IMPROVED FINGERCODE FOR FILTERBANK-BASED FINGERPRINT MATCHING

Lifeng Sha, Feng Zhao, and Xiaoou Tang

Department of Information Engineering

The Chinese University of Hong Kong
Shatin, N.T., Hong Kong
{lfsha1, fzhao0, xtang}@ie.cuhk.edu.hk
ABSTRACT the problems with the minutiae-based matching
algorithms.
FingerCode has been shown to be an effective However, the original feature extraction algorithm [5]
representation to capture both the local and global assumes that the fingerprints are captured in a vertical
information in a fingerprint. However, the performance of position. While the reference point location method works
FingerCode is influenced by the reference point detection well for vertically oriented fingerprints, it fails to
process, and the AAD features cannot fully extract the precisely locate the reference point after the original
discriminating information in fingerprints. In this paper, fingerprint is rotated. In this paper, we propose a new
we first propose a new rotation-invariant reference point rotation-invariant reference point location method.
location method, and then combine the direction features Furthermore, the AAD feature in the previous work [5]
with the AAD features to form an oriented FingerCode. does not utilize the direction information that
Experiments conducted on a large fingerprint database characterizes the oriented flow pattern of ridges and
(NIST-4) show that the proposed method produces a valleys in a fingerprint. In order to improve the overall
much improved matching performance. matching performance, we combine the AAD features
with the corresponding direction features to generate an
oriented FingerCode. Experimental results obtained from
a standard fingerprint database (NIST-4) confirm the
1. INTRODUCTION effectiveness of the proposed method.

The fingerprint matching algorithms can be broadly 2. ORIGINAL FINGERCODE

classified into two categories: (i) minutiae-based, and (ii)
filterbank-based. The minutiae-based matching algorithms The original FingerCode generation and matching
first extract the local minutiae such as ridge endings and process [5] can be summarized in the following steps:
ridge bifurcations from the thinning image [2][7] or the
1. Locate the reference point and determine the region
grayscale image [6], and then match their relative
of interest for the fingerprint image;
placement in a given fingerprint with the stored template.
2. Tessellate the region of interest centered at the
A number of matching techniques are available in the
reference point. The region of interest is divided into
literature including point-based matching [2], graph-based
a series of B concentric bands and each band is sub-
matching [1], and string-based matching [7]. Although the
divided into k sectors (B = 5 or 7, depending on the
minutiae-based matching is widely used in fingerprint
image size; k = 16);
verification, it has problems in efficiently matching two
3. Normalize the region of interest. The gray values in
fingerprint images containing different number of
every sector are normalized to a specified constant
unregistered minutiae points. Further, it does not utilize a
mean M0 and variance V0, thus to remove the effects
significant portion of the rich discriminatory information
of sensor noise and gray level deformation due to the
available in the fingerprints. The filterbank-based
finger pressure differences;
matching algorithm [3][4][5] uses a bank of Gabor filters
4. Filter the region of interest in eight different
to capture both local and global information in a
directions using a bank of Gabor filters, thus produce
fingerprint as a compact fixed-length FingerCode, which
a set of eight filtered images. The tuned Gabor filters
is suitable for matching and storage. The fingerprint
enhance the ridges and valleys that are oriented at
matching is based on the Euclidean distance between two
the same angle as the filter and suppress the ridges
corresponding FingerCodes. Thus, it overcomes some of
and valleys oriented at other angles;

0-7803-7750-8/03/$17.00 2003 IEEE. ICIP 2003

5. Compute the average absolute deviation from the of the line M12 as the reference point; otherwise, we
mean (AAD) of gray values in each of the 80 sectors assign the coordinates of M1 as the reference point.
for every filtered image to form the FingerCode;
6. Rotate the features in the FingerCode cyclically to
generate five templates corresponding to five (i, j)
rotations (45o, 22.5o, 0o) of the original fingerprint
image, thus to approximate the rotation-invariance;
7. Rotate the original fingerprint image by an angle of t
11.25o and generate its FingerCode. Another five (ic, jc) d
templates corresponding to five rotations are
generated in the same way as Step 6;
8. Match the FingerCode of the input fingerprint with
each of the ten templates stored in the database to
obtain ten matching scores. The final matching score
Figure 1. Region S is centered around pixel (ic, jc). Pixel (i, j) is
is the minimum of the ten matching scores, which
within the region S. t and d represent the tangent direction and
corresponds to the best matching of the two the local ridge direction of pixel (i,j), respectively.
fingerprints.
Figure 2 and 3 show some results of our reference
3. REFERENCE POINT LOCATION point location algorithm. The new algorithm can locate
the reference points for fingerprint images of different
The reference point is the center of the region of interest. classes. However, the algorithm may fail to locate the
Locating the reference point is an essential step that can reference point for the fingerprint with very poor quality
influence the matching accuracy. The reference point as shown in Fig. 3 (the last image). The examples shown
detection algorithm developed in [5] has been shown to be in Fig. 4 illustrate the rotation-invariance of our algorithm
more effective than most previous methods. However, we compared with the algorithm in [5].
find that it is sensitive to fingerprint rotation. In this work,
we develop a new reference point location algorithm that
is rotation invariant. Our reference point location
algorithm includes the following steps:

1. Compute the directional image using the same

algorithm as in [5]. Let D be the smoothed
directional image with D(i, j) representing the local
ridge direction at pixel (i, j);
2. Initialize a label image L with the same size of D. L
will be used to indicate the reference point;
3. As shown in Fig. 1, for each pixel (ic, jc) in D, define
a local region S centered around (ic, jc). Assign the
corresponding pixel in L the value of the following
summation:

L(ic , jc ) = cos( (i, j)

(i , j )S
t d (i, j ))

where d(i, j) represents the local ridge direction of a

pixel (i, j) in S, t(i, j) is the direction perpendicular
to the line linking (i, j) and (ic, jc). If (ic, jc) is the
center of the ridge curves, d and t will be similar at Figure 2. Examples of the reference point location results for
most points within S, thus will produce a high value different fingerprint classes.
in L;
4. Search for the maximum value M1 and the secondary 4. ORIENTED FINGERCODE
maximum value M2 in L. If M2 is over 0.95 times of
M1, there may be a double loop in the fingerprint The normalized region of interest is filtered using a bank
image, so we assign the coordinates of the midpoint of Gabor filters. The typically used Gabor filter [3][4][5]
is an even symmetric real-valued Gabor filter. It has the
following general form in the spatial domain:

1 x' 2 y ' 2
G ( x, y; f , ) = exp 2 + 2 cos(2fx' ) (1)
2 x ' y '
x' = x sin + y cos ( 2)
y ' = x cos y sin (3)

where f is the frequency of the sinusoidal plane wave

along the direction with respect to the x-axis, and x and
y are the standard deviations of the Gaussian envelope
along x- and y-axis, respectively. The filterbank consists
of eight Gabor filters, each differing in orientation only
(f = 0.1, reverse of the average inter-ridge distance; x, y
= 4.0, about half of the average inter-ridge distance.). The
eight values of are 0o, 22.5o, 45o, 67.5o, 90o, 112.5o,
135o, and 157.5o. Thus, the filtering process produces a
set of eight filtered images.

Figure 4. Examples to illustrate the rotation-invariance of our

algorithm (top row) compared with the algorithm in [5] (bottom
row).
Let Fm (m = 0, 1, , 7) be the mth filtered image. We
define a new 2-D feature vector for each sector si (i = 0, 1,
, 79) in Fm as Vim. The magnitude value of Vim is the
average absolute deviation from the mean (AAD) as
defined in [5]; and the angle of Vim is the dominant local
ridge direction in sector si. The features in all the sectors
constitute a sub-FingerCode for the filtered image Fm. The
collection of all the eight sub-FingerCodes is defined as
the oriented FingerCode. Similar to [5], to approximate
the rotation invariance, we cyclically rotate the features
and the image to generate ten templates. The matching is
based on the Euclidean distance between two
corresponding FingerCodes. Let Vim and Vjm represent the
vector features for the two corresponding sectors in the
mth sub-FingerCode. Then, the Euclidean distance
between two FingerCodes is defined as follows:
7 79
ED = V
m =0 i , j =0
im V jm

Note that there is an ambiguity of in the local ridge

direction, i.e., local ridges oriented at /4 and 5/4 cannot
be differentiated from each other. If the intersecting angle
between Vim and Vjm is larger than ninety degrees, we
reverse one of them to obtain a smaller distance.
Experimental results show that using the oriented
FingerCode gives a better performance than using AAD
Figure 3. Reference point location results of the fingerprint pairs features only. Since the local ridge direction information
of different quality. (Top row: fingerprint pairs of high quality; is computed at the reference point location stage, the
middle row: fingerprint pairs of moderate quality; bottom row: addition of the local ridge direction as a new feature will
fingerprint pairs of poor quality.) have almost no effect on the matching time.
 5. EXPERIMENTS 7. ACKNOWLEDGMENT

Experiments are conducted on a standard fingerprint The work described in this paper was fully supported by a
database NIST-4, which contains a set of 2000 fingerprint grant from the Research Grants Council of the Hong
image pairs (512x512, 256 gray levels, 500 dpi). Each Kong SAR (Project no. CUHK 4190/01E).
fingerprint pair has two different rolled impressions of the
same finger. We remove some fingerprints if their
reference points are too close to the edge of the image.
To test the performance of the oriented FingerCode,
we first implement a filterbank-based fingerprint matching
system using the original approach [5]. Here, we replace
the reference point location algorithm with the new
algorithm discussed in Section 3. The dataset is then
processed using the implemented system.
We then replace the FingerCode generation module (f = 0.1; x, y = 6) (f = 0.1; x, y = 4)
with the new algorithm to compute the oriented
FingerCode. The same dataset is again processed. Each
test fingerprint image is matched with all the other
fingerprints in the database. If the Euclidean distance
between two FingerCodes is below a threshold, we make
the decision that the two fingerprint images are
matched; otherwise, we say, the two fingerprint images
come from different fingers. The matching is labeled
correct if the matched pair is from the same finger and (f = 0.05; x, y = 4) (f = 0.2; x, y = 4)
incorrect, otherwise. The genuine acceptance rate (GAR)
and the false acceptance rate (FAR) corresponding to a Figure 5. ROC curves comparing the performance of the new
distance threshold are computed according to these FingerCode with the original FingerCode on the NIST-4
database.
labeled matching results. The overall matching
performance can be measured by a receiver operating
8. REFERENCES
characteristic (ROC) curve, which plots GAR against
FAR at different operating points (distance thresholds).
[1] A. K. Hrechak and J. A. Mchugh, Automated fingerprint
Each point on the curve corresponds to a special distance recognition using structural matching, Pattern
threshold. In the experiments, we use the Gabor filters Recognition, vol. 23, no. 8, pp. 893-904, 1990.
with various f and x, y. Figure 5 illustrates the [2] A.K. Jain, L. Hong, S. Pankanti, and R. Bolle, " An
performance improvement of the new approach. The identity authentication system using fingerprints'', Proc.
dashed line represents the performance of the original IEEE, Vol. 85, No. 9, pp. 1365-1388, 1997.
approach [5], while the solid line shows the performance [3] A. K. Jain, S. Prabhakar, and L. Hong, A multichannel
of the new approach. From the results, we can see that the approach to fingerprint classification, IEEE Trans.
new approach outperforms the original approach over a Pattern Anal. Machine Intell., vol. 21, no. 4, pp. 348-359,
1999.
wide range of FAR values, especially at low FAR values.
[4] A. K. Jain, S. Prabhakar, L. Hong, and S. Pankanti,
FingerCode: a filterbank for fingerprint representation and
6. CONCLUSION matching, Proc. IEEE Computer Society Conf. Computer
Vision and Pattern Recognition(CVPR), vol. 2, pp. 187-
In this paper, we have presented an improved method for 193, 1999.
filterbank-based fingerprint matching, which utilize both [5] A. K. Jain, S. Prabhakar, L. Hong, and S. Pankanti,
the AAD features and the direction features available in Filterbank-based fingerprint matching, IEEE Trans.
the fingerprints. Experimental results obtained from a Image Processing, vol. 9, no. 5, pp. 846-859, 2000.
large fingerprint database (NIST-4) show that the addition [6] D. Maio and D. Maltoni, Direct gray-scale minutiae
detection in fingerprints, IEEE Trans. Pattern Anal.
of the direction features leads to a substantial
Machine Intell., vol. 19 no. 1 , pp. 27-40, 1997.
improvement in the overall matching performance.
[7] N. Ratha, S. Chen, K. Karu, and A.K. Jain, " A real-time
Moreover, the proposed reference point location method matching system for large fingerprint databases", IEEE
is robust and rotation-invariant for fingerprint images. Trans. Pattern Anal. Machine Intell., vol. 18, no. 8, pp.
799-813, 1996.