Ten Lectures on Statistical and Structural Pattern Recognition

531 Pages·2002·17.018 MB·English

by  Michail I. Schlesinger, Václav Hlaváč (auth.)

Ten Lectures on Statistical and Structural Pattern Recognition
Computational Imaging and Vision
Managing Editor
MAX A. VIERGEVER
Utrecht University, Utrecht, The Netherlands
Editorial Board
RUZENA BAJCSY, University ofP ennsylvania, Philadelphia, USA
MIKE BRADY, Oxford University, Oxford, UK
OLIVIER D. FAUGERAS, INRIA, Sophia-Antipolis, France
JAN J. KOENDERINK, Utrecht University, Utrecht, The Netherlands
STEPHEN M. PIZER, University ofN orth Carolina, Chapel Hill, USA
SABURO TSUJI, Wakayama University, Wakayama, Japan
STEVEN W. ZUCKER, McGill University, Montreal, Canada
Volume24
Ten Lectures on Statistical
and Structural Pattern Recognition
by
Michaill. Schlesinger
Ukranian Academy of Sciences,
Kiev, Ukraine
and
Vaclav Hlavac
Czech Technical University,
Prague, Czech Republic
SPRINGER-SCIENCE+BUSINESS MEDIA, B.V.
A C.I.P. Catalogue record for this book is available from the Library of Congress.
ISBN 978-90-481-6027-3 ISBN 978-94-017-3217-8 (eBook)
DOI 10.1007/978-94-017-3217-8
This is a completely revised and updated translation of Deset prednasek z teorie
statistickeho a strukturniho rozpoznavani, by M.I. Schlesinger and V. Hlavac. Published
by Vydavatelstvi CVUT, Prague 1999. Translated by the authors. This book was typeset
by Vit Zyka in Latex using Computer Modern 10/12 pt.
Printed on acid-free paper
AII Rights Reserved
© 2002 Springer Science+Business Media Dordrecht
Origina\ly published by Kluwer Academic Publishers in 2002
Softcover reprint ofthe hardcover Ist edition 2002
No part of this work may be reproduced, stored in a retrieval system, or transmitted
in any form or by any means, electronic, mechanical, photocopying, microfilming, recording
or otherwise, without written permission from the Publisher, with the exception
of any material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work.
Contents
Preface xi
Preface to the English edition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
A letter from the doctoral student Jifi Pecha prior to publication of
the lectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
A letter from the authors to the doctoral student Jifi Pecha . . . . . . . xiii
Basic concepts and notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
Acknowledgements xix
Lecture 1 Bayesian statistical decision making 1
1.1 Introduction to the analysis of the Bayesian task . . . . . . . . . . . . 1
1.2 Formulation of the Bayesian task . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Two properties of Bayesian strategies . . . . . . . . . . . . . . . . . . . . 3
1.4 Two particular cases of the Bayesian task . . . . . . . . . . . . . . . . . 7
1.4.1 Probability of the wrong estimate of the state . . . . . . . . . 7
1.4.2 Bayesian strategy with possible rejection . . . . . . . . . . . . . 9
1.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Bibliographical notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Lecture 2 Non-Bayesian statistical decision making 25
2.1 Severe restrictions of the Bayesian approach . . . . . . . . . . . . . . 25
2.1.1 Penalty function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1.2 A priori probability of situations . . . . . . . . . . . . . . . . . 26
2.1.3 Conditional probabilities of observations . . . . . . . . . . . . 27
2.2 Formulation of the known and new non-Bayesian tasks. . . . . . . 28
2.2.1 Neyman-Pearson task . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.2 Generalised task with two dangerous states . . . . . . . . . . 31
2.2.3 Minimax task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2.4 Wald task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.5 Statistical decision tasks with non-random interventions . 33
v
vi Contents
2.3 The pair of dual linear programming tasks, properties and
solutions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.4 The solution of non-Bayesian tasks using duality theorems . . . . 40
2.4.1 Solution of the Neyman-Pearson task . . . . . . . . . . . . . . 41
2.4.2 Solution of generalised Neyman-Pearson task with two
dangerous states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4.3 Solution of the minimax task . . . . . . . . . . . . . . . . . . . . 46
2.4.4 Solution of Wald task for the two states case . . . . . . . . . 48
2.4.5 Solution of Wald task in the case of more states . . . . . . . 50
2.4.6 Testing of complex random hypotheses . . . . . . . . . . . . . 52
2.4. 7 Testing of complex non-random hypotheses . . . . . . . . . . 53
2.5 Comments on non-Bayesian tasks . . . . . . . . . . . . . . . . . . . . . . 53
2.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2. 7 Bibliographical notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Lecture 3 Two statistical models of the recognised object 73
3.1 Conditional independence of features . . . . . . . . . . . . . . . . . . . 73
3.2 Gaussian probability distribution . . . . . . . . . . . . . . . . . . . . . . 75
3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.4 Bibliographical notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Lecture 4 Learning in pattern recognition 101
4.1 Myths about learning in pattern recognition .............. 101
4.2 Three formulations of learning tasks in pattern recognition . . . . 102
4.2.1 Learning according to the maximal likelihood ........ 104
4.2.2 Learning according to a non-random training set ...... 105
4.2.3 Learning by minimisation of empirical risk ........... 106
4.3 Basic concepts and questions of the statistical theory of learning 108
4.3.1 Informal description of learning in pattern recognition .. 108
4.3.2 Foundations of the statistical learning theory according to
Chervonenkis and Vapnik ............................ 113
4.4 Critical view of the statistical learning theory . . . . . . . . . . . . . 120
4.5 Outlines of deterministic learning . . . . . . . . . . . . . . . . . . . . . . 122
4.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4. 7 Bibliographical notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Lecture 5 Linear discriminant function 137
5.1 Introductory notes on linear decomposition ............... 137
5.2 Guide through the topic of the lecture . . . . . . . . . . . . . . . . . . . 138
5.3 Anderson tasks ................................... 141
5.3.1 Equivalent formulation of generalised Anderson task .... 141
5.3.2 Informal analysis of generalised Anderson task ........ 142
5.3.3 Definition of auxiliary concepts for Anderson tasks ..... 145
5.3.4 Solution of Anderson original task ................. 147
5.3.5 Formal analysis of generalised Anderson task ......... 150
Contents VII
5.3.6 Outline of a procedure for solving generalised Anderson
task ........................................... 157
5.4 Linear separation of finite sets of points . . . . . . . . . . . . . . . . . 159
5.4.1 Formulation of tasks and their analysis ............. 159
5.4.2 Algorithms for linear separation of finite sets of points .. 163
5.4.3 Algorithm for c:-optimal separation of finite sets of points
by means of the hyperplane .......................... 167
5.4.4 Construction of Fisher classifiers by modifying Kozinec
and perceptron algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
5.4.5 Further modification of Kozinec algorithms . . . . . . . . . . 171
5.5 Solution of the generalised Anderson task ................ 175
5.5.1 c:-solution of Anderson task ..................... 175
5.5.2 Linear separation of infinite sets of points ........... 179
5.6 Discussion ...................................... 182
5.7 Link to a toolbox ................................. 213
5.8 Bibliographical notes ............................... 213
Lecture 6 Unsupervised learning 215
6.1 Introductory comments on the specific structure of the lecture 215
6.2 Preliminary and informal definition of unsupervised learning . . . 217
6.3 Unsupervised learning in a perceptron ................... 219
6.4 Empirical Bayesian approach after H. Robbins ............. 226
6.5 Quadratic clustering and formulation of a general clustering task 232
6.6 Unsupervised learning algorithms and their analysis ......... 238
6.6.1 Formulation of a recognition task ................. 238
6.6.2 Formulation of a learning task ................... 238
6.6.3 Formulation of an unsupervised learning task ......... 240
6.6.4 Unsupervised learning algorithm .................. 241
6.6.5 Analysis of the unsupervised learning algorithm ....... 242
6.6.6 Algorithm solving Robbins task and its analysis ....... 251
6. 7 Discussion ...................................... 253
6.8 Link to a toolbox ................................. 273
6.9 Bibliographical notes ......................... ; ..... 274
Lecture 7 Mutual relationship of statistical and structural
recognition 275
7.1 Statistical recognition and its application areas ............ 275
7.2 Why is structural recognition necessary for image recognition? . 277
7.2.1 Set of observations ............................ 277
7.2.2 Set of hidden parameter values for an image ......... 280
7.2.3 The role of learning and unsupervised learning in image
recognition ...................................... 281
7.3 Main concepts necessary for structural analysis ............ 284
7.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
7.5 Bibliographical notes ............................... 305
VIII Contents
Lecture 8 Recognition of Markovian sequences 307
8.1 Introductory notes on sequences ....................... 307
8.2 Markovian statistical model of a recognised object .......... 308
8.3 Recognition of the stochastic automaton ................. 312
8.3.1 Recognition of the stochastic automaton; problem
formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
8.3.2 Algorithm for a stochastic automaton recognition ...... 313
8.3.3 Matrix representation of the calculation procedure ..... 314
8.3.4 Statistical interpretation of matrix multiplication ...... 316
8.3.5 Recognition of the Markovian object from incomplete data 318
8.4 The most probable sequence of hidden parameters .......... 321
8.4.1 Difference between recognition of an object as a whole
and recognition of parts that form the object ............. 321
8.4.2 Formulation of a task seeking the most probable sequence
of states ........................................ 321
8.4.3 Representation of a task as seeking the shortest path in
a graph ........................................ 321
8.4.4 Seeking the shortest path in a graph describing the task . 323
8.4.5 On the necessity of formal task analysis . . . . . . . . . . . . . 326
8.4.6 Generalised matrix multiplications ................ 327
8.4. 7 Seeking the most probable subsequence of states . . . . . . 330
8.5 Seeking sequences composed of the most probable hidden
parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
8.6 Markovian objects with acyclic structure . . . . . . . . . . . . . . . . . 338
8.6.1 Statistical model of an object .................... 338
8.6.2 Calculating the probability of an observation . . . . . . . . . 339
8.6.3 The most probable ensemble of hidden parameters ..... 343
8. 7 Formulation of supervised and unsupervised learning tasks . . . . 344
8.7.1 The maximum likelihood estimation of a model during
learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
8.7.2 Minimax estimate of the model ................... 345
8.7.3 Tuning of the recognition algorithm ............... 346
8.7.4 Task of unsupervised learning .................... 347
8.8 Maximum likelihood estimate of the model ............... 347
8.9 Minimax estimate of a statistical model ................. 353
8.9.1 Formulation of an algorithm and its properties ........ 353
8.9.2 Analysis of a minimax estimate .................. 356
8.9.3 Proof of the minimax estimate algorithm of a Markovian
mode .......................................... 366
8.10 Tuning the algorithm that recognises sequences ............ 366
8.11 The maximum likelihood estimate of statistical model ........ 368
8.12 Discussion ...................................... 372
8.13 Link to a toolbox ................................. 395
8.14 Bibliographical notes ............................... 395
Contents IX
Lecture 9 Regular languages and corresponding pattern
recognition tasks 397
9.1 Regular languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
9.2 Other ways to express regular languages ................. 399
9.2.1 Regular languages and automata ................. 399
9.2.2 Regular languages and grammars ................. 400
9.2.3 Regular languages and regular expressions ........... 401
9.2.4 Example of a regular language expressed in different ways 402
9.3 Regular languages respecting faults; best and exact matching .. 404
9.3.1 Fuzzy automata and languages ................... 405
9.3.2 Penalised automata and corresponding languages ...... 406
9.3.3 Simple best matching problem ................... 407
9.4 Partial conclusion after one part of the lecture ............. 409
9.5 Levenstein approximation of a sentence .................. 410
9.5.1 Preliminary formulation of the task ................ 410
9.5.2 Levenstein dissimilarity ........................ 411
9.5.3 Known algorithm calculating Levenstein dissimilarity ... 412
9.5.4 Modified definition of Levenstein dissimilarity and its
properties ....................................... 414
9.5.5 Formulation of the problem and comments to it ....... 417
9.5.6 Formulation of main results and comments to them .... 418
9.5.7 Generalised convolutions and their properties ......... 420
9.5.8 Formulation of a task and main results in convolution
form ........................................... 427
9.5.9 Proof of the main result of this lecture ............. 429
9.5.10 Nonconvolution interpretation of the main result ...... 440
9.6 Discussion ...................................... 443
9.7 Link to a toolbox ................................. 477
9.8 Bibliographical notes ............................... 477
Lecture 10 Context-free languages, their 2-D generalisation,
related tasks 4 79
10.1 Introductory notes ................................. 479
10.2 Informal explanation of two-dimensional grammars and languages 480
10.3 Two-dimensional context-free grammars and languages ....... 484
10.4 Exact matching problem. Generalised algorithm of C-Y- K .... 486
10.5 General structural construction ....................... 489
10.5.1 Structural construction defining observed sets ........ 490
10.5.2 Basic problem in structural recognition of images ...... 493
10.5.3 Computational procedure for solving the basic problem . 494
10.6 Discussion ...................................... 498
10.7 Bibliographical notes ............................... 505
Bibliography 507
Index 514
Preface
Preface to the English edition
This monograph Ten Lectur,es on Statistical and Structural Pattern Recognition
uncovers the close relationship between various well known pattern recognition
problems that have so far been considered independent. These relationships
became apparent when formal procedures addressing not only known prob
lems but also their generalisations were discovered. The generalised problem
formulations were analysed mathematically and unified algorithms were found.
The book unifies of two main streams ill pattern recognition-the statisti
cal a11d structural ones. In addition to this bridging on the uppermost level,
the book mentions several other unexpected relations within statistical and
structural methods.
The monograph is intended for experts, for students, as well as for those
who want to enter the field of pattern recognition. The theory is built up
from scratch with almost no assumptions about any prior knowledge of the
reader. Even when rigorous mathematical language is used we make an effort
to keep the text easy to comprehend. This approach makes the book suitable
for students at the beginning of their scientific career. Basic building blocks are
explained in a style of an accessible intellectual exercise, thus promoting good
practice in reading mathematical text. The paradoxes, beauty, and pitfalls
of scientific research are shown on examples from pattern recognition. Each
lecture is amended by a discussion with an inquisitive student that elucidates
and deepens the explanation, providing additional pointers to computational
procedures and deep rooted errors.
We have tried to formulate clearly and cleanly individual pattern recognition
problems, to find solutions, and to prove their properties. We hope that this
approach will attract mathematically inclined people to pattern recognition,
which is often not the case if they open a more practically oriented literature.
The precisely defined domain and behaviour of the method can be very substan
tial for the user who creates a complicated machine or algorithm from simpler
modules.
XI