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.
1
5
,
No
.
2
,
A
p
r
il 2
0
2
6
,
p
p
.
1
6
4
6
~
1
6
6
3
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
2
.
p
p
1
6
4
6
-
1
6
6
3
1646
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Unified
vo
ting
-
ba
sed ensem
ble
lear
ning
f
o
r ric
e leaf
disea
se
detec
tion usin
g
i
mpro
v
ed pret
ra
ined mo
dels
G
o
v
ind
a
ra
j
a
n Sub
bu
ra
m
a
n,
M
a
ry
Vennila
Selv
a
du
ra
i
D
e
p
a
r
t
me
n
t
o
f
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
,
P
r
e
si
d
e
n
c
y
C
o
l
l
e
g
e
,
U
n
i
v
e
r
si
t
y
o
f
M
a
d
r
a
s,
C
h
e
n
n
a
i
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ap
r
1
0
,
2
0
2
5
R
ev
is
ed
J
an
1
3
,
2
0
2
6
Acc
ep
ted
Feb
6
,
2
0
2
6
As
a
sta
p
le
fo
o
d
fo
r
a
larg
e
p
o
rti
o
n
o
f
t
h
e
g
l
o
b
a
l
p
o
p
u
latio
n
,
rice
is
p
a
rti
c
u
larly
su
sc
e
p
t
ib
le
to
lea
f
d
ise
a
se
s
th
a
t
a
d
v
e
rse
ly
a
ffe
c
t
it
s
y
ield
a
n
d
o
v
e
ra
ll
q
u
a
li
t
y
.
Th
is
stu
d
y
u
ti
li
z
e
s
fo
u
r
p
re
train
e
d
c
o
n
v
o
lu
t
io
n
a
l
n
e
u
ra
l
n
e
two
rk
(CNN
)
m
o
d
e
ls
to
c
o
n
stru
c
t
a
u
n
ifi
e
d
v
o
ti
n
g
-
b
a
se
d
e
n
se
m
b
le
a
p
p
ro
a
c
h
f
o
r
rice
lea
f
d
ise
a
se
c
l
a
ss
ifi
c
a
ti
o
n
.
Th
e
m
o
d
e
ls
in
c
lu
d
e
VG
G
1
6
,
De
n
se
Ne
t1
2
1
,
I
n
c
e
p
ti
o
n
V
3
,
a
n
d
Xc
e
p
ti
o
n
.
T
h
e
d
a
tas
e
t
u
se
d
i
n
th
is
st
u
d
y
wa
s
c
o
ll
e
c
ted
fro
m
Ka
g
g
le
a
n
d
f
u
rth
e
r
e
n
rich
e
d
wi
th
ima
g
e
s
o
b
tain
e
d
fr
o
m
G
o
o
g
le
so
u
rc
e
s.
It
c
o
m
p
rise
s
a
to
tal
o
f
4
,
0
0
0
ima
g
e
s
c
a
teg
o
rize
d
in
t
o
six
c
las
se
s:
b
a
c
teria
l
lea
f
b
li
g
h
t,
b
r
o
wn
sp
o
t,
lea
f
b
las
t,
lea
f
sc
a
ld
,
n
a
r
ro
w
b
r
o
wn
sp
o
t,
a
n
d
h
e
a
lt
h
y
lea
v
e
s
.
It
w
a
s
sp
li
t
i
n
t
o
trai
n
in
g
(
3
2
7
ima
g
e
s/c
las
s),
v
a
li
d
a
ti
o
n
(1
4
0
ima
g
e
s/c
las
s),
a
n
d
tes
ti
n
g
(
2
0
0
ima
g
e
s/c
las
s).
Im
a
g
e
s
we
re
n
o
rm
a
li
z
e
d
to
[0
,
1
]
a
n
d
a
u
g
m
e
n
te
d
th
ro
u
g
h
ro
tati
o
n
,
fli
p
p
in
g
,
sh
if
ti
n
g
,
sh
e
a
r,
z
o
o
m
,
b
rig
h
t
n
e
ss
,
a
n
d
c
h
a
n
n
e
l
a
d
j
u
stm
e
n
ts
t
o
imp
ro
v
e
g
e
n
e
ra
li
z
a
ti
o
n
.
In
d
i
v
id
u
a
ll
y
,
t
h
e
fin
e
-
t
u
n
e
d
m
o
d
e
ls
a
c
h
iev
e
d
a
c
c
u
ra
c
ies
o
f
9
1
.
3
%
(
V
GG
1
6
),
9
5
.
6
%
(De
n
se
Ne
t1
2
1
),
9
2
.
1
%
(I
n
c
e
p
ti
o
n
V3
),
a
n
d
8
9
.
8
%
(Xc
e
p
t
io
n
).
T
h
e
e
n
se
m
b
le
lev
e
ra
g
e
d
m
a
jo
ri
ty
v
o
ti
n
g
(9
3
.
6
%
),
we
ig
h
te
d
v
o
ti
n
g
(9
6
.
5
%
),
a
n
d
so
ft
v
o
ti
n
g
(9
7
%
),
y
ield
i
n
g
a
n
a
b
so
lu
te
g
a
i
n
o
f
1
.
4
%
o
v
e
r
t
h
e
b
e
st
in
d
i
v
id
u
a
l
m
o
d
e
l
a
n
d
4
.
8
%
o
v
e
r
t
h
e
a
v
e
ra
g
e
o
f
a
ll
m
o
d
e
ls.
T
o
o
u
r
k
n
o
wle
d
g
e
,
t
h
is
i
s
th
e
first
e
n
se
m
b
le
c
o
m
b
in
in
g
th
e
s
e
fo
u
r
a
rc
h
it
e
c
t
u
re
s
with
u
n
if
ied
v
o
ti
n
g
f
o
r
id
e
n
ti
f
y
in
g
d
ise
a
se
s
in
rice
lea
v
e
s
,
d
e
li
v
e
ri
n
g
a
sc
a
lab
le
a
n
d
c
o
m
p
u
tatio
n
a
ll
y
e
fficie
n
t
so
l
u
ti
o
n
su
it
a
b
le
i
n
a
d
v
a
n
c
e
d
ia
g
n
o
sis
a
n
d
ti
m
e
ly
e
x
e
c
u
ti
o
n
i
n
a
g
ricu
lt
u
ra
l
se
tt
in
g
s with
li
m
it
e
d
r
e
so
u
rc
e
s.
K
ey
w
o
r
d
s
:
Data
au
g
m
en
tatio
n
tec
h
n
iq
u
es
E
n
s
em
b
le
d
ee
p
lear
n
in
g
I
m
ag
e
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es
Plan
t d
is
ea
s
e
clas
s
if
icatio
n
R
ice
leaf
d
is
ea
s
es
T
r
an
s
f
er
lear
n
i
n
g
m
o
d
els
Vo
tin
g
en
s
em
b
le
m
eth
o
d
s
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
:
Go
v
in
d
ar
ajan
Su
b
b
u
r
am
an
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
,
Pre
s
id
en
cy
C
o
lleg
e,
Un
i
v
er
s
ity
o
f
Ma
d
r
as
C
h
en
n
ai,
I
n
d
ia
E
m
ail:
cscr
esear
ch
2
0
2
4
@
g
m
a
il.c
o
m
1.
I
NT
RO
D
UCT
I
O
N
R
ice
is
o
n
e
o
f
th
e
wo
r
ld
’
s
m
o
s
t
v
ital
cr
o
p
s
,
s
u
s
tain
in
g
a
lar
g
e
p
o
r
tio
n
o
f
t
h
e
g
lo
b
al
p
o
p
u
l
atio
n
an
d
h
o
ld
in
g
s
ig
n
if
ica
n
t
ec
o
n
o
m
ic
an
d
cu
ltu
r
al
im
p
o
r
tan
ce
.
At
p
r
esen
t,
it
co
n
s
titu
tes
th
e
s
t
ap
le
f
o
o
d
f
o
r
o
v
er
2
.
7
b
illi
o
n
p
eo
p
le,
an
d
th
is
n
u
m
b
er
is
p
r
o
jecte
d
to
in
cr
ea
s
e
s
u
b
s
tan
tially
,
r
ea
ch
in
g
n
ea
r
ly
3
.
9
b
illi
o
n
b
y
2
0
2
5
[
1
]
.
Desp
ite
r
ice
p
lan
t
d
is
ea
s
es
ca
u
s
in
g
an
an
n
u
al
3
7
%
lo
s
s
in
p
r
o
d
u
ctio
n
d
u
e
to
in
s
u
f
f
icien
t
d
is
ea
s
e
id
en
tific
atio
n
an
d
m
an
a
g
em
en
t
k
n
o
wled
g
e,
ef
f
ec
tiv
e
d
ia
g
n
o
s
tic
an
d
m
an
ag
em
en
t
ap
p
licat
io
n
s
r
em
ain
s
ca
r
ce
[
2
]
.
Am
o
n
g
th
e
v
ar
i
o
u
s
d
is
ea
s
es
th
at
af
f
ec
t
r
ice
cu
ltiv
atio
n
,
b
r
o
wn
s
p
o
t,
b
last
,
an
d
b
ac
ter
ia
l
leaf
b
lig
h
t
ar
e
th
e
m
o
s
t
wid
esp
r
ea
d
an
d
ca
u
s
e
t
h
e
g
r
ea
test
ec
o
n
o
m
ic
lo
s
s
es.
T
h
ese
d
is
ea
s
es
s
ev
er
ely
h
in
d
er
r
ice
p
lan
t
g
r
o
wth
an
d
p
r
o
d
u
ctiv
ity
,
o
f
ten
r
esu
l
tin
g
in
s
u
b
s
tan
tial
ec
o
n
o
m
ic
an
d
en
v
ir
o
n
m
en
tal
lo
s
s
es.
E
ar
ly
an
d
ac
cu
r
ate
d
etec
tio
n
o
f
th
ese
d
is
ea
s
es
with
in
a
s
h
o
r
t
tim
e
f
r
am
e
is
ess
en
tial,
as
it
ca
n
h
elp
m
in
im
iz
e
cr
o
p
d
a
m
ag
e
an
d
s
af
eg
u
ar
d
f
ar
m
er
s
f
r
o
m
s
u
b
s
t
an
tial
f
in
an
cial
s
etb
ac
k
s
[
3
]
.
On
e
o
f
I
n
d
ia'
s
m
o
s
t
p
o
p
u
lar
s
tap
le
cr
o
p
s
is
r
ice,
an
d
th
e
c
o
u
n
tr
y
'
s
ag
r
icu
ltu
r
al
in
d
u
s
tr
y
ac
c
o
u
n
ts
f
o
r
ab
o
u
t 1
9
.
9
% o
f
g
r
o
s
s
d
o
m
esti
c
p
r
o
d
u
ct.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Un
ified
vo
tin
g
-
b
a
s
ed
en
s
emb
l
e
lea
r
n
in
g
fo
r
r
ice
lea
f d
is
ea
s
e
d
etec
tio
n
…
(
Go
vi
n
d
a
r
a
ja
n
S
u
b
b
u
r
a
ma
n
)
1647
R
ice
cr
o
p
y
ield
an
d
q
u
ality
ar
e
o
f
ten
r
ed
u
ce
d
b
y
p
lan
t
d
is
ea
s
es,
ca
u
s
in
g
ec
o
n
o
m
ic
lo
s
s
es
to
f
ar
m
er
s
.
Sin
ce
d
is
ea
s
e
id
en
tific
atio
n
b
ased
o
n
v
is
u
al
ex
p
er
ien
ce
is
u
n
r
eliab
le,
an
au
to
m
ate
d
an
d
ac
cu
r
ate
ea
r
l
y
d
iag
n
o
s
is
s
y
s
tem
is
ess
en
tial
[
4
]
.
C
r
o
p
h
ea
lth
p
lay
s
a
v
i
tal
r
o
le
in
s
u
p
p
o
r
tin
g
g
lo
b
al
f
o
o
d
s
u
p
p
ly
an
d
s
u
s
tain
ab
le
f
ar
m
in
g
p
r
ac
tices.
Ho
wev
er
,
v
ar
i
o
u
s
f
ac
to
r
s
ca
n
lead
to
th
e
r
a
p
id
s
p
r
ea
d
o
f
d
is
ea
s
es
in
cr
o
p
s
,
ca
u
s
in
g
s
ig
n
if
ican
t
s
o
cial
an
d
ec
o
n
o
m
ic
ch
allen
g
es.
T
h
e
s
e
d
is
ea
s
es
n
o
t
o
n
ly
h
in
d
er
p
lan
t
g
r
o
wth
an
d
d
ev
elo
p
m
e
n
t
b
u
t
also
r
ed
u
ce
cr
o
p
y
ield
an
d
q
u
ality
,
m
a
k
in
g
th
em
a
m
ajo
r
co
n
tr
ib
u
to
r
to
p
r
o
d
u
ctiv
ity
lo
s
s
.
E
ar
ly
d
etec
tio
n
o
f
s
u
ch
illn
e
s
s
es
an
d
th
e
tim
ely
u
s
e
o
f
a
p
p
r
o
p
r
iate
p
esti
cid
es
ar
e
cr
u
cial
to
p
r
e
v
en
t
s
o
il
p
o
llu
tio
n
a
n
d
m
itig
ate
its
im
p
ac
t
[
5
]
.
T
im
ely
d
is
ea
s
e
d
iag
n
o
s
is
in
cr
o
p
s
is
cr
u
cial
f
o
r
s
u
s
tain
ab
le
ag
r
ic
u
ltu
r
al
d
ev
elo
p
m
e
n
t,
as
it
lo
wer
s
in
p
u
t
co
s
ts
an
d
p
r
ev
en
ts
y
ield
lo
s
s
.
I
n
r
ice
p
r
o
d
u
ctio
n
,
ea
r
ly
d
etec
tio
n
o
f
p
ests
an
d
d
is
ea
s
es
r
ed
u
ce
s
d
ep
en
d
en
cy
o
n
ch
em
ical
tr
ea
tm
en
ts
an
d
c
o
n
tr
ib
u
tes
to
im
p
r
o
v
ed
p
r
o
d
u
ctiv
ity
.
T
r
ad
itio
n
al
m
eth
o
d
s
,
s
u
ch
as
m
an
u
al
o
b
s
er
v
atio
n
,
ar
e
n
eith
e
r
f
ea
s
ib
le
n
o
r
ef
f
ec
tiv
e
f
o
r
lar
g
e
-
s
ca
le
f
ar
m
in
g
.
Ho
wev
e
r
,
ad
v
an
ce
m
e
n
ts
in
in
f
o
r
m
atio
n
t
ec
h
n
o
lo
g
y
h
av
e
s
ig
n
if
ican
tl
y
co
n
tr
ib
u
ted
to
im
p
r
o
v
i
n
g
cr
o
p
p
r
o
d
u
ctiv
ity
wh
ile
o
p
tim
izin
g
f
er
tili
ze
r
u
s
e.
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs),
a
cr
u
cial
ar
ea
o
f
ar
tific
ial
in
tellig
en
ce
,
p
r
o
v
id
e
a
u
s
ef
u
l
m
eth
o
d
f
o
r
cl
ass
if
y
in
g
p
lan
t
d
is
ea
s
es,
an
d
im
ag
e
p
r
o
ce
s
s
in
g
m
eth
o
d
s
h
av
e
g
r
ea
t
p
o
ten
tial
t
o
s
o
lv
e
p
r
o
b
le
m
s
in
th
e
ag
r
icu
ltu
r
al
in
d
u
s
tr
y
[
6
]
.
W
h
il
e
d
e
ep
l
ea
r
n
in
g
h
a
s
s
ig
n
if
i
ca
n
tl
y
im
p
r
o
v
ed
v
i
s
u
a
l
p
l
an
t
d
i
s
ea
s
e
r
e
co
g
n
it
io
n
,
th
e
s
c
ar
c
ity
o
f
s
o
m
e
d
i
s
e
as
e
in
s
t
an
c
e
s
le
ad
s
to
d
a
ta
im
b
a
lan
ce
r
e
la
ti
v
e
t
o
h
e
al
th
y
s
am
p
le
s
.
D
at
a
co
l
l
ec
t
ed
u
n
d
er
n
a
tu
r
al
co
n
d
i
tio
n
s
ty
p
ic
al
ly
ex
h
ib
it
l
im
i
ted
s
am
p
le
s
o
f
r
ar
e
d
is
ea
s
e
cl
a
s
s
e
s
,
le
ad
in
g
to
cl
a
s
s
i
m
b
al
an
c
e
i
s
s
u
es
i
n
m
ac
h
in
e
le
ar
n
i
n
g
s
y
s
te
m
s
.
T
h
i
s
i
m
b
a
lan
ce
ca
u
s
e
s
s
u
p
e
r
v
i
s
ed
le
ar
n
i
n
g
m
o
d
e
l
s
to
o
v
er
f
i
t,
a
s
d
e
ci
s
i
o
n
b
o
u
n
d
ar
ie
s
b
e
co
m
e
b
i
a
s
ed
t
o
war
d
th
e
d
o
m
in
an
t
c
la
s
s
e
s
[
7
]
,
[
8
]
.
B
y
b
r
o
ad
en
in
g
th
e
tr
ain
in
g
d
at
a
s
e
t
's
d
iv
er
s
ity
,
d
a
ta
au
g
m
en
ta
tio
n
ap
p
r
o
a
ch
e
s
ca
n
b
e
u
s
e
d
to
i
m
p
r
o
v
e
m
o
d
e
l
g
e
n
er
al
iz
at
io
n
an
d
o
v
er
co
m
e
th
i
s
d
if
f
icu
lty
.
U
s
in
g
tr
an
s
f
o
r
m
a
t
io
n
s
lik
e
r
o
t
at
io
n
(
u
p
to
4
0
°),
w
id
th
an
d
h
e
ig
h
t
s
h
if
t
s
(
u
p
to
3
0
%)
,
s
h
ea
r
tr
an
s
f
o
r
m
a
tio
n
(
0
.
2
)
,
zo
o
m
in
g
(
u
p
to
3
0
%),
b
r
ig
h
tn
e
s
s
ad
j
u
s
tm
en
t
(
b
et
we
en
0
.
7
an
d
1
.
3
)
,
ch
an
n
e
l
s
h
i
f
t
in
g
(
u
p
to
3
0
.
0
)
,
h
o
r
i
zo
n
t
al
f
lip
p
in
g
,
an
d
n
ea
r
e
s
t
-
n
e
ig
h
b
o
r
f
i
l
l
in
g
f
o
r
em
p
ty
r
eg
io
n
s
,
th
e
s
e
te
ch
n
iq
u
e
s
cr
e
at
e
d
if
f
e
r
en
t
v
er
s
io
n
s
o
f
ex
i
s
t
in
g
i
m
ag
e
s
.
T
h
e
s
e
t
ec
h
n
iq
u
es
ar
e
cr
u
c
ia
l
f
o
r
b
a
lan
cin
g
d
ata
s
et
s
,
p
r
ev
en
tin
g
o
v
er
f
i
t
tin
g
,
an
d
en
ab
l
in
g
m
o
d
el
s
to
le
ar
n
m
o
r
e
r
o
b
u
s
t
a
n
d
u
n
b
ia
s
ed
d
ec
i
s
io
n
b
o
u
n
d
ar
i
es
f
o
r
im
p
r
o
v
ed
d
i
s
ea
s
e
c
la
s
s
i
f
i
ca
tio
n
.
Dee
p
lear
n
in
g
m
et
h
o
d
s
h
av
e
b
ee
n
in
v
esti
g
ated
r
ec
e
n
tly
f
o
r
r
ice
lea
f
d
is
ea
s
e
d
etec
tio
n
,
with
v
er
y
en
co
u
r
a
g
in
g
r
esu
lts
.
Go
g
o
i
et
a
l.
[
9
]
f
o
r
in
s
tan
ce
,
r
ep
o
r
ted
an
ac
cu
r
ac
y
o
f
9
3
.
9
9
%
o
n
a
d
ataset
in
clu
d
in
g
8
,
8
8
3
im
a
g
es
u
s
in
g
a
th
r
ee
-
s
tag
e
C
NN
wi
th
tr
an
s
f
er
lear
n
in
g
u
s
in
g
E
f
f
icien
tNetB
5
.
L
ev
er
ag
in
g
a
Mo
b
ileNet
b
ac
k
b
o
n
e,
W
an
g
et
a
l.
[
1
0
]
c
r
ea
ted
an
atten
tio
n
-
b
ased
d
ep
th
wis
e
s
ep
ar
ab
le
n
e
u
r
al
n
etw
o
r
k
o
p
tim
ized
with
B
ay
esian
ap
p
r
o
ac
h
es,
wh
ich
ac
h
iev
ed
9
4
.
6
5
%
ac
cu
r
ac
y
o
n
a
p
u
b
lic
r
ice
illn
ess
d
atase
t
with
f
o
u
r
d
is
ea
s
e
class
if
icatio
n
s
.
Du
r
in
g
a
d
if
f
er
en
t
s
tu
d
y
,
Kr
is
h
n
am
o
o
r
th
y
et
a
l.
[
4
]
u
s
ed
I
n
ce
p
tio
n
R
esNetV2
an
d
tr
an
s
f
e
r
lear
n
in
g
t
o
class
if
y
p
ad
d
y
d
is
ea
s
e
o
f
th
e
leav
es
with
9
5
.
6
7
%
ac
cu
r
ac
y
.
T
h
ese
wo
r
k
s
h
ig
h
lig
h
t
p
o
ten
tial
o
f
d
ee
p
lear
n
in
g
m
o
d
els;
h
o
we
v
er
,
m
o
s
t
r
ely
o
n
s
in
g
le
p
r
e
tr
ain
ed
m
o
d
els
with
o
u
t
en
s
em
b
le
s
tr
ateg
ies.
T
o
ad
d
r
ess
th
is
g
ap
,
o
u
r
s
tu
d
y
in
t
eg
r
ates
th
e
m
u
ltip
le
p
r
etr
ain
ed
C
NNs
with
u
n
if
ied
v
o
tin
g
to
ac
h
iev
e
r
o
b
u
s
t
r
ice
leaf
d
is
ea
s
e
class
if
icat
io
n
.
W
h
ile
cu
r
r
en
t
r
esear
ch
p
r
im
ar
ily
f
o
cu
s
es
o
n
s
in
g
le
m
ac
h
in
e
lear
n
in
g
m
o
d
els
f
o
r
p
lan
t
d
is
ea
s
e
d
etec
tio
n
[
1
1
]
,
Fig
u
r
e
1
illu
s
t
r
ates
th
e
co
m
p
ar
is
o
n
b
etwe
en
tr
ad
itio
n
al
an
d
en
s
em
b
le
clas
s
if
icatio
n
m
eth
o
d
s
u
s
ed
in
th
is
s
tu
d
y
.
Fig
u
r
e
1
(
a
)
s
h
o
ws th
at
tr
ad
itio
n
al
class
if
ic
atio
n
s
tr
u
g
g
les with
co
m
p
lex
,
m
u
lti
-
d
is
ea
s
e
d
ata,
lead
in
g
to
lo
wer
ac
cu
r
ac
y
.
Fig
u
r
e
1
(
b
)
illu
s
tr
ates
h
o
w
en
s
em
b
le
class
if
icatio
n
en
h
an
ce
s
r
eliab
ilit
y
an
d
g
en
er
aliza
tio
n
b
y
in
teg
r
atin
g
m
u
ltip
le
m
o
d
els.
Ho
wev
er
,
d
ee
p
lear
n
in
g
m
eth
o
d
s
s
till
f
ac
e
in
ter
p
r
etab
ilit
y
ch
allen
g
es
in
u
n
d
er
s
tan
d
i
n
g
d
is
ea
s
e
p
atter
n
s
.
Fo
r
th
e
p
u
r
p
o
s
e
o
f
en
h
an
cin
g
th
e
ac
cu
r
ate
d
etec
tio
n
o
f
n
u
m
er
o
u
s
r
ice
illn
ess
o
f
th
e
leav
es
an
d
h
ig
h
lig
h
t
th
e
p
r
im
ar
y
ca
u
s
es
o
f
th
eir
o
cc
u
r
r
en
ce
,
t
h
is
s
tu
d
y
p
r
esen
ts
a
n
o
v
el
ap
p
r
o
ac
h
th
at
in
co
r
p
o
r
ates
en
s
em
b
le
lear
n
in
g
tech
n
iq
u
es.
E
n
s
em
b
le
lear
n
in
g
e
n
h
an
ce
s
p
r
e
d
ictiv
e
p
er
f
o
r
m
an
ce
b
y
in
te
g
r
atin
g
m
u
ltip
le
m
o
d
els,
an
d
in
th
is
s
tu
d
y
,
a
d
ee
p
lear
n
i
n
g
en
s
em
b
le
is
co
n
s
tr
u
cted
u
s
in
g
VGG1
6
,
Den
s
eNe
t1
2
1
,
I
n
ce
p
tio
n
V3
,
an
d
Xce
p
tio
n
.
T
o
en
s
u
r
e
r
o
b
u
s
t
an
d
ac
cu
r
ate
r
esu
lts
,
u
n
if
ied
v
o
tin
g
tech
n
iq
u
es,
m
ajo
r
ity
v
o
tin
g
,
s
o
f
t
v
o
tin
g
,
an
d
weig
h
ted
v
o
tin
g
,
a
r
e
u
tili
ze
d
to
co
n
s
o
lid
at
e
p
r
ed
ictio
n
s
f
r
o
m
in
d
iv
id
u
al
m
o
d
els,
h
ar
n
ess
in
g
th
eir
co
m
b
in
e
d
s
tr
en
g
th
s
f
o
r
i
m
p
r
o
v
e
d
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
T
h
e
k
ey
c
o
n
tr
ib
u
tio
n
s
o
f
th
is
r
esear
ch
in
clu
d
e:
–
C
o
m
p
iled
a
b
alan
ce
d
d
ataset
o
f
4
,
0
0
0
r
ice
leaf
im
ag
es
f
r
o
m
Kag
g
le
an
d
Go
o
g
le,
co
v
er
in
g
s
ix
ca
teg
o
r
ies.
–
F
in
e
-
tu
n
ed
f
o
u
r
p
r
etr
ain
ed
C
NN
m
o
d
els
(
VGG1
6
,
Den
s
eNe
t1
2
1
,
I
n
ce
p
tio
n
V3
,
an
d
Xce
p
tio
n
)
f
o
r
d
is
ea
s
e
class
if
icatio
n
.
–
T
o
en
h
a
n
ce
m
o
d
el
g
en
er
aliza
tio
n
an
d
a
d
d
r
ess
d
ata
im
b
ala
n
ce
,
d
ata
au
g
m
en
tatio
n
was
ap
p
lied
u
s
in
g
tr
an
s
f
o
r
m
atio
n
s
s
u
c
h
as
r
o
tatio
n
(
u
p
t
o
4
0
°),
wid
th
an
d
h
e
ig
h
t
s
h
if
ts
(
u
p
to
3
0
%),
s
h
ea
r
(
0
.
2
)
,
zo
o
m
(
u
p
to
3
0
%),
b
r
i
g
h
tn
ess
ad
ju
s
tm
en
t
(
0
.
7
–
1
.
3
)
,
ch
a
n
n
el
s
h
if
tin
g
(
u
p
to
3
0
.
0
)
,
h
o
r
izo
n
tal
f
lip
p
in
g
,
an
d
n
ea
r
est
-
n
eig
h
b
o
r
f
illi
n
g
.
–
P
r
o
p
o
s
e
th
e
f
ir
s
t u
n
if
ied
v
o
tin
g
en
s
em
b
le
f
o
r
r
ice
leaf
d
is
ea
s
e
d
etec
tio
n
u
s
in
g
t
h
ese
f
o
u
r
m
o
d
els.
–
Ou
r
en
s
em
b
le
(
s
o
f
t
v
o
tin
g
)
ac
h
iev
ed
9
7
% a
cc
u
r
ac
y
,
o
u
t
p
er
f
o
r
m
in
g
i
n
d
iv
id
u
al
m
o
d
els.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
6
4
6
-
1
6
6
3
1648
–
T
h
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
d
em
o
n
s
tr
ated
s
u
p
er
io
r
p
er
f
o
r
m
an
c
e,
s
h
o
wca
s
in
g
th
e
ef
f
ec
tiv
en
es
s
o
f
en
s
em
b
le
lear
n
in
g
an
d
a
u
n
if
ied
v
o
tin
g
s
ch
em
e
f
o
r
ac
cu
r
ate
class
if
icatio
n
o
f
r
ice
d
is
ea
s
es o
f
th
e
leav
e
s
.
(
a)
(
b
)
F
i
g
u
r
e
1
.
C
o
m
p
a
r
i
s
o
n
o
f
c
l
a
s
s
i
f
i
c
a
t
i
o
n
a
p
p
r
o
a
c
h
e
s
f
o
r
(
a
)
t
r
a
d
i
t
i
o
n
a
l
c
l
a
s
s
i
f
i
c
a
t
i
o
n
a
n
d
(
b
)
e
n
s
e
m
b
l
e
c
l
a
s
s
i
f
i
c
a
t
i
o
n
2.
RE
L
AT
E
D
WO
RK
S
T
h
is
s
ec
tio
n
p
r
esen
ts
a
s
u
m
m
ar
y
o
f
ea
r
lier
s
tu
d
ies
f
o
cu
s
ed
o
n
th
e
d
etec
tio
n
o
f
r
ice
leaf
d
i
s
ea
s
es.
T
o
p
r
ec
is
ely
d
etec
t
v
ar
io
u
s
r
ice
leaf
d
is
ea
s
es,
p
r
io
r
r
esear
ch
h
as
u
s
ed
a
v
ar
iety
o
f
m
ac
h
in
e
lear
n
in
g
an
d
d
ee
p
lear
n
in
g
tech
n
iq
u
es.
On
e
a
p
p
r
o
ac
h
co
m
b
in
ed
m
u
ltip
le
C
NN
ar
ch
itectu
r
es
in
to
an
en
s
em
b
le
f
r
am
ewo
r
k
t
o
im
p
r
o
v
e
class
if
icatio
n
p
er
f
o
r
m
an
ce
an
d
d
em
o
n
s
tr
ated
t
h
e
f
ea
s
ib
ilit
y
o
f
r
ea
l
-
tim
e
d
e
p
lo
y
m
en
t
th
r
o
u
g
h
s
m
ar
tp
h
o
n
e
ap
p
licatio
n
s
[
1
2
]
.
An
o
th
er
s
tu
d
y
in
tr
o
d
u
ce
d
a
l
ar
g
e
-
s
ca
le
d
ataset
o
f
r
ice
leaf
im
ag
es
an
d
ap
p
lied
en
s
em
b
le
C
NN
m
o
d
els,
s
h
o
win
g
th
at
lig
h
tweig
h
t
ar
c
h
itectu
r
es
ca
n
b
e
o
p
tim
ized
f
o
r
u
s
e
in
m
o
b
ile
an
d
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
e
n
v
ir
o
n
m
en
ts
[
1
3
]
.
C
o
m
p
ar
ativ
e
an
a
ly
s
es
h
av
e
also
b
ee
n
ca
r
r
ie
d
o
u
t
u
s
in
g
s
tan
d
ar
d
C
NN
m
o
d
els.
Fo
r
in
s
tan
ce
,
e
x
p
er
im
en
ts
with
I
n
ce
p
tio
n
V3
,
VGG1
6
,
Alex
Net,
Mo
b
ileNetV2
,
an
d
R
esNet1
8
o
n
m
o
r
e
th
a
n
7
,
0
0
0
im
ag
es
ac
h
iev
ed
h
ig
h
class
if
icatio
n
ac
cu
r
ac
y
[
1
4
]
.
Similar
ly
,
s
ev
er
al
C
NN
m
o
d
els
,
in
clu
d
in
g
VGG1
6
,
R
esNet5
0
,
an
d
Den
s
eNe
t1
2
1
wer
e
ev
al
u
ated
ac
r
o
s
s
m
u
ltip
le
d
atasets
co
n
tain
in
g
th
r
ee
to
f
iv
e
d
is
ea
s
e
ca
teg
o
r
ies,
with
Den
s
eNe
t1
2
1
f
o
u
n
d
to
b
e
t
h
e
m
o
s
t
ef
f
ec
tiv
e
[
1
5
]
.
Hy
b
r
id
f
r
am
ewo
r
k
s
h
a
v
e
also
g
ain
ed
atten
tio
n
.
An
en
s
em
b
l
e
o
f
C
NN
an
d
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
s
(
SVM)
was
d
e
v
elo
p
ed
to
class
if
y
f
iv
e
m
ajo
r
r
ice
d
is
ea
s
es,
ac
h
iev
in
g
co
m
p
etitiv
e
p
er
f
o
r
m
a
n
ce
[
1
6
]
.
T
r
a
n
s
f
er
lear
n
in
g
with
p
r
e
-
t
r
ain
ed
m
o
d
els
h
as
b
ee
n
wid
ely
ex
p
l
o
r
ed
as
well.
Fo
r
ex
am
p
le,
Xce
p
tio
n
Net,
R
esNet5
0
,
Den
s
eNe
t,
VG
G1
9
,
an
d
Sq
u
ee
ze
Net
wer
e
test
ed
o
n
a
m
u
lti
-
class
d
ataset
o
f
ten
d
is
ea
s
es,
with
Xce
p
tio
n
Net
s
h
o
win
g
t
h
e
b
est
r
e
s
u
lts
[
1
7
]
.
An
o
t
h
er
f
r
am
ewo
r
k
in
te
g
r
ated
VGG1
6
with
lig
h
t
g
r
ad
ien
t
b
o
o
s
tin
g
m
ac
h
in
e
(
L
ig
h
tGB
M
)
f
o
r
f
o
u
r
-
class
r
ice
d
is
ea
s
e
class
if
icatio
n
an
d
d
ep
lo
y
ed
t
h
e
s
y
s
tem
v
ia
a
clo
u
d
-
b
ased
m
o
b
ile
in
ter
f
ac
e
[
1
8
]
.
Oth
er
r
es
ea
r
ch
h
as
a
p
p
lied
ad
v
an
ce
d
a
r
ch
itectu
r
es
an
d
o
p
tim
izatio
n
m
eth
o
d
s
.
I
n
ce
p
tio
n
R
esNetV2
was
em
p
lo
y
ed
f
o
r
au
to
m
atic
d
is
ea
s
e
class
if
icatio
n
u
s
in
g
tr
an
s
f
er
le
ar
n
in
g
,
y
ield
in
g
h
ig
h
ac
cu
r
ac
y
o
n
a
d
ataset
o
f
th
r
ee
d
is
ea
s
e
ty
p
es
[
4
]
.
A
m
o
r
e
co
m
p
lex
p
ip
elin
e
in
teg
r
ated
p
r
ep
r
o
ce
s
s
in
g
,
s
eg
m
en
tatio
n
with
Seg
Net,
an
d
m
u
lti
-
f
ea
tu
r
e
e
x
tr
ac
tio
n
,
f
o
llo
wed
b
y
a
d
ee
p
r
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
(
Dee
p
R
NN)
class
if
ie
r
o
p
tim
ized
with
a
m
etah
eu
r
is
tic
alg
o
r
ith
m
f
o
r
im
p
r
o
v
e
d
d
is
ea
s
e
r
ec
o
g
n
itio
n
[
1
9
]
.
T
ab
le
1
p
r
o
v
id
es
a
co
m
p
a
r
ativ
e
o
v
er
v
iew
o
f
e
x
is
tin
g
ap
p
r
o
ac
h
es
,
d
etailin
g
th
e
m
o
d
els
im
p
lem
e
n
ted
,
th
e
s
ize
o
f
th
e
d
atasets
,
th
e
en
s
em
b
le
tech
n
i
q
u
es
ap
p
lie
d
,
an
d
th
e
ac
cu
r
ac
y
r
esu
lts
r
ep
o
r
ted
in
ea
c
h
s
tu
d
y
.
W
h
ile
in
d
iv
id
u
al
an
d
h
y
b
r
id
m
eth
o
d
s
h
av
e
s
h
o
wn
p
o
ten
tia
l,
th
er
e
r
em
ain
s
a
n
ee
d
f
o
r
m
o
r
e
r
o
b
u
s
t
an
d
a
d
ap
tab
le
m
o
d
els
f
o
r
r
ice
leaf
d
is
ea
s
e
d
etec
tio
n
.
Ma
n
y
e
x
is
tin
g
s
tu
d
ies
ar
e
lim
ited
t
o
a
f
ew
d
is
ea
s
e
ty
p
es
an
d
s
tr
u
g
g
le
with
g
en
er
aliza
t
io
n
ac
r
o
s
s
d
if
f
er
en
t
en
v
i
r
o
n
m
en
tal
co
n
d
itio
n
s
.
T
h
e
r
ec
o
m
m
en
d
ed
en
s
em
b
le
m
eth
o
d
e
n
h
an
ce
s
th
e
p
r
ec
is
io
n
an
d
r
esil
ien
ce
o
f
r
ice
leaf
d
i
s
ea
s
e
d
etec
tio
n
b
y
i
n
teg
r
atin
g
th
e
k
e
y
f
ea
tu
r
es
o
f
VGG1
6
,
Den
s
eNe
t1
2
1
,
I
n
ce
p
tio
n
V3
,
an
d
Xce
p
tio
n
.
T
h
e
m
o
d
el
is
ex
p
o
s
ed
to
a
v
ar
iety
o
f
tr
ain
in
g
s
am
p
les
th
r
o
u
g
h
th
e
u
s
e
o
f
v
a
r
io
u
s
d
ata
au
g
m
en
tatio
n
ap
p
r
o
ac
h
es,
s
u
ch
as
r
o
tatio
n
,
f
lip
p
in
g
,
a
n
d
zo
o
m
i
n
g
,
wh
ich
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Un
ified
vo
tin
g
-
b
a
s
ed
en
s
emb
l
e
lea
r
n
in
g
fo
r
r
ice
lea
f d
is
ea
s
e
d
etec
tio
n
…
(
Go
vi
n
d
a
r
a
ja
n
S
u
b
b
u
r
a
ma
n
)
1649
en
h
an
ce
s
th
e
m
o
d
el'
s
g
en
er
aliza
tio
n
an
d
ab
ilit
y
to
c
o
r
r
ec
tly
id
en
tify
v
ar
ie
d
illn
ess
p
atter
n
s
.
T
h
r
o
u
g
h
a
u
n
if
ied
v
o
tin
g
ap
p
r
o
ac
h
,
in
clu
d
in
g
m
ajo
r
ity
,
s
o
f
t,
an
d
weig
h
ted
v
o
tin
g
,
th
e
m
o
d
el
en
h
an
ce
s
p
r
e
d
ictio
n
ac
cu
r
ac
y
b
y
ag
g
r
eg
atin
g
th
e
o
u
tp
u
ts
f
r
o
m
m
u
ltip
le
m
o
d
els,
en
s
u
r
in
g
th
e
f
in
al
d
ec
is
io
n
r
ef
lects
th
e
co
n
s
en
s
u
s
o
f
d
iv
er
s
e
ch
ar
ac
ter
is
tics
.
T
h
is
m
eth
o
d
e
n
h
an
ce
s
n
o
t
ju
s
t
th
e
ac
cu
r
ac
y
o
f
th
e
m
o
d
el
b
u
t
also
its
r
o
b
u
s
tn
ess
an
d
ab
ilit
y
to
g
en
er
alize
,
en
a
b
lin
g
m
o
r
e
r
eli
ab
le
d
etec
tio
n
o
f
r
ice
leaf
d
is
ea
s
es u
n
d
er
d
iv
er
s
e
e
n
v
ir
o
n
m
e
n
tal
co
n
d
itio
n
s
.
T
ab
le
1
.
Su
m
m
a
r
izes
a
co
m
p
a
r
is
o
n
o
f
c
u
r
r
en
t
m
eth
o
d
s
u
s
ed
f
o
r
d
etec
tin
g
r
ice
leaf
d
is
ea
s
es
.
A
u
t
h
o
r
M
o
d
e
l
s/
met
h
o
d
s
En
se
mb
l
e
V
o
t
i
n
g
m
e
c
h
a
n
i
sm
D
a
t
a
s
e
t
s
i
z
e
A
c
c
u
r
a
c
y
(
%)
D
e
n
g
e
t
a
l
.
[1
2
]
D
e
n
seN
e
t
1
2
1
,
S
E
-
R
e
sN
e
t
5
0
,
R
e
sN
e
S
t
5
0
Y
e
s
Y
e
s
(
3
3
,
0
2
6
i
m
a
g
e
s,
6
d
i
se
a
ses)
91
P
a
i
e
t
a
l
.
[1
3
]
M
o
b
i
l
e
N
e
t
V
2
,
G
o
o
g
L
e
N
e
t
,
Ef
f
i
c
i
e
n
t
N
e
t
,
R
e
sN
e
t
3
4
,
D
e
n
seN
e
t
1
2
1
,
V
G
G
1
6
Y
e
s
Y
e
s
(
1
8
,
5
6
3
i
m
a
g
e
s,
6
d
i
se
a
ses)
9
3
.
4
D
e
b
e
t
a
l
.
[1
4
]
I
n
c
e
p
t
i
o
n
V
3
,
V
G
G
1
6
,
A
l
e
x
N
e
t
,
M
o
b
i
l
e
N
e
t
V
2
,
R
e
sN
e
t
1
8
No
No
(7
,
0
9
6
i
ma
g
e
s
,
5
d
i
s
e
a
ses)
9
6
.
2
3
I
sl
a
m
e
t
a
l
.
[1
5
]
V
G
G
1
6
,
R
e
sN
e
t
5
0
,
D
e
n
seN
e
t
1
2
1
No
No
(
D
a
t
a
se
t
1
:
3
8
6
i
ma
g
e
s
,
5
d
i
s
e
a
se
s
;
D
a
t
a
set
2
:
1
2
0
i
ma
g
e
s,
3
d
i
se
a
ses
;
D
a
t
a
s
e
t
3
:
2
7
8
i
ma
g
e
s
,
3
d
i
s
e
a
s
e
s)
9
1
.
6
7
,
8
5
.
4
5
,
8
9
.
8
5
H
a
r
i
d
a
sa
n
e
t
a
l
.
[1
6
]
C
N
N
+
S
V
M
Y
e
s
No
(
P
l
a
n
t
d
a
t
a
se
t
,
5
d
i
se
a
ses)
9
1
.
4
5
M
a
n
d
w
a
r
i
y
a
a
n
d
Jo
t
w
a
n
i
[1
7
]
X
c
e
p
t
i
o
n
N
e
t
,
R
e
sN
e
t
5
0
,
D
e
n
seN
e
t
1
2
1
,
V
G
G
1
9
,
S
e
q
u
e
e
z
e
N
e
t
a
n
d
C
N
N
No
No
(4
,
1
9
3
i
ma
g
e
s
,
1
0
d
i
se
a
ses)
9
3
.
1
3
B
h
o
w
mi
k
e
t
a
l
.
[1
8
]
V
G
G
1
6
(
t
r
a
n
sf
e
r
l
e
a
r
n
i
n
g
)
+
L
i
g
h
t
G
B
M
(
e
n
sem
b
l
e
)
Y
e
s
No
(3
,
3
5
5
i
ma
g
e
s
,
4
d
i
s
e
a
ses)
9
6
.
4
9
K
r
i
s
h
n
a
m
o
o
r
t
h
y
e
t
a
l
.
[
4
]
I
n
c
e
p
t
i
o
n
R
e
sN
e
t
V
2
No
No
(5
,
2
0
0
i
ma
g
e
s
,
3
d
i
s
e
a
ses)
9
5
.
6
7
B
h
o
w
mi
k
[
1
8
]
S
e
g
N
e
t
,
+
D
e
e
p
R
N
N
Y
e
s
No
(
1
2
0
i
ma
g
e
s
,
3
d
i
s
e
a
ses
9
0
.
5
3.
P
RO
P
O
SE
D
M
E
T
H
O
D
T
h
e
p
r
o
p
o
s
e
d
a
p
p
r
o
a
c
h
e
m
p
l
o
y
s
a
n
e
n
s
e
m
b
le
l
e
a
r
n
i
n
g
cl
a
s
s
if
i
e
r
f
o
r
d
e
t
e
c
t
i
n
g
r
ic
e
l
e
a
f
d
i
s
ea
s
es
.
T
h
e
p
r
o
c
e
s
s
i
n
v
o
l
v
es
s
e
v
e
r
al
s
ta
g
es
,
i
n
c
l
u
d
i
n
g
d
a
t
a
c
o
l
le
c
t
i
o
n
,
im
a
g
e
p
r
e
p
r
o
c
e
s
s
i
n
g
s
u
c
h
as
r
es
i
z
i
n
g
,
a
p
p
l
ic
a
t
i
o
n
o
f
d
a
t
a
a
u
g
m
e
n
t
a
ti
o
n
t
e
c
h
n
i
q
u
es,
t
r
a
i
n
i
n
g
,
a
n
d
e
v
a
l
u
at
i
n
g
d
e
ep
l
e
a
r
n
i
n
g
m
o
d
e
l
s
,
a
n
d
f
i
n
a
l
ly
c
o
m
b
i
n
i
n
g
t
h
e
i
r
p
r
e
d
i
c
t
i
o
n
s
t
h
r
o
u
g
h
a
u
n
i
f
i
e
d
v
o
t
i
n
g
-
b
a
s
e
d
e
n
s
e
m
b
l
e
s
t
r
a
te
g
y
.
T
o
e
n
h
a
n
c
e
t
h
e
o
v
e
r
a
l
l
p
e
r
f
o
r
m
a
n
c
e
o
f
t
h
e
m
o
d
e
l
,
t
h
e
r
i
c
e
l
e
a
f
d
is
e
as
e
d
a
t
as
et
u
n
d
e
r
w
e
n
t
p
r
e
p
r
o
c
e
s
s
i
n
g
a
n
d
d
a
t
a
a
u
g
m
e
n
t
a
t
i
o
n
.
T
h
e
f
o
l
lo
w
i
n
g
s
u
b
s
e
ct
i
o
n
s
d
e
s
c
r
i
b
e
e
a
c
h
s
t
e
p
o
f
t
h
e
p
r
o
p
o
s
e
d
m
e
t
h
o
d
o
l
o
g
y
,
w
h
i
c
h
is
i
ll
u
s
t
r
a
t
e
d
i
n
F
i
g
u
r
e
2
.
Fig
u
r
e
2
.
W
o
r
k
f
lo
w
o
f
p
r
o
p
o
s
ed
en
s
em
b
le
lear
n
i
n
g
with
u
n
i
f
ied
v
o
tin
g
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
6
4
6
-
1
6
6
3
1650
3
.
1
.
Da
t
a
s
et
des
cr
iptio
n
T
h
e
r
ice
lea
f
d
is
ea
s
e
d
ataset
was
m
ain
ly
s
o
u
r
ce
d
f
r
o
m
Kag
g
le
an
d
s
u
p
p
lem
en
ted
with
ad
d
itio
n
al
im
ag
es
f
r
o
m
Go
o
g
le,
c
o
m
p
r
is
in
g
s
ix
class
es:
f
iv
e
d
is
ea
s
e
ty
p
es
—
n
ar
r
o
w
b
r
o
wn
s
p
o
t,
le
af
s
ca
ld
,
leaf
b
last
,
b
r
o
wn
s
p
o
t,
an
d
b
ac
te
r
ial
leaf
b
lig
h
t
—
an
d
o
n
e
h
ea
lth
y
ca
teg
o
r
y
,
as
illu
s
tr
ated
in
Fi
g
u
r
e
3
.
A
to
tal
o
f
4
,
0
0
0
im
ag
es
wer
e
in
cl
u
d
ed
an
d
d
iv
id
e
d
i
n
to
tr
ain
i
n
g
,
v
alid
atio
n
,
a
n
d
test
in
g
s
ets
u
s
in
g
a
7
0
:3
0
r
atio
.
T
r
ain
in
g
s
et
c
o
n
tain
ed
3
2
7
im
ag
es
p
er
class
,
v
alid
atio
n
s
et
1
4
0
im
a
g
es
p
er
class
,
an
d
test
in
g
s
et
2
0
0
im
ag
es
p
er
class
.
I
m
ag
es
we
r
e
r
escal
ed
ac
co
r
d
in
g
to
t
h
e
in
p
u
t
d
i
m
en
s
io
n
s
r
eq
u
ir
e
d
b
y
ea
c
h
p
r
et
r
ain
ed
ar
ch
itectu
r
e,
u
s
in
g
2
2
4
×2
2
4
p
ix
els
[
2
0
]
f
o
r
VGG1
6
an
d
Den
s
eNe
t1
2
1
,
an
d
2
9
9
×
2
9
9
p
ix
els f
o
r
I
n
ce
p
tio
n
V3
an
d
Xce
p
tio
n
.
Fig
u
r
e
3
.
E
x
am
p
les o
f
r
ice
lea
f
d
is
ea
s
e
im
ag
es
3
.
2
.
P
re
pro
ce
s
s
ing
a
nd
a
ug
m
ent
a
t
io
n
T
h
is
s
tag
e
aim
s
to
p
r
ep
ar
e
im
ag
e
d
ata
s
o
th
at
it
is
m
o
r
e
co
n
s
is
ten
t
an
d
v
ar
ied
to
s
u
p
p
o
r
t
th
e
m
o
d
el
tr
ain
in
g
p
r
o
ce
s
s
.
−
N
o
r
m
aliza
tio
n
:
p
ix
el
v
alu
es
wer
e
r
escaled
b
y
a
f
ac
to
r
o
f
1/
2
5
5
,
en
s
u
r
in
g
th
e
y
f
all
with
in
t
h
e
[
0
,
1
]
r
an
g
e
f
o
r
s
tab
le
tr
ain
in
g
an
d
f
aster
c
o
n
v
er
g
en
ce
.
−
Au
g
m
en
tatio
n
:
to
in
cr
ea
s
e
th
e
v
ar
iety
o
f
tr
ain
in
g
s
am
p
les
an
d
lim
it
o
v
er
f
itti
n
g
,
s
ev
er
al
im
ag
e
tr
an
s
f
o
r
m
atio
n
m
eth
o
d
s
wer
e
em
p
lo
y
ed
.
T
h
ese
in
clu
d
e
d
co
n
tr
o
lled
r
o
tatio
n
s
o
f
u
p
t
o
4
0
°
,
h
o
r
izo
n
tal
a
n
d
v
er
tical
tr
an
s
latio
n
s
r
ea
ch
in
g
3
0
%,
g
eo
m
etr
ic
d
is
to
r
tio
n
u
s
i
n
g
a
s
h
ea
r
f
ac
to
r
o
f
0
.
2
,
zo
o
m
o
p
er
atio
n
s
u
p
to
3
0
%,
in
ten
s
ity
v
ar
iatio
n
th
r
o
u
g
h
b
r
ig
h
tn
ess
s
ca
lin
g
b
etw
ee
n
0
.
7
a
n
d
1
.
3
,
co
l
o
r
c
h
an
n
e
l
ad
ju
s
tm
en
ts
u
p
to
3
0
.
0
,
an
d
h
o
r
iz
o
n
tal
m
ir
r
o
r
in
g
o
f
im
ag
es.
−
Sp
litt
in
g
:
th
e
av
ailab
le
d
ata
wer
e
o
r
g
an
ized
in
to
s
ep
ar
a
te
lear
n
in
g
,
tu
n
in
g
,
an
d
ev
al
u
atio
n
g
r
o
u
p
s
f
o
llo
win
g
a
p
r
ed
e
f
in
ed
r
atio
,
wh
er
e
ea
ch
class
co
n
tr
ib
u
ted
3
2
7
im
a
g
es
f
o
r
m
o
d
el
lea
r
n
in
g
,
1
4
0
im
ag
es
f
o
r
p
a
r
am
eter
v
alid
atio
n
,
an
d
2
0
0
im
a
g
es f
o
r
f
i
n
al
ass
ess
m
e
n
t.
3
.
3
.
B
a
s
e
mo
dels
I
n
th
is
s
tu
d
y
,
f
o
u
r
p
r
etr
ain
ed
d
ee
p
v
is
io
n
ar
c
h
itectu
r
es
wer
e
ad
ap
ted
th
r
o
u
g
h
f
in
e
-
t
u
n
in
g
to
class
if
y
r
ice
leaf
co
n
d
itio
n
s
,
an
d
th
e
k
ey
ch
ar
ac
ter
is
tics
o
f
th
ese
ch
o
s
en
n
etwo
r
k
s
ar
e
s
u
m
m
ar
ized
in
T
ab
le
2
.
T
r
an
s
f
er
lear
n
in
g
:
tr
a
n
s
f
er
lear
n
in
g
wa
s
ap
p
lied
b
y
s
tar
tin
g
ea
ch
n
et
wo
r
k
with
weig
h
ts
lear
n
e
d
f
r
o
m
I
m
a
g
eNe
t
an
d
th
en
ad
a
p
tin
g
th
e
m
o
d
els
t
o
th
e
r
ice
leaf
d
is
ea
s
e
d
ata
th
r
o
u
g
h
ad
d
itio
n
al
tr
ai
n
in
g
.
Ar
ch
itec
tu
r
al
m
o
d
if
icatio
n
s
:
th
e
ar
ch
itectu
r
e
in
co
r
p
o
r
ates
a
f
ea
tu
r
e
s
u
m
m
ar
izatio
n
s
tag
e,
f
o
llo
wed
b
y
a
n
o
n
lin
ea
r
f
u
lly
co
n
n
ec
te
d
lay
er
,
a
r
eg
u
lar
izatio
n
co
m
p
o
n
e
n
t
with
a
0
.
4
d
r
o
p
r
ate,
a
n
d
a
f
in
al
class
if
icat
io
n
lay
er
th
at
o
u
tp
u
ts
p
r
o
b
ab
ilit
ies
f
o
r
s
ix
class
e
s
.
T
ab
le
2
.
C
h
ar
ac
ter
is
tics
o
f
p
r
e
tr
ain
ed
C
NN
ar
ch
itectu
r
es
M
o
d
e
l
P
a
r
a
me
t
e
r
s
N
o
t
a
b
l
e
f
e
a
t
u
r
e
s
I
n
p
u
t
si
z
e
V
G
G
1
6
~
1
3
8
M
D
e
e
p
b
u
t
si
m
p
l
e
s
e
q
u
e
n
t
i
a
l
a
r
c
h
i
t
e
c
t
u
r
e
2
2
4
×
2
2
4
D
e
n
seN
e
t
1
2
1
~
8
M
D
e
n
se
c
o
n
n
e
c
t
i
v
i
t
y
,
p
a
r
a
m
e
t
e
r
e
f
f
i
c
i
e
n
c
y
2
2
4
×
2
2
4
I
n
c
e
p
t
i
o
n
V
3
~
2
4
M
M
u
l
t
i
-
sca
l
e
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
2
9
9
×
2
9
9
X
c
e
p
t
i
o
n
~
2
2
M
D
e
p
t
h
w
i
s
e
s
e
p
a
r
a
b
l
e
c
o
n
v
o
l
u
t
i
o
n
s
2
9
9
×
2
9
9
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Un
ified
vo
tin
g
-
b
a
s
ed
en
s
emb
l
e
lea
r
n
in
g
fo
r
r
ice
lea
f d
is
ea
s
e
d
etec
tio
n
…
(
Go
vi
n
d
a
r
a
ja
n
S
u
b
b
u
r
a
ma
n
)
1651
3
.
4
.
E
ns
em
ble
lea
rning
s
t
ra
t
eg
ies
T
o
en
h
an
ce
class
if
icatio
n
ac
cu
r
ac
y
a
n
d
g
en
e
r
aliza
tio
n
in
r
i
ce
leaf
d
is
ea
s
e
d
etec
tio
n
,
t
h
r
e
e
en
s
em
b
le
m
eth
o
d
s
wer
e
a
p
p
lied
with
in
a
u
n
if
ied
v
o
tin
g
s
tr
ateg
y
,
wh
e
r
e
ea
ch
m
et
h
o
d
co
n
tr
i
b
u
tes
to
th
e
f
in
al
d
ec
is
io
n
th
r
o
u
g
h
co
m
b
in
ed
p
r
ed
ictio
n
s
.
T
h
is
ap
p
r
o
ac
h
im
p
r
o
v
es
r
o
b
u
s
tn
ess
b
y
lev
er
ag
in
g
th
e
s
tr
en
g
th
s
o
f
m
u
ltip
le
m
o
d
els
in
s
tead
o
f
r
el
y
in
g
o
n
a
s
in
g
le
class
if
ier
.
As
a
r
esu
lt
,
th
e
o
v
er
all
p
er
f
o
r
m
an
ce
in
i
d
en
tify
in
g
r
ice
lea
f
d
is
ea
s
es b
ec
o
m
es m
o
r
e
s
tab
le
an
d
r
eliab
le.
3
.
4
.
1
.
M
a
j
o
rit
y
v
o
t
ing
Un
d
er
th
is
ap
p
r
o
ac
h
,
ea
ch
tr
a
in
ed
m
o
d
el
p
r
o
v
id
es
its
p
r
ed
ictio
n
,
an
d
th
e
o
u
tc
o
m
e
is
d
eter
m
in
ed
b
y
th
e
lab
el
th
at
r
ec
eiv
es th
e
g
r
ea
test
o
v
er
all
s
u
p
p
o
r
t
[
2
1
]
as in
(
1
)
.
̂
=
ℎ
=
∑
(
,
)
=
1
(
1
)
W
h
er
e
̂
is
r
ep
r
e
s
en
ts
s
elec
ted
o
u
tp
u
t
ca
te
g
o
r
y
[
2
1
]
,
is
d
en
o
tes
to
tal
n
u
m
b
er
o
f
p
ar
ticip
ati
n
g
m
o
d
els
,
is
in
d
icate
s
a
c
lass
lab
el
,
an
d
(
,
)
is
eq
u
als
1
wh
en
t
h
e
ℎ
m
o
d
el
p
r
e
d
icts
clas
s
j,
an
d
0
o
t
h
er
wis
e.
3
.
4
.
2
.
So
f
t
v
o
t
ing
I
n
th
is
a
p
p
r
o
ac
h
,
p
r
o
b
ab
ilit
y
s
co
r
es
p
r
o
d
u
ce
d
b
y
ea
c
h
m
o
d
el
ar
e
c
o
m
b
in
e
d
b
y
co
m
p
u
tin
g
th
eir
m
ea
n
v
alu
es,
an
d
th
e
ca
teg
o
r
y
wi
th
th
e
s
tr
o
n
g
est
o
v
er
all
co
n
f
id
en
ce
is
s
elec
ted
.
B
y
r
ely
in
g
o
n
co
n
f
id
e
n
ce
in
f
o
r
m
atio
n
r
ath
er
th
an
o
n
ly
d
is
cr
ete
lab
els,
th
e
en
s
em
b
l
e
b
en
ef
its
f
r
o
m
th
e
r
eliab
ilit
y
o
f
all
in
d
iv
id
u
al
p
r
ed
ictio
n
s
[
2
1
]
,
[
2
2
]
as d
escr
ib
ed
in
(
2
)
.
̂
=
1
∑
=
1
(
)
(
2
)
W
h
er
e
(
)
is
p
r
o
b
ab
ilit
y
o
f
class
p
r
ed
icted
b
y
m
o
d
el
3
.
4
.
3
.
Weig
hte
d
v
o
t
ing
I
n
th
is
m
eth
o
d
,
m
o
d
el
weig
h
ts
wer
e
d
eter
m
in
ed
ac
co
r
d
in
g
to
v
alid
atio
n
p
er
f
o
r
m
an
ce
a
n
d
o
v
er
all
ac
cu
r
ac
y
.
Den
s
eNe
t1
2
1
an
d
I
n
ce
p
tio
n
V3
,
wh
ich
s
h
o
wed
s
u
p
er
io
r
r
esu
lts
am
o
n
g
th
e
s
elec
ted
m
o
d
els,
wer
e
ea
ch
g
iv
e
n
a
weig
h
t
o
f
0
.
3
0
,
wh
ile
VGG1
6
an
d
Xce
p
tio
n
wer
e
ass
ig
n
ed
lo
wer
weig
h
ts
o
f
0
.
2
0
d
u
e
t
o
th
ei
r
r
elativ
ely
wea
k
er
p
er
f
o
r
m
an
ce
.
T
h
e
f
in
al
class
p
r
ed
icti
o
n
was
p
r
o
d
u
ce
d
b
y
ca
lcu
latin
g
a
weig
h
ted
co
m
b
in
atio
n
o
f
th
e
class
p
r
o
b
ab
ilit
y
o
u
t
p
u
ts
f
r
o
m
all
m
o
d
els,
allo
win
g
h
i
g
h
er
-
p
er
f
o
r
m
in
g
m
o
d
els
to
h
a
v
e
g
r
ea
ter
im
p
ac
t w
h
ile
s
till
b
en
e
f
itin
g
f
r
o
m
th
e
d
i
v
er
s
ity
o
f
t
h
e
en
s
em
b
le
as in
(
3
)
[
2
2
]
.
̂
=
∑
=
1
(
)
(
3
)
W
h
er
e
is
w
eig
h
t
as
s
ig
n
ed
to
m
o
d
el
(
e.
g
.
,
b
ased
o
n
ac
cu
r
ac
y
)
,
(
)
is
p
r
o
b
ab
ilit
y
o
f
class
p
r
ed
icted
b
y
m
o
d
el
.
T
h
e
s
eq
u
e
n
ce
o
f
s
tep
s
o
u
tl
in
ed
is
u
s
ed
to
ev
alu
ate
th
e
ex
p
la
n
atio
n
s
f
o
r
u
n
if
ie
d
v
o
tin
g
in
en
s
em
b
le
lear
n
in
g
.
−
Step
1
:
in
p
u
t: m
o
d
els
{M
1
,
M
2
,
.
.
.
,
M
n
},
test
d
ataset
D,
tr
u
e
la
b
els Y
(
o
p
tio
n
al)
,
v
o
tin
g
ty
p
e,
wei
g
h
ts
W
−
Step
2
:
in
itialize
:
P
=
s
o
f
t p
r
ed
ictio
n
s
(
p
r
o
b
ab
ili
ties
)
,
C
=
m
ajo
r
ity
p
r
ed
ictio
n
s
(
class
lab
els).
−
Step
3
:
co
llect
p
r
ed
ictio
n
s
:
Fo
r
ea
ch
m
o
d
el
:
Pre
d
ict
p
r
o
b
a
b
ilit
ies
=
.
(
)
.
C
o
m
p
u
te
m
ajo
r
ity
p
r
ed
ictio
n
s
=
(
)
.
Sto
r
e
in
an
d
in
.
−
Step
4
:
ag
g
r
e
g
ate
p
r
e
d
ictio
n
s
:
So
f
t
v
o
tin
g
:
c
o
m
p
u
te
=
(
)
,
th
en
=
(
)
.
W
eig
h
ted
v
o
tin
g
:
n
o
r
m
aliz
e
,
co
m
p
u
te
ℎ
=
∑
[
]
.
[
]
,
an
d
=
(
ℎ
)
.
Ma
jo
r
ity
v
o
tin
g
: f
o
r
ea
ch
s
am
p
le,
d
eter
m
in
e
t
h
e
m
o
s
t f
r
e
q
u
e
n
t c
lass
f
r
o
m
.
−
Step
5
:
f
in
al
p
r
e
d
ictio
n
s
.
−
Step
6
:
en
d
f
o
r
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
6
4
6
-
1
6
6
3
1652
3
.
5
.
T
ra
ini
ng
co
nfig
ura
t
io
n
T
h
e
m
o
d
els
wer
e
tr
ain
ed
with
th
e
Ad
am
o
p
tim
izatio
n
alg
o
r
ith
m
u
s
in
g
a
lear
n
in
g
r
ate
o
f
0
.
0
0
0
0
5
,
ca
teg
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
,
a
b
atch
s
ize
o
f
3
2
,
a
n
d
a
to
tal
o
f
5
0
tr
ain
i
n
g
ep
o
ch
s
.
T
o
im
p
r
o
v
e
g
en
er
aliza
tio
n
,
d
r
o
p
o
u
t
lay
e
r
s
an
d
d
ata
au
g
m
en
tatio
n
wer
e
ap
p
lied
d
u
r
in
g
tr
ai
n
in
g
.
T
h
e
ex
p
er
im
en
tal
s
etu
p
u
tili
ze
d
m
u
ltip
le
h
ar
d
war
e
ac
ce
ler
ato
r
s
to
en
s
u
r
e
ef
f
icien
t
p
r
o
ce
s
s
in
g
,
with
C
PUs
h
an
d
lin
g
d
ata
p
r
ep
r
o
ce
s
s
in
g
,
T
4
GPUs
s
u
p
p
o
r
tin
g
co
m
p
u
tatio
n
ally
in
te
n
s
iv
e
o
p
er
atio
n
s
,
an
d
T
PU
v
2
-
8
r
eso
u
r
ce
s
u
s
ed
f
o
r
h
ig
h
ly
p
ar
allel
wo
r
k
lo
a
d
s
.
4.
E
NS
E
M
B
L
E
M
O
DE
L
–
VG
G
1
6
,
DE
NSE
N
E
T
1
2
1
,
I
NCE
P
T
I
O
NV3
,
AND
XC
E
P
T
I
O
N
T
h
is
r
esear
ch
aim
ed
to
in
v
esti
g
ate
th
e
ef
f
ec
tiv
en
ess
o
f
co
m
b
i
n
in
g
VGG1
6
,
Den
s
eNe
t1
2
1
,
I
n
ce
p
tio
n
V3
,
an
d
Xce
p
tio
n
i
n
to
a
s
in
g
le
en
s
em
b
le
f
r
am
e
wo
r
k
to
en
h
a
n
ce
class
if
icatio
n
p
er
f
o
r
m
an
ce
an
d
m
o
d
el
r
o
b
u
s
tn
ess
b
y
ex
p
lo
itin
g
th
e
co
m
p
lem
en
tar
y
ca
p
ab
ilit
ies o
f
ea
ch
ar
ch
itectu
r
e.
4
.
1
.
VG
G
1
6
VGG1
6
is
a
well
-
k
n
o
wn
im
a
g
e
r
ec
o
g
n
itio
n
a
r
ch
itectu
r
e
d
e
v
elo
p
ed
b
y
an
ac
a
d
em
ic
r
ese
ar
ch
team
,
ch
ar
ac
ter
ized
b
y
a
s
im
p
le
lay
e
r
ed
s
tr
u
ctu
r
e
t
h
at
r
elies
o
n
s
m
all
3
×3
co
n
v
o
lu
tio
n
f
ilter
s
.
I
t a
ch
iev
ed
s
ig
n
if
ican
t
s
u
cc
ess
in
th
e
2
0
1
4
I
m
ag
eNe
t
lar
g
e
s
ca
le
v
is
u
al
r
ec
o
g
n
itio
n
ch
allen
g
e
(
2
0
1
4
I
L
SVR
C
)
,
d
em
o
n
s
tr
atin
g
t
h
e
p
o
wer
o
f
d
ee
p
n
etwo
r
k
s
f
o
r
h
ier
ar
ch
ical
f
ea
tu
r
e
ex
tr
ac
tio
n
[
2
3
]
.
W
ith
1
6
weig
h
t
lay
er
s
,
VGG1
6
in
f
lu
en
ce
d
th
e
d
ev
elo
p
m
en
t
o
f
d
ee
p
er
n
etwo
r
k
s
an
d
its
m
o
d
u
lar
d
esig
n
h
as
in
s
p
ir
ed
m
an
y
s
u
b
s
eq
u
en
t
m
o
d
els
i
n
co
m
p
u
ter
v
is
io
n
.
I
n
th
is
s
tu
d
y
,
th
e
m
o
d
if
ied
VGG1
6
n
etwo
r
k
p
r
o
ce
s
s
es c
o
lo
r
im
ag
es r
esized
to
a
f
ix
ed
s
q
u
ar
e
r
eso
lu
tio
n
s
u
itab
le
f
o
r
its
in
p
u
t
lay
er
.
T
h
e
ar
c
h
itectu
r
e
c
o
n
tain
s
th
ir
teen
co
n
v
o
lu
tio
n
s
tag
es
ar
r
an
g
ed
in
co
n
s
ec
u
tiv
e
b
lo
c
k
s
,
with
a
n
o
n
lin
ea
r
ac
tiv
atio
n
a
p
p
lied
af
te
r
ea
ch
s
tag
e.
T
h
e
n
etwo
r
k
b
e
g
in
s
with
an
in
itial
s
tag
e
th
at
ap
p
lies
two
f
ea
tu
r
e
ex
tr
ac
tio
n
lay
er
s
u
s
in
g
a
s
m
all
k
er
n
el,
af
ter
wh
ich
s
p
atial
r
eso
lu
tio
n
is
r
ed
u
ce
d
th
r
o
u
g
h
p
o
o
lin
g
.
I
n
t
h
e
n
ex
t
s
tag
e,
th
e
d
ep
th
o
f
th
e
f
ea
tu
r
e
m
ap
s
is
ex
p
an
d
e
d
wh
ile
m
ain
tain
in
g
th
e
s
am
e
k
er
n
el
d
im
en
s
io
n
s
,
a
g
ain
f
o
llo
wed
b
y
a
d
o
wn
s
am
p
lin
g
o
p
er
a
tio
n
.
T
h
e
th
ir
d
s
tag
e
f
u
r
th
er
d
e
ep
en
s
th
e
n
etwo
r
k
b
y
in
tr
o
d
u
cin
g
th
r
ee
s
u
cc
ess
iv
e
f
ea
tu
r
e
ex
tr
ac
tio
n
lay
er
s
b
ef
o
r
e
r
e
d
u
cin
g
s
p
atial
s
ize.
T
h
e
f
in
al
two
s
tag
es
f
o
llo
w
a
s
im
ilar
d
esig
n
p
atte
r
n
,
ea
ch
em
p
lo
y
i
n
g
th
r
ee
f
ea
tu
r
e
ex
tr
ac
tio
n
lay
er
s
with
a
h
ig
h
er
n
u
m
b
er
o
f
ch
an
n
els,
an
d
ea
ch
s
tag
e
co
n
clu
d
es
with
a
p
o
o
lin
g
o
p
e
r
atio
n
to
p
r
o
g
r
ess
iv
ely
co
n
d
e
n
s
e
th
e
f
ea
tu
r
e
m
ap
s
.
Fo
llo
win
g
th
e
f
ea
tu
r
e
ex
tr
ac
ti
o
n
s
tag
es,
th
e
m
o
d
el
m
o
v
es
i
n
to
a
s
er
ies
o
f
d
en
s
e
la
y
er
s
,
wh
er
e
th
e
f
ir
s
t
two
h
av
e
a
h
ig
h
n
u
m
b
er
o
f
u
n
its
with
n
o
n
lin
ea
r
ac
tiv
atio
n
,
an
d
th
e
f
in
al
lay
er
r
ed
u
ce
s
th
e
r
ep
r
esen
tatio
n
to
a
s
m
aller
s
et
o
f
o
u
tp
u
t
u
n
its
.
Fin
ally
,
a
So
f
tMa
x
o
u
tp
u
t
lay
e
r
is
ap
p
lied
to
p
e
r
f
o
r
m
class
if
icatio
n
ac
r
o
s
s
th
e
tar
g
et
ca
teg
o
r
ies,
as illu
s
tr
ated
in
Fig
u
r
e
4
.
4
.
2
.
DenseNet
1
2
1
Den
s
eNe
t1
2
1
u
s
es
d
en
s
e
co
n
n
ec
tio
n
s
wh
er
e
ea
ch
lay
e
r
co
n
n
ec
ts
to
all
p
r
ec
ed
in
g
lay
er
s
,
im
p
r
o
v
in
g
in
f
o
r
m
atio
n
f
lo
w,
g
r
ad
ien
t
p
r
o
p
ag
atio
n
,
an
d
tr
ai
n
in
g
e
f
f
icien
cy
.
W
ith
1
2
1
la
y
er
s
,
Den
s
eNe
t1
2
1
ef
f
icien
tl
y
ca
p
tu
r
es
f
in
e
-
g
r
ain
e
d
f
ea
tu
r
es
wh
ile
k
ee
p
in
g
t
h
e
n
u
m
b
e
r
o
f
p
ar
am
eter
s
m
a
n
ag
ea
b
le,
m
a
k
in
g
it
well
-
s
u
ited
f
o
r
v
is
u
al
r
ec
o
g
n
itio
n
ap
p
licatio
n
s
[
2
3
]
.
I
n
th
is
r
esear
ch
,
th
e
m
o
d
if
ied
Den
s
eNe
t1
2
1
m
o
d
el
ac
ce
p
ts
co
lo
r
im
ag
es
r
esized
to
a
f
ix
e
d
r
eso
lu
tio
n
o
f
2
2
4
b
y
2
2
4
p
ix
els.
T
h
e
n
et
wo
r
k
b
e
g
in
s
with
a
co
n
v
o
lu
ti
o
n
al
o
p
e
r
atio
n
u
s
in
g
6
4
f
ilter
s
with
a
r
elativ
ely
la
r
g
e
k
er
n
el,
co
m
b
in
ed
with
a
d
o
wn
s
am
p
lin
g
s
tep
an
d
a
n
o
n
lin
ea
r
ac
tiv
atio
n
,
f
o
llo
wed
b
y
a
p
o
o
lin
g
la
y
er
th
at
f
u
r
t
h
er
r
ed
u
c
es
th
e
s
p
atial
d
im
en
s
io
n
s
o
f
th
e
f
ea
tu
r
e
m
ap
s
.
T
h
e
m
o
d
e
l
co
n
tin
u
es
with
m
u
ltip
le
d
e
n
s
ely
co
n
n
ec
ted
b
lo
ck
s
in
ter
lea
v
ed
with
tr
an
s
itio
n
lay
er
s
.
T
h
e
f
ir
s
t
d
en
s
e
b
lo
ck
co
n
s
is
ts
o
f
f
o
u
r
s
eq
u
en
tial
lay
er
s
,
ea
ch
em
p
lo
y
i
n
g
co
n
v
o
lu
ti
o
n
s
with
s
m
all
an
d
m
ed
iu
m
-
s
i
ze
d
k
er
n
els,
alo
n
g
with
b
atch
n
o
r
m
aliza
tio
n
a
n
d
n
o
n
lin
ea
r
ac
tiv
atio
n
f
u
n
ctio
n
s
.
T
h
is
is
f
o
llo
wed
b
y
a
tr
a
n
s
itio
n
lay
e
r
with
a
1
×1
co
n
v
o
l
u
tio
n
o
f
1
2
8
f
ilter
s
an
d
2
×2
a
v
er
ag
e
p
o
o
lin
g
f
o
r
d
o
w
n
s
am
p
lin
g
.
Den
s
e
b
lo
ck
s
2
,
3
,
an
d
4
f
o
llo
w
th
e
s
am
e
s
tr
u
ctu
r
e,
ea
ch
with
f
o
u
r
lay
er
s
o
f
3
2
f
ilter
s
,
s
ep
ar
ated
b
y
tr
an
s
itio
n
lay
er
s
.
Af
ter
th
e
f
in
al
d
en
s
e
b
lo
ck
,
g
lo
b
al
av
e
r
ag
e
p
o
o
lin
g
r
ed
u
ce
s
ea
ch
f
ea
tu
r
e
m
ap
t
o
1
×1
,
f
o
llo
we
d
b
y
a
f
u
lly
co
n
n
ec
ted
lay
e
r
with
1
,
0
2
4
n
eu
r
o
n
s
an
d
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U
)
ac
tiv
atio
n
,
an
d
f
in
ally
a
So
f
tMa
x
lay
er
f
o
r
class
if
icatio
n
,
as
s
h
o
wn
in
Fig
u
r
e
5
.
4
.
3
.
I
ncept
io
nV3
I
n
ce
p
tio
n
V3
is
an
ef
f
icien
t
C
NN
d
ev
elo
p
ed
b
y
Go
o
g
le,
k
n
o
wn
f
o
r
its
in
n
o
v
ativ
e
"
in
ce
p
tio
n
m
o
d
u
le
,
"
wh
ic
h
p
er
f
o
r
m
s
p
ar
a
llel c
o
n
v
o
lu
tio
n
s
with
m
u
ltip
l
e
k
er
n
el
s
izes f
o
r
m
u
lti
-
s
ca
le
f
ea
tu
r
e
ex
tr
ac
tio
n
.
I
t
o
p
tim
izes
d
ep
th
an
d
co
m
p
u
t
atio
n
al
co
s
t,
ac
h
iev
in
g
h
ig
h
class
if
icatio
n
ac
cu
r
ac
y
an
d
ef
f
icien
cy
,
s
ettin
g
b
en
ch
m
ar
k
s
in
co
m
p
u
ter
v
is
io
n
[
2
4
]
.
I
n
th
is
wo
r
k
,
th
e
r
e
f
in
ed
I
n
ce
p
tio
n
V3
ar
ch
itectu
r
e
p
r
o
ce
s
s
es
co
lo
r
im
ag
es
r
esized
to
a
h
ig
h
er
s
q
u
ar
e
r
eso
lu
tio
n
s
u
itab
le
f
o
r
its
in
p
u
t
lay
er
.
T
h
e
d
esig
n
b
eg
in
s
with
an
in
itial
co
n
v
o
l
u
tio
n
s
tag
e
a
n
d
th
e
n
a
p
p
lies
a
s
eq
u
en
ce
o
f
i
n
ce
p
tio
n
u
n
its
th
at
p
r
o
ce
s
s
in
f
o
r
m
ati
o
n
in
p
ar
allel
u
s
in
g
f
ilter
s
o
f
d
if
f
er
en
t
s
p
atial
ex
te
n
ts
,
co
m
b
in
ed
with
p
o
o
lin
g
o
p
er
atio
n
s
to
lear
n
f
ea
tu
r
es
at
m
u
ltip
le
s
ca
les.
T
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Un
ified
vo
tin
g
-
b
a
s
ed
en
s
emb
l
e
lea
r
n
in
g
fo
r
r
ice
lea
f d
is
ea
s
e
d
etec
tio
n
…
(
Go
vi
n
d
a
r
a
ja
n
S
u
b
b
u
r
a
ma
n
)
1653
f
ir
s
t
in
ce
p
tio
n
m
o
d
u
le
in
clu
d
e
s
co
n
v
o
lu
tio
n
s
with
6
4
,
1
2
8
,
a
n
d
3
2
f
ilter
s
,
f
o
llo
wed
b
y
m
ax
-
p
o
o
lin
g
,
wh
ile
th
e
s
ec
o
n
d
m
o
d
u
le
ap
p
lies
co
n
v
o
lu
tio
n
s
with
1
2
8
,
1
9
2
,
an
d
9
6
f
ilter
s
b
ef
o
r
e
p
o
o
lin
g
.
Su
b
s
eq
u
en
t
in
ce
p
tio
n
m
o
d
u
les
ex
p
an
d
in
co
m
p
lex
ity
,
in
co
r
p
o
r
atin
g
wid
er
f
ilte
r
co
n
f
ig
u
r
atio
n
s
f
o
r
r
ich
e
r
f
ea
tu
r
e
ex
tr
ac
tio
n
.
T
o
r
e
d
u
ce
s
p
atial
d
im
en
s
io
n
s
,
two
r
e
d
u
ctio
n
m
o
d
u
les
ar
e
e
m
p
lo
y
ed
:
th
e
f
ir
s
t
ap
p
lies
d
o
wn
-
s
am
p
lin
g
u
s
in
g
3
×3
co
n
v
o
lu
tio
n
s
an
d
m
a
x
-
p
o
o
lin
g
af
ter
th
e
in
itial
in
ce
p
t
io
n
m
o
d
u
les,
wh
ile
th
e
s
ec
o
n
d
f
u
r
th
er
r
e
d
u
ce
s
d
im
en
s
io
n
s
at
d
ee
p
er
le
v
els
o
f
th
e
n
etwo
r
k
.
Af
ter
th
e
in
ce
p
tio
n
s
tag
es,
th
e
lear
n
ed
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
ar
e
p
ass
ed
to
a
h
ig
h
-
ca
p
ac
ity
d
e
n
s
e
s
tag
e
f
o
r
f
u
r
th
er
ab
s
tr
ac
t
io
n
,
f
o
llo
wed
b
y
a
r
e
g
u
lar
iz
atio
n
s
tep
to
lim
it
o
v
er
f
itti
n
g
.
T
h
e
m
o
d
el
th
en
p
r
o
d
u
ce
s
p
r
o
b
ab
ilit
y
-
b
ased
o
u
tp
u
ts
to
d
is
tin
g
u
is
h
am
o
n
g
th
e
d
if
f
er
en
t
leaf
co
n
d
itio
n
s
,
as illu
s
tr
ated
in
Fi
g
u
r
e
6
.
4
.
4
.
Xce
ptio
n
Xce
p
tio
n
,
also
k
n
o
wn
as
ex
t
r
em
e
in
ce
p
tio
n
,
is
an
ad
v
a
n
c
ed
v
is
io
n
m
o
d
el
d
ev
el
o
p
ed
t
o
ac
h
iev
e
s
tr
o
n
g
r
esu
lts
in
a
wid
e
r
an
g
e
o
f
v
is
u
al
r
ec
o
g
n
itio
n
ap
p
licatio
n
s
[
2
4
]
.
I
t
b
u
ild
s
u
p
o
n
th
e
I
n
ce
p
tio
n
V3
d
esig
n
b
y
ad
o
p
tin
g
a
r
e
f
in
ed
s
tr
u
ctu
r
e
th
at
s
ep
ar
ates
s
p
atial
an
d
ch
an
n
el
-
wis
e
p
r
o
ce
s
s
in
g
.
T
h
is
ap
p
r
o
ac
h
lead
s
t
o
lo
wer
co
m
p
u
tatio
n
al
d
em
a
n
d
s
an
d
im
p
r
o
v
e
d
ac
cu
r
ac
y
,
as illu
s
tr
ated
in
Fig
u
r
e
7
.
Fig
u
r
e
4
.
Ar
c
h
itectu
r
e
o
f
th
e
i
m
p
r
o
v
e
d
VGG1
6
m
o
d
el
Fig
u
r
e
5
.
Stru
ctu
r
e
o
f
th
e
en
h
a
n
ce
d
Den
s
eNe
t1
2
1
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
6
4
6
-
1
6
6
3
1654
Fig
u
r
e
6
.
Stru
ctu
r
e
o
f
th
e
en
h
a
n
ce
d
I
n
ce
p
tio
n
V3
m
o
d
el
Fig
u
r
e
7
.
Stru
ctu
r
e
o
f
th
e
en
h
a
n
ce
d
Xce
p
tio
n
m
o
d
el
T
h
e
im
p
r
o
v
ed
Xce
p
tio
n
m
o
d
el
ac
ce
p
ts
in
p
u
t
im
ag
es
o
f
s
ize
2
9
9
×2
9
9
×3
(
h
eig
h
t,
wid
th
,
an
d
R
GB
ch
an
n
els).
T
h
e
m
o
d
el
b
eg
i
n
s
with
two
in
itial
f
ea
tu
r
e
e
x
tr
ac
tio
n
s
tag
es,
wh
er
e
th
e
f
ir
s
t
r
ed
u
ce
s
s
p
atial
r
eso
lu
tio
n
u
s
in
g
a
s
m
aller
n
u
m
b
er
o
f
f
ilter
s
an
d
n
o
n
lin
e
ar
ac
tiv
atio
n
,
f
o
llo
wed
b
y
a
s
ec
o
n
d
s
tag
e
th
at
in
cr
ea
s
es
ch
an
n
el
d
e
p
th
wh
il
e
p
r
eser
v
in
g
f
in
e
r
s
p
atial
d
eta
il.
T
h
e
d
esig
n
is
o
r
g
a
n
ized
in
to
th
r
ee
s
u
cc
ess
iv
e
p
r
o
ce
s
s
in
g
p
h
ases
th
at
h
a
n
d
le
ea
r
ly
f
ea
tu
r
e
ex
tr
ac
tio
n
,
in
ter
m
ed
iate
r
ep
r
esen
tatio
n
lear
n
in
g
,
an
d
f
in
al
f
ea
t
u
r
e
r
ef
in
em
en
t.
T
h
e
ea
r
ly
f
ea
t
u
r
e
ex
tr
ac
tio
n
p
h
ase
is
m
ad
e
u
p
o
f
r
ep
ea
ted
co
m
p
u
tatio
n
al
b
lo
ck
s
th
at
s
ep
ar
ate
ch
an
n
el
-
wis
e
an
d
s
p
atial
f
ilter
in
g
,
with
s
h
o
r
tcu
t
co
n
n
ec
tio
n
s
in
clu
d
e
d
to
s
u
p
p
o
r
t
s
tab
le
a
n
d
ef
f
icien
t
lear
n
in
g
.
T
h
is
s
tag
e
co
n
s
is
ts
o
f
th
r
ee
m
ain
b
lo
ck
s
with
1
2
8
,
2
5
6
,
an
d
7
2
8
f
ilter
s
,
r
esp
ec
tiv
ely
.
T
h
e
im
m
ed
iat
e
r
ep
r
esen
tatio
n
lear
n
in
g
c
o
m
p
r
is
es
eig
h
t
id
e
n
tical
b
lo
c
k
s
,
ea
ch
c
o
n
tain
in
g
th
r
ee
d
ep
th
wis
e
s
ep
ar
ab
le
co
n
v
o
l
u
tio
n
s
with
7
2
8
f
ilter
s
an
d
R
eL
U
ac
tiv
atio
n
s
,
wi
th
o
u
t
an
y
d
o
w
n
-
s
am
p
lin
g
.
I
n
th
e
f
in
al
f
ea
tu
r
e
r
ef
in
em
en
t,
d
ep
th
wis
e
s
ep
ar
ab
le
co
n
v
o
lu
tio
n
s
with
1
,
0
2
4
f
ilter
s
ar
e
a
p
p
lied
,
ac
co
m
p
an
ied
b
y
r
esid
u
al
co
n
n
ec
tio
n
s
,
an
d
th
e
r
esu
ltin
g
f
ea
tu
r
e
m
ap
s
ar
e
th
en
r
ed
u
ce
d
to
1
×1
d
im
en
s
io
n
u
s
in
g
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
.
T
h
e
f
in
al
s
tag
es
em
p
l
o
y
a
d
en
s
e
r
ep
r
esen
tatio
n
lay
er
f
o
r
ad
v
an
ce
d
f
ea
tu
r
e
ab
s
tr
ac
tio
n
,
ap
p
ly
r
eg
u
lar
izatio
n
to
lim
it o
v
er
f
itti
n
g
,
a
n
d
p
r
o
d
u
ce
p
r
o
b
a
b
ilit
y
-
b
ased
o
u
tp
u
ts
to
d
i
s
tin
g
u
is
h
am
o
n
g
t
h
e
v
ar
i
o
u
s
r
i
ce
leaf
co
n
d
itio
n
s
.
4
.
5
.
E
ns
em
ble
wo
rk
f
lo
w
wit
h uni
f
ied v
o
t
ing
T
h
e
p
r
o
p
o
s
ed
en
s
em
b
le
wo
r
k
f
lo
w
is
illu
s
tr
ated
in
F
ig
u
r
e
8
.
I
n
th
is
f
r
am
ewo
r
k
,
p
r
e
p
r
o
ce
s
s
ed
r
ice
leaf
im
ag
es
ar
e
f
ir
s
t
f
ed
in
to
f
o
u
r
p
r
etr
ain
ed
d
ee
p
lear
n
in
g
m
o
d
els
VGG1
6
,
Den
s
eNe
t1
2
1
,
I
n
ce
p
tio
n
V3
,
an
d
Xce
p
tio
n
to
e
x
tr
ac
t
f
ea
tu
r
es
an
d
g
en
er
ate
in
itial
class
p
r
ed
ictio
n
s
.
T
h
e
p
r
ed
ictio
n
s
f
r
o
m
ea
ch
in
d
i
v
id
u
al
m
o
d
el
ar
e
s
u
b
s
eq
u
en
tly
c
o
m
b
in
ed
with
in
a
u
n
if
ied
v
o
tin
g
f
r
am
ewo
r
k
,
wh
ich
in
teg
r
at
es
th
r
ee
en
s
em
b
le
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Un
ified
vo
tin
g
-
b
a
s
ed
en
s
emb
l
e
lea
r
n
in
g
fo
r
r
ice
lea
f d
is
ea
s
e
d
etec
tio
n
…
(
Go
vi
n
d
a
r
a
ja
n
S
u
b
b
u
r
a
ma
n
)
1655
s
tr
ateg
ies:
m
ajo
r
ity
v
o
tin
g
,
s
o
f
t
v
o
tin
g
,
a
n
d
weig
h
ted
v
o
tin
g
.
I
n
th
e
m
ajo
r
ity
ap
p
r
o
ac
h
,
th
e
o
u
tp
u
t
co
r
r
esp
o
n
d
s
to
th
e
class
s
u
p
p
o
r
ted
b
y
th
e
h
i
g
h
est
n
u
m
b
er
o
f
m
o
d
els,
wh
ile
th
e
s
o
f
t
ap
p
r
o
ac
h
r
elies
o
n
th
e
av
er
ag
e
co
n
f
i
d
en
ce
s
co
r
es
f
r
o
m
all
m
o
d
els
to
d
eter
m
i
n
e
th
e
s
elec
ted
class
.
I
n
weig
h
ted
v
o
tin
g
,
m
o
d
els
ar
e
ass
ig
n
ed
p
er
f
o
r
m
an
ce
-
b
ased
weig
h
ts
,
an
d
th
e
f
in
al
p
r
ed
ict
io
n
is
d
e
r
iv
ed
f
r
o
m
th
e
weig
h
ted
co
m
b
in
atio
n
o
f
th
eir
o
u
tp
u
ts
,
allo
win
g
t
h
e
en
s
em
b
le
to
u
tili
ze
th
e
s
tr
en
g
th
s
o
f
ea
ch
m
o
d
el
f
o
r
m
o
r
e
r
eliab
le
an
d
ac
cu
r
ate
class
if
icatio
n
.
I
n
th
e
p
r
o
p
o
s
ed
weig
h
ted
v
o
ti
n
g
s
tr
ateg
y
,
th
e
weig
h
ts
[
0
.
3
,
0
.
3
,
0
.
2
,
0
.
2
]
wer
e
ass
ig
n
ed
t
o
th
e
f
o
u
r
p
r
etr
ain
ed
m
o
d
els
(
VGG1
6
,
Den
s
eNe
t1
2
1
,
I
n
ce
p
tio
n
V
3
,
an
d
Xce
p
tio
n
)
b
ased
o
n
th
eir
in
d
iv
id
u
al
p
er
f
o
r
m
an
ce
ac
c
u
r
ac
ies
an
d
co
m
p
lem
en
tar
y
s
tr
en
g
t
h
s
.
Den
s
eNe
t1
2
1
(
9
5
.
6
%)
an
d
I
n
ce
p
tio
n
V3
(
9
2
.
1
%)
d
em
o
n
s
tr
ated
r
elativ
ely
h
ig
h
er
an
d
m
o
r
e
s
tab
le
class
if
ic
atio
n
p
er
f
o
r
m
a
n
ce
,
an
d
th
er
e
f
o
r
e
wer
e
ass
ig
n
ed
s
lig
h
tly
h
ig
h
er
weig
h
ts
(
0
.
3
e
ac
h
)
.
VGG1
6
(
9
1
.
3
%)
an
d
Xce
p
tio
n
(
8
9
.
8
%)
ac
h
iev
ed
co
m
p
etitiv
e
b
u
t
s
lig
h
tly
lo
wer
ac
cu
r
ac
ies,
s
o
th
ey
wer
e
ass
ig
n
ed
s
m
aller
weig
h
ts
(
0
.
2
ea
ch
)
.
Fig
u
r
e
8
.
E
n
s
em
b
le
wo
r
k
f
lo
w
o
f
p
r
etr
ai
n
ed
m
o
d
els with
u
n
if
ied
v
o
tin
g
f
o
r
r
ice
leaf
d
is
ea
s
e
class
if
icatio
n
T
h
is
weig
h
tin
g
en
s
u
r
es
th
at
m
o
d
els
with
s
tr
o
n
g
e
r
p
r
e
d
i
ctiv
e
ca
p
ab
ilit
ies
h
av
e
a
p
r
o
p
o
r
tio
n
ally
g
r
ea
ter
in
f
lu
e
n
ce
o
n
th
e
f
in
al
d
ec
is
io
n
,
wh
ile
s
till
r
etain
in
g
th
e
d
iv
er
s
ity
b
en
e
f
its
o
f
in
clu
d
i
n
g
all
f
o
u
r
m
o
d
els.
E
n
s
em
b
le
lear
n
in
g
im
p
r
o
v
es
g
en
er
aliza
tio
n
b
y
ag
g
r
eg
ati
n
g
p
r
e
d
ictio
n
s
f
r
o
m
m
u
ltip
l
e
m
o
d
els,
th
er
eb
y
r
ed
u
cin
g
th
e
r
is
k
t
h
at
th
e
f
in
a
l
o
u
tp
u
t
is
o
v
er
ly
d
ep
e
n
d
en
t
o
n
th
e
b
iases
o
r
wea
k
n
ess
es
o
f
an
y
s
in
g
le
m
o
d
el
.
Ma
jo
r
ity
an
d
s
o
f
t
v
o
tin
g
alr
e
ad
y
p
r
o
v
i
d
e
r
o
b
u
s
tn
ess
b
y
b
a
lan
cin
g
p
r
ed
ictio
n
s
,
b
u
t
weig
h
ted
v
o
tin
g
f
u
r
th
e
r
en
h
an
ce
s
p
er
f
o
r
m
an
ce
b
y
em
p
h
asizin
g
s
tr
o
n
g
er
m
o
d
els.
T
h
is
r
ed
u
ce
s
th
e
lik
elih
o
o
d
o
f
o
v
er
f
itti
n
g
,
s
in
ce
th
e
en
s
em
b
le
ca
p
tu
r
es
a
b
r
o
ad
er
r
ep
r
esen
tatio
n
o
f
f
ea
tu
r
es
lear
n
ed
ac
r
o
s
s
ar
ch
itectu
r
es
an
d
av
o
id
s
r
ely
in
g
ex
ce
s
s
iv
ely
o
n
a
s
in
g
le
m
o
d
el
th
at
m
ay
h
av
e
o
v
e
r
f
it
to
th
e
tr
ain
in
g
s
et.
As
a
r
esu
lt,
th
e
u
n
if
ie
d
v
o
tin
g
en
s
em
b
le
ac
h
iev
es m
o
r
e
s
tab
le
an
d
ac
cu
r
ate
p
r
e
d
ictio
n
s
ac
r
o
s
s
d
iv
er
s
e
test
ca
s
es.
4
.
6
.
Unifie
d
v
o
t
ing
T
h
e
g
o
al
o
f
u
n
i
f
ied
v
o
ti
n
g
in
en
s
em
b
le
lear
n
in
g
is
to
c
o
m
b
i
n
e
o
u
tp
u
ts
f
r
o
m
s
ev
er
al
c
o
m
p
lem
en
tar
y
ar
ch
itectu
r
es,
allo
win
g
th
eir
in
d
iv
id
u
al
s
tr
en
g
th
s
to
co
llecti
v
ely
en
h
an
ce
o
v
er
all
class
if
icatio
n
ac
cu
r
ac
y
.
T
h
is
ap
p
r
o
ac
h
im
p
r
o
v
es
ac
cu
r
ac
y
b
y
r
ed
u
cin
g
t
h
e
lik
elih
o
o
d
o
f
er
r
o
r
s
m
a
d
e
b
y
in
d
iv
id
u
al
m
o
d
els
an
d
m
itig
ates
b
iases
in
h
er
en
t
in
ea
ch
m
o
d
el,
as
it
ag
g
r
e
g
ates
th
eir
p
r
ed
icti
o
n
s
to
c
o
u
n
ter
ac
t
th
e
wea
k
n
ess
es
o
f
o
n
e
with
th
e
s
tr
en
g
th
s
o
f
o
th
er
s
.
Un
if
ie
d
v
o
tin
g
en
h
an
ce
s
r
o
b
u
s
tn
ess
to
n
o
is
e
an
d
o
u
tlier
s
b
y
r
ed
u
cin
g
th
e
in
f
lu
en
ce
o
f
an
o
m
alies
th
r
o
u
g
h
th
e
co
m
b
i
n
ed
d
ec
is
io
n
s
o
f
all
m
o
d
els.
Fig
u
r
e
9
d
e
p
icts
th
e
th
r
ee
m
ain
ap
p
r
o
ac
h
es
f
o
r
co
m
b
in
in
g
m
o
d
el
p
r
ed
ictio
n
s
.
Fig
u
r
e
9
(
a)
d
em
o
n
s
tr
ates
th
e
m
ajo
r
ity
-
b
ased
s
tr
ateg
y
,
in
wh
ich
th
e
o
u
tc
o
m
e
co
r
r
esp
o
n
d
s
to
th
e
lab
el
m
o
s
t
f
r
eq
u
en
tly
s
u
p
p
o
r
ted
b
y
t
h
e
p
ar
ticip
atin
g
m
o
d
els.
Fig
u
r
e
9
(
b
)
illu
s
tr
ates
Evaluation Warning : The document was created with Spire.PDF for Python.