Towards integrating level-3 Features with perspiration pattern for robust fingerprint recognition

Abstract

Level-3 fingerprint features from fingerprint images like pores are difficult to capture detect, and involve high resolution scanners with higher ppi count. However, these features provide finer information about a fingerprint characteristics. Furthermore, fingerprint pores may be useful in determining liveness of fingerprint in order to prevent spoofing of fingerprint devices. In this study fingerprint pores along the ridges are used for fingerprint matching. Wavelet based fingerprint enhancement techniques are implemented to ease detection of the level-3 features. Delaunay triangulation based alignment and matching of the fingerprints is performed. The pores are checked for the liveness by perspiration activity in the time series captures. The developed matching scheme is tested for the high resolution data (686 ppi) for 114 live and spoof fingerprint classes. ROC is plotted and EER of 2.97% is obtained.

Full text

Page 1

Towards integrating level-3 Features with
perspiration pattern for robust fingerprint
recognition
Aditya Abhyankar and Stephanie Schuckers
Electrical and Computer Eng Dept.
Clarkson University
Potsdam, NY, USA
{abhyanas,sschucke}@clarkson.edu
Abstract—Level-3 fingerprint features from fingerprint images
like pores are difficult to capture detect, and involve high reso-
lution scanners with higher ppi count. However, these features
provide finer information about a fingerprint characteristics.
Furthermore, fingerprint pores may be useful in determining
liveness of fingerprint in order to prevent spoofing of fingerprint
devices. In this study fingerprint pores along the ridges are used
for fingerprint matching. Wavelet based fingerprint enhancement
techniques are implemented to ease detection of the level-3
features. Delaunay triangulation based alignment and matching
of the fingerprints is performed. The pores are checked for the
liveness by perspiration activity in the time series captures. The
developed matching scheme is tested for the high resolution data
(686 ppi) for 114 live and spoof fingerprint classes. ROC is plotted
and EER of 2.97% is obtained.
Index Terms—fingerprints, level-3 features, pores, perspiration
pattern, liveness, wavelets.
I. INTRODUCTION
Among all biometric identifiers, fingerprints are most widely
used [1], [2], [3]. Fingerprint recognition is performed by
matching fingerprint features which can be divided into level-
1, level-2 and level-3 categories, representing coarser to finer
representation. Level-2 features, mainly minutiae points i.e.
ridge endings and bifurcations are the most studied finger-
print pointers and are used by most fingerprint matchers for
recognition.
Level-3 features are the least studied features and include
pores, incipient ridges, line shapes, deformations, warts, scars
and so forth [4], [5]. Although level-3 features require the
capturing resolution to be very high for accurate detection,
they effectively provide finer information about fingerprints.
Level-3 features may provide more information in order to per-
form accurate authentication and may be useful for difficult-
to-classify fingerprints and for large database of fingerprints,
or for applications requiring high performance. In addition,
previous work has shown that it is possible to spoof fingerprint
devices [6] and analysis of level-3 features may provide
additional information in order to prevent spoof attacks.
In previous approaches, characteristics of the fingerprint
image across two fingerprint images taken in series have been
used which quantify the moisture changes in time due to
perspiration [6]. In this research, the liveness algorithm is
integrated with the matching algorithm. That is, liveness has
a distinct individual pattern associated with the pores which
is useful for both characterization of liveness and matching.
Thus, it would be difficult to separate the two such that a
would-be spoofing would have to replicate the exact fingerprint
in addition to the exact pore pattern changes over two images.
DATA BASE
Enhancemen
Pore Detection/
Selection
Delaunay
Triangulation
Alignment
Liveness Score
Level-3 Matching
Liveness Enbedded
Level-3 Match Score
Receiver Operating
Characteristics
0 sec
0 sec
2 sec
2 sec
Enhancement
Pore
Detection
Triangulation
Triangulation
Pore
Detection
Enhancement
Alignment Plane
Fig. 1.
compared with the database representation of the fingerprint classes. The raw
fingerprint images are enhanced and subjected to pore detection. Delaunay
triangulation is used for alignment and verification match score and liveness
score is used to detect live, genuine classes.
Entire algorithm in a snap-shot. The query fingerprint class is
Level-3 features in the form of sweat pores are detected
and matched for the recognition purposes. This is integrated
with the perspiration based liveness detection. Delaunay tri-
angulation based methods are used to align fingerprints [7],
match the level-3 features and calculate the liveness score by
measuring the perspiration activity across the pores.
3085978-1-4244-7993-1/10/$26.00 ©2010 IEEE ICIP 2010
Proceedings of 2010 IEEE 17th International Conference on Image ProcessingSeptember 26-29, 2010, Hong Kong

Page 2

II. FINGERPRINT DATA AND ENHANCEMENT
This section presents the data considered for this study and
wavelet based fingerprint enhancement techniques. Presently
fingerprint recognition systems rarely use level-3 features and
hence fingerprint scanners with capturing strength of up to 500
ppi (pixels per inch) have been standardized. In order to be
able to successfully extract level-3 features, a database with
higher resolution has been collected for this study [8]. Also,
wavelet based enhancement is performed in order to differen-
tiate between noise and subtle fingerprint characteristics.
A. Data management
The data
Identix DFR 2100 fingerprint scanner. The scanner is used
in the maximum active image area mode, by which image
resolution of 686 ppi and image size of 720×720 is obtained.
The data is collected from 30 subjects with 3 fingers of each
subjects in 1 session. For each finger class 5 time-series
captures were obtained. The SDK obtained along with the
scanner is modified in order to capture the fingerprint at zeroth
second and then again after 2 seconds. The captured data is
presented in Table (I).
used for thisstudy iscollectedusing
TABLE I
DATA DISTRIBUTION
Live
30
90
450
Gummy
12
12
60
Play-Doh
12
12
60
Subjects
Classes
TS Captures
B. Fingerprint Enhancement
A wavelet-based fingerprint enhancement steps are shown
in figure (2). A cubic spline wavelet is used to perform
phase-based denoising to obtain fingerprint images with more
prominent level-3 features [9]. First the raw fingerprint image
is transformed into the wavelet domain, second the wavelet
coefficients are analyzed using different selection and rule-
based mechanisms, and finally denoised fingerprint image is
synthesized using processed wavelet coefficients.
In this implementation separable cubic-spline function
φ(x,y) is at the base and corresponding oriented wavelets are
represented as,
ψ1(x,y) =
∂
∂xφ(x,y)
ψ2(x,y) =
∂
∂xφ(x,y)
(1)
The gradient and the edges in the fingerprint ridge patterns
are represented by the convolution function and smoothed by
the function φ(x,y) at dyadic scales.
ρ2jf(x,y)=
=
|∇2jf(xi,yi)|
{[W1
(2)
2jf(xi,yi)]2+ [W2
2jf(xi,yi)]2}1/2(3)
λ2jf(xi,yi) = arctan[[W2
2jf(xi,yi)]2
[W1
2jf(xi,yi)]2]
(4)
Original Fingerprint Image
Wavelet level-4 decomposition
Enhanced
fingerprint
Image. Sweat
pores more
distinctly
observed along
the ridges
Fig. 2.
detection. Wavelet sub bands are shown and represent multi resolution anal-
ysis. Enhanced image clearly shows the improved detectability of fingerprint
pores.
Wavelet based fingerprint enhancement for effective level-3 feature
A point (xi,yi) is considered an edge if the magnitude in
equation (2) of gradient ρ2jf reaches local maximum along
gradient direction given in equation (4). Thus, we can identify
all the excess ridge points and then define a subset of points
that satisfy the phase conditions.
For a j-level 2 − D wavelet transform (Figure (2)), the
discrete low-pass approximation of f(x,y) at the coarsest
scale 2jis presented as,
{S2jf(xi,yi),|I2(f),G2j(f,I2j)|1≤j}
This dominated the multi-scale representation of the image
f, where the two data sets, I2j(f) and G2j(f,I2j) describe
the ridge locations and the multi-scale gradients respectively
and are a reduced subset of significant information.
Phase-based denoising is used to remove noisy areas. The
criteria is based on the fact that gradients of real edges have
almost the same angles in a certain regions [9]. Neighboring
gradients with different phases strongly represent noise and
should be removed. Consider a search region Q(i) around
pixel i. The pixel with no similar phases in the search region
are removed,
(5)
I2j(f) =
{i|λ2jf(xp,yp) ≈ λ2j+1f(xq,yq),
∀p,q ∈ Q(i) ⊂ I2j(f),∀i ∈ I2j(f)}
After retaining only significant informative regions, the
contrast stretching in the spatial domain is performed as shown
in figure (2).
(6)
(7)
3086

Page 3

Enhanced fingerprint image with distinctly marked sweat pores
Fig. 3.
with all the marked pores is shown. Only very few open pores are not correctly
detected.
Phase consistency based pore detection. An enhanced image along
III. PORE DETECTION AND DELAUNAY TRIANGULATION
For detecting the pores along the fingerprint ridges, the
phase-based approach adopted for the enhancement is ex-
tended. The pores along the ridges give rise to points where the
wavelet components of the fingerprint image are maximally
in phase. The use of phase consistency has an advantage
over the gradient consistency that being dimensionless, phase
consistency doesn’t depend on spatial factors like brightness
and contrast, thus producing accurate pore detection [9].
Let Ifp be the fingerprint image under consideration and
Meven
n
and Modd
n
be the even and odd-symmetric wavelets at
scale n, then the filter response would be,
[ern(x),orn(x)] = [I(x) ∗ Meven
and the phase is given as,
n
,I(x) ∗ Modd
n
]
(8)
φn(x) = arctan2(ern(x),orn(x))
(9)
Since all the images subjected for phase consistency are
enhanced images, the underlying noise levels are poorly spread
and an accurate consistent pore pattern is detected by simply
looking at the phase inconsistencies. The pore detection is
presented in figure (3).
Firstly, only those pores are retained which have consistent
directional values (in the spatial domain) along the ridges.
Secondly, if all the pores are retained then further processing
slows down and hence only active pores are retained. A pore is
decided as an active pore if it experiences a gradient change
of more than 40% of the maximum value from first image
captured at zeroth second to the second image captured after
two seconds. For further processing only the second image
pores are used.
Once the pore detection and selection procedures are over,
two set of detected pores are obtained, for the query and the
template images respectively. Using Delaunay triangulation,
each fingerprint is represented as a special connected graph
with each node being an active pore along the ridge and each
edge connecting two active pores. The approach taken in this
work closely follows the work presented in [10], [7]. Since
(a2) Selected Node Points
(a3) Selected Pore Points
(a1) Enhanced FP Image
(a4) Delaunay Triangulation
(b) Edge guided alignment
Fig. 4.
image and selected pore points image respectively from left to right. The
bottom row shows Delaunay triangulation mesh on the top of the fingerprint
image. (b) Edge-guided triangle alignment in the projected plane. This plane is
used to project the delaunay information of the verification fingerprint images.
(a1-a4) Upper row shows the original image, selected node point
Delaunay triangulation gets affected only locally, incremental
active pores do not have any effect on the further processing.
The intuitional reasoning for choosing Delaunay triangulation
is as follows:
• Produces unique pore representation for the selected N
pores
• O(N) triangles can be formulated in O(N logN) time
• Inaccuracies in the pore detection can only locally affect
the triangulation procedures
IV. MATCH SCORE AND LIVENESS SCORE CALCULATIONS
Once the meshes across the selected pores are created,
two fingerprints can be matched using Local-structure-based
aligned matching, as given in [10]. For aligning the fingerprint
images, the mesh information of query image is projected on
the edge-guided triangles of the template image as shown in
figure (4-b).
Let MI
q
be the selected pores for query and
template images respectively and NI
responding number of neighboring pores. As given in [10],
similarity measure between an edge− →
and− →
pand MT
pand NT
q
be the cor-
pi in the query image
qj in template image is calculated as follows:
SE(− →
pi,− →
qj) =
?
TH1−WT|Vpi−Vqj|
TH1
0,
,
otherwise
(10)
This condition applies if WT|Vpi−Vqj| < TH1, where W
is the weighing factor, TH1 is the threshold value, MI
MT
iand
jare neighboring pores of MI
pand MT
qrespectively.
3087

Page 4

The matching score of two local structures with central
pores of MI
pand MT
qis then calculated as,
SL(MI
p,MT
q) =
?NI
p
i=1
?NT
q
j=1SE(− →
min{NI
pi,− →
q}
qj)
p,NT
(11)
Similarly, liveness score is calculated by measuring activity
of the active pores from zeroth second to two seconds [6]. Let
ML1I
for query image pair and template image pair respectively and
NL1I
number of neighboring pores. As given in [10], similarity
measure between an edge− →
template image is calculated as follows:
pand ML2I
p, ML1T
qand ML2T
qbe the selected pores
pand NL2I
p, NL1T
qand NL2T
qbe the corresponding
pliin the query image and− →
qlj in
SEL(− →
pli,− →
qlj) =
?
TH2−WT|V lpi−V lqj|
TH2
0,
,
otherwise
(12)
This condition applies if WT|V lpi− V lqj| < TH2, where
W is the weighing factor, TH2is the liveness threshold value,
MI
and MT
The matching score of two local structures with central
pores of MI
i(1/2) and MT
q(1/2) respectively.
j(1/2) are neighboring pores of MI
p(1/2)
p(1/2) and MT
q(1/2) is then calculated as,
SLL(MI
p,MT
q) =
?NI
p(1/2)
i=1
min{NI
?NT
q(1/2)
j=1
p(1/2),NT
SEL(− →
q(1/2)}
pli,− →
qlj)
(13)
V. RESULTS
This section presents the results of the methodology pre-
sented in the last two sections applied to the data tabulated
in section (2). The genuine scores are the live-genuine scores,
while live-imposter, not live-genuine, and not live-imposter
form the imposter class. This is different than traditional
biometric applications which do not include spoof data in the
imposter class. Considering this, total 322050 inter class and
1140 intra class combinations are worked out.
Receiver operating characteristics curve is plotted using the
recognition match scores. EER of 2.97% is obtained for the
given set up. Figure (5) shows the ROC curve.
VI. DISCUSSION, FUTURE WORK AND CONCLUSION
A new liveness embedded level-3 features based fingerprint
recognition system is designed. This method uses Delaunay
triangulation of the detected pores along the ridges to align
and match the fingerprints. Based on the pore level activity
in the time series images, a liveness score is generated. Both
scores are compared against heuristically decided threshold
values, and a decision is made about whether the fingerprint
is of a live-genuine class. Since this study involves level-3
features detection, high resolution data is specially prepared
for this study. When ROC was plotted, an equal error rate of
2.97% was obtained.
Previously there has been algorithms for matching based on
level-3 features [5], [4], and for spoof detection based on l-3
features [6]. A fully integrated fingerprint liveness matching
10
−2
10
−1
10
0
10
1
0
10
20
30
40
50
60
70
80
90
100
False Accept Rate(%)
Genuine Accept Rate(%)
Liveness embedded level−3 FP recognition ROCs
Fig. 5.
obtained where genuine class is represented by live and genuine fingerprints.
ROC of combined level-3 and liveness scores. EER of 2.97% was
algorithm based on l-3 features is designed and implemented
in this work. The designed integrated system when tested for
a 686 ppi scanner and for 1140 intra and 322050 inter class
combinations gave a better EER of 2.97% (match time 1.02
min) as against EER of 4.92% for a system designed in [4]
when tested for 1000 ppi scanner, for 2460 intra and 83845
inter class combinations (match time 45 sec).
The study presented in this paper facilitates research in
many different directions. Future studies are desired in: fusion
of the liveness score and the match score, deriving physical
meaning from the pore level activity in order to better un-
derstand the phenomenon of perspiration, categorical study of
pores to understand which pores (open, closed etc.)contribute
more for liveness detection, adaptive wavelet filtering for
enhancing only desired pores.
REFERENCES
[1] A. Jain, R. Bolle, and S. Pankanti, Biometrics: Personal Identification
in Networked Society. Kluwer Academic Publisher, 1999.
[2] D. Maltonie, D. Maio, A. K. Jain, and S. Prabhakar, Handbook of
Fingerprint Recognition. Springer-Verlag New York, Inc., 2003.
[3] J. D. Woodward, N. M. Orlans, and P. T. Higgins, Biometrics. McGraw-
Hill/Osborne, 2003.
[4] A. Jain, Y. Chen, and M. Demirkus, “Pores and ridges: High resolution
fingerprint matching using level 3 features,” IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol. 29, no. 1, pp. 15–27,
January 2007.
[5] J. Stosz and L. Alyea, “Automated system for fingerprint authentication
using pores and ridge structure,” in SPIE, 1994, pp. 210–223.
[6] A. Abhyankar and S. A. C. Schuckers, “Integrating a wavelet based
perspiration liveness check with fingerprint recognition,” Pattern Recog-
nition, vol. 42, no. 3, pp. 452–464, 2009.
[7] G. Debis, T. Deaconu, and M. Georgiopoulos, “Fingerprint identification
using delaunay triangulation,” in Int. Conf. on Information Intelligence
and Systems, 1999, pp. 452–459.
[8] A. V. Ceguerra and I. Koprinska, “Integrating local and global features
in automatic fingerprint verification,” in ICPR, 2002, pp. 347–350.
[9] S. G. Mallat and W. L. Hwang, “Characterization of signals from
multiscale edges,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 14, pp. 710–732, 1992.
[10] H. Deng and Q. Huo, “Minutiae matching based fingerprint verification
using delaunay triangulation and aligned-edge-guided triangle match-
ing.” in AVBPA, 2005, pp. 270–278.
3088