TELKOM
NIKA
, Vol.12, No
.3, Septembe
r 2014, pp. 5
97~604
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v12i3.80
597
Re
cei
v
ed Ma
rch 4, 2
014;
Re
vised J
une
14, 2014; Accepte
d
Ju
ne
30, 2014
Batik Image Retrieval Based on Color Difference
Histogram and Gray Level Co-Occurrence Mat
r
ix
Agus Eko Mi
narno*
1
, Nan
i
k Suciati
2
1
T
e
knik Informatika, Univ
ersit
a
s Muhamm
a
d
i
yah Ma
lan
g
, Jl.Ra
y
a T
l
ogoma
s
No. 246 Mal
a
ng, Kampus
3
Gedun
g Kul
i
ah
Bersama 3, T
e
lp. 03
41-
464
3
1
8
2
T
e
knik Informatika, Institut
T
e
kno
l
og
i Sep
u
l
uh No
pemb
e
r Jl.
T
e
knik Kimi
a, Gedun
g T
e
knik Informatika,
Kampus IT
S Sukoli
lo, Sura
ba
ya, 6
011
1 T
e
lp: 031 – 59
392
1
4
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: aguseko
@
u
mm.ac.id
1
, nanik@if.its.ac.id
2
A
b
st
r
a
ct
Study in b
a
tik imag
e retrieva
l
is still chal
le
n
g
in
g today. On
e of the metho
d
s for this pro
b
le
m is
usin
g a
Col
o
r
Differenc
e Hist
ogra
m
(
CDH),
w
h
ich is b
a
se
d
on th
e differ
e
nce of c
o
lor fe
atures a
nd
ed
ge
orie
ntation f
eat
ures. How
e
ver,
CDH is
only
u
t
ilisin
g l
o
cal
fe
atures i
n
stea
d
of glo
bal
f
eatur
es; conse
q
u
e
n
t
ly
it cann
ot repr
e
s
ent i
m
a
ges
gl
oba
lly.
W
e
sug
gest that by
a
ddi
ng g
l
o
bal fe
atures for b
a
ti
k imag
e retriev
a
l,
precisi
on w
ill i
n
creas
e. Ther
efore, in this
study, w
e
combin
e the
use
of mo
difie
d
C
DH to defin
e l
o
cal
features
an
d th
e us
e of
Grey
Leve
l
C
o
-occur
r
ence
Ma
trix (
G
LCM) to
defi
ne
glo
b
a
l
feat
u
r
es. T
he
modifi
e
d
CDH is
perfor
m
e
d
by c
h
a
ngi
ng the
si
z
e
of imag
e qu
antis
ation, so
it can
reduc
e
the
nu
mb
er of feat
ur
es
.
F
eatures
th
at are detecte
d b
y
GLCM
are
e
nergy, entro
py,
contrast an
d co
rrelati
on. I
n
t
h
is stu
d
y, w
e
use
300 b
a
tik i
m
a
g
e
s w
h
ich cons
i
s
t of 50 classe
s and six i
m
a
g
e
s in eac
h clas
s. T
he experi
m
ent result sh
o
w
s
that the
pro
pos
ed
metho
d
is
a
b
le
to ra
ise
96.
5%
of th
e
prec
i
s
ion
rate w
h
ich
is
3.5%
hi
gher
than
the
us
e o
f
CDH on
ly. The propos
ed
met
hod is extracti
n
g
a sma
ller
n
u
m
b
e
r of featur
es; how
ever it perfor
m
s better
for
batik i
m
a
ge ret
r
ieval. T
h
is i
ndi
cates that the
use of GLCM i
s
effective combin
ed w
i
th CD
H.
Ke
y
w
ords
: ba
tik, imag
e retri
e
val, col
o
r diffe
rence
h
i
stogra
m
, gray lev
e
l c
o
-occurr
ence
matrix
1. Introduc
tion
On 2 O
c
tobe
r 200
9, UNE
SCO an
noun
ced that bati
k
is on
e of the
intangi
ble
cultural
heritag
es of
humanity. Batik can be d
e
fined as a tra
d
itional meth
od to write so
me pattern
s and
dots on fa
bri
c
and
other
material. Bati
k patterns
consi
s
t of on
e or mo
re
motifs whi
c
h
are
repe
atedly written in an o
r
de
rly sequ
e
n
ce o
r
a
disorde
r
ly seq
u
ence. Based
on these m
o
tifs,
batik images
can be
classi
fied in
order to facilitate the document
at
ion. Several studies on batik
have be
en p
r
opo
sed
su
ch
as, u
s
ing
a cardin
al spi
ne
curve
to extract bati
k
feat
ure
s
by Fa
na
ni
[1], and
conte
n
t-ba
sed
ima
ge retrieval
u
s
ing
an
en
h
a
n
ce
d mi
cro st
ructu
r
e
de
scri
ptor by
Mina
rno
[2]. Content-b
ase
d
ima
ge
retrieval (CBI
R) i
s
a me
th
o
d
u
s
ed fo
r
se
arching
rel
e
vant imag
es from
a colle
ction o
f
images. The
r
e are many tech
niqu
es u
s
ed for conten
t-based ima
g
e
retrieval
su
ch
as Z
e
rnike m
o
ment [3],
Li,
improving
Relevan
c
e F
e
e
dba
ck [4], bi
partite g
r
a
p
h
model
[5], a
n
d
discre
et co
si
nes d
o
main [
6
].
Usually, th
ere i
s
a calcu
l
ation of dista
n
ce b
e
twe
en
an imag
e qu
ery
and a
data
set. Currently,
the calculati
on la
rgely util
ize
s
featu
r
e
s
su
ch
a
s
col
o
r,
texture,
sh
a
p
e
and sp
atial la
yout. To be
more
spe
c
ific, severa
l pre
v
ious re
se
arches were based on col
o
r a
nd
texture only. For in
stan
ce there i
s
GL
CM which
is propo
sed to extract featu
r
e
s
based on the
co-
occurre
n
ce of grey intensi
t
y [7] and edge dire
cti
on
histog
ram (E
DH) whi
c
h is used to retri
e
ve
logo im
age
s [8]. However, EDH i
s
invariant fo
r rotated,
scaled a
nd tra
n
slate
d
ima
g
e
s.
Manjun
ath al
so p
r
op
oses
a method
kn
own a
s
e
dge
histog
ram d
e
scripto
r
(E
HD) [9]. Anoth
e
r
approa
ch i
s
p
r
opo
se
d by
Julesz u
s
in
g t
e
xton to
re
co
gnize texture,
based
on a
grey
scale im
age
[10]. Guang
-Hai Liu
and
Jing-Yu Ya
ng
[11] also u
s
e
five types of texton combi
ned with
GL
CM
feature
s
, such as energy,
contra
st,
entropy and homogenity with angles of 0
°
, 45
°
, 90
°
a
nd
135
°
. The fea
t
ures
are
extracted from RGB colo
r
spa
c
e. Thi
s
met
hod is
kn
own
as texton co-
occ
u
rrence matrix (TCM).
Two yea
r
s l
a
ter, in 2010
Guang
-hai
Liu
et al. [1
2] improved
the TCM me
thod by
prop
osi
ng a n
e
w feature extraction p
r
o
c
e
ss u
s
in
g
a m
u
lti-texton histogram
(MT
H
). Comp
are
d
to
TCM, the mo
st signifi
cant
improvem
ent
of MTH is
a
faster featu
r
e extractio
n
pro
c
e
ss th
at can
avoid overla
p
p
ing. Thi
s
is
facilitated by
the
use
of four textons in
stead of o
ne
texton and two
pixel frictio
n
s
from left to
ri
ght an
d fro
m
top to b
o
ttom
.
In 201
1, Gu
ang-Hai
Liu
e
t
al. pro
p
o
s
ed
a
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 59
7 – 604
598
micro-structure de
scriptor
(MSD) [13] whi
c
h
utilizes various si
zes of filters for micro-struct
ure
map dete
c
tio
n
. They used
2x2, 3x
3, 5x5 and 7x7 filt
er si
ze
s for t
h
is p
u
rp
ose. Usi
ng a p
a
rti
c
ula
r
filter, they ide
n
tified co-o
ccurren
ce val
u
e
s
from the
ce
ntre of t
he filter to
othe
r p
a
r
ts
of the filte
r
.
The d
e
tecte
d
features a
c
tually are q
u
a
n
tized
valu
e
s
of colo
r inte
nsity an
d e
d
ge o
r
ientatio
n on
HSV col
o
r
sp
ace. T
he fri
c
ti
on filter of M
S
D is
den
ot
e
d
by four
sta
r
ting point
s at
(0, 0),
(0, 1
)
,
(1,
0) a
nd
(1, 1
)
.
The
different
startin
g
p
o
in
t purp
o
se i
s
t
o
avoid
missed pixel
s
. Co
mpared to
T
C
M
and MT
H, MSD uses th
ree pixel fricti
ons in
stea
d
of one or t
w
o pixel frictio
n
s to dete
c
t
co-
occurre
n
ce pi
xel intensity.
Colo
r
Differe
nce
Hi
stog
ra
m (CDH) is
a modifi
catio
n
of MSD o
b
tained
by p
e
rformi
ng
colo
r differen
c
e to e
dge
map an
d ed
ge ori
entatio
n to col
o
r
a
map [14]. Ho
wever,
CDH
only
utilizes local f
eatures to represent
images. In fact, the use of glob
al information is also important
for this purpose. One of the met
hods for global features
extraction that is still reliable i
s
Grey
Level Co-o
ccurren
ce M
a
tri
x
(GL
C
M). T
herefo
r
e, in
t
h
is study
we combi
ne
the
use of
CDH as
a
local featu
r
e
s
extractor a
n
d
the use of
G
L
CM a
s
a glo
bal feature ex
tracto
r.
2. Datase
t
Batik ima
g
e
dataset is
co
llected
by
ca
pturing
5
0
types of
bat
ik fabric.
Each fabric
i
s
captu
r
ed to a
s
much as si
x random im
age
s and the
n
resi
ze
d to 128x12
8 pixels size in JP
EG
format. Thu
s
the total nu
mber
of ima
ges i
n
a d
a
taset i
s
30
0
and
con
s
i
s
ts of 50 cl
asses.
Examples of batik imag
es
are sho
w
n in Figure 1.
In g
eneral, there
are two patte
rns of captured
batik imag
es;
geometri
c an
d non-geo
me
tric patterns.
Figure 1. Example of Batik image
s
3.
Color Differ
e
n
ce His
t
ogr
a
m
CDH
co
nsi
s
t
s
of five
mai
n
sectio
ns.
Fi
rst, tra
n
sfo
r
m
i
ng a
n
RGB i
m
age to
L*
a*
b*
color
spa
c
e. Se
cond, dete
c
ti
ng ed
ge o
r
ientation
a
nd qu
antizi
n
g the valu
es to
m
bins,
m
=
{6,12,18,2
4
,30,36}
. Third, quantizin
g each com
pon
ent of the image
L*a*
b*
color into
n
bins,
n={
54,72,1
0
8
}
. Furth
e
rm
o
r
e, calculatin
g the diffe
re
nce
of the
colo
r q
uanti
z
ation an
d e
dge
orientatio
n m
ap; and the
differen
c
e of
edge q
uant
i
z
ation and th
e
colo
r inten
s
i
t
y map. Final
ly,
combi
n
ing th
e colo
r histo
g
r
am an
d edg
e orientatio
n histog
ram.
3.1
RG
B image to
L*a*b*
Color Space
RGB
colo
r sp
ace i
s
a
com
m
on colo
r sp
ace th
at is u
s
ed in g
ene
ral
appli
c
ation.
Thoug
h
this col
o
r
spa
c
e is
simpl
e
and st
raig
htforwa
r
d,
it ca
n
not mimic h
u
man color
pe
rce
p
tion. On
the
other si
de,
L*
a*b*
col
o
r sp
ace was d
e
si
gned to be p
e
rceptually u
n
iform [15], and highly uniform
with respe
c
t to hum
an
colo
r pe
rception,
thus
L
*
a*b*
color sp
ace
is particula
rly
a better choi
ce in
this case. In this sectio
n, we a
r
e transf
o
rmin
g an
RGB image to
an XYZ col
o
r spa
c
e b
e
fore
transfo
rmin
g to an
L*a*
b*
color
spac
e.
3.2
Edge Orientation Dete
cti
on in
L*a*b*
C
o
lo
r
Sp
a
ce
Edge o
r
ientat
ion play
s an
i
m
porta
nt rol
e
and
i
n
fluen
ce in h
u
man i
m
age
perce
ption. Thi
s
can
provide
a de
scriptio
n
of obje
c
t bo
unda
rie
s
. Thi
s
al
so
provid
es
sem
antic i
n
formatio
n of
an
image an
d re
pre
s
ent
s feat
ure
s
of texture and shap
e simultan
eou
sl
y. In this paper we cal
c
ula
t
e
edge
ori
entati
on for ea
ch
compon
ent of
L*a*b*
color
space. Comm
on e
dge
ori
e
ntation d
e
tect
ion
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Batik Im
age Retrieval Based on Color
Differenc
e
Hi
stogram
and Gray .... (Agus Eko Minarno)
599
is sta
r
ted by
conve
r
ting a
colo
r imag
e
to a gr
ey
-scale imag
e. T
hen, gradie
n
t magnitud
e
a
nd
orientatio
n are detected b
a
se
d on
the grey-scale im
age. Ho
weve
r, this metho
d
suffers fro
m
the
loss of a num
ber of chrom
a
tic inform
ation.
To p
r
event th
e lo
ss of
ch
romatic inform
ation,
we utili
sed
the
Zen
z
o [16] m
e
tho
d
to g
e
t
the g
r
adie
n
t
from the
col
o
r im
age. In
this
se
ction,
ea
ch
L*a*
b*
co
mpon
ent
is d
e
tecte
d
with
hori
z
ontal
ori
entation
(g
xx
)
and ve
rtical
orientatio
n
(gy
y
)
,
an
d then
the compute
d
dot p
r
od
uct
of
gxx and gyy, resulting in
gxy, using Sobel op
erat
o
r
s. Furthe
rmo
r
e, the gra
d
i
ent con
s
i
s
ts
of
magnitud
e
co
mpone
nt and
directio
n. In
orde
r to
get the maximum rate of orient
ation cha
nge
of
gradi
ent, the followin
g
form
ula is u
s
ed:
,
ar
ctan
(1)
Whe
r
e the rate of orientatio
n cha
nge for
,
given by equ
ation (2
) and
(3)
1
,
cos
2
2
2
/
(2)
2
,
cos
2
2
2
/
2
/
(3)
More
over, we
need to find the maximum
value of gradi
ent (
Fm
ax
)
max 1,
2
(4)
Whe
r
e:
1
,
2
,
(5)
After computi
ng the edge orientatio
n for each pixel
,
, then we qu
ant
ise the
,
into
m
bins, with variatio
n
m
=
{6,12,18,
24,30,36
}
. Th
e interval
size of each bin
is cal
c
ul
ated
by
dividing 36
0
with
m
. For e
x
ample, for
m=
6
, the inte
rval si
ze i
s
6
0
, so all
edg
e orie
ntation
s
are
uniformly qu
a
n
tized to intervals of 0, 60, 120, 180, 24
0
,
300.
3.3 Quan
tiza
tion
L*a*b*
Color Space
Colo
r is
one
of the impo
rtant asp
e
ct
s
for im
ag
e ret
r
ieval, be
cau
s
e it provide
s
hig
h
ly
reliabl
e spati
a
l inform
ation. This aspect
is usua
lly pe
rforme
d a
s
fe
ature
s
in
a color
histo
g
ra
m.
Comm
on ap
p
lication
s
typically use RGB color
spa
c
e
f
o
r practi
cal re
aso
n
s, ho
we
ver, this ca
nn
ot
rep
r
e
s
ent hu
man visual p
e
rception.
Therefore, i
n
this p
ape
r
we u
s
e
L*a
*
b*
color spa
c
e. Additionall
y
, each
co
m
pone
nt in
L*a*b*
is qua
ntized
to
n
bi
ns, whe
r
e
n
=
{54,63,7
2
,81,
90}.
Giv
e
n
a
colo
r ima
ge o
f
size
M x
N,
f
o
r
inst
an
ce,
if
w
e
set
n=
7
2
, i
t
is equal
L
=
8, *a=3, *b
=3
.Den
ote by
C(
x,y
)
the qu
antize
d
imag
e,
w
h
er
e
0<
x<
M,0
<
y<N
.
3.4 Map
detec
t
ion
After edge o
r
ientation d
e
tection a
nd color
qu
antiza
t
ion, the nex
t process is micro
-
stru
cture d
e
tection
on
bot
h ed
ge
qua
ntization
an
d
color qua
ntizat
ion with
fri
c
tion
o
ne pixel and
filter 3x3. We compa
r
ed
the center value of th
e f
ilter to its ei
ght neig
hbo
rs. Thi
s
a
c
tivity
prod
uced a
set of edge m
aps a
nd
colo
r maps.
Utilizi
ng these map
s
, the differe
n
c
e value
∆
o
f
colo
r int
ensit
y and
edg
e
orientatio
n of
ea
ch
co
mp
onent
L*a*
b*
ca
n
b
e
cal
c
ulated and
then
store
d
as
col
o
r feature an
d edge featu
r
es. The final
step is p
r
e
s
e
n
ting these two features i
n
to a
combi
ned
hi
stogram. A
n
ill
ustratio
n
of
map
dete
c
tio
n
is p
r
e
s
ente
d
in
Figu
re
2.
The
value
s
of a
quanti
z
ed im
age
C
(
x,
y)
a
r
e denote
d
as
∈
0,1,
…
,
1
. Denote neig
hbori
ng pixel
locatio
n
s by
(x,
y
)
an
d
(x
’
,
y
’
)
and their colo
r index value
s
as
C(
x,
y)
=w
1
and
C(
x
’
,y
’
)
=
w
2
. Th
e values of a
n
edge o
r
ientat
ion image
,
are denoted by
∈
0
,
1
…
,
1
. The angles at
(x,y)
an
d
(x
’
,
y
’
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 59
7 – 604
600
are de
noted
by
,
1
and
′,
′
2
. For neigh
bori
ng p
i
xels, whe
r
e the distan
ce i
s
D
and re
sp
ectiv
e
quantizatio
n numbe
rs fo
r the colo
r an
d edge ori
ent
ations a
r
e
W
and
V
, we c
a
n
define the col
o
r differe
nce histog
ra
m a
s
equatio
n (6
) and eq
uation
(7).
Color Quantization
Color Co-
O
ccurr
ence
Color Map
Color Feat
ure
10
40
90
20
10
20
10
40
90
20
10
20
10
40
90
20
10
20
90
90
50
10
20
10
90
90
50
10
20
10
90
90
50
10
20
10
50
50
90
20
10
20
50
50
90
20
10
20
50
50
90
20
10
20
34
34
56
70
60
80
34
34
56
70
60
80
34
34
56
70
60
80
78
56
56
70
50
80
78
56
56
70
50
80
78
56
56
70
50
80
90
90
90
70
60
80
90
90
90
70
60
80
90
90
90
70
60
80
Edge Qua
n
tization
Edge Co-
O
ccurr
ence
Edge Map
Edge Featu
r
e
10
12
10
4
5
3
10
12
10
4
5
3
10
12
10
4
5
3
11
10
7
5
3
13
11
10
7
5
3
13
11
10
7
5
3
13
10
5
10
10
16
10
10
5
10
10
16
10
10
5
10
10
16
10
8
5
2
10
5
10
8
5
2
10
5
10
8
5
2
10
5
10
5
10
3
17
10
5
5
10
3
17
10
5
5
10
3
17
10
5
10
6
10
10
16
10
10
6
10
10
16
10
10
6
10
10
16
10
Figure 2. Map detection illustration
,
∑
∆
∆
∆
,
′,
′
;m
a
x
|
|
,
|
|
(6)
,
∑
∆
∆
∆
,
′,
′
;m
a
x
|
|
,
|
|
(7)
Whe
r
e
∆
,
∆
and
∆
are differen
c
e value bet
ween two
col
o
r pixels. In th
is pap
er
we
use
D
=1. If the ed
ge
orie
ntation is
W
and q
uanti
z
at
ion of
colo
r i
s
V,
then th
e
feature
of CDH
denote
d
as fo
llows:
0
,
1
…
1
,
0,
1
…
1
(8)
Whe
r
e
Hc
olor
is th
e
colo
r histog
ram
a
nd
Hor
i
i
s
th
e edg
e o
r
ien
t
ation histo
g
ram. For
instan
ce, if we set the
dim
ensi
on of
col
o
r q
uant
izatio
n = 7
2
an
d e
dge o
r
ientatio
n = 1
8
, thu
s
, the
total feature
s
for im
age
ret
r
ieval a
r
e
72
+ 1
8
= 9
0
di
mensi
onal
fe
ature
s
. Fu
rth
e
rmo
r
e, thi
s
t
o
tal
feature is d
e
n
o
ted as
H.
4.
Gra
y
Le
v
e
l Co-oc
c
urre
nc
e Matrix
The u
s
e
of G
r
ay Level
Co
-occurre
nce M
a
trix
is
perfo
rmed to
dete
c
t four fe
ature
s
, these
are e
n
e
r
gy, e
n
tropy, co
ntrast an
d corre
l
ation in fou
r
dire
ction
s
; 0°
, 45°, 90°
an
d 135
°; so th
ere
are 16 featu
r
es totally. The first step of
GLCM
is tra
n
sforming th
e RGB imag
e into gray scale
image. The seco
nd step i
s
creatin
g a co-o
ccurren
c
e
matrix. The third ste
p
is d
e
cidi
ng a sp
a
t
ial
relation
shi
p
b
e
twee
n
the re
feren
c
e pixel and
the nei
g
h
bor pixel.
Parameters
that are co
nsid
ere
d
are ed
ge o
r
ie
ntation
(
θ
)
an
d distan
ce
(d)
. Furthe
rmo
r
e, the next step
is creatin
g
a symmetrical
matrix by ad
ding
co
-o
ccu
rre
nce matrix
with it
s t
r
a
n
sp
osed m
a
trix. Then, it
is follo
wed
b
y
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Batik Im
age Retrieval Based on Color
Differenc
e
Hi
stogram
and Gray .... (Agus Eko Minarno)
601
norm
a
lizi
ng the symmet
r
i
c
al matrix b
y
comput
in
g
the prob
abi
lity of each
matrix eleme
n
t.
More
over, th
e final step
is co
mputin
g
GLCM
feat
ure
s
. Each f
eature i
s
co
mputed by o
n
e
distan
ce
pixel
in fou
r
di
re
ctions, th
ose a
r
e 0°,
4
5
°,
9
0
°
and 135
°,
to detect co
-o
ccurren
c
e. Whe
n
GLCM
ha
s a
matrix
with
a
siz
e
Lx
L
, in
which
L
i
s
the
numbe
r of
gray levels
of the o
r
iginal
im
age
and
whe
n
th
e probability
of pixel
i
is the neig
hbo
r of pixel
j
within
di
stan
ce
d
and
ed
ge
orientatio
n
θ
is
P
, the energy featu
r
e, the
entro
py feature, the co
ntra
st feature an
d
the
correl
ation fe
ature can be
cal
c
ulate
d
by equatio
ns (14
)
, (15
)
, (16
)
a
nd (17
)
.
∑
,
,
,
,
(9)
∑
,
,
,
.l
o
g
,
,
,
,
(10
)
∑
.
,
,
,
,
(11
)
∑
,
,
,
,
(12
)
Whe
r
e
∑
.
,
,
,
,
,
∑
.
,
,
,
,
,
∑
,
2.
,,,
,
,012.
,,,
.
Energy, also
called a
s
Angula
r
Seco
nd
Moment, i
s
t
he rep
r
e
s
ent
ation of im
ag
e hom
oge
nei
ty. When th
e
value
of en
ergy i
s
hi
gh,
the
relation
shi
p
s
betwe
en pixe
ls are highly
homog
eno
us.
Entropy is t
he opp
osite
of energy wh
ich
rep
r
e
s
ent the
ran
domn
e
ss value b
e
twe
en ima
g
e
s
. A highe
r valu
e
of entropy in
dicate
s that t
h
e
relation
s bet
wee
n
pixels
are hi
ghly ra
ndom. Contra
st
is a variati
on of an ima
ge’s g
r
ay lev
e
l. An
image with a smooth
texture
ha
s
a
lo
w contra
st
value
and an
im
ag
e
with a roug
h
texture ha
s
a
high co
ntra
st value. Correl
ation is a line
a
r rela
tio
n
shi
p
betwee
n
pi
xels. Let, each GLCM feat
ure
0°, 45°, 90° a
nd 135
° as
H
asm
(
θ
), Hent
(
θ
), Hcont
(
θ
), Hc
or
r(
θ
)
w
her
e
∈
0
,45
,
90,
135
. So, th
e
feature of GL
CM is de
note
d
as follo
ws:
0°
…
135°
,
0°
…
135°
,
0°
…
135°
,
0°
...
135°
(13
)
So, the final feature of a b
a
tik image i
s
a co
mbi
natio
n of CDH featur
e
s
and G
L
CM features
denote
d
as fo
llows :
,
(14
)
For exampl
e, if we
set th
e
quanti
z
ation
of col
o
r =
7
2
, it con
s
ist
s
of
R=8,
G=3, B
=
3; q
uanti
z
ati
on
edge o
r
ientati
on = 18 a
nd the total GLCM feat
ure
s
are 16, so the total feature = 106.
5. Performan
c
e
Measure
For e
a
ch template im
a
ge in the
dataset, an M-dim
e
n
s
i
onal featu
r
e
vector
T=[T1,T2,…
T
M
]
is extra
c
te
d an
d
stored
in the
data
b
a
se.
Let
Q=[
Q
1,Q2,…Q
M
]
be th
e fe
atu
r
e
vector of a qu
ery image a
n
d
the distan
ce betwe
en th
em is sim
p
ly cal
c
ulate
d
as
,
∑
|
|
|
|
|
|
(15
)
Whe
r
e
∑
and
∑
. The class label
s of the
template ima
ge that yield
the
smalle
st di
stance will be
assigne
d to
the quer
y i
m
age.In this experime
n
t, performan
ce
wa
s
measured u
s
i
ng pre
c
i
s
ion
and re
call
whi
c
h are define
d
as follo
ws:
P(N) =
I
N
/ N
(16
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 59
7 – 604
602
R(
N) =
I
N
/ M
(17
)
Whe
r
e
I
N
is th
e num
ber of
retrieve
d ima
ges,
N
i
s
th
e
numb
e
r
of relevant ima
g
e
s, a
nd
M
i
s
the
numbe
r of all relevant data
in the dataset.
6. Resul
t
s
and
Discus
s
ion
Feature extra
c
tion u
s
in
g CDH
re
sults i
n
90 f
eatures
whi
c
h
con
s
ist
of 72 color f
eature
s
and 1
8
e
dge
orientatio
n fe
ature
s
. Th
en,
feature
s
extraction usi
ng GLCM
re
sult
s
in 16
featu
r
es
whi
c
h
con
s
i
s
t of feature
s
inclu
d
ing
e
nergy,
e
n
tro
p
y, contrast
and
co
rrel
a
tion. Each G
L
CM
feature
is co
mputed
in fo
ur
dire
ction
s
,
these
a
r
e
0°
, 45°,
90°
an
d 13
5°. T
o
tall
y, there
are
106
feature
s
whi
c
h can b
e
re
pre
s
ente
d
on
a histog
ram
,
as sho
w
n i
n
Figure 3. Figure 3a is
an
example
of a
com
puted
i
m
age
with
Fi
gure
3b
a
s
it
s hi
stog
ram
repre
s
e
n
tation
. The h
o
ri
zo
n
t
a
l
axis re
pre
s
e
n
t
s 106 featu
r
es, of whi
c
h
the first
72 f
eature
s
a
r
e
color featu
r
e
s
, the se
con
d
18
feature
s
are edge o
r
ientat
ion f
eature
s
and the la
st 16 are G
L
CM
feature
s
. By utilizing these
feature
s
, the
simila
rity of image
s i
s
co
mputed
usi
n
g
the Canb
erra ba
sed
on
e
quation
(1
8);
after
whi
c
h, the
p
r
eci
s
ion
an
d t
he
re
call of
the imag
e
re
trieval p
r
o
c
e
s
s a
r
e
com
p
uted b
a
sed
on
equatio
ns
(1
9) a
nd
(20
)
.
For eval
uatio
n pu
rpo
s
e
s
,
one im
age i
s
ran
domly
se
lected f
r
om
e
a
ch
cla
ss, resulting in 50 imag
es a
s
a data
test. The
eval
uation is
run
to retrieve fo
ur, six and ei
ght
image
s ba
se
d on e
a
ch i
m
age in t
he
data test. CDH,
by d
e
fa
ult extract
s
1
08 featu
r
es
whi
c
h
con
s
i
s
t of 90 colo
r features and
18 ed
ge
orientatio
n feature
s
.
In this
study,
we va
ried
the
si
ze
of colo
r
quanti
z
ation
and fe
ature
q
uantization, a
nd the
n
combi
ned
th
e re
sult
with
the GL
CM f
eature
s
. T
h
e
aim of thi
s
activity wa
s to find the
b
e
st
quanti
z
ation
size su
ch
th
at
the num
b
e
r of
features
is decrea
s
e
d
wh
ile in
crea
sin
g
p
r
e
c
isio
n
at the
same time. T
he evaluation
is perfo
rmed
in four
sche
mes. First, we use
CDH with 108 features
whi
c
h
co
nsi
s
t
of 9
0
colo
r f
eature
s
(1
0x3
x
3)
a
nd 18 e
dge ori
entatio
n
featu
r
e
s
.
S
e
co
nd,
u
s
e
CDH
with 90 featu
r
es
whi
c
h co
nsi
s
t of 72 color f
eatu
r
e
s
(8x3x3) a
nd 18 edge o
r
ie
ntation features.
Third,
use CDH and
GL
CM with
124
feature
s
wh
i
c
h con
s
ist of
90
colo
r feat
ure
s
(10x3x3
)
, 1
8
edge
ori
entati
on featu
r
e
s
a
nd 1
6
G
L
CM
feature
s
.
Fin
a
lly,
use CDH and GL
CM wi
th
106
featu
r
es
whi
c
h
con
s
i
s
t of 72
colo
r f
eature
s
(8x3x
3
), 18
edg
e o
r
ientation
feat
ure
s
a
nd 1
6
GLCM
features.
The re
sult is
sho
w
n in Ta
ble 1. The averag
e preci
s
i
on value
s
for retrieving fo
ur, six and ei
ght
image
s in th
e four
sche
mes a
r
e
81.
92%, 81.
53
%, 85.44% and 8
5
.44%
respe
c
tively. The
evaluation
was
cond
ucte
d
in Matlab 20
13a an
d Wi
n
dows 7 o
p
e
r
ation sy
stem
with Co
re i5
2.3
Hz an
d 4
GB
memory. T
h
e
feature
extra
c
tion
of ea
ch
image i
n
e
a
ch sch
e
me to
o
k
1.0
91, 1.08
4,
1.097 an
d 1.
086 second
s
respe
c
tively. The re
sult
sh
ows that the redu
ction
of fe
ature
s
from 1
0
8
to 90 may d
e
crea
se p
r
e
c
i
s
ion by 0.3
9
%
. Though t
here i
s
a lo
wer preci
s
io
n, this differe
nce is
relatively sm
all; on the
op
posite
scal
e,
a sm
alle
r nu
mber of featu
r
es pot
e
n
tiall
y
decrea
s
e
s
the
compl
e
xity of
the pro
c
e
ss.
Furthe
r, new
feature
s
that are abl
e to incr
e
a
se pre
c
i
s
ion signifi
cant
ly can be add
ed. The
feature
s
u
s
e
d
are
extra
c
te
d usi
ng
GL
CM. GLCM
fea
t
ures ad
d 16
new fe
atures.
The eval
uati
o
n
of the effect
of the ad
ditio
n
to 90
CDH feat
ure
s
a
n
d
108
CDH fe
ature
s
sho
w
s an in
crea
sin
g
pre
c
isi
on
by 3.35% a
n
d
3.92% resp
ectively co
m
pare
d
to th
e
use of
CDH featu
r
e
s
o
n
ly.
Ho
wever,
the
average
pre
c
isi
on
of bot
h ad
diti
on
s
remain
the
sa
me, 85.4
4
%. The
r
efo
r
e,
we
con
c
lu
de that
use of co
rre
c
t feat
ure
s
contribute
s
hi
gher to p
r
e
c
i
s
ion tha
n
the
use of a large
numbe
r of f
eature
s
. F
r
o
m
all evalu
a
t
ion sche
me
s, the fou
r
th
schem
e is the mo
st effi
cient
scheme b
e
cause it use
s
smalle
r
nu
mber of feat
ure
s
with 0.0
05 extra time. An example of
retrieval for f
our ima
g
e
s
retrieval is sh
own in
Fig
u
re
4. Figure 4a
is an imag
e q
uery and Fig
u
re
4b-e i
s
retrie
val image
s b
a
se
d on the
Can
berra m
e
asu
r
em
ent. Figure
4b i
s
the mo
st simi
lar
image to th
e
image
que
ry, while
Figu
re
4e is the fou
r
th most
simil
a
r ima
ge to t
he ima
ge q
u
ery.
The pe
rform
a
nce
comp
ari
s
on of all sche
mes is
pre
s
e
n
ted in Figu
re
5.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Batik Im
age Retrieval Based on Color
Differenc
e
Hi
stogram
and Gray .... (Agus Eko Minarno)
603
(a)
(b)
Figure 3. An example of Batik image a
n
d
it
s histog
ra
m feature: (a
) Batik image
and (b
)
Histo
g
ra
m CDH of Batik i
m
age.
(a)
(b)
(c)
(d)
(e)
Figure 4. An example of image retrieva
l: (a) T
he q
u
e
r
y image, (b
-e) the simil
a
r
image
s retu
rn
ed
.
Table 1.Average preci
s
io
n
and re
call
co
mpari
s
o
n
feature sele
ction
Retrieval
Performa
nce
CDH
108
CDH
90
CDH
124
Pr
op
osed
Met
h
o
d
4
Pr
ec
ision
93,
0
0
93,
5
0
95,
5
0
96,
5
0
Recal
l
62,
0
0
62,
3
3
63,
6
7
64,
3
3
6
Pr
ec
ision
85,
0
0
83,
3
3
90,
3
3
89,
3
3
Recal
l
85,
0
0
83,
3
3
90,
3
3
89,
3
3
8
Pr
ec
ision
67,
7
5
67,
7
5
70,
5
0
70,
5
0
Recal
l
90,
3
3
90,
3
3
94,
0
0
94,
0
0
Figure 5. The
performan
ce
compa
r
i
s
on
of CDH and p
r
opo
se
d meth
od
0
50
100
150
200
250
300
1
1
12
13
14
15
16
17
18
19
1
1
0
1
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 59
7 – 604
604
7. Conclu
sion
This
study p
r
ese
n
ts ima
g
e
rep
r
e
s
entati
on
u
s
ing
a m
odified
Colo
r Differen
c
e Hi
stogram
and GL
CM. T
he modification is pe
rform
ed in ord
e
r to
find the optimal colo
r qu
a
n
tization
size, so
that the number of col
o
r feature
s
can
be red
u
c
ed. The re
du
ced
feature
s
, then, are repl
aced by
GLCM
features to im
prov
e image
ret
r
ieval perfo
rm
ance. Base
d
on evalu
a
tion result, th
e
prop
osed me
thod is abl
e to sho
w
hig
h
e
r
pre
c
i
s
ion
when ret
r
ievin
g
four imag
e
s
. The p
r
opo
sed
method
usi
n
g
106
CDH fe
ature
s
a
nd
1
8
GL
CM fe
ature
s
, ha
s
96.
50% p
r
e
c
isio
n; wh
ere
a
s the
use of 1
08 CDH featu
r
e
s
has 9
3
.00%
pre
c
isi
on.
Th
e avera
ge p
r
eci
s
ion of the
prop
osed m
e
thod
whe
n
retri
e
ving four, six a
nd eight imag
es is 3.
8
5
% h
i
gher tha
n
th
e origin
al CDH. Therefore, we
can
co
ncl
ude
that the prop
ose
d
meth
od
is abl
e to pe
rf
orm b
e
tter, while it ha
s a
smaller
numb
e
r
of feature
s
co
mpared to the origin
al CDH for bati
k
image
s retri
e
val
.
Referen
ces
[1]
F
anan
i A, Yun
i
arti A, and S
u
ciati
N. Geom
etric F
eature
Extractio
n
of
Batik Image U
s
ing C
a
rdi
n
a
l
Spli
ne C
u
rve
Repr
esentati
on.
TEL
K
OMN
I
KA Te
l
e
commu
n
i
ca
ti
o
n
,
Com
p
u
t
i
n
g
,
El
e
c
tro
n
i
cs and
Contro
l.
201
4; 12(2).
[2]
Minarn
o
Ag
us-
E
ko, Munark
o
Yuda, Bim
anto
r
o
F
i
tri, Kurnia
w
a
rd
han
i Arrie
and Suc
i
ati N
anik.
Batik
Imag
e Retri
e
v
a
l Base
d on
Micro-Structur
e Descri
p
tor
.
2
nd Asia-P
acifi
c
Confere
n
ce
on Com
puter
Aide
d S
y
stem
Engi
neer
in
g–A
PCASE. Ba
li, Indo
nesi
a
. 201
4; 2: 63-64.
[3]
Nugr
oho, S
apt
adi,
and
Darm
a
w
a
n
U. R
o
tati
on Inv
a
ria
n
t In
de
xi
ng F
o
r
Ima
ge Us
in
g Z
e
rn
i
k
e Mome
nts
and R-T
r
ee.
T
E
LKOMNIKA T
e
lec
o
mmun
icat
ion, Co
mputin
g
,
Electronics a
nd Co
ntrol.
20
11; 9(2).
[4]
Li, Guizhi. Improvin
g Re
l
e
vanc
e F
eed
back in Imag
e Retriev
a
l b
y
Incorp
orati
n
g Unl
abe
lle
d
Images.
T
E
LKOMNIKA Indon
esia
n Journ
a
l o
f
Electrical Eng
i
ne
erin
g.
201
3
;
11(7): 363
4-3
640.
[5]
Z
u
kuan, W
.
E. I., et al. An
Effici
ent Con
t
ent Based I
m
age R
e
triev
a
l Schem
e.
TELKOMNIKA
Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
ng.
2013; 1
1
(11):
698
6-69
91.
[6]
Irianto, Su
he
n
d
ro Y. Se
gme
n
t
ation T
e
chni
q
ue fo
r Ima
ge I
nde
xing
an
d R
e
trieva
l o
n
Dis
crete Cos
i
n
e
s
Domai
n
.
TELKOMNIKA Telecommunication, Com
puti
ng, El
ectronics a
nd
Contro
l.
201
2; 11(1): 1
1
9
-
126.
[7]
RM H
a
ral
i
ck, K
Sh
an
gmug
a
m
, Dinste
i
n.
T
e
xt
ural
featur
e
for ima
g
e
clas
sificatio
n
.
IEEE
T
r
ansacti
on
on Syste
m
s, Man an
d Cyb
e
r
m
atics.
1
973;
SMC-2(6): 61
0
-
621.
[8]
AK. Jain, A. Vaila
ya
. Image r
e
trieva
l usin
g
color a
nd sh
ap
e.
Journ
a
l of
P
a
ttern Rec
ogn
i
t
ion
. 19
96;
29(8): 12
33-
12
44.
[9]
BS manju
nath
,
JR. Ohm,
VV. Vasudev
a
n
,
A. Yamad
a
. Color a
n
d
texture
desc
r
iptors.
IEEE
Confer
ence
on
Circuit an
d Systems for vid
e
o
technol
ogy.
20
01; 11(6): 70
3
-
715.
[10]
B. Julesz, T
e
xtons. T
he ele
m
ents of te
xtu
r
e perc
ept
i
on and
t
heir inter
a
ctions.
Na
tu
re
.
1
9
8
1
;
290
(580
2): 91-9
7
.
[11]
Guang-
Hai L
i
u,
Jing Yu Yang.
Image Retrie
v
a
l bas
ed on th
e texto
n
co-oc
c
urrenc
e matrix.
Journ
a
l of
Pattern Reco
g
n
itio
n.
200
8;
41: 3521-
35
27.
[12]
Guang-
Hai
Li
u
,
Lei Z
h
a
ng, Y
i
ng-K
un H
ou,
Z
uo Yon
g
L
i
, Jing Y
u
Ya
ng.
Image R
e
triev
a
l b
a
sed
on
multi-te
xton h
i
s
t
ogram.
Journ
a
l
of Pattern Re
cogn
ition.
2
010
; 43: 2380-
238
9.
[13]
Guang-
Hai
Li
u, Z
uo-Yon
g
L
i
,
Lei Z
h
an
g, Yo
ng
Xu.
201
1. Image
Retr
iev
a
l
base
d
o
n
mic
r
o-structur
e
descri
por.
Jour
nal of Pattern
Reco
gniti
on
. 2
011; 44: 2
123-
213
3.
[14]
Guang-
Hai
Liu
,
Jing-Yu Y
a
n
g
. Conte
n
t-ba
sed
ima
ge r
e
trieval
usi
ng c
o
lor d
i
fferenc
e
histogr
am.
Journ
a
l of Pattern Rec
ogn
itio
n
. 2013; 4
6
: 18
8-19
8.
[15]
RC. Gonzal
es, RE. W
oods.
Di
gital i
m
age pr
o
c
essin
g
.
3rd ed
ition. Prentic
e Hall. 2
007: 4
2
4
:
447.
[16]
SD Z
enzo. A
n
o
te on t
he gr
a
d
ie
nt of multi
i
m
age.
C
o
mput
er Visi
on, Grap
hic, an
d Imag
e
Processi
ng.
199
6; 33(1): 11
6-12
5.
Evaluation Warning : The document was created with Spire.PDF for Python.