Int
ern
at
i
onal
Journ
al of
El
e
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
4593~
4602
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp4593
-
46
02
4593
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Hybrid
Mu
ltilevel Thr
esh
oldin
g and Imp
roved H
ar
m
ony
Search Alg
or
ith
m for Se
gment
ation
Er
w
in
1
, S
apar
udin
2
, Wul
andari S
aputri
3
1,3
Depa
rtment
of
Com
pute
r
Eng
i
nee
ring
,
Univ
ersity
of
Sriwijay
a
,
Indone
sia
2
Depa
rtment of I
nform
at
ic
Engi
n
ee
ring
,
Univ
ersity
of
Sriwij
a
y
a
,
I
ndonesia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
9
, 2
01
8
Re
vised
Ju
l
12
,
201
8
Accepte
d
J
ul
22
, 2
01
8
Thi
s
pape
r
pro
poses
a
new
m
et
hod
for
image
segm
ent
a
tion
is
h
y
brid
m
ult
il
evel
thre
s
holdi
ng
and
improved
har
m
on
y
sea
rch
al
go
rit
hm
.
Im
prove
d
har
m
on
y
sea
r
ch
al
gorit
hm
which
is
a
m
et
hod
for
findi
ng
ve
ct
or
s
olut
ions
b
y
inc
re
asing
it
s
accura
c
y
.
The
pro
posed
m
et
hod
looks
for
a
ran
dom
ca
ndida
t
e
soluti
on,
th
en
its
qual
ity
is
ev
aluate
d
throug
h
t
he
Otsu
obje
cti
ve
func
ti
on
.
Furthermore,
th
e
oper
a
tor
conti
nues
to
evol
v
e
t
he
soluti
on
c
and
ida
t
e
ci
r
cui
t
unti
l
the
opti
m
al
soluti
on
is
foun
d.
Th
e
d
at
ase
t
us
ed
in
thi
s
stud
y
i
s
the
r
et
in
a
dat
ase
t,
tongu
e,
le
nna
,
baboon,
and
ca
m
era
m
a
n.
The
exp
eri
m
ent
a
l
result
s
show
tha
t
thi
s
m
et
hod
produc
e
s
the
high
per
f
orm
anc
e
as
see
n
from
pea
k
signal
-
to
-
nois
e
r
at
io
anal
y
s
is
(PN
SR
).
The
PN
SR
result
for
re
ti
nal
imag
e
ave
rag
ed
40.
34
2
dB
while
for
the
av
era
g
e
to
ngue
image
35
.
340
dB.
For
le
nn
a
,
baboon
a
nd
ca
m
era
m
en
p
roduc
e
an
av
erage
of
33.
781
dB
,
33.
499
dB,
and
34.
869
dB.
Furthermore,
the
proc
ess
of
obje
ct
r
ec
o
gnit
ion
an
d
ide
nti
f
icati
on
is
expe
c
te
d
to
us
e
thi
s
m
et
hod
t
o
produc
e
a
hig
h
degr
e
e
of
ac
cur
acy
.
Ke
yw
or
d:
Im
age seg
m
entat
ion
Im
pr
ov
e
d
ha
r
m
on
y sea
rch
Algorithm
Mult
il
evel thre
sh
ol
ding
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Erw
i
n
,
Dep
a
rtm
ent o
f C
om
pu
te
r
E
ng
i
neer
i
ng,
Un
i
ver
sit
y o
f S
riwij
ay
a,
Jl. Ray
a
Pale
m
bang
-
Pr
a
bum
ulih K
M.
32 In
dral
ay
a O
ga
n Ili
r
3066
2,
Pale
m
bang, I
ndonesi
a
.
Em
a
il
:
erw
in@
un
s
ri.ac.i
d
1.
INTROD
U
CTION
Segm
entat
ion
is
a
proc
ess
in
im
age
processi
ng
that
process
es
the
or
i
gin
al
i
m
age
into
c
onsti
tuent
or
obj
ect
area
s.
T
he
pu
rpose
of
segm
entat
ion
is
to
sepa
rate
a
n
ob
j
ect
f
ro
m
the
w
ho
le
im
a
ge.
C
urren
tl
y,
i
m
age
processi
ng
ca
n
be
app
li
e
d
very
widely
in
var
io
us
f
ie
lds,
for
exam
ple
in
t
he
fiel
ds
of
ast
ron
om
y,
arch
eolo
gy,
and
e
ven
bio
m
edical
.
Im
age
processi
ng
on
bio
m
edici
ne
ha
s
been
widely
us
ed
,
inclu
ding
face
detect
io
n
[
1],
iris
[
2],
ear
,
a
nd
to
ngue
.
Us
ing
im
age
pr
oc
essing
te
ch
ni
qu
e
s
uch
as
l
evel
set
a
nd
r
egio
n
gro
wing
,
an
ophth
al
m
olo
gi
st
m
ay
kn
ow
t
he
disease
t
hro
ugh
ey
e
reti
na
and
t
he
te
ch
no
l
og
y
ca
n
kn
ow
the
disease
in
t
he
ey
e
reti
na
[
3].
O
ne
of
res
earc
h
co
nducted
by
[
4]
aim
s
to
cl
assif
y
ty
pes
of
dise
ases
thr
ough
th
e
color
of
t
he
tong
ue
with a
su
cces
s
rate o
f 91.9
9
%
accu
racy.
Anothe
r
resea
r
ch
on
tong
ue
im
age
segm
en
t
at
ion
was
do
ne
by
[5
]
with
70%
accu
racy
and
[
6]
us
in
g
act
ive
co
ntour
m
od
el
m
et
ho
d
w
hich
gi
ves
75%
of
acc
uracy
.
Additi
ona
ll
y,
[7
]
com
bin
es
re
gion
-
bas
ed
an
d
edg
e
-
based
m
e
thods
in
segm
enting
im
ages.
Detect
i
on
a
nd
cl
assifi
cat
ion
of
t
he
reti
nal
c
hanges
for
Diabeti
c
Re
ti
no
pat
hy
m
on
it
ori
ng
we
re
perform
ed
by
[8
]
.
T
his
rese
arch
e
xtract
re
trospect
ive
cha
ng
e
s
in
lo
ng
it
ud
i
nal
crack
a
nd
ty
r
osi
neto
pathy
showi
ng
97%
de
te
ct
ion
rate
an
d
99.
3%
cl
assifi
cat
ion
rate.
Re
seach
from
[9
]
has
been co
nducte
d
to
au
t
om
at
ic
segm
entat
ion
a
nd ide
ntific
at
ion
of d
ia
betic
s thro
ugh reti
nal
vessels.
The
m
et
ho
d
us
ed
for
se
gm
entat
ion
is
Ga
bor
wav
el
et
tra
nsf
or
m
at
ion
.
T
herefo
re,
t
he
res
ults
obta
ined
that
tradit
ion
al
featur
es
do
no
t
detect
early
pr
olife
rati
ve
reti
nopathy.
For
the
cl
assifi
cat
io
n
m
et
ho
d
us
e
d
is
the
wav
el
et
m
et
ho
d
that
is
able
t
o
gr
oup
the
reti
nal
blood
vess
el
s
in
accor
da
nc
e
with
the
pr
e
sence
or
a
bs
e
nc
e
of
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4593
-
4602
4594
proliferati
ve
re
ti
no
pat
hy.
The
su
ccess
rate
of
this
m
et
ho
d
decr
ease
d
to
50%
f
o
r
the
ide
ntific
at
ion
f
rom
the
pr
e
vious
sta
ge
s.
Ba
se
d
on
t
he
desc
riptio
n
a
bove
it
is
nece
ssary
to
opti
m
i
ze
segm
entat
ion
process
,
in
orde
r
t
o
i
m
pr
ove the
qu
al
it
y of
seg
m
entat
ion
res
ults.
The
th
res
ho
l
d
m
et
ho
d
is
on
e
of
t
he
m
os
t
widely
us
e
d
m
et
ho
ds
of
im
age
se
gm
ention
[10],[
11]
.
Segm
entat
ion
process
f
or
tw
o
segm
en
of
im
age
is
called
bilevel
thresholding
an
d
m
or
e
than
tw
o
segm
en
is
cal
le
d
m
ulti
lev
el
thres
holdi
ng.
Ots
u
an
d
Kapur
a
re
two
cl
assic
al
m
et
ho
ds
of
m
ulti
le
vel
thres
ho
l
ding.
Segm
entat
ion
us
in
g
m
ult
ilevel
t
hr
es
ho
l
di
ng
re
quires
a
lon
g
c
om
pu
ta
ti
on
ti
m
e
and
in
volves
a
la
rg
e
cal
culat
ion
.
I
n
or
de
r
to
so
l
ve
this
pro
ble
m
,
op
tim
iz
ation
m
et
ho
ds
sh
oul
d
be
ap
plied
[
12
]
.
S
ever
al
op
ti
m
iz
ation
m
et
ho
ds
that
hav
e
bee
n
su
c
cessf
ully
app
li
ed
in
m
ulti
le
v
el
thresholdi
ng,
incl
ud
i
ng
Gen
et
i
c
Algorithm
s
(G
A
)
an
d
Par
ti
cl
e
Sw
arm
Op
ti
m
iz
ation
(P
S
O).
On
e
i
nteresti
ng
exa
m
ple
of
m
ult
il
evel
thres
ho
l
ding
w
it
h
GA
is
sho
wn
i
n
[
13]
,
[
14]
.
F
ur
the
rm
or
e,
[
15
]
us
in
g
m
ul
ti
le
vel
thresh
ol
ding
a
nd
I
m
pr
ov
e
d
Diff
e
re
ntial
Search
Algorith
m
(I
DSA)
to
the
se
gm
ent.
Im
age
seg
m
entat
ion
with
P
SO
-
base
d
m
ul
ti
le
vel
thres
ho
l
ding
w
as do
ne by [
16]
-
[18].
Har
m
on
y
sear
ch
al
gorithm
(H
S
A
)
is
an
op
ti
m
iz
ation
al
gorithm
insp
ired
by
the
im
pr
ov
isa
ti
on
process
of
ja
zz
m
us
ic
ia
ns
,
fro
m
the
ph
en
ome
non
of
ope
ra
m
us
ic
that
consi
sts
o
f
va
rio
us
m
us
ic
al
instru
m
ents
and
pr
oduces
beau
ti
f
ul
m
el
od
ie
s.
T
he
al
go
rithm
was
intr
oduce
d
by
Ge
e
m
et
al
.
[19]
[
20
]
.
I
n
pr
e
vi
ou
s
researc
h,
m
ult
il
evel
threshol
ding
with
H
S
A
was
us
ed
i
n
segm
enting
the
i
m
age
with
two
m
et
ho
ds
of
thres
ho
l
ding,
ie
Otsu
,
an
d
K
ap
ur.
Mult
il
ev
el
thresholding
with
Ots
u
that
has
bee
n
op
ti
m
iz
ed
us
ing
H
S
A
sh
ows
bette
r
r
esults
com
par
e
d
to
Ka
pur
[
19]
.
Im
pr
ove
d
har
m
on
y
searc
h
al
gorithm
(IHS
A)
has
al
s
o
bee
n
app
li
ed
i
n
seg
m
entat
ion
pro
blem
s
fo
r
br
ai
n
im
ages
[2
1].
The
stu
dy
us
e
s
IH
S
A
a
nd
c
om
b
ines
it
wit
h
f
uzzy
cl
us
te
rin
g
al
gorithm
s.
Ho
we
ve
r,
there
rem
ai
ns
a
weakness
in
the
stud
y
is
sh
ow
n
with
th
e
PSN
R
value
that
is
no
t
high e
noug
h.
Re
search
on
re
ti
nal
i
m
age
segm
entat
ion
ha
s
been
done
by
[2
2],
[23]
ai
m
ing
to
sim
pli
fy
or
c
ha
ng
e
i
m
age
rep
rese
nt
at
ion
into
so
m
et
hin
g
t
hat
is
easi
er
to
analy
ze.
[24]
who
ha
ve
co
nducte
d
a
reti
nal
i
m
age
stud
y
by
pro
posin
g
a
com
pu
te
rized
te
chn
iq
ue
for
e
xtracti
ng
reti
na
l
vessels.
In
a
dd
it
io
n,
[25]
a
nd
[
26
]
co
nduc
te
d
a
reti
nal
i
m
age
stud
y
f
or
cl
as
sific
at
ion
se
gm
entat
ion
of
r
et
inal
disease
ty
pes
us
in
g
diff
e
ren
t
m
et
ho
ds.
[
27
]
segm
enting
the
r
et
inal
vessels
us
in
g
a
sin
gle
or
ie
nted
m
ask fil
te
r.
The
ex
pe
rim
e
ntal
resu
lt
s
show
th
at
the
pro
posed
m
et
ho
d
ou
t
perf
or
m
s
a
sing
le
ori
ented
m
ask
filt
er
[28]
.
S
eg
m
enting
the
ar
ea
around
the
r
et
ina
by
us
in
g
adap
ti
ve
s
uper
pix
al
at
io
n
that
is
us
ed
to
dete
ct
the
disease
a
rou
nd
the
reti
na
a
rea.
E
xperim
e
ntal
evaluati
on
giv
es
bette
r
res
ults
with
96%
acc
ur
acy
[29]
.
C
onduct
ed
a
s
tud
y
to
i
den
ti
f
y
the
early
di
agnosis
of
e
pil
eptic
diseases
of
glauc
om
a,
di
abeti
c
reti
no
pathy,
m
acular
de
ge
ne
rati
on,
hype
rtensive
reti
no
pa
thy,
a
nd
arter
ioscle
rosis.
T
he
re
a
re
tw
o
m
et
hods
of
do
i
ng
this
segm
entat
ion
by
us
in
g
the
m
et
ho
d
of
e
xt
racti
on
of
blood
vessel
ce
nt
erli
ne
pix
el
s
and
it
erati
ve
r
egio
n
grow
i
ng.
This
reseac
h
pro
po
se
d
a
novel
m
et
ho
d
t
o
i
m
pr
ove
seg
m
entat
ion
perf
or
m
ance
usi
ng
m
ultilevel
thres
ho
l
ding
w
it
h
HSA
an
d
I
HSA.
This
pa
pe
r
int
rod
uces
a
ne
w
m
et
ho
d
of
hybr
i
d
m
ulti
l
evel
th
res
ho
l
din
g
a
nd
i
m
pr
oved
harm
on
y
search
al
gorithm
(MT
-
IH
S
A
).
The
pa
ram
et
ers
us
ed
diff
e
r
f
ro
m
th
e
HS
A
that
li
es
in
the
adjustm
ent
of
pitch
ad
justi
ng
rate
(P
AR
)
a
nd
ba
ndwidt
h
(B
W)
[
21
]
.
W
her
e
this
m
et
ho
d
c
om
bin
es
I
HSA
m
et
ho
d
a
nd
t
hresh
old
i
ng
us
i
ng
Otsu.
This
researc
h
is
ex
pe
ct
ed
to
sho
w
m
or
e
op
ti
m
a
l
resu
lt
s
of
PS
N
R
than
pr
e
vious
resea
r
ch.
2.
OTSU M
ULT
I
LE
VEL TH
RESH
OLDI
N
G FO
R
I
M
A
GE SEG
MEN
TATION
Im
age
segm
en
ta
ti
on
is
a
pro
cess
to
sepa
rat
e
the
i
m
age
to
the
fore
groun
d
an
d
ba
ck
gro
und
s
o
it
is
easi
er
to
analy
ze
[
30
]
.
The
proces
s
of
im
ag
e
segm
entat
ion
is
ve
ry
im
po
r
ta
nt,
the
hi
gh
e
r
the
accu
racy
le
vel
gen
e
rated
a
t
th
e
segm
entat
ion
sta
ge
t
he
bette
r
the
ob
j
ect
rec
ogniti
on
pr
oce
ss
[
31
]
.
T
hr
es
holdin
g
is
kn
own
as
a
non
-
li
near
ope
rati
on
t
hat
is
i
m
po
rtant
in
im
age
se
gm
entat
i
on
[
32
]
.
The
ba
sic
idea
of
t
hresh
old
in
g
is
t
o
choose
an
opti
m
a
l
gr
ay
-
le
vel
thr
e
sh
ol
d
val
ue
to
sepa
rate
obj
ect
s
a
nd
ba
ckgr
ounds
ba
sed
on
gray
-
le
ve
l
distrib
ution [
33]
.
Ther
e
ar
e
two
ty
pes
of
thres
holdin
g,
ie
glob
al
and
local
.
O
tsu
is
a
glo
bal
thres
ho
l
ding
in
tro
du
ce
d
by
Otsu
i
n
19
79
[
34
]
.
T
his
m
et
ho
d
is
widely
use
d
becau
s
e
it
include
s
a
sim
ple
an
d
ef
fecti
ve
m
et
ho
d.
Ot
su
us
es
the
m
axi
m
u
m
var
ia
nce
value
of
cl
ass
dif
fere
nces
as
the
i
m
age
segm
enting
crit
eria.
By
ta
king
the
int
ensity
le
vel
(L)
of
th
e
gr
ay
scal
e
or
RGB
i
m
age,
the
pro
ba
bili
ty
distrib
ution
of
the
intensit
y
va
lue
of
the
im
age
can
be
cal
culat
e
d
a
s fo
ll
ow
s
E
qu
a
ti
on
1
[19]:
ℎ
=
ℎ
,
∑
ℎ
=
1
,
=
1
(1)
wh
e
re;
i=
inten
sit
y l
evel (0 ≤
i
≤
L
-
1)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Hyb
ri
d
M
ulti
le
vel
Th
res
holdi
ng an
d Impr
ov
ed Ha
r
m
ony
Se
ar
c
h
….
(
Erw
in
)
4595
c
=
com
po
ne
nt
of the im
age,
de
pendin
g o
n gra
ysc
al
e o
r
R
G
B
NP
=
the
num
ber
of
pix
el
s
in
t
he
i
m
age
ℎ
=
histo
gr
am
(
th
e num
ber
of
pix
el
s
c
orrespo
ndin
g
to
the i
nte
ns
it
y l
evel in c
)
ℎ
=
pro
bab
il
it
y of distrib
utio
n
The
sim
plest b
il
evel segm
entat
ion
ca
n be
de
fine
d
as
f
ollows
Eq
uatio
n 2
:
1
=
ℎ
1
0
(
ℎ
)
,
…
ℎ
ℎ
0
(
ℎ
)
,
2
=
ℎ
ℎ
+
1
1
(
ℎ
)
,
…
ℎ
1
(
ℎ
)
(2)
Wh
e
re
(
ℎ
)
an
d
1
(
ℎ
)
is
the d
ist
ri
bu
ti
on proba
bili
ty
f
or
C
1
an
d
C
2
w
he
re :
0
(
ℎ
)
=
∑
ℎ
,
1
(
ℎ
)
=
∑
ℎ
=
ℎ
+
1
ℎ
=
1
0
=
∑
ℎ
0
(
ℎ
)
,
1
=
∑
ℎ
1
(
ℎ
)
=
ℎ
+
1
ℎ
=
1
Otsu
Va
riance
betwee
n
cl
asse
s c
an
b
e
calc
ul
at
ed
by
f
ollow
i
ng E
qu
at
io
ns
3 an
d 4 as
foll
ows:
2
=
1
+
2
,
(3)
1
=
0
(
0
+
)
2
,
2
=
1
(
1
+
1
)
2
(4)
Wh
e
re
=
0
+
1
1
an
d
0
+
1
=
1
.
The follo
wing
is an
obj
ect
ive
functi
on
based
on the
value
of
1
an
d
2
:
(
)
=
ma
x
(
2
(
)
)
(5)
W
it
h
0
≤
ℎ
≤
−
1
,
=
1
,
2
,
.
.
.
,
,
w
her
e
ℎ
=
ℎ
1
,
ℎ
2
,
…
,
ℎ
−
1
is a vect
or co
ntainin
g
se
ver
al
t
hresh
olds
and the
n
the
va
riance is c
om
pu
te
d
a
s E
qu
at
i
on 6 as:
2
=
∑
=
1
,
(6)
3.
IMP
ROVED
HARM
ONY
SEARCH
A
L
GORIT
HM
HSA
is
a
new
m
et
aheu
risti
c
op
ti
m
iz
er
intro
du
ce
d
by
Z
ong
Woo
Geem
,
Jo
on
g
H
oon
Ki
m
,
and
G.V.
Lo
gan
at
han
i
n
2001,
this
m
eth
od
yi
el
de
d
ve
ry
good
re
su
lt
s
in
the
fiel
d
of
op
ti
m
iz
ation
[
32
]
.
HSA
is
in
sp
ire
d
by
i
m
pr
ovise
d
j
azz
m
us
ic
ia
ns
,
f
ro
m
the
pheno
m
enon
of
op
e
ra
m
us
ic
and
pro
duces
be
autiful
m
el
od
ie
s.
T
he
adv
a
ntage
s
of
HSA
com
par
e
d
to
ot
her
op
ti
m
iz
at
ion
te
chni
qu
es
ar
e:
HSA
is
a
m
et
aheu
risti
c
al
gorith
m
and
do
e
s
no
t
re
qu
i
r
e
co
nf
i
gurati
on
val
ues
base
d
on
de
te
rm
inant
va
riables,
H
S
A
us
es
stoc
has
ti
c
rando
m
searches
,
HSA
does
not
require
de
riva
ti
ve
inform
at
io
n,
ha
s
seve
ral
par
am
et
ers,
an
d
can
be
easi
ly
ado
pt
e
d
in
a
wide
range
of opti
m
iz
at
ion
pr
ob
le
m
s.
The
ste
ps
i
n
the
HSA
pr
oc
ess are
as
fo
ll
ow
s
[
21]
:
a.
Mi
ni
m
al
iz
e
(
)
subj
ect
t
o
∈
=
1,2, ..
.,N
wh
e
re;
(
)
=
obj
ect
ive fu
nction
=
colle
ct
ion
of
decisi
on
var
ia
bl
es
N
=
total
of
deci
sion va
riables
=
colle
ct
ion
of
pro
bab
il
it
y ran
ge fo
r
eac
h dec
isi
on
var
ia
bles
In
this
ste
p,
t
he
HSA
pa
ram
e
te
rs
are
s
pecifi
ed.
HSA
pa
ra
m
et
ers
co
ns
ist
of
the
num
ber
of
s
olu
ti
on
vecto
rs
in
harm
on
y
m
e
m
or
y
(H
M)
cal
le
d
ha
rm
on
y
m
e
m
or
y
siz
e
(H
MS
),
har
m
on
y
m
e
mo
ry
c
on
si
der
at
i
on
rate
(H
MC
R),
pitc
h
a
dju
sti
ng
rat
e
(PAR)
,
a
nd
te
rm
inati
on
crit
eria
cal
le
d
a
num
ber
of
im
pr
ov
isa
ti
ons
(
NI)
.
He
re
are the
p
a
ram
e
te
rs
in t
he HS
A
[
33]
:
HMS: tota
l
vec
tors
sim
ultaneo
us
ly
in Harm
on
y M
e
m
or
y (HM
).
Values
v
a
r
y from
1
to 1000.
HMCR
: t
he
le
vel or
per
ce
nta
ge of
H
S
A val
ues
ta
ken f
ro
m
H
M, t
he value
v
a
ries f
ro
m
0
.
7
to
0.99
PA
R:
poi
nter
a
t t
he
le
vel
or pe
rcen
ta
ge of
th
e close
value
, t
he value
va
ries
f
r
om
0
.1
-
0.5
Nu
m
ber
of NI:
ind
ic
at
es t
he
it
erati
on num
ber in the
opti
m
izati
on
alg
ori
thm
.
b.
Har
m
on
y M
em
or
y
(H
M
)
I
niti
al
iz
at
ion
At
this
sta
ge,
the
HM
m
at
r
ix
is
fill
ed
wi
th
HMS
w
hic
h
is
th
e
s
olu
ti
on
vect
or
ra
ndom
iz
ed
by
Eq
uation 7
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4593
-
4602
4596
=
[
1
1
2
1
⋯
−
1
1
1
1
2
2
2
⋯
−
1
2
2
⋮
⋮
⋯
⋮
⋮
1
−
1
2
−
1
⋯
−
1
−
1
−
1
1
2
⋯
−
1
]
(7)
c.
Im
pr
ov
isa
ti
on
of N
e
w Harm
on
y
Im
pr
ov
is
at
io
n
is
pr
oc
essed
t
o
obta
in
ne
w
har
m
on
y.
T
he
new
vector
ha
rm
on
ie
s
can
be
obta
ine
d
unde
r
the
foll
owin
g ru
le
s:
a)
Choose
one
of
the v
al
ues of H
SA
Har
m
on
y
(
HMCR
)
b)
Sele
ct
o
ne
val
ue
closest to
HS
A
m
e
m
or
y (to
ne
a
dju
stm
ent)
c)
Sele
ct
a r
a
ndom
v
al
ue
from
a r
a
ng
e
of
poss
ible val
ues (ra
ndom
iz
at
ion
)
I
n
m
e
m
or
y
co
nsi
der
at
io
ns
,
fir
st
dete
rm
ine
the
val
ue
of
the
var
ia
ble
(
1
1
)
f
or
a
new
vecto
r
ta
ke
n
from
on
e
of
the
valu
es
in
the
pr
e
de
fine
d
HM
rang
e
(
′
1
1
−
′
1
)
.
T
he
value
of
t
he
oth
e
r
det
erm
inant
var
ia
bles
is
picke
d
in
th
e
s
a
m
e
way.
HMCR
(0
to
1)
is
the
ste
p
of
sel
e
ct
ing
a
ra
ndom
value
f
r
om
a
po
s
sible
ra
nge
value
.
At
this
sta
ge
,
HM
co
ns
ide
rat
ion
s
,
to
ne
ad
ju
st
m
ents,
or
ra
ndom
sel
ect
ion
are
ap
plied
al
t
ern
at
el
y
f
or
ea
ch
ne
w
har
m
on
y
vecto
r variable.
d.
Update
HM
In
this
ste
p,
if
the
new
har
m
on
y
vecto
r
is
be
tt
er
than
the
ex
ist
ing
har
m
on
y
in
HM
rather
than
ba
se
d
on
the
val
ue
of
the
ob
j
ect
ive
f
un
ct
io
n,
the
new
ha
rm
on
y
can
e
nter
HM,
an
d
t
he
worst
ha
rm
on
y
will
no
t
be
include
d
i
n H
M.
e.
Check
for t
erm
inati
on
c
rite
ria
If
the
te
rm
inatio
n
c
ri
te
ria
are
m
et
(
m
axi
m
u
m
NI
)
of
t
he
c
om
pu
ta
ti
on
process
will
be
s
topped
.
I
f
no
t
,
rep
eat
ste
ps
3
and
4.
T
he
m
a
in
dif
fer
e
nce
be
tween
HS
A
a
nd
IHSA
is
on
the
PA
R
an
d
B
W
ad
j
ust
m
en
t
path.
IH
S
A
im
pr
ov
e
s
the
perform
a
nce
of
HSA
al
gorithm
s
and
el
i
m
inate
s
weak
po
ints
.
This
m
et
hod
us
es
PAR
a
nd
B
W
in
step
3 (i
m
pr
ov
isa
ti
on).
Pseudoc
ode
for origi
nal IHS
A
al
go
rithm
:
1.
In
it
ia
li
ze pa
ram
et
ers
HMS, HMC
R,
,
,
,
, and N
I.
2.
In
it
ia
li
ze HM a
nd cal
culat
e f(x
) of
e
ach
har
m
on
y
ve
ct
or
.
3.
Im
pr
ov
ise
n
e
w
h
a
rm
on
y.
for
it
erati
on
≤ nu
m
ber
of v
a
ri
able
PA
R
=
+
(
−
)
x g
n
c
=
in(
/
)
/ NI
BW
=
x
e
xp(c
x gn)
f
or(all
v
a
ria
ble)
i
f
ra
nd(
) ≤
HMCR
′
=
(j
=
1, 2, .
.., H
MS) (
c
hoose
va
lue from
H
M)
if
rand()
≤
P
AR
′
=
′
±
rand
()
x
B
W
en
dif
el
s
e
(cho
ose
a r
a
ndom
v
a
lue of
var
ia
ble)
′
=
+
rand()
x (
-
)
e
nd
i
f
e
ndf
or
endf
or
4.U
pdat
e
HM.
if(n
e
w sol
ution ≤
w
or
st s
olu
ti
on)
re
place t
he w
orst ha
rm
on
y i
n HM with
the
new
har
m
on
y
end
i
f
S
5.
C
hec
k st
opping c
rite
ria. If
NI is com
plete
d,
te
rm
ina
te
co
m
pu
ta
ti
on; othe
rw
ise
go
back to
Step
3.
4.
PROP
OSE
D MET
HO
D: H
YBRID
M
ULTIL
EVEL
TH
RESH
OLDI
N
G AN
D
I
MP
R
OVED
HARM
ONY
SEARCH
A
L
GORIT
HM
(MT
-
I
HS
A)
This
m
et
ho
d
is
a
hybri
d
betw
een
tw
o
strat
if
ie
d
thres
holdi
ng
m
et
ho
ds
by
Otsu
a
nd
I
HSA,
wh
ic
h
is
MT
-
I
HSA.
T
he
pro
posed
m
e
thod
is
to
sear
ch
r
an
dom
l
y
i
n
the
histo
gra
m
as
a
cand
id
at
e,
then
eval
ua
te
it
s
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Hyb
ri
d
M
ulti
le
vel
Th
res
holdi
ng an
d Impr
ov
ed Ha
r
m
ony
Se
ar
c
h
….
(
Erw
in
)
4597
qu
al
it
y
by
us
ing
the
obj
ect
i
ve
f
unct
io
n
by
Otsu
.
F
ur
th
erm
or
e,
I
HSA
operato
rs
will
evo
l
ve
on
c
and
i
dat
e
strings
un
ti
l t
he
m
os
t op
ti
m
a
l
so
l
ution i
s fo
und.
Pse
ud
ocod
e f
or
MT
-
I
HSA alg
ori
thm
:
1.
Ob
ta
in
h
ist
ogr
a
m
s.
2.
Ca
lc
ulate
the proba
bili
ty
d
ist
ribu
ti
on.
3.
In
it
ia
li
ze the IHS
A param
et
e
rs: H
MS
, HM
CR
,
,
,
,
, a
nd
NI
.
4.
In
it
ia
li
ze a Ha
r
m
on
y M
e
m
or
y (H
M
).
Step
5
Ca
lc
ulate
O
tsu
v
a
riance.
6.
Evaluate
obj
ec
ti
ve
f
un
ct
io
n
7.
Im
pr
ov
ise
a
ne
w harm
on
y.
8
Update the
HM
9.
If
NI is c
om
ple
te
d
or t
he
sto
p crit
eria i
s sati
sf
ie
d,
the
n j
um
p
to Step
10, ot
he
rw
ise
go
back to Ste
p 6.
10.
Sele
ct
the
har
m
on
y t
hat h
a
s the
b
est
obj
ect
iv
e f
unc
ti
on
value.
11
Apply t
he best
thres
ho
l
ds
valu
es to t
he
im
age.
5.
RESU
LT
A
N
D DIS
CUSS
I
ON
The
dataset
use
d
in
this
e
xp
erim
ent
con
sist
ed
of
tw
o
dat
aset
s,
the
reti
na
l
dataset
ob
ta
ined
from
STruct
ured
A
na
ly
sis
of
t
he
R
et
ina
(S
ta
re)
w
it
h
45
0
reti
na
i
m
ages
an
d
to
ngue
dataset
obt
ai
ned
f
ro
m
biom
et
ric
researc
h
cente
r
(BRC
)
with
12
jp
g
f
or
m
at
the
i
m
age
of
the
to
ngue
.
In
ad
diti
on,
le
nn
a,
babo
on,
an
d
ca
m
era
m
an
i
m
ages a
re also
used
for
te
sti
ng
the prop
os
e
d m
et
ho
d. T
he
pa
ram
et
ers
and
values use
d
i
n t
he
MT
-
IH
S
A
pr
es
ente
d
in
Ta
ble
1
c
on
sist
of
NI,
H
MS,
HMCR
,
P
AR
Mi
n,
P
AR
Ma
x,
B
W
Mi
n,
an
d
B
W
Ma
x.
The
value
of the
p
a
ram
et
ers
us
ing
[19],
nam
ely:
Table
1.
Param
et
ers
use
d i
n
M
T
-
I
HSA
Para
m
eters
Valu
es
NI
2
,00
0
HMS
5
HMCR
0
.9
PAR Min
0
.01
PAR Max
0
.99
BW
Min
0
.00
1
BW
Max
0
.1
Table 2
is t
he
r
esult of
im
age
segm
entat
ion
usi
ng
MT
-
IHSA
w
it
h
thres
ho
l
d
value
5,
a h
ist
ogram
o
f
5
ty
pes
of
the
im
age
s
hows
a
ve
ry
sig
nificant
di
ff
ere
nce.
I
n
th
e
or
i
gin
al
im
age,
the
resu
lt
in
g
histo
gr
am
sti
ll
has
red,
gree
n,
blue
(RGB
).
F
or
the
gr
ay
scal
e
i
m
age,
the
res
ul
ti
ng
histo
gra
m
has
a
gr
ay
c
olor,
bu
t
t
he
re
su
lt
ing
colo
r
st
il
l
has
a
ve
ry
hi
gh
col
or
so
it
is
sti
ll
diff
ic
ult
to
disti
nguish
bet
wee
n
foregr
ound
a
nd
bac
kgr
ound
.
T
he
third
histo
gra
m
is
the
res
ulti
ng
histo
gr
am
for
im
age
i
m
ple
m
entat
ion
of
the
MT
-
IHSA
.
The
col
or
s
pr
oduce
d
after
goi
ng
t
hroug
h
the
MT
-
I
HSA
pr
ocess
ha
ve
ve
ry
fe
w
c
olor
com
ponent
s.
Be
cause
the
colo
rs
with
ap
plie
d
MT
-
I
HSA c
olors a
pp
li
ed
are
m
or
e li
kely
to
bin
a
ry.
Table
2
s
hows
the
res
ults
of
the
ap
plica
ti
on
of
se
gm
entat
i
on
us
in
g
m
ultilevel
thres
ho
l
di
ng
.
Be
fore
app
ly
in
g
the
MT
-
I
HSA
first
the
histogram
value
of
the
or
iginal
i
m
age
is
ta
ken
in
orde
r
to
see
the
colo
r
pix
el
values
co
ntain
ed
in
the
or
i
gin
al
im
age.
The
n
do
the
gr
ay
s
cal
e
process
t
o
reduce
t
he
pixe
ls
con
ta
i
ned
in
RG
B
colo
r.
T
he
histogram
ob
ta
ined
from
the
gr
ay
scal
e
ha
s
few
e
r
pix
el
values
.
T
hen
the
ap
plic
at
ion
of
segm
entat
ion
usi
ng
MT
-
I
HSA
.
The
pi
xels
ob
ta
ined
are
ver
y
low.
Ba
ck
gro
und
an
d
f
oreg
r
ound
are
com
plete
l
y
separ
at
e
d.
Althou
gh
t
he
r
esu
lt
ing
pi
xels
ar
e
ver
y
l
ow
,
th
e
resu
lt
in
g
im
age
qu
al
it
y
is
ver
y
good
an
d
is
see
n
m
or
e cle
arly
usi
ng MT
-
IH
S
A
.
The
th
re
shold
value
s
pecified
in
this
te
sti
ng
process
is
th
=
2,3,4,
5.
P
SN
R
i
s
the
value
of
com
par
iso
n
betwee
n
t
he
m
axim
u
m
pix
el
value
of
t
he
im
age u
sin
g
the m
ean
squa
re
e
r
ror
(MSE)
.
MS
E
is
the
a
ver
a
ge
er
r
or
value
be
twee
n
the
segm
ented
i
m
age
and
the
or
i
gin
al
i
m
age.
The
great
er
th
e
PSN
R
res
ults
sh
ow
bette
r
i
m
age
qu
al
it
y.
PSN
R
is
exp
resse
d
in
decibels
(d
B
).
The
value
of
PSN
R
can
be
cat
egorized
we
ll
if
>
=
30
dB,
it
can
be fo
rm
ulate
d
as foll
ows
with
Equati
on
8
a
nd
root m
ean squar
e
er
ror (RM
SE) value
w
it
h Eq
uatio
n 9:
=
20
10
(
255
)
(8)
=
√
∑
∑
(
0
(
,
)
−
ℎ
(
,
)
)
=
1
=
1
(9)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4593
-
4602
4598
wh
e
re;
0
=
or
igi
na
l im
age
ℎ
=
segm
ented
im
age
row x c
ol
=
Tot
al
a
m
ou
nt
of i
m
age row
s
and
co
lum
ns
Anothe
r
goal
of
usi
ng
PSNR
is
to
evalua
te
the
si
m
il
arity
between
the
segm
ented
i
m
age
and
the
or
i
gin
al
im
age.
Com
par
iso
n
of
PNSR
re
su
l
ts
us
in
g
reti
nal
i
m
age
an
d
t
ongue
im
age
in
Table
3.
P
NSR
on
a
reti
nal im
age i
s great
er t
han
with P
NS
R
of t
ongue im
age.
In
T
a
ble
3,
th
e
resu
lt
s
of
th
e
app
li
cat
io
n
of
se
gm
entat
ion
on
reti
nal
im
age
and
t
on
gu
e
ha
ve
be
en
ob
ta
ine
d.
PS
N
R
gen
e
rated
a
bove
30
dB.
T
hat
is,
the
res
ul
t
of
segm
entat
ion
is
do
ne
s
uc
cessf
ully
becau
se
it
has
e
xceed
e
d
30
dB.
The
bi
gg
e
r
the
PNS
R
gets,
the
bet
te
r
the
pix
el
ge
ts.
Segm
entat
ion
perf
or
m
ed
on
the
tongu
e
im
age r
ecei
ves
a l
ow
e
r
PS
NR
value
than that
of the
reti
nal im
age.
Table
2.
Res
ults o
f
Im
age S
e
gm
entat
ion
a
nd
Histo
gr
am
w
it
h
MT
-
I
HSA
w
it
h
Th
res
ho
l
d Value
5
Table
3.
C
om
par
iso
n of PS
N
R Value
and
S
egm
entat
ion
T
hr
es
hold Res
ul
ts wit
h
MT
-
I
H
SA
Using Ret
ina
I
m
age and
T
on
gu
e
I
m
age
I
m
ag
e
th
PNSR
i
m
ag
e
PNSR
Retin
a I
m
ag
e
Ton
g
u
e I
m
ag
e
Retin
a I
m
ag
e
Ton
g
u
e I
m
ag
e
1
2
4
3
.36
3
5
.55
8
7
3
5
.57
8
3
4
.88
3
4
5
.12
1
3
6
.07
4
3
9
.68
3
5
.61
7
4
4
5
.12
1
3
7
.19
8
4
0
.34
9
3
6
.06
1
5
4
8
.13
1
3
8
.01
5
4
1
.14
1
3
7
.17
3
2
2
4
1
.14
1
3
3
.74
5
8
3
9
.68
3
1
.99
2
3
4
2
.11
3
4
.76
1
4
2
.11
3
4
.85
7
4
4
3
.36
3
5
.48
2
4
5
.12
1
3
6
.42
6
5
4
5
.12
1
3
6
.01
3
4
8
.13
1
3
6
.42
6
3
2
4
0
.34
9
3
4
.67
4
9
3
8
.58
8
3
2
.03
7
3
3
9
.68
3
5
.46
8
3
9
.68
3
4
.27
9
4
4
2
.11
3
6
.03
5
4
0
.34
9
3
5
.28
8
5
4
0
.34
9
3
7
.13
8
3
9
.68
3
5
.93
6
4
2
4
1
.14
1
3
4
.76
10
3
9
.09
3
1
.59
5
3
4
2
.11
3
5
.65
7
3
7
.33
9
3
2
.90
8
4
4
3
.36
3
6
.37
7
3
8
.13
1
3
5
.30
6
5
4
5
.12
1
3
7
.30
7
3
9
.68
3
5
.92
6
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Hyb
ri
d
M
ulti
le
vel
Th
res
holdi
ng an
d Impr
ov
ed Ha
r
m
ony
Se
ar
c
h
….
(
Erw
in
)
4599
Table
3.
C
om
par
iso
n of PS
N
R Value
and
S
egm
entat
ion
T
hr
es
hold Res
ul
ts wit
h
MT
-
I
H
SA
Using Ret
ina
I
m
age and
T
on
gu
e
I
m
age
I
m
ag
e
th
PNSR
i
m
ag
e
PNSR
Retin
a I
m
ag
e
Ton
g
u
e I
m
ag
e
Retin
a I
m
ag
e
Ton
g
u
e I
m
ag
e
5
2
3
3
.07
9
3
5
.10
9
11
3
8
.13
1
3
3
.11
5
3
3
4
.51
4
3
6
.01
2
3
8
.13
1
3
4
.59
4
4
3
6
.99
1
3
6
.80
6
3
9
.68
3
5
.65
3
5
3
7
.33
9
3
7
.63
5
4
1
.14
1
3
6
.08
2
6
2
3
9
.1
3
3
.67
3
12
3
3
.98
1
3
4
.41
9
3
4
0
.34
9
3
3
.82
2
3
4
.70
7
3
5
.03
2
4
4
1
.14
1
3
4
.81
4
3
6
.67
3
6
.03
2
5
4
2
.11
3
5
.56
3
7
.33
9
3
6
.98
1
Segm
entat
ion
from
[28]
is
usi
ng
sin
gle
or
i
ented
m
ask
filt
er,
the
res
ult
of
the
segm
entat
ion
has
increase
d
bu
t
the
res
ulti
ng
weig
ht
has
no
t
been
m
axim
iz
ed.
T
he
disadv
a
ntage
of
t
his
m
et
ho
d
is
wh
e
n
processi
ng
th
e
resu
lt
s,
th
e
tim
e
is
done
lo
ng
enou
gh
s
o
that
the
process
is
slow
e
r.
As
f
or
the
process
of
MT
-
IH
S
A,
pr
ocess
data
is
proces
sed
faster
an
d
the
res
ults
ob
t
ai
ned
excee
ds
30
dB
.
T
he
ne
xt
m
et
ho
d
us
e
d
f
or
segm
entat
ion
is
by
us
i
ng
the
Gabor
wa
velet
transfor
m
at
i
on.
But
f
or
t
he
resu
lt
s
obta
ined
t
hat
tradit
ion
al
featur
e
s
do
not
detect
early
prolife
rati
ve
reti
nopathy.
Pe
rce
ntage
of
s
ucce
ss
is
on
ly
50%
.
O
f
al
l
the
m
e
thods
descr
i
bed,
we c
an
see
the c
om
par
ison
that
MT
-
I
HSA is a
n
e
xcell
e
nt m
eth
od
f
or
s
egm
entat
ion
proce
ss
.
Figure
1
is
th
e
res
ult
of
t
he
segm
entat
ion
com
par
iso
n
by
us
in
g
t
he
m
ulti
thres
holdi
ng
ha
rm
on
y
search
al
go
rith
m
(MT
-
HSA)
perform
ed
by
[19]
with
the
MT
-
I
HSA
pro
po
s
ed
m
et
ho
d
us
in
g
le
nn
a
,
ba
boon,
and
cam
era
m
e
n
im
ages.
Re
su
lt
s
from
MT
-
HSA,
P
NS
R
values
obta
ine
d
belo
w
30
dB
.
The
pi
xel
values
ob
ta
ine
d
do
not
m
at
ch
the
def
a
ult
value
of
30
dB.
T
his
m
eans
that
t
he
im
ple
m
entat
ion
us
i
ng
MT
-
HSA
segm
entat
ion
is
no
t
ap
propria
te
and
is
sti
ll
below
the
ave
ra
ge.
H
oweve
r,
f
or
se
gm
entat
io
n
res
ults
us
ing
MT
-
IH
S
A,
t
he
seg
m
entat
ion
obta
ined
ex
ceeds
t
he
30
dB
li
m
it.
Segm
entat
ion
with
the
ap
plica
ti
on
of
MT
-
I
HSA
sh
o
ws
go
od qu
al
it
y and
s
ucce
ss b
eca
us
e
the
pix
el
s
pro
du
ce
d
a
re e
xcell
ent
and also
ex
cee
d 30 dB
.
To
com
par
e
it
us
in
g
othe
r
m
e
thods
pe
rfor
m
ed
by
[35]
u
sin
g
the
m
e
tho
d
of
T
ongue
Color,
Textu
re
,
and
Ge
om
et
ry
Feat
ur
es
(CT
G
F)
,
[19]
with
Mult
il
evel
Thresh
old
in
g
harm
on
y
search
a
lgorit
hm
(MT
-
HSA)
m
et
ho
d,
a
nd
[36]
wit
h
M
ul
ti
le
vel
Thr
es
ho
l
ding
Firefl
y
Algorithm
m
et
ho
d
(MT
-
FA
)
a
nd
Mul
ti
le
vel
Thr
e
sholdi
ng S
ocial
Sp
i
der Al
gorithm
(
MT
-
S
SA
)
are
prese
nt
ed
as
Fig
ur
e
2.
Figure
1. Dia
gra
m
PSN
R
c
omparis
on
MT
-
HS
A
a
nd
MT
-
I
H
SA
So
the
com
par
iso
n
of
P
NSR
resu
lt
s
between
t
he
propose
d
m
et
ho
d
(
MT
-
I
HSA)
wi
th
MT
-
HSA
,
Color,
Te
xture
,
an
d
Geo
m
et
ry
(CTGF),
MT
-
FA
an
d
MT
-
S
SA
m
et
ho
ds
re
su
lt
ed
in
t
he
hi
gh
est
P
NS
R
s
cor
e
.
In
c
reasin
g
the
value
of
P
N
SR
sh
ows
tha
t
the
resu
lt
s
of
i
m
age
segm
entat
ion
with
the
pro
posed
m
et
hod
pro
du
ce
the
be
st segm
entat
ion
qu
al
it
y com
par
ed
w
it
h ot
her m
e
tho
ds.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4593
-
4602
4600
Figure
2. Dia
gra
m
PS
NR c
omparis
on of
c
omparis
on of t
he PNSR
val
ue of
seg
m
entat
ion
m
et
ho
d o
f
MT
-
IH
S
A, M
T
-
HSA, CT
GF, MT
-
FA
,
MT
-
S
SA
with MT
-
I
HSA
6.
CONCL
US
I
O
N
In
this
stu
dy,
we
pro
pose
a
new
m
et
ho
d
of
m
ult
il
evel
threshold
hybri
d
with
i
m
pr
ove
d
ha
rm
on
y
search
al
gorithm
.
This
m
eth
od
is
a
hy
bri
d
betwee
n
th
e
IH
S
A
al
go
rithm
and
the
obj
ect
ive
func
ti
on
of
m
ul
ti
le
vel
t
hr
e
sh
ol
ding
us
in
g
the
Otsu
m
et
ho
d.
T
he
i
nd
ic
a
tor
us
e
d
i
n
this
stu
dy
to
e
val
ua
te
the
perfor
m
ance
of
MT
-
IHSA
i
s
PSN
R.
T
he
r
esults
of
MT
-
I
HSA
experim
e
nts
i
m
ple
m
ent
ed
on
the
reti
na
i
m
age
are
hi
gh
e
r
than
the
im
age
of
the
tongu
e
i
m
age,
bu
t
for
the
resu
lt
s
obta
in
ed,
the
im
a
ge
of
the
to
ngue
pro
du
ce
s
ex
cel
le
nt
segm
entat
ion
com
par
ed
t
o
t
he
reti
nal
im
a
ge.
Like
wise,
for
th
e
im
age
of
Le
na,
ba
boon,
a
nd
cam
era
m
an,
PN
SR
pro
du
c
ed
after
a
pply
ing
MT
-
I
HSA
increase
d.
P
r
evio
us
ly
for
t
he
im
age
of
Len
na,
babo
on,
a
nd
ca
m
era
m
en
app
l
ie
d
us
i
ng
M
T
-
H
SA
with
P
NS
R
res
ult
below
30dB.
T
he
com
par
ison
of
segm
entat
ion
us
ing
oth
e
r
m
et
ho
d
yi
el
ds
the
PNS
R
value
us
in
g
the
highest
M
T
-
I
HSA.
T
he
l
evel
of
il
lum
in
at
ion
of
a
n
ob
je
ct
is
ver
y
in
flue
ntial
fo
r
segm
entat
ion
s
o
that
th
e
resu
lt
s
ob
ta
i
ned
m
or
e
c
le
a
rly
.
For
the
to
ngue
im
age
resu
lt
s
sh
owe
d bett
er
PSN
R
res
ults than
pre
vious st
ud
ie
s
h
i
gh
e
r
th
an 30 dB
.
The
PN
SR
res
ult
f
or
reti
nal
i
m
age
ave
rag
e
d
40,
342
dB
w
hi
le
fo
r
the
ave
r
age
to
ngue
im
age
35.34
0
dB.
F
or
le
nna
,
ba
boon
a
nd
ca
m
era
m
an
produce
a
ve
rag
e
PNSR
33.
78
1
dB,
33.49
9
dB
a
nd
34.
869
dB
resp
ect
ively
.
F
ur
t
her
m
or
e,
t
he
process
of
ob
j
ect
rec
ogniti
on
a
nd
ide
ntific
at
ion
is
e
xpect
ed
to
us
e
this
m
et
ho
d
to pr
oduce a
h
i
gh d
e
gree
of
ac
cur
acy
ACKN
OWLE
DGE
MENTS
This
pa
per
is
pa
rtly
su
pport
ed
by
Direk
t
or
at
Ri
set
d
an
Peng
abd
ia
n
Ma
sya
r
akat,
Dire
ktora
t
Jend
eral
Penguata
n
Ri
s
et
da
n
Pe
nge
m
ban
gan
,
Ke
m
enterian
Ri
s
et
,
Tek
nolo
gi
dan
Pe
nd
i
dik
a
n
Ti
nggi
Ind
onesi
a
a
nd
Re
ct
or
of Univ
ersit
y of Sri
wija
ya
.
REFERE
NCE
S
[1]
M.
Fachr
urroz
i,
Erwin,
Sap
aru
din,
and
Mardi
ana
,
“
Multi
-
obj
ec
t
f
ace
re
cogn
it
ion
using
Con
te
nt
B
ase
d
Im
a
ge
Ret
ri
eva
l
(CBIR
),
”
in
2017
Inte
r
nati
onal
Confe
r
enc
e
on
Elec
trical
Engi
ne
ering
and
Computer
S
ci
en
ce
(
ICECOS)
,
2017,
pp
.
193
–
1
97.
[2]
M.
Fachr
urroz
i
and
M.
Mujta
h
i
d,
“
Iris
Im
age
Rec
ogni
ti
on
Base
d
on
Inde
pend
en
t
Com
ponent
Anal
y
s
is
and
Supp
ort
Vec
tor
Ma
chi
n
e,”
Tel
komnik
a
,
v
ol.
13
,
no
.
2
,
pp
.
597
–
603,
2015
.
[3]
Y.
Qian
Zh
ao,
X
.
Hong
W
ang,
X
.
Fang
W
ang,
an
d
F.
Y.
Shih,
“
R
et
in
al
vessels
se
gm
ent
at
ion
b
ase
d
on
le
ve
l
set
an
d
reg
ion
growing
,
”
Pa
tt
ern
Recog
nit
.
,
vol
.
4
7,
no.
7,
pp
.
2437
–
244
6,
2014
.
[4]
D.
Zha
ng,
H
.
Z
hang,
and
B
.
Zh
ang,
“
Tongue
C
olor
Anal
y
s
is
for
Medic
a
l
Appli
ca
t
ion,
”
Tongue
Image
Anal
.
,
v
ol.
2013,
pp
.
207
–
2
23,
2017
.
[5]
H.
Murat
et
a
l.
,
“
Feat
ure
Ext
r
act
ion
and
Anal
y
sis
on
{CT}
Im
age
of
Xinji
ang
Lo
c
al
Li
v
er
H
y
datid
b
y
Us
ing
Gra
y
-
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Hyb
ri
d
M
ulti
le
vel
Th
res
holdi
ng an
d Impr
ov
ed Ha
r
m
ony
Se
ar
c
h
….
(
Erw
in
)
4601
sca
le Histogra
m
s,”
Sc
i.
&
a
mp; T
ec
hnol
.
Re
v
.
,
vol
.
30
,
no
.
6
,
pp
.
6
6
–
70,
2010
.
[6]
Sapar
udin,
Erwi
n,
and
M.
Fachr
urroz
i,
“
Tongue
Segm
ent
at
ion
U
sing
Acti
ve
Contour
Model,
”
in
IAE
S
Inte
rnat
io
nal
Confe
renc
e
on
E
le
c
tric
al
Engi
n
e
ering,
Comput
er
Scienc
e
and
Inf
orm
ati
cs
IOP Pu
bli
shing
,
2017,
p
p.
1
–
6.
[7]
K.
W
u
and
D.
Zha
ng,
“
Robust
tongue
segm
en
ta
ti
on
b
y
fusing
reg
ion
-
b
ase
d
a
nd
edge
-
b
ase
d
a
pproa
che
s,
”
Ex
p
ert
Syst.
Appl.
,
vol.
42,
no
.
21
,
pp
.
8
027
–
8038,
2015
.
[8]
H.
Nara
sim
ha
-
I
y
er
et
al
.
,
“
Robust
Dete
c
ti
on
an
d
Cla
ss
ifi
cation
of
Longi
tudi
n
al
Change
s
in
Col
or
Ret
inal
Fundus
Im
age
s for
Moni
tori
ng
Di
abe
t
ic
Ret
inopath
y
,
”
I
E
EE
Tr
ans.
B
iomed.
Eng.
,
vo
l. 53
,
no
.
6
,
pp
.
1084
–
1098,
2006
.
[9]
H.
F.
Jeli
nek,
M.
J.
Cree
,
J.
J.
G.
Le
andr
o,
J.
V.
B.
Soare
s,
R.
M.
Cesar
,
and
a
Lu
cki
e
,
“
Autom
at
ed
segm
ent
at
ion
o
f
ret
in
al
blood
v
e
ss
el
s
and
ide
nt
if
ic
a
ti
on
of
pro
li
f
era
t
ive
di
abetic
ret
inop
at
h
y
.
,
”
J.
Opt.
Soc
.
Am.
A.
Opt
.
Image
S
ci.
Vi
s.
,
vol
.
24
,
no
.
5,
pp
.
14
48
–
145
6,
2007
.
[10]
D.
Zhou
and
H.
Zhou,
“
Minim
isat
ion
of
local
wi
thi
n
-
class
var
ia
n
ce
for
image
seg
m
ent
at
ion
,
”
IET
Image
Proce
ss
.
,
vol.
10
,
no
.
2
,
pp
.
608
–
615
,
2016
.
[11]
X.
Zhe
ng
,
H
.
Ye
,
and
Y.
Ta
ng
,
“
I
m
age
bi
-
l
eve
l
th
resholdi
ng
b
ase
d
on
gra
y
level
-
lo
ca
l
var
i
ance
hist
ogra
m
,
”
Ent
ropy
,
vol.
19
,
no
.
5
,
pp
.
1
–
8
,
2017
.
[12]
K.
Chen,
Y.
Zh
ou,
Z
.
Zha
ng
,
M
.
Dai
,
Y.
Chao
,
and
J.
Shi
,
“
Mul
ti
le
v
el
Im
age
Se
gm
ent
at
ion
Base
d
on
an
Im
prove
d
Firefly
Algor
it
h
m
,
”
Math
.
Probl.
Eng
.
,
vol
.
2016
,
pp
.
1
–
12
,
2016
.
[13]
S.
Manika
n
dan
,
K.
Ramar,
M.
W
il
lj
uice
Iruth
a
y
araja
n
,
and
K.
G.
Sriniva
sa
gan,
“
Multi
l
evel
thre
sholding
f
or
segm
ent
at
ion
of
m
edi
ca
l
br
ai
n
i
m
age
s
using
rea
l
code
d
genetic
a
lgori
thm,”
M
eas.
J.
Int
.
Me
as.
C
onfe
d.
,
vol.
47
,
no.
1,
pp
.
558
–
568
,
2014.
[14]
S.
Kotte
,
P.
Raje
sh
Kum
ar,
and
S.
K.
Inj
et
i
,
“
An
e
ffic
i
ent
app
roa
c
h
for
opti
m
al
m
ult
ilevel
thre
shol
ding
select
ion
fo
r
gra
y
sca
le i
m
ages
base
d
on
improved
diff
ere
n
ti
a
l
s
ea
rch
a
lgori
thm,
”
A
in
Shams
Eng
.
J
.
,
2015.
[15]
M.
H.
Moza
ffa
ri
and
W
.
-
S.
Le
e,
“
Converge
nt
he
te
roge
n
eous
par
t
ic
l
e
sw
arm
opti
m
isat
ion
al
gori
t
hm
for
m
ult
il
ev
el
image
thr
esholding segm
ent
ation
,
”
I
ET
Image
Pr
oce
ss
.
,
vol. 11, n
o.
8
,
pp
.
605
–
61
9,
2017
.
[16]
D.
Mishra,
I
.
B
ose,
U.
Chandr
a
De,
and
B.
P
rad
han,
“
A
m
ul
ti
le
v
el
image
th
resholdi
ng
usin
g
par
ticle
sw
arm
opti
m
iz
ation,
”
I
nt.
J. E
ng
.
Te
chn
ol.
,
vol
.
6
,
no
.
2
,
pp.
1204
–
1211
,
2014.
[17]
Y.
Li
u,
C.
Mu,
W
.
Kou,
and
J.
Li
u,
“
Modifie
d
par
ticle
sw
arm
opti
m
iz
ation
-
base
d
m
ult
il
evel
thres
holdi
ng
for
image
segm
ent
at
ion
,
”
S
oft
Comput
.
,
vol. 19, no. 5, pp. 13
11
–
1327,
2015
.
[1
8]
H.
Mar
y
am,
A.
Mus
ta
pha,
and
J.
Younes,
“
A
m
ult
il
evel
thre
s
holdi
ng
m
et
hod
for
image
segm
ent
at
ion
base
d
on
m
ult
iobj
ectiv
e
p
art
i
cl
e
sw
arm opt
imiza
t
ion,”
I
EEE
,
pp
.
0
–
5
,
2017
.
[19]
D.
Oliva
,
E.
Cue
vas,
G.
Paja
res,
D.
Za
ldi
v
ar,
and
M.
Pere
z
-
ci
sner
os,
“
Multi
le
vel
Thre
sholding
Se
gm
ent
at
ion
Base
d
on
Harm
on
y
Sea
rch
Optimiz
at
ion
,
”
J. A
pp
l. Mat
h.
,
vol
.
2013
,
2013
.
[20]
Z.
W
.
Ge
em,
J.
H.
Kim
,
and
G.
Loga
n
at
han
,
“
A
New
Heuri
stic
Optimiz
at
io
n
Algorit
hm
:
Harm
ony
Sear
c
h,
”
Simulat
ion
,
vol
.
76,
no
.
2
,
pp
.
60
–
68,
2001
.
[21
]
M.
Mahda
vi,
M
.
Fesangha
r
y
,
an
d
E.
Dam
angi
r
,
“
An
improved
h
armon
y
sea
r
ch
a
lgori
thm
for
solving
opti
m
izati
o
n
proble
m
s,”
Appl
.
Math. Comput.
,
vol.
188
,
no
.
2
,
p
p.
1567
–
1579
,
2
007.
[22]
B.
Shuai,
Z
.
Z
uo,
B.
W
ang,
G.
W
ang,
and
S.
Mem
ber
,
“
S
ce
ne
Segm
ent
a
t
ion
w
it
h
DA
G
-
Rec
urre
n
t
Neur
al
Networks,”
vo
l.
8828,
no
.
c
,
201
7.
[23]
Z.
Yang
,
Y.
Sh
ufa
n,
G.
Li,
an
d
D.
W
ei
fen
g
,
“
Segm
ent
at
ion
of
MRI
Brai
n
I
m
age
s
with
an
Im
prove
d
Harm
on
y
Sear
chi
ng
Algor
it
hm
,
”
Bi
omed
R
es.
Int
.
,
vol
.
201
6,
2016
.
[24]
J.
Dash
and
N.
Bhoi,
“
A
Thr
es
holdi
ng
Based
Te
chn
ique
to
E
xtra
c
t
Re
ti
na
l
B
lood
Vess
el
s
fro
m
Fundus
Im
ag
es,
”
Fut
ur.
Comput
.
I
nformatic
s J
.
,
20
17.
[25]
C.
Zhu
et
al
.
,
“
Ret
inal
vessel
segm
ent
ation
in
col
our
fundus
images
using
Ext
reme
Learni
ng
Mac
hine,”
Com
put.
Me
d.
Imaging
G
raph.
,
vol
.
55
,
p
p.
68
–
77
,
2017
.
[26]
J.
H.
Ta
n
,
U.
R.
Acha
r
y
a
,
S.
V.
Bhanda
r
y
,
K
.
C.
Chua,
and
S
.
Si
vapr
asa
d,
“
Segm
ent
at
ion
of
op
t
ic
disc
,
fove
a
an
d
ret
in
al
v
asc
ul
at
u
re
using
a
sing
le
convol
uti
on
al neural
n
et
work,
”
J. Comput.
S
ci.
,
v
ol.
20
,
pp
.
70
–
79
,
2017
.
[27]
F.
Kasm
in,
A.
Abdulla
h,
and
A.
S.
Prabuwono,
“
Ret
inal
Blo
od
Vess
el
Segme
ntation
Us
ing
Ensemble
of
Single
Orien
te
d
Mask F
il
te
rs
,
”
IJE
C
E
In
t.
J. E
l
ectr.
Com
put.
Eng.
,
vo
l. 7, no. 3, p. 1414, 2
017.
[28]
N.
Singh
and
R.
Gill,
“
Ret
inal
Area
Segm
ent
ation
using
A
daptive
Superpixala
t
ion
and
it
s
Cla
s
sific
a
ti
on
using
RBF
N,”
IJE
C
E
I
nt.
J. E
l
ectr.
Co
mput.
Eng
.
,
vol
.
6,
no
.
6
,
p
.
2674
,
2016.
[29]
R.
Akhava
n
and
K.
Faez
,
“
Two
novel
retina
l
blo
od
vessel
segm
ent
at
ion
al
gori
th
m
s,”
IJE
CE
Int.
J.
Elec
tr
.
Comput.
Eng.
,
vol
.
4
,
no
.
3,
pp
.
398
–
410
,
2014.
[30]
R.
Dass
and
S. D
evi
,
“
Im
age Se
gm
ent
at
ion
Tech
nique
s,”
IJ
ECT
,
vol.
3
,
no
.
1
,
pp
.
66
–
70,
2012
.
[31]
R.
C.
Gonza
le
s
a
nd
R.
E. Wood,
Digit
al Im
age P
roce
ss
ing
.
Prent
i
ce
Ha
ll,
2005
.
[32]
D.
Manja
rre
s
et
al.
,
“
A
surve
y
o
n
appl
icati
ons
of
the
har
m
on
y
se
arc
h
al
gor
it
hm
,
”
Eng.
Appl
.
Arti
f
.
Inte
l
l.
,
vol.
26,
no.
8
,
pp
.
1818
–
1831,
2013
.
[33]
M.
-
A.
Al
-
Betar and
A.
-
T
.
Khade
r,
“
A ha
rm
on
y
s
ea
rch
al
gor
it
hm
for
unive
rsit
y
co
urse
ti
m
et
ab
li
ng
,
”
Ann
.
Oper. Re
s
.
,
vol.
194
,
no
.
1
,
p
p.
3
–
31
,
2012
.
[3
4]
M.
H.
J.
Val
a
a
nd
A.
B
axi
,
“
A
rev
i
ew
on
Otsu
image
s
egment
at
ion
al
gor
it
hm
,
”
Int
.
J.
Ad
v.
R
es.
Comput
.
En
g.
Technol
.
,
vo
l. 2,
no.
2
,
pp
.
387
–
3
89,
2013
.
[35]
B.
Zh
ang,
B
.
V.
K.
V.
Kum
ar,
and
D.
Zhang,
“
Dete
c
ti
ng
Diabe
t
es
Mellit
us
and
Non
-
Prolife
r
at
iv
e
Dia
b
eti
c
Ret
inopath
y
usin
g
Tongue Col
or
,
T
ext
ur
e , and
G
eometr
y
Fea
tures
,
”
no
.
c
,
20
13.
[36]
R.
Singh,
P.
Ag
arwa
l,
M
.
Kash
y
ap,
and
M
.
Bhatt
ac
har
y
a
,
“
Kapur
’
s
And
Otsu
’
s
Based
Optimal
Multi
le
v
el
Im
ag
e
Thre
sholding
Us
ing
Soci
al
Spid
e
r
and
Fir
efly
A
lg
orit
hm
,
”
no
.
v
,
p
p.
2220
–
2224
,
2
016.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4593
-
4602
4602
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Erwin
was
bor
n
in
Pale
m
bang
,
Indone
sian,
in
1971.
He
rec
ei
ved
the
Bac
h
elor
degr
ee
in
Mathe
m
at
i
cs
fro
m
the
Univer
sit
y
of
Sriwi
jay
a
,
Indone
sian,
in
1
994,
and
the
M.
Sc.
degr
ee
s
in
Actua
r
i
al
from
the
Bandung
Instit
ute
of
Technol
og
y
(IT
B)
,
Band
ung,
Indone
sian,
in
2002.
He
is
stud
y
ing
Ph.D.
d
egr
ee
s
in
Inform
at
i
cs
Engi
n
ee
rin
g
at Unive
rsit
y
o
f
Sriwijay
a
.
In
1
994,
he
joi
n
ed,
Univer
sit
y
of
Sr
iwij
a
y
a
,
as
a
Lectu
rer
.
Since
De
ce
m
ber
2006,
he
has
bee
n
with
the
Depa
rtment
of
Inform
at
ic
s
Engi
ne
eri
ng,
U
nive
rsit
y
of
Sri
wijay
a
,
where
he
was
an
As
sistant
Profess
or,
bec
ame
an
As
sociate
Profess
or
in
2011.
Sinc
e
2012,
h
e
has
bee
n
with
th
e
Depa
rtment
of
Com
pute
r
Enginee
ring
,
Univer
sit
y
of
Sriwij
a
ya
His
cur
r
ent
r
ese
arc
h
intere
sts
inc
lud
e
imag
e
proc
essing,
and computer
vision
.
Erwin, S.
Si.
,
M
.
Si.
is
a
m
ember of IAENG a
nd
I
EE
E
.
Sapar
udin
was
b
orn
in
Pangka
l
P
ina
ng,
Indone
sia
n,
in
1969
.
He
r
ec
e
ive
d
the
B
achel
or
d
egr
ee
in
m
at
hemathi
c
ed
uca
t
ion
from
the
Univer
sit
y
of
Sriwijay
a
,
Indone
sian,
in
1993,
an
d
the
M.T
ec
h
.
degr
ee
s in
infor
m
at
ic
s
from
the Ba
ndung
Instit
u
t
e
of
Te
chno
log
y
(IT
B),
Bandung
,
Indone
si
an,
in
2000
and
Ph.D.
degr
ee
s
in
co
m
pute
r
scie
n
ce
from
the
Malays
ia
n
Univer
sit
y
of
Te
chno
lo
g
y
(UTM),
Johor
Bahru,
Mal
a
y
s
ian,
in
2012.
In
1995,
he
jo
ine
d
,
Univer
sit
y
of
Sriwijay
a
,
as
a
Le
c
ture
r
.
Since
Dec
ember
2006,
he
has
bee
n
with
the
Depa
rtme
nt
of
Inform
at
ics
Engi
nee
ring
,
Univer
sit
y
of
Sri
wijay
a
,
wher
e
h
e
was
an
As
sista
nt
Profess
or,
be
c
ame
an
As
sociat
e
Profess
or
in
2011,
and
a
Profess
or
in
2017.
His
cur
ren
t
rese
arc
h
in
te
rests
i
ncl
ude
image
p
roc
essing,
and
computer
vision.
Drs
.
Sapar
udin,
M.T
.
,
P
h.
D.
is
a
m
ember
of
Instit
ute
of
Advanc
e
d
Engi
nee
r
ing
and
Sci
ence
(IA
ES)
and
I
EEE.
W
ula
ndar
i
Sapu
t
ri
was
born
in
P
al
embang,
Indon
esia
n,
in
1996
.
S
he
is
studen
t
at
Depa
rtment
of
Com
pute
r
Enginee
ring
,
Fa
cul
t
y
of
Com
pute
r
S
ci
en
ce,
Univer
si
t
y
of
Sriwij
a
y
a
,
Indone
sia
.
In
2017,
she
joi
n
e
d
the
La
bor
at
or
y
of
Im
age
Proce
ss
ing,
Univ
er
sit
y
of
Sriwij
a
ya,
as
As
sistant
Le
c
ture
r
.
Her
cur
ren
t
rese
ar
ch
int
ere
sts
inc
lu
de
art
ifici
al
intell
ig
ence,
patter
n
rec
ognit
ion
,
computer
vision
,
and im
age
pr
oc
essing.
Evaluation Warning : The document was created with Spire.PDF for Python.