2 FINGERPRINT VERIFICATION
Lawrence O'Gorman
Veridicom Inc.
Chatham, NJ
log@veridicom.com
1. Introduction
1
The use of fingerprints as a biometric is both the oldest mode of computer-aided,
personal identification and the most prevalent in use today. However, this widespread
use of fingerprints has been and still is largely for law enforcement applications.
There is expectation that a recent combination of factors will favor the use of
fingerprints for the much larger market of personal authentication. These factors
include: small and inexpensive fingerprint capture devices, fast computing hardware,
recognition rate and speed to meet the needs of many applications, the explosive
growth of network and Internet transactions, and the heightened awareness of the need
for ease-of-use as an essential component of reliable security.
This chapter contains an overview of fingerprint verification methods and related
issues. We first describe fingerprint history and terminology. Digital image processing
1
Portions of this chapter have previously appeared in, L. O'Gorman, "Overview of fingerprint
verification technologies," Elsevier Information Security Technical Report, Vol. 3, No. 1, 1998.
Abstract The use of fingerprints for identification has been
employed in law enforcement for about a century. A much
broader application of fingerprints is for personal
authentication, for instance to access a computer, a network, a
bank-machine, a car, or a home. The topic of this chapter is
fingerprint verification, where "verification" implies a user
matching a fingerprint against a single fingerprint associated
with the identity that the user claims. The following topics are
covered: history, image processing methods, enrollment and
verification procedures, system security considerations,
recognition rate statistics, fingerprint capture devices,
combination with other biometrics, and the future of fingerprint
verification.
Keywords: fingerprint verification, fingerprint matching,
biometric, image enhancement, image filtering, feature detection,
minutiae, security, fingerprint sensor.
2 O'Gorman
methods are described that take the captured fingerprint from a raw image to match
result. Systems issues are discussed including procedures for enrollment, verification,
spoof detection, and system security. Recognition statistics are discussed for the
purpose of comparing and evaluating different systems. We describe different
fingerprint capture device technologies. We consider fingerprints in combination with
other biometrics in a multi-modal system and finally look to the future of fingerprint
verification.
It is necessary to state at the onset that there are many different approaches used
for fingerprint verification. Some of these are published in the scientific literature,
some published only as patents, and many are kept as trade secrets. We attempt to
cover what is publicly known and used in the field, and cite both the scientific and
patent literature. Furthermore, while we attempt to be objective, some material is
arguable and can be regarded that way.
2. History
There is archaeological evidence that fingerprints as a form of identification have
been used at least since 7000 to 6000 BC by the ancient Assyrians and Chinese. Clay
pottery from these times sometimes contain fingerprint impressions placed to mark
the potter. Chinese documents bore a clay seal marked by the thumbprint of the
originator. Bricks used in houses in the ancient city of Jericho were sometimes
imprinted by pairs of thumbprints of the bricklayer. However, though fingerprint
individuality was recognized, there is no evidence this was used on a universal basis
in any of these societies.
In the mid-1800's scientific studies were begun that would established two critical
characteristics of fingerprints that are true still to this day: no two fingerprints from
different fingers have been found to have the same ridge pattern, and fingerprint ridge
patterns are unchanging throughout life. These studies led to the use of fingerprints
for criminal identification, first in Argentina in 1896, then at Scotland Yard in 1901,
and to other countries in the early 1900's.
Computer processing of fingerprints began in the early 1960s with the introduction
of computer hardware that could reasonably process these images. Since then,
automated fingerprint identification systems (AFIS) have been deployed widely
among law enforcement agencies throughout the world.
In the 1980s, innovations in two technology areas, personal computers and optical
scanners, enabled the tools to make fingerprint capture practical in non-criminal
applications such as for ID-card programs. Now, in the late 1990s, the introduction of
inexpensive fingerprint capture devices and the development of fast, reliable matching
algorithms has set the stage for the expansion of fingerprint matching to personal use.
Why include a history of fingerprints in this chapter? This history of use is one that
other types of biometric do not come close to. Thus there is the experience of a
century of forensic use and hundreds of millions of fingerprint matches by which we
can say with some authority that fingerprints are unique and their use in matching is
extremely reliable. For further historical information, see [2].
Fingerprint Verification 3
3. Matching: Verification and Identification
Matching can be separated into two categories: verification and identification.
Verification is the topic of this chapter. It is the comparison of a claimant fingerprint
against an enrollee fingerprint, where the intention is that the claimant fingerprint
matches the enrollee fingerprint. To prepare for verification, a person initially enrolls
his or her fingerprint into the verification system. A representation of that fingerprint
is stored in some compressed format along with the person's name or other identity.
Subsequently, each access is authenticated by the person identifying him or herself,
then applying the fingerprint to the system such that the identity can be verified.
Verification is also termed, one-to-one matching.
Identification is the traditional domain of criminal fingerprint matching. A
fingerprint of unknown ownership is matched against a database of known
fingerprints to associate a crime with an identity. Identification is also termed, one-tomany
matching.
There is an informal third type of matching that is termed one-to-few matching.
This is for the practical application where a fingerprint system is used by "a few"
users, such as by family members to enter their house. A number that constitutes
"few" is usually accepted to be somewhere between 5 and 20.
4. Feature Types
The lines that flow in various patterns across fingerprints are called ridges and the
spaces between ridges are valleys. It is these ridges that are compared between one
fingerprint and another when matching. Fingerprints are commonly matched by one
(or both) of two approaches. We describe the fingerprint features as associated with
these approaches.
The more microscopic of the approaches is called minutia matching. The two
minutia types that are shown in Figure 2.1 are a ridge ending and bifurcation. An
ending is a feature where a ridge terminates. A bifurcation is a feature where a ridge
splits from a single path to two paths at a Y-junction. For matching purposes, a
minutia is attributed with features. These are type, location (x, y), and direction (and
some approaches use additional features).
Figure 2.1 Fingerprint minutiae: ending and bifurcation.
bifurcationending
4 O'Gorman
The more macroscopic approach to matching is called global pattern matching or
simply pattern matching. In this approach, the flow of ridges is compared at all
locations between a pair of fingerprint images. The ridge flow constitutes a global
pattern of the fingerprint. Three fingerprint patterns are shown in Figure 2.2.
(Different classification schemes can use up to ten or so pattern classes, but these
three are the basic patterns.)
Two other features are sometimes used for matching: core and delta. (Figure 2.2.)
The core can be thought of as the center of the fingerprint pattern. The delta is a
singular point from which three patterns deviate. The core and delta locations can be
used as landmark locations by which to orient two fingerprints for subsequent
matching ­ though these features are not present on all fingerprints.
There may be other features of the fingerprint that are used in matching. For
instance, pores can be resolved by some fingerprint sensors and there is a body of
work (mainly research at this time) to use the position of the pores for matching in the
same manner that the minutiae are used. Size of the fingerprint, and average ridge and
valley widths can be used for matching, however these are changeable over time. The
Figure 2.2 Fingerprint patterns: arch, loop, and whorl. Fingerprint landmarks are also
shown: core and delta. (No delta locations fall within the captured area of the whorl
here.)
positions of scars and creases can also be used, but are usually not used because they
can be temporary or artificially introduced.
core
core
core
delta
delta
ARCH WHORLLOOP
Fingerprint Verification 5
5. Image Processing and Verification
Following image capture to obtain the fingerprint image, image processing is
performed. The ultimate objective of image processing is to achieve the best image by
which to produce the correct match result. The image processing steps are the
following: image noise reduction and enhancement, feature detection, and matching.
This section is organized to describe first the sequence of processing and
verification via a "common" minutia-based approach. This is described without
variants and optional methods (of which there are many) for the sake of reading flow
and simplicity. It is important to note that, though many researchers and product
developers follow this approach, all do not, and even the choice of what constitutes
"common" may be contentious. In the final subsections of this section, variations of
this approach, both minutia-based and non-minutia-based, are described.
Image Specifications
Depending upon the fingerprint capture device, the image can have a range of
specifications. Commonly, the pixels are 8-bit values, and this yields an intensity
range from 0 to 255. The image resolution is the number of pixels per unit length, and
this ranges from 250 dots per inch (100 dots per centimeter) to 625 dots per inch (250
dots per centimeter), with 500 dots per inch (200 dots per centimeter) being a
common standard. The image area is from 0.5 inches square (1.27 centimeter) to 1.25
inches (3.175 centimeter), with 1 inch (2.54 centimeter) being the standard. We
discuss more on image capture devices in Section 8.
Image Enhancement
A fingerprint image is one of the noisiest of image types. This is due predominantly to
the fact that fingers are our direct form of contact for most of the manual tasks we
perform: finger tips become dirty, cut, scarred, creased, dry, wet, worn, etc. The
image enhancement step is designed to reduce this noise and to enhance the definition
of ridges against valleys. Two image processing operations designed for these
purposes are the adaptive, matched filter and adaptive thresholding. The stages of
image enhancement, feature detection, and matching are illustrated in Figure 2.3.
There is a useful side to fingerprint characteristics as well. That is the
"redundancy" of parallel ridges. Even though there may be discontinuities in
particular ridges, one can always look at a small, local area of ridges and determine
their flow. We can use this "redundancy of information" to design an adaptive,
matched filter. This filter is applied to every pixel in the image (spatial convolution is
the technical term for this operation). Based on the local orientation of the ridges
around each pixel, the matched filter is applied to enhance ridges oriented in the same
direction as those in the same locality, and decrease anything oriented differently. The
latter includes noise that may be joining adjacent ridges, thus flowing perpendicular to
the local flow. These incorrect "bridges" can be eliminated by use of the matched
filter. Figure 2.3(b) shows an orientation map where line sectors represent the
orientation of ridges in each locality. Thus, the filter is adaptive because it orients
6 O'Gorman
itself to local ridge flow. It is matched because it should enhance ­ or match ­ the
ridges and not the noise.
After the image is enhanced and noise reduced, we are ready to extract the ridges.
Though the ridges have gradations of intensity in the original grayscale image, their
true information is simply binary: ridges against background. Simplifying the image
to this binary representation facilitates subsequent processing. The binarization
operation takes as input a grayscale image and returns a binary image as output. The
image is reduced in intensity levels from the original 256 (8-bit pixels) to 2 (1-bit
pixels).
The difficulty in performing binarization is that all the fingerprint images do not
have the same contrast characteristics, so a single intensity threshold cannot be
chosen. Furthermore, contrast may vary within a single image, for instance if the
finger is pressed more firmly at the center. Therefore, a common image processing
tool is used, called locally adaptive thresholding. This operation determines thresholds
adaptively to the local image intensities. The binarization result is shown in Figure
2.3(c).
The final image processing operation usually performed prior to minutia detection
is thinning. Thinning reduces the widths of the ridges down to a single pixel. See
Figure 2.3(d). It will be seen in the next section how these single-pixel width ridges
facilitate the job of detecting endings and bifurcations. A good thinning method will
reduce the ridges to single-pixel width while retaining connectivity and minimizing
the number of artifacts introduced due to this processing. These artifacts are
comprised primarily of spurs, which are erroneous bifurcations with one very short
branch. These artifacts are removed by recognizing the differences between legitimate
and erroneous minutiae in the feature extraction stage described below.
Image enhancement is a relatively time-consuming process. A 500x500-pixel
fingerprint image has 250,000 pixels; several multiplications and other operations are
applied at each pixel. Both matched filtering and thinning contribute largely to this
time expenditure. Consequently, many fingerprint systems are designed to conserve
operations at this stage to reach a match result more quickly. This is not a good
tradeoff. The results of all subsequent operations depend on the quality of the image
as captured by the sensor and as processed at this stage. Economizing for the sake of
speedup will result in degraded match results, which in turn will result in repeated
attempts to verify or false rejections. Therefore, it is our contention that a system
offering reasonable speed with a correct answer is much better than a faster system
that yields poorer match results.
Feature Extraction
The fingerprint minutiae are found at the feature extraction stage. Operating upon the
thinned image, the minutiae are straightforward to detect. Endings are found at
termination points of thin lines. Bifurcations are found at the junctions of three lines.
See Figure 2.3(e).
Fingerprint Verification 7
a) b)
c) d)
e) f)
Figure 2.3 Sequence of fingerprint processing steps: a) original, b) orientation, c)
binarized, d) thinned, e) minutiae, f) minutia graph.
8 O'Gorman
There will always be extraneous minutiae found due to a noisy original image or
due to artifacts introduced during matched filtering and thinning. These extraneous
features are reduced by using empirically determined thresholds. For instance, a
bifurcation having a branch that is much shorter than an empirically determined
threshold length is eliminated because it is likely to be a spur. Two endings on a very
short isolated line are eliminated because this line is likely due to noise. Two endings
that are closely opposing are eliminated because these are likely to be on the same
ridge that has been broken due to a scar or noise or a dry finger condition that results
in discontinuous ridges. Endings at the boundary of the fingerprint are eliminated
because they are not true endings but rather the extent of the fingerprint in contact
with the capture device.
Feature attributes are determined for each valid minutia found. These consist of:
ridge ending or bifurcation type, the (x,y) location, and the direction of the ending or
bifurcation. Although minutia type is usually determined and stored, many fingerprint
matching systems do not use this information because discrimination of one from the
other is often difficult.
The result of the feature extraction stage is what is called a minutia template. This
is a list of minutiae with accompanying attribute values. An approximate range on the
number of minutiae found at this stage is from 10 to 100. If each minutia is stored
with type (1 bit), location (9 bits each for x and y), and direction (8 bits), then each
will require 27 bits ­ say 4 bytes ­ and the template will require up to 400 bytes. It is
not uncommon to see template lengths of 1024 bytes.
Verification
At the verification stage, the template from the claimant fingerprint is compared
against that of the enrollee fingerprint. This is done usually by comparing
neighborhoods of nearby minutiae for similarity. A single neighborhood may consist
of three or more nearby minutiae. Each of these is located at a certain distance and
relative orientation from each other. Furthermore, each minutia has its own attributes
of type (if it is used) and minutia direction, which are also compared. If comparison
indicates only small differences between the neighborhood in the enrollee fingerprint
and that in the claimant fingerprint, then these neighborhoods are said to match. This
is done exhaustively for all combinations of neighborhoods and if enough similarities
are found, then the fingerprints are said to match. Template matching can be
visualized as graph matching, that is comparing the shapes of graphs joining
fingerprint minutiae. This is illustrated in Figure 2.3(f).
Note that the word, "similar" is used in the paragraph above instead of "same".
Neighborhoods will rarely match exactly because of two factors. One is the noisy
nature of a fingerprint image. The other is that the skin is an elastic surface, so
distances and minutia directions will vary.
One result of the verification stage is a match score, usually a number between 0
and 1 (or 10 or 100). Higher values in the range indicate higher confidence in a match.
This match score is then subject to a user-chosen threshold value. If the score is
greater than the threshold, the match result is said to be true (or 1) indicating a correct
Fingerprint Verification 9
verification, otherwise the match is rejected and the match result is false (or 0). This
threshold can be chosen to be higher to achieve greater confidence in a match result,
but the price to pay for this is a greater number of false rejections. Conversely, the
threshold can be chosen lower to reduce the number of false rejections, but the price
to pay in this case is a greater number of false acceptances. The trade-off between
false acceptance and false rejection rates is further discussed in Section 7.
The user has control of only one parameter, the threshold, for most commercial
verification products. This customization procedure is called back-end adjustment,
because a match score is calculated first and a threshold can be chosen after to
determine the match result. There are systems that, in addition to offering back-end
adjustment, offer front-end adjustment as well. This enables the user to adjust some of
the parameter values before the match score is calculated, then to adjust the threshold
after. Systems with front-end adjustment offer more versatility in obtaining the best
results for different conditions, but are more complex for the user to adjust. This is
why, for most systems, the vendor sets the optimum front-end parameter values and
the user has control only of the matching threshold value via back-end adjustment.
Identification and One-to-Few Matching
Although the emphasis in this chapter is verification, we briefly mention identification
and one-to-few matching methods. For identification, the objective is to determine a
match between a test fingerprint and one of a database of fingerprints whose size may
be as high as 10,000 to tens of millions. One cannot simply apply the verification
techniques just described to all potential matches because of the prohibitive
computation time required. Therefore, identification is usually accomplished as a twostep
process. Fingerprints in the database are first categorized by pattern type, or
binned. The same is done for the test fingerprint. Pattern comparison is done between
test fingerprint and database fingerprints. This is a fast process that can be used to
eliminate the bulk of non-matches. For those fingerprints that closely match in
pattern, the more time-consuming process of minutia-based verification is performed.
One-to-few matching is usually accomplished simply by performing multiple
verifications of a single claimant fingerprint against the 5 to 20 potential matches.
Thus the execution time is linear in the number of potential matches. This time
requirement becomes prohibitive if "few" becomes too large, then an approach akin to
identification must be used.
Variations on the Common Approach: Other Methods
Since one of the most vexing challenges of fingerprint processing is obtaining a clean
image upon which to perform matching, there are various methods proposed to
perform image enhancement. Most of these involve filtering that is adaptively
matched to the local ridge orientations [23, 19, 25, 22, 24, 37, 26, 27, 14]. The
orientation map is first determined by dividing the image into windows (smaller
regions) and calculating the local ridge orientations within these. The orientation can
10 O'Gorman
be determined in each window by spatial domain processing or by frequency domain
processing after transformation by a 2-dimensional fast Fourier transform.
After image enhancement and binarization of the fingerprint image, thinning is
usually performed on the ridges. However, a different approach eliminates the
binarization and thinning stages (both computationally expensive and noise
producing) [20]. This approach involves tracing ridges not from the binary or thinned
image, but from the original grayscale image. The result of grayscale ridge-following
is the endpoint and bifurcation minutiae similar to the common approach.
Instead of using only a single size window to determine the orientation map,
multiple window sizes can be used via a multi-resolution approach [24, 15]. Local
orientation values are determined first throughout the image at a chosen, initial
resolution level ­ that is a chosen window size of pixels within which the orientation
is calculated. A measure of consistency of the orientation in each window is
calculated. If the consistency is less than a threshold, the window is divided into four
smaller sub-windows and the same process is repeated until consistency is above
threshold for each window or sub-window. This multi-resolution process is performed
to avoid smoothing over small areas of local orientation, as will be the case especially
at the fingerprint core.
Because of the difficulty of aligning minutiae of two fingerprints, neighborhood
matching was one of the earliest methods of facilitating a match [28, 1, 42]. Groups of
neighboring minutiae are identified in one fingerprint, usually two to four minutiae to
a neighborhood, and each of these is compared against prospective neighborhoods of
another fingerprint. There are two levels to matching. One is matching the
configurations of minutiae within a neighborhood against another neighborhood. The
other is matching the global configurations formed by the separate neighborhoods
between enroll and verify fingerprints.
Because it is time-consuming to compare all neighborhood combinations between
enroll and verify fingerprints, methods have been proposed to align the fingerprints to
reduce the number of comparisons. A common method, and also a traditional method
used for visual matching, is to locate a core and delta and align the fingerprints based
on these landmarks [29]. The core and delta are usually found on the basis of their
position with respect to the ridge flow, therefore the orientation map is determined
and used for this [41]. An elegant method to locate singular points in a flow field is
the Poincaré index [17, 36, 16]. For each point in the orientation map, the orientation
angles are summed for a closed curve in a counter-clockwise direction around that
point. For non-singular points, the sum is equal to 0 degrees; for the core, the sum is
equal to 180 degrees; for a delta, the sum is equal to ­180 degrees.
Other methods have been proposed to reduce the computational load of minutia
matching. One approach is to sort the list of minutiae in some order conducive to
efficient comparisons prior to matching. (This is especially appropriate for one-tomany
matching, since sorting is done once per fingerprint, but matching many times.)
A linearly sorted list of minutiae can be compiled by scanning the fingerprint from a
selected center point outward by a predetermined scanning trajectory such as a spiral
[39]. In this way, one-dimensional vectors of minutiae, including their characteristics,
can be compared between enroll and verify fingerprints. Another method to linearize
the minutia comparison is the "hyperladder" matcher [11]. This hyperladder is
constructed sequentially by comparing minutia pairs in enroll and verify fingerprints
Fingerprint Verification 11
and adding more rungs as consecutive neighboring minutiae match. In another
approach, an attributed graph can be constructed where branches constitute nearestneighbor
minutiae and these emanate like "stars" on the graph [10]. These stars are
compared between fingerprint pairs, the number of matching branches constituting the
degree of confidence in the match.
Because there is so little discriminating information at a single minutia (even the
type is unreliable), a different approach is to describe minutiae by more features [47,
40]. For instance, a minutia can be described by the length and curvature of the ridge
it is on and of similar features on neighboring ridges.
Variations on the Common Approach: Correlation Matching
This discussion of matching has been minutia-focused to this point, to the exclusion
of the global pattern matching approach mentioned in Section 4. Instead of using
minutiae, some systems perform matches on the basis of the overall ridge pattern of
the fingerprint. This is called global matching, correlation, or simply image
multiplication or image subtraction.
It is visibly apparent that a pair of fingerprints of different pattern types, for
instance whorl and arch, does not match. Global matching schemes go beyond the
simple (and few) pattern categories to differentiate one whorl from a different whorl,
for instance. Simplistically, this can be thought of as a process of aligning two
fingerprints and subtracting them to see if the ridges correspond. There are four
potential problems (corresponding to three degrees of freedom and another factor).
1. The fingerprints will likely have different locations in their respective images
(translational freedom). We can establish a landmark such as a core or delta by
which to register the pair, however if these are missing or not found reliably,
subsequent matching steps will fail.
2. The fingerprints may have different orientations (rotational freedom). If a proper
landmark has been found in (1), the fingerprint can be rotated around this, but this
is error-prone, computationally expensive, or both. It is error-prone because the
proper center of rotation depends on a single, reliably determined landmark. It is
computationally expensive because performing correlation for many orientations
involves repeatedly processing the full image.
3. Because of skin elasticity (non-linear warping), even if matching fingerprints are
registered in location and orientation, all sub-regions may not align.
4. Finally, there is the inevitable problem of noise. Two images of matching
fingerprints will have different image quality, ridges will be thicker or thinner,
discontinuities in ridges will be different depending on finger dryness, the portion
of the fingerprint captured in each image will be different, etc.
The descriptions below are more sophisticated modifications and extensions to the
basic correlation approach to deal with the problems listed.
Strictly speaking, correlation between two images involves translating one image
over another and performing multiplication of each corresponding pixel value at each
translation increment [38]. When the images correspond at each pixel, the sum of
these multiplications is higher than if they do not correspond. Therefore, a matching
12 O'Gorman
pair will have a higher correlation result than a non-matching pair. A threshold is
chosen to determine whether a match is accepted, and this can be varied to adjust the
false acceptance rate versus false rejection rate tradeoff similarly to the case for
minutia matching.
Correlation matching can be performed in the spatial frequency domain instead of
in the spatial domain as just described [12]. The first step is to perform a 2dimensional
fast Fourier transform (FFT) on both the enrollee and claimant images.
This operation transforms the images to the spatial frequency domain. The two
transformed images are multiplied pixel-by-pixel, and the sum of these
multiplications is equivalent to the spatial domain correlation result. An advantage of
performing frequency domain transformation is that the fingerprints become
translation-independent; that is, they do not have to be aligned translationally because
the origin of both transformed images is the zero-frequency location, (0,0). There is a
trade-off to this advantage however, that is the cost of performing the 2-dimensional
FFT.
Frequency domain correlation matching can be performed optically instead of
digitally [43, 44, 21]. This is done using lenses and a laser light source. Consider that
a glass prism separates projected light into a color spectrum, that is it performs
frequency transformation. In a similar manner, the enrollee and claimant images are
projected via laser light through a lens to produce their Fourier transform. Their
superposition leads to a correlation peak whose magnitude is high for a matching pair
and lower otherwise. An advantage of optical signal processing is that operations
occur at the speed of light, much more quickly than for a digital processor. However,
the optical processor is not as versatile ­ as programmable ­ as a digital computer,
and because of this few or no optical computers are used in commercial personal
verification systems today.
One modification of spatial correlation is to perform the operation not upon image
pixels but on grids of pixels or on local features determined within these grids [8, 6].
The enrollee and claimant fingerprint images are first aligned, then (conceptually)
segmented by a grid. Ridge attributes are determined in each grid square: average
pixel intensity, ridge orientation, periodicity, or number of ridges per grid.
Corresponding grid squares are compared for similar attributes. If enough of these are
similar, then this yields a high match score and the fingerprints are said to match.
The relative advantages and disadvantages between minutia matching and
correlation matching differ between systems and algorithmic approaches. In general,
minutia matching is considered by most to have a higher recognition accuracy.
Correlation can be performed on some systems more quickly than minutia matching,
especially on systems with vector-processing or FFT hardware. Correlation matching
is less tolerant to elastic, rotational, and translational variances of the fingerprint and
of extra noise in the image.
6. Systems Issues
The effectiveness of a complete fingerprint verification system depends on more than
the verification algorithms just described. There are other, higher level considerations,
which we will call systems issues. These include enrollment and verification
Fingerprint Verification 13
procedures, speed and ergonomics, user-feedback, anti-spoofing, and security
considerations.
It is essential to the goal of high recognition rate that the enrollment procedure
results in the capture of the highest quality fingerprint image(s) obtainable because
enrollment occurs once while verification occurs many times. Therefore, a welldesigned
verification system will require the user to go through more time and effort
for enrollment than for verification. A fingerprint may be captured multiple times and
the best taken or some combination of each taken as the enrolled fingerprint.
There are options in the design of the verification procedure as well. The
fingerprint can be captured once or a few times until a positive match is made. A
procedure such as this will decrease false rejections, but increase false acceptances.
Verification can be performed on not just one, but two or more fingers. This will
enhance the recognition rate, however it will also cause the user to expend more time.
System ergonomics are important. For instance, there are limits to the amount of
time that a person is willing to wait in personal authentication applications. That
amount of time varies with the particular application and depends on what the person
is also doing during processing, for instance swiping a bankcard or entering an
identification number. Between 0.5 and 1 second are usually regarded as an
acceptable range for processing time. Other user ergonomics considerations include:
the number of repeated attempts in case of false rejections, the procedures for
enrollment and verification, the design of the capture device, and the recognition
setting that determines the trade-off between false acceptance and false rejection.
Quality feedback is useful when an image is captured to indicate to the user how to
place the finger for the best possible image quality. The type of feedback includes:
"finger is placed too high", "finger is not pressed hard enough", etc.
Anti-spoofing deterrents must be built into a fingerprint system to prevent use of
an artificial fingerprint, a dead finger, or latent fingerprint. A latent fingerprint
sometimes remains on a sensor surface due to skin oil residue from the previously
applied fingerprint. Countermeasures are built into some sensors, such as the ability to
distinguish true skin temperature, resistance, or capacitance.
Since the fingerprint system is only as secure as its weakest link, a complete,
secure system must be designed. For instance, minutia templates must be secured by
some means such as encryption to prevent impostors from inserting their templates
into the database in place of properly enrolled users. The end result of fingerprint
verification is a "yes" or "no" that is used to gain access. If it is simple just to
circumvent the fingerprint system to send a "yes", then the system provides little
security. A solution to this problem is to ensure that the host receiving the recognition
decision knows that this is from the trusted client, such as by digitally signing the
information passed to the host. (For further information on encryption, see reference
[33].)
14 O'Gorman
7. Recognition Rate
Terminology and Measurement
The ultimate measure of utility of a fingerprint system for a particular application is
recognition rate. This can be described by two values. The false acceptance rate
(FAR) is the ratio of the number of instances of pairs of different fingerprints found to
(erroneously) match to the total number of match attempts. The false rejection rate
(FRR) is the ratio of the number of instances of pairs of the same fingerprint are found
not to match to the total number of match attempts. FAR and FRR trade off against
one another. That is, a system can usually be adjusted to vary these two results for the
particular application, however decreasing one increases the other and vice versa.
FAR is also called, false match rate or Type II error, and FRR is also called false
non-match rate or Type I error. These are expressed as values in [0, 1] interval or as
percentage values.
The ROC-curve plots FAR versus FRR for a system. (ROC stands for Receiver
Operating Curve for historical reasons. Yes, "ROC-curve" is redundant, but this is the
common usage.) ROC-curves are shown in Figure 2.4. The FAR is usually plotted on
the horizontal axis as the independent variable. The FRR is plotted on the vertical axis
as the dependent variable. Because of the range of FAR values, this axis is often on a
logarithmic scale. Figure 2.4 contains two solid curves and three dotted curves. The
solid curves do not represent any particular data; they are included for illustrative
purposes to show better and worse curve placements. The typical ROC-curve has a
shape whose "elbow" points toward (0,0) and whose asymptotes are the positive xand
y-axes. The sharper the elbow and (equivalently) the closer is the ROC-curve to
the x- and y-axes, the lower is the recognition error and the more desirable is the
result.
Figure 2.4 ROC-curves. The 2 solid curves are of hypothetical data illustrating
desirable and less desirable recognition performance. The 3 dotted curves are of real
data measuring the performance of 3 commercial AFIS [46].
FRR [%]
FAR [log %]
0
4
12
0.001 0.01 0.1 1.0 10.0
20
8
16
better
recognition
performance
worse
recognition
performance
AFIS data
Fingerprint Verification 15
The procedure for using the ROC-curve is as follows. Choose an acceptable level
of FAR. On Figure 2.4, a dashed line is shown at 0.01% FAR. The FRR
corresponding to this choice is the attainable FRR, in this example about 4%.
Alternatively, the FRR can be specified and the FAR found on the curve.
There is no single set of FAR and FRR specifications useful for all different
applications. If the fingerprint system is specified for very high security situations
such as for military installations, then the FAR will be chosen to be very low (e.g.,
<0.001%). However, this results in higher FRR, sometimes in the range from 5% to
20%. Typical customer applications such as for automatic teller machines cannot
afford to alienate users with such a high FRR. Therefore, the choice in these
applications is low FRR (e.g., <0.5%), at the sacrifice of higher FAR. (An FRR
specification that is sometimes quoted for automatic teller machines is less than 1 per
100,000 false rejection.)
Third-Party Benchmarking
In Figure 2.4, we include three ROC-curves of AFIS data from third-party
benchmarking [46]. The database was compiled from employees of the Philippine
Social Security System, mostly white-collar workers. The database consists of 600
people, 8 fingers per person, and two sets per person, where enrollment and
verification sets were captured with an intervening interval of 2 to 8 weeks. From this
database, 3278 matching fingerprint pairs and 4129 non-matching pairs were tested.
These images were captured with an Identicator DF-90 optical scanner at 500dpi,
512x512 pixels, 1x1" image size.
We include these AFIS data from a respected third-party tester for the reader to
compare against other data whose validity may be suspect. There is much misleading
information in the commercial biometric industry regarding recognition rates. In
general, these AFIS can be expected to yield better recognition results than most
verification systems (though AFIS generally have higher cost, they are slower for 1to-1
matching, and require more memory). Note that this AFIS test is only for single
image comparisons. A verification system can take advantage of the real-time nature
of its application to perform multiple verification attempts so as to improve the
recognition rate.
Specifying and Evaluating Recognition Rate Statistics
For statistical results to be properly evaluated, they must be accompanied by the
following information: sample size, description of population, and testing description.
The sample size should contain the following information: the number of subjects
(people), the number of fingers, the number of images per finger, and the total number
of fingerprint images. In addition, the image number should be broken out into
number of match and non-match images. For example, a test might consist of 100
subjects, 2 fingers per subject and 4 images per finger. The total number of images is
100x2x4=800. If each finger (200) is compared against each of the other images from
the same finger (4 choose 2 = 6 pairs per unique finger), there are 200x6=1200
matching pairs. If one image from each finger (200) is compared against all images
from different subjects (99x4x2=792), there are 200x792=158,400 non-match pairs.
16 O'Gorman
The description of population states the type of subjects included in the sample. Of
particular importance in judging fingerprint statistics is the type of work engaged in
by the subjects. A study involving masons will have different statistical results than
that involving white-collar workers whose hands are subject to less abuse. The age
statistics should be described, at least stating a relative breakdown on the number of
children, adults, and elderly people included in the sample. The proportion of males
and females should also be stated.
Finally, the test design should be described. Of particular interest is who performed
the tests. The strong preference is that a reputable third-party conducts and reports the
test. Was the capture procedure supervised or not? Were the subjects given training or
visual feedback to place the finger correctly on the fingerprint capture device? Was
the sample manually filtered in any way to remove "goats" (people whose fingerprints
are very difficult to capture and match with reliable quality)? Was the procedure
adjusted using a practice sample of fingerprints, then tested separately on different
images to yield the published results? What was the range of rotational and
translational variance allowed, or were the fingerprints manually centered in the
image? What were the make and specifications of the capture device? Where and
when were the tests conducted (e.g., Florida humid summer or Minnesota dry winter)?
What components of the system were involved in the test: just matching algorithms,
just sensor, full system? Most test results do not list all these conditions, but the most
possible information enables more valid evaluation.
It is important to emphasize that results cannot be compared if determined under
different test conditions. It is a misrepresentation of test data to state that a matcher
achieved certain results for test design A, so it can be compared against the results
from test design B. Valid comparisons between results can be done only for the same
database under the same conditions.
8. Image Capture Devices
We organize image capture devices into three categories: optical, solid-state, and
other. There is yet another category, fingerprint acquisition via inking, which is the
traditional mode of criminal fingerprint capture. It is evident that this is inappropriate
for fingerprint verification due to the inconvenience involved with ink, the need for
subsequent digitization, and perhaps the stigma of this type of capture. The type of
image acquisition for fingerprint verification is also called "live-scan fingerprint
capture".
Optical fingerprint capture devices have the longest history and use of these
categories, dating back to the 1970s. These operate on the principal of frustrated total
internal reflection (FTIR). A laser light illuminates a fingerprint placed on a glass
surface (platen). The reflectance of this light is captured by a CCD array (solid-state
camera). The amount of reflected light is dependent upon the depth of ridges and
valleys on the glass and the finger oils between the skin and glass. The light that
passes through the glass into valleys is not reflected to the CCD array, whereas light
that is incident upon ridges on the surface of the glass (more precisely, the finger oils
on the ridges that constitute the ridge-to-glass seal) is reflected.
Fingerprint Verification 17
Innovations in optical devices have been made recently, primarily in an effort to
reduce the size of these devices. Whereas an optical sensor was housed in a box about
6x3x6 inches as recently as the mid-1990s, smaller devices have recently appeared
that are in the order of 3x1x1 inches. Different optical technologies than FTIR have
also been developed. For instance, fiber optics has been proposed to capture the
fingerprint [7]. A bundle of optical fibers is aimed perpendicularly to the fingerprint
surface. These illuminate the fingerprint and detect reflection from it to construct the
image. Another proposal is a surface containing an array of microprisms mounted
upon an elastic surface [4]. When a fingerprint is applied to the surface, the different
ridge and valley pressures alter the planar surfaces of the microprisms. This image is
captured optically via the reflected light (or absence of it) from the microprisms.
Solid-state sensors have appeared on the marketplace recently, though they have
been proposed in the patent literature for almost two decades. These are microchips
containing a surface that images the fingerprint via one of several technologies.
Capacitive sensors have been designed to capture the fingerprint via electrical
measurements [45, 18, 48, 13]. Capacitive devices incorporate a sensing surface
composed of an array of about 100,000 conductive plates over which is a dielectric
surface. When the user places a finger on this surface, the skin constitutes the other
side of an array of capacitors. The measure of voltage at a capacitor drops off with the
distance between plates, in this case the distance to a ridge (closer) or a valley
(further). Pressure-sensitive surfaces have been proposed where the top layer is of an
elastic, piezoelectric material to conform to the topographic relief of the fingerprint
and convert this to an electronic signal [30, 31, 9, 34]. Temperature sensitive sensors
have been designed to respond to the temperature differential between the ridges
touching the surface of the device and the valleys more distant from them [9].
Ultrasonic scanning falls into the final category of fingerprint capture technologies
[32]. An ultrasonic beam is scanned across the fingerprint surface much like laser
light for optical scanners. In this case, it is the echo signal that is captured at the
receiver, which measures range, thus ridge depth. Ultrasonic imaging is less affected
by dirt and skin oil accumulation than is the case for optical scanning, thus the image
can be a truer representation of the actual ridge topography.
Two of the three most important factors that will decide when fingerprint
verification will be commercially successful in the large-volume personal verification
market are low cost and compact size. (The other factor is recognition rate, discussed
in Section 7.) Capture device prices have fallen over an order of magnitude between
the early to late 1990s (from approximately $1500 (US) to $100), and manufacturers
promise close to another order of magnitude decrease in the next few years. As far as
size, we have mentioned the reduction of optical sensor size from 6x3x6 inches to
3x1x1. Solid-state sensor systems are this size or smaller, and as further integration
places more circuitry on the chip (such as digitizer circuitry to convert the fingerprint
measurements to digital intensities), these systems are becoming even smaller. Solidstate
sensors are approaching the lower limit of size needed to capture the surface area
of the finger, about 1x1 inch with a fraction of an inch depth.
A functionality that has not been available before solid-state sensors is locally
adjustable, software-controlled, automatic gain control (AGC). For most optical
devices, gain can be adjusted only manually to change the image quality. Some solidstate
sensors, however, offer the capability to automatically adjust the sensitivity of a
18 O'Gorman
pixel or row or local area to provide added control of image quality. AGC can be
combined with feedback to produce high quality images over different conditions. For
instance, a low-contrast image (e.g., dry finger) can be sensed and the sensitivity
increased to produce an image of higher contrast on a second capture. With the
capability to perform local adjustment, a low-contrast region in the fingerprint image
can be detected (e.g., where the finger is pressed with little pressure) and sensitivity
increased for those pixel sensors on a second capture.
Optical scanners also have advantages. One advantage of larger models is in image
capture size. It is costly to manufacture a large, solid-state sensor, so most current
solid-state products have sub-1 inch square image area, whereas optical scanners can
be 1 inch or above. However, this advantage is not true for some of the smaller optical
scanners. The small optical scanners also have smaller image capture areas because a
larger area would require a longer focal length, thus larger package size. Optical
scanners are subject to linear distortion at the image edges when larger image capture
area is combined with smaller package size.
9. Multi-Modal Biometrics
Multi-modal biometrics refers to the combination of two or more biometric modalities
into a single system. The most compelling reason to combine different modalities is to
improve recognition rate. This can be done when features of different biometrics are
statistically independent. For the different modalities listed in Table 2.1, it is likely
that each is largely independent from the other (though we know of no research study
to date that confirms this).
There are other reasons to combine biometrics. One is that different modalities are
more appropriate in different situations. For a home banking application for instance,
a customer might enroll both with fingerprint and voice. Then, the fingerprint can be
used from a home or laptop sensor; while voice and a PIN (personal identification
number) can be used over the phone. Another reason is simply customer preference.
For instance, an automatic teller machine could offer eye and fingerprint and face
biometrics, or a combination of two of these for the customer to choose.
Although fingerprints can be combined with other modalities, there are reasons to
suggest that this would not be the first biometric to require complementing. One
reason is that, along with eye systems, fingerprint systems already have very high
recognition rates. This contrasts with less reliable modalities where combining one
with another or with a PIN is more advantageous. Another reason is that a single
person has up to ten statistically independent samples in ten fingers, compared to two
for eye and hand, and one for face, voice, and signature.
Table 2.1 shows selected features of each modality and can be used to determine
complementary modalities for multi-modal systems. A few notes on this table:
´ Biometric technologies are changing rapidly; for the most up-to-date information,
check company literature and industry reports such as at reference [3] and review
issues such as [35, 5].
Fingerprint Verification 19
´ The row for the eye biometric describes features applying to either iris or retinal
scanning technologies.
´ In the matching column, whereas all technologies are appropriate for 1-to-1
matching, only fingerprint and eye technologies are proven to have acceptable
recognition rates to be practical for 1-to-many matching. This is an indication that
these two modalities provide the highest recognition rates for verification as well.
´ Variation of the salient features used for recognition is very different for different
modalities. Fingerprint and eye features remain consistent for a lifetime, whereas
the others change with growth. On a day-to-day basis, there is far less variation
for all modalities, though voice can change with illness and signature with
demeanor.
´ As far as sensor cost, eye systems are currently more costly than the others; voice
systems can be zero cost to the user if a telephone is used.
´ Fingerprint and voice systems have the smallest comparative sizes with eye
systems currently the largest.
Biometric
Matching
1-to-1,
1-to-many
Variation:
Lifetime,
Day-to-Day
Maximum
Independent
Samples per
Person
Sensor
Cost
[$US]
Sensor
Size
fingerprint yes, yes none,
little
10 10-102
very
small
eye yes, yes none,
very little
2 102
-103
medium
hand yes, no much,
very little
2 102
medium
face yes, no much,
medium
1 102
small
voice yes, no much,
medium
1 0-102
very
small
signature yes, no much,
medium
1 102
medium
Table 2.1 Features of different biometric modalities.
10. Future
Where is biometric technology going? System price will continue to decrease along
with size, while recognition rates will improve (at a slower rate than price and size
20 O'Gorman
changes). Recognition rate will be a deciding factor in acceptance for demanding
applications such as automatic teller machines (requiring a very low rate of false
rejections), and military (requiring a very low rate of false acceptances). For
especially demanding applications, multi-modal systems will evolve to combine
biometrics to provide an optimum level of security and convenience to users.
Alternatively, multiple verifications, such as by using multiple fingers, will be used to
enhance recognition reliability. If costs plummet as the industry projects, personal use
of biometric systems will grow to replace the current reliance on passwords, PINs,
and door keys that are used for computers, home security systems, restricted entry,
ATMs, credit cards, Internet access, corporate networks, confidential databases, etc.
The biometrics promise is to make access much simpler while at the same time
providing a higher level of security.
References
[1] K. Asai, H. Izumisawa, H. Owada, S. Kinoshita, and S. Matsuno, "Method and Device for
Matching Fingerprints with Precise Minutia Pairs Selected from Coarse Pairs," US Patent
4646352, 1987.
[2] J. Berry, "The history and development of fingerprinting," in Advances in Fingerprint
Technology, (H. C. Lee and R. E. Gaensslen, ed.s), CRC Press, Florida, 1994, pp. 1-38,
1994.
[3] Biometric Consortium Web page: www.biometrics.org.
[4] W. S. Chen and C. L. Kuo, "Apparatus for Imaging Fingerprint or Topgraphic Relief
Pattern on the Surface of an Object," US Patent 5448649, 1995.
[5] C. Ciechanowicz, Special issue on biometric technologies, Elsevier Information Security
Technical Report, Vol. 3, No. 1, 1998.
[6] L. Coetzee and E. C. Botha, "Fingerprint Recognition in Low Quality Images," Pattern
Recognition, Vol. 26, No. 10, pp. 1441-1460, 1993.
[7] R. F. Dowling Jr. and K. L. Knowlton, "Fingerprint Acquisition System With a Fiber Optic
Block," US Patent 4785171, 1988.
[8] E. C. Driscoll Jr., C. O. Martin, K. Ruby, J. J. Russell, and J. G. Watson, "Method and
Apparatus for Verifying Identity Using Image Correlation," US Patent 5067162, 1991.
[9] D. G. Edwards, "Fingerprint Sensor," US Patent 4429413, 1984.
[10] M. A. Eshera and R. E. Sanders, "Fingerprint Matching System," US Patent 5613014,
1997.
[11] S. Ferris, R. L. Powers, and T. Lindh, "Hyperladder Fingerprint Matcher," US Patent
5631972, 1997.
[12] R. C. Gonzalez and Richard E. Woods, Digital Image Processing, Addison-Wesley,
Massachusetts, 1992.
[13] D. Inglis, L. Manchanda, R. Comizzoli, A. Dickinson, E. Martin, S. Mendis, P. Silverman,
G. Weber, B. Ackland, and L. O'Gorman, ``A robust, 1.8V, 250 microWatt, direct contact
500dpi fingerprint sensor,''IEEE Solid State Circuits Conference, San Francisco, 1998.
[14] A. K. Jain, L. Hong, and R. Bolle, "On-line fingerprint verification," IEEE Trans. Pattern
Analysis and Machine Intelligence, Vol. 19, No. 4, pp. 302-313, 1997.
[15] A. K. Jain, L. Hong, S. Pankanti, and R. Bolle, "An Identity-Authentication System Using
Fingerprints," Proceedings of the IEEE, Vol. 85, No. 9, pp. 1365-1388, 1997.
[16] K. Karu and A. K. Jain, "Fingerprint Classification," Pattern Recognition, Vol. 29, No. 3,
pp. 389-404, 1996.
Fingerprint Verification 21
[17] M. Kawagoe and A. Tojo, "Fingerprint Pattern Classification," Pattern Recognition, Vol.
17, pp. 295-303, 1984.
[18] A. G. Knapp, "Fingerprint Sensing Device and Recognition System Having Predetermined
Electrode Activation," US Patent 5325442, 1994.
[19] H. E. Knutsson, R. Wilson, and G. H. Granlund, "Anisotropic Nonstationary Image
Estimation and its Applications: Part I ­ Restoration of Noisy Images," IEEE Trans.
Communications, Vol. 31, pp. 388-397, 1983.
[20] D. Maio and D. Maltoni, "Direct Gray-Scale Minutiae Detection in Fingerprints," IEEE
Trans. Pattern Analysis and Machine Intelligence, Vol. 19, No. 1, pp. 27-40, 1997.
[21] R. A. Marsh and George S. Petty, "Optical Fingerprint Correlator," US Patent 5050220,
1991.
[22] B. M. Mehtre, N. N. Murthy, and S. Kapoor, "Segmentation of Fingerprint Images Using
the Directional Image," Pattern Recognition, Vol. 20, No. 4, pp. 429-435, 1987.
[23] O. Nakamura, K. Goto, and T. Minami, "Fingerprint Classification by Directional
Distribution Patterns," Systems, Computers, and Controls, Vol. 13, pp. 81-89, 1982.
[24] L. O'Gorman and J. V. Nickerson, "An approach to fingerprint filter design", Pattern
Recognition, Vol. 22, No. 1, pp. 29-38, 1989.
[25] E. Peli, "Adaptive Enhancement Based on a Visual Model," Optical Engineering, Vol. 26,
No. 7, pp. 655-660, 1987.
[26] N. K. Ratha, S. Chen, and A. K. Jain, "Adaptive Flow Orientation-Based Feature
Extraction in Fingerprint Images," Pattern Recognition, Vol. 28, No. 11, pp. 1657-1672,
1995.
[27] N. K. Ratha, K. Karu, S. Chen, and A. K. Jain, "A Real-Time Matching System for Large
Fingerprint Databases", IEEE Trans. Pattern Analysis and Machine Intelligence, Vol. 18,
No. 8, pp. 799-813, 1996.
[28] J. P. Riganati and V. A. Vitols, "Minutiae Pattern Matcher," US Patent 4135147, 1979.
[29] J. P. Riganati and V. A. Vitols, "Automatic Pattern Processing System," US Patent
4151512, 1979.
[30] H. Ruell, "Input Sensor Unit for a Fingerprint Identification System," US Patent 4340300,
1982.
[31] H. Ruell, "Fingerprint Sensor," US Patent 4394773, 1983.
[32] J. K. Schneider and W. E. Glenn, "Surface Feature Mapping Using High Resolution Cspan
Ultrasonography," US Patent 5587533, 1996.
[33] B. Schneier, Applied Cryptography, John Wiley and Sons, Inc., New York, 1996.
[34] T. Scheiter, M. Biebl, and H. Klose, "Sensor for Sensing Fingerprints and Method for
Producing the Sensor," US Patent 5373181, 1994.
[35] W. Shen and R. Khanna (eds.), "Special issue on automated biometrics," Proceedings of
the IEEE, Vol. 85, No. 9, Sept., pp. 1343-1492, 1997.
[36] B. G. Sherlock and D. M. Munro, "A Model for Interpreting Fingerprint Topology,"
Pattern Recognition, Vol. 26, No. 7, pp. 1047-1055, 1993.
[37] B. G. Sherlock, D. M. Munro, and K. Millard, "Algorithm for Enhancing Fingerprint
Images," Electronics Letters, Vol. 28, No. 18, pp. 1720-1721, 1992.
[38] A. Sibbald, "Method and Apparatus for Fingerprint Characterization and Recognition
Using Auto-correlation Pattern," US Patent 5633947, 1997.
[39] M. K. Sparrow, "Fingerprint Recognition and Retrieval System," US Patent 4747147,
1988.
[40] M. K. Sparrow, "Vector Based Topological Fingerprint Matching," US Patent 5631971,
1997.
[41] V. S. Srinivasan and N. N. Murthy, "Detection of Singular Points in Fingerprint Images,"
Pattern Recognition, Vol. 25, No. 2, pp. 139-153, 1992.
[42] K. E. Taylor and J. B. Glickman, "Apparatus and Method for Matching Image
Characteristics Such as Fingerprint Minutiae," US Patent 4896363, 1990.
22 O'Gorman
[43] C. E. Thomas, "Method and Apparatus for personal Identification," US Patent 3704949,
1972.
[44] G. J. Tomko, "Method and Apparatus for Fingerprint Verification," US Patent 4876725,
1989.
[45] C. Tsikos, "Capacitive Fingerprint Sensor," US Patent 4353056, 1982.
[46] J. L. Wayman, "Biometric Identification Standards Research," Report to U.S. Federal
Highway Administration (FHWA), San Jose State University, December 1997.
[47] M. Yamada, A. Kodata, and H. Tominaga, "A Method of Describing Fingerprint Structure
and Identification Algorithm Using Geometric Characteristics," Systems and Computers in
Japan, Vol. 25, No. 5, pp. 100-112, 1994.
[48] N. D. Young, G. Harkin, R. M. Bunn, D. J. McCulloch, R. W. Wilks, and A. K. Knapp,
"Novel fingerprint scanning arrays using polysilicon TFT's on glass and polymer
substrates," IEEE Electronic Device Letters, Vol. 18, No. 1, pp. 19-20, 1997.