TELKOM
NIKA
, Vol.11, No
.1, March 2
0
1
3
, pp. 119~1
2
6
ISSN: 1693-6
930
accredited by D
G
HE (DIKTI
), Decree No: 51/Dikti/Kep/2010
119
Re
cei
v
ed Se
ptem
ber 18, 2012; Revi
se
d Jan
uary 21,
2013; Accept
ed Feb
r
ua
ry
3, 2013
Segmentation for Image Indexing and Retrieval on
Discrete Cosines Domain
Suhendro Y
Irianto
Dep
a
rtment of Informatics, F
a
cult
y
of Compu
t
er Science
Darmaj
a
ya Info
rmatics and Bu
siness Institute
,
Jl. Z
.
A. Pagar Alam 93A, Ba
n
dar Lam
pu
ng, Lamp
ung, Ind
o
nesi
a
e-mail: s
y
ira
n
to
196
2@gm
ail.c
o
m, pasca@
d
a
rmaja
ya.
a
c.id,
A
b
st
r
a
k
Dala
m
artikel
i
n
i di
gu
nak
an t
e
knik re
gi
on gr
ow
ing u
n
tuk pr
oses se
g
m
ent
asi p
ada c
i
tra
DC. Den
g
a
n
me
ng
gun
aka
n
teknik re
gio
n
g
r
ow
ing p
a
d
a
ci
tra DC aka
n
meng
uran
gi
ju
ml
ah re
gio
n
ya
ng
akan
dig
u
n
a
k
a
n
dal
a
m
pros
es
pen
gin
deks
an
citra. T
e
knik in
i akan
men
gur
ang
i w
a
ktu pe
mros
esa
n
da
la
m
me
mbuat ku
nc
i
ind
e
ks. T
e
knik recursive re
gio
n
grow
ing b
u
kan
meru
pak
an metod
e
ata
u
teknik yan
g
baru, aka
n
tetap
i
apliks
i
pa
da cit
r
a DC unt
uk membu
a
t kunci i
ndeks cuk
up
b
a
ru da
n
masi
h
jaran
g
di
gu
na
kan ol
eh p
e
n
e
l
i
ti-
pen
eliti l
a
in. H
a
sil pe
ne
litia
n me
nu
njuk
an b
ahw
a citr
a DC
yang tela
h dis
e
g
m
etas
i men
unj
ukan ef
ektiv
i
tas
lebi
h bes
ar dib
and
ingk
an d
e
n
gan citra DCT
yang tid
a
k dil
a
kukan pr
oses
seg
m
e
n
tasi di
ma
na pr
ecisi
o
nya
mas
i
n
g
-masi
ng
sebes
ar 0.75
dan
0.59. Di s
a
mpi
ng it
u, kar
ena
dal
a
m
pe
neliti
an
meng
g
unak
an citra
D
C
untuk
pe
ncari
a
n citra
(CBIR)
mak
a
hany
a
1/64 s
a
ja
d
a
ri
8
x
8 ya
ng
di
gun
akan
unt
uk i
n
d
e
xin
g
. H
a
l y
a
n
g
pali
ng
pe
ntin
g
dal
a
m
pe
ne
liti
an i
n
i
ada
la
h
hany
a d
i
gu
nak
an satu
keo
e
fi
sien
dari
64 k
oefisi
en D
C
T
yan
g
ada ya
itu koefi
s
ien DC, se
hin
gga b
any
ak meng
he
mat stor
age.
Ka
ta
k
unc
i:
DC coefficie
n
ts, regi
on grow
in
g
,
segmentatio
n
,
DC domai
n
A
b
st
r
a
ct
T
h
is pap
er us
ed reg
i
on
gro
w
ing seg
m
e
n
tation tech
ni
qu
e to seg
m
ent
the Discrete C
o
sin
e
s (DC)
imag
e. T
he cla
ssic prob
le
m o
f
content Base
d i
m
ag
e re
triev
a
l (CBIR) is th
e lack of acc
u
r
a
cy in
match
i
n
g
betw
een i
m
a
g
e
query a
nd i
m
a
ge i
n
the d
a
tabas
e. By
using re
gio
n
gr
ow
ing techn
i
q
ue on DC i
m
a
ge,it
reduc
ed th
e n
u
mber
of i
m
a
g
e
reg
i
o
n
s i
nde
xed. T
he
pro
p
o
sed
of rec
u
rs
ive re
gi
on
gro
w
ing is
not n
e
w
techni
qu
e b
u
t i
t
s appl
icati
on
o
n
DC
i
m
a
ges t
o
bu
il
d i
ndex
i
ng keys
is
quit
e
new
a
nd
not
yet pres
ente
d
by
ma
ny
auth
o
rs. T
he ex
peri
m
e
n
tal res
u
lts sh
ow
that
the
pro
pose
d
meth
od
s on s
e
g
m
e
n
te
d i
m
a
ges
pres
en
t
goo
d
pr
ecisi
o
n
w
h
ich are hig
her
th
an 0.60
on all
class
e
s. So, it c
oul
d b
e
co
nclu
de
d th
at reg
i
on
grow
i
n
g
seg
m
e
n
ted
ba
sed CBIR
more efficie
n
t co
mp
are
d
to
DC
imag
es in ter
m
of their pr
ecisi
on 0.5
9
a
nd 0.
75
,
respectiv
e
ly. M
o
reov
er, DC
b
a
sed
CBIR c
a
n sav
e
ti
me
an
d si
mp
lify a
l
g
o
r
ithm co
mp
are
d
to D
C
T
i
m
a
g
es.
T
he most sig
n
i
f
icant findi
ng from this w
o
rk is
instead
of usi
ng 64 D
C
T
coefficients this researc
h
only u
s
e
d
1/64 coeffici
ent
s w
h
ich is DC coefficie
n
t.
Ke
y
w
ords
:
D
C
coefficie
n
ts, regi
on grow
in
g
,
segmentatio
n
,
DC domai
n
1. Introduction
In the field of digital imagi
n
g
, image
seg
m
entation pl
a
y
s a vital role
as a p
r
elimi
n
ary step
for hi
gh l
e
vel
image
processing.
To
und
ersta
n
d
an
i
m
age,
one
n
eed
s to i
s
ol
ate the
obje
c
ts
in it
and find rel
a
tion amon
g them. The process of im
age
partition refe
rre
d as ima
g
e
segm
entati
o
n
[1]. In other word
s, se
gme
n
t
ation is u
s
ed
to pull
out th
e sig
n
ificant o
b
ject
s from th
e image.
Den
g
[2] propo
se
d
a JSEG alg
o
rithm to se
gment the i
m
age b
a
sed
on multi scale ‘Jim
age
s’
. The
image
s whi
c
h corre
s
po
nd
to the measurem
ents of
local ho
mog
e
neities at different scal
es
are
called as ‘J-i
mages’. The syst
em
has the ability to segment
color textured
images without
sup
e
rvisi
on.
First th
e colo
ur in
sid
e
the
image i
s
qu
antize
d
to
se
veral
cla
s
ses. The pixel
s
are
then repla
c
e
d
by their
co
rresp
ondi
ng
co
lour
cla
s
s lab
e
l whi
c
h
form
s the
cla
s
s m
ap of the
ima
ge.
A region g
r
o
w
ing meth
od
is then u
s
ed t
o
segm
ent the image ba
se
d on multiscal
e
‘J-im
age
s’.
Histo
g
ra
m thresholdi
ng is one of the
commo
n techni
que
s for monochrom
e image
segm
entation
[2], [3]. This te
chni
que
con
s
id
ers
th
at an im
age
co
nsi
s
t of
different
regi
on
s
corre
s
p
ondin
g
to the
grey level range
s. The
hi
stog
ram of a
n
im
age
can
be
sep
a
rate
d u
s
ing
pea
ks (m
ode
s) corre
s
pon
ding to
the
di
fferent regio
n
s
. A th
reshol
d value
corre
s
po
ndin
g
to t
he
valley betwe
en two adja
c
ent pea
ks can be used
to
sepa
rate these obje
c
t. But one of
the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 11, No. 1, March 2
013 : 119 – 1
2
6
120
wea
k
n
e
sse
s
of this meth
o
d
is th
at, it ignore
s
th
e sp
atial rel
a
tionship info
rmatio
n of the pixel
s
.
Adolfo
[4]
p
r
opo
sed
a ne
ural network based ada
ptive thre
shol
di
ng
segm
enta
t
ion alg
o
rithm
for
mono
ch
rome image.
The main a
d
v
antage of this metho
d
is that, it does n
o
t requi
re a p
r
iori
kno
w
le
d
ge abo
ut
numbe
r of o
b
j
e
cts i
n
the im
age. To
hum
ans, a
n
ima
g
e
is
not ju
st a
ran
dom
colle
ction of
pixels; it
is a
mea
n
ingf
ul arra
ngem
e
n
t of re
gion
s
and
obje
c
ts
.
There al
so
e
x
its a vari
ety of image
s
su
ch
as n
a
tural
scene
s, and
pai
ntings.
De
spit
e the la
rge va
riation
s
of the
s
e ima
g
e
s
, h
u
man
s
have
no
probl
em to in
terpret them.
Image
seg
m
entation i
s
th
e first
step i
n
image
analy
s
is
and
patte
rn
recognitio
n
. It is a criti
c
al a
nd esse
ntial comp
one
nt of image an
alysis
system, is one of the most
difficult tasks in image proce
s
sing, an
d determin
e
s the quality
of the final result of anal
ysis.
Image
segm
entation i
s
th
e process of
dividing a
n
i
m
age i
n
to dif
f
erent
regio
n
s
su
ch that
e
a
ch
regio
n
is ho
m
ogen
eou
s.
Many conte
n
t
-based ima
g
e
retrieval (CBI
R) syste
m
s have been
develop
ed si
nce the
early ninetie
s. A recent arti
cle pu
blished
by
Smeulders [5], reviewe
d
more tha
n
200 refe
re
nces
in this ever
changi
ng field.
Read
ers are
referred
to t
hat articl
e an
d som
e
additi
onal refere
nces
[4] for more i
n
formatio
n. Most of the CBIR pr
oje
c
ts aimed at g
eneral-purpo
se imag
e ind
e
xing
and retri
e
val system
s
fo
cu
s
o
n
sea
r
chin
g
imag
es
visually simil
a
r to the
que
ry i
m
age
or a
qu
ery
sketch. They do not hav
e the capability of assigning co
mprehensive text
ual descripti
o
n
automatically to images, be
cau
s
e of the
great difficu
lt
y in recog
n
izi
ng a larg
e nu
mber of obj
ects.
Re
sea
r
che
s
done
by [5]-[7] in the la
st few
years u
s
ed
all coefficient
s DCT f
o
r ima
g
e
indexing an
d
retrieval. This metho
d
cau
s
e
d
low
spe
ed retri
e
ving pro
c
e
s
s due to more
coeffici
ents p
r
ocesse
d co
mpare to
the purp
o
sed me
thod whi
c
h i
s
only 1/64 (a
DC coefficie
n
ts)
of DCT
co
efficient
s. T
herefor
e
the
purposed
metho
d
pe
rforms 6
4
time
s fa
ste
r
in
ind
e
xing
and
retrievin
g
co
mpared to DCT ba
se
d im
age retrieval.
Beside
s tha
t
the prop
osed metho
d
a
l
so
provide
s
a
n
u
mbe
r
of
pot
ential a
d
vant
age
s a
nd fe
a
t
ures,
which
can
be
b
r
iefly su
mma
rized
as:
(i) lo
w
com
p
l
e
xity and low com
puting
cost; (ii) high
pro
c
e
ssi
ng
speed,
(iii) e
a
s
y to implem
ent
insid
e
the JPEG comp
re
ssed dom
ain
,
and t
hus
providin
g the
additional
advantag
e that
comp
re
ssed i
m
age
s ca
n b
e
dire
ctly retri
e
ved witho
u
t full decom
pre
ssi
on.
Some re
sea
r
ch
ers h
a
ve attempted to
use m
a
chin
e-lea
r
ni
ng te
chni
que
s for image
indexing
and
retrieval [2], [5]. The sy
ste
m
intern
ally g
enerated m
u
ch
s
egm
entat
ion or groupi
ngs
of ea
ch i
m
a
ge’s re
gion
s ba
sed
on
different fe
ature
s
com
b
in
ation, then
l
earn
ed
whi
c
h
combi
nation
s
best rep
r
e
s
ented the se
mantic cate
g
o
rie
s
given a
s
example
s
by the user.
The
system
req
u
ires the
su
pe
rvised tr
aining
of variou
s p
a
rts
of the
im
age. Mo
st of
CBIR u
s
e
s
all
regions in the image to
match between image
qu
ery and image in the database [7],[8].
Therefore the
i
r wo
rks are inefficient an
d
time
con
s
um
ing due to th
e use m
o
re
DCT
coeffi
cie
n
ts
use
d
. In order to wo
rk o
u
t of these pro
b
l
em
s, this pa
per u
s
ed
Reg
i
on Gro
w
in
g Segmentatio
n to
sea
r
ch a
nd
retrieve ima
g
e
s
fro
m
data
b
a
se.
The
re
mainde
r of th
e pap
er i
s
organi
zed
as fo
llows.
Section 2 d
e
s
cribe
s
the
rese
arch met
hod
s used in
this wo
rk. S
e
ction 3
discusse
s
the re
sults
and di
scussio
n
. Section 4 descri
b
e
s
the
con
c
lu
sion
s
and some re
mark for the future works.
2. Res
earc
h
Method
s
2.1. Datab
a
s
e
and JPEG
In this
work
more th
an
5,000 im
age
s u
s
ed
as g
r
o
und truth (im
age d
a
taba
se
) which a
r
e
colle
cted fro
m
Internet an
d other
sou
r
ces. This
wo
rk al
so
us
ed
J
PEG
images
,
as
it
has
man
y
advantag
es compa
r
ed
to o
t
her fo
rmat. I
n
JPEG
im
a
g
es, to
comput
e the
DCT i
m
age
s of
a li
st
of
length n
=
8
a
nd the 2
D
DCT of a
n
8 x
8 array. Ra
ther than
taki
ng the tran
sformatio
n
of the
image a
s
a whole, the DCT is applie
d separately to
8 x 8 blocks of the image, this called a
s
a
DCT
blo
c
k. In cal
c
ul
ating
a DCT bl
ock,
actually
the
work
doe
s n
o
t actually h
a
v
e to divide t
h
e
image into bl
ocks. Sin
c
e t
he 2D
DCT i
s
sepa
rabl
e,
it can be
parti
tioned ea
ch
row into li
sts
o
f
length 8,
appl
y the DCT to
them, rej
o
in t
he result
ing li
sts, a
nd th
en
tran
spo
s
e
th
e whole i
m
ag
e
and re
peat th
e pro
c
e
ss.
DCT
-b
ased image comp
ression relies on two tech
nique
s to re
duce the data requi
red.
First i
s
q
u
a
n
tization, a
n
d
the
se
con
d
is
entro
py codi
ng
of the qua
ntize
d
co
efficient
s.
Quanti
z
ation
is the
proce
s
s of
red
u
cin
g
the nu
mb
e
r
of po
ssibl
e v
a
lue
s
of a
q
uantity, there
b
y
redu
cin
g
the
numbe
r of
bits nee
de
d to rep
r
e
s
e
n
t it. Entropy coding i
s
a techni
que
for
rep
r
e
s
entin
g
the
quanti
z
ed d
a
ta a
s
comp
actly
a
s
p
o
ssibl
e
.
A functio
n
t
hen
develo
p
ed to
quanti
z
e ima
ges
and to
calcul
ate the l
e
vel of co
m
p
ressio
n provided by different deg
ree
s
of
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Segm
entation for Im
age Inde
xing a
nd
Retrie
val o
n
Discrete Co
si
nes
Dom
a
in (Suhend
ro YI)
121
quanti
z
ation.
JPEG
u
s
e
s
a combin
atio
n of
spatial
-
d
o
main and
freque
ncy-dom
ain codi
ng.
T
he
image i
s
divid
ed into 8 x 8
blocks, ea
ch
of whi
c
h i
s
transfo
rme
d
int
o
the fre
quen
cy domai
n u
s
i
ng
the discrete
co
sine tra
n
sf
orm (DCT). Each blo
c
k of the image
is thus re
prese
n
ted by 64
freque
ncy
co
mpone
nts. T
he si
gnal te
nds to
con
c
entrate in th
e lower
spat
ial frequ
en
ci
es,
enabli
ng hig
h
-
freq
uen
cy compon
ents,
many of wh
i
c
h are u
s
u
a
lly zero, to be discarded
without
sub
s
tantially affecting
the
appe
ara
n
ce of
the
imag
e.
The m
a
in
so
urce of lo
ss
of informatio
n in
JPEG is a
qu
antizatio
n of the DCT
coeff
i
cient
s.
A table of quanti
z
at
ion co
efficien
ts is u
s
ed, o
n
e
per
coeffici
e
n
t, usually re
lated to hum
an pe
rc
eptio
n of different
freque
nci
e
s.
The qu
antized
coeffici
ents
a
r
e ord
e
red
i
n
a zig
-
zag se
quen
ce, st
a
r
ting at
the
up
per left (th
e
DC comp
one
nt),
sin
c
e mo
st of
the ene
rgy lies in the fi
rst
few co
e
fficie
n
ts. The final
step is
entro
py codin
g
of the
coeffici
ents, u
s
ing eith
er Huffman codi
n
g
or arithm
etic co
ding
.
2.2 DC Imag
es Tech
niqu
e
This
wo
rk u
s
ed o
n
ly DC
coeffici
ents i
n
stea
d
of usi
ng
all DCT coefficient
s
to
co
nstruct
image i
ndexi
ng a
nd
retri
e
ving.
Discrete Cosi
ne
s
co
efficient i
s
o
n
e
of 6
4
DCT
coeffici
ents a
s
a
n
image
ca
n be
con
s
tru
c
ted
as
8 x 8
arra
ys or bl
o
c
k. T
o
build
ind
e
xing
keys, a
s
ill
ustratio
n give
n
the
N-blo
cks of
an
ima
g
e
to
con
s
truct indexing
ke
ys by u
s
in
g
only DC
co
efficients of ev
ery
image in the
databa
se u
s
i
ng the followi
ng equ
ation:
64
1
i
hi
H
(
1
)
whe
r
e
N
i
i
i
N
DC
h
1
,
i
s
t
h
e
i
th
DC co
efficient of every block an
d
N
is
th
e
nu
mb
er
o
f
b
l
oc
k
in
the
image. In building ind
e
xing key
s
, this appro
a
ch only consi
d
e
r
s one DC
coe
fficient in every
block, which
mean
s it only
need
s1/64
o
f
the ti
mes n
eede
d by the
full extractio
n
and in
dexi
n
g
pro
c
e
ss.
Du
e to the limitations
of hardwa
r
e
spe
e
d
,
the work o
n
ly use
d
a d
a
taba
se of 5,
000
image
s.
2.3
Region Gro
w
in
g Segmentation Technique
In this
wo
rk,
Regi
on
gro
w
i
ng Imag
e
se
gmentation
u
s
ed
a
s
first
key pro
c
e
s
s in
nume
r
o
u
s
appli
c
ation
s
of compute
r
vision. It partitions the im
age into different meani
ng
ful region
s with
homog
ene
ou
s cha
r
a
c
teri
stics usi
ng di
scontinuitie
s
or simila
rities
of
imag
e compon
ents,
t
he
sub
s
e
que
nt processe
s d
e
p
end on it
s pe
rforma
nce.
In most
ca
se
s, the segm
en
tation of colo
ur
image de
mon
s
trate
s
to be more u
s
eful t
han t
he seg
m
entation of mono
ch
rome
image, beca
u
se
colo
ur im
age
expre
s
ses m
u
ch
more ima
ge featu
r
e
s
th
an mo
no
chro
me imag
e. In
fact, ea
ch pix
e
l
is
cha
r
a
c
teri
zed
by a g
r
eat num
ber
of com
b
i
nati
ons
of R,
G
,
B chromati
c
comp
one
nts.
Ho
wever, m
o
re
compli
cate
d se
gme
n
tation techniq
u
e
s
a
r
e
requi
re
d to deal
wit
h
ri
ch
chrom
a
tic
informatio
n in
the se
gme
n
tation of colo
ur ima
g
e
s
. A variety of se
gmentation t
e
ch
niqu
es
ha
ve
been p
r
o
p
o
s
ed in the literature. However, mo
st
tech
nique
s a
r
e ki
nd of dimen
s
ional exten
s
i
on
dire
ctly inheri
t
ed from the
segm
entation
of m
onochro
m
e image [9]
.
The spatial
comp
actn
ess
and colou
r
ho
mogen
eity are two de
sirab
l
e pro
per
tie
s
i
n
unsupe
rvised se
gme
n
tation, whi
c
h lea
d
to image-dom
ain and featu
r
e-spa
c
e ba
se
d segm
entati
on tech
niqu
e
s
.
The segme
n
tation of imag
es ha
s al
way
s
bee
n a key probl
em in compute
r
visio
n
. Up to
the early nin
e
ties bottom
-
up tech
niqu
e
s
like
edg
e detectio
n
an
d split-and
-m
erge
algo
rith
ms
were the
pri
m
ary fo
cu
s
of re
sea
r
ch. Ho
weve
r, b
y
that
time
peopl
e
realized
that
p
e
rf
ect
segm
entation
wo
uld n
o
t b
e
po
ssible
wi
thout in
co
rpo
r
ation
of hig
h
e
r level
kno
w
ledge. T
h
u
s
t
h
e
f
o
cu
s shif
t
e
d
t
o
war
d
s
mo
d
e
l
ba
se
d
t
e
ch
nique
s
li
ke
snakes
an
d
method
s
ba
sed o
n
g
eom
e
t
ric
model
s [10].
Regi
on g
r
o
w
i
ng algo
rithm
starts f
r
om
an initial, incom
p
lete
segmentatio
n
and try to
aggregate th
e yet unlab
ell
ed pixel
s
to
one of the
gi
ven regi
on
s. The initial
re
gion
s are u
s
ually
calle
d seed
region
s o
r
see
d
s. Th
e de
ci
sion whethe
r
a
pixel sh
ould
join a
regio
n
or n
o
t is b
a
se
d
on some fitn
ess fun
c
tion
whi
c
h reflect
s
the si
milari
ty between t
he re
gion a
n
d the ca
ndid
a
te
pixel. As
pro
posed i
n
[11]
, the o
r
de
r i
n
whi
c
h
the
pi
xel is
processed
is dete
r
mined
by a
g
l
obal
prio
rity queu
e whi
c
h
sort
s all can
d
idate
pixels by
their fitness val
ues. Thi
s
ap
proa
ch
elega
ntly
mixes local (fitness) an
d gl
obal (pixel o
r
der) info
rmati
on. There is
an abun
dan
ce of literature
on
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 11, No. 1, March 2
013 : 119 – 1
2
6
122
image
seg
m
entation, an
d
a numb
e
r
of review
arti
cl
es hi
ghlightin
g them. Met
hod
s al
so h
a
v
e
been d
e
fine
d for post
pro
c
e
ssi
ng the low-level
segme
n
tation to furthe
r reg
u
lari
ze
the
segm
entation
output, such as Ma
rkov Random Fi
eld
s
[12].
Automatic im
age
seg
m
ent
ation is on
e
of the p
r
imary probl
ems o
f
early
comp
uter visi
on,
has
bee
n int
ensively
stud
ied in th
e p
a
st [11]. Th
e
existing
aut
omatic im
ag
e se
gme
n
tation
techni
que
s
can
b
e
cl
a
ssifie
d
into
four
app
ro
ach
e
s,
na
m
e
ly:
thresh
ol
ding
techniq
ues,
boun
dary
-
ba
sed m
e
thod
s,
re
gion
- b
a
sed m
e
thod
s,
and
hyb
r
id
tech
niqu
es
.
Re
gion
-ba
s
ed
techni
que
s
re
ly on the a
ssumption th
at adja
c
ent pixe
ls in th
e sam
e
re
gion
hav
e simil
a
r visu
al
feature
s
su
ch
as g
r
ey level
,
co
lo
urs val
u
e, or texture.
A well
-k
no
wn
tech
niqu
e of
this
app
ro
ach
is split a
nd
merg
es [13].
Obviou
sly, the perfo
rma
n
ce of this app
roach largely depe
nd
s on t
h
e
sele
cted
hom
ogen
eity crite
r
ion. In
stead
of tuni
ng
ho
mogen
eity pa
ramete
rs, th
e
see
ded
re
gi
on
gro
w
ing
(SRG) te
chniq
ue
is controll
ed
by a numb
e
r
of initial se
ed
s [14]. Given
the se
ed
s, SRG
tries to find
an a
c
curate
segm
entation
of imag
es
i
n
to re
gion
s
with the
pro
perty that
ea
ch
con
n
e
c
ted
compon
ent of
a re
gion
m
eets exa
c
tly one of th
e
see
d
s. M
o
reover, hi
gh-l
e
vel
kno
w
le
dge o
f
the image comp
one
nts
can b
e
expl
oi
ted throu
g
h
the choi
ce
of seed
s. Th
is
prop
erty is v
e
ry attra
c
tive for sem
antic obj
ect extra
c
tion towa
rd content-b
ased
image data
b
a
se
appli
c
ation
s
.
Ho
wever, S
R
G
suffe
rs f
r
om
anoth
e
r pro
b
lem:
ho
w to
sel
e
ct t
he initial
se
e
d
s
automatically for providin
g more a
c
cu
rate s
egme
n
tation of ima
ges. Th
e alg
o
rithm of re
g
i
on
gro
w
ing
seg
m
entation techniqu
e ca
n b
e
descri
b
e
s
a
s
follows:
Inpu
t
:
imag
e I
cre
a
te
an (empty) set S of se
gments
stage 0: i:=
0
;
for
all DC coefficients P in I
create
a n
e
w
segment R
p
of level 0
(
consisting on
ly of P)
pu
t
R
p
in S
repeat
stage i:
for
a
l
l
se
gm
en
ts R
i
of lev
e
l i in S
repeat
fi
nd
a segme
n
t
Ŕ
j
of level j
≤
i in S,
R
i
and
Ŕ
j
are ne
igh
b
o
u
re
d and
R
i
Ú
Ŕ
j
is ho
mo
gen
eo
us eno
ug
h
remove
R
i
an
d
Ŕ
j
from
S
re
de
fi
ne
R
i
:= R
i
Ú
Ŕ
j
of level i+1
unti
l
no
such
Ŕ
j
ca
n be
found
ad
d
R
i
to S
i:=i+1
u
n
til
stage i-1 has create
d
no new
seg
m
e
n
t
In Markov Random Fiel
d
s
[12] algori
t
hm for image seg
m
enta
t
ion has be
en dra
w
n
considerable
attention due to its
ability
to integrate t
e
xture,
colo
ur, and edge i
n
fo
rmation in an
optimal m
ann
er to
devi
s
e
a
ro
bu
st lab
e
li
ng of th
e im
a
ge into
ho
mo
gene
ou
s regi
ons [15]. Th
e
s
e
method
s
still
depe
nd o
n
th
e a
s
sumption
that the
pixels b
e
lon
g
ing t
o
the
obje
c
t o
f
intere
st sha
r
e
a co
mmon
set of low-lev
e
l imag
e attri
butes, th
er
e
b
y
allowin
g
th
e obje
c
t to
b
e
extra
c
ted
a
s
a
singl
e entity.
If an object is compo
s
e
d
of multiple
regi
ons of differi
ng texture or colo
ur then the
obje
c
t is divid
ed into re
gio
n
s corre
s
po
n
d
ing to ea
ch
of these, an
d
these
sub
re
gion
s mu
st then
be re
-a
ssem
bled thro
ugh
some
conte
x
tual-ba
s
ed
post p
r
ocessi
ng to segm
e
n
t the compl
e
te
obje
c
t from the image. By employing
an additi
on
al con
s
trai
nt upon the segmentatio
n that
encourage
s it
to find
a h
u
m
an, it would
be
po
ssi
ble
t
o
only
extract
the regio
n
s correspon
ding
to
the huma
n
in the image
. This additi
onal con
s
trai
nt can b
e
p
r
ovided th
ro
ugh info
rmati
on
rega
rdi
ng the
desired shap
e of the final retained regio
n
.
The p
r
op
ose
d
method
of
this research
is
ap
plying
region
gro
w
in
g se
gmentati
on on
DC
images. The
region growi
n
g segmentati
on is
not new method to segm
ent an i
m
age [2],[9],
but
applying it to
DC ima
ge i
s
a quite n
e
w
and
have
not done
bef
ore by othe
r authors. This
prop
osed met
hod ha
s thre
e significant merits n
a
mel
y
reduce storage u
s
age, m
o
re a
c
curate
and
effective in matchin
g
of image re
gion.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Segm
entation for Im
age Inde
xing a
nd
Retrie
val o
n
Discrete Co
si
nes
Dom
a
in (Suhend
ro YI)
123
Figure 1. The query p
r
o
c
e
ss of region
g
r
owi
ng ba
se
d
image retri
e
val system
The q
u
e
r
y proce
s
s h
a
s be
en e
s
tabli
s
h
e
d
a
s
follo
ws:
Firstly, use
r
queri
e
s an
RGB imag
e
in the
system
, RGB im
age
then i
s
conv
erted to
g
r
ay
scale im
age.
Seco
ndly, b
y
utilizing
re
g
i
on
gro
w
ing
alg
o
rithm this ima
ge
will b
e
seg
m
ented
into
meanin
g
ful
re
gion
s. Finally
ba
sed
on
the
s
e
regio
n
s, the
minimum di
stance betwe
e
n
them
will be cal
c
ulated
and compa
r
e
d
to the image
regions in the database.
Figu
re.1 illustrates
query process
diagram
of
the
image ret
r
iev
a
l
system p
r
op
o
s
ed. On
ce
a query i
s
sp
ecified,
it score
s ea
ch
segm
ented ima
ge
based on h
o
w
clo
s
ely and
satisfie
s to t
he imag
es i
n
the datab
ase. Th
e score
i
for
ea
ch atomi
c
q
uery
(se
g
me
nted i
m
age
) is calculated by usi
ng the followi
ng equ
ation.
63
0
i
hki
hqi
64
1
Hk)
d(Hq
(
2
)
whe
r
e
H
q
and
H
k
a
r
e
que
ry indexing
ke
y and ima
ge i
ndexing
key
s
, re
spe
c
tively. The di
stan
ce
is
equal to 0, if the image is identical in all the
region
s. We then rank the imag
es acco
rdi
n
g
to
overall sco
r
e
and retu
rn to
the twenty be
st match
e
s.
3. Resul
t
s
and
Discus
s
ion
In this experi
m
ent, 5,000 of JPEG image
s used a
nd co
nsi
s
t of 10 classes which
con
s
i
s
t
of
bear, bike, buil
d
ing, car, cat, flower, m
odel
/celeb
rity, m
ountain, sky
, a
nd
tex
t
ure
. The
work
evaluat
es
only the
to
p twe
n
ty ima
ges ra
nked i
n
term
s of th
e
simila
rity me
asu
r
e
s
by
usi
ng
pre
c
isi
on an
d
re
call
pa
ra
meters. Pre
c
isi
on i
s
the
ratio
of the
numb
e
r
of
relevant im
a
ges
retrieve
d to the total numbe
r of retrieve
d
(both
irreleva
nt and releva
nt) image
s
ret
r
ieved. Whilst
,
Re
call is the
ratio of the n
u
mbe
r
of rel
e
vant
image
s retrieved to
the total num
ber of rel
e
va
n
t
image
s in the
databa
se.
retrieved
images
of
number
retrieved
images
relevant
Precision
class
in the
images
relevant
retrieved
images
relevant
Recall
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 11, No. 1, March 2
013 : 119 – 1
2
6
124
Figure 2. The
effectiveness of image retri
e
val by grayscale a
nd seg
m
ented ima
g
e
s
based metho
d
s produ
ce
d by the work
The excellent precision
of 0.98 has been demonstrated by applyin
g
region
growing
technique f
o
r bear class, and worst precision
of 0.25 for texture cla
ss.
Interesting r
e
sult, gra
y
scale meth
od shows that the best precision o
f
0.88 also for
bear, and t
he lowest
precision
of
0.32 for
cat cla
ss a
s
shown in
table 1.
T
h
e
experimenta
l
result
sho
w
s that the
proposed method on se
gmented images prese
n
ts
good precision which a
r
e higher th
an 0.50 on
all classe
s except for texture class.
Figure 2. shows that all precision an
d recall values of segmented images greater than
grayscale image (no segmentation applied).
This re
sult
s demonstr
ate that re
gion
growing seg
m
entation offers more effective method for image indexing and retrieving
compared to un-segmented image or grayscale
image with full DCT
coefficie
n
ts
as
some previ
ous researchers done [1], [3], [6], [14]. Since the proposed method u
s
ed
only object without its background as well as used 1/64 of image size, so that it wil
l
save time and huge of
storage in image processing a
s
ge
neral, especially for Content
Based Image Retrieval. For further precision and
recall deta
i
l , it can be
seen in table
1.
Table. 1. Precisi
on and recall
on grayscale and
segm
ented images
Table 1 al
so
illustrates the aver
age (mean) of preci
s
i
on 0.60 an
d
0.75 for grayscale and
segm
ented
i
m
age
s, resp
ectively. The
experim
ent
re
sult
s
sho
w
that ap
plying
of re
gion
g
r
o
w
ing
techni
que (segm
ented i
m
age
s) give
s better avera
ge
preci
s
ion
of ten classe
s in the data
base.
The most si
gnifica
nt of this wo
rk, it
use
d
only
DC coeffici
ent
s instea
d of using all DCT
0
0.2
0.4
0.6
0.8
1
0.03
0.08
0.1
0
.12
0
.13
PRE
C
ISION
RECALL
PRECISION/RECALL
grayscale
segmented
Class
Gra
y
scale i
m
ages
Seg
m
ented i
m
ages
Precision Recall
Precision
Recall
Bear 0.
88
0.
03
0.
98
0.
01
Bike 0.
70
0.
08
0.
92
0.
02
Build
0.
60
0.
10
0.
88
0.
03
Car
s
0.
58
0.
11
0.
93
0.
04
Cat 0.
25
0.
19
0.
78
0.
06
Flower 0.
47
0.
13
0.
72
0.
07
M
odel 0.
67
0.
08
0.
70
0.
08
M
ount 0.
48
0.
13
0.
68
0.
09
Sky
0.
72
0.
07
0.
53
0.
12
T
e
xt 0.
52
0.
12
0.
33
0.
17
Average 0.
60
0,
75
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Segm
entation for Im
age Inde
xing a
nd
Retrie
val o
n
Discrete Co
si
nes
Dom
a
in (Suhend
ro YI)
125
coeffici
ents o
r
pixels
whi
c
h saved 1/6
4
stora
ge
(im
a
ge si
ze
) to construct ima
g
e
indexing a
n
d
retrieval a
n
d
to segme
n
t images
whi
c
h some a
u
thors u
s
ed p
r
eviously [6], [11], [16].
DC
coeffici
ent is
one of 64 DCT coefficie
n
ts as an imag
e can b
e
co
nstructed a
s
8 x 8 array or blo
ck.
To build i
n
dexing
keys,
as illustration given the
N-bl
ocks of
an image, it can
be used to
con
s
tru
c
t in
d
e
xing keys
b
y
using
only
DC
co
efficien
ts of eve
r
y image i
n
data
base u
s
ing
this
equatio
n:
64
1
i
hi
H
(
3
)
whe
r
e
N
i
N
DCi
hi
1
, is the
i
th
DC coefficient
of eve
r
y block an
d
N
i
s
th
e n
u
m
ber of
block
i
n
the
image
s. In
building in
de
xing keys, this app
roa
c
h o
n
ly consi
d
e
r
s one DC coe
fficient in every
block, which
mean
s it only
need
s1/6
4
th
of the time ta
ken
by the ful
l
extraction, i
ndexing
pro
c
ess.
Due
to the
li
mitations
of
hard
w
a
r
e
sp
eed, the
work o
n
ly u
s
ed
a data
b
a
s
e
o
f
5,000
im
a
ges
.
Comp
are to
the previo
us work, this
rese
arch
h
a
s two me
rits
simple
r al
gorithm and fa
ster
pro
c
e
ssi
ng.
Whil
st, the effectivene
ss image retrie
va
l of this work
and m
o
st other ima
g
e
retrieval
re
se
arche
s
i
s
n
o
t
comp
arable
due to
the
resu
lt
of
ima
ge
retrieval
has
very
mu
ch
influen
ced by
characte
ri
stics of ima
ge d
a
taba
se u
s
ed
in the rese
arch.
4. Conclusion and Futur
e Work
s
Ne
w ap
pro
a
ch ha
s be
en p
r
opo
sed fo
r an
image
retriev
a
l system
ba
sed on
regi
on
gro
w
ing
segm
entation
on
DCT
co
mpre
ss d
o
m
a
in. It is p
r
e
s
ented
as a
di
fferent
way t
o
devel
op i
m
age
indexing by usin
g of DCT descri
p
tors. The me
thod ha
s been
carried out
for com
p
re
ssed
image
s d
a
ta
base to ve
rify its pe
rform
ance in
JPE
G
sta
nda
rd
strea
m
lin
e.
The
propo
sed
method
of re
gion g
r
o
w
ing
segm
entation
on DC im
ag
es offe
rs
hug
e sto
r
ag
e an
d time saving
for
Image indexi
ng and retriev
i
ng.
From thi
s
wo
rk, it co
uld b
e
con
c
lu
ded th
at s
egm
entati
on, while i
m
p
e
rfect, is
an
essential
step
and
very useful
in
b
u
ilding
indexi
ng
keys.
In
su
mma
ry, this i
ndexin
g
key m
e
thod
i
s
a
promi
s
in
g
m
e
thod
fo
r
im
age retri
e
val on seg
m
ente
d
imag
e on comp
re
ss
d
o
m
ain.
This new
approa
ch
co
uld be u
s
ed
for image i
ndexing by
other
segm
e
n
tation meth
ods. Fo
r the
near
future, it will be use
d
anoth
e
r se
gme
n
tation app
roa
c
h
e
s
su
ch a
s
Su
pport Ve
ctor
Machi
ne, Fu
zzy
logic, an
d Split Merge to im
prove spee
d of image inde
xing and Retrieval.
Quer
y
G
r
ay
scale
Segmented
Rank 0
Rank 1
Rank 2
Rank 3
Rank 4
Rank 5
Rank 6
Rank 7
Rank 8
Rank 9
Figure 3. Result exampl
es of the syst
em for segm
ented imag
es retrieved wit
h
RGB
image qu
ery,
RGB image
converte
d in
to grayscale,
the grayscal
e image
then partition
ed by regio
n
gro
w
ing te
ch
nique.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 11, No. 1, March 2
013 : 119 – 1
2
6
126
Referen
ces
[1]
N. A. Mat-Isa, M. Y. Mashor,
N. H. Othman.
S
eed
ed
Re
gio
n
Gro
w
i
ng F
e
a
t
ures E
x
tractio
n
Alg
o
rithm:
Its Potential U
s
e In Improvi
n
g Scree
n
in
g F
o
r Cervic
al C
a
ncer.
Internati
o
nal J
ourn
a
l of t
he C
o
mputer,
T
he Internet And Man
a
g
e
m
e
n
t.
2005; 13(
1): 61-70.
[2]
T
.
P.
Minka, R.W
.
Picard.199
7. Inte
ractive Learn
i
ng Us
ing
a Societ
y
of Mode
ls.
Pattern Recognition
.
199
7; 30(3): 56
5-57
3.
[3]
Mehmet Sezg
i
n
, Bu¨lent Sa
nkur. Surve
y
over
Image T
h
resh
old
i
n
g
T
e
chn
i
qu
es an
d
Quantitative
Performanc
e Evalu
a
tion.
Jo
ur
nal of Electro
n
i
c
Imagi
ng
. 20
0
4
;13(1): 14
6–
1
65.
[4]
A
dolfo Mart
ı
´
n
ez-Uso, F
i
l
i
bert
o
Pla, P
edr
o G
a
rc
ı
´a-S
evil
la. Unsu
pervis
ed Colo
ur
Ima
ge Segme
n
tatio
n
b
y
Lo
w
-
Lev
el Perceptu
a
l
Gr
oup
ing.
Pa
tte
rn
An
al
ysi
s an
d Ap
l
i
c
a
t
i
o
n
.
S
p
ring
er. 201
1. 1
23-1
32. DOI
10.10
07/s1
00
4
4
-01
1
-02
59-
1.
[5]
A
.W
.M. Smeulders, M. W
o
rri
ng, S. Sa
ntin
i,
A. Gupta a
n
d
R. Jai
n
. Co
nte
n
t-Based
Imag
e R
e
trieva
l at
the End of the
Earl
y
Ye
ars.
IEEE Transacti
on on Patter
n
Analys
is and
Machi
ne Intell
i
genc
e
. 200
0;
22(1
2
): 134
9-1
380.
[6]
Jun Su
n, Yan
W
ang1,
Xi
ao
ho
ng W
u
,
Xia
o
d
ong Z
h
an
g, H
ong
ya
n Ga
o. A Ne
w
Ima
ge
Segme
n
tatio
n
A
lgorit
hm and
Its Application
in
Lettuce Obj
e
ct Segmentati
on.
TELKOMNIKA.
2012; 10(
3): 557-5
6
3
.
e-ISSN: 2087-
278
X
[7]
Nidh
i
. Sing
ha
i, Shishir K. S
han
dil
y
a. A S
u
rve
y
On: Co
ntent Base
d Image R
e
triev
a
l S
y
stems.
Internatio
na
l Journ
a
l of
Co
mputer App
lic
ations
. 201
0; 4(2)
: 22-26.
[8]
Riten
dra Datta,
Dhiraj Jos
h
i, Jia Li, James Z
.
W
ang. Image Retriev
a
l: Idea
s, Influences, and T
r
ends
of the Ne
w
Ag
e.
ACM Comp
uting Surv
eys
. 200
8; 40(2): Article 5.
[9]
J. Philbin,
M.
Isard, J. Sivic,
A. Zisserman.
Descri
p
tor L
earn
i
ng
F
o
r Ef
ficient
Retriev
a
l
. Euro
pe
an
Confer
ence
on
Computer Vi
s
i
on. 201
0; 5(2): 213-
217.
[10]
A
.K. Jain. F
undame
n
tals of D
i
gital Ima
ge Pr
ocessi
ng.
Upp
e
r Sadd
le Riv
e
r
. NJ: Prentice Hall. 1
989.
[11]
A
nn
y Y
uni
arti, Anin
dh
ita Si
git Nu
gro
ho,
Bilqis
Am
al
ia
h
,
Agus Z
a
i
n
a
l
Arifin. Cl
assi
fication
an
d
Numb
erin
g of Denta
l
R
a
d
i
og
raphs
for an
A
u
tomated
H
u
m
an Id
entific
atio
n S
y
stem.
TEL
K
OMNIKA
.
201
2; 10(1): 13
7-14
6.
[12]
N. A. Mat-Isa, M. Y. Mashor
,
N. H. Othman.
Seed
ed
Re
gio
n
Gro
w
i
ng F
e
a
t
ures E
x
tractio
n
Alg
o
rithm:
Its Potential U
s
e In Improvin
g Screen
in
g fo
r Cervical
Can
c
er.
Internatio
n
a
l Jour
nal
of the Co
mp
uter,
T
he Internet an
d Mana
ge
ment
.
2005: 13;(1):
61 -70.
[13]
Kanchan Deshmukh,G. N. Shinde
. An Adaptive Neuro-Fuzzy
Sy
st
em
For
Color Imag
e Segmentation.
Journ
a
l Indi
an
Institute of Science
. 20
06;8
6
(
1
): 493–
50
6.
[14]
R. Adams, L. Bischof. See
d
e
d
Reg
i
o
n
Grow
i
ng.
IEEE Transaction on P
a
ttern Analysis
and
Ma
ch
i
ne
Intelli
gent
. 19
9
4
;16(6): 64
1-6
47.
[15]
Z
.
T
u
,S.-C. Z
hu.
Image
Seg
m
entatio
n b
y
Data-driv
en M
a
rkov Cha
i
n M
onte Carl
o.
IEEE Transaction
on Pattern Analysis
. 200
2; 24(
5): 657–
67
3.
[16]
Mustafa Ozde
n, Ediz
Pol
a
t. A Co
lor Ima
ge
S
egme
n
tati
on A
ppro
a
ch
F
o
r Cont
ent-B
ased
Imag
e
Retriev
a
l.
Pattern Rec
ogn
itio
n
. 2007; 4
0
(2): 131
8 – 13
25.
Evaluation Warning : The document was created with Spire.PDF for Python.