This document analyzes the interoperability performance of fingerprint sensors from different vendors and technologies. It collected fingerprint images from 190 subjects using 8 different sensors. The sensors used different technologies like optical, thermal, and capacitive, and different interaction types like swipe and touch. The study analyzed minutiae count, image quality scores, and false non-match rates (FNMR) to evaluate interoperability between native and cross-sensor datasets. Statistical tests found significant differences in minutiae counts and image quality scores between datasets. Tests of FNMR proportions found some interoperable dataset pairs had significantly different error rates, while others did not, without a clear trend based on sensor technology or interaction type. The lowest observed interoper
Deploy with confidence: VMware Cloud Foundation 5.1 on next gen Dell PowerEdg...
(2009) Statistical Analysis Of Fingerprint Sensor Interoperability
1. Statistical Analysis of Fingerprint Sensor Interoperability
Performance
Shimon K. Modi, Stephen J. Elliott, and Hale Kim
Abstract—The proliferation of networked authentication distributed & multi-vendor architectures become more
systems has put focus on the issue of interoperability. pervasive.
Fingerprint sensors are based on a variety of different This paper examines interoperability from the perspective
technologies that introduce inconsistent distortions and of fingerprint sensors. This research defined fingerprint sensor
variations in the feature set of the captured image, which makes interoperability as the ability to match fingerprints of the same
the goal of interoperability challenging. The motivation of this individual collected from different sensors. Fingerprint
research was to examine the effect of fingerprint sensor
interoperability on the performance of a minutiae based
sensors are based on a variety of different technologies like
matcher. A statistical analysis framework for testing electrical, optical, thermal etc. The physics behind these
interoperability was formulated to test similarity of minutiae technologies introduces inconsistent distortions and variations
count, image quality and similarity of performance between in the feature set of the captured image, which makes the goal
native and interoperable datasets. False non-match rate (FNMR) of interoperability even more challenging. A fingerprint
was used as the performance metric in this research. recognition system deployed in a distributed architecture can
Interoperability performance analysis was conducted on each benefit from gaining a deeper insight into interoperability of
sensor dataset and also by grouping datasets based on the sensors and its effect on error rates. The purpose of this
acquisition technology and interaction type of the acquisition
research was to examine the effect of sensor dependent
sensor. The lowest interoperable FNMR observed was 0.12%.
variations and distortions, and characteristics of the sensor on
the interoperability matching error rates of minutiae based
I. INTRODUCTION
fingerprint recognition systems. This study focused on an
Authentication of individuals is a process that has been exclusive aspect of the problem space - the acquisition
performed in one form or another since the beginning of subsystem. The end objective of this study was to provide
recorded history. Whilst establishing and maintaining the greater insight into the effect of a fingerprint dataset acquired
identity of individuals has been ongoing since this time from various sensors on performance measured in terms of
accurate automated recognition is becoming increasingly false non match rates (FNMR).
important in today’s networked world. As technology
advances, the complexity of these tasks has also increased. II. LITERATURE REVIEW
Digital identities and electronic credentialing have changed
Performance of a biometric system can be affected by the
the way authentication architectures are designed. Instead of
errors introduced in the data capture subsystem of the general
stand-alone and monolithic authentication architectures of the
biometric model [1]. These factors can be attributed to either
past, today’s networked world offers the advantage of
the user or sensor. The human interaction error – that is the
distributed and federated authentication architectures. The
interaction between the individual and the sensor is described
development of distributed authentication architectures can be
in depth in [2]. Whilst [2] discusses the human biometric
seen as an evolutionary step, but also raises the issue always
sensor interaction, this paper examines the variability
accompanied by an attempt to mix disparate systems:
introduced by the sensor. Sensor variability results in moving
Interoperability. This issue is of relevance to all kinds of
the distribution of genuine scores away from the origin,
authentication mechanisms, and biometric recognition
thereby increasing error rates and negatively impacting the
systems in particular. The last decade has witnessed a huge
performance of the system [3]. There are many challenges that
increase in deployment of biometric systems, and while most
occur when matching two fingerprints, which are outlined in
of these systems have been single vendor, monolithic
[4]. These include existence of spurious features and missing
architectures the issue of interoperability is bound to arise as
features compared to the database template, transformation or
rotation of features, and elastic deformation of features.
Manuscript received June 07, 2009. Ko and Krishana presented a methodology for measuring
S.K. Modi,Ph.D, is the Director of Research of the Biometric Standards
Performance and Assurance Laboratory, Purdue University, West Lafayette,
and monitoring quality of fingerprint database and fingerprint
IN 47907 USA (phone: 765-494-0298; e-mail: modis@purdue.edu). match performance of the Department of Homeland
H. Kim, Ph.D., is a professor with School of Information and Security’s Biometric Identification System [5]. They
Communication Engineering, INHA, Incheon, Korea. (e-mail: proposed examining fingerprint image quality not only as a
hikim@inha.ac.kr).
S.J. Elliott, Ph.D., is an associate professor with the Industrial Technology
predictor of matcher performance but also at different stages
Department, and Director of the Biometric Standards Performance and in the fingerprint recognition system. They pointed out the
Assurance Laboratory, Purdue University, West Lafayette, IN 47907 USA importance of understanding the impact on performance if
(e-mail: Elliott@purdue.edu).
Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:29 from IEEE Xplore. Restrictions apply.
2. fingerprints captured by a new fingerprint sensor were III. METHODOLOGY
integrated into an existing identification application. Their
observations and recommendations were primarily aimed at A. Experiment Setup
facilitating maintenance and matcher accuracy of large scale A wide cross-section of commercially available fingerprint
applications. sensors was chosen for this study. Three different
Jain and Ross investigated the problem related to technologies and two different interaction types were selected
fingerprint sensor interoperability, and defined sensor – Table I outlines the technical details. 190 subjects were
interoperability as the ability of a biometric system to adapt recruited, consisting of 131 males and 59 females – other
raw data obtained from different sensors [6]. They defined the demographic characteristics are shown in Table II. Each
problem of sensor interoperability as variability introduced in subject provided six images of their index finger on their
the feature set by different sensors. They collected fingerprint dominant hand. The order of the sensors was randomized for
images on an optical sensor manufactured by Digital each subject to minimize the learning effect for interaction
Biometrics and a solid state capacitive sensor manufactured with the sensors.
by Veridicom. The fingerprint images were collected on both TABLE I
sensors from 160 different individuals who provided four SENSOR CHARACTERISTICS
impressions each for right index, right middle, left index, and Sensor Technology Type Interaction Capture
Type Area (mm)
left middle finger. Their results showed that EER of 6.14%for
D1 Thermal Swipe 14 X .4
matching images collected from Digital Biometrics sensor and
D2 Capacitive Swipe 13.8 X 5
EER of 10.39% for matching images collected from D3 Optical Touch 30.5 X 30.5
Veridicom sensor. The EER for the matching images D4 Optical Touch 14.6 X 18.1
collected from Digital Biometrics sensor to Veridicom sensor D5 Capacitive Touch 12.8X15
was 23.13%. Their results demonstrated the impact of sensor D6 Optical Touch 16 X 24
interoperability on matching performance of a fingerprint D7 Optical Touch 15 X 15
system. Nagdir and Ross proposed a non-linear calibration D8 Capacitive Touch 12.8 X 18
scheme based on thin plate splines to facilitate sensor
interoperability for fingerprints [7]. Their calibration model TABLE II
was designed to be applied to the minutiae dataset and to the DATASET DESCRIPTION
fingerprint image itself. They used the same fingerprint Total 190
Subjects
dataset used in the study conducted by [6] but used the
Gender Male Female
VeriFinger minutiae based matcher and BOZORTH3
minutiae based matcher for the matching fingerprints. They 131 59
applied the minutiae and image calibration schemes to Occupation Manual Office Worker
fingerprints collected from Digital Biometrics sensor and Laborer
Veridicom sensor and matched the calibrated images from the 17 173
two sensors against each other. Their results showed an
Number of 190*6 = 1140
increase in Genuine Accept Rate (GAR) from approximately samples
30% to 70% for the VeriFinger matcher after applying the
minutiae calibration model. For the BOZORTH3 matcher an
increase in GAR from approximately 35% to 65% was B. Data Processing Methodology
observed. The variables analyzed for this study were minutiae count
The National Institute of Standards and Technology (NIST) and image quality score of the fingerprint image. Minutiae
performed an evaluation test to assess the feasibility of count was generated for each dataset using the VeriFinger 5.0
interoperability of INCITS 378 templates [8]. The minutiae extractor. Image quality scores were generated using NFIQ.
template interoperability test was called the MINEX 2004 Minutiae count scores were always greater than 0, and NFIQ
test. The MINEX report identified quality of the datasets as a scores ranged from one to five where one indicated best
factor which affected level of interoperability. The DOS and possible score and five indicated the worst possible score.
DHS datasets were of lower quality and did not exhibit a level The VeriFinger 5.0 matcher was used to generate genuine
of interoperability as that of POEBVA and POE databases. match and imposter non-match scores.
Another paper proposed to solve the problem of acquisition
mismatch using class conditional score distributions specific C. Data Analysis Methodology
to biometric devices [9]. Their problem formulation depended Hybrid testing, which combined live acquisition and offline
on class conditional score distributions, but these probabilities matching, was performed to analyze the data collected in this
were not known a-priori. They devised an approach which experiment [10]. A hybrid testing scenario was necessary for
used observed quality measures for estimating probabilities of this experiment because live subjects would not want to sit
matching using specific devices. Their device specific score through combinatorial use of multiple sensors. Genuine and
normalization procedure reduced the probability of mismatch imposter match scores were generated offline after all 190
for samples collected from different devices. subjects had completed their data collection sessions. The six
fingerprint images provided by the subjects were split into two
Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:29 from IEEE Xplore. Restrictions apply.
3. groups: the first three images were placed in an enrollment H20: μs score = μr score for all s (2)
dataset and the last three images were placed in the test H2A: μs score r score for all s
dataset. Enrollment and test template datasets were created for where s = 1..8, r = 1..8 datasets
all eight sensors using VeriFinger 5.0. The resulting
enrollment templates from each dataset were compared Previous sensor interoperability studies have compared the
against templates from each test dataset resulting in a set of numerical FNMR of native and interoperable datasets. The
scores S, where test of homogeneity of proportions was used to test similarity
of FNMR. The main objective of this test was to examine if
S = {(Ei,Vj,scoreij)} the difference in FNMR among the datasets was statistically
i= 1,..,number of enrolled templates significant. This test aided in statistically testing equality of
j = 1,.., number of test templates FNMR for native datasets and interoperable datasets. The
scoreij = match score between enrollment template and test template critical value of 2 was computed at a significance level of
0.05 and degrees of freedom (n-1), and then compared to the
Match score analysis was conducted by computing test statistic 2. If the test statistic exceeded the critical value,
performance matrices, consisting of FNMR for all datasets the null hypothesis was rejected. If the null hypothesis was
collected in the experiment at the fixed FMR of 0.1% [11] as rejected, further analysis was performed to examine which
shown in Fig. 1. interoperable datasets caused the rejection. This was done
using the Marascuillo procedure (for a significance level of
Sensor 1 Sensor n
0.05) simultaneously tests the differences of all pairs of
E proportions for all groups under investigation [13].
N
R Template IV. ANALYSIS OF RESULTS
O Generator
L
L A. Minutiae Count
Template
M Table III shows the descriptive statistics for minutiae count.
E
N The omnibus test for main effect of sensor dataset was
T DB DB
statistically significant p < 0.001 at = 0.05. Tukey’s HSD
Match
test was performed to determine the statistical significance of
Matcher Score difference between each possible pair in the group of datasets.
All pairwise comparisons were found to be statistically
significant in their differences.
DB DB
T TABLE III
E MINUTIAE COUNT
S Template Sensor Mean Standard
T Deviation
I Template D1 41.72 10.57
N Generator
G
D2 28.32 10.73
D3 40.25 10.12
D4 30.74 8.06
Sensor 1 Sensor n D5 24.38 6.87
D6 38.62 9.18
Fig. 1. Score Generation Methodology D7 27.53 7.69
D8 26.15 6.73
The fingerprint feature analysis involved examination of
B. Image Quality
minutiae count and image quality scores. Statistical tests were
performed to test similarities in minutiae counts and image Table IV shows the descriptive statistics for image quality
quality between all of the fingerprint datasets. If a statistically scores generated by NFIQ.
significant difference was observed for the test, all possible TABLE IV
pairs of means were compared. Tukey’s Honestly Significant NFIQ SCORES
Difference (HSD) was used to test all pairwise mean Sensor Mean Median
comparisons. Tukey’s HSD is effective at controlling the D1 1.29 1
overall error rate at significance level , and thus preferred D2 1.91 2
over other pairwise comparison methods [12]. D3 1.75 1
D4 2.00 2
H10: μs minutiae_count = μr minutiae_count for all s r (1) D5 2.18
Eq. 3.1 2
H1 A: μs minutiae_count μr minutiae_count for all s r D6 1.77 2
D7 2.03 2
D8 1.58 2
Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:29 from IEEE Xplore. Restrictions apply.
4. NFIQ uses a 3-layer feed forward nonlinear perceptron The first cell in the rows in Table VI indicated the native
model to predict the image quality values based on the input dataset and the remaining cells in that row indicated its
feature vector of the fingerprint image [14]. Neural networks corresponding interoperable datasets. A p value of less than
are non-parametric processors, which implies that the results 0.05 indicated a statistically significant difference, indicated
produced would have non-parametric characteristics [15]. by S in the table. NS indicates a non significant difference.
This property of NFIQ values precluded the use of parametric The pairwise comparisons test showed an even distribution of
based approach for detecting differences in quality scores similarity of FNMR with interoperable datasets without any
between all fingerprint datasets. Instead an analysis of apparent trend among acquisition technologies and interaction
variance on rank of the response variable was performed using types of the sensors that the datasets were collected from. The
the Kruskal Wallis test. The test for main effect of the sensor test of {D2, D1} was interesting since it did not show a
was statistically significant at = 0.05. Follow up tests were difference in the pairwise test of proportions but showed a
performed on pairwise comparisons of sensor datasets to difference in minutiae count and image quality scores for all
determine which pairs of sensors were statistically significant software used. D2 was collected a capacitive swipe sensor and
in the differences of NFIQ scores. Tukey’s HSD pairwise D7 was collected using an optical touch sensor. The test of
comparisons were performed on the ranks of the observations {D3,D4} was interesting since it did not show a difference in
for each dataset. The test of pairwise comparison for pairwise test of proportions, and also did not show a
{D4,D7}; {D5,D8}; and {D7,D8} was found to be not difference in minutiae count and image quality scores. D3 and
statistically significant at = 0.05. As a group, neither optical D4 were both collected using optical touch sensors. These two
touch sensors nor capacitive touch sensors showed a high tests indicated that impact of minutiae count similarity and
level of similarity of quality scores. image quality score similarity was not consistent on the
pairwise test of proportions. Interoperable datasets which
C. FNMR Analysis
contained D8 as its second dataset showed a high level of
The interoperability FNMR matrices are shown in Table V. similarity to the native datasets which were used to create the
The cells along the diagonal indicate enrollment and test interoperable dataset. This indicated that D8 did not degrade
fingerprint images from the same fingerprint sensor. The cells the performance of an interoperable dataset compared to the
off the diagonal indicate fingerprint images from different performance of native datasets.
sensors. The sensor dataset in the rows indicate the source of
the enroll sensor and the sensor dataset in the columns indicate TABLE VI
the source of the test sensor. All the native FNMR were found PAIRWISE COMPARISON OF PROPORTIONS
to be lower than interoperable datasets except for the D1 D2 D3 D4 D5 D6 D7 D8
interoperable dataset {D5,D8} where D5 and D8 were both D1 S S S S S S S
collected from a capacitive touch type sensor. The D2 NS NS NS S S S NS
interoperable FNMR for {D5, D8} = 0.49% and native D3 S S NS NS S S NS
D4 S S NS NS S S NS
FNMR for D5 = 0.78% as shown in Table V. The dataset for
D5 S S NS NS S NS NS
D8 had a mean minutiae count of 26.15 while dataset for D5
D6 S S S NS S NS NS
had a mean minutiae count of 24.38. This result indicated the D7 S S NS NS NS NS NS
interoperable dataset {D5, D8} had a larger number of D8 S S NS NS NS S NS
minutiae points to match compared to native dataset of D5.
The capture area of the sensor used to for D8 dataset was D. Impact of Image Quality
larger compared to capture area used for D5 dataset. A cross This section describes the impact of removing low quality
reference analysis of the FNMR matrix with similarity of fingerprint images and recalculating the interoperable FNMR.
minutiae count and quality scores from the previous section The fingerprint images which had NFIQ quality scores of four
did not show any specific relations. and five were not used for the matching operation and the
The Marascuillo procedure was used to simultaneously test results are shown in Table VII. NFIQ provides five categories
the differences of all pairs. The results for pairwise of quality scores and experiments have been performed by
comparison for each native dataset are shown in Table VI. NIST to predict the impact of NFIQ quality scores on
matching operations. The NIST MINEX test report indicated
TABLE V that a reduction in interoperability FNMR would be observed
FNMR AT FIXED FMR 0.1%
since poorly performing images were not used [8]. The
D1 D2 D3 D4 D5 D6 D7 D8
analysis performed in this section, examined the impact of
D1 0.47 6.79 5.60 4.44 5.61 31.03 11.30 3.76
removing low quality images on variance of interoperability
D2 7.33 5.05 6.49 7.44 10.92 16.18 10.83 6.11
FNMR.
D3 8.57 4.78 0.24 1.07 1.28 2.73 1.79 0.54
D4 5.10 4.44 0.95 0 1.39 1.60 1.83 1.08 Note that the interoperable FNMR for {D1,D6} increased
D5 5.56 7.64 0.85 0.85 0.78 4.42 2.23 0.42 to 0.267% compared to the full dataset interoperability FNMR
D6 33.86 13.42 2.99 1.30 4.47 0.17 1.41 2.15 of 0.25%. D1 dataset was collected using a thermal swipe
D7 14.92 9.89 1.73 2.12 2.35 2.71 0.94 1.85 sensor and D6 dataset was collected using an optical touch
D8 2.87 2.96 0.12 0.90 0.49 1.73 0.77 0.11 sensor. This was an interesting result since all the other FNMR
Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:29 from IEEE Xplore. Restrictions apply.
5. reduced compared to FNMR calculated for full datasets. The FMR of 0.1%. The scatter plot only contained interoperable
Marascuillo procedure was repeated on interoperable FNMR and interoperable dataset core overlap percentage
fingerprint datasets which contained fingerprint images with since interoperable FNMR were the data points of interest. An
NFIQ score of one, two or three. The overall test of inverse relation between FNMR and percentage of pairs with
proportions was found to be significant in their difference at core overlap was observed. There were two data points of
= 0.05. interest which represented interoperable datasets {D1,D6}
TABLE VII and {D6,D1}. The relation between percentage of images with
FNMR AT FIXED FMR 0.1% FOR IMAGE WITH NFIQ < 4
D1 D2 D3 D4 D5 D6 D7 D8
cores and FNMR of these two interoperable datasets did not
follow the trend observed in the other interoperable datasets,
D1 0.001 0.050 0.038 0.033 0.038 0.267 0.097 0.028
where a higher percentage of core overlap between images
D2 0.055 0.046 0.057 0.060 0.098 0.150 0.098 0.051 was correlated to a lower FNMR. D1 was collected using
D3 0.065 0.033 0.000 0.008 0.006 0.022 0.007 0.002 thermal swipe sensor and D6 was collected using optical touch
D4 0.037 0.036 0.007 0.001 0.01 0.014 0.011 0.007 sensor. These points indicate existence of other underlying
D5 0.046 0.066 0.005 0.006 0.004 0.040 0.015 0.003 factors which affected the interoperable FNMR.
D6 0.306 0.116 0.019 0.011 0.032 0.000 0.012 0.012
D7 0.119 0.084 0.007 0.007 0.012 0.019 0.004 0.007
D8 0.025 0.026 0.000 0.007 0.004 0.013 0.007 0.001
The results from pairwise comparisons are shown in Table
VIII. The recalculated pairwise comparisons showed that the
test of proportions for {D3,D7} dataset did not show any
statistically significant difference compared to results from
Table VI. D3 and D7 both collected using optical touch
sensors. This test indicated that removing the low quality Fig. 2. Core overlap scatterplot
images reduced the difference of variance of FNMR between F. Interoperability at Interaction and Technology Level
these datasets. A follow up evaluation of interoperability was performed
The recalculated results for {D6,D4}, {D6,D7}, and by grouping the datasets into three categories: datasets
{D6,D8} datasets showed a statistically significant difference. collected using swipe interaction type sensors, datasets
The results from Table VI of the same test showed no collected using optical touch type sensors, and datasets
statistically significant difference between the native and collected using capacitive touch type sensors. The evaluation
interoperable datasets. This result was interesting as it of different groups allowed for examination of interoperability
indicated that difference of variance of FNMR increased for at the acquisition and interaction level, not the sensor level.
these interoperable datasets when the lowest NFIQ score D1 and D2 were placed in the Swipe group. D3, D4, D6 and
images were removed. It should also be noted that D4, D6, and D7 were placed in the Optical Touch group. D5 and D8 were
D7 were all optical touch sensors. Removal of low quality placed in the Capacitive Touch group.
images did not have a consistent effect on FNMR of
interoperable datasets which were all collected using optical TABLE IX
touch sensors. ACQUISITION AND INTERACTION LEVEL INTEROPERABILITY
AT FIXED FMR 0.1%
TABLE VIII TEST
PAIRWISE COMPARISON OF PROPORITIONS Swipe Optical Capacitive
D1 D2 D3 D4 D5 D6 D7 D8 Touch Touch
D1 S S S S S S S E Swipe 4.81 11.75 6.58
D2 NS NS NS S S S NS N
R Optical 11.93 1.53 1.89
D3 S S NS NS S NS NS
O Touch
D4 S S NS NS S NS NS L Capacitive 4.76 1.47 0.45
D5 S S NS NS S NS NS L Touch
D6 S S S S S S S
D7 S S NS NS NS NS NS The {Capacitive Touch, Optical Touch} interoperable
D8 S S NS NS NS S NS dataset showed the lowest interoperable FNMR indicating a
high level of interoperability between optical touch and
E. Core Overlap capacitive touch technologies. The {Capacitive Touch,
A scatter plot (Fig. 2) was created where the x-axis Swipe} interoperable dataset had a lower FNMR than the
represented the percentage of fingerprint image pairs with Swipe native dataset which indicated that the interoperable
core overlap from each interoperable dataset and the y-axis dataset performed better than the native dataset. The {Swipe,
represented its corresponding interoperable FNMR at fixed Optical Touch} dataset had the highest FNMR and indicated
Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:29 from IEEE Xplore. Restrictions apply.
6. the lowest degree of interoperability. The results showed that interoperable datasets performance. The dataset collected for
Capacitive Touch dataset had the lowest interoperable FNMR this research can be used for evaluating interoperability of
compared to Optical Touch and Swipe interoperable datasets. sensors, feature extractors, and feature matchers as part of the
The interoperable datasets generated with D8 dataset showed same experiment. A combined analysis of fingerprint sensors,
a high level of similarity of FNMR with the other native feature extractors and feature matchers would further the
datasets. Since D8 was part of the Capacitive Touch group it general understanding of this field. Further work into
showed a better interoperable FNMR with other groups. In transforming the image so that the distortion of fingerprint
combination with the sensor level interoperability analysis, images would be reduced without having any a-priori
these results show the effect of interoperability at the knowledge about the fingerprint sensors would be of immense
acquisition and interaction level. benefit to the field of fingerprint recognition. Sensor agnostic
transformation methods, even if limited in its capabilities,
V. CONCLUSIONS & FUTURE WORK would significantly augment the current methodologies being
It was observed that similarity of minutiae count of the investigated.
different sensor datasets did not show a relation to a specific
acquisition technology or interaction type. Fingerprint images REFERENCES
collected from optical touch sensors showed a higher level of
[1] ISO, ISO/IEC 19795-1: Information technology - Biometric
similarity in quality scores with fingerprints collected from
performance testing and reporting - Part 1: Principles and framework,
other optical touch sensors. The combination of similarity of ISO/IEC, Editor. 2006, ISO/IEC JTC 1/SC37: Geneva.
minutiae count and image quality scores did not have an [2] Kukula, E.P., et al., Effect of Human Interaction on Fingerprint
impact on similarity of FNMR for native and interoperable Matching Performance, Image Quality, and Minutiae Count.
datasets. Higher minutiae count also did not have an impact on International Journal of Computer Applications in Technology, 2009.
34(4): p. 270-277.
FNMR of its corresponding interoperable datasets. From the [3] Wayman, J. A Generalized Biometric Identification System Model. in
observed results performance of interoperable datasets could 31st Conference on Signals, Systems and Computers. 1997. Pacific
not be predicted by separately analyzing performance of Grove, California.
native datasets. Interoperable datasets which had a higher [4] Bolle, R. and N.K. Ratha. Effect of Controlled Image Acquisition on
Fingerprint Matching. in 14th International Conference on Pattern
percentage of pairs of fingerprint images in which a core was Recognition. 1998. Brisbane, Australia.
detected had a positive relation with lower FNMR, with the [5] Ko, T. and R. Krishnan. Monitoring and Reporting of Fingerprint
only exception of {D1,D6} and {D6,D1} interoperable Image Quality and Match Accuracy for a Large User Application. in
datasets. Consistent interaction of the finger with a sensor is Applied Imagery Pattern Recognition Workshop. 2004. Washington,
D.C.: IEEE Computer Society.
an important factor in improving performance, and it becomes [6] Jain, A. and A. Ross, eds. Biometric Sensor Interoperability. BioAW
even more important when the fingerprints are captured from 2004, ed. A. Jain and D. Maltoni. Vol. 3067. 2004, Springer-Verlag:
different sensors. A simple metric for consistent placement is Berlin. 134-145.
required for collecting fingerprints from different sensors, and [7] Nagdir, R. and A. Ross. A Calibration Model for Fingerprint Sensor
Interoperability. in SPIE Conference on Biometric Technology for
this research showed that core overlap provides a metric Human Identificaiton III. 2006. Orlando, USA.
which has an impact on FNMR of interoperable datasets. [8] Grother, P., et al., MINEX Performance and Interoperability of the
Removing low quality images from interoperable datasets did INCITS 378 Fingerprint Template. 2006, National Institute of
not lead to a reduction in statistical variance of FNMR for Standards and Technology: Gaithersburg, Maryland. p. 48.
[9] Poh, N., J. Kittler, and T. Bourlai. Improving Biometric Device
interoperable datasets, although the absolute FNMR was Interoperability by Likelihood Ratio-based Quality Dependent Score
reduced for all native and interoperable datasets. As a general Normalization. in BTAS 2007. 2007. Washington, D.C.
operational procedure checking quality of fingerprint images [10] Grother, P. An Overview of ISO/IEC 19795-4. in Biometric Consotium.
is a good practice and should definitely be followed for 2006. Baltimore.
[11] Campbell, J. and M. Madden, ILO Seafarers' Identity Documents
interoperable datasets as well. The analysis of interaction
Biometric Interoperability Test Report 2006, International Labour
types indicated that sensors of the same interaction type had a Organization: Geneva. p. 170.
lower interoperability FNMR. This can be attributed to similar [12] Montgomery, D.C., Design and Analysis of Experiments. 4th ed. 1997,
distortion occurring in images from the same interaction type. New York: John Wiley & Sons. 704.
There are several avenues for future work in this area. The [13] NIST. Engineering Statistics Handbook. [cited 2008; Available from:
http://www.itl.nist.gov/div898/handbook/prc/section4/prc474.htm.
dataset collected as part of this research can be used to analyze [14] Garris, M.D., et al., User's Guide to NIST Fingerprint Image Software
False Match Rates (FMR) if an identification scenario is of 2 (NFIS2). 2004, National Institute of Standards and Technology:
interest. The statistical analysis framework can be modified to Gaithesburg, MD. p. 217.
test for FMR of interoperable datasets, and it would be [15] Shu-Long, J., S. Zhong-Kang, and W. Yan-Yan. The Non-Parametric
Detection with Neural Network. in International Conference on
interesting to understand the impact of interoperability on Circuits and Systems. 1991. China: IEEE.
FMR of fingerprint datasets collected from multiple sensors.
The impact of removing low quality images from
interoperable datasets did not lead to a higher level of
similarity in FNMR between interoperable datasets and native
datasets. The results indicated that further work is required to
investigate the impact of quality on the variance of
Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:29 from IEEE Xplore. Restrictions apply.