Yanushkevich S.N., Wang P.S.P., Gavrilova M.L., Srihari S.N. (eds.) Image Pattern Recognition. Synthesis and Analysis in Biometrics
Подождите немного. Документ загружается.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
318 Carlos A. Coello Coello
Adaptive Systems. PhD thesis, University of Michigan.
61. De Jong, K.A. (1993). Genetic Algorithms are NOT Function Optimizers.
In L. D. Whitley, editor, Foundations of Genetic Algorithms 2, pp. 5–17.
Morgan Kaufmann Publishers, San Mateo, California.
62. Kennedy, J. and Eberhart, R.C. (1995). Particle Swarm Optimization. Proc.
IEEE Int. Conf. Neural Networks, pp. 1942–1948, IEEE Service Center,
Piscataway, New Jersey.
63. Kennedy, J. and Eberhart, R.C. (1997). A Discrete Binary Version of
the Particle Swarm Algorithm. Proc. IEEE Conf. Systems, Man, and
Cybernetics, pp. 4104–4109, IEEE Service Center, Piscataway, New Jersey.
64. Kennedy, J. and Eberhart, R.C. (2001). Swarm Intelligence. Morgan
Kaufmann Publishers, San Francisco, California.
65. Kharma, N., Suen, C.Y. and Guo, P.F. (2003). PalmPrints: A Novel
Co-evolutionary Algorithm for Clustering Finger Images. In E. Cant´u-Paz
(editor) 2003 Genetic and Evolutionary Computation Conference, pp. 322–
331, Springer, Lecture Notes in Computer Science Vol. 2723, Chicago, Illinois.
66. Koza, J.R. (1992). Genetic Programming. On the Programming of Computers
by Means of Natural Selection. The MIT Press, Cambridge, Massachusetts.
67. Kung, S.Y., Mak, W.M. and Lin, S.H. (2004). Biometric Authentication: A
Machine Learning Approach. Prentice Hall.
68. Leung, K.F., Leung, F.H.F., Lam, H.K. and Ling, S.H. (2004). On
interpretation of graffiti digits and characters for ebooks: neural-fuzzy
network and genetic algorithm approach. IEEE Trans. Industrial Electronics,
51, 2, pp. 464–471.
69. Michalewicz, Z. (1996). Genetic Algorithms + Data Structures = Evolution
Programs. Third edition. Springer.
70. Miettinen, K.M. (1998). Nonlinear Multiobjective Optimization.Kluwer.
71. Mitchell, M. (1996). An Introduction to Genetic Algorithms. The MIT Press,
Cambridge, Massachusetts.
72. Morita, M.E., Sabourin, R., Bortolozzi, F., and Suen, C.Y. (2003).
Unsupervised feature selection using multi-objective genetic algorithm for
handwritten word recognition. Proc. 7th Int. Conf. Document Analysis and
Recognition, pp. 666–670, Edinburgh, Scotland.
73. Nunes de Castro, L. and Timmis, J. (2002). Artificial Immnue System: A
New Computational Intelligence Approach. Springer.
74. Nunes de Castro, L. and Von Zuben, F.J. (2002). Learning and
Optimization Using the Clonal Selection Principle. IEEE Trans. Evolutionary
Computation, 6, 3, pp. 239–251.
75. Osyczka, A. (2002). Evolutionary Algorithms for Single and Multicriteria
Design Optimization. Physica Verlag.
76. Price, K.V. An Introduction to Differential Evolution. In D. Corne, M. Dorigo
and F. Glover, Eds., New Ideas in Optimization, pp. 79–108, McGraw-Hill,
London, UK.
77. Qi, Y., Tian, J. and Dai, R.-W. (1998). Fingerprint classification system
with feedback mechanism based on genetic algorithm. Proc. 14th Int. Conf.
Pattern Recognition, Vol. 1, pp. 163–165, IEEE.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Evolutionary Algorithms: Basic Concepts and Applications in Biometrics 319
78. Rao, S.S. (1996). Engineering Optimization. Theory and Practice. 3rd edition.
John Wiley & Sons.
79. Rechenberg, I. (1973). Evolutionsstrategie: Optimierung technischer Systeme
nach Prinzipien der biologischen Evolution. Frommann–Holzboog, Stuttgart,
Germany.
80. Reinders, M.J.T., Sankur, B., and van der Lubbe, J.C.A. (1992).
Tranformation of a general 3D facial model to actual scene face. Proc. 11th
IAPR Int. Conf. Pattern Recognition, Vol. III, C: Image, Speech and Signal
Analysis, pp. 75–78.
81. Rothlauf, F. (2002). Representations for Genetic and Evolutionary
Algorithms. Physica-Verlag, New York.
82. Rudolph, G. (1994). Convergence Analysis of Canonical Genetic Algorithms.
IEEE Trans. Neural Networks, 5, pp. 96–101.
83. Schwefel, H.-P. (1981). Numerical Optimization of Computer Models. Wiley,
Chichester, UK.
84. Soldek, J., Shmerko, V., Phillips, P., Kukharev, G., Rogers, W.,
and Yanushkevich, S. (1997). Image Analysis and Pattern Recognition
in Biometric Technologies. Proc. Int. Conf. on the Biometrics: Fraud
Prevention, Enhanced Service, pp. 270–286, Las Vegas, Nevada.
85. Stadnyk, I. (1987). Schema Recombination in a Pattern Recognition Problem.
In J.J. Grefenstette, Ed., Genetic Algorithms and Their Applications:
Proc. 3nd Int. Conf. Genetic Algorithms, pp. 27–35, Lawrence Erlbraum
Associates, Hillsdale, New Jersey.
86. Sweldens, W. (1996). The lifting scheme: A custom-design construction of
biorthogonal wavelets. J. Applied and Computational Harmonic Analysis, 3,
2, pp. 186–120.
87. Syswerda, G. (1989). Uniform Crossover in Genetic Algorithms. In J.D.
Schaffer, editor, Proc. 3rd Int. Conf. on Genetic Algorithms, pp. 2–9, Morgan
Kaufmann Publishers, San Mateo, California.
88. Talay, A.C. (2005). An Approach for Eye Detection Using Parallel Genetic
Algorithm. In V.S. Sunderam, G.D. van Albada, P.M.A. Sloot, and J.J.
Dongarra, Eds., (2005). Computational Science: Proc. 5th Int. Conf., Part
III, pp. 1004, Springer. Lecture Notes in Computer Science Vol. 3516,
Atlanta, Georgia.
89. Teller, A. and Veloso, M. (1995). Algorithm Evolution for Face Recognition:
What Makes a Picture Difficult. Int. Conf. on Evolutionary Computation,
pp. 608–613, IEEE, Perth, Australia.
90. Teredasai, A., Park, J., and Govindaraju, V. (2001). Active Handwritten
Character Recognition Using Genetic Programming. In J. Miller, M.
Tomassini, P. L. Lanzi, C. Ryan, A.G.B. Tettamanzi, and W.B. Langdon,
editors, (2001), Genetic Programming. 4th European Conf., pp. 371–379,
Springer, Lake Como, Italy.
91. Trellue, R.E. (1973). The Recognition of Handprinted Characters Through
Evolutionary Programming. Master’s thesis, New Mexico State University,
Las Cruces, New Mexico.
92. Whitley, D. (1989). The GENITOR Algorithm and Selection Pressure: Why
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
320 Carlos A. Coello Coello
Rank-Based Allocation of Reproductive Trials is Best. In J.D. Schaffer,
editor, Proc. 3rd Int. Conf. on Genetic Algorithms, pp. 116–121. Morgan
Kaufmann Publishers, San Mateo, California.
93. Wilson, S.W. (1985). Adaptive “Cortical” Pattern Recognition. In J.J.
Grefenstette, Ed., Proc. 1st Int. Conf. Genetic Algorithms and Their
Applications, pp. 188–196, Lawrence Erlbraum Associates, Hillsdale, New
Jersey.
94. Yanushkevich, S. N., Stoica, A., Shmerko, V. P. and Popel, D. V. (2005).
Biometric Inverse Problems, Taylor & Francis/CRC Press.
95. Yokoo, Y. and Hagiwara, M. (1996). Human faces detection method using
genetic algorithm. Proc. IEEE Int. Conf. Evolutionary Computation, pp.
113–118.
96. Yu, E. and Cho, S. (2004). Keystroke dynamics identity verification-its
problems and practical solutions. Computers & Security, 23, 5, pp. 428–440.
97. Zhang, D.D. (2000). Automated Biometrics: Technologies and Systems.
Kluwer.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Chapter 13
Some Concerns on the Measurement for
Biometric Analysis and Applications
PatrickS.P.Wang
Image Processing Group,
College of Computer and Information Science,
Northeastern University, Boston, MA02115, USA
pwang@ccs.neu.edu
www.ccs.neu.edu/home/pwang
Some concerns of measurement for biometric analysis and synthesis
are investigated. This research tries to reexamine the nature of the
basic definition of “measurement” of distance between two objects or
image patterns, which is essential for comparing the “similarity” of
patterns. According to a recent International Workshop on Biometric
Technologies: Modeling and Simulation at University of Calgary, June,
2004, Canada, biometric refers to the studies of analysis, synthesis,
modeling and simulation of human behavior by computers, including
mainly recognition of hand printed words, machines printed characters,
handwriting, fingerprint, signature, facial expression, speech, voice,
emotion and iris etc.
The key idea is the “measurement” that defines the similarity
between different input data that can be represented by image data. This
paper deals with the very fundamental phenomena of “measurement” of
these studies and analysis. Preliminary findings and observations show
that the concepts of “segmentation” and “disambiguation” are extremely
important, which have been long ignored. Even while computer and
information professionals and researchers have spent much effort, energy,
and time, trying very hard and diligently to develop methods that
may reach as high as 99.9999% accuracy rate for character and symbol
recognition, a poorly or ill considered pre-designed board poster or
input pattern could easily destroy its effectiveness and lower the overall
performance accurate rate to less than 50%. The more data it handles,
the worse the results. Its overall performance accuracy rate will be
proportionally decreasing. Take road safety for example. If street
direction signs are poorly designed, then even a most intelligent robot
driver with perfect vision and 100% accurate rate of symbol recognition,
321
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
322 Synthesis and Analysis in Biometrics
still has at least 50% error rate. That is, it can achieve at most only
50% accuracy rate in determining which correct direction to follow. It’s
not just a matter of being slow or losing time.
However, in more serious and urgent life threatening situation like
fire, accidents, terrorists attack or natural disaster, it’s a matter of life
and death. So the impact is enormous and widespread. The idea
and concept of “ambiguity,” “disambiguation,” “measurement,” and
“learning,” and their impacts on biometrics image pattern analysis and
recognition, as well their applications are illustrated through several
examples.
Contents
13.1. Overview(Introduction) ............................ 322
13.2. The Problem of “Ambiguity,” and Its Solutions . . . . . . . . . . . . . . . 325
13.3. DiscussionsandConclusions .......................... 335
Bibliography ....................................... 336
13.1. Overview (Introduction)
We first briefly review what’s happening in biometric research and their
results. According to
[
5
]
, a majority of biometric studies fall into the
categories of fingerprint, facial and symbol analysis and recognition, about
74%. It can be shown in the Fig. 13.1.
Finger Scan
52.1%
Keystroke Scan
0.3%
Facial Scan
11.4%
Hand Scan
10.0%
Middleware
12.4%
Iris Scan
7.3%
Voice Scan
4.1%
Signature Scan
2.4%
Fig. 13.1. Percentage of biometric applications in real life [Int. biometric group (2003)],
in which biometrics are defined as automated methods of recognizing a person based on
the acquired physiological or behavioural characteristics.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Some Concerns on the Measurement for Biometric 323
Identification
is based
on analysis
Analysis Synthesis
Inverse methods in analysis:
g
eneration of variations
(morphed/distorted signals)
Inverse Biometrics:
g
eneration of synthetic
biometric data
Fig. 13.2. Analysis and synthesis of biometric images.
Inverse methods in analysis examples are shown in
[
10
]
,
[
11
]
,
[
12
]
,
[
21
]
(Fig. 13.2):
(a) Modeling, or tampering (e.g. signature forgery in the task of signature
recognition,
(b) Voice synthesis (telecommunication services), and
(c) Facial image synthesis (animation, “talking heads”).
There are also some developments of generating synthetic biometric
data, using computer graphics, computational geometry methodologies,
with applications to the following fields, which are being implemented
successfully
[
6
]
,
[
7
]
,
[
9
]
,
[
13
]
,
[
23
]
,
[
24
]
:
Collecting large databases for testing the robustness of identification
systems,
Training personnel on a system that deploys simulation of biometric
data,
Cancelable biometrics,
Biometric data hiding, and
Biometric encryption.
Further, according to a most recent book regarding advanced aspects
of biometrics Handbook of Multibiometrics
[
14
]
, consistent advances in
biometrics help to address problems that plague traditional human
recognition methods and offer significant promise for applications in
security as well as general convenience. In particular, newly evolving
systems can measure multiple physiological or behavioral traits and thereby
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
324 Synthesis and Analysis in Biometrics
increase overall reliability that much more. Multimodal Biometrics provides
an accessible, focused examination of the science and technology behind
multimodal human recognition systems, as well as their ramifications for
security systems and other areas of application. After clearly introducing
multibiometric systems, it demonstrates the noteworthy advantages of these
systems over their traditional and unimodal counterparts. In addition, the
work describes the various scenarios possible when consolidating evidence
from multiple biometric systems and examines multimodal system design
and methods for computing user-specific parameters.
In another book about general biometrics systems which was published
by Springer
[
20
]
, which provides an overview of the principles and
methods needed to build reliable biometric systems. It covers 3 main
topics: key biometric technologies, testing and management issues, and
the legal and system considerations of biometric systems for personal
verification/identification. It focuses on the four most widely used
technologies — speech, fingerprint, iris and face recognition. It includes:
(a) In-depth coverage of the technical and practical obstacles which are
often neglected by application developers and system integrators and
which result in shortfalls between expected and actual performance,
(b) Detailed guidelines on biometric system evaluation, and
(c) Protocols and benchmarks which will allow developers to compare
performance and track system improvements.
The book of Biometric Systems — Technology, Design and Performance
Evaluation is intended as a reference book for anyone involved in the design,
management or implementation of biometric systems.
According to a recent paper regarding metric learning for text
documentation analysis and understanding
[
8
]
, many algorithms in machine
learning rely on being given a good distance metric over the input space.
Rather than using a default metric such as the Euclidean metric, it is
desirable to obtain a metric based on the provided data.
We consider the problem of learning a Riemannian metric associated
with a given differentiable manifold and a set of points. Their
approach to the problem involves choosing a metric from a parametric
family that is based on maximizing the inverse volume of a given
data set of points. From a statistical perspective, it is related to
maximum likelihood under a model that assigns probabilities inversely
proportional to the Riemannian volume element.
We discuss in detail learning a metric on the multinomial simplex
where the metric candidates are pull-back metrics of the Fisher information
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Some Concerns on the Measurement for Biometric 325
under a Lie group of transformations. When applied to text document
classification the resulting geodesic distance resemble, but outperform, the
tfidf cosine similarity measure.
In the literature, there is another recent book regarding some guidance
to biometrics published in 2004
[
1
]
. This complete, technical guide offers
some rather detailed descriptions and analysis of the principles, methods,
technologies, and core ideas used in biometric authentication systems. It
explains the definition and measurement of performance and examines
the factors involved in choosing between different biometrics. It also
delves into practical applications and covers a number of topics critical
for successful system integration. These include recognition accuracy, total
cost of ownership, acquisition and processing speed, intrinsic and system
security, privacy and legal requirements, and user acceptance.
From above investigations, it can be seen that most recent development
and applications largely rely on the fundamental definition of pattern
matching, which in turn depends on the measurement of distances between
two input images. It has been observed that the concept of “ambiguity”
plays a very important fundamental roles of all these distance measurement
and image pattern processing in both learning (analysis) and recognition
(synthesis), which is also the most difficult obstacles of essentially all
recognition problems in the real daily life. Therefore in the next section,
we are going to discuss these problems and how input images (patterns)
can and should be “disambiguated” in dealing with real life problem
applications.
13.2. The Problem of “Ambiguity,” and Its Solutions
The main purpose of “Pattern Recognition” is trying to achieve high
accuracy, fast speed, and saving memory and cost. It basically has
two portions: learning (analysis) and recognition (synthesis), as can
be shown in Fig. 13.3.
Notice that here we adapt a general diagram that incorporate the
capability to handle general 3D objects, including “articulated” objects,
which is a compromise between two extreme categories of “rigid” objects
such as bricks, of which the object has fixed shape and “deformable”
objects such as newspaper and handkerchief, of which the shape can easily
vary arbitrarily. Interesting enough, this category of “articulated” objects
also include some human body movements such as hands, fingers, arms,
elbows, arms stretch, legs running, body language, hand sign language for
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
326 Synthesis and Analysis in Biometrics
handicapped people gymnastic movements, and swimming diving etc
[
16
]
,
[
17
]
.
We also include “interactive” dynamics in which human beings are
involved in the process interactively with the machine, such as making
decision as for which input pattern should be included in the compiled
dictionary in “learning” as well as “recognition stage”
[
15
]
.
Fig. 13.3. Learning/recognition general diagram.
Recently, through rapid advancements and developments of past years,
there have been tremendous achievements in reaching these goals. For
example, take character recognition, there have been methods that can
reach as high as almost 99.99% accuracy rate
[
2
]
. However, even with
100% accuracy rate, there still might be more work need to be done to
overcome serious difficulties in solving real life problems. This is related
to “syntax,” “semantics,” and “pragmatics” problems. It is well known
that “ambiguity” refers to the same “syntax” with multiple “meanings” or
“interpretations,” also known as “semantics”.
Let’s call this “1-st level of syntax-to-semantics ambiguity.” Likewise,
even with the same “semantics,” there might be different interpretations,
therefore, different implementations accordingly, also known as
“pragmatics”. Let’s call this 2nd level semantics-to-pragmatics ambiguity.
The 1st level “semantics ambiguity” is mainly due to the following three
categories and their variations:
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Some Concerns on the Measurement for Biometric 327
Syntax (pattern, shape) of the input images are very similar but with
different semantics (interpretations, meanings), like English capital
letter ‘O’ and numerical number ‘0’; English small letter ‘l’ and
numerical number one ‘1’ etc. Some hand printed English characters
examples are shown in Fig. 13.4.
Segmentation difficulty, like the famous statement,
Read my lips, NO NEW TAXES!
(by former President George Bush), which can be interpreted in at least
two different ways, namely:
NO NEW TAXES,
and
NO NEWT AXES!
If the word strings were segmented differently. And there is other
example and commonly misinterpreted hand printed mathematical
formula of X = H +G as X =1−1+CT. It is particularly troublesome
with those which are in the immediate neighborhood of “continuous
transformation”
[
18
]
,
[
19
]
, for examples, O and C, C and |, H and A
etc., as shown in Fig. 13.5.
Grammatical rule tree parsing, for example,
TIME FLIES LIKE AN ARROW,
which has at least three different interpretations, each corresponds a
parsing tree. It also deals with “learning” experience. For example,
let’s take a look at the following scenario in a typical elementary school
classroom:
Teacher: Let’s learn an English letter Oh "O", got it?
Students: yes, (it is easy)
Teacher: Now let’s learn another numerical symbol zero
"O", got it?
Students: yes, (it is easy)
Teacher: Now that you have learned two symbols, let me
give you a quick quiz: I’ll write a circle on the
blackboard like "O", can you recognize what it is?
Students: ? ? ?
I think by now, all students will pause for a while, because there are
two possible answers.