TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 15, No. 2, August 201
5, pp. 313 ~
320
DOI: 10.115
9
1
/telkomni
ka.
v
15i2.837
7
313
Re
cei
v
ed Ap
ril 17, 2015; Revi
sed
Jun
e
24, 2015; Accepted July 1
0
,
2015
New Method of Content Based Image Retrieval based
on 2-D ESPRIT Method and the Gabor Filters
Cha
w
k
i
Youness*, El As
naoui Khalid
, Ouanan Mo
hammed, Ak
sass
e Bra
h
im
ASIA
T
eam, M2I Labor
ator
y, Dep
a
rtm
ent of Comp
uter Scie
nces,
F
a
cult
y
of Scie
nce an
d T
e
chniqu
es, Moula
y
I
s
mail Un
iversit
y
, Morocco
*Corres
p
o
ndi
n
g
author, em
ail
:
you
ness.ch
a
w
k
i
@gm
a
il.co
m
A
b
st
r
a
ct
W
e
prop
ose, in
this pa
per, a n
e
w
meth
od for
Cont
e
n
t Base
d
Imag
e R
e
triev
a
l (CBIR) by
e
x
ploiti
n
g
the di
gita
l i
m
a
ge co
ntent. O
u
r
metho
d
is
b
a
sed
on
t
he r
epres
entati
on
of the
di
g
i
tal i
m
a
ge
c
onte
n
t by
a
character
i
stics
vector of th
e
ind
e
xe
d
i
m
a
g
e
.
Indee
d, w
e
h
a
ve ex
pl
oite
d
the i
m
ag
e texture to
extract its
character
i
stics
and for c
o
n
s
tructing a
ne
w
descripto
r
vector by co
mb
ini
ng th
e
Bidi
me
nsi
ona
l
Hig
h
Resolution Spectral Analysis
2-D ESPRIT
(Estim
ation
of Signal Parameters vi
a Rotationnal
Invariance
T
e
chni
ques)
method
a
n
d
Ga
bor fi
lter. T
o
e
v
alu
a
te
th
e pe
rforma
nce,
w
e
hav
e
teste
d
our appr
oac
h on
Brodat
z
i
m
age
datab
ase. T
h
e results
sh
ow
that the repr
e
s
ent
atio
n of th
e dig
i
tal i
m
ag
e
content a
p
p
e
a
rs
signific
ant in re
search of i
m
a
g
i
ng infor
m
ation.
Ke
y
w
ords
:
in
d
e
xin
g
, search i
m
a
ges by co
ntent, high res
o
l
u
ti
on, spectra
l
ana
lysis, 2-d e
s
prit, Gabor filter
Copy
right
©
2015 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
The techniq
u
e
s for in
dexi
ng and info
rmation ret
r
ie
val have bee
n develop
ed
sin
c
e the
70s. First, Te
xt-Based Ima
ge Ret
r
ieval (TBIR) me
thods
were the firs
t methods
us
ed. Howev
e
r,
these m
e
thod
s have maj
o
r dra
w
ba
cks: the first is
rela
ted to the am
ount of wo
rk i
n
volved in the
pro
c
e
ss,
and
the se
co
nd i
s
the lack of
a
c
cura
cy in
th
e de
scription
in larg
e data
b
a
se
s. Th
erefo
r
e,
to overco
me this problem, i
t
is important
to dev
elop others syste
m
s using inh
e
re
nt information
of
the conte
n
t. Hen
c
e
Conte
n
t-Based Ima
ge Retri
e
va
l (CBIR) Syste
m
base
d
on t
he image
s di
gital
conte
n
t is de
veloped. Inde
ed, these a
p
p
r
oa
che
s
con
s
ist in rep
r
e
s
e
n
ting every i
m
age by a se
t o
f
visual fe
ature
s
su
ch
a
s
th
e colo
r, the
shape
an
d the
texture. T
h
e
s
e vi
sual
cha
r
acte
ri
stics a
r
e
then used to comp
are and
to find a query image in a databa
se.
Texture i
s
the se
con
d
visual attribute
widel
y u
s
ed i
n
image
sea
r
ch by content
. It allows
filling the gaps that the
color i
s
unable
to
do, especially when t
he color di
stributions
are
very
clo
s
e. More spe
c
ifically, the texture can be view
e
d
as a set
of pixels (grayscale)
spat
ially
arrang
ed in
a numbe
r of
spatial rel
a
tionshi
p
s
, an
d creating a
homoge
neo
us region. T
hus,
several app
ro
ach
e
s a
nd m
odel
s [1, 2] have been p
r
o
posed for m
o
deling the tex
t
ure, whi
c
h
we
quote:
statisti
cal a
p
p
r
oa
ch
es, g
eometri
c a
pproa
che
s
a
nd the
fre
quen
cy ap
proache
s. Fo
r
the
latter, the Ga
bor filters are
the mo
st kno
w
n an
d most
use
d
method.
Gabo
r filters
are i
n
tro
d
u
c
e
d
by
Hun
gari
an Physi
c
i
s
t
Den
n
is Ga
bo
r in
194
6 [3].
They
prove
to b
e
a
n
inte
re
sting t
ool fo
r texture an
al
ysi
s
a
n
d
a
r
e
wid
e
ly
adopte
d
in
th
e literature.
T
h
e
advantag
e of
these filters
is that
their f
unctio
n
in
g
is
clo
s
e to
the
human
visual
treatme
nts,
and
they have the advantage of
being p
r
og
ra
mmable in
fre
quen
cy and i
n
orient
atio
n. Indeed, Ga
bo
r
filters find
th
eir pl
ace in
several a
r
ea
s
su
ch
a
s
:
segmentatio
n
[4], pattern
reco
gnition [5
, 6],
cla
ssifi
cation
[7, 8], con
t
ent based i
m
age
retriev
a
l [9, 10]. The texture
para
m
eters are
determi
ned b
y
calcul
ating the avera
ge a
nd the varian
ce of the ima
ge filtered by Gabo
r filter.
The ai
m for
spe
c
tral
an
al
ysis i
s
to
ch
ara
c
teri
ze
th
e freq
uen
cy
conte
n
t of im
age
s o
r
sign
als in g
e
neral.
We of
ten use, for
image
s,
the freque
ncy inf
o
rmatio
n to extract texture
conte
n
t. The
high
re
solutio
n
spe
c
tral
me
thods called
para
m
etri
c m
e
thod
ca
n b
e
divided i
n
to t
w
o
categ
o
rie
s
: t
he first cont
ains p
s
eud
o-spe
c
tru
m
me
thods while t
he
se
cond
family group
s the
sub
s
p
a
ce me
thods a
nd a
r
e
base
d
on the
exploitati
on
of the inheren
t structu
r
e of
the model. Th
e
2-D ESPRIT
method [11-13] is the most fa
mous and the used method i
n
bidimensional
freque
ncy
e
s
timation. It
provides t
he p
a
ir
f
r
eq
uen
cy an
d i
t
s corre
s
po
n
d
ing
orie
ntation
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 313 –
320
314
automatically with
good preci
s
ion.
These last param
e
ters
will be
i
n
jected into the Gabor filt
er.
Since it is de
pendi
ng on th
eme.
This
work p
r
e
s
ent
s a ne
w
approa
ch for i
ndexing a
nd
conte
n
t base
d
image retrie
val. The
followin
g
sect
ions
are arra
nged
as foll
o
w
s:
Se
ction
2 de
scrib
e
s t
he 2
-
D ESP
RIT metho
d
,
and
then the G
a
bor filters a
r
e pre
s
e
n
ted
in Sectio
n
3.
Section 4
e
x
plains our new app
roa
c
h.
Subse
que
ntly the experim
e
n
tal re
sults a
r
e pre
s
e
n
t
ed i
n
Section 5.
Finally we
co
nclu
de ou
r work
with co
ncl
u
si
ons a
nd pe
rspective
s.
2.
The 2-D ESPRIT Method [12]
The fre
quen
cy extracted b
y
the 2-D ES
PRIT me
thod
is in the form of two co
mpone
nts
whi
c
h a
r
e a
freque
ncy
co
uple: one
co
mpone
nt is
on the first frequ
en
cy axis and th
e ot
her
comp
one
nt is on the
se
cond fre
que
ncy axis; theref
ore the f
r
equ
ency mo
dule
is define
d
a
s
follow:
2
2
2
1
i
i
i
f
f
F
(
1
)
The co
rrespo
nding rotation
angle
i
is giving by:
)
arctan(
1
2
i
i
i
f
f
(
2
)
Figure 1
depi
cts th
e relatio
n
shi
p
bet
wee
n
the frequ
en
cy compo
nen
ts
i
f
1
,
i
f
2
, and the
modul
e
i
F
, and rotation
angle
i
:
Figure 1. Rel
a
tionship bet
wee
n
the two
compo
nent
s and the rotati
on angl
e
Thus ESPRI
T
2-D provides
us th
e frequencies and its orientat
ions that will
be the input
para
m
eters o
f
Gabor filters.
In our
ca
se,
the num
ber of frequ
en
ci
es
i
F
and
ori
entation
s
i
are the
same
as
numbe
r of the freque
nci
e
s compo
nent
s.
3.
The Gabo
r F
ilters
Gabo
r
filters are multi-cha
nnel
filteri
ng t
e
ch
niqu
es th
at allo
w the
d
e
scriptio
n of t
e
xtures
locali
ze
d in
freque
ncy
an
d o
r
ientation.
In othe
r
word
s, the
ch
ara
c
teri
stic calcul
ation
s
a
r
e
operated o
n
each of the
pixels. The
s
e
contai
n t
he intensity
vari
ations on
a smalle
r scale
of
freque
ncy an
d orientatio
n. The gab
or
fu
nction
can b
e
written a
s
follows:
)
2
cos(
))
(
2
1
exp(
)
,
,
,
(
2
2
2
2
fx
y
x
f
y
x
G
y
x
(
3
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
New Method
of Content Based Im
age Retrie
val based on 2-D ESPRIT
… (Chawki Youness)
315
Whe
r
e:
cos
sin
sin
cos
y
x
y
y
x
x
(
4
)
is the ori
enta
t
ion,
F
the frequ
ency an
d
x
(resp
e
ctively
y
)
the stand
ard deviation
of the gaussian acco
rdin
g the x axis (y axis
respectively). By applying this fun
c
tion as a
convol
ution
mask, we defi
ne a co
nvol
ut
ion filter calle
d Gabo
r filter.
For a give
n frequ
en
cy
F
an
d a given o
r
i
entation
, we have a Ga
bo
r filter, allowi
ng
towards the e
nd to have
several filters
by cha
ngin
g
the o
r
ientation
and al
so th
e
frequ
ency, a
n
d
finally to have a set of filt
ers call
ed ba
nc
of filters.
After con
s
tru
c
ting the
Ga
bor filters, th
ey are a
pplie
d to a
MxN
ima
ge in o
r
de
r to
extract the a
v
erage a
nd the varian
ce
of f
iltered image by ea
ch
Gabo
r. The
average a
n
d
varian
ce hav
e the followin
g
expre
ssi
on
s re
spe
c
tively:
M
x
N
y
x
G
xy
n
m
n
m
)
,
(
,
,
(
5
)
MxN
y
x
G
xy
n
m
n
m
n
m
2
,
,
,
)
)
,
(
(
(
6
)
The value
s
of
n
m
,
and
n
m
,
will re
prese
n
t the co
mpone
nts
of our de
scri
ptor vector.
4.
The Propos
e
d
Appro
ach
Here, we p
r
o
pose a ne
w method for in
dexing an
d content ba
sed
image retri
e
val using
the bidimensi
onal high resolution spectral analys
i
s
m
e
thods especially
the 2-D ESPRIT method
and
Gab
o
r filt
ers. In
dee
d,
each ima
ge i
s
cha
r
a
c
teri
zed by th
e pai
r given
by the
Gab
o
r filte
r
s
for
each o
r
ientati
on a
nd
ea
ch f
r
equ
en
cy, this la
st p
a
ra
me
ters are give
n
by the
2-D E
SPRIT metho
d
applie
d on th
e origin
al ima
ge.
Indexing a
n
d
conte
n
t ba
sed imag
e retrieval
sy
stem
inclu
d
e
s
two main
pha
ses, the
indexing an
d
the sea
r
ch ph
ase, Figu
re 2
sho
w
s these two ph
ases.
Figure 2. Overall archite
c
tu
re
of co
ntent based imag
e retrieval
The first invol
v
es extra
c
tin
g
the
cha
r
a
c
t
e
risti
c
s of e
a
c
h im
age
an
d they a
r
e
st
ored
in
a
databa
se. Thi
s
step i
s
ru
nn
ing in Offline.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 313 –
320
316
The seco
nd
pha
se is run
n
ing in Onli
n
e
, the syste
m
take
s a q
uery imag
e, extract its
digital
cha
r
a
c
teri
stics, an
d then
mea
s
ur
e the
di
stance
bet
we
en the
que
ry image ve
ctor
descri
p
tor a
n
d
those in the
databa
se.
The main
steps in features extraction of
textures image usi
ng the
2-D ESPRIT
method
and Ga
bor filters a
r
e illu
strated and
sum
m
ari
z
ed in Fi
gure 3.
Figure 3. The
Main step
s o
f
the pr
opo
se
d approa
che
d
are cited b
e
ll
ow
1) Fr
eque
nc
y
pair and orienta
t
ion
extr
action
:
The first ste
p
is to extract
from the
original image by using the 2-
D ESPRIT method the frequency pairs
)
,
(
2
1
i
i
f
f
. Further
m
o
re, w
e
cal
c
ulate th
e
frequ
en
cy module
and
the ori
entat
io
n.
The
s
e
p
a
rameters will be
inje
cted
i
n
to
Gabo
r filters.
2) Av
erage
calculation
:
After applyin
g
the G
abo
r
filters o
n
the
image,
with
different
orientatio
ns
m
and fre
que
n
c
y
m
f
, we calculate the av
erag
e
m
th
a
t
c
h
ar
ac
te
r
i
ze
s
th
e
luminou
s inte
nsity of the image corre
s
p
ondin
g
to t
he averag
e grayscale of all pi
xels in the image
(Equatio
n 5).
3) S
t
and
a
rd
dev
i
ation calculation
:
T
h
e s
t
an
d
a
r
d
de
via
t
io
n
m
ch
ara
c
te
rize
s the
variation of th
e average i
n
tensity of all
pi
xels.
It corre
s
pond
s to the
cha
nge
of the
image
co
ntra
st
(Equatio
n 6).
After extrac
ting the charac
teris
t
ics
of e
a
ch
imag
e,
we con
s
tru
c
t th
e vecto
r
de
scriptor in
whi
c
h ea
ch
compon
ent re
pre
s
ent
s two
values: the a
v
erage a
nd th
e varian
ce. In other wo
rd
s, for
each freq
uen
cy and e
a
ch
orientatio
n, we have a p
a
ir (
,
). Finally, the descriptor vector
will
be define
d
as follows:
)
,
,.....,
,
,
,
(
1
1
0
0
m
m
V
(
7
)
The algo
rithm
of our app
ro
ach i
s
the followin
g
:
Algorithm
: 2
-
D_ESP
RIT_
Gabo
r metho
d
1:
D: datab
ase
2: For
image i in D
do
3:
i
image
read
C
)
(
s
,
i
ize
N
M
4:
)
(
_
2
,
,
,
,
,
3
3
2
2
1
1
C
ESPRIT
D
F
F
F
5
3
3
2
2
1
1
,
,
,
,
,
GaborFiltr
F
F
F
A
6:
i
A
n
Convolutio
FFG
,
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
New Method
of Content Based Im
age Retrie
val based on 2-D ESPRIT
… (Chawki Youness)
317
7:
N
a
M
b
b
a
FFG
N
M
Moy
11
,
*
1
8:
2
11
)
,
(
*
1
Moy
b
a
FFG
N
M
N
a
M
b
9:
Moy
DV
,
12:
Return DV
13: EndFo
r
The simil
a
rity measu
r
e bet
wee
n
a que
ry image
Q and
a target image T is ca
rri
e
d
out by
cal
c
ulatin
g, for each value
of the ve
ctor, of the Swain distan
ce [14]
)
,
(
T
Q
D
defined by:
)
,
(
)
,
(
T
Q
d
T
Q
D
m
(
8
)
Whe
r
e,
)
(
),
(
T
m
Q
m
T
m
Q
m
m
d
(
9
)
5.
Experemen
t
al Results
In our expe
riment, we
p
e
rform
detail
ed pe
rforma
nce
co
mpa
r
i
s
on
to evalu
a
te the
efficien
cy of
the p
r
opo
se
d
app
roa
c
h.
Indee
d,
we h
a
ve teste
d
o
u
r
app
roa
c
h
on the
texture
Brodat
z data
base [15] in
cludi
ng 13 t
e
xture cl
a
sses, 16
sam
p
les e
a
ch. All image
s are
in
grayscal
e BMP format, each 256 by
256 pixels. Figure 4 sho
w
s exam
ple
s
of images from
Brodat
z data
base.
Figure 4. Examples of ima
ges fro
m
Bro
datz data
b
a
s
e
5.1. Implementa
tion
In the implem
entation ph
ase, we have
u
s
ed
a
co
mput
er, Pro
c
e
s
sor:
Intel(R) Core
(TM)
2 CPU T5
870
@ 2,00 G
H
z,
2,00 GH
z, 2Go RAM, Win
dows 7.
We have
dev
elope
d an u
s
er graphi
c int
e
rface for
di
splaying the re
sults.
We ju
st display
the 14 image
s simila
r to the query ima
g
e
.
Figure 5 sho
w
s a
n
exam
ple of query
image (D16
_001 ima
ge),
our app
roa
c
h ha
s
returned
14 relevant simil
a
r imag
es
whil
e figure
6 illu
strate
s an
an
other exa
m
pl
e of que
ry im
age
(D8
4_0
01
) with 12 imag
e
s
simil
a
r im
a
ges in
clu
d
ing
two not rele
vant image
s
(D9
4_0
01 a
n
d
D38
_01
0).
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 313 –
320
318
Table1. Th
e pre
c
isi
on en
d
the recall for each cla
s
s on
Brodatz ima
g
es d
a
taba
se
Classes
Fou
nd i
m
ag
es
Precison
(
%
)
Recall (
%
)
D9 169
68,75
60,16
D12
163
71,43
62,50
D15
98
70,09
61,89
D16
55
77,68
87,97
D19
15
84,38
65,23
D24
127
78,57
68,75
D29
174
89,29
78,13
D38
163
77,23
67,58
D68
38
100,00
87,50
D84
167
77,68
64,09
D92
178
82,14
71,88
D94
165
81,25
71,09
D112
167
83,93
73,44
A
v
e
r
age Preci
si
on (
%
)
A
v
e
r
age
Recall
(
%
)
80,19
69,24
The Tabl
e 1 sho
w
s the preci
s
ion a
nd recall of
ea
ch
cla
ss of Bro
d
a
tz datab
ase. We ca
n
see th
at the
pre
c
isi
on i
s
more th
an 7
0
%
for all cl
asse
s an
d the
averag
e p
r
e
c
ision i
s
e
q
ual
to
80,19%, whi
c
h sho
w
s that our ap
proa
ch
is accu
rate.
Figure 5. Similar imag
es fo
r D16
_00
1 cl
ass
Figure 6. Similar imag
es fo
r D84
_00
1 cl
ass
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
New Method
of Content Based Im
age Retrie
val based on 2-D ESPRIT
… (Chawki Youness)
319
5.2. Ev
aluation Protocol
In the i
n
form
ation
retrieval
syst
em, the
use
r
i
s
i
n
tere
sted to
releva
nt sy
stem
re
spon
se
s.
So the info
rm
ation
sea
r
ch
system
s
req
u
i
re the
p
r
e
c
isi
on eval
uation
of the
re
spo
n
s
e. T
h
is type
of
evaluation i
s
con
s
id
ere
d
as the
re
search p
e
rfo
r
mance evalu
a
tion. Thu
s
,
to evaluate
the
efficien
cy of our sy
stem, we are intere
sted to
cal
c
ul
ate the two most com
m
onl
y used mea
s
ure
s
.
Namely: the
pre
c
isi
on a
n
d
the recall,
these t
w
o p
a
ram
e
ters m
easure th
e relevan
c
e of
our
sy
st
em.
Rec
a
ll
is the
ratio betwee
n
the numb
e
r of relevant i
m
age
s in the
set of image
s found
and the num
b
e
r of releva
nt image in the i
m
age data
b
a
s
e.
R
R
a
Recall
(
1
0
)
Whe
r
e:
a
R
: Numbe
r
of relevant imag
es in the set of resp
on
se
s.
R
: Number of relevant imag
es in the ima
ge datab
ase.
Precision
is the ratio
be
tween th
e nu
mber
of rel
e
vant image
s
in the set of
image
s
found an
d the
numbe
r of re
levant image.
A
R
a
Precision
(
1
1
)
Whe
r
e:
A
: Numbe
r
of image
s in the set of re
spon
se
s.
The curve
7
and 8
sh
ow t
he preci
s
io
n
averag
e an
d
the re
call av
erag
e respe
c
tively
o
f
our app
roa
c
h
com
pared wi
th
the
Ga
bo
r method. We
see th
at ou
r
method i
s
m
o
re a
c
curacy t
han
the Gabo
r me
thod.
Figure 7. Pre
c
isi
on curve o
f
our app
roa
c
h
comp
ared wit
h
the Gabo
r
Filter
Figure 8. Re
call curve of o
u
r app
ro
ach
comp
ared wit
h
the Gabo
r
Filter method
6.
Conclu
sion & Perspe
ctiv
es
In this work, we have p
r
e
s
ented a ne
w
method for in
dexing and
content ba
sed
image
retrieval
usin
g Gab
o
r filte
r
s p
o
wered
or feed
by the freq
uen
cy
and o
r
ientat
ion given by
the
bidimen
s
io
nal
high resoluti
on spe
c
tral a
nalysi
s
e
s
pe
cially 2-D ESP
RIT. This
app
roa
c
h i
s
ap
pli
e
d
to the Brod
at
z data
b
a
s
e
containin
g
20
8
image
s a
nd
has
given g
o
od results ma
king it p
o
ssibl
e
to
say that our a
ppro
a
ch is a
c
curate.
In terms of p
e
rspe
ctives,
we
will try to apply ou
r ap
proa
ch i
n
itiall
y on image d
a
taba
se
s
other than te
xtured datab
ase
s
, and the
n
we will co
m
b
ine the Hig
h
Resolution Spectral Analysis
method
s with
other visual
cha
r
a
c
teri
stics su
ch a
s
the
shap
e and
color.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 313 –
320
320
Referen
ces
[1]
J Z
han
g, T
T
an. Brief r
e
vie
w
of
i
n
vari
ant te
xture
a
nal
ys
is
methods.
Pattern Recognition
. 20
02;
35:
735-
747.
[2]
D Gabor. T
heor
y
of commu
nic
a
tion.
Jour
na
l of the Institute
of Electrical En
gin
eers
. 19
46; 93: 429-
45
7.
[3]
J Z
han
g, T
T
an, L Ma.
Invariant Texture Se
gm
entation Via
Circular Gabor
Filters
. Pattern
Recognition,
Procee
din
g
s 1
6
th Internati
o
n
a
l
Co
nferenc
e. 200
2; 2: 901-9
04.
[4]
CS Avil
a, RS
Reil
lo. T
w
o
different a
ppr
oac
hes for
ir
is
rec
ogn
ition usi
ng Gabor
filters
a
nd mu
ltiscal
e
zero-crossi
ng r
epres
entati
on.
Pattern Reco
g
n
itio
n
. 200
5; 38: 231-2
40.
[5]
K T
hangair
u
la
ppa
n, JBV Je
yasin
gh. F
a
ce
Repr
es
entati
o
n
Usin
g Com
b
i
n
ed Me
th
od of Gabor
F
ilters
,
W
a
velet T
r
ans
formation
an
d
DCV
an
d R
e
cog
n
itio
n Us
i
ng R
B
F
.
Jour
nal
of Intel
lig
ent L
earn
i
n
g
Systems a
nd A
pplic
atio
ns
. 20
12; 4: 266-
273.
[6]
S Arivazha
gan
, L Ganesan, SP Pri
y
a
l
.
T
e
xture classificati
on usi
ng Gab
o
r
w
a
vel
e
ts ba
sed rotatio
n
invari
ant featur
es.
Pattern Re
cogn
ition
Lette
rs
. 2006; 27: 1
976-
198
2.
[7]
Rahm
ad
w
a
t
i
,
G Nag
h
d
y
, M
Ros, C T
odd,
E Nor
ahma
w
a
t
i.
Cervic
al Ca
n
c
er
Cl
assific
a
ti
on Usin
g
Ga
bo
r
Filters
. Health
care Informati
cs, Imaging a
nd S
y
stems
Biol
og
y
(
H
ISB
)
, F
i
rst
IEEE
International
Confer
ence. 2
011: 48- 5
2
.
[8]
N Bourkac
he,
S Sidh
om, M Lag
hrouc
he.
A
ppre
n
tissag
e
n
u
mériq
ue
pour
la rech
erche
d
’
inf
o
rmatio
n
s
en i
m
ag
eri
e
médic
a
le: Mod
é
li
sation d
e
s filtre
s de Gabor
. Internati
ona
l s
y
m
posi
u
m ISKO-Maghr
eb
̓
13
,
Conc
epts an
d T
ools for Kno
w
led
ge Man
a
g
e
m
ent (KM). Marrakech, Moroc
c
o. 2013.
[9]
M Agar
w
a
l, RP Mah
e
sh
w
a
ri. Cont
ent
Based Ima
g
e
Retrieva
l Ba
sed o
n
Lo
g
Gabor W
a
ve
le
t
T
r
ansform.
Ad
vance
d
Materi
als Res
earch,
T
r
ans T
e
ch Pu
blicati
ons
. 2
0
1
2
: 871-8
78.
[10]
Y Hua. Estima
ting T
w
o-D
i
me
nsio
nal F
r
e
q
u
e
n
cies
b
y
Matri
x
En
ha
ncem
en
t and Matri
x
P
ence
l
.
IEEE
T
r
ansactio
n
s o
n
Sign
al Proc
e
ssing.
19
97; 40
(9).
[11]
P Vanp
ouck
e
, M Moon
en, Y
Berthoum
ieu.
An efficie
n
t su
bspac
e al
gor
ithm for
2D
har
mo
nic retri
e
va
l
.
IEEE Internatio
nal C
onfere
n
ce
on Acoustics,
Speec
h an
d Si
gna
l Processi
n
g
. 1994: 4
61-4
64.
[12]
Y Z
hou, DZ
F
eng, JQ Liu.
A novel alg
o
r
ithm for t
w
o-
dime
nsio
nal fr
equ
enc
y estim
a
tion.
Si
gn
al
Processi
ng.
20
07; 87: 1-1
2
.
[13]
S Ro
uq
uette,
M Naj
i
m. F
r
eq
uenc
ies
an
d
d
a
mpi
ng f
a
ctors
b
y
t
w
o
-
dim
e
n
s
ion
a
l ES
PRIT
t
y
pe m
e
tho
d
s.
IEEE Transactions on
Signal Processing.
20
01; 49(1): 2
37-
245.
[14]
Y Berthoum
ie
u
,
M El Ansari,
B Aksasse, M
Doni
as
, M Na
ji
m. A 2-D Ro
bu
st high r
e
sol
u
ti
on freq
ue
nc
y
estimatio
n
ap
p
r
oach.
Sig
n
a
l
Processi
ng
. 20
0
5
; 85(6): 11
65-
118
8.
[15] USC-SIPI
imag
e datab
ase. Av
aila
bl
e onl
in
e a
t: http://sipi.usc.edu/dat
abas
e/. 1977
Evaluation Warning : The document was created with Spire.PDF for Python.