A 3D face and hand biometric system for robust user-friendly authentication

nauseatingcynicalSecurity

Feb 22, 2014 (3 years and 1 month ago)

71 views

A 3D face and hand biometric system for robust
user-friendly authentication
Filareti Tsalakanidou
a,
*
,Sotiris Malassiotis
a
,Michael G.Strintzis
a,b
a
Informatics and Telematics Institute,Centre for Research and Technology Hellas,1st km Thermi-Panorama Road,P.O.Box 60361,
Thermi 57001,Thessaloniki,Greece
b
Information Processing Laboratory,Electrical and Computer Engineering Department,Aristotle University of Thessaloniki,Thessaloniki 54124,Greece
Received 17 January 2007;received in revised form 22 June 2007
Available online 21 July 2007
Communicated by N.Pears
Abstract
A complete authentication system based on fusion of 3D face and hand biometrics is presented and evaluated in this paper.The sys-
tem relies on a low cost real-time sensor,which can simultaneously acquire a pair of depth and color images of the scene.By combining
2D and 3D facial and hand geometry features,we are able to provide highly reliable user authentication robust to appearance and envi-
ronmental variations.The design of the proposed system addresses two basic requirements of biometric technologies:dependable per-
formance under real-world conditions along with user convenience.Experimental evaluation on an extensive database recorded in a real
working environment demonstrates the superiority of the proposed multimodal scheme against unimodal classifiers in the presence of
numerous appearance and environmental variations,thus making the proposed system an ideal solution for a wide range of real-world
applications,from high-security to personalization of services and attendance control.
 2007 Elsevier B.V.All rights reserved.
Keywords:Multimodal biometrics;3D face;3D hand geometry;Color images;Classifier fusion;Score normalization
1.Introduction
The use of biometric technologies for personal authenti-
cation has been an ongoing topic of research,mainly due to
the increasing number of real-world applications requiring
reliable authentication of humans.Although biometric
authentication has been extensively used in high-security
systems,its use in everyday applications,such as access
control in buildings,tracking attendance or user personal-
ization,was until recently rather limited.This can be
mainly attributed to the fact that high performance biomet-
rics such as the iris or fingerprint do not enjoy user accep-
tance,since they are highly obtrusive,while other friendlier
biometrics,such as the face,may not be reliable under real-
world conditions.
Since user acceptance and convenience is a decisive
parameter in designing a biometric system suitable for a
wide range of applications,much effort has been devoted
to personal authentication based on images of the human
face and many algorithms to this end have been proposed
(Zhao et al.,2003).The majority of these techniques rely
on 2D gray-scale images.Recent public face recognition
benchmarks demonstrated that the performance of the best
2D face recognition algorithms is similar to that of finger-
print recognition,when frontal homogeneously illuminated
views are used,but degrades significantly for images sub-
ject to pose,illumination or facial expressions variations
(Phillips et al.,2003).To improve performance under these
conditions,the use of 3D facial images was proposed,
based on the fact that the 3D structure of the human face
can be highly discriminatory and is inherently insensitive to
0167-8655/$ - see front matter  2007 Elsevier B.V.All rights reserved.
doi:10.1016/j.patrec.2007.07.005
*
Corresponding author.Tel.:+30 2310 464160;fax:+30 2310 464164.
E-mail addresses:filareti@iti.gr (F.Tsalakanidou),malasiot@iti.gr
(S.Malassiotis),strintzi@eng.auth.gr (M.G.Strintzis).
www.elsevier.com/locate/patrec
Pattern Recognition Letters 28 (2007) 2238–2249
illumination variations and face pigment.Moreover,3D
information can significantly aid pose estimation (Bowyer
et al.,2006).
The earliest approach towards 3D face recognition is
based on computation of the surface curvature,which is
subsequently used for the localization of facial features
(Kim et al.,2001),the construction of Extended Gaussian
Images (Tanaka et al.,1998) or the extraction of local
Point Signatures (Chua et al.,2000).Although,high recog-
nition rates were reported for these techniques,in practice
curvature-based methods are very sensitive to image noise
and occlusions of the face.
The iterative closest point (ICP) algorithm,which is
widely used for registration of 3Dmodels,is also employed
by many 3D face recognition techniques (Medioni and
Waupotitsch,2003).The matching efficiency of the ICP is
improved by considering additional features,such as color
or curvature,or by using a weighted distance (Lu et al.,
2006).In (Chang et al.,2005),a recognition rate of 92%
is claimed in a database of 4000 images belonging to 477
individuals.
Appearance based methods like PCA or Fisherfaces
have also been proposed (Heseltine et al.,2004;Chang
et al.,2003).The main problemof these techniques is align-
ment of 3D images and pose variations,usually solved by
manual detection of facial features such as the eyes,nose
or mouth.In (Chang et al.,2003),a recognition rate of
93% is obtained in a test set comprised of 950 frontal
views of 200 people.Other approaches include the use of
deformable 3D models (Passalis et al.,2005) or techniques
based on the isometry assumption and use of geodesic
distances (Bronstein et al.,2005).In (Passalis et al.,
2005),a 90% recognition is reported in a database of
3500 images.
The combination of 2D and 3D data for face recogni-
tion was also investigated and significant improvements
have been reported (Tsalakanidou et al.,2005b;Chang
et al.,2003).Multimodal techniques usually rely on fusion
of scores obtained by unimodal classifiers,disregarding the
actual information conveyed by the two modalities.In an
attempt to exploit the main advantage of 3D face geome-
try,i.e.relative robustness to viewpoint and illumination
changes,a novel face authentication system integrating
2Dand 3Dimages was proposed by the authors in (Tsalak-
anidou et al.,2005a;Malassiotis and Strintzis,2005).The
combination of 2D and 3D facial data,along with treat-
ment of illumination and pose variations resulted in signif-
icant gains in terms of system performance,but still did
not cope with other sources of variation,e.g.presence of
facial expressions,change of appearance due to time and
occlusions.
To improve authentication accuracy,we propose the
combination of the above 3D face authentication system
with another user-friendly biometric modality:hand geom-
etry.Systems based on hand geometry are very popular
and are widely implemented for their ease of use,public
acceptance and integration capabilities.However,such sys-
tems present several shortcomings,the most important
being that hand geometry features are not highly distinc-
tive,thus preventing the use of this modality for high-secu-
rity applications.
Another argument against the use of hand recognition
systems for personal identification is their obtrusiveness
imposed by the use of special devices for placing the user’s
palm,such as platters with knobs or pegs (Sanchez-Reillo
et al.,2000).These devices,however,simplify the process
of feature extraction,which is usually based on the analysis
of image contours of hand views.Various features such as
the width of the fingers,length of the fingers and width of
the palm have been proposed.Satisfactory recognition
results are obtained (96% for recognition and less than
5% EER for authentication).Hand silhouettes have also
been employed (Jain and Duta,1999).
Document scanners have also been proposed for the
acquisition of hand images (Oden et al.,2003;Bulatov
et al.,2004).In (Bulatov et al.,2004) a feature-based
approach is used and an FRR close to 3% was achieved
for an FAR of 1% on a database of 70 people.In (Oden
et al.,2003),implicit polynomials are fitted on hand con-
tours and geometric invariants are subsequently computed
fromthese polynomials.An FRR of 1%for an FAR of 1%
on a small database (45 images) is reported.
A touch-free technique is proposed in (Zheng et al.,
2004).It is based on the localization of finger creases on
the front size of the palm.Zero FAR for an FRR of
2.8% is claimed on a small dataset.In (Woodard and
Flynn,2004),a 3Dscanner is used to acquire a range image
of the back of the hand placed on a dark surface.Recogni-
tion is based on curvature measurements of fingers.For
images acquired on the same week the recognition rate
was 99.4%,but dropped to 75% when probe and gallery
images were acquired with one week lapse.
A novel,unobtrusive hand recognition system based on
3D images of the user’s hand acquired using the low-cost
3D sensor of Tsalakanidou et al.(2005a) was proposed
by the authors in (Malassiotis et al.,2006).The experimen-
tal evaluation showed that the system’s performance was
comparable to that of other state-of-the-art hand geometry
systems.
In this paper,we propose the integration of the afore-
mentioned 3D face and hand geometry recognition algo-
rithms in a novel multimodal biometric system,which
combines high authentication reliability with user conve-
nience and acceptance.Multimodal authentication based
on face and hand geometry was also investigated elsewhere
(Jain et al.,2005;Ross and Jain,2003),but it was only
tested under ideal conditions regarding both appearance
variations and environmental conditions.Moreover,face
recognition was based on a CCD camera,while hand rec-
ognition relied on placement of the user’s hand on a special
platter with knobs or pegs constraining the hand’s posi-
tioning on the platter.2D images of the face and the hand
were used for extracting facial and hand geometry features,
thus making the system sensitive to illumination changes.
F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249 2239
Both methods obtain satisfactory results in a database of
50 people.In (Ross and Jain,2003),a FRR of 4.7% is
reported for a FAR of 1%,while in (Jain et al.,2005) a
FRR of 1.4% is claimed for a FAR of 0.1% when fusing
face,hand and fingerprint scores.
Our systemon the other hand,offers a low-cost solution
by using a single sensor and emphasizes on user conve-
nience by employing a non-contact approach.The addi-
tional use of 3D information for the hand and the face
offers increased robustness in illumination and pose varia-
tions,as well as face pigment,while it greatly simplifies
face/hand detection,localization and pose estimation.
The proposed system was evaluated with extensive experi-
ments in a large database comprised of images depicting
numerous variations in facial appearance and pose,as well
as hand posture.
The paper is organized as follows.In Section 2,we
describe the application scenario and acquisition set-up.
The face and hand geometry authentication algorithms
are briefly presented in Sections 3 and 4,respectively.
The normalization and fusion techniques deployed for
the experimental evaluation of the proposed multimodal
system are outlined in Section 5.The performance of the
system is extensively evaluated in Section 6,while Section
7 draws the conclusions and proposes ideas for future
work.
2.Application scenario and acquisition set-up
The proposed system relies on a novel sensor capable
of quasi-synchronous acquisition of color and 3D images.
The sensor consists of a low cost CCTV camera and a
standard multimedia projector,both embedded in a
mechanical construction.A color-coded light pattern is
projected on object surfaces.By measuring its deforma-
tion in the images captured by the camera,a 3D image
of the scene can be generated using an active triangulation
principle.Switching rapidly between the colored pattern
and a white light,a color image may be captured as well,
approximately synchronized with the depth image.A fra-
merate of 14 fps is achieved in a 3.2 GHz PC.More
details on the sensor can be found in (Tsalakanidou
et al.,2005a).
For the experimental evaluation of the proposed multi-
modal system,an access-control application scenario was
adopted and the sensor was optimized under this assump-
tion.The user stands in front of the camera at about one
meter distance in an effective working volume of 60 cm·
40 cm· 50 cm (width · height · depth) (see Fig.1).The
depth accuracy of the 3D sensor is 0.5 mm standard
deviation.The statistical depth error is quantified by
acquiring several depth maps of a given static scene,i.e.
by repeating the same measurement over and over,and
analyzing the scatter of depth measurements (Tsalakani-
dou et al.,2005a).The spatial resolution of the depth image
is almost equal to the color camera resolution for low-
bandwidth surfaces such as the face and the hand.The gen-
erated color and depth images have a resolution of
780 · 580 pixels.
For the proposed access-control application,the
authentication system consists of the 3D sensor,a monitor
and a standard PC,where the software runs.The user
stands in front of the sensor so that her face is inside the
effective working volume and she looks at the color cam-
era.The monitor shows a live video of the 2D image
sequence recorded by the sensor,while it also provides
direct feedback to the user and displays directions for cor-
rect placement of the face and hand inside the working
volume (see Section 3).After a pair of face images is
acquired,the user is asked to place her hand in front of
her face with the back of the hand facing the sensor,keep-
ing her fingers straight.This posture is most convenient for
all users and provides the best resolution of hand images
(see Fig.1).After capturing several pairs of hand images,
the system permits or prohibits user entrance based on
fusion of the matching scores provided by the face and
hand classifiers.
Although there are some limitations on the working
conditions under which the system operates,e.g.large
facial poses (more than ±40) or large finger bending are
not allowed,the proposed authentication system does not
impose any strict constraints on user movement;at the
same time special care is taken to provide a user-friendly
interface to the user.
For the evaluation of the performance of the proposed
multimodal system,we used the above set-up in an office
environment and recorded a face and hand database com-
prised of 50 subjects in 2 recording sessions.The test pop-
ulation contains 15 female and 35 male subjects between
19 and 36 years old.In each session,multiple pairs of
color and depth face and hand images depicting several
appearance variations were acquired for each subject.
First,the user was asked to look at the camera and make
several expressions (neutral face,smile,laugh).Face
images depicting illumination variations,pose variations
and images with/without wearing eyeglasses were also
acquired.
Then,the user was asked to place her hand in front of
her face with the back of the hand facing the camera.Sev-
eral pairs of hand images were acquired depicting varia-
tions of this posture.First,the user was asked to keep
her fingers straight,then relax and bend them a little.
She was also asked to wear a ring and vary the orientation
of her palm with respect to the sensor.The second session
was recorded 10 days after the first.The variations of the
recorded face and hand images are similar to those
encountered in real-world applications,thus allowing us
to evaluate system performance under real-world
conditions.
In each session,approximately 70 pairs of color
and depth face images were recorded for each user,and
about 50 pairs of hand images,thus resulting in a total
of 3500 face recordings and 2500 hand recordings per
session.
2240 F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249
3.3D face authentication system
In this section,we briefly describe the various steps of
the face authentication algorithm.A more detailed descrip-
tion can be found in (Tsalakanidou et al.,2005a;Malassi-
otis and Strintzis,2005).First,we detect the face in the
input 3Dimage using global moment descriptors and a pri-
ori knowledge of the geometry and relevant dimensions of
the head and other body parts.Then,we localize the face
position by using a knowledge-based 3D technique that
allows us to detect the nose with high accuracy.By fitting
a 3D line on the ridge of the nose and exploiting the inher-
ent bilateral symmetry of the face,we can reliably estimate
the face orientation (Malassiotis and Strintzis,2005).
The above technique has near real-time (about 10 fps)
performance and is used to provide the user with feedback
regarding the distance from the sensor and head pose.For
the latter,a low-resolution depth sub-image of the center of
the face is warped into canonical pose and projected on a
face-subspace created by a set of frontal 3D views,thus
permitting the computation of a measure of ‘‘face-ness’’
(Moghaddam et al.,1998),i.e.a measure of the extent to
which this sub-image resembles a frontal view of a human
face.Using the ‘‘face-ness’’ score,the computed distance
from the camera and the estimated face pose,the software
provides feedback to the user regarding the correct place-
ment of her head inside the working volume.This offers,
in addition,a means to cope with mimicry attacks (e.g.
putting a printed high resolution color image of the face
in front of the camera).
Using the pose information,input images are subse-
quently warped into canonical images depicting a frontal
view (see Fig.2).Then,an illumination compensation step
is applied.This consists of determining the illumination
direction and using this together with 3D depth map ren-
derings of the face to produce illumination compensated
images depicting frontal illumination (see Malassiotis and
Strintzis,2005 for details).
The resulting images are subsequently used as input to
the face classifier.Note,that the same normalization proce-
dure was applied to gallery images as well.For user authen-
tication,we employ a simpler version of the well-known
Probabilistic Matching algorithm (Moghaddam et al.,
1998),which is based on extra-personal eigenfaces.The
PM algorithm is applied to both color and depth images,
independently.
Fig.1.Color and depth face and hand images acquired using the 3Dsensor (see Section 2).In the range image,warmer colors correspond to points closer
to the camera,while black pixels correspond to undetermined depth values.(For interpretation of the references to colour in figure,the reader is referred to
the web version of this article.)
Fig.2.Pose compensation examples.The original pair of images and the resulting frontal views are shown.
F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249 2241
4.3D hand geometry
In this section,we briefly outline the algorithms
employed by the hand geometry recognition system.A
more detailed presentation can be found in (Malassiotis
et al.,2006).The segmentation of the hand from the body
is achieved by assuming that the hand does not move;
thus we may exploit the results of the face pose estimation
to form a plane that separates the face from the hand.
Then,segmentation of the hand from the forearm is
achieved by a 3D blob-based approach.Next,we localize
the center and radius of the palm using the estimated cen-
ter of the hand blob and the chamfer distance algorithm.
Finger detection relies on estimating the finger skeletons
from the location of the mid-points of circular finger seg-
ments,which result after drawing homocentric circular
arcs around the estimated palm center (see Fig.3a–c).
This last step may not be very accurate,since there is
missing depth information along finger boundaries.
Therefore,the color image is also used to localize finger
boundaries more accurately by means of a deformable
model.The fingertip of each finger j and a sequence of
K linear segments s
j
i
¼ fs
j
l
;s
j
r
g
i
;i ¼ 1;...;K,which are
perpendicular to the finger axis,are efficiently estimated
(see Fig.3d).
For each detected finger,we define two signature func-
tions,W
j
(x) and C
j
(x),parameterized by the 3D distance
x from the fingertip along the finger’s ridge.The first func-
tion corresponds to the width of the finger in 3D and it is
computed by fitting a 3Dline on the 3Dpoints correspond-
ing to each segment s
j
i
,projecting the end-points of this seg-
ment on this line and computing their Euclidian distance.
The second signature corresponds to the mean curvature
of the curve,which is defined by the 3Dpoints correspond-
ing to each finger segment.
Hand classification is based on 3Dgeometric features of
the user’s fingers extracted using these signature functions.
In total,12 width and 12 curvature measurements are cal-
culated for each of the four fingers of the user (the thumb is
excluded),resulting in a concatenated vector of 96 mea-
surements.The similarity score between two feature vectors
is based on their L
1
-norm.
5.Fusion of 3D face and hand modalities
Our implementation relies on fusion at the matching
score level,i.e.fusion of the matching scores computed
by different classifiers for the same enrolled user.Since
the output of different classifiers may not be in the same
range or follow the same statistical distribution,it is
essential that the matching scores provided by unimodal
classifiers be transformed in a common domain.This
procedure is known as score normalization.After the uni-
modal scores are normalized,several fusion schemes can be
used to generate a single multimodal score.In the follow-
ing,we briefly present the normalization and fusion tech-
niques deployed for the experimental evaluation of the
proposed multimodal biometric system.
5.1.Score normalization
For the normalization of different modality scores we
have tested five well-known normalization techniques:
min–max,z-score,median,tanh and quadric-line-quadric
normalization (Jain et al.,2005;Snelick et al.,2005).We
denote with s the original score and s
n
the score resulting
after normalization.The various normalization parameters
are computed using a bootstrap set of matching scores
S ={s
1
,s
2
,...,s
M
} produced by a unimodal classifier.
(1) Min–max (MM).This technique maps the original
scores into [0,1]:
s
n
¼
s min
max min
:ð1Þ
max and min are the maximum and minimum values
of S,i.e.the bounds of the scores produced by the
classifier.
(2) z-score (ZS).This method transforms the original
scores into a distribution with zero mean and unit
variance:
s
n
¼
s l
r
:ð2Þ
Fig.3.Finger detection and localization.(a) Original depth image,(b) circular arcs used to detect finger segments,(c) 2D lines fitted on skeleton points
and boundaries and (d) estimated finger boundary segments.
2242 F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249
l and r are the mean and standard deviation of
S.z-score normalization does not guarantee a com-
mon numerical range for the normalized scores of dif-
ferent classifiers.
(3) Median–MAD (MED).This method transforms the
raw scores in a way similar to z-score normalization:
s
n
¼
s median
MAD
:ð3Þ
median is the median of S and MAD=
median(js
i
medianj),s
i
2 S.
(4) Tanh.This technique maps the original scores into
[0,1]:
s
n
¼
1
2
tanh 0:05
s l
k
r
k
  
þ1
 
:ð4Þ
l
k
and r
k
are the mean value and standard deviation
of the genuine transaction score distribution of user
k.These are obtained using the Hampel robust esti-
mator (Jain et al.,2005).Specifically,let S ={s
i
,
i =1,...,N} be a set of N genuine transaction scores,
obtained from a bootstrap set of images of a given
subject.Then,a robust estimate of their mean value
l is obtained by means of the iterative re-weighted
least squares technique l
tþ1
¼
P
N
i¼1
w
t
i
s
i
,where w
t
i
are given by the Hampel influence function:
w
t
i
ðuÞ ¼
u;06juj <a;
a signðuÞ;a6juj <b;
a signðuÞ 
cjuj
cb
 
;b6juj <c;
0;juj Pc;
8
>
>
>
>
<
>
>
>
>
:
u ¼s
i
l
t
ð5Þ
and a,b,c are chosen as the 70th,85th and 95th per-
centiles of js lj.l
0
is initially set equal to the med-
ian of the scores.For the standard deviation r,the
value of a was shown to be a good estimate.
(5) Quadric-line-quadric (QLQ).QLQ normalization is
an adaptive normalization procedure proposed by
Snelick et al.(2005) aiming to decrease the overlap
of the genuine and impostor distributions,while still
mapping the scores in [0,1].First the scores are nor-
malized using the min–max normalization.Then the
following mapping function is applied to them:
s
n
¼
1
c
w
2
s
2
mm
;s
mm
6 c 
w
2
;
s
mm
;c 
w
2
< s
mm
6 c þ
w
2
;
c þ
w
2
þ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð1 c 
w
2
Þðs
mm
c 
w
2
Þ
p
;
otherwise;
8
>
>
>
>
<
>
>
>
>
:
ð6Þ
where c and w are the center and width of the overlap
zone of the min–max normalized scores s
mm
.
The parameters of min–max,z-score,median and QLQ
normalization can be calculated using two different
approaches:a global and a user-specific one.In the global
approach,the matching score set S includes genuine and
impostor scores produced by test images of all enrolled
users,i.e.the normalization parameters are the same for
all users.In the user-specific approach,these parameters
are computed for each user:the matching score set S
k
for
user k consists of genuine and impostor scores obtained
when the ID of user k is claimed.As experimental results
in Section 6.2 demonstrate,the second approach is more
efficient for user authentication.Note here,that the user-
specific approach requires the recording of several gallery
images during the user’s enrolment (e.g.more than 5).
These images are used for the computation of the user’s
normalization parameters.For the global approach on
the other hand,a single gallery image per user is enough.
5.2.Score fusion
For combining the scores of different classifiers we have
experimented with four well-known fusion techniques:sim-
ple sum,product,max-score and weighted sum (Kittler
et al.,1998;Snelick et al.,2005).We denote with s
m
i
the nor-
malized score computed by classifier mfor user i and with f
i
the fused score of user i.
(1) Simple sum (SS).This is the simplest fusion method.
The combined score is the sum of the normalized
scores of different classifiers:f
i
¼
P
M
m¼1
s
m
i
.
(2) Product (P).The combined score is the product of the
normalized scores of different classifiers:f
i
¼
Q
M
m¼1
s
m
i
.
(3) Max-score (MAS).The combined score is equal to
the maximum of the normalized scores produced by
different classifiers for the same user:f
i
¼ max
M
m¼1
s
m
i
.
(4) Weighted-sum (WS).The combined score is equal to
the weighted sumof the normalized scores of different
classifiers:f
i
¼
P
M
m¼1
w
m
s
m
i
,where
P
M
m¼1
w
m
¼ 1.The
weights w
m
are assigned to the individual classifiers
so that the total Equal Error Rate is minimized.This
way,more robust classifiers are assigned higher
weights,while less accurate classifiers are assigned
lower weights.Based on this fact,it is logical to con-
clude that the WS fusion technique is equal or supe-
rior to the SS method.
6.Experimental evaluation
The focus of the experimental evaluation was to investi-
gate the efficiency of the proposed multimodal system in
conditions similar to those encountered in real-world appli-
cations and show that it achieves superior performance
compared to the unimodal face and hand recognition sys-
tems,as well as to the combination of 2D and 3D facial
data proposed in (Tsalakanidou et al.,2005a;Malassiotis
et al.,2006).
Unlike other previous works that constructed a multi-
modal database by merging two or more separate dat-
abases,i.e.by combining face and hand images of
different individuals thus resulting to the creation of the
so-called chimeric users (Jain et al.,2005;Ross and Jain,
F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249 2243
2003;Snelick et al.,2005),we have recorded a new data-
base comprised of both modalities for a single set of users,
using the 3D sensor of Tsalakanidou et al.(2005a).In
order to make the experimental results comparable to pub-
lic evaluations of biometric systems,we followed well-
established practices regarding the evaluation concept
and database recording methodology (Mansfield and Way-
man,2002;Phillips et al.,2003).
The experiments were performed on the face and hand
database described in Section 2.The database consists of
two sessions.The first was used for training and the other
for testing.Three different modalities and their combina-
tions were tested using this dataset:face reflectance data
(color face images – FC),3D face geometry (depth face
images – FD),hand geometry (hand geometry measure-
ments extracted from color and depth images – H) and
combinations of the three,i.e.FC + FD,FC + H,
FD+ H and FC + FD+ H.
6.1.Training and testing
For the training of the face classifiers,two frontal neu-
tral views per subject were selected from the first recording
session and one feature vector was computed for each of
them.Note that in previous work (Tsalakanidou et al.,
2005a),the training set also contained images depicting
facial expressions.For the training of the hand classifier,
we used four pairs of hand images for each enrolled user,
depicting an ‘‘ideal’’ posture,i.e.wide open palm approxi-
mately parallel to the camera,with extended fingers,not
touching each other,no rings worn.For each image pair,
a feature vector containing 3D finger measurements was
extracted.
For a probe face image,computation of the matching
score consists in calculating a similarity score between this
image and the two gallery images corresponding to the
claimed ID and then selecting the maximum of the two
scores.For a hand probe image on the other hand,first
the L
1
distance between the extracted feature vector and
the four previously estimated gallery feature vectors is
computed,and then the matching score is set equal to the
minimum distance.
To improve the performance of the hand classifier,we
use a sequence of four input frames of the same individual
instead of just one probe image,and then combine the
scores computed on the four frames using a simple averag-
ing as proposed in (Malassiotis et al.,2006).For testing,4-
tuples are generated randomly from the set of probe hand
images belonging to the same individual.In total,the hand
test database consists of 24898 such 4-tuples,while the face
test database consists of 3457 face image pairs.
6.2.Experiments
For each probe (face and/or hand) image the identity of
all enrolled users is claimed in turn,thus resulting in 1 gen-
uine (user to whom this image actually belongs to) and 49
impostor claims per image.In total,3457 genuine and
169393 (3457 · 49) impostor matching scores are com-
puted from the face test database and 24898 genuine and
1220002 (24898 · 49) impostor scores from the hand
database.
For the evaluation of the proposed face + hand multi-
modal system,each pair of probe face images (depth +
color) is associated with five randomly selected pairs of
hand images belonging to the same person,thus resulting
in 17285 (3457 · 5) ‘‘pairs’’ of hand + face images,that is
17285 score vectors.A score vector is a triplet
hs
FC
,s
FD
,s
H
i (or pair hs
FD
,s
H
i,hs
FC
,s
H
i),where s
FC
,s
FD
and s
H
are the matching scores obtained by the FC,
FD and H classifiers,respectively.From each score vec-
tor,a multimodal score is computed in the following
way:first we normalize the matching scores provided by
the unimodal classifiers using one of the normalization
techniques of Section 5.1.Then,we consolidate the nor-
malized scores using one of the fusion methods described
in Section 5.2.
Using the above procedure,17285 genuine and 846965
(17285 · 49) impostor fusion scores were produced for
evaluating the performance of the FC + H,FD+ H and
FC + FD+ H authentication systems.For the evaluation
of the 2D+ 3D face authentication system,the 3457 image
pairs of the face database were used,resulting in 172850
(3457 · 50) fusion scores.
The performance of the unimodal and multimodal
authentication systems is presented in terms of EER(Equal
Error Rate) values and the Receiver Operating Character-
istics (ROC) curves.In every user or impostor transaction,
the computed similarity score is compared to an authenti-
cation threshold and accordingly the claimed identity is
either verified or rejected.By varying this threshold,pairs
of FAR (False Acceptance Rate) and FRR (False Rejec-
tion Rate) values are computed,defining points in the
ROC curve.
Following the paradigm of Tsalakanidou et al.
(2005b),we estimated the c =95% confidence interval of
the FAR and FRR values,based on the aforementioned
numbers of user and impostor transactions.For our
experiments,we get a confidence interval of 0.0464% for
an estimated FAR of 5% and 0.3249% for an estimated
FRR of 5%.The corresponding confidence intervals for
an estimated FAR and FRR of 2% are 0.0298% and
0.2087%,respectively.It is clear that these confidence
intervals are very low,thus allowing us to assume with
relative certainty that the EERs presented in this section
are reliable.
For determining the various normalization and fusion
parameters of different modalities,we used a large set of
genuine and impostor matching scores generated by a
bootstrap set of images belonging to the first database ses-
sion.In the following,a normalization technique will be
referred to as e.g.QLQ-I or QLQ-II,depending on whether
the global or user-specific approach is used for parameter
calculation.
2244 F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249
6.2.1.Unimodal classifiers
The matching scores produced by unimodal classifiers
were normalized prior to comparison with the authentica-
tion threshold using the normalization methods of Section
5.1.User-specific normalization is only tested,since global
normalization does not affect the relative ranking of scores.
Table 1 tabulates the obtained authentication/identification
rates.The ZS-II and MED-II normalization techniques
demonstrate the best performance for the face classifiers,
while QLQ-II seems to be a better solution for the hand
geometry classifier.The highest identification rates are
obtained when min–max I normalization is employed.
6.2.2.Multimodal classifiers
Table 2 summarizes the EER of the FC + FD+ H mul-
timodal system,along with its identification rate (IR),for
different normalization and fusion schemes.It is evident
that the multimodal classifier combining facial and hand
data exhibits better authentication rates (lower error rates)
than the best unimodal system or the combination of 2D
and 3D facial data.Obviously,combining facial features
with hand geometry features can be more efficient,since
these features are considerably less correlated than,for
example,2D and 3D facial data.
Next,we study the effect of each normalization tech-
nique in the performance of the proposed multimodal sys-
tem.Fig.4 shows the ROC curves of the system for
different but fixed fusion methods.The ROC curves of
the unimodal systems and the FC + FD system are also
shown for comparison.It is evident that QLQ-II normali-
zation outperforms all other normalization techniques for
all fusion methods.In all cases,user-specific techniques
behave more robustly than global ones,significantly
enhancing the system’s performance.All the previous also
stand for the FD+ H and FC + H multimodal systems.
For the FC + FD system on the other hand,ZS-II and
MED-II normalization methods produce the lowest EERs.
As far as the recognition performance is concerned,we
have observed that min–max I,QLQ-I and QLQ-II provide
the highest rank-1 identification rates.
ROCs in Fig.5 illustrate how each fusion technique
affects the system’s performance when different but fixed
normalization methods are used.It can be seen that the
weighted sum and simple sum fusion methods outperform
Table 1
EER values and rank-1 identification rates (%) of unimodal classifiers obtained using different normalization methods
Normalization method Modality
2D face (FC) 3D face (FD) Hand geometry (H)
Min–max I 6.46 (97.75) 7.55 (97.86) 5.44 (97.80)
Min–max II 6.25 (93.78) 7.95 (92.03) 6.31 (84.34)
z-score II 5.08 (97.51) 6.47 (96.52) 7.30 (83.1)
Median II 4.99 (96.70) 6.66 (95.95) 6.45 (78.42)
Tanh II 6.43 (91.46) 8.07 (92.24) 6.27 (75.21)
QLQ II 5.52 (96.44) 7.75 (93.69) 4.88 (94.6)
The lowest EERs and highest recognition rates achieved for each modality are indicated with bold typeface.
Table 2
EERvalues and rank-1 identification rates (%) of the 2D+3Dface authentication systemof Tsalakanidou et al.(2005a) (F) and the proposed face +hand
geometry authentication system (F +H) for possible combinations of various normalization/fusion methods
Normalization method Modality Fusion method
SS P MAS WS
MM-I F 4.79 (98.73) 4.89 (98.72) 7.45 (97.61) 4.69
*
(98.75
*
)
F +H 2.70 (99.86) 3.16 (99.39) 3.58 (100.0) 1.35
*
(100.0
*
)
ZS-I F 4.74 (98.75) 5.90 (90.63) 4.88 (98.19) 4.74
*
(98.75
*
)
F +H 3.10 (99.73) 3.77 (95.64) 3.90 (99.80) 1.45
*
(100.0
*
)
QLQ-I F 4.78 (98.73) 4.91 (98.73) 7.45 (97.63) 4.70
*
(98.75
*
)
F +H 2.70 (99.86) 3.17 (99.45) 3.13 (100.0) 1.28
*
(100.0
*
)
MM-II F 4.93 (98.19) 4.92
*
(97.90) 7.69 (92.51) 4.93 (98.23
*
)
F +H 2.47 (99.68) 2.99 (99.31) 3.13 (97.70) 1.44
*
(99.93
*
)
ZS-II F 4.26 (98.64) 5.24 (94.30) 4.48 (97.90) 4.21
*
(98.64
*
)
F +H 2.88 (99.68) 3.48 (96.12) 2.44 (98.73) 1.69
*
(99.78
*
)
QLQ-II F 4.40 (98.43) 4.42 (98.28) 6.97 (94.27) 4.39
*
(98.51
*
)
F +H 1.43 (100.0) 2.16 (99.76) 1.87 (99.50) 0.82
*
(100.0
*
)
Tanh F 5.34 (96.06) 5.43 (96.09) 5.9 (93.4) 5.24
*
(96.15
*
)
F +H 2.19 (98.28) 2.22 (98.28) 2.75 (95.17) 2.14
*
(98.28
*
)
The lowest EERs and highest recognition rates achieved for each column are indicated with bold typeface and for each row with the (
*
) symbol.
F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249 2245
all other methods.Weighted sum has afar the best perfor-
mance among all fusion techniques.The high authentica-
tion rates achieved when the WS technique is employed,
can be mainly ascribed to the effective weighing of the
scores obtained by the unimodal classifiers according to
their actual authentication performance.
Figs.4 and 5 clearly demonstrate the superiority of the
proposed multimodal system compared to the unimodal
systems and the 2D+ 3D face authentication system of
Tsalakanidou et al.(2005a),under different score normali-
zation and score fusion schemes.More specifically an EER
of 0.82%and an IR of 100%are reported when the QLQ-II
normalization method and the WS fusion technique are
employed by the multimodal classifier,while the best
EER reported for the FC + FD classifier is 4.22% (ZS-
II,WS).The best EERs for the FC,FD and H classifiers
are 5%,6.47% and 4.9%,respectively.Given that the test-
ing set consists of face and hand images depicting signifi-
cant variations (expressions,poses,hand postures,etc.),it
is easily perceived that the combination of face and hand
geometry features for personal authentication offers high
reliability and increased robustness.More specifically,the
system described in (Tsalakanidou et al.,2005a),although
fairly robust in facial pose and illumination changes,over-
comes the problem of facial expressions by using training
images depicting several expressions.Regarding the hand
recognition system of Malassiotis et al.(2006),the main
problem is finger bending and palm orientation,which
can be addressed in the same manner,i.e.by including such
variations in the training set.The proposed multimodal
scheme on the other hand,is trained using only frontal face
images and ideal hand postures.
We have also performed experiments using the FD+ H
and FC + H classifiers.These combinations demonstrate
performance which is very close to that of the FC + FD+ H
fusion.This suggests that we can reduce computational
costs by omitting the depth or color face classifier.
Next,we study our system’s performance for user recog-
nition applications (closed-set 1:N identification).The
cumulative recognition rates vs.rank are illustrated in
Fig.6,where a recognition rate of 100% is depicted for
the multimodal system,while the best recognition rate
reported for the FC + FD system is 98.75%.
Finally,we evaluate the systemunder the open-set 1:Nor
watch-list identification scenario (Grother,2004).Accord-
ing to this,there are probe images belonging to individuals
0
0.05
0.1
0.15
0
0.05
0.1
0.15
FAR
FRR
ROC diagram. Fusion method: Simple Sum (SS)
FC, ZS–II
FD, ZS–II
H, QLQ–II
FC+FD, ZS–II
FC+FD+H, TANH
FC+FD+H, QLQ–I
FC+FD+H, MM–II
FC+FD+H, ZS–II
FC+FD+H, QLQ–II
0
0.05
0.1
0.15
0
0.05
0.1
0.15
FAR
FRR
ROC diagram. Fusion method: Product (P)
FC, ZS

II
FD, ZS

II
H, QLQ

II
FC+FD, ZS

II
FC+FD+H, TANH
FC+FD+H, QLQ–I
FC+FD+H, MM

II
FC+FD+H, ZS

II
FC+FD+H, QLQ

II
0
0.05
0.1
0.15
0
0.05
0.1
0.15
FAR
FRR
ROC diagram. Fusion method: Max Score (MAS)
FC, ZS–II
FD, ZS–II
H, QLQ–II
FC+FD, ZS–II
FC+FD+H, TANH
FC+FD+H, QLQ–I
FC+FD+H, MM–II
FC+FD+H, ZS–II
FC+FD+H, QLQ–II
0
0.05
0.1
0.15
0
0.05
0.1
0.15
FAR
FRR
ROC diagram. Fusion method: Weighted Sum (WS)
FC, ZS–II
FD, ZS–II
H, QLQ–II
FC+FD, ZS–II
FC+FD+H, TANH
FC+FD+H, QLQ–I
FC+FD+H, MM–II
FC+FD+H, ZS–II
FC+FD+H, QLQ–II
a b
c d
Fig.4.ROC diagrams of the proposed multimodal system (FC + FD+H) obtained when using (a) simple sum,(b) product,(c) max-score and
(d) weighted sum fusion rule.In each figure,the ROC curves produced by the employment of different normalization techniques are drawn.The ROC
curves of the unimodal systems and the FC +FD system are also shown for comparison.
2246 F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249
not known to the system(impostors).In this case,the recog-
nition task involves (a) negative identification (rejection) of
people not belonging to the gallery and (b) correct identifi-
cation of people that make up the watch list.
System performance is measured over the impostor and
watch-list populations using the correct rejection rate and
the correct identification rate,respectively.The correct
rejection rate is defined as the fraction of impostor probe
images for whom the computed similarity measures,
derived from their comparison to enrolled users,are below
threshold t.The correct identification rate is the fraction of
enrolled user test images,for whom the similarity measure
corresponding to their own gallery image (same id) is
greater than t.The accuracy of the system is defined as
(average number of correct rejections + average number
of correct recognitions)/N,in a probe set of N images.
In Fig.7,the average correct recognition rate is plotted
against the average correct rejection rate for varying thresh-
old values.These rates were obtained over 200 randomized
runs of the algorithm.In each run,the watch-list subjects
are randomly selected from the 50 individuals of the data-
base.It is clearly seen that the proposed multimodal system
leads to increased identification accuracy.More specifically,
for a watch list of 15 subjects,the accuracy of the
FC + FD+ H system is 98.9%,while the accuracy of the
FC + FD system is 95.6%.For a watch list of 25 subjects,
the corresponding values are 98.4%and 93.2%,respectively.
0
0.05
0.1
0.15
0.2
0
0.05
0.1
0.15
0.2
FAR
FRR
ROC diagram. Normalization method: Min–max–II
FC
FD
H
FC+FD, WS
FC+FD+H, SS
FC+FD+H, P
FC+FD+H, MAS
FC+FD+H, WS
0
0.05
0.1
0.15
0.2
0
0.05
0.1
0.15
0.2
FAR
FRR
ROC diagram. Normalization method: Tanh
FC
FD
H
FC+FD, WS
FC+FD+H, SS
FC+FD+H, P
FC+FD+H, MAS
FC+FD+H, WS
0
0.05
0.1
0.15
0.2
0
0.05
0.1
0.15
0.2
FAR
FRR
ROC diagram. Normalization method: QLQ–I
FC
FD
H
FC+FD, WS
FC+FD+H, SS
FC+FD+H, P
FC+FD+H, MAS
FC+FD+H, WS
0
0.05
0.1
0.15
0.2
0
0.05
0.1
0.15
0.2
FAR
FRR
ROC diagram. Normalization method: QLQ–II
FC
FD
H
FC+FD, WS
FC+FD+H, SS
FC+FD+H, P
FC+FD+H, MAS
FC+FD+H, WS
a b
c d
Fig.5.ROC diagrams of the proposed multimodal system (FC +FD+ H) obtained when using (a) MM-II,(b) Tanh,(c) QLQ-I and (d) QLQ-II score
normalization methods.In each figure,the ROC curves produced by the employment of different score fusion methods are drawn.The ROC curves of the
unimodal systems and the FC +FD system are also shown for comparison.
1
10
20
30
40
50
0.975
0.98
0.985
0.99
0.995
1
Rank
Cumulative Recognition Rate
Cumulative Recognition Rates vs. Rank
FC
FD
H
FC+FD, MM–I, WS
FC+FD+H, QLQ–II, WS
Fig.6.Cumulative recognition rates vs.rank for the unimodal systems,
the 2D+3D face recognition system of Tsalakanidou et al.(2005a) and
the proposed face +hand geometry recognition system.
F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249 2247
7.Conclusions
An end-to-end authentication system based on integra-
tion of 2D and 3D face and hand biometrics was presented
in this paper.Unlike other state-of-the art techniques,the
proposed multimodal scheme does not impose any strict
constraints on face and hand placement or on background
structure,while at the same time assisting the user through-
out the authentication process.Thus,it is suitable for
everyday applications,where unobtrusiveness and user
convenience are top priorities.At the same time,experi-
mental results obtained by extensively evaluating the sys-
tem’s performance in a large image database including
numerous appearance and environmental variations,
clearly demonstrate the system’s robustness to such varia-
tions and its superiority against unimodal classifiers and
the combination of 2D and 3D facial data.Specifically,
using the QLQ-II normalization method and the weighted
sumfusion technique,an EERequal to 0.82%and a rank-1
identification rate equal to 100% were reported for a test-
set comprised of 17285 pairs of face and hand images
depicting significant variations.Conclusively,the proposed
multimodal authentication scheme effectively combines
high dependability with user-friendliness to ensure suitabil-
ity for a wide range of real-world applications,from high
security to personalization of services.
Future work will involve investigation of alternative
fusion techniques and particularly fusion of face and hand
features at an earlier stage,as well as full system deploy-
ment for testing in real scale applications.
Acknowledgement
This work was supported by Research Project PASION
FP6-027654 (‘‘Psychologically Augmented Social Interac-
tion over Networks’’) under the Information Society Tech-
nologies (IST) priority of the 6th Framework Programme
of the European Community.
References
Bowyer,K.W.,Chang,K.,Flynn,P.J.,2006.A survey of approaches and
challenges in 3Dand multi-modal 3D+2Dface recognition.Comput.
Vision Image Understanding 101 (1),1–15.
Bronstein,A.M.,Bronstein,M.M.,Kimmel,R.,2005.Three-dimensional
face recognition.Internat.J.Comput.Vision 64 (1),5–30.
Bulatov,Y.,Jambawalikar,S.,Kumar,P.,Sethia,S.,2004.Hand
recognition using geometric classifiers.In:Proc.1st Internat.Conf.
on Biometric Authentication (ICBA),Hong Kong,July,pp.753–
759.
Chang,K.I.,Bowyer,K.W.,Flynn,P.J.,2003.Face recognition using 2D
and 3D facial data.In:Proc.ACM Workshop on Multimodal User
Authentication,Santa Barbara,California,December,pp.25–32.
Chang,K.I.,Bowyer,K.W.,Flynn,P.J.,2005.Adaptive rigid multi-region
selection for handling expression variation in 3D face recognition.In:
Proc.IEEE Computer Society Conf.on Computer Vision and Pattern
Recognition (CVPR’05),vol.3,June,pp.157– 166.
Chua,C.-S.,Han,F.,Ho,Y.-K.,2000.3D human face recognition using
point signature.In:Proc.4th IEEE Internat.Conf.on Automatic Face
and Gesture Recognition (FG00),March,pp.233–238.
Grother,P.,2004.Face recognition vendor test 2002.Supplemental
Report NISTIR 7083,February,National Institute of Standards and
Technology,US.
Heseltine,T.,Pears,N.,Austin,J.,2004.Three-dimensional face
recognition:An eigensurface approach.In:Proc.Internat.Conf.on
Image Processing (ICIP’04),vol.2,October,pp.1421–1424.
Jain,A.,Duta,N.,1999.Deformable matching of hand shapes for
verification.In:Proc.Internat.Conf.on Image Processing (ICIP),vol.
2,Kobe,Japan,October,pp.857–861.
Jain,A.,Nandakumar,K.,Ross,A.,2005.Score normalization in
multimodal biometric systems.Pattern Recognition 38 (12),2270–
2285.
Kim,T.K.,Kee,S.C.,Kim,S.R.,2001.Real-time normalization and
feature extraction of 3D face data using curvature characteristics.In:
Proc.10th IEEE Internat.Workshop on Robot and Human Interac-
tive Communications,September,pp.74–79.
Kittler,J.,Hatef,M.,Duin,R.P.,Matas,J.,1998.On combining
classifiers.IEEE Trans.Pattern Anal.Machine Intell.20 (3),226–
239.
Lu,X.,Colbry,D.,Jain,A.K.,2006.Matching 2.5D face scans to 3D
models.IEEE Trans.Pattern Anal.Machine Intell.28 (1),31–43.
Malassiotis,S.,Strintzis,M.G.,2005.Robust face recognition using 2D
and 3D data:Pose and illumination compensation.Pattern Recogni-
tion 38 (12),2537–2548.
0.4
0.5
0.6
0.7
0.8
0.9
1
0.4
0.5
0.6
0.7
0.8
0.9
1
Correct Reco
g
nition Rate
Correct Rejection Rate
Open–set 1:N identification performance (Watch list size = 15 subjects)
FC (ZS–II)
FD (ZS–II)
H (QLQ–II)
FC+FD (ZS–II, WS)
FC+FD+H (QLQ–II, WS)
0.4
0.5
0.6
0.7
0.8
0.9
1
0.4
0.5
0.6
0.7
0.8
0.9
1
Correct Reco
g
nition Rate
Correct Rejection Rate
Open–set 1:N identification performance (Watch list size = 25 subjects)
FC (ZS–II)
FD (ZS–II)
H (QLQ–II)
FC+FD (ZS–II, WS)
FC+FD+H (QLQ–II, WS)
a b
Fig.7.Correct recognition rate vs.correct rejection rate for a watch list of (a) 15 and (b) 25 subjects.The performance of the proposed face + hand
identification system is compared against the performance of the unimodal systems and the 2D+3D face recognition system of Tsalakanidou et al.
(2005a).
2248 F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249
Malassiotis,S.,Aifanti,N.,Strintzis,M.G.,2006.Personal authentication
using 3D finger geometry.IEEE Trans.Inf.Forensics Security 1 (1),
12–21.
Mansfield,A.J.,Wayman,J.L.,2002.Best practices in testing and
reporting performance of biometric devices v.2.01.Evaluation report,
August,Centre for Mathematics and Scientific Computing,National
Physical Laboratory,UK.
Medioni,G.,Waupotitsch,R.,2003.Face modeling and recognition in
3D.In:Proc.IEEE Internat.Workshop on Analysis and Modeling of
Faces and Gestures (AMFG 2003),October,pp.232–233.
Moghaddam,B.,Wahid,W.,Pentland,A.,1998.Beyond eigenfaces:
Probabilistic matching for face recognition.In:Proc.Internat.Conf.
on Automatic Face and Gesture Recognition (FGR’98),vol.1,Nara,
Japan,April,pp.30–35.
Oden,C.,Ercil,A.,Buke,B.,2003.Hand recognition using implicit
polynomials and geometric features.Pattern Recognition Lett.24 (13),
2145–2152.
Passalis,G.,Kakadiaris,I.A.,Toderici,G.,Murtuza,N.,2005.Evaluation
of 3D face recognition in the presence of facial expressions:An
annotated deformable model approach.In:Proc.IEEE Computer
Society Conf.on Computer Vision and Pattern Recognition
(CVPR’05),vol.3,June,pp.171–171.
Phillips,P.J.,Grother,P.,Micheals,R.J.,Blackburn,D.M.,Tabassi,E.,
Bone,J.M.,2003.Face recognition vendor test 2002.Evaluation
report,Defense Advanced Research Projects Agency and National
Institute of Justice.
Ross,A.,Jain,A.,2003.Information fusion in biometrics.Pattern
Recognition Lett.24 (13),2115–2125.
Sanchez-Reillo,R.,Sanchez-Avila,C.,Gonzalez-Marcos,A.,2000.
Biometric identification through hand geometry measurements.IEEE
Trans.Pattern Anal.Machine Intell.22 (10),1168–1171.
Snelick,R.,Uludag,U.,Mink,A.,Indovina,M.,Jain,A.,2005.Large-
scale evaluation of multimodal biometric authentication using state-of-
the-art systems.IEEE Trans.Pattern Anal.Machine Intell.27 (3),
450–455.
Tanaka,H.T.,Ikeda,M.,Chiaki,H.,1998.Curvature-based face surface
recognition using spherical correlation.Principal directions for curved
object recognition.In:Proc.3rd IEEE Internat.Conf.on Automatic
Face and Gesture Recognition (FGR’98),April,pp.372–377.
Tsalakanidou,F.,Forster,F.,Malassiotis,S.,Strintzis,M.G.,2005a.
Real-time acquisition of depth and color images using structured light
and its application to 3D face recognition.Real-Time Imaging 11 (5–
6),358–369,Special Issue on Multi-Dimensional Image Processing.
Tsalakanidou,F.,Malassiotis,S.,Strintzis,M.G.,2005b.Face localiza-
tion and authentication using color and depth images.IEEE Trans.
Image Process.14 (2),152–168.
Woodard,D.L.,Flynn,P.J.,2004.3Dfinger biometrics.In:Proc.8th Eur.
Conf.on Computer Vision (ECCV 2004),Biometric Authentication
Workshop (BioAW),Prague,Czech Republic,May,pp.238–247.
Zhao,W.,Chellappa,R.,Rosenfeld,A.,Phillips,P.J.,2003.Face
recognition:A literature survey.ACMComput.Surv.35 (4),399–459.
Zheng,G.,Wang,C.,Boult,T.E.,2004.Personal identification by cross
ratios of finger features.In:Proc.1st Internat.Conf.on Pattern
Recognition (ICPR),Workshop on Biometrics,August,Cambridge,
UK.
F.Tsalakanidou et al./Pattern Recognition Letters 28 (2007) 2238–2249 2249