélie Rives, was
revived in 1920 by William Faversham, who appeared as Miles
Hendon. The revival met with a reception as warm as that which had
greeted the original production.
In one respect this professional dramatization was inferior to Mrs.
Clemens’s amateur arrangement; it was so devised that one
performer should assume two characters, the little Prince and the
little Pauper; and this necessitated the omission of the culminating
moment in the tale when the Prince and the Pauper stand face to
face. And in both the amateur and the professional performances
these two lads were impersonated by girls. This may have been
necessary, since it is almost impossible to find competent boy actors,
while there are girl actors aplenty; but none the less was it
unfortunate, since a girl is never entirely satisfactory in boy’s clothes.
Very rarely can she conceal from us the fact that she is a girl, doing
her best to be a boy. Curiously enough, boys can act girls’ parts and
make us forget for the moment that they are not what they seem.
Five years after Mrs. Richardson had dramatized the ‘Prince and
the Pauper,’ Frank Mayo made a most effective play out of
‘Pudd’nhead Wilson.’ He arranged the title-part for his own vigorous
and impressive acting. He simplified Mark’s story and he amplified it;
he condensed it and he heightened it; he preserved the ingenious
incidents and the veracious characters; he made his profit out of the
telling dialog; and he was skilful in disentangling the essentially
dramatic elements of Mark’s rather rambling story. He produced it in
New York in the spring of 1895. Mark was then in Europe; but when
he returned he made haste to see the piece. He was discovered by
the audience and called upon for a speech, in which he
congratulated the player-playwright on a “delightful play.” He ended
by saying, “Confidentially I have always had an idea that I was well
equipt to write plays, but I have never encountered a manager who
has agreed with me”—which was not strictly accurate since two
different managers had accepted the ‘Gilded Age’ and ‘Ah Sin.’

When the ‘Gilded Age’ was brought out in New York in the fall of
1874, Mark climbed the eighty steps which led to the editorial offices
of the New York World, then in the control of Manton Marble. He
asked for the city editor and he was shown into the cubicle occupied
by William C. Brownell. He explained that he had come to ask the
editor to puff his play; whereupon Brownell inquired if it was a good
play. “No,” was Mark’s drawling answer, “it isn’t a good play. It’s a
bad play, a damned bad play. I couldn’t write a good play. But it has
a good character. I can write character; and that character is the best
I can do. If it was a good play, I shouldn’t have had to climb up here
to ask you to puff it.”
Here Mark was unconsciously revealing his agreement with
Aristotle, the master of all who know. Aristotle declared that in a
tragedy—and the remark is even more applicable to comedy—plot is
more important than character, since you can have an appealing
drama without character but you cannot have it without plot. Lowell
said the same thing in more detail, in one of his lectures on the ‘Old
English Dramatists.’

In a play we not only expect a succession of scenes, but that each scene should
lead by a logic more or less stringent, if not to the next, at any rate to something
that is to follow and that all should contribute their fraction of impulse to the
inevitable catastrophe. That is to say, the structure should be organic, with a
necessary and harmonious connection and relation of parts, and not merely
mechanical with an arbitrary or haphazard joining of one part to another.

It was this constructive skill that Mark lacked. He could create

characters; he could make them reveal themselves in appropriate
situations; he could carry on a story which in the library would delight
all of us, but which was without the compact directness demanded
by us when we are in the theater. He possessed all the qualifications
of the dramatist except the one thing needful, without which the rest