An Overview on Multi-biometric Score-level Fusion

Veriﬁcation and Identiﬁcation

Naser Damer, Alexander Opel and Andreas Shahverdyan

Competence Center Identiﬁcation and Biometrics,

Fraunhofer Institute for Computer Graphics Research IGD, Darmstadt, Germany

Keywords:

Multi-biometrics, Score-level Fusion, Biometric Identiﬁcation.

Abstract:

Multi-biometrics is the use of multiple biometric recognition sources to provide a more dependable veriﬁcation

or identiﬁcation decision. Fusion of multi-biometric sources can be performed on different levels, such as

the data, feature, or score level. This work presents an overview of the multi-biometric score-level fusion

problem, along with the proposed solution in the literature. A discussion is made to provide a comparison

between multi-biometric fusion in both scenarios. This discussion aims at providing a clearer view of future

developments especially under the identiﬁcation scenario where many related applications are rapidly growing

such as forensics and ubiquitous surveillance.

1 INTRODUCTION

Biometrics is a rapidly growing technology that aims

to identify or verify people identities based on their

physical or behavioral properties. Multi-biometrics

use more than one biometric recognition approach in

a uniﬁed frame in an effort to solve problems faced

by the conventional uni-modal biometrics. The multi-

biometric approach aims at improving biometrics by

increasing accuracy, and robustness to intra-person

variations and to noisy data. It also aims to solve uni-

modal biometrics problems with non-universality and

vulnerability to spoof attacks.

Information fusion in multi-biometrics is used to

build an identiﬁcation/veriﬁcation decision based on

the information collected from different biometric

sources. The fusion can be done on different lev-

els such as data-level, feature-level, score-level, rank-

level or decision-level. In this work, score-level fu-

sion will be inspected as it is widely used to integrate

different modalities (based on different biometrics, al-

gorithms and manufacturers) through fusion. Score

here refers to the comparison score (similarity) be-

tween each captured biometric property and a stored

reference.

Biometrics recognition technologies are usually

developed under one of two scenarios, veriﬁcation or

identiﬁcation. Biometric veriﬁcation is the use of bio-

metrics information to verify a persons claimed iden-

tity. Identiﬁcation, on the other hand, can be deﬁned

as the process of assigning a previously registered

identity to a person based on the captured biometrics

information of the person.

The different nature between veriﬁcation and

identiﬁcation scenarios effects the implementation of

multi-modal biometrics solutions, especially the fu-

sion process. This is due to the different available

information in both scenarios, as well as, the different

nature of the expected fusion decision.

Figure 1 presents an overview of multi-biometric

score-level fusion. Scores from different sources (al-

gorithms and modalities) are normalized then passed

into a fusion algorithm. The fusion then results in a

fused score.

Figure 1: General structure of multi-biometric score-level

fusion (Identiﬁcation). Comparison results are produced by

different algorithms and modalities then processed by the

normalization stage. The sets of normalized scores feed the

fusion algorithm to produce a ﬁnal set ranked by fused com-

parison scores.

647

Damer N., Opel A. and Shahverdyan A..

An Overview on Multi-biometric Score-level Fusion - Veriﬁcation and Identiﬁcation.

DOI: 10.5220/0004358306470653

In Proceedings of the 2nd International Conference on Pattern Recognition Applications and Methods (BTSA-2013), pages 647-653

ISBN: 978-989-8565-41-9

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

In Section 2 below, a literature review is presented

covering some of the most interesting works dealing

with multi-biometric fusion under both the veriﬁca-

tion and identiﬁcation scenarios. The literature re-

view is presented along with the explanation of the

different steps of the multi-biometric score-level fu-

sion process. In Section 3, the fusion process under

the two scenarios is discussed. A comparison between

the natures of the fusion process under both scenarios

is also presented. Finally, in Section 4, conclusions

and recommendations are drawn.

2 LITERATURE SURVEY

In the following, topics relating to different stages of

the multi-biometric fusion process are discussed and

connected to cutting edge literature. The topics cov-

ered are multi-biometrics schemes, score normaliza-

tion, fusion algorithms, identiﬁcation and veriﬁcation

scenarios, consideration of biometric sample quality,

available datasets, and ﬁnally the robustness to possi-

bly missed data.

2.1 Multi-biometrics

Multi-biometrics is categorized into several ap-

proaches depending on the source of multi-decisions

from which a uniﬁed ﬁnal decision is built on.

The main approaches are multi-modalities, multi-

algorithmic, multi-instance, multi-sensorial, and

multi-presentation.

Multi-modalities is the use of more than one

biometric characteristic as an identity measure.

Some works combined ﬁngerprints and face images

(K. Nandakumar and Ross, 2009; Tong et al., 2010;

Kim et al., 2010), others fused ﬁngerprints and iris

biometrics (Baig et al., 2009). Using face images

along with iris biometrics was also introduced (Wang

et al., 2003). One of the most interesting multi-modal

approaches is the use of ear and face biometrics, as

they can be easily and non-intrusively captured us-

ing same or similar devices (Chang et al., 2003; Yan,

2006). Many other combinations were also intro-

duced, such as ear and ﬁngerprint biometrics (Rattani

et al., 2006).

Many works dealt with multi-algorithmic biomet-

rics, such as using multiple face matchers (K. Nan-

dakumar and Ross, 2009; Basak et al., 2010; Kim

et al., 2010), or ear identiﬁers (Moreno et al., 1999;

Yan and Bowyer, 2005). Multi-instance fusion

was also studied, such as two different ﬁngerprints

(Yan and Bowyer, 2005; Basak et al., 2010; Kim

et al., 2010), multi-sensorial (Arandjelovic and Ham-

moud, 2006), and multi-presentation biometric fusion

(Cheng et al., 2011) were also discussed thoroughly

in the literature.

2.2 Score Normalization

The scores processed by the fusion algorithm are usu-

ally not homogeneous as they are produced by differ-

ent sources. Those scores have to be brought into a

common comparable range by a normalization pro-

cess. Some of the most common normalization tech-

niques are min-max normalization, z-score normal-

ization, double sigmoid, tanh-estimator, and median

absolute normalization. The parameters that rule the

normalization process are determined based on the

statistical properties of the training data. The per-

formances of normalization techniques are not di-

rectly comparable as they depend on the overall multi-

biometric system.

Here, presented in more details are four of those

normalization techniques. The min-max normal-

ization, the z-score, the median absolute deviation

(MAD), and the double sigmoid function normaliza-

tion.

Given comparison scores set S

, k = 1,2,..., N

the normalized score is a function of the score f (S).

The min-max normalization depends on the range

that the scores span regardless of the distribution

properties and aims to map the scores into a range

of [0,1]. This normalization scheme was used

successfully by many works dealing with score-level

multi-biometric fusion (Nisha Srinivas, 2009; Vajaria

et al., 2007). The normalized score by min-max

normalization is given by:

f (S) =

S−min{S

}

max{S

}−min{S

}

The min-max normalization scheme is highly sen-

sitive to outliers (Jain et al., 2005) as it depends on

single minimum and maximum values.

Z-score normalization is a more sophisticated

score normalization method and was used in several

works (Nisha Srinivas, 2009; Vajaria et al., 2007).

Here, the arithmetic mean (µ) and the standard

deviation (σ) of the score values are considered. The

normalized score by z-score normalization is given

by:

f (S) =

S−µ

This method assumes a Gaussian distribution of

the score values. This normalization method has low

robustness as well, as the parameters µ and σ are sen-

sitive to outliers.

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

648

The median absolute deviation normalization

method is similar to the z-score normalization

method but uses the median and median absolute

deviation instead of the mean and standard deviation.

This method is more robust to outliers but it also

assumes a near Gaussian distribution of the compar-

ison score values. The median absolute deviation

normalization is given by:

f (S) =

(S−median)

MAD

MAD = median(|{S

} − median)

The double sigmoid function is another normal-

ization method. This normalization maps the scores

into a range of [0,1] and requires ﬁne tuning of its

parameters. The normalization is given by:

f (S) =

(

1+e

(−2((S−t)/r

))

i f S < t

1+e

(−2((S−t)/r

))

otherwise.

The double sigmoid normalization was used to

combine ﬁnger print comparison scores by Cappelli

et al. (Cappelli et al., 2000).

Marsico et al. recently proposed a novel normal-

ization approach named the Quasi-Linear Sigmoid

(QLS) (De Marsico et al., 2011). This approach aims

at overcoming the limitations of the traditional nor-

malization algorithms.

The selection of a proper normalization method

is a tradeoff between efﬁciency and robustness, and

it depends largely on the nature of the application.

Methods like min-max normalization and z-score nor-

malization tend to be more efﬁcient. On the other

hand, median absolute deviation normalization and

double sigmoid function normalization are usually

more robust but require higher computational effort

(Jain et al., 2005).

2.3 Fusion Algorithms

Fusion algorithms can be categorized into two main

types, combination rules and classiﬁcation based fu-

sion.

Combination rules are simple operations per-

formed on the normalized scores. Those rules

produce a combined score and a classiﬁcation de-

cision is made based on this combined score value.

Main combination rules are the sum rule, weighted

sum rule, product rule, max rule, min rule and median

rule. Assuming the combined score is a function of

the N biometric scores S inputs and is given by C(S).

The combination rules can be formulated as follows:

The sum rule: C(S) =

∑

i=1

The weighted sum rule: C(S) =

∑

i=1

, where

is the i

weight factor.

The product rule: C(S) =

∏

i=1

The max rule: C(S) = max

The min rule: C(S) = min

The median rule: C(S) = median

Some works discussed the difference in perfor-

mance between the combination rules. Most studies

showed the superiority of sum and product combi-

nation rules (Hariri and Shokouhi, 2012; Nandaku-

mar et al., 2006; Chang et al., 2004). One must

keep in mind that different combination of sum rules

and normalization techniques produce different re-

sults (L. Latha, 2011).

Classiﬁcation based fusion considers the input

score values as a feature vector. Given those vectors, a

classiﬁer is trained to classify a new given vector into

genuine or imposter class. Different types of classi-

ﬁers can be used, just as neural networks (Alsaade,

2010), K-NN (Jin et al., 2004), SVM (Singh et al.,

2007; Garcia-salicetti et al., 2005), Adaboost (Ichino

et al., 2010; Moin and Parviz, 2009), or as likelihood

ratio (Nandakumar et al., 2008; Islam and Rahman,

2010) classiﬁers. Some works showed comparable

results between combination rules and classiﬁcation

based fusion (Rodr

ıguez et al., 2008; Mehrotra et al.,

2012). Other works showed the superiority of combi-

nation rules (Singh et al., 2007).

2.4 Identiﬁcation and Veriﬁcation

Veriﬁcation is to conﬁrm or reject a claimed identity

of a person based on his/her biometric characteris-

tics. Identiﬁcation is to assign a pre-registered iden-

tity to an unknown (no identity claim) person based

on his/her biometric characteristics.

The identiﬁcation process itself can be divided

into two different operations, the open-set identiﬁca-

tion and the closed-set identiﬁcation. The closed-set

identiﬁcation refers to the situation where the uniden-

tiﬁed individual is known to be enrolled in the bio-

metrics reference database. Here, the identiﬁcation

process assigns an ID to the unknown captured indi-

vidual.

The open-set identiﬁcation (watchlist) refers to the

scenario where the unidentiﬁed individual is not def-

initely enrolled in the biometrics reference database.

This case requires verifying the existence of the in-

dividual record in the reference database, as well as

identifying the individual.

Most of the works dealing with multi-biometrics

fusion consider the case of veriﬁcation (Rodr

ıguez

AnOverviewonMulti-biometricScore-levelFusion-VerificationandIdentification

649

et al., 2008; Jin et al., 2004; Poh and Kittler, 2008;

Nisha Srinivas, 2009). Nonetheless, some of the re-

cent works dealt explicitly with the fusion problem

under the identiﬁcation scenario (Basak et al., 2010;

K. Nandakumar and Ross, 2009).

A detailed discussion about the differences and in-

teractions between the veriﬁcation and identiﬁcation

scenarios is presented in Section 3.

2.5 Biometric Sample Quality and

Missing Data

In order to compensate for missing information when

moving from feature-level fusion to score-level fu-

sion and therefore improve accuracy, the quality of

the biometric samples is considered. Beside accuracy,

robustness of the biometric system is important. In

practice, some of the comparison scores can be miss-

ing because of a missing modality or a low quality

capture. To build a robust multi-biometric system, the

possibility of missing score values must be considered

and dealt with so that a reliable biometric decision is

made.

2.5.1 Biometric Sample Quality

The quality of the captured biometric sample (image

or scan) has an effect on the comparison score values

and the conﬁdence of those values as the features ex-

tracted from those samples are not reliable (Nandaku-

mar et al., 2006). This reﬂects on the role a certain

score value plays in the fusion process and therefore,

taking the quality measures into account will enhance

the performance of the fused multi-biometric system

(Nandakumar et al., 2006; Poh and Kittler, 2008; Poh

et al., 2009).

Nandakumar et al. (Nandakumar et al., 2006) pro-

posed a fusion algorithm that takes into account the

sample quality into their likelihood ratio-based fusion

scheme. Their experiments proved a highly positive

impact of considering the quality measures.

Poh et al. (Poh et al., 2009) proposed a classiﬁer

based fusion algorithm that considers the biometric

sample quality, as well as, the biometric capture de-

vice information. Their experiments clearly showed

the effect of including the quality measures on perfor-

mance.

2.5.2 Missing Data

Multi-biometric systems make a decision based on a

set of scores. A case where one or more of those score

values are missing may occur, especially in large scale

identiﬁcation systems. Missing data can occur be-

cause of the non-universality of a certain biometric

modality, or a poorly captured modality in uncon-

trolled and ubiquitous biometric systems.

Many works considered the problem of missing

data and proposed solutions for robust fusion algo-

rithms (Poh et al., 2010b; K. Nandakumar and Ross,

2009; Dinerstein et al., 2007). However, most of those

works dealt with the fusion problem under the veriﬁ-

cation scenario.

Nandakumar et al. proposed a robust fusion so-

lution for multi-biometric fusion under the identiﬁ-

cation scenario that aims to produce an identiﬁca-

tion decision regardless of the partially missing data

(K. Nandakumar and Ross, 2009). The authors ex-

tended the likelihood ratio-based score fusion (origi-

nally designed for veriﬁcation problems) to perform

under the identiﬁcation scenario.

2.6 Databases

In order to build and evaluate an optimized multi-

biometric fusion system, a diverse and informative

dataset must be available. In this case, a database

that contains matching scores from multi-biometrics

resources for imposters and genuine matches is re-

quired. In the following, four publicly available

databases are considered.

The XM2VTS Score-level Fusion Benchmark

Dataset (Poh and Bengio, 2006) that contains a

database of scores taken from experiments carried

out on the XM2VTS face and speaker veriﬁcation

database. Another database is the BANCA score

database, a free database that contains 1186 baseline

face and speech experiments (Poh, ).

Another multi-biometric database was released as

a part of the Multiple Biometrics Grand Challenge

(Phillips et al., 2009). This database includes face and

iris biometric information and aims at the evaluation

and improvement of face and iris biometrics captured

under uncontrolled conditions.

The Biosecure DS2 database (Poh et al., 2010a)

contains biometric comparison scores, as well as the

quality measures of the captured biometric samples.

The Biosecure DS2 database is considered to be the

one of the ﬁrst databases to build a quality-based fu-

sion benchmark.

3 VERIFICATION VS.

IDENTIFICATION

In this section, a comparison between the fusion pro-

cesses of multi-biometrics under both scenarios is de-

livered. This comparison will focus on the available

information for the fusion algorithm under different

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

650

scenarios. The expected fusion decision is also dis-

cussed. The importance of dealing with image qual-

ity and ﬂexibility to missing data is also discussed in

this section along with the evaluation measure for the

identiﬁcation and veriﬁcation problems.

3.1 Available Information

The veriﬁcation process starts by supplying the bio-

metric system with a claimed identity and one or more

captured biometrics characteristics. The claimed

identity has one or more corresponding biometrics

references stored in a dataset. The biometric veriﬁca-

tion system matches the captured biometrics with the

corresponding stored biometric reference. In a sum,

the available information in the veriﬁcation scenario

is the claimed identity, captured biometrics of the sub-

ject and the corresponding biometrics references of

the claimed subject. Those information leads to only

one set (belongs to one identity) of biometrics com-

parison scores to be fed into the fusion algorithm.

Identiﬁcation aims to compare the subject’s cap-

tured biometric characteristics with the available ref-

erences in the dataset to build an identiﬁcation deci-

sion. Therefore, the available information are the cap-

tured biometrics of the subject and the references of

all the enrolled subjects. This leads to a number of

comparison scores sets. This number can be equal to

the number of enrolled subjects or a subset of them

(if initial guess was made). Those sets of comparison

scores are then fed into the fusion algorithm to build

the identiﬁcation decision.

While building the fusion algorithm, the informa-

tion available for both identiﬁcation and veriﬁcation

scenarios are similar. As the identities are known

while training, one can get access to the comparison

scores between all captured biometrics and all refer-

ences. Therefore, a set of genuine/imposter compari-

son scores is available to help optimize the fusion al-

gorithm in both scenarios. However, the availability

of multiple comparisons in the identiﬁcation scenario

can provide more information about the genuinity of a

certain capture. It is expected that a larger difference

between the ﬁrst and second ranked matches indicates

a higher conﬁdence that the ﬁrst rank is the genuine

match. This kind of information is only available un-

der the identiﬁcation scenario.

3.2 Expected Decision

The expected output of the fusion algorithm varies be-

tween veriﬁcation and identiﬁcation scenarios. Given

the set of comparison scores between the captured

biometrics and the references of the claimed iden-

tity, the veriﬁcation process output is a binary deci-

sion. This decision marks the claimed identity either

as genuine (true) or imposter (false).

In identiﬁcation, given the sets of biometrics com-

parison scores between the captured subject and all

stored references (sometimes partial set), the fusion

decision must rank the stored identities by similarity

to the captured subject.

It must be mentioned that open-set identiﬁcation

operation must be followed by a veriﬁcation process

for the top ranked identity. This veriﬁcation ensures

that the top ranked identity is the correct identity. Es-

pecially when the database may not include the iden-

tity of interest. If the veriﬁcation of the top ranked

identity resulted in an imposter decision, the captured

person is believed not to be registered in the reference

database.

3.3 Quality Measures, Missed Data and

Evaluation

The accuracy of the fusion decision largely depends

on the quality of the comparison processes of the dif-

ferent modalities. This quality of the comparison in

each modality depends on different factors. It is af-

fected by the quality of the captured biometric infor-

mation, the quality of the stored biometric reference

and the quality of the capturing device itself. It is also

dependent on the quality of the comparison algorithm

and the features used to represent the biometrics. The

accuracy of the fusion algorithm is also affected by

the quality of the preprocessing of comparison scores

i.e. normalization.

The biometric decision in veriﬁcation is based

only on one set of captured biometrics and one set

(ID) of references (1:1 match). However, under iden-

tiﬁcation, the decision is based on more than one set

of biometric references (1:N match). This fact leads

to the believe that the identiﬁcation scenario is more

affected by the quality measures, especially in the

cases where the quality of biometrics references vary

largely.

One of the main reasons to use multi-modal bio-

metrics is the pursuit of higher robustness in biomet-

ric systems. This appears usually when considering

the universality of biometric systems, as well as de-

signing ubiquitous biometric systems. Multi-modal

biometric systems must be designed to be functional

even when some information are missed, this directly

affect the fusion algorithm design and have different

effects under the veriﬁcation and identiﬁcation sce-

narios.

A missing biometric measure in a veriﬁcation sce-

nario can occur because of a missed captured bio-

AnOverviewonMulti-biometricScore-levelFusion-VerificationandIdentification

651

metric characteristic or a missed reference within

the claimed identity references set. In identiﬁcation,

a missed captured biometric measure, as well as a

missed reference can occur. However, under identi-

ﬁcation where the system depends on a large number

of references sets, the missed modality can be differ-

ent in each comparison between each ID pair. This

situation argues the development of more advanced

and ﬂexible solutions for missing data under the iden-

tiﬁcation scenario.

The performance evaluation of a multi-biometric

systems is not different than that of a conventional

uni-modal biometric systems. Identiﬁcation results

are usually represented in a Cumulative Match Char-

acteristic (CMC) curve, especially when dealing with

closed-set identiﬁcation. The veriﬁcation perfor-

mance is shown usually as Receiver Operating Char-

acteristic (ROC) curve and as an equal error rate

(EER).

4 CONCLUSIONS

This work presented an overview on the score-level

multi-biometric fusion problem. Some of the most

interesting works in this ﬁeld were discussed along

with the structured steps of the multi-biometric fusion

process. A comparison between the multi-biometric

systems under identiﬁcation and veriﬁcation scenar-

ios was drawn. The discussion presented aims at pro-

viding a clear view on multi-biometric system devel-

opment especially under the relatively understudied

identiﬁcation scenario.

ACKNOWLEDGEMENTS

This work is funded by the Federal Ministry of Edu-

cation and Research (BMBF) of Germany in the con-

text of the research programme for public safety and

security.

REFERENCES

Alsaade, F. (2010). A study of neural network and its prop-

erties of training and adaptability in enhancing accu-

racy in a multimodal biometrics scenario. Information

Technology Journal.

Arandjelovic, O. and Hammoud, R. (2006). Multi-sensory

face biometric fusion (for personal identiﬁcation). In

Proceedings of the 2006 Conference on Computer Vi-

sion and Pattern Recognition Workshop, CVPRW ’06,

pages 128–, Washington, DC, USA. IEEE Computer

Society.

Baig, A., Bouridane, A., Kurugollu, F., and Qu, G. (2009).

Fingerprint - iris fusion based identiﬁcation system

using a single hamming distance matcher. In Proceed-

ings of the 2009 Symposium on Bio-inspired Learning

and Intelligent Systems for Security, BLISS ’09, pages

9–12, Washington, DC, USA. IEEE Computer Soci-

ety.

Basak, J., Kate, K., Tyagi, V., and Ratha, N. (2010). A

gradient descent approach for multi-modal biometric

identiﬁcation. In Pattern Recognition (ICPR), 2010

20th International Conference on, pages 1322 –1325.

Cappelli, R., Maio, D., and Maltoni, D. (2000). Com-

bining ﬁngerprint classiﬁers. In Proceedings of the

First International Workshop on Multiple Classiﬁer

Systems, MCS ’00, pages 351–361, London, UK, UK.

Springer-Verlag.

Chang, K., Bowyer, K. W., Sarkar, S., and Victor, B. (2003).

Comparison and combination of ear and face images

in appearance-based biometrics. IEEE Trans. Pattern

Analysis and Machine Intelligence, 25:1160–1165.

Chang, K. I., Bowyer, K. W., Flynn, P. J., and Chen, X.

(2004). Multi-biometrics using facial appearance,

shape and temperature. In Proceedings of the Sixth

IEEE international conference on Automatic face and

gesture recognition, FGR’ 04, pages 43–48, Washing-

ton, DC, USA. IEEE Computer Society.

Cheng, X., Tulyakov, S., and Govindaraju, V. (2011).

Multiple-sample fusion of matching scores in biomet-

ric systems. In Computer Vision and Pattern Recogni-

tion Workshops (CVPRW), 2011 IEEE Computer So-

ciety Conference on, pages 120 –125.

De Marsico, M., Nappi, M., Riccio, D., and Tortora,

G. (2011). Nabs: Novel approaches for biomet-

ric systems. Systems, Man, and Cybernetics, Part

C: Applications and Reviews, IEEE Transactions on,

41(4):481 –493.

Dinerstein, S., Dinerstein, J., and Ventura, D. (2007). Ro-

bust multi-modal biometric fusion via multiple svms.

In Proceedings of the IEEE International Conference

on Systems, Man and Cybernetics, Montral, Canada,

7-10 October 2007, pages 1530–1535. IEEE.

Garcia-salicetti, S., Mellakh, M. A., Allano, L., and

Dorizzi, B. (2005). Multimodal biometric score fu-

sion: The mean rule vs. support vector classiﬁers. In

Proceedings of the EUSIPCO.

Hariri, M. and Shokouhi, S. B. (2012). Robustness of

multi biometric authentication systems against spoof-

ing. Computer and Information Science, pages 77–86.

Ichino, M., Komatsu, N., Wang, J.-G., and You, W. Y.

(2010). Speaker gender recognition using score level

fusion by adaboost. In 11th International Confer-

ence on Control, Automation, Robotics and Vision,

ICARCV 2010, Singapore, 7-10 December 2010, Pro-

ceedings, pages 648–653. IEEE.

Islam, M. R. and Rahman, M. F. (2010). Article:likelihood

ratio based score fusion for audio-visual speaker iden-

tiﬁcation in challenging environment. International

Journal of Computer Applications, 6(7):6–11. Pub-

lished By Foundation of Computer Science.

Jain, A., Nandakumar, K., and Ross, A. (2005). Score nor-

malization in multimodal biometric systems. Pattern

Recogn., 38(12):2270–2285.

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

652

Jin, A. T. B., Samad, S. A., and Hussain, A. (2004). Near-

est neighbourhood classiﬁers in a bimodal biometric

veriﬁcation system fusion decision scheme. Journal

of Research and Practice in Information Technology,

36(1):47–62.

K. Nandakumar, A. K. J. and Ross, A. (2009). Fusion in

multibiometric identiﬁcation systems: What about the

missing data? In Proceedings of the 3rd International

Conference on Biometrics, Alghero, Italy.

Kim, Y., Toh, K.-A., and Teoh, A. (2010). An online learn-

ing algorithm for biometric scores fusion. In Biomet-

rics: Theory Applications and Systems (BTAS), 2010

Fourth IEEE International Conference on, pages 1 –6.

L. Latha, S. T. (2011). On improving the performance

of multimodal biometric authentication through ant

colony optimization. WSEAS Transactions on Infor-

mation Science and Applications.

Mehrotra, H., Vatsa, M., Singh, R., and Majhi, B. (2012).

Biometric match score fusion using rvm: A case study

in multi-unit iris recognition. In Computer Vision and

Pattern Recognition Workshops (CVPRW), 2012 IEEE

Computer Society Conference on, pages 65 –70.

Moin, M. S. and Parviz, M. (2009). Exploring auc boosting

approach in multimodal biometrics score level fusion.

In Proceedings of the 2009 Fifth International Confer-

ence on Intelligent Information Hiding and Multime-

dia Signal Processing, IIH-MSP ’09, pages 616–619,

Washington, DC, USA. IEEE Computer Society.

Moreno, B., Sanchez, A., and Velez, J. (1999). On the use of

outer ear images for personal identiﬁcation in security

applications. In Security Technology, 1999. Proceed-

ings. IEEE 33rd Annual 1999 International Carnahan

Conference on, pages 469 –476.

Nandakumar, K., Chen, Y., Dass, S. C., and Jain, A. (2008).

Likelihood ratio-based biometric score fusion. IEEE

Trans. Pattern Anal. Mach. Intell., 30(2):342–347.

Nandakumar, K., Chen, Y., Jain, A. K., and Dass, S. C.

(2006). Quality-based score level fusion in multibio-

metric systems. In Proceedings of the 18th Interna-

tional Conference on Pattern Recognition - Volume

04, ICPR ’06, pages 473–476, Washington, DC, USA.

IEEE Computer Society.

Nisha Srinivas, Kalyan Veeramachaneni, L. O. (2009). Fus-

ing correlated data from multiple classiﬁers for im-

proved biometric veriﬁcation. In 12th International

Conference on Information Fusion.

Phillips, P., Flynn, P., Beveridge, J., Scruggs, W., OToole,

A., Bolme, D., Bowyer, K., Draper, B., Givens, G.,

Lui, Y., Sahibzada, H., Scallan, JosephA., I., and

Weimer, S. (2009). Overview of the multiple bio-

metrics grand challenge. In Tistarelli, M. and Nixon,

M., editors, Advances in Biometrics, volume 5558 of

Lecture Notes in Computer Science, pages 705–714.

Springer Berlin Heidelberg.

Poh, N. Banca score database.

Poh, N. and Bengio, S. (2006). Database, protocols and

tools for evaluating score-level fusion algorithms in

biometric authentication. Pattern Recogn., 39(2):223–

233.

Poh, N., Bourlai, T., and Kittler, J. (2010a). A multi-

modal biometric test bed for quality-dependent, cost-

sensitive and client-speciﬁc score-level fusion algo-

rithms. Pattern Recogn., 43(3):1094–1105.

Poh, N. and Kittler, J. (2008). A family of methods for

quality-based multimodal biometric fusion using gen-

erative classiﬁers. In ICARCV, pages 1162–1167.

IEEE.

Poh, N., Merati, A., and Kittler, J. (2009). Making bet-

ter biometric decisions with quality and cohort infor-

mation: A case study in ﬁngerprint veriﬁcation. In

Proc. 17th European Signal Processing Conf. (Eu-

sipco), pages 70–74, Glasgow.

Poh, N., Windridge, D., Mottl, V., Tatarchuk, A., and

Eliseyev, A. (2010b). Addressing missing values in

kernel-based multimodal biometric fusion using neu-

tral point substitution. Trans. Info. For. Sec., 5(3):461–

469.

Rattani, A., Kisku, D. R., Bicego, M., and Tistarelli, M.

(2006). Robust Feature-Level Multibiometric Classi-

ﬁcation. Biometric Consortium Conference, 2006 Bio-

metrics Symposium: Special Session on Research at

the, pages 1–6.

Rodr

ıguez, L. P., Crespo, A. G., Lara, M. J. P., and Mezcua,

B. R. (2008). Study of different fusion techniques

for multimodal biometric authentication. In Proceed-

ings of the 2008 IEEE International Conference on

Wireless & Mobile Computing, Networking & Com-

munication, WIMOB ’08, pages 666–671, Washing-

ton, DC, USA. IEEE Computer Society.

Singh, R., Vatsa, M., and Noore, A. (2007). Intelligent

biometric information fusion using support vector ma-

chine. In Nachtegael, M., Van der Weken, D., Kerre,

E., and Philips, W., editors, Soft Computing in Image

Processing, volume 210 of Studies in Fuzziness and

Soft Computing, pages 325–349. Springer Berlin Hei-

delberg.

Tong, Y., Wheeler, F., and Liu, X. (2010). Improving bio-

metric identiﬁcation through quality-based face and

ﬁngerprint biometric fusion. In Computer Vision and

Pattern Recognition Workshops (CVPRW), 2010 IEEE

Computer Society Conference on, pages 53 –60.

Vajaria, H., Islam, T., Mohanty, P., Sarkar, S., Sankar, R.,

and Kasturi, R. (2007). Evaluation and analysis of a

face and voice outdoor multi-biometric system. Pat-

tern Recogn. Lett., 28(12):1572–1580.

Wang, Y., Tan, T., and Jain, A. (2003). Combining face and

iris biometrics for identity veriﬁcation. In Kittler, J.

and Nixon, M., editors, Audio- and Video-Based Bio-

metric Person Authentication, volume 2688 of Lecture

Notes in Computer Science, pages 805–813. Springer

Berlin Heidelberg.

Yan, P. (2006). Ear biometrics in human identiﬁcation. PhD

thesis, University of Notre Dame, Notre Dame, IN,

USA. AAI3406950.

Yan, P. and Bowyer, K. W. (2005). Multi-biometrics 2d and

3d ear recognition. In Kanade, T., Jain, A. K., and

Ratha, N. K., editors, AVBPA, volume 3546 of Lecture

Notes in Computer Science, pages 503–512. Springer.

AnOverviewonMulti-biometricScore-levelFusion-VerificationandIdentification

653