Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
10
,
No.
1
,
A
pr
il
201
8
, p
p.
154
~
167
IS
S
N:
25
02
-
4752
, DO
I: 10
.11
591/
ijeecs
.
v
10
.i
1
.pp
154
-
167
154
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Multi
-
Level
of F
eatur
e Ex
traction
and Clas
sificatio
n for X
-
Ray
Medic
al Image
M.M
Ab
d
ulra
zz
aq
1
, I
m
ad
F
T Yaseen
2
,
SA Noah
3
, M
oayad
A. Fa
dhil
4
1
,2
Com
pute
r
Science
Dep
art
m
ent,
Inte
rn
at
ion
al Isl
amic
Univer
si
t
y
Malay
s
ia,
IIUM
,
Malay
si
a
3
Facul
t
y
of
Infor
m
at
ion
Sci
ence &
T
ec
hnolog
y
,
Univer
siti
Keba
ngsaa
n
Mal
a
y
s
ia,
UK
M,
Mal
a
y
s
i
a
4
Facul
t
y
of
Infor
m
at
ion
T
ec
hnolo
g
y
,
Univer
si
t
y
of
Phila
d
el
phi
a, Jorda
n
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n
9
, 201
8
Re
vised
Ma
r
2
,
201
8
Accepte
d
Ma
r
18
, 201
8
The
re
ha
s
be
en
a
rise
in
deman
d
for
digitized
m
edi
ca
l
images
over
th
e
l
ast
two
decade
s.
Medical
images'
p
ivot
al
rol
e
in
surgical
pl
anni
ng
is
al
so
an
essenti
a
l
source
of
informati
on
for
disea
ses
and
as
m
edi
ca
l
re
f
er
enc
e
as
wel
l
as
for
th
e
purpo
se
of
r
ese
arc
h
a
nd
tr
a
ini
ng.
Th
e
re
fore
,
eff
ective
techniques
for
m
edi
ca
l
ima
ge
re
trieva
l
and cl
assifi
ca
t
ion
are re
quire
d
to
prov
ide
accurate
sea
rc
h
through
s
ubstant
ial
amount
of
imag
es
in
a
ti
m
ely
m
anne
r
.
Given
the
amount
of
images
tha
t
are
re
qu
ir
ed
to
de
al
with
,
it
is
a
non
-
v
i
able
pra
ctice
to
m
anua
lly
anno
t
at
e
th
ese
m
edica
l
imag
es.
Additi
ona
lly
,
re
tr
ie
ving
an
d
inde
xing
the
m
with
image
vis
ual
f
ea
tur
e
ca
n
not
c
apt
ure
hig
h
le
v
el
of
sem
ant
ic
conc
ep
ts,
which
are
n
e
ce
ss
ar
y
for
ac
cu
ra
te
re
tr
ie
va
l
an
d
eff
e
ct
i
v
e
cl
assifi
ca
t
ion
of
m
edi
ca
l
imag
e
s.
The
re
for
e,
a
n
aut
om
at
ic
m
e
cha
nism
is
re
quire
d
to
addr
ess
the
se
li
m
it
ations.
Address
ing
thi
s,
thi
s
stud
y
form
ula
te
d
an
eff
e
ct
iv
e
class
ifi
cation
for
X
-
ra
y
m
edica
l
images
using
diffe
r
ent
fe
a
tur
e
ext
ra
ct
ions
and
cl
assifi
ca
t
ion
t
e
chni
ques.
Spec
if
ic
a
ll
y
,
thi
s
stud
y
p
roposed
per
ti
n
ent
fe
at
ur
e
ext
ra
ct
ion
algorithm
for
X
-
ra
y
m
edica
l
i
m
age
s
and
det
ermined
m
achine
learni
ng
m
et
hods
for
aut
om
at
ic
X
-
ra
y
m
e
dic
a
l
image
cl
assifi
ca
t
ion.
T
his
stud
y
al
so
eva
lu
at
ed
diff
er
ent
image
fe
at
u
re
s
(c
hief
l
y
globa
l
,
local
,
a
nd
combined)
a
nd
cl
assifi
ers.
Consequent
l
y
,
t
he
obtained
re
sults
from
thi
s
study
improved
re
sults
obtai
ned
from
pre
vious
re
la
ted
studie
s.
Ke
yw
or
d
s
:
X
-
ray m
edical
i
m
age
SV
M
K
-
NN
Feat
ur
e
Ex
t
racti
on
Cl
assifi
cat
ion
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Eng
in
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Moh
am
m
ed
Muay
ad A
bdulra
zzaq,
Com
pu
te
r
Scie
nce
Dep
a
rtm
ent,
Ku
ll
iy
ya
h
of
I
nfor
m
at
ion
a
nd Com
m
un
ic
at
i
on Tec
hnology
(
K
ICT
),
In
te
r
natio
nal Is
lam
ic
U
niv
er
sit
y M
al
ay
sia
E
-
m
ail: eng
.al
ob
ay
dee81
@g
m
ai
l.co
m
1.
INTROD
U
CTION
The
pro
du
ct
io
n
a
nd
relat
ive
ly
strai
gh
tf
orward
m
anag
e
m
ent
of
dig
it
a
l
visu
al
c
on
te
nt
ha
ve
bee
n
increasin
gly
in
dem
and
ov
e
r
these
rece
nt
ye
ars.
Sp
eci
fical
ly
in
the
m
edical
do
m
ai
n
f
or
dig
it
al
inf
orm
a
ti
on
,
the
con
ti
nuous
dev
el
opm
ent
of
m
edical
i
mages
su
c
h
as
X
-
ray,
Com
pute
d
Tom
og
ra
phy
(CT)
sca
ns,
and
Ma
gn
et
ic
Re
sonance
Im
age
(MRI)
scans
c
ontrib
uted
s
ubsta
ntial
a
m
ou
nt
of
im
ages
daily
.
Fo
r
e
xam
ple,
the
Dep
a
rtm
ent
of
Ra
dio
lo
gy
in
Un
ive
rsity
H
os
pital
of
Ge
nev
a
pro
du
ce
d
12,
000
to
15,00
0
im
ag
es
daily
in
20
02
[
1].
T
he
nu
m
ber
of
pro
du
ce
d
a
nd
store
d
im
ages
daily
for
t
his
dep
a
rtm
ent
con
ti
nu
e
d
t
o
i
ncrea
se
t
o
50,00
0
i
m
ages
in
2007
[2
]
and
114,0
00
i
m
ages
in
2009
[
3].
Essenti
al
ly
,
these
i
mages
re
veal
cr
it
ic
al
inf
or
m
at
ion
of
visu
al
ly
inacce
ssible
body
part
s,
w
hich
a
re
e
ssentia
l
for
m
e
dical
diag
nosis
,
m
edical
edu
cat
ion
,
and m
edical
st
ud
ie
s
.
Ther
e
f
or
e,
ef
f
ect
ive
te
ch
niques
t
o
na
vig
a
te
and
search
su
bs
ta
ntial
a
m
ou
nt
of
m
e
dical
i
m
ages
accuratel
y
are
necessa
ry.
T
he
co
nv
e
ntio
nal
im
age
retrieval
syst
e
m
dep
en
d
s
on
keyw
ord
search
,
in
w
hic
h
th
e
keyw
ords
or
annotat
ed
im
a
ge
desc
riptio
ns
are
m
anu
al
ly
assigne
d
f
or
ind
e
xing
purpose.
Subse
quently
,
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Multi
-
Level
of
Feature
Extr
ac
ti
on
and
Cl
as
si
fi
cation
f
or
X
-
Ray
Me
dical
Image
(
M.
M A
bdulr
az
z
aq
)
155
releva
nt
i
m
ages
are
retrieve
d
us
in
g
this
inde
xing
syst
em
,
wh
ic
h
is
know
n
as
Text
Ba
s
ed
Im
age
Re
trie
val
(TBIR)
.
Howe
ver,
the
TBIR
m
e
tho
d
is
dis
reg
a
rd
e
d
due
t
o
the
pr
ese
nce
of
th
ousa
nd
s
or
eve
n
m
il
li
o
ns
of
i
m
age
in
the
database
.
The
process
of
e
nterin
g
m
et
ad
at
a
to
each
of
these
i
m
ages
is
costly
and
tim
e
-
consum
ing
[4].
Con
se
quently
,
rather
tha
n
de
pe
nd
i
ng
on
TBI
R,
the
Con
te
nt
Ba
se
d
Im
age
Re
trie
val
(CBIR)
m
et
hod
is
op
te
d,
in
w
hi
ch
the
i
m
age
retrieval
proces
s
dep
e
nds
on
f
eat
ur
es
ext
racted
from
the
i
mage
it
sel
f
(the
visu
a
l
con
te
nt
of
an
i
m
age)
.
S
pecif
ic
al
ly
,
low
-
le
ve
l
featur
es
s
uc
h
as
col
or,
te
xt
ur
e,
a
nd
sh
a
pe
are
co
ns
ide
r
ed
as
featur
e v
ect
ors
,
w
hich
are
a
ut
om
atical
ly
extr
act
ed
in
the p
r
ocess
of
sea
rchi
ng
f
or
sp
eci
fic
i
m
ages
with
r
espec
t
to
the
query
i
m
age.
Acc
ordi
ng
ly
,
this
te
ch
nique
is
le
ss
tim
e
-
con
s
um
ing
com
par
e
d
to
the
te
chn
i
qu
e
that
dep
e
nds
on
te
xts
f
or
t
he
pur
po
s
es
of
in
de
xi
ng
an
d
retriev
ing
[
5]
.
Howe
ver,
CB
IR
doe
s
not
inte
rpret
data
in
the
sam
e
way
t
hat
a
hu
m
an
does.
Additi
on
al
ly
,
it
is
inexp
e
dient
f
or
the
sy
stem
to
el
ucidate
hig
h
pi
xel
im
ages
as
how
a
hum
an
pe
rceives
i
m
ages.
Su
c
h
l
i
m
i
t
at
ion
is
kn
ow
n
as
sem
a
ntic
gap
[6
]
,
wh
ic
h
is
sp
eci
fical
ly
def
i
ned
as
the
diff
e
re
nce
bet
ween
how
a
hum
an
per
cei
ve
s
an
i
m
age
bas
ed
on
a
high
-
le
vel
sem
antic
con
ce
pt
and
how
a
c
om
pu
te
r
cl
assifi
es
an
im
age
based
on
lo
w
-
le
ve
l
featur
es.
Ne
ver
t
heless,
in
pract
ic
e,
CB
IR
cannot
be
achie
ved
ba
sed
on
on
ly
si
m
pl
e
ind
epe
nd
ent
vis
ual
featur
es
.
Va
rio
us
m
edical
i
m
age
cl
assifi
cat
ion
m
et
ho
ds
us
in
g
m
achine
le
arn
in
g
are
de
velo
ped
t
o
re
du
ce
t
he
issue
s
of
sem
antic
gap
.
W
it
h
that,
t
his
stu
dy
f
or
m
ulate
d
an
e
ff
ect
ive
cl
assifi
cat
ion
syst
e
m
fo
r
X
-
ray
m
edical
i
m
ag
es
bas
ed
on
m
ul
ti
-
le
vel
feat
ure
e
xtracti
on,
f
eat
ur
e
reducti
on,
a
nd
m
ulti
-
cl
assifi
cat
ion
te
ch
niques.
T
he
e
valuati
on
of
this
integrati
on
was
pe
r
form
e
d
us
i
ng
Im
ageCLEF2005
database
.
A
tt
e
m
pts
to
utili
ze
glo
bal
or
lo
cal
featur
es
wi
th
ei
ther
Suppor
t
Vect
or
Ma
chin
e
(S
VM
)
cl
assi
fi
er
or
k
-
Nea
re
st
Neig
hbor
(
k
-
NN)
cl
assifi
er
f
or
X
-
ray
m
edical
i
m
ages
we
re
pe
rfo
r
m
ed
in
var
i
ou
s
relat
ed
stud
ie
s,
as
s
umm
arized
in
Table
1
[7]
-
[
9]
.
For
this
st
ud
y,
t
he
eval
ua
ti
on
wa
s
bas
ed
on
correct
ness
rat
e.
T
he
c
orrectness
rate,
as
s
how
n
in
E
qu
a
ti
on
1,
i
s
the
resu
lt
of
di
vidi
ng
t
he
nu
m
ber
of
correct
ly
classi
fied
im
ages b
y
the total
nu
m
ber
of
im
ages
.
(1)
2.
ANALY
SIS
AND PR
OPO
S
ED SO
L
UTI
ON
Re
al
ist
ic
ally,
it
is
a
chall
en
ge
to
re
du
ce
the
sem
antic
gap
because
vis
ual
featu
res
of
im
ages
do
not
pr
ese
nt
high
-
le
vel
se
m
antic
c
on
ce
pts
an
d
in
ste
ad
of
util
iz
ing
the
co
nte
nt
of
im
ages,
us
ers
op
t
f
or
te
xt
-
base
d
qu
e
ry.
W
it
h
th
at
, th
is has fur
t
her
in
sti
gated
s
tud
ie
s to devel
op
e
ff
ect
ive m
edical
i
m
age cl
assifi
cat
ion
m
e
thods
.
Howe
ver,
the
fam
i
li
arizat
ion
process
with
the
sem
antic
m
od
el
in
cl
assify
ing
im
ages
an
d
e
nh
a
nci
ng
th
e
retrieval
perf
orm
ance is co
m
plex.
Conver
sel
y,
the
resu
lt
s
obta
ined
from
the
pr
e
vious
stu
di
es
to
cl
assify
X
-
ray
m
edical
i
m
ages
,
as
sh
ow
n
in
Ta
ble
1,
util
iz
ing
gl
ob
al
or
l
ocal
f
eat
ur
es
with
ei
ther
S
VM
or
k
-
N
N
cl
assifi
er
s
are
no
t
regar
de
d
as
the
fi
nest
s
olu
t
ion
s
to
t
he
iss
ue
of
reducin
g
t
he
sem
antic
ga
p.
Thes
e
r
esult
s
rem
ai
ned
vary
from
on
e
a
nothe
r.
Fo
r
exam
ple,
ref
e
rr
in
g
t
o
T
able
1,
R
W
T
H
-
i
6
te
am
ac
hieve
d
er
r
or
r
at
e
of
12.6%
wh
il
e
Mo
ntrea
l
te
a
m
achieve
d
e
rror
rate
of
55.0%
for
th
e
sam
e
dataset
.
Me
an
wh
il
e,
M
ueen
[10]
com
bin
ed
featu
re
e
xtrac
ti
on
s
of
global,
local
,
a
nd
pix
el
f
or
X
-
ray
m
edical
i
mage
cl
assifi
cat
ion
a
nd
a
nnotat
ion
usi
ng
b
oth
SV
M
cl
assifi
er
and
k
-
NN
cl
assifi
e
r.
The
res
ultant
ou
tc
om
e
of
this
com
bine
d
featu
re
e
xtracti
on,
c
onsist
ing
of
57
cl
asse
s
(I
m
ageCLEF2
005
database
)
rev
eal
e
d
that
the
pe
rfor
m
anc
e
of
S
VM
exc
eeded
t
he
pe
rfor
m
ance
of
k
-
NN
i
n
m
os
t
of
the
classes
(sp
e
ci
fical
ly
,
48
cl
asses)
w
hile
the
per
f
or
m
ance
of
k
-
NN
e
xcee
de
d
the
pe
rfor
m
ance
of
SV
M
in
the
r
e
m
ai
nin
g
nin
e
cl
asses
on
ly
.
W
it
h
that,
SVM
was
con
si
de
red
f
or
a
nnot
at
ion
pur
pose.
Thr
ee
hierar
c
hical
le
vels
of
im
age a
nnotati
on w
e
re
appli
ed
to
r
e
duce the
sem
antic
g
ap
.
Ap
a
r
t
f
ro
m
th
at
,
in
a
no
t
her
stud
y
on
4,9
37
X
-
ray
m
edical
i
m
ages,
Fes
ha
rak
i
&
P
our
ghassem
[11
]
achieve
d
accu
racy
rate
of
82.
8%
usi
ng
fe
at
ur
e
ext
racti
on
of
sh
a
pe
a
nd
Ba
ye
sia
n
cl
assifi
er.
Co
nv
ersely
,
Ghofra
ni
[12]
achieve
d
hi
gh
e
r
accu
racy
rate
(90.8%)
us
in
g
featu
re
e
xtract
ion
of
s
ha
pe
and
e
dges
as
w
el
l
as
SV
M
cl
assifi
er
on
a
dataset
of
1,1
69
X
-
ray
m
edical
i
m
ages.
The
acc
ur
ac
y
rate
increased
to
94.
2%
wit
h
the
integrati
on
of
featur
e
extract
ion
of
s
ha
pe
a
nd
te
xture
an
d
SV
M
cl
assifi
er
(
rather
tha
n
neural
cl
assifi
cat
io
n
te
chn
iq
ue
)
on
a
dataset
of
4,402
X
-
r
ay
m
e
dical
i
m
ages
[13]
.
Za
re
[
14
]
util
iz
ed
feature
extracti
ons
of
Gray
Level
Co
-
occurre
nce
Ma
trix
(G
LCM
)
,
Ca
nny,
pix
el
,
B
oW,
a
nd
LPB
d
as
well
as
S
V
M
and
k
-
NN
c
la
ssifie
rs,
wh
e
re
SV
M ac
hieve
d hig
her
accurate
rate (
90%) b
ase
d o
n Im
ageCLEF2
007 databa
se.
In
c
on
cl
us
io
n,
there
is
a
nee
d
to
util
iz
e
an
eff
ect
ive
cl
assif
ic
at
ion
that
integrates
m
ulti
-
l
evel
featu
re
extracti
on
(
glo
bal
an
d
loc
al
featur
es)
a
nd
m
ulti
-
cl
assifi
cat
ion
te
ch
niques
f
or
X
-
ray
m
edical
i
m
age
cl
assifi
cat
ion
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
10
, N
o.
1
,
A
pr
il
201
8
:
154
–
167
156
Table
1
.
Su
m
m
aries o
f rel
at
ed
work
Au
th
o
r
Featu
res
Clas
sif
ier
Databas
e
Se
m
an
ticg
ap
Res
u
lts
RWT
H
-
i6
IDM
with X3
2
thu
m
b
n
ails
sizes
,
So
b
el f
ilter
1
-
NN
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
1
2
.6%
RWT
H
-
mi
CC
F and
I
DM
1
-
NN
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
1
3
.3%
Ulg
.ac.
b
e
1
6
X1
6
r
an
d
o
m
ly
ex
tracted patch
es f
ro
m
i
m
ag
es.
Decisio
n
T
ree
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
1
4
.1%
Gen
ev
a
-
g
if
t
Gab
o
r
Textu
re
Filt
ers
5
-
NN
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
2
0
.6%
Inf
o
co
m
m
Textu
re
f
eatu
res
SVM
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
2
0
.6%
MI
RAC
LE
weig
h
tin
g
f
u
n
ctio
n
f
o
r
2
0
-
NN
20
-
NN
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
2
1
.4%
NTU
Gra
y
valu
es
1
-
NN
2
-
NN
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
2
1
.7%
NCTU
-
DBLAB
Scalin
g
SVM
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
2
4
.7%
CEA
So
b
el f
ilter
3
-
NN
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
3
6
.9%
Mtho
ly
o
k
e
Ta
m
u
r
a textu
re
f
ea
tu
res
an
d
Gabo
r
k
-
NN
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
3
7
.8%
CINDI
Can
n
y
E
d
g
e detect
o
r
SVM
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
4
5
.3%
Mon
treal
F
o
u
rier
sh
ap
e and
co
n
to
u
r
d
escript
o
rs
-
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Er
ror
Rate
5
5
.7%
Qiu
&
Xu
(
2
0
0
5
)
Blo
b
Gra
y
level t
ex
tu
re
an
d
co
n
trast
SVM
I
m
ag
eCL
EF20
0
5
No
t add
ressed
Accurac
y
:
89%
Mueen
(
2
0
0
9
)
Textu
re,
Shap
e,
L
o
cal &
g
lo
b
al f
eatu
res.
k
-
NN
SVM
I
m
ag
eCL
E
F2
0
0
5
Ad
d
ressed
Accurac
y
:
8
2
% f
o
r
k
-
NN,
8
9
% f
o
r
SVM
Fesh
araki
&
Po
u
rgh
ass
e
m
(20
1
2
)
Sh
ap
e f
eatu
res
Bayesian
4
9
3
7
X
-
ray
No
t add
ressed
8
2
.87
%
Gh
o
f
rani et al.
,
(20
1
2
)
ed
g
es an
d
sh
ap
e
SVM
1
1
6
9
X
-
ray
No
t add
ressed
9
0
.88
%
Moh
a
m
m
ad
i
et
al
.,
(20
1
2
)
Sh
ap
e and
te
x
tu
re
SVM,
Euclid
ean
d
istan
ce,
an
d
n
eu
ral
n
etwo
rk
4
4
0
2
X
-
ray
No
t add
ressed
8
8
.77
%
9
4
.2 %
Fesh
araki &
Po
u
rgh
ass
e
m
(20
1
3
)
Sh
ap
e and
tex
tu
re
k
-
NN and
n
eu
ral
n
etwo
rk
2
1
5
8
X
-
ray
Ad
d
ressed
9
3
.6%
Zar
e
et al
.,
(20
1
3
)
GLCM
,
Can
n
y
Pix
el,
Bo
W
,
LPB
SVM
k
-
NN
I
m
ag
e
CL
EF20
0
7
No
t add
ressed
9
0
% f
o
r
SVM
8
6
% f
o
r
k
-
NN
3.
METHO
DOL
OGY
This
pr
e
sent
stud
y
pro
pose
d
a
fr
am
ewo
rk
t
o
cl
assify
X
-
r
ay
m
edical
i
mages
base
d
on
m
ulti
-
le
vel
featur
e
e
xtract
ion
usi
ng
the
Im
ageCLEF2005
database
.
I
n
this
stud
y
,
the
de
velo
pm
e
nt
of
the
pro
pose
d
fr
am
ewo
r
k
wa
s
base
d
on
f
eat
ur
e
e
xtracti
on,
c
om
bin
at
ion
an
d
sel
ect
ion,
a
nd
cl
assifi
cat
ion
,
w
hic
h
a
re
sp
eci
fical
ly
d
is
cusse
d
in
the
f
ollow
i
ng secti
ons.
3.1
Fea
tu
re E
xt
r
act
i
on
This
stu
dy
extracte
d,
c
om
bin
ed,
an
d
util
iz
e
d
va
rio
us
featu
res
to
ex
plore
diff
e
re
nt
aspec
ts
of
X
-
ra
y
m
edical
i
m
age
s.
As
pr
ese
nte
d
in
Ta
ble
1,
s
ever
al
feat
ur
e
extracti
ons
we
re
util
iz
ed,
where
gl
ob
al
feat
ur
e
a
nd
local
featu
re
w
ere
c
on
si
der
e
d
in
ce
rtai
n
stu
dies.
Me
a
nwhi
le
,
f
or
t
his
stu
dy,
the
f
ollo
wing
feat
ur
e
e
xtr
act
io
n
al
gorithm
s
were
c
onside
re
d:
(
1)
gl
ob
al
featu
re,
(2)
l
ocal
fe
at
ur
e,
(
3)
pi
xel
featu
re,
an
d
(
4)
sp
ee
de
d
up
rob
us
t
featur
e
s (SUR
F)
.
In
par
ti
cula
r,
global
featu
res
wer
e
e
xtracte
d
from
each
im
age
by
app
l
yi
ng
featu
re
te
chn
i
qu
e
s
of
sh
a
pe
an
d
te
xt
ur
e
,
w
hich
ge
ne
rated
282
feat
ur
es
.
T
hese
fe
at
ur
es
incl
ud
e
d
13
0
dim
ension
s
of
sh
a
pe
f
eat
ur
es
and
152
dim
en
sion
s
of
te
xtu
r
e
featur
es
.
Th
e
local
featur
es
,
on
the
oth
e
r
ha
nd,
we
re
extr
act
ed
by
segm
entin
g
the
in
pu
t
im
age
into
fou
r
no
n
-
overlap
ping
blo
c
ks
of
pi
xe
ls,
res
ulti
ng
t
o
the
e
xtracti
on
of
28
2
dim
en
sion
s
from
each
patch.
T
he
pix
el
fe
at
ur
e
was
e
xtra
ct
ed
after
resiz
ing
each
im
age
to
15
x
15
p
ix
el
s,
wh
ic
h
ge
ne
rated
225 feat
ures.
S
URF tec
hn
i
qu
e
subse
qu
e
ntly
ex
tract
ed
15
0 f
eat
ur
es
from
each
im
age.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Multi
-
Level
of
Feature
Extr
ac
ti
on
and
Cl
as
si
fi
cation
f
or
X
-
Ray
Me
dical
Image
(
M.
M A
bdulr
az
z
aq
)
157
3.2
Te
xture F
eature
Essentia
ll
y,
te
xture
feat
u
res
ref
e
r
to
t
he
unde
rly
ing
str
uc
tural
ar
ra
ng
e
m
ent
of
t
he
s
urfaces
i
n
the
input
i
m
age.
Ther
e
a
re
tw
o
ty
pes
of
te
xtu
re
featu
res,
wh
ic
h
are
(1)
Gr
ay
Le
vel
Co
-
occurre
nce
Ma
trix
(G
LCM
)
a
nd
(
2)
W
a
velet
Tr
ansfo
rm
(
W
T)
.
GLCM
was
f
irstl
y
intro
duc
ed
by
[15]
.
It
is
m
ai
nly
util
i
zed
to
com
pu
te
the
se
cond
-
orde
r
te
xt
ur
e
c
har
act
e
risti
cs
in
so
l
ving
the
issues
of
c
at
egorizat
ion
e
ff
ic
ie
ntly
.
F
or
N
x
N
i
m
age,
it
inclu
des
pix
el
s
with
gray
le
vels
of
0,
1,
2,
….
(
G
-
1)
an
d
re
pr
es
ented
by
m
at
rix
w
he
re
e
a
c
h
m
at
rix
el
em
en
t
sta
nds
f
or
the
j
oi
nt
inci
de
nce
of
intensit
y
le
vels
a
nd
with
pros
pects
at
a
certai
n
distanc
e,
d
(which
r
e
fer
s
to
the
r
el
at
ed
d
i
sta
nce
betwee
n ea
ch pair
of
pi
xels a
nd a r
el
at
ed orie
ntati
on
ang
le
)
[
16]
.
In
order
t
o
ob
ta
in
enh
a
nce
d
ou
t
pu
ts
,
seve
r
al
co
-
oc
c
urre
nc
e
m
at
rices
m
us
t
be
c
on
si
de
red
;
one
f
or
each
relat
ed
l
ocati
on
offer
s
va
rio
us
te
xture
fe
at
ur
e
s
or
si
m
il
ar
featur
e
s
at
va
rio
us
s
cal
es.
Se
ver
al
te
xtu
re
m
easur
es
of
G
LCM
co
uld
be
directl
y
cal
culat
ed
[15]
[
17]
[
18
]
[19]
[
20
]
[21]
[
22]
[23]
.
Gen
e
rall
y,
θ
i
s
qu
a
ntize
d
i
nto
four d
i
ff
e
ren
t
di
recti
on
s:
0o, 4
5o, 90o,
and
135o.
In
t
his
st
ud
y,
22
co
-
occurre
nce
m
at
rices
fo
r
each
of
the
se
f
our
directi
on
s
we
re
obta
ined,
w
hi
c
h
include
d
(
1)
a
uto
c
orrelat
ion,
(2)
cl
us
te
r
prom
inence,
(
3)
cl
us
te
r
sh
a
de
,
(
4)
co
ntrast,
(5)
c
orrelat
io
n,
(
6)
diff
e
re
nce
entr
op
y,
(
7)
dif
fere
nce
va
riance,
(8
)
dissim
il
arit
y,
(9
)
ene
rg
y
,
(10)
e
ntropy,
(1
1)
hom
og
e
neity
,
(12)
in
ver
se
di
ff
e
ren
ce
,
(
13)
inv
e
rse
dif
fer
e
nc
e
norm
al
iz
ed,
(14
)
in
ve
rse
di
ff
ere
nce
m
ome
nt
no
rm
alized,
(
15)
inf
or
m
at
ion
m
easur
es
of
correla
ti
on
1,
(16)
inf
orm
a
ti
on
m
easur
es
of
c
orrelat
io
n
2,
(
17)
m
axim
u
m
pro
bab
il
it
y,
(
18)
m
axi
m
u
m
pr
oba
bili
ty
,
(19
)
s
um
aver
a
ge,
(
20)
s
um
entr
op
y,
(
21)
su
m
of
squa
res,
an
d
(22)
su
m
var
ia
nce
.
Con
se
quently
,
88
dim
ensions
we
re
ob
ta
in
ed
.
T
he
fo
ll
owin
g
sect
io
n
re
vea
ls
the
no
ta
ti
on
s
us
e
d
to
descr
i
be
the
var
io
us
feat
ures
of
GLCM
a
nd
e
qu
at
io
ns
ut
il
iz
ed
fo
r
te
xt
ur
e
sta
ti
sti
cs
in
i
m
ages,
as
sh
own
in
Eq
uations 2
–
23
[
15]
[
17
]
[18]
[
19
]
[20]
[
21]
[
22
]
[23]
.
: i
t rep
re
sents
the “e
ntry in
a
norm
al
iz
e
d
GL
CM
”
it
r
epr
e
sents t
he
gray
levels
num
ber
for
k=
2,3,...,2
Ng
for
k=
0,1,...,N
g
-
1
Me
an of
Me
an of p
x
a
nd
py r
es
pecti
ve
ly
They re
present
the “sta
nd
a
rd
dev
ia
ti
ons
of px a
nd p
y”
,
r
es
pe
ct
ively
”
They re
present
“the e
ntr
op
ie
s
of px a
nd
py r
e
sp
e
ct
ively
”
The follo
wing
equ
at
io
ns we
re
u
ti
li
zed to co
m
pu
te
the pres
ented
t
wen
ty
-
t
wo textu
re s
ta
t
ist
ic
s:
(2)
(3)
(4)
,,
th
c
i
j
i
j
g
N
1
,
g
N
x
j
c
i
c
i
j
1
,
g
N
y
i
c
j
c
i
j
,:
,
xy
i
j
i
j
k
c
k
c
i
j
,:
,
ij
xy
i
j
k
c
k
c
i
j
,
c
i
j
,
xy
1
0
g
N
x
y
x
y
i
ip
i
,
xy
,
M
H
M
Y
1
,
l
o
g
xy
ij
H
X
Y
c
i
j
c
i
c
j
2
l
o
g
x
y
x
y
ij
H
X
Y
c
i
c
j
c
i
c
j
2
E
n
e
r
g
y
=
,
ij
c
i
j
E
n
t
r
o
p
y
=
,
l
o
g
,
ij
c
i
j
c
i
j
D
i
s
s
i
m
i
l
a
r
i
t
y
=
.
c
,
ij
i
j
i
j
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
10
, N
o.
1
,
A
pr
il
201
8
:
154
–
167
158
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
(17)
(18)
(19)
Ma
xim
u
m
Cor
relat
ion
C
oeffi
ci
ent = (
sec
on
d l
arg
est
ei
ge
n v
al
ue
of
Q )0.5
(20)
(21)
(22)
2
C
o
n
t
r
a
s
t
=
.
,
ij
i
j
c
i
j
c,
C
o
r
r
e
l
a
ti
o
n
=
xy
ij
i
j
i
j
xy
,
H
o
m
o
g
e
n
e
i
t
y
=
1
ij
c
i
j
ij
A
u
t
o
c
o
r
r
e
l
a
t
i
o
n
=
.
,
ij
i
j
c
i
j
3
C
l
u
s
t
e
r
S
h
a
d
e
=
,
ij
i
j
x
y
c
i
j
4
C
l
u
s
t
e
r
P
r
o
m
i
n
e
n
c
e
=
,
ij
i
j
x
y
c
i
j
M
a
x
i
m
u
m
P
r
o
b
a
b
i
l
i
t
y
=
m
a
x
,
,
i
j
c
i
j
2
S
u
m
o
f
S
q
u
a
r
e
s
=
,
ij
i
c
i
j
22
0
S
u
m
A
v
e
r
a
g
e
=
G
xy
i
i
c
i
2
22
0
S
u
m
A
v
e
r
a
g
e
S
u
m
V
a
r
i
a
n
c
e
=
xy
Ng
i
ii
c
22
0
S
u
m
E
n
t
r
o
p
y
=
l
o
g
G
x
y
x
y
i
c
i
c
i
2
0
D
i
f
f
e
r
e
n
c
e
v
a
r
i
a
n
c
e
=
x
y
x
y
i
i
c
i
D
i
f
f
e
r
e
n
c
e
E
n
t
r
o
p
y
=
-
l
o
g
x
y
x
y
i
c
i
c
i
1
I
n
f
o
r
m
a
t
i
o
n
M
e
a
s
u
r
e
s
o
f
C
o
r
r
e
l
a
t
i
o
n
1
=
m
a
x
,
E
n
t
r
o
p
y
H
X
Y
H
X
H
Y
I
n
f
o
r
m
a
t
i
o
n
M
e
a
s
u
r
e
s
o
f
C
o
r
r
e
l
a
t
i
o
n
2
=
1
-
e
x
p
2
2
H
X
Y
E
n
t
r
o
p
y
,,
W
h
e
r
e
Q
(
i
,
j
)
=
k
xy
c
i
k
c
j
k
c
i
c
k
,
I
n
v
e
r
s
e
D
i
f
f
e
r
e
n
c
e
N
o
r
m
a
l
i
z
e
d
=
1
ij
c
i
j
ij
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Multi
-
Level
of
Feature
Extr
ac
ti
on
and
Cl
as
si
fi
cation
f
or
X
-
Ray
Me
dical
Image
(
M.
M A
bdulr
az
z
aq
)
159
(23)
Me
anwhil
e,
on
e
of
the
m
os
t
com
m
on
ly
us
ed
m
e
tho
ds
for
m
ulti
-
reso
luti
on
im
age
descr
ipti
on
an
d
analy
sis
is
the
WT.
It
s
pecific
al
ly
of
fe
rs
a
n
e
ff
ic
ie
nt
set
of
t
oo
ls
f
or
va
rio
us
ap
plica
ti
on
s
su
c
h
as
com
pressi
on
of
im
ages
or
sign
al
s,
detect
ion
of
obj
ect
s
,
i
m
pr
ovem
ent
of
im
ages,
and
noise
rem
ov
al
.
Wa
velet
s
ar
e
functi
ons
,
sat
is
fyi
ng
a
li
nea
r
com
bin
at
ion
of
var
io
us
c
onve
rsion
an
d
scal
ing
processe
s
of
a
wa
ve
f
unct
ion.
It
util
iz
es
wav
el
et
transf
orm
,
sp
eci
fical
ly
the
Haar
wa
velet
,
to
extract
te
xture
featu
re.
Thi
s
first
known
wav
el
et
is
con
si
der
e
d
as
the
sim
ples
t
wav
el
et
basi
s
,
w
hich
was
util
iz
ed
for
or
thonorm
al
wav
el
et
trans
for
m
wit
h
com
pact su
pp
ort
[24
]
.
Equati
on
24 r
e
pr
e
sents
the
Haar f
un
ct
ion
e
quat
io
n us
ing
a
step
fun
ct
ion
,
.
(24)
The
Haa
r
wa
ve
le
t
wa
s
app
li
e
d
in
this
stu
dy
since
it
is
the
m
os
t
eff
ic
ie
nt
te
chn
iq
ue
to
c
al
culat
e
the
featur
e
vecto
r
[25]
.
This
was
per
f
or
m
ed
by
app
ly
ing
the
Haar
wa
velet
f
or
f
our
tim
es
i
n
orde
r
to
div
i
de
the
input i
m
age in
to
16 sub
-
im
ag
es, as
il
lustrate
d
in
Fig
ure
2.
Fi
gure
2
.
A
pply
ing
Haar
W
a
ve
le
t Four
Tim
es
Each
im
age
I
of
t
he
siz
e
of
was
init
ia
ll
y
resized
int
o
10
0
x
100
pix
el
s.
The
Haar
wa
ve
le
t
was
su
bse
que
ntly
app
li
ed
to
eac
h
i
m
age
befor
e
di
vid
in
g
it
into
four
sub
-
im
ages,
w
her
e
each
im
age
has
siz
e
-
L1
0L
10, L
10H
10, H1
0L
10, a
nd
H10
H10.
In
t
he
s
ub
-
im
age o
f
L
10L
10, low f
r
eq
uen
ci
es
wer
e
present in
bo
t
h
ho
rizo
ntal
directi
on
a
nd
ver
ti
cal
direct
ion
.
Sp
eci
fical
ly
,
low
fr
e
qu
e
nc
ie
s
wer
e
pr
es
ent
in
the
hori
zon
ta
l
directi
on
w
hile
high
f
reque
ncies
we
re
pr
esent
in
the
ve
rtic
al
directi
on.
Howe
ver
,
i
n
the
s
ub
-
im
a
ge
of
H10L
10,
high
fr
e
qu
e
ncies
we
re
pr
ese
nt
in
th
e
horiz
on
ta
l
di
recti
on
w
hile
l
ow
f
reque
ncies
we
re
pr
ese
nt
i
n
the
ver
ti
cal
directi
on.
an
d
i
n
the
H10
H10
s
ub
-
i
m
age,
there
a
re
hi
gh
fr
e
que
ncies
in
bo
t
h
directi
ons
.
Fo
l
lowing
that,
the
Haa
r
wav
el
et
was
a
pp
li
ed
on
the
i
m
age
of
L
10L
10
with
the
siz
e
of
to
ob
ta
i
n
four
ne
w
sub
-
i
m
ages,
w
her
e
each
im
age
ha
s
siz
e
-
L
11L1
1,
L1
1H1
1,
H
11L
11,
a
nd
H
11
H11.
Sim
i
la
r
proces
s
w
as
rep
eat
e
d
twic
e
to
obta
in
s
ub
-
im
age
s
of
and
,
res
pecti
ve
ly
,
as
il
lustra
te
d
in
Fig
ure
2.
Additi
on
al
ly
,
f
our
feat
ur
es
w
ere
com
pu
te
d
f
or
eac
h
of
the
pr
ese
nted
f
our
proce
dures,
w
hich
a
re
(
1)
e
nt
ropy,
(2)
e
nergy,
(3)
m
ean,
and
(
4)
sta
nd
a
r
d
dev
i
at
ion
.
With
t
ha
t,
there
we
re
64
feat
ur
es
c
om
pu
te
d
from
al
l
su
b
-
i
m
ages.
Figure
3
il
lustrate
s
a
sam
ple
i
m
age
us
e
d
as
an
in
put
for
the
Haa
r
WT,
w
hich
w
as
ob
ta
ine
d
f
rom
the
Im
ageCLEF2005 d
at
a
base.
Figure
3
.
Im
age sam
ple f
ro
m
I
m
ageCLEF2005
2
,
I
n
v
e
r
s
e
D
if
f
e
r
e
n
c
e
M
o
m
e
n
t
N
o
r
m
a
l
iz
e
d
=
1
ij
c
i
j
ij
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
10
, N
o.
1
,
A
pr
il
201
8
:
154
–
167
160
The
obta
ine
d
su
b
-
im
ages
after
app
ly
in
g
the
seco
nd
,
third,
and
f
ourt
h
Ha
ar
wav
el
et
are
dep
ic
te
d
i
n
Figure
4,
base
d
on
the
a
f
or
e
m
entioned.
H
a
ar
wa
velet
wa
s
ap
plied
f
our
tim
es
in
or
de
r
to
div
ide
t
he
input
i
m
age
into
16
su
b
-
im
ages
to
get
the
m
os
t
inf
or
m
at
ion
a
bout
the
im
age,
ap
plyi
ng
Haa
r
wa
velet
f
or
th
e
fiv
e
tim
e g
et
s the sub i
m
ages L1
3L
13 e
qu
al
t
o
z
ero, the
refor
e
was
a
ppli
ed onl
y four
ti
m
es.
Figure
4
.
The
ob
ta
ine
d su
b
-
i
m
ages af
te
r
apply
ing
t
he
sec
ond, thi
rd, a
nd fourth
H
aa
r wa
velet
3.3
Sh
ap
e
Fe
at
u
re
The
s
ha
pe
featur
e
offe
r
s
ge
om
et
rical
info
r
m
at
ion
co
ncernin
g
a
n
im
age
obj
ect
,
w
hich
do
e
s
no
t
vary
with
the
va
riat
ion
s
in
th
e
or
ie
ntati
on
,
scal
e,
and
locat
io
n.
F
or
this
pro
cess,
the
sh
ape
in
form
ation
of
a
n
im
age
was
ex
plored
base
d
on
ed
ge
s.
Th
us
,
the
histogram
of
edge
te
chn
iq
ues
a
nd
S
URF
te
ch
nique
we
re
ap
plied
in
this
stud
y
to
e
xtract
the
sha
pe
featur
e
of
im
ages.
Histo
gr
a
m
of
ed
ge
was
util
iz
ed
to
expl
or
e
the
s
hape
featu
r
e
for
each
im
age
.
I
n
pa
rtic
ular,
bo
t
h
gra
dient
hi
stog
ram
and
e
dg
e
ori
entat
ion
histo
gr
am
were
app
li
e
d.
T
he
firs
t
e
dg
e
hist
ogra
m
te
chn
iq
ue
was
util
iz
ed
to
e
xtract
50
f
eat
ur
es
from
each
im
age
w
hile
the
sec
on
d
e
dg
e
histo
gr
am
techn
iq
ue was
util
iz
ed wit
h
a Ca
nny fil
te
r
to
e
xtr
act
8
0 feat
ur
es
from
each
im
a
ge
[26]
.
The
S
URF
te
chn
i
qu
e
has
a
scal
e
and
r
otati
on
in
va
riance
pr
op
e
rty
,
wh
ic
h
facil
it
at
es
obj
ect
identific
at
ion
with
no
re
gard
s
to
the
im
age's
resize
or
re
presentat
io
n
of
ro
ta
ti
on
ar
ound
a
ce
rtai
n
a
xi
s
[
27
]
.
Re
al
ist
ic
ally,
var
ia
nce
occ
urs
becau
se
no
t
al
l
inform
at
i
on
c
ou
l
d
be
c
aptu
red
f
r
om
a
sp
eci
fic
rec
ordin
g.
Invar
ia
nce
is
a
n
esse
ntial
pro
per
ty
of
im
age
since
the
sim
ilarity
m
easur
em
ent
is
prob
a
bl
e
based
on
the
feature
betwee
n
tw
o
im
ages
that
can
no
t
be
du
plica
te
d.
T
hus,
the
S
URF
te
ch
nique
was
ap
plied
to
extract
15
0
fe
at
ur
es
from
each
im
a
ge.
3.4
C
omb
in
ati
on
an
d
Select
io
n
Com
bin
ed
featur
e
re
fer
s
to
th
e
com
bin
at
ion
of
gl
ob
al
featu
re,
l
ocal
feat
ure,
pi
xel
featur
e
,
an
d
S
UR
F
into
one
vect
or.
Figure
5
dep
ic
ts
the
ov
erall
pr
oce
ss
of
feat
ur
e
e
xtr
act
ion
as
well
as
co
m
bin
at
ion
a
nd
sel
ect
ion
. I
n
or
der
t
o
e
xtract p
ixel
featu
res,
im
a
ges
we
re
re
siz
ed
to 15
x
15, which
c
ontribu
te
d
a v
ect
or o
f
22
5
pix
el
featu
res.
The
global
fea
tures
re
fer
to
t
he
featu
res
of
sh
a
pe
a
nd
te
xt
ur
e
,
wh
ic
h
were
extracte
d
fro
m
the
whole
i
m
age;
t
hu
s
the
resu
lt
a
nt
ou
tc
om
e
of
this
com
bin
ed
vecto
r
was
28
2
featu
res,
s
pe
ci
fical
ly
13
0
f
eat
ur
es
from
the
ed
ge
histo
gr
am
,
64
featur
e
s
from
t
he
WT
an
d
88
featu
res
f
ro
m
the
GLCM.
Conve
rsely
,
the
local
featur
e
s
wer
e
extracte
d
by
s
egm
enting
the
i
m
age
into
f
our
no
n
-
ov
e
rlap
patch
es,
w
hich
s
har
e
d
sim
ilar
282
featur
e
s.
T
his
le
d
to
1,1
28
f
eat
ur
es,
com
bin
ed
in
one
loc
al
featur
e
vect
or.
Me
an
w
hile
,
15
0
feat
ur
es
were
ob
ta
ine
d f
or
t
he
SU
RF
.
Figure
5
.
Feat
ure e
xtracti
on, c
om
bin
at
ion
a
nd selec
ti
on
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Multi
-
Level
of
Feature
Extr
ac
ti
on
and
Cl
as
si
fi
cation
f
or
X
-
Ray
Me
dical
Image
(
M.
M A
bdulr
az
z
aq
)
161
As
a
res
ult,
the
ov
e
rall
featur
e
vector
dim
ension
al
it
y
fo
r
eac
h
im
age
equ
al
s
to
1,785
feat
ure
vecto
rs.
Give
n
the
s
ubs
ta
ntial
nu
m
ber
of
featu
re
vec
tors
in
volve
d,
a
certai
n
dim
e
ns
io
nalit
y
red
uc
ti
on
te
ch
niqu
e
m
us
t
be
perform
ed
to d
ecrease
the f
eat
ur
e
vect
or
s
.
The
m
os
t
com
m
on
ly
us
ed
dim
ension
al
it
y
reducti
on
te
c
hniq
ue
i
s
the
pri
nci
pal
com
po
ne
nt
anal
ysi
s
(P
CA
)
[28]
.
This
sim
ple
te
chn
iq
ue
e
ff
e
ct
ively
decr
eas
es
the
dim
ensi
on
al
it
y
of
data.
W
it
h
t
he
ap
plica
ti
on
of
this
te
ch
nique,
the
feat
ur
e
vecto
rs
we
re
r
edu
ce
d
f
r
om
1,
785
into
25,
50,
an
d
100 o
ne
to
stu
dy
an
d ch
oose t
he op
ti
m
al
p
recisi
on
outc
om
e
s.
3.5
Cla
s
sific
ati
on
The
cl
assifi
cat
ion
of
im
age
is
pr
ese
nted
as
the
m
a
in
aspect
of
this
pr
e
sent
stud
y
with
res
pect
to
the
obj
ect
ives
of
t
his
stu
dy.
Fou
r
disti
nct
feat
ures
wer
e
init
ia
ll
y
extracte
d
f
ro
m
the
in
pu
t
i
m
age,
w
hich
wer
e
global
feat
ur
e,
local
featu
re,
pi
xe
l
featu
re,
a
nd
S
URF.
Fo
ll
owin
g
that,
thes
e
extracte
d
feat
ur
es
we
re
c
ombine
d
into
one
featu
r
e
vector.
PCA
was
subse
qu
e
nt
ly
per
form
ed
t
o
dec
rease
the
dim
ension
al
it
y
of
featu
re
vec
tors.
The
de
velo
pe
d
i
m
age
cl
assifi
cat
ion
syst
em
fr
om
this
stud
y
was
e
va
luate
d
u
si
ng
the
Im
ageCLEF2
005
database
[29]
.
This
data
base
was
segm
ented
into
trai
ning
s
et
and
te
sti
ng
s
et
.
The
trai
ning
set
was
cat
eg
or
iz
e
d
into
57 kn
own cl
asses,
wh
ic
h were
pre
-
def
i
ne
d.
4.
EVAL
UA
TI
O
N
A
series
of
e
xperim
ents
was
cond
ucted
to
e
valuate
the
pe
r
form
a
nce
of
the
pr
op
os
e
d
m
e
thod
in
this
stud
y.
I
n
pa
rtic
ular,
this
is
to
validat
e
the
propose
d
m
e
tho
d
and
it
s
sign
ifi
cance
for
X
-
ra
y
m
edical
i
m
a
ge.
Th
e
i
m
ple
m
entat
io
n
of
the
pro
po
sed
m
et
ho
d
in
cl
ud
e
d
featu
re
extracti
on,
feat
ur
e
com
bin
at
io
n
a
nd
re
duct
io
n,
a
nd
X
-
r
ay
m
edical
i
m
age
cl
assifi
c
at
ion
us
i
ng
S
MV
and
k
-
N
N
cl
assifi
ers,
w
hi
ch
wer
e
e
valu
at
ed
to
determ
i
ne
it
s
perform
ance
ba
sed
on
the
res
ults
of
acc
ur
ac
y
rate. A
s
a res
ult,
f
our
e
xperi
m
ents
wer
e
co
nducted
.
T
he
s
pecifi
c
m
et
ho
ds
a
nd
s
et
ti
ng
s
of
thes
e
exp
e
rim
ents
ar
e
desc
ribe
d
and
ob
ta
ine
d
re
su
lt
s
in
this
stud
y
are
pr
ese
nt
ed
an
d
discusse
d i
n
t
he
foll
ow
i
ng sub
-
sect
io
ns.
Essentia
ll
y,
th
e
Im
ageCLEF200
5
database
was
util
iz
ed
in
this
stu
dy
[29
]
,
wh
ic
h
c
onta
ined
10,
000
X
-
ray
i
m
ages,
div
ide
d
int
o
9,
000
trai
ni
ng
im
ages
and
1
,
000
te
sti
ng
im
a
ges.
T
hese
im
a
ges
we
re
in
gray
scal
e
with
dif
fer
e
nt
reso
l
utions,
w
hich
wer
e
ob
ta
ined
us
in
g
dif
f
eren
t
im
aging
te
chn
iq
ues
.
Th
ere
we
re
57
cl
asses
,
con
ta
ini
ng
dif
fer
e
nt
nu
m
ber
of
sam
ple
i
m
ages.
Durin
g
the
evaluati
on
sta
ge
,
the
trai
nin
g
da
ta
set
was
rand
om
l
y
par
ti
ti
on
ed
into
tw
o
set
s.
The
fir
st
dataset
was
div
ide
d
i
nto
80%
of
t
rainin
g
i
m
ages
an
d
20%
of
te
sti
ng
im
ages
wh
il
e
the
sec
ond
dataset
was
div
i
ded
into
90%
of
trai
ni
ng
i
m
ages
an
d
10%
of
te
sti
ng
i
m
ages.
This
ens
ur
e
d
that
each
cl
ass
con
ta
ine
d
t
rain
ing
im
ages
and
corres
pondin
g
te
st
i
m
ages.
The
first
datas
et
was
sel
ect
ed
to
c
om
par
e
with
th
e
obta
ined
res
ults
from
pr
e
vi
ou
s
relat
ed
st
ud
ie
s
w
hile
th
e
seco
nd
datas
et
was
sel
ect
ed
in
that
sp
eci
fic
rati
o
t
hat
it
was
si
m
i
la
r
to
the
Im
ageCLEF2
005
da
ta
base,
w
he
re
it
con
ta
ine
d
sim
il
ar
per
ce
ntages
of
bo
t
h
trai
ning
s
et
(90%
)
an
d
t
est
ing
set
(10
%).
T
he
num
ber
of
trai
ning
i
m
ages
an
d
te
st
i
m
ages
for
eac
h
cl
ass
in
these
t
wo
dat
aset
s
was
ta
ble
d
in
Ta
ble
3
(
80:2
0)
a
nd
Ta
ble
4
(
90:1
0)
(m
entione
d
i
n
col
um
ns
with
fine da
s
he
d
sty
le
).
It s
hould
be n
oted
t
ha
t t
he
se
qu
e
nce
r
efe
rs
t
o
the
cl
ass num
ber
.
Table
3
.
N
um
ber
of
im
ages, 8
0%
trai
ning a
nd
20% test
ing
Clas
s
No
.
Of
I
m
ag
es
80%
20%
Clas
s
No
.
Of
I
m
ag
es
80%
20%
Clas
s
No
.
Of
I
m
ag
es
80%
20%
1
336
269
67
20
31
25
6
39
38
38
8
2
32
26
6
21
194
155
39
40
51
41
10
3
215
172
43
22
48
38
10
41
65
52
13
4
102
82
20
23
79
63
16
42
74
59
15
5
225
180
45
24
17
14
3
43
98
78
20
6
576
461
115
25
284
230
57
44
193
154
39
7
77
62
15
26
170
136
34
45
35
28
7
8
48
38
10
27
109
87
22
46
30
24
6
9
69
55
14
28
228
182
46
47
147
118
29
10
32
26
6
29
86
69
17
48
79
63
16
11
108
86
22
30
59
47
12
49
78
62
16
12
2563
2050
513
31
60
48
12
50
91
73
18
13
93
74
19
32
78
62
16
51
9
7
2
14
152
122
30
33
62
50
12
52
9
7
2
15
15
12
3
34
880
70
4
176
53
15
12
3
16
23
18
5
35
18
14
4
54
46
37
9
17
217
174
43
36
94
75
19
55
10
8
2
18
205
164
41
37
22
18
4
56
15
12
3
19
137
110
27
38
116
93
23
57
57
46
11
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
10
, N
o.
1
,
A
pr
il
201
8
:
154
–
167
162
Table
4
.
N
um
ber
of
im
ages, 9
0%
trai
ning a
nd
10% test
ing
Clas
s
No
.
Of
I
m
ag
es
90%
10%
Cl
ass
No
.
Of
I
m
ag
es
90%
10%
Clas
s
No
.
Of
I
m
ag
es
90%
10%
1
336
302
34
20
31
28
3
39
38
34
4
2
32
29
3
21
194
175
19
40
51
46
5
3
215
193
22
22
48
43
5
41
65
58
7
4
102
92
10
23
79
71
8
42
74
67
7
5
225
202
23
24
17
15
2
43
98
88
10
6
576
518
58
25
284
2
56
28
44
193
174
19
7
77
69
8
26
170
153
17
45
35
31
4
8
48
43
5
27
109
98
11
46
30
27
3
9
69
62
7
28
228
205
23
47
147
132
15
10
32
29
3
29
86
77
9
48
79
71
8
11
108
97
11
30
59
53
6
49
78
70
8
12
2563
2307
256
31
60
54
6
50
91
82
9
13
93
84
9
32
7
8
70
8
51
9
8
1
14
152
137
15
33
62
56
6
52
9
8
1
15
15
13
2
34
880
792
88
53
15
13
2
16
23
21
2
35
18
16
2
54
46
41
5
17
217
195
22
36
94
85
9
55
10
9
1
18
205
184
21
37
22
20
2
56
15
13
2
19
137
123
14
38
116
104
12
57
57
51
6
4.1
E
xp
eri
me
nt
1
-
Fe
ature
Redu
c
tio
n
This
e
xperim
e
nt
was
c
onduc
te
d
to
eval
uate
an
d
i
nv
e
sti
gate
the
per
f
or
m
ance
of
the
ou
r
pr
op
os
e
d
syst
e
m
after
red
uci
ng
t
he
nu
m
ber
of
feat
ures
us
in
g
PC
A,
wh
ic
h
wa
s
ess
entia
l
to
deter
m
ine
the
accur
acy
rate
of
t
he
syst
em
.
Feat
ure
vect
ors
wer
e
re
du
c
ed
f
r
om
1,
785
featu
res
int
o
25
feat
ur
es
,
50
feat
ur
e
s,
a
nd
10
0
featur
e
s,
wh
ic
h
w
as
te
rm
ed
as
PC1,
PC2
a
nd
PC3
,
re
sp
e
ct
ively
.
The
re
su
lt
s
of
acc
ur
a
cy
rate
with
optim
al
featur
e
re
duct
ion
we
re
obta
ined
with
a
nd
without
a
thre
sh
ol
d
for
both
dataset
s
c
onta
ining
s
pecific
rati
o
of
trai
ning set a
nd test
in
g
set
(80:2
0; 90:1
0) a
nd
wer
e
subse
quently
co
m
pared am
on
g t
hes
e obtai
ned va
lu
es.
The
PC
A
is
c
onside
red
com
petent
an
d
ef
fec
ti
ve
in
re
duci
ng
the
dim
ension
al
it
y
of
data.
Both
S
V
M
(w
it
h
RB
F
kernel)
a
nd
k
-
N
N
(w
it
h
k
=
1)
w
ere
em
plo
ye
d
f
or
each
e
valuat
ion
sta
ge
i
n
t
his
ex
per
im
ent.
Table
5
re
veals
th
e
ob
ta
ine
d
res
ults
us
i
ng
k
-
N
N
cl
assifi
er
w
hile
Table
6
re
ve
al
s
the
obta
ine
d
re
su
lt
s
us
in
g
SV
M
cl
assifi
er.
C
onseq
uen
tl
y,
PC
2
(
50
feat
ur
es
)
a
chieve
d
t
he
highest
acc
ur
acy
r
at
e.
Spec
ific
al
ly
,
PC2
obta
ine
d
th
e
highest
per
ce
ntages, wit
h an
d wit
hout the
th
r
esh
old
us
in
g b
oth
classi
fiers.
Ba
sed on t
his e
xp
e
rim
ent, P
C2
was
consi
der
e
d
f
or
the
su
bse
que
nt
experim
ent
s.
It
shou
l
d
be
no
te
d
that
unde
niably,
it
is
essenti
al
t
o
ha
ve
su
f
fici
en
t nu
m
ber
of f
eat
ur
e
s f
or d
isc
rim
inatio
n
a
nd f
or h
ig
h
acc
ur
acy
r
at
e
.
Ha
ving f
e
w
f
eat
ur
es
m
igh
t
l
ead
to
low
accu
racy
rate
an
d
ina
de
qu
at
e
num
ber
of
feat
ur
es
s
ubse
qu
e
ntly
affe
ct
s
the
discri
m
inati
on
am
on
g
the
featur
e
s
of
ot
he
r
im
ages.
Ne
ve
rtheless,
hi
gh
accurac
y
rate
i
s
not
warrante
d
with
high
nu
m
ber
of
feat
ures
due
to
the
high
oc
currence
of
c
om
m
on
featu
res
,
w
hich
a
ff
ect
s
the
disc
rim
i
nation
am
on
g
the
featu
res
of
oth
e
r
i
m
ages
as
well
.
Con
se
quentl
y,
PC2
wa
s
pr
ov
e
n
to
ac
hiev
e
the
highest
accuracy
rate
,
rather
t
han
PC
1
(
25
featur
e
s)
a
nd P
C3 (1
00 f
eat
ures).
Table
5
.
Acc
uracy
r
esults
for PC
A by u
sin
g k
-
NN
80%
-
20%
W
ith
thresh
o
ld
80%
-
20%
W
ith
o
u
t thresh
o
ld
90%
-
10%
W
ith
thresh
o
ld
90%
-
10%
W
ith
o
u
t thresh
o
ld
PCA 1
6
0
.95
2
8
5
.90
1
8
4
.85
3
9
1
.62
0
PCA 2
6
1
.11
8
8
7
.01
1
8
7
.88
1
92
.13
8
PCA 3
6
0
.78
8
8
6
.50
3
8
6
.03
6
9
1
.73
1
Ta
bl
e
6
.
Ac
cur
a
c
y
re
sul
ts
for
PC
A b
y
using SVM
80%
-
20%
W
ith
thresh
o
ld
80%
-
20%
W
ith
o
u
t thresh
o
ld
90%
-
10%
W
ith
thresh
o
ld
90%
-
10%
W
ith
o
u
t thresh
o
ld
PCA 1
9
0
.04
7
9
1
.81
0
8
9
.89
9
9
4
.20
1
PCA 2
9
0
.45
0
99
.20
2
9
1
.92
4
9
5
.3
68
PCA 3
9
0
.29
0
99
.06
6
9
0
.76
1
9
5
.12
2
4.2
E
xp
eri
me
nt
2
-
Fe
ature
Co
m
bina
tio
n
This
ex
pe
rim
e
nt
aim
ed
to
inve
sti
gate
the
perform
ance
of
a
sing
le
feat
ur
e
e
xtracti
on
f
ro
m
each
of
the
four
featu
res
(
global
featu
re,
local
featur
e
,
pix
el
featu
re,
and
S
URF
)
a
n
d
t
he
com
bina
ti
on
of
these
four
featur
e
e
xtract
ion
s
.
T
he
res
ul
ta
nt
outc
om
e
of
this
e
xperi
m
ent
was
cr
uc
ia
l
in
te
rm
s
of
accu
racy
ra
te
an
d
ind
e
xing.
T
hes
e
resu
lt
s
wer
e
com
par
ed
with
tho
se
of
relat
e
d
pre
vious
stu
dies
in
te
rm
of
featu
re
set
s.
Most
of
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Multi
-
Level
of
Feature
Extr
ac
ti
on
and
Cl
as
si
fi
cation
f
or
X
-
Ray
Me
dical
Image
(
M.
M A
bdulr
az
z
aq
)
163
the
m
edica
l
c
on
te
nt
-
based
i
m
age
retrieval
syst
e
m
s
utiliz
e
global
featu
res.
T
he
m
ai
n
adv
a
ntage
of
global
featur
e
s
is
the
com
pu
ta
ti
on
s
peed
w
her
e
th
e
featur
e
e
xtra
ct
ion
an
d
m
at
c
hing
sim
il
arit
y
are
com
pu
ta
ti
on
al
ly
faster.
H
oweve
r,
they
m
ay
fai
l
to
identify
per
ti
nen
t
vi
s
ual
char
act
erist
ic
s.
The
cl
assifi
cat
ion
process
of
global
featur
e
s
incl
udes
tw
o
ph
a
ses
,
w
hich
are
tr
ai
nin
g
a
nd
te
s
ti
ng
.
I
n
the
tr
ai
nin
g
ph
ase
of
this
st
ud
y,
global
featur
e
s
we
re
extracte
d
fro
m
al
l
trai
nin
g
i
m
ages
and
t
he
cl
assifi
er
was
s
ub
se
que
ntly
trai
ne
d
on
these
extracte
d
featu
res
to
create
a
m
od
el
.
In
ord
e
r
to
cl
assify
th
e
te
st
i
m
ages,
featur
e
s
we
re
init
ia
ll
y
extract
ed
in
the sam
e w
ay
as in
t
he
trai
ning
ph
a
se. T
he
m
od
el
was
the
n uti
li
zed to
clas
sify t
est
i
m
ages.
Howe
ver,
loca
l
featur
es
a
re
i
nh
e
re
ntly
r
obust
against
tra
nsl
at
ion
.
I
n
this
exp
e
rim
ent,
local
featur
es
wer
e
e
xtracted
fr
om
fo
ur
s
quare
i
m
ages,
w
hich
w
ere
ta
ke
n
from
or
igina
l
on
es
afte
r
di
vid
in
g
the
im
a
ge
int
o
four
blo
c
ks
.
Si
m
il
ar
cl
assifi
c
at
ion
proces
s
that
was
ap
plied
f
or
gl
ob
al
fe
at
ur
es
wa
s
sub
seq
uen
tl
y
app
l
ie
d
for
local
featur
e
s,
excep
t
that
loc
al
featur
es
were
extracte
d
fro
m
each
su
b
-
im
age.
T
he
pi
xel
value
c
om
par
ison
is
al
so
an
e
ff
ect
i
ve
ap
proac
h
to
seek
sim
i
la
r
im
ages
in
the
da
ta
base.
F
or
m
os
t
ap
plica
ti
ons,
this
ap
proac
h
is
not
feasible
beca
u
se
the
diff
e
rence
between
th
e
pix
el
s
of
on
e
i
m
age
to
an
oth
e
r
is
no
t
ev
ident.
H
oweve
r,
it
is
feasible f
or
the
p
ixel value c
om
par
ison
to
id
entify
o
nly o
ne
sp
eci
fic obj
ect
o
f
eq
ual size
an
d
locat
e
d
at
sim
il
ar
po
sit
io
n
(sim
ilar
row
an
d
col
um
n
of
an
im
a
ge
m
a
tri
x)
bet
ween
im
ages
with
sm
all
reso
luti
ons.
F
or
t
his
stud
y,
Ex
per
im
ent 2
a
lso
util
iz
ed
pixe
l i
nfor
m
at
ion
.
The
S
URF,
a
de
scripto
r
feat
ure,
is
al
so
a
scale
and
r
otati
on
i
nv
a
riant
detect
or.
The
scal
e
and
r
otati
on
inv
a
riance
denotes
that
an
obj
ect
co
uld
be
ident
ifie
d
eve
n
wh
e
n
it
is
scal
ed
in
siz
e
or
r
ot
at
ed.
The
S
U
RF
was
app
li
ed
i
n
this
exp
e
rim
ent
as
well
,
but
it
was
no
t
util
iz
ed
a
s
on
e
of
the
l
oc
al
featur
es
giv
en
t
hat
the
ex
tract
ion
of
t
hese
featu
r
es
is
a
ti
m
e
-
con
su
m
ing
proce
ss.
I
n
t
he
trai
ni
ng
phase,
al
l
i
m
age
s
wer
e
r
esi
zed
to
100
x
100
pix
el
s,
w
her
e
the
res
ultant
l
arg
e
featu
re
ve
ct
or
c
on
ta
i
nin
g
1,7
85
feat
ures
was
re
duc
ed
to
25
f
eat
ures,
50
featur
e
s,
a
nd
100
featu
res
us
i
ng
PCA
.
Re
ferrin
g
to
the
obta
ined
resu
lt
of
Ex
per
im
ent
1,
PC2
(
50
feat
ures)
was
c
onside
red f
or
Ex
per
im
ent 2
.
Fo
r
t
he
ge
ner
a
ti
on
of
m
od
el
,
bo
t
h
SV
M
cl
as
sifie
r
an
d
k
-
N
N
cl
assifi
er
we
re
com
par
ed
.
The
S
VM
i
s
widely
use
d
f
or
sta
ti
sti
cal
lear
ni
ng
an
d
cl
a
ssific
at
ion
.
P
rim
aril
y,
the
S
VM
deals
wit
h
bin
a
ry
cl
assifi
cat
io
n
issues.
The
re
are
prese
ntly
two
m
ul
ti
ple
cl
assifi
cat
ion
appr
oach
es
in
us
e,
s
pecific
al
ly
on
e
-
a
gain
st
-
one
appr
oach
a
nd
on
e
-
a
gainst
-
al
l
appro
ac
h.
T
he
on
e
-
agai
ns
t
-
al
l
app
r
oach
was
sp
eci
fical
ly
con
side
red
f
or
this
exp
e
rim
ent
becau
se
it
is
c
om
pu
ta
ti
on
al
ly
faster
t
han
the
oth
e
r
a
ppr
oac
h.
Acc
ordin
g
ly
,
the
RB
F
kernel
wa
s
app
li
ed
with
g
=
0.062
5,
a
nd
a
trade
-
off
between
t
he
trai
ni
ng
e
rror
a
nd
m
arg
i
n,
c
=
8.
It
sh
oul
d
be
no
te
d
that
these
values
w
ere
ob
ta
ine
d
f
r
om
an
em
pirical
stud
y.
T
he
s
econd
m
os
t
widely
us
e
d
cl
as
sific
at
ion
m
et
ho
d
is
the
k
-
N
N
(
k
=
1),
w
hich
was
us
e
d
f
or
furthe
r
com
par
iso
ns
(d
et
ai
ls
on
the
par
am
et
ers
of
SV
M
a
nd
k
-
N
N
a
r
e
furthe
r
disc
us
s
ed
f
or
E
xp
e
ri
m
ent
4)
.
Re
sul
ts
wer
e
cal
cu
la
te
d
after
pe
r
form
ing
random
sa
m
pling
on
the
dataset
for 1
0
t
i
m
es in o
r
de
r
t
o pro
du
ce
r
el
ia
ble r
es
ults.
The
re
su
l
ts
s
hown
i
n
Table
7
and
Table
8
re
fer
to
t
he
co
rr
e
ct
ness
rate
of
di
ff
ere
nt
feat
ur
e
set
s
us
in
g
bo
t
h
SV
M
cl
a
ssifie
r
an
d
k
-
NN
cl
assifi
er
,
resp
ect
ively
.
It
cou
ld
be
obse
rv
e
d
in
Table
8
that
in
the
X
MIAR
prototype
,
the
com
bin
ed
f
eat
ur
es
of
al
l
f
our
featu
res
us
i
ng
the
S
VM
cl
as
sifie
r
achie
ve
d
the
hi
ghest
ac
cur
acy
rate
(95.3
68%)
by
ap
plyi
ng
th
e
second
set
of
evaluati
on
(
90
%
of
trai
ning
im
ages
an
d
10
%
of
te
sti
ng
im
ages)
without
a
thre
sh
ol
d.
The
c
om
bin
ed
featu
r
es
of
al
l
fe
at
ures
c
on
ta
i
ned
pix
el
i
nfor
m
ation
,
gl
obal
fe
at
ur
es
(f
eat
ures
of
s
ha
pe
an
d
te
xt
ure),
local
feat
ures
(
featur
e
s
of
sh
a
pe
an
d
te
xture),
a
nd
S
URF.
T
her
e
f
ore,
the
app
li
cat
io
n
of
the
SV
M
usi
ng
c
om
bin
ed
featur
e
s
outpe
rfor
m
ed
the
ot
her
a
pp
li
cat
io
ns
us
i
ng
each
of
the
featur
e
set
s
se
par
at
el
y
as
fo
l
lows
:
(
1)
gl
ob
al
featur
e
set
,
(2)
local
featu
re
set
,
(3)
pi
xe
l
value
set
,
and
(4)
SU
RF
set
.
T
he
com
par
iso
n
of
these
disti
nct
featur
e
s
al
s
o
re
vealed
that
t
he
us
e
of
pixe
l
fe
at
ur
es
outpe
rfo
rm
ed
the
us
es
of
bo
th
global
feat
ures
an
d
local
f
eat
ur
es
f
or
al
l
evaluati
on
set
s
wh
il
e
the
loc
al
featur
es
pro
vid
e
d
resu
lt
s
of h
i
gher acc
ur
acy
rate t
han the
gl
obal
f
eat
ures f
or
al
l evaluati
on s
et
s.
Table
7
.
Acc
uracy
r
esults
of e
xtracted
f
eat
ures b
y
us
in
g k
-
NN
80%
-
20%
with
thresh
o
ld
80%
-
20%
with
o
u
t thresh
o
ld
90%
-
10%
with
thresh
o
ld
90%
-
10%
with
o
u
t thresh
o
ld
SURF
5
4
.27
6
8
4
.53
2
8
0
.65
2
8
6
.82
3
Glo
b
al
6
1
.11
1
8
2
.75
1
8
6
.31
7
8
8
.31
3
Local
6
1
.56
4
8
5
.59
0
8
7
.59
3
8
8
.86
8
Pix
el
6
3
.88
7
8
7
.59
6
8
6
.67
9
9
1
.29
3
All f
eatu
res
6
9
.38
1
8
9
.32
7
8
7
.88
5
9
2
.71
2
Evaluation Warning : The document was created with Spire.PDF for Python.