Access Control System with High Level Security Using Fingerprints

Younhee Gil, Dosung Ahn, Sungbum Pan, Yongwha Chung1

Information Security Research Division
Electronics and Telecommunications Research Institute
{yhgil, dosung, sbpan}@etri.re.kr
1
Dept. of Computer and Information Science
Korea University
[email protected]

Abstract based, token based security methods have crucial

weakness that can be lost, stolen, or forgotten. In recent
Biometric based applications guarantee for years, there is an increasing trend of using biometrics.
resolving numerous security hazards. As a method of X9.84 [1] standard defines terminology of biometrics
preserving of privacy and the security of sensitive as ‘A measurable biological or behavioral
information, biometrics has been studied and used for characteristic, which reliably distinguishes one person
the past few decades. Fingerprint is one of the most from another, used to recognize the identity, or verify
widely used biometrics. A number of fingerprint the claimed identity, of an enrollee’. The fingerprint is
verification approaches have been proposed until now. one of widely used biometrics satisfying uniqueness
However, fingerprint images acquired using current and permanency [2]. Thus a number of fingerprint
fingerprint input devices that have small field of view verification approaches have been proposed until now.
are from just very limited areas of whole fingertips. Jain et al. [3] presented a minutiae-based verification,
Therefore, essential information required to which aligns minutiae using Hough transform and
distinguish fingerprints could be missed, or extracted performs minutiae matching by bounding box. Ross et
falsely. The limited and somewhat distorted al. [5] proposed hybrid matching method of local-based
information are detected from them, which might matching and global-based matching to enhance the
reduce the accuracy of fingerprint verification systems. performance. Pan et al. proposed an alignment
In the systems that verify the identity of two algorithm using limited processing power and memory
fingerprints using fingerprint features, it is critical to space to be executed in a smart card, and showed the
extract the correct feature information. In order to possibility of match-on-card [4]. In the match-on-card
deal with these problems, compensation of imperfect system, as entire verification operation is executed on
information can be performed using multiple the smart card, the system doesn’t have to maintain
impressions of enrollee's fingerprints. central database and the biometric template is
In this paper, additional three fingerprint images prevented from being streamed out of the smart card.
are used in enrollment phase of fingerprint verification Therefore, it can prevent biometric templates from
system. Our experiments using FVC 2002 databases being misused by the fraud.
show that the enrollment using multiple impressions Although it is true that technical improvement has
improves the performance of the whole fingerprint been achieved, there still exist challenging problems
verification system. relating to the quality of fingerprint images and
reliability of extracted minutiae. Most of input devices
1. Introduction get fingerprint images having fingerprints being
pressed on it not rolled, as a result, the area of
Traditionally, verified users have gained access to fingerprint images can not help being very limited.
their property or service via dozens of PIN/password, Fingerprint mosaicking [6] uses multiple fingerprint
smart cards and so on. However, these knowledge images to generate template, augmenting minutiae sets

Proceedings of the 32nd Applied Imagery Pattern Recognition Workshop (AIPR’03)

0-7695-2029-4/03 $ 17.00 © 2003 IEEE
 from plural fingerprint images on enrollment stage. Image preprocessing refers to the refinement of the
But, it does not check the reliability of each fingerprint fingerprint image against the image distortion occurred
images. during the image acquisition and transmission.
We proposed enrollment using multiple fingerprint Minutiae extraction refers to the detection of features in
images to extend enrolled fingerprint image and also the fingerprint image and finding out of their
guarantee the reliability of each fingerprint image. And information, i.e., position, direction and type
we have tested our algorithm on the first FVC 2002 Based on the minutiae, the claimed fingerprint is
database [7,8]. compared with the enrolled fingerprint. Generic
This paper is organized as follows. Section 2 minutiae matching is composed of alignment stage and
describes fingerprint verification system and the matching stage. In order to match two fingerprints
enrollment using plural fingerprints briefly. Section 3 captured with unknown direction and position, the
explains proposed methods. The experimental results differences of direction and position between two
are shown in Section 4. Finally, Section 5 contains fingerprints should be evaluated, and alignment
conclusion. between them needs to be preceded. In the alignment
stage, transformations such as translation and rotation
2. User Verification Using Fingerprint between two fingerprints are estimated, and two
minutiae sets are aligned according to the estimated
It is widely known that the fingerprint is unique, and alignment parameters. If the alignment procedure is
invariant with aging, which implies that user performed accurately, the remaining matching stage is
authentication can be relied on the comparing two referred to point matching simply. In matching stage,
fingerprints [2]. In general, a professional fingerprint two minutiae are compared based on their position,
examiner relies on details of ridge structures of the direction, and type. Then, a matching score is
fingerprint in order to make fingerprint identifications. computed.
And the structural features are composed of the points
where ridges end or bifurcate, that are called minutiae. Enrolled
PRE-
Figure 1 shows small part of an enlarged fingerprint fingerprint
PROCESSING
EXTRACTION
STORE
Enrolled
image and two types of minutiae pointed by square Minutiae
marker and circle marker. The minutia marked by Enrollment
square is bifurcation, and that by circle is ending point, Verificatoin
and the branch from minutia represents the direction of Input
fingerprint PRE-
the minutiae. Usually, each minutia is described by the PROCESSING EXTRACTION MATCH Yes
No
Input
position in the coordinate, the direction it flows and the Minutiae
type, whether it is ridge ending or bifurcation.
Figure 2:User authentication system using fingerprint

2.1. Minutiae Extraction

Our minutiae extraction algorithm mainly consists

of four components: generation of direction map,
binarization of fingerprint image, detection of minutiae,
Figure 1:Fingerprint minutiae and removal of false minutiae.
G e n e r a t e D ir e ct ion M a p
Figure 2 presents a fingerprint verification system,
which consists of two phases: enrollment and
B in a r ize I m a g e
verification. In the off-line enrollment phase, at first,
the fingerprint image of an enrollee is acquired and
preprocessed. Then, the minutiae are extracted from the D e t e ct M in u t ia e

raw image and stored as enrolled template. And in the

on-line verification phase, it reads the fingerprint from R e m ov e F a ls e M in u t ia e
a claimer, and detects the minutiae information through
the same procedure as in the enrollment phase. Then, it Figure 3:Minutiae Extraction
estimates the similarity between the enrolled minutiae
and the input minutiae.

Proceedings of the 32nd Applied Imagery Pattern Recognition Workshop (AIPR’03)

0-7695-2029-4/03 $ 17.00 © 2003 IEEE
 Besides these steps, it is critical to analyze the in regions of poor image quality, hooks, overlaps,
image and determine the areas that are degraded and minutiae that are too wide, and minutiae that are too
likely to cause problem because the image quality of a narrow. The more details about minutiae extraction can
fingerprint may vary. Several characteristics can be be found in [2].
measured that convey information regarding the quality
of localized regions in the image. These include 2.2. Enrollment with Plural Fingerprint
detecting regions of low contrast and determining Images
directional flow of ridges. Using this information, the
unstable areas in the image where minutiae detection is As automatic fingerprint identification and
unreliable can be distinguished. authentication systems rely on representing the two
most prominent minutiae, i.e., bifurcation and ridge
Generation of Direction Map. One of the ending, a reliable minutiae extraction algorithm is
fundamental steps in minutiae extraction is deriving a critical to the performance of the system. However,
directional ridge flow map. The purpose of this map is although minutiae are detected through not only
to represent areas of the image with sufficient ridge extraction process but also false minutiae removal
structure. Well-formed and clearly visible ridges are process, it is still possible to detect false minutiae and
essential to reliably detecting minutiae. In addition, the miss true ones. These faults can cause matching failure.
direction map records the general direction of the In particular, if they occur during enrollment phase and
ridges as they flow across the image. The directional are stored as enrolled template, it will be serious
information are estimated based on the 8 8 pixel sized problem, because they will affect the matching phase
windows and their range is 0~15. And background that continuously. In the other word, if the system ensures
has no variation of intensity is set as –1. This there are neither false minutiae nor missed minutiae, its
information can be used to segment images. reliability will be increased.
We suggested using plural fingerprint images on the
Binarization of Image. As our minutiae detection enrollment phase to discard the false minutiae and
algorithm is working on bi-level image, every pixel in compensate the missed minutiae. The system that
the grayscale input image must be binarized. A pixel is adopts the enrollment using multiple fingerprint images
assigned as a binary value based on the ridge flow is shown in figure 4. As this figure presents, enrolled
direction associated with the block the pixel is within. minutiae are generated from several genuine fingerprint
In order to determine whether current pixel should be images.
set to black or white, the pixel intensities of 7 9 pixel Base
fingerprint
grid rotated according to the orientation of it, which
surround the current pixel, are analyzed. Grayscale PRE- EXTRACT- ENROLLED
pixel intensities are accumulated along each rotated PROCES-
SING
ION MINUTIAE
GENERATION
Enrolled
row in the grid, forming a vector of row sums. The Genuine Minutiae
fingerprints STORE
binary value to be assigned to the center pixel is • • • •
determined by multiplying the center row sum by the •
•
•
•
•
•
•
•
number of rows in the grid and comparing this value to
the accumulated grayscale intensities within the entire
grid. If the multiplied center row sum is lower than the
total intensity of the grid, the center pixel is set to Enrollment

black. Otherwise, it is set to white. Verificatoin

Input
fingerprint PRE- Yes
EXTRACTION MATCH
Detection of Minutiae and Removal of False PROCESSING
Input
No
Minutiae
Minutiae. Before minutiae are detected, the binarized
image should be thinned, and the detection step scans Figure 4:User Authentication system that adopts the
the thinned image with some kernels that can detect enrollment using plural fingerprint images
minutiae. After detection of minutiae, candidate
minutiae points are detected. Usually, many false 2.3. Minutiae Matching Algorithm
minutiae are included in the candidate list, therefore,
removal of them are necessary to increase the As mentioned before, minutiae matching is
performance of the fingerprint verification system. The composed of alignment stage and matching stage. In
step includes removing islands, lakes, holes, minutiae order to make the explanation easier, we define the

Proceedings of the 32nd Applied Imagery Pattern Recognition Workshop (AIPR’03)

0-7695-2029-4/03 $ 17.00 © 2003 IEEE
 notation of two minutiae sets extracted from enrolled are the same. Their results affect the matching score. It
and claimed fingerprint images two minutiae sets as P depends on the size of bound box that how many
and Q respectively. candidates match. The smaller the size is, the stricter
P = {( p , p
1
x
1
y ,α 1
),..., ( p x
p
, p yp , α p
)} matching stage gets and it can be happened that the rate
= {(q , q ),..., (q )} (1) of falsely non matching is increased, and the opposite
Q 1
x
1
y ,β 1 Q
x ,q ,β
Q
y
Q

where ( p xi , p iy , α i ) and ( q xj , q yj , β j ) are the three case is the same way likewise. It is possible that there is
more than one candidate for one enrolled minutia. In
features (spatial position and direction) associated with this case, the candidate with the largest score is
the ith and jth minutia in the set P and Q respectively, selected as mated minutiae. Figure 5 shows an example
and P and Q are the number of elements in the P and Q of mating of two minutiae. Assume that m1, m2 are
set. enrolled minutiae and n1, n2 are input minutiae to be
The alignment stage gets two minutiae sets, P and Q, mated and the kinds of four minutiae are same. The
as input and estimates how their differences of position direction of each minutia is marked as branch from
and orientation were when the two fingerprints were minutia, and they are assumed to be similar enough so
captured. And then, transforms minutiae set Q for the that their difference is within the tolerance. Then,
claimed fingerprint to have same locality as the consider the mating result. In this case, it is reasonable
enrolled fingerprint according to the estimated that n1 is mated with m1 and n2 with m2. Although both
a
difference. We denote aligned minutiae set Q as Q . n1 and n2 are candidates of the mated minutiae with m2,
For the purpose of proper alignment, the estimation of n2 must be regarded as mated minutia with m2 because
the rotation and translation parameters must precede. n2 is nearer to m2 than n1. Also, correlation of n1 with
In order to estimate transformation parameters, we find m1 should be considered. Even if n1 is a mated
out ( ∆ x , ∆ y ) and ∆ θ satisfying formula (2) according candidate of not only m1 but also m2 and it is nearer to
to the formula (3). m2 than m1, it should be mated with m2, as m2 has been
Fθ ,∆x ,∆y (( q x , q y , β ) T ) = ( p x , p y , α ) T (2) mated with n1 already.
§ x · § cos ∆ θ sin ∆ θ 0 ·§ x · § ∆ x ·
¨ ¸ ¨ ¸¨ ¸ ¨ ¸
Fθ , ∆ x , ∆ y ¨ y ¸ = ¨ − sin ∆ θ cos ∆ θ 0 ¸¨ y ¸ + ¨ ∆ y ¸ (3)
¨θ ¸ ¨ 0 0 1 ¸¹ ¨© θ ¸¹ ¨© ∆ θ ¸¹
© ¹ ©
where ( ∆ x , ∆ y ) and ∆ θ are the translation and
rotation parameters; ( p x , p y ,α ) T represents the enrolled
minutiae and ( q x , q y , β ) T represents the claimed
minutiae.
The evaluated transformation parameters
Figure 5:Mating of minutiae
( ∆ x , ∆ y , ∆ θ ) T are used to align claimed minutiae by
§q a
· § cos ∆ θ sin ∆ θ 0 ·§ q x · § ∆ x ·
¨ x
¸ ¨ ¸¨ ¸ ¨ ¸ 3. Multiple Fingerprint Images and
¨q a
y ¸ = ¨ − sin ∆ θ cos ∆ θ 0 ¸¨ q y ¸ + ¨ ∆ y ¸ (4)
¨β a ¸ ¨ 0 0 1 ¸¹ ¨© β ¸¹ ¨© ∆ θ ¸¹ Enhanced Matching Method
© ¹ ©
where ( q , q , β ) is the aligned of the claimed
a
x
a
y
a T
Usually, even superior extraction process including
minutiae. removal of false minutiae have some false minutiae
After alignment step, the comparison of the remained still. And it can miss the true ones. The
information of two minutiae sets, P and Qa, is performance of fingerprint verification is influenced by
accomplished by point pattern matching in the polar both of them. Figure 6 is enlarged part of fingerprint
coordinate system with respect to the center of images and shows some extracted minutiae. In Figure 6
foreground. Such a point matching can be (a), examples of false minutiae are presented being
accomplished by placing bounding box [3] around pointed out with black arrows. They are neither the
enrolled minutiae. When an aligned minutia is placed bifurcation nor the ending of ridges, are just points on
within the bounding box, the minutia is considered as a the middle of ridges. They should not have been
candidate of the mated one with the enrolled minutiae. detected. These are caused by couple of reasons such
And two additional conditions will be checked: as failure in extraction stage or noise of fingerprint
whether the difference between their directions is image itself. However, these can be eliminated based
below predetermined tolerance, and whether their types on the fact that they are temporal and no false minutia

Proceedings of the 32nd Applied Imagery Pattern Recognition Workshop (AIPR’03)

0-7695-2029-4/03 $ 17.00 © 2003 IEEE
 on the same position in one image as another area is used. This method first measures the overlapped
fingerprint image may merely be detected. Therefore, area of two fingerprint images, and then finds out
plural fingerprint images are used on the enrollment minutiae pairs in the area. To estimate the intersected
phase in order to discard the false minutiae. area, segmentation of two images should precede.
Next, it estimates the ratio of minutiae pairs to total
enrolled minutiae. It regards two fingerprint images as
ones from the same finger when the ratio is over the
predetermined threshold. And in the case when two
minutiae sets are determined to be identical, it appends
the enrolled minutiae in the non-overlapped area to the
claimed minutiae set. Then more mated pairs can be
(a) (b) found and it will improve matching performance.
Figure 6:Extracted minutiae including false minutiae
superimposed on the enlarged fingerprint image I n pu t base fi n ger pr i n t i m age, I (B )
and discarded result of false minutiae
E x t r act m i n u t i ae of I (B )
M (B )= {(x i ,y i ,t i )}, i = 1,¡ ,L
In the enrollment phase using multiple impressions, F i n d cen t er , C T = (x c,y c)
at first, one fingerprint image is acquired and set as
base image, and minutiae from it are extracted. Then Con v er t m i n u t i ae t o pol ar coor di n at e
sy st em w i t h r espect t o C T ,
after segmentation of image, find out the center of P(B )= {(l i ,p i ,t i )}, i = 1,¡ ,L

foreground, which will be the reference point of polar I n pu t gen u i n e i m age, I (G ) n , n = 1,¡ ,N
coordinate system. And then, minutiae set from base
fingerprint image are transformed into polar coordinate E x t r act m i n u t i ae of I (G ) n ,
M (G ) n = {(x i n ,y i n ,t i n )}, i n = 1,¡ ,L n
system. Once finishing above processes to the base
image, a few genuine fingerprint impressions are A l i gn M (G ) n t o M (B ),
M (G ) n ¡= {(x ¡i n ,y ¡i n ,t ¡i n )}, i n = 1,¡ ,L n
acquired. And minutiae from each fingerprint image are
extracted, and each minutiae set is aligned with base Con v er t m i n u t i ae t o pol ar coor di n at e
sy st em w i t h r espect t o C T ,
fingerprint image. Then, they are converted into polar P(G ) n = {(l ¡i n ,p ¡i n ,t ¡i n )}, i n = 1,¡ ,L n
coordinate system with respect to the center of base
Sel ect v al i d base t em pl at e com par i n g
image. Now, each minutia from one fingerprint image P(B ) an d P(G ) n
is examined if it can be mated with the minutia from
the other fingerprint image. To do this, it is required to St or e v al i d base t em pl at e

compute similarity between two minutiae. The minutia Figure 7:Enrollment using multiple impressions
is regarded as false minutia if it has been mated with no
minutia during whole examinations, and discarded.
Figure 6 (b) shows the false minutiae have been
4. Experimental Results
discarded through our false minutiae discard step. The
We have tested our fingerprint verification
simple flowchart of false minutiae discard algorithm is
algorithm using one of the FVC 2002 databases [7,8].
shown in Figure 7.
There are four different databases in the FVC 2002
Another factor reducing the accuracy of the
databases, each of which were collected by different
verification is limited contact area of fingerprint
scanners or generated by using SFinGE software.
scanner. As input device is getting smaller, the size of
Hence, they have different image size and resolution.
window acquiring image has been shrunk. And owing
Every database consists of two classes, set A for
to the tiny window, the verification system use only
evaluation and set B for training. Set A is composed of
partial regions of fingerprint images. Therefore, two
eight fingerprint images per one finger from 100
identical fingerprint images can be misjudged they are
individuals for a total of 800 fingerprint images and set
from different fingerprints if two images are scanned
B from 10 individuals for a total of 80 images. The
from the opposite parts of the fingertip. Besides the
details about FVC 2002 databases are in [8]. Among
extreme case like this, when the overlapped area of two
them, A set of DB1 was used in our experiment. The
images is small, it causes low matching score.
In order to compensate one minutiae set from the size of fingerprint images of DB1 was 388 374 at
identical fingerprint images that have little common 500dpi.
area, padding method of minutiae in non-overlapped For the enrollment, the first four fingerprint images
among eight were used, and the remaining four were

Proceedings of the 32nd Applied Imagery Pattern Recognition Workshop (AIPR’03)

0-7695-2029-4/03 $ 17.00 © 2003 IEEE
 used for verification. Therefore, when a matching was on-card system is shown in Figure 9. In the match-on-
labeled GENUINE if the matching was performed card system, the whole matching step should be
between fingerprint images from same finger, and executed in the smart card and the processing power of
IMPOSTER otherwise, 400 GENUINEs were able to smart card is very limited [4]. Therefore, it is essential
be performed. And 9900 IMPOSTERs were performed, to extract reliable minutiae in order to make matching
i.e., 99 IMPOETERs have been tested for one more accurate even though matching step uses only
fingerprint using the other 99 fingerprints. simple information of minutiae such as the position,
We performed two matching tests in order to show direction and type.
the effect of the adoption of enrollment using multiple Although the performance of matching adopting the
impressions, matching after enrollment using single proposed enrollment was improved considerably, the
impression and matching after enrollment using result is not satisfied yet. Thus, it is required to enhance
multiple impressions. Figure 8 presents the distribution the algorithm and study new method.
of false match rate and false non-match rate. Vertical
axis represents the normalized distribution of matching
scores, and horizontal axis represents the score ranging
from 0 to 100. As shown in the curve of Figure 8 (b), it Enrolled PRE-
Fingerprint EXTRACTION
is observed that equal error rate is lower by 1.38% PROCESSING
Enrolled STORE
Minutiae
when using multiple impressions on enrollment than Enrollment

when using single impression. Verification

When false non-match rate is set 1%, false matches Input PRE- MATCH Yes
happen at the rate of about 40% in the former. In the Fingerprint PROCESSING EXTRACTION
Input
No

other hand, the false match rate of the latter is 6.15%. Minutiae

Host PC Smart card

Figure 9:Fingerprint based Match-on-Token

References
[1] ANSI web site, http://www.ansi.org/
[2] Jain, L.C., Halici, U., Hayashi, I., Lee, S.B., Tsutsui, S.:
(a) (b) Intelligent Biometric Techniques in Fingerprint and Face
Figure 8:FMR/FNMR curve, (a) case when using one Recognition, CRC Press LLC, (1999)
fingerprint image on the enrollment, (b) case when [3] Jain, A., Hong, L., Bolle, R.: On-line Fingerprint
using four fingerprint images on the enrollment Verification. IEEE Trans. on Pattern Analysis and
Machine Intelligence, Vol.19, No.4 (1997) 302–313
[4] Pan, S.B., Gil, Y.H., Moon, D., Chung, Y., Park, C.H.: A
5. Conclusion and Future Work Memory-Efficient Fingerprint Verification Algorithm
using A Multi-Resolution Accumulator Array, ETRI
This article shows the comparison between matching Journal, Vol. 25, No. 3, (2003) 179–186
performances when using single fingerprint image on [5] Ross, A., Jain, A. K., Reisman, J.: A Hybrid Fingerprint
enrollment phase and when using multiple fingerprint Matcher, Pattern Recognition, Vol. 36, No. 7, (2003)
images. If plural fingerprint images are used during 1661-1673
enrollment, extracted and stored minutiae can be [6] Jain, A. K., Ross, A.: Fingerprint Mosaicking, Proc.
quaranteed becuase false minutiae are discarded and ICASSP, (2002)
missed ones are added in the final minutiae set. And we [7] FVC 2002 web site, http://bias.csr.unibo.it/fvc2002
[8] Maio, D., Maltoni, D., Cappelli, R., Wayman, J.L., Jain,
confirmed this by experiments. On using multiple
A.K.: FVC2002:Second finger-print verification
impressions on enrollment, equal error rate decreased competition, Proc. ICPR, (2002)
by 1.38% as compared with the case when using single
impression, and its FMR 100[7] is 6.15%.
The enrollment phase using multiple impressions
can be part of match-on-cord [4] system. The match-

Proceedings of the 32nd Applied Imagery Pattern Recognition Workshop (AIPR’03)

0-7695-2029-4/03 $ 17.00 © 2003 IEEE