Centrality of the Fingerprint Core Location

Laurenz Ruzicka

1 a

, Bernhard Strobl

1 b

, Bernhard Kohn

1 c

and Clemens Heitzinger

2 d

DSS, Austrian Institute of Technology, Gieﬁnggasse 4, Vienna, Austria

Mathematics and Geoinformation, TU Wien, Karlsplatz 13, Vienna, Austria

Keywords:

Biometrics, Fingerprint, Core, Bayesian Information Criterion, Monte Carlo Goodness-of-Fit, NFIQ 2.

Abstract:

Fingerprints have long been recognized as a unique and reliable means of personal identiﬁcation. Central to

the analysis and enhancement of ﬁngerprints is the concept of the ﬁngerprint core. Although the location of the

core is used in many applications, to the best of our knowledge, this study is the ﬁrst to investigate the empirical

distribution of the core over a large, combined dataset of rolled, as well as plain ﬁngerprint recordings. We

identify and investigate the extent of incomplete rolling during the rolled ﬁngerprint acquisition and investigate

the centrality of the core. After correcting for the incomplete rolling, we ﬁnd that the core deviates from the

ﬁngerprint center by 5.7% ± 5.2% to 7.6% ± 6.9%, depending on the ﬁnger. Additionally, we ﬁnd that the

assumption of normal distribution of the core position of plain ﬁngerprint recordings cannot be rejected, but for

rolled ones it can. We ﬁnd the non-central Fischer distribution best describes the cores’ horizontal positions.

Finally, we investigate the correlation between mean core position offset and the NFIQ 2 score and ﬁnd a weak

preference of the NFIQ 2 towards rolled recordings with a lower than central core.

1 INTRODUCTION

Fingerprints have long been recognized as a unique

and reliable means of personal identiﬁcation, play-

ing a pivotal role in forensic investigations and bio-

metric authentication systems. Central to the analy-

sis and enhancement of ﬁngerprints is the concept of

the ﬁngerprint core. The core represents a key fea-

ture within the ridge structure, and is often assumed

to be located near the center of the ﬁngerprint [Man-

hua et al., 2005]. Understanding the location of the

core and its relationship to the ﬁnger position is of

great importance for various purposes, such as tem-

plate alignment [Manhua et al., 2005], ﬁnger type

classiﬁcation [Karu and Jain, 1996], and pose correc-

tion [Tan and Kumar, 2020, Ruzicka et al., 2023].

Another application that requires empirical

knowledge of the core distribution is the process

of generating synthetic ﬁngerprints. The goal of

generating a synthetic ﬁngerprint is to mimic the

natural properties of a real ﬁngerprint to an extend

that it is no longer possible to differentiate the two.

Therefore, the research question addressed in this

https://orcid.org/0000-0002-0823-9601

https://orcid.org/0000-0002-8920-0468

https://orcid.org/0000-0002-3177-3159

https://orcid.org/0000-0003-1613-5164

paper is: What is the distribution of the ﬁngerprint

core? To the best of our knowledge, this study is the

ﬁrst to analyse the empirical distribution of the ﬁnger-

print core on a large dataset of both rolled and plain

ﬁngerprint recordings. By investigating the distribu-

tion and therefore also the deviation of the core from

the ﬁngerprint center, we aim to gain a deeper un-

derstanding of the inherent structure of ﬁngerprints.

Additionally, we examine the linkage between in-

complete rolling and deviations of the core from the

segmented ﬁngerprint image center to gain insights

into the reliability of rolled ﬁngerprint representa-

tions. Furthermore, we investigate the distribution of

the cores’ horizontal and vertical positions. Finally,

we analyze the correlation between the NFIQ 2 score

and the mean core offset.

In the following sections, we will present related

work, our methodology, datasets used for analysis,

core detection technique, and the results obtained.

Also, we will discuss the implications of our ﬁnd-

ings and their signiﬁcance for ﬁngerprint analysis and

identiﬁcation. By shedding light on the deviation of

the ﬁngerprint core from the center, this study aims to

contribute to the advancement of forensic science and

biometric technologies.

Ruzicka, L., Strobl, B., Kohn, B. and Heitzinger, C.

Centrality of the Fingerprint Core Location.

DOI: 10.5220/0012309300003657

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 1, pages 713-720

ISBN: 978-989-758-688-0; ISSN: 2184-4305

713

1.1 Related Work

In the ﬁeld of ﬁngerprint analysis, there has been ex-

tensive research devoted to the deﬁnition and detec-

tion of the ﬁngerprint core as well as to the classiﬁca-

tion of ﬁngerprints.

In the existing literature, two primary methods for

deﬁning the core point in ﬁngerprints are prominent.

The ﬁrst method, proposed by Srinivasan and Murthy

[Srinivasan and Murthy, 1992] utilizes ridge orienta-

tion to determine the core. According to their deﬁ-

nition, the core is identiﬁed as the location with the

maximal ridge curvature. The second way of deﬁning

a core uses the notion of the innermost closed ridge

loop and is described in [Bahgat et al., 2013] by Bah-

gat, Khalil, Abdel and Mashali as well as in ISO/IEC

19794-1:2011 3.33. They deﬁne the core as being

the topmost point of the innermost ridge line. Note

that there can also be two cores inside one ﬁngerprint

in the case of a ﬁngerprint of the double loop type,

which has two innermost closed ridge loops. Also,

there can be none for the case of a ﬁngerprint of the

arch type [Ametefe et al., 2022].

In addition to the detection of the core position

within a ﬁngerprint, it is crucial to consider the con-

text of the ﬁngerprint within the image. Fingerprint

images may not always be centered or aligned uni-

formly, which necessitates image segmentation tech-

niques to extract the relevant ﬁngerprint region and

make the core position relative to the ﬁngerprint it-

self, rather than the image as a whole.

A variety of image segmentation approaches have

been employed in ﬁngerprint analysis to address this

challenge. Tomaz, Candeias and Shahbazkia analyse

the pixel color information in the ST space [Tomaz

et al., 2004], while Bazen and Gerez use custom pixel

features to segment the ﬁngerprint image [Bazen and

Gerez, 2001]. More recent work has used a deep

learning approach to ﬁngerprint segmentation [Mur-

shed et al., 2022, Grosz et al., 2021].

An established open-source tool for ﬁngerprint

segmentation is the nfseg tool from the NBIS toolset.

This tool was developed by the National Institute of

Standards and Technology (NIST) and provides a re-

liable, fast and accurate way of segmenting ﬁngerprint

images.

2 METHODS

2.1 Core Detection

In this work, we follow the ISO core deﬁnition to

locate the core positions and use the state-of-the-art,

commercially available, automated ﬁngerprint identi-

ﬁcation system IDKit from Innovatrix.

2.2 Fingerprint Segmentation &

Relative Core Offset

To evaluate the stability of the core’s centrality in a

ﬁngerprint, a multi-step approach is employed. The

ﬁrst step involves segmenting and cutting out the ﬁn-

gerprint, which is accomplished using the nfseg tool.

Additionally, nfseg rotates the ﬁngerprint image, such

that the major symmetry axis is aligned with the y-

axis.

Once the ﬁngerprint image is appropriately seg-

mented and the image is cropped to the ﬁngerprint

region, the core points are detected within the im-

age. If multiple cores are detected, we calculate the

mean position of the cores and use this mean for fur-

ther calculations. To assess the position of the core

within the ﬁngerprint, its offset from the central point

of the cropped image is measured. We call this dis-

tance core offset o

core

and o

core

, where x is the index

for the horizontal component and y the index for the

vertical component.

To facilitate the evaluation of core positions in a

standardized manner, a normalization process is em-

ployed. In order to increase readability, we depict

analogous expressions for the y-axis in this section

in parenthesis. The measured core offsets o

core

)

are divided by half the width (height) of the cropped

ﬁngerprint w

cropped

, resulting in a score that repre-

sents the core’s centrality. This score ranges from -1

to 1, indicating the relative position of the core within

the ﬁngerprint. A score of -1 suggests that the core

is located at the left (upper) edge of the ﬁngerprint,

while a score of 1 indicates its proximity to the right

(lower) edge. We call this score the relative core off-

set ro

core

(ro

core

), as shown in Equation 1.

core

cropped

(1)

and analogous for the y-coordinate.

In order to investigate the variability of the cen-

trality of the core, we calculate the mean of the abso-

lute values of the relative core offsets aro

core

(aro

core

)

for each ﬁnger position f , as shown in Equation 2.

aro

core

( f ) =

∑

|ro

core;i

|, (2)

where N is the number of cores for a given ﬁnger

position in the database and i is the index of the core.

The calculation for the y-coordinate is analogous.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

714

2.3 Incomplete Rolling

While recording rolled ﬁngerprints, we observed that

the user often does not roll the ﬁnger completely

from nail-to-nail, but misses the last part of the ﬁn-

ger. This implies that the distance along the x-axis

from the core in the collected ﬁngerprint to the ﬁn-

gerprint boarders is biased towards the starting side

of the rolling process. Especially for the case of op-

erator assisted rolling, where an experienced opera-

tor controls the acquisition process and routinely rolls

from a preferred side for a given left or right hand, a

bias b

core

for the left vs right side can be expected. If

there is no preferred rolling direction and the com-

plete nail-to-nail rolling process was not enforced,

the bias vanishes but larger relative core offset val-

ues are expected. Note that the x-axis is deﬁned for

the rotated and cropped ﬁngerprint image and there-

fore aligns with the rolling direction and not the sen-

sor recording surface. This bias b

core

can be measured

by the mean relative core offset, as seen in equation 3.

core

∑

core;i

(3)

A non vanishing relative core offset implies one of

two things: Either the incomplete rolling is the cause

for the bias or there is an anatomic tendency for a ﬁn-

gerprint core for a given ﬁnger to sit further on one

side than on the other.

In this work, we assume that the core for a

given participant is sampled from a distribution with

zero mean, i.e. that the core is positioned evenly

around the ﬁngerprint center, and that therefore a non-

vanishing bias is an indication for a statistical error

induced by the incomplete rolling. We analyse this

claim by comparing biases for plain and rolled ﬁnger-

prints.

In order to correct for a non-vanishing bias, we

calculate the corrected relative core offset caro

core

for

each ﬁnger position f . For this calculation, we sub-

tract the found bias, i.e. the mean relative core offset,

from the calculated relative core offset, resulting in

the corrected mean of absolve values of the relative

core offsets caro

core

, as shown in Equation 4.

caro

core

( f ) =

∑

|ro

core;i

− b

core

( f )| (4)

2.4 Distribution of Core Positions

In order to determine the underlying distribution

of the core positions for each ﬁnger, the initial

step was assessing the normality of the data. The

Anderson-Darling test [Anderson and Darling, 1952]

was speciﬁcally chosen as the preferred method due to

its high power and sensitivity in detecting departures

from normality. Should the assumption of normal-

ity not be met, an alternative approach was employed

to identify a suitable set of distributions that best de-

scribe the data. This was achieved by employing

the Bayesian Information Criterion (BIC) [Schwarz,

1978], which combines the complexity of a model

with its performance into one score. It chose from

a set of over 110 distributions and selected the overall

best matches. In the next step, a Generalized Monte

Carlo goodness-of-ﬁt procedure was used to ﬁnd the

best distribution of the previously selected set of dis-

tributions for each ﬁnger position. This procedure

is often described as the parametric bootstrap test in

literature [Stute et al., 1993, Kojadinovic and Yan,

2012].

2.5 NFIQ 2 Evaluation

The NIST Fingerprint Image Quality (NFIQ) is a soft-

ware tool designed for assessing the quality of ﬁnger-

print images in the context of biometric identiﬁcation

and veriﬁcation systems [Tabassi et al., 2021].

To assess the relationship between the NFIQ 2

score and the core offset in both the X and Y di-

rections, we employed Spearman’s correlation coef-

ﬁcient [Spearman, 1904].

In our analysis, the null hypothesis (H0) posits

that there is no signiﬁcant correlation between the

NFIQ 2 scores and the core offset. The alternative hy-

pothesis (Ha) suggests that there exists a signiﬁcant

correlation between the NFIQ 2 scores and the core

offset. We investigate the hypothesis separately for

horizontal and vertical core offset. Additionally, we

also investigate the correlation between the aro

core

aro

core

and the NFIQ 2.

For the hypothesis test, the H0 can be transformed

to a Student’s t distribution with the number of sam-

ples minus two as the degrees of freedom [Howell,

2013, p. 280]. Note that this is only accurate for over

500 observations [Community, ].

To determine the statistical signiﬁcance of our

ﬁndings, we set the signiﬁcance level (α) for the hy-

pothesis test to 10

−3

2.6 Datasets

This study uses 6 different datasets to investigate the

centrality of the core. Some of the datasets provide

a ﬁnger position description for each of the recorded

ﬁngerprints, which we denote in the FGP scheme in-

troduced by NIST [National Institute of Standards and

Technology, 2000, p. 18]:

Centrality of the Fingerprint Core Location

715

Table 1: FGP values mapped to ﬁnger names.

Finger Right Left

Thumb 1 6

Index 2 7

Middle 3 8

Ring 4 9

Little 5 10

Plain Thumb 11 12

Unknown 0 0

Table 2: Number of rolled ﬁngerprints with found cores.

FGP AIT 300a 302a 302b

1 1083 815 935 521

2 875 801 981 542

3 528 819 981 605

4 523 832 984 632

5 941 803 952 616

6 1077 797 889 505

7 512 806 963 575

8 522 813 942 609

9 525 816 993 630

10 525 791 940 603

Table 3: Number of plain ﬁngerprints with found cores.

FGP 300a 302b Neurotechnology PolyU

0 0 0 840 2776

1 0 226 0 0

2 704 228 0 0

3 662 255 0 0

4 768 270 0 0

5 653 244 0 0

11 779 11 0 0

6 0 236 0 0

7 771 236 0 0

650 234 0 0

9 779 257 0 0

10 675 231 0 0

12 748 151 0 0

Of the 6 datasets used, the NIST Special Datasets

300a [Fiumara et al., 2018], 302a and 302b [Fiumara

et al., 2021], the Neurotechnology dataset and the

PolyU contact-based dataset [Lin and Kumar, 2018]

are publicly available, while the AIT Dataset [Weis-

senfeld et al., 2022] is an in-house dataset.

2.6.1 Rolled Combination

For the combined results of the rolled ﬁngerprints, we

used the datasets: AIT Dataset, NIST 300a (rolled

images only), 302a and 302b (rolled images only).

Since we use the normalized core offset for our calcu-

lations, we were able to combine datasets in various

resolutions. Table 2 shows the number of ﬁngerprint

recordings for each ﬁnger position used. In total, we

had 30602 rolled ﬁngerprint recordings.

2.6.2 Plain Combination

For the combined results of the plain ﬁngerprints we

used the following datasets: NIST Special Dataset

300a (plain images only), NIST Special Dataset 302b

(slap-segmented images only), Neurotechnology and

PolyU Contact-based. A ﬁnger-wise overview of the

number of recordings per ﬁnger can be seen in table

3. In total, we had 9568 plain ﬁngerprint recordings.

3 RESULTS & DISCUSSION

3.1 Bias & Incomplete Rolling

For the rolled ﬁngerprints, we observed the follow-

ing bias values, described as percentages of the ro

core

score in Table 4. We ﬁnd that the right hand with ﬁn-

gers 1-5 shows a positive bias, i.e. a tendency of the

core to sit closer to the right side. This implies that

we measured incomplete rolling where the ﬁnger was

rolled starting with the right ﬁnger-nail edge. This

effect was especially strong for middle, ring and little

ﬁnger, with the strongest effect on the ring ﬁnger. One

explanation for this could be obstruction of the non-

rolled ﬁngers in the rolling process. Another expla-

nation could be the inter-connectivity of the tendons

passing the carpal tunnel as well as the connection of

the ﬂexor digitorum superﬁcialis to all ﬁngers except

the thumb. This makes it challenging to lift only the

ring ﬁnger and to a lesser extend also the middle ﬁn-

ger for most people. Therefore, the effort of separat-

ing the ring or middle ﬁnger from the other ﬁngers is

increased, which could lead to a faster termination of

the rolling process.

On the other side, the left hand also shows the

pattern that the ring ﬁnger and the middle ﬁnger cre-

ate the largest bias. The same reasoning of the hand

anatomy, leading to the observed bias, could apply

here. Also note the sign ﬂip, indicating that the pre-

ferred starting edge of the rolling process changed. In

future work, the effect of inverted roll directions and

left versus right handiness on the bias could be ex-

plored with a dedicated dataset.

Additionally, also the plain ﬁngerprint recordings

showed a non-vanishing mean of the core x-position,

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

716

Table 4: Rolled Fingerprint Bias for Average Core X-

Positions.

FGP Bias [%] FGP Bias [%]

1 2.90 6 0.42

2 0.67 7 0.33

3 3.27 8 -1.11

4 5.17 9 -3.62

5 4.89 10 -0.50

Table 5: Mean Core X-Position Plain Recordings for Aver-

age Core Positions.

FGP Offset [%] FGP Offset [%]

1 -1.50 6 2.35

2 -6.43 7 11.69

3 -5.67 8 9.79

4 10.01 9 -1.64

5 -1.53 10 15.49

11 7.21 12 -0.72

0 2.32 - -

similar to the bias of the rolled ﬁngerprints. The mag-

nitude of this can be seen in Table 5.

One observation of the mean ﬁngerprint core in

the plain recording setting is the mirrored mean core

x-position shift. The sign of the offset of the core

is ﬂipped when comparing the left to the right hand.

This could be an indication that the hand is placed

tilted on the sensor, i.e. rotated along the ﬁnger axis,

where the tilt between left and right hand is inverted.

Another key observation is the comparison of the

bias of the rolled ﬁngerprints and the mean core po-

sition for plain ﬁngerprint recordings. The tendency

of the core in rolled ﬁngerprints of the right hand was

to sit closer to the right edge, while for plain record-

ings, only the thumb and the ring ﬁnger cores had the

tendency to sit closer to the right edge. Also for the

left hand, only the thumb and index ﬁnger cores both

in plain and rolled ﬁngerprints tend towards the right

side of the ﬁnger. This is an indication that the reason

for the non-zero core position offset is not an effect

of human ﬁnger anatomy, but rather of the recording

process.

Finally, also the y-position of the mean ﬁngerprint

core offset can be calculated. For plain/rolled ﬁnger-

print recordings, we ﬁnd the following offsets from

the image center in percentages of the total ro

core

, -

100% depicts the upper ﬁnger edge and 100% the

lower ﬁnger edge: 3.37%/15.94% for right thumbs,

-8.39%/-5.28% for right index ﬁngers, -16.80%/-

6.93% for right middle ﬁngers, -14.70%/-8.36% for

right ring ﬁngers, -13.10%/-13.33% for right little ﬁn-

gers, 5.15%/15.03% for left thumbs, -9.66%/-6.83%

for left index ﬁngers, -19.75%/-8.37% for left ring ﬁn-

gers and -12.60%/-13.15% for left little ﬁngers. For

the unknown ﬁnger position we found an offset of

5.18%/-.

The direction of the y-offset of the mean core

position agrees for all 10 ﬁngers between plain and

rolled recordings. This could be due to the fact that

plain recordings in the observed dataset are taken

from the same participants as the rolled recordings,

with the exception of the PolyU and Neurotechnology

datasets, which had no ﬁnger information.

3.2 CARO & ARO

The caro

core

from equation 4 is an important mea-

sure for how strongly the core is scattered around the

bias position. It is the central indication of how sta-

ble the centrality of the core position is. Its non-bias

corrected counterpart, the aro

core

from equation 2 is

an important measurement for the expected variabil-

ity in a real recording environment, where incomplete

rolling and therefore a non-vanishing bias is to be ex-

pected.

We found the following caro

core

scores for

plain/rolled recordings, which are written as percent-

ages of the ro

core

score. This score can vary from 0%

(no deviation, every core in center) to 100% (maximal

deviation, every core at the ﬁnger edge). Additionally,

we report the aro

core

scores for comparison in paren-

thesis. This can be seen in Table 6.

Table 6: CARO (ARO) Scores for Rolled Fingerprints.

FGP Scores [%] FGP Scores [%]

1 6 ± 5 (7 ± 5) 6 7 ± 6 (7 ± 6)

2 7 ± 6 (8 ± 7) 7 7 ± 6 (7 ± 6)

3 6 ± 5 (7 ± 6) 8 7 ± 5 (7 ± 6)

4 5 ± 5 (6 ± 5) 9 6 ± 5 (6 ± 5)

5 6 ± 5 (7 ± 5) 10 6 ± 5 (7 ± 5)

One key remark is that the scattering of the core

around the ﬁngerprint center is similar between all

ﬁngers and falls within a range of 5.7% ± 5.2% to

7.6% ± 6.9 for the different ﬁngers. This quantiﬁes

the centrality of the core as a landmark for various

ﬁngerprint enhancement techniques, such as pose cor-

rection for contactless ﬁngerprint recordings, where

the centrality of the core is the assumption for calcu-

lating the viewing angle.

For the aro

core

for rolled ﬁngerprints, we found

scores from 9% ± 6% for the right index ﬁnger to

12% ± 8% for the right thumb. Both thumbs had the

highest values and the index ﬁngers the lowest. The

other ﬁngers were all in the range of around 10%.

Centrality of the Fingerprint Core Location

717

This indicates that the core varies in height roughly

equally for all ﬁngers, with a slight increase for the

thumbs and decrease for the index ﬁngers.

3.3 Distribution of the Core

The Anderson-Darling test revealed that the assump-

tion of normality of the caro

core

data could be rejected

with a signiﬁcance level of under 10

−16

for all rolled

ﬁngerprints. For the plain recordings, normality of

caro

core

could only be rejected for the unlabeled ﬁn-

gers with a FGP of 0 and the caro

core

for all ﬁngers

except FGP 12. Therefore, we continued the testing

with the BIC for the rolled ﬁngerprint recordings in

both axis, as well as the vertical core position of the

plain ﬁngerprints.

The overall best BIC scores could be achieved

with the Logistic distribution. Other well perform-

ing distributions that were added to the set of candi-

date distributions are: Laplace [Abramowitz and Ste-

gun, 1972, p. 930], Cauchy [Abramowitz and Ste-

gun, 1972, p. 930], Dagum (also known as Mielke)

[Dagum, 1999], non-central Fischer (NCF) [Caba

na,

2011, Ramirez, 2004], Burr [Burr, 1942], Normal and

the Lognormal distribution [Galton, 1997].

Of the eight candidate distributions, Logistic,

Laplace and Cauchy distributions could be rejected

with a signiﬁcance level of 5% for all ﬁngers. Both

Dagum and NCF distribution achieved high p-values

of up to 0.94 for the NCF and 0.70 for Dagum.

The results for all the ﬁngers can be seen in Ta-

ble 7. In most cases, the NCF distribution performed

best. For rolled ﬁngerprints, all ﬁngers could best be

described with either the NCF or the Burr distribution.

For both plain as well as rolled ﬁngerprint recordings,

all cases where the NCF distribution did not have the

highest goodness-of-ﬁt p-value, it was the runner up

with the second highest p-value.

We therefore conclude that the NCF is best suited

to describe the core position and suggest to implement

those for the generation of synthetic ﬁngerprint sam-

ples.

3.4 NFIQ 2 Correlation

In this section, we present the results of our analysis,

which aimed to investigate the correlation between

the mean ro

core

, mean ro

core

and the NFIQ 2 scores,

as well as the correlation between both aro

core

scores

and the NFIQ 2 scores.

For rolled ﬁngerprint recordings, we found a sig-

niﬁcant, positive correlation of 0.23 between the

mean ro

core

and the NFIQ 2 scores. For the x compo-

nent of the mean core offsets of the rolled ﬁngerprint

0.1 0 0.0 5 0.00 0 .05 0.1 0 0.1 5

Correlat ion Mean Core Offset X - NFIQ 2

0.1 5

0.1 0

0.0 5

0.0 0

0.0 5

0.1 0

0.1 5

0.2 0

0.2 5

Correlat ion Mean Core Offset Y - NFIQ 2

-1, -1

-20 , -20

-16 , -39

-11 , -39

-4, -14

-2, -2

-8, -15

-2, -17

-2, -20

-1, -7

-2, -2

-1, -2

-1, -3

-2, -1

-1, -2

-1, -3

-2, -4

-3, -1

-3, -2

-1, -5

-1, -3

-7, -30

-6, -6

Mean Core Offset Correlation w it h NFIQ 2

Mode

Rolled

Plain

FGP

All

Figure 1: Correlations between the Mean Core Offset in X

and Y direction and the NFIQ 2 score for rolled (blue) and

plain (orange) ﬁngerprint recordings, for each ﬁnger posi-

tion (see markers in legend).

recordings, as well as both components of the mean

core offsets of plain ﬁngerprint recordings, no indica-

tive correlation with a magnitude of over 0.06 could

be found.

Therefore, we conclude that for rolled ﬁngerprint

recordings, a non-zero core offset in the vertical ﬁn-

gerprint direction where the core is shifted to the

lower part of the image is favoured by the NFIQ 2

score. Furthermore, we observed a very small cor-

relation of -0.09 between the aro

core

and the NFIQ 2

score. Therefore, there seems to be a preference of the

NFIQ 2 score for core locations on the lower halve of

the image, with a tendency to prefer cores closer to

the center.

Interestingly, we did not observe a relevant corre-

lation in x direction between the mean core offset and

the NFIQ 2 score. Also the correlation of the aro

core

is not statistically signiﬁcant. This implies that there

is either no preferred side by the NFIQ 2 score or that

the horizontal component of the core position does

not strongly inﬂuence the NFIQ 2 score.

A more detailed analysis on a ﬁnger-wise basis

can be seen in Figure 1. Here, the correlation results

for plain and rolled ﬁngerprint recordings for each ﬁn-

ger are depicted. The labels are the exponents of the

p-values of the Spearman’s correlation test with 10 as

their basis. The ﬁrst number describes the p value of

the correlation in x direction and the second in y di-

rection.

Notable, with the exception of both thumbs in

plain recordings, the correlation between the mean

core

and the NFIQ 2 score is close to zero or pos-

itive. And for the thumbs in plain recordings, the re-

sults are not statistically signiﬁcant with their p val-

ues of above 10

−3

. More data is required to come to

a conclusion regarding the thumbs in plain ﬁngerprint

recordings.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

718

Table 7: Best ﬁtting distributions - Dagum (Dag), Lognorm (LNo), Burr (Bur), Norm (Nor).

- Plain Rolled

- x y x y

FGP Dist. p [%] Dist. p [%] Dist. p [%] Dist. p [%]

0 Dag 70 NCF 78 - - - -

1 LNo 67 NCF 20 Bur 11 NCF 95

2 LNo 56 Bur 60 NCF 32 NCF 71

3 Bur 47 NCF 80 NCF 37 Bur 37

4 NCF 94 NCF 55 NCF 85 Bur 29

5 NCF 50 NCF 73 NCF 84 Bur 42

6 NCF 84 NCF 44 NCF 65 Bur 36

7 NCF 56 Bur 55 NCF 73 NCF 74

8 NCF 28 NCF 54 NCF 33 NCF 30

9 NCF 67 Bur 31 NCF 60 NCF 78

10 Nor 75 NCF 67 Bur 41 NCF 61

11 NCF 33 Bur 43 - - - -

12 Nor 72 NCF 23 - - - -

4 CONCLUSION

We found a non-vanishing bias in our combined rolled

dataset. This effect was especially strong for middle

(right 3%, left -1%), ring (right 5%, left -4%) and

little (right 5%, left -1 %) ﬁnger, with the strongest

effect on the ring ﬁnger. Furthermore, we also ob-

served a non-vanishing bias for the plain recording

setting, but comparing rolled with plain recordings,

we found that the preferred side is not linked to the

ﬁnger position value. This indicates that the ﬁnger-

print core position does not have an anatomically pre-

ferred side but rather that the non-vanishing bias orig-

inates from incomplete rolling in the case of rolled

ﬁngerprint recordings and placing that hand tilted on

the sensor for the plain ﬁngerprint recordings.

In addition to this ﬁnding, we measured the vari-

ability of the core position, given as caro

core

and

aro

core

. We found the caro

core

to be around 6% to

8%, depending on the ﬁnger position. This indicates

that the core point is stable enough to be used as a

reference point for various applications.

Both of those ﬁndings are especially interesting

for the ﬁeld of contactless ﬁngerprint analysis, where

the perspective of the camera has to be accounted

for. A non-vanishing core offset originating from an

anatomical preference of the core towards one side

would require a re-calibration of the role of the ﬁnger-

print core as a reference point. Also, the magnitude at

which the core scatters around the central position of

the segmented ﬁngerprint image is crucial. A too un-

stable core position would make the core unreliable

as reference point.

Furthermore, we found the non-central Fischer

distribution to be the best matching distribution for

the core’s x and y position. We think that this ﬁnding

can help improve the quality of synthetically gener-

ated ﬁngerprints and further increase the understand-

ing of the human ﬁngerprint.

Finally, we found a correlation between the verti-

cal offset of the core from the ﬁngerprint center and

the NFIQ 2 score. The NFIQ 2 scores favors core po-

sitions in rolled ﬁngerprint recordings where the core

sits close to the image center, but slightly off towards

the lower image edge. Interestingly, we did not ﬁnd

any correlation of the NFIQ 2 score and the core off-

set in the horizontal direction.

ACKNOWLEDGEMENTS

We gratefully acknowledge Dr. Schmid from the

Austrian BMI for the continuous support during the

recording sessions as well as the Bundesamt f

Sicherheit in der Informationstechnik (BSI) and Jan-

nis Priesnitz from Hochschule Darmstadt for their

valuable discussions and support in this research.

REFERENCES

Abramowitz, M. and Stegun, I. (1972). Handbook of Math-

ematical Functions With Formulas, Graphs and Math-

ematical Tables. Number 55 in Applied Mathematics.

National Bureau of Standards, Washington, DC, USA,

10 edition.

Centrality of the Fingerprint Core Location

719

Ametefe, D., Seroja, S., mohd ali, D., and Muhammad, Z.

(2022). Fingerprint pattern classiﬁcation using deep

transfer learning and data augmentation. The Visual

Computer, 39:1–14.

Anderson, T. W. and Darling, D. A. (1952). Asymptotic

Theory of Certain ”Goodness of Fit” Criteria Based

on Stochastic Processes. The Annals of Mathemati-

cal Statistics, 23(2):193–212. Publisher: Institute of

Mathematical Statistics.

Bahgat, G. A., Khalil, A. H., Abdel Kader, N. S., and

Mashali, S. (2013). Fast and accurate algorithm for

core point detection in ﬁngerprint images. Egyptian

Informatics Journal, 14(1):15–25.

Bazen, A. M. and Gerez, S. H. (2001). Segmentation of

Fingerprint Images.

Burr, I. W. (1942). Cumulative Frequency Functions.

The Annals of Mathematical Statistics, 13(2):215–

232. Publisher: Institute of Mathematical Statistics.

Caba

na, E. M. (2011). F Distribution. In Lovric, M., edi-

tor, International Encyclopedia of Statistical Science,

pages 499–501. Springer, Berlin, Heidelberg.

Community, S. scipy.stats.spearmanr — SciPy v1.11.3

Manual.

Dagum, C. (1999). A Study on the Distributions of Income,

Wealth, and Human Capital. Revue europ

eenne des

sciences sociales, 37(113):231–268. Publisher: Li-

brairie Droz.

Fiumara, G., Flanagan, P., Grantham, J., Bandini, B., Ko,

K., and Libert, J. (2018). NIST special database 300:

uncompressed plain and rolled images from ﬁnger-

print cards. Technical Report NIST TN 1993, National

Institute of Standards and Technology, Gaithersburg,

MD.

Fiumara, G., Schwarz, M., Heising, J., Peterson, J., Flana-

gan, P., and Marshall, K. (2021). NIST Special

Database 302: Supplemental Release of Latent Anno-

tations. Technical report, National Institute of Stan-

dards and Technology.

Galton, F. (1997). XII. The geometric mean, in vital and

social statistics. Proceedings of the Royal Society of

London, 29(196-199):365–367. Publisher: Royal So-

ciety.

Grosz, S. A., Engelsma, J. J., Liu, E., and Jain, A. K. (2021).

C2CL: Contact to Contactless Fingerprint Matching.

arXiv:2104.02811 [cs, eess].

Howell, D. C. (2013). Statistical methods for psychology.

Cengage, Boston, MA, eighth edition edition.

Karu, K. and Jain, A. K. (1996). Fingerprint classiﬁcation.

Pattern Recognition, 29(3):389–404.

Kojadinovic, I. and Yan, J. (2012). Goodness-of-ﬁt testing

based on a weighted bootstrap: A fast large-sample al-

ternative to the parametric bootstrap. Canadian Jour-

nal of Statistics, 40.

Lin, C. and Kumar, A. (2018). Matching Contactless and

Contact-Based Conventional Fingerprint Images for

Biometrics Identiﬁcation. IEEE Transactions on Im-

age Processing, 27(4):2008–2021.

Manhua, L., Jiang, X., and Kot, A. (2005). Fingerprint

Reference-Point Detection. EURASIP Journal on Ad-

vances in Signal Processing, 2005.

Murshed, M. G. S., Kline, R., Bahmani, K., Hussain, F., and

Schuckers, S. (2022). Deep Slap Fingerprint Segmen-

tation for Juveniles and Adults. arXiv:2110.04067 [cs,

eess].

National Institute of Standards and Technology (2000).

American national standard for information systems.

Technical Report NIST SP 500-245, National Insti-

tute of Standards and Technology, Gaithersburg, MD.

Edition: 0.

Ramirez, D. (2004). Noncentral generalized F distributions

with applications to joint outlier detection. Computing

Science and Statistic, 36:801–825.

Ruzicka, L., Kohn, B., S

ollinger, D., Heitzinger, C., Strobl,

B., and Uhl, A. (2023). Improving Sensor Inter-

operability between Contactless and Contact-based

Fingerprints using Pose Correction & Unwarping.

[Manuscript submitted for publication].

Schwarz, G. (1978). Estimating the Dimension of a Model.

The Annals of Statistics, 6(2):461–464. Publisher: In-

stitute of Mathematical Statistics.

Spearman, C. (1904). The Proof and Measurement of Asso-

ciation between Two Things. The American Journal

of Psychology, 15(1):72–101. Publisher: University

of Illinois Press.

Srinivasan, V. S. and Murthy, N. N. (1992). Detection of

singular points in ﬁngerprint images. Pattern Recog-

nition, 25(2):139–153.

Stute, W., Manteiga, W. G., and Quindimil, M. P. (1993).

Bootstrap based goodness-of-ﬁt-tests. Metrika,

40(1):243–256.

Tabassi, E., Olsen, M., Bausinger, O., Busch, C., Figlarz,

A., Fiumara, G., Henniger, O., Merkle, J., Ruhland,

T., Schiel, C., and Schwaiger, M. (2021). NIST Fin-

gerprint Image Quality 2. NIST. Last Modiﬁed: 2022-

11-29T09:11-05:00 Publisher: Elham Tabassi, Martin

Olsen, Oliver Bausinger, Christoph Busch, Andrew

Figlarz, Gregory Fiumara, Olaf Henniger, Johannes

Merkle, Timo Ruhland, Christopher Schiel, Michael

Schwaiger.

Tan, H. and Kumar, A. (2020). Towards More Accu-

rate Contactless Fingerprint Minutiae Extraction and

Pose-Invariant Matching. IEEE Transactions on Infor-

mation Forensics and Security, 15:3924–3937. Con-

ference Name: IEEE Transactions on Information

Forensics and Security.

Tomaz, F., Candeias, T., and Shahbazkia, H. (2004). Fast

and accurate skin segmentation in color images. In

First Canadian Conference on Computer and Robot

Vision, 2004. Proceedings., pages 180–187.

Weissenfeld, A., Schmid, R., Kohn, B., Strobl, B., and

Dom

ınguez, G. F. (2022). Case study of the acquisi-

tion of contactless ﬁngerprints in a real police setting.

In 2022 International Conference of the Biometrics

Special Interest Group (BIOSIG), pages 1–5. ISSN:

1617-5468.

BIOSIGNALS 2024 - 17th International Conference on Bio-inspired Systems and Signal Processing

720