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
This page intentionally left blankThis page intentionally left blank
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Chapter 14
Issues Involving the Human Biometric Sensor Interface
Stephen J. Elliott
∗
, Eric P. Kukula, Shimon K. Modi
Biometrics Standards, Performance, and Assurance Laboratory,
Knoy Hall Room 372, 401 N. Grant Street,
West Lafayette, IN- 47906, USA
∗
elliottpurdue.edu
This chapter outlines the role of Human Biometric Sensor Interaction
(HBSI) which examines topics such as ergonomics, the environment,
biometric sample quality, and device selection, and how these factors
influence the successful implementation of a biometric system. Research
conducted at Purdue University’s Biometric Standards, Performance,
and Assurance Laboratory has shown that the interaction of these
factors has a significant impact on the performance of a biometric
system. This chapter examines the impact on biometric system
performance of a number of modalities including two-dimensional and
three-dimensional face recognition, fingerprint recognition, and dynamic
signature verification. By applying ergonomic principles to device design
as well as understanding the environment in which the biometric sensor
will be placed, can positively impact, the quality of a biometric sample,
thus resulting in increased performance.
Contents
14.1. Introduction ................................... 340
14.2. Ergonomics and an Introduction to the Human Biometric
SensorInteraction................................ 343
14.3. Influence of Environmental Factors on Face Recognition . . . . . . . . . . 346
14.3.1. 2DFaceRecognition ......................... 346
14.3.2. 3DFaceRecognition ......................... 348
14.4. FingerprintImageQuality ........................... 350
14.5. DigitalSignatureVerification ......................... 354
14.6. Summary..................................... 359
Bibliography ....................................... 360
339
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
340 Synthesis and Analysis in Biometrics
Glossary
FRVT — The Face Recognition Vendor Test
HBSI — Human Biometric Sensor Interaction
IEA — The International Ergonomics Association
NIST — The National Institute of Standards and Technology
14.1. Introduction
Biometric technologies are defined as the automated recognition of
behavioural and biological characteristics of individuals. Using various
techniques of pattern recognition, specific behavioural and biological
traits of individuals are collected, extracted, and matched to produce
a result, which will either validate the claim of identity or reject it.
Biometric technologies are used currently for a multitude of applications
for authenticating an individual either as a stand alone application, or
as part of a multi-factor authentication system (i.e. combined with other
possessions such as an ID card, or knowledge such as a PIN). When
deciding whether to implement a biometric system, many factors have to
be examined to evaluate the potential performance of the system against
other non-biometric solutions. One important factor in the evaluation of a
biometric system is to understand how the target population will react to
that system, and what their resulting performance will be.
As biometrics become more pervasive in society, understanding
the interaction between the human and the biometric sensor becomes
imperative, and each biometric modality have their strengths and
weaknesses depending on the application and target population. Biometrics
can be split into two different types, commonly termed
Behavioural and
Biological (also sometimes referred to as physiological).
Behavioural biometrics include signature and voice, and biological
or physiological biometrics include face, finger, hand geometry and iris.
Although to some extent these categories overlap as some modalities are
functions of both behavioural and biological voice, face, and signature can
have influences from both behavioural and physiological.
Biometric modalities can also be classified by five properties outlined
by
[
9
]
:
(a) Universality,
(b) Uniqueness,
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Issues Involving the Human Biometric Sensor Interface 341
(c) Permanence,
(d) Collectibility, and
(e) Acceptability.
Herein lies one of the challenges associated with large scale deployments
of biometrics and the topic of the paper the majority of biometrics are
challenged by these five categories.
Example 8. Not everybody has a particular biometric trait or that
individuals biometric trait maybe significantly different from the common
expected trait for example in hand geometry an individual may have a
missing digit which may prevent them from providing an image to the
sensor.
Another issue is that a biometric characteristic should be invariant over
time, but in the case of fingerprints, the scarring of the print over time, may
affect its character (a measure of image quality) which will subsequently
affect the performance of the biometric matcher. And as biometrics become
deployed in a variety of locations, the positioning of the sensor, and its
surrounding environment become important.
As biometric technologies become more pervasive in society, research
must be undertaken to assess the impact of those subjects who have non-
normal biometric traits, in order to predict the impact of such samples
on the biometric system performance, and with such research emerges
the field of Human Biometric Sensor Interaction (HBSI). The research
described below examines some of the problems associated with biometric
authentication, namely with the environment, population and devices.
These research areas are fundamental, because the widespread adoption
of biometric technologies requires an understanding of how each of these
factors can affect the performance of the device. Both government and
industry need to understand
(a) How biometric samples perform over time,
(b) How changes to biometric samples affect system performance, and
(c) How the image quality of a biometric sample changes with age.
The Face Recognition Vendor Test (FRVT), designed by the National
Institute of Standards and Technology (NIST) and conducted in 2000 and
2002 in support of the U.S.A. Patriot Act, assessed if advancements in
face recognition algorithms actually improved performance. Ten vendors
participated in this test in 2002; test results proved that variation in outdoor
lighting conditions, even for images collected on the same day, drastically
reduced system performance.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
342 Synthesis and Analysis in Biometrics
Example 9. The verification performance for the best face recognition
systems drops from 95 percent to 54 percent when the test environment is
moved from indoors to outdoors
[
36,38
]
.
This poses the fundamental question:
Is the photograph of an individual who enrolls indoors at a drivers
license bureau represent an environment of controlled illumination?
(a) If controlled lighting is used, then facial recognition systems will
verify that particular individual 95 percent of the time in that same
environment.
(b) If verification is attempted outdoors, then the facial recognition system
performance would fail about 46 percent of the time, even on the same
day as enrollment.
This scenario of failure in different environments but on the same day
as enrollment eliminates the factor of template aging. In most real-world
scenarios, an individual will not be identified
(a) In the same environment or
(b) On the same day as enrollment.
Therefore, more research must be conducted to assess system
performance in a variety of conditions.
It is important to consider the effect of poor quality biometric samples
on the performance of the overall system.
Image quality for fingerprints of the elderly population has been shown
to have poor ridge definition and lower moisture content, resulting in lower
image quality
[
44
]
. Fingerprint recognition systems that will be used by an
elderly population need to take into account these factors so that impact
on performance is minimal.
Example 10. A research study was conducted to examine image qualities
of younger group (18–25 years old) and elderly group (62 years and above).
The research compared the number of minutiae and image quality from
the two different age groups, and the research showed that there was a
significant difference between the image qualities of the fingerprints from
the two age groups.
How biometric samples differ when collected on different devices.
Another important consideration involves how biometric samples differ
when collected on different devices
[
51
]
.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Issues Involving the Human Biometric Sensor Interface 343
Example 11. In
[
12
]
, more than 15,000 dynamic signatures were collected
on different mobile computing devices and the resulting signatures analyzed
to see which variables are stable among different digitizers. The purpose of
this research was to test signature verification software on both traditional
digitizer tables and on wireless/mobile computing devices, in order to assess
how the dynamics of the signature signing process change on such devices.
Example 12. The research examined the differences on table-based
digitizers (Wacom Technologies Intuos and Interlink Electronics, Inc.’s ePad
electronic signature devices) and mobile computing devices (Palm, Inc.’s
Palm IIIxeTM, Symbol 1500 and 1740 digitizers). The research showed
significant differences in specific variables across different digitizers.
Therefore, when deciding to implement dynamic signature verification
using different digitizers, the device type and identity needs to be embedded
in the signature data. Therefore, the study of the environment, population,
device variables, and how the human interacts with the sensor is important
to the field of biometrics, and will be discussed in this chapter.
14.2. Ergonomics and an Introduction to the Human
Biometric Sensor Interaction
Traditionally, biometric hardware and software research and development
have focused on increasing performance, increasing throughput, and
decreasing the size of the sensor or hardware device. However,
understanding how system design impacts both the user interaction and
system performance is of utmost importance and is one of the objectives of
Ergonomics.
Ergonomics is a derivative of the Greek words “ergon,” or work, and
“nomos,” meaning laws. In 2000, the International Ergonomics Association
(IEA) defined ergonomics or human factors as:
The scientific discipline concerned with the understanding of
interactions among humans and other elements of a system, and the
profession that applies theory, principles, data and methods to
design in order to optimize human well-being and overall system
performance
[
18
]
.
In 2003, leading biometric experts met under the sponsorship of
the National Science Foundation to “develop a rational and practical
description of crucial scholarly research to support the development of
biometric systems”
[
41
]
. Within the proposed research agenda was an
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
344 Synthesis and Analysis in Biometrics
item on ergonomic design of the capture system and usability studies to
evaluate the effect on system performance
[
41
]
. Andatthetimeofwriting,
the authors’ have found minimal literature or research in the fusion of
biometrics and ergonomics.
Fig. 14.1. The fit of ergonomics in the human and machine interaction [8].
There is potential for merging the disciplines of ergonomics and
biometrics, creating an interdisciplinary research opportunity called the
HBSI. In design, ergonomics attempts to achieve an optimal relationship
between humans and machines in a particular environment. The goal of
ergonomics according to
[
48
]
is:
The goal of ergonomics is to fit, or adapt, the work to individuals, as
opposed to fitting the individuals to the work or device.
Figure 14.1 presents the model proposed by
[
48
]
to show where
ergonomics fits in the human-machine interaction. If we apply Tayyari and
Smith’s model for biometrics — which includes the human-machine (sensor)
interaction (Fig. 14.1), and add ergonomics to the model, the overlap of all
three areas is defined as the HBSI, which can be seen in Fig. 14.2.
The HBSI applies ergonomic principles in the design of a biometric
device or system to minimize unneeded stressors on the human body to
simplify the way individuals interact with the sensor, system, etc. This
paper examines the assertion that by improving the HBSI overall system
performance could be improved through a higher rate of repeatability of
samples by humans using the system. This is important as atypical users
such as those with musculoskeletal disorders, disabilities such as missing
digits, or the elderly for example may have problems either using the
device to produce an image/sample or have problems producing repeatable
samples or images.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Issues Involving the Human Biometric Sensor Interface 345
Fig. 14.2. Interaction of the human, biometric sensor/device, and ergonomics to form
the HBSI.
Example 13. An example of a “normal” and “problematic” user using a
hand geometry reader can be seen in Fig. 14.3. The “problematic” user
has difficulty aligning their hand on the platen as the ring finger does not
completely reach the guide pin, which causes highly variable placements,
ultimately affecting the ability for this individual to be recognized by the
system. The flat shape of the platen, wrist extension angles, and motion
of the fingers may also cause discomfort or awkwardness, resulting in users
not being able to use the device, or even worse individuals not being able
to repeat hand placements in the same manner over time. The impact of
non-repeatable hand placements is sizeable, as this challenges the algorithm
and system to match hand images from users with varying characteristics
and attributes.
Thus:
(a) By incorporating ergonomic principles into biometric sensor/device
research and development an optimal relationship between the device
and users can potentially be reached.
(b) Through optimization, system performance benefits a greater number
of users that could successfully use the device and produce repeatable
samples.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
346 Synthesis and Analysis in Biometrics
Fig. 14.3. Example of a normal user (left) and a problematic user with a partial missing
digit (right).
14.3. Influence of Environmental Factors on Face Recognition
Another factor that can impact the human biometric interface is
environment. The discussion in this section involves the impact of
illumination on both 2D and 3D face. It must be noted the studies cannot
be directly compared as they were conducted at different times, on different
populations, using different cameras, different light levels, but did follow
similar testing protocols. Thus the two stories are presented here to give
specific examples of how one environmental factor illumination impacted
the human biometric interface and the system performance.
14.3.1. 2D Face Recognition
Numerous experts, reports and papers have stated that illumination has a
significant impact on facial recognition performance. There is, however,
only anecdotal evidence to support this claim
[
1
]
,
[
32
]
,
[
37
]
,
[
38
]
,
[
40
]
,
[
42
]
,
[
45
]
,
[
46
]
,
[
47
]
.In
[
44
]
evaluated the performance of a commercially
available 2D facial recognition algorithm across three illumination levels.
The illumination levels were determined by:
(a) Polling levels in two locations, with the third light level (medium light)
determined by calculating the mean of the high and
(b) Low light levels.
Figure 14.4 shows sample images illustrating the differences between the
three illumination levels.
The study consisted of:
(a) Thirty individuals, of which 73 percent were males, Caucasian and
between the ages of 20 and 29.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Issues Involving the Human Biometric Sensor Interface 347
Fig. 14.4. 2D face images captured at varying light levels [17].
(b) Other ethnicities represented in the study included Hispanic and
Asian/Pacific Islander.
There were:
1. Six enrollment failures out of ninety-six attempts, representing an overall
failure to enroll (FTE) rate of 6.25 percent.
2. The failure to acquire (FTA) rates for low light (7–12 lux), medium light
(407–412 lux) and high light (800–815 lux) were 0.92%, 0.65 %, and
0%, respectively
[
23
]
.
The 2D evaluation compared the face recognition algorithm’s ability
to verify each enrollment illumination condition template to the sample
verification images taken at each illumination level.
Example 14. Sample verification images (3 attempts at each illumination
condition in Fig. 14.4: low (left), medium (middle), and high (right)) were
compared to the low illumination enrollment template (Fig. 14.4, leftmost
image). The process was then repeated for the other two enrollment
conditions and took place over three visits.
The statistical analysis revealed that when the high illumination
enrollment template was used, the illumination of verification attempts
was not statistically significant, based on the three tested illumination
levels. So the results of this study reveal that when illumination conditions
during verification are expected to be variable, the illumination level at
enrollment should be as high as practically possible. For the low and
medium enrollment templates, the illumination used for the verification
attempts was statistically significant, meaning that enrollments in a darker
environment do not perform well when environmental lighting conditions
for verification vary. In summary, the lessons learned from
[
23
]
show that
the enrollment illumination level is a better indicator of performance than
the illumination level of the verification attempts. Thus, stringent quality