TELKOM
NIKA
, Vol.14, No
.4, Dece
mbe
r
2016, pp. 16
08~161
6
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v14i4.4289
1608
Re
cei
v
ed
Jul
y
18, 201
6; Revi
sed
No
ve
m
ber 7, 2016
; Accepte
d
Novem
b
e
r
22, 2016
Region Based Image Retrieval Using Ratio of
Proportional Overlapping Object
Agus Zainal Arifin, Rizka
Wakhidatus Sholikah
*, Dimas Fann
y
H. P., Dini Adni Nav
astar
a
Informatics De
partment, F
a
cu
lt
y
of Informati
on T
e
chnol
og
y, Institut
T
e
knologi Se
pul
uh N
opem
ber,
Jl. Ra
ya IT
S Kampus Suk
o
li
lo
, Suraba
ya, 6
0
111, Ind
ones
ia
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: ri
zka
.w
akhi
da
tu
s1
5@
mhs.if
.its.ac.id
A
b
st
r
a
ct
In Reg
i
on B
a
s
ed Imag
e Retri
e
val (RBIR), d
e
termin
a
tion
of the relev
ant b
l
ock in
qu
ery regi
on
i
s
base
d
on th
e p
e
rcenta
ge of i
m
a
ge o
b
j
e
cts that overl
ap w
i
th each s
ub-b
l
o
ck. But in some imag
es, the si
z
e
s
of releva
nt obj
ects are small.
It may cause t
he ob
ject
to be
ignor
ed i
n
det
ermini
ng the r
e
leva
nt sub-bl
oc
ks.
Therefore,
in t
h
is stu
d
y w
e
p
r
opos
ed
a syst
em of
RBIR b
a
sed
on
the
p
e
rcenta
g
e
of p
r
oporti
ona
l
obj
ects
that overla
p w
i
th sub-bl
ocks. Each sub-
bloc
k is sele
cted a
s
a query regi
o
n
. The co
lor an
d texture featur
es
of the
qu
ery re
gio
n
w
ill
b
e
ext
r
acted
by
usin
g
HSV
histo
g
ra
m
an
d L
o
ca
l Bi
nary P
a
ttern (
L
BP), respectiv
e
ly.
W
e
also use
d
shap
e
as glo
b
a
l
featur
e
by a
pplyi
ng in
v
a
ria
n
t mo
ment
as
descri
p
tor. Exp
e
ri
ment
al res
u
l
t
s
show
that the propos
ed
meth
o
d
has
aver
ag
e precisi
on w
i
th 74%.
Ke
y
w
ords
: Pe
rcentag
e overl
app
ing
obj
ect, RBIR, Invariant
mo
me
nt, HSV, LBP
Copy
right
©
2016 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
In the last few de
cad
e
s,
Conte
n
t Based
Image
Re
trieval (CBIR) has be
co
me
popula
r
resea
r
ch. CBIR is o
ne of th
e tech
nique
s
to retrieve im
age fro
m
larg
e image
data
bases
ba
sed
on
their visual
si
milarity [1]. Searchin
g ima
ges
u
s
in
g the
i
r conte
n
t has many advantage
s than u
s
i
ng
their an
notati
on text, beca
u
se
not every
image
s have
annotation
a
nd not
eve
r
y annotatio
ns
can
resembl
e
the
image
s well.
Therefo
r
e,
CBIR is
a
b
le
to overco
me
the wea
k
n
e
ss
of text-ba
s
ed
image retriev
a
l method.
Query
by
Ex
ample
(QBE
) is one of the query
techni
que
s in CBIR system that gives an
example ima
ge as q
uery. The features
of quer
y will
be extracte
d and compa
r
e
d
with feature
s
from im
age
d
a
taba
se
s. Th
e featu
r
e
ca
n
be
divided
a
s
gl
obal
feat
ure
and
lo
cal
feature. Glo
bal
feature
uses overall im
ag
e to extra
c
t their fe
at
ure
without
con
s
i
der th
e u
s
ers inte
re
st [2, 3].
While th
e lo
cal feature
onl
y use
s
pa
rt o
f
image that
use
r
s nee
d.
Some of the
resea
r
che
s
find
that using l
o
cal feature b
a
s
ed o
n
their
region i
s
mo
re
effective to satisfy what u
s
ers
requi
re [
1
],
[4-6]. L
o
cal f
eature
extra
c
tion from
qu
e
r
y imag
e b
a
sed o
n
regi
on
is kn
own a
s
Re
gion
Ba
sed
Image Ret
r
ie
val (RBIR).
In RBIR, som
e
of the regi
o
n
s a
r
e cre
a
te
d in
que
ry image. Not eve
r
y region i
s
rel
e
vant to
determi
ne th
e use
r
intere
st. Therefo
r
e
use
r
s h
a
ve
to define the
Regio
n
of Intere
st (ROI) in
image q
uery
so that the irrelev
ant regi
on ca
n be eli
m
inated. ROI cho
s
en
by the user i
s
m
o
re
c
a
pa
b
l
e to
e
x
p
r
es
s us
er
in
te
r
e
s
t, b
u
t
it
be
c
o
me
s les
s
effective if u
s
ers h
a
ve to
d
eal
with a
lot
of
query. Anoth
e
r a
pproa
ch i
s
cre
a
ting
ROI by sy
stem
[1, [4]. The method
divides im
age i
n
to
sub
-
blocks
in
cert
ain size (e.g. 3x3)
th
an det
ermin
e
ROI from eve
r
y sub
-
blo
c
ks whi
c
h
is
overl
ap
wit
h
the obje
c
t [1]
.
The obj
ect
is obtai
ned f
r
om s
egme
n
ting bet
wee
n
foreg
r
o
und
an
d ba
ckgroun
d.
Another
re
se
arch to auto
m
atically defi
ned ROI
is
usin
g Re
gion
Important Index (RII) a
n
d
Saliency
Re
g
i
on Ove
r
lap
p
i
ng Blo
c
k (S
ROB) [4]. Ho
wever if the
image
co
nte
n
t small
obj
e
c
t,
sometim
e
s th
e sub
-
blo
c
k cannot be d
e
tected a
nd be
co
me i
rrel
e
va
nt sub
-
blo
ck.
That ca
n lead
to
define
wro
n
g
ROI o
r
sele
cted
regio
n
. Therefore
it
is ne
ce
ssary t
o
have a
met
hod that
can
be
adapte
d
well i
n
the different
size of obj
ect
.
In this pape
r, we prop
osed system
RBIR
base
d
o
n
the percen
t
age of prop
ortiona
l
obje
c
ts that o
v
erlap with
sub-bl
ock to d
e
termin
e sele
cted re
gion.
To find the si
milarity, we u
s
e
colo
r
and
texture
as lo
cal
feature a
n
d
sh
ape
a
s
gl
obal fe
ature. Thi
s
m
e
tho
d
is exp
e
cte
d
to
improve ima
g
e
relevan
c
y compa
r
ed to e
x
isting metho
d
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Regi
on Base
d Im
age Retri
e
val
Usi
ng Ratio of Pr
oportional Overla
pping
…
(Agu
s Zainal Arifin
)
1609
2. Rese
arch
Metho
d
The
Wan
g
’s image
dat
aset
s a
r
e u
s
ed i
n
thi
s
pape
r
(can
be d
o
wnl
o
aded i
n
http://wang.ist.psu.edu/docs/related
/). This dataset usually used
to measure perform
a
nce of
image retriev
a
l
metho
d
.
Th
ere are 100
0 image
s,
which
co
nsi
s
ts of 10
catego
rie
s
.
Each categ
o
r
y
has 1
00 ima
g
e
s. Step by step of pro
p
o
s
ed metho
d
a
s
we ca
n see
in Figure 1
will be d
e
scri
bed
in the next su
bse
c
tion.
Figure 1. De
sign System
2.1. Preproc
essing
The first
step
, image que
ry is blurred b
y
Gaussi
an fi
lter to red
u
ce
noise. Th
en
conve
r
t
the re
sult im
age to g
r
ay
scale. In this rese
a
r
ch, segmentatio
n
is do
ne by
usin
g the ed
ge
descri
p
tor.
T
he u
s
e
of e
d
ge d
e
scri
ptor to si
mp
lify th
e
pr
oc
es
s of s
e
gme
n
t
a
t
ion
h
a
s
a
l
so
be
e
n
done in
previous
studi
es [
1
]. Sobel filter is u
s
e
d
to e
x
tract the p
r
o
per e
dge
of o
b
ject
s from t
he
gray
scale im
age. Thi
s
sta
ge pr
odu
ce
s
black a
nd
wh
ite image
with edg
es
of o
b
ject
s, the re
sult
are sho
w
n in
Figure 2.
Figure 2. Pre
p
ro
ce
ssi
ng: Sobel filter, (a)
Orig
in
al imag
e, (b) g
r
ay scale image a
n
d
(c) image
after applying
Sobel filter
However,
there
are
still any gap
s between edges i
n
t
h
e
sam
e
object. To
overcome thi
s
issue, dilation
is use
d
, so that the gap b
e
twee
n
edg
e
s
ca
n be redu
ced. To p
e
rfo
r
m dilation, li
ne-
sha
pe
of st
ru
cturin
g
eleme
n
t whi
c
h
is 5
pixels
wid
e
i
s
used.
Due
t
o
spa
c
e
between
edg
es th
at
have vary in
positio
n, we
use fou
r
kind
of line st
ru
ct
uring
eleme
n
t with different
angel
s, i.e. 0°,
45°, 90° a
nd
135°. After dil
a
tion pro
c
e
ss, filling are pe
rform to get fully segme
n
ted image.
Dilation p
r
o
c
ess by appl
ying 4 stru
ct
uring
eleme
n
t
has si
de e
ffect in obje
c
t si
ze.
Comp
are to
origin
al o
b
je
ct, the re
sult
h
a
s
bigg
er
obj
ect tha
n
the
origin
al ima
g
e
. Ero
s
ion
u
s
ing
circle
with
si
ze 3
as st
ru
cturing
ele
m
en
t are
pe
rform
e
d to
minimi
ze tho
s
e
effects a
s
sho
w
n
i
n
Figure 3.
2.2. Dete
rmine Proportio
nal Ov
erlapp
ing Sub-bloc
ks
To dete
r
mine
the regi
on a
s
the que
ry, image q
uery i
s
divided into fi
xed si
ze n x n
.
In this
pape
r we u
s
e 3 x 3 as
shown in Figu
re 4. From th
e previo
us
re
sea
r
ch [4, 6], the best
size to
divide image
s is 5 x 5 but has hig
her
co
mputation tha
n
3 x 3.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1608 – 161
6
1610
Figure 3. Ero
s
ion u
s
in
g circle, (a
) Ori
g
in
al image, (b
) Image after dil
a
tion and fillin
g and (c)
Image after e
r
osi
o
n
Figure 4. Det
e
rmin
ation re
gion a
s
que
ry, (a) Divi
de
d image into fixed si
ze 3 x 3, (b) Give the
identity numb
e
r in ea
ch su
b-blo
c
ks
Comm
only, there i
s
relev
ant and irrele
vant
sub-blo
c
k in an imag
e. The irrel
e
vant sub
-
block
will be eliminated,
because
they do not have i
n
formation to
descri
be what user needs and
can
cau
s
e
error in im
age
retrieval. To
determi
ne wh
ether the
sub
-
blo
ck i
s
rele
vant or not, o
u
r
prop
osed met
hod are used.
Black
and
white image th
at alrea
d
y se
gmented
bet
wee
n
obje
c
t
and ba
ckg
r
o
und, than
label a
r
e giv
en in eve
r
y obje
c
t. Howe
ver, not all image
s can b
e
se
gmente
d
well. The
r
e
are
some im
age
s that still contain noise. Fo
r exampl
e if the ba
ckgro
u
n
d
image is n
o
t
homogen
ou
s,
sometim
e
s in
segme
n
tatio
n
pro
c
e
ss a p
a
rt of bac
kgro
und ca
n be d
e
tected a
s
ob
ject. So filtering
relevant o
b
je
ct of the ima
ge is n
e
cessary in th
is
ste
p
. The ide
a
i
s
ba
se
d on a
s
sumption th
at if
the dete
c
ted
obje
c
t ha
s
smaller
area t
han ave
r
a
ge
area
of obj
e
ct
in an i
m
age
and its lo
cati
on is
far from th
e center of im
ag
e, t
hen it ca
n
be cl
assified
as n
o
ise or
i
r
relevant obj
e
ct. By calculati
n
g
area
of o
b
je
ct and
minimu
m dista
n
ce b
e
twee
n eve
r
y
pixel
in obje
ct
and cente
r
of
image, we
ca
n
determi
ne
th
e
rel
e
vancy o
f
object.
Th
e distan
ce bet
wee
n
pixel
s
,
in obje
c
t
and
cente
r
of
image
,
are do
ne by usin
g Euclid
ean di
sta
n
ce a
s
sho
w
n
in Equation 1
.
a
r
g
m
i
n
,
(
1
)
Whe
r
e
,
can b
e
cal
c
ulate
d
from Equatio
n
2.
,
(2)
Area in
every object
s
is
cal
c
ulate
d
b
a
se
d on th
ei
r label, fo
r e
x
ample if th
e image
c
ontains
two objec
ts
the res
u
lt w
ill
s
h
ow
tw
o ar
ea
siz
e
. Then w
e
c
o
mbine the r
e
sult betw
een
area a
nd pix
e
l distan
ce o
f
every objects to the cen
t
er of image.
From imag
e
, the object
=
(
1,2,
3,...,
k
) having a
r
ea
more th
an
α
and
di
stan
ce
less than
β
,
can be determined a
s
relevant obj
e
c
t. Whe
r
e
α
can be calcula
t
ed by Equation 3 and
β
ha
s a value 0.4.
∑
5
⁄
(
3
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Regi
on Base
d Im
age Retri
e
val
Usi
ng Ratio of Proportional Overla
pping
… (Agu
s Zainal Arifin
)
1611
After obtainin
g
the relevan
t
object, the
nex
t step i
s
comp
ute the
percenta
ge o
f
object
area in every
sub-bl
ocks. From area
,
that describe a
s
area obj
ect
overlap with
sub
-
blo
c
k
, the ratio of propo
rtion
a
l area overla
ppin
g
obje
c
t
,
can
be cal
c
ul
ated
by Equation 4.
,
,
(4)
This m
e
thod
use
d
thre
sh
ol
d
,
0.1. So tha
t
if sub-block contai
ns
are
a
overla
ppin
g
obje
c
t mo
re
than 0.1
it can b
e
in
dicated a
s
re
leva
nt su
b-bl
ock.
By usi
ng th
e ratio of
are
a
overlapping object, the re
l
e
vant obj
ect t
hat has
small
si
ze
can still
be detected.
Every
rel
e
vant
sub
-
blo
c
ks
will be i
n
cl
u
ded into li
st
of que
ry sub-bl
ock that
we
calle
d
Saliency
Re
gion
Overlap
p
ing
Block
(SROB
)
. For exampl
e in Figure
4,
the SROB of image que
ry is sub
-
blo
c
k 2,
5, 6, 8 a
nd 9.
Then
the
sel
e
cted
que
ry
sub-bl
ock i
s
u
s
ed
in the
ne
xt step to extract lo
cal fe
ature
of query imag
e.
2.3. Featur
e Extrac
tion
Local featu
r
e
and
glo
bal
feature
are
combin
e
d
to
get bette
r
re
trieval results. Local
feature
of im
age q
u
e
r
y an
d imag
e data
base a
r
e
extracted
ba
sed
on
the sel
e
ct
ed sub
-
blo
cks.
This
re
se
arch
, use
s
colo
r a
nd texture
a
s
local fe
ature, while glob
al
f
eat
ure
u
s
e
s
shape of
obj
ect
as de
scripto
r
.
2.3.1. Color
C
o
lor
s
ar
e
c
o
mmo
n
l
y us
ed a
s
fe
a
t
ur
e de
sc
r
i
p
t
or
, be
ca
u
s
e
na
tu
ra
lly h
u
m
an
visua
liz
a
t
ion
can ea
sily distinguish imag
e by its color.
In th
is paper, we use HS
V histrog
r
am
to extract col
o
r
feature. HSV
is ch
osen be
cause this mo
del ca
n
be su
perio
r color
space, comp
are to RGB [7].
HSV colo
r m
odel divide
s color into 3 co
mpone
nt, H (Hue
), S (satu
r
ation
)
, and V (Value).
The Hu
e co
mpone
nt rep
r
ese
n
ts the type of co
lor e.
g. red, yellow, green, etc. Hue a
r
e
descri
bed
by spe
c
ific
po
si
tion in color
whe
e
l,
red
st
arting at
0 d
egre
e
s, g
r
e
e
n
at 120
deg
ree
s
and blue
at 240 deg
ree
s
.
Compl
e
me
ntary colo
urs
are in
-b
etwe
en: yell
ow i
s
at
60
de
grees,
cyan is at 18
0 degree
s, and mage
nta is at 300
de
grees. Saturation com
pon
en
t repre
s
e
n
ts
ho
w
white th
e
col
o
r i
s
. Satu
rati
on h
a
s value
betwe
en
0 a
n
d
1. T
he
col
o
r
sho
w
s p
u
re
col
o
r when
the
saturation value is 1, and will be dil
u
ted by
whit
e when saturation is decreased. Value
comp
one
nt al
so
de
scribe
a
s
b
r
ightn
e
ss.
This
co
mpon
ent mea
s
u
r
e
how bla
c
k th
e color is.
Wh
en
value com
p
o
nent are d
e
crease, the bla
c
kne
ss of the
color a
r
e in
creased.
Every com
p
o
nent of HSV
model
can
be
com
puted from RGB (Re
d
, Gre
en, Blu
e
) mo
del
by cal
c
ul
ating
maximum
RGB
M
, minim
u
m RGB
m,
and
delta
d
b
e
twee
n
M
an
d
m
as
sh
own i
n
Equation 5
-
1
0
[7].
From thi
s
pu
rpo
s
e, color i
s
extra
c
ted fr
om every sel
e
cted
sub
-
bl
ocks in im
ag
e que
ry
and imag
e d
a
taba
se. The
feature is re
pre
s
ente
d
by histogram from each com
pone
nt (H, S, V).
Than the valu
e is normali
zed between 0
and 1.
m
a
x
,
,
(
5
)
m
i
n
,
,
(
6
)
(
7
)
(
8
)
(
9
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1608 – 161
6
1612
.
,
1
,
2
,
3
,
4
,
5
,
(10)
2.3.2. Textur
e
Texture
ha
s
an imp
o
rtant
role to
de
scri
be t
he
su
rfa
c
e of an
obje
c
t and its rel
a
tionship
with the surro
undin
g
area [
4
]. Local Bin
a
r
y Pattern
(L
BP) [8] is kn
o
w
n a
s
go
od t
e
xture de
scri
ptor
with high
er p
e
rform
a
n
c
e result. Some o
f
the applic
ati
ons a
r
e u
s
ed
LBP as de
scriptor [4], [6],
[9].
LBP de
scribe
s a
pixel valu
e ba
se
d o
n
it
s n
e
igh
bor g
r
ay level pixel.
Given
a
cent
ral pixel
c
that h
a
s
gra
y
value
g
c
an
d gray value f
r
om its nei
gh
bours
g
p
, the
LBP can
be
calcul
ated u
s
i
ng
Equation 11.
,
∑
.
2
,
1,
0
0,
0
(
1
1
)
P
is the total
numbe
r of ne
ighbo
urs and
R
is ra
diu
s
of the neighbo
u
r
hoo
d. After the LBP
of each pixel i
n
sel
e
cte
d
su
b-blo
c
k a
r
e calcul
ated,
the
n
the hi
stogra
m
is cre
a
ted t
o
rep
r
e
s
e
n
t the
texture in every sele
cted sub-bl
ock.
2.3.3. Shape
Moment inva
riants are im
portant
sh
ap
e de
scri
pto
r
s in comput
er vision. T
here are t
w
o
types of
sh
ape
s d
e
scri
p
t
ors:
conto
u
r-ba
sed
shap
e de
scripto
r
s a
nd
regi
o
n
-ba
s
e
d
sha
p
e
descri
p
tors. Reg
u
larly, in
the most po
p
u
lar ty
pe inva
riant mom
ent
used i
s
cont
our-ba
s
e
d
sh
ape
desc
r
iptor [10].
Hu invari
ant moment
s is b
a
si
c to measure si
mil
a
rity betwe
en the template or d
a
taba
se
image.
Hu’
s
Seven Mom
e
nts Invari
ants are
invari
ant
unde
r tran
sl
ation, chang
e
s
in
scale,
a
n
d
a
l
s
o
r
o
ta
tion
.
So
it d
e
sc
r
i
bes
th
e imag
e
de
s
p
ite o
f
it
s
lo
cat
i
on,
si
ze,
a
nd
rot
a
tion. In this
res
e
arc
h
we u
s
e i
n
vari
ant mome
nt to extra
c
t the
sha
pe featu
r
e. The
seven
momen i
n
va
riant a
r
e u
s
e
d
based on
normalize
d
central moment
s [11]. They ar
e
useful fo
r im
age scali
ng, tran
slation
an
d
rotation. To calcul
ate them
, formula in Equation 1
2
ca
n be used.
∅
∅
4
∅
3
(
1
2
)
∅
∅
3
3
3
3
3
∅
4
∅
3
3
3
3
Tran
slatio
n invarian
ce
ca
n be achieve
d
by
shifting the polynomi
a
l basi
s
to the obje
c
t
centroid a
s
sown in Equ
a
tion 13.
,
∞
∞
∞
∞
(
1
3
)
In Equation 1
3
, image d
e
scrib
ed a
s
I
(
x,
y)
a
s
a pi
ece
w
ise co
ntinuo
us b
ound
ed f
unctio
n
,
variable
p
and
q
i
s
po
sitive intege
r [10]
. Whe
r
e
,
are the
coo
r
din
a
tes of th
e
obje
c
t centroi
d
. Scaling invariant
s is obt
ained by
normalizatio
n of each momen
t. For the sca
le
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Regi
on Base
d Im
age Retri
e
val
Usi
ng Ratio of Proportional Overla
pping
… (Agu
s Zainal Arifin
)
1613
invariant
s the formula is re
pre
s
ent
e
d
as Equation 14. The moment
ca
n be com
puted usi
n
g
the ce
ntroid
of the image
I(x
,
y)
that e
quivalent to the
whose center h
a
s
be
en shifted to
centroid of the image.
,
1
(
1
4
)
,
0,1
,
2,3
,
…
2.4. Similarity
Measurem
e
nt
Similarity bet
wee
n
imag
e
query
and im
age d
a
taba
se is
com
pute
d
by u
s
ing E
u
clid
ean
distan
ce fo
r every feature
.
For local fe
ature,
the di
stance i
s
cal
c
ulate for ea
ch sele
cted
sub-
blocks of
qu
ery a
nd i
m
a
ge d
a
taba
se.
Assume
tha
t
image
qu
ery has
n
sele
cted
su
b-blo
c
ks.
Selected
sub
-
blo
c
ks
que
ry are
re
pre
s
e
n
t
ed by
={
,
, .
. . . ,
} and
,
,
....,
rep
r
e
s
ente
d
as selecte
d
sub-bl
ocks in image dat
a
b
a
s
e. The di
sta
n
ce only com
pute in the sa
me
sub
-
blo
c
ks in
dex. For exa
m
ple, to find distan
ce
in
sub
-
blo
c
k 1
,
the euclide
an dista
n
ce i
s
perfo
rmed be
tween
,
and
,
.
So that
the distan
ce betwe
en sub
-
blo
c
k query and su
b-blo
c
k
image data
b
a
s
e
d
(
,
,
,
) can b
e
cal
c
ulate u
s
ing Equation
15.
,
,
,
(
1
5
)
The final dist
ance of local feature is descri
bed as average di
stance
for every sub-blocks
as sho
w
n in
Equation 16.
,
∑
,
,
,
(16)
For
glob
al fe
ature, th
e
si
milarity is
cal
c
ul
ate
d
by
a
pplying Eu
cli
dean
di
stan
ce bet
wee
n
image q
uery
and ima
ge
databa
se di
rectly. The to
tal distan
ce f
r
om ima
ge q
uery an
d im
age
databa
se i
s
a
combi
nation
of three featu
r
e di
stan
ce
s. For eve
r
y fea
t
ure, we
assi
gn weight tha
t
are multiplie
d
by distance
value.
Given weig
ht of feature color
, weight of feature texture
,
and weight of
feature shap
e
, the total distan
ce can b
e
cal
c
ulate by
Equation 17.
,
.
.
.
(
1
7
)
In this rese
arch the optimal distan
ce
that we used is 0.1, 0.4, 0.5 for
,
,
respe
c
tively. After obtaining all the dist
ance to image data
base, then perfo
rmed th
e
squ
e
n
c
ing by
ascen
d
ing o
r
de
r.
The sm
allest
the
dist
ance, indi
cat
e
that two im
age
s ha
s hi
g
her
s
i
milarity.
3. Results a
nd Discu
ssi
on
3.1. Results
1. Performa
n
c
e Mea
s
u
r
e
Usi
ng Pre
c
i
s
i
on-Re
call
The
pro
p
o
s
e
d
meth
od i
s
evaluated
u
s
i
ng p
r
e
c
isi
on.
Pre
c
isi
on i
s
co
mmonly
u
s
ed
to
evaluate the
performan
ce
of IR syste
m
. Based
o
n
Table 1, preci
s
ion
can
be cal
c
ul
ated
by
Equation 18.
(18
)
W
h
er
e
TP
(True
Po
sitive)
sho
w
s the
num
ber of
rele
vant
ima
ges
that ca
n
be retrieved
by
sy
st
em,
TN
(True Negati
v
e) is the n
u
mbe
r
of rel
e
vant image
s that ca
nno
t be retrieve
d by
sy
st
em,
FP
(False
Po
sitive) i
s
the
nu
mber of irre
l
e
vant imag
es that sy
stem
retri
e
ve a
n
d
FN
(Fal
se Negati
v
e) is the nu
mber of irrele
vant
images t
hat doe
sn’t re
trieve by the system.
For e
a
ch qu
e
r
y, the re
sult
s are
sh
own b
a
se
d on th
eir
ran
k
ing. T
he
ran
k
ing
are compute
from the dist
ance betwe
e
n
query and
image in databa
se. Tab
l
e 2 sho
w
s
pre
c
isi
on wit
h
a
different
k
, where
k
de
scri
be as n
u
mbe
r
of retrieval image
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1608 – 161
6
1614
Table 1. Co
nfusio
n Matrix
Rele
v
a
n
t
Not
rele
v
a
n
t
Retrie
v
e
TP FP
Unretrie
v
e
TN
FN
Table 2. Pre
c
ision in Va
riet
y of k
Label quer
y
Precision
k=5 k=10
k=15
k=20
Africa 0,60
0,80
0,67
0,70
Beaches 0,40
0,40
0,40
0,45
Buildings 0,80
0,50
0,47
0,45
Bus
1,00 1,00
1,00
1,00
Dinosaur
1,00 1,00
1,00
1,00
Elephant
1,00
0,90 0,80
0,70
Flowers
1,00
0,90 0,87
0,75
Horse
1,00 1,00
1,00
1,00
Mountains 0,60
0,70
0,60
0,50
Food
1,00
0,90 0,80
0,80
Average
0,84
0,81
0,76
0,74
(a)
(b)
(c
)
Figure 5. Example que
ry a
nd re
sults fo
r retrie
v
a
l
k=
20
,
(a) Que
r
y
a
nd re
sult
s of
cat
e
g
o
ry
“Din
osau
r”, (b
) Que
r
y and result
s of cate
gory
“Horse
s”, (c) Qu
ery an
d results of category
“Mou
ntain”
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Regi
on Base
d Im
age Retri
e
val
Usi
ng Ratio of Proportional Overla
pping
… (Agu
s Zainal Arifin
)
1615
From
Tabl
e 2
,
it can
be
se
en the
avera
ge p
r
e
c
isi
on f
o
r
k
with a va
lue of 5,
10,
15, an
d
20 a
r
e
0.84,
0.81, 0.76,
a
nd 0.7
4
. Whe
n
k
in
crea
se
s the valu
es o
f
pre
c
i
s
ion
wi
ll de
cre
a
se.
For
all variatio
ns
of do
cume
nt
return, Bus,
Dino
sa
ur
and
Ho
rse h
a
s the hi
ghe
st p
r
eci
s
ion
with
1
.
00
(100%
). Example of the query an
d re
sult ca
n be
see
n
in Figu
re 5. Experi
m
ents
were also
con
d
u
c
ted
by
usin
g stand
ard 11 re
call.
The re
su
lts
can
be
se
en
in Fig
u
re
6
that sh
ows t
he
pre
c
isi
o
n
-re
call of the prop
ose
d
method.
2. Weight Parameter
The mea
s
u
r
e
m
ent of the simila
rity betwee
n
the qu
ery image
wi
th the image
s in the
databa
se i
s
use
d
three
ki
nd of
featu
r
e
s
, where ea
ch feature ha
s a weight. Th
e wei
ght of t
h
e
feature
s
a
r
e
used to i
n
d
i
cate th
e pe
rcenta
ge
of
the effe
ct of
these
feature
s
to
determi
ne
simila
rity. The weig
hts for the colo
r, te
xture, and sh
ape featu
r
e
s
are
symboli
z
ed by
w
c
,
w
t
, a
nd
w
s
, re
sp
ectiv
e
ly. We
p
e
rformed
expe
riments u
s
ing
differe
nt we
ights, to
obt
ain the
opti
m
al
para
m
eter val
ues. T
able
3
sho
w
s the
averag
e valu
e o
f
pre
c
isi
on fo
r
k
=20 with different weig
hts.
Based o
n
the
Table 3, the value of the highe
st pre
c
isi
on is obtain
e
d
when the
weights a
r
e
w
c
,
w
t
and
w
s
of
0.1
0
, 0.40
and
0
.
50, re
spe
c
tively, and al
so
whe
n
the val
ues are 0.3
5
, 0.35
and
0.3
0
.
Between th
ose two
set-wei
ght value
s
, in
this re
se
arch
we
cho
o
se the first o
ne b
e
ca
use in
k=5,
10 and 1
5
the
first value-se
t is out perform than the se
con
d
one.
Figure 6. Performa
nce of the prop
osed
method
Table 3. Weight Paramete
r
w
c
w
t
w
s
Precision
0.40 0.50 0.10 0.73
0.10 0.40 0.50
0.74
0.50 0.10 0.40 0.61
0.10 0.50 0.40 0.73
0.35 0.35 0.30
0.74
0.30 0.35 0.35 0.72
0.35 0.30 0.35 0.72
3.2. Discussi
on
From the ex
perim
ental re
sults, the
r
e a
r
e se
ve
ral p
o
ints that co
uld be di
scu
s
sed. In
Table 2, it ca
n be se
en th
at for categ
o
ry “Dino
s
au
r” i
n
Figure 5(a) has a p
r
e
c
isi
on value of 1
.
0
(100%
). Thi
s
i
s
b
e
cau
s
e the
“Di
n
o
s
au
r” ha
s
a
ba
ckgro
und
image
that
is
relativel
y
homog
ene
ou
s, con
s
istin
g
only of
on
e colo
r. It affects t
he
segm
entation
re
sult
s in
the
determi
nation
of the
sele
ct
ed regio
n
for
local
f
eatures extractio
n
. T
he
segm
entat
ion results
ca
n
also
affect th
e determinati
on of the
sh
ape
whi
c
h
is use
d
a
s
a g
l
obal featu
r
e.
Categ
o
ry which
also
pro
d
u
c
e
s
preci
s
io
n
value 1.0 i
s
“Ho
r
se” i
n
Figure 5(b).
Although th
e Ho
rse ha
s a
backg
rou
nd
con
s
i
s
ting
of a va
riety of
col
o
rs, b
u
t
have a
different colo
r
co
mbination
of
othe
r
colo
rs in the
dataset, as well as texture.
Another
cat
e
gory
is ”B
u
s
”,
in t
h
is cat
e
g
o
ry
t
he B
u
s h
a
s
simila
r sh
ape
in every image on this
cat
egory that
ca
n be identify well u
s
ing the
feature shap
e.
This m
e
thod
can
also be
use
d
in a
het
erog
ene
ou
s
backg
rou
nd i
m
age fo
r exa
m
ple in
Figure 5 (b).
Ho
wever, be
cau
s
e the ba
ckgro
und i
s
h
e
terog
ene
ou
s, the selected
salien
c
y regi
on
is the ove
r
all
picture of a
horse ima
ge.
It is
beca
u
se whe
n
seg
m
enting u
s
in
g edge
de
scriptor,
gra
ss i
s
also detecte
d as o
b
ject.
The othe
r
ca
tegory
such
as
“Bea
ch”
a
nd “M
ountai
n
”
, both h
a
ve
simila
rities i
n
col
o
r
feature. In a
dataset that we u
s
e, b
o
th
categ
o
rie
s
a
r
e domin
ated
by white a
nd
blue
colo
rs. T
h
i
s
can
cau
s
e
errors
wh
en
re
trieve the
im
age
s, a
s
can
be
seen
in
Figure 5
(
c).
Beside
that t
h
e
scana
ry ima
g
e
ha
s
no fixe
d shap
e, so t
hat the
glob
al
feature
that
has hig
e
r
wei
ght cann
ot gi
ve
muc
h
effec
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1608 – 161
6
1616
This meth
od
has a
wea
k
ne
ss in the
image
s that has n
o
cle
a
r bo
unda
ry
betwee
n
foreg
r
ou
nd a
nd ba
ckgro
u
n
d
, as in imag
es of scena
ry
. Where
a
s in
the image
s with clear o
b
je
ct,
the method
can produ
ce g
ood p
r
e
c
isi
o
n
value. Th
is
wea
k
n
e
ss is
based on th
e
sele
ction of t
he
sele
cted regi
on
are
taken from
the resu
lts
of
image
segmentatio
n. For o
b
je
cts t
hat se
gment
ed
properly, the select
ed regi
on obtained
will be better
i
n
representing the object.
Image
con
s
i
s
ting of multipl
e
obje
c
ts
at the same time
can
also affe
ct the d
e
term
ination
of the
sele
cte
d
region. S
o
me of th
e im
age
s
that
ha
ve not h
o
mo
gene
ou
s b
a
ckgroun
d a
r
e
often
detecte
d a
s
obje
c
ts. T
herefore, th
e d
e
termin
ati
on
of dete
c
ted
obj
ect
releva
ncy
is impo
rtant
to
do.
In image
retri
e
val, sele
cte
d
image
que
ry can a
ffe
ct the value of p
r
eci
s
io
n. In rese
arch
con
d
u
c
ted by
Vemina a
nd
Ja
cob [1] m
e
ntioned th
e result
s of testi
ng ag
ain
s
t se
veral meth
od
s.
Test
s were
p
e
rform
ed u
s
i
ng the
same
dataset with
t
he data
s
et u
s
ed in thi
s
stu
d
y [1]. For so
me
categ
o
ry like
“Africa
”
, “Bu
s
”, “Elepha
nt”,
“D
ino
s
a
u
r”, and “F
ood
” o
u
r propo
se
d method p
r
od
uce
s
better p
r
e
c
isi
on value. But
for the
other
categ
o
ry
the
pre
c
isi
on val
ue of o
u
r
pro
posed m
e
tho
d
i
s
lowe
r than the Vemina a
nd Ja
cob M
e
thod [1]. Nonet
hel
ess q
uery that is use
d
in previ
ous
method [1] i
s
un
kno
w
n, th
us
allo
wing f
o
r the
differe
nce
bet
wee
n
the imag
e q
u
e
ry cond
ucte
d in
this study wit
h
the experim
ents that
hav
e been
con
d
u
c
ted by Vemi
na and
Ja
cob
.
In this study, the determi
n
a
tion of the weight
pa
rame
ter for ea
ch f
eature
ha
s a
n
impa
ct
on the value
of precisi
o
n
of the search re
sult
s. In Table 3 it can be se
en that the weig
ht
variation
s
re
sult the diffe
rent p
r
e
c
isio
n.
For this
prop
osed me
thod, the sh
ape an
d textu
r
e
feature
s
can
be better
de
scripto
r
tha
n
the color
fe
ature in
som
e
categ
o
ry. Thi
s
is
be
cau
s
e
in
some
cate
go
ry the obje
c
t has
comm
on
in colo
r,
so
that in this case
colo
r ca
nnot be u
s
e
d
as
good de
scri
ptor. We al
so find that texture can give
a l
a
rge
r
effect than the sh
ap
e, beca
u
se when
the
w
t
is set
with low valu
e the pre
c
isi
o
n can b
e
dro
p
about 0.13.
4. Conclusio
n
The p
r
o
p
o
s
e
d
metho
d
b
a
s
ed
on
prop
ortional
overl
appin
g
regio
n
to
cho
o
se
relevant
sele
cted
regi
on h
a
s ave
r
age
preci
s
io
n 74
% in
k
=
2
0 a
n
d
ha
s ma
ximu
m pr
e
c
is
ion
10
0%.
Ho
wever, thi
s
metho
d
ha
s a wea
k
ne
ss in t
he ima
ges that h
a
s no cle
a
r b
o
unda
ry between
foreg
r
ou
nd a
nd ba
ckgroun
d, as in im
ag
es of
sce
ne
ry
. For furthe
r rese
arch, it wil
l
be optimi
z
e
d
in
determi
ning t
he weigh
parameter. And will al
so optimize edge
descri
p
tor segmentation by
prun
ning u
n
n
e
ce
ssary ed
g
e
from un
rele
vant object.
Referen
ces
[1]
R Vemi
na, K
P
Jaco
b. A
Sub-b
l
ock B
a
s
ed Ima
ge
Ret
r
ieval
Usi
n
g
Modifi
ed Inte
g
r
ated
Reg
i
o
n
Matching.
IJCSI: International Journa
l of Co
mputer Scie
nce.
201
3; 10(1): 68
6-69
2.
[2]
XY W
a
n
g
, YJ Yu, HY Yang.
An Effective Im
age R
e
triev
a
l
Scheme Usi
n
g Col
o
r, T
e
xture and S
h
a
p
e
F
eatures.
Co
mputer Stan
dard
s
& Interfaces.
201
1; 33(1): 59
-68.
[3]
XY W
a
n
g
, Z
F
Chen, JJ Yun. An effective
method for
color ima
ge r
e
trieva
l base
d
on textur
e.
Co
mp
uter Stan
dards & Interfa
c
es.
2012; 3
4
(
1
): 31-35.
[4]
S Nisha
nt, V
T
y
a
g
i. Co
ntent
base
d
imag
e retrieva
l
base
d
on rel
a
tive loc
a
tions of multi
p
le re
gio
n
s of
interest usi
ng s
e
lectiv
e regi
on
s matching.
Information Sciences.
2014: 2
1
2
-
224.
[5]
M Subr
ahma
n
y
a
n
, RP M
a
h
e
s
h
w
ari,
R Ba
la
subram
ani
an.
Loca
l
ma
ximu
m ed
ge
bin
a
r
y
patterns: A
ne
w
d
e
scriptor
for image retri
e
val an
d ob
ject trackin
g
.
Signa
l
Processin
g
.
2012; 92(
6): 146
7-14
79.
[6]
Guizhi L
i
. Improvin
g Rel
e
va
n
c
e F
eedb
ack i
n
Image R
e
tri
e
val b
y
Inc
o
rp
oratin
g Unl
a
b
e
lled Ima
ges.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2013; 1
1
(7): 3
634-
364
0
[7]
V Chern
o
v, J
Alan
der, V Bochko. Integer-b
as
ed acc
u
rate
convers
i
on b
e
tw
e
e
n
RGB an
d HSV colo
r
spaces.
Co
mp
uter and El
ectri
c
al Eng
i
ne
eri
n
g.
2015; 4
6
: 32
8-33
7.
[8]
T
Ojala, M Pietikai
ne
n, T
Maen
paa. Mul
t
ires
ol
ution gr
a
y
sca
le a
n
d
rotation inv
a
riant text
ur
e
classification w
i
th
loc
a
l bi
nar
y
patterns.
IEE
E
T
r
ans. o
n
P
a
ttern An
alysis
an
d Mac
h
in
e
Intelli
genc
e.
200
2; 24(7): 97
1-98
7.
[9]
Z
Guo, L Z
h
a
ng, D Z
h
a
ng.
A compl
e
ted
mode
lin
g of
local
bi
nar
y p
a
ttern op
erato
r
for textur
e
classification.
IEEE Transactions on
I
m
age Processing.
20
10; 19(6): 1
657
-166
3.
[10]
M Rizon, H Yazid, P Saad,
AYM Shakaff, AR Saad, MR Mamat,
S Yaacob, H
Desa,
M
Karthigay
an.
Object detecti
on us
ing
ge
o
m
etric invar
i
a
n
t
moment.
America
n
Jour
na
l
of Appl
ie
d Sc
ienc
es.
20
08;
2(6): 187
6-1
8
7
8
.
[11]
MK Hu. Visua
l
Pattern Reco
gniti
on b
y
M
o
ment Invari
ant
s.
IRE T
r
ansations o
n
Infor
m
ation th
eory.
196
2: 179-
187.
Evaluation Warning : The document was created with Spire.PDF for Python.