Li S.Z., Jain A.K. (eds.) Encyclopedia of Biometrics

Подождите немного. Документ загружается.

the performa nce of using this Bayes supervisor in

a recognition system relying on a ﬁngerprint and a

signature expert is shown. The quality scores are gen-

erated by human experts independent of the experi-

ment. To automatically ﬁnd qualit y scores p

for

biometric samples is an emerging ﬁeld of study

[8–11]. The results of Bayes supervisor in connection

with machine-delivered quality scores are illustrated by

Fig. 3 where three ﬁngerprint (machine) experts’ opi-

nions are weighted to yield the supervisor opinion. The

experts are called A, B, and C and the quality measures

used were (1) no quality used, i.e. p

¼ 1 (2) An

automatic quality measure [10], (the method is pub-

licly made available by NIST), (3) another automatic

quality measure based on local symmetries [11], (4)

Quality measures provided by human experts. At each

experiment, one of the four quality measures is at-

tached to the scores of A (so that this expert is called

qA) in a two or three expert conﬁgurations to evaluate

the effect of using sample adaptive quality measures in

machine supervisors. As can be seen, using quality

measures does improve the recognition performance.

It is not surprising that human experts perform better

in quality estimation, as this is a very complex task in

which human exper ts still excel. However, the ma-

chine-delivered quality estimates are fairing quite

well, not too far away behind human assessments of

the quality. It is also worth noting that the ﬁnal deci-

sions are suggested by the machine supervisor which

processed both human and machine delivered opinion

parameters transparently.

Multiple Experts. Figure 2 The graphs, from [12], illustrate the recognition performance of two supervisors on the same

data-set. In a probabilistic supervisor (Bayesian) and in (b) a discrimination based supervisor (SVM) were used. The

used experts were common, F: ﬁngerprint, and S: signature. In (b) there were two different ﬁngerprint experts as well as

two different signature experts. Q-Fusion represents quality based fusion.

Multiple Experts. Figure 3 The graphs, adapted from

[11], illustrate the recognition performance of two (qA, B)

and three (qA, B, C) ﬁngerprint experts combined by the

Bayesian supervisor with automatically extracted quality

measures (attached to qA).

990

Multiple Experts

Discrimination Functions Based

Fusion

Discrimination functions are frequently used in pattern

classiﬁcation and can also be used to fuse decisions of

biometric experts. Discriminative statistics differs from

Bayesian approach in that distributions of random

variables are not necessary for decision making. In-

stead, modeling the decision boundaries is the focus

of attention. Here we exemplify the use of this approach

in fusion by Support Vector Machines, SVMs [3].

Assume that we are given a set of observations

; y

g; fx

; y

gfx

; y

gð16Þ

where x

¼ (x

1,j

,...,x

1,j

)

is a feature vector of dimen-

sion m, and y

is the class-label of the latter (relative to

the classes, ‘‘client’’ and ‘‘impostor’’), respectively. As-

sume further that the two classes are separable by a

hyperplane. Then there is an optimal hyperplane in a

high dimensional space to which x is mapped. For

simplicity we assume that the mapping is the (trivial)

identity transformation but other tranformations

using e.g. polynomials, or radial basis functions, can

be used with little impact on the discussion that fol-

lows next [3], provided that appropriate kernel func-

tions are used whenever scalar products are utilized in

computations. The separation hyperplane

f ðxÞ¼w  x þ b ¼ 0 ð17Þ

can be made to have maximal distance d

max

to samples

belonging to the two classes, Fig. 4. The equation

can be multiplied by a nonzero constant such that

kwk¼1∕d

max

. We can (using this freedom) represent

thereby the two class-labels asþ1 and1, to follow the

convention of SVM litterature. Then, w e have

f ðx

Þ1ify

¼ 1

f ðx

Þ1ify

¼1



ð18Þ

Equivalently, the distance is maximized if

kwk

minimized under the constraints given by (18).Ifwe

know w and b, the function f will be a discrimination

function, e.g. f(x)  0 prompts for a decision y ¼ 1.

The parameters w and b can be found by solving a

quadratic problem with linear constraints.

In case the classes are not separable by a hyper-

plane, slack variables x

are introduced so as to allow

a classiﬁcation that makes an error, but that this error

is the smallest on the training/observation set. The

corresponding problem is still a quadratic optimiza-

tion problem

min

kwk

ð19Þ

subject to the constraints

f ðx

Þ1ify

¼ 1  x

f ðx

Þ1ify

¼1 þ x



ð20Þ

The constant C assures that there is a limit on the

amount of change the training vectors can introduce

to the solution.

The SVM formulation allows one to construct a

supervisor that is able to assign a class label y

to the

score vector of m experts x

¼ (x

1,j

, x

2,j

... x

m,j

)

(at j ¼ n

). However, such a supervisor would not be

quality adaptive yet, because the contribution of each

sample to the total cost function would be uniform

due to C ’s being a constant. This can be changed such

that the cost depends on the quality of the biometric

sample by means of a heuristically chosen function.

In the following section we follow the description

of [8 ] to obtain such a sample adaptation and a ﬁnal

supervisor.

Multiple Experts. Figure 4 Illustration of two classes

(circle:client and square:impostor) that are separable by

a hyperplane with direction w. The support vectors that

deﬁne the separation hyperplane are represented by the

outlined square and the circle on dashed hyperplanes. The

width of the separation zone is 2∕kwk which is maximized

by SVM.

Multiple Experts

991

Let the original quality measure delivered by the

expert i be p

2 [0, p

max

] and

ﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ





ð21Þ

where



is the average quality measure of the signals

that expert i delivered for the training samples, which

is also in the range of [0, p

max

]. Then one can train m

SVMs, each having its own discrimination function

, by deﬁning the cost coefﬁcients for the respective

function as follows.

¼ C

6¼i

m1

max



ð22Þ

The coefﬁcient C

represents now the cost of not

inﬂuencing the biometric sample j by expert i and it

is measured as the product of the quality measures of

other exp erts (excluding expert i) on the current sam-

ple j. The training samples of f

are x

, j :1...n, which

equals to x

except that its component corresponding

to expert i has been removed.

¼ðx

1;j

; ; x

i1;j

; x

iþ1;j

; x

m;j

ð23Þ

Here the use i superscript is in the sense of label, not

exponent, signiﬁes that the data of expert i is lacking.

The exponent g

is an empirically chosen constant the

purpose of which is to adjust the overall inﬂuence of

quality based discrimination on the ﬁnal decision. In a

similar fashion an additional discrimination functio n

can be computed, except that the cost coefﬁcients are

now deﬁned as

¼ C

i¼1

i;j

max

ð24Þ

This represents the alternative cost of using all individ-

ual quality measures including those delivered by

expert i. The discrimination function f

is obtained

by an SVM training on full length expert score vectors

, j :1...n, as opposed to f

,i :1,..., m which trains on

, lacking the opinion of expert i . When the system is

operational at time j ¼ n

, the m quality scores q

i;n

well as m expert scores x

i;n

are available. The quality

measures q

i;n

, as well as the corresponding scores q

i;n

and the discrimination functions f

, i :1,...,m, are

re-indexed such that q

1;n

 q

m;n

. A ﬁnal super-

visor can then be obtained by aggregating f

with f

,...

as follows:

¼ b

m1

i¼1

m1

¼1

ðx

Þþð1  b

Þf

ðx

Þð25Þ

where

m;n

 q

i;n

max



ð26Þ

The results of this supervisor is shown in Fig. 2 (2)

where ﬁngerprint and signature traits are fused using

human expert opinions. Again, one can conclude that

skill and sample adaptation do help to improve the

recognition performance.

Related Entries

▶ Quality Measures

References

1. Ross, A., Jain, A.K.: Information fusion in biometrics. Pattern

Recogn. Lett. 24(13), 2115–2125, (2003)

2. Lindley, D.V.: Making decisions. Wiley, London (1990)

3. Cortes, C., Vapnik, V.: Support-vector networks. Mach. Learn.

20, 273–297 (1995)

4. Bigun, E., Bigun, J., Duc, B., Fischer, S.: Expert conciliation for

multi modal person authentication systems by bayesian statis-

tics. In: Bigun, J. Chollet, G. Borgefors G. Audio and Video based

Person Authentication – AVBPA97, pp. 291–300. Springer, Hei-

delberg (1997)

5. Kittler, J., Li, Y., Matas, J., Sanchez, M.R.: Combining evidence in

multimodal personal identity recognition systems. In: Bigun, J.

Chollet, G. Borgefors G. Audio and Video based Person Authen-

tication - AVBPA97, pp. 327–334. Springer, Heidelberg (1997)

6. Ben-Yacoub, S., Abdeljaoued, Y., Mayoraz, E. Fusion of face and

speech data for person identity veriﬁcation. IEEE Trans. Neural

Networks 10(5), 1065 (1999)

7. Gutschoven, B., Verlinde, P.: Multi-modal identity veriﬁcation

using support vector machines (SVM). In: Proceedings of the

Third International Conference on Information Fusion. IEEE

Press (2000). http://citeseer.ist.psu.edu/303146.html;ftp://ftp.

elec.rma.ac.be/user/verlinde/pub/fusion_2000.ps.gz

8. Fierrez-Aguilar, J., Ortega-Garcia, J., Gonzalez-Rodriguez, J.,

Bigun, J.: Discriminative multimodal biometric authentication

based on quality measures. Pattern Recogn. 38, 777–779 (2005)

9. Lim, E., Toh, K.A., Saganthan, P.N., Jiang, X.D., Yan, W.Y.:

Fingerprint image qualit y analysis. In: International Conference

on Image Processing , pp. II: 1241–1244 (2004). URL http: //dx.

doi.org/10.1109/ICIP.2004.1419530

10. Tabassi, E., Wilson, C.L.: A novel approach to ﬁngerprint image

quality. In: ICIP (2), pp. 37–40 (2005). URL http://dx.doi.org/

10.1109/ICIP.2005.1529985

992

Multiple Experts

11. Fronthaler, H., Kollreider, K., Bigun, J., Fierrez, J., Alonso-

Fernandez, F., Ortega-Garcia, J.: Automatic image quality esti-

mation and its application to multi-algorithm veriﬁcation. IEEE

Transactions on Information Forensics and Security p. 3(2):

331–338 (2008)

12. Fierrez-Aguillar, J.: Adapted fusion schemes for multimodal

biometric authentication. Phd thesis, Universidad Politecnica

de Madrid, ETSIT (2006)

Multiple View Geometry

Multiple view geometry is the scientiﬁc ﬁeld which

studies the geometric relations between images of

the same objects taken from different views. It relies

heavily on afﬁne- and projective-geometry.

▶ Palm Vein Image Sensor

Multispectral and Hyperspectral

Biometrics

BESMA ROUI-ABIDI,MONGI ABIDI

The University of Tennessee, Knoxville, TN, USA

Synonyms

Imaging spectroscopy; Multiband biometrics; Fusion,

Wavelet-Based

Definition

Multispectral biometrics are based on data consisting

of four to ten separate images of the same biometric

trait and representing sensors’ responses in different

wavelengths of the electromagnetic spectrum. In con-

trast to conventional images, which generally represent

an integrated response of a single sensor over a wide

range of bands in the same spectral zone, multispectral

data usually refers to multiple separate sensor(s) res-

ponses in relatively narrow spectral bands. The word

multispectral was ﬁrst used in space-based imaging to

denote data acquired in the visible and infrared spec-

tra. In biometrics, the word multispectral has been used

to describe responses in multiple narrow bands either

all in the visible, or all in the infrared, or a mixture

of both. Even though the words hyperspectral and mul-

tispectral have often been used interchangeably, hyper-

spectral imaging usually refers to cases where the

number of bands are higher than 10 and when these

bands encompass more than one region of the electro-

magnetic spectrum, such as the visible and the infra-

red. Without lack of generality, the word multispectral

will be used in the reminder of this article to refer to

any biometric work involving more than the three

customary red, green, and blue channels.

Introduction

Multispectral imaging provides data in the form of a

three dimensional image cube, with two spatial dimen-

sions (horizontal and vertical coordinates) and one

spectral dimension. Also called imaging spectroscopy,

this type of data details the spectral content of each pixel

in the 2D image, therefore providing more than what the

human eye can captur e with its receptors in the red, green,

and blue. The separation of a pixel’s content into infor -

mation within multiple very narr ow wavelengths allows

the material imaged to be identiﬁed based on its spectral

characteristics in addition to its spatial characteristics.

The high number and different characteristics of the

bands make it easier to differentiate between objects that

would look similar in regular intensity images or even in

conventional color images. In other words, a multispec-

tral dataset is a higher resolution image in the spectral

dimension that makes it possible to resolve information

and details non resolvable in conventional images.

The higher dimensionality of multispectral data

presents a desirable feature for biometric systems, i.e.,

the uniqueness of a material based on its spectral

characteristics, including material of a person’s skin,

iris, and vasculature. A conventional sensor’s response

to the combination of mate rial and illumination is not

always most informative in the visible domain. For

instance, conditions such as ni ght time and low inten-

sity light are better dealt within the infrared domain,

while the visible spectrum generally reveals better in-

formation in uniform well lit situations. Concealment

and disguise is another issue that is usually hard to deal

with in the visible domain alone. Figure 1 shows two

images of a person wearing sunglasses, one in the

visible light and the other in the near infrared (NIR).

Multispectral and Hyperspectral Biometrics

993

The eyes, which are fundamental features for any face

recognition algorithm, appear completely blocked out

in the visible image, while visible in the NIR image.

Versatility, usability, and security are some of the

required characteristics of any biometric system. Such

system must have the capability to acquire and process

biometric data at different times of day and night, in a

variety of weather and environmental conditions, and

be resistant to spooﬁng attempts. Multispectral bio-

metrics is one of the few technologies shown to solve

many of the aforementioned issues. Multispectral anal-

ysis has been used to improve recognition rates and

detect spooﬁng attempts for various biometric modal-

ities, including face [1], iris [2], ﬁngerprint [3], and

vasculature [4]. Wavelengths covering the visible spec-

trum all the way to the long wave infrared have been

used in the analysis of Biometric data. Figure 2 displays

the distribution and values of wavelengths along the

electromagnetic spectrum, with emphasis on the most

used bands in biometrics. The visible spectrum com-

prises wavelengths between 400 and 750 nm. Bands of

lengths from 750 to 1,400 nm represent the NIR, while

wavelengths between 1.4 and 3 mm are called short wave

infrared (SWIR). Bands from 3 to 8 mmaresaidtobe

in the mid wav e infrared (MWIR) and bands between

8and15mm are called long wave infrared (LWIR). Mid

wave and long wave are also referred to as thermal

infrared.

Multispectral Biometric Data Acquisition

Most multispectral instruments use a 2D detector array

and are scanned over time to acquire the third dimen-

sion of the cube. Some common techniques for

acquiring multispectral data are: (1) Spatially scanning

split spectrometer, (2) wavelet tunable spectral ﬁlter,

(3) two-dimensional Fourier transform imager, or (4)

a combination of sensors with responses in different

regions of the spectrum. A number of instruments

were used in Biometrics to acquire multispectral data

of face, iris, ﬁngerprint, and vasculature. Separate but

aligned/calibrated visible and infrared cameras were

used to obtain registered face images [5, 6]. Liquid

crystal tunable ﬁlters attached to monochromatic

video cameras were also deployed to acquire multiple

spectral bands (25 bands between 400 and 800 nm [7]

and 31 bands between 700 and 1,000 nm [8]) of face

images over time. Tunable ﬁlters allow the band to be

changed electronically and continuously with select-

able resolution and interval. A spectropolarimetric

camera was used for hyperspec tral face data acquisition

in the range of 450–1,100 nm [9]. Iris images in the

visible and NIR were obtained using a multispectral

camera consisting of three charge coupled devices and

three band-pass prisms, resulting into a 4 band output

[2]. Multispectral ﬁngerprint data has been collected

using a single placement of the ﬁnger on a sensor that

combines a conventional optical ﬁngerprint reader and

multiple illuminators with different wavelengths, light

orientations, and polarization conditions to yield in-

formation on both the surface and subsurface region of

the ﬁnger’s skin [3].

Multispectral Face

Changes in conditions between gallery and probe ima-

ges have often resulted in major performance degrada-

tions of face recognition algorithms. Such conditions

Multispectral and Hyperspectral Biometrics. Figure 1 Visible and NIR images of a subject wearing sun glasses.

Images are acquired using an IR-Enabled Sony DSC-S30 Digital Camera & X-Nite780 nm filter under incandescent lighting

http://www.maxmax.com/aXRayIRExamplesSecurity.htm.

994

Multispectral and Hyperspectral Biometrics

include illumination, pose, time lapse, and expression

[10]. A number of studies were conducted in recent

years using multispectral face data in an attempt to

establish invariance of face biometric systems to

changes in these conditions. Researchers and scientists

addressed various multispectral aspects of face recog-

nition. These included restricting the bands to cer tain

zones of the spectrum, such as using just visible bands

or only infrared wavelengths, as well as combining

spectral bands from the visible, NIR, and/or thermal

infrared. Figure 3 displays a sample set of face images

acquired in different bands of the electromagnetic

spectrum.

A study conducted at the University of Tennessee

used 25 bands in the visible spectrum to analyze the

effect of change in illumination conditions, especially

outdoor versus indoor lighting, on face recognition

performance. Chang et al. analyzed the multispectral

face data in light of the measured spectral distributions

of the illuminants and addressed spectral band selec-

tion as well as the fusion of multispectral data to

improve invariance to illumination [7, 11]. Fusion

methods such as averaging, Principal Components

Analysis,

▶ wavelet analysis, physics based, and illumi-

nation adjustment were implemented and tested. Rank-

based decision fusion was also tested on various bands

Multispectral and Hyperspectral Biometrics. Figure 2 Bands of the electromagnetic spectrum and their wavelengths

www.itcnewsletter.com [Courtesy: Infrared Training Center].

Multispectral and Hyperspectral Biometrics. Figure 3 A set of multispectral face images acquired in 8 single narrow

bands of the visible spectrum, followed by the three R, G, B components from a conventional color camera, the composite

color image, a NIR, and a thermal infrared image of the same subject- (Courtesy of Imaging, Robotics, and Intelligent

Systems lab, University of Tennessee).

Multispectral and Hyperspectral Biometrics

995

and shown to considerably improve recognition perfor-

mance. The authors of the study demonstrated that

fused face data outperforms conventional images espe-

cially under severe changes in illumination conditions

and long time lapses between gallery and probe images.

Using a database of 200 face images spanning 31

bands of the NIR spectrum, Pan et al. demonstrated

that the subsurface information provided in these

bands can be unique to each individual, is relatively

stable over time, and invariant to pose and expression

variations [8, 12].

Other studies on face biometrics involved a combi-

nation of spectral data from different spectral zones.

One of the ﬁrst studies combining visible and infrared

face images was conducted by Sokolinsky et al. [5],

who were able to show an increase in performance by

fusing visible and infrared spectral bands. Kong et al.

demonstrated that wavelet-based fusion of visible and

thermal infrared imagery considerably imp roved face

recognition under a variety of illumination directions

and intensities [1].

Buddharaju et al. combined data in the visible with

data in the thermal spectra and used the physiological

characteristics of the thermal face vasculature to im-

prove recognition rates [4] of face images with var ying

expressions. A score level fusion approach was used.

The authors also studied the effect of change in tem-

perature and sweating on the performance of the algo-

rithm and found that extreme conditions resulted in

non linear variations in thermal data and therefore

deterioration in performance. Kakadiaris et al. com-

bined 3D face reconstructed from multiple visible sen-

sors with the textu re and vasculature emanating from

a calibrated infrared camera to cancel out the effect of

expression changes on recognition performance [6].

Skin temperature was encoded in the metadata records

used for recognition and the algorithm showed a min-

imal drop in performance in the presence of changes

in facial expression compared to the baseline algori-

thms used on the Face Recognition Grand Challenge

database.

Multispectral Iris

Multispectral iris processing and recognition is a rela-

tively new area of biometrics. Early work by Imai was

conducted for a purely clinical application. It describes

experiments for the spectral characterization of the

human iris using an ophthalmic microscope, a di gital

camera, and a spectroradiometer under halogen light.

The spectral bands studied were all in the visible range

from 400 to 700 nm and the study showed that the

recovered iris’ spectra of different subjects presented

clear distinctive differences. No database was built and

no biometric processing was conducted within the

study [13]. Boyer et al. acquired multisp ectral data of

the iris in the visible red, green, and blue and in the

NIR in an effort to demonstrate that different charac-

teristics of the iris will show in different bands based

on the color of the eye [2]. Figure 4 shows an example

of a blue eye iris in the various bands. The authors used

an adaptive color histogram equalization technique to

enhance the iris structure and reveal the needed infor-

mation in the individual channels. They then evaluated

the performance of iris recognition in the various

bands using a database of 24 subjects and 5 samples

per subject. A before and after performance evaluation

showed a substantial improvement in the blue channel.

The authors then conducted a series of experiments of

cross matching between individual channels and

showed that the performance degrades with the in-

crease in distance between the matched channels.

This indicates a decrease in correlation between the

information revealed in the more distant ban ds. The

study went on to show potential improvements to iris

segmentation using color based clustering.

Park and Kang processed infrared multispectral iris

data using a gradient-based fusion scheme. The purpose

of the study was to detect spooﬁng attempts by using

the speciﬁc complementary information contained

in the various bands of a real iris. Images with n o

spectral variation (fake irises) resulted in an errone-

ous fused image that could not be m atched against

actual irises [14].

Multispectral Fingerprint

Fingerprint sensors are widely used in law enfor-

cement, cyber, and physical security applications.

However, the performance of these sensors is easily

degraded by physiological and environmental condi-

tions, such as dry or wet skin, heat or cold, and ambi-

ent lighting conditions. These factors can result in

failure to enroll, a high rate of false rejec tion, and

dissatisﬁed users. Multispectral ﬁngerprint Biometrics

is a relatively new technology aiming to solve some of

996

Multispectral and Hyperspectral Biometrics

the aforementioned problems. Multispectral ﬁnger-

print w as ﬁrst introduced by Rowe and Nixon [3],

who designed and tested a sensor able to acquire ﬁn-

gerprint data in ﬁve visible spectral bands (445, 500,

574, 610, and 660 nm). The purpose of the effort was

to improve the usability of ﬁngerprint biometric sys-

tems by improving the robustness of the sampling

process to physiological and environmental conditions

such as dry skin, rain, heat, lig hting conditions, cold,

etc. The authors analyzed ﬁngerprint recognition on a

number of samples from a baseline database of 4,105

samples, a cold ﬁnger study of 300 samples, and a wet

ﬁnger study of 1,186 samples using multiple sensors of

the same design. A dramatic improvement in Equal

Error Rate was achieved by using the new multispectral

acquisition technique compared to the customar y

total internal reﬂectance technique (IRT). An example

of multispectral ﬁnger data and an IRT op tical sample

of a ﬁngerprint are shown in Fig. 5. The subject was

asked to apply too little pressure on the sensors. In

addition to a better and more successful enrollment,

the new multispectral sensor also provides information

on the subsurface of the ﬁnger and other attributes that

makes spooﬁng by using an artiﬁcial rubber or other

fake ﬁngers difﬁcult to achieve [15].

A diffe rent form of multispectral analysis called

Fourier Transform Infrared (FTIR) chemical imaging

was used in forensic studies to recover and identify

latent ﬁngerprints from traditionally hard to analyze

backgrounds [16]. In this approach, the infrared spec-

trum of the sample is collected by passing a beam

of infrared light through the sample. A systematic

methodolog y is used for each surface by optimizing

the spectral resolution, number of scans, and pixel

aggregation. Examination of the transmitted light

reveals how much energy was absorbed at each wave-

length. A Fourier transform instrument measures all

Multispectral and Hyperspectral Biometrics. Figure 5

Fingerprint scans (left) and their enhanced versions (right)

using a conventional IRT optical sensor (top) and a

multispectral sensor (bottom). The subject was asked to

apply very little pressure on the sensor [15].

Multispectral and Hyperspectral Biometrics. Figure 4 An example of a blue iris showing the false color NIR, the RGB

composite, and each individual channel [2].

Multispectral and Hyperspectral Biometrics

997

wavelengths at once. From this, a transmittance or

absorbance spectrum is produced. Analysis of these

absorption characteristics reveals details about the mo-

lecular structure of the sample. The use of this ap-

proach was shown to improve sample analysis from

polymer bank notes and aluminum drink cans. Crane

et al. used various processing techniques on FTIR

images of ﬁngerprints on a number of challenging

porous and nonporous substrates to extract the ridge

patterns. Techniques used include basic infrared spec-

troscopic band intensities, addition and subtraction of

band intensity measurements, principal components

analysis, and calculation of second derivatives band

intensities. Trace evidence within the ﬁngerprints

were also recovered and identiﬁed.

Related Entries

▶ Biometric System

▶ Face Recognition

▶ Fingerprint Recognition

▶ Iris Recognition

▶ Liveness and Anti Spooﬁng

▶ Near Infrared Based Face Recognition

▶ Skin Spectroscopy

References

1. Kong, S.G., Heo, J., Boughorbel, F., Zheng, Y., Abidi, B.R.,

Koschan, A., Abidi, M.A.: Multiscale fusion of visible and ther-

mal IR images for illumination-invariant face recognition. Int.

J. Comput. Vis. 71(2), 223–253 (2007)

2. Boyce, C., Ross, A., Monaco, M., Hornak, L., Li, X.: Multispec-

tral iris analysis: a preliminary study. In: Proceedings of IEEE

International Conference on Computer Vision and Pattern Rec-

ognition Workshop on Biometrics (CVPRW), New York. IEEE

publisher, New York (2006)

3. Rowe, R., Nixon, K.A., Corcoran, S.P.: Multi spectral ﬁngerprint

biometrics. In: Proceedings of IEEE Workshop on Information

Assurance and Security, United States Military Academy, West

Point, NY, pp. 14–20 (2005)

4. Buddharaju, P., Pavlidis, I.: Multi-spectral face recognition –

fusion of visual imagery with physiological information. In:

Hammoud, R., Abidi, B., Abidi, M. (eds.) Face Biometrics for

Personal Identiﬁcation. Springer, Berlin, pp. 91–108 (2007)

5. Socolinsky, D.A., Wolff, L.B., Neuheisel, J.D., Eveland, C.K.:

Illumination invariant face recognition using thermal

infrared imagery. In: Proceedings of the IEEE ICPR, CVPR,

Kauai, HI, pp. 527–534 (2001)

6. Kakadiaris,I.A.,Passalis,G.,Toderici,G.,Lu,Y.,Karampatziakis,N.,

Murtuza, N., Theoharis, T.: Expression-invariant multispectral

face recognition: you can smile now! In: Flynn, P.J.,

Sharath Pankanti. (eds.) Proceedings of SPIE. Biometric Tech-

nology for Human Identiﬁcation III, vol. 6202, pp. 620 204.1–

620 240.7 (2006)

7. Chang, H., Harishwaran, H., Yi, M., Koschan, A., Abidi, B.,

Abidi, M.: An indoor and outdoor, multimodal, multispectral

and multi-illuminant database for face recognition. In: Proceed-

ings of the IEEE International Conference on Computer Vision

and Pattern Recognition, Biometrics Work Shop, New York,

NY (2006)

8. Pan, Z., Healey, Z., Prasad, M., Tromberg, B.: Face recognition

in hyperspectral images. IEEE Trans. Pattern Anal. Mach. Intell.

25(12), 1552–1560 (2003)

9. Denes, L.J., Metes, P., Liu, Y.: Hyperspectral face database. Tech-

nical report CMU-RI-TR-02-25, Robotics Institute, Carnegie

Mellon University (2002)

10. Phillips, P.J., Grother, P., Micheals, R.J., Blackburn, D.M.,

Tabassi, E., Bone, J.M.: FRVT 2002: Evaluation Report (2003).

http://www.frvt.org/FRVT2002/documents.htm

11. Chang, H., Koschan, A., Abidi, B., Abidi, M.: Physics-based

fusion of multispectral data for improved face recognition. In:

Proceedings of the International Conference on Pattern Recog-

nition, Hong Kong (2006)

12. Pan Z., Healey G., Prasad M., Tromberg B.: Multiband and

spectral eigenfaces for face recognition in hyperspectral images.

In: Proceedings of SPIE, San Jose, CA, vol. 5779, pp. 144–151

(2005)

13. Imai, F.H.: Preliminary experiment for spectral reﬂectance esti-

mation of human iris using a digital camera, Munsell Color

Science Laboratory Technical Report (2002)

14. Park, J.H., Kang, M.G.: Iris recognition against counterfeit attack

using gradient based fusion of multi-spectral images. Advances

in Biometric Person Authentication, LNCS vol. 3781. Springer,

Berlin, pp. 1611–3349 (2005)

15. Nixon, K.A., Rowe, R.K.: Multispectral ﬁngerprint imaging for

spoof detection. In: Jain, A.K., Ratha, N.K. (ed.) Proceedings of

SPIE. Biometric Technology for Human Identiﬁcation II, vol.

5779, pp. 214–225 (2005)

16. Crane, N.J., Bartick, E.G., Perlman, R.S., Huffman, S.: Infrared

Spectroscopic Imaging for Noninvasive Detection of Latent Fin-

gerprints. Forensic Sci. 52(1), 48–53 (2007)

Multistage Matching

Multistage matching is a technique used in order to

simultaneously achieve high accuracy and high speed

during the matching stage: A fast initial algorithm is

used to compare the query to each template of the

998

Multistage Matching

database, and a decision is made to disregard or keep

the template for the next stage. A slower, more accu-

rate, algorithm is then applied to the surviv ing tem-

plates, eliminating the additional database entries. This

process is repeated until the last stage, where the ﬁnal

decision will be made. In multistage matching, the

slowest algorithms are applied only to a few numbers

of templates and, hence, have a small impact on the

overall matching speed.

▶ Biometric Algorithms

Mutual Authentication

Mutual authentication or two-way authentication is

a process in which two entities in communication

authenticates each other before any application data

is transferred. This is typically achieved by exchange

of digital certiﬁcates issued by trusted entities.

Mutual authentication helps in eliminating the

man-in-the-middl e attack, where an adversary estab-

lishes independent links with both the victims and

relays messages between them. The victims are led to

believe that they are in direct communication, while

infact, the entire communication between them is con-

trolled by the adversary.

▶ Security Issues, System Design

Mutual information

The mutual information I(x1; x2) of two random

variables x1 and x2 is the relative entropy between

the joint distri bution p(x1, x2) and the product distri-

bution p(x1)p(x2),

Iðx1; x2Þ¼

pðx1; x2Þlogðpðx1; x2ÞÞ

pðx1Þpðx2Þ

dx1dx

▶ Independent Component Analysis

Mutual information

999