T
E
L
K
O
M
NIKA
T
elec
o
mm
un
ica
t
io
n,
Co
m
pu
t
ing
,
E
lect
ro
nics
a
nd
Co
ntr
o
l
Vo
l.
1
8
,
No
.
4
,
A
u
g
u
s
t
2020
,
p
p
.
1
9
3
4
~
1
9
4
1
I
SS
N:
1
6
9
3
-
6
9
3
0
,
ac
cr
ed
ited
First Gr
ad
e
b
y
Ke
m
e
n
r
is
te
k
d
i
k
ti,
Dec
r
ee
No
: 2
1
/E/KPT
/2
0
1
8
DOI
: 1
0
.
1
2
9
2
8
/
T
E
L
KOM
NI
KA
.
v
1
8
i
4
.
1
4
8
6
4
1934
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//jo
u
r
n
a
l.u
a
d
.
a
c.
id
/in
d
ex
.
p
h
p
/TELK
OM
N
I
K
A
H
a
ndw
riting identif
ica
tion
using
deep
co
nv
o
l
utiona
l neural n
et
w
o
rk
m
ethod
O
k
a
Su
da
na
,
I
Wa
y
a
n G
un
a
y
a
,
I
K
et
ut
G
ede
Da
r
m
a
P
ut
ra
De
p
a
rtme
n
t
o
f
In
f
o
rm
a
ti
o
n
T
e
c
h
n
o
lo
g
y
,
Ud
a
y
a
n
a
Un
iv
e
rsit
y
,
In
d
o
n
e
sia
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Dec
8
,
2019
R
ev
i
s
ed
Feb
2
8
,
2
0
2
0
A
cc
ep
ted
Ma
r
1
3
,
2
0
2
0
Ha
n
d
w
rit
in
g
is
a
u
n
iq
u
e
th
i
n
g
t
h
a
t
p
ro
d
u
c
e
d
d
if
f
e
re
n
tl
y
f
o
r
e
a
c
h
p
e
rso
n
.
Ha
n
d
w
rit
in
g
h
a
s
a
c
h
a
ra
c
teristic
th
a
t
re
m
a
in
th
e
sa
m
e
w
it
h
sin
g
le
w
rit
e
r,
so
a
h
a
n
d
w
rit
in
g
c
a
n
b
e
u
se
d
a
s
a
v
a
riab
le
in
b
io
m
e
tri
c
sy
ste
m
s.
Eac
h
p
e
rso
n
h
a
v
e
a
d
iffere
n
t
f
o
r
m
o
f
h
a
n
d
w
r
it
in
g
sty
le
b
u
t
w
it
h
a
s
m
a
ll
p
o
ss
ib
il
it
y
th
a
t
sa
m
e
c
h
a
ra
c
ter
s
h
a
v
e
so
m
e
th
in
g
c
o
m
m
o
n
s.
W
e
p
ro
p
o
se
a
h
a
n
d
w
rit
in
g
id
e
n
ti
f
ica
ti
o
n
m
e
th
o
d
u
si
n
g
se
n
ten
c
e
se
g
m
e
n
ted
h
a
n
d
w
rit
in
g
f
o
rm
s
.
S
e
n
ten
c
e
f
o
r
m
is
u
se
d
to
g
e
t
m
o
re
c
o
m
p
lete
h
a
n
d
w
rit
in
g
c
h
a
ra
c
teristics
t
h
a
n
u
si
n
g
a
sin
g
le
c
h
a
ra
c
t
e
rs
o
r
w
o
rd
s.
Da
tas
e
t
u
se
d
is
d
iv
id
e
d
in
to
t
h
re
e
c
a
teg
o
ries
o
f
i
m
a
g
e
s,
b
in
a
r
y
,
g
r
a
y
s
c
a
le,
a
n
d
in
v
e
rted
b
in
a
ry
.
A
ll
d
a
tas
e
ts
h
a
v
e
sa
m
e
im
a
g
e
w
it
h
d
iff
e
re
n
t
in
c
o
lo
r
a
n
d
c
o
n
sist
o
f
1
0
0
c
las
s.
T
ra
n
sfe
r
lea
r
n
in
g
u
se
d
in
th
is
p
a
p
e
r
a
re
p
re
-
train
e
d
m
o
d
e
l
V
G
G
1
9
.
T
ra
in
in
g
w
a
s
c
o
n
d
u
c
te
d
in
1
0
0
e
p
o
c
h
s.
Hig
h
e
st
re
su
lt
is
g
ra
y
sc
a
le
i
m
a
g
e
s
w
it
h
g
e
n
u
in
c
e
a
c
c
e
p
tan
c
e
ra
te
o
f
9
2
.
3
%
a
n
d
e
q
u
a
l
e
rro
r
ra
te o
f
7
.
7
%
.
K
ey
w
o
r
d
s
:
B
io
m
etr
ic
s
C
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
e
t
w
o
r
k
T
r
an
s
f
er
lear
n
in
g
W
r
iter
id
en
tif
icatio
n
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
:
I
W
ay
a
n
Gu
n
a
y
a,
Dep
ar
t
m
en
t o
f
I
n
f
o
r
m
atio
n
T
e
ch
n
o
lo
g
y
,
Ud
a
y
an
a
U
n
i
v
er
s
it
y
,
R
a
y
a
Ka
m
p
u
s
U
n
u
d
J
i
m
b
ar
an
St.,
Kec
.
Ku
ta
Sel.,
Kab
u
p
ate
n
B
ad
u
n
g
,
B
ali,
I
n
d
o
n
esia.
E
m
ail:
w
a
y
a
n
g
u
n
a
y
a
@
s
tu
d
e
n
t
.
u
n
u
d
.
ac
.
id
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
h
an
d
w
r
iti
n
g
id
en
ti
f
icatio
n
is
a
p
r
o
b
le
m
t
h
at
i
s
c
u
r
r
en
tl
y
s
ti
ll
b
ein
g
s
tu
d
ied
.
Ha
n
d
w
r
i
tin
g
is
ea
s
y
to
r
ec
o
g
n
ize
b
ec
au
s
e
o
f
a
n
u
m
b
er
o
f
t
h
i
n
g
s
,
li
k
e
t
h
e
f
o
r
m
o
f
letter
s
a
n
d
d
i
f
f
er
e
n
t
s
t
y
les
o
f
w
r
iti
n
g
b
et
w
ee
n
ea
ch
p
er
s
o
n
.
Ha
n
d
w
r
iti
n
g
i
d
en
tific
atio
n
ca
n
b
e
v
er
y
u
s
ef
u
l
e
s
p
ec
iall
y
f
o
r
s
ec
u
r
it
y
s
ec
to
r
,
b
an
k
in
g
,
an
d
f
o
r
en
s
ics.
T
h
e
b
an
k
i
n
g
s
ec
to
r
r
eq
u
ir
es
h
an
d
w
r
iti
n
g
r
e
co
g
n
itio
n
to
v
er
if
y
r
ec
eip
ts
o
r
w
h
e
n
d
is
b
u
r
s
in
g
ch
ec
k
m
o
n
e
y
.
I
n
f
o
r
en
s
ic
f
iel
d
,
h
an
d
w
r
iti
n
g
id
en
ti
f
icat
io
n
ca
n
b
e
u
s
ed
as
a
b
e
n
ch
m
ar
k
f
o
r
a
cr
i
m
i
n
al
ca
s
e
r
elate
d
to
a
h
an
d
w
r
iti
n
g
[
1
]
.
Han
d
w
r
iti
n
g
id
e
n
ti
f
icatio
n
h
as
b
ee
n
d
o
n
e
u
s
i
n
g
s
i
m
p
le
an
d
lo
n
g
-
s
tan
d
i
n
g
m
e
th
o
d
s
.
Ma
ch
in
e
lear
n
in
g
is
a
n
e
w
f
ield
o
f
s
cie
n
ce
w
h
er
e
th
is
f
ield
u
s
e
a
co
ll
ec
tio
n
o
f
d
ata
an
d
u
s
e
t
h
at
d
ata
to
s
tu
d
y
it
p
atter
n
s
to
g
et
a
ce
r
tain
k
n
o
w
led
g
e.
W
h
en
th
e
lear
n
i
n
g
p
r
o
ce
s
s
ab
o
u
t
ex
is
tin
g
p
atter
n
s
d
ee
p
en
s
,
it
w
ill
b
ec
o
m
e
d
ee
p
lear
n
in
g
.
Dee
p
lear
n
in
g
i
s
t
h
e
p
r
o
ce
s
s
o
f
lear
n
in
g
ar
ti
f
icial
n
eu
r
al
n
et
w
o
r
k
s
w
h
er
e
p
atter
n
s
ar
e
s
tu
d
ied
m
o
r
e
d
ee
p
ly
to
p
r
o
d
u
ce
b
etter
r
esu
l
t
s
.
Dee
p
lear
n
in
g
m
i
m
ics t
h
e
a
b
ilit
y
o
f
h
u
m
a
n
b
r
ain
to
r
e
m
e
m
b
er
t
h
i
n
g
s
.
I
m
p
le
m
e
n
tatio
n
o
f
d
ee
p
lear
n
i
n
g
w
ill
b
e
ab
le
to
lear
n
e
v
er
y
f
ea
t
u
r
e
th
a
t e
x
is
t
s
i
f
th
e
n
et
w
o
r
k
is
m
ad
e
d
ee
p
er
.
Dee
p
lear
n
in
g
i
s
d
es
ig
n
ed
to
u
s
e
s
e
v
er
al
la
y
er
s
o
f
ar
tif
icia
l
n
e
u
r
al
n
et
w
o
r
k
t
o
lear
n
ab
o
u
t
ea
ch
f
ea
t
u
r
e.
A
r
ch
itec
tu
r
e
in
d
ee
p
lear
n
i
n
g
m
o
d
el
s
p
la
y
an
i
m
p
o
r
tan
t
r
o
le
in
th
e
ab
ilit
y
o
f
n
et
w
o
r
k
to
s
tu
d
y
ex
is
t
in
g
d
ata.
T
h
er
e
a
r
e
s
ev
er
al
m
o
d
els
in
d
ee
p
lear
n
in
g
,
s
u
ch
as
d
ee
p
n
eu
r
al
n
et
w
o
r
k
s
,
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
et
w
o
r
k
,
an
d
r
ec
u
r
r
en
t n
e
u
r
al
n
et
w
o
r
k
[
2
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
K
A
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
Ha
n
d
w
r
itin
g
id
en
tifi
ca
tio
n
u
s
i
n
g
d
ee
p
co
n
v
o
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k
meth
o
d
(
Oka
S
u
d
a
n
a
)
1935
C
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
et
w
o
r
k
(
C
NN)
is
a
d
ee
p
lear
n
in
g
alg
o
r
ith
m
th
at
u
s
e
co
n
v
o
lu
tio
n
la
y
er
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
a
f
u
ll
y
co
n
n
ec
ted
la
y
er
f
o
r
its
class
i
f
icatio
n
.
T
r
ain
in
g
p
r
o
ce
s
s
in
th
e
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
et
w
o
r
k
u
s
e
b
ac
k
p
r
o
p
ag
atio
n
a
lg
o
r
it
h
m
to
u
p
d
ate
t
h
e
w
ei
g
h
t
an
d
b
ia
s
f
o
r
ea
c
h
ep
o
ch
.
E
ac
h
n
e
u
r
o
n
f
o
u
n
d
o
n
C
NN
w
ill
b
e
co
n
n
e
cted
to
th
e
i
n
p
u
t
i
m
ag
e,
s
o
ar
ch
itect
u
r
es
o
f
C
NN
w
i
ll
n
ee
d
s
ev
er
al
d
ee
p
la
y
er
s
.
T
h
e
class
i
f
icatio
n
p
r
o
ce
s
s
w
i
ll
u
s
e
f
u
ll
y
co
n
n
ec
ted
la
y
er
s
to
class
i
f
y
ex
is
ti
n
g
f
ea
tu
r
e
s
w
it
h
p
r
ed
eter
m
i
n
ed
class
es.
On
e
o
f
t
h
e
d
is
ad
v
an
ta
g
es
o
f
t
h
e
C
NN
m
et
h
o
d
is
t
h
a
t
th
e
a
m
o
u
n
t
o
f
d
ata
n
ee
d
ed
is
q
u
ite
lar
g
e
to
g
et
o
p
tim
a
l r
esu
l
ts
.
T
h
e
h
ar
d
w
ar
e
n
ee
d
ed
m
u
s
t a
ls
o
b
e
s
u
f
f
icie
n
t
en
o
u
g
h
[
3
]
.
Han
d
w
r
iti
n
g
id
e
n
ti
f
icatio
n
b
y
u
t
ilizi
n
g
d
ee
p
lear
n
i
n
g
alg
o
r
ith
m
s
w
ill
b
e
ab
le
to
p
r
o
d
u
ce
h
i
g
h
er
ac
cu
r
ac
y
v
a
lu
e
s
co
m
p
ar
ed
to
o
th
er
m
et
h
o
d
s
.
Han
d
w
r
i
tin
g
c
o
llected
f
r
o
m
v
ar
io
u
s
a
u
t
h
o
r
s
th
at
w
ill
b
e
u
s
ed
as
d
ata
to
b
e
s
tu
d
ied
.
T
h
e
u
s
e
o
f
d
ee
p
lear
n
in
g
al
g
o
r
ith
m
is
ex
p
ec
ted
to
b
e
ab
le
to
i
m
p
r
o
v
e
th
e
r
esu
lts
o
f
h
a
n
d
w
r
i
tin
g
id
en
tific
atio
n
s
o
th
at
it
ca
n
b
e
u
s
ed
p
r
o
p
er
ly
[
4
]
.
T
h
er
e
ar
e
s
ev
er
al
s
tu
d
ies
an
d
r
esear
ch
p
ap
er
th
at
r
elate
d
to
th
is
r
esear
c
h
t
h
at
h
av
e
b
ee
n
ca
r
r
ied
o
u
t.
So
m
e
o
f
t
h
e
m
i
n
t
h
e
f
o
r
m
o
f
c
h
ar
ac
ter
r
ec
o
g
n
it
io
n
an
d
h
an
d
w
r
it
in
g
id
en
ti
f
ica
tio
n
u
s
in
g
o
th
er
o
r
o
ld
er
m
et
h
o
d
.
R
esear
ch
ab
o
u
t
ch
ar
ac
ter
r
ec
o
g
n
itio
n
h
a
s
b
ee
n
co
n
d
u
cted
b
y
De
w
a
e
n
ti
tle
d
“
C
o
n
v
o
l
u
tio
n
a
l
Ne
u
r
al
N
et
w
o
r
k
s
f
o
r
Han
d
w
r
itte
n
J
av
an
e
s
e
C
h
ar
ac
ter
s
R
ec
o
g
n
itio
n
”
[
5
]
.
T
h
is
s
t
u
d
y
u
s
ed
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
et
w
o
r
k
m
eth
o
d
to
r
ec
o
g
n
ize
J
av
an
ese
tr
ad
itio
n
al
ch
ar
ac
ter
s
i
n
t
h
e
f
o
r
m
o
f
h
a
n
d
w
r
i
tin
g
.
T
h
e
y
u
s
ed
2
0
0
0
d
ata
w
h
ic
h
i
s
i
n
cl
u
d
ed
in
2
0
class
es
o
f
J
av
a
n
ese
ch
ar
ac
ter
s
.
T
h
is
r
esear
c
h
u
s
e
s
O
p
en
C
V
lib
r
ar
y
w
i
th
a
s
el
f
-
m
ad
e
C
NN
ar
c
h
itect
u
r
e.
T
h
e
r
esu
lt
s
s
h
o
w
t
h
at
th
e
ac
cu
r
ac
y
r
ate
o
b
tain
ed
is
9
0
%.
R
esear
ch
r
elate
d
to
h
a
n
d
w
r
i
ti
n
g
id
e
n
ti
f
icatio
n
h
as
b
ee
n
ca
r
r
ied
o
u
t
b
y
s
e
v
er
al
p
r
ev
io
u
s
r
esear
ch
er
s
,
o
n
e
o
f
th
e
m
is
a
s
t
u
d
y
co
n
d
u
cted
b
y
Dh
a
n
d
r
a
[
6
]
.
I
n
th
is
r
esear
ch
,
h
a
n
d
w
r
i
tin
g
id
en
t
if
ic
atio
n
u
s
e
s
Kan
n
ad
a
ch
ar
ac
ter
s
a
n
d
u
s
e
r
an
d
o
m
tr
a
n
s
f
o
r
m
m
et
h
o
d
an
d
d
is
cr
ete
c
o
s
in
e
tr
a
n
s
f
o
r
m
w
h
ic
h
is
co
m
b
in
ed
to
id
en
ti
f
y
t
h
e
au
th
o
r
.
T
h
e
r
esu
lt
o
f
th
i
s
s
tu
d
y
i
n
d
icate
th
at
t
h
e
lev
el
o
f
ac
cu
r
ac
y
r
ea
ch
es
1
0
0
%.
I
n
ad
d
itio
n
,
it
w
as
also
co
n
clu
d
ed
t
h
at
s
en
ten
ce
s
,
v
ar
y
in
g
s
tr
u
ctu
r
al
v
ar
iat
io
n
s
,
an
d
/o
r
a
co
m
b
in
atio
n
o
f
t
w
o
o
r
m
o
r
e
w
o
r
d
s
h
ad
a
s
ig
n
if
ican
t
i
m
p
ac
t
o
n
th
e
i
d
en
tific
atio
n
o
f
h
an
d
w
r
iti
n
g
’
s
au
th
o
r
.
An
o
th
er
r
esear
ch
w
a
s
co
n
d
u
cted
w
h
er
e
h
an
d
w
r
iti
n
g
id
en
ti
f
icat
io
n
is
d
o
n
e
o
n
lin
e
[
7
]
.
T
h
is
s
tu
d
y
u
s
ed
d
ata
ac
q
u
ir
ed
o
n
lin
e
t
h
at
w
il
l
b
e
p
r
o
ce
s
s
ed
d
ir
ec
tl
y
.
T
h
e
m
e
th
o
d
u
s
ed
i
n
th
i
s
r
esear
ch
i
s
C
NN
w
i
th
d
r
o
p
s
eg
m
e
n
t
m
et
h
o
d
.
T
h
is
m
eth
o
d
u
s
ed
d
ata
au
g
m
e
n
tatio
n
b
y
d
r
o
p
p
in
g
ea
c
h
s
e
g
m
en
t
o
f
th
e
e
n
ter
ed
ch
ar
ac
ter
.
T
h
e
r
esu
lts
o
f
t
h
is
s
t
u
d
y
ar
e
9
8
%
ac
cu
r
ac
y
f
o
r
E
n
g
l
is
h
h
a
n
d
w
r
iti
n
g
a
n
d
9
5
% f
o
r
C
h
i
n
ese
h
a
n
d
w
r
i
tin
g
.
T
h
is
p
ap
er
ai
m
s
to
a
n
al
y
ze
h
an
d
w
r
iti
n
g
id
e
n
ti
f
icatio
n
b
y
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
et
w
o
r
k
m
et
h
o
d
.
Data
s
et
i
n
t
h
is
s
tu
d
y
is
a
co
llectio
n
o
f
h
a
n
d
w
r
i
tten
i
m
a
g
e
s
f
r
o
m
I
AM
h
a
n
d
w
r
iti
n
g
d
at
ab
ase
[
8
]
.
W
e
w
il
l
cr
ea
te
th
r
ee
s
ep
ar
ated
d
ataset
s
w
it
h
d
i
f
f
er
e
n
t
i
m
a
g
e
co
lo
r
s
s
u
c
h
a
s
,
g
r
a
y
s
ca
le,
b
i
n
ar
y
,
a
n
d
in
v
er
ted
b
in
ar
y
.
T
h
e
h
an
d
w
r
it
ten
f
ea
tu
r
e
s
th
at
w
il
l
b
e
a
d
if
f
er
en
tiato
r
is
t
h
e
th
ick
n
e
s
s
,
s
lo
p
e,
an
d
d
is
tan
ce
b
et
w
ee
n
th
e
letter
s
.
T
h
e
f
ea
tu
r
e
s
,
i
m
a
g
es,
a
n
d
m
et
h
o
d
s
u
s
ed
i
n
t
h
i
s
s
tu
d
y
w
il
l
b
e
a
d
if
f
er
en
tiato
r
co
m
p
ar
ed
to
p
r
ev
io
u
s
r
esear
ch
.
T
h
e
m
et
h
o
d
u
s
ed
in
th
i
s
p
ap
er
is
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
et
w
o
r
k
.
C
o
n
v
o
l
u
tio
n
al
n
eu
r
al
n
et
w
o
r
k
(
C
NN)
is
in
c
lu
d
ed
in
t
h
e
t
y
p
e
o
f
d
ee
p
n
eu
r
al
n
et
w
o
r
k
,
b
ec
au
s
e
o
f
t
h
e
n
et
w
o
r
k
d
ep
th
a
n
d
ap
p
lied
m
ai
n
l
y
to
i
m
a
g
er
y
.
C
NN
m
et
h
o
d
p
r
o
v
ed
s
u
cc
ess
f
u
ll
y
i
n
s
u
r
p
ass
i
n
g
o
th
er
m
ac
h
i
n
e
lear
n
i
n
g
m
eth
o
d
s
,
s
u
ch
as
SVM.
E
ac
h
n
eu
r
o
n
in
C
N
N
is
p
r
esen
ted
in
t
w
o
-
d
i
m
en
s
io
n
a
l f
o
r
m
.
T
r
an
s
f
er
lear
n
i
n
g
is
o
n
e
o
f
t
h
e
m
et
h
o
d
s
th
at
b
ein
g
u
s
ed
f
o
r
co
n
v
o
l
u
tio
n
al
n
e
u
r
a
l
n
e
t
w
o
r
k
.
T
r
an
s
f
er
lear
n
in
g
o
n
co
n
v
o
l
u
tio
n
al
n
e
u
r
al
n
et
w
o
r
k
m
ea
n
t
h
e
r
e
-
u
s
e
o
f
m
o
d
el
t
h
at
h
as
b
ee
n
tr
ain
ed
u
s
i
n
g
a
h
u
g
e
a
m
o
u
n
t
o
f
d
ata
an
d
alr
ea
d
y
p
er
f
o
r
m
ed
w
ell
i
n
a
ce
r
ta
in
tas
k
[9
-
11]
.
T
h
is
p
r
e
-
tr
ain
ed
m
o
d
el
is
a
r
es
u
lt
f
r
o
m
a
co
m
p
etitio
n
to
cr
ea
te
a
n
e
w
al
g
o
r
ith
m
f
o
r
o
b
j
ec
ts
d
ete
ctio
n
an
d
class
i
f
icat
io
n
[
1
2
,
13]
.
R
esu
lt
o
f
u
s
i
n
g
p
r
e
-
tr
ain
ed
m
o
d
el
d
ep
en
d
o
n
w
h
ic
h
ar
c
h
itect
u
r
e
t
h
at
w
il
l
b
e
u
s
ed
.
T
h
er
e
ar
e
s
e
v
er
al
ar
ch
itect
u
r
es
l
ik
e
Den
s
eNe
t,
R
e
s
n
e
t,
an
d
VGG
[
1
4
]
.
E
ac
h
o
f
th
is
ar
c
h
itect
u
r
e
w
ill
h
a
v
e
a
d
i
f
f
er
en
t
m
o
d
el
an
d
t
h
e
r
es
u
lt
w
ill
h
av
e
d
i
f
f
er
en
t
ac
c
u
r
ac
y
v
al
u
e
.
C
o
n
ce
p
t
o
f
tr
a
n
s
f
er
lear
n
i
n
g
is
to
u
ti
lize
k
n
o
w
led
g
e
ac
q
u
i
r
ed
f
o
r
o
n
e
task
to
s
o
lv
e
a
n
o
th
er
r
elate
d
tas
k
.
T
h
er
e
ar
e
s
ev
e
r
al
b
en
ef
its
o
f
u
s
i
n
g
tr
a
n
s
f
er
lear
n
in
g
,
s
u
ch
a
s
h
ig
h
er
s
tar
t,
h
i
g
h
er
s
lo
p
e,
an
d
h
ig
h
er
as
y
m
p
to
te
[
1
1
,
1
5
]
.
2.
RE
S
E
ARCH
M
E
T
H
O
D
T
h
is
r
esear
ch
w
ill
m
ai
n
l
y
u
s
e
C
NN
w
i
th
o
u
t
an
y
ad
d
itio
n
i
n
f
ea
t
u
r
es
ex
tr
ac
tio
n
m
et
h
o
d
o
r
ad
d
itio
n
in
cla
s
s
i
f
icatio
n
m
et
h
o
d
.
W
e
w
o
u
ld
li
k
e
to
u
s
e
C
NN
o
n
l
y
w
it
h
p
r
e
-
tr
ain
ed
m
o
d
el
to
s
h
o
w
th
e
p
er
f
o
r
m
a
n
ce
in
ac
cu
r
ac
y
a
n
d
p
r
o
ce
s
s
in
g
tim
e
th
at
w
e
w
ill
g
et.
W
e
u
s
e
d
VGG
o
r
Vis
u
al
Geo
m
e
tr
y
Gr
o
u
p
p
r
e
-
tr
ain
ed
m
o
d
el
th
at
h
a
v
e
1
9
d
ee
p
lay
er
s
,
o
r
VGG1
9
[
1
6
]
.
T
h
is
ar
ch
itect
u
r
e
s
h
o
w
s
t
h
at
a
d
e
ep
la
y
er
in
C
NN
is
an
i
m
p
o
r
tan
t
f
ac
to
r
to
cr
ea
te
a
class
i
f
icatio
n
s
y
s
te
m
th
a
t
h
a
v
e
h
ig
h
r
es
u
lt
ac
cu
r
ac
y
.
VGG1
9
h
as
b
ee
n
tr
ain
ed
u
s
i
n
g
I
m
a
g
eNe
t
d
ataset
s
w
h
ic
h
h
a
v
e
1
0
0
0
class
es
o
f
i
m
a
g
e
s
an
d
ca
n
h
elp
o
v
er
co
m
e
t
h
e
p
r
o
b
lem
o
f
li
m
i
ted
n
u
m
b
er
o
f
i
m
ag
e
s
in
t
h
e
d
ataset
th
at
b
ei
n
g
u
s
ed
in
tr
ai
n
i
n
g
p
r
o
ce
s
s
[
1
0
]
.
W
e
u
s
ed
V
GG1
9
as
p
r
e
-
tr
ain
ed
m
o
d
el
a
n
d
r
ep
lace
th
e
to
p
lay
er
w
it
h
o
u
r
o
w
n
to
p
lay
er
w
it
h
to
tal
n
u
m
b
er
o
f
o
u
r
clas
s
es
.
W
e
u
s
ed
1
0
0
class
es
in
th
is
r
esear
ch
w
i
th
d
ata
s
et
o
f
s
en
te
n
ce
s
i
m
a
g
e
w
it
h
d
i
m
e
n
s
i
o
n
o
f
1
0
0
x
1
2
0
0
p
ix
el.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6930
T
E
L
KOM
NI
K
A
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
,
Vo
l.
1
8
,
No
.
4
,
A
u
g
u
s
t
2020:
1
9
3
4
-
1
9
4
1
1936
Fig
u
r
e
1
s
h
o
w
s
th
e
ar
ch
itect
u
r
e
o
f
C
NN
th
at
b
ein
g
u
s
ed
in
t
h
is
r
esear
ch
.
W
e
u
s
e
VG
G1
9
b
ase
la
y
er
an
d
f
r
ee
ze
it
to
k
ee
p
it
s
v
al
u
e
in
f
ir
s
t
tr
ai
n
in
g
s
tep
.
VGG1
9
u
s
ed
5
b
lo
ck
s
o
f
co
n
v
o
l
u
tio
n
l
a
y
er
,
w
ith
f
ir
s
t
t
w
o
b
lo
ck
s
h
av
e
t
w
o
co
n
v
o
l
u
tio
n
la
y
er
a
n
d
la
s
t
th
r
ee
b
lo
c
k
s
h
av
e
f
o
u
r
la
y
er
s
o
f
co
n
v
o
lu
t
io
n
.
W
e
f
r
ee
ze
t
h
es
e
b
lo
ck
s
to
k
ee
p
its
d
ata
f
r
o
m
p
r
ev
io
u
s
tr
ai
n
i
n
g
.
Af
ter
t
h
e
b
ase
la
y
er
p
r
o
d
u
ce
a
f
ea
tu
r
e
m
ap
,
w
e
ap
p
l
y
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
to
g
et
a
s
i
n
g
le
f
ea
tu
r
e
m
ap
.
T
o
p
lay
er
s
o
f
VGG1
9
r
ep
lace
d
b
y
m
o
d
if
ied
to
p
la
y
er
w
it
h
1
0
0
class
es a
n
d
t
w
o
f
u
ll
y
co
n
n
ec
ted
la
y
er
.
Fig
u
r
e
1
.
VGG1
9
w
it
h
Mo
d
if
i
ed
to
p
lay
er
ar
ch
itect
u
r
e
T
o
p
lay
er
th
at
u
s
ed
in
th
i
s
ar
ch
i
tect
u
r
e
is
o
b
tain
ed
b
y
t
u
n
i
n
g
its
p
ar
a
m
eter
.
P
ar
am
et
er
tu
n
i
n
g
is
i
m
p
o
r
tan
t
b
ec
au
s
e
it
ca
n
a
f
f
ec
t
th
e
r
esu
lt
s
o
f
class
i
f
icatio
n
p
r
o
ce
s
s
[
1
7
]
.
T
u
n
in
g
p
ar
am
eter
in
th
is
r
esear
c
h
is
d
o
n
e
m
an
u
all
y
b
y
tr
ial
a
n
d
er
r
o
r
p
r
o
ce
s
s
,
u
n
t
il
w
e
g
et
t
h
e
b
est
p
ar
a
m
eter
to
tr
ain
o
u
r
d
at
a.
First,
w
e
r
ep
lace
f
latte
n
w
it
h
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
o
r
G
A
P
to
g
en
er
ate
o
n
e
f
ea
tu
r
e
m
ap
f
o
r
ea
ch
co
r
r
esp
o
n
d
in
g
o
u
tp
u
t
f
r
o
m
b
lo
ck
co
n
v
o
l
u
tio
n
la
y
er
[
1
8
]
.
T
h
e
u
s
e
o
f
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
is
to
r
ed
u
ce
t
h
e
o
v
er
f
itti
n
g
th
a
t
p
r
o
n
e
w
h
e
n
w
e
u
s
e
f
u
ll
y
co
n
n
ec
ted
.
Fu
r
t
h
er
m
o
r
e,
g
lo
b
al
av
er
ag
e
p
o
o
li
n
g
ta
k
es
t
h
e
s
u
m
s
o
u
t
v
a
lu
e
o
f
s
p
atial
in
f
o
r
m
a
tio
n
,
s
o
it
w
i
ll sp
ee
d
u
p
th
e
tr
ai
n
i
n
g
p
r
o
ce
s
s
[
1
9
]
.
W
e
co
m
b
i
n
e
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
w
i
th
f
u
ll
y
co
n
n
ec
ted
la
y
er
to
o
b
tain
ed
b
etter
r
esu
l
t.
Af
ter
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
,
w
e
u
s
e
t
w
o
f
u
ll
y
co
n
n
ec
ted
la
y
er
as
h
id
d
en
la
y
er
w
ith
a
d
r
o
p
o
u
t.
Fu
l
l
y
co
n
n
ec
ted
la
y
er
h
a
s
a
f
u
n
ctio
n
to
tak
e
t
h
e
f
ea
tu
r
e
m
ap
an
d
u
s
e
t
h
e
m
to
cla
s
s
i
f
y
th
e
i
m
a
g
e
i
n
to
co
r
r
esp
o
n
d
in
g
lab
el,
b
u
t
f
u
ll
y
co
n
n
ec
ted
la
y
er
i
s
p
r
o
n
e
to
o
v
er
f
itti
n
g
e
s
p
ec
iall
y
w
h
en
l
ar
g
e
n
eu
r
al
n
et
s
is
tr
ain
ed
o
n
r
elati
v
el
y
s
m
all
d
atasets
.
B
ec
au
s
e
o
f
t
h
is
,
w
e
u
s
e
d
r
o
p
o
u
t
to
o
v
er
ca
m
e
th
i
s
p
r
o
b
lem
[
2
0
,
2
1
]
.
Dr
o
p
o
u
t
h
av
e
a
f
u
n
ctio
n
to
r
an
d
o
m
l
y
s
e
tti
n
g
t
h
e
o
u
t
g
o
in
g
ed
g
es
o
f
h
id
d
en
la
y
er
u
n
it
s
t
o
0
at
ea
ch
u
p
d
ate
o
f
tr
ain
in
g
p
h
ase.
B
ec
au
s
e
o
f
th
is
,
w
e
s
et
a
ce
r
tain
a
m
o
u
n
t o
f
d
r
o
p
o
u
t in
th
e
f
u
ll
y
co
n
n
ec
te
d
lay
er
to
av
o
id
o
v
er
f
itti
n
g
.
Op
ti
m
izer
h
a
v
e
a
r
o
le
to
u
p
d
at
e
th
e
w
ei
g
h
t
p
ar
a
m
eter
s
to
m
i
n
i
m
ize
t
h
e
lo
s
s
f
u
n
c
tio
n
.
I
n
t
h
i
s
r
esear
ch
w
e
u
s
e
S
GD
o
r
s
to
ch
ast
ic
g
r
a
d
ien
t d
escen
t a
s
o
p
ti
m
izer
b
ec
au
s
e
SG
D
p
er
f
o
r
m
co
m
p
u
tatio
n
s
n
o
t o
n
th
e
w
h
o
le
d
ataset,
b
u
t
co
m
p
u
te
o
n
s
m
al
l
s
u
b
s
et
o
r
e
v
en
r
an
d
o
m
s
ele
ctio
n
o
f
d
ata
e
x
a
m
p
le
s
,
s
o
i
t
w
il
l
m
ak
e
tr
ai
n
in
g
p
r
o
ce
s
s
m
o
r
e
e
f
f
iecie
n
t
[
2
2
,
23]
.
A
ls
o
,
SGD
ca
n
p
r
o
d
u
ce
t
h
e
s
a
m
e
p
er
f
o
r
m
a
n
ce
a
s
r
e
g
u
la
r
g
r
ad
ie
n
t
d
escen
t
w
h
e
n
th
e
lear
n
i
n
g
r
ate
ar
e
q
u
ite
lo
w
.
T
h
e
s
u
m
m
ar
ize
o
f
C
NN
ar
ch
itectu
r
e
ca
n
b
e
s
ee
n
i
n
T
ab
le
1
.
T
ab
le
1
s
h
o
w
s
t
h
e
s
u
m
m
ar
ize
o
f
C
N
N
ar
ch
itectu
r
e
t
h
at
u
s
ed
in
t
h
is
r
esear
ch
.
T
h
is
ar
ch
itec
tu
r
e
th
an
tr
ain
ed
u
s
i
n
g
Go
o
g
le
C
o
lab
o
r
ato
r
y
b
ec
a
u
s
e
its
co
m
p
atib
ili
t
y
a
n
d
t
h
e
r
es
o
u
r
ce
av
ailab
le
o
n
its
clo
u
d
s
er
v
ice
is
f
as
ter
th
at
p
h
y
s
ical
r
eso
u
r
ce
a
n
d
C
o
lab
o
r
ato
r
y
p
r
o
v
id
e
a
f
r
ee
-
of
-
ch
a
r
g
e
r
eso
u
r
ce
th
a
t
is
e
n
o
u
g
h
t
o
s
o
lv
e
d
e
m
a
n
d
in
g
r
ea
l
-
w
o
r
l p
r
o
b
le
m
s
[
2
4
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
K
A
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
Ha
n
d
w
r
itin
g
id
en
tifi
ca
tio
n
u
s
i
n
g
d
ee
p
co
n
v
o
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k
meth
o
d
(
Oka
S
u
d
a
n
a
)
1937
T
ab
le
1.
Su
m
m
ar
ize
o
f
C
NN
a
r
ch
itect
u
r
e
L
a
y
e
r
Ty
p
e
O
u
t
p
u
t
S
h
a
p
e
L
a
y
e
r
Ty
p
e
O
u
t
p
u
t
S
h
a
p
e
I
n
p
u
t
1
0
0
,
1
2
0
0
,
3
B
l
o
c
k
1
C
o
n
v
o
l
u
t
i
o
n
1
1
0
0
,
1
2
0
0
,
6
4
B
l
o
c
k
4
C
o
n
v
o
l
u
t
i
o
n
1
1
2
,
1
5
0
,
5
1
2
C
o
n
v
o
l
u
t
i
o
n
2
1
0
0
,
1
2
0
0
,
6
4
C
o
n
v
o
l
u
t
i
o
n
2
1
2
,
1
5
0
,
5
1
2
M
a
x
P
o
o
l
i
n
g
5
0
,
6
0
0
,
6
4
C
o
n
v
o
l
u
t
i
o
n
3
1
2
,
1
5
0
,
5
1
2
B
l
o
c
k
2
C
o
n
v
o
l
u
t
i
o
n
1
5
0
,
6
0
0
,
1
2
8
C
o
n
v
o
l
u
t
i
o
n
4
1
2
,
1
5
0
,
5
1
2
C
o
n
v
o
l
u
t
i
o
n
2
5
0
,
6
0
0
,
1
2
8
M
a
x
P
o
o
l
i
n
g
6
,
7
5
,
5
1
2
M
a
x
P
o
o
l
i
n
g
2
5
,
3
0
0
,
1
2
8
B
l
o
c
k
5
C
o
n
v
o
l
u
t
i
o
n
1
6
,
7
5
,
5
1
2
B
l
o
c
k
3
C
o
n
v
o
l
u
t
i
o
n
1
2
5
,
3
0
0
,
2
5
6
C
o
n
v
o
l
u
t
i
o
n
2
6
,
7
5
,
5
1
2
C
o
n
v
o
l
u
t
i
o
n
2
2
5
,
3
0
0
,
256
C
o
n
v
o
l
u
t
i
o
n
3
6
,
7
5
,
5
1
2
C
o
n
v
o
l
u
t
i
o
n
3
2
5
,
3
0
0
,
2
5
6
C
o
n
v
o
l
u
t
i
o
n
4
6
,
7
5
,
5
1
2
C
o
n
v
o
l
u
t
i
o
n
4
2
5
,
3
0
0
,
2
5
6
M
a
x
P
o
o
l
i
n
g
3
,
3
7
,
5
1
2
M
a
x
P
o
o
l
i
n
g
1
2
,
1
5
0
,
2
5
6
G
l
o
b
a
l
A
v
e
r
a
g
e
P
o
o
l
i
n
g
5
1
2
D
e
n
se
+
D
r
o
p
o
u
t
1
0
2
4
D
e
n
se
+
D
r
o
p
o
u
t
1
0
2
4
S
o
f
t
max
1
0
0
3.
RE
SU
L
T
S
A
ND
AN
AL
Y
SI
S
I
n
th
is
s
e
ct
i
o
n
,
w
e
s
h
o
w
r
e
s
u
l
ts
th
a
t
w
e
g
et
b
a
s
e
d
o
n
m
e
t
h
o
d
th
a
t
w
e
u
s
e
.
F
i
r
s
t
,
w
e
g
iv
e
b
r
i
ef
e
x
p
l
an
at
i
o
n
a
b
o
u
t
th
e
im
ag
e
d
a
t
a
s
et
u
s
e
d
in
th
is
r
es
e
a
r
ch
.
T
h
en
,
t
r
a
in
in
g
an
d
t
es
t
in
g
r
esu
l
ts
w
il
l
b
e
ev
alu
a
t
e
d
w
ith
c
o
m
p
a
r
is
o
n
t
o
o
v
e
r
l
o
o
k
th
e
e
f
f
e
ct
o
f
d
if
f
e
r
en
t
p
r
e
p
r
o
c
e
s
s
in
g
im
ag
e
m
et
h
o
d
an
d
u
s
e
o
f
p
r
e
-
t
r
ain
e
d
m
o
d
el
.
3
.
1
.
Da
t
a
s
et
T
h
is
r
esear
ch
i
s
d
o
n
e
u
s
i
n
g
d
ataset
f
r
o
m
I
A
M
Ha
n
d
w
r
iti
n
g
Data
b
ase
[
8
]
.
T
h
e
i
m
a
g
e
u
s
ed
in
t
h
i
s
p
ap
er
is
an
i
m
a
g
e
o
f
h
a
n
d
w
r
i
ti
n
g
f
r
o
m
v
ar
io
u
s
w
r
iter
.
T
h
e
i
m
ag
e
f
r
o
m
I
A
M
Ha
n
d
w
r
iti
n
g
Data
b
ase
is
a
lr
ea
d
y
s
eg
m
e
n
ted
in
s
en
ten
ce
s
f
o
r
m
,
an
d
in
g
r
a
y
s
ca
le.
T
h
is
r
esear
ch
u
s
e
5
4
6
7
i
m
ag
e
s
w
it
h
1
0
0
class
es
o
f
d
i
f
f
er
e
n
t
w
r
iter
s
an
d
4
0
-
1
0
0
i
m
a
g
e
s
i
n
ea
ch
clas
s
f
o
r
tr
ain
i
n
g
p
r
o
ce
s
s
,
an
d
f
i
v
e
i
m
a
g
es
f
o
r
te
s
ti
n
g
p
r
o
ce
s
s
.
W
r
iter
th
en
id
en
ti
f
ied
w
ith
id
,
li
k
e
0
0
0
,
0
5
2
,
an
d
6
7
0
,
th
at
w
e
g
et
f
r
o
m
I
A
M
h
an
d
w
r
it
tin
g
d
ataset
.
T
h
is
i
m
a
g
e
s
ep
ar
ated
in
to
th
r
ee
ca
teg
o
r
ies,
g
r
a
y
s
ca
le
i
m
a
g
es,
b
in
ar
y
i
m
ag
e
s
,
an
d
in
v
er
t
ed
b
in
ar
y
i
m
a
g
es.
B
i
n
ar
y
i
m
ag
e
is
cr
ea
ted
u
s
in
g
g
r
a
y
s
ca
le
i
m
a
g
e
w
i
th
O
ts
u
’
s
th
r
es
h
o
ld
in
g
[
2
5
]
,
a
n
d
i
n
v
er
ted
i
m
a
g
e
cr
ea
ted
u
s
in
g
b
in
ar
y
i
m
a
g
e.
Fi
g
u
r
e
2
s
h
o
w
e
x
a
m
p
le
o
f
t
h
r
ee
ca
teg
o
r
ies
i
m
a
g
e
th
a
t
b
ein
g
u
s
e
in
t
h
is
r
esear
ch
.
T
h
ese
th
r
ee
ca
teg
o
r
ies
h
a
v
e
th
e
s
a
m
e
i
m
a
g
e
w
it
h
d
if
f
er
en
t
co
lo
r
,
an
d
tr
ain
ed
u
s
i
n
g
t
h
e
s
a
m
e
C
NN
A
r
ch
itect
u
r
e.
(
a)
(
b
)
(
c)
Fig
u
r
e
1
.
I
m
a
g
e
d
ataset
u
s
ed
w
it
h
f
o
r
m
at:
(
a)
g
r
a
y
s
ca
le,
(
b
)
b
in
ar
y
,
an
d
(
c)
in
v
er
s
e
b
in
ar
y
3
.
2
.
T
ra
ini
ng
re
s
u
lt
T
r
ain
in
g
p
r
o
ce
s
s
d
o
n
e
b
y
u
s
in
g
Go
o
g
le
C
o
lab
o
r
ato
r
y
[
2
4
]
w
it
h
n
u
m
b
er
o
f
ep
o
ch
u
s
ed
in
th
i
s
tr
ain
i
n
g
is
1
0
0
ep
o
ch
.
T
r
ain
in
g
p
r
o
ce
s
s
also
u
s
e
2
0
%
o
f
tr
ain
i
n
g
d
ata
as
v
al
id
atio
n
d
ata
to
v
alid
ate
tr
ain
in
g
af
ter
o
n
e
ep
o
ch
.
I
t
tak
e
al
m
o
s
t
1
0
h
o
u
r
s
to
co
m
p
lete
th
e
tr
ai
n
in
g
p
r
o
ce
s
s
.
Fig
u
r
e
3
s
h
o
w
c
o
m
p
ar
is
o
n
b
et
w
ee
n
tr
ain
i
n
g
ac
cu
r
ac
y
f
r
o
m
t
h
e
u
s
es
o
f
b
in
ar
y
i
m
a
g
e,
g
r
a
y
s
ca
l
e
i
m
a
g
e,
a
n
d
i
n
v
er
ted
b
i
n
ar
y
i
m
a
g
e
as
tr
ain
in
g
d
ataset.
B
ased
o
n
th
is
f
ig
u
r
e,
t
r
ain
in
g
ac
cu
r
ac
y
f
r
o
m
in
v
er
te
d
b
in
ar
y
i
m
a
g
e
h
a
v
e
s
teep
er
cu
r
v
e
t
h
an
t
h
e
o
th
er
t
w
o
i
m
a
g
es
d
ataset,
a
n
d
h
av
e
a
m
o
r
e
s
tab
le
lear
n
i
n
g
c
u
r
v
e.
I
n
v
er
ted
b
i
n
ar
y
i
m
a
g
e
a
n
d
g
r
a
y
s
ca
le
i
m
a
g
e
h
a
v
e
a
tr
ain
i
n
g
ac
c
u
r
ac
y
t
h
at
r
ea
c
h
9
9
%,
g
r
a
y
s
ca
le
i
m
a
g
e
h
av
e
t
r
ain
in
g
ac
c
u
r
ac
y
o
f
9
7
%
an
d
b
in
ar
y
i
m
a
g
e
h
a
v
e
tr
ain
i
n
g
ac
c
u
r
ac
y
o
f
9
7
%.
F
ig
u
r
e
4
s
h
o
w
v
al
id
atio
n
ac
c
u
r
ac
y
b
et
w
ee
n
t
h
e
t
h
r
ee
d
atasets
.
Valid
atio
n
p
r
o
ce
s
s
is
u
s
ed
to
ev
al
u
ate
t
h
e
m
o
d
el
af
ter
o
n
e
ep
o
ch
is
d
o
n
e
an
d
is
u
s
ed
to
g
iv
e
a
n
u
n
b
iased
esti
m
ate
o
f
t
h
e
tr
ain
ed
m
o
d
el
[
2
6
]
.
Valid
atio
n
d
ata
s
et
u
s
es
in
t
h
is
r
esear
ch
i
s
2
0
%
f
r
o
m
tr
ain
i
n
g
d
atase
t,
w
it
h
to
t
al
o
f
1
0
5
6
i
m
ag
e
s
.
B
ased
o
n
th
is
f
i
g
u
r
e,
v
alid
atio
n
o
f
g
r
a
y
s
ca
le
i
m
ag
e
d
ata
s
et
i
s
a
m
o
r
e
s
tab
le,
w
it
h
9
0
%
o
f
ac
cu
r
ac
y
co
m
p
ar
ed
to
th
e
o
th
er
t
w
o
i
m
a
g
e
d
ataset
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6930
T
E
L
KOM
NI
K
A
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
,
Vo
l.
1
8
,
No
.
4
,
A
u
g
u
s
t
2020:
1
9
3
4
-
1
9
4
1
1938
Fig
u
r
e
3.
C
o
m
p
ar
is
o
n
o
f
tr
ai
n
i
n
g
ac
c
u
r
ac
y
Fig
u
r
e
4.
C
o
m
p
ar
is
o
n
o
f
v
alid
atio
n
ac
cu
r
ac
y
3.
3
.
T
esting
re
s
ult
a
nd
a
na
ly
s
is
T
h
i
s
r
e
s
e
a
r
c
h
te
s
ts
th
e
m
e
th
o
d
u
s
e
d
w
it
h
th
e
v
a
lu
e
o
f
E
E
R
(
e
q
u
a
l
e
r
r
o
r
r
a
t
e
)
.
W
e
s
o
u
g
h
t
t
o
o
b
t
a
i
n
t
h
e
l
o
w
e
s
t
e
q
u
a
l
e
r
r
o
r
r
a
t
e
b
e
tw
ee
n
th
e
th
r
e
e
d
a
t
as
e
ts
,
an
d
th
en
c
o
m
p
a
r
e
d
i
t
w
ith
o
th
e
r
m
e
t
h
o
d
.
T
o
f
in
d
e
q
u
a
l
e
r
r
o
r
r
a
t
e
,
f
i
r
s
t
w
e
f
in
d
f
a
ls
e
ac
c
e
p
t
an
c
e
r
a
t
e
(
FA
R
)
an
d
f
a
ls
e
r
e
je
c
t
e
d
r
a
t
e
(
FR
R
)
.
FA
R
o
c
c
u
r
s
w
h
en
th
e
te
s
t
i
d
e
n
tif
i
e
d
h
a
n
d
w
r
i
ti
n
g
o
w
n
e
r
t
o
a
n
o
th
e
r
w
r
i
te
r
,
o
r
f
a
l
s
e
p
o
s
i
ti
v
e
.
F
R
R
o
c
cu
r
s
w
h
en
h
an
d
w
r
it
in
g
i
s
r
e
je
c
t
e
d
b
e
c
a
u
s
e
o
f
a
f
ai
lu
r
e
t
o
i
d
en
t
if
iy
,
o
r
f
a
ls
e
n
eg
at
iv
e
[
2
7
]
.
T
o
f
i
n
d
E
E
R
,
FA
R
an
d
F
R
R
i
s
p
lo
t
t
e
d
t
o
g
e
th
e
r
i
n
a
g
r
a
p
h
.
E
E
R
i
s
th
e
p
o
i
n
t
w
h
e
r
e
FA
R
(
1
)
an
d
F
R
R
(
2
)
m
e
et
[
1
0
]
.
E
E
R
is
o
f
t
en
u
s
e
d
t
o
m
e
asu
r
e
b
i
o
m
et
r
ic
s
y
s
tem
s
.
Ge
n
u
i
n
e
a
c
c
e
p
ta
n
c
e
r
a
t
e
(
GA
R
)
(
3
)
is
d
e
f
in
e
d
a
s
p
e
r
c
e
n
ta
g
e
o
f
g
e
n
u
in
e
w
r
i
t
e
r
ac
c
ep
t
e
d
b
y
th
e
s
y
s
tem
.
=
(
1
)
=
(
2
)
=
1
−
(
3
)
W
e
test
o
u
r
tr
ain
ed
m
o
d
el
u
s
i
n
g
5
0
0
i
m
a
g
es (
f
i
v
e
i
m
a
g
es
f
o
r
ea
ch
class
)
.
T
h
e
r
esu
lt
s
s
h
o
w
th
at
u
s
i
n
g
g
r
a
y
s
ca
le
im
ag
es
d
ataset
w
e
o
b
tain
ed
th
at
g
r
a
y
s
ca
le
i
m
a
g
e
s
d
ataset
h
av
e
th
e
lo
w
e
s
t
F
AR
an
d
FR
R
,
an
d
t
h
e
h
ig
h
e
s
t
G
A
R
v
al
u
e.
T
h
e
r
esu
lt
th
e
n
f
o
llo
w
ed
b
y
in
v
er
ted
b
in
ar
y
i
m
ag
e
s
d
ataset
an
d
b
in
ar
y
i
m
a
g
es
d
ataset
.
Fig
u
r
e
5
s
h
o
w
s
th
at
g
r
a
y
s
ca
l
e
i
m
ag
e
y
ield
ed
lo
w
est
f
alse
ac
ce
p
t
an
ce
r
ate
co
m
p
ar
e
to
th
e
o
th
er
t
w
o
.
Fro
m
5
0
0
im
a
g
es
w
e
te
s
ted
,
o
n
l
y
3
8
i
m
a
g
es
w
er
e
id
en
t
if
ied
a
s
w
r
o
n
g
w
r
iter
.
W
e
f
o
u
n
d
th
a
t
s
o
m
e
o
f
t
h
e
3
8
i
m
a
g
es
h
av
e
a
s
i
m
ilar
it
y
w
i
th
t
h
e
cla
s
s
es
t
h
e
w
er
e
id
en
ti
f
ied
a
s
,
w
it
h
ex
a
m
p
le
ca
n
b
e
s
ee
n
i
n
Fi
g
u
r
e
6
.
Fi
g
u
r
e
6
il
lu
s
tr
ates
t
h
e
s
i
m
ilar
it
y
b
et
wee
n
th
e
t
w
o
w
r
iter
class
e
s
.
T
h
e
s
i
m
ilar
it
y
ca
n
b
e
s
ee
n
i
n
s
o
m
e
w
a
y
alp
h
ab
et
s
an
d
/o
r
w
o
r
d
s
is
w
r
i
tten
.
B
ase
d
o
n
tes
t
r
es
u
lt,
i
m
a
g
e
(
a)
id
e
n
ti
f
ied
as
clas
s
3
8
7
w
it
h
8
9
%
ac
cu
r
ac
y
.
I
t
s
h
o
w
s
th
at
s
i
m
ilar
it
y
in
t
h
e
w
a
y
alp
h
ab
ets
is
w
r
itte
n
,
ev
e
n
o
n
l
y
s
o
m
e
o
f
t
h
e
m
,
h
a
v
e
a
g
r
ea
t
i
m
p
ac
t
in
id
en
tific
atio
n
p
r
o
ce
s
s
.
Fig
u
r
e
5
also
s
h
o
w
s
t
h
at
g
r
a
y
s
ca
le
d
ataset
h
as
lo
w
e
s
t
f
a
ls
e
r
ej
ec
ted
r
ate.
B
ased
o
n
t
h
e
test
s
,
o
n
l
y
3
9
i
m
a
g
es
r
ej
ec
ted
.
So
m
e
o
f
it
g
o
t
id
en
tif
ied
as
t
h
e
r
ig
h
t
clas
s
b
u
t
w
ith
lo
w
ac
c
u
r
ac
y
a
n
d
s
o
m
e
g
o
t
id
en
tifie
d
as
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
K
A
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
Ha
n
d
w
r
itin
g
id
en
tifi
ca
tio
n
u
s
i
n
g
d
ee
p
co
n
v
o
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k
meth
o
d
(
Oka
S
u
d
a
n
a
)
1939
w
r
o
n
g
cla
s
s
b
u
t
w
it
h
lo
w
ac
cu
r
ac
y
.
Fi
g
u
r
e
5
also
s
h
o
w
s
g
en
u
i
n
e
ac
ce
p
tan
ce
r
ate
o
f
all
th
r
ee
ca
te
g
o
r
ies
d
ataset.
T
h
e
h
ig
h
est
v
al
u
e
o
f
GAR
w
as
g
r
a
y
s
ca
le
i
m
ag
e
s
d
ataset
w
it
h
v
al
u
e
o
f
9
2
.
3
%.
Fig
u
r
e
5.
C
h
ar
t o
f
F
AR
,
F
R
R
,
an
d
G
A
R
v
alu
e
(
a)
(
b
)
Fig
u
r
e
6.
C
o
m
p
ar
is
o
n
h
an
d
w
r
i
tin
g
o
f
d
if
f
er
en
t
w
r
iter
:
(
a)
w
r
i
ter
w
it
h
id
3
8
5
an
d
,
(
b
)
w
r
iter
w
it
h
id
3
8
7
Fig
u
r
e
7
s
h
o
w
s
th
e
R
OC
o
r
r
ec
eiv
er
o
p
er
atio
n
ch
ar
ac
ter
is
tics
o
f
g
r
a
y
s
ca
le
i
m
a
g
es
d
ataset.
T
h
is
g
r
ap
h
s
h
o
w
s
w
h
er
e
F
AR
an
d
FR
R
v
al
u
e
m
ee
t.
B
ased
o
n
th
is
f
i
g
u
r
e,
eq
u
al
er
r
o
r
r
ate
ca
n
b
e
f
o
u
n
d
o
n
th
r
es
h
o
ld
5
7
.
T
h
e
co
m
p
lete
r
esu
lt
ca
n
b
e
s
ee
n
in
T
ab
le
2
.
B
ased
o
n
t
h
is
r
es
u
lt,
g
r
a
y
s
ca
l
e
i
m
a
g
e
d
ata
s
et
h
as
th
e
b
est
r
es
u
lt
co
m
p
ar
e
w
it
h
th
e
o
t
h
er
d
ataset
s
.
Gr
a
y
s
ca
le
d
ataset
g
e
t
eq
u
al
er
r
o
r
r
ate
(
E
E
R
)
o
f
7
.
7
%
t
h
a
t
cr
o
s
s
ed
at
th
r
esh
o
ld
5
1
.
I
t
s
h
o
w
s
t
h
at
p
r
e
-
tr
ai
n
ed
m
o
d
el
V
GG1
9
h
av
e
a
g
o
o
d
co
m
p
atib
i
lit
y
w
it
h
g
r
a
y
s
ca
le
i
m
a
g
es.
T
h
e
p
o
o
r
est
p
er
f
o
r
m
an
ce
i
s
w
it
h
b
i
n
ar
y
d
ataset
w
it
h
h
i
g
h
est
E
E
R
v
al
u
e
o
f
1
3
.
2
%
th
at
cr
o
s
s
ed
at
th
r
es
h
o
ld
5
2
.
I
n
v
er
ted
b
in
ar
y
h
a
s
a
g
o
o
d
p
er
f
o
r
m
a
n
ce
w
it
h
E
E
R
v
a
lu
e
o
f
9
.
9
% t
h
at
cr
o
s
s
ed
at
th
r
es
h
o
ld
6
0
.
Fig
u
r
e
7.
R
ec
eiv
er
o
p
er
atio
n
ch
ar
ac
ter
is
tic
s
g
r
ap
h
o
f
g
r
a
y
s
ca
le
d
ataset
T
ab
le
2.
C
o
m
p
lete
r
es
u
lt o
f
t
h
e
th
r
ee
d
atasets
D
a
t
a
se
t
F
A
R
(
%)
F
R
R
(
%)
T
h
r
e
sh
o
l
d
G
A
R
(
%)
B
i
n
a
r
y
1
3
.
4
%
1
3
.
0
%
52
8
6
.
8
%
G
r
a
y
sca
l
e
7
.
6
%
7
.
8
%
51
9
2
.
3
%
I
n
v
e
r
t
e
d
B
i
n
a
r
y
1
0
.
2
%
9
.
6
%
52
9
0
.
1
%
T
h
e
w
o
r
s
t
r
esu
l
t
w
e
g
et
is
w
h
e
n
u
s
in
g
b
in
ar
y
i
m
a
g
e
as
d
ataset.
T
h
is
r
esu
lt
m
a
y
b
e
ca
u
s
ed
b
y
th
e
n
o
is
e
a
n
d
d
ata
lo
s
s
in
t
h
e
b
in
ar
y
i
m
a
g
es
th
at
m
a
k
e
i
t
d
if
f
ic
u
lt
f
o
r
th
e
C
NN
to
lear
n
.
W
h
en
b
i
n
ar
y
i
m
a
g
e
s
g
en
er
ated
f
r
o
m
g
r
a
y
s
ca
le
i
m
a
g
es,
it
s
lo
s
s
s
e
v
er
al
f
ea
t
u
r
es
f
r
o
m
o
r
ig
i
n
al
i
m
a
g
es.
I
n
v
er
ted
b
in
ar
y
d
atase
t
h
a
v
e
a
h
ig
h
r
esu
lt
o
f
G
AR
ca
u
s
ed
b
y
it
s
b
lack
b
ac
k
g
r
o
u
n
d
r
ec
o
g
n
ized
as
a
f
ea
tu
r
es
b
y
C
NN.
I
f
w
e
c
o
m
p
ar
e
th
i
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6930
T
E
L
KOM
NI
K
A
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
,
Vo
l.
1
8
,
No
.
4
,
A
u
g
u
s
t
2020:
1
9
3
4
-
1
9
4
1
1940
m
et
h
o
d
w
it
h
o
t
h
er
r
esear
ch
[6
,
7]
,
w
e
ca
n
s
a
y
t
h
at
th
is
m
et
h
o
d
h
av
e
s
e
v
er
al
ad
v
an
tag
e.
F
i
r
s
t,
th
e
m
eth
o
d
t
h
at
w
e
u
s
ed
d
o
esn
’
t
n
ee
d
m
u
ch
o
f
p
r
ep
r
o
ce
s
s
in
g
m
et
h
o
d
an
d
an
o
th
er
s
e
g
m
e
n
tat
io
n
m
e
th
o
d
.
T
h
is
m
ak
e
ti
m
e
t
h
at
n
ee
d
ed
f
o
r
p
r
ep
a
r
in
g
th
e
d
a
taset
f
a
s
ter
th
a
n
th
e
o
th
er
m
et
h
o
d
.
Seco
n
d
,
w
e
u
s
e
h
a
n
d
w
r
i
tin
g
i
n
f
o
r
m
o
f
s
en
te
n
ce
s
,
s
o
it
h
as
m
o
r
e
ch
a
r
ac
ter
is
tics
an
d
f
ea
tu
r
es
t
h
at
ca
n
b
e
ex
tr
ac
ted
an
d
lear
n
b
y
C
NN.
B
u
t
w
e
ca
n
’
t
co
m
p
ar
e
ti
m
e
n
ee
d
ed
f
o
r
tr
ain
in
g
b
ec
a
u
s
e
t
h
e
o
th
er
r
esear
ch
d
o
esn
’
t sp
ec
i
f
y
it.
4.
CO
NCLU
SI
O
N
W
e
u
s
ed
tr
an
s
f
er
lear
n
in
g
to
id
en
ti
f
ied
w
r
iter
b
ased
o
n
th
e
ir
h
an
d
w
r
iti
n
g
u
s
i
n
g
p
r
e
-
tr
ain
ed
C
NN.
W
e
tr
ain
ed
m
o
d
el
u
s
i
n
g
I
A
M
h
an
d
w
r
it
in
g
d
ataset
co
m
p
r
is
i
n
g
o
f
1
0
0
class
e
s
o
f
w
r
iter
.
B
ased
o
n
th
e
r
es
u
l
t
o
f
tr
ain
in
g
a
n
d
tes
tin
g
,
p
r
e
-
tr
ain
ed
m
o
d
el
VGG1
9
h
a
s
t
h
e
b
est
p
er
f
o
r
m
a
n
ce
w
h
e
n
u
s
i
n
g
g
r
a
y
s
ca
le
i
m
a
g
es
th
an
u
s
i
n
g
b
in
ar
y
o
r
i
n
v
er
ted
b
in
ar
y
i
m
a
g
es.
O
u
r
f
in
d
i
n
g
al
s
o
in
d
icate
s
t
h
at
h
a
n
d
w
r
iti
n
g
i
m
a
g
e
i
n
s
en
ten
ce
s
f
o
r
m
ca
n
b
e
u
s
ed
d
ir
ec
tl
y
w
i
th
o
u
t
a
n
o
th
er
s
eg
m
e
n
tatio
n
o
r
f
ea
t
u
r
e
ex
tr
ac
tio
n
m
et
h
o
d
n
ee
d
ed
.
T
h
e
d
is
ad
v
an
tag
e
o
f
th
i
s
r
ese
ar
ch
is
th
e
ti
m
e
it
to
o
k
to
f
i
n
i
s
h
tr
ai
n
in
g
p
r
o
ce
s
s
.
T
r
ain
in
g
p
r
o
ce
s
s
to
o
k
al
m
o
s
t
1
0
h
o
u
r
s
to
co
m
p
lete
1
0
0
ep
o
ch
s
.
RE
F
E
R
E
NC
E
S
[1
]
K.
Ch
a
u
d
h
a
ri
a
n
d
A
.
T
h
a
k
k
a
r,
“
S
u
rv
e
y
o
n
h
a
n
d
w
rit
in
g
-
b
a
se
d
p
e
r
so
n
a
li
ty
trait
id
e
n
ti
f
ica
ti
o
n
,
”
Exp
e
rt
S
y
ste
ms
wit
h
Ap
p
li
c
a
ti
o
n
s
,
v
o
l
.
1
2
4
,
p
p
.
2
8
2
–
3
0
8
,
J
u
n
2
0
1
9
.
d
o
i:
1
0
.
1
0
1
6
/j
.
e
sw
a
.
2
0
1
9
.
0
1
.
0
2
8
.
[2
]
A
.
Kriz
h
e
v
s
k
y
,
I.
S
u
tsk
e
v
e
r,
a
n
d
G
.
E.
Hin
to
n
,
“
Im
a
g
e
N
e
t
c
las
si
f
ic
a
ti
o
n
w
it
h
d
e
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s,”
Co
mm
u
n
ica
ti
o
n
s
ACM
,
v
o
l.
6
0
,
n
o
.
6
,
p
p
.
8
4
–
9
0
,
M
a
y
2
0
17
.
d
o
i:
1
0
.
1
1
4
5
/3
0
6
5
3
8
6
.
[3
]
K.
H.
Jin
,
M
.
T
.
M
c
Ca
n
n
,
E.
F
ro
u
ste
y
,
a
n
d
M
.
Un
se
r,
“
De
e
p
Co
n
v
o
l
u
ti
o
n
a
l
Ne
u
ra
l
Ne
two
rk
f
o
r
In
v
e
rse
P
r
o
b
lem
s
in
Im
a
g
in
g
,
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
Im
a
g
e
Pr
o
c
e
ss
in
g
,
v
o
l.
2
6
,
n
o
.
9
,
p
p
.
4
5
0
9
–
4
5
2
2
,
S
e
p
2
0
1
7
.
d
o
i:
1
0
.
1
1
0
9
/T
IP
.
2
0
1
7
.
2
7
1
3
0
9
9
.
[4
]
J.
S
c
h
m
id
h
u
b
e
r,
“
De
e
p
L
e
a
rn
in
g
i
n
Ne
u
ra
l
Ne
tw
o
rk
s:
A
n
Ov
e
r
v
ie
w
,
”
Ne
u
ra
l
Ne
two
rk
s
,
v
o
l.
6
1
,
p
p
.
8
5
–
1
1
7
,
Ja
n
2
0
1
5
.
d
o
i:
1
0
.
1
0
1
6
/j
.
n
e
u
n
e
t.
2
0
1
4
.
0
9
.
0
0
3
.
[5
]
C
.
K
.
D
e
w
a
,
A
.
L
.
F
a
d
h
i
l
a
h
,
a
n
d
A
.
A
f
i
a
h
a
y
a
t
i
,
“
C
o
n
v
o
l
u
t
i
o
n
a
l
N
e
u
r
a
l
N
e
tw
o
rk
s
f
o
r
H
a
n
d
w
r
i
t
t
e
n
J
a
v
a
n
e
s
e
C
h
a
r
a
c
t
e
r
R
e
c
o
g
n
i
t
i
o
n
,
”
I
J
C
C
S
(
I
n
d
o
n
e
s
i
a
n
J
o
u
r
n
a
l
o
f
C
o
m
p
u
t
i
n
g
a
n
d
C
y
b
e
r
n
e
t
i
c
s
S
y
s
t
e
ms
)
,
v
o
l
.
1
2
,
n
o
.
1
,
p
p
.
8
3
-
94
,
J
a
n
2
0
1
8
.
d
o
i
:
1
0
.
2
2
1
4
6
/
i
j
c
c
s
.
3
1
1
4
4
.
[6
]
B
.
V
.
D
h
a
n
d
r
a
a
n
d
M
.
B.
V
i
j
a
y
a
l
a
x
m
i
,
“
A
N
o
v
e
l
Ap
p
r
o
a
c
h
to
Te
x
t
D
e
p
e
n
d
e
n
t
Wr
i
t
e
r
Id
e
n
ti
fi
c
a
t
io
n
o
f
K
a
n
n
a
d
a
H
a
n
d
wr
i
t
in
g
,
”
P
r
o
c
e
d
ia
Co
mp
u
te
r
S
c
i
e
n
c
e
,
v
o
l
.
4
9
,
p
p
.
3
3
–
4
1
,
2
0
1
5
.
d
o
i:
1
0
.
1
0
1
6
/
j.
p
r
o
c
s
.
2
0
1
5
.
0
4
.
2
2
4
.
[7
]
L
.
X
in
g
a
n
d
Y.
Qia
o
,
“
De
e
p
Wr
it
e
r:
A
M
u
lt
i
-
S
tr
e
a
m
D
e
e
p
CN
N
f
o
r
T
e
x
t
-
in
d
e
p
e
n
d
e
n
t
W
rit
e
r
Id
e
n
ti
f
ica
ti
o
n
,
”
a
rX
iv
:1
6
0
6
.
0
6
4
7
2
,
Ju
n
2
0
1
6
.
[8
]
U.
V
.
M
a
rti
a
n
d
H.
Bu
n
k
e
,
“
T
h
e
IA
M
-
d
a
tab
a
se
:
a
n
En
g
li
sh
se
n
ten
c
e
d
a
tab
a
se
f
o
r
o
ff
li
n
e
h
a
n
d
w
rit
in
g
re
c
o
g
n
it
io
n
,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
n
Do
c
u
me
n
t
An
a
lys
is
a
n
d
Rec
o
g
n
i
ti
o
n
,
v
o
l.
5
,
n
o
.
1
,
p
p
.
3
9
–
4
6
,
N
o
v
2
0
0
2
.
d
o
i:
1
0
.
1
0
0
7
/s
1
0
0
3
2
0
2
0
0
0
7
1
.
[9
]
A
.
Re
h
m
a
n
,
S
.
Na
z
,
M
.
I.
Ra
z
z
a
k
,
a
n
d
I.
A
.
Ha
m
e
e
d
,
“
A
u
to
m
a
ti
c
V
isu
a
l
F
e
a
t
u
re
s
f
o
r
W
rit
e
r
Id
e
n
ti
fica
ti
o
n
:
A
De
e
p
L
e
a
rn
in
g
A
p
p
ro
a
c
h
,
”
IEE
E
Acc
e
ss
,
v
o
l.
7
,
p
p
.
1
7
1
4
9
–
1
7
1
5
7
,
2
0
1
9
.
d
o
i
:
1
0
.
1
1
0
9
/A
CCES
S
.
2
0
1
8
.
2
8
9
0
8
1
0
.
[1
0
]
P
.
Hrid
a
y
a
m
i,
“
F
ish
S
p
e
c
ies
R
e
c
o
g
n
it
io
n
Us
in
g
V
GG
1
6
De
e
p
Co
n
v
o
l
u
ti
o
n
a
l
Ne
u
ra
l
Ne
tw
o
rk
,
”
J
o
u
rn
a
l
o
f
Co
mp
u
t
in
g
S
c
ien
c
e
a
n
d
E
n
g
i
n
e
e
rin
g
,
v
o
l.
1
3
,
n
o
.
3
,
p
p
.
1
2
4
-
1
3
0
,
2
0
1
9
.
[1
1
]
D.
Ha
n
,
Q.
L
iu
,
a
n
d
W
.
F
a
n
,
“
A
n
e
w
i
m
a
g
e
c
las
sif
i
c
a
ti
o
n
m
e
t
h
o
d
u
si
n
g
CNN
tran
sf
e
r
lea
rn
in
g
a
n
d
w
e
b
d
a
ta
a
u
g
m
e
n
tatio
n
,
”
Exp
e
rt
S
y
ste
ms
wit
h
A
p
p
l
ica
ti
o
n
s
,
v
o
l.
9
5
,
p
p
.
4
3
–
5
6
,
A
p
r
2
0
1
8
.
d
o
i:
1
0
.
1
0
1
6
/j
.
e
sw
a
.
2
0
1
7
.
1
1
.
0
2
8
.
[1
2
]
O.
Ru
ss
a
k
o
v
sk
y
e
t
a
l.
,
“
Im
a
g
e
N
e
t
L
a
rg
e
S
c
a
le V
isu
a
l
Re
c
o
g
n
it
i
o
n
Ch
a
ll
e
n
g
e
,
”
a
rX
iv
:1
4
0
9
.
0
5
7
5
,
S
e
p
.
2
0
1
4
.
[1
3
]
Y.
Yo
u
,
Z.
Zh
a
n
g
,
C.
-
J.
Hs
ieh
,
J.
De
m
m
e
l,
a
n
d
K.
Ke
u
tze
r,
“
I
m
a
g
e
Ne
t
T
r
a
in
in
g
in
M
in
u
tes
,
”
a
rX
iv
:1
7
0
9
.
0
5
0
1
1
,
Ja
n
2
0
1
8
.
[1
4
]
S
.
H.
S
.
Ba
sh
a
,
S
.
R.
D
u
b
e
y
,
V
.
P
u
lab
a
ig
a
ri,
a
n
d
S
.
M
u
k
h
e
rjee
,
“
I
m
p
a
c
t
o
f
F
u
ll
y
Co
n
n
e
c
ted
L
a
y
e
rs
o
n
P
e
rf
o
rm
a
n
c
e
o
f
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
t
w
o
rk
s
f
o
r
I
m
a
g
e
Clas
sifica
ti
o
n
,
”
Ne
u
ro
c
o
mp
u
ti
n
g
,
v
o
l.
3
7
8
,
p
.
1
1
2
-
1
1
9
,
Oc
t
2
0
1
9
.
d
o
i:
1
0
.
1
0
1
6
/
j.
n
e
u
c
o
m
.
2
0
1
9
.
1
0
.
0
0
8
.
[1
5
]
S
.
G
iri
a
n
d
B.
Jo
sh
i,
“
T
ra
n
sf
e
r
L
e
a
rn
in
g
Ba
s
e
d
I
m
a
g
e
V
isu
a
li
z
a
ti
o
n
Us
in
g
CNN
,
”
I
n
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
A
rtif
icia
l
I
n
telleg
e
n
c
e
a
n
d
A
p
p
li
c
a
ti
o
n
,
v
o
l
.
1
0
,
n
o
.
4
,
p
p
.
4
7
–
5
5
,
J
u
l
2
0
1
9
.
d
o
i:
1
0
.
5
1
2
1
/
ij
a
ia.2
0
1
9
.
1
0
4
0
4
.
[1
6
]
H.
Ye
,
H.
Ha
n
,
L
.
Zh
u
,
a
n
d
Q.
Du
a
n
,
“
V
e
g
e
tab
le
P
e
st
I
m
a
g
e
Re
c
o
g
n
it
io
n
M
e
th
o
d
Ba
s
e
d
o
n
Im
p
ro
v
e
d
V
GG
Co
n
v
o
lu
ti
o
n
Ne
u
ra
l
Ne
tw
o
rk
,
”
J
o
u
rn
a
l
o
f
P
h
y
sic
s
Co
n
fer
e
n
c
e
S
e
rie
s
,
v
o
l.
1
2
3
7
,
J
u
n
2
0
1
9
.
d
o
i:
1
0
.
1
0
8
8
/1
7
4
2
-
6
5
9
6
/
1
2
3
7
/
3
/0
3
2
0
1
8
.
[1
7
]
T
.
S
in
h
a
,
B
.
V
e
rm
a
,
a
n
d
A
.
Ha
id
a
r,
“
Op
ti
m
iza
ti
o
n
o
f
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
p
a
ra
m
e
ters
f
o
r
i
m
a
g
e
c
las
si
f
ica
ti
o
n
,
”
2
0
1
7
IEE
E
S
y
mp
o
siu
m
S
e
rie
s
o
n
C
o
mp
u
ta
ti
o
n
a
l
I
n
telli
g
e
n
c
e
(
S
S
CI)
,
Ho
n
o
l
u
l
u
,
HI,
p
p
.
1
–
7
,
2
0
1
7
.
d
o
i:
1
0
.
1
1
0
9
/S
S
CI.
2
0
1
7
.
8
2
8
5
3
3
8
.
[1
8
]
M
.
L
in
,
Q.
C
h
e
n
,
a
n
d
S
.
Ya
n
,
“
Ne
tw
o
rk
in
Ne
tw
o
rk
,
”
a
rX
iv
:1
3
1
2
.
4
4
0
0
,
M
a
r
2
0
1
4
.
[1
9
]
L
.
A
l
z
u
b
a
id
i,
H.
A
.
A
l
w
z
w
a
z
y
,
a
n
d
Z.
M
.
A
rk
a
h
,
“
Co
n
v
o
l
u
ti
o
n
a
l
Ne
u
ra
l
Ne
tw
o
rk
w
i
th
G
lo
b
a
l
A
v
e
ra
g
e
P
o
o
li
n
g
f
o
r
Im
a
g
e
Cl
a
ss
i
f
ica
ti
o
n
,
”
In
ter
n
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
o
n
El
e
c
trica
l,
C
o
mm
u
n
ic
a
ti
o
n
,
El
e
c
tro
n
ics
,
In
str
u
me
n
ta
t
io
n
a
n
d
Co
mp
u
t
in
g
(
ICECE
IC)
,
2
0
2
0
.
[2
0
]
Jin
g
Ya
n
g
a
n
d
G
u
a
n
c
i
Ya
n
g
,
“
M
o
d
if
ied
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
tw
o
rk
Ba
s
e
d
o
n
Dro
p
o
u
t
a
n
d
th
e
S
to
c
h
a
stic
G
ra
d
ien
t
De
sc
e
n
t
Op
ti
m
iz
e
r,
”
Al
g
o
rith
ms
,
v
o
l
.
1
1
,
n
o
.
3
,
M
a
r
2
0
1
8
.
d
o
i
:
1
0
.
3
3
9
0
/a1
1
0
3
0
0
2
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
K
A
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l
C
o
n
tr
o
l
Ha
n
d
w
r
itin
g
id
en
tifi
ca
tio
n
u
s
i
n
g
d
ee
p
co
n
v
o
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k
meth
o
d
(
Oka
S
u
d
a
n
a
)
1941
[2
1
]
N.
S
riv
a
sta
v
a
,
G
.
Hin
to
n
,
A
.
Kriz
h
e
v
sk
y
,
I.
S
u
tsk
e
v
e
r,
a
n
d
R.
S
a
la
k
h
u
td
i
n
o
v
,
“
Dro
p
o
u
t:
A
S
i
m
p
le
W
a
y
to
P
re
v
e
n
t
Ne
u
ra
l
Ne
t
w
o
rk
s
f
ro
m
Ov
e
rfi
tt
i
n
g
,
”
T
h
e
J
o
u
rn
a
l
o
f
M
a
c
h
in
e
L
e
a
rn
in
g
Res
e
a
rc
h
(
J
M
L
R)
,
v
o
l.
15
,
n
o
.
56
,
p
p
.
1
9
2
9
−
1
9
5
8
,
2
0
1
4
.
[2
2
]
N.
S
.
Ke
sk
a
r
a
n
d
R.
S
o
c
h
e
r,
“
I
m
p
ro
v
in
g
Ge
n
e
ra
li
z
a
ti
o
n
P
e
rf
o
rm
a
n
c
e
b
y
S
w
it
c
h
in
g
f
ro
m
Ad
a
m
to
S
G
D,”
a
rX
iv
:1
7
1
2
.
0
7
6
2
8
,
De
c
2
0
1
7
.
[2
3
]
A
.
C.
W
il
so
n
,
R.
Ro
e
lo
f
s,
M
.
S
te
rn
,
N.
S
re
b
r
o
,
a
n
d
B.
Re
c
h
t
,
“
T
h
e
M
a
rg
in
a
l
V
a
lu
e
o
f
A
d
a
p
ti
v
e
G
r
a
d
ien
t
M
e
th
o
d
s
in
M
a
c
h
i
n
e
L
e
a
rn
in
g
,
”
a
rX
iv
:1
7
0
5
.
0
8
2
9
2
,
M
a
y
2
0
1
8
.
[2
4
]
T
.
Ca
rn
e
iro
,
R.
V.
M
e
d
e
ir
o
s
Da
No
b
re
g
a
,
T
.
Ne
p
o
m
u
c
e
n
o
,
G
.
B.
Bian
,
V.
H.
C
.
De
A
lb
u
q
u
e
rq
u
e
,
a
n
d
P
.
P
.
R
.
F
il
h
o
,
“
P
e
rf
o
rm
a
n
c
e
A
n
a
l
y
sis
o
f
G
o
o
g
le
Co
lab
o
ra
to
ry
a
s
a
T
o
o
l
f
o
r
A
c
c
e
l
e
ra
ti
n
g
De
e
p
L
e
a
rn
in
g
A
p
p
li
c
a
ti
o
n
s,
”
IEE
E
Acc
e
ss
,
v
o
l.
6
,
p
p
.
6
1
6
7
7
–
6
1
6
8
5
,
2
0
1
8
.
d
o
i:
1
0
.
1
1
0
9
/A
CCES
S
.
2
0
1
8
.
2
8
7
4
7
6
7
.
[2
5
]
C
.
L
v
,
T
.
Z
h
a
n
g
,
C
.
L
i
u
,
a
n
d
S
c
h
o
o
l
o
f
A
u
t
o
m
a
t
i
o
n
,
Be
i
j
i
n
g
I
n
s
t
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
,
“
A
n
Im
p
r
o
v
e
d
O
ts
u
’
s
T
h
r
e
s
h
o
l
d
i
n
g
A
lg
o
r
i
t
h
m
o
n
G
e
s
t
u
r
e
S
e
g
m
e
n
t
a
t
i
o
n
,
”
J
o
u
r
n
a
l
o
f
A
d
v
a
n
c
e
d
C
o
m
p
u
t
a
t
i
o
n
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
I
n
t
e
l
l
i
g
e
n
t
I
n
f
o
r
m
a
t
i
c
s
,
v
o
l
.
2
1
,
n
o
.
2
,
p
p
.
2
4
7
–
2
5
0
,
M
a
r
2
0
1
7
.
d
o
i
:
1
0
.
2
0
9
6
5
/
j
a
c
i
i
i
.
2
0
1
7
.
p
0
2
4
7
.
[2
6
]
R
.
Y
a
m
a
s
h
i
t
a
,
M
.
N
i
s
h
i
o
,
R
.
K
.
G.
D
o
,
a
n
d
K
.
T
o
g
a
s
h
i
,
“
C
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
tw
o
r
k
s
:
a
n
o
v
e
rv
i
e
w
a
n
d
a
p
p
l
i
c
a
t
i
o
n
i
n
r
a
d
i
o
l
o
g
y
,
”
I
n
s
i
g
h
t
s
I
m
a
g
i
n
g
,
v
o
l
.
9
,
n
o
.
4
,
p
p
.
6
1
1
–
6
2
9
,
A
u
g
2
0
1
8
.
d
o
i
:
1
0
.
1
0
0
7
/
s
1
3
2
4
4
-
018
-
0
6
3
9
-
9.
[2
7
]
NK
A
W
ird
ian
i,
T
.
L
a
tt
i
f
ia,
I.
K.
S
u
p
a
d
m
a
,
B.
J.
K.
M
a
h
a
r,
D.
A
.
N.
T
a
ra
d
h
it
a
,
“
Re
a
l
-
T
i
m
e
F
a
c
e
R
e
c
o
g
n
it
io
n
w
it
h
Ei
g
e
n
f
a
c
e
M
e
th
o
d
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
Im
a
g
e
,
Gr
a
p
h
ics
a
n
d
S
ig
n
a
l
Pr
o
c
e
ss
in
g
(
IJ
IGS
P)
,
v
o
l.
1
1
,
n
o
.
1
1
,
p
p
.
1
–
9
,
N
o
v
2
0
1
9
.
d
o
i:
1
0
.
5
8
1
5
/i
ji
g
sp
.
2
0
1
9
.
1
1
.
0
1
.
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
A.
A.
K
o
m
p
ia
n
g
O
k
a
S
u
d
a
n
a
,
c
u
rre
n
tl
y
a
L
e
c
tu
re
r
in
th
e
In
f
o
r
m
a
ti
o
n
T
e
c
h
n
o
l
o
g
y
S
tu
d
y
P
r
o
g
ra
m
,
Ud
a
y
a
n
a
Un
iv
e
rsit
y
,
Ba
li
.
A
re
a
s
o
f
in
tere
st:
I
m
a
g
e
P
r
o
c
e
ss
in
g
,
In
f
o
rm
a
ti
o
n
S
y
st
e
m
s,
IT
I
m
p
le
m
e
n
tatio
n
in
Cu
lt
u
re
a
n
d
Re
li
g
io
n
.
S
1
in
In
f
o
rm
a
ti
c
s
De
p
a
rt
m
e
n
t
o
f
IT
S
S
u
ra
b
a
y
a
,
M
a
ste
rs
in
El
e
c
tri
c
a
l
En
g
in
e
e
rin
g
UG
M
Yo
g
y
a
k
a
rta,
c
u
rre
n
tl
y
in
th
e
p
ro
c
e
ss
o
f
stu
d
y
in
g
S
3
a
t
Ud
a
y
a
n
a
Un
iv
e
rsity
.
I
W
a
y
a
n
G
u
n
a
y
a
w
a
s
b
o
r
n
o
n
S
e
p
t
e
m
b
e
r
3
,
1
9
9
8
i
n
B
a
n
g
l
i
,
B
a
l
i
,
I
n
d
o
n
e
s
i
a
.
H
e
i
s
a
s
t
u
d
e
n
t
i
n
D
e
p
a
r
tm
e
n
t
o
f
I
n
f
o
rm
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
U
d
a
y
a
n
a
U
n
i
v
e
r
s
i
ty
,
B
a
li
,
I
n
d
o
n
e
s
i
a
.
H
i
s
re
s
e
a
r
c
h
i
n
t
e
r
e
s
t
s
a
r
e
i
n
im
a
g
e
p
r
o
c
e
s
s
i
n
g
,
c
o
m
p
u
t
e
r
v
i
s
i
o
n
,
d
a
t
a
sc
i
e
n
c
e
,
a
n
d
m
a
c
h
i
n
e
le
a
r
n
i
n
g
.
I
K
e
tu
t
G
e
d
e
Da
r
m
a
P
u
tr
a
c
u
r
re
n
tl
y
a
L
e
c
tu
re
r
in
De
p
a
rt
m
e
n
t
o
f
El
e
c
tri
c
a
l
En
g
in
e
e
rin
g
a
n
d
In
f
o
rm
a
ti
o
n
T
e
c
h
n
o
l
o
g
y
,
Ud
a
y
a
n
a
Un
iv
e
rsit
y
Ba
li
,
In
d
o
n
e
sia
.
He
re
c
e
iv
e
d
h
is
S
.
Ko
m
d
e
g
re
e
in
In
f
o
rm
a
ti
c
s
En
g
in
e
e
rin
g
f
ro
m
In
stit
u
te
o
f
S
e
p
u
l
u
h
No
v
e
m
b
e
r
T
e
c
h
n
o
lo
g
y
S
u
ra
b
a
y
a
,
In
d
o
n
e
sia
in
1
9
9
7
.
He
re
c
e
i
v
e
d
h
is
m
a
ste
r
d
e
g
re
e
o
n
In
f
o
rm
a
ti
c
s
a
n
d
Co
m
p
u
ter
En
g
in
e
e
rin
g
f
ro
m
El
e
c
tri
c
a
l
En
g
in
e
e
rin
g
De
p
a
rt
m
e
n
t,
G
a
d
jah
M
a
d
a
Un
iv
e
rsit
y
,
In
d
o
n
e
sia
in
2
0
0
0
a
n
d
a
c
h
iev
e
d
h
is
d
o
c
to
ra
t
e
d
e
g
re
e
o
n
In
f
o
rm
a
ti
c
s
a
n
d
Co
m
p
u
ter
En
g
in
e
e
rin
g
f
ro
m
El
e
c
tri
c
a
l
En
g
in
e
e
rin
g
De
p
a
rtme
n
t,
G
a
d
jah
M
a
d
a
Un
iv
e
rsity
,
In
d
o
n
e
sia
in
2
0
0
6
.
His
re
se
a
rc
h
in
tere
sts a
re
in
b
io
m
e
tri
c
s,
im
a
g
e
p
ro
c
e
ss
in
g
,
d
a
ta m
in
in
g
,
a
n
d
so
f
t
c
o
m
p
u
ti
n
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.