I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
10
,
No
.
4
,
Dec
em
b
er
202
1
,
p
p
.
1
0
6
9
~
1
0
78
I
SS
N:
2252
-
8938
,
DOI
: 1
0
.
1
1
5
9
1
/i
j
ai.
v
10
.i
4
.
p
p
1
0
6
9
-
1
0
78
1069
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
A per
for
m
a
n
ce e
v
a
lua
tion o
f
co
nv
o
lutiona
l neural n
et
w
o
rk
a
rchitec
t
ure
for c
la
ss
ificatio
n of ric
e leaf disea
se
Af
is
J
uli
a
nto
,
Andi
Su
ny
o
t
o
M
a
g
ister o
f
In
f
o
r
m
a
ti
c
s E
n
g
in
e
e
ri
n
g
,
Un
iv
e
rsitas
A
M
IKO
M
Yo
g
y
a
k
a
rta,
De
p
o
k
,
S
lem
a
n
,
Yo
g
y
a
k
a
rt
a
,
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
Mar
28
,
2
0
2
1
R
ev
i
s
ed
A
u
g
28
,
2
0
2
1
A
cc
ep
ted
Sep
17
,
2
0
2
1
P
la
n
t
d
ise
a
se
is
a
c
h
a
ll
e
n
g
e
in
th
e
a
g
ricu
lt
u
ra
l
se
c
to
r,
e
sp
e
c
ially
f
o
r
rice
p
ro
d
u
c
ti
o
n
.
I
d
e
n
ti
f
y
in
g
d
ise
a
se
s
i
n
rice
lea
v
e
s
is
th
e
f
irst
ste
p
to
wip
e
o
u
t
a
n
d
trea
t
d
ise
a
se
s
to
re
d
u
c
e
c
ro
p
f
a
il
u
re
.
W
it
h
th
e
ra
p
id
d
e
v
e
lo
p
m
e
n
t
o
f
th
e
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
t
w
o
rk
(C
NN
),
rice
le
a
f
d
ise
a
se
c
a
n
b
e
r
e
c
o
g
n
ize
d
w
e
ll
w
it
h
o
u
t
t
h
e
h
e
lp
o
f
a
n
e
x
p
e
rt.
In
th
is
re
se
a
rc
h
,
th
e
p
e
rf
o
rm
a
n
c
e
e
v
a
lu
a
ti
o
n
o
f
CNN
a
rc
h
it
e
c
tu
re
w
il
l
b
e
c
a
rried
o
u
t
t
o
a
n
a
ly
z
e
th
e
c
las
sif
ic
a
ti
o
n
o
f
rice
lea
f
d
ise
a
se
i
m
a
g
e
s
b
y
c
las
si
fy
in
g
5
9
3
2
im
a
g
e
d
a
t
a
w
h
ich
a
re
d
iv
i
d
e
d
in
t
o
4
d
ise
a
se
c
las
se
s.
T
h
e
c
o
m
p
a
riso
n
o
f
train
in
g
d
a
ta,
v
a
li
d
a
ti
o
n
,
a
n
d
tes
ti
n
g
a
re
6
0
:
2
0
:
2
0
.
A
d
a
m
o
p
ti
m
iza
ti
o
n
with
a
lea
rn
in
g
ra
te
o
f
0
.
0
0
0
9
a
n
d
so
f
tm
a
x
a
c
ti
v
a
ti
o
n
w
a
s
u
se
d
in
th
is
st
u
d
y
.
F
ro
m
th
e
e
x
p
e
ri
m
e
n
tal
re
su
lt
s,
th
e
In
c
e
p
ti
o
n
V
3
a
n
d
In
c
e
p
t
io
n
Re
sn
e
tV
2
a
rc
h
i
tec
tu
re
s
g
o
t
th
e
b
e
st
a
c
c
u
ra
c
y
,
n
a
m
e
l
y
1
0
0
%
,
Re
sN
e
t5
0
a
n
d
De
n
se
Ne
t2
0
1
g
o
t
9
9
.
8
3
%
,
M
o
b
il
e
N
e
t
9
9
.
3
3
%
,
a
n
d
Ef
f
icie
n
tNe
tB3
9
0
.
1
4
%
a
c
c
u
r
a
c
y
.
K
ey
w
o
r
d
s
:
C
NN
ar
ch
i
tectu
r
e
I
m
ag
e
cla
s
s
i
f
icatio
n
P
er
f
o
r
m
a
n
ce
ev
al
u
atio
n
R
ice
leaf
d
i
s
ea
s
e
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
:
An
d
i S
u
n
y
o
to
Ma
g
is
ter
o
f
I
n
f
o
r
m
atics E
n
g
in
ee
r
in
g
Un
i
v
er
s
ita
s
Am
ik
o
m
Yo
g
y
ak
a
r
ta
Dep
o
k
,
Sle
m
a
n
,
Yo
g
y
ak
ar
ta
5
5
2
8
1
,
I
n
d
o
n
esia
E
m
ail: a
n
d
i@
a
m
i
k
o
m
.
ac
.
id
1.
I
NT
RO
D
UCT
I
O
N
R
ice
i
s
o
n
e
o
f
t
h
e
f
o
o
d
co
m
m
o
d
itie
s
f
o
r
m
o
s
t
p
eo
p
le
in
w
o
r
ld
.
Fo
r
e
x
a
m
p
le
I
n
d
o
n
esia
,
r
ic
e
p
r
o
d
u
ctio
n
in
2
0
1
9
am
o
u
n
ted
to
5
4
.
6
0
m
illi
o
n
to
n
n
e
s
o
f
m
illed
d
r
y
u
n
h
u
l
led
w
i
th
a
h
ar
v
est
ar
ea
o
f
1
0
.
6
8
m
illi
o
n
h
ec
tar
es
[
1
]
.
Ho
w
ev
e
r
,
r
ice
p
lan
ts
ar
e
o
f
te
n
e
x
p
o
s
ed
to
th
e
d
i
s
ea
s
e
d
u
r
in
g
t
h
ei
r
g
r
o
w
in
g
p
er
io
d
.
Dis
ea
s
e
in
r
ice
f
r
eq
u
en
t
l
y
a
tt
ac
k
s
th
e
leav
e
s
.
F
u
n
g
i
an
d
b
ac
ter
ia
ar
e
b
eliev
ed
to
b
e
t
h
e
m
ai
n
ca
u
s
e
s
o
f
t
h
e
d
is
ea
s
e
[
2
]
.
So
m
e
o
f
th
e
d
i
s
ea
s
es
t
h
at
o
f
te
n
attac
k
r
ice
leav
e
s
ar
e
B
r
o
w
n
Sp
o
t,
T
u
n
g
r
o
,
B
ac
ter
ial
L
ea
f
B
lig
h
t,
an
d
B
last
.
R
ice
lea
f
d
is
ea
s
e
ca
n
r
esu
lt in
lo
w
l
y
r
ice
p
lan
t
g
r
o
w
t
h
,
r
es
u
lti
n
g
i
n
d
ec
r
ea
s
ed
p
r
o
d
u
ctio
n
.
Fro
m
o
n
e
s
t
u
d
y
,
it
i
s
e
s
ti
m
at
ed
th
at
th
e
d
ec
r
ea
s
e
i
n
r
ice
p
r
o
d
u
ctio
n
d
u
e
to
b
last
d
i
s
ea
s
e
ca
n
r
ea
c
h
100%
[
3
]
,
b
ac
ter
ial
leaf
b
lig
h
t
1
5
-
25%
[
4
]
,
an
d
b
r
o
w
n
lea
f
s
p
o
t
4
0
%
[
5
]
.
I
f
n
o
t
h
an
d
le
d
s
er
io
u
s
l
y
,
it
w
ill
b
r
in
g
h
u
g
e
ec
o
n
o
m
ic
lo
s
s
e
s
f
o
r
r
ice
f
ar
m
er
s
.
So
m
e
f
ar
m
er
s
,
in
g
en
er
a
l,
d
o
n
o
t
k
n
o
w
e
n
o
u
g
h
ab
o
u
t
r
ice
leaf
d
is
ea
s
es,
esp
ec
iall
y
y
o
u
n
g
f
ar
m
er
s
w
h
o
lack
e
x
p
er
tis
e
i
n
ag
r
icu
lt
u
r
e
s
o
th
at
i
t
is
d
i
f
f
icu
l
t
to
id
en
tify
t
h
e
t
y
p
e
o
f
d
is
ea
s
e.
W
ith
o
u
t
k
n
o
w
i
n
g
th
e
t
y
p
e
o
f
d
is
ea
s
e,
it
w
ill
b
e
d
if
f
icu
lt
to
ch
o
o
s
e
s
u
itab
l
e
d
r
u
g
s
an
d
g
r
ea
t
h
an
d
li
n
g
p
r
o
ce
d
u
r
es.
T
h
er
ef
o
r
e,
it is
n
ec
es
s
ar
y
to
b
e
ab
le
to
class
i
f
y
th
e
t
y
p
es o
f
d
is
ea
s
es i
n
r
ice
leav
es.
I
n
r
ec
en
t
y
ea
r
s
,
m
an
y
s
tu
d
i
es
h
a
v
e
b
ee
n
co
n
d
u
cted
o
n
th
e
cla
s
s
i
f
icatio
n
o
f
r
ice
le
af
d
is
ea
s
e.
R
esear
ch
h
as
b
ee
n
co
n
d
u
c
ted
u
s
i
n
g
th
e
k
-
n
ea
r
est
n
e
ig
h
b
o
r
(
KNN)
m
et
h
o
d
to
class
if
y
B
la
s
t
an
d
B
r
o
w
n
Sp
o
t
d
is
ea
s
e
i
n
r
ice
lea
v
es
o
b
tain
e
d
an
ac
c
u
r
ac
y
o
f
7
6
.
5
9
%
[
6
]
.
Oth
er
s
tu
d
ie
s
h
a
v
e
b
ee
n
co
n
d
u
cted
to
d
etec
t
lea
f
s
m
u
t,
b
ac
ter
ial
leaf
b
lig
h
t,
an
d
b
r
o
w
n
s
p
o
ts
o
n
r
ice
leav
es
.
T
h
e
s
tu
d
y
co
n
d
u
c
ted
an
ex
p
er
im
e
n
t
u
s
i
n
g
t
h
e
lo
g
is
tic
r
e
g
r
ess
io
n
al
g
o
r
ith
m
,
o
b
tain
ed
an
ac
cu
r
ac
y
o
f
7
0
.
8
3
%,
KNN
9
1
.
6
7
%,
9
7
.
9
1
% d
ec
is
io
n
tr
ee
(
DT
)
,
an
d
5
0
%
n
aïv
e
b
a
y
e
s
(
NB
)
[
7
]
.
Fu
r
th
er
r
esear
ch
w
as
ca
r
r
ied
o
u
t
u
s
i
n
g
th
e
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
m
eth
o
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
10
,
No
.
4
,
Dec
em
b
er
202
1
:
1
0
6
9
-
1
0
7
8
1070
to
d
iag
n
o
s
e
B
r
o
w
n
Sp
o
t
an
d
B
last
d
is
ea
s
e
i
n
r
ice
lea
v
es
with
a
to
tal
d
ata
o
f
1
2
0
,
o
b
tain
ed
an
ac
cu
r
ac
y
o
f
9
5
.
5
%
[
8
]
.
S
VM
is
also
u
s
ed
f
o
r
th
e
clas
s
i
f
icatio
n
p
r
o
ce
s
s
with
t
h
e
ad
d
itio
n
o
f
th
e
k
-
m
ea
n
s
clu
s
ter
i
n
g
m
et
h
o
d
an
d
o
b
tain
ed
an
ac
cu
r
ac
y
o
f
9
2
.
0
6
%
[
9
]
.
T
h
e
ar
tif
icial
n
eu
r
al
n
et
w
o
r
k
(
ANN)
m
et
h
o
d
h
a
s
b
ee
n
u
s
ed
f
o
r
th
e
class
i
f
icatio
n
o
f
r
ice
leaf
d
is
ea
s
e
b
y
s
e
g
m
e
n
ti
n
g
an
d
f
ea
tu
r
e
ex
tr
ac
tio
n
o
n
i
m
a
g
es
[
1
0
]
.
Ho
w
e
v
er
,
to
g
et
a
g
o
o
d
lev
el
o
f
ac
cu
r
ac
y
s
t
ill
d
ep
en
d
s
o
n
th
e
f
ea
tu
r
e
s
elec
tio
n
tech
n
iq
u
e
w
h
e
n
u
s
i
n
g
th
is
m
et
h
o
d
.
R
ec
e
n
t
r
esear
ch
o
n
co
n
v
o
l
u
tio
n
al
n
e
u
r
al
n
et
w
o
r
k
(
C
NN)
h
as
co
n
tr
ib
u
ted
g
r
e
atl
y
to
i
m
ag
e
-
b
ased
id
en
ti
f
icatio
n
b
y
eli
m
i
n
ati
n
g
t
h
e
n
ee
d
f
o
r
p
r
e
-
p
r
o
ce
s
s
i
n
g
i
m
ag
e
s
a
n
d
h
a
v
i
n
g
b
u
ilt
-
i
n
f
e
atu
r
e
s
elec
t
io
n
[
1
1
]
.
R
esear
ch
u
s
in
g
t
h
e
C
NN
m
et
h
o
d
h
as
b
ee
n
[
1
2
]
ca
r
r
ied
o
u
t
to
id
en
ti
f
y
1
0
d
is
ea
s
e
c
lass
e
s
in
r
ice
lea
v
es
a
n
d
s
te
m
s
u
s
i
n
g
5
0
0
i
m
a
g
e
d
ata,
f
r
o
m
th
e
tes
t
r
es
u
lts
o
b
tain
e
d
an
ac
c
u
r
ac
y
o
f
9
5
.
4
8
%
u
s
i
n
g
1
0
Fo
ld
C
r
o
s
s
-
Valid
atio
n
.
Oth
er
s
t
u
d
ies
w
er
e
also
co
n
d
u
cted
u
s
i
n
g
C
NN
ar
ch
itect
u
r
al
m
o
d
els
s
u
ch
a
s
VG
G
1
6
w
it
h
9
2
.
4
6
%
ac
cu
r
ac
y
[
1
1
]
,
A
le
x
Net
9
1
.
2
3
%
ac
cu
r
ac
y
[
1
3
]
,
an
d
VGG
1
9
w
it
h
9
2
%
ac
cu
r
ac
y
[
1
4
]
.
I
n
r
ec
en
t
y
ea
r
s
,
m
a
n
y
r
esear
ch
er
s
h
av
e
b
ee
n
d
e
v
elo
p
in
g
n
e
w
C
N
N
ar
ch
i
tectu
r
e.
So
it
i
s
n
ec
ess
ar
y
to
r
esear
c
h
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
n
e
w
C
NN
ar
c
h
itect
u
r
al
m
o
d
el
f
o
r
th
e
class
i
f
ica
tio
n
o
f
r
ic
e
leaf
d
is
ea
s
e.
T
h
is
p
ap
er
w
ill
co
n
d
u
ct
a
s
t
u
d
y
t
h
at
f
o
cu
s
e
s
o
n
th
e
cla
s
s
i
f
icatio
n
o
f
4
t
y
p
es
o
f
r
ice
leaf
d
is
ea
s
es,
n
a
m
e
l
y
B
ac
ter
ial
L
ea
f
B
lig
h
t,
B
last
,
T
u
n
g
r
o
,
an
d
B
r
o
w
n
S
p
o
t
u
s
i
n
g
6
t
y
p
e
s
o
f
C
NN
ar
ch
itect
u
r
es,
n
a
m
el
y
I
n
ce
p
tio
n
V3
,
R
e
s
Net5
0
,
I
n
ce
p
tio
n
R
es
n
etV2
,
De
n
s
e
Net2
0
1
,
Mo
b
ileNet,
an
d
E
f
f
icien
tNe
tB
3
.
B
y
co
m
p
ar
i
n
g
all
th
e
C
N
N
ar
ch
i
tectu
r
e
s
,
p
er
f
o
r
m
an
ce
ev
al
u
atio
n
w
i
ll
b
e
ca
r
r
ied
o
u
t
to
s
ee
t
h
e
b
est
w
a
y
to
r
ec
o
g
n
ize
r
ice
leaf
d
is
ea
s
e.
2.
RE
S
E
ARCH
M
E
T
H
O
D
T
h
is
s
ec
tio
n
ex
p
lai
n
s
t
h
e
r
ese
ar
ch
f
lo
w
r
eg
ar
d
in
g
t
h
e
c
lass
i
f
icatio
n
o
f
r
ice
lea
f
d
is
ea
s
e
s
t
ar
tin
g
f
r
o
m
th
e
ac
q
u
i
s
itio
n
o
f
d
atasets
,
p
r
e
-
p
r
o
ce
s
s
in
g
d
ata,
th
e
u
s
e
o
f
th
e
co
n
v
o
lu
t
io
n
al
n
eu
r
a
l
n
et
w
o
r
k
(
C
N
N)
ar
ch
itect
u
r
al
m
o
d
el,
an
d
ev
a
l
u
atio
n
o
f
t
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
C
NN
ar
ch
itect
u
r
al
m
o
d
el
as
s
h
o
w
n
in
F
ig
u
r
e
1
.
Fig
u
r
e
1
.
R
esear
ch
f
lo
w
2
.
1
.
Da
t
a
s
et
I
n
t
h
is
s
t
u
d
y
,
th
e
d
ataset
u
s
e
d
w
a
s
ta
k
en
f
r
o
m
Me
n
d
ele
y
Data
w
h
ich
is
a
co
llectio
n
o
f
p
ictu
r
es
o
f
r
ice
leaf
d
i
s
ea
s
e
[
1
5
]
.
T
h
is
d
ataset
h
as
5
9
3
2
r
ice
lea
f
d
is
ea
s
e
i
m
ag
e
d
ata
co
n
s
is
t
i
n
g
o
f
4
d
is
ea
s
e
cla
s
s
es
in
cl
u
d
in
g
B
ac
ter
ial
B
li
g
h
t,
B
last
,
B
o
w
n
Sp
o
t,
an
d
T
u
n
g
r
o
d
is
ea
s
e
a
s
s
h
o
w
n
i
n
F
ig
u
r
e
2
.
A
ll
i
m
ag
e
d
ata
o
b
tain
ed
in
th
i
s
s
t
u
d
y
ar
e
s
to
r
e
d
in
J
P
G
f
o
r
m
at.
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
2
.
T
y
p
es o
f
d
is
ea
s
e
o
n
r
ice
leav
es;
(
a)
B
last
,
(
b
)
B
ac
t
er
ia
B
lig
h
t
,
(
c)
T
u
n
g
r
o
,
(
d
)
B
r
o
w
n
Sp
o
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
A
p
erfo
r
ma
n
ce
ev
a
lu
a
tio
n
o
f
co
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
a
r
ch
itectu
r
e
fo
r
.
.
.
(
A
fis
Ju
lia
n
to
)
1071
2
.
2
.
P
re
-
pro
ce
s
s
ing
da
t
a
T
h
e
d
ataset
is
d
iv
id
ed
in
to
6
0
%
tr
ain
i
n
g
,
2
0
%
v
alid
atio
n
,
an
d
2
0
%
test
in
g
.
T
h
en
ea
ch
i
m
ag
e
w
ill
b
e
r
esized
to
3
0
0
x
3
0
0
p
ix
els.
I
n
th
e
tr
ai
n
i
n
g
d
ata,
a
n
a
u
g
m
e
n
ta
tio
n
p
r
o
ce
s
s
w
ill
b
e
ca
r
r
ied
o
u
t
b
y
r
o
tati
n
g
u
p
to
4
0
d
eg
r
ee
s
,
s
h
i
f
ti
n
g
t
h
e
i
m
ag
e
to
a
s
ca
le
o
f
0
.
2
,
en
lar
g
i
n
g
t
h
e
i
m
a
g
e
to
a
s
ca
le
o
f
0
.
2
,
an
d
f
lip
p
in
g
t
h
e
i
m
a
g
e
v
er
ticall
y
a
n
d
h
o
r
izo
n
tall
y
.
S
o
th
at
th
e
n
u
m
b
er
o
f
i
m
a
g
es
in
th
e
tr
ain
i
n
g
d
ata
in
cr
ea
s
e
s
6
tim
e
s
in
cl
u
d
i
n
g
i
m
a
g
es
t
h
at
ap
p
l
y
au
g
m
en
tat
i
o
n
.
Fig
u
r
e
3
s
h
o
w
s
t
h
e
s
a
m
p
l
e
d
ata
f
r
o
m
t
h
e
au
g
m
en
ta
tio
n
r
esu
lt
s
.
T
h
e
d
is
ea
s
e
n
a
m
e
a
n
d
th
e
d
ata
a
m
o
u
n
t u
s
e
d
in
th
i
s
s
t
u
d
y
ca
n
b
e
s
ee
n
i
n
T
ab
le
1
.
Fig
u
r
e
3
.
Au
g
m
e
n
tatio
n
r
esu
l
t
s
f
r
o
m
t
h
e
i
m
ag
e
o
f
b
last
d
is
ea
s
e
T
ab
le
1
.
Deta
ils
o
f
r
ice
leaf
d
is
ea
s
e
d
ataset
L
e
a
f
d
i
se
a
se
s
O
r
i
g
i
n
a
l
i
mag
e
T
r
a
i
n
i
n
g
(
6
0
%)
V
a
l
i
d
a
t
i
o
n
(
2
0
%)
T
e
st
i
n
g
(
2
0
%)
B
e
f
o
r
e
a
u
g
me
n
t
a
t
i
o
n
A
f
t
e
r
a
u
g
me
n
t
a
t
i
o
n
B
a
c
t
e
r
i
l
B
l
i
g
h
t
1
5
8
4
1
2
6
7
7
6
0
2
3
1
7
3
1
7
B
l
a
st
1
4
4
0
1
1
5
2
6
9
1
2
2
8
8
2
8
8
B
r
o
w
n
S
p
o
t
1
6
0
0
1
2
8
0
7
6
8
0
3
2
0
3
2
0
T
u
n
g
r
o
1
3
0
8
1
0
4
6
6
2
7
6
2
6
2
2
6
2
T
o
t
a
l
5
9
3
2
4
7
4
5
2
8
4
7
0
1
1
8
7
1
1
8
7
2
.
3
.
M
o
del C
NN
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
et
w
o
r
k
s
(
C
NNs)
ar
e
t
h
e
m
o
s
t
p
o
p
u
lar
d
ee
p
lear
n
in
g
al
g
o
r
ith
m
s
f
o
r
r
esear
ch
er
s
to
test
[
1
6
]
.
C
NN
w
a
s
f
ir
s
t in
tr
o
d
u
ce
d
in
1
9
8
8
b
y
Yan
n
L
eC
u
n
[
1
7
]
.
T
h
e
in
tr
o
d
u
ctio
n
o
f
C
NN
h
as
ch
an
g
ed
th
e
w
a
y
p
r
o
b
le
m
s
ar
e
s
o
lv
ed
in
th
e
clas
s
i
f
icatio
n
o
f
an
i
m
a
g
e
[
1
8
]
.
I
n
th
is
s
tu
d
y
,
s
ix
t
y
p
e
s
o
f
C
N
N
ar
ch
itect
u
r
e
w
ill
b
e
u
s
ed
to
co
n
d
u
ct
e
x
p
er
i
m
e
n
ts
in
c
lass
if
y
i
n
g
d
i
s
ea
s
es
in
r
ice
lea
v
e
s
,
in
cl
u
d
in
g
I
n
ce
p
tio
n
V3
[
1
9
]
,
R
esNet5
0
[
2
0
]
,
I
n
ce
p
tio
n
R
e
s
n
et
V2
[
2
1
]
,
Den
s
eNe
t2
0
1
[
2
2
]
,
M
o
b
ileNet
[
2
3
]
,
an
d
E
f
f
icien
tNetB
3
[
2
4
]
.
T
h
e
f
o
llo
w
i
n
g
w
ill b
r
ief
l
y
ex
p
lain
t
h
e
C
NN
ar
c
h
itect
u
r
e
u
s
e
d
in
th
i
s
s
t
u
d
y
.
2
.
3
.
1
.
I
ncept
io
nV3
I
n
ce
p
tio
n
V3
i
s
a
C
N
N
ar
ch
itectu
r
e
d
e
v
elo
p
ed
b
y
Go
o
g
le
at
t
h
e
i
m
ag
e
n
et
lar
g
e
s
c
ale
v
i
s
u
a
l
r
ec
o
g
n
itio
n
ch
alle
n
g
e
(
I
L
SV
R
C
)
in
2012
[
1
9
]
.
I
n
ce
p
tio
n
V3
w
a
s
d
ev
elo
p
ed
to
p
a
r
s
e
co
n
v
o
lu
tio
n
[
1
9
]
.
T
h
is
m
ea
n
s
t
h
at
ea
c
h
co
n
v
o
l
u
tio
n
ca
n
b
e
r
ep
lace
d
b
y
a
co
n
v
o
l
u
tio
n
f
o
llo
w
ed
b
y
a
co
n
v
o
lu
t
io
n
,
th
i
s
ca
n
p
ar
s
e
m
an
y
p
ar
a
m
e
ter
s
,
av
o
id
t
h
e
p
r
o
b
lem
o
f
r
ed
u
n
d
a
n
t
f
itti
n
g
a
n
d
s
tr
en
g
t
h
en
th
e
ab
ili
t
y
o
f
n
o
n
lin
ea
r
e
x
p
r
ess
io
n
s
[
2
5
]
.
2
.
3
.
2
.
ResNet
5
0
R
esNet5
0
is
a
C
NN
ar
ch
itec
tu
r
e
th
at
i
n
tr
o
d
u
ce
s
t
h
e
co
n
c
ep
t
o
f
s
h
o
r
tcu
t
co
n
n
ec
tio
n
s
[
2
0
]
.
T
h
e
co
n
ce
p
t
em
er
g
e
n
ce
o
f
s
h
o
r
tcu
t
co
n
n
ec
tio
n
s
in
th
e
R
esNet5
0
ar
ch
itectu
r
e
is
r
elate
d
to
th
e
v
an
i
s
h
in
g
g
r
ad
ie
n
t
p
r
o
b
lem
th
at
o
cc
u
r
s
w
h
e
n
e
f
f
o
r
ts
to
d
ee
p
en
th
e
s
tr
u
ctu
r
e
o
f
a
n
et
w
o
r
k
ar
e
ca
r
r
ied
o
u
t.
Ho
w
e
v
er
,
d
ee
p
en
in
g
a
n
et
w
o
r
k
w
it
h
t
h
e
ai
m
o
f
i
m
p
r
o
v
in
g
its
p
er
f
o
r
m
an
ce
ca
n
n
o
t
b
e
d
o
n
e
s
i
m
p
l
y
b
y
s
tack
i
n
g
la
y
er
s
.
T
h
e
d
ee
p
er
a
n
et
w
o
r
k
ca
n
lead
to
a
v
an
is
h
i
n
g
g
r
ad
ien
t
p
r
o
b
le
m
th
at
ca
n
m
ak
e
th
e
s
m
all
g
r
ad
ie
n
t
w
h
ic
h
r
esu
lt
s
in
d
ec
r
ea
s
ed
p
er
f
o
r
m
a
n
ce
o
r
ac
cu
r
ac
y
[
20]
.
2
.
3
.
3
.
I
ncept
io
nResn
et
V2
I
n
ce
p
tio
n
R
e
s
NetV2
w
a
s
f
ir
s
t
i
n
tr
o
d
u
ce
d
b
y
Sze
g
ed
y
et
a
l
.
[
2
1
]
.
I
n
ce
p
tio
n
R
esNetV2
ar
ch
i
tectu
r
e
is
a
co
m
p
o
u
n
d
o
f
th
e
I
n
ce
p
tio
n
s
tr
u
ct
u
r
e
an
d
r
esid
u
al
m
o
d
u
l
e.
T
h
e
co
n
v
o
lu
tio
n
f
ilter
is
co
m
b
i
n
ed
w
ith
t
h
e
r
esid
u
al
co
n
n
ec
tio
n
w
h
ic
h
ai
m
s
to
a
v
o
id
th
e
p
r
o
b
le
m
s
ca
u
s
ed
b
y
th
e
d
ee
p
er
s
tr
u
ct
u
r
e.
R
esid
u
al
co
n
n
ec
tio
n
ca
n
r
ed
u
ce
t
h
e
ti
m
e
d
u
r
i
n
g
t
h
e
tr
ai
n
in
g
p
r
o
ce
s
s
[
2
6
]
.
I
n
ce
p
tio
n
R
es
n
etV1
a
n
d
I
n
ce
p
tio
n
R
es
n
etV2
h
av
e
th
e
s
a
m
e
o
v
er
all
s
tr
u
ct
u
r
e,
b
u
t d
if
f
er
en
t
m
o
d
u
le
s
in
t
h
e
n
e
t
w
o
r
k
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
10
,
No
.
4
,
Dec
em
b
er
202
1
:
1
0
6
9
-
1
0
7
8
1072
2
.
3
.
4
.
DenseNet
2
0
1
Den
s
eNe
t
is
a
C
NN
ar
ch
itect
u
r
e
th
at
in
tr
o
d
u
ce
s
s
h
o
r
t
co
n
n
e
ctio
n
s
t
h
at
co
n
n
ec
t
ea
c
h
la
y
er
d
ir
ec
tl
y
to
th
e
o
t
h
er
la
y
er
s
i
n
a
f
ee
d
-
f
o
r
war
d
m
an
n
er
.
De
n
s
eNe
t
h
a
s
a
n
ar
r
o
w
la
y
er
w
it
h
a
s
m
all
s
et
o
f
f
ea
t
u
r
e
m
ap
s
t
h
at
b
elo
n
g
to
th
e
n
et
w
o
r
k
[
2
2
]
.
2
.
3
.
5
.
M
o
bil
eNe
t
Mo
b
ileNet
w
as
f
ir
s
t
i
n
tr
o
d
u
c
ed
b
y
r
esear
ch
er
s
f
r
o
m
Go
o
g
le
[
2
3
]
,
to
o
v
er
co
m
e
t
h
e
n
ee
d
f
o
r
lar
g
e
co
m
p
u
ti
n
g
r
eso
u
r
ce
s
s
o
th
at
t
h
is
ar
c
h
itect
u
r
e
ca
n
b
e
u
s
ed
f
o
r
m
o
b
ile
p
h
o
n
es.
Mo
b
ileNet
u
s
e
s
a
co
n
v
o
lu
tio
n
s
cr
ee
n
w
it
h
a
f
il
ter
th
ic
k
n
e
s
s
th
at
m
atch
e
s
th
e
t
h
ic
k
n
ess
o
f
th
e
in
p
u
t.
T
h
e
Mo
b
ileNet
ar
ch
itect
u
r
e
is
d
iv
id
ed
in
to
2
t
y
p
es o
f
co
n
v
o
lu
tio
n
,
n
a
m
el
y
d
ep
t
h
w
is
e
co
n
v
o
l
u
tio
n
a
n
d
p
o
in
t
w
is
e
co
n
v
o
l
u
tio
n
.
2
.
3
.
6
.
E
f
f
icienNet
B
3
E
f
f
icien
Net
w
as
f
ir
s
t
p
r
o
p
o
s
ed
b
y
T
an
an
d
L
ee
i
n
2
0
1
9
[
2
4
]
an
d
an
ar
ch
itect
u
r
e
u
s
ed
t
o
o
p
tim
iz
e
class
i
f
icatio
n
n
et
w
o
r
k
s
.
I
n
g
e
n
er
al,
th
er
e
ar
e
3
in
d
icato
r
s
u
s
ed
b
y
m
o
s
t
n
et
w
o
r
k
s
,
in
cl
u
d
in
g
w
id
en
i
n
g
th
e
n
et
w
o
r
k
,
d
ee
p
en
i
n
g
th
e
n
et
wo
r
k
an
d
in
cr
ea
s
in
g
t
h
e
r
eso
l
u
tio
n
q
u
alit
y
.
T
h
er
ef
o
r
e,
th
e
ap
p
licatio
n
o
f
t
h
e
co
m
b
i
n
ed
s
ca
li
n
g
m
o
d
el
i
s
ap
p
lied
to
o
p
tim
ize
th
e
n
et
w
o
r
k
w
id
t
h
,
d
ep
th
a
n
d
n
et
w
o
r
k
r
es
o
lu
tio
n
to
i
m
p
r
o
v
e
ac
cu
r
ac
y
[
2
5
]
.
P
r
ev
io
u
s
l
y
,
t
h
e
m
o
d
el
u
s
ed
i
n
th
is
e
x
p
er
i
m
e
n
t
w
a
s
tr
ain
ed
u
s
i
n
g
a
d
ataset
f
r
o
m
I
m
a
g
eN
et.
A
ll
p
r
e
-
tr
ain
ed
m
o
d
els
o
n
t
h
e
C
N
N
a
r
ch
itect
u
r
e
b
y
d
ef
a
u
lt
h
a
v
e
1
0
0
0
f
u
ll
y
co
n
n
ec
ted
(
F
C
)
la
y
er
o
u
tp
u
t
n
o
d
es.
Fo
r
th
e
o
u
tp
u
t
F
C
la
y
er
,
it
w
i
ll
b
e
r
ep
lace
d
w
ith
4
n
o
d
es
ac
co
r
d
in
g
to
th
e
n
u
m
b
er
o
f
class
es
i
n
r
ice
leaf
d
is
ea
s
e
an
d
ad
d
ed
w
it
h
So
f
t
m
ax
ac
t
iv
atio
n
.
2
.
4
.
E
v
a
lua
t
i
o
n o
f
CNN
a
rc
hite
ct
ure
m
o
del
T
o
ev
alu
ate
an
d
co
m
p
ar
e
th
e
p
er
f
o
r
m
a
n
ce
o
f
a
tr
ied
C
NN
a
r
ch
itect
u
r
e.
T
h
e
f
ir
s
t
ev
a
lu
at
io
n
i
s
ca
r
r
ied
o
u
t
o
n
tr
ai
n
i
n
g
d
ata
an
d
v
alid
atio
n
b
y
ca
lcu
la
tin
g
ac
c
u
r
ac
y
an
d
lo
s
s
a
n
d
ca
lc
u
lati
n
g
t
h
e
c
o
m
p
u
tatio
n
ti
m
e
o
f
ea
ch
C
NN
ar
c
h
itect
u
r
e.
Fu
r
t
h
er
m
o
r
e,
an
ev
a
lu
at
io
n
o
f
t
h
e
ar
ch
itectu
r
al
m
o
d
el
t
h
at
h
a
s
b
ee
n
tr
ain
ed
w
it
h
test
i
n
g
d
ata
u
s
i
n
g
t
h
e
C
o
n
f
u
s
i
o
n
Ma
tr
ix
w
il
l
b
e
ca
r
r
ied
o
u
t,
in
clu
d
i
n
g
ca
lcu
la
tin
g
ac
c
u
r
ac
y
,
p
r
ec
i
s
io
n
,
r
ec
all,
an
d
F1
Sco
r
e.
T
h
e
f
o
llo
w
in
g
c
o
n
f
u
s
io
n
m
atr
i
x
m
u
lticlas
s
u
s
e
d
in
th
i
s
s
t
u
d
y
is
s
h
o
w
n
i
n
T
ab
le
2
.
T
ab
le
2
.
C
o
n
f
u
s
io
n
m
atr
i
x
f
o
r
m
u
lticla
s
s
P
r
e
d
i
c
t
e
d
N
u
m
b
e
r
C
l
a
ss
1
C
l
a
ss
2
…
C
l
a
ss
n
A
c
t
u
a
l
N
u
m
b
e
r
C
l
a
ss
1
X
11
X
12
…
X
1n
C
l
a
ss
2
X
21
X
22
…
X
2n
…
…
…
…
.
C
l
a
ss
n
X
n1
X
n2
…
X
nn
Fro
m
t
h
e
T
ab
le
2
,
w
e
w
i
ll
g
et
th
e
n
u
m
b
er
o
f
tr
u
e
p
o
s
iti
v
es
(
T
T
P
)
f
o
r
all
clas
s
es,
tr
u
e
n
e
g
ati
v
e
(
T
T
N)
,
f
alse p
o
s
itiv
e
(
T
FP
)
,
a
n
d
f
al
s
e
n
e
g
ati
v
e
(
T
FN)
f
o
r
ea
ch
class
i
w
h
ic
h
is
ca
lc
u
lated
u
s
i
n
g
(
1
)
-
(
4
)
[
2
7
]
.
=
∑
=
1
(
1
)
=
∑
∑
=
1
≠
=
1
≠
(
2
)
=
∑
=
1
≠
(
3
)
=
∑
=
1
≠
(
4
)
Fro
m
th
e
T
PP
,
T
T
N,
T
F
P
an
d
T
FN
r
esu
lts
,
th
e
n
it
w
il
l
b
e
u
s
ed
to
ca
lcu
late
th
e
o
v
er
all
ac
cu
r
ac
y
(
OA
)
,
p
r
ec
is
io
n
(
P
)
,
r
ec
all
(
R
)
,
f
-
m
ea
s
u
r
e
s
co
r
e
(
F1
)
o
f
ea
ch
class
i c
alc
u
lated
u
s
in
g
(
5
)
-
(
8
)
[
2
7
]
,
[
2
8
]
.
=
(
)
(
5
)
=
(
)
(
+
)
(
6
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
A
p
erfo
r
ma
n
ce
ev
a
lu
a
tio
n
o
f
co
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
a
r
ch
itectu
r
e
fo
r
.
.
.
(
A
fis
Ju
lia
n
to
)
1073
=
(
)
(
+
)
(
7
)
=
2
(
)
(
)
+
(
8
)
I
n
th
is
s
t
u
d
y
,
t
h
e
ex
p
er
i
m
e
n
t
w
as
ca
r
r
ied
o
u
t
b
y
ap
p
l
y
in
g
b
atch
s
ize
3
2
t
o
th
e
tr
ain
ed
m
o
d
el
a
n
d
b
atch
s
ize
1
to
th
e
test
.
T
h
e
d
ata
r
an
d
o
m
izatio
n
p
r
o
c
ess
w
a
s
u
s
ed
d
u
r
i
n
g
th
e
tr
ain
in
g
.
A
d
a
m
o
p
ti
m
izer
[
2
9
]
w
it
h
a
lear
n
i
n
g
r
ate
o
f
0
.
0
0
0
9
w
a
s
u
s
ed
to
m
i
n
i
m
ize
t
h
e
lo
s
s
f
u
n
ctio
n
o
n
t
h
e
C
NN
m
o
d
el
d
u
r
in
g
t
h
e
tr
ai
n
i
n
g
p
r
o
ce
s
s
.
A
ll e
x
p
er
i
m
e
n
t
s
co
n
d
u
cted
in
t
h
i
s
s
t
u
d
y
u
s
e
T
en
s
o
r
Flo
w
w
h
ic
h
i
s
o
n
Go
o
g
le
C
o
la
b
o
r
ato
r
y
as a
clo
u
d
co
m
p
u
ti
n
g
p
r
o
v
id
er
p
latf
o
r
m
t
h
at
h
a
s
1
2
Gb
R
A
M
s
p
ec
if
icat
io
n
s
an
d
a
n
Nv
id
ia
K8
0
s
GP
U
.
3.
RE
SU
L
T
S
A
ND
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
w
i
ll
d
escr
ib
e
th
e
r
esu
lt
s
o
f
ea
c
h
ex
p
er
i
m
en
t
was
co
n
d
u
cted
.
E
x
p
er
i
m
en
ts
wer
e
ca
r
r
ied
o
u
t
o
n
tr
ain
i
n
g
d
ata
an
d
v
alid
atio
n
d
ata
u
s
i
n
g
t
h
e
C
N
N
I
n
ce
p
tio
n
V3
,
R
esNet5
0
,
I
n
ce
p
tio
n
R
es
n
etV2
,
Den
s
eNe
t2
0
1
,
Mo
b
ileNet
,
an
d
E
f
f
icie
n
tNe
tB
3
ar
ch
itect
u
r
al
m
o
d
els.
T
h
is
ex
p
er
i
m
en
t
ai
m
s
to
f
i
n
d
t
o
ac
cu
r
ac
y
,
lo
s
s
,
a
n
d
co
m
p
u
tati
o
n
ti
m
e
r
eq
u
ir
ed
o
f
ea
c
h
ar
ch
itectu
r
al
m
o
d
el
d
u
r
i
n
g
t
h
e
d
at
a
tr
ain
i
n
g
p
r
o
ce
s
s
.
T
h
is
ex
p
er
i
m
en
t
w
ill
b
e
ca
r
r
i
ed
o
u
t
u
s
in
g
5
0
tr
ain
in
g
ep
o
ch
s
.
T
h
e
r
esu
lts
o
f
t
h
e
ac
cu
r
ac
y
,
lo
s
s
,
an
d
tr
ain
i
n
g
co
m
p
u
tatio
n
ti
m
e
o
f
ea
ch
C
N
N
ar
ch
itec
t
u
r
e
ca
n
b
e
s
ee
n
i
n
T
ab
le
3
.
T
ab
le
3
.
A
cc
u
r
ac
y
,
lo
s
s
a
n
d
ti
m
e
co
m
p
u
tin
g
tr
ai
n
i
n
g
d
ata
an
d
v
alid
atio
n
o
f
C
NN
ar
ch
itec
t
u
r
e
af
ter
5
0
tr
ain
i
n
g
ep
o
ch
s
C
N
N
A
r
si
t
e
k
t
u
r
T
r
a
i
n
A
c
c
(
%)
V
a
l
A
c
c
(
%)
T
r
a
i
n
L
o
ss
V
a
l
L
o
ss
T
i
me
(
M
i
n
u
t
e
)
I
n
c
e
p
t
i
o
n
V
3
9
9
,
3
4
9
9
,
9
2
0
,
0
2
1
5
0
,
0
0
3
9
77
R
e
sN
e
t
5
0
9
8
,
6
5
9
9
,
7
5
0
,
0
4
0
8
0
,
0
1
0
4
86
I
n
c
e
p
t
i
o
n
R
e
sn
e
t
V
2
9
9
,
6
1
9
9
,
8
3
0
,
0
1
4
2
0
,
0
0
4
3
1
1
7
D
e
n
se
N
e
t
2
0
1
9
9
,
1
2
9
9
,
6
6
0
,
0
2
4
2
0
,
0
1
1
1
98
M
o
b
i
l
e
N
e
t
9
7
,
8
4
9
9
,
2
4
0
,
0
5
7
3
0
,
0
2
3
8
72
Ef
f
i
c
i
e
n
t
N
e
t
B
3
8
5
,
4
8
8
9
,
8
1
0
,
5
3
8
7
0
,
3
4
6
5
1
1
2
Fro
m
th
e
r
esu
lt
o
f
ex
p
er
i
m
e
n
t
s
th
at
h
a
v
e
b
ee
n
d
o
n
e,
I
n
ce
p
tio
n
R
es
n
etV2
g
et
s
th
e
b
est
r
esu
lts
f
r
o
m
al
l
C
NN
ar
ch
itect
u
r
al
m
o
d
els
t
h
at
ar
e
tr
ain
ed
w
it
h
an
ac
cu
r
ac
y
v
alu
e
o
f
9
9
.
6
1
%.
T
h
en
f
o
llo
wed
b
y
I
n
ce
p
tio
n
V3
ar
ch
itect
u
r
e
w
it
h
9
9
.
3
4
%
ac
cu
r
ac
y
,
De
n
s
eNe
t2
0
1
w
it
h
9
9
.
1
2
%
ac
cu
r
ac
y
,
R
e
s
Net5
0
w
it
h
9
8
.
6
5
%
ac
cu
r
ac
y
,
Mo
b
ileNet
w
i
th
9
7
.
8
4
%
ac
cu
r
ac
y
a
n
d
E
f
f
icie
n
tNe
tB
3
w
it
h
8
5
.
4
8
%
ac
cu
r
ac
y
.
A
lt
h
o
u
g
h
I
n
ce
p
tio
n
R
e
s
n
e
tV2
g
ets
th
e
b
est
ac
c
u
r
ac
y
v
al
u
e
o
f
all
ar
c
h
itect
u
r
es,
it
r
eq
u
ir
e
s
t
h
e
lo
n
g
es
t
co
m
p
u
tatio
n
ti
m
e
a
m
o
n
g
a
ll
tr
ai
n
ed
ar
ch
itect
u
r
es
to
ac
h
iev
e
5
0
tr
ain
i
n
g
ep
o
ch
s
o
f
1
1
7
m
i
n
u
te
s
.
Mo
b
ileNet
g
ets
t
h
e
b
est
co
m
p
u
te
ti
m
e
to
r
ea
ch
5
0
ep
o
ch
,
w
h
ic
h
is
7
2
m
i
n
u
tes,
f
o
llo
w
ed
b
y
I
n
ce
p
tio
n
V3
w
it
h
7
7
m
i
n
u
tes
o
f
co
m
p
u
tin
g
ti
m
e
,
R
esNet5
0
w
it
h
8
6
m
i
n
u
te
s
o
f
co
m
p
u
tatio
n
ti
m
e,
Den
s
eNe
t2
0
1
w
ith
9
8
m
i
n
u
t
es
o
f
co
m
p
u
ta
tio
n
ti
m
e,
an
d
E
f
f
icien
tNetB
3
w
it
h
1
1
2
m
in
u
te
s
o
f
co
m
p
u
tatio
n
t
i
m
e.
T
ab
le
3
also
d
is
p
lay
s
th
e
tr
ain
i
n
g
lo
s
s
o
f
ea
c
h
tr
ai
n
ed
C
NN
a
r
ch
itect
u
r
al
m
o
d
el.
T
h
e
ar
ch
it
ec
tu
r
e
th
a
t
h
as
t
h
e
lo
w
est
tr
ai
n
i
n
g
lo
s
s
i
s
I
n
ce
p
tio
n
R
es
n
etV2
w
it
h
a
v
alu
e
o
f
0
.
0
1
4
2
,
f
o
llo
w
ed
b
y
I
n
ce
p
tio
n
V3
w
it
h
a
v
alu
e
o
f
0
.
0
2
1
5
,
Den
s
eNe
t2
0
1
w
it
h
a
v
al
u
e
o
f
0
.
0
2
4
2
,
R
esNet5
0
w
it
h
a
v
al
u
e
o
f
0
.
0
4
0
8
,
Mo
b
ileNet
w
it
h
a
v
alu
e
o
f
0
.
0
5
7
3
,
an
d
E
f
f
icie
n
NetB
3
w
i
th
a
v
al
u
e
o
f
0
.
5
3
8
7
.
T
h
e
ac
cu
r
ac
y
an
d
lo
s
s
c
h
ar
ts
in
th
e
tr
ain
i
n
g
pr
o
ce
s
s
ca
n
b
e
s
ee
n
in
Fi
g
u
r
e
s
4
-
9.
Fig
u
r
e
4
.
A
cc
u
r
ac
y
d
an
lo
s
s
tr
ain
i
n
g
I
n
ce
p
tio
n
V3
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
10
,
No
.
4
,
Dec
em
b
er
202
1
:
1
0
6
9
-
1
0
7
8
1074
Fig
u
r
e
5
.
A
cc
u
r
ac
y
d
an
lo
s
s
tr
ain
i
n
g
R
esNet5
0
Fig
u
r
e
6
.
A
cc
u
r
ac
y
d
an
lo
s
s
tr
ain
i
n
g
I
n
ce
p
tio
n
R
e
s
NetV2
Fig
u
r
e
7
.
A
cc
u
r
ac
y
d
an
lo
s
s
tr
ain
i
n
g
De
n
s
e
Net2
0
1
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
A
p
erfo
r
ma
n
ce
ev
a
lu
a
tio
n
o
f
co
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
a
r
ch
itectu
r
e
fo
r
.
.
.
(
A
fis
Ju
lia
n
to
)
1075
Fig
u
r
e
8
.
A
cc
u
r
ac
y
d
an
lo
s
s
tr
ain
i
n
g
Mo
b
ileNet
Fig
u
r
e
9
.
A
cc
u
r
ac
y
d
an
lo
s
s
tr
ain
i
n
g
E
f
f
icie
n
tNetB
3
T
ab
le
4
p
r
esen
ts
t
h
e
e
v
al
u
atio
n
m
atr
ix
b
et
w
ee
n
th
e
tr
ain
ed
m
o
d
el
an
d
t
h
e
te
s
ti
n
g
d
ata.
T
h
e
r
esu
lt
s
o
f
th
e
te
s
t
e
x
p
er
i
m
e
n
ts
co
n
d
u
c
te
d
,
I
n
ce
p
tio
n
R
e
s
n
et
V2
an
d
I
n
c
ep
tio
n
V3
g
o
t
t
h
e
b
es
t
r
es
u
lt
s
f
r
o
m
al
l
t
h
e
C
NN
ar
ch
itect
u
r
al
m
o
d
els
tes
ted
w
ith
1
0
0
%
ac
cu
r
ac
y
,
1
0
0
%
p
r
e
cisi
o
n
,
1
0
0
%
R
ec
all,
an
d
1
0
0
%
F1
Sco
r
e.
T
h
en
f
o
llo
w
ed
b
y
t
h
e
R
es
n
et5
0
an
d
Den
s
eNe
t2
0
1
ar
ch
itect
u
r
es
g
o
t
an
ac
cu
r
ac
y
v
al
u
e
o
f
9
9
.
8
3
,
p
r
ec
is
io
n
9
9
.
8
3
%,
R
ec
all
9
9
.
8
3
%,
an
d
F1
Sco
r
e
9
9
.
8
3
%.
Mo
b
ileNet
g
ets
an
ac
cu
r
ac
y
v
al
u
e
o
f
9
9
.
3
3
%,
p
r
ec
is
io
n
9
9
.
3
6
%,
Re
ca
ll
9
9
.
3
2
%,
F1
Sco
r
e
9
9
.
3
4
%
an
d
E
f
f
icie
n
tNetB
3
g
ets
an
ac
cu
r
ac
y
v
a
lu
e
o
f
9
0
.
1
4
%,
p
r
ec
is
io
n
9
0
.
4
6
%,
R
ec
all
9
0
.
2
3
%,
F1
Sco
r
e
9
0
.
2
8
%.
T
ab
le
4
.
E
v
alu
atio
n
o
f
m
o
d
el
tr
ain
in
g
w
it
h
d
ata
test
i
n
g
C
N
N
a
r
c
h
i
t
e
c
t
u
r
e
T
e
st
A
C
C
(
%)
P
r
e
c
i
si
o
n
(
%)
R
e
c
a
l
l
(
%)
F
I
s
c
o
r
e
(
%)
I
n
c
e
p
t
i
o
n
V
3
1
0
0
1
0
0
1
0
0
1
0
0
R
e
sN
e
t
5
0
9
9
,
8
3
9
9
,
8
3
9
9
,
8
3
9
9
,
8
3
I
n
c
e
p
t
i
o
n
R
e
sn
e
t
V
2
1
0
0
1
0
0
1
0
0
1
0
0
D
e
n
se
N
e
t
2
0
1
9
9
,
8
3
9
9
,
8
3
9
9
,
8
3
9
9
,
8
3
M
o
b
i
l
e
N
e
t
9
9
,
3
3
9
9
,
3
6
9
9
,
3
2
9
9
,
3
4
Ef
f
i
c
i
e
n
t
N
e
t
B
3
9
0
,
1
4
9
0
,
4
6
9
0
,
2
3
9
0
,
2
8
Fig
u
r
e
1
0
s
h
o
w
s
th
e
r
es
u
lts
o
f
th
e
I
n
ce
p
tio
n
V3
ar
ch
itec
tu
r
e
m
o
d
el
co
n
f
u
s
io
n
m
atr
i
x
w
it
h
d
ata
test
i
n
g
.
Fro
m
t
h
e
1
1
8
7
s
a
m
p
l
e
d
ata
test
ed
,
n
o
d
ata
w
er
e
m
is
clas
s
if
ied
.
A
ll
d
ata
w
er
e
c
la
s
s
i
f
ied
co
r
r
ec
tl
y
a
s
s
h
o
w
n
in
F
ig
u
r
e
1
0
.
T
h
e
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
s
co
r
e
v
alu
es c
a
n
b
e
s
ee
n
i
n
T
ab
le
4
.
Fig
u
r
e
1
1
s
h
o
w
s
th
e
r
es
u
lt
s
o
f
th
e
R
esNet5
0
ar
ch
itect
u
r
e
m
o
d
el
co
n
f
u
s
io
n
m
a
tr
ix
w
it
h
d
ata
test
in
g
.
Fro
m
t
h
e
1
1
8
7
s
a
m
p
le
d
ata
te
s
ted
,
th
er
e
ar
e
2
s
a
m
p
le
d
ata
w
h
ic
h
ar
e
m
is
cla
s
s
i
f
ied
.
1
s
a
m
p
le
o
f
m
i
s
clas
s
i
f
ied
b
r
o
w
n
s
p
o
t
class
d
ata
an
d
1
m
is
c
lass
if
ie
d
t
u
n
g
r
o
class
d
at
a
s
a
m
p
le
as
s
h
o
w
n
i
n
Fi
g
u
r
e
1
1
.
T
h
e
v
alu
es
o
f
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all
an
d
F1
Sco
r
e
ca
n
b
e
s
ee
n
in
T
ab
le
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
10
,
No
.
4
,
Dec
em
b
er
202
1
:
1
0
6
9
-
1
0
7
8
1076
Fig
u
r
e
1
2
s
h
o
w
s
t
h
e
r
e
s
u
l
ts
o
f
t
h
e
co
n
f
u
s
io
n
m
atr
i
x
o
f
t
h
e
I
n
ce
p
tio
n
R
e
s
n
e
tV2
ar
ch
itec
tu
r
al
m
o
d
el
w
it
h
d
ata
te
s
ti
n
g
.
Fro
m
t
h
e
1
1
8
7
s
am
p
le
d
ata
test
ed
,
n
o
d
ata
w
er
e
m
i
s
clas
s
i
f
ied
.
A
ll
d
ata
w
er
e
c
lass
if
ied
co
r
r
ec
tly
as
s
h
o
w
n
i
n
Fi
g
u
r
e
1
2
.
T
h
e
v
alu
e
o
f
ac
c
u
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all
an
d
F1
Sco
r
e
ca
n
b
e
s
ee
n
in
T
ab
le
4
.
Fig
u
r
e
1
3
s
h
o
w
s
t
h
e
r
es
u
lt
s
o
f
th
e
De
n
s
eNe
t2
0
1
ar
ch
itect
u
r
al
m
o
d
el
co
n
f
u
s
io
n
m
a
tr
ix
w
it
h
d
ata
test
i
n
g
.
Fro
m
t
h
e
1
1
8
7
s
am
p
l
e
d
ata
test
ed
,
th
er
e
a
r
e
2
d
ata
s
a
m
p
les
in
t
h
e
m
i
s
clas
s
i
f
ied
b
r
o
w
n
s
p
o
t
class
as
s
h
o
w
n
in
F
ig
u
r
e
1
3
.
T
h
e
v
alu
e
s
f
o
r
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
al
l a
n
d
F1
Sco
r
e
ca
n
b
e
s
ee
n
in
T
ab
le
4
.
Fig
u
r
e
1
4
s
h
o
w
s
th
e
r
e
s
u
lts
o
f
th
e
Mo
b
ileNet
ar
c
h
itect
u
r
e
m
o
d
el
co
n
f
u
s
io
n
m
atr
i
x
w
it
h
d
a
ta
test
i
n
g
.
Fro
m
th
e
1
1
8
7
s
am
p
le
d
ata
te
s
ted
,
8
s
a
m
p
le
d
ata
w
er
e
m
is
c
lass
i
f
ied
.
T
w
o
s
a
m
p
le
s
o
f
m
i
s
class
i
f
ied
b
ac
ter
ial
b
lig
h
t
cla
s
s
d
ata,
an
d
6
m
is
c
la
s
s
i
f
ied
b
last
clas
s
d
ata
s
a
m
p
le
s
as
s
h
o
w
n
i
n
Fi
g
u
r
e
1
4
.
T
h
e
v
alu
e
s
f
o
r
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
Sco
r
e
ca
n
b
e
s
ee
n
i
n
T
ab
le
4
.
Fig
u
r
e
1
5
s
h
o
w
s
t
h
e
r
es
u
lts
o
f
th
e
co
n
f
u
s
io
n
m
atr
ix
m
o
d
el
f
o
r
th
e
E
f
f
icie
n
tNe
tB
3
ar
ch
itectu
r
e
w
it
h
d
ata
test
in
g
.
Fro
m
t
h
e
1
1
8
7
s
a
m
p
le
d
ata
test
ed
,
1
1
7
s
am
p
le
d
ata
w
er
e
m
is
cla
s
s
i
f
ied
.
T
h
e
3
6
s
a
m
p
les
o
f
m
is
c
lass
if
ied
b
ac
ter
ial
b
lig
h
t
class
d
ata,
5
7
m
i
s
clas
s
i
f
ied
b
last
clas
s
d
ata
s
a
m
p
les,
1
8
m
is
clas
s
if
ied
b
r
o
w
n
s
p
o
t
class
d
ata
s
a
m
p
le
s
a
n
d
6
m
i
s
clas
s
i
f
ied
t
u
n
g
r
o
cla
s
s
d
ata
s
a
m
p
les
a
s
s
h
o
w
n
i
n
Fi
g
u
r
e
1
5
.
T
h
e
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
Sco
r
e
v
alu
e
s
ca
n
b
e
s
ee
n
in
T
ab
le
4
.
Fig
u
r
e
10
.
C
o
n
f
u
s
io
n
m
a
tr
ix
o
f
I
n
ce
p
tio
n
V3
Fig
u
r
e
1
1
.
C
o
n
f
u
s
io
n
m
a
tr
ix
o
f
R
e
s
Net5
0
Fig
u
r
e
1
2
.
C
o
n
f
u
s
io
n
m
a
tr
ix
o
f
I
n
ce
p
tio
n
R
es
n
etV2
Fig
u
r
e
1
3
.
C
o
n
f
u
s
io
n
m
a
tr
ix
o
f
Den
s
eNe
t2
0
1
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
A
p
erfo
r
ma
n
ce
ev
a
lu
a
tio
n
o
f
co
n
vo
lu
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
a
r
ch
itectu
r
e
fo
r
.
.
.
(
A
fis
Ju
lia
n
to
)
1077
Fig
u
r
e
14
.
C
o
n
f
u
s
io
n
m
a
tr
ix
o
f
Mo
b
ileNet
Fig
u
r
e
15
.
C
o
n
f
u
s
io
n
m
a
tr
ix
o
f
E
f
f
icie
n
tNetB
3
Fu
r
t
h
er
m
o
r
e,
a
f
ter
co
m
p
ar
i
s
o
n
s
o
f
th
e
r
es
u
lts
o
f
th
is
e
x
p
er
i
m
e
n
t
w
i
th
s
e
v
er
al
p
r
ev
io
u
s
s
tu
d
ie
s
/r
esear
ch
.
B
y
u
s
in
g
th
e
C
NN
ar
ch
itect
u
r
e,
th
e
b
est
p
er
f
o
r
m
an
ce
ca
n
b
e
o
b
tain
ed
f
r
o
m
t
h
is
e
x
p
er
i
m
e
n
t
f
o
r
class
i
f
y
in
g
d
is
ea
s
e
s
i
n
r
ice
leav
es
b
e
y
o
n
d
co
n
v
e
n
tio
n
al
m
eth
o
d
s
s
u
c
h
as
KNN
[
6
]
,
lo
g
is
tic
r
eg
r
e
s
s
io
n
,
d
ec
is
io
n
tr
ee
(
DT
)
,
n
aïv
e
b
a
y
es
(
NB
)
[
7
]
,
SVM
[
8
]
,
an
d
ANN
[
1
0
]
.
T
h
e
ex
p
er
i
m
e
n
ts
i
n
th
is
s
t
u
d
y
al
s
o
h
a
v
e
b
etter
p
er
f
o
r
m
a
n
ce
th
a
n
o
th
er
C
NN
ar
ch
i
tectu
r
es,
li
k
e
VG
G
1
6
[
1
1
]
,
A
lex
Net
[
1
3
]
,
an
d
VGG1
9
[
1
4
]
.
4.
CO
NCLU
SI
O
N
Af
ter
e
v
alu
a
tin
g
t
h
e
e
x
p
er
i
m
en
ts
o
f
ea
ch
C
N
N
ar
ch
itect
u
r
al
m
o
d
el,
th
e
b
es
t
ar
c
h
itec
tu
r
es
f
o
r
t
h
e
class
i
f
icatio
n
o
f
r
ice
leaf
d
is
e
ase
ar
e
I
n
ce
p
tio
n
V3
a
n
d
I
n
ce
p
tio
n
R
es
n
etV2
,
w
i
th
a
n
ac
c
u
r
ac
y
o
f
1
0
0
%.
T
h
en
f
o
llo
w
ed
b
y
t
h
e
R
es
Net5
0
ar
ch
itect
u
r
e
w
it
h
a
n
ac
c
u
r
ac
y
o
f
9
9
.
8
3
%,
Den
s
e
Net2
0
1
9
9
.
8
3
%,
Mo
b
ileNet
9
9
.
3
3
%
an
d
E
f
f
ec
ien
t
NetB
3
9
0
.
1
4
%.
T
h
is
e
x
p
er
i
m
e
n
t
w
a
s
ca
r
r
ied
o
u
t
u
s
in
g
A
d
a
m
's
o
p
ti
m
izat
io
n
a
n
d
m
o
d
i
f
y
i
n
g
t
h
e
b
atch
s
ize
an
d
lear
n
in
g
r
ate.
T
h
e
r
esu
lt
s
h
o
w
s
t
h
e
p
r
o
p
o
s
ed
C
NN
m
o
d
el
e
x
ce
ed
s
th
e
co
n
v
e
n
tio
n
al
m
et
h
o
d
s
an
d
o
th
er
C
NN
ar
ch
itectu
r
es
f
o
u
n
d
in
p
r
ev
io
u
s
s
tu
d
ie
s
in
th
e
clas
s
i
f
icatio
n
o
f
r
ice
leaf
d
is
ea
s
e.
I
t
is
i
m
p
o
r
tan
t
to
d
ev
elo
p
a
C
NN
m
o
d
el
f
o
r
f
u
r
th
er
r
esear
ch
t
h
at
h
as
b
etter
tr
ain
i
n
g
ti
m
e
an
d
ac
cu
r
ac
y
.
I
t
is
n
ec
e
s
s
ar
y
to
ad
d
m
o
r
e
d
ata
o
n
th
e
t
y
p
e
s
o
f
d
i
s
ea
s
es
o
n
r
ice
lea
v
es
a
n
d
m
o
r
e
t
y
p
es
o
f
p
ests
o
n
r
ice
leav
es.
RE
F
E
R
E
NC
E
S
[1
]
Ce
n
tral
Bu
re
a
u
o
f
S
tatisti
c
s,
“
A
r
e
a
o
f
Ric
e
Ha
r
v
e
st
a
n
d
P
r
o
d
u
c
ti
o
n
in
I
n
d
o
n
e
sia
2
0
1
9
,
”
Ber
it
a
Res
mi
S
ta
ti
st
ik
,
n
o
.
1
6
,
2
0
2
0
.
[2
]
A
.
Leo
n
a
rd
a
n
d
P
.
G
ian
e
ss
i,
“
Im
p
o
rtan
c
e
o
f
P
e
sticid
e
s
f
o
r
G
ro
w
in
g
Rice
in
S
o
u
t
h
a
n
d
S
o
u
th
Eas
t,
”
Cro
p
li
fe
Fo
u
n
d
a
ti
o
n
,
p
p
.
3
0
–
3
3
,
2
0
1
4
.
[3
]
A
.
Na
su
ti
o
n
a
n
d
B.
Nu
ry
a
n
to
,
“
Blas
P
y
ricu
laria
g
rise
a
Dise
a
se
i
n
Rice
P
lan
ts
a
n
d
C
o
n
tr
o
l
S
trat
e
g
y
,
”
Pl
a
n
ti
n
g
Fo
o
d
,
v
o
l
.
9
,
n
o
.
2
,
p
p
.
8
5
–
9
6
,
2
0
1
5
.
[4
]
Ag
ricu
lt
u
ra
l
Re
se
a
rc
h
a
n
d
De
v
e
lo
p
m
e
n
t
Ag
e
n
c
y
,
“
T
e
c
h
n
ica
l
Gu
id
e
li
n
e
s
f
o
r
P
lan
t
Dise
a
se
P
e
sts,”
In
d
o
n
e
si
a
n
Ce
n
ter
fo
r R
ice
Res
e
a
rc
h
e
rs
,
v
o
l.
1
,
p
p
.
1
–
5
,
2
0
0
7
.
[5
]
N.
L
a
k
sh
it
a
,
S
.
H.
P
o
r
o
m
a
rto
,
a
n
d
H.
Ha
d
iw
i
y
o
n
o
,
“
Re
sista
n
c
e
o
f
S
o
m
e
Rice
V
a
rieties
to
Ce
rc
o
sp
o
ra
o
ry
z
a
e
,
”
Ag
ro
tec
h
n
o
l
o
g
y
Res
e
a
rc
h
J
o
u
rn
a
l.
,
v
o
l.
3
,
n
o
.
2
,
p
.
7
5
,
2
0
1
9
,
d
o
i:
1
0
.
2
0
9
6
1
/ag
ro
tec
h
re
sj.v
3
i
2
.
2
9
9
7
6
.
[6
]
M
.
S
u
re
sh
a
,
K.
N.
S
h
re
e
k
a
n
th
,
a
n
d
B.
V
.
T
h
iru
m
a
les
h
,
“
Re
c
o
g
n
it
io
n
o
f
d
ise
a
se
s
in
p
a
d
d
y
le
a
v
e
s
u
sin
g
k
n
n
c
las
si
f
ier,”
in
In
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
fo
r
Co
n
v
e
rg
e
n
c
e
in
T
e
c
h
n
o
l
o
g
y
(
I2
CT
)
,
2
0
1
7
,
p
p
.
6
6
3
-
6
6
6
,
d
o
i:
1
0
.
1
1
0
9
/I
2
CT
.
2
0
1
7
.
8
2
2
6
2
1
3
.
[7
]
K.
A
h
m
e
d
,
T
.
R.
S
h
a
h
id
i,
S
.
M
.
I
rf
a
n
u
l
A
la
m
,
a
n
d
S
.
M
o
m
e
n
,
“
Rice
lea
f
d
ise
a
s
e
d
e
tec
ti
o
n
u
sin
g
m
a
c
h
in
e
lea
rn
in
g
tec
h
n
iq
u
e
s,”
in
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
S
u
sta
in
a
b
le
T
e
c
h
n
o
l
o
g
ies
fo
r
In
d
u
stry
4
.
0
(
S
T
I
)
,
v
2
0
1
9
,
p
p
.
1
-
5
,
d
o
i:
1
0
.
1
1
0
9
/
S
T
I4
7
6
7
3
.
2
0
1
9
.
9
0
6
8
0
9
6
.
[8
]
S
.
P
a
v
it
h
ra
,
A
.
P
riy
a
d
h
a
rsh
in
i,
V
.
P
ra
v
e
e
n
a
,
a
n
d
T
.
M
o
n
ik
a
,
“
P
a
d
d
y
L
e
a
f
Dis
e
a
se
De
t
e
c
ti
o
n
Us
in
g
S
v
m
,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
f
c
o
mm
u
n
i
c
a
ti
o
n
a
n
d
c
o
mp
u
ter
T
e
c
h
n
o
lo
g
ie
s
,
v
o
l.
3
,
n
o
.
1
,
p
p
.
1
6
–
2
0
,
2
0
1
5.
[9
]
F
.
T
.
P
in
k
i,
N.
K
h
a
tu
n
,
a
n
d
S
.
M
.
M
.
Isla
m
,
“
Co
n
ten
t
b
a
se
d
p
a
d
d
y
l
e
a
f
d
ise
a
se
re
c
o
g
n
it
io
n
a
n
d
re
m
e
d
y
p
re
d
ictio
n
u
sin
g
su
p
p
o
r
t
v
e
c
to
r
m
a
c
h
in
e
,
”
in
In
ter
n
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
o
f
Co
mp
u
ter
a
n
d
In
f
o
rm
a
ti
o
n
T
e
c
h
n
o
l
o
g
y
(
ICCIT
)
2
0
1
7
,
p
p
.
1
-
5
,
d
o
i:
1
0
.
1
1
0
9
/ICCI
T
ECHN.2
0
1
7
.
8
2
8
1
7
6
4
.
[1
0
]
R.
De
sh
m
u
k
h
a
n
d
M
.
De
sh
m
u
k
h
,
“
De
tec
ti
o
n
o
f
P
a
d
d
y
Lea
f
Dise
a
se
s,”
In
ter
n
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
Co
mp
u
ter
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
10
,
No
.
4
,
Dec
em
b
er
202
1
:
1
0
6
9
-
1
0
7
8
1078
Ap
p
li
c
a
ti
o
n
s
,
p
p
.
9
7
5
–
8
8
8
7
,
2
0
1
5
.
[1
1
]
S
.
G
h
o
sa
l
a
n
d
K.
S
a
rk
a
r,
“
Rice
Lea
f
Dise
a
se
s Cla
ss
i
f
ica
ti
o
n
Us
in
g
CNN
w
it
h
T
ra
n
s
f
e
r
L
e
a
rn
in
g
,
”
in
IEE
E
Ca
lcu
t
ta
Co
n
fer
e
n
c
e
(
CAL
CON)
,
2
0
2
0
,
p
p
.
2
3
0
-
2
3
6
,
d
o
i:
1
0
.
1
1
0
9
/CA
L
CO
N4
9
1
6
7
.
2
0
2
0
.
9
1
0
6
4
2
3
.
[1
2
]
B.
Kn
y
a
z
e
v
,
R.
S
h
v
e
tso
v
,
N.
E
fre
m
o
v
a
,
a
n
d
A
.
Ku
h
a
re
n
k
o
,
“
Co
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s
p
re
train
e
d
o
n
larg
e
f
a
c
e
re
c
o
g
n
it
io
n
d
a
tas
e
ts f
o
r
e
m
o
ti
o
n
c
las
sif
ica
ti
o
n
f
ro
m
v
id
e
o
,
”
a
rXiv
,
p
p
.
2
–
5
,
2
0
1
7
.
[1
3
]
R.
R.
A
to
le
a
n
d
D.
P
a
rk
,
“
A
m
u
l
ti
c
las
s
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
c
la
ss
i
f
ier
f
o
r
d
e
tec
ti
o
n
o
f
c
o
m
m
o
n
rice
p
lan
t
a
n
o
m
a
li
e
s,”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
A
d
v
a
n
c
e
d
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
Ap
p
li
c
a
ti
o
n
s
,
v
o
l.
9
,
n
o
.
1
,
p
p
.
6
7
–
7
0
,
2
0
1
8
,
d
o
i:
1
0
.
1
4
5
6
9
/IJA
CS
A
.
2
0
1
8
.
0
9
0
1
0
9
.
[1
4
]
J.
Ch
e
n
,
J.
Ch
e
n
,
D.
Zh
a
n
g
,
Y.
S
u
n
,
a
n
d
Y.
A
.
Na
n
e
h
k
a
ra
n
,
“
Us
in
g
d
e
e
p
tra
n
sf
e
r
lea
rn
in
g
f
o
r
ima
g
e
-
b
a
se
d
p
lan
t
d
ise
a
se
id
e
n
ti
f
ica
ti
o
n
,
”
J
o
u
rn
a
l
Co
m
p
u
ter
s
a
n
d
El
e
c
tro
n
i
c
s
in
Ag
ric
u
lt
u
re
,
v
o
l
.
1
7
3
,
2
0
2
0
,
d
o
i
:
1
0
.
1
0
1
6
/
j.
c
o
m
p
a
g
.
2
0
2
0
.
1
0
5
3
9
3
.
[1
5
]
P
.
K.
S
e
t
h
y
,
“
Rice
L
e
a
f
Dise
a
se
I
m
a
g
e
S
a
m
p
les
,
”
M
e
n
d
e
ley
Da
ta
,
V1
,
2
0
2
0
,
d
o
i:
d
o
i
:
1
0
.
1
7
6
3
2
/f
wc
j
7
stb
8
r
.
1
.
[1
6
]
J.
G
u
e
t
a
l.
,
“
Re
c
e
n
t
a
d
v
a
n
c
e
s
in
c
o
n
v
o
lu
t
io
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s,”
J
o
u
rn
a
l
Pa
tt
e
rn
Rec
o
g
n
it
io
n
,
v
o
l.
7
7
,
p
p
.
3
5
4
–
3
7
7
,
2
0
1
8
,
d
o
i:
1
0
.
1
0
1
6
/
j.
p
a
tco
g
.
2
0
1
7
.
1
0
.
0
1
3
.
[1
7
]
Y.
L
e
Cu
n
,
L
.
Bo
tt
o
u
,
Y.
Be
n
g
io
,
a
n
d
P
.
Ha
ff
n
e
r,
“
G
ra
d
ien
t
-
b
a
se
d
lea
rn
in
g
a
p
p
li
e
d
to
d
o
c
u
m
e
n
t
re
c
o
g
n
it
io
n
,
”
in
Pro
c
e
e
d
in
g
s
o
f
t
h
e
IEE
E
,
v
o
l.
8
6
,
n
o
.
1
1
,
p
p
.
2
2
78
–
2
3
2
3
,
1
9
9
8
,
d
o
i
:
1
0
.
1
1
0
9
/5
.
7
2
6
7
9
1
.
[1
8
]
K.
A
n
n
a
p
u
ra
n
i
a
n
d
D.
Ra
v
il
la,
“
CNN
b
a
se
d
i
m
a
g
e
c
la
ss
i
f
ica
ti
o
n
m
o
d
e
l,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
In
n
o
v
a
t
ive
T
e
c
h
n
o
l
o
g
y
a
n
d
Exp
l
o
rin
g
En
g
i
n
e
e
rin
g
(
IJ
IT
EE
)
,
v
o
l.
8
,
n
o
.
1
1
,
p
p
.
1
1
0
6
–
1
1
1
4
,
2
0
1
9
,
d
o
i
:
1
0
.
3
5
9
4
0
/i
ji
tee
.
K1
2
2
5
.
0
9
8
1
1
S
1
9
.
[1
9
]
C.
S
z
e
g
e
d
y
,
V
.
V
a
n
h
o
u
c
k
e
,
S
.
Io
ff
e
,
J.
S
h
len
s,
a
n
d
Z.
W
o
jn
a
,
“
Re
th
in
k
in
g
th
e
In
c
e
p
t
io
n
A
rc
h
it
e
c
tu
re
f
o
r
Co
m
p
u
ter
V
isio
n
,
”
in
I
EE
E
C
o
n
fer
e
n
c
e
o
n
Co
mp
u
ter
Vi
sio
n
a
n
d
P
a
tt
e
rn
Rec
o
g
n
i
ti
o
n
(
CVP
R)
,
v
o
l.
1
,
p
p
.
2
8
1
8
–
2
8
2
6
,
2
0
1
6
,
d
o
i:
1
0
.
1
1
0
9
/C
VP
R.
2
0
1
6
.
3
0
8
.
[2
0
]
K
K.
He
,
X
.
Zh
a
n
g
,
S
.
Re
n
,
a
n
d
J.
S
u
n
,
“
De
e
p
re
sid
u
a
l
lea
rn
i
n
g
f
o
r
i
m
a
g
e
re
c
o
g
n
it
io
n
,
”
in
IEE
E
Co
n
fer
e
n
c
e
o
n
Co
mp
u
ter
V
isio
n
a
n
d
Pa
t
ter
n
Rec
o
g
n
it
io
n
(
CVP
R)
,
2
0
1
6
,
p
p
.
7
7
0
-
7
7
8
,
d
o
i:
1
0
.
1
1
0
9
/C
VPR.
2
0
1
6
.
9
0
.
[2
1
]
C.
S
z
e
g
e
d
y
,
S
.
Io
f
f
e
,
V
.
V
a
n
h
o
u
c
k
e
,
a
n
d
A
.
A
.
A
le
m
i,
“
In
c
e
p
ti
o
n
-
v
4
,
in
c
e
p
t
io
n
-
Re
sN
e
t
a
n
d
th
e
imp
a
c
t
o
f
re
sid
u
a
l
c
o
n
n
e
c
ti
o
n
s o
n
lea
rn
i
n
g
,
”
in
Co
n
f
e
re
n
c
e
o
n
Arti
fi
c
i
a
l
I
n
telli
g
e
n
c
e
,
p
p
.
4
2
7
8
–
4
2
8
4
,
2
0
1
7
.
[2
2
]
G
.
Hu
a
n
g
,
Z.
L
iu
,
L
.
V
a
n
De
r
M
a
a
ten
,
a
n
d
K.
Q.
W
e
in
b
e
rg
e
r,
“
De
n
se
l
y
c
o
n
n
e
c
ted
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
tw
o
rk
s,”
in
IEE
E
Co
n
fer
e
n
c
e
o
n
Co
mp
u
ter
Vi
sio
n
a
n
d
Pa
t
ter
n
Rec
o
g
n
it
i
o
n
(
CVP
R)
,
v
o
l.
1
,
p
p
.
2
2
6
1
–
2
2
6
9
,
2
0
1
7
,
d
o
i
:
1
0
.
1
1
0
9
/C
VP
R
.
2
0
1
7
.
2
4
3
.
[2
3
]
A
.
G
.
Ho
w
a
rd
e
t
a
l.
,
“
M
o
b
il
e
Ne
ts:
Eff
icie
n
t
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s
f
o
r
m
o
b
il
e
v
isio
n
a
p
p
li
c
a
ti
o
n
s,”
a
rXiv
,
2
0
1
7
.
[2
4
]
M
.
T
a
n
a
n
d
Q.
V
.
L
e
,
“
E
ff
icie
n
tNe
t:
Re
th
in
k
in
g
m
o
d
e
l
sc
a
li
n
g
f
o
r
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s,”
in
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
M
a
c
h
i
n
e
L
e
a
rn
in
g
(
ICM
L
)
,
p
p
.
1
0
6
9
1
–
1
0
7
0
0
,
2
0
1
9
.
[2
5
]
W
.
Jia
,
J.
G
a
o
,
W
.
X
ia,
Y.
Z
h
a
o
,
H.
M
in
,
a
n
d
J.
T
.
L
u
,
“
A
P
e
rf
o
rm
a
n
c
e
Ev
a
lu
a
ti
o
n
o
f
Clas
sic
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
tw
o
rk
s
f
o
r
2
D
a
n
d
3
D
P
a
lm
p
rin
t
a
n
d
P
a
lm
V
e
in
Re
c
o
g
n
it
i
o
n
,
”
I
n
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
A
u
to
ma
ti
o
n
a
n
d
Co
mp
u
t
in
g
,
v
o
l
.
1
8
,
n
o
.
1
,
p
p
.
1
8
–
4
4
,
2
0
2
1
,
d
o
i:
1
0
.
1
0
0
7
/s
1
1
6
3
3
-
0
2
0
-
1
2
5
7
-
9.
[2
6
]
D.
Krisn
a
n
d
i
e
t
a
l.
,
“
Dise
a
se
s
Clas
sif
ic
a
ti
o
n
f
o
r
T
e
a
P
lan
t
Us
in
g
Co
n
c
a
ten
a
ted
Co
n
v
o
lu
ti
o
n
Ne
u
ra
l
Ne
t
w
o
rk
,
”
J
o
u
rn
a
l
Co
mm
IT
(
C
o
mm
u
n
ic
a
t
io
n
a
n
d
In
f
o
rm
a
ti
o
n
T
e
c
h
n
o
lo
g
y
)
,
v
o
l.
1
3
,
n
o
.
2
,
p
p
.
6
7
–
7
7
,
2
0
1
9
,
d
o
i:
1
0
.
2
1
5
1
2
/co
m
m
it
.
v
1
3
i2
.
5
8
8
6
.
[2
7
]
C.
M
a
n
li
g
u
e
z
,
“
G
e
n
e
ra
li
z
e
d
Co
n
f
u
sio
n
M
a
tri
x
f
o
r
M
u
lt
ip
le Cl
a
ss
e
s
,
”
p
p
.
5
–
7
,
2
0
1
6
.
[2
8
]
D.
Zh
a
n
g
,
J.
W
a
n
g
,
a
n
d
X
.
Z
h
a
o
,
“
Esti
m
a
ti
n
g
th
e
u
n
c
e
rtain
ty
o
f
a
v
e
ra
g
e
F
1
sc
o
re
s,”
in
Pro
c
e
e
d
in
g
s
o
f
t
h
e
2
0
1
5
In
ter
n
a
t
io
n
a
l
Co
n
fer
e
n
c
e
o
n
T
h
e
T
h
e
o
ry
o
f
In
fo
rm
a
ti
o
n
Retrie
v
a
l
,
p
p
.
3
1
7
–
3
2
0
,
2
0
1
5
,
d
o
i:
1
0
.
1
1
4
5
/
2
8
0
8
1
9
4
.
2
8
0
9
4
8
8
.
[2
9
]
D.
P
.
Ki
n
g
m
a
a
n
d
J.
L
.
Ba
,
“
A
d
a
m
:
A
M
e
th
o
d
f
o
r
S
t
o
c
h
a
stic Op
ti
m
iz
a
ti
o
n
,
”
a
rXiv
,
p
p
.
1
–
1
5
,
2
0
1
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.