Indonesian
J
our
nal
of
Electrical
Engineering
and
Computer
Science
V
ol.
20,
No.
3,
December
2020,
pp.
1332
1341
ISSN:
2502-4752,
DOI:
10.11591/ijeecs.v20i3.pp1332-1341
r
1332
Biometric
authentication
using
cur
v
elet
transf
orm
K
errache
Soumia
1
,
Beladgham
Mohammed
2
,
Hamza
A
ymen
3
,
Kadri
Ibrahim
4
1,2,4
L
TIT
Laboratory
,Department
of
Electrical
Engineering,
T
ahri
Mohammed
Bechar
Uni
v
ersity
,
Algeria
3
Department
of
Electrical
Engineering,
T
ahri
Mohammed
Bechar
Uni
v
ersity
,
Algeria
Article
Inf
o
Article
history:
Recei
v
ed
Jan
25,
2020
Re
vised
May
1,
2020
Accepted
Jun
5,
2020
K
eyw
ords:
Biometric
authentication
Classification
algorithmes
Curv
elet
transform
Feature
e
xtraction
Multiresolution
analysis
ABSTRA
CT
In
this
paper
,
we
propose
a
feature
e
xtraction
method
f
or
tw
o-dimensional
image
authentication
algorithm
using
curv
elet
transform
and
principal
component
analysis
(PCA).
Since
w
a
v
elet
transform
can
not
adequately
describe
f
acial
curv
es
features,
the
proposed
approach
in
v
olv
es
image
denoising
applying
a
2D-Curv
elet
transform
to
achie
v
e
compact
representations
of
curv
es
singularities.
T
o
assess
the
performance
of
the
presented
method,
we
ha
v
e
emplo
yed
three
classification
techniques:
Neural
netw
orks
(NN),
K-Nearest
Neighbor
(KNN)
and
Support
V
ector
machines
(SVM).
Extensi
v
e
e
xperimental
res
ults
and
comparison
with
the
e
xisting
methods
sho
w
the
ef-
fecti
v
eness
of
the
proposed
recognition
method
in
the
ORL
f
ace
database
and
CASIA
iris
database.
Copyright
c
2020
Insitute
of
Advanced
Engineeering
and
Science
.
All
rights
r
eserved.
Corresponding
A
uthor:
K
errache
Soumia,
Department
of
Electrical
Engineering,
Information
Processing
and
T
elecommunication
Laboratory
(L
TIT),
T
ahri
Mohammed
Bechar
Uni
v
ersity
,
Algeria.
Email:
s
guemana@yahoo.com
1.
INTR
ODUCTION
In
the
past
20
years,
biometric
recognition
technology
has
Quickly
de
v
eloped.
Biometric
aut
hentica-
tion
is
a
set
of
procedures
of
comparing
data
to
determine
resemblance
for
the
characteristics
of
the
indi
vidual
to
the
biometric
”template”
of
that
person.
Fingerprints,
f
acial
features,
tone,
hand
mechanics,
handwriting,
retina,
and
iris
ha
v
e
formed
biometric
frame
w
orks.
W
e
ha
v
e
seen
ne
w
techniques
such
as
Principal
component
analysis
(PCA)
and
Linear
discriminant
analysis
(LD
A)
and
Independent
component
analysis
(ICA)
emer
ge
o
v
er
the
past
fe
w
years
[1-3].
Multiresolution
multidirectional
transforms
with
the
w
a
v
elet
transform
for
pat-
tern
recognition
w
as
compared
in
[4]
and
since
the
con
v
entional
w
a
v
elet
transformation
can
only
describe
the
singularity
of
the
point
in
the
image
that
af
fects
the
w
a
v
elet
coef
ficients,
it
is
dif
ficult
for
the
w
a
v
elet
to
achie
v
e
satisf
actory
curv
e
e
xpression
results
[5-7].
Recently
a
number
of
ne
w
multiresolution
analysis
tools
lik
e,
ridgelet
[8],
contourlet
[9-12],
etc.
were
de
v
eloped
to
solv
e
the
abo
v
e
problem.
These
tools
ha
v
e
better
directional
decomposition
capabili
ties
and
bet-
ter
ability
to
represent
edges
and
other
singularities
along
curv
es
than
w
a
v
elets.
F
ollo
wing
the
introduction
of
Curv
elet
transform
theories
by
Emmanuel
and
Donoho,
multi-scale
analysis
in
image
processing
has
been
widely
applied
[13-16].
The
de
v
eloped
continuous
curv
elet
transform
can
represent
image
objects
with
edges
and
other
singularities
along
the
curv
e
which
were
not
captured
by
w
a
v
elets.
The
curv
elet
transforma
tion
[17]
is
implemented
in
the
proposed
method
in
order
to
capture
f
acial
features
at
v
arious
angles
and
scales.
F
ace
and
Iris
recognition
e
xperiments
and
ha
v
e
been
car
ried
out
on
ORL
and
CASIA
database.
The
curv
el
et
trans-
form
with
classifiers
such
as
neural
netw
ork
(NN)
and
support
v
ector
machine
(SVM)
and
k-nearest
neighbors
(KNN)
has
been
used
to
yield
better
recognition
results
as
compared
to
e
xisting
methods.
Rest
of
the
paper
is
structured
as
follo
ws:
Section
2
describes
the
curv
elet
transform,
Principal
Component
Analysis
and
classifi-
cation
algorithms
in
more
details.
Section
3
and
4
gi
v
es
a
discussion
and
e
xperimental
results
with
conclusion.
J
ournal
homepage:
http://ijeecs.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1333
2.
THE
PR
OPOSED
METHOD
In
order
to
de
v
elop
a
practical
approach
to
Biometric
authentication,
we
proposed
se
v
eral
m
ethods
based
on
the
combined
Curv
elet
and
PCA
and
three
classification
algorithms
(SVM-KNN-NN),
the
accurac
y
of
these
approaches
are
carried
out
by
simulation
and
comparati
v
e
study
.
In
Figure
1,
the
proposed
system
starts
with
applying
the
Curv
elet
transform
to
handle
curv
es
using
only
a
small
number
of
coef
ficients.
Hence
the
Curv
elet
handles
curv
e
discontinuities
well.
After
that,
the
image
is
sent
to
the
PC
A
step,
based
on
the
creation
of
lo
w
dimensional
representation.
Ne
xt,
we
select
the
eigen
v
ectors
with
the
higher
v
alue
of
eigen
v
alue.
Finally
,
we
ha
v
e
used
SVM
and
KNN
then
NN
for
feature
classification,
based
on
these
methods
we
can
present
three
recognition
approach,
the
first
one
is
curv
elet+PCA+SVM,
the
second
one
is
based
on
curv
elet
+PCA+KNN,
and
the
last
one
is
curv
elet+PCA+NN.
The
dif
ferent
used
methods
for
these
approaches
are
educed
in
the
follo
wing
subsections.
Figure
1.
Proposed
algorithm
2.1.
Cur
v
elet
transf
orm
The
curv
elet
transform
w
as
first
introduced
by
Candes
and
Donoho
1999
to
o
v
ercome
the
dra
wbacks
and
limitations
Of
widely
used
multiresolution
methods
such
as
the
w
a
v
elet
transform
and
Ridglet
transform.
The
multiscale
transform
principle
is
a
property
common
to
curv
elet
and
w
a
v
elet
transform
where
each
has
multiple
frames
inde
x
ed
by
location
and
scale
parameters.
Ho
we
v
er
,
the
Curv
elet
transform,
unlik
e
the
w
a
v
elet
transform,
has
a
v
ery
high
de
gree
of
directional
fle
xibility
,
and
the
frame
size
is
subject
to
the
anisotropic
scaling
principle.
Curv
elet
transform
ha
v
e
tw
o
possible
implementations,
the
first
well-kno
wn
Implementations
is
Called
Curv
elet
G1
and
the
second
one
is
called
Curv
elet
G2.
In
this
paper
we
will
co
v
er
only
the
first
one
since
it’
s
the
one
we
w
ork
ed
with
[18,
19].
2.1.1.
First
generation
cur
v
elets
(DCTG1)
The
first
generation
discrete
curv
elet
transform
(DCTG1)
of
a
continuum
function
f
(
x
)
mak
es
use
of
a
dyadic
sequence
of
scales,
and
a
bank
of
filters
with
the
property
that
the
bandpass
fil
ter
j
is
concentrated
near
the
frequencies
[2
2
j
;
2
2
j
+2
]
,e.g.
j
(
f
)
=
2
j
f
;
2
j
(
v
)
=
(2
2
j
v
)
(1)
In
w
a
v
elet
theory
,
one
uses
a
decomposition
into
dyadic
sub-bands
[2
j
;
2
j
+1
]
.
In
contrast,
the
sub-
bands
used
in
the
discrete
curv
elet
transform
of
continuum
functions
ha
v
e
the
nonstandard
form
[2
2
j
;
2
2
j
+2
]
.
Biometric
authentication
using
curvelet
tr
ansform
(K
err
ac
he
Soumia)
Evaluation Warning : The document was created with Spire.PDF for Python.
1334
r
ISSN:
2502-4752
This
is
nonstandard
feature
of
the
DCTG1
well
w
orth
remembering
(this
is
where
the
approximate
parabolic
scaling
la
w
comes
into
play).
The
DCTG1
decomposition
is
the
sequence
of
the
follo
wing
steps:
-
Sub-band
Decomposition,
(The
object
f
is
decomposed
into
sub-bands).
-
Smooth
P
artitioning,
(Each
sub-band
is
smoothly
windo
wed
into
“squares”
of
an
appropriate
sca
le
(of
side-length
2
2
j
)).
-
Ridgelet
Analysis,
(
Each
square
is
analyzed
via
the
DR
T).
In
this
definition,
the
tw
o
dyadic
sub-bands
[2
2
j
;
2
2
j
+1
]
and
[2
2
j
+1
;
2
2
j
+2
]
are
mer
ged
before
applyi
ng
the
ridgelet
transform.
As
sho
wn
in
Figure
2.
Figure
2.
First
generation
discrete
curv
elet
transform
(DCTG1)
flo
wchart.
The
figure
illustrates
the
decomposition
of
the
original
image
into
sub-bands
follo
wed
by
the
spatial
partitioning
of
each
sub-band.
The
ridgelet
transform
is
then
applied
to
each
block
2.2.
Principals
component
analysis
(PCA)
Principal
component
analysis
is
suggested
by
T
urk
and
Pent
land
in
1991
[20],
which
is
often
used
for
e
xtracting
features
of
the
image.
Principal
Component
Analysis
is
the
most
widely
used
method
considering
the
f
acial
feature
e
xtraction
in
image
processing.
The
basic
idea
behind
the
PCA
is,
the
set
of
images
are
initially
transformed
into
Eigenf
aces
i.e.
lo
wer
data
space
by
using
the
K-L
transform
method.
This
method
includes
the
linear
transformation
of
the
higher
data
space
i
nto
the
lo
wer
data
space
using
linear
transformation
method.
This
e
xtracted
lo
wer
-dimensional
image
preserv
es
most
of
the
data
or
information
from
the
original
higher
-
dimensional
f
acial
image.
This
mapped
lo
wer
data
space
is
called
as
the
Eigenf
ace.
Then
the
test
Eigenf
aces
v
ector
from
the
database
is
projected
on
the
trainee
Eigenf
aces
v
ector
to
get
the
correct
match.
F
or
PCA,
the
tw
o-dimensional
image
matrix
must
be
first
transformed
into
a
one-dimensional
v
ector
with
high
order
.
While
the
number
of
trai
ning
samples
is
small,
it
is
challenging
to
calculate
the
co
v
ariance
m
atrix
of
the
training
sample
accurat
ely
.
Furthermore,
structure
information
will
be
lost
during
processing.
The
Eigenf
aces
v
ector
as
considered
as
the
v
ector
for
constructing
the
co
v
ariance
matrix.
Here,
the
pix
el
information
of
each
image
is
used
to
construct
the
Eigen
v
ector
.
This
Eigen
v
ector
information
is
used
to
select
the
Principal
Component
ha
ving
a
higher
Eigen
v
alue.
Each
image
location
Contrib
utes
to
each
Eigen
v
ector
so
that
we
can
display
the
Eigen
v
ector
as
a
sort
of
f
ace.
Computing
PCA:
-
First,we
tak
e
a
set
of
images
in
column
matrix
or
the
ro
w
matrix
form,named
=
(
1
;
2
;
3
;
:::::
;
M
)
(2)
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
20,
No.
3,
December
2020
:
1332
–
1341
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1335
where,
M
is
the
totale
of
objects
present
in
total
database.
[-]
Find
the
a
v
erage
of
the
defined
matrix
=
1
M
M
X
n
=1
n
(3)
Here,
n
is
t
he
total
number
of
im
ages
in
s
ingle
object
of
Database
is
the
Mean
of
the
defined
matrix
-
Then
find
the
dif
ferential
distance
between
the
trainee
images
and
the
mean
calculated
=
j
(4)
2.3.
Support
v
ector
machines
(SVM)
A
linear
model
for
classification
and
re
gression
tasks
is
the
SVM
or
Support
V
ector
Machine.
It
can
handle
linear
and
non-linear
problems
and
function
well
for
man
y
practical
issues.
SVM’
s
idea
is
simple:
the
algorithm
generates
a
line
or
h
yperplane
that
di
vides
data
into
groups.
~
w
:
~
x
b
=
0
(5)
where
~
w
is
the
(not
necessarily
normalized)
normal
v
ector
to
the
h
yperplane.
This
is
much
lik
e
Hesse
normal
form,
e
xcept
that
~
w
is
not
necessarily
a
unit
v
ector
.
The
parameter
b
k
~
w
k
determines
the
of
fset
of
the
h
yper
p
l
ane
from
the
origin
along
the
normal
v
ector
~
w
.
From
both
sets,
we
consider
the
points
nearest
to
the
line
according
to
the
SVM
algorithm.
Such
points
are
kno
wn
as
v
ectors
of
support.
No
w
we
measure
the
distance
between
the
v
ectors
of
the
line
and
the
support.
The
mar
gin
is
called
this
g
ap.
Our
goal
is
to
optimize
the
mar
gin.
The
ideal
h
yperplane
is
the
h
yperplane
for
which
the
mar
gin
is
a
peak.
SVMs
are
essentially
classifiers
for
tw
o
classes.
The
traditional
w
ay
to
classify
multi-class
SVMs
is
to
construct
j
C
j
one-v
ersus-rest
classifiers
(commonly
referred
to
as
“one-
v
ersus-all”
or
O
V
A
classification),
And
the
select
the
class
that
classifies
the
maximum
mar
gin
of
the
test
data.
Another
technique
is
to
create
a
set
of
one-v
ersus-one
classifie
rs
and
to
pick
the
class
selected
by
the
most
classifiers.
While
this
in
v
olv
es
b
uilding
j
C
j
(
j
C
j
1)
=
2
classifiers,
As
the
training
data
set
for
each
classifier
is
much
smaller
,
the
time
for
training
classifiers
may
actually
decrease.
2.4.
K-Near
est
neighbors
(KNN)
As
with
most
adv
ances
in
technology
in
the
early
1900s,
the
KNN
algorithm
w
as
born
out
of
armed
forces
w
ork.T
w
o
of
fices
of
the
USAF
School
of
A
viat
ion
Medicine
Fix
and
Hodges
(1951)
published
a
technical
report
proposing
a
non-parametric
method
for
pattern
identification,
which
has
since
become
popular
as
the
nearest
neighbor
a
lgorithm.
KNN
f
alls
in
the
super
vised
lear
ning
f
amily
of
algorithms.
This
means
that
gi
v
en
a
label
led
dataset
consisting
of
training
observ
ations
(x,y)
,we
w
ould
lik
e
to
capture
the
relationship
between
x
!
the
data
and
y
!
the
l
abel
.
More
formally
,
we
w
ant
to
learn
a
function
g
:
X
!
Y
so
that
gi
v
en
an
unseen
observ
ation
x;
g
(
x
)
can
confidently
predict
the
corresponding
output
y
.
KNN’
s
objecti
v
e
is
to
label
the
test
set.
T
o
label
a
test
point,
in
its
neighbourhood,
we
search
for
e
xisting
labels.
The
latter
are
kno
wn
as
the
training
set.
W
e
choose
the
k
labeled
points
that
lie
closest
to
our
test
point.
Then
we
assign
the
label
to
the
majority
of
these
k
neighbours
[21,
22].
2.5.
Artificial
neural
netw
orks
(ANN)
W
arren
McCullough
and
W
alter
Pitts,
tw
o
scientists
at
the
Uni
v
ersity
of
Chicago
who
mo
v
ed
to
MIT
in
1952
as
founding
members
of
what
is
often
considered
(the
first
department
of
cogniti
v
e
science),
were
the
first
who
suggested
neural
netw
orks
in
1944.
The
idea
is
to
tak
e
a
wide
range
of
training
e
xamples
and
then
de
v
elop
a
system
that
can
learn
from
them.
Neural
netw
orks
are
dif
ferent
from
the
w
ay
con
v
entional
machine-
learning
algorithms
lik
e
SVM
and
KNN
are
implemented.
T
o
see
ho
w
neural
netw
orks
are
approach
solving
problems,
we
start
by
defining
a
fe
w
notations.
Let’
s
be
gin
with
a
notation
which
lets
us
refer
to
weights
in
the
netw
ork
in
an
unambiguous
w
ay
.
W
e’
ll
use
w
l
j
k
to
denote
the
weight
for
the
connection
from
the
k
th
neuron
in
the
(
l
1)
th
layer
to
the
j
th
neuron
in
the
l
th
layer
.
W
e
use
a
similar
notation
for
the
netw
ork’
s
biases
and
acti
v
ations.
Explicitly
we
use
b
l
j
for
the
bias
of
the
j
th
neuron
in
the
l
th
layer
.
And
we
use
a
l
j
for
the
acti
v
ation
of
the
j
th
neuron
in
the
l
th
layer
.
The
follo
wing
diagram
e
xamples
of
notations
as
sho
wn
in
Figure
3.
Biometric
authentication
using
curvelet
tr
ansform
(K
err
ac
he
Soumia)
Evaluation Warning : The document was created with Spire.PDF for Python.
1336
r
ISSN:
2502-4752
Figure
3.
Multilayer
perceptron
(MLP)
W
ith
these
notations
the
acti
v
ation
a
l
j
of
the
j
th
neuron
in
the
l
th
layer
is
related
to
the
acti
v
ations
in
the
(
l
1)
th
layer
by
the
equation:
a
l
j
=
X
k
w
l
j
k
a
l
1
k
+
b
l
j
!
(6)
3.
RESUL
T
AND
DISCUSSION
3.1.
Experiment
on
ORL
database
The
ORL
F
aces
Database
includes
a
collection
of
f
acial
images
tak
en
in
the
A
T&T
Laboratories
in
Cambridge,
from
April
1992
to
April
1994.
There
are
10
dif
ferent
images
of
e
v
ery
40
indi
viduals.
The
images
were
tak
en
at
v
arious
times,
for
some
subjects,
with
dif
ferent
illuminations,
f
acial
e
xpressions
(open/closed
e
yes,
smiling
/
no-smiling)
and
f
ace
specifics
(glasses
/
no
glasses).
All
images
were
tak
en
with
the
subjects
in
an
upright,
frontal
posture
(with
tol
erance
for
some
side
mo
v
ement)
ag
ainst
a
dark
homogeneous
background
as
sho
wn
in
Figure
4.
Firstly
,
the
curv
elet
transform
is
applied
to
the
ORL
images
where
the
vital
information
are
e
xtracted
from
the
original
images
as
sho
wn
in
Figure
5.
Figure
4.
ORL
images
database
Figure
5.
Curv
elet
applied
image
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
20,
No.
3,
December
2020
:
1332
–
1341
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1337
Then,
the
essential
features
are
e
xtracted
from
the
ne
w
images
by
using
the
PCA
algorithm,
the
gi
v
en
forth
first
important
eigen
v
alues
are
represented,
in
Figure
6.
Finally
,
the
obtained
eigen
v
ectors
are
classified
by
classification
algorithms
the
accurac
y
of
each
algorithm
is
sho
wn
in
the
follo
wing
Figure
7.
Figure
7
sho
w
the
e
xperiment
results
of
recognition
rate
obtained
for
ORL
f
aces
database
using
PCA+Curv
elet
for
feature
e
xtraction
step,
A
three
weak
classifiers
are
use
d
and
we
ha
v
e
achie
v
ed
a
v
erage
recognition
rates
of
97.2,
98.75
and
92.5
respecti
v
ely
KNN,
NN
and
SVM.
Figure
6.
The
first
important
eigen
v
alues
e
xtracted
by
PCA
Figure
7.
Recognition
rate
for
ORL
database
with
three
classification
algorithms
3.2.
CASIA
Iris
image
database
W
ith
a
homemade
iris
camera,
iris
images
of
CASIA
V1.0
were
captured.
Eight
NIR
illuminators
of
850
nm
are
arranged
circularly
around
the
sensor
to
ensure
that
the
iris
is
illuminated
uniformly
and
correctly
.
T
o
protect
our
IPR
in
the
design
of
the
iris
camera
(especially
the
NIR
lighting
system)
before
patents
are
issued,
Pupil
areas
of
all
CASIA
V1.0
iris
imaging
were
automatically
detected
and
replaced
with
a
constant
circular
area,
masking
the
specular
NIR
illuminators
reflections
in
adv
ance
of
their
publication.
Such
processing
may
af
fect
the
detection
of
pupils
b
ut
does
not
af
fect
other
components
of
an
iris
recognition
system
such
as
iris
e
xtraction,
only
the
pupil
and
Sclera
area,
i.e.
the
ring-shaped
iris
area
[23].
-
CASIA
Iris
Image
Database
(v1).0)
includes
108
e
ye
pictures
with
756
iris,
and
in
2
s
essions
with
three
and
four
samples
collected
during
the
first
and
second
session,
se
v
en
pictures
in
total
are
captured
for
each
e
ye
as
sho
wn
in
Figure
8.
-
It
is
recommended
that
when
you
w
ant
to
measure
the
in-class
v
ariance,
you
compare
tw
o
samples
from
the
same
e
ye
tak
en
from
dif
ferent
sessions.
F
or
e
xample,
the
iris
images
in
the
first
session
can
be
used
as
the
training
dataset,
and
those
from
the
second
session
can
be
used
for
testing.
Biometric
authentication
using
curvelet
tr
ansform
(K
err
ac
he
Soumia)
Evaluation Warning : The document was created with Spire.PDF for Python.
1338
r
ISSN:
2502-4752
Figure
8.
CASIA
iris
image
database
V1
Firstly
,
we
apply
the
curv
elet
transform
to
e
xtract
curv
ed
singularity
information
from
original
images
then
se
gmentation
and
normalization
of
the
iris
are
used
on
the
ne
w
images
as
sho
wn
in
Figure
9.
Ne
xt,
the
essential
features
are
e
xtracted
from
the
ne
w
images
by
using
the
PCA
algorithm,
lik
e
we
did
the
pre
vious
ORL
database.
Finally
,
the
obtained
eigen
v
ectors
are
classified
by
classification
algorit
hms.
The
accurac
y
of
each
algorithm
is
sho
wn
in
the
follo
wing
Figure
10.
Figure
9.
Curv
elet
iris
image
with
normalization
applied
Figure
10
sho
ws
the
e
xperiment
results
of
recognition
rate
obtained
for
CASIA
iris
images
using
PCA+Curv
elet
for
feature
e
xtraction
step.
The
same
three
weak
classifiers
are
used,
and
we
ha
v
e
achie
v
ed
a
v
erage
recognition
rates
of
92.0%,
93.0%
and
97.0%
respecti
v
ely
KNN,
NN
and
SVM.
Figure
10.
Recognition
rate
for
CASIA
database
with
three
classification
algorithms
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
20,
No.
3,
December
2020
:
1332
–
1341
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1339
3.3.
Comparison
of
experiment
r
esults
In
this
paper
,
the
recognition
system
PCA+curv
elet
ha
v
e
been
used
with
three
classification
algorit
hms
for
comparison.
These
methods
ha
v
e
been
check
ed
both
ORL
and
CASIA
database,
and
the
testing
protocols
used
in
the
e
xperiment
are
almost
the
same.
T
able
1,
sho
ws
a
comparison
of
this
tw
o
recognition
systems.
This
comparison
sho
ws
that
the
best
recognition
rate
on
ORL
database
w
as
presented
for
PCA+curv
elet
using
SVM
with
98.7%,
and
the
best
recognition
rate
on
the
CASIA
database
w
as
presented
using
NN
with
97.0
%.
T
able
1.
Recognition
rate
for
dif
ferent
algorithms
Approaches
Used
algorithms
Database
RecognitionRate
(%)
[24]
PCA+SVM
ORL
90.24
[25]
D
WT+PCA+SVM
ORL
96.00
[26]
PCA+
ANFIS
ORL
96.66
[27]
F
D+Manhattan
Distance
CASIA
96.00
[27]
PCA+Euclidean
Distance
CASIA
92.00
Pr
oposed
appr
oach
DCTG1+PCA+SVM
98.70
DCTG1+PCA+KNN
ORL
97.5
DCTG1+PCA+NN
93.7
//
DCTG1+PCA+SVM
93.00
DCTG1+PCA+KNN
CASIA
91.0
DCTG1+PCA+NN
97.0
4.
CONCLUSION
In
this
paper
,
an
ef
ficient
and
po
werful
f
acial
feature
e
xtraction
approach,
such
as
DCTG1/PCA
i
s
proposed.
The
latter
is
selected
as
a
f
ast
and
strong
technique
in
representing
edges
and
curv
es,
and
reducing
the
dimensionality
of
the
images
f
ace/iris.
As
a
case
study
of
use,
we
presented
an
automatic
2D
f
ace/iris
recognition
system
using
(curv
elet+pca)
feature
e
xtraction
algorithm.in
the
classification
step,
we
ha
v
e
used
the
SVM,
KNN,
NN.
The
results
were
implemented
on
tw
o
well-kno
wn
image
datasets
(ORL
f
ace
database
and
CASIA
iris
database).
The
results
sho
w
that
the
access
speed
feature
e
xtraction
and
the
accurac
y
for
the
recognition
system
of
the
(Curv
elet+pca)
are
more
accurate
than
that
of
the
only
PCA.
REFERENCES
[1]
S.
A.
F
.
A.
I.
M.
Rahman,
S.
Naim,
”Curv
elet
te
xture
based
f
ace
recognition
using
principal
compo-
nent
analysis,
”
Proceedings
of
13th
International
Conference
on
Computer
and
Information
T
echnology
,
pp.
45-50,
2010.
[2]
J.
W
.
Q.
Y
.
Y
.
Mandal,
T
.,
”Curv
elet
based
f
ace
recognition
via
dimension
reduction,
”
Signal
Processing,
pp.
2345-2353,
2009.
[3]
e.
a.
Xu
X.-B.
Zhang
D.-Y
.
Zhang
X.-M.,
”An
algorithm
for
multimodal
biometric
recognition
based
on
feature
le
v
el
and
the
second-generation
curv
elet
transform,
”
Journal
of
Xi’an
Jiaotong
Uni
v
ersity
,
pp.
32-36,
2009.
[4]
B.
A.
Majumdar
,
A.,
”A
comparati
v
e
study
i
n
w
a
v
elets,
curv
elets
and
contour
-lets
as
feature
sets
for
pattern
recognition,
”
Int.
Arab
J.
of
Information
T
echnology
,
2009.
[5]
F
.
P
.
Amira,
A.,
”An
automatic
f
ace
recognition
system
based
on
w
a
v
elet
transforms,
”
IEEE
International
Symposium
on
Circuits
and
Systems,
pp.
6252-6255,
2005.
[6]
W
.
H.
Liu,
C.-J.,
”Independent
component
analysis
of
g
abor
features
for
f
ace
recognition,
”
IEEE
T
rans-
actions
on
Neural
Netw
orks,
pp.
919-928,
2003.
[7]
Z.
R.-C.
Su,
H.-T
.,
”F
ace
recognition
based
on
w
a
v
elet
transform
and
multiple
classifi
er
,
”
Computer
Applications,
pp.
25-27,
2002.
[8]
V
.
M.
Do,
M.N.,
”The
finite
ridgelet
transform
for
image
representation,
”
IEEE
T
rans.
on
Image
Process-
ing,
pp.
16-28,
2003.
[9]
L.
X.
W
.-S.
L.
H.
Li,
A.,
”A
multiscale
and
multidirectional
image
denoising
algo-rithm
based
on
con-
tourlet
transform,
”
Int.
Conf.
on
Intelligent
Information
Hiding
and
Multimedia,
pp.
635-638,
2006.
[10]
G.
P
.
Belbachir
,
N.,
”The
contourlet
transform
for
image
compression,
”
IEEE
Conf.
on
Ph
ysics
in
Signal
and
Image
Processing,
2005.
[11]
C.
A.
D.-M.
Zhou,
J.,
”Nonsubsampled
contourlet
transform:
filter
design
and
application
in
denoising,
”
Proc.
of
Int.
Conf.
on
Image
Processing
(ICIP),
pp.
469-472,
2005.
Biometric
authentication
using
curvelet
tr
ansform
(K
err
ac
he
Soumia)
Evaluation Warning : The document was created with Spire.PDF for Python.
1340
r
ISSN:
2502-4752
[12]
G.
B.
N.-W
.
Y
ang,
L.,
”Multimodality
medical
image
fusion
based
on
multiscale
geometric
analysis
of
contourlet
transform,
”
Neurocomputing,
pp.
203-211,
2008.
[13]
M.-G.
W
.
S.
e.
a.
Jiao,
L.-C.
Gong,
”Adv
ances
in
nat
ural
computation,
”
Machine
Learning
and
Image
Understanding,
pp.
224-243,
2008.
[14]
A.
S.
Nikam,
S.B.,
”Curv
elet-based
fingerprint
anti-spoofing,
”
Signal,
Image
and
V
ideo
Processing,
pp.
75-87,
2010.
[15]
R.
Lakshmi,
A.,
”Ne
w
curv
elet
features
for
image
inde
xing
and
retrie
v
al,
”
Computer
Netw
orks
and
Intel-
ligent
Computing,
pp.
492-501,
2011.
[16]
S.
M.-X.
W
uX.-X.,
Zhao
J.-Y
.,
”F
ace
processing
in
human-computer
interaction
using
curv
elet
analysis,
”
Image
and
Graphics,
pp.
1309-1317,
2010.
[17]
D.
L.-D.
D.
Y
.-L.
Candes,
E.J
.,
”F
ast
discrete
curv
elet
t
ransform,
”
SIAM
Multiscale
Modelling
and
Simu-
lations,
pp.
861-899,
2005.
[18]
J.
F
adili
and
J.
L.
Starck,
”Curv
elets
and
ridgelets,
”
Computational
Comple
xity:
Theory
,
T
echniques,
and
Applications,
pp.
754-773,
2012.
[19]
L.
Y
.
Emmanuel
Candes,
Laurent
Demanet,
Da
vid
Donoho,
”Curv
eLab
http://www
.curv
elet.or
g/.
”
[On-
line].
A
v
ailable:
http://www
.curv
elet.or
g/
[20]
F
.
Bellakhdhar
,
K.
Loukil,
and
M.
Abid,
”F
ace
recognition
approach
using
Gabor
W
a
v
elets,
PCA
and
SVM,
”
International
Journal
of
Computer
Science
Issues,
v
ol.
10,
no.
2,
pp.
201-207,
2013.
[21]
D.
Machanje,
J.
Orero,
and
C.
Marsala,
”A
2D-approach
to
w
ards
the
detection
of
distress
using
fuzzy
K-
nearest
neighbor
,
”
Communications
in
Computer
and
Information
Science,
v
ol.
853,
pp.
762-773,
2018
[22]
Z.
Y
ao
and
W
.
L.
Ruzzo,
”A
re
gression-based
K
nearest
neighbor
algorithm
for
gene
functi
on
prediction
from
heterogenous
data,
”
2018
BMC
Bioinformatics,
v
ol.
7,
2006.
[23]
CASIA
-
Center
for
Biometrics
and
Se
curity
[Online].
A
v
ailable:
http://www
.cbsr
.ia.ac.cn/IrisDatabase.html,
”CASIA
V1.0database.
”
[24]
A.
Bouzalmat,
J.
Kharroubi,
and
A.
Zar
ghili,
”Comparati
v
e
study
of
PCA,
ICA,
LD
A
using
SVM
classi-
fier
,
”
Journal
of
Emer
ging
T
echnologies
in
W
eb
Intelligence,
2014.
[25]
Z.
B.
Laha
w
,
D.
Essaidani,
and
H.
Seddik,
”Rob
ust
F
ace
Recognition
Approaches
Using
PCA,
ICA,
LD
A
Based
o
n
D
WT
,
and
SVM
Algorithm
s,
”
41st
International
Conference
on
T
elecommunications
and
Signal
Processing,
pp.
1-5,
2018.
[26]
R.
Sharma
and
M.
S.
P
atte
rh,
”Pose
In
v
ariant
F
ace
Recognition
using
Neuro-Fuzzy
Approach,
”
IOSR
Journal
of
Computer
Engineering
V
er
.
II,
v
ol.
17,
no.
3,
pp.
2278-661,
2016.
[27]
M.
H.
Hamd
and
S.
K.
Ahmed,
”Biometric
system
design
for
iris
recognition
using
intelligent
al-
gorithms,
”
Int.
Journal
of
Modern
Education
and
Computer
Science,
v
ol.
10,
no.
3,
pp.
9-16,
2018.
BIOGRAPHIES
OF
A
UTHORS
Soumia
K
errache
w
as
born
in
Medea,Algeria
She
recei
v
ed
the
dipl.El-Ing
from
the
uni
v
ersity
of
Medea
,Alger
ia
in
2009,and
master’
s
de
gree
in
Instrumentation
and
microelectronics
from
the
uni
v
ersity
of
Medea
in
2014,currently
she
prepares
the
doctoral
de
gree
Es-science
at
uni
v
er
-
sity
of
bechar
,Algeria.her
main
interests
are
image
processing
,microelectronics,
Embedded
sys-
tems,Correspondence
address:Information
Processing
and
T
elecommunication
Laboratory
(L
TIT),
T
ahri
Mohammed
Uni
v
ersity
,
Bechar
08000,Algeria.
Email:
s
guemana@yahoo.com
Mohammed
Beladgham
w
as
born
in
Tlemcen,
Algeria.
He
recei
v
ed
the
electri
cal
engineering
diploma
from
uni
v
ersity
of
Tlemcen,
Algeria,
and
then
a
Master
in
signals
and
systems
from
Uni
v
er
-
sity
of
Tlemcen,
Algeria
and
the
PhD
de
gree
in
Electronics
from
the
Uni
v
ersity
of
Tlemcen
(Algeria),
in
2012.
He
w
as
an
Associate
Professor
at
the
Uni
v
ersity
of
Bechar
,
Algeria.
Since
2015.
He
is
cur
-
rently
a
Professor
at
Uni
v
ersity
of
Bechar
in
the
department
of
Electical
Engineering,
and
does
his
research
at
the
L
TIT
Laboratory
,
T
ahri
Mohammed
Uni
v
ersity
,
Bechar
.
His
research
interests
are
Image
and
video
processing,
Image
se
gmentation
Medical
image
compression,
Biomedical
imaging,
biometric
systems,
w
a
v
elets
transform
and
optimal
encoder
,
Bechar
08000,Algeria.
Email:
beladgham.tlm@gmail.com
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
20,
No.
3,
December
2020
:
1332
–
1341
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1341
A
ymen
Hamza
w
as
born
in
Bechar
,
Algeria;he
recei
v
ed
bachelor’
s
de
gree
from
the
uni
v
ersity
of
bechar
,currently
he
is
an
automation
engineer
master’
s
student
at
the
uni
v
ersity
of
bechar
.His
re-
search
interests
are
Computer
vision
and
Reinforcement
learning
and
Control
theory
.Correspondence
address:Department
of
Electrical
Engineering,
T
ahri
Mohammed
Bechar
Uni
v
ersity
,
Algeria,Bechar
08000,Algeria.
Email:
aimen
hamza@hotmail.com
Ibrahim
Kadri
is
a
third-year
PhD
student
in
Data
Processing
and
T
elecommunication
at
T
ahri
Mo-
hamed
of
Bechar
,
Algeria.
His
reasearch
interests
are
in
Embedded
Systems
and
T
elecommunication.
He
holds
a
master’
s
de
gree
in
Digital
Communication
Systems
from
the
Uni
v
ersity
of
Bechar
,
Alge-
ria,
in
2017.
Correspondence
address
:
Information
Processing
and
T
elecommunication
Laboratory
(L
TIT),
T
ahri
Mohammed
Uni
v
ersity
of
Bechar
,
Algeria.
Email
:
hayamoto11@gmail.com
Biometric
authentication
using
curvelet
tr
ansform
(K
err
ac
he
Soumia)
Evaluation Warning : The document was created with Spire.PDF for Python.