Minutiae Persistence among Multiple Samples of the

Same Person’s Fingerprint in a Cooperative User Scenario

Vedrana Krivokuća, Waleed Abdulla and Akshya Swain

Department of Eletrical and Computer Engineering, The University of Auckland, Auckland, New Zealand

Keywords: Biometrics, Fingerprints, Fingerprint Recognition, Partial Fingerprints, Minutiae, Missing Minutiae.

Abstract: A significant challenge in the development of automated fingerprint recognition algorithms is dealing with

missing minutiae. While it is generally assumed that some minutiae will always be missing between

multiple samples of the same fingerprint, this assumption has never been empirically evaluated. An

important factor influencing minutiae persistence in civilian fingerprint recognition applications is the

consistency with which a user places their finger on the fingerprint scanner during fingerprint image

acquisition. This paper investigates the probability of a reference minutia repeating in another sample of the

same person’s fingerprint, when that probability depends on user consistency alone. The investigation

targets cooperative users in a civilian fingerprint recognition application. To simulate this scenario, a

database of 800 fingerprint samples from 100 participants was collected. Analysis of the database showed

that the median probability of a reference minutia repeating in another sample of the same fingerprint is

0.95 with an interquartile range of 0.04. Combining multiple samples of the same fingerprint to filter out

only the most reliable reference minutiae was shown to improve this probability. A complementary study

demonstrated that automatic feature extractors and matchers may lower minutiae repeatability, but that user

consistency is nevertheless the most influential factor.

1 INTRODUCTION

Fingerprint matching is usually based on small ridge

discontinuities called minutiae (Maltoni et al.,

2009a). The most common minutiae types are the

bifurcation and the termination (see Figure 1). The

more minutiae that two fingerprints have in

common, the greater the probability that they

originated from the same finger.

Figure 1: The most common fingerprint minutiae types.

One of the biggest problems in automated

fingerprint matching is that of missing minutiae. A

minutia may be considered “missing” if it is present

in the reference fingerprint but its corresponding

minutia cannot be found in the query fingerprint,

when both fingerprints come from the same finger.

There are four main reasons why a reference minutia

may be missing from the query fingerprint:

1. The part of the fingerprint in which that

particular minutia exists has not been captured in

the query fingerprint; so, the minutia is literally

not present in the query fingerprint.

2. The minutia is physically present in the query

fingerprint, but the quality of this fingerprint is

poorer than that of the reference fingerprint, so

the minutia cannot be noticed.

3. The minutia is present in the query fingerprint

and the fingerprint is of sufficiently good quality

for the minutia to be noticed by a human expert,

but the automated feature extractor fails to detect

it.

4. The minutia is present in the query fingerprint

and it has been detected by the feature extractor,

but the matcher does not consider this minutia to

match its corresponding reference minutia (even

though the two minutiae do match).

The likelihood of minutiae missing due to reasons 2

to 4 can be reduced by improving the robustness of

the feature extractor and matcher, as well as by

incorporating quality control during fingerprint

Bifurcation

Termination

Krivoku

ca V., Abdulla W. and Swain A..

Minutiae Persistence among Multiple Samples of the Same Person’s Fingerprint in a Cooperative User Scenario.

DOI: 10.5220/0004816500760086

In Proceedings of the 3rd International Conference on Pattern Recognition Applications and Methods (ICPRAM-2014), pages 76-86

ISBN: 978-989-758-018-5

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

capture in civilian fingerprint recognition

applications. The probability of minutiae missing

due to reason 1, however, is more difficult to

control, since it mainly depends on how consistent

the owner of the fingerprint is in presenting that

fingerprint for image capture. The problem of

minutiae missing due to reason 1 falls into the

category of partial fingerprint matching.

Partial fingerprint matching refers to the situation

where we are required to match two fingerprints that

come from the same finger but may not have a large

area of overlap. The area of overlap is usually

defined in terms of the number of minutiae that are

common between both fingerprints. Partial

fingerprint matching has had a considerable amount

of attention in the literature since the early days of

fingerprint recognition. The most popular minutiae-

based methods of matching partial fingerprints rely

on using local minutiae structures; for example

(Hrechak and McHugh, 1990, Chen and Kuo, 1991,

Jea and Govindaraju, 2005). Use of local minutiae

structures avoids the need for fingerprint alignment

using singular points, such as the core and delta,

which may not be present in partial fingerprints.

To improve partial fingerprint matching, several

researchers have proposed the use of additional

fingerprint features to increase the ‘uniqueness’ of

small fingerprint portions; e.g., dots (isolated ridges)

and incipients (thin, immature ridges between the

regular ridges) (Yi and Jain, 2007), the coordinates

and orientations of representative ridge points (Fang

et al., 2007), sweat pores (Kryszczuk et al., 2004,

Zhao et al., 2010), etc. These additional features

introduce supplementary information to make up for

the typically few minutiae that are present in a

partial fingerprint, thereby improving the

performance of partial fingerprint matchers.

Partial fingerprints are most commonly

encountered in forensics, because latent prints left at

crime scenes are usually not planned. In civilian

fingerprint recognition applications, where

fingerprint acquisition is deliberate, there are two

main reasons why a captured fingerprint may be

partial: (i) inconsistency in the placement of the

finger on the fingerprint scanner, and (ii) size of the

scanning surface being smaller than the fingerprint.

In this paper, we investigate reason (i) in terms of

the captured fingerprint minutiae. In particular, we

empirically quantify the probability of a reference

minutia being present in another sample of the same

fingerprint, when the only thing that probability

depends on is the consistency with which a person

places their finger onto a fingerprint scanner. Such

an evaluation is important for determining the

amount of influence that a legitimate user of a

fingerprint recognition system is likely to have on

the final authentication decision.

This investigation targets cooperative users in

civilian fingerprint recognition applications;

therefore, we were unable to use public fingerprint

databases, such as the Fingerprint Verification

Competition (FVC) series (Biometric System

Laboratory, 2013), for testing. This is because most

of those databases were created by asking the

participants to deliberately exaggerate the

inconsistency with which they place their finger on

the provided scanner, so the resulting fingerprint

images are not representative of cooperative users in

a civilian fingerprint recognition application. For

this reason, we collected our own database of 800

fingerprint samples from 100 cooperative users in a

simulated civilian fingerprint recognition scenario.

Although minutiae persistence (repeatability)

among cooperative users would naturally be

expected to be high, an empirical evaluation of this

assumption has not previously been undertaken.

Analysis of our database indicates that cooperative

users in a civilian fingerprint recognition application

may be expected to be consistent enough in the

placement of their fingers onto the fingerprint

scanner to ensure that the median probability of a

reference minutia being present in another sample of

the same fingerprint is 0.95 with an interquartile

range of 0.04. Additional analysis suggests that this

probability may be improved by combining multiple

fingerprints during enrolment to filter out only the

most reliable reference minutiae.

While user consistency is important in ensuring

that the same minutiae are captured during each

scan, minutiae repeatability is also affected by

additional factors, of which errors in automatic

feature extraction and matching are prominent. The

effect of a commercial feature extractor and matcher

on minutiae persistence was thus studied. Results

from this study show that these modules lower

minutiae repeatability, but that user consistency is

nevertheless the most influential factor. This study

serves as an example of how our results on user

consistency may be applied towards honing in on the

most problematic areas in a fingerprint recognition

system, which would be helpful in the development

of the constituent algorithms.

Section 2 of this paper provides details on the

database collection procedure. Section 3 analyses

the database to obtain the probability of a reference

minutia repeating in another sample of the same

fingerprint, when the minutiae persistence depends

only on the consistency with which a user presents

MinutiaePersistenceamongMultipleSamplesoftheSamePerson'sFingerprintinaCooperativeUserScenario

their fingerprint to the scanner. The minutiae

persistence is analysed in two scenarios: one where

the reference minutiae are extracted from a single

reference fingerprint and one where multiple

reference fingerprints are combined to select only

the most reliable reference minutiae. Section 4

illustrates the effect of a commercial minutiae

extractor and matcher on minutiae persistence, and

suggests how the results of our investigation on user

consistency can be applied in the development of

fingerprint recognition algorithms. Section 5

concludes this investigation and recommends venues

for future work in this direction.

2 FINGERPRINT DATABASE

COLLECTION

Public fingerprint databases, such as those provided

for the Fingerprint Verification Competitions (FVC)

(Biometric System Laboratory, 2013), have

generally been constructed by asking the participants

to deliberately exaggerate the inconsistency with

which they place their finger on the provided

fingerprint scanner, e.g., (Maio et al., 2002). Figure

2 shows three samples of the same fingerprint from

the FVC2002 DB1_A database: the first image was

acquired when the user placed their finger on the

scanner in a cooperative manner, and the second and

third images are deliberately rotated and translated

samples of the same fingerprint, respectively.

Figure 2: Three samples of the same fingerprint from

FVC2002 DB1_A.

The nature of these databases makes them suitable

for testing fingerprint recognition algorithms

designed for deployment in uncooperative user

scenarios, e.g., forensics, where the latent prints are

usually partial and of poor quality; border security,

where a criminal may attempt to avoid recognition;

etc. However, they are not representative of

fingerprint samples that would be acquired in

cooperative civilian fingerprint authentication

applications. In such applications, it is in the users’

best interests to be recognised, so it is fair to assume

that they would be fairly consistent in presenting

their fingers to the fingerprint scanner.

The aim of this investigation was to quantify the

consistency of cooperative users in a civilian

fingerprint recognition application. This consistency

was measured in terms of the probability of a

reference minutia being present across multiple

samples of the same person’s fingerprint. At first,

the FVC2006 public fingerprint database (Biometric

System Laboratory, 2006), which was collected by

asking the participants to place their fingers on the

scanner naturally, appeared suitable for our

purposes. However, the construction of this

database did not involve a quality check on the

acquired fingerprint images. In our investigation, a

quality check was important for two reasons.

Firstly, since we were interested in evaluating

minutiae repeatability based on user consistency

alone, we had to eliminate the fingerprint quality

factor from the database. This means that

fingerprint images acquired from the same finger

had to be of approximately the same quality.

Secondly, our investigation targets civilian

fingerprint recognition applications, which usually

perform a quality check on the captured fingerprint

images (Maltoni et al., 2009b). This helps to

improve the chances of a correct authentication

decision by ensuring that the acquired fingerprint

images are all of a sufficiently high quality for

subsequent processing. For this reason, using

fingerprint images of very variable quality was

irrelevant to our investigation. Hence, the FVC2006

database was an unsuitable testing platform for our

purposes and it was necessary to collect our own

fingerprint database. Sections 2.1 to 2.3 describe

our database collection procedure in detail.

2.1 Scanner Specifications

The images in our fingerprint database were

acquired using the Futronic FS88 fingerprint scanner

(Futronic, 2013). The FS88 is an optical scanner,

which produces 8-bit grey level fingerprint images

with a resolution of 320x480 pixels, 500dpi.

A crucial property of electronic fingerprint

scanners, which sets them apart, is their underlying

sensor technology. Since optical sensors are a

popular choice in fingerprint scanner design (Jain et

al., 2011) and since these types of scanners generally

exhibit similar user interfaces, the FS88 scanner may

be considered to be “typical”. This means that the

results of our investigation are not limited to this

particular scanner.

2.2 Participant Selection

Our fingerprint database was constructed using

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

fingerprints provided by volunteers. The fact that

participation was voluntary was the first step in

ensuring that the database would represent

cooperative users. The participants consisted of

adults of both genders, from diverse ethnic

backgrounds and of various ages in the range [18,

60] (though the majority were young adults). In

total, 100 participants were used in this study.

2.3 Methodology

The participants were invited to play the part of

cooperative users in a fingerprint-based computer

typical computer station with the scanner positioned

on the desk approximately where the computer

mouse would be. Each user was free to move the

scanner around and position it in whichever way was

most comfortable for them (as long as it stayed flat

on the desk). Users were asked to choose a finger

that they would use to authenticate themselves in a

fingerprint-based computer login application. The

only guidance that the users received regarding the

proper placement of their finger on the scanner was

that the line of the first joint from the top of the

finger should roughly lie on the line just below the

glass platen on the fingerprint scanner, such that the

maximum fingerprint area is captured (see Figure 3).

Figure 3: Guide on the proper placement of a finger on the

FS88 scanner: align the lines inside the red rectangles.

The participants were then asked to find a

comfortable position on the scanner, which they feel

they could naturally repeat for future scans. Each

participant’s chosen fingerprint was scanned 8 times.

To ensure that a fingerprint image was of

sufficiently good quality for subsequent processing

and that the quality across multiple samples of the

same person’s fingerprint was approximately

consistent, the quality of the fingerprints was

visually examined by the investigator. Users with

dry skin were asked to rub their fingers on the side

of their noise or onto their forehead to apply some

grease to the fingerprint, and users with very wet or

greasy fingers were asked to dab their finger onto a

piece of clothing. A fingerprint image was deemed

to be of sufficiently good quality when the

difference between the ridges and valleys was clear.

Note that fingerprint databases are often

constructed by acquiring multiple samples of the

same person’s fingerprint over several days. The

purpose of this is to simulate natural variability

between the samples; e.g., on some days a person’s

finger may be drier than on other days. However,

since our investigation required elimination of the

quality factor, simulating this natural variability was

unnecessary. So, we elected to collect each of a

participant’s 8 fingerprint samples on the same day.

To simulate multiple authentication attempts, after

each scan the participant was asked to remove their

finger from the scanner while their previous

fingerprint image was saved by a human operator.

The images were saved manually to deliberately

introduce some delay in between the scans and to

‘distract’ the participant, thereby mimicking

different authentication attempts. Once the scanning

started, the human operator did not guide the user in

the placement of their finger on the scanner.

The participants were observed to be careful in

the way in which they placed their fingers on the

scanner. They also became very aware of what a

good quality fingerprint image should look like after

the initial quality check, and most controlled this

quality on their own for subsequent scans, without

prompting by the operator. This suggests that users

are both capable and willing to be cooperative in a

scenario in which they want to be recognised.

3 ANALYSIS OF MINUTIAE

PERSISTENCE BASED ON

USER CONSISTENCY ALONE

The collected database was analysed to gain insight

into the expected persistence (repeatability) of

reference minutiae in a cooperative civilian

fingerprint recognition application, when that

persistence depends on user consistency alone. This

persistence was quantified in terms of the probability

of a reference minutia being physically present in

another sample of the same fingerprint.

To ensure that we were evaluating the baseline

minutiae repeatability, based on user consistency

alone, it was necessary to use ground truth minutiae

information, free from the errors of automatic

fingerprint feature extractors and matchers. For this

reason, the minutiae from each fingerprint were

extracted manually and correspondences between

the minutiae in all 8 samples of each fingerprint

MinutiaePersistenceamongMultipleSamplesoftheSamePerson'sFingerprintinaCooperativeUserScenario

were also established manually. All 8 samples of a

person’s fingerprint were scrutinised simultaneously

to find matching minutiae. Once all the minutiae

were thought to have been identified and matched, a

final, careful check of all 8 samples was made to

ensure that no minutiae were missed out. Note that

minutiae identification and matching in good quality

fingerprint images is fairly simple for an informed

human, as people are naturally good at pattern

recognition. Since a quality check was performed

during image acquisition (see Section 2.3), the

images were of sufficiently good quality to make the

process of identifying minutiae reasonably

straightforward; it just took a lot of patience to

ensure that they were all found! Therefore, we may

conclude that, if any human error crept into this

process, it was insignificant compared to the total

number of minutiae extracted for the entire database.

Reference minutiae repeatability was analysed in

two different scenarios: one in which the reference

minutiae are extracted from a single reference

fingerprint, and one in which multiple reference

fingerprints are combined to filter out the reliable

minutiae. Sections 3.1 and 3.2, respectively, detail

the analysis in each of these scenarios.

3.1 Scenario 1: Single Reference

Fingerprint

In this scenario, the reference minutiae were

extracted from only one reference fingerprint and all

the reference minutiae were considered reliable. For

every person in the database, each of their 8

fingerprint samples had a turn at being the reference,

while their remaining 7 samples were used as the

test fingerprints. The number of test samples in

which each of the reference minutiae appears was

counted, and the probability of a reference minutia

repeating in another sample of the same fingerprint

was then calculated using Equations (1) and (2):









,17

(1)





























(2)

In Equation (1), 



is a fraction representing the

number of test samples out of 7. In Equation (2), i

represents the index of the person whose fingerprints

we are currently analysing; since there are 100

people in our database, 1    100. The subscript

j represents the index of the fingerprint sample that

is currently being used as the reference fingerprint;

since there are 8 fingerprint samples per person,

18. So, 





denotes the probability of a

reference minutia repeating in another sample of

person i’s fingerprint, when the reference minutiae

are extracted from person i’s fingerprint sample j.

The total number of reference minutiae in person i’s

fingerprint sample j is denoted by 





. The number

of reference minutiae that appear in k test samples is

represented by 





The probability of a reference minutia repeating

in another sample of the same fingerprint was

calculated for each of a person’s reference

fingerprints in turn, so there were 8 probabilities per

person. This was repeated for all 100 people in the

database, so there were 800 probabilities in total.

These 800 probabilities were used to plot a

distribution of the probabilities of a reference

minutia repeating in another sample of the same

fingerprint; this distribution is depicted in Figure 4.

Figure 4: Distribution of the probabilities of a reference

minutia repeating in another sample of the same

fingerprint when one reference fingerprint is used.

It is immediately evident that the distribution in

Figure 4 is highly skewed to the left. Calculating the

skewness in MATLAB produced a value of -1.5924,

which confirms this observation. When a

distribution is skewed, the median is a better

indicator of the distribution’s central tendency than

is the mean. The box and whisker plot in Figure 5

provides a visual analysis of the distribution in

Figure 4 in terms of the median, interquartile range

and range of the data.

From Figure 5, the median of 0.95 indicates the

typical probability of a reference minutia repeating

in another sample of the same fingerprint. The

lower quartile tells us that, 75% of the time, we

may expect the probability of a reference minutia

0.65 0.7 0.75 0.8 0.85 0.9 0.95 1

0.02

0.04

0.06

0.08

0.1

0.12

Distribution of the Probabilities of a Reference Minutia Repeating

in Another Sample of the Same Fingerprint

(1 Reference Fingerprint Used)

Probability of a Reference Minutia Repeating

Probability of Obtaining a Certain Probability

for a Reference Minutia Repeating

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

Figure 5: Box and whisker plot of the reference minutiae

repeatability distribution when one reference fingerprint is

used.

repeating in another sample of the same fingerprint

to be 0.93 and above. The upper quartile suggests

that, 25% of the time, the probability of a reference

minutia repeating will be 0.97 and above. So, we

may conclude that the probability of a reference

minutia repeating in another sample of the same

fingerprint when 1 reference fingerprint is used is

typically 0.95 with an interquartile range of 0.04.

This means that cooperative users in a civilian

fingerprint recognition application can be expected

to be consistent enough in placing their finger onto

the fingerprint scanner to ensure that, typically, 95%

of the same minutiae are captured across multiple

samples of the fingerprint, and, 50% of the time, 93-

97% of the same minutiae are captured.

The whiskers of the plot in Figure 5 represent the

range of our distribution, ignoring outliers. (Note

that outliers are those data values that lie more than

1.5 times the height of the box away from either side

of the box, which is a commonly applied rule of

thumb.) From Figure 5, we can see that our data lies

in the range [0.86, 1]. Considering Figure 4, we

may conclude that the probability of a reference

minutia repeating in another sample of the same

fingerprint is likely to lie in the range [0.86, 1] about

96% of the time, since around 96% of the

distribution in Figure 4 lies in this range.

While these results are certainly promising, it

appears logical that using multiple reference

fingerprints to filter out only the most reliable

reference minutiae would increase the probability of

a reference minutia repeating in another sample of

the same fingerprint. We investigate this claim in

Section 3.2.

3.2 Scenario 2: Multiple Reference

Fingerprints

In this scenario, instead of using only a single

reference fingerprint at a time, multiple reference

fingerprints were combined. The idea was to filter

out only the most reliable minutiae to use as the

reference minutiae. If N reference fingerprints are

combined, then the most reliable minutiae are those

minutiae that appear in all N reference fingerprints.

Logically, we would expect that using more

reference fingerprints would improve the chances of

a reference minutia repeating in a test sample of the

same fingerprint. This is because our confidence in

a reference minutia repeating in another sample of

the same fingerprint grows with every sample it

appears in. To verify this expected trend, the

number of reference fingerprints was varied from 1

to 7 for each person. If we let N denote the number

of reference fingerprints used, then the probability of

a reference minutia repeating in another sample of

the same fingerprint was calculated for each N.

Every possible combination of a person’s N

fingerprint samples was used in turn as the reference

sample set. Let 





denote the number of N-

reference-fingerprint combinations per person. For

each value of N, 





was calculated via Equation (3):









!



8



(3)

Let 



denote the total number of N-reference-

fingerprint combinations for the entire database of

100 people. For each N, 



was computed using

Equation (4):











100

(4)

Table 1 lists the values of 





and 



as the number of

reference fingerprints, N, varies from 1 to 7.

Table 1: Values of 





and 



as N varies from 1 to 7.











1 8 800

2 28 2,800

3 56 5,600

4 70 7,000

5 56 5,600

6 28 2,800

7 8 800

To ensure fairness in the comparison between the

probabilities at different values of N, the same

number of test fingerprints was used for each N.

Since only one fingerprint remains to be used as the

test sample when N = 7, one test fingerprint was

0.7

0.75

0.8

0.85

0.9

0.95

Box and Whisker Plot of the

Reference Minutiae Repeatability Distribution

(1 Reference Fingerprint Used)

(MANUAL Minutiae Extraction and Matching)

Probability of a Reference Minutia Repeating

in Another Sample of the Same Fingerprint

MinutiaePersistenceamongMultipleSamplesoftheSamePerson'sFingerprintinaCooperativeUserScenario

used for all values of N. Note that each of a person’s

fingerprint samples had one or more turns

(depending on N) at being the test fingerprint. The

probability of a reference minutia repeating for each

N was then calculated using Equation (5):





















(5)

In Equation (5), 





denotes the total number of

reference minutiae resulting from person i’s

reference fingerprint combination j, when N

reference fingerprints are used. 





denotes the

number of reference minutiae repeating in the test

fingerprint, so 





is the probability of a reference

minutia repeating in the test fingerprint. Note that

1    100, 17, and 1





For each N, Equation (5) was used to calculate







for every N-reference-fingerprint combination

out of 





total combinations, for each person. In the

end, 



probabilities were obtained for each value

of N (see Table 1). For every N, its total set of 



probabilities was used to construct a distribution of

the probabilities of a reference minutia repeating in

another sample of the same fingerprint when N

reference fingerprints are used. Each distribution

was converted into a box and whisker plot for

analysis. Figure 6 shows the box and whisker plots

of the distributions corresponding to each value of

N, side by side for easy comparison.

Figure 6: Box and whisker plots comparing the minutiae

repeatability distributions as the number of reference

fingerprints increases.

The first observation to note from Figure 6 is that, as

the number of reference fingerprints increases, the

median of the minutiae repeatability distribution also

increases, reaching a value of 1 when 3 or more

reference fingerprints are used. This suggests that

increasing the number of reference fingerprints

increases the likelihood of a reference minutia

repeating in another sample of the same fingerprint.

The second observation is that the interquartile range

and range both decrease with an increase in the

number of reference fingerprints, reaching a value of

0 when 6 or more reference fingerprints are used.

This suggests that increasing the number of

reference fingerprints gives us greater confidence

that the minutiae repeatability will be equal to the

median probability.

As an additional measure of the significance of

these results, we calculated the 5

and 1

percentile

for the minutiae repeatability distribution as the

number of reference fingerprints increases. The

results are illustrated in Figure 7.

Figure 7: 5th and 1st percentiles of the minutiae

repeatability distribution as the number of reference

fingerprints increases.

The trends in Figure 7 are expected based on the

analysis of Figure 6. From Figure 7, we can see

that, when only 1 reference fingerprint is used, 95%

of the time we may expect the probability of a

reference minutia repeating in another sample of the

same fingerprint to be 0.83 and above, and 99% of

the time we may expect this probability to be 0.75

and above. Using 7 reference fingerprints increases

these probabilities to 0.94 and 0.86, respectively.

These results are extremely encouraging, because

they suggest that it is possible to improve the

probability of a reference minutia repeating in a test

sample of the same fingerprint simply by using more

reference fingerprints to filter out only the most

reliable reference minutiae. However, we must also

consider the effect that this improvement strategy

has on the total number of reference minutiae

remaining for fingerprint matching purposes. Since

using more reference fingerprints effectively gets rid

of more (unreliable) minutiae, it makes sense to

0.4

0.5

0.6

0.7

0.8

0.9

1234567

Number of Reference Fingerprints Used

Box and W hisker Plots Comparing the

Distributions of the Probabilities of a Reference Minutia Repeating

when Different Numbers of Reference Fingerprints are Used

(MANUAL Minutiae Extraction and Matching)

Probability of a Reference Minutia Repeating

in Another Sample of the Same Fingerprint

1 2 3 4 5 6 7

0.74

0.76

0.78

0.8

0.82

0.84

0.86

0.88

0.9

0.92

0.94

Plot Showing the 5th and 1st Percentiles of the Minutiae Repeatability Probability Distribution

as the Number of Reference Fingerprints Increases

(MANUAL Minutiae Extraction and Matching)

Number of Reference Fingerprints Used

Probability above which 95% and 99% of the

Minutiae Repeatability Distribution Lies

5th

1st

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

conclude that this filtering operation will result in

fewer reference minutiae remaining. To show what

happens to the total number of reference minutiae as

the number of reference fingerprints increases, we

took the value of every





, which was used in

Equation (5) to calculate every 





for Figure 6.

The number of 





values for each N was thus the

same as the number of 





values for each N (see 



values in Table 1). The 





values for each N were

then used to construct the distribution of the total

numbers of reference minutiae when N reference

fingerprints were used. The median of each

distribution was calculated. Figure 8 is a plot

showing the trend in the typical (median) number of

reference minutiae as N increases from 1 to 7.

Figure 8: Typical (median) number of reference minutiae

as the number of reference fingerprints increases.

The trend in Figure 8 is as expected; namely, as the

number of reference fingerprints increases, the

number of reference minutiae decreases. This is

because the idea behind using multiple reference

fingerprints is to filter out only the most reliable

reference minutiae. The most reliable reference

minutiae are those minutiae that are present in all the

reference fingerprints. So, the more reference

fingerprints that are used, the less probable it

becomes that a minutia will be present in all these

fingerprints. Consequently, increasing the number

of reference fingerprints has the effect of removing a

larger number of (unreliable) reference minutiae.

From Figure 8, we can see that, for our database,

the typical number of reference minutiae decreases

from 48 when 1 reference fingerprint is used to 40

when 7 reference fingerprints are used. This means

that the typical number of reference minutiae

resulting from using 7 reference fingerprints was



less than the typical number of reference

minutiae resulting from using 1 reference

fingerprint. This is not a significant difference,

which may be attributed to the fact that the

participants in our database collection were very

consistent in the placement of their fingers onto the

fingerprint scanner. The more consistent a user is in

placing their finger onto a scanner, the more similar

multiple samples of their same fingerprint will be.

Consequently, most of the minutiae should be the

same across all their samples. Relating this

observation to Figure 8, we may conclude that,

typically, about 40 out of 48 minutiae (over 83%)

will be present in all the samples of the same

fingerprint (for cooperative users in a civilian

fingerprint recognition application). This means that

combining multiple samples to filter out only the

most reliable minutiae should not result in the loss of

many minutiae, as is proven in Figure 8. Note that,

traditionally, 12 matching minutiae have been

considered sufficient evidence for a positive

fingerprint match (e.g., see (Kingston, 1964)). This

means that 40 reference minutiae provide ample

opportunity for reliable fingerprint recognition;

therefore, using 7 reference fingerprints would

ensure a high probability of a reference minutia

repeating in another sample of the same fingerprint,

whilst maintaining satisfactory recognition accuracy.

4 EFFECT OF AUTOMATIC

FEATURE EXTRACTOR AND

MATCHER ON MINUTIAE

PERSISTENCE

The analysis in section 3 shows that, when minutiae

persistence (repeatability) for cooperative users

relies only on the user’s consistency in placing their

finger on the fingerprint scanner, the persistence is

typically well over 90%. Unfortunately, while user

consistency is very important for ensuring that the

same fingerprint features are captured during every

authentication attempt, there is often a discrepancy

between what features are actually present in a

fingerprint and what features the automatic

fingerprint recognition system ‘thinks’ are present.

In other words, minutiae persistence is influenced

not only by the user’s consistency in capturing the

same fingerprint area, but also by the robustness of

the subsequent image processing and pattern

recognition algorithms in the fingerprint recognition

system.

We may logically expect minutiae repeatability

to be quite heavily influenced by the robustness of

1 2 3 4 5 6 7

Typical (Median) Number of Reference Minutiae

as Number of Reference Fingerprints Increases

Number of Reference Fingerprints Used

Typical (Median) Number of Reference Minutiae

MinutiaePersistenceamongMultipleSamplesoftheSamePerson'sFingerprintinaCooperativeUserScenario

the automatic feature extractor and matcher. This is

because, even if the same minutiae are captured in

every scan of a fingerprint, if the feature extractor

does not detect a minutia or the matcher cannot find

a match for it in the query fingerprint, then, as far as

the recognition system is concerned, that minutia

‘does not exist’ in the query fingerprint.

The effect of automated feature extractors and

matchers on minutiae persistence is practically

impossible to evaluate universally, because the

results are dependent upon which feature extraction

and matching algorithms are applied. For this

reason, we chose to conduct this study using the

latest version (6.7) of VeriFinger, a well-known and

easily available commercial feature extractor and

corresponding matcher (Neurotechnology, 2013).

The experiments described in Section 3 were

repeated for this study. The difference was that, this

time, the minutiae were extracted automatically and

the correspondences between minutiae across

different samples of the same fingerprint were also

established automatically. Figure 9 illustrates the

minutiae repeatability as the number of reference

fingerprints increases.

Figure 9: Box and whisker plots comparing the minutiae

repeatability distributions as the number of reference

fingerprints increases, when the minutiae are extracted and

matched automatically instead of manually.

Comparing Figure 9 to Figure 6, it is immediately

evident that minutiae repeatability is worse in the

case where the automatic feature extractor and

matcher are used in place of their manual

counterparts. In particular, two important

distinctions may be drawn. Firstly, while the

median in Figure 6 reaches a probability of 1 when 3

reference fingerprints are used, in Figure 9 the

median reaches is highest value of 0.99 when 7

reference fingerprints are used. Secondly, while the

interquartile range drops to 0 when 6 reference

fingerprints are used in Figure 6, the lowest

interquartile range in Figure 9 is 0.03 when 7

reference fingerprints are used. These observations

suggest that the automatic feature extractor and

matcher are not as consistent as a human expert in

identifying the minutiae and their correspondences.

For this reason, more filtering (i.e., a larger number

of reference fingerprints) is required to filter out

those minutiae that are most consistently identified.

For the sake of completeness, Figure 10 shows

the 5

and 1

percentiles of the minutiae

repeatability distributions used to generate Figure 9.

Figure 10: 5th and 1st percentiles of the minutiae

repeatability distribution for automatically extracted and

matched minutiae as the number of reference fingerprints

increases.

From Figure 10, it is evident that, when 1 reference

fingerprint is used, the probability of a minutia

repeating in another sample of the same fingerprint

is 0.78 and above 95% of the time, and 0.65 and

above 99% of the time. Compare these values to the

probabilities of 0.83 and 0.75, respectively, from

Figure 7. When 7 reference fingerprints are used,

the 5

and 1

percentiles from Figure 10 are 0.90

and 0.83, respectively. Contrast these probabilities

with 0.94 and 0.86, respectively, from Figure 7.

This analysis confirms the fact that using automatic

minutiae extraction and matching is likely to

decrease the probability of a minutia repeating in

another sample of the same fingerprint. This is

expected, because automated feature extractors and

matchers generally introduce errors of their own; so,

errors from user inconsistency, the feature extractor

and the matcher all combine to adversely affect

minutiae repeatability.

An important reason for conducting this study

was to illustrate how the results of our investigation

on user consistency can be applied in the

development and testing of automated fingerprint

0.4

0.5

0.6

0.7

0.8

0.9

1234567

Number of Reference Fingerprints Used

Box and Whisker Plots Comparing the

Distributions of the Probabilities of a Reference Minutia Repeating

when Different Numbers of Reference Fingerprints are Used

(AUTOMATIC Minutiae Extraction and Matching)

Probability of a Reference Minutia Repeating

in Another Sample of the Same Fingerprint

1 2 3 4 5 6 7

0.65

0.7

0.75

0.8

0.85

0.9

Plot Showing the 5th and 1st Percentiles of the Minutiae Repeatability Probability Distribution

as the Number of Reference Fingerprints Increases

(AUTOMATIC Minutiae Extraction and Matching)

Number of Reference Fingerprints Used

Probability above which 95% and 99% of the

Minutiae Repeatability Distribution Lies

5th

1st

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

recognition systems. Let us consider the scenario in

which only 1 reference fingerprint is used during

enrolment. From Figures 5 and 6, we can see that

there is typically (median) a 0.95 probability of a

reference minutia repeating in another sample of the

same fingerprint, when that repeatability depends

only on the user’s consistency in placing their finger

on the scanner. So, we may deduce that the

probability of a reference minutia missing in another

sample of the same fingerprint as a result of user

inconsistency alone is about 0.05. Turning our

attention to Figure 9, we can see that the typical

(median) probability of a reference minutia

repeating in another sample of the same fingerprint

when 1 reference fingerprint is used is around 0.93.

So, we may deduce that the probability of a

reference minutia missing in another sample of the

same fingerprint as a result of user inconsistency

and feature extractor errors and matcher errors is

around 0.07. Since our analysis of Figures 5 and 6

shows that user inconsistency may typically be

expected to account for about 5% of the reason for a

missing minutia, we could reasonably conclude that

the remaining 2% (or probability of 0.02 = 0.07 –

0.05) is due to errors in automated feature extraction

and matching. This tells us that, when this particular

fingerprint minutiae extractor and matcher are used,

minutiae repeatability is most heavily influenced by

user consistency. Analysis of this sort would be

extremely useful in zoning in on the most

problematic modules in a fingerprint recognition

system, which would help the designers of these

systems identify and then focus on the most crucial

area(s) of concern.

5 CONCLUSIONS

This paper investigates the probability of a reference

minutia repeating in another sample of the same

fingerprint, when the only thing that probability

depends on is the consistency with which a user

places their finger onto a fingerprint scanner. The

investigation specifically targets cooperative users in

civilian fingerprint recognition applications. To

simulate this scenario, a database of 800 fingerprint

samples from 100 cooperative users was collected.

Analysis of the database showed that, when the

reference minutiae are extracted from a single

reference fingerprint, the median probability of a

reference minutia repeating in another sample of the

same fingerprint is 0.95 with an interquartile range

of 0.04. When multiple reference fingerprints are

combined to filter out only the most reliable

reference minutiae, the probability of a reference

minutia repeating in another sample of the same

fingerprint is improved. The best result was

obtained using 7 reference fingerprints, in which

case it was found that the probability of a reference

minutia repeating in another sample of the same

fingerprint can be expected to be 0.94 and above

95% of the time, with a median probability of 1 and

an interquartile range and range of 0.

An analysis of what happens to the number of

reference minutiae as the number of reference

fingerprints increases showed a decreasing trend.

This is because using more reference fingerprints

has the effect of removing a larger number of

(unreliable) reference minutiae, so fewer reference

minutiae remain for fingerprint recognition

purposes. Our results indicate that, when users are

consistent in the placement of their finger onto a

fingerprint scanner, this loss of reference minutiae is

not very significant. Specifically, the median

number of reference minutiae dropped from 48 to 40

when 1 and 7 reference fingerprints were used,

respectively. Since 40 reference minutiae are

sufficient for a convincing fingerprint match, this

loss in the number of reference minutiae is fairly

insignificant.

While user consistency is extremely important in

ensuring that the same fingerprint features are

captured during each scan, errors in automatic

feature extraction and matching may also contribute

to minutiae persistence (repeatability). A study on a

commercial fingerprint feature extractor and matcher

confirmed that this is indeed the case, but that user

consistency is nevertheless the most influential

contributor to minutiae repeatability. This study was

used to illustrate how the results of our investigation

on user consistency can be applied towards more

rigorous development and testing of automated

fingerprint recognition systems. In particular,

knowing the likelihood of a minutia missing due to

user inconsistency will be useful for establishing the

most likely cause of a false non-match. This will

help to tease out the most problematic modules in an

automatic fingerprint recognition system.

Future work in this direction should primarily

focus on separately evaluating the influence of other

factors (e.g., fingerprint quality, feature extractor,

matcher) on minutiae persistence in the same

application scenario. The results of that work should

then be used in conjunction with the results of our

investigation to pinpoint the areas of concern in

automatic fingerprint recognition systems designed

to operate in such environments. Minutiae

persistence could also be evaluated in a number of

MinutiaePersistenceamongMultipleSamplesoftheSamePerson'sFingerprintinaCooperativeUserScenario

other fingerprint acquisition contexts (e.g.,

uncooperative user scenarios). The results would be

useful for the development of fingerprint recognition

algorithms suited to those conditions.

ACKNOWLEDGEMENTS

The authors wish to express their gratitude to the

volunteers who donated their fingerprints to our

database, without whom this investigation would not

have been possible. We would also like to thank

Vytautas Pranckenas from Neurotechnology’s

technical support team for his assistance in getting

various components of the VeriFinger 6.7 SDK up

and running. Finally, the first author wishes to

acknowledge the efforts of Robert Bowmaker,

Robert Dunn, Thomas Hahn and Maja Krivokuća in

proofreading this paper and offering constructive

criticism on its content.

REFERENCES

Biometric System Laboratory. 2006. FVC 2006

Fingerprint Verification Competition. [Online].

Available: http://bias.csr.unibo.it/fvc2006/

databases.asp. (2 December 2013).

Biometric System Laboratory. 2013. Databases. [Online].

Available:

http://biolab.csr.unibo.it/databasesoftware.asp.

[7 October 2013].

Chen, Z. & Kuo, C. H. 1991. A topology-based matching

algorithm for fingerprint authentication. In

Proceedings of the 25h Annual 1991 IEEE

International Carnahan Conference on Security

Technology. IEEE Xplore.

Fang, G., Srihari, S. N., Srinivasan, H. & Phatak, P. 2007.

Use of ridge points in partial fingerprint matching. In

Proceedings of SPIE: Biometric Technology for

Human Identification IV. SPIE.

Futronic. 2013. FS88 FIPS201/PIV Compliant USB2.0

Fingerprint Scanner. (Online). Available:

http://www.futronic-tech.com/product_fs88.html.

(2 October 2013).

Hrechak, A. K. & Mchugh, J. A. 1990. Automated

fingerprint recognition using structural matching.

Pattern Recognition, 23(8): 893-904.

Jain, A. K., Ross, A. A. & Nandakumar, K. 2011.

Fingerprint Recognition, in Introduction to Biometrics.

New York: Springer Science+Business Media LLC:

51-96.

Jea, T.-Y. & Govindaraju, V. 2005. A minutia-based

partial fingerprint recognition system. Pattern

Recognition, 38(10): 1672-1684.

Kingston, C. R. 1964. Probabilistic Analysis of Partial

Fingerprint Patterns. D.Crim. 6503017, University of

California, Berkeley.

Kryszczuk, K., Morier, P. & Drygajlo, A. 2004. Study of

the Distinctiveness of Level 2 and Level 3 Features in

Fragmentary Fingerprint Comparison, in Biometric

Authentication, edited by Maltoni, D. & Jain, A. K.

Springer Berlin Heidelberg: 124-133.

Maio, D., Maltoni, D., Cappelli, R., Wayman, J. L. & Jain,

A. K. 2002. FVC2002: Second Fingerprint

Verification Competition. In Proceedings of the 16th

International Conference on Pattern Recognition.

IEEE Xplore.

Maltoni, D., Maio, D., Jain, A. K. & Prabhakar, S. 2009a.

Fingerprint Matching, in Handbook of Fingerprint

Recognition. New York: Springer-Verlag: 167–233.

Maltoni, D., Maio, D., Jain, A. K. & Prabhakar, S. 2009b.

Introduction, in Handbook of Fingerprint Recognition.

New York: Springer-Verlag: 1-56.

Neurotechnology. 2013. VeriFinger SDK. (Online).

Available: http://www.neurotechnology.com/

verifinger. html. (25 October 2013).

Yi, C. & Jain, A. K. 2007. Dots and Incipients: Extended

Features for Partial Fingerprint Matching. In

Biometrics Symposium, 2007. IEEE Xplore.

Zhao, Q., Zhang, D., Zhang, L. & Luo, N. 2010. High

resolution partial fingerprint alignment using pore–

valley descriptors. Pattern Recognition, 43(3): 1050-

1061.

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods