I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
36
,
No
.
2
,
N
o
v
em
b
e
r
20
24
,
p
p
.
1
1
1
4
~1
126
I
SS
N:
2
502
-
4
7
52
,
DOI
: 1
0
.
1
1
5
9
1
/ijee
cs
.v
36.
i
2
.
p
p
1
1
1
4
-
1
1
2
6
1114
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs
.
ia
esco
r
e.
co
m
Dete
c
tion o
f
co
lo
r
iza
tion ba
sed ima
g
e f
o
rg
eries using
co
nv
o
lutiona
l auto
enco
der met
ho
d
So
um
y
a
s
hree
M
ura
lid
ha
r
P
a
ncha
l
1
,
Ash
a
K
et
ha
g
a
na
ha
lli
H
a
nu
m
a
ntha
ia
h
2
,
B
ind
us
hree
Cha
nn
a
ba
s
a
v
a
ra
j
u Do
dd
a
s
idd
a
v
a
na
ha
lli
3
,
M
a
nju
M
o
re
E
s
hwa
r
Ra
o
4
,
Am
bik
a
B
elek
er
e
J
a
y
a
r
a
m
u
1
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
M
a
n
i
p
a
l
I
n
st
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
B
e
n
g
a
l
u
r
u
,
M
a
n
i
p
a
l
A
c
a
d
e
m
y
o
f
H
i
g
h
e
r
E
d
u
c
a
t
i
o
n
,
M
a
n
i
p
a
l
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
D
o
n
B
o
sc
o
I
n
st
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
,
B
e
n
g
a
l
u
r
u
,
I
n
d
i
a
3
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
R
EV
A
U
n
i
v
e
r
si
t
y
,
B
e
n
g
a
l
u
r
u
,
I
n
d
i
a
4
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
P
ES U
n
i
v
e
r
s
i
t
y
,
B
e
n
g
a
l
u
r
u
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Mar
15
,
2
0
2
4
R
ev
is
ed
J
u
l
26
,
2
0
2
4
Acc
ep
ted
J
u
l
29
,
2
0
2
4
Re
c
e
n
tl
y
,
i
t
h
a
s
b
e
c
o
m
e
d
iffi
c
u
lt
to
re
c
o
g
n
ize
a
n
d
e
a
sie
r
t
o
m
isu
se
d
ig
i
tal
ima
g
e
s
d
u
e
to
t
h
e
larg
e
n
u
m
b
e
r
o
f
e
d
i
ti
n
g
t
o
o
ls
a
v
a
il
a
b
le.
De
tec
ti
n
g
fo
rg
e
ries
in
ima
g
e
s
is
c
ru
c
ial
fo
r
se
c
u
rit
y
a
n
d
f
o
re
n
sic
p
u
rp
o
se
s.
Th
e
re
fo
re
,
th
is
re
se
a
rc
h
imp
lem
e
n
ts
a
d
e
e
p
l
e
a
rn
i
n
g
(DL)
m
e
th
o
d
o
f
c
o
n
v
o
l
u
ti
o
n
a
l
a
u
to
e
n
c
o
d
e
r
(CAE)
wh
ic
h
imp
ro
v
e
s
c
o
lo
riza
ti
o
n
-
b
a
se
d
ima
g
e
fo
rg
e
r
y
d
e
tec
ti
o
n
b
y
lev
e
ra
g
in
g
sp
a
ti
a
l
a
n
d
c
o
l
o
r
in
f
o
rm
a
ti
o
n
,
in
c
re
a
sin
g
th
e
d
e
tec
ti
o
n
a
c
c
u
ra
c
y
.
At
first,
th
e
p
re
-
p
ro
c
e
ss
e
d
in
p
u
t
f
o
r
g
e
ry
ima
g
e
s
a
re
u
se
d
with
t
h
e
wie
n
e
r
fil
teri
n
g
-
c
o
n
tras
t
re
stricte
d
imp
ro
v
e
d
h
ist
o
g
ra
m
e
q
u
a
li
z
a
ti
o
n
(W
E
-
CLAHE)
tec
h
n
i
q
u
e
.
H
y
b
ri
d
d
u
a
l
-
tree
c
o
m
p
lex
wa
v
e
let
tri
g
o
n
o
m
e
tri
c
tran
sfo
rm
(H‑D
TCW
T)
a
n
d
VG
G
-
1
6
a
re
u
se
d
t
o
e
x
trac
t
e
ffe
c
ti
v
e
fe
a
tu
re
s
fro
m
th
e
c
l
u
ste
re
d
d
a
ta.
Im
p
r
o
v
e
d
h
o
rse
h
e
rd
o
p
ti
m
iza
ti
o
n
(IHH
)
is
e
m
p
lo
y
e
d
t
o
re
d
u
c
e
th
e
d
ime
n
si
o
n
a
li
ty
o
f
a
fe
a
tu
re
.
At
las
t,
t
h
e
C
AE
m
o
d
e
l
is i
m
p
lem
e
n
ted
to
si
g
n
ifi
c
a
n
tl
y
re
c
o
g
n
ize
th
e
ima
g
e
fo
rg
e
ry
.
Th
e
a
c
c
u
ra
c
y
o
f
CASIA
V1
a
n
d
G
RIP
d
a
tas
e
ts
o
f
9
9
.
9
5
%
a
n
d
9
9
.
9
7
%
,
re
sp
e
c
ti
v
e
ly
is
a
c
h
iev
e
d
.
He
n
c
e
,
th
is
imp
lem
e
n
t
e
d
m
e
th
o
d
o
b
tai
n
s
a
h
ig
h
fo
r
g
e
ry
d
e
tec
ti
o
n
p
e
rfo
rm
a
n
c
e
th
a
n
th
e
e
x
ist
in
g
m
e
th
o
d
s.
K
ey
w
o
r
d
s
:
B
lin
d
f
o
r
g
er
y
d
etec
tio
n
C
o
n
v
o
lu
tio
n
al
a
u
to
en
co
d
er
Dee
p
lear
n
in
g
Dig
ital im
ag
es
I
m
ag
e
f
o
r
g
er
y
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Am
b
ik
a
B
elek
er
e
J
ay
ar
am
u
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
E
n
g
in
ee
r
in
g
,
M
a
n
ip
al
I
n
s
titu
te
o
f
T
ec
h
n
o
lo
g
y
B
en
g
alu
r
u
Ma
n
ip
al
Aca
d
em
y
o
f
Hig
h
e
r
E
d
u
ca
tio
n
Man
ip
al,
5
7
6
1
0
4
,
Kar
n
ata
k
a,
I
n
d
ia
E
m
ail: a
m
b
ik
a.
b
j@
m
an
ip
al.
ed
u
1.
I
NT
RO
D
UCT
I
O
N
C
o
p
y
-
m
o
v
e
f
o
r
g
er
y
(
C
MF)
is
a
p
r
ev
alen
t
tech
n
iq
u
e
em
p
l
o
y
ed
in
im
a
g
e
m
an
ip
u
latio
n
,
aim
ed
at
d
ec
eiv
in
g
v
iewe
r
s
.
I
ts
p
r
im
ar
y
o
b
jectiv
e
is
to
eith
er
d
r
aw
atten
tio
n
to
ce
r
tain
elem
en
ts
o
r
co
n
ce
al
s
p
ec
i
f
ic
in
f
o
r
m
atio
n
with
in
an
im
ag
e
[
1
]
.
As
th
e
r
ep
ea
ted
s
ec
tio
n
s
ar
e
tak
en
f
r
o
m
th
e
s
am
e
im
ag
e,
th
ey
s
h
ar
e
m
an
y
attr
ib
u
tes
s
u
ch
as
c
o
lo
r
p
atter
n
s
,
n
o
is
e
lev
els
a
n
d
d
y
n
a
m
ic
r
an
g
es.
B
ec
au
s
e
th
ese
r
ep
ea
ti
n
g
p
o
r
tio
n
s
b
en
d
-
in
with
th
e
r
est
o
f
th
e
im
ag
e,
it
b
ec
o
m
es
ch
allen
g
in
g
f
o
r
h
u
m
a
n
s
to
d
if
f
er
en
t
iate
b
etwe
en
t
h
e
d
u
p
licate
p
ar
ts
an
d
th
e
o
r
ig
in
al
im
ag
e
[
2
]
,
[
3
]
.
Sp
licin
g
,
r
eto
u
ch
in
g
an
d
co
p
y
-
m
o
v
e
ar
e
th
e
m
o
s
t
p
o
p
u
lar
m
et
h
o
d
s
u
s
ed
in
im
ag
e
f
o
r
g
er
ies.
A
s
ec
tio
n
o
f
an
im
ag
e
is
p
asted
an
d
co
p
ied
to
a
n
ew
p
o
s
itio
n
in
th
e
C
M
F [
4
]
.
T
h
is
ap
p
r
o
ac
h
r
ep
ea
ts
so
m
e
elem
en
ts
th
at
ar
e
clea
r
f
r
o
m
th
e
im
a
g
e
b
y
u
s
in
g
n
o
is
e,
co
lo
r
,
co
n
tr
ast
o
r
o
t
h
er
f
ac
to
r
s
.
As
a
co
n
s
eq
u
en
ce
,
it
b
ec
o
m
es
in
cr
ea
s
in
g
ly
d
if
f
icu
lt
to
d
etec
t
o
th
er
ty
p
es
o
f
f
o
r
g
er
ies
s
u
ch
as
s
p
licin
g
an
d
r
eto
u
ch
in
g
[
5
]
,
[
6
]
.
I
t
is
h
e
n
ce
n
ec
ess
ar
y
to
a
p
p
ly
ce
r
t
a
in
p
r
o
ce
s
s
in
g
tech
n
iq
u
es
lik
e
r
o
tatio
n
,
s
ca
lin
g
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
Dete
ctio
n
o
f c
o
lo
r
iz
a
tio
n
b
a
s
ed
ima
g
e
fo
r
g
eries
u
s
in
g
co
n
v
o
lu
tio
n
a
l
… (
S
o
u
mya
s
h
r
ee
Mu
r
a
lid
h
a
r
P
a
n
ch
a
l
)
1115
d
o
wn
s
am
p
lin
g
,
J
PEG
co
m
p
r
e
s
s
io
n
an
d
n
o
is
e
ad
d
itio
n
t
o
i
n
cr
ea
s
e
tr
u
s
t
in
co
p
y
-
m
o
v
e
a
lter
ed
im
ag
es.
T
h
is
wo
r
k
f
o
c
u
s
es
o
n
co
p
y
-
m
o
v
e
f
o
r
g
er
y
d
etec
tio
n
(
C
MFD
)
m
eth
o
d
s
as
im
ag
e
C
MF
d
etec
ti
o
n
is
a
ch
allen
g
in
g
is
s
u
e
[7
]
,
[
8
]
.
Sin
ce
J
PEG
f
o
r
m
at
is
u
s
ed
b
y
a
m
ajo
r
ity
o
f
d
ig
ital
ca
m
er
as
an
d
i
m
ag
e
p
r
o
ce
s
s
in
g
s
o
f
twar
e
to
en
co
d
e
d
ig
ital
im
ag
es,
it
is
cr
u
cial
to
d
etec
t
th
e
c
o
m
p
r
es
s
in
g
h
is
to
r
y
o
f
J
PEG
im
ag
es
.
Ad
d
itio
n
ally
,
th
e
co
m
p
r
ess
io
n
h
is
to
r
y
o
f
an
im
ag
e
r
ev
ea
ls
p
o
ten
tial
ch
a
n
g
es
in
a
p
ar
ticu
lar
J
PEG
im
ag
e
[
9
]
,
[
1
0
]
.
C
o
p
y
-
p
aste
an
d
s
p
licin
g
/im
ag
e
co
m
p
o
s
ite
ar
e
th
e
two
f
o
r
g
er
ies
th
at
t
y
p
ically
af
f
ec
t
d
ig
ital
im
ag
er
y
.
T
o
id
e
n
tify
c
o
p
y
-
p
aste
f
o
r
g
er
ies,
ex
is
tin
g
ac
ad
em
ics
p
r
o
v
id
e
m
an
y
tech
n
iq
u
es
th
at
all
f
o
llo
w
an
ea
s
y
p
r
o
ce
d
u
r
e
o
f
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
f
ea
tu
r
e
v
ec
to
r
co
m
p
ar
is
o
n
.
H
o
wev
er
,
im
a
g
e
s
p
licin
g
p
r
o
v
i
d
es
s
o
m
e
in
h
er
en
t
alliter
atio
n
s
to
th
e
im
ag
es th
at
in
cr
ea
s
e
v
ar
iat
io
n
s
alr
ea
d
y
p
r
esen
t.
An
y
f
o
r
g
er
y
d
etec
tio
n
m
et
h
o
d
’
s
m
ain
o
b
jectiv
e
is
to
id
e
n
tify
th
ese
in
co
n
s
is
ten
cies
in
th
e
im
ag
es
[
1
1
]
.
I
t is ea
s
y
to
p
r
o
d
u
ce
f
ak
e
im
ag
es with
o
u
t le
av
in
g
a
r
ac
e
with
h
elp
o
f
a
n
im
ag
e
e
d
itin
g
s
o
f
twar
e,
wh
ich
is
av
ailab
le
f
o
r
f
r
ee
in
th
e
m
ar
k
et.
T
h
ese
elem
en
ts
h
av
e
in
cr
ea
s
ed
th
e
tr
u
s
t
is
s
u
e
r
eg
ar
d
in
g
th
e
r
eliab
ilit
y
o
f
d
ig
ital
im
ag
es
in
r
e
ce
n
t
y
ea
r
s
.
T
h
er
ef
o
r
e,
th
er
e
is
a
co
n
s
t
an
tly
in
cr
ea
s
in
g
n
ee
d
f
o
r
ef
f
i
cien
t
d
ig
ital
im
ag
e
f
o
r
g
er
y
d
etec
tio
n
m
eth
o
d
s
[
1
2
]
,
[
1
3
]
.
T
h
e
p
r
o
ce
s
s
o
f
p
r
o
d
u
cin
g
C
MT
I
F
in
clu
d
es
p
asti
n
g
an
d
c
o
p
y
i
n
g
a
s
ec
tio
n
o
f
th
e
im
ag
e
a
n
d
th
e
n
p
o
s
t
-
p
r
o
ce
s
s
in
g
th
e
r
esu
ltan
t
im
ag
e
[
1
4
]
.
T
h
e
ess
en
tial
p
r
o
p
er
ties
in
clu
d
i
n
g
illu
m
in
atio
n
b
eh
a
v
io
r
,
n
o
is
e
in
cu
r
r
e
d
,
an
d
c
o
lo
r
s
tate
m
ar
e,
ar
e
ess
en
tially
m
ain
tain
ed
in
b
o
th
th
e
s
o
u
r
ce
an
d
tam
p
er
ed
im
ag
es
[
1
5
]
.
Du
e
to
a
lack
o
f
d
ig
ital
m
ed
ia,
m
alici
o
u
s
im
ag
e
-
r
aisi
n
g
h
as
s
er
io
u
s
n
eg
ativ
e
ef
f
ec
t
s
o
n
m
ilit
ar
y
,
p
o
liti
cs,
ac
ad
em
ics
an
d
th
e
r
ea
l
wo
r
ld
.
T
h
u
s
,
th
e
n
ee
d
f
o
r
a
n
ef
f
ec
tiv
e
m
et
h
o
d
o
f
id
en
tify
in
g
im
ag
e
tam
p
er
in
g
a
n
d
f
r
a
u
d
is
cr
u
cial
[
1
6
]
,
[
1
7
]
.
Ho
wev
er
,
th
er
e
is
a
m
ajo
r
d
if
f
e
r
en
ce
in
l
o
w
-
lev
e
l
s
em
an
tic
n
o
tio
n
s
,
b
ec
au
s
e
m
o
s
t
C
MF
d
etec
tio
n
s
y
s
tem
s
p
r
im
ar
ily
r
ely
o
n
s
ca
le
-
in
v
ar
ian
t
f
ea
tu
r
e
tr
a
n
s
f
o
r
m
(
SIFT
)
,
a
lo
w
-
lev
el
v
is
u
al
r
ep
r
esen
tatio
n
o
f
d
ig
it
al
im
ag
es
[
1
8
]
.
Mo
d
er
n
tech
n
iq
u
es
lik
e
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN
)
,
m
o
b
ile
n
et
an
d
R
esNet5
0
v
2
ar
e
tr
ain
ed
o
n
lar
g
e
d
atasets
,
an
d
au
to
m
atica
lly
e
x
tr
ac
t
th
e
p
o
s
s
ib
le
f
ea
tu
r
es
a
c
c
o
r
d
i
n
g
t
o
t
h
e
d
e
v
el
o
p
m
e
n
t
o
f
d
e
e
p
l
e
a
r
n
i
n
g
(
D
L
)
[
1
9
]
.
Dee
p
f
e
a
t
u
r
e
s
a
r
e
u
s
e
d
f
o
r
p
e
r
s
o
n
i
d
e
n
t
i
f
i
ca
t
i
o
n
,
s
k
i
n
l
e
s
i
o
n
cl
a
s
s
i
f
ic
a
t
i
o
n
a
n
d
i
m
a
g
e
q
u
a
l
i
t
y
a
s
s
es
s
m
e
n
t
,
o
r
a
f
e
w
i
n
s
t
a
n
c
es
o
f
C
N
N
-
b
as
e
d
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
s
[
2
0
]
.
W
alia
et
a
l.
[
2
1
]
im
p
lem
e
n
ted
s
ca
le
an
d
d
ir
ec
tio
n
lo
ca
l
b
in
a
r
y
p
atter
n
(
SD
-
L
B
P),
u
tili
ze
d
f
o
r
d
ig
ital
im
ag
e
m
an
ip
u
latio
n
d
etec
tio
n
.
T
h
e
im
p
lem
en
ted
m
eth
o
d
w
as
em
p
lo
y
ed
to
lear
n
d
ee
p
an
d
co
m
p
lex
f
ea
tu
r
es
f
r
o
m
p
r
e
-
p
r
o
ce
s
s
ed
im
ag
es
f
o
r
class
if
icatio
n
in
to
au
t
h
e
n
tic
an
d
f
o
r
g
ed
im
ag
es.
I
MD
2
0
2
0
,
C
ASI
A
v
1
,
DVM
M,
an
d
C
ASI
A
v
2
wer
e
f
o
u
r
d
atasets
em
p
lo
y
ed
to
ev
alu
ate
th
is
m
eth
o
d
th
at
ac
h
iev
ed
h
ig
h
d
etec
tio
n
ac
cu
r
ac
y
u
s
in
g
R
esNet5
0
.
Du
e
to
less
im
ag
e
f
o
r
g
e
r
y
lo
c
aliza
tio
n
,
th
is
m
eth
o
d
r
eq
u
ir
ed
an
ex
ten
s
io
n
in
lo
ca
lize
d
f
o
r
g
er
y
th
r
o
u
g
h
p
r
o
v
id
in
g
in
p
u
t
to
th
e
n
eu
r
al
n
e
two
r
k
.
Sriv
astav
a
an
d
Yad
a
v
[
2
2
]
im
p
lem
en
ted
m
u
ltip
le
lo
ca
l
b
in
ar
y
p
atter
n
s
(
L
B
P)
u
tili
ze
d
f
o
r
th
e
d
etec
tio
n
o
f
b
o
th
s
p
licin
g
an
d
C
MF
.
An
R
GB
im
ag
e
was
co
n
v
er
ted
in
to
a
C
b
an
d
C
r
im
ag
e,
wh
ile
YC
b
C
r
im
ag
e
co
m
p
o
n
en
ts
wer
e
ex
tr
ac
ted
i
n
th
is
im
p
lem
en
ted
m
u
ltip
le
L
B
P
m
eth
o
d
.
C
ASI
A
v
2
.
0
,
C
o
lu
m
b
ia
an
d
C
ASI
A
v
1
.
0
wer
e
th
r
ee
d
atasets
u
tili
ze
d
f
o
r
ef
f
ec
tiv
ely
ev
alu
atin
g
tam
p
er
in
g
d
etec
tio
n
,
em
p
l
o
y
ed
with
SVM
to
p
r
o
v
id
e
g
o
o
d
r
esu
lts
in
b
o
th
s
m
al
l
an
d
lar
g
e
d
atasets
.
Du
e
to
th
eir
im
p
ac
t
o
n
lo
ca
liz
in
g
f
o
r
g
ed
p
ar
ts
in
s
p
liced
im
ag
es,
th
is
m
eth
o
d
n
ee
d
e
d
to
c
o
m
b
in
e
DL
an
d
ML
m
eth
o
d
s
,
in
o
r
d
er
to
ef
f
icien
tl
y
lo
ca
lize
f
o
r
g
ed
p
ar
ts
in
s
p
liced
im
ag
es.
R
ao
et
a
l.
[
2
3
]
im
p
lem
en
te
d
a
k
n
o
wled
g
e
-
b
ased
f
u
zz
y
ap
p
r
o
x
im
atio
n
(
KB
FA)
an
d
h
y
b
r
id
g
r
ey
wo
lf
an
t
lio
n
o
p
tim
izatio
n
(H
-
GW
AL
)
(
KB
FA
with
H
-
G
W
AL
)
ap
p
r
o
ac
h
.
H
-
GW
AL
was
im
p
lem
en
ted
to
d
etec
t
s
p
liced
im
ag
es,
wh
ile
lo
ca
liz
ed
tam
p
e
r
ed
r
eg
io
n
s
p
liced
im
ag
es
wer
e
u
tili
ze
d
b
y
th
e
KB
FA
m
o
d
el.
T
h
e
KB
FA
alg
o
r
ith
m
was
u
s
ed
to
ef
f
ec
tiv
ely
d
etec
t
an
d
ca
teg
o
r
ize
im
ag
es,
wh
e
r
ea
s
th
e
d
etec
tio
n
ac
cu
r
ac
y
p
er
f
o
r
m
an
ce
was
im
p
r
o
v
ed
b
y
th
e
H
-
GW
AL
m
eth
o
d
.
H
o
wev
er
,
th
e
i
m
p
lem
en
ted
m
o
d
el
o
n
l
y
lo
ca
lized
s
p
liced
im
ag
es
an
d
d
i
d
n
o
t
p
r
o
p
er
l
y
class
if
y
th
e
im
ag
es.
Su
s
h
ir
et
a
l.
[
2
4
]
im
p
lem
e
n
ted
a
h
y
b
r
id
d
ee
p
co
n
v
o
l
u
tio
n
al
ca
p
s
u
le
au
to
e
n
co
d
er
(
h
y
b
r
i
d
DC
C
AE
)
,
wh
ich
was
u
tili
ze
d
f
o
r
s
ig
n
if
ica
n
t
im
ag
er
f
o
r
g
er
y
r
ec
o
g
n
itio
n
.
A
wien
er
f
ilter
in
g
-
co
n
tr
ast
lim
ited
in
cr
e
ased
h
i
s
to
g
r
am
eq
u
aliza
tio
n
was
u
s
ed
to
ef
f
ec
tiv
el
y
p
r
e
-
p
r
o
ce
s
s
th
e
im
a
g
e.
T
h
e
im
p
lem
en
ted
m
et
h
o
d
attain
ed
h
i
g
h
p
er
f
o
r
m
an
ce
o
f
th
e
s
y
s
tem
,
b
ec
a
u
s
e
o
f
th
e
ef
f
icien
t
r
eliab
ilit
y
an
d
tr
ain
in
g
m
o
d
el
ca
p
ab
ilit
y
,
th
e
r
eb
y
ef
f
ec
tiv
ely
d
etec
tin
g
b
lin
d
im
ag
e
f
o
r
g
er
ies.
No
n
eth
eless
,
th
e
im
p
lem
en
te
d
h
y
b
r
id
DC
C
AE
m
eth
o
d
h
a
d
lo
wer
f
o
r
g
er
y
d
etec
tio
n
a
n
d
im
ag
e
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
.
Diwan
et
a
l.
[
2
5
]
im
p
lem
en
ted
th
e
Su
p
er
p
o
i
n
t
ap
p
r
o
ac
h
f
o
r
d
etec
tin
g
C
M
F
in
d
ig
ital
im
ag
es,
u
s
ed
as
a
k
ey
p
o
in
t
d
etec
to
r
in
s
elf
-
s
u
p
er
v
is
ed
im
ag
es.
GR
I
P,
C
M
FD,
MI
C
C
-
F2
2
0
,
MI
C
C
-
F2
0
0
0
,
MI
C
-
F2
2
0
,
C
o
Mo
Fo
D,
C
ASI
A
V2
.
0
,
an
d
C
OVE
R
AGE
wer
e
th
e
eig
h
t d
atasets
u
tili
ze
d
to
d
etec
t c
o
p
y
-
m
o
v
e
im
a
g
e
f
o
r
g
er
y
.
T
h
e
im
p
lem
en
ted
m
et
h
o
d
p
r
o
d
u
ce
d
h
i
g
h
r
esu
lts
in
i
m
ag
es
with
d
if
f
er
en
t
attac
k
s
,
an
d
also
ef
f
ec
tiv
ely
d
etec
ted
C
MF
in
a
d
iv
er
s
e
r
an
g
e
o
f
f
o
r
g
e
d
im
a
g
es.
Du
e
to
less
ef
f
ec
tiv
en
ess
in
f
o
r
g
er
y
d
etec
tio
n
,
th
is
Su
p
er
p
o
i
n
t
ap
p
r
o
ac
h
f
ac
ed
co
m
p
u
tatio
n
al
co
m
p
le
x
ity
,
alo
n
g
with
ch
allen
g
es
in
r
ea
l
-
tim
e
p
r
o
ce
s
s
in
g
o
r
lar
g
e
-
s
ca
le
ap
p
licatio
n
s
.
B
ib
i
et
a
l.
[
2
6
]
p
r
esen
ted
m
u
ltip
le
s
tr
u
ctu
r
es
s
tack
ed
au
to
en
co
d
er
s
(
SAE)
m
eth
o
d
,
wh
ich
was
b
ased
o
n
C
NN
u
tili
ze
d
f
o
r
f
o
r
g
er
y
d
etec
tio
n
.
T
h
e
SAE
b
ased
C
NN
m
eth
o
d
was
em
p
lo
y
ed
with
m
u
ltip
le
s
tr
u
ctu
r
es
wh
ich
co
n
tain
s
two
an
d
th
r
e
e
s
tack
ed
au
to
e
n
co
d
e
r
s
f
o
r
d
ec
r
ea
s
in
g
th
e
f
ea
tu
r
e
d
im
en
s
io
n
s
.
T
wo
C
ASI
A
d
atasets
wer
e
u
tili
ze
d
to
ev
al
u
ate
th
e
p
r
esen
ted
m
eth
o
d
,
w
h
ich
p
r
o
v
id
e
d
h
ig
h
f
o
r
g
er
y
d
e
tectio
n
ac
cu
r
ac
y
o
n
J
PEG
im
ag
e.
No
n
eth
eless
,
th
e
p
r
esen
ted
m
eth
o
d
h
a
d
o
v
er
f
itti
n
g
is
s
u
e,
d
u
e
to
th
e
co
m
p
lex
ar
ch
itectu
r
e
o
f
SAE.
Alh
aid
er
y
et
a
l.
[
2
7
]
i
m
p
lem
en
ted
r
eg
io
n
g
r
o
win
g
-
m
er
g
in
g
s
eg
m
en
tatio
n
(
R
GM
S)
ap
p
r
o
ac
h
,
wh
ich
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
36
,
No
.
2
,
No
v
em
b
er
20
24
:
1
1
1
4
-
1
1
2
6
1116
was
u
tili
ze
d
to
d
etec
t
C
MF
in
d
ig
ital
im
ag
es.
T
h
e
im
p
lem
en
ted
R
GM
S
u
tili
ze
d
am
o
n
g
b
lo
ck
-
b
ased
an
d
k
ey
p
o
i
n
t
-
b
ased
m
eth
o
d
s
ef
f
ec
tiv
ely
lo
ca
lized
th
e
d
u
p
licated
r
eg
io
n
s
.
I
t
ad
d
itio
n
ally
im
p
r
o
v
ed
th
e
d
etec
tio
n
an
d
r
eg
io
n
o
f
in
ter
est
ass
is
t
ed
in
en
h
an
ci
n
g
lo
ca
lizatio
n
f
o
r
s
m
all
an
d
lar
g
e
r
e
g
io
n
s
.
Yet
,
th
is
m
eth
o
d
p
o
s
s
ess
ed
lo
w
ca
p
ab
ilit
y
to
r
ec
o
g
n
ize
th
e
d
if
f
er
e
n
ce
b
etwe
en
th
e
au
th
en
tic
an
d
f
o
r
g
ed
r
eg
io
n
s
.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
s
o
f
th
is
wo
r
k
ar
e
g
iv
en
as f
o
llo
ws:
−
A
DL
-
b
ased
f
ea
tu
r
e
ex
t
r
ac
tio
n
ap
p
r
o
ac
h
is
u
s
ed
to
ac
cu
r
atel
y
d
etec
t
f
o
r
g
er
y
im
a
g
es,
alo
n
g
s
id
e
r
ed
u
cin
g
r
ed
u
n
d
an
t
d
ata
a
n
d
p
r
o
v
id
in
g
v
alu
a
b
le
f
ea
tu
r
es
,
e
v
en
in
ar
ea
s
with
lo
wer
c
o
n
tr
ast
a
n
d
h
u
g
e
-
s
ized
im
ag
es a
lo
n
g
VGGN
et
an
d
D
T
T
.
−
T
h
is
s
tu
d
y
p
r
esen
ts
a
tech
n
iq
u
e
f
o
r
p
r
e
-
p
r
o
ce
s
s
in
g
im
ag
es u
s
in
g
a
W
E
-
C
L
AHE
wh
ich
en
h
a
n
ce
s
co
n
tr
ast
an
d
n
o
is
e
,
a
n
d
im
p
r
o
v
es sy
s
tem
ef
f
icien
cy
t
h
r
o
u
g
h
f
ea
tu
r
e
e
x
tr
ac
tio
n
.
−
C
o
n
v
o
lu
tio
n
al
au
to
e
n
co
d
e
r
(
C
AE
)
is
a
DL
-
b
ased
clas
s
if
icatio
n
f
r
am
ewo
r
k
th
at
ac
cu
r
ately
d
etec
ts
f
o
r
g
er
y
,
wh
ile
en
h
an
cin
g
class
if
icatio
n
ac
cu
r
ac
y
d
u
e
t
o
its
en
h
an
ce
d
ca
p
ab
ilit
y
.
T
h
e
r
est
o
f
th
e
p
ap
er
is
s
tr
u
ctu
r
ed
f
o
llo
ws:
s
ec
tio
n
2
d
escr
i
b
es
th
e
p
r
o
p
o
s
ed
m
eth
o
d
.
Sec
tio
n
3
ex
p
lain
s
th
e
p
r
o
ce
s
s
o
f
C
AE
-
C
NN
,
wh
ile
s
ec
tio
n
4
s
h
o
ws
th
e
r
esu
lts
an
d
d
is
cu
s
s
io
n
,
an
d
co
n
clu
s
io
n
o
f
th
is
wo
r
k
is
g
iv
en
in
s
ec
tio
n
5.
2.
P
RO
P
O
SE
D
M
E
T
H
O
D
I
n
th
is
wo
r
k
,
a
C
AE
m
o
d
el
i
s
im
p
lem
en
ted
f
o
r
c
o
lo
r
izatio
n
-
b
ased
im
ag
e
f
o
r
g
e
r
y
d
etec
t
io
n
.
T
h
is
wo
r
k
in
cl
u
d
es
C
ASI
A
V1
a
n
d
GR
I
P
d
atasets
f
o
r
c
o
llectin
g
d
ata,
an
d
W
E
-
C
L
AHE
u
s
ed
in
th
e
p
r
e
-
p
r
o
ce
s
s
in
g
p
h
ase
to
ef
f
ec
tiv
ely
elim
in
ate
n
o
is
e
f
r
o
m
th
e
in
p
u
t
im
ag
e.
H
‑
(
DT
C
W
T
–
DC
T
)
an
d
VGG
-
1
6
m
eth
o
d
s
ar
e
u
s
ed
to
ex
tr
ac
t
th
e
o
p
tim
al
f
ea
tu
r
es
an
d
th
en
,
I
HH
is
em
p
lo
y
ed
to
r
ed
u
ce
t
h
e
f
ea
tu
r
es'
d
im
en
s
io
n
ality
.
T
h
e
im
p
lem
en
ted
C
AE
m
o
d
el
is
u
tili
ze
d
to
s
ig
n
if
ican
tly
r
e
co
g
n
ize
im
a
g
e
f
o
r
g
er
y
.
Fig
u
r
e
1
d
e
n
o
tes
th
e
im
p
lem
en
ted
m
eth
o
d
’
s
b
l
o
ck
d
iag
r
am
.
Fig
u
r
e
1
.
I
m
p
lem
en
te
d
m
o
d
el’
s
b
lo
ck
d
iag
r
am
2
.
1
.
Da
t
a
s
et
s
T
h
e
wo
r
k
u
tili
ze
s
two
d
atasets
in
clu
d
in
g
C
ASI
A
V1
an
d
GR
I
P
to
ev
alu
ate
th
e
im
p
lem
en
te
d
m
eth
o
d
.
C
ASI
A
V1
[
2
8
]
an
d
GR
I
P
[
2
9
]
d
atasets
ar
e
em
p
lo
y
ed
t
o
co
llect
th
e
in
f
o
r
m
atio
n
.
T
h
ese
t
wo
d
atasets
ar
e
u
s
e
d
to
d
etec
t c
o
lo
r
izatio
n
-
b
ased
i
m
ag
e
f
o
r
g
er
ie
s.
2
.
1
.
1
.
CASI
A
V1
da
t
a
s
et
T
h
is
C
ASI
A
V1
d
ataset
co
n
s
is
ts
o
f
b
o
th
f
a
k
e
an
d
p
r
is
tin
e
im
ag
es,
wh
e
r
e
th
e
r
e
a
r
e
9
2
1
tam
p
er
e
d
im
ag
es
an
d
8
0
0
au
th
en
tic
im
ag
es.
T
h
is
d
ataset
h
as
1
,
7
2
1
J
PEF
co
m
p
r
ess
ed
im
ag
es
in
t
o
tal,
o
f
th
e
s
ize
o
f
384
×
2
5
6
.
T
h
u
s
,
t
h
e
9
2
1
tam
p
er
ed
im
ag
es
a
r
e
d
i
v
id
ed
i
n
to
4
6
9
s
p
liced
an
d
4
5
2
c
o
p
y
-
m
o
v
e
k
in
d
o
f
tam
p
e
r
ed
im
ag
es.
Ad
o
b
e
Ph
o
to
s
h
o
p
s
o
f
twar
e
is
u
s
ed
to
m
ak
e
s
p
lice
i
m
ag
es,
wh
ich
ar
e
s
ec
tio
n
s
o
f
o
n
e
im
ag
e
th
at
ar
e
co
n
n
ec
ted
t
o
o
th
er
s
.
Fig
u
r
e
2
s
h
o
ws th
e
s
am
p
le
im
ag
es o
f
th
e
C
ASI
A
V1
d
ataset.
Fig
u
r
e
2
.
Sam
p
le
im
a
g
es o
f
th
e
C
ASI
A
V1
d
ataset
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
Dete
ctio
n
o
f c
o
lo
r
iz
a
tio
n
b
a
s
ed
ima
g
e
fo
r
g
eries
u
s
in
g
co
n
v
o
lu
tio
n
a
l
… (
S
o
u
mya
s
h
r
ee
Mu
r
a
lid
h
a
r
P
a
n
ch
a
l
)
1117
2
.
1
.
2
.
G
RIP
da
t
a
s
et
T
h
is
GR
I
P
d
ataset
was
cr
ea
t
ed
b
y
C
o
zz
o
lin
o
,
co
n
tain
i
n
g
8
0
in
p
u
t
im
ag
es
an
d
8
0
p
la
in
f
o
r
g
e
r
y
im
ag
es,
wh
ich
ar
e
ac
ce
s
s
ib
le
in
1
0
2
4
×
7
6
8
in
th
e
f
o
r
m
at
o
f
PNG.
I
n
th
e
f
o
r
m
at
o
f
PNG,
th
e
tam
p
er
ed
p
o
r
tio
n
s
ar
e
s
ep
ar
ately
d
ep
o
s
ited
.
T
h
e
tam
p
er
ed
p
o
r
tio
n
’
s
lo
ca
tio
n
is
s
to
r
ed
in
th
e
f
o
r
m
at
o
f
tex
t in
th
e
f
o
r
g
er
y
im
ag
es.
T
h
is
d
ataset
is
d
iv
id
ed
in
to
3
p
ar
ts
with
5
0
r
ea
l
im
ag
es
wi
th
o
u
t
f
o
r
g
er
y
im
a
g
es,
5
0
tam
p
er
ed
im
ag
es
af
ter
tr
an
s
latio
n
,
an
d
s
ca
lin
g
2
0
s
ets o
f
an
im
a
g
es a
f
ter
r
o
tatio
n
.
Fig
u
r
e
3
r
ep
r
esen
ts
th
e
s
am
p
le
i
m
ag
es o
f
th
e
GR
I
P
d
ataset.
Fig
u
r
e
3
.
Sam
p
le
im
a
g
es o
f
G
R
I
P d
ataset
2
.
2
.
P
re
-
pro
ce
s
s
ing
Fo
llo
win
g
th
e
co
llectio
n
o
f
d
ata,
p
r
e
-
p
r
o
c
ess
in
g
is
p
er
f
o
r
m
ed
u
s
in
g
th
e
wien
er
f
ilter
‑
C
L
AHE
.
T
h
e
in
co
m
e
im
ag
e
s
ize
o
f
b
o
th
d
atasets
co
n
tain
s
7
6
8
×
1
0
2
4
×
3
f
o
r
th
e
GR
I
P
d
ataset
an
d
2
5
6
×
3
8
4
×
3
f
o
r
th
e
C
ASI
A
V1
d
ataset.
An
im
p
o
r
tan
t
aim
o
f
p
r
e
-
p
r
o
ce
s
s
in
g
im
ag
e
is
to
in
cr
ea
s
e
th
e
q
u
ality
o
f
im
ag
es
an
d
p
r
o
v
id
e
s
u
p
e
r
io
r
ev
alu
atio
n
s
.
T
o
im
p
r
o
v
e
th
e
ac
cu
r
ac
y
o
f
f
o
r
g
er
y
d
etec
tio
n
,
th
e
u
n
wa
n
ted
d
is
to
r
tio
n
s
ar
e
co
m
p
r
ess
ed
an
d
th
e
f
ea
tu
r
es
ar
e
in
cr
ea
s
ed
in
th
e
im
ag
e.
At
f
ir
s
t,
to
elim
in
ate
n
o
is
e
p
r
e
-
p
r
o
ce
s
s
in
g
is
p
er
f
o
r
m
ed
,
an
d
th
e
im
a
g
e
c
o
n
tr
ast
is
in
cr
ea
s
ed
.
T
h
is
wo
r
k
u
tili
ze
s
wien
er
f
ilter
‑
C
L
AHE
[
3
0
]
t
o
p
e
r
f
o
r
m
p
r
e
-
p
r
o
ce
s
s
in
g
.
W
ien
er
f
ilter
wo
r
k
s
o
n
s
tatis
tical
m
o
d
els
h
as
a
lo
w
p
ass
f
ilter
wh
ich
en
s
u
r
es
th
at
th
e
s
ig
n
als
an
d
n
o
is
e
ar
e
s
tatio
n
ar
y
.
Neig
h
b
o
u
r
h
o
o
d
m
ea
n
s
an
d
v
a
r
ian
ce
ar
e
ca
lcu
lated
,
an
d
s
tr
o
n
g
s
m
o
o
th
in
g
is
u
s
ed
wh
en
th
e
v
ar
iatio
n
is
at
its
m
in
im
u
m
,
wh
ile
th
e
m
in
im
al
s
m
o
o
th
in
g
is
ap
p
lied
wh
en
v
ar
iatio
n
is
at
it
s
m
ax
im
u
m
.
T
h
is
f
ilter
d
ec
r
ea
s
es
th
e
er
r
o
r
b
etwe
en
th
e
o
r
ig
in
al
s
ig
n
al
an
d
esti
m
ated
s
ig
n
al.
I
f
it
is
ass
u
m
ed
th
at
′
is
th
e
esti
m
ated
im
ag
e,
an
d
is
th
e
o
r
ig
in
al
im
ag
e,
th
e
e
r
r
o
r
is
m
ea
s
u
r
ed
u
s
in
g
th
e
(
1
)
,
wh
er
e
t
h
e
u
n
ex
p
ec
te
d
ter
m
is
,
th
en
(
2
)
s
h
o
ws th
e
m
in
im
u
m
er
r
o
r
f
u
n
ctio
n
.
2
=
{
(
−
′
)
2
}
(
1
)
′
(
1
,
2
)
=
[
∗
(
1
,
2
)
′
(
1
,
2
)
|
′
(
1
,
2
)
|
2
(
1
,
2
)
+
′
(
1
,
2
)
]
(
1
,
2
)
(
2
)
W
h
er
e,
a
d
eg
r
ad
ed
im
ag
e
tr
a
n
s
f
o
r
m
is
(
1
,
2
)
,
an
esti
m
ated
im
a
g
e
in
a
f
r
e
q
u
en
c
y
is
′
(
1
,
2
)
,
an
d
a
d
eg
r
ad
e
d
f
u
n
ctio
n
tr
an
s
f
o
r
m
i
s
(
1
,
2
)
.
T
h
en
,
a
n
o
n
-
g
r
ad
in
g
im
ag
e
p
o
wer
s
p
ec
tr
al
v
alu
e
is
′
(
1
,
2
)
,
an
d
(
1
,
2
)
co
n
ju
g
ate
ter
m
is
∗
(
1
,
2
)
.
Pre
-
pr
o
ce
s
s
ed
o
u
tp
u
t
im
ag
es
a
r
e
p
a
s
s
ed
as
in
p
u
t
to
th
e
f
ea
t
u
r
e
ex
tr
ac
tio
n
p
r
o
ce
s
s
.
2
.
3
.
F
e
a
t
ure
ex
t
r
a
ct
io
n
Fo
llo
win
g
th
e
im
ag
e
p
r
e
-
p
r
o
c
ess
in
g
,
VGG
-
1
6
an
d
H‑
(
DT
C
W
T
–
DC
T
)
ar
e
u
tili
ze
d
in
th
is
p
r
o
ce
s
s
o
f
f
ea
tu
r
e
ex
tr
ac
tio
n
.
T
h
is
p
r
o
ce
s
s
o
f
f
ea
tu
r
e
e
x
tr
ac
tio
n
is
p
e
r
f
o
r
m
ed
to
d
ec
r
ea
s
e
d
im
en
s
io
n
ality
wh
er
e
a
lar
g
e
p
ix
el’
s
n
u
m
b
e
r
is
ef
f
ec
tiv
ely
ex
p
r
ess
ed
,
s
o
th
at
th
e
i
m
p
o
r
tan
t
p
ar
ts
o
f
th
e
im
ag
e
ar
e
c
o
llected
.
C
u
r
r
en
tly
,
d
is
cr
ete
wav
elet
tr
an
s
f
o
r
m
is
in
cr
ea
s
ed
with
DT
C
W
T
f
o
r
s
ig
n
if
ican
t
p
r
o
p
er
ties
.
T
h
e
C
NN
’
s
m
o
s
t
r
ec
en
t
v
is
io
n
,
VVG
-
1
6
h
as
a
h
ig
h
le
ar
n
in
g
ca
p
ac
ity
an
d
is
s
im
p
le
to
ass
em
b
le.
Fro
m
VGG
1
6
,
t
h
e
m
o
s
t
im
p
o
r
tan
t
f
ea
tu
r
es
ar
e
ex
tr
ac
ted
with
d
ee
p
n
eu
r
al
n
etwo
r
k
s
p
r
o
ce
s
s
in
g
th
e
1
6
lay
er
s
.
T
h
e
ab
s
tr
a
ctio
n
o
f
im
p
o
r
ta
n
t
f
ea
tu
r
es
f
r
o
m
VGG
-
1
6
an
d
DT
W
C
T
i
s
in
clu
d
ed
in
th
is
f
ea
tu
r
e
ex
tr
ac
tio
n
p
r
o
ce
s
s
.
L
o
w
co
n
tr
ast
ar
ea
s
ar
e
en
h
an
ce
d
wh
ile
th
e
q
u
ality
o
f
th
e
im
ag
e
is
im
p
r
o
v
ed
u
s
in
g
th
e
p
r
e
-
p
r
o
ce
s
s
in
g
tech
n
i
q
u
e.
T
h
e
im
p
o
r
tan
t
f
ea
tu
r
es
ar
e
th
en
ef
f
ec
tiv
ely
ex
tr
ac
ted
f
r
o
m
lo
w
c
o
n
tr
as
t
r
eg
io
n
s
u
s
in
g
t
h
e
ef
f
ec
tiv
e
f
ea
tu
r
e
ex
tr
ac
ti
o
n
tech
n
iq
u
es
VGG
-
1
6
an
d
H‑
(
DT
C
W
T
–
D
C
T
)
,
an
d
t
h
e
p
e
r
f
o
r
m
an
ce
is
en
h
an
ce
d
b
y
d
ec
r
ea
s
in
g
th
e
f
ea
tu
r
e
di
m
en
s
io
n
ality
.
2
.
3
.
1
.
H
‑
(
DT
CWT
–
DCT)
Dir
ec
tio
n
al
s
elec
tiv
ity
in
2
o
r
m
o
r
e
d
im
en
s
io
n
s
an
d
s
h
i
f
t
v
ar
ian
ce
is
s
u
es
ar
e
ad
d
r
ess
ed
u
s
in
g
DT
C
W
T
[
3
1
]
.
W
ith
th
e
ass
is
tan
ce
o
f
a
n
aly
tic
wav
elets,
DT
C
W
T
ac
h
iev
es
d
ir
ec
tio
n
al
s
elec
tiv
ity
.
Six
d
ir
ec
tio
n
al
s
u
b
b
a
n
d
s
ar
e
g
en
e
r
ated
b
y
th
is
m
o
d
el
e
x
p
o
s
ed
i
n
th
e
d
ir
ec
tio
n
s
in
clu
d
in
g
+1
2
,
+4
5
,
+
7
5
r
ea
l
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
36
,
No
.
2
,
No
v
em
b
er
20
24
:
1
1
1
4
-
1
1
2
6
1118
im
ag
in
ar
y
p
a
r
ts
.
L
et
f
ilter
s
b
e
d
en
o
ted
as
ℎ
(
)
,
an
d
(
)
.
T
h
e
f
ir
s
t
(
)
(
)
an
d
s
ec
o
n
d
(
)
(
)
ar
e
ℎ
p
h
ase
f
ilter
b
an
k
r
esp
o
n
s
e
e
x
p
r
ess
ed
in
(
3
)
.
T
o
d
ec
o
m
p
o
s
e
th
e
im
ag
e
(
,
)
u
s
in
g
DT
C
W
T
f
o
r
th
e
tr
an
s
latio
n
’
s
s
er
ies,
co
m
p
lex
s
ca
lin
g
d
ilatio
n
s
an
d
f
u
n
ctio
n
s
o
f
6
C
W
T
ar
e
p
r
o
v
id
ed
in
(
4
)
.
(
)
(
)
=
{
′
(
)
(
)
}
(
3
)
(
,
)
=
∑
∈
2
,
,
(
,
)
+
∑
∑
∑
,
,
(
,
)
∈
2
≥
∈
2
(
4
)
W
h
er
e,
th
e
s
ca
lin
g
ter
m
is
d
en
o
ted
as
,
(
,
)
,
an
d
6
wav
elet
ter
m
s
ar
e
r
ep
r
esen
ted
as
,
(
,
)
,
wh
ile
th
e
wav
elets
an
d
s
ca
lin
g
co
ef
f
icien
ts
ar
e
s
h
o
wn
as
an
d
.
At
last
,
(
,
)
im
ag
e
is
s
ep
ar
ated
in
to
s
u
b
-
b
an
d
s
o
f
n
o
n
-
o
v
e
r
lap
p
in
g
.
Ne
x
t,
th
e
h
ig
h
est
an
d
ce
n
ter
f
r
eq
u
en
cy
s
u
b
-
b
an
d
s
ar
e
d
iv
id
e
d
i
n
to
b
lo
ck
s
o
f
4
×
4
.
Nex
t,
th
e
DC
T
tr
an
s
f
o
r
m
f
u
n
c
tio
n
is
ap
p
ea
led
t
o
ea
ch
s
u
b
-
b
an
d
b
lo
c
k
,
an
d
th
e
DC
T
tr
an
s
f
o
r
m
is
g
iv
en
in
(
5
)
.
At
last
,
H‑
(
DT
C
W
T
–
DC
T
)
c
o
ef
f
icien
ts
ar
e
ex
tr
ac
ted
f
r
o
m
p
r
e
-
p
r
o
ce
s
s
ed
im
ag
es
u
s
in
g
H‑
(
DT
C
W
T
–
D
C
T
)
.
(
,
)
=
√
2
√
2
∑
∑
(
,
)
(
2
+
1
)
∏
2
−
1
=
0
−
1
=
0
(
2
+
1
)
∏
2
(
5
)
2
.
3
.
2
.
VG
G
-
16
T
h
e
VGG
-
1
6
’
s
d
ee
p
ar
c
h
itectu
r
e
en
h
an
ce
s
f
ea
tu
r
e
ex
tr
ac
tio
n
f
o
r
co
lo
r
izatio
n
-
b
ased
im
a
g
e
f
o
r
g
e
r
ies
d
etec
tio
n
,
wh
ile
it
is
p
r
e
-
tr
ai
n
ed
o
n
lar
g
e
d
atasets
,
h
en
ce
in
cr
ea
s
in
g
th
e
m
o
d
el
p
e
r
f
o
r
m
an
ce
with
m
i
n
im
al
tr
ain
in
g
d
ata.
T
h
e
VGG
-
1
6
[
3
2
]
h
as
1
6
lay
e
r
s
,
an
d
it
is
an
e
x
ce
llen
t
v
is
io
n
n
etwo
r
k
.
T
h
r
e
e
3
f
u
lly
co
n
n
ec
ted
(
FC
)
lay
er
s
,
1
3
co
n
v
o
lu
tio
n
al
lay
er
s
,
an
d
1
6
lear
n
ab
le
weig
h
ts
lay
er
s
co
n
s
is
t
o
f
4
1
la
y
er
s
o
f
3
×
224
×
224
p
ix
el
VGG
in
p
u
t.
I
t
h
as
two
ty
p
es
o
f
f
ilter
s
:
2
×
2
with
m
a
x
p
o
o
lin
g
lay
er
’
s
s
tr
id
e
2
a
n
d
3
×
3
wit
h
co
n
v
o
l
u
tio
n
al
lay
er
s
’
s
tr
id
e
1
.
T
h
e
6
4
an
d
1
2
8
f
ilter
s
ar
e
i
n
th
e
f
ir
s
t
a
n
d
s
ec
o
n
d
co
n
v
o
l
u
tio
n
al
lay
er
s
.
2
5
6
,
5
1
2
,
an
d
5
1
5
f
ilter
s
ar
e
in
o
th
er
co
n
v
o
lu
tio
n
al
lay
er
s
.
1
st
an
d
2
nd
FC
h
av
e
4
,
0
9
6
n
eu
r
o
n
s
,
an
d
at
th
e
en
d
,
3
FC
lay
er
s
ar
e
p
r
o
d
u
ce
d
.
L
ast
FC
u
tili
ze
d
f
o
r
f
ea
tu
r
es
is
r
ed
u
ce
d
to
a
th
o
u
s
an
d
d
im
en
s
io
n
s
.
T
h
is
co
n
tain
s
1
3
8
m
illi
o
n
p
ar
am
eter
s
ap
p
r
o
x
im
a
tely
,
an
d
is
a
lar
g
e
n
etwo
r
k
.
Fig
u
r
e
4
illu
s
tr
ates th
e
VGG
-
1
6
ar
ch
itectu
r
e.
H‑
(
DT
C
W
T
–
DC
T
)
h
as
a
h
ig
h
d
ir
ec
tio
n
al
s
elec
tiv
ity
f
ea
tu
r
e
th
at
d
ec
o
m
p
o
s
es
im
ag
e
te
x
tu
r
e
f
r
o
m
d
if
f
er
en
t
d
ir
e
ctio
n
s
an
d
ac
q
u
ir
es
d
escr
ip
tio
n
s
o
f
m
u
ltip
le
f
ea
tu
r
e
in
f
o
r
m
atio
n
.
Ad
d
itio
n
ally
,
m
ax
-
p
o
o
lin
g
a
n
d
co
n
v
o
l
u
tio
n
ar
e
ap
p
ea
led
f
o
r
t
h
e
f
ir
s
t 4
b
lo
ck
s
in
VGG
-
1
6
.
T
h
is
is
b
ec
au
s
e
th
e
o
u
tp
u
t o
f
th
e
ac
h
iev
ed
f
ea
tu
r
es
h
as a
s
ize
o
f
16
×
16
×
5
1
2
an
d
i
n
p
u
t im
ag
e
s
ize
is
3
×
224
×
2
2
4
.
T
h
e
ex
tr
ac
te
d
im
ag
e
f
ea
t
u
r
es a
r
e
p
ass
ed
as in
p
u
t
to
th
e
f
ea
tu
r
e
s
elec
tio
n
p
h
ase.
Fig
u
r
e
4
.
Ar
c
h
itectu
r
e
o
f
VGG
-
16
2
.
4
.
F
e
a
t
ure
s
elec
t
io
n
Fo
llo
win
g
th
e
f
ea
tu
r
e
e
x
tr
ac
ti
o
n
,
th
e
f
ea
tu
r
e
d
im
e
n
s
io
n
ality
is
d
ec
r
ea
s
ed
u
s
in
g
th
e
I
HH.
T
h
is
I
HH
m
eth
o
d
i
n
cr
ea
s
es
th
e
d
etec
tio
n
o
f
co
lo
r
izatio
n
-
b
ased
im
a
g
e
f
o
r
g
e
r
y
b
y
e
f
f
ec
tiv
ely
e
x
p
lo
r
in
g
s
o
lu
tio
n
s
p
ac
e,
r
esu
ltin
g
in
f
aster
co
n
v
er
g
en
ce
an
d
h
ig
h
ac
cu
r
ac
y
.
On
e
i
m
p
o
r
tan
t
p
r
o
ce
d
u
r
e
th
at
en
s
u
r
es
o
p
tim
al
f
ea
tu
r
e
co
llectio
n
to
im
p
r
o
v
e
d
etec
tio
n
ac
cu
r
ac
y
,
s
p
ee
d
u
p
d
etec
tio
n
,
lo
wer
er
r
o
r
,
an
d
d
ec
r
ea
s
e
f
ea
tu
r
e
d
im
en
s
io
n
ality
in
f
ea
tu
r
e
s
elec
tio
n
.
Me
tah
eu
r
is
tic
alg
o
r
ith
m
s
ex
tr
ac
t
im
p
o
r
tan
t
f
ea
tu
r
es
f
r
o
m
th
e
f
ea
tu
r
e
s
ets.
T
h
is
o
p
tim
izatio
n
d
ep
en
d
s
o
n
th
e
h
o
r
s
e'
s
ch
ar
ac
ter
is
tics
,
an
d
it
h
as
p
h
ases
th
at
in
clu
d
e
I
m
itatio
n
(
I
)
,
Gr
az
in
g
(
G)
,
So
ciab
ilit
y
(
S),
Def
en
ce
(
D)
an
d
R
o
am
(
R
)
.
B
ased
o
n
th
e
f
o
llo
win
g
(
6
)
,
th
e
m
o
v
e
m
en
t
is
p
r
o
v
id
ed
t
o
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
Dete
ctio
n
o
f c
o
lo
r
iz
a
tio
n
b
a
s
ed
ima
g
e
fo
r
g
eries
u
s
in
g
co
n
v
o
lu
tio
n
a
l
… (
S
o
u
mya
s
h
r
ee
Mu
r
a
lid
h
a
r
P
a
n
ch
a
l
)
1119
h
o
r
s
e
at
ea
ch
iter
atio
n
.
W
h
er
e
a
h
o
r
s
e
v
elo
city
v
ec
to
r
is
r
ep
r
esen
ted
as
⃗
⃗
⃗
,
,
th
e
p
r
esen
t
iter
atio
n
is
d
en
o
ted
as
,
h
o
r
s
e’
s
ag
e
is
s
h
o
wed
as
,
an
d
ℎ
h
o
r
s
e
p
o
s
itio
n
is
d
en
o
ted
as
,
.
,
=
⃗
⃗
⃗
,
+
(
−
1
)
,
(
6
)
Gr
az
in
g
(
G)
:
t
h
e
g
r
az
in
g
f
i
eld
is
d
ev
elo
p
e
d
b
y
I
HH
o
p
tim
izatio
n
ar
o
u
n
d
ea
ch
h
o
r
s
e
with
ℎ
co
ef
f
icien
t,
an
d
g
r
az
i
n
g
m
o
d
e
l
is
ex
p
r
ess
ed
a
s
(
7
)
an
d
(
8
)
.
W
h
er
e
,
th
e
h
o
r
s
e
m
o
tio
n
v
ar
i
ab
le
is
d
en
o
ted
as
,
,
an
d
th
is
e
x
p
r
ess
io
n
is
u
ti
lized
f
o
r
d
ec
r
ea
s
in
g
lin
ea
r
ity
with
iter
atio
n
s
o
f
.
Fu
r
th
er
,
th
e
u
p
p
er
a
n
d
lo
wer
lim
it sp
ac
e
in
g
r
az
in
g
is
r
ep
r
esen
ted
as
an
d
.
,
=
ℎ
(
+
)
+
(
−
1
)
(
7
)
ℎ
,
=
ℎ
(
−
1
)
,
×
(
8
)
Hier
ar
ch
y
(
H)
:
h
o
r
s
es
o
b
s
er
v
e
h
ier
ar
c
h
y
law
in
th
e
m
id
d
le
a
g
e
o
f
an
d
,
as
p
r
o
v
en
b
y
r
esear
ch
.
I
t
is
p
r
o
v
id
ed
in
(
9
)
an
d
(
1
0
)
.
T
h
e
h
o
r
s
e’
s
b
etter
lo
ca
tio
n
is
d
e
n
o
ted
as
∗
(
−
1
)
,
an
d
th
e
ef
f
o
r
t
o
f
h
o
r
s
e’
s
b
ette
r
lo
ca
tio
n
is
r
ep
r
esen
ted
as
⃗
⃗
,
.
⃗
⃗
,
=
(
−
1
)
,
[
∗
(
−
1
)
−
(
−
1
)
]
(
9
)
,
=
ℎ
(
−
1
)
,
×
(
1
0
)
So
ciab
ilit
y
(
S):
i
t
is
d
ef
in
ed
as
th
e
n
u
m
b
er
o
f
o
b
s
er
v
ed
h
o
r
s
es
ag
ed
b
etwe
en
y
ea
r
s
5
t
o
1
5
th
at
ex
p
r
ess
in
ter
est
in
a
h
er
d
,
as
ex
p
r
ess
ed
in
(
1
1
)
an
d
(
1
2
)
.
W
h
er
e
,
th
e
h
o
r
s
e
o
r
ie
n
tatio
n
is
d
en
o
ted
as
,
an
d
s
o
cial
v
ec
to
r
m
o
tio
n
is
s
h
o
wn
as
⃗
,
.
W
ith
th
e
f
ac
to
r
,
,
is
r
ed
u
ce
d
in
ea
ch
cy
cle.
⃗
,
=
,
[
(
1
∑
(
−
1
)
=
1
)
−
(
−
1
)
]
(
1
1
)
,
=
(
−
1
)
,
×
(
1
2
)
I
m
itatio
n
(
I
)
:
t
h
e
h
o
r
s
e
ch
ar
ac
ter
is
tic
is
co
n
s
id
er
ed
as
ter
m
in
th
e
alg
o
r
ith
m
.
Oth
e
r
h
o
r
s
es
an
d
th
e
b
eh
av
io
r
o
f
s
m
all
h
o
r
s
es
d
o
n
o
t
co
n
v
er
t
th
eir
liv
es
wh
ich
is
ex
p
r
ess
ed
as
(
1
3
)
an
d
(
1
4
)
.
W
h
er
e
,
th
e
d
ec
r
ea
s
ed
f
ac
to
r
p
er
c
y
cle
is
d
en
o
ted
as
,
m
o
tio
n
v
ec
to
r
is
d
e
n
o
ted
as
⃗
,
an
d
th
e
p
o
s
itio
n
is
s
h
o
wn
a
s
,
.
T
h
e
b
est p
o
s
itio
n
o
f
th
e
to
tal
h
o
r
s
e
is
s
h
o
wn
as
.
,
=
,
[
(
1
∑
⃗
(
−
1
)
=
1
)
−
(
−
1
)
]
(
1
3
)
,
=
(
−
1
)
,
×
(
1
4
)
Def
en
ce
(
D)
:
u
s
in
g
th
e
n
eg
at
iv
e
co
ef
f
icien
ts
s
y
s
tem
o
f
d
ef
en
s
e
in
I
HH
is
d
eter
m
in
ed
in
(
1
5
)
an
d
(
1
6
)
,
u
tili
ze
d
to
m
o
v
e
a
h
o
r
s
e
b
ey
o
n
d
a
n
u
n
s
u
ited
p
lace
.
W
h
er
e
,
th
e
h
o
r
s
e
with
th
e
wo
r
s
t
p
o
s
itio
n
is
r
ep
r
esen
ted
as
.
T
h
e
escap
e
v
e
cto
r
is
d
en
o
ted
b
y
⃗
⃗
,
.
⃗
⃗
,
=
−
,
[
(
1
∑
⃗
(
−
1
)
=
1
)
−
(
−
1
)
]
(
1
5
)
,
=
(
−
1
)
,
×
(
1
6
)
R
o
am
(
R
)
:
t
h
e
h
o
r
s
e
is
r
o
am
i
n
g
an
d
g
r
az
in
g
to
f
in
d
f
o
o
d
f
r
o
m
o
n
e
p
lace
to
an
o
th
e
r
.
T
h
e
r
o
am
in
g
b
eh
av
io
r
is
co
m
m
o
n
ly
v
iewe
d
in
y
o
u
n
g
h
o
r
s
es,
an
d
r
ep
r
esen
ted
b
y
(
1
7
)
an
d
(
1
8
)
.
W
h
er
e
,
t
h
e
r
ed
u
ce
d
ter
m
is
s
h
o
wn
as
,
an
d
th
e
lo
ca
l
s
ea
r
ch
v
elo
city
v
ec
to
r
is
r
e
p
r
esen
ted
by
⃗
,
.
I
n
I
HH,
(
1
9
)
s
h
o
ws
t
h
e
h
o
r
s
e
v
elo
city
am
o
n
g
ag
es 0
to
5
y
e
ar
s
.
⃗
,
=
(
,
)
(
−
1
)
(
1
7
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
36
,
No
.
2
,
No
v
em
b
er
20
24
:
1
1
1
4
-
1
1
2
6
1120
,
=
(
−
1
)
,
×
(
1
8
)
=
×
(
(
−
1
)
,
(
+
)
[
(
−
1
)
]
)
+
(
−
1
)
×
[
(
−
1
)
,
]
[
(
1
∑
(
−
1
)
=
1
)
−
(
−
1
)
]
+
[
(
−
1
)
,
(
−
1
)
]
(
1
9
)
Fo
r
s
elec
tin
g
th
e
m
ea
n
in
g
f
u
l
f
ea
tu
r
es,
th
e
r
eg
u
latio
n
p
a
r
am
e
ter
s
an
d
ar
e
u
s
ed
.
T
h
e
r
ef
o
r
e,
I
HH
is
em
p
lo
y
ed
to
s
el
ec
t
th
e
o
p
tim
al
ch
ar
ac
ter
is
tics
.
Fig
u
r
e
5
illu
s
tr
ates
th
e
f
lo
wch
ar
t
f
o
r
I
HH.
T
h
e
f
lo
w
ch
ar
t
r
ep
r
esen
ts
th
e
f
ea
tu
r
e
s
elec
ti
o
n
m
eth
o
d
’
s
d
etailed
d
escr
ip
t
io
n
.
Fo
llo
win
g
th
e
p
ar
am
eter
u
p
d
ate,
it
s
atis
f
ies
ev
alu
atio
n
cr
iter
ia
f
o
r
th
e
im
p
lem
en
ted
p
er
f
o
r
m
an
ce
;
t
h
e
p
r
o
ce
d
u
r
e
is
r
ep
ea
ted
u
n
til
ac
h
iev
in
g
th
e
o
p
tim
al
s
o
lu
tio
n
.
T
h
e
o
p
tim
al
f
ea
tu
r
es
en
h
an
ce
th
e
d
etec
tio
n
s
p
ee
d
an
d
p
er
f
o
r
m
an
ce
o
f
o
v
er
all
b
lin
g
f
o
r
g
er
y
d
etec
tio
n
.
T
h
e
o
u
tp
u
t o
f
s
elec
t
ed
f
ea
tu
r
es is
p
ass
ed
as in
p
u
t t
o
th
e
class
if
icatio
n
p
r
o
ce
ss.
Fig
u
r
e
5
.
I
HH
o
p
tim
izatio
n
f
l
o
wch
ar
t
3.
CO
NVO
L
U
T
I
O
NA
L
AU
T
O
E
NCO
D
E
R
-
CO
NVO
L
U
T
I
O
NAL N
E
URA
L
NE
T
WO
RK
Af
ter
th
e
f
ea
tu
r
e
s
elec
tio
n
,
th
e
im
p
lem
en
ted
C
AE
-
C
NN
(
C
AE
-
C
NN)
is
u
tili
ze
d
f
o
r
th
e
ca
teg
o
r
izatio
n
p
r
o
ce
s
s
.
B
y
u
s
in
g
co
n
v
o
lu
ti
o
n
al
lay
e
r
s
to
c
ap
tu
r
e
c
o
m
p
lex
co
lo
r
p
atter
n
s
in
im
ag
es,
C
AE
-
C
NN
im
p
r
o
v
es
th
e
ac
cu
r
ac
y
o
f
f
o
r
g
er
y
d
etec
tio
n
an
d
allo
ws
f
o
r
s
tr
o
n
g
d
if
f
er
e
n
tiatio
n
b
etwe
en
f
o
r
g
e
r
y
an
d
au
th
en
tic
im
a
g
es.
At
f
ir
s
t,
th
e
tr
ain
in
g
d
ataset
is
tr
ain
ed
u
s
in
g
C
AE
.
T
h
e
d
ec
o
d
er
co
m
p
o
n
en
t
is
d
is
ca
r
d
ed
af
ter
th
e
C
AE
co
m
p
letes
its
tr
ain
in
g
p
r
o
ce
d
u
r
e
,
wh
e
r
ea
s
c
o
m
p
r
ess
in
g
f
ir
s
t
h
ig
h
-
d
im
e
n
s
io
n
al
im
a
g
e
d
ataset
in
to
a
co
m
p
r
ess
ed
d
ataset
o
f
im
ag
e
is
ca
r
r
ied
o
u
t
b
y
u
s
in
g
an
en
c
o
d
er
.
At
last
,
th
e
C
AE
’
s
en
co
d
e
r
o
u
tp
u
t
(
co
m
p
r
ess
ed
d
ataset
im
ag
e)
i
s
u
tili
ze
d
f
o
r
tr
ain
in
g
a
n
d
f
ee
d
in
g
ca
teg
o
r
izatio
n
o
f
C
NN
in
clu
d
in
g
VGG,
an
d
R
esNet
.
I
n
th
is
co
n
s
eq
u
en
ce
,
r
ep
r
esen
t
s
th
e
m
o
d
el
o
f
C
NN
class
if
ic
atio
n
an
d
=
{
1
,
2
,
…
}
,
wh
er
e
∈
{
0
,
1
}
∀
∈
ℕ
,
as
N
r
ef
er
s
to
tal
class
o
u
tp
u
t
tar
g
et
in
class
if
icatio
n
is
s
u
e.
o
f
in
itial
tr
ain
in
g
d
at
aset
i
s
co
n
v
er
ted
th
r
o
u
g
h
en
co
d
er
in
to
an
en
co
d
e
d
co
m
p
r
ess
ed
2
D
r
ep
r
esen
tatio
n
.
̂
=
{
̂
1
,
̂
2
,
…
̂
}
is
th
e
r
aw
o
u
tp
u
t
o
f
C
NN
class
if
icat
io
n
m
o
d
el
,
as
g
iv
en
i
n
(
2
0
)
.
T
h
e
r
ec
o
n
s
tr
u
ctio
n
er
r
o
r
in
(
2
1
)
is
u
s
ed
to
m
ea
s
u
r
e
th
e
C
NN
clas
s
if
icatio
n
m
o
d
el’
s
p
er
f
o
r
m
a
n
ce
.
̂
=
(
)
(
2
0
)
=
(
̂
(
)
,
(
)
)
(
2
1
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
Dete
ctio
n
o
f c
o
lo
r
iz
a
tio
n
b
a
s
ed
ima
g
e
fo
r
g
eries
u
s
in
g
co
n
v
o
lu
tio
n
a
l
… (
S
o
u
mya
s
h
r
ee
Mu
r
a
lid
h
a
r
P
a
n
ch
a
l
)
1121
T
h
e
d
i
f
f
er
en
ce
in
m
ea
s
u
r
em
e
n
ts
lik
e
wid
e
an
d
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
f
u
n
ctio
n
s
ar
e
r
ep
r
esen
ted
b
y
t
h
e
f
u
n
ctio
n
.
Nex
t,
(
2
2
)
s
h
o
ws
th
e
g
en
er
al
f
o
r
m
o
f
th
e
co
s
t
f
u
n
ctio
n
.
At
last
,
o
p
tim
al
weig
h
t
p
ar
am
eter
s
ar
e
attain
ed
f
o
r
t
h
e
ca
teg
o
r
izatio
n
m
o
d
el
o
f
C
NN
b
y
r
ed
u
cin
g
th
e
co
s
t
f
u
n
ctio
n
c
o
n
ce
r
n
in
g
task
class
if
icatio
n
.
=
1
∑
(
(
(
(
)
)
)
,
(
)
)
=
1
(
2
2
)
3
.
1
.
I
m
ple
m
ent
ed
CAE’s
a
rc
hite
ct
ura
l desig
n
T
h
e
im
p
lem
e
n
ted
t
o
p
o
lo
g
y
a
r
ch
itectu
r
e
o
f
C
AE
is
r
e
p
r
es
en
ted
in
Fig
u
r
e
6
.
At
last
,
t
h
e
C
AE
’
s
p
ar
am
eter
s
ettin
g
s
co
n
f
ig
u
r
ati
o
n
s
etu
p
is
d
e
n
o
ted
in
T
ab
le
1
.
T
h
e
s
y
m
m
et
r
ic
ar
ch
itectu
r
e
o
f
th
e
im
p
lem
en
ted
C
AE
with
2
D
co
n
v
o
lu
tio
n
al
an
d
d
ec
o
n
v
o
l
u
tio
n
al
lay
er
s
o
f
4
b
atch
es
co
m
es
af
ter
an
ac
t
iv
atio
n
f
u
n
ctio
n
o
f
r
ec
tifie
d
lin
ea
r
u
n
it
(R
e
L
U)
.
T
h
e
d
ec
o
n
v
o
lu
tio
n
al
(
o
r
tr
an
s
p
o
s
ed
co
n
v
o
lu
tio
n
)
is
p
er
f
o
r
m
s
th
e
co
n
v
o
lu
tio
n
lay
er
’
s
r
ev
e
r
s
e
o
p
e
r
atio
n
.
I
t
p
ar
ticu
lar
ly
c
o
n
v
e
r
ts
in
f
o
r
m
a
tio
n
f
r
o
m
a
lo
w
-
d
im
e
n
s
io
n
al
s
p
ac
e
to
a
h
i
g
h
-
d
im
en
s
io
n
al
o
n
e
.
Mo
s
t
p
ar
ticu
lar
ly
,
r
aw
in
co
m
e
d
im
en
s
io
n
s
×
×
3
ar
e
p
ass
ed
to
1
st
lay
er
(
2
D
co
n
v
1
-
R
eL
U1
)
,
wh
ich
is
also
an
in
co
m
e
o
f
C
AE
’
s
E
n
co
d
er
.
T
h
e
d
im
e
n
s
io
n
s
2
⁄
×
2
⁄
d
ev
el
o
p
3
2
d
o
wn
s
am
p
led
s
p
atial
f
ea
tu
r
e
m
ap
s
,
em
p
lo
y
in
g
3
2
f
i
lter
s
o
f
4
×
4
k
e
r
n
el
s
ize.
Af
ter
war
d
,
th
e
o
u
tco
m
e
is
p
ass
ed
in
to
2
nd
lay
er
(
2
D
C
o
n
v
2
-
R
eL
U2
)
wh
ic
h
is
th
e
o
u
tp
u
t
o
f
E
n
co
d
er
with
d
im
en
s
io
n
s
2
⁄
×
2
⁄
×
3
r
ep
r
esen
tatio
n
o
f
co
m
p
r
ess
ed
im
ag
e
,
em
p
l
o
y
in
g
a
d
im
e
n
s
io
n
with
3
f
ilter
s
o
f
2
×
2
k
er
n
el
s
ize.
Usi
n
g
a
s
m
aller
k
er
n
el
s
ize
in
2
nd
lay
er
m
ak
es
it
clea
r
b
ec
au
s
e
f
ea
tu
r
e
m
ap
s
f
r
o
m
1
st
lay
er
’
s
o
u
tp
u
t
h
av
e
s
m
aller
d
im
en
s
io
n
al
s
izes
th
an
th
e
in
p
u
t
im
ag
e.
L
ik
ewise,
th
ir
d
a
n
d
f
o
u
r
th
-
lay
e
r
C
AE
’
s
d
ec
o
d
e
r
co
m
p
o
n
en
t
s
p
er
f
o
r
m
(
2
D
Dec
o
n
v
4
-
R
eL
U4
,
2
D
Dec
o
n
v
3
-
R
eL
U3
)
i
n
an
e
n
co
d
er
’
s
r
ev
er
s
e
o
p
e
r
atio
n
s
y
m
m
etr
ic
way
.
Fig
u
r
e
6
.
C
AE
to
p
o
lo
g
y
’
s
ar
c
h
itectu
r
al
p
r
esen
tatio
n
T
ab
le
1
.
C
AE
to
p
o
l
o
g
y
’
s
p
ar
a
m
eter
s
ettin
g
s
La
y
e
r
s
I
n
p
u
t
S
i
z
e
K
e
r
n
e
l
S
i
z
e
S
t
r
i
d
e
O
u
t
p
u
t
S
i
z
e
En
c
o
d
e
r
2
D
C
o
n
v
(
E.
I
n
p
u
t
)
×
×
3
4
×
4
×
32
2
×
2
×
1
2
⁄
×
2
⁄
×
32
R
e
LU
2
⁄
×
2
⁄
×
32
-
-
2
⁄
×
2
⁄
×
32
2
D
C
o
n
v
2
⁄
×
2
⁄
×
32
2
×
2
×
3
2
×
2
×
1
4
⁄
×
4
⁄
×
3
S
i
g
m
o
i
d
(
E
.
O
u
t
p
u
t
)
4
⁄
×
4
⁄
×
3
-
-
4
⁄
×
4
⁄
×
3
D
e
c
o
d
e
r
2
D
D
e
c
o
n
v
3
(
D
.
I
n
p
u
t
)
4
⁄
×
4
⁄
×
3
2
×
2
×
32
2
×
2
×
1
2
⁄
×
2
⁄
×
32
R
e
LU
2
⁄
×
2
⁄
×
32
-
-
2
⁄
×
2
⁄
×
32
2
D
D
e
c
o
n
v
4
2
⁄
×
2
⁄
×
32
4
×
4
×
3
2
×
2
×
1
×
×
3
S
i
g
m
o
i
d
(
D
.
O
u
t
p
u
t
)
×
×
3
-
-
×
×
3
T
h
er
ef
o
r
e,
ef
f
ec
tiv
e
d
etec
tio
n
ac
cu
r
ac
y
is
ac
h
iev
e
d
b
y
u
s
in
g
th
e
ap
p
r
o
ac
h
p
r
o
v
i
d
ed
,
em
p
l
o
y
in
g
two
d
if
f
er
en
t
d
atasets
.
I
n
co
m
p
a
r
i
s
o
n
to
th
e
ex
is
tin
g
m
eth
o
d
s
,
t
h
e
d
im
en
s
io
n
ality
f
ailu
r
e
an
d
d
etec
tio
n
tim
e
ar
e
lo
w.
Hen
ce
,
t
h
e
im
p
lem
e
n
ted
wo
r
k
is
an
al
y
ze
d
f
o
r
th
e
h
ig
h
l
y
ac
cu
r
ate
u
s
e
in
b
lin
d
im
ag
e
f
o
r
g
er
ies
e
f
f
icien
t
d
etec
tio
n
.
T
h
e
f
o
llo
win
g
s
ec
tio
n
p
r
o
v
id
es a
clea
r
a
n
aly
s
is
o
f
th
e
p
er
f
o
r
m
an
ce
o
f
t
h
e
im
p
le
m
en
ted
s
y
s
tem
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
52
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
,
Vo
l.
36
,
No
.
2
,
No
v
em
b
er
20
24
:
1
1
1
4
-
1
1
2
6
1122
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
im
p
lem
e
n
ted
m
et
h
o
d
is
tr
ain
ed
on
two
d
atasets
,
C
ASI
A
V1
an
d
GR
I
P.
I
n
tel
c
o
r
e
i7
p
r
o
ce
s
s
o
r
,
W
in
d
o
ws
1
0
(
6
4
-
b
it)
o
p
er
atin
g
s
y
s
tem
an
d
1
6
GB
R
AM
ar
e
t
h
e
s
y
s
tem
r
eq
u
ir
em
e
n
ts
f
o
r
im
p
lem
en
tin
g
t
h
e
s
u
g
g
ested
m
eth
o
d
.
T
h
is
m
et
h
o
d
’
s
e
f
f
ec
tiv
en
ess
is
ca
lcu
l
ated
in
ter
m
s
o
f
s
p
ec
if
icity
,
r
ec
all,
F
1
-
s
co
r
e,
p
r
ec
is
io
n
,
s
en
s
itiv
ity
an
d
ac
c
u
r
ac
y
,
ex
p
lain
ed
b
el
o
w.
4
.
1
.
E
v
a
lua
t
i
o
n
pa
ra
m
e
t
er
s
T
h
e
p
a
r
am
eter
s
lik
e
s
p
ec
if
icit
y
,
r
ec
all,
F
1
-
s
co
r
e,
p
r
ec
is
io
n
,
s
en
s
itiv
ity
an
d
ac
cu
r
ac
y
ar
e
u
tili
ze
d
to
co
m
p
u
te
m
o
d
el
p
er
f
o
r
m
an
ce
.
T
h
e
p
ar
am
eter
s
ar
e
ex
p
r
ess
ed
m
ath
em
atica
lly
in
th
e
f
o
llo
win
g
(
2
3
)
,
(
2
4
)
,
(
2
5
)
,
(
2
6
)
,
(
2
7
)
an
d
(
2
8
)
.
W
h
er
e
,
,
,
an
d
ar
e
d
en
o
ted
as
f
alse
p
o
s
itiv
e,
f
alse
n
eg
ativ
e
,
tr
u
e
n
e
g
ativ
e
an
d
tr
u
e
p
o
s
itiv
e.
=
+
(
2
3
)
=
+
(
2
4
)
=
+
+
+
+
(
2
5
)
−
=
∗
+
(
2
6
)
=
(
+
)
(
2
7
)
Sp
e
c
ifi
c
ity
=
(
+
)
(
2
8
)
4
.
2
.
Q
ua
ntit
a
t
iv
e
a
nd
qu
a
lita
t
iv
e
a
na
ly
s
is
T
h
is
s
ec
tio
n
p
r
o
v
id
es
an
ex
a
m
in
atio
n
o
f
th
e
C
AE
m
eth
o
d
to
d
eter
m
in
e
th
e
F
1
-
s
co
r
e,
s
p
ec
if
icity
,
r
ec
all,
p
r
ec
is
io
n
,
s
en
s
itiv
ity
an
d
ac
cu
r
ac
y
.
T
h
e
C
AE
m
eth
o
d
’
s
p
er
f
o
r
m
an
ce
is
ev
alu
ated
u
s
in
g
two
d
atasets
of
C
ASI
A
V1
an
d
GR
I
P.
T
h
e
ef
f
ec
tiv
en
ess
o
f
th
e
im
p
lem
e
n
ted
m
o
d
el
is
r
ep
r
esen
te
d
in
d
if
f
er
en
t ta
b
les b
elo
w.
4
.
2
.
1
.
P
er
f
o
r
m
a
nce
a
na
l
y
s
is
us
ing
CA
SI
A
V1
da
t
a
s
et
T
ab
le
2
r
ep
r
esen
ts
th
e
p
r
e
s
en
ted
f
ea
tu
r
e
ex
tr
ac
tio
n
m
eth
o
d
’
s
p
er
f
o
r
m
an
ce
.
T
h
e
p
r
esen
ted
H‑
(
DT
C
W
T
–
DC
T
)
an
d
VGG
-
16
m
et
h
o
d
s
ar
e
co
m
p
ar
e
d
with
th
e
cu
r
r
e
n
t
f
ea
t
u
r
e
ex
tr
ac
tio
n
m
et
h
o
d
s
in
clu
d
in
g
DT
C
W
T
,
VGG
-
1
6
,
an
d
co
m
p
lex
wav
elet
tr
an
s
f
o
r
m
(
C
W
T
)
u
s
in
g
th
e
C
ASI
A
V1
d
ataset
.
T
h
e
p
r
esen
ted
H‑
(
DT
C
W
T
–
DC
T
)
an
d
VGG
-
16
m
eth
o
d
s
ac
h
iev
e
th
e
h
ig
h
est
v
al
u
es
o
f
9
0
%
p
r
e
cisi
o
n
,
9
1
%
r
ec
all
,
9
3
%
ac
cu
r
ac
y
,
8
5
%
F
1
-
s
co
r
e,
9
2
%
s
en
s
itiv
ity
,
an
d
8
9
%
s
p
ec
if
icity
,
in
co
n
tr
ast
to
t
h
e
ex
is
tin
g
f
ea
tu
r
e
ex
tr
ac
tio
n
m
eth
o
d
s
.
T
ab
le
3
d
en
o
tes
th
e
p
r
esen
ted
f
ea
tu
r
e
s
elec
tio
n
m
eth
o
d
’s
p
er
f
o
r
m
an
ce
.
T
h
e
p
r
esen
ted
I
HH
tech
n
iq
u
e
is
co
m
p
ar
ed
with
cu
r
r
en
t
f
ea
t
u
r
e
s
elec
tio
n
m
o
d
els
lik
e
th
e
f
r
u
it
f
ly
o
p
tim
izatio
n
alg
o
r
ith
m
(
FOA)
,
ar
tific
ial
b
ee
co
lo
n
y
(
AB
C
)
,
an
d
cr
o
w
s
ea
r
ch
alg
o
r
ith
m
(
C
SA)
o
n
th
e
C
ASI
A
V1
d
ataset.
T
h
e
p
r
e
s
en
ted
I
HH
m
eth
o
d
ac
h
iev
e
s
th
e
h
ig
h
est
v
alu
es
o
f
9
2
.
2
1
%
p
r
ec
is
io
n
,
9
3
.
7
9
%
r
e
ca
ll,
9
5
.
6
2
%
ac
cu
r
ac
y
,
9
2
.
8
7
%
F
1
-
s
co
r
e,
9
0
.
9
8
%
s
en
s
itiv
ity
an
d
9
4
.
2
6
% sp
ec
if
icity
,
wh
en
co
m
p
ar
ed
with
o
th
er
f
ea
tu
r
e
s
elec
tio
n
m
et
h
o
d
s
.
T
ab
le
2
.
Per
f
o
r
m
an
ce
o
f
f
ea
tu
r
e
ex
tr
ac
tio
n
m
eth
o
d
s
F
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
m
e
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
R
e
c
a
l
l
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
C
W
T
71
65
75
68
79
70
D
TC
W
T
65
72
66
75
75
68
VGG
-
16
75
74
81
66
71
65
H
‑
(
D
TC
W
T
–
D
C
T)
a
n
d
V
G
G
-
16
93
90
85
91
89
92
T
ab
le
3
.
Per
f
o
r
m
an
ce
o
f
f
ea
tu
r
e
s
elec
tio
n
m
eth
o
d
s
F
e
a
t
u
r
e
s
e
l
e
c
t
i
o
n
me
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
R
e
c
a
l
l
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
F
O
A
7
9
.
8
9
8
0
.
7
8
8
3
.
4
6
8
1
.
7
1
8
6
.
7
4
8
5
.
5
1
C
S
A
8
2
.
5
6
8
9
.
1
2
8
5
.
1
3
9
0
.
8
2
9
1
.
6
3
8
7
.
6
2
A
B
C
9
1
.
2
3
8
6
.
4
5
8
9
.
7
9
8
8
.
9
3
8
7
.
5
2
8
8
.
8
4
I
H
H
9
5
.
6
2
9
2
.
2
1
9
2
.
8
7
9
3
.
7
9
9
4
.
2
6
9
0
.
9
8
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2
5
0
2
-
4
7
52
Dete
ctio
n
o
f c
o
lo
r
iz
a
tio
n
b
a
s
ed
ima
g
e
fo
r
g
eries
u
s
in
g
co
n
v
o
lu
tio
n
a
l
… (
S
o
u
mya
s
h
r
ee
Mu
r
a
lid
h
a
r
P
a
n
ch
a
l
)
1123
T
ab
le
4
s
h
o
ws
th
e
im
p
lem
e
n
ted
m
o
d
el
’s
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
T
h
e
p
r
esen
ted
C
AE
m
o
d
el
is
co
m
p
ar
ed
with
cu
r
r
en
t
class
if
icatio
n
m
o
d
els
s
u
ch
as
VGG
-
1
6
,
I
n
ce
p
tio
n
an
d
Xce
p
tio
n
Net
u
s
in
g
th
e
C
ASI
A
V1
d
ataset.
T
h
e
im
p
lem
en
ted
C
AE
m
eth
o
d
attain
s
th
e
h
ig
h
est
v
alu
es
o
f
9
9
.
5
3
%
p
r
ec
is
io
n
,
9
9
.
6
2
%
r
ec
all,
9
9
.
9
5
%
ac
cu
r
ac
y
,
9
9
.
8
2
%
F
1
-
s
co
r
e,
9
9
.
7
0
%
s
en
s
itiv
ity
,
an
d
9
9
.
8
9
%
s
p
ec
if
icity
wh
en
co
m
p
ar
ed
to
th
e
o
t
h
er
class
if
icatio
n
m
o
d
els.
T
ab
le
4
.
Per
f
o
r
m
an
ce
o
f
class
if
icatio
n
m
o
d
els
C
l
a
s
si
f
i
c
a
t
i
o
n
me
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
R
e
c
a
l
l
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
VGG
-
16
9
1
.
8
4
8
9
.
8
2
8
8
.
9
4
9
2
.
3
1
8
5
.
9
9
9
0
.
8
9
I
n
c
e
p
t
i
o
n
9
5
.
5
2
9
2
.
7
1
9
3
.
6
1
9
0
.
6
4
9
4
.
8
1
9
1
.
5
7
X
c
e
p
t
i
o
n
N
e
t
9
3
.
2
8
9
0
.
9
3
9
4
.
3
7
8
8
.
9
7
9
5
.
7
8
9
5
.
2
6
C
A
E
9
9
.
9
5
9
9
.
5
3
9
9
.
8
2
9
9
.
6
2
9
9
.
8
9
9
9
.
7
0
4
.
2
.
2
.
P
er
f
o
r
m
a
nce
a
na
l
y
s
is
us
ing
G
RIP
da
t
a
s
et
T
ab
le
5
r
ep
r
esen
ts
th
e
p
r
o
p
o
s
ed
m
eth
o
d
’
s
f
ea
tu
r
e
e
x
tr
ac
tio
n
p
er
f
o
r
m
a
n
ce
.
T
h
e
p
r
esen
ted
H‑
(
DT
C
W
T
–
DC
T
)
an
d
VGG
-
16
m
eth
o
d
s
ar
e
co
m
p
ar
e
d
with
cu
r
r
en
t m
et
h
o
d
s
o
f
f
ea
t
u
r
e
e
x
tr
ac
tio
n
,
in
clu
d
in
g
DT
C
W
T
,
VGG
-
16
an
d
CWT
o
n
th
e
GR
I
P
d
ataset
.
T
h
e
p
r
esen
ted
H‑
(
DT
C
W
T
–
DC
T
)
an
d
VGG
-
16
m
eth
o
d
s
attain
s
u
p
er
io
r
v
al
u
es
o
f
9
1
%
p
r
ec
is
io
n
,
9
2
%
r
ec
all,
9
3
%
ac
cu
r
ac
y
,
9
0
%
F
1
-
s
co
r
e,
9
1
%
s
en
s
itiv
ity
an
d
9
2
%
s
p
ec
if
icity
,
in
co
n
tr
ast to
th
e
e
x
is
tin
g
m
eth
o
d
s
o
f
f
ea
tu
r
e
ex
tr
ac
tio
n
.
T
ab
le
5
.
Per
f
o
r
m
an
ce
o
f
f
ea
tu
r
e
ex
tr
ac
tio
n
m
eth
o
d
s
F
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
m
e
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
R
e
c
a
l
l
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
C
W
T
89
70
88
79
85
87
D
TC
W
T
90
89
75
84
89
79
VGG
-
16
78
85
86
90
76
81
H
‑
(
D
TC
W
T
–
D
C
T)
a
n
d
V
G
G
-
16
93
91
90
92
92
91
Fu
r
th
er
m
o
r
e
,
T
ab
le
6
r
ep
r
e
s
en
ts
th
e
p
r
o
p
o
s
ed
m
eth
o
d
’s
f
ea
tu
r
e
s
elec
tio
n
p
er
f
o
r
m
an
ce
.
T
h
e
p
r
esen
ted
I
HH
m
eth
o
d
is
co
m
p
ar
ed
with
cu
r
r
en
t
f
ea
tu
r
e
s
elec
tio
n
tech
n
iq
u
es
lik
e
th
e
f
r
u
it
FOA
,
A
B
C
,
an
d
C
SA
o
n
th
e
GR
I
P
d
ataset.
T
h
e
p
r
esen
ted
I
HH
m
eth
o
d
ac
h
iev
e
s
th
e
h
ig
h
est
v
alu
es
o
f
9
3
.
8
9
%
p
r
ec
is
io
n
,
9
4
.
4
9
%
r
ec
all,
9
5
.
3
5
%
ac
cu
r
ac
y
,
9
3
.
9
5
%
F
1
-
s
co
r
e,
9
3
.
9
5
%
s
en
s
itiv
ity
,
an
d
9
2
.
6
2
%
s
p
ec
if
icity
,
co
n
tr
ar
y
to
th
e
ex
is
tin
g
f
ea
tu
r
e
s
elec
tio
n
m
eth
o
d
s
.
T
ab
le
6
.
Per
f
o
r
m
an
ce
o
f
f
ea
tu
r
e
s
elec
tio
n
m
eth
o
d
s
F
e
a
t
u
r
e
s
e
l
e
c
t
i
o
n
me
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
R
e
c
a
l
l
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
F
O
A
8
7
.
3
6
8
8
.
3
5
8
2
.
9
3
8
5
.
9
6
7
9
.
9
8
9
0
.
4
5
C
S
A
9
2
.
2
5
8
2
.
5
9
8
5
.
8
1
8
6
.
8
5
8
6
.
8
9
8
3
.
9
7
A
B
C
8
9
.
1
4
9
0
.
7
8
8
9
.
5
0
9
1
.
7
4
8
1
.
7
2
8
5
.
8
4
I
H
H
9
5
.
3
5
9
3
.
8
9
9
2
.
7
5
9
4
.
4
9
9
2
.
6
2
9
3
.
9
5
T
a
b
l
e
7
s
h
o
w
s
t
h
e
i
m
p
l
e
m
e
n
te
d
m
o
d
e
l
’s
c
l
a
s
s
i
f
i
c
a
ti
o
n
p
e
r
f
o
r
m
a
n
c
e
.
T
h
e
p
r
e
s
e
n
t
e
d
C
AE
m
e
t
h
o
d
i
s
c
o
m
p
a
r
e
d
w
i
t
h
c
u
r
r
e
n
t
c
l
as
s
i
f
ic
a
t
i
o
n
VG
G
-
1
6
,
I
n
c
e
p
t
i
o
n
,
a
n
d
X
c
e
p
t
i
o
n
N
et
m
o
d
e
ls
o
n
t
h
e
G
R
I
P
d
a
t
as
et
.
T
h
e
i
m
p
l
e
m
e
n
t
e
d
C
A
E
m
e
t
h
o
d
a
ch
i
e
v
e
s
t
h
e
h
i
g
h
e
s
t
v
a
l
u
es
o
f
9
9
.
5
9
%
p
r
e
c
i
s
i
o
n
,
9
9
.
7
0
%
r
e
c
a
ll
,
9
9
.
9
7
%
a
c
c
u
r
a
c
y
,
9
9
.
4
8
%
F
1
-
s
c
o
r
e
,
9
9
.
6
2
%
s
e
n
s
i
t
i
v
i
t
y
a
n
d
9
9
.
5
8
%
s
p
e
c
i
f
i
ci
t
y
,
c
o
n
t
r
a
r
y
t
o
t
h
e
e
x
i
s
t
i
n
g
c
l
as
s
i
f
ic
a
t
i
o
n
m
o
d
e
l
s
.
T
ab
le
7
.
Per
f
o
r
m
an
ce
o
f
class
if
icatio
n
m
o
d
els
C
l
a
s
si
f
i
c
a
t
i
o
n
me
t
h
o
d
s
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
F1
-
sc
o
r
e
(
%)
R
e
c
a
l
l
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
VGG
-
16
8
5
.
9
0
8
0
.
9
7
9
1
.
6
8
8
9
.
9
8
9
5
.
1
7
8
7
.
8
4
I
n
c
e
p
t
i
o
n
9
3
.
8
4
9
2
.
5
0
9
2
.
1
8
9
0
.
6
5
9
3
.
2
8
8
9
.
9
5
X
c
e
p
t
i
o
n
N
e
t
8
9
.
9
9
9
0
.
6
4
9
3
.
3
5
9
2
.
3
2
9
2
.
9
3
9
4
.
8
6
C
A
E
9
9
.
9
7
9
9
.
5
9
9
9
.
4
8
9
9
.
7
0
9
9
.
5
8
9
9
.
6
2
Evaluation Warning : The document was created with Spire.PDF for Python.