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

Подождите немного. Документ загружается.

Definition

Linear dimension reduction technique reduces the di-

mension of biometric data using a linear transform. The

linear transform is always learned by optimization of a

criterion. Biometric data are then projected on to the

range space of this transform. Subsequent processing is

then performed in that lower-dimensional space.

Introduction

In biometrics, data are invariably represented in vectors

and the dimensionality is consistently very high. It would

be computationally expensive to process them directly by

using many algorithms. Moreo ver, it is sometimes desir-

able to exact robust, informativ e or discriminative facts

contained in the data. For these reasons, a lower dimen-

sional subspace is found such that the most important

part of the data is retained for linear representation.

Among the techniques for learning such subspace, linear

dimension reduction methods are popular.

Given a set of N data samples {x

,..., x

}, where x

. Linear dimension reduction technique ﬁnds a

linear transform matrix W ¼ (w

, ..., w

ℓ

) such that

data are projected on to the range space span{w

,...,

ℓ

}by

¼ W

; ð1Þ

where

denotes the transpose and y

is the representa-

tion of x

in the lower-dimensional space.

The linear dimension reduction technique is equal

to extraction of linear features w

,...,w

ℓ

. Different

linear dimension reduction techniques lie in different

goals of the information retained by these linear fea-

tures. Generally, they can be categorized into three

classes, which address the linear dimension problem

in the

▶ unsupervised, ▶ supervised, and ▶ semi-

supervised cases respectively. Some representative

algorithms of these classes are des cribed here.

When biometric data are represented in a matrix

form, line ar dimension reduction techniques can also

be extended. This kind of extension is called the two-

dimensional linear dimension reduction technique.

Unsupervised Linear Dimension

Reduction

Unsupervised linear dimension reduction aims at the

extraction of reconstructive features. Biometric data are

then linearly approximated in a lower-dimensional sub-

space spanned by the reconstructive features. Among the

popular linear algorithms used in biometrics for dimen-

sion reduction, the Principal Component Analysis (PCA)

[1] is the most representative one. The PCA ﬁnds a

lower-dimensional space that preserves the greatest

variations of data, that is, an optimal transform is

learned by maximization of the following criterion:

opt

¼ arg max

W¼I

traceðW

WÞ; ð2Þ

where C

is the total-class covariance matrix deﬁned by

i¼1

ðx

 uÞðx

 uÞ

; u ¼

i¼1

: ð3Þ

Eigen-decomposition is used to ﬁnd w

,...,w

ℓ

, which

are eigenvectors corresponding to the largest ℓ eigen-

values. However, preserving the largest eigenvectors

may not be the best strategy for other applications of

the PCA, such as recognition and clustering . In these

cases, the selection of proper principal components

may be useful.

Features extracted by the PCA are statistically

uncorrelated, but they are not ensured to be statistically

independent. If it is assumed that data are approximately

linear representations of some independent sources,

then it is useful to ﬁnd these independent components

for representation of data in an intrinsic subspace. The

Independent Component Analysis (ICA) [2] pursues

features in this aspect. Let X ¼ [x

,..., x

] be the

data matrix. In the IC A, data matrix X is approximated

by a multiplication of mixing matrix A and matrix S

which consists of independent components as follows

X  AS: ð4Þ

Several approaches have been proposed for the estima-

tion of the ICA and the most popular of these is

FastICA [2].

In biometric learning, especially facial image anal-

ysis, has been experimentally demonstrated that the

extraction of localized features is useful for recognition

and for the interpretation and understanding of the

structure of data as well. Unlike the PCA and the ICA,

non-negative matrix factorization (NMF) [3] is a novel

technique for this purpose. In the NMF, data are

approximated only by additive combination of non-

negative components. The NMF, therefore, ﬁnds two

non-negative matrices, namely, the component matrix

opt

and the coefﬁcient matrix H

opt

, which are mini-

mums of the following criterion

900

Linear Dimension Reduction

ðW

opt

; H

opt

Þ¼arg min

ðW;HÞ

jjX  WHjj

;

s:t: W  0 & H  0:

ð5Þ

where jjjj

is the Frobenius norm. Other than using

the Euclidean distance, the following criterion based

on the generalized Kullback-Leibler divergence that is

lower bounded by Zoro is also popular

ðW

opt

; H

opt

Þ¼arg min

ðW;HÞ

i;j

log

ðWHÞ

 X

þðWHÞ

; s:t: W  0 & H  0:

ð6Þ

Closed forms of W

opt

and H

opt

in the NMF are difﬁcult

to obtain. At present, the multiplicative update method

[4] has been widely used for ﬁnding a locally optimal

solution.

The former three methods do not take the geomet-

ric relationship between data into account. Assume

data are actuall y distributed on a manifold, it is natural

that the neighboring data in the input space would also

be close to each other when they are represented in the

lower rank subspace. The locality preserving projection

(LPP) [5] is a typical algorithm developed for this

purpose. It learns an optimal transform W

opt

such

that the locality relationship between data in the

input data space are preserved after dimension reduc-

tion. The cost function of a transform W for locality

preserving is modeled by

i;j

jjy

 y

traceðW

i;j

ðx

 x

Þðx

 x

WÞ;

where g

the afﬁnity between samples x

and x

deﬁned by

1ifx

ðx

Þ or x

ðx

Þ;

0 otherwise;



ð7Þ

and N

ðx

Þ is the set of s nearest neighbors of data x

Then the optimal transform explored by the LPP is

learned as follows

opt

¼ arg min

XDX

¼1

traceðW

WÞ; ð8Þ

where D¼diag(∑

,...,∑

), S

¼ XLX

and L is

termed the Laplacian matrix (He and Niyogi [5]) for-

mulated by

L ¼ D  G; G

¼ g

: ð 9Þ

Supervised Linear Dimension

Reduction

PCA, the ICA, the NMF and the LPP are methods for

learning unlabeled data. Since the features extracted

by these methods are not discriminative, they may not

be useful for recognition. When class labels of data

are known, supervised techniques for dimension re-

duction can be developed by additionally using label

information. Suppose data x

, ..., x

are drawn from

L classes, namely, classes C

,...,C

, and N

be the

number of samples of class C

, the supervised dimen-

sion reduction technique then ﬁnds a transform such

that data of the same class are close to each other

while data of different classes are scattered as far as

possible. A popular and widely applied technique is

(Generalized) Fisher’s linear discriminant analysis

(LDA) [6]. The LDA ﬁnds a lower-dimensional sub-

space in which the ratio between between-class vari-

ance and within-class variance is maximized. That is, a

discriminative linear transform W

opt

is the maximum

of the following criterion

opt

¼ arg max

traceðW

WÞ

traceðW

WÞ

; ð10Þ

where S

and S

are between-class and within-class

covariance matrices respectively, which are deﬁned as

follows:

k¼1

ðx

u

Þðx

u

; u

;

k¼1

ðu

uÞð u

uÞ

To obtain the optimum of a criterion [10], the

eigenvectors of S

1

have to be determined. As

dimensionality of data is high and training samples

are in invariably limited in biometric learning, with-

in-class covariance matrix S

will be singular. This

kind of singularity problem is referred to as the small

sample size problem in the LDA. Techniques used to

address this problem include, PCA+LDA [7], the Null-

space LDA (N-LDA) (Chen et al. [8]), and the regular-

ized LDA [9]. In the PCA+LDA, LDA is performed in a

principal component subspace in which within-class

covariance matrix is of full rank. In the N-LDA, the

nullspace of within-class covariance matrix S

is ﬁrst

learned; then data are projected on to that subspace;

ﬁnally the discriminant transform is found for

Linear Dimension Reduction

901

maximization of the variance of between-class data.

Though both the PCA+LDA and the N-LDA are

lower-rank methods that ﬁrst project data onto a

lower-dimensional subspace before implementation

of any discriminant processing, they are different in

that S

in the PCA+LDA will be of full rank after

dimension reduction by the PCA while trace

W) ¼ 0 after dimension reduction in the N-

LDA. Unlike the lower-rank approach, in the regular-

ized LDA (R-LDA), a regularized term, such as aI

where a > 0, is added to matrix S

in Eq. (10),

so that Fisher criterion is well-conditioned. Regu-

larization is a straightforward way to solve this

singularity problem, but the regularized parameter

will have a signiﬁcant impact on the performance

of the R-LDA.

An alternative way to maximize between-class var-

iance as well as minimize within-class variance can be

modeled as follows:

opt

¼ arg max

traceðW

WÞl traceðW

WÞ;

ð11Þ

where l > 0. The advantage of using this model is that

computation of the inverse of S

is not required, but

how to determine the importance weight l could be a

problem. This criterion is known as the maximum

margin criterion [10] when l ¼ 1.

At times, unsupervised linear dimension reduction

techniques can be used as a preprocessing step before

applying supervised techniques. Methods driven in

this way are two-step dimension reduction techniques.

The PCA+LDA is typically in line with this approach.

In addition, though supervised methods are always

preferred for recognition, it is difﬁcult to ascertain

which kind of linear dimension reduction technique

is the best. For example, the PCA may be better than

the LDA for face recognition if the number of training

samples for each class is sm all [11].

Semi-Supervised Linear Dimension

Reduction

Linear dimension reduction for partially labeled data

would be highly demanded for large scale problems,

since labeling data is an expensive task. Therefore, it is

desirable to utilize unlabeled data for extraction of

supervised features for dimension reduction. Among

the developed techniques to achieve this goal, a special

regularized LDA for performing linear dimension re-

duction on partially labeled data can be formulated as

follows [12]:

opt

¼ arg max

traceðW

WÞ

traceðW

ðS

þ bS

ÞWÞ

; ð12Þ

where S

¼ S

þ S

and b > 0. In this criterion, labeled

data are used to estimate superv ised class covariance

information, and the effect of unlabeled data is

reﬂected by the Laplacian term S

. The underusing

idea is that labeled data are separated in the same

way as done in the LDA while the neighboring data

including unlabeled data are nearby after dimension

reduction.

Two-Dimensional Linear Dimension

Reduction Techniques

Many well-known linear dimension reduction techni-

ques assume that input patterns are represented in vec-

tors. However, biometric data are captured in images,

and the dimensionality is very high when they are

reshaped into vectors. Unlike traditional techniques,

some linear dimension reduction techniques are devel-

oped by performing linear transformation directly

on matrix form data, such as image matrices. This is

advantageous in tackling large scale problems. Suppose

, ..., X

are the corresponding matrix representa-

tions of vector form data x

,...,x

, then a transform

W for dimension reduction of X

would perform as

follows

¼ W

; ð13Þ

where Y

is the representatio n after dimension reduc-

tion. Among the developed techniques are two repre-

sentative algorithms: two-dimensional principal

component analysis (2D-PCA) [13] and two-dimen-

sional linear discriminant analysis (2D-LDA) [14]. The

vector-based PCA and LDA are also referred to as the

1D-PCA and 1D-LDA respectively.

The main difference between 2D-PCA and 1D-PCA

as well as between 2D-LDA and 1D-LDA lies in their

different means of covariance matrix estimation. In

two-dimensional linear reduction techniques, the co-

variance matrices are calculated directly based on data

represented in matrix form. Apart from this, the main

902

Linear Dimension Reduction

ideas of 2D-PCA and 2D-LDA are almost similar to

those of 1D-PCA and 1D-LDA respectively. More spe-

ciﬁcally, 2D-PCA learns the optimal transform via the

following criterion

opt

¼ arg max

W¼I

traceðW

WÞ; ð14Þ

where

i¼1

ðX

 UÞðX

 UÞ

; U ¼

i¼1

For 2D-LDA, the criterion is modiﬁed in a simil ar

manner as follows:

opt

¼ arg max

traceðW

WÞ

traceðW

WÞ

; ð15Þ

where

k¼1

ðX

 U

ÞðX

 U

;

k¼1

ðU

 UÞðU

 UÞ

The above 2D-PCA and 2D-LDA are unilateral,

which mean s only one transform matrix multiplied

on one side of the data matrix is available. To overcome

this limitation, there are generalizations such as bilat-

eral 2D-PCA [15] and bilateral 2D-LDA [16], which

learn transform matrices W

and W

on both sides of

the data matrix and perform dimension reduction for

data X

as follows

¼ W

: ð16Þ

However, it would be difﬁcult to obtain a closed solu-

tion for bilateral techniques, and alternating

optimization methods are used for ﬁnding a locally

optimized solution.

In general, two-dimensional linear dimension re-

duction techniques gain lower cost of computation,

and 2D-L DA in particular avoids the small sample

size problem. However, it is cannot be deﬁnitely said

that two-dimensional techniques are better. An in-

sightful analysis and extensive comparisons between

2D-LDA and 1D-LDA have been made by [17]. Besides

the theoretical comparisons, Zheng et al. noted that

2D-LDA is not always better than 1D-LDA when the

number of training samples for each class or the num-

ber of extracted features is small.

Summary

Linear dimension reduction is an important step for

processing of biometric data. It is equal to extraction

of a set of linear feature vectors, which span a lower-

dimensional subspace. Linear techniques for ﬁnding

the most robust, informative and discriminative

information are fundamental technologies in pattern

recognition. Besides, the development of new linear

techniques for large scale problems in biometric learning

is an important topic.

Related Entries

▶ Feature Extraction

▶ Kernel Methods

▶ LDA (Linear Discriminant Analysis)

▶ Manifold Learning

▶ PCA (Principal Comp onent Analysis)

References

1. Turk, M., Pentland, A.: Eigenfaces for recognition. J. Cogn.

Neurosci. 3(1), 71–86 (1991)

2. Hyva

rinen, A., Oja, E.: Independent component analysis:

algorithms and applications. Neural Networks 13, 411–430

(2000)

3. Lee, D.D., Seung, H.S.: Learning the parts of objects by

non-negative matrix factorization. Nature 401, 788–791

(1999)

4. Lee, D.D., Seung, H.S.: Algorithms for non-negative matrix

factorization. In: Advances in Neural Information Processing

Systems, Denver, Colorado pp. 556–562 (2000)

5. He, X., Niyogi, P.: Locality preserving projections. In: Advances

in Neural Information Processing Systems, Vancauver, Canada

pp. 153–160 (2003)

6. A., F.R.: The Use of Multiple Measures in Taxonomic Problems.

Ann. Fisher, RA Eugenics 7, 179–188 (1936)

7. Belhumeour, P.N., Hespanha, J.P., Kriegman, D.J.: Eigenfaces vs.

ﬁsherfaces: recognition using class speciﬁc linear pro-

jection. IEEE Trans. Pattern Analy. Mach. Intell. 19(7),

711–720 (1997)

Linear Dimension Reduction

903

8. Chen, L.F., Liao, H.Y.M., Ko, M.T., Lin, J.C., Yu, G.J.: A

new LDA-based face recognition system which can solve the

small sample size problem. Pattern Recogn. 33, 1713–1726

(2000)

9. Webb, A.R. (ed.): Statistical Pattern Recognition, 2nd ed. 2002.

Wiley, England (2002)

10. Li, H., Jiang, T., Zhang, K. Efﬁcient and robust feature extraction

by maximum margin criterion. IEEE Trans. Neural Networks

17(1), 157–165 (2006)

11. Martı

nez, A.M., Kak, A.C.: PCA versus LDA. IEEE Trans. Pattern

Analy. Mach. Intell. 23(2), 228–233 (2001)

12. Cai, D., He, X., Han, J.: Semi-supervised Discriminant

Analysis. In: IEEE Int. Conf. Comput. Vis. Rio de Janeiro, Brazil

(2007)

13. Yang, J., Zhang, D., Frangi, A.F., Yang, J.y.: Two-dimensional

PCA: a new approach to appearance-based face representation

and recognition. IEEE Trans. Pattern Analy. Mach. Intell. 26(1),

131–137 (2004)

14. Xiong, H., Swamy, M.N.S., Ahmad, M.O.: Two-

dimensional FLD for face recognition. Pattern Recogn. 38,

1121–1124 (2005)

15. Kong, H., Li, X., Wang, L., Teoh, E.K., J.-G., W., Venkateswarlu, R.:

Generalized 2D principal component analysis. In: IEEE Interna-

tional Joint Conference on Neural Networks, vol. 1, pp. 108–113.

Oxford, UK (2005)

16. Ye, J.P., Janardan, R., Li, Q.: Two-dimensional linear discrimi-

nant analysis. In: Advances in Neural Information Processing

Systems, pp. 1569–1576 (2004)

17. Zheng, W.S., Lai, J.H., Li, S.Z.: 1D-LDA versus 2D-

LDA: when is vector-based linear discriminant analysis

better than matrix-based? Pattern Recogn. 41(7), 2156–2172

(2008)

Linear Feature Extraction

▶ Linear Dimension Reduction

Linearly Symmetric Image

An image described by f(x, y) is linearly sy mmetric if its

isocurves have a common direction, i.e., there exists a

1D function h(t) such that f(x, y)=h(kxx + kyy) for a

certain (constant) direction vector k =(kx, ky)T.

▶ Fingerprint Features

Lip Movement Recognition

PETAR S. ALEKSIC

Google Inc ., New York, NY, USA

Synonyms

Audio–visual-dynamic speaker recognition; Visual-

dynamic speaker recognition

Definition

Lip movement recognition is a speaker recognition

technique, where the identity of a speaker is deter-

mined/veriﬁed by exploiting information contained in

dynamics of changes of visual features extracted from

the mouth region. The visual features usually consist of

appropriate representations of the mouth appearance

and/or shape. This dynamic visual information can

also be used in addition to the acoustic information

in order to improve the performance of audio-only

speaker recognition systems and increase their resil-

ience to spooﬁng, therefore giving rise to audio–visual-

dynamic speaker recognition systems.

Introduction

Speech contains information about identity, emotion,

location, as well as linguistic information, and plays a

signiﬁcant role in the development of human computer

interaction (HCI) systems, including speaker recogni-

tion systems. However, audio-only systems can per-

form poorly even at typical acoustic background

signal-to-noise ratio levels (10 to 20 dB). In addition,

they can be sensitive to microphone types (desktop,

telephone, etc.), acoustic environment (car, plane,

factory, babble, etc.), channel noise (telephone lines,

VoIP, etc.), or complexity of the scenario (speech under

stress, Lombard speech, whispered speech).

Similarly, the visual modality can be exploited to

improve HCI [1, 2]. Visual facial features extracted

from the mouth region are both correlated to the

produced audio signal [3] and also contain comple-

mentary information to it [4]. Lip movement recogni-

tion systems, also known as visual-dynamic speaker

904

Linear Feature Extraction

recognition systems, exploit information contained in

dynamics of changes of visual features extracted from

video sequences of the mouth area (see Fig. 1). Although

speaker recognition systems that utilize only visua l-

dynamic information can achieve high recognition

rates, visual-dynamic information is typically utilized

together with acoustic information to improve speaker

recognition performa nce, resulting in audio–visual-

dynamic speaker recognition systems (see Fig. 1).

There is a number of audio–visual fusion techniques

[5], that combine audio and v isual information to

achieve higher recognition performance than both

audio-only and visual-only system s. The use of visual

information improves speaker recognition perfor-

mance even in noise-free environments [ 5, 6]. The

potential for such improvements is even greater in

acoustically noisy environments, since v isual features

are typically much less affected by acoustic nois e than

the acoustic features. In addition, audio-only, as well as

static-image-based (face recognition) pers on recogni-

tion systems are vulnerable to impostor attacks (spoof-

ing), if the impostor possesses a photograph and/o r

speech recordings of the client. However, it is signiﬁ-

cantly more difﬁcult for an impostor to impersonate

both acoustic and dynamical-visual (lip movements)

information simultaneously. The main advantage of

audio–visual-dynamic speaker recognition systems

lies in their robustness, since each modality can pro-

vide independent and complementary information

and therefore prevent performance degradation due

to the noise present in one or both of the modalities,

as well as increase resilience to spooﬁng. Another ad-

vantage of audio–visual speaker recognition systems is

that they use unobtrusive and low-cost methods for

extracting biometric features, thus enabling natural

speaker recognition and reducing inconvenience.

Operation of a Lip-Movement

Recognition System

The block diagram of a visual-dynamic speaker recog-

nition system is shown in Fig. 2. It consists of three

main blocks, corresponding to prepr ocessing, visual

feature extraction, and classiﬁcation phases.

In the preprocessing phase, detection and tracking of

the face and the important facial features in the input

video are performed. These tasks represent challenging

problems, especially in cases where the background,

head pose, and lighting are varying [7]. The most com-

monly used face detection techniques use statistical

modeling of the face appearance to obtain a classiﬁcation

Lip Movement Recognition. Figure 2 Block diagram of a visual-dynamic speaker recognition system.

Lip Movement Recognition. Figure 1 Speaker recognition systems.

Lip Movement Recognition

905

of image regions into face and non face classes. These

regions are typically represented as vectors of grayscale

or color image pixel intensities. They are often projected

onto lower dimensional spaces, and are deﬁned over a

number of possible locations, scales, and orientations in

the video frame. One or more techniques, such as artiﬁ-

cial neural networks (ANNs), support vector machines

(SVMs), Gaussian mixture models (GMMs), clustering

algorithms, and linear discriminants, are usually utilized

to classify these regions. Typically, face detection goes

hand-in-hand w ith tracking in which the temporal cor-

relation is taken into account.

After successful face detection, facial features, such as

mouth corners, eyes, nostrils, and chin, are detected by

utilizing prior knowledge of their relative position on the

face to simplify the search. If color information is avail-

able, it can be utilized to detect certain facial features

(especially lips). Subsequently, the mouth region-of-

interest (ROI), containing useful visual information is

extracted. The ROI is typically a rectangle containing

the mouth, possibly including larger parts of the lower

face, such as the jaw and cheeks (see Fig. 3). The nor-

malization is usually performed with respect to head-

pose information and lighting. The ROI can also be

extended into a three-dimensional rectangle, contain-

ing adjacent frame ROIs, thus capturing dynamic

visual information.

In the visual feature extraction phase, relevant visual

features from the ROI are extracted. Investigating choice

of visual features for speaker recognition is a relatively

new research topic. The various sets of the choice facial

features proposed in the literature are generally grouped

into three categories [8]: (1) appearance-based features;

(2) shape-based features; and (3) features that are a

combination of both appearance and shape features

[5]. The appearance-based visual features are extracted

from the ROI using image transforms, such as princi-

pal component analysis (PCA,) generati ng ‘‘eigenlips’’

[9], the discrete cosine transform (DCT) [10], the

discrete wavelet transform (DWT) [10], linear discrim-

inant analysis (LDA) [2, 5, 11], Fish er linear discrimi-

nant (FLD), etc. (see Fig. 3). These transforms are

applied on a feature vector, created by ordering the

grayscale pi xel values inside the ROI, and are com-

monly applied in series to cope with the ‘‘curse of

dimensionality’’ problem.

Shape-based visual mouth features are divided into

geometric, parametric, and statistical (see Fig. 3). With

shape-based features it is assumed that most of the

information is contained in the shape of the speaker’s

lips [5, 12]. Hence, such features achieve a compact

representation of mouth images using low-dimen-

sional vectors. Geometric features, such as the height,

width, perimeter of the mouth, etc., are meaningful to

humans and can be readily extracted from the mouth

images. Alternatively, model-based visual features are

typically obtained in conjunction with a parametric or

statistical facial feature extraction algorithm. With

model-based approaches, the model parameters are

directly used as visual features [5, 12]. Examples of

Lip Movement Recognition. Figure 3 Various visual facial features divided into appearance- and shape-based features.

906

Lip Movement Recognition

statistical models are active shape and active appear-

ance models [13].

The combination of appearance- and shape-based

visual features has also been utilized in expectation of

improving the performance of the recognition system,

since they contain respectively low- and high-

level information about the person’s lip movements.

Appearance- and shape-based features are usually just

concatenated, or a single model of face shape and

appearance is created [13]. The dynamics of the

changes of v isual features are usually captured by aug-

menting the visual feature vector by its ﬁrst- and

second-order time derivatives, computed over a short

temporal window centered at the current video frame.

Finally, in the classiﬁcation phase, a num ber of

classiﬁers can be used to model prior knowledge of

how the visual features are generated by each speaker.

They are usually statistical in nature and utilize ANNs,

SVMs, GMMs, HMMs, etc. The parameters of the

prior model are estimated during training. During

testing, based on the trained model the posterior prob-

ability is maximized, and identiﬁcation/veriﬁcation

decision is made.

Summary

Lip movement recognition and audio–visual-dynamic

speaker recognition are speaker recognition technolo-

gies that are user-friendly, low-cost, and resilient to

spooﬁng. There are many biometric applications,

such as, sport venue entrance check, access to desktop,

building access, etc., in which it is very impor tant to

use unobtrusive methods for extracting biometric fea-

tures, thus enabling natural person recognition and

reducing inconvenience. Low cost of audio and video

biometric sensors and the ease of acquiring audio and

video signals (even without assistance from the client),

makes biometric technology more socially acceptable

and accelerates its integration into every day life.

Related Entries

▶ Face Recognition

▶ Face Tracking

▶ Multibiometrics

▶ Multibiometrics and Data Fusion

▶ Multimodal Systems

▶ Speaker Matching

▶ Session Effects on Speaker Modeling

▶ Speaker Recognition, Overview

▶ Speech Analysis

▶ Speech Production

▶ Spooﬁng

References

1. Chen, T., Rao, R.R.: Audio-visual integration in multimodal

communication. Proc. IEEE 86(5), 837–852 (1998)

2. Aleksic, P.S., Potamianos, G., Katsaggelos, A.K.: Exploiting visual

information in automatic speech processing. In: Bovik, A.L.

(ed.) Handbook of Image and Video Processing. Academic,

London (2005)

3. Aleksic, P.S., Katsaggelos, A.K.: Speech-to-video synthesis using

MPEG-4 compliant visual features. IEEE Trans CSVT, Special

Issue on Audio and Video Analysis for Multimedia Interactive

Services, pp. 682–692, May (2004)

4. Summerﬁeld, A.Q.: Some preliminaries to a comprehensive

account of audio-visual speech perception. In: Campbell, R.,

Dodd, B. (eds.) Hearing by Eye: The Psychology of Lip-Reading,

pp. 3–51. Lawrence Erlbaum, London, United Kingdom (1987)

5. Aleksic, P.S. Katsaggelos, A.K.: Audio-visual biometrics. IEEE

Proc 94(11), 2025–2044 (2006)

6. Chaudhari, U.V., Ramaswamy, G.N., Potamianos, G., Neti, C.:

Audio-visual speaker recognition using time-varying stream re-

liability prediction. IEEE Proc. Int. Conf. Acoustics Speech Sig-

nal Process. (Hong Kong, China) 5, V-712–15 (2003)

7. Hjelmas, E., Low, B.K.: Face detection: A survey. Computer

Vision. Image Understand. 83(3), 236–274 (2001)

8. Hennecke, M.E., Stork, D.G., Prasad, K.V.: Visionary speech:

Looking ahead to practical speechreading systems. In: Stork,

D.G., Hennecke, M.E. (eds.) Speechreading by Humans and

Machines, pp. 331–349. Springer, Berlin (1996)

9. Aleksic, P.S., Katsaggelos, A.K.: Comparison of low- and high-

level visual features for audio-visual continuous automatic

speech recognition. IEEE Proc. Int. Conf. Acoustics Speech

Signal Process. (Montreal, Canada) 5, 917–920 (2004)

10. Potamianos, G., Graf, H.P., Cosatto, E.: An image transform

approach for HMM based automatic lipreading. Paper presented

at the Proceedings of the International Conference on Image

Processing, vol. 1, pp. 173–177. Chicago, IL, 4–7 Oct. 1998

11. Wark, T., Sridharan, S., Chandran, V.: Robust speaker veriﬁca-

tion via fusion of speech and lip modalities. Proc. Int. Conf.

Acoustics Speech Signal Process. Phoenix 6, 3061–3064 (1999)

12. Aleksic, P.S., Katsaggelos, A.K.: An audio-visual person identiﬁ-

cation and veriﬁcation system using FAPs as visual features.

Paper presented at the Proceedings of Works. Multimedia User

Authentication, pp. 80–84. Santa Barbara, CA (2003)

13. Cootes, T.F., Edwards, G.J., Taylor, C.J.: Active appearance mod-

els. Paper presented at the Proceedings of European Conference

on Computer Vision, pp. 484–498. Freiburg, Germany (1998)

Lip Movement Recognition

907

Lip-Radiation Effect

The lip-radiation effect corresponds to changing the

volume velocity waveform at the lips to a speech pressure

signal in a free ﬁeld at a certain distance from the speaker.

▶ Speech Production

Liquid Crystal Displays (LCD)

Liquid crystal display (LCD) is a digital display that

uses liquid crystal cells whose reﬂectiv ity varies accord-

ing to the voltage applied to them. Liquid crystal is a

substance that ﬂows like liquid but maintains some of

the ordered structure characteristic of the crystals.

▶ Digitizing Table t

Liveness Detection

▶ Fingerprint Fake Detection

▶ Liveness Iris

Liveness Assurance in Face

Authentication

MIC HAEL WAGN E R ,GIR IJA CHETTY

School of Information Sciences, University of

Canberra, Australia

Synonyms

Face authentication; Face recognition; Face veriﬁcation

Definition

The process of verifying whether the face image pre-

sented to an authentication system is real (i.e., alive),

or whether it is reproduced or synthetic, and thus

fraudulent. When a face authentication system is to

recognize the face of a person by means of an electronic

camera and associated image recognition software, it is

important to be sure that the person seeking the au-

thentication actually presents his or her face to the

camera at the time and place of the authentication

request. The face is presented live as on live television

as distinct from a movie or cartoon programme. In

contrast, an impostor could try to present a photo-

graph, or a video recording, of a legitimate client to the

camera in order to be falsely authenticated by the

system as that client. That kind of threat to aut henti-

cation systems is known generally as

▶ replay attack.In

turn, liveness assurance uses a range of measures to

reduce the vulnerability of face authentication systems

to the threats of replay attack.

Introduction

The primary design objective for a face recognition

system is its ability to distinguish, as clearly as possible,

between different persons on the basis of their facial

images. As is described in detail in sections

▶ Face

Recognition, Overv iew, a good face recognition

system utilizes a suitable feature set, employs sophisti-

cated pattern recognition algorithms and sets decision

thresholds appropriate for the speciﬁc application

context. Nevertheless, current face recognition tech-

nology is vulnerable on several fronts: on one hand,

different persons like twins, especially identical twins,

other siblings, parents and children can have quite

similar facial appearance, while on the other hand the

same person can appear quite dissimilar at different

times owing to facial hair and hairstyle, make-up,

cosmetic surgery, eye glasses, and even just due to

their physical or emotional state. Figure 1 shows an

example of the faces of two identical twins being almost

indistinguishable and Fig . 2, in contrast, shows the

large difference between two images of the same person

who changed his facial app earance drastically. Face

recognition also has robustness issues unless environ-

ment variables such as lighting of the face and pose

with respect to the camera are controlled meticulously.

In addition, face authenticati on systems are vulner-

able to impostors who present a photograph of a

legitimate client to the system camera and may be

falsely accepted as that client by the system [1]. Gener-

ally, such an attack on a biometric authentication

908

Lip-Radiation Effect

system is known as a replay attack. Replay attacks can

be carried out by presenting a printed photograph to

the system camera or by holding a computer screen

showing a photo or video recording in front of the

camera. However, replay attacks are also possible by

injecting a suitable recorded signal or data ﬁle at other

points within the authentication system. All replay

attacks have in common that, at the time of the au-

thentication they play back to the system a signal that

was recorded from the client at an earlier time.

Liveness Assurance in Face Authentication. Figure 2 Dissimilarity of two facial images of the same person

(downloaded from http://news.bbc.co.uk).

Liveness Assurance in Face Authentication. Figure 1 Similarity of the facial images of two different persons

(downloaded from http://www.mary-kateandashley.com).

Liveness Assurance in Face Authentication

909