Fingerprint Matching using Gabor Filters

Abstract

We present a fingerprint matching scheme that utilizes a ridge feature map to match fingerprint images. The technique described here obviates the need for extracting minutiae points to match fingerprint images. The proposed scheme uses a set of 16 Gabor filters, whose spatial frequencies correspond to the average inter-ridge spacing in fingerprints, is used to capture the ridge strength at equally spaced orientations. A circular tessellation of filtered image is then used to construct the ridge feature map. This ridge feature map contains both global and local details in a fingerprint as a compact fixed length feature vector. The fingerprint matching is based on the Euclidean distance between two corresponding feature vectors. The genuine accept rate of the Gabor filter based matcher is observed to be ~ 10% to 15% higher than that of minutiae-based matcher at low false accept rates. Fingerprint feature extraction and matching takes ~ 7.1 seconds on a Pentium IV, 2.4 GHz processor.

Full text

National Conference on Emerging Technologies 2004 147
Fingerprint Matching using Gabor Filters
Muhammad Umer Munir and Dr. Muhammad Younas Javed
College of Electrical and Mechanical Engineering, National University of Sciences and Technology
Rawalpindi, Pakistan.
Abstract: We present a fingerprint matching scheme that
utilizes a ridge feature map to match fingerprint images.
The technique described here obviates the need for
extracting minutiae points to match fingerprint images.
The proposed scheme uses a set of 16 Gabor filters, whose
spatial frequencies correspond to the average inter-ridge
spacing in fingerprints, is used to capture the ridge
strength at equally spaced orientations. A circular
tessellation of filtered image is then used to construct the
ridge feature map. This ridge feature map contains both
global and local details in a fingerprint as a compact fixed
length feature vector. The fingerprint matching is based
on the Euclidean distance between two corresponding
feature vectors. The genuine accept rate of the Gabor
filter based matcher is observed to be ~ 10% to 15%
higher than that of minutiae-based matcher at low false
accept rates. Fingerprint feature extraction and matching
takes ~ 7.1 seconds on a Pentium IV, 2.4 GHz processor.
Keywords: Biometrics, Gabor filters, fingerprints,
matching, verification, core point
1. INTRODUCTION
Fingerprint-based identification is one of the most
important biometric technologies which has drawn a
substantial amount of attention recently. Humans have
used fingerprints for personal identification for centuries
and the validity of fingerprint identification has been well
established. In fact, fingerprint technology is so common
in personal identification that it has almost become the
synonym of biometrics. Fingerprints are believed to be
unique across individuals and across fingers of same
individual. Even identical twins having similar DNA, are
believed to have different fingerprints. These observations
have led to the increased use of automatic fingerprint-
based identification in both civilian and law-enforcement
applications.
A fingerprint is the pattern of ridges and furrows on the
surface of a fingertip. Ridges and valleys are often run in
parallel and sometimes they bifurcate and sometimes they
terminate. When fingerprint image is analyzed at global
level, the fingerprint pattern exhibits one or more regions
where ridge lines assume distinctive shapes. These shapes
are characterized by high curvature, terminations,
bifurcations, cross-over etc. These regions are called
singular regions or singularities. These singularities may
be classified into three topologies; loop, delta and whorl.
At local level, there are other important features known as
minutiae can be found in the fingerprint patterns. Minutiae
means small details and this refers to the various ways that
the ridges can be discontinuous. A ridge can suddenly
come to an end which is called termination or it can divide
into two ridges which is called bifurcations (Figure 1).
Figure 1: Ridge ending, core point and ridge bifurcation is
shown.
2. FINGERPRINT MATCHING
Fingerprint matching techniques can be broadly classified
as minutiae based and correlation based [1]. Minutiae-
based technique first locates the minutiae points in a given
fingerprint image and matches their relative placements in
a stored template fingerprint. A good quality fingerprint
contains between 60 and 80 minutiae, but different
fingerprints have different number of minutiae. The
performance of minutiae-based techniques rely on the
accurate detection of minutiae points and the use of
sophisticated matching techniques to compare two
minutiae fields which undergo non-rigid transformations.
Correlation based techniques compare the global pattern of
ridges and valleys to see if the ridges in the two
fingerprints align. The global approach to fingerprint
representation is typically used for indexing [2] and does
not offer reliable fingerprint discrimination.
The ridge structure in a fingerprint can be viewed as an
oriented texture patterns having a dominant spatial
frequency and orientation in a local neighborhood. The
frequency is due to inter ridge-spacing present in a
fingerprint and the orientation is due to the flow pattern
exhibited by ridges. Most textured images contain a
narrow range of spatial frequencies. For a typical
fingerprint images scanned at 500 dpi, there is a little
variation in the spatial frequencies among different
fingerprints. This implies that there is an optimal scale

National Conference on Emerging Technologies 2004 148
(spatial frequency) for analyzing the fingerprint texture.
By capturing the frequency and orientation of ridges in
local regions in the fingerprint, a distinct representation of
the fingerprint is possible [3].
The proposed scheme first detects the core point in a
fingerprint image using two different techniques. Core
point is defined as the north most point of inner-most ridge
line. In practices, the core point corresponds to center of
north most loop type singularity. Some fingerprints do not
contain loop or whorl singularities, therefore it is difficult
to define core. In that kind of images, core is normally
associated with the maximum ridge line curvature.
Detecting a core point is not a trivial task; therefore two
different techniques have been used to detect optimal core
point location. A circular region around the core point is
located and tessellated into 128 sectors. The pixel
intensities in each sector are normalized to a constant
mean and variance. The circular region is filtered using a
bank of sixteen Gabor filters to produce a set of sixteen
filtered images. Gabor filter-banks are a well known
technique to capture useful information in specific band
pass channels. Two such techniques have been discussed
in [3] and [4]. The average absolute deviation with in a
sector quantifies the underlying ridge structure and is used
as a feature. The feature vector (2048 values in length) is
the collection of all the features, computed from all the
128 sectors, in every filtered image. The feature vector
captures the local information and the ordered
enumeration of the tessellation captures the invariant
global relationships among the local patterns. The
matching stage computes the Euclidean distance between
the two corresponding feature vectors.
It is desirable to obtain representations for fingerprints
which are translation and rotation invariant. In the
proposed scheme, translation is taken care of by a
reference point which is core point during the feature
extraction stage and the image rotation is handled by a
cyclic rotation of the feature values in the feature vector.
The features are cyclically rotated to generate feature
vectors corresponding to different orientations to perform
the matching.
3. CORE POINT DETECTION
Two different methods are used to detect core point in a
fingerprint image (Figure 2). The core point location is
more accurately detected by using multiple techniques.
3.1 Core point detection using Poincare index
1) Estimate the orientation field O using the least square
orientation estimation algorithm [5]. Orientation field
O is defined as an M x N image, where O(i,j)
represents the local ridge orientation at pixel (i,j). An
image is divided into a set of w x w non-overlapping
blocks and a single orientation is defined for each
block.
2) Initialize A, a label image used to indicate the core
point.
3) For each pixel (i,j) in O, compute Poincare index [2]
and assign the corresponding pixels in A the value of
one if Poincare index is between 0.45 and 0.51.The
Poincare index at pixel (i,j) enclosed by a digital
curve, which consists of sequence of pixels that are on
or within a distance of one pixel apart from the
corresponding curve, is computed as follows:
Np-1
Poincare (i ,j) = 1/(2π) ∑ ∆(k) (1)
k=0
δ(k) if |δ(k)| < π/2
∆(k) = π + δ(k) if δ(k) ≤ −π/2 (2)
π − δ(k) otherwise
δ(k) = θ(x
(k+1) mod Np
, y
(k+1) mod Np
) − θ(x
k
, y
k
)
(3)
For our method, Np is selected as 8.
4) The center of block with the value of one is
considered to be the center of fingerprint. If more than
one block has value of one, then calculate the average
of coordinates of these blocks.
3.2 Core point detection using slope
1) Estimate the orientation field O using the least square
orientation estimation algorithm [5].
2) Smooth the orientation field in local neighborhood.
Let the smoothed orientation field be represented as
O’.
3) Initialize A, a label image used to indicate the core
point.
4) In O’(i , j) , start from first row (0, 0), find the block
whose angle is between 0 and π /2 and then trace
down vertically until a block with a slope not with in
Figure 2: Optimal core point location

National Conference on Emerging Technologies 2004 149
Figure 3: Filtered images and their corresponding feature vectors for orientations
ο
0 ,
ο
5.22 , and
ο
45 are shown.
that range (0 and π /2) is encountered. That block is
then marked [6] in A. This procedure is performed on
all the rows of orientation field O’(i,j).
5) The center of block with the highest number of marks
is considered to be the center of fingerprint.
3.3 Optimal Core Point
Now we have two core point locations obtained from
above two techniques. Optimal core point is then
calculated by taking average of x-coordinate values and
taking the maximum of two y-coordinate values as
maximum y-coordinate is more precise location of core
point.
4. TESSELLATION
The spatial tessellation of fingerprint image which consists
of the region of interest is defined by a collection of
sectors. We use four concentric bands around the core
point. Each band is 20 pixels wide and segmented into
thirty two sectors. Thus we have a total of 32 x 4 = 128
sectors and the region of interest is a circle of radius 100
pixels, centered at the core point.
5. NORMALIZATION
Normalization is performed to remove the effects of
sensor noise and gray level background due to finger
pressure differences. For all the pixels in sector Si, where i
∈
(0,1,2….127), the normalized image is defined as:
i
io
V
MyxIV
M
2
0
)),(( −×
+
if I(x,y) > Mi
Ni(x, y) = (4)
i
io
V
MyxIV
M
2
0
)),(( −×
−
otherwise
Mi and Vi are estimated mean and variance of grey levels
in sector Si respectively. M
o
and V
o
are the desired mean
and variance values, respectively. For our experiments,
both M
o
and V
o
are set to a value of 50.
6. FILTERING
Gabor filters optimally capture both local orientation
and frequency information from a fingerprint image. By
tuning a Gabor filter to specific frequency and direction,
the local frequency and orientation information can be
obtained [2][3] as shown in Figure 3. Thus, they are suited
for extracting texture information from images [4].
Daugman has successfully used these filters to extract
discriminatory features from human iris [7].
An even symmetric Gabor filter has the following
general form in the spatial domain:
G(x, y ; f,
θ
) =
)'2cos(
'
'
2
1
exp
2
'
2
2
'
2
fx
y
x
yx
π
δδ


















+
−
(5)
x’ = x sin
θ
+ y cos
θ
(6)
y’ = x cos
θ
- y sin
θ
(7)

National Conference on Emerging Technologies 2004 150
Figure 4: The ROC curve comparing the performance of the Gabor filter based approach with the minutiae based approach
where f is the frequency of the sinusoidal plane wave
along the direction
θ
from the x-axis, and
x
δ
and
y
δ
are the space constants of the Gaussian envelope along x
and y axes, respectively.
The filtering is performed in the spatial domain with a
mask size of 17x17. The frequency f is the average ridge
frequency (1/K), where K is the average inter ridge
distance. The average inter ridge distance is approximately
10 pixels in a 500 dpi fingerprint image. Hence, f = 1/10.
Sixteen different orientations are examined. These
correspond to
θ
values of 0, 11.25, 22.5, 33.75, 45,
56.25, 67.5, 78.75, 90, 101.25, 112.5, 123.75, 135, 146.25,
157.5 and 168.75 degrees. The values for
'x
δ
and
'y
δ
were empirically determined and each is set to 4 (about
half the average inter ridge distance).
7. FEATURE VECTOR
Let
),( yxF
i
θ
be the
θ
-direction filtered image for sector
Si. Now,
∀
i
∈
(0,1,2….127) and
θ
∈
(0, 11.25,
22.5, 33.75, 45, 56.25, 67.5, 78.75, 90, 101.25, 112.5,
123.75, 135, 146.25, 157.5, 168.75 degrees) the feature
value,
θ
i
V
, is the average absolute deviation from the
mean defined as:
θ
i
V
=






−
∑
i
n
ii
i
PyxF
n
θθ
),(
1
(8)
where n
i
is the number of pixels in S
i
and
θ
i
P is the mean
of pixel values of
),( yxF
i
θ
in sector S
i
. The average
absolute deviation of each sector in each of the sixteen
filtered images defines the components of our 2048-
dimensional feature vector [4]. The average absolute
deviation from the mean of each sector in each of the
sixteen filtered images defines the components of our
feature vector.
The rotation invariance is achieved by cyclically rotating
the features in a feature vector itself. A single step cyclic
rotation of the features corresponds to a feature vector
which would be obtained if the image was rotated by
11.25 degrees.
Fingerprint matching is based on finding the Euclidean
distance between the corresponding feature vectors. This
minimum score corresponds to the best alignment of the
two fingerprints being matched. If the Euclidean distance
between two feature vectors is less than a threshold, then
the decision that “the two images come from the same
finger” is made, otherwise a decision that “the two images
come from different fingers” is made. Since the template
generation for storage in the database is an off-line
process, the verification time still depends on the time
taken to generate a single template
.
8. EXPERIMENTAL RESULTS
The database of fingerprint images contains 180 images.
There are eight different impressions per finger. The
performance of biometric system can be shown as a
Receiver Operating Characteristic (ROC) curve that plots
the Genuine Accept Rate (GAR) against the False Accept
Rate (FAR) at different thresholds on the matching score.
We compare this performance with a minutiae-based
approach [4][8]. As can be seen in Figure 4, our approach
outperforms the minutiae based approach over wider range
of FAR values. For example, at 1% FAR, the Gabor filter
based fingerprint matcher gives a GAR of 91% while the
minutiae based matcher gives a GAR of 73%.
The Gabor filter based fingerprint technique takes ~ 7.1
seconds on Pentium – IV, 2.4 GHz processor, for feature
extraction and matching. About 95% of the total time i.e. ~
Minutiae based matching
60
65
70
75
80
85
90
0.112345610
False Accept Rate (%)
Genunie Accept Rare (%)
Gabor filter based matching
80
85
90
95
100
0.1 0.9 3.5 5.3 8 16
False Accept Rate (%)
Genunie Accept Rate (%)

National Conference on Emerging Technologies 2004 151
6.7 seconds, is taken by the convolution of the input image
with 16 Gabor filters. The convolution operation can be
made significantly faster by dedicated DSP processors. If
the core point is correctly located, the features are
translation invariant and the rotation handled in the
matching stage is very fast. As a result, the matching
process is extremely fast.
9. SUMMARY AND FUTURE WORK
We have presented a fingerprint matching scheme that
utilizes both the frequency and orientation information
available in a fingerprint. Sixteen Gabor filters are used to
extract features from the template and input images. The
primary advantage of our approach is improved translation
and rotation invariance. The following areas of
improvement are also being studied:
(1) New matching methods for comparing the ridge
feature maps of two images
(2) Constructing the ridge feature maps using adaptive
methods for optimal selection of Gabor filters
REFERENCES
[1] D. Maltoni, D. Maio, A. K. Jain, and S. Prabhakar,
Handbook of Fingerprint Recognition, Springer-
Verlag, June 2003.
[2] A. K. Jain, S. Prabhakar and L. Hong, "A
Multichannel Approach to Fingerprint
Classification", IEEE Transactions on PAMI,
Vol.21, No.4, pp. 348-359, April 1999.
[3] A. Ross, A. K. Jain, and J. Reisman, "A Hybrid
Fingerprint Matcher", Pattern Recognition, Vol.
36, No. 7, pp. 1661-1673, 2003.
[4] A. K. Jain, A. Ross, and S. Prabhakar, "Fingerprint
Matching Using Minutiae and Texture Features",
Proc International Conference on Image
Processing (ICIP), pp. 282-285, Greece, October
7-10, 2001.
[5] L. Hong, Y. Wan and A.K. Jain, "Fingerprint
Image Enhancement: Algorithms and Performance
Evaluation", IEEE Transactions on PAMI, Vol. 20,
No. 8, pp.777-789, August 1998.
[6] A. R. Rao, A Taxonomy for Texture Description
and Identification, New York: Springer-Verlag,
1990.
[7] J. Daugman, Recognizing persons by their iris
patterns, in: A. K. Jain, R. Bolle, S. Pankanti
(Eds.), Biometrics: Personal Identification in a
Networked Society, Kluwer Academic Publishers,
1999, pp. 103-121.
[8] Anil K. Jain, Lin Hong, Sharat Pankanti, and Ruud
Bolle, “An identity authentication system using
fingerprints,” Proceedings of the IEEE, vol. 85,
no. 9, pp. 1365–1388, 1997