INTER
N
A
TIO
N
AL
CONFERENCE
ON
ELECT
R
ONICS
AND
OIL

ICEO2013
5

6
MARCH
2013,
O
U
ARGLA,
ALGERIA
1
P
On

line
Finge
r

Knuckle

Print
Identification
Using
Gaussian
Mixture
Models
&
Discrete
Cosine
T
ransform
Abdallah
Meraoumi
a
1
,Salim
Chitroub
2
and
Ahmed
Bouridan
e
3
,
4
1
Kasdi
Merbah
un
i
v
ersity
of
Oua
r
gla,
Electrical
engineering
Laboratory
Sciences
and
T
echn
ology
and
Mater
Sciences
F
acult
y
,
Oua
r
gla,
30000,
Algeria
2
Signal
and
Image
Processing
Laboratory,
Electronics
and
Computer
Science
F
acult
y
,
USTHB.
P
.O.
box
32,
El
Alia,
Bab
Ezzoua
r
,
16111,
Algiers,
Algeria
3
School
of
Computing,
Engineering
and
Informati
on
Sciences,
Northumbria
Un
i
v
ersit
y
,
P
andon
Building,
N
e
wcastle
upon
T
yne,
UK.
4
Department
of
Computer
Science,
King
Saudi
Un
i
v
ersit
y
,
P
.O.
Box
2454,
Riyadh,
11451,
Saudi
Arabia
Email:
Ame
r
aoumia@gmail.com,
S
c
hit
r
oub@hotmail.com,
Ahmed.Bouridane@northumbria.ac.uk,
ABouridan
e
.c@ksu.edu.sa
Abstrac
t
—
Biometric system
has
been
act
iv
e
ly
eme
r
ging
in
v
arious
industries
f
or
the
past
few
y
ears,
and
it
is
continuing to
r
oll
to
p
r
o
vide higher
security featu
r
es
f
or
access
cont
r
ol
system.
In
the
r
ecent
y
ears,
hand
based
biometrics
is
extens
iv
ely
used
f
or
personal
r
ecognition. In
this
pape
r
,
we
p
r
opose
an
efficient
online
personal
identification system
based
on
Finge
r

Knuckle

Print
(FKP)
using
the
Gaussian
Mixtu
r
e
Model
(
GMM
)
and
t
w
o

dimensional Block
Based
Disc
r
ete
Cosine
T
rans
f
orm
(
2D

BDCT
).
In
this
stud
y
,
a
segmented
FKP
is
firstly
d
i
vided
into
non

ov
erlapping
and
equal

sized
blocks,
and
then,
applies
the
2D

BDCT
ov
er
each
block. By
using
zigzag scan
order each
trans
f
orm block
is
r
eorde
r
ed to
p
r
oduce
the
featu
r
e
v
ecto
r
.
Subsequentl
y
,
we
use
the
GMM
f
or
modeling
the
featu
r
e
v
ector
of
each
FK
P
.
Finall
y
,
Log

li
k
elihood sco
r
es
a
r
e
used
f
or
FKP
matchin
g
.
Experimental
r
esults
sh
o
w
that
our
p
r
oposed
method
yields
the
best
per
f
ormance
f
or
identifying FKPs
and
it
is
able
to
p
r
o
vide
an
excellent
identification
rate
and
p
r
o
vide
mo
r
e
securit
y
.
Index
T
erms
—
Biometrics,
identification,
Finge
r

Knuckle

Print,
2D

BDCT,
GMM,
Data
fusion.
I
.
I
NT
R
ODUCTIO
N
ERSO
N
AL
identification
plays
a
critical
role
in
our
soci

et
y
.
T
raditional
kn
o
wledge
based
or
to
k
en

based
personal
identification
systems
are
time

consuming,
ine
f
ficient and
e
x

pens
i
v
e.
Biometrics
o
f
fers
a
natural
and
reliable
solution
to
the
problem
of
identity
determination
by
recognizing
ind
i
viduals
based
on
some
characteristics
that
are
inherent
to
the
person
[1].
Biometrics
is
a
study
of
metho
ds
for
uniquely
recogniz

ing
ind
i
viduals
based
on
one
or
more
intrinsic
p
h
ysical
or
beh
a
vioral
traits,
including
the
e
xtens
i
v
ely
studied
fingerprint,
iris,
speech,
hand
geometr
y
,
and
palmprint.
One
of
the
most
popular
biometric
systems
is
based
on
the
han
d
due
to
its
ease
of
use.
Recentl
y
, a
n
ov
el
hand

based biometric
feature,
finge
r

knuckle

print
(FKP),
has
attracted
an
increasing amount
of
at

tention
[2].
The
t
e
xture
pattern
produced
by
the
finger
knuckle
bending
is
highly
unique
and
ma
k
es
the
sur
f
ace
a
distinct
i
v
e
biometric
identifie
r
.
Li
k
e
a
n
y
other
biometric
identifiers,
FKPs
are
beli
e
v
ed to
h
a
v
e
the
critical
properties
of
un
i
v
ersalit
y
,
uniqueness,
and
permanence
for
personal
recognition
[3].
An
important
issue
in
FKP
identification is
to
e
xtr
act
FKP
features
that
can
discriminate an
ind
i
vidual
from
the
othe
r
.
Based
on
t
e
xture
analysis,
our
biometric
identification
system
used
the
2D

BDCT
for
features
e
xtracted
from
FKP
images.
In
this
method,
a
FKP
is
firstly d
i
vided
into
non

ov
erlapping
and
equalized blocks,
and
then,
applies
the
2D
discrete
cosine
transform
ov
er
each
block.
By
using
zigzag
scan
orde
r
,
each
transform
block
is
reordered
to produce
the
feature
v
ector
and
then
concatenated all
v
ectors
for
produce
an
obser
v
ation
v
ecto
r
.
Subsequ
entl
y
,
we use the GMM
for modeling
this
v
ector
(for
each
FKP).
Finall
y
,
log

li
k
elihood scores
are
used
for
matching.
In
this
w
ork,
a
series
of
e
xperiments were
carried
out
using a
FKP
database.
T
o
ev
aluate
the
e
f
ficien
c
y
of
this
technique,
th
e
e
xperiments were
designed
as
foll
o
w:
the
performances
under
di
f
ferent
finger types
were
compared
to
each
othe
r
,
in
order
to
determine
the
best
finger
type
at
which
the
FKP
identification
system
performs.
H
o
w
e
v
e
r
,
because
our
database
contains
FKPs
from
f
our
types
of
fingers, an
ideal
FKP
identification system
should
be
based
on
the
fusion
of
these
fingers
at
di
f
ferent
fusion
l
e
v
els.
The
rest
of
the
paper
is
o
r
g
anized as
foll
o
ws:
The
proposed
FKP
recognition scheme
is
presented
in
section
2.
Section
3
g
i
v
e
s
a
brief
description of
the
method
used
for
e
xtracting the
R
e
gion
Of
Interest
(
R
OI
).
The
feature
e
xtraction
and
modeling
process,
including
an
ov
ervi
e
w
of
the
t
w
o

dimensional block
based
discrete
cosine
transform and
the
g
aussian
mixture
model,
is
pres
ented
in
section
4.
A
sections 5
is
d
e
v
oted
to
describe
the
ev
aluation
and
normalization method.
The
obtained results,
prior
to
fusion
and
after
fusion,
are
ev
aluated
and
commented in
section
6.
Finall
y
,
conclusions and
future
w
ork
are
g
i
v
en
in
sect
ion
7.
II
.
S
YSTE
M
O
VE
R
VIE
W
Fig.
1
illustrates the
v
arious
modules
of
our
proposed uni

modal FKP
identification
system
(single finger).
The
pro

posed
system
consists
of
preprocessing,
feature
e
xtraction
and
modeling, matching
and
decision
stages.
T
o
enroll
into
the
system
database,
the
user
has to
pr
o
vide
a
set of
training
FKP
images.
T
ypicall
y
,
an
obser
v
ation
v
ector
is
e
xtracted
from
each
finger
which
describes
certain
characteristics
of
the
FKP
images
using
Discrete
Cosine
T
ransform
(DCT
)
technique
and
modeling using
g
aussian
mixture
model.
Finall
y
,
the
models
parameters are
stored
as
references
models.
F
or
identification,
the
same
obser
v
ation
v
ectors
are
e
xtracted
from
the
test
FKP
images and
the
log

li
k
elihood
is
computed
using
all
of
mo
dels
INTER
N
A
TIO
N
AL
CONFERENCE
ON
ELECT
R
ONICS
AND
OIL

ICEO2013
5

6
MARCH
2013,
O
U
ARGLA,
ALGERIA
2
½
GMM

Modeling
D
AT
A
B
ASE
Preprocessing
Feature
v
ector
Matching
Module
Decision
Fig.
1.
Block

diagram
of
the
FKP
identification
system
based
on
the
g
aussian
mixtu
re
model.
Image
original
Determine
the
X

axis
R
OI
Localisation
R
OI
Extraction
Fig.
2.
R
OI
e
xtraction
process.
(
a
)
Image
original;
(
b
)
X

axis
of
the
coordinate
system;
(
c
)
R
OI
coordinate
system
and
(
d
)
R
e
gion
of
interest
(
R
OI)
references. Our
database
contains
FKPs
from
four
types
of
fingers,
for
this
raison, each
FKP
modalities
are
used
as
inputs
of the
matcher
modules
(sub

system).
F
or
the
multi

modal
system,
each
sub

system compute
its
o
wn
matc
hing
score
and
these
ind
i
vidual
scores
are
finally combined
into
a
total
score
(using
fusion
at
the
matching
score
l
e
v
el),
which
is
used
by
the
and
a
1D

DCT
on
the
r
o
ws.
G
i
v
en
an
image
f
,
where
H
,
W
represent
their
size,
the
DCT
coe
f
ficients
of
the
spati
al
block
are
then
determined
by
the
foll
o
wing
formula:
M
−
1
M
−
1
F
ij
(
u,
v
)
=
C
(
v
)
C
(
u
)
X
X
f
ij
(
n,
m
)
ψ
(
n,
m,
u,
v
)
(1)
m
=0
n
=0
decision
module.
W
e
h
a
v
e
also
tried
the
fusion
at
the
decision
l
e
v
el
to
choose
the
best
one
for
FKPs
classification.
ψ
(
n,
m,
u,
v
)
=
cos
h
(2
n
+
1)
uπ
i
2
M
cos
h
(2
m
+
1)
v
π
i
2
M
(2)
u,
v
=
0
,
1
,
·
·
·
,
M
−
1,
i
=
1
,
·
·
·
,
η
1
,
j
=
1
,
·
·
·
,
η
2
with
III
.
R
EGIO
N
O
F
INTERES
T
EXTR
A
CTIO
N
After
the
image
is
captured,
it
is
pre

processed
to
obtain
only
the
area
information
of
the
FK
P
.
The
detailed
steps
η
1
=
H
W
M
,
η
2
=
M
and
F
ij
(
u,
v
)
are
the
DCT
coe
f
ficients
of
the
B
ij
block,
f
ij
(
n,
m
)
is
the
luminance
v
alue
of
the
pi
x
el
(
n,
m
)
of
the
B
ij
block,
and
for
pre

processing
process
are
as
foll
o
ws
[4]:
First,
apply
a
Gaussian
smoothing
ope
ration
to
the
original
image.
Second,
C
(
u
)
=
1
√
2
if
u
=
0
(3)
determine
the
X

axis
of
the
coordinate
system
fitted from
the
bottom
boundary
of
the
finger; the
bottom
boundary
of
the
finger
can
be
easily
e
xtracted
by
a
Can
n
y
edge
de
tecto
r
.
Third,
determine
the
Y

axis
of
the
coordinate
system
by
applying
a
Can
n
y
edge
detector
on
the
cropped
sub

image
e
xtracted
from
the
image
original
base
on
X

axis,
then
find the
co
n
ve
x
direction coding
scheme.
Finall
y
,
e
xtract
the
R
OI
coordinate
syst
em,
where
the
rectangle indicates the
area
of
the
R
OI
that
will
be
e
xtracted.
The
pre

processing
steps
are
sh
o
wn
in
Fig.
2
.
I
V
.
F
E
A
TUR
E
EXTR
A
CTIO
N
AN
D
M
ODELIN
G
A.
2D
Blo
c
k
based
disc
r
ete
cosine
t
r
ansform
Discrete
cosine
transform
is
a
p
o
werful
transform
to
e
xtract
1
if
u
=
0
After
transformation process,
if
M=8
,
there
will
be
64
DCT
coe
f
ficients contained
within
each
transformed
block,
where
the
coe
f
ficient
at
the
top

left
is
called
DC
h
F
ij
(0
,
0)
i
coe
f
fi

cient
and
the
rest
is
call
ed
A
C
coe
f
ficients.
B.
Observation
vector
The
block

based approach
partitions
the
input
image,
with
size
H
×
W
,
when
H
=
220
and
W
=
110,
into
small
non

ov
erlapped
blocks;
each
of
them
is
then
mapped
into
a
block
of
coe
f
ficients via
the
2D

DCT.
Most
popular
block
size
is
commonly
set
to
M
×
M
with
M=8
.
The
number
of
blocks
e
xtracted
from
each
FKP
image
equals
to:
proper
features
for
FKP
identification.
The
DCT
is
the
most
widely
used
transform
in
image
processing
algorithms,
such
η
=
b
η
1
c
∗
b
η
2
c
=
b
220
8
c
∗
b
110
8
c
=
27
∗
13
=
351
blocks
(4)
as
image/video compression and
pattern
recognition.
Its
pop

ularity
is
due
mainly
to
the
f
act
that
it
achi
e
v
es a
good
data
compaction,
that
is,
it
concentrates
the
information
content
in
a
rela
t
i
v
ely
f
e
w
transform
coe
f
ficients
[5].
In
the
2D

BDCT
formulation,
the
input
image
is
first
d
i
vided
into
η
1
×
η
2
blocks,
and
the
2D

DCT
of
each
block
is
determined.
The
2D

DCT
can
be
obtained
by
performing
a
1D

DCT
on
the
columns
Then,
we
form
a
feature
v
ector
from
the
2D

DCT
coe
f
ficients
of
each
image
block
(see
Fig.
3
).
The
2D

DCT
concentrates
the
information content
in
a
relat
i
v
ely
f
e
w
transform
coe
f
fi

cients
top

left
zone
of
block,
for
this,
the
coe
f
ficients, where
the
information
is
concentrated,
ten
d
to
be
grouped
together
at
the
start
of
the
reordered
arra
y
, Thus, a
suitable
scan
order
is
a
zigzag
starting
from
the
DC
(top

left)
coe
f
ficient
[6].
Starting
INTER
N
A
TIO
N
AL
CONFERENCE
ON
ELECT
R
ONICS
AND
OIL

ICEO2013
5

6
MARCH
2013,
O
U
ARGLA,
ALGERIA
3


j
=1
Fig.
3.
Obser
v
ation
v
ector
e
xtraction.
(a)
blocks
e
xtraction,
(b)
block
feature
e
xtract
ion
and
(c)
Obser
v
ation
v
ecto
r
.
with
the
DC
coe
f
ficient,
each
coe
f
ficient
is
copied
into
a
one

dimensional arra
y
.
So,
each
block
can
be
represented
by
a
v
ector
of
coe
f
ficients:
O
ij
=
[
F
ij
(0
,
0)
F
ij
(0
,
1)
F
ij
(1
,
0)
·
·
·
·
·
·
F
ij
(
U
,
V
)]
T
(5)
U,
V
are
chosen
as
well
as
the
identification
rate
w
as
max

imum.
Thus,
U,
V
∈
[0
·
·
7]
and
the
size
of
O
ij
is
τ
with
τ
∈
[1
·
·
64].
Finall
y
,
the
results
o
ij
of
a
blocks
image
are
combined
in
the
single
template
as
foll
o
ws:
V
O
bs
=
[
O
11
O
12
O
13
O
14
·
·
·
O
η
1
η
2
]
(6)
Where
the
size
of
result
ing
obser
v
ation
v
ector
is
[
τ
η
]
.
C.
Gaussian
mixtu
r
e
model
Gaussian
mixture
model
is
pattern
recognition
technique
that
uses
an
approach of
the
statistical methods
[7].
The
obser
v
ation
v
ector
of
each
class
measurement
can
be
described
by
normal
distri
b
ution,
also
called
Gaussian
distri
b
ution.
Each
class
measurement
may
be
then
defined by
t
w
o
parameters:
mean
(
a
v
erage)
and
standard
d
e
viation
(
v
ariability). Suppose
that
the
obser
v
ation
v
ector
is
the
discrete
random
v
ariable
V
O
bs
.
F
or
the
general
case,
w
here
v
ector
is
multidimensional,
the
probability density
function
of
the
normal
distri
b
ution is
a
g
aussian
function:
P
(
V
O
bs

µ,
Σ
)
=
The
EM
is
the
ideal
candidate
for
solving
parameter
estimation
problems for
the
GMM.
Each
of
the
EM
iterations consists
of
t
w
o
steps
Estimation
(E)
and
Maximization (M).
The
M

step
maximizes
a
li
k
elihood
function
that
is
refined
in
each
iteration
by
the
E

step.
V
.
F
E
A
TUR
E
M
A
TCHIN
G
AN
D
NORMALIZ
A
TIO
N
After
e
xtracting
the
obser
v
ation
v
ectors
corresponding
to
the
test
images,
the
probability
of
the
obser
v
ation sequence
g
i
v
en
a
GMM
model
is
computed. The
model
with
the
highest
log

li
k
elihood i
s
selected
and
this
model
r
e
v
eals
the
identity
of
the
unkn
o
wn
finge
r
.
Thus,
during
the
identification
process,
the
characteristics of
the
test
image
are
e
xtraction
by
the
2D

BDCT
corresponding to
each
person.
Then
the
Log

li
k
elihood
score
of
the
obser
v
ati
on
v
ectors
g
i
v
en
each
model,
P
(
V
O
bs

θ
i
)
=
`
(
V
O
bs
,
θ
i
),
is
computed
[8].
Therefore,
the
score
v
ector
is
g
i
v
en
by:
L
(
V
O
bs
)
=
[
`
(
V
O
bs
,
θ
1
)
`
(
V
O
bs
,
θ
2
)
·
·
·
`
(
V
O
bs
,
θ
D
)]
(9)
Where
D
represents
the
size
of
model
database.
An
important
aspect
that
has
to
be
addressed in
identifi

cation
process
is
the
normalization of
the
scores
obtained.
Normalization typically
i
n
v
ol
v
es
mapping
the
scores
obtained
into
a
common domain.
Thus,
a
Min

Max
normalization
scheme
w
as
empl
o
yed
to
transform
the
Log

li
k
elihood scores
computed
into
similarity
scores
in
the
same
range.
1
e
xp
h
p
(2
π
)
d
Σ
−
1
(
V
O
bs
2
−
µ
)
T
Σ
−
1
(
V
O
bs
i
−
µ
)
(7)
L
N
=
L
−
min
(
L
)
(10)
max
(
L
)
−
min
(
L
)
where
µ
is
the
mean,
Σ
is
the
c
o
v
ariance
matrix
and
d
is
Where
L
N
denotes
the
normalized
Log

li
k
elihood
scores.
the
dimension
of
fea
ture
v
ecto
r
.
C
o
v
ariance matrix
is
the
natural
generalization to
higher
dimensions
of
the
concept
of
the
v
ariance of
a
random
v
ariable. If
we
suppose the
random
v
ariable
measurement is
not
characterized only
with
simple
g
aussian
distri
b
ution, we
can
the
n
define
it
with
multiple
g
aussian
components.
GMM
is
a
probability
distri
b
ution
that
is
a
co
n
ve
x
combination
of
other
g
aussian
distri
b
utions:
N
P
(
V
O
bs
)
=
X
π
j
P
(
V
O
bs

µ
j
,
Σ
j
)
(8)
j
=1
where
N
is
the
number
of
Gaussian mixtures and
π
j
is
the
weight
of
each
of
the
mixture.
After
GMM is
trained,
the
model
of
each
user
will
be
the
final
v
alues
of
π
j
,
µ
j
and
Σ
j
.
Thus,
the
compact
notation
θ
,
such
that
θ
=
{
π
j
,
µ
j
,
Σ
j
}
N
,
is
used
to
represent a
model.
T
o
estimate
the
density
para

meters
of
a
GMM
statistic
model,
cluster
estimation method
called
Expectation

maximization
algorithm
(EM)
is
adopted.
H
o
w
e
v
e
r
,
these
scores
are
com
pared,
and
the
highest
score
is
selected.
Therefore,
the
best
score
is
D
o
and
its
equal
to:
D
o
=
max(
L
N
)
(11)
i
Finall
y
,
this
score
is
used
for
decision
making.
VI
.
E
XPERI
M
EN
T
A
L
RESU
L
T
S
AN
D
DISCUSSIO
N
A.
Exper
imental
database
W
e
e
xperimented our
method
on
Hong
K
ong
polytechnic
un
i
v
ersity
(PolyU)
FKP
Database [9].
The
database has
a
total
of
7920
images
obtained from
165
persons. this
database
including 125
males
and
40
females.
Among
them,
143
subjects a
re
20
∼
30
years
old
and
the
others
are
30
∼
50
years
old.
these
images
are
collected
in
t
w
o
separate
sessions.
In
each
session, the
subject
w
as
as
k
ed
to
pr
o
vide 6
images
for
each
of
Left
Ind
e
x
Fingers (LIF),
Left
Middle Fingers (LMF),
Right
Ind
e
x
Fingers
(RIF
)
and
Right
Middle
Fingers (RMF).
Therefore,
48
images
were
collected
from
each
subject.
INTER
N
A
TIO
N
AL
CONFERENCE
ON
ELECT
R
ONICS
AND
OIL

ICEO2013
5

6
MARCH
2013,
O
U
ARGLA,
ALGERIA
4
D
A
T
A
B
ASE
T
o
F
AR
FRR
T
o
F
AR
FRR
T
o
F
AR
FRR
T
o
F
AR
FRR
0.9600
8.808
1.556
0.9600
7.970
1.630
0.6500
8.777
2.963
0.9400
7.820
1.482
165
Persons
0.9740
3.445
3.445
0.9717
3.874
3.874
0.9644
4.173
4.173
0.9586
3.318
3.318
0.9850
1.158
7.630
0.9800
1.866
6.593
0.9800
1.332
7.630
0.9800
0.731
7.259
(
a
)
(
b
)
(
c
)
F
ig.
4.
Uni

modal
open
set
identification
system
performance.
(a)
System
performance
under
di
f
ferent
2D

BDCT
coe
f
ficients
number
in
each
block
and
v
arious
GMM,
(b)
The
R
OC
cur
v
es
for
all
GMMs
and
(c)
The
R
OC
cur
v
es
for
all
finger
types.
T
ABLE
1
:
OP
EN
SET
IDENTIFIC
A
TION
TEST
RESU
L
T
IN
THE
CASE
OF
UNI

MO
D
AL
SYSTEM
LEFT
INDEX
FINGER
LEFT
MIDDLE
FINGER
RIGHT
INDEX
FINGER
RIGHT
MIDDLE
FINGER
B.
Evaluation
criteria
The
measure
of
utility
of
a
n
y
biometric
recognition system
for
a
particular
application
can
be
e
xplained
by
t
w
o
v
alues
[10].
The
v
alue
of
the
F
alse
Acceptance Rate
(
F
AR)
criterion,
which
is
the ratio
of
the number
of
instances
of
di
f
ferent
feature
pairs
of
the
t
raits
found
do
match
to
the
total
number
of
counterpart attempts,
and
the
v
alue
of
the
F
alse
Rejection
Rate
(FRR
)
criterion,
which
is
the
ratio
of
the
number
of
instances
of
same
feature
pairs
of
the
traits
found
do
not
match
to
the
total
number
of
counte
rpart attempts.
It
is
clear
that
the
system
can
be
adjusted to
v
ary
the
v
alues
of
these
t
w
o
criteria
for
a
particular
application. H
o
w
e
v
e
r
,
decreasing
one
i
n
v
ol
v
es
increasing
the
other
and
vice
v
ersa.
The
system
threshold
v
alue
is
obtained
using
Equal
Erro
r
Rate
(EER
)
criteria
when
F
AR
=
FRR.
This
is
based
on
the
rationale
that
both
rates
must
be
as
l
o
w
as
possible
for
the
biometric
system
to
w
ork
e
f
fect
i
v
el
y
.
Another
performance
measurement is
obtained
from
F
AR
and
FRR,
which
is
the
Genuine
Acc
eptance
Rate
(GAR
).
It
represents
the
identification
rate
of
the
system.
In
order
to
visually
describe
the
performance of
a
biometric
system,
Rece
i
v
er
Operating
Characteristics (
R
OC)
cur
v
es
are
usually
g
i
v
en. A
R
OC
cur
v
e
sh
o
ws h
o
w
the
F
AR
v
alues are
chan
ged
relat
i
v
ely
to
the
v
alues
of
the
GAR
and
vice

v
ersa [11].
Biometric
recognition systems
generate
matching
scores
that
represent
the
d
e
gree
of
similarity
(or
dissimilarity)
between
the
input
and
the
stored
template.
C.
2D

BDCT
coe
f
ficients
selection
in
ea
c
h
blo
c
k
The
2D

BDCT coe
f
ficients
reflect
the
compact
ene
r
gy
of
di
f
ferent
frequencies.
Most
of
the
higher
frequen
c
y
coe
f
ficients
are
small
and
th
e
y
become
n
e
gligible, as
result,
the
features
der
i
v
ed
from
the
2D

BDCT
computation is
limited
to
an
array
of
summed
spectral ene
r
gies
within
a
block
in
frequen
c
y
domain.
In
this
section,
we
present
the
identification
accura
c
y
of
our
system
as
a
function of
the
number
of
2D

BDCT
coe
f
ficients
(in
each
block)
used.
The
performance
ev
aluation
w
as
repeated f
or
v
arious
numbers of
2D

BDCT
coe
f
ficients
and
v
arious
numbers
of
GMM,
and
the
results
are
as
sh
o
wn
in
Fig.
4
.
(a)
.
The
reason
Fig.
4
.
(a)
w
as generated
w
as
to
sh
o
w
h
o
w
the
number of
2D

BDCT
coe
f
ficients
selection
in
each
block
and
the
number
of
GMMs
used
might
h
a
v
e
an
e
f
fect
on
the
performance of
our
system.
W
e
obser
v
e
that
the
identification accura
c
y
becomes
v
ery
high
at
certain
coe
f
ficients
and
slight
decrease
in
identification
accura
c
y
as
we
go
to
higher
numbers
of
coe
f
ficients.
F
or
e
xample,
if
1

GMM
with
28
coe
f
ficients
in
each
block,
is
used
for
the
identification,
we
h
a
v
e
a
GAR
equal
to
91.480 %.
In
the
case
of
using
2

GMM with
24
coe
f
ficients
in
each
block,
GAR
w
as
95.291
%.
3

GMM
with
22
coe
f
ficients in
each
block,
impr
ov
es
th
e
result
(GAR
=
96.467
%)
for
a
database size
equal
to
165
persons. In
Fig.
4
.
(b)
,
we
compare the
system
performance
under
di
f
ferent
GMMs.
The
results
sh
o
w
the
benefits
of
using
3

GMM.
Thus,
the
performance of
the
open
set
uni

modal
identification
system
i
s
significantly
impr
ov
ed
by
using
the
3

GMM
with
22
coe
f
ficients
in
each
block.
D.
Uni

modal
system
test
r
esults
T
o
ev
aluate
the
e
f
ficien
c
y of
the
uni

modal
biometric
method,
the
e
xperiments were
designed
as
foll
o
w:
three
sam

ples
(for
each
finger) of
e
ach
person
is
randomly
selected
for
enrollment,
and
the
rest
nine
finger images
are
used
as
test
samples
for
identification. Thus,
123255
comparisons
were
generated
for
performance
ev
aluation
(165
persons).
In
this
section
we
compare
the
performance
of
all
finger
types.
In
the
case
of
open
set
identification, Fig.
4
.
(c)
compares
the
performance
of
the
system
for
deferent
finger types.
It
can
safely
be
see
the
benefits
of
using
the
RMF
finger
than
the
LI
F
,
LMF
and
RIF
fingers
in
terms
of
EER.
It
can
be
achi
e
v
e
an
INTER
N
A
TIO
N
AL
CONFERENCE
ON
ELECT
R
ONICS
AND
OIL

ICEO2013
5

6
MARCH
2013,
O
U
ARGLA,
ALGERIA
5
SUM
WH
T
MIN
MAX
MUL
T
o
EER
T
o
EER
T
o
EER
T
o
EER
T
o
EER
0.9795
1.108
0.9795
1.115
0.9876
1.601
0.9723
1.562
0.9596
1.092
0.9735
1.519
0.9749
1.333
0.9851
1.874
0.9642
2.177
0.9483
1.489
0.9708
1.482
0.9701
1.510
0.9848
1.512
0.9627
1.770
0.9434
1.442
0.9664
1.577
0.9672
1.466
0.9833
1.447
0.9568
2.000
0.9354
1.556
0.9791
0.376
0.9792
0.375
0.9964
0.869
0.9607
1.239
0.9190
0.370
T
ABLE
2
:
OPEN
SET
IDENTIFIC
A
TION
TEST
RESU
L
T
IN
THE
CASE
OF
THE
FUSION
A
T
M
A
TCHING
SCORE
LEVEL
COMBI
N
A
TION
LIF

LMF
LIF

RIF
LMF

RMF
RIF

RMF
All
Fingers
T
ABLE
3
:
OPEN
SET
IDENTIFIC
A
TION
TEST
RESU
L
T
IN
THE
CASE
OF
THE
FUSION
A
T
DECISION
LEVEL
LIF

LMF

RIF
LIF

LMF

RMF
RIF

RMF

LIF
RIF

RMF

LMF
All
Fingers
F
AR
FRR
GAR
F
AR
FRR
GAR
F
AR
FRR
GAR
F
AR
FRR
GAR
F
AR
FRR
GAR
2.341
1.111
97.676
2.470
0.815
97.
552
0.705
1.037
99.291
1.047
1.333
98.949
2.997
2.074
97.016
EER
equal
to
3.318
%
at the
threshold
T
o
= 0.9586.
Therefore,
the
system
can
achi
e
v
e
higher
accura
c
y at
the
RMF
finger
compared
with
the
other
finger types.
Final
l
y
,
T
able
1
sh
o
ws
the
F
AR
and
FRR
with
percentage using
LI
F
,
LM
F
,
RIF
and
RMF
at
deferent
thresholds.
E.
Multi

modal
system
test
r
esults
A
ro
b
ust
identification
system may
require fusion
of
s
e
v
eral
finger
types
for
the
reason
that
the
limitation
presente
d
in
one
finger
may
be
compensated
by
another
finge
r
.
The
goal
of
this
e
xperiment
w
as
to
i
n
v
esti
g
ate the
systems
performance when
we
fuse
information
from
s
e
v
eral
finger
types
of
a
person.
In
f
act,
at
such
a
case
the
system
w
orks
as
a
kind
of
multi

modal
system
with
a
single
biometric trait
b
ut
multiple
units.
Therefore,
information presented
by
di
f
ferent
biometrics
(LI
F
,
LM
F
,
RIF
and
RMF)
is
fused
to
ma
k
e
the
system
e
f
ficient.
1)
Fusion
at
mat
c
hing sco
r
e
l
e
vel:
Fusion
at
the
matching

score
l
e
v
el
is
prefe
rred in
the
field
of
biometrics
because
there
is
su
f
ficient information
content
and
it
is
easy
to
access
and
combine
the
matching
scores
[12]. At
the
matching
score l
e
v
el
fusion,
the
matching scores
output
by
multiple
matchers (sub

system)
are
int
e
grated.
In
our
system,
di
f
ferent
combinations
of
finger types
and
di
f
ferent
fusion
rules,
such
as
Sum

sco
r
e
(
SU
M
),
Min

sco
r
e
(
MI
N
),
Max

sco
r
e
(
MA
X
),
Mul

sco
r
e
(
MU
L
)
and
Sum

weighting
sco
r
e
(
WHT
),
were
tested
to
find
the
combination that
optimizes
the
s
ystem
accura
c
y
.
Thus,
to
find
the
better
of
the
all
fusion
rules
and
combinations,
with
the
l
o
west
EER,
T
able
2
ta
b
ulates EER
for
v
arious
combinations
and
fusion rules.
As
can
be
seen,
the
best
result
w
as
obtained
with
the
combination
of
all
fingers and
th
e
fusion
rule
w
as
MUL
rule,
it
can
achi
e
v
e
e
v
en
higher
precision,
an
EER
of
0.370
%
and
a
T
o
of
0.9190.
The
performance of
the
open
set
identification system
is
significantly
impr
ov
ed
by
using
the
fusion
and
it
is
comparable with
other
hand
based
biometr
ics,
such
as
hand
geometry
and
fingerprint
identification
[13],
[14].
2)
Fusion
at
decision
l
e
vel:
Designing
a
suitable
method
of
decision
combinations is
a
k
e
y
point
for
the
ensembles
performance. In
this
pape
r
,
a
simplest
method
for
the
v
oting
schemes,
plu
r
ality
vote
,
is
used
[15].
H
o
w
e
v
e
r
,
in
this
method,
just
count
the
number
of
decision
for
each
class
and
assign
the
sample
to
the
class
that
obtained the
highest
number
of
v
otes.
Note
that,
the
number
of
sub

systems
should
be
ab
ov
e
or
equal
to
3.
F
or
t
he
ev
aluation
of
the
system
performance,
in
the
case
of
multi

modal system
based
on
decision
l
e
v
el,
a
series
of
e
xperiments were
carried
out
using
a
di
f
ferent
finger
type
combinations and
the
results
are
sh
o
wn
in
T
able
3.
From
T
able
3,
it
can
be
seen
that
our
identification
system
achi
e
v
es
a
best
performance
when
using
RI
F
,
RMF
and
LIF
(
F
AR
=
0.705
%,
FRR
=
1.037%
and
GAR
=
99.291%).
Finall
y
,
in
Fig.
5
.
(a)
,
we
compare
the
performance of
di
f
ferent systems
(uni

modal
and
multi

modal
based
on
fusion
at matchin
g
score
l
e
v
el).
The
results
sh
o
w
the
benefits
of
using
the
multi

modal system
with
matching
score
l
e
v
el
fusion.
Therefo
r
,
the
distance
distri
b
utions of
genuine
and
imposter
matchings obtained
by
the
proposed
scheme,
if
the
all
fingers
are
fused
in
the
case
of
matching score
l
e
v
el
by
MUL
rule
and
the
results
e
xpressed
as
a
F
AR and
FRR
depending
on
the
threshold,
are
plotted
in
Fig.
5
.
(b)
and
Fig.
5
.
(c)
,
respect
i
v
el
y
.
VII
.
C
ONCLUSIO
N
AN
D
FU
R
THE
R
W
OR
K
In
this
pape
r
,
a
multi

modal
biometric
identification
s
ystem,
using
FKP
biometric, based
on
fusion
of
s
e
v
eral
biometric
traits,
four
finger types,
has
been
proposed.
Fusion
of
these
biometric traits
is
carried
out
at
the
matching score
l
e
v
el
and
decision
l
e
v
el.
The
proposed
system
use 2D

BDCT
for feature
e
xtra
cted,
GMM
for
modeling
and
log

li
k
elihood
for
matching
process.
T
o
compare
the
proposed
multi

modal
system
with
the
uni

modal
systems,
a
series
of
e
xperiments
has
been
performed
in
the
case
of
open
set
identification
and
it
has
been
found
that
the
proposed
multi

modal system
g
i
v
es
a
considerable
performance
g
ain
ov
er
the
uni

modal
systems.
Our
future
w
ork
will
focus
on
the
performance
ev
aluation
in
both
phases
(
v
erification
and
identification)
by
using
a
la
r
ge
size
database
and
int
e
gration
of
other
b
iometric
traits
such
as
fingerprint
or
f
ace
to
get
the
system
performances
with
a
high
accura
c
y
.
R
EFERENCE
S
[1]
Arun
A.
Ross,
K.
Nandakumar
and
A.
K.
Jain,
“Handbook
of
Multibio

metrics”,
Springer
Science+Business
Media,
LLC,
N
e
w
Y
ork,
2006.
[2]
Rui
Zhao,
K
unlun
Li,
Ming
Liu,
Xue
Sun,
“
A
N
ov
el
Approach
of
Personal
Identification
Based
on
Single
Knuckle

print
Image”,
Asia

P
acific
Conference
on
Information
Processing,
APCI
P
,
2009.
[3]
Lin
Zhang, Lei
Zhang, and
D
a
vid Zhang, “Finge
r

Knuc
kle

Print
V
erifica

tion Based
on Band

Limited
Phase

Only
Correlation”,
CAIP
2009,
LNCS
5702,
pp.
141

148,
2009.
[4]
L.
Zhang,
L.
Zhang,
D.
Zhang,
“Finge
r

knuckle

print:
a
n
e
w
biometric
identifier”,
in:
Proceedings
of
the
ICIP09,
2009.
INTER
N
A
TIO
N
AL
CONFERENCE
ON
ELECT
R
ONICS
AND
OIL

ICEO2013
5

6
MARCH
2013,
O
U
ARGLA,
ALGERIA
6
(
a
)
(
b
)
(
c
)
Fig.
5.
Multi

modal
system
performance
in
the
case
of
fusion
at
matching
score
l
e
v
el
(all
fingers)
with
MUL
rule.
(a)
The
comparison between
the
uni

modal
and
multi

modal
systems,
(b)
The
genuine
and
the
imposter
distri
b
ution
and
(c)
The
dependen
c
y
of
the
F
AR
and
the
FRR
on
the
v
alue
of
the
threshold.
[5]
Ahmed
N,
Natarajan
T
,
and
Rao
K,
“Discrete
cosine
t
ransform”, IEEE
T
rans,
on
Computers,
23(1):9093.
[6]
Abdallah
Meraoumia, Salim
Chitroub
and
Bouridane
ahmed,
“Gaussian
Modeling
And
Discrete
Cosine
T
ransform
F
or
E
f
ficient And
Automatic
P
almprint
Identification”,
International
Conference
on
Machine
and
W
e
b
Intelligence

ICMWI’2010,
USTHB,
Algiers,
October
3

5,
2010.
[7]
Peter
V
archol,
D
ˇ
usan
L
e
vic
k
´
y,
“Using
of
Hand
Geometry in
Biometric
Security
Systems”,
Radioengineering,
V
OL.
16,
NO.
4,
pp:82

87,
De

cember
2007
[8]
Dimitrios
V
er
v
eridis
and Constantine
K
otropoulos,
“Gaussian
mixture
modeling
by
e
xploiting
the
Mahalanobis distance”,
IEEE
T
rans.
Signal
Processing,
v
ol.
56,
issue
7B,
pp.
2797

2811,
2008.
[9]
PolyU
Finger
KnucklePrint
Database.
A
v
ailable:
http://www4.comp.polyu.edu.
hk
/
∼
biometrics/
FK
P
.htm
.
[10]
Connie
T
.,
T
eoh
A.,
Goh
M.
and
Ngo
D,
“
P
almprint
recognition
with
PCA
and
IC
A
”,
N
e
w
Zealand,
P
almerston
North,
2003.
[11] Jain
A.
K.
Ross,
A.
and
Prabhaka
r
,
S,
“
An
introduction
to
biometric
recognition”, IEEE
T
ransactions on
Circuits
and
Systems
for
V
ideo
T
echnolog
y
,
14(1),
420,
2004.
[12]
Abdallah
Meraoumia,
Salim
Chitroub
and
Ahmed
Bouridane,
“2D
and
3D
P
almprint
Information
and
Hidden
Mar
k
o
v
Model for
Impr
ov
ed
Iden

tification Performance”,
International
conference
on
inte
lligent
systems
design
and
applications

IS
D
A
2011,
C
o
`
rdoba,
spain,
N
ov
ember
22

24,
2011
[13]
S.
M.
Prasad,
V
.
K.
G
o
vindan,
P
.
S.
Sathid
e
vi,
“
P
almprint
Authentication
Using
Fusion
of
W
a
v
elet
Based
Representations”, Procedinng
by
IEEE,
pp.520

525,
2009.
[
14]
Richa
Singh,
Mayank
V
atsa,
Afzel
Noore,
“Hierarchical fusion
of
multispectral
f
ace
images
for
impr
ov
ed
recognition
performance”,
science
direct,
Information
Fusion
9
(2008)
200

210.
[15]
Francisco
Moreno

Seco,
Jos
e
´
M.
I
˜
nesta,
Pedro
J.
Ponce
de
Le
o
´
n
and
Luisa
Mic
o
´
,
“Comparison
of
classifier
Fusion
Methodes
for Classification
in
P
attern
Recognition
T
asks”,
LNCS
4109,
pp:
705

713,
2006.
Σχόλια 0
Συνδεθείτε για να κοινοποιήσετε σχόλιο