T
E
L
KO
M
N
I
KA
T
e
lec
om
m
u
n
icat
ion
,
Com
p
u
t
i
n
g,
E
lec
t
r
on
ics
an
d
Cont
r
ol
Vol.
18
,
No.
3
,
J
une
2020
,
pp.
1
376
~
1
381
I
S
S
N:
1693
-
6930,
a
c
c
r
e
dit
e
d
F
ir
s
t
G
r
a
de
by
Ke
me
nr
is
tekdikti
,
De
c
r
e
e
No:
21/E
/KP
T
/2018
DO
I
:
10.
12928/
T
E
L
KO
M
NI
KA
.
v18i3.
14840
1376
Jou
r
n
al
h
omepage
:
ht
tp:
//
jour
nal.
uad
.
ac
.
id/
index
.
php/T
E
L
K
OM
N
I
K
A
C
on
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
or
k
f
or
m
ai
z
e
l
e
a
f
d
is
e
as
e
image
c
la
ssi
f
ic
at
io
n
M
oh
am
m
ad
S
yar
ief
,
Wah
yu
d
i
S
e
t
iawan
In
fo
rma
t
i
c
s
D
ep
ar
t
men
t
,
U
n
i
v
er
s
i
t
y
o
f
T
ru
n
o
j
o
y
o
Ma
d
u
ra,
In
d
o
n
es
i
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
Aug
17,
2019
R
e
vis
e
d
J
a
n
4,
2020
Ac
c
e
pted
F
e
b
26,
2020
T
h
i
s
art
i
cl
e
d
i
s
cu
s
s
e
s
t
h
e
mai
ze
l
eaf
d
i
s
ea
s
e
i
mag
e
cl
a
s
s
i
fi
ca
t
i
o
n
.
T
h
e
ex
p
eri
me
n
t
a
l
i
mag
e
s
co
n
s
i
s
t
o
f
2
0
0
i
mag
e
s
w
i
t
h
4
cl
as
s
e
s
:
h
ea
l
t
h
y
,
cerco
s
p
o
ra,
co
mm
o
n
ru
s
t
a
n
d
n
o
rt
h
ern
l
eaf
b
l
i
g
h
t
.
T
h
er
e
are
2
s
t
ep
s
:
fea
t
u
re
e
x
t
r
a
c
t
i
o
n
a
n
d
c
l
a
s
s
i
f
i
c
a
t
i
o
n
.
F
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
o
b
t
a
i
n
s
f
e
a
t
u
r
e
s
a
u
t
o
m
a
t
i
c
a
l
l
y
u
s
i
n
g
c
o
n
v
o
l
u
t
i
o
n
a
l
n
eu
ra
l
n
et
w
o
r
k
(CN
N
).
Sev
e
n
CN
N
mo
d
e
l
s
w
ere
t
es
t
ed
i
.
e
A
l
e
x
N
e
t
,
v
i
r
t
u
a
l
g
eo
me
t
ry
g
r
o
u
p
(V
G
G
)
1
6
,
V
G
G
1
9
,
G
o
o
g
l
e
N
et
,
In
cep
t
i
o
n
-
V
3
,
res
i
d
u
al
n
e
t
w
o
rk
5
0
(Res
N
et
5
0
)
an
d
Res
N
et
1
0
1
.
W
h
i
l
e
t
h
e
cl
as
s
i
f
i
cat
i
o
n
u
s
i
n
g
mach
i
n
e
l
earn
i
n
g
met
h
o
d
s
i
n
cl
u
d
e
k
-
N
eare
s
t
n
ei
g
h
b
o
r,
d
eci
s
i
o
n
t
re
e
an
d
s
u
p
p
o
r
t
v
ect
o
r
mach
i
n
e
.
Bas
ed
o
n
t
h
e
t
e
s
t
i
n
g
res
u
l
t
s
,
t
h
e
b
e
s
t
c
l
as
s
i
f
i
cat
i
o
n
w
as
A
l
e
x
N
e
t
a
n
d
s
u
p
p
o
rt
v
ect
o
r
mach
i
n
e
w
i
t
h
accu
racy
,
s
e
n
s
i
t
i
v
i
t
y
,
s
p
ec
i
fi
c
i
t
y
o
f
9
3
.
5
%
,
9
5
.
0
8
%
,
an
d
9
3
%
,
res
p
ec
t
i
v
el
y
.
K
e
y
w
o
r
d
s
:
Ale
xNe
t
C
las
s
if
ica
ti
on
C
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
k
-
ne
a
r
e
s
t
ne
ighbor
M
a
ize
lea
f
im
a
ge
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e
.
C
or
r
e
s
pon
din
g
A
u
th
or
:
W
a
hyudi
S
e
ti
a
wa
n
,
I
nf
or
matics
De
pa
r
tm
e
nt
,
Unive
r
s
it
y
of
T
r
unojoyo
M
a
dur
a
,
R
a
ya
T
e
lang
S
t.
,
P
e
r
umaha
n
T
e
lang
I
nda
,
T
e
lang,
Ka
mal,
B
a
ngka
lan,
J
a
wa
T
im
ur
69162
,
I
ndone
s
ia
.
E
mail:
ws
e
ti
a
wa
n@tr
unojoyo.
a
c
.
id
1.
I
NT
RODU
C
T
I
ON
C
onvolut
ional
ne
ur
a
l
ne
two
r
k
(
C
NN
)
is
a
de
ve
lop
ment
of
the
a
r
ti
f
icia
l
ne
ur
a
l
ne
twor
k
that
c
ons
is
ts
of
tens
to
hundr
e
ds
o
f
laye
r
s
[
1]
.
C
NN
is
a
meth
od
in
de
e
p
lea
r
ning
that
c
a
n
pe
r
f
or
m
va
r
ious
tas
ks
s
uc
h
a
s
im
a
ge
c
las
s
if
ica
ti
on
[2
,
3]
,
s
e
gmenta
ti
on
[4
,
5]
,
r
e
c
ognit
ion
[6
,
7]
,
a
nd
objec
ts
de
tec
ti
on
[8
,
9]
.
C
NN
tec
hnology
ha
s
gr
own
wide
ly
including
f
ields
o
f
medic
a
l
im
a
ge
[
10
,
11]
,
a
utonom
ous
dr
ive
r
s
[
12
,
13
]
,
r
oboti
c
s
[
14
,
15]
,
a
nd
a
gr
icultu
r
a
l
im
a
ge
[
16]
.
M
a
ny
im
a
ge
s
tudi
e
s
ha
ve
be
e
n
c
a
r
r
ied
out,
s
uc
h
a
s
dis
e
a
s
e
c
las
s
if
ica
ti
on
in
15
f
ood
c
r
ops
us
i
ng
5
c
onvolut
i
ona
l
laye
r
s
[
17]
,
c
las
s
if
ica
ti
on
o
f
d
is
e
a
s
e
s
in
9
c
l
a
s
s
plant
im
a
ge
s
us
ing
googleN
e
t
[
18]
.
M
oha
nty
e
t
a
l
.
c
las
s
if
ied
14
types
of
f
ood
c
r
ops
,
including
maiz
e
.
T
he
r
e
wa
s
26
c
las
s
of
dis
e
a
s
e
s
.
T
he
tes
ti
ng
us
e
d
im
a
ge
s
in
va
s
t
number
s
,
i
.
e
54,
306
im
a
ge
s
.
De
e
p
lea
r
ning
c
onv
e
nti
ona
l
ne
ur
a
l
ne
twor
k
wi
th
two
a
r
c
hit
e
c
tur
e
(
Ale
xNe
t
a
nd
Google
Ne
t)
wa
s
us
e
d
f
or
c
las
s
if
ica
ti
on.
T
he
c
las
s
if
ica
ti
on
r
e
s
ult
s
s
howe
d
a
n
a
c
c
ur
a
c
y
of
3
1.
4%
[
19]
.
I
n
thi
s
s
tudy,
c
las
s
if
ica
ti
o
n
wa
s
c
a
r
r
ied
out
to
de
tec
t
dis
e
a
s
e
s
in
maiz
e
lea
ve
im
a
ge
s
us
ing
C
NN
.
One
of
the
pr
e
vious
s
tudi
e
s
that
c
a
r
r
ied
out
d
is
e
a
s
e
s
c
las
s
if
ica
ti
on
of
maiz
e
lea
ve
s
us
ing
C
NN
wa
s
S
ibi
ya
&
S
umbwa
nya
mbe
[
20]
.
T
he
y
us
ing
3
c
las
s
e
s
of
dis
e
a
s
e
c
la
s
s
if
ica
ti
on:
nor
ther
n
lea
f
b
l
ight
,
c
om
mon
r
us
t,
a
nd
c
e
r
c
os
por
a
.
C
NN
a
r
c
hit
e
c
tur
e
us
e
d
wa
s
not
e
xplaine
d
in
de
tail
,
but
it
only
mentioned
us
ing
5
0
hidden
laye
r
s
c
ons
is
ti
ng
of
c
onvolut
ion
laye
r
s
with
f
il
te
r
ke
r
ne
ls
that
ha
ve
a
media
n
of
24,
r
e
c
ti
f
ied
l
in
e
a
r
unit
s
(
R
e
L
U)
a
nd
pooli
ng
la
ye
r
s
.
One
hundr
e
d
im
a
ge
s
p
e
r
c
las
s
wa
s
us
e
d
with
a
r
a
ti
o
of
70
%
f
or
tr
a
ini
ng
a
nd
30%
f
or
tes
ti
ng.
T
he
tes
ti
ng
r
e
s
ult
s
s
howe
d
a
n
a
c
c
ur
a
c
y
of
92.
85
%
[
20]
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
C
onv
olut
ional
ne
ur
al
ne
tw
or
k
for
maiz
e
leaf
dis
e
as
e
image
c
las
s
if
ication
(
M
ohamm
ad
Sy
ar
ief
)
1377
Z
ha
ng
c
las
s
if
ied
8
dis
e
a
s
e
s
in
maiz
e
lea
f
im
a
ge
s
:
s
outher
n
lea
f
bli
ght
,
br
own
s
pot,
c
ur
vular
ia
lea
f
s
pot,
r
us
t,
dwa
r
f
mos
a
ic,
gr
a
y
lea
f
s
pot,
r
ound
s
pot
,
a
nd
nor
ther
n
lea
f
bli
ght
[
18
]
.
C
NN
a
r
c
hit
e
c
tur
e
u
s
e
d
wa
s
googleN
e
t
or
I
nc
e
pti
on
-
V1.
T
he
e
xpe
r
i
ments
we
r
e
c
onduc
ted
us
ing
3,
672
im
a
ge
s
,
80
%
f
or
t
r
a
ini
ng
a
nd
20%
f
or
tes
ti
ng.
C
las
s
if
ica
ti
on
r
e
s
ult
s
s
howe
d
a
n
a
c
c
ur
a
c
y
of
98.
9%
[
18]
.
Hida
y
a
t
c
las
s
if
ied
thr
e
e
dis
e
a
s
e
s
in
maiz
e
lea
f
im
a
ge
s
:
c
omm
on
r
us
t,
c
e
r
c
os
por
a
,
a
nd
nor
ther
n
lea
f
bli
ght
[
21]
.
T
he
e
xpe
r
im
e
nts
us
e
d
300
maiz
e
lea
f
im
a
ge
s
.
Ave
r
a
ge
a
c
c
ur
a
c
y
r
e
s
ult
wa
s
93
.
67%
[
21]
.
B
oth
types
of
r
e
s
e
a
r
c
h,
S
ibi
ya
&
S
u
mbwa
nya
mbe
[
20]
a
nd
Hida
ya
t
e
t
a
l.
[2
1
]
,
only
e
xplaine
d
the
type
of
C
NN
laye
r
s
,
but
the
nu
mber
of
e
a
c
h
la
ye
r
type
a
nd
de
tailed
pa
r
a
mete
r
s
we
r
e
no
t
e
xplain
e
d,
while
Z
ha
ng’
s
[
18
]
us
e
d
e
xis
ti
ng
C
NN
a
r
c
hit
e
c
tur
e
,
i.
e
Google
Ne
t
that
c
ons
is
ts
of
177
laye
r
s
.
T
he
r
e
wa
s
a
nove
lt
y
in
thi
s
s
tudy.
F
ir
s
t,
us
e
of
7
C
NN
a
r
c
hit
e
c
tur
e
s
:
Ale
xNe
t
[
22]
,
VG
G16
,
VG
G19
[
23
]
,
Google
L
e
t
[
23]
,
I
nc
e
pti
on
-
V3
[
24]
,
R
e
s
Ne
t50
a
nd
R
e
s
Ne
t101
[
25]
a
nd
mac
hine
lea
r
ning
c
las
s
if
ica
ti
on
method
(
kNN
,
d
e
c
is
ion
tr
e
e
,
S
VM
)
to
c
las
s
if
y
maiz
e
lea
f
dis
e
a
s
e
s
.
S
e
c
ond,
the
pe
r
c
e
ntage
of
a
c
c
ur
a
c
y
incr
e
a
s
e
d
while
c
ompar
e
d
to
the
p
r
e
vious
s
tudy.
2.
RE
S
E
AR
CH
M
E
T
HO
D
T
he
s
teps
f
or
c
las
s
if
ica
ti
on
pr
oc
e
s
s
us
ing
C
NN
a
r
e
s
hown
in
F
igu
r
e
1.
M
a
ize
lea
f
im
a
ge
s
a
s
da
ta
a
r
e
divi
de
d
int
o
2
pa
r
ts
:
t
r
a
ini
ng
a
nd
tes
ti
ng
da
ta.
F
ur
t
he
r
mor
e
,
C
NN
is
a
ppli
e
d
,
the
f
unc
ti
on
o
f
C
NN
a
s
a
f
e
a
tur
e
e
xtr
a
c
ti
on
pr
oc
e
s
s
without
de
ter
mi
ning
type
of
f
e
a
tur
e
e
xtr
a
c
ti
on
a
s
in
c
onve
nti
ona
l
mac
hine
l
e
a
r
ning.
T
he
ne
xt
p
r
oc
e
s
s
is
c
las
s
if
ica
ti
on
us
ing
k
-
Ne
a
r
e
s
t
Ne
ighbor
,
s
uppor
t
ma
c
hine
a
nd
de
c
is
ion
tr
e
e
.
F
igur
e
1.
T
he
r
e
s
e
a
r
c
h
method
o
f
maiz
e
lea
f
dis
e
a
s
e
im
a
ge
c
las
s
if
ica
ti
on
2.
1
.
M
aize
leaf
im
a
ge
I
mage
da
ta
us
e
d
maiz
e
lea
ve
s
that
s
ize
of
256x2
56
pixels
.
Da
ta
c
ons
is
ts
of
200
im
a
ge
s
whic
h
a
r
e
divi
de
d
int
o
4
c
las
s
e
s
,
50
im
a
ge
s
pe
r
c
las
s
.
E
xpe
r
im
e
nt
da
ta
obtaine
d
f
r
om
M
oha
nty
plant
v
il
l
a
ge
[
19]
.
E
xa
mpl
e
s
of
im
a
ge
da
ta
on
maiz
e
lea
ve
s
a
r
e
s
ho
w
n
in
F
igur
e
2
.
W
he
n
tr
a
ini
ng
a
nd
tes
ti
ng
us
ing
C
NN
,
im
a
ge
s
ize
is
a
djus
ted
to
de
f
a
ult
s
ize
of
e
a
c
h
C
NN
a
r
c
hit
e
c
tur
e
.
T
a
ble
1
s
how
the
de
f
a
ult
inpu
t
s
ize
of
C
NN
model.
F
igur
e
2
.
(
a
)
N
o
r
mal,
(
b)
C
e
r
c
os
por
a
,
(
c
)
N
or
ther
n
lea
f
bli
ght
,
(
d)
C
omm
on
r
us
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
3
,
J
une
2020:
1
376
-
1
381
1378
Ta
ble
1.
De
f
a
ult
input
s
ize
of
C
NN
C
N
N
D
e
f
a
ul
t
I
nput
S
iz
e
A
le
xN
e
t
227×
227
V
G
G
16
224×
224
V
G
G
19
224×
224
G
oogl
e
N
e
t
224×
224
I
nc
e
pt
io
n
-
V3
299×
299
R
e
s
N
e
t5
0
224×
224
R
e
s
N
e
t1
01
224×
224
2.
2.
Convol
u
t
ion
al
n
e
u
r
al
n
e
t
wor
k
C
NN
c
ons
is
t
s
of
2
main
pa
r
ts
:
f
e
a
tur
e
e
xtr
a
c
ti
on
a
nd
c
las
s
if
ica
ti
on.
T
he
f
e
a
tur
e
e
xtr
a
c
ti
on
s
e
c
ti
on
include
s
input
laye
r
,
c
onvolut
ional
laye
r
with
s
t
r
ide
a
nd
pa
dding
,
r
e
c
ti
f
ied
li
ne
a
r
unit
(
R
e
L
U)
,
pooli
ng,
a
nd
ba
tch
nor
maliza
ti
on
laye
r
.
W
hil
e
the
c
las
s
if
i
c
a
ti
on
pa
r
t
c
ons
is
ts
of
f
u
ll
y
c
onne
c
ted
laye
r
,
s
of
t
max
da
n
output
laye
r
.
C
NN
a
r
c
hit
e
c
tur
e
c
a
n
ha
ve
mor
e
tha
n
one
type
of
laye
r
[
26]
.
C
NN
a
r
c
hit
e
c
tur
e
s
a
na
lyze
d
in
thi
s
pa
pe
r
we
r
e
Ale
xNe
t,
VG
G16,
VG
G19,
Googl
e
Ne
t,
I
nc
e
pti
on
-
V3,
R
e
s
Ne
t50,
a
nd
R
e
s
Ne
t101.
T
hos
e
a
r
c
hit
e
c
tur
e
s
ha
ve
25,
41
,
177
a
nd
144
laye
r
s
,
r
e
s
pe
c
ti
ve
ly.
F
i
g
u
r
e
3
s
h
o
w
s
a
s
i
m
p
l
e
C
N
N
m
o
d
e
l
t
h
a
t
h
a
s
1
3
l
a
y
e
r
s
:
1
i
n
p
u
t
l
a
y
e
r
,
3
c
o
n
v
o
l
u
t
i
o
n
a
l
l
a
y
e
r
s
w
i
t
h
s
t
r
i
d
e
a
n
d
p
a
d
d
i
n
g
,
3
R
e
L
U
l
a
y
e
r
s
,
p
o
o
l
i
n
g
l
a
y
e
r
,
2
n
o
r
m
a
l
i
z
a
t
i
o
n
l
a
y
e
r
,
F
C
L
,
s
o
f
t
m
a
x
,
a
n
d
o
u
t
p
u
t
l
a
y
e
r
.
F
igur
e
3.
S
im
ple
C
NN
model
Ale
xNe
t
a
r
c
hit
e
c
tur
e
ha
s
twe
nty
-
f
ive
laye
r
s
[
22]
:
I
nput
laye
r
,
5
c
onvolut
ional
laye
r
s
,
f
ir
s
t
co
nvolut
ional
laye
r
ha
s
a
11×
11
f
il
ter
,
s
e
c
ond
laye
r
ha
s
a
5×
5
f
il
ter
,
a
nd
thi
r
d
,
to
f
i
f
t
h
laye
r
ha
ve
3
×
3
f
il
ter
s
.
F
ur
ther
mor
e
,
7
R
e
L
U
laye
r
s
,
2
nor
maliza
ti
on
laye
r
s
,
3
max
-
pooli
ng
laye
r
s
,
3
f
ull
y
c
onne
c
ted
laye
r
,
2
dr
opouts
0.
5
,
S
of
tm
a
x
a
nd
output
laye
r
.
Vis
ua
l
Ge
ometr
y
Gr
oup
(
VG
G)
f
r
om
Oxf
o
r
d
Unive
r
s
it
y
c
r
e
a
tes
a
VG
G16
ne
twor
k
a
r
c
hit
e
c
tur
e
with
41
laye
r
s
.
VG
G
s
im
pli
f
ies
the
pr
oc
e
s
s
e
s
by
c
r
e
a
ti
ng
a
3×
3
f
il
ter
f
or
e
a
c
h
laye
r
.
E
quivale
nt
a
nd
s
maller
f
il
ter
s
ize
us
e
d
in
VG
G
c
a
n
pr
oduc
e
mor
e
c
ompl
e
x
f
e
a
tur
e
s
a
nd
lowe
r
c
omput
ing
than
Ale
xNe
t’
s
.
VG
G16
a
r
c
hit
e
c
tu
r
e
c
ons
is
ts
of
[
23]
:
the
input
laye
r
s
ize
is
224×
224
pixels
.
13
c
onvolut
ional
laye
r
s
.
F
i
r
s
t
a
nd
s
e
c
ond
c
onvolut
ional
laye
r
s
ha
ve
f
il
ter
s
ize
of
64
p
ixels
,
thi
r
d
a
n
d
f
our
th
ha
ve
f
il
ter
s
ize
o
f
128
pixels
,
f
if
th
to
s
e
ve
nth
ha
ve
f
il
te
r
s
ize
of
256
pixels
a
nd
e
ig
ht
to
thi
r
tee
nth
h
a
ve
f
il
ter
s
ize
of
512.
F
if
tee
n
R
e
L
U.
5
max
-
pooli
ng,
3
f
ull
y
c
onne
c
ted
laye
r
s
.
T
wo
dr
opout
0.
5
,
S
of
t
max
a
nd
o
utput
laye
r
W
hil
e
VG
G19
a
r
c
hit
e
c
tur
e
c
ons
is
ts
of
[
23]
:
inpu
t
laye
r
is
224×
224
pixels
.
S
ixt
e
e
n
c
onvolut
ional
laye
r
s
.
F
ir
s
t
a
nd
s
e
c
ond
c
onvolut
ional
laye
r
s
ha
ve
f
il
te
r
s
ize
o
f
64
pixels
,
thi
r
d
a
nd
f
our
th
ha
ve
f
il
te
r
s
ize
of
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
C
onv
olut
ional
ne
ur
al
ne
tw
or
k
for
maiz
e
leaf
dis
e
as
e
image
c
las
s
if
ication
(
M
ohamm
ad
Sy
ar
ief
)
1379
128
pixels
,
5
to
8
ha
ve
f
il
ter
s
ize
of
256
pixels
a
nd
the
9
to
16
ha
v
e
f
il
ter
s
ize
of
512
pixels
,
1
8
R
e
L
U,
5
max
-
pooli
ng,
3
f
u
ll
y
c
onne
c
ted
laye
r
s
,
2
dr
o
pouts
of
0
.
5
S
of
tm
a
x
a
nd
outpu
t
laye
r
.
R
e
s
Ne
t50
da
n
R
e
s
Ne
t101
incr
e
a
s
ing
number
of
laye
r
s
is
d
ir
e
c
tl
y
pr
opor
ti
ona
l
to
the
incr
e
a
s
e
in
lea
r
n
ing,
but
it
c
a
n
lea
d
to
lea
r
ning
mor
e
a
nd
mo
r
e
dif
f
icult
a
nd
a
c
c
ur
a
c
y
de
c
r
e
a
s
e
s
.
R
e
s
idual
lea
r
ning
pr
ovides
s
olut
ions
to
thes
e
pr
oblems
.
R
e
s
idual
Ne
twor
k
(
R
e
s
Ne
t)
is
a
C
NN
n
e
twor
k
a
r
c
hit
e
c
tur
e
f
or
r
e
s
idual
lea
r
ning.
R
e
s
idual
lea
r
ning
s
kips
laye
r
c
onne
c
ti
on.
R
e
s
Ne
t50
a
r
c
hit
e
c
tur
e
ha
s
177
laye
r
s
,
while
R
e
s
Ne
t101
ha
s
347
laye
r
s
[
26]
.
GoogL
e
Ne
t
(
I
nc
e
pti
on
-
V1)
is
a
C
NN
a
r
c
hit
e
c
tur
e
that
ha
s
144
laye
r
s
.
GoogL
e
Ne
t
c
or
r
e
c
ts
de
f
icie
nc
ies
in
VG
G
that
r
e
quir
e
high
c
omput
ing,
both
memor
y
a
nd
ti
me.
T
he
wo
r
king
pr
inciple
of
I
nc
e
pti
on
is
that
the
ne
twor
k
will
a
utom
a
t
ica
ll
y
c
hoos
e
the
b
e
s
t
c
onvolut
ion
r
e
s
ult
s
us
ing
a
c
e
r
tain
s
ize
.
F
il
ter
s
ize
us
e
d
in
thi
s
a
r
c
hit
e
c
tur
e
is
1×
1
p
ixels
,
3×
3
pixels
,
5×
5
pixels
a
nd
max
-
pooli
ng
3×
3
pixels
.
Anothe
r
va
r
ia
nt
us
e
d
in
thi
s
s
tudy
wa
s
I
nc
e
pti
on
-
V3.
I
nc
e
pti
on
-
V3
a
r
c
hit
e
c
tur
e
c
ons
is
ts
of
316
laye
r
s
[
27
-
29]
2.
3.
Clas
s
if
icat
ion
m
e
t
h
od
s
I
n
thi
s
s
tudy,
we
us
e
d
thr
e
e
c
las
s
if
ica
ti
on
me
thods
f
or
tes
ti
ng:
s
uppor
t
ve
c
tor
mac
hine
[
30
]
,
k
-
Ne
a
r
e
s
t
Ne
ighbor
[
31]
a
nd
de
c
is
ion
t
r
e
e
[
3
2]
.
F
or
e
a
c
h
tes
ti
ng
,
we
c
onf
igu
r
e
the
ne
twor
k
laye
r
,
e
xtr
a
c
t
the
f
e
a
tur
e
s
,
a
nd
make
c
las
s
if
ica
ti
on
us
ing
e
a
c
h
method
a
bove
.
3.
RE
S
UL
T
AN
D
DI
S
CU
S
S
I
ON
T
he
e
xpe
r
im
e
nt
divi
de
d
int
o
3
s
c
e
na
r
ios
,
output
of
7
C
NN
models
c
las
s
if
ied
with
S
VM
,
kNN
a
nd
De
c
is
ion
T
r
e
e
.
T
e
s
ti
ng
r
e
s
ult
s
us
ing
the
C
NN
a
r
c
hit
e
c
tur
e
we
r
e
f
ound
in
T
a
ble
2
to
T
a
ble
4
.
T
a
ble
2
r
e
pr
e
s
e
nted
c
las
s
if
ica
ti
on
tes
ti
ng
u
s
ing
the
S
VM
method,
while
T
a
ble
3
a
nd
T
a
ble
4
f
or
c
las
s
if
ica
ti
on
us
ing
k
-
Ne
a
r
e
s
t
Ne
ighbor
a
nd
de
c
is
ion
tr
e
e
method
s
,
r
e
s
pe
c
ti
ve
ly.
T
a
ble
2.
T
e
s
ti
ng
r
e
s
ult
s
us
ing
S
VM
C
N
N
mode
l
S
e
ns
it
iv
it
y
(%)
S
pe
c
if
ic
it
y
(%)
A
c
c
ur
a
c
y
(%)
A
le
xN
e
t
95.83
100
95
V
gg16
88.4
92.03
88.3
V
gg19
88.4
92.03
88.3
R
e
s
N
e
t5
0
87.28
90.63
86.7
R
e
s
N
e
t1
01
90
92.85
90
G
oogl
e
N
e
t
83.75
89.13
83.3
I
nc
e
pt
io
n
-
V3
87.55
91.32
86.7
A
ve
r
a
ge
88.74
92.57
88.33
T
a
ble
3.
T
e
s
ti
ng
r
e
s
ult
s
us
ing
kNN
C
N
N
mode
l
S
e
ns
it
iv
it
y
(%)
S
pe
c
if
ic
it
y
(%)
A
c
c
ur
a
c
y
(%)
A
le
xN
e
t
94.72
94.715
93.3
V
gg16
82.23
89.65
76.7
V
gg19
88.4
91.94
88.3
R
e
s
N
e
t5
0
73.68
87.2
75
R
e
s
N
e
t1
01
81.98
88.89
80
G
oogl
e
N
e
t
84.55
89.41
83.3
I
nc
e
pt
io
n
-
V3
80.98
88.33
80
A
ve
r
a
ge
83.79
90.02
82.37
T
a
ble
4.
T
e
s
ti
ng
r
e
s
ult
s
us
ing
de
c
is
ion
tr
e
e
C
N
N
mode
l
S
e
ns
it
iv
it
y (
%
)
S
pe
c
if
ic
it
y (
%
)
A
c
c
ur
a
c
y (
%
)
A
le
xN
e
t
74.33
84.19
73.3
V
gg16
77.73
84.58
75
V
gg19
76.93
86.89
76.7
R
e
s
N
e
t5
0
83.58
88.22
83.3
R
e
s
N
e
t1
01
76
83.7
73.3
G
oogl
e
N
e
t
73.85
85.61
75
I
nc
e
pt
io
n
-
V3
66.53
79.93
65
A
ve
r
a
ge
75.56
84.73
74.51
B
a
s
e
d
on
the
tes
ti
ng
r
e
s
ult
s
a
bove
,
the
be
s
t
c
las
s
if
i
c
a
ti
on
wa
s
pr
oduc
e
d
by
Ale
xNe
t
a
r
c
hit
e
c
tur
e
with
S
uppor
t
Ve
c
tor
M
a
c
hine
c
las
s
if
ica
ti
on.
I
t
s
how
e
d
the
be
s
t
pe
r
f
o
r
manc
e
mea
s
ur
e
s
ba
s
e
d
on
s
e
ns
it
ivi
ty,
s
pe
c
if
icity,
a
nd
a
c
c
ur
a
c
y
of
95.
83
%
,
100%
,
a
nd
95%
,
r
e
s
pe
c
ti
ve
ly.
B
e
s
t
a
ve
r
a
ge
a
c
c
ur
a
c
y
of
88.
33
%
us
ing
S
VM
.
F
u
r
ther
mor
e
,
to
do
va
li
da
ti
on
us
e
d
10
-
f
old
c
r
os
s
-
va
li
da
ti
on.
T
his
method
will
d
ivi
de
da
ta
int
o
10
e
qua
l
pa
r
ts
.
T
he
c
ompl
e
te
s
tage
s
a
r
e
a
s
f
oll
ows
:
a)
F
ir
s
t,
9
da
ta
s
e
c
ti
ons
a
r
e
us
e
d
f
or
tr
a
ini
ng
,
one
f
ina
l
da
ta
s
e
c
ti
on
is
us
e
d
f
or
tes
ti
ng.
b)
S
e
c
ond,
the
da
ta
s
e
c
ti
on
unti
l
the
da
ta
is
us
e
d
f
or
t
r
a
ini
ng,
the
f
i
r
s
t
da
ta
s
e
c
ti
on
is
us
e
d
f
o
r
tes
ti
ng,
a
nd
s
o
on.
c)
And
s
o
on
unti
l
the
da
ta
pa
r
t
1
to
8
a
nd
10
a
r
e
us
e
d
a
s
tr
a
ini
ng
da
ta,
while
da
ta
s
e
c
ti
on
9
is
u
s
e
d
a
s
tes
ti
ng
da
ta.
d)
F
ind
the
a
ve
r
a
ge
va
lue
of
a
ll
r
ounds
.
F
o
r
mor
e
de
tails
,
a
n
il
lus
tr
a
ti
on
of
10
-
f
old
c
r
os
s
-
va
li
da
ti
on
is
s
hown
in
F
igur
e
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
18
,
No
.
3
,
J
une
2020:
1
376
-
1
381
1380
T
he
r
e
s
ult
s
of
10
-
f
old
c
r
os
s
-
va
li
da
ti
on
a
r
e
s
hown
in
T
a
ble
5.
I
t
us
ing
Ale
xNe
t
a
nd
S
VM
a
s
c
las
s
if
ica
ti
on.
I
n
the
f
inal
s
e
c
ti
on,
the
r
e
s
ult
s
of
10
k
c
r
os
s
-
va
li
da
ti
o
n
we
r
e
c
ompar
e
d
with
p
r
e
vious
s
tudi
e
s
.
T
a
ble
6
r
e
pr
e
s
e
nts
a
c
ompar
is
on
be
twe
e
n
thi
s
s
tudy
a
nd
p
r
e
vious
s
tudi
e
s
us
ing
maiz
e
lea
f
im
a
ge
s
f
or
dis
e
a
s
e
c
la
s
s
if
ica
ti
on.
F
igur
e
4.
10
-
f
old
c
r
os
s
-
va
li
d
a
ti
on
T
a
ble
5.
P
e
r
f
o
r
manc
e
m
e
a
s
ur
e
of
10
-
f
old
c
r
os
s
-
va
l
idation
R
ound
S
e
ns
it
iv
it
y (
%
)
S
pe
c
if
ic
it
y
(%)
A
c
c
ur
a
c
y (
%
)
1
95.83
95.83
95
2
90
92.86
90
3
100
100
100
4
87.5
88.69
85
5
100
100
100
6
95.83
95.83
95
7
100
100
100
8
80
88.89
80
9
90
92.86
90
10
95.83
95.83
95
A
ve
r
a
ge
93.5
95.08
93
T
a
bl
e
6.
R
e
s
ult
s
c
ompar
is
on
of
maiz
e
lea
f
c
las
s
if
ica
ti
on
A
ut
hor
s
N
umbe
r
of
c
la
s
s
e
s
S
e
ns
it
iv
it
y
(%)
S
pe
c
if
ic
it
y
(%)
A
c
c
ur
a
c
y
(%)
S
ib
iy
a
&
S
umbwa
nya
mbe
[
20
]
3
-
-
92.85
Z
ha
ng e
t
a
l.
[
18]
8
-
-
98.9
H
id
a
ya
t
e
t
a
l.
[
21]
3
-
-
93.67
P
r
opos
e
d me
th
od
4
93.5
95.08
93
4.
CONC
L
USI
ON
T
his
s
tudy
a
na
lyze
d
maiz
e
lea
f
im
a
ge
c
las
s
if
ica
ti
on
us
ing
7
C
NN
a
r
c
hit
e
c
tur
e
s
(
Ale
xNe
t,
VG
G16,
VG
G19,
R
e
s
Ne
t50,
R
e
s
Ne
t110.
Google
Ne
t,
a
nd
I
nc
e
pti
on
-
V3)
a
nd
the
c
las
s
if
ica
ti
on
methods
(
S
V
M
,
kNN
,
a
nd
De
c
is
ion
T
r
e
e
)
.
T
he
be
s
t
c
las
s
if
ica
ti
on
wa
s
g
e
ne
r
a
ted
by
Ale
xNe
t
a
r
c
hit
e
c
tur
e
with
S
VM
.
T
h
is
s
howe
d
that
Ale
xNe
t
a
nd
S
VM
methods
we
r
e
be
s
t
s
uit
e
d
f
or
f
e
a
tu
r
e
e
xtr
a
c
ti
on
a
nd
i
mage
c
las
s
if
ica
ti
on
of
maiz
e
lea
ve
s
dis
e
a
s
e
.
F
ur
ther
mor
e
,
we
c
ould
incr
e
a
s
e
the
pe
r
c
e
ntage
of
a
c
c
ur
a
c
y
by
a
dding
opti
mi
z
a
ti
on
me
thods
in
C
NN
a
r
c
hit
e
c
tur
e
s
.
R
E
F
E
RE
NC
E
S
[1
]
C.
St
eg
er,
M.
U
l
ri
ch
,
an
d
C.
W
i
ed
eman
n
,
"
Mach
i
n
e
V
i
s
i
o
n
A
l
g
o
r
i
t
h
ms
an
d
A
p
p
l
i
cat
i
o
n
s
,
"
1
st
E
d
i
t
i
o
n
.
W
ei
n
h
e
i
m
:
W
i
l
ey
-
V
C
H
,
2
0
0
7
.
[2
]
D
.
H
an
a,
Q
.
L
i
u
,
an
d
W
.
Fa
n
,
“A
n
e
w
i
ma
g
e
cl
as
s
i
f
i
cat
i
o
n
me
t
h
o
d
u
s
i
n
g
CN
N
t
ra
n
s
fer
l
earn
i
n
g
an
d
w
e
b
d
at
a
au
g
me
n
t
a
t
i
o
n
,
”
E
x
p
er
t
S
ys
t
.
A
p
p
l
.
,
v
o
l
.
9
5
,
p
p
.
4
3
-
5
6
,
2
0
1
8
.
[3
]
C.
Z
h
an
g
et
a
l
.
,
“A
h
y
b
r
i
d
ML
P
-
CN
N
c
l
as
s
i
f
i
er
fo
r
v
ery
fi
n
e
res
o
l
u
t
i
o
n
remo
t
el
y
s
en
s
ed
i
mag
e
c
l
as
s
i
f
i
cat
i
o
n
,
”
IS
P
R
S
J.
P
h
o
t
o
g
r
a
m
m
.
R
e
m
o
t
e
S
e
n
s
.
,
v
o
l
.
1
4
0
,
p
p
.
1
3
3
-
1
4
4
,
2
0
1
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
C
onv
olut
ional
ne
ur
al
ne
tw
or
k
for
maiz
e
leaf
dis
e
as
e
image
c
las
s
if
ication
(
M
ohamm
ad
Sy
ar
ief
)
1381
[4
]
W
.
Set
i
aw
a
n
,
M.
I.
U
t
o
y
o
,
an
d
R.
Ru
l
an
i
n
g
t
y
as
,
“V
es
s
el
s
s
eman
t
i
c
s
eg
me
n
t
a
t
i
o
n
w
i
t
h
g
rad
i
en
t
d
es
c
en
t
o
p
t
i
mi
zat
i
o
n
,
”
In
t
e
r
n
a
t
i
o
n
a
l
A
ca
d
em
i
c
Jo
u
r
n
a
l
s
.
,
v
o
l
.
7
,
n
o
.
4
,
p
p
.
4
0
6
2
-
4
0
6
7
,
2
0
1
8
.
[5
]
W
.
Set
i
a
w
an
,
M.
I
.
U
t
o
y
o
,
an
d
R.
Ru
l
an
i
n
g
t
y
as
,
“Se
man
t
i
c
s
eg
me
n
t
a
t
i
o
n
o
f
art
ery
-
v
en
o
u
s
ret
i
n
a
l
v
es
s
e
l
u
s
i
n
g
s
i
mp
l
e
co
n
v
o
l
u
t
i
o
n
a
l
n
eu
ra
l
n
et
w
o
r
k
,
”
IO
P
Co
n
f
e
r
en
ce
S
er
i
es
:
E
a
r
t
h
a
n
d
E
n
vi
r
o
n
m
e
n
t
a
l
S
c
i
en
ce
,
v
o
l
.
2
4
3
,
n
o
.
1
,
p
p
.
1
-
10
,
2
0
1
9
.
[6
]
Y
.
L
i
,
J
.
Z
en
g
,
an
d
S.
Sh
an
,
“O
ccl
u
s
i
o
n
aw
are
faci
al
ex
p
res
s
i
o
n
reco
g
n
i
t
i
o
n
u
s
i
n
g
CN
N
w
i
t
h
at
t
en
t
i
o
n
mech
an
i
s
m,
”
IE
E
E
Tr
a
n
s
.
Im
a
g
e
P
r
o
ces
s
.
,
v
o
l
.
2
8
,
n
o
.
5
,
p
p
.
2
4
3
9
-
2
4
5
0
,
2
0
1
8
.
[7
]
Y
.
X
,
Y
an
g
,
et
a
l
.
,
“Face
reco
g
n
i
t
i
o
n
u
s
i
n
g
t
h
e
SR
-
C
N
N
Mo
d
el
,
”
S
e
n
s
o
r
s
,
v
o
l
.
1
8
,
n
o
.
12
,
2
0
1
8
.
[8
]
Y
.
Ch
en
,
et
al
.
,
“D
o
mai
n
ad
ap
t
i
v
e
fas
t
er
R
-
CN
N
f
o
r
o
b
j
ect
d
et
ec
t
i
o
n
i
n
t
h
e
w
i
l
d
,
”
p
p
.
3
3
3
9
-
3
3
4
8
,
2
0
1
8
.
[9
]
H
.
G
ao
,
et
al
.
,
“O
b
j
ect
cl
a
s
s
i
fi
ca
t
i
o
n
u
s
i
n
g
CN
N
-
b
a
s
ed
fu
s
i
o
n
o
f
v
i
s
i
o
n
an
d
L
ID
A
R
i
n
au
t
o
n
o
mo
u
s
v
e
h
i
cl
e
en
v
i
ro
n
men
t
,
”
IE
E
E
T
r
a
n
s
.
In
d
.
In
f
o
r
m
a
t
i
c
s
,
v
o
l
.
14
,
n
o
.
9
,
p
p
.
4
2
2
4
-
4
2
3
1
,
2
0
1
8
.
[1
0
]
A
.
K
h
at
am
i
,
et
a
l
.
,
“A
s
eq
u
e
n
t
i
al
s
earch
-
s
p
ace
s
h
ri
n
k
i
n
g
u
s
i
n
g
CN
N
t
ra
n
s
fe
r
l
earn
i
n
g
an
d
a
Rad
o
n
p
r
o
j
ec
t
i
o
n
p
o
o
l
fo
r
med
i
ca
l
i
ma
g
e
ret
ri
e
v
a
l
,
”
E
xp
e
r
t
S
y
s
t
.
A
p
p
l
.
,
v
o
l
.
1
0
0
,
p
p
.
2
2
4
-
2
3
3
,
2
0
1
8
.
[1
1
]
M.
Fri
d
-
ad
ar,
et
a
l
.
,
“G
A
N
-
b
a
s
ed
s
y
n
t
h
e
t
i
c
me
d
i
ca
l
i
m
ag
e
au
g
me
n
t
a
t
i
o
n
f
o
r
i
n
cre
a
s
ed
C
N
N
p
erf
o
rma
n
ce
i
n
l
i
v
er
l
es
i
o
n
cl
a
s
s
i
f
i
cat
i
o
n
,
”
Neu
r
o
co
m
p
u
t
i
n
g
,
v
o
l
.
3
2
1
,
p
p
.
3
2
1
-
3
3
1
,
2
0
1
8
.
[1
2
]
S.
O
.
A
.
Ch
i
s
h
t
i
,
e
t
al
.
,
“Sel
f
-
d
ri
v
i
n
g
car
s
u
s
i
n
g
C
N
N
an
d
Q
-
l
earn
i
n
g
,
”
IE
E
E
2
1
st
I
n
t
e
r
n
a
t
i
o
n
a
l
M
u
l
t
i
-
T
o
p
i
c
Co
n
f
er
e
n
ce
(INM
IC)
,
p
p
.
1
-
7
,
2
0
1
8
.
[1
3
]
A
.
D
h
a
l
l
,
D
.
D
a
i
,
a
n
d
L
.
V
a
n
G
o
o
l
,
“
R
e
a
l
-
t
i
m
e
3
D
T
r
a
f
f
i
c
c
o
n
e
d
e
t
e
c
t
i
o
n
f
o
r
a
u
t
o
n
o
m
o
u
s
d
r
i
v
i
n
g
,
”
p
p
.
4
9
4
-
5
0
1
,
2
0
1
9
.
[1
4
]
K
.
Mo
t
t
,
“St
at
e
cl
a
s
s
i
fi
ca
t
i
o
n
o
f
co
o
k
i
n
g
o
b
j
ect
s
u
s
i
n
g
a
V
G
G
CN
N
,
”
arX
i
v
p
rep
r
i
n
t
arX
i
v
:
1
9
0
4
.
1
2
6
1
3
,
2
0
1
8
.
[1
5
]
A
.
Mi
l
i
o
t
o
a
n
d
C.
S
t
ach
n
i
s
s
,
“Bo
n
n
e
t
:
A
n
o
p
e
n
-
s
o
u
rce
t
rai
n
i
n
g
a
n
d
d
ep
l
o
y
men
t
framew
o
rk
f
o
r
s
ema
n
t
i
c
s
eg
me
n
t
a
t
i
o
n
i
n
ro
b
o
t
i
c
s
u
s
i
n
g
CN
N
s
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
R
o
b
o
t
i
c
s
a
n
d
A
u
t
o
m
a
t
i
o
n
,
p
p
.
7
0
9
4
-
7
1
0
0
,
2
0
1
9
.
[1
6
]
A
.
K
ami
l
ar
i
s
an
d
F.
X
.
Pren
afet
a
-
Bo
l
d
u
,
“A
rev
i
e
w
o
f
t
h
e
u
s
e
o
f
co
n
v
o
l
u
t
i
o
n
al
n
e
u
ral
n
et
w
o
r
k
s
i
n
ag
ri
c
u
l
t
u
r
e,
”
J.
A
g
r
i
c.
S
ci
.
,
v
o
l
.
1
5
6
,
n
o
.
3
,
p
p
.
3
1
2
-
3
2
2
,
2
0
1
8
.
[1
7
]
S.
Sl
ad
o
j
ev
i
c,
et
a
l
.
,
“D
eep
n
e
u
ral
n
et
w
o
r
k
s
b
as
e
d
re
co
g
n
i
t
i
o
n
o
f
p
l
an
t
d
i
s
eas
e
s
b
y
l
e
a
f
i
ma
g
e
cl
a
s
s
i
fi
cat
i
o
n
,
”
Co
m
p
u
t
.
In
t
e
l
l
.
Neu
r
o
s
ci
.
,
v
o
l
.
2
0
1
6
,
p
p
.
1
-
1
1
,
2
0
1
6
.
[1
8
]
X
.
Z
h
an
g
,
e
t
a
l
.
,
“Id
e
n
t
i
fi
ca
t
i
o
n
o
f
M
ai
ze
l
eaf
d
i
s
eas
es
u
s
i
n
g
i
mp
r
o
v
e
d
d
ee
p
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
ral
n
et
w
o
r
k
s
,
”
IE
E
E
A
cces
s
,
v
o
l
.
6
,
p
p
.
3
0
3
7
0
-
3
0
3
7
7
,
2
0
1
8
.
[1
9
]
S.
P.
Mo
h
an
t
y
,
D
.
H
u
g
h
e
s
,
an
d
M.
Sal
at
h
é,
“U
s
i
n
g
d
eep
l
earn
i
n
g
fo
r
i
ma
g
e
-
b
a
s
ed
p
l
an
t
d
i
s
eas
e
d
et
e
c
t
i
o
n
,
”
F
r
o
n
t
.
P
l
a
n
t
S
ci
.
,
v
o
l
.
7
,
n
o
.
1
0
,
p
p
.
1
-
7
,
2
0
1
6
.
[2
0
]
M
.
S
i
b
i
y
a
a
n
d
M
.
S
u
m
b
w
a
n
y
a
m
b
e
,
“
A
c
o
m
p
u
t
a
t
i
o
n
a
l
p
r
o
c
e
d
u
r
e
f
o
r
t
h
e
r
e
c
o
g
n
i
t
i
o
n
a
n
d
c
l
a
s
s
i
f
i
c
a
t
i
o
n
o
f
m
a
i
z
e
l
e
a
f
d
i
s
e
a
s
e
s
o
u
t
o
f
h
e
a
l
t
h
y
l
e
a
v
e
s
u
s
i
n
g
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
,
”
A
g
r
i
E
n
g
i
n
e
e
r
i
n
g
,
v
o
l
.
1
,
n
o
.
1
,
p
p
.
1
1
9
-
1
3
1
,
2
0
1
9
.
[2
1
]
A
.
H
i
d
ay
a
t
,
U
,
D
aru
s
al
am,
a
n
d
I,
Irma
w
at
i
.
,
“D
et
ec
t
i
o
n
o
f
d
i
s
ea
s
e
o
n
co
r
n
p
l
a
n
t
s
u
s
i
n
g
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
al
,
”
J
u
r
n
a
l
I
l
m
u
K
o
m
p
u
t
e
r
d
a
n
I
n
f
o
r
m
a
s
i
.,
v
o
l
.
1
2
,
n
o
.
1
,
p
p
.
5
1
-
5
6
,
2
0
1
9
.
[2
2
]
A
.
K
ri
zh
e
v
s
k
y
,
I.
Su
t
s
k
e
v
er,
an
d
G
.
E
.
H
i
n
t
o
n
,
“Imag
e
N
et
cl
a
s
s
i
fi
ca
t
i
o
n
w
i
t
h
d
eep
c
o
n
v
o
l
u
t
i
o
n
a
l
n
eu
ra
l
n
et
w
o
rk
s
,
”
A
d
v
a
n
c
e
s
i
n
n
e
u
r
a
l
i
n
f
o
r
m
a
t
i
o
n
p
r
o
c
e
s
s
i
n
g
s
y
s
t
e
m
s
,
p
p
.
1
0
9
7
-
1
1
0
5
,
2
0
1
2
.
[2
3
]
K
.
Si
mo
n
y
an
an
d
A
.
Z
i
s
s
erma
n
,
“V
ery
d
eep
co
n
v
o
l
u
t
i
o
n
al
n
e
t
w
o
rk
s
fo
r
l
arg
e
-
s
cal
e
i
ma
g
e
reco
g
n
i
t
i
o
n
,
”
ar
X
i
v
p
rep
r
i
n
t
arX
i
v
:
1
4
0
9
.
1
5
5
6
,
2
0
1
4
.
[2
4
]
K
.
H
e,
X
.
Z
h
an
g
,
S.
Ren
,
an
d
J
.
Su
n
,
“D
eep
res
i
d
u
al
l
earn
i
n
g
f
o
r
i
mag
e
s
t
e
g
an
a
l
y
s
i
s
,
”
M
u
l
t
i
m
e
d
.
To
o
l
s
A
p
p
l
.
,
p
p
.
7
7
0
-
7
7
8
,
2
0
1
5
.
[2
5
]
K
.
H
e
an
d
J
.
Su
n
,
“
D
ee
p
res
i
d
u
a
l
l
ear
n
i
n
g
fo
r
i
mag
e
rec
o
g
n
i
t
i
o
n
,
”
IE
E
E
X
p
l
o
r
e
,
p
p
.
1
-
9
,
2
0
1
5
.
[2
6
]
W
.
Set
i
aw
a
n
,
M.
I
,
U
t
o
y
o
,
an
d
R.
Ru
l
an
i
n
g
t
y
a
s
,
“Cl
as
s
i
fi
cat
i
o
n
o
f
n
eo
v
a
s
cu
l
ari
za
t
i
o
n
u
s
i
n
g
co
n
v
o
l
u
t
i
o
n
a
l
n
e
u
ral
n
et
w
o
r
k
mo
d
el
,
”
TE
LK
O
M
NIK
A
Tel
eco
m
m
u
n
i
ca
t
i
o
n
Co
m
p
u
t
i
n
g
E
l
ect
r
o
n
i
c
s
a
n
d
Co
n
t
r
o
l
,
v
o
l
.
1
7
,
n
o
.
1
,
p
p
.
4
6
3
-
4
7
2
,
2
0
1
9
.
[2
7
]
C.
Szeg
ed
y
,
V
.
V
a
n
h
o
u
k
e,
S.
Io
ffe,
an
d
J
.
S
h
l
e
n
s
,
“
R
et
h
i
n
k
i
n
g
t
h
e
i
n
ce
p
t
i
o
n
arch
i
t
ec
t
u
re
fo
r
c
o
mp
u
t
er
v
i
s
i
o
n
ch
ri
s
t
i
an
,
”
IE
E
E
E
x
p
l
o
r
.
,
p
p
.
2
8
1
8
-
2
8
2
6
,
2
0
1
6
[2
8
]
C.
Szeg
ed
y
,
et
a
l
.
,
“In
ce
p
t
i
o
n
-
v
4
,
In
cep
t
i
o
n
-
Re
s
N
e
t
an
d
t
h
e
i
m
p
act
o
f
res
i
d
u
a
l
co
n
n
ect
i
o
n
s
o
n
l
ear
n
i
n
g
,
”
Th
i
r
t
y
-
f
i
r
s
t
A
A
A
I
co
n
f
e
r
en
ce
o
n
a
r
t
i
f
i
ci
a
l
i
n
t
e
l
l
i
g
e
n
ce,
p
p
.
1
-
1
2
,
2
0
1
7
.
[2
9
]
C.
Szeg
ed
y
et
a
l
.
,
“G
o
i
n
g
d
ee
p
er
w
i
t
h
c
o
n
v
o
l
u
t
i
o
n
s
,
”
IE
E
E
c
o
n
f
er
e
n
ce
o
n
c
o
m
p
u
t
er
v
i
s
i
o
n
a
n
d
p
a
t
t
er
n
r
eco
g
n
i
t
i
o
n
,
p
p
.
1
-
9
,
2
0
1
5
.
[3
0
]
A
.
K
o
w
al
cz
y
k
,
"S
u
p
p
o
r
t
v
ec
t
o
r
mach
i
n
e,
"
Mo
rri
s
v
i
l
l
e,
N
o
r
t
h
Caro
l
i
n
a:
s
y
n
cfu
s
i
o
n
,
2
0
1
7
.
[3
1
]
A
.
S.
Pras
at
h
,
et
al
.
,
“
D
i
s
t
a
n
ce
an
d
s
i
mi
l
ari
t
y
mea
s
u
re
s
e
ffect
o
n
t
h
e
p
erf
o
rman
ce
o
f
K
-
N
eare
s
t
n
ei
g
h
b
o
r
c
l
as
s
i
f
i
e
r,
”
E
l
s
evi
e
r
,
p
p
.
1
-
53
,
2
0
1
9
.
[3
2
]
J
.
M.
Mart
í
n
ez
-
O
t
zet
a,
et
a
l
.
,
“K
N
eare
s
t
n
e
i
g
h
b
o
r
ed
i
t
i
o
n
t
o
g
u
i
d
e
cl
a
s
s
i
fi
ca
t
i
o
n
t
ree
l
earn
i
n
g
:
Mo
t
i
v
at
i
o
n
an
d
ex
p
er
i
men
t
al
res
u
l
t
s
,
”
Lect
u
r
e
N
o
t
e
s
i
n
C
o
m
p
u
t
er
S
c
i
en
c
e
,
2
0
1
8
,
p
p
.
5
3
-
63
,
2
0
1
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.