Proceedings of 2010 IEEE 17th International Conference on Image Processing September 26-29, 2010, Hong Kong

A HYBRID FUSION METHOD OF FINGERPRINT IDENTIFICATION

FOR HIGH SECURITY APPLICATIONS

Yilong Yin, Yanbin Ning, Zhiguo Yang

School of Computer Science and Technology

Shandong University, Jinan 250101, China
ylyin@[Link] , ningyanbin009@[Link] , yzhgdd@[Link]
ABSTRACT These sources include multiple sensors, multiple features,
multiple matchers, multiple fingers, multiple impressions of
Though fingerprint identification is widely used now, its a same finger, etc. (iii) Combining fingerprint with other
imperfect performance for some high security applications, biometrics traits to construct a more robust and effective
such as ATM, the access control of nuclear power stations biometrics system. These efforts mainly aim to decrease
and exchequers, etc, is still a challenge. In high security equal error rate (EER) of a biometrics system and can
applications, an extremely low false accept rate and as low indeed improve the performance of identification.
as possible false reject rate are desired at the same time, EER is used as the most important parameter to
which is called Double Low problem in this paper. It is to evaluate the performance of a fingerprint system now and it
be noted that even a fingerprint system with very low equal can indicate the general performance of a fingerprint system
error rate can not achieve such a Double Low goal. It is fairly well. A very low EER usually denotes that a
difficult to solve Double Low problem only by improving fingerprint system has very high performance [2, 3].
the performance of a certain individual fingerprint However, it is not always suitable to mainly use EER to
identification algorithm, and the fusion of various evaluate the performance of a fingerprint system. Some high
fingerprint identification algorithms becomes a promising security applications, such as ATM, the access control of
way. In this paper, a hybrid fusion method of fingerprint nuclear power stations and exchequers [4, 5], etc, have
identification is proposed to solve Double Low problem. special demands to the performance of a fingerprint system.
Firstly, minutiae-based and ridge-based matching In these applications, if an authentic person is mistakenly
algorithms are used orderly, which is a kind of serial fusion rejected, it just is troublesome. While, if an impostor is
strategy. Secondly, a rank-level fusion is used, which is a mistakenly accepted, it may be a disaster. Two kinds of
kind of parallel fusion strategy. Experiment results on errors will cause different amount of losses. The second
FVC2002DB1 and FVC2002DB2 indicate that only 6.6% error is far more serious than the first one. It is a cost-
fingerprints are falsely rejected on the average under zero sensitive problem in fact. To meet the demand of such high
false accept rate with our method, while 14.8%, 9.4% security applications, an extremely low false accept rate
fingerprints are falsely rejected under zero false accept rate (FAR) and as low as possible false reject rate (FRR) are
with the serial fusion strategy and the parallel fusion desired at the same time, which is called Double Low
strategy, respectively. problem in this paper.
Even a fingerprint system with a very low EER could
Index Terms—fingerprint identification, hybrid fusion, not achieve such a Double Low goal. A system with low
serial fusion, parallel fusion, high security application EER might have a bad performance in high security
applications for its FRR will rise acutely when its FAR
1. INTRODUCTION AND MOTIVATION becomes very low. For example, to a minutiae-based
fingerprint system with EER of 2.7 %, its FRR can reach
During the recent years, fingerprint identification has 15.8% when its FAR is zero on FVC2002DB2! This can not
received more and more attention and been widely used in meet the demand of high security applications obviously.
various fields due to its university, distinctiveness, Fig.1 shows a ROC curve of a fingerprint system.
permanence and acceptability [1]. Unlike common applications, only the part of ROC curve in
Efforts for fingerprint identification are mainly focused the shadow is concerned for high security applications.
on: (i) Improving the performance of one or more steps of In fact, Double Low problem exists all the time and has
automatic fingerprint verification system. The steps include important applications. However, we pay little attention to it
segmentation, enhancement and matching, etc. (ii) Using and few efforts are taken about it until now.
multiple sources of a fingerprint to get a higher accuracy.

978-1-4244-7994-8/10/$26.00 ©2010 IEEE 3101 ICIP 2010

 It is difficult to solve Double Low problem only by matching algorithm are used orderly. The second part is a
improving the performance of a certain individual rank-level parallel fusion which fuse the two matching
fingerprint identification algorithm for the limitation of algorithms. Firstly minutiae features are extracted after a
techniques, and fusion of various fingerprint identification query fingerprint image is acquired. Then minutiae-based
algorithms becomes a promising way. matching is used to match the query fingerprint with all
fingerprints in the template database and multiple matching
scores are acquired. The maximum matching score is
compared with a threshold T1. The identification is
FRR successful if the maximum score is higher than T1.
Otherwise, we will deal with the query fingerprint with
ridge-based matching algorithm. In the same way, the
identification is successful if the maximum matching score
is higher than another threshold T2. Otherwise, we identify
FAR the query fingerprint with rank-level fusion method. In this
part, for every query fingerprint, the minutiae-based
Fig. 1. ROC curve.
matching score and the ridge-based matching score are all
To solve Double Low problem, a hybrid fusion method required. The results of identification are obtained
is proposed and two popular fingerprint matching according to the rule of rank-level fusion. The following is
algorithms, minutiae-based algorithm and ridge-based detailed descriptions of the two fingerprint matching
algorithm [3], are used in this paper. algorithms and the rule of rank-level fusion used in the
proposed hybrid fusion method.
2. THE PROPOSED HYBRID FUSION METHOD
2.1. Minutiae-based matching algorithm
The framework of the proposed method is shown in Fig. 2.
Query Fingerprint Template Database Fingerprint Minutiae Local structure
preprocessing extraction computation
Best matched
Global structure Minutiae structures Local structure
matching correspondences matching
Minutiae Global matching score Local matching score
Extraction
Fusion
Ridge Ridge
>=T Final matching <T
Success Failure
Extraction Extraction score
Fig.3. Framework of minutiae-based matching algorithm.
Minutiae-Based S1>=T1
Matching Success There are already some minutiae-based matching algorithms.
In our proposed hybrid fusion method, we choose a typical
minutiae-based matching algorithm which matches the
S1<T1 fingerprint minutiae using both the local and global
structures of minutiae [6], whose framework is shown in Fig.
3. The local structure of a minutia is rotation and translation
invariant because it consists of the direction and location
S2>=T2
Ridge- Based relative to some other minutiae. It is used to find the
Success
Matching correspondence of two minutiae sets and to increase the
reliability of the global matching. Moreover, the local
S2<T2 structure can tolerate some deformation because it is formed
from only a small area of the fingerprint. So the local
Serial fusion part Rank-Level Fusion Parallel fusion part structures can be directly used for matching and the best
matched local structures will provide the correspondences
for aligning the global structure of the minutiae. The global
Fig. 2. Framework of the proposed hybrid fusion method. structure of minutiae reliably determines the uniqueness of
fingerprint. The aligned global structure together with the
As shown in Fig. 2, the proposed hybrid fusion method result of the local structure matching finally determines
consists of two parts. The first part is a serial fusion in whether the two fingerprints are acquired from the same
which minutiae-based matching algorithm and ridge-based

3102
finger. Therefore, the local and global structures of minutiae Unlike a fingerprint verification system, a fingerprint
together provide a solid basis for reliable and robust identification system typically outputs a ranking or a
minutiae matching. candidate list instead of a match score or a Boolean value.
So, the rank 1 template fingerprint is more similar with the
2.2. Ridge-based matching algorithm query fingerprint than the rank 2 template fingerprint, and
so forth. Rank-level fusion [8] is a kind of parallel fusion
The ridge-based algorithm [7] chosen in this paper consists and is used when the output of matchers is the rank of the
of three stages: preprocessing, alignment and matching, “candidates” in a template database. This kind of method is
whose framework is shown in Fig. 4. In the preprocessing expected to assign a higher rank to a template fingerprint
stage, ridges are extracted by sampling equidistantly from that is more similar to the query fingerprint. The goal of
the thinned image. The relations between ridges and rank-level fusion method is to combine ranks assigned by
minutiae are established. In the alignment stage, a set of N various fingerprint matchers to derive an “integrative” rank
initial substructure pairs is found using a novel approach. In for each identity. There are three rules usually used to
the matching stage, for each of the N initial substructure combine ranks assigned by different matchers, namely, the
pairs, ridge matching is performed to produce a matching highest rank rule, the Borda count rule, and the logistic
score. Finally, the maximum of the N scores is used as the regression rule. However, the above three rank fusion rules
final matching score of the two fingerprints. The alignment are relatively loose and they are unsuitable for applications
algorithm focuses on how to choose a reliable local feature of high security because the demand of Double Low is very
pair as the datum mark of matching. This is accomplished strict. Therefore in this paper, we increase the restricting
firstly by defining a substructure that contains as much local condition of fusion and use a far more rigorous rank fusion
information (one minutia and several ridges) as possible, rule with which the query fingerprint is regarded to be
and secondly by finding the substructure pair which have successfully identified only when its identities
the most consistent substructure pairs around. In the corresponding with the highest rank of different matchers
matching algorithm, during the process of ridge matching, are same. As to two matchers of minutiae-based and ridge-
minutiae are also paired, and the matching score is based used in this paper, the rule can be called the double
computed according to both the matched minutiae and the highest rank fusion.
matched ridges.
4XHU\)LQJHUSULQW 7HPSODWH)LQJHUSULQW
3. EXPERIMENT RESULTS

In this section, we design an experiment to testify the effect

3UHSURFHVVLQJ 3UHSURFHVVLQJ of the proposed method for the high security applications.
We select two fingerprint databases, namely, FVC2002DB1
&OHDUBXS &OHDUBXS and FVC2002DB2, used in this experiment. In each selected
fingerprint database above, there are one hundred fingers
and eight hundred fingerprint images, and every finger
'LVSHUVLQJ 6XEVWUXFWXUH 6XEVWUXFWXUH 'LVSHUVLQJ corresponds to eight fingerprint images. We select one
image of each finger to constitute the template database and
the other seven images to constitute the query database for
above two selected databases. As to each selected
6XEVWUXFWXUH0DWFKLQJ fingerprint database, the template database has one hundred
fingerprints and the query database has seven hundred
fingerprints. We do the experiment with two individual
,QLWLDO6XEVWUXFWXUH3DLU methods, two fusion methods and the proposed hybrid
fusion method on template databases and query databases.
The detailed explanation of the front four methods is as
,QLWLDO5LGJH3DLU
follows:
(i)Two individual methods: One is the minutiae-based
matching algorithm and the other is the ridge-based
5LGJH0DWFKLQJ
matching algorithm. They are called individual method 1
and individual method 2 in this paper, respectively.
(ii)Two fusion methods: One is the serial fusion method
0DWFK6FRUH of the minutiae-based matching algorithm and the ridge-
Fig.4. Framework of ridge-based matching algorithm. based matching algorithm, and the other is the rank-level
parallel fusion method of the two individual methods. They
2.3. Rank-level fusion

3103
are called fusion method 1 and fusion method 2 in this paper, as a datum mark to compare the performance of above five
respectively. methods. So the threshold T1 and T2 are the minimum
Extremely low FAR is a relative and theoretical thresholds to assure zero FAR for minutiae-based matching
conception. For the sake of comparing the performance of algorithm and ridge-based matching algorithm respectively.
above five methods quantificationally, a certain reference Experiment results are given in table 1.
value of FAR must be fixed above all. We select zero FAR

The individual The individual The fusion The fusion The hybrid fusion
Database method 1 method 2 method 1 method 2 method
FRR FRR FRR FRR FAR FRR FAR
FVC2002DB1 22.9% 56.1% 17.6% 12.1% 0 8.3% 0
FVC2002DB2 15.8% 47.4% 12.0% 6.7% 0 5.0% 0
Total 19.1% 51.8% 14.8% 9.4% 0 6.6% 0
Table 1. The performance of five methods.

As shown in Table 1, using the individual method 1 5. ACKNOWLEDGEMENTS

and individual method 2, 19.1%, 51.8% fingerprints are
falsely rejected under zero FAR on the average on This work was supported in part by Shandong Province
FVC2002DB1 and FVC2002DB2, respectively. Using the Natural Science Foundation under No. Z2008G05,
fusion method 1 and the fusion method 2, 14.8%, 9.4% Shandong University Independent Innovation Foundation
fingerprints are falsely rejected under zero FAR, 2009TS034 and 2009TS035.
respectively. Using the proposed hybrid fusion method,
only 6.6% fingerprints are falsely rejected under zero FAR. 6. REFERENCES
Though FARs of five methods in Table 1 are all zero,
it is to be noted that, for the individual method 1, the [1] A. A. Ross, K. Nandakumar, and A. K. Jain, Handbook of
individual method 2 and the fusion method 1, thresholds T1 multibiometrics, Springer-Verlag, New York, 2006.
and T2 themselves are selected under zero FAR, while for
the fusion method 2 and the proposed method, zero FARs [2] Jianjiang Feng, “Combining minutiae descriptors for
fingerprint matching,” Pattern Recognition, vol. 41, issue 1, pp.
and FRRs of 9.4% and 6.6% are practically acquired by the
342-352, 2008.
rigorous fusion rule, respectively. The uniform
precondition of zero FARs for five methods makes it [3] D. Maltoni, D. Maio, A. K. Jain, and S. Prabhakar, Handbook
feasible to compare their performances. of fingerprint recognition, Springer-Verlag, New York, 2009.G

4. CONCLUSIONS AND FUTURE WORKS [4] A. K. Jain , S. Prabhakar, and S. Chen, “Combining multiple
matchers for a high security fingerprint verification system,”
In this paper, Double Low problem is illustrated and its Pattern Recognition Letters, vol. 20, issues 11-13, pp. 1371-1379,
characteristics are analyzed. A hybrid fusion method is 1999.
proposed and its implementation is described. Experimental
[5] L. Marcialis, and F. Roli, “High security fingerprint
results indicate that the proposed hybrid fusion method has verification by perceptron-based fusion of multiple matchers,”
better performance than existed methods and it can solve Multiple Classifier Systems, vol. 3077, pp. 364-373, 2004.
Double Low problem to some degree. We consider that the
proposed hybrid fusion method can not only be used to fuse [6] X. D. Jiang, and W. Y. Yau, “Fingerprint minutiae matching
different fingerprint identification methods but also can be based on the local and global structures,” 15th International
used as a framework to fuse different biometrics to achieve Conference on Pattern Recognition, vol. 2, pp. 1038-1041, 2000.
Double Low goal. It has considerable value for high
security applications in biometrics field. [7] J. J. Feng, Z. Y. Ouyang, and A. Cai, “Fingerprint matching
using ridges,” Pattern Recognition, vol. 39, issue 11, pp. 2131 -
Future work will focus on two aspects. One is to fuse
2140, 2006.
more different fingerprint algorithms with the proposed
hybrid fusion method to solve the Double Low problem [8] M. M. Monwar, and M. L. Gavrilova, “Multimodal Biometric
more thoroughly for fingerprint identification. The other is System Using Rank-Level Fusion Approach,” IEEE Trans. On
to try the proposed hybrid fusion method with different Sysgtems, Man And Cybernetics-Part B: Cybernetics, vol. 39, no.
biometrics for high security applications. 4, pp. 867 - 878, 2009.

3104