You are on page 1of 12

An Idea of Biometric Traits Selection for Multibiometrics

Li Weijie Xu Yong
Bio-computing Research Center, Shenzhen Graduate School, Harbin Institute of Technology
E-MAIL: weijiekaoyan@163.com, laterfall2@yahoo.com.cn

Abstract

In this paper, we propose an idea of biometric traits selection for multibiometrics. “Biometric
traits selection” means that selection a given number of biometric traits from those people can
obtain for constructing a multibiometric system which has the best performance. We utilize the
correlation coefficients between different biometric traits to analysis relations of them and select
those most irrelevant to construct a multibiometric system and value its performance. The results
of experiments show that the method proposed by us is rational and effective. The multibiometric
system constructed by method in this paper is in its best performance.

1. Introduction

The pronounced need for reliably identifying or verifying the identity of a person in electronic
society has spurred active research in the field of biometric authentication, which is also called as
biometrics simply. Biometrics is the science of establishing an identity based on the physical or
behavioral attributes of an individual including fingerprint, face, palmprint, hand geometry, voice,
gait, iris, signature, ear and so on. In about half century, there are many applications based on
biometrics used in domain of military and civil affairs [1, 2, 3, 4 and 5]. But also we can find that
those traditional systems, unimodal biometrics systems, which applies a single biometric trait for
identifying or verifying the identity of a person, is not sufficient to meet a number of system
requirements [6]. Then researchers turn to multibiometric systems in all over the world.
Multibiometric systems seek to alleviate some of the drawbacks encountered by unibiometric
systems by consolidating the evidence presented by multiple biometric sources [6, 7]. Since
proposed in 1990s, there have been many papers on the multibiometric systems [6, 7]. From an
academic perspective, research in multibiometrics has several different facets: identifying the
sources of multiple biometric information, determining the type of information to be fused,
designing optimal fusion methodologies, evaluating and comparing different fusion
methodologies, and building robust multimodal interfaces that facilitate the efficient acquisition of
multibiometric data [7]. In this paper, we only give our attention to biometric traits selection,
selecting some ones from optional biometric traits for constructing a multibiometric system which
has the best performance, belonging to the second facet.
In this paper we utilize the correlation coefficients between different biometric traits to
analysis the relations of them and select those most irrelevant to construct a multibiometric system
and value its performance. The paper is organized as follows. Section 1 introduces some basic
information about multibiometrics and the organization of the paper. In section 2 we propose our
method for biometric traits selection. The test of method proposed in section 2 is in section 3. We
conclude the paper in the section 4.

2. Biometric traits selection

“Biometric traits selection” means that selection a given number of biometric traits from those that
people can obtain for constructing a multibiometric system which has the best performance. We
can’t get a criterion for selection traits now [8, 9]. Many researchers just fuse what they can obtain
[10-14]. There may be some drawbacks.

We can suppose that there is a multibiometric system having n classifiers and si is the score

from the ith-classifier. And we also assume that the genuine posterior probability,
P ( genuine | si ) , represents the probability of someone being genuine, given a score for a

particular classifier while the imposter posterior probability, P (imposter | si ) , represents the

probability of someone being imposter, given a score for a particular classifier. There are a number
of well-know techniques applied in many systems [15], shown in Table 1. The simple sum rule
adds the scores of each classifier to estimate the fused score. The minimum score fusion method
selects the score having the least value of classifiers while the maximum score fusion method
selects the score having the greatest value of the classifiers. The sum of probabilities and product
of probabilities by adding and multiplying respectively these probabilities for all classifies.
Table 1. Summary of Fusion Techniques
Rule Fused score of multibiometric system s f

Simple n

s
i 1
i
Sum
Minimum min{s1 , s2 ,..., sn }

Score
Maximum max{s1 , s2 ,..., sn }
Score
Sum n

Probabilities
 p( genuine | s )
i 1
i

Product n

 p( genuine | s )
i 1
i
Probabilities

We turn to one of drawbacks now. It is assumed that all of scores, s1 , s2 ,..., sn , are absolute

relative, such as
s j   j sk (1)
where k , j  {1, 2,..., n  1} . And  j  1 when j  k . So we can convert sum rule, minimum

score, maximum score, the sum probabilities and the product probabilities into follows:
n n
s f   si (1  2  ...  n ) sk  sk  i (2)
i 1 i 1

s f  min{s1 , s2 ,..., sn }  min{1 , 2 ,...n }sk (3)

s f  max{s1 , s2 ,..., sn }  max{1 , 2 ,...n }sk (4)

n n
s f   p ( genuine | si )   p ( genuine |  j sk ) (5)
i 1 i 1

n n
s f   p ( genuine | si )   p( genuine |  j sk ) (6)
i 1 i 1

From Equ. 2, Equ. 3, Equ. 4, Equ. 5 and Equ. 6, it is noticed that the fused score s f can be

expressed analytically as a function of sk from one of n classifiers. So we conclude that we can

replace the multibiometric system by a unimodal biometrics system which utilizes the k-th
classifier. In other word, we can get small performance by a unimodal biometrics system as the
multibiometric system.

For a sub set {sk1 , sk2 ,..., skm } of {s1 , s2 ,..., sn } , if elements in it are relevant each other, then

we can say that we can replace the fused score obtained using some fusion techniques as a
function of one element in the subset. In this paper, we call this sub set whose elements are
correlative each other as correlative subset. The biometric traits selection can be stated as follows:

given a score set {s1 , s2 ,..., sn } , where si is the score from the i-th classifier using some biometric

trait, i  1, 2,..., n and n is the number of classifiers, find out all correlative sets of {s1 , s2 ,..., sn } .

It is obvious that when elements in {s1 , s2 ,..., sn } are uncorrelative to each other, there are n

correlative subsets and each set has only one element and when all elements in {s1 , s2 ,..., sn } are

relevant, there is only one relevant subset which is also equal to {s1 , s2 ,..., sn } . When all elements

are correlative to each other, the performance of the multibiometric system fusing all the match
scores may be not better than that of a unimodal biometrics system applying one of them.
On selecting biometric traits, besides criteria for applying a biometric trait such as
universality, uniqueness, permanence, measurability, performance, acceptability and
circumvention and so on [3, 6], we think we should also stand by following criteria:
1. The technology of the biometric trait must be mature one. If it is on researching state,
we should keep it in domain of biometrics. For immature technology, it will introduce
some unknown problem in a multibiometric system.
2. The match scores from classifiers being fused should be irrelevant or independent.
The gain in the performance of a multibiometric system applying score level fusion is
affected by the correlation between match scores emitted by the different classifiers
[16, 17, and 18].
Now we turn to how to calculation the correlation coefficient of biometric traits for
multibiometric system. We assume that the size of train dataset is M and the size of test dataset is
N . It is known that there are n classifiers and n biometric traits. Each classifier refers to only one
biometric trait. So the match score between the i-th sample of train and the j-the one of test dataset
i, j
using the k-th classifier can be represented as sk . Given the i-the test sample, we get a match

score matrix after matching:


 s1i1 s1i 2 ... s1iM 
 i1 
 s2 s2i 2 ... s2iM 
(7)
 ... ... ... ... 
 i1 
 sn sni 2 ... sniM  
After match, we can utilize the matrix to estimate the correlation coefficient between the u-th and
i i
the v-th biometric traits using the u-th row vector su and the v-row vector sv :

cov( sui , svi )


ruvi  (8)
D( sui ) D( svi )
i i i i i i i i
where cov( su , sv ) is the covariance of su and sv , D ( su ) and D ( sv ) is the variance of su and sv

i i i i
respectively. cov( su , sv ) , D ( su ) and D ( sv ) are estimated using following equations:

M
1
s  i
u
M
s j 1
ij
u (9)

M
1
svi 
M
sj 1
ij
v (10)

M
1
D( sui ) 
M
 (sj 1
ij
u  sui ) 2 (11)

M
1
D( svi ) 
M
 (sj 1
ij
v  svi ) 2 (12)

M
1
cov( sui , svi ) 
M
 (s
j 1
ij
u  sui )( svij  svi ) (13)
1 2 N
Then we apply the mean of ruv , ruv ,..., ruv to measure the relationship of the u-th and the v-th

biometric traits:
N
1
ruv 
N
r
i 1
i
uv (14)

Now we can state biometric traits selection as follows in tails:


It is assumed that we want to select m biometric traits from n ones to construct a
multibiometric system where m must be less than n . The size of test and train dataset is N and M
respectively. And there are n classifiers every one of which refers to only a biometric trait in the

system. Then the problem is to find m correlative subsets from a match score set {s1 , s2 ,..., sn } .

Let X  {s1 , s2 ,..., sn } and m correlative subsets be X 1 , X 2 ,..., X m . After biometric traits

selection, the sets should satisfy with the following properties:

1. X  , X 1  X 2  ...  X m  X ;

2. X i  X j  ,1  i, j  m

For constructing the i-th correlative subset, the processes are as follows:

1. Let X i   ;

2. When X   which means that we can’t construct the i-th correlative subset, stop;

3. When X   ,select a score sk from X randomly , and X  X  {sk } , X i  X i  {sk } ;

4. For any element su  X , if ruk , the correlation coefficient between su and sk , satisfies with

ruk  r0 , then X  X  {su } , X i  X i  {su } , else stop.

where r0 is the threshold of correlation coefficients. When the correlation coefficient between the

u-th and v-th biometric traits ruv  r0 , we think the u-th biometric trait and the v-th one are

relative and the two biometric traits should be in the same correlative subset. When ruv  r0 , we

think the u-th biometric trait and the v-th biometric trait are uncorrelative and the two biometric

traits should not be in the same correlative subset. The threshold r0 should be 1 ideally.

3. Experiments

In this paper we use the multibiometric database acquired from 42 persons who are members
of our Bio-computing Research Center to test our idea. In the multibiometric database, there are
facie images, left and right palmprint images, fingerprint images including index, middle and third
fingers from left and right hands, left and right hand vein images, left and right hand back vein
images, tongue images and so on. Before acquiring the biometric traits, each one is given a
number as his or her identity number from 01 to 42. Each biometric trait is acquired 10 times. And
each trait also is given an identity number. So we distinguish traits using the person number, type
of traits, and the trait number. For example, the serial number of the fifth of traits of left index
fingerprint image from the third person is “03LI05”. Figure 1 shows a multibiometric trait.

Figure 1. A multibiometric trait


In this paper, we only consider facial images and left and right palm images. And we assume
that we need to construct a multibiometric system which utilizes two of the three traits. So at first,
we should analyze the relationship of traits and select ones uncorrelative or independent. Then we
construct and value the multibiometric system applying traits selected. In the first experiment,
correlative analysis, we apply the method proposed in Section 2 to get relationship between traits
after obtaining the match scores and select traits we need. Then, in the second experiment,
constructing a multibiometric system, we build a multibiometric system and value its
performance.
Several face recognition algorithms have been proposed in the literature [19]. We apply the
most mature one, principal component analysis (PCA). At first we need to convert the RGB facie
images to gray images. Then each of the H × W grey-level face images from every subject is
represented by a vector of HW × 1dimension using column ordering. The normalized set of such
training vectors is subjected to principal component analysis. The PCA generates a set of
orthonormal vectors, also known as eigenfaces, which can optimally represent the grey-level
information in the training dataset. The projection of subjects training face image on eigenfaces is
used to compute the characteristic features. The match score for every test face image is generated
by computing the similarity score between the feature vectors from the examples in testing dataset
and samples in training dataset.

Let y be the feature vector from an example in testing dataset and yi be the feature vector

from the i-th sample in training dataset. Then the similarity score is
In( y , yi )
 ( y, yi )  (14)
In( y , y ) In( yi , yi )

where In(.,.) is inner product. After that, we normalize the similarity to match score using

following equation:
1
s face ( y, yi )  ( ( y, yi )  1) (15)
2
So match score “1” means that one example and one sample is most similar while match score “0”
means the pair is the least similar.
Palmprint contains several complex features, e.g. minutiae, principal lines, wrinkles and
texture, which have been suggested for personal identification [20, 21, and 22]. The palmprint
matching approach used in this paper is same as detailed in [20]. After obtained the interest of
region of palm images, the palm features are extracted by 2-D Gabor filters. Then a palm image
can be represented using real and imaginary parts of the texture image. A Hamming distance is
adopted to determine the similarity measurement for palmprint matching. Let P and Q be the two
palmprint feature vectors. The Hamming distance can be described as
H W

 ( P (h, w)  Q (h, w)  P (h, w)  Q (h, w))


r r i i
(16)
D0 ( P, Q)  h 1 w 1

2  H W

where Pr (Qr ) and Pi (Qi ) are the real part and imaginary part respectively. The result of Boolean

operator  is equal to zero if and only if Pr (h, w)  Qr (h, w)( Pi (h, w)  Qi (h, w)) . H  W is

the size of the feature vectors. We also need to normalize the similarity to match scores for be
consistent with the face match score. For palm match, we adopt following technology:
s palm ( P, Q)  1  D0 ( P, Q) (17)

After normalized, match score “1” means that one example and one sample is most similar while
match score “0” means the pair is the least similar.

3.1 Correlative analysis and biometric traits selection Experiment

For obtaining the real relationship between facial and palm traits, we should construct
training and testing datasets at first. The method for constructing datasets is as follows. A serial of
random numbers from 1 to 10 is generated at first. Then we need to set the rate between testing
and training traits, choosing from 1:9 to 9:1. At last, the traits are decided as testing examples or
training samples according to the serial of numbers and the rate. So we can use the rate to identify
the datasets. For example, when the rate is equal to 3:7, after getting the serial of numbers and
ordering the original traits by the serial, there are 3 examples for testing datasets and 7 samples for
training datasets of each biometric trait, a facial image or left or right palm image.
In this work, the size of facial images is 350 × 380. In all the processes of the experiments,
the rate of principal component is 95%. The size of ROI for palm images is 128 × 128. The result
of correlative analysis is showed as Table 2.
Table 2. The results of relative analysis of face and palm traits
Dataset Face + Left palm Face + Right palm Left + Right palm
1:9 0.136564 0.175874 0.334065
2:8 0.182517 0.175684 0.358408
3:7 0.165532 0.185360 0.380081
4:6 0.176079 0.178896 0.359579
5:5 0.172698 0.186380 0.354181
6:4 0.158769 0.179943 0.341226
7:3 0.167272 0.182323 0.337660
8:2 0.164234 0.201124 0.349248
9:1 0.182366 0.171449 0.325921

From the Table 2, it is obvious that it is weakly correlative that the relationship between face
and palm traits and middle correlative that the one between left and right palm. And we can
conclude that it is more correlative that the relationship between face and right palm than that
between face and left palm. According to the method proposed in Section 3, if the threshold of

correlation coefficients r0 is set as 0.3, we can select face and left palm traits. Then we can

construct a multibiometric system using them and value it.

3.2 Multibiometric System Experiment

In this section, we give our attention to construction and value of multibiometric systems. For
there are threes traits, face, right palm and left palm, it is three choice to build a multibiometric
system which applies two of them. They are the ones applying face and left palm, face and right
palm and left and right palm. As the result in Section 3.1, it is known that the one applying face
and left palm should have the best performance.
In this work, we utilize the framework showed in Figure 2 for construction multibiometric
systems. In the fusion module, it is the sum rule that be applied. For being sure the fusion score be
in the range of [0, 1] and “0” meaning the least match score and “1” meaning the most score, we
modify the sample sum rule to the weighting sum rule,
sm   * s1  (1   ) * s2 (18)

where s1 and s2 are the two match scores from two different traits, and  is the weighting

coefficient.
Figure 2. The framework of the multibiometric system
In the experiment, we choose the dataset whose rate between testing and training traits is 8:2.
At first, we need to set the weighting coefficient. Figure 3.a shows the receiver operator
characteristic curves of biometric systems applying face, right palm and left palm. From the
figure, we can know that the performance of system which applies face is best. So the weighting
coefficient of facial traits in multibiometric systems should be more than other ones’. In the
following processes of experiment, the weighting coefficient of facial traits is set to 8/9.
The Figure 2.b shows the receive operator characteristic curves of multibiometric systems.
We con build three multibiometric systems which apply facial and left palm traits, facial and right
palm traits and right and left palm traits. From the Figure 2.b we can know the performance of
multibiometric system which apply facial and left palm traits is best, which is according to the
assumption based the correlative analysis and traits selection. What’s more, the performance of
multibiometric systems utilizing left and right palm traits is least, which is more correlative than
other pairs. The performance of multibiometric system using facial and left traits and that of one
using facial and right palm traits is similar with each other. It may be for that the relationship
between facial and left palm traits and one between facial and right palm traits are similar with
each other as list in Table 2. Although the gap of the correlative coefficient between facial and
right palm traits and that between facial and left palm traits on condition of the dataset whose
training and testing rate is 8:2 is huge, the grad of the average of the correlative coefficient
between facial and left palm traits and that between facial and left palm is small which is 0.0146
(0.0146 =0.1819 - 0.1673).
From Figure 3.c, we can know that the performance of multibiometric system fusing the
facial and left palm traits is better than biometric systems applying facial traits, left palm traits or
right palm traits, which means that our method is effective.
(a) (b)

(c)
Figure 3. Experiment results. (a) the receiver operator characteristic curves of biometric systems
using one trait, (b) the receiver operator characteristic curves of multibiometric systems using two
traits and (c) the receiver operator characteristic curves of a multibiometric system and biometric
systems

4. Conclude and future work

We have proposed the idea of biometric traits selection for multibiometrics. The results of
experiments show that the method proposed is rational and effective. The multibiometric system
constructed by method in this paper is in its best performance. In this paper, we assume that all the
match scores are correlative at first. Then we analyze the relationship between biometric traits and
the performance of multibiometric systems applying different combination of traits. At last we
value the multibiometric systems and get the conclusion. But we think there is still some place
where we can go on our research.
The first place which is needed to research is to set the threshold of correlation coefficients.
In this paper, after obtaining all the correlation coefficients between pairs of traits, we select one
of them and think it is the threshold. In other word, we need to know all the correlation
coefficients before setting the threshold. There may be some relationship between the threshold
and the requirement of constructing multibiometric systems. Maybe we can know the threshold at
knowing the expectation of the performance of multibiometric systems we need to build before
calculating all the correlation coefficients.
In this paper we only consider the relationship between one trait and the other one. But in
practice a match score may be a linearity equation of other match scores, such as
k
s  a0  a1s1  ...  ak sk  a0   ai sk . In this case, the traits selection may be more complex.
i 1

If we still apply the method proposed in this paper to select traits, then we would not obtain the
real relationship of traits and performance of the multibiometric system utilizing traits selected
may be not better than others.
Acknowledgment

We thank ones that help us and give us some useful advices. They are Ajay Kumar,
Robert Snelick, Giorgio Fumera, Gerard Chollet, Zhenan Sun, Arun Ross, Stephanie Schuckers, Thomas Huang
and Dennis J Lin. We also thank the anonymous reviewers for their constructive advice.

Reference

[1] Biometric: Personal Identification in Networked Society. A. Jain, R. Bolle, and S. Pankanti, eds.
Boston: Kluwer Academic, 1999.
[2] D. Zhang, Automated Biometrics-Technologies and Systems. Boston: Kluwer Academic, 2000.
[3] Anil K.Jain. An Introduction to Biometric Recognition. IEEE Transactions on Circuits and System
for Video Technology, 14: 4~2. January 2004.
[4] Anil K.Jain. Biometric Recognition: How Do I Know Who You Are ? Signal Processing and
Communications Applications Conference, 2004. Proceedings of the IEEE 12th, 3~5, 2004.
[5] Anil K.Jain. Biometrics: A Grand Challenge. Proceedings of International Conference on Pattern
Recognition, Cambridge, U.K, 2:935~942. Aug.2004.
[6] A. Ross and Anil. K. Jain. Multimodal Biometrics: An Overview. Proc. Of 12 th European Signal
Processing Conference, pp. 1221-1224, September 2004.
[7] A. Ross, K. Nandakumar and A. K. Jain. Handbook of Multibiometrics. Springer Publishers, 1st
edition, 2006.
[8] Robert Snelick, Mike Indovina, James Yen and Alan Mink. Multimodal Biometrics: Issues in
Design and Testing. ICMI’03, Vancouver, British Columbia, Canada, November 5-7, 2003.
[9] M. Indovina. Multimodal Biometric Authentication Methods: A cost Approach. Workshop on
Multimodal User Authentication (MMUA 2003), 99–106, Santa Barbara, 2003.
[10] R. Brunelli and D. Falavigna. Person Identification Using Multiple Cues. IEEE Transactions on
PAMI. 12: 955~966. 1995.
[11] Souheil Ben-Yacoub, Yousri Abdeljaoued and Eddy Mayorza. Fusion of face and speech data for
person identity verification. IEEE Transactions on Neural Networks, 10(5):1065~1074. 1999.
[12] . Hong and A. K. Jain. Integrating faces and fingerprints for personal identification. IEEE
Transactions on PAMI, vol. 20, pp. 1295-1307, December 1998.
[13] Sonia Garcia-Salicetti. Biomet:A Multimodal Person Authentication Database Including Face,
Voice, Fingerprint, Hand and Signature Modalities. Springer LNCS-2688, Fourth International
Conference on Audio- and Video-Based Biometric Person Authentication (AVBPA’03) 845~853.
Guildford, 2003.
[14] A. Kumar, D. C. M. Wong, H. C. Shen, and Anil K. Jain .Personal Verification Using Palmprint
and Hand Geometry Biometric. Proc. of 4th Int’l Conf. on Audio and Video~based Biometric
Person Authentication (AVBPA), 668~678. Guildford, UK .2003.
[15] J. Kittler, M. Hatef, R. P. Duin, and J. G. Matas. On Combining Classifiers. IEEE Transactions on
PAMI, 20: 226~239. 1998.
[16] Kuncheva, L. I., Whitaker, C. J., Shipp, C.A., and Duin, R. P. W. Is Independence Good for
Combining Classifiers? In Proceedings of International Conference on Pattern Recognition, Vol.
2, pp. 168-171, Barcelona, Spain, 2000.
[17] Prabhaker, S. and Jain, A. K. Decision-level Fusion in Fingerprint Verification. Pattern
Recognition, 35(4), pp. 861-874, 2002.
[18] Poh, N. and Bengio, S. How Do Correlation and Variance of Base-Experts Affect Fusion in
Biometric Authentication Tasks? IEEE Transaction On Signal Processing, 53(11), pp. 4383-4396,
2005.
[19] M.-H. Yang, D. J. Kriegman, and N. Ahuja. Detecting faces in images: a survey. IEEE Transaction
Pattern Anal. Machine Intell., vol. 24, pp. 34-58, Jan. 2002.
[20] David Zhang, W. Kong, J. You, M. Wong. Online Palmprint Identification. Transactions on PAMI,
25(9):1041~1050. 2003.
[21] Guangming Lu, David Zhang, Kuanquan Wang. Palmprint Recognition Using Eigenpalms
Features. Pattern Recognition Letters, 24: 1463~1467. 2003.
[22] Xiangqian Wu, David Zhang, Kuanquan Wang. Fisherpalms Based Palmprint Recognition.
Pattern Recognition Letters, 24:2829~2838. 2003.

You might also like