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
7
9
0
~
1
8
0
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
7
9
0
-
1
8
0
3
1790
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
A sequen
tial a
tt
e
ntion
-
enha
nc
ed d
eep learning
fram
ewo
rk f
o
r
ro
bust pota
to lea
f
disea
se dia
g
no
sis
under
real fi
eld c
o
ndition
s
Wa
t
cha
rk
o
rn
Yo
o
cho
m
bo
o
n
1
,
Nit
hizet
he
M
h
ua
dtho
ng
o
n
1
,
P
iy
a
po
rn
K
ra
cho
dn
o
k
2
1
D
i
g
i
t
a
l
Te
c
h
n
o
l
o
g
y
,
S
c
h
o
o
l
o
f
S
c
i
e
n
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
S
u
k
h
o
t
h
a
i
Th
a
m
mat
h
i
r
a
t
O
p
e
n
U
n
i
v
e
r
si
t
y
,
N
o
n
t
h
a
b
u
r
i
,
T
h
a
i
l
a
n
d
2
S
c
h
o
o
l
o
f
T
e
l
e
c
o
mm
u
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
I
n
st
i
t
u
t
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
S
u
r
a
n
a
r
e
e
U
n
i
v
e
r
s
i
t
y
o
f
T
e
c
h
n
o
l
o
g
y
,
N
a
k
h
o
n
R
a
t
c
h
a
s
i
ma
,
T
h
a
i
l
a
n
d
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
No
v
2
6
,
2
0
2
5
R
ev
is
ed
Feb
7
,
2
0
2
6
Acc
ep
ted
Ma
r
5
,
2
0
2
6
Dia
g
n
o
si
n
g
p
o
tato
lea
f
d
ise
a
se
s
fro
m
ima
g
e
s
c
o
ll
e
c
ted
i
n
re
a
l
-
li
fe
field
se
tt
in
g
s
is
c
h
a
ll
e
n
g
in
g
,
m
a
in
ly
b
e
c
a
u
se
o
f
u
n
e
v
e
n
li
g
h
ti
n
g
,
c
o
m
p
lex
b
a
c
k
g
ro
u
n
d
s,
a
n
d
d
ise
a
se
sy
m
p
to
m
s
th
a
t
a
re
o
fte
n
s
u
b
t
le
o
r
v
is
u
a
ll
y
in
c
o
n
siste
n
t.
In
t
h
is
stu
d
y
,
a
d
e
e
p
lea
rn
in
g
-
b
a
se
d
fra
m
e
wo
rk
wa
s
d
e
v
e
lo
p
e
d
to
su
p
p
o
rt
p
o
tato
lea
f
d
ise
a
se
d
i
a
g
n
o
sis,
with
p
a
rti
c
u
lar
a
tt
e
n
ti
o
n
g
i
v
e
n
t
o
imp
ro
v
i
n
g
g
e
n
e
ra
li
z
a
ti
o
n
a
n
d
i
n
terp
re
tatio
n
.
S
e
v
e
ra
l
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
(CNN
)
a
rc
h
it
e
c
tu
re
s
we
re
first
e
x
a
m
in
e
d
u
n
d
e
r
th
e
sa
m
e
e
x
p
e
rime
n
tal
c
o
n
d
it
io
n
s,
a
n
d
Re
sN
e
Xt
-
50
sh
o
we
d
t
h
e
m
o
st
sta
b
le
o
v
e
ra
ll
p
e
rfo
rm
a
n
c
e
.
Th
e
m
o
d
e
l
wa
s
th
e
n
e
x
ten
d
e
d
b
y
a
p
p
ly
in
g
e
fficie
n
t
c
h
a
n
n
e
l
a
tt
e
n
ti
o
n
(ECA),
fo
l
lo
we
d
b
y
sp
a
ti
a
l
a
tt
e
n
ti
o
n
a
d
a
p
ted
fro
m
th
e
c
o
n
v
o
lu
ti
o
n
a
l
b
lo
c
k
a
tt
e
n
ti
o
n
m
o
d
u
le
(CBAM)
.
Tes
t
re
su
l
ts
in
d
ica
te
th
a
t
t
h
is
se
q
u
e
n
ti
a
l
a
tt
e
n
ti
o
n
d
e
sig
n
p
e
rfo
rm
s
b
e
tt
e
r
th
a
n
th
e
b
a
se
li
n
e
m
o
d
e
l
a
s
we
ll
a
s
v
a
rian
ts
u
sin
g
o
n
ly
a
si
n
g
le
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
.
Ad
d
it
io
n
a
l
e
v
a
lu
a
ti
o
n
u
sin
g
3
0
0
re
a
l
-
field
ima
g
e
s
c
o
ll
e
c
ted
u
n
d
e
r
d
if
fe
re
n
t
fiel
d
c
o
n
d
i
ti
o
n
s
su
g
g
e
sts
im
p
ro
v
e
d
r
o
b
u
stn
e
ss
,
wh
il
e
v
is
u
a
li
z
a
ti
o
n
re
su
lt
s
fro
m
g
ra
d
ie
n
t
-
we
ig
h
ted
c
las
s
a
c
ti
v
a
ti
o
n
m
a
p
p
in
g
(G
ra
d
-
CAM)
sh
o
w
c
lea
re
r
fo
c
u
s
o
n
les
io
n
-
re
late
d
re
g
io
n
s.
Ov
e
ra
ll
,
th
e
fin
d
in
g
s
s
u
g
g
e
st
t
h
a
t
c
o
m
b
in
i
n
g
c
h
a
n
n
e
l
-
wise
a
n
d
sp
a
ti
a
l
a
tt
e
n
ti
o
n
c
a
n
imp
ro
v
e
b
o
th
p
re
d
icti
o
n
re
li
a
b
il
it
y
a
n
d
in
terp
re
tab
il
it
y
,
m
a
k
in
g
t
h
e
a
p
p
r
o
a
c
h
su
it
a
b
le f
o
r
p
ra
c
ti
c
a
l
a
g
ric
u
lt
u
ra
l
u
se
.
K
ey
w
o
r
d
s
:
Atten
tio
n
m
ec
h
an
is
m
Dee
p
lear
n
in
g
E
f
f
icien
t c
h
an
n
el
atten
tio
n
C
o
n
v
o
lu
tio
n
al
b
lo
ck
atten
tio
n
m
o
d
u
le
Po
tato
leaf
d
is
ea
s
e
R
esNeX
t
-
50
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
:
Nith
izeth
e
Mh
u
ad
th
o
n
g
o
n
D
i
g
i
t
al
T
e
c
h
n
o
l
o
g
y
,
S
c
h
o
o
l
o
f
S
c
i
e
n
c
e
a
n
d
T
ec
h
n
o
l
o
g
y
,
S
u
k
h
o
t
h
a
i
T
h
a
m
m
a
t
h
i
r
at
O
p
e
n
U
n
iv
e
r
s
i
t
y
9
/
9
M
u
9
,
C
h
a
e
n
g
w
a
tt
a
n
a
R
d
.
B
a
n
g
p
o
o
d
,
P
a
k
k
r
e
t
N
o
n
t
h
a
b
u
r
i,
T
h
a
i
l
a
n
d
E
m
ail: N
ith
izeth
e.
Mh
u
@
s
to
u
.
ac
.
th
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
p
o
tato
(
S
o
la
n
u
m
tu
b
ero
s
u
m
L
.
)
is
th
e
w
o
r
ld
’
s
m
o
s
t
im
p
o
r
tan
t
f
o
o
d
cr
o
p
af
ter
wh
e
at,
r
ice,
an
d
m
aize
.
T
h
e
p
o
tato
is
k
n
o
wn
w
o
r
ld
wid
e
f
o
r
its
n
u
tr
itio
n
al
v
al
u
e
[
1
]
.
I
t
is
a
v
er
s
atile
cr
o
p
wi
th
m
an
y
in
d
u
s
tr
ial
u
s
es.
Mo
r
eo
v
er
,
it
p
lay
s
an
i
m
p
o
r
tan
t
r
o
le
in
g
lo
b
al
f
o
o
d
s
ec
u
r
ity
.
I
t
is
p
ar
ticu
lar
ly
im
p
o
r
tan
t
f
o
r
d
e
v
elo
p
in
g
ec
o
n
o
m
ies.
T
h
e
co
n
s
u
m
p
tio
n
o
f
p
r
o
ce
s
s
ed
p
o
tato
p
r
o
d
u
cts
i
s
in
cr
ea
s
in
g
r
ap
i
d
ly
[
2
]
.
I
n
T
h
ailan
d
,
f
a
r
m
er
s
in
th
e
n
o
r
th
er
n
an
d
n
o
r
th
ea
s
ter
n
r
eg
io
n
s
co
n
tr
ib
u
te
to
th
e
n
atio
n
al
p
r
o
d
u
ctio
n
o
f
o
v
er
1
2
0
,
0
0
0
to
n
s
o
f
p
o
tato
es
ev
er
y
y
ea
r
ac
r
o
s
s
m
o
r
e
t
h
an
4
2
,
0
0
0
r
ai
o
f
f
a
r
m
lan
d
.
Alth
o
u
g
h
a
h
ig
h
ly
n
u
tr
itio
u
s
an
d
ec
o
n
o
m
ically
im
p
o
r
tan
t
cr
o
p
,
p
o
tato
p
r
o
d
u
ctiv
ity
is
u
n
d
er
co
n
s
tan
t
th
r
ea
t
f
r
o
m
f
o
liar
d
is
ea
s
es,
n
am
ely
late
b
lig
h
t
an
d
ea
r
ly
b
lig
h
t
.
Acc
o
r
d
in
g
to
r
esear
c
h
co
n
d
u
cted
in
2
0
2
2
,
s
m
allh
o
ld
er
f
a
r
m
er
s
m
ay
l
o
s
e
u
p
t
o
7
0
%
o
f
th
eir
h
ar
v
est
to
P
h
yto
p
h
th
o
r
a
in
fest
a
n
s
an
d
A
lter
n
a
r
ia
s
o
la
n
i
,
d
r
i
v
en
b
y
f
a
v
o
r
ab
le
h
u
m
id
ity
an
d
tem
p
er
a
tu
r
e
co
n
d
itio
n
s
[
3
]
.
A
m
o
r
e
r
ec
en
t
s
tu
d
y
s
u
g
g
e
s
ts
th
at
m
an
ag
em
en
t
f
ac
to
r
s
r
elate
d
to
f
ield
p
r
ac
tices
a
n
d
in
teg
r
ated
p
est
m
an
ag
em
en
t
ca
n
lo
wer
th
e
s
e
v
er
ity
o
f
ea
r
ly
b
lig
h
t
an
d
im
p
r
o
v
e
d
is
ea
s
e
co
n
tr
o
l
ef
f
icien
c
y
[
4
]
,
[
5
]
.
I
n
a
n
id
e
al
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
A
s
eq
u
en
tia
l a
tten
tio
n
-
en
h
a
n
c
ed
d
ee
p
le
a
r
n
in
g
fr
a
mewo
r
k
fo
r
r
o
b
u
s
t
…
(
Wa
tch
a
r
ko
r
n
Yo
o
ch
o
mb
o
o
n
)
1791
s
ce
n
ar
io
,
th
e
d
etec
tio
n
o
f
p
o
t
ato
leaf
d
is
ea
s
es
wo
u
ld
b
e
ti
m
ely
,
ac
cu
r
ate,
an
d
clo
s
ely
alig
n
ed
with
r
ea
l
f
ield
co
n
d
itio
n
s
,
en
a
b
lin
g
f
ar
m
er
s
t
o
r
esp
o
n
d
b
ef
o
r
e
ir
r
e
v
er
s
ib
le
y
ield
lo
s
s
es o
cc
u
r
.
I
n
an
id
ea
l
ag
r
icu
ltu
r
al
ec
o
s
y
s
tem
,
p
r
o
d
u
ctiv
ity
wo
u
ld
b
e
co
n
s
is
ten
t
with
s
u
s
ta
in
ab
ilit
y
,
p
r
o
d
u
ci
n
g
h
ig
h
y
ield
s
with
o
u
t e
x
ce
s
s
iv
e
u
s
e
o
f
p
esti
cid
es.
I
n
r
ea
lity
,
f
r
eq
u
en
t d
is
ea
s
e
o
u
tb
r
ea
k
s
h
av
e
led
T
h
ai
f
ar
m
er
s
to
r
ely
h
ea
v
ily
o
n
c
h
em
ical
f
u
n
g
icid
es.
W
h
ile
th
ese
p
r
o
v
id
e
s
h
o
r
t
-
ter
m
p
r
o
tectio
n
,
th
ey
i
n
cr
ea
s
e
p
r
o
d
u
ctio
n
co
s
ts
,
d
eg
r
ad
e
s
o
il
an
d
wate
r
q
u
ality
,
an
d
r
aise
co
n
ce
r
n
s
o
v
er
f
o
o
d
s
af
ety
an
d
e
x
p
o
r
t
p
o
t
en
tial
[
6
]
.
T
h
ailan
d
p
r
o
d
u
ce
s
o
v
er
4
.
5
b
illi
o
n
-
b
a
h
t
wo
r
th
o
f
p
o
tato
es
ea
ch
y
ea
r
,
y
et
im
p
o
r
tin
g
c
o
u
n
tr
ies
s
u
ch
as
J
ap
an
,
So
u
th
Ko
r
ea
,
an
d
Sin
g
ap
o
r
e
c
o
n
tin
u
e
to
tig
h
ten
r
esid
u
e
s
tan
d
ar
d
s
[
7
]
,
[
8
]
.
Alth
o
u
g
h
s
u
s
tain
ab
le
p
est
-
m
an
ag
em
e
n
t
p
r
ac
tices
ca
n
im
p
r
o
v
e
e
n
v
ir
o
n
m
en
tal
p
er
f
o
r
m
an
ce
[
9
]
,
[
1
0
]
th
eir
ad
o
p
tio
n
in
T
h
ailan
d
r
e
m
ain
s
s
lo
w,
lar
g
ely
d
u
e
to
lim
itatio
n
s
in
d
is
ea
s
e
d
iag
n
o
s
is
an
d
tech
n
o
lo
g
y
[
1
1
]
–
[
1
3
]
.
Dis
ea
s
e
id
en
tific
atio
n
in
r
u
r
al
ar
ea
s
is
s
til
l
b
ased
m
ain
ly
o
n
v
is
u
al
o
b
s
er
v
atio
n
,
wh
ich
is
o
f
te
n
in
co
n
s
is
t
en
t
an
d
tim
e
-
c
o
n
s
u
m
in
g
[
1
4
]
,
lead
in
g
f
ar
m
er
s
to
o
v
er
u
s
e
f
u
n
g
icid
es
as
a
p
r
ec
au
tio
n
.
T
h
is
s
itu
atio
n
r
ef
lects
a
p
r
ac
tical
b
o
ttlen
ec
k
,
as
th
e
lack
o
f
r
eliab
le,
f
ield
-
r
ea
d
y
d
iag
n
o
s
tic
m
eth
o
d
s
p
r
ev
en
t
s
u
s
tain
ab
ilit
y
m
ea
s
u
r
es f
r
o
m
b
ei
n
g
ef
f
ec
tiv
ely
ap
p
lied
o
n
f
a
r
m
s
.
Ag
r
o
-
b
ased
r
esear
ch
o
n
ar
tifi
cial
in
tellig
en
ce
an
d
d
ee
p
lea
r
n
in
g
h
as
witn
ess
ed
in
cr
ea
s
in
g
atten
tio
n
o
v
er
th
e
p
ast
d
ec
ad
e
f
o
r
em
p
lo
y
in
g
leaf
im
a
g
es
f
o
r
d
is
e
ase
d
iag
n
o
s
is
[
1
5
]
,
[
1
6
]
.
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN)
ar
e
also
ef
f
ec
tiv
e
in
class
if
y
in
g
m
an
y
ty
p
es
o
f
p
lan
t
d
is
ea
s
es,
o
f
ten
ac
h
iev
in
g
s
tr
o
n
g
ac
cu
r
ac
y
in
ex
p
e
r
im
en
tal
s
ettin
g
s
[
1
7
]
.
Ho
wev
er
,
m
u
c
h
o
f
th
is
d
ev
elo
p
m
en
t
h
as
b
ee
n
d
r
iv
en
b
y
m
o
d
els
tr
ain
ed
o
n
p
u
b
lic
d
atasets
a
n
d
ev
alu
ated
in
co
n
tr
o
lled
e
n
v
ir
o
n
m
en
ts
,
s
u
ch
as
Plan
t
Villag
e
o
r
Kag
g
le.
Alth
o
u
g
h
th
ese
d
atasets
ar
e
co
m
m
o
n
l
y
u
s
ed
f
o
r
b
en
c
h
m
ar
k
in
g
im
ag
e
-
b
ased
m
o
d
els,
th
ey
d
o
n
o
t
ca
p
t
u
r
e
th
e
v
ar
iab
ilit
y
an
d
co
m
p
lex
ity
s
e
en
in
f
ield
im
ag
es
[
1
8
]
,
[
1
9
]
.
C
o
n
s
eq
u
en
tly
,
m
o
d
els
th
at
d
is
p
lay
r
o
b
u
s
tn
ess
in
lab
o
r
ato
r
y
s
ettin
g
s
ten
d
to
p
er
f
o
r
m
p
o
o
r
ly
w
h
en
f
ac
e
d
with
t
h
e
v
is
u
al
u
n
ce
r
tain
ty
o
f
r
ea
l f
a
r
m
en
v
ir
o
n
m
e
n
ts
.
W
h
en
m
o
d
els
tr
ain
ed
o
n
co
n
t
r
o
lled
d
atasets
ar
e
d
e
p
lo
y
ed
in
r
ea
l
f
ield
s
,
th
e
s
tr
u
ctu
r
al
li
m
itatio
n
s
o
f
th
e
d
ata
b
ec
o
m
e
o
b
v
io
u
s
q
u
ick
ly
.
T
h
e
clar
ity
o
f
f
ield
i
m
ag
es
is
im
p
air
ed
d
u
e
to
v
a
r
iatio
n
s
in
lig
h
tin
g
,
b
ac
k
g
r
o
u
n
d
clu
tter
,
p
a
r
tial
leaf
o
cc
lu
s
io
n
,
an
d
d
is
ea
s
e
s
y
m
p
to
m
s
.
T
h
ese
s
y
m
p
to
m
s
d
o
n
o
t
co
n
f
o
r
m
to
an
y
s
p
ec
if
ic
p
atter
n
.
Fu
r
th
er
m
o
r
e,
th
ey
d
o
n
o
t
s
tan
d
o
u
t
d
is
tin
c
tly
f
r
o
m
th
e
n
atu
r
al
tex
tu
r
es
f
o
u
n
d
o
n
th
e
leav
es
th
em
s
elv
es.
T
h
e
ch
a
r
ac
ter
is
tics
o
f
th
ese
im
ag
es
a
r
e
f
u
n
d
am
e
n
tally
d
if
f
er
en
t
f
r
o
m
t
h
o
s
e
p
r
e
s
en
t
in
th
e
m
o
d
el'
s
tr
ain
in
g
im
ag
es.
As
a
r
esu
lt,
th
e
in
ter
n
al
p
r
io
r
itizatio
n
o
f
f
ea
tu
r
es
is
d
is
r
u
p
ted
.
T
h
e
d
r
o
p
in
class
if
icatio
n
ac
cu
r
ac
y
is
n
o
t
th
e
o
n
l
y
is
s
u
e.
I
t
also
r
ef
lects
th
e
d
eg
r
ad
atio
n
o
f
g
en
e
r
aliza
tio
n
ab
ilit
y
wh
en
f
ac
ed
with
u
n
s
ee
n
d
ata
[
2
0
]
–
[
2
2
]
.
T
h
is
in
s
tab
ilit
y
s
ev
er
ely
er
o
d
es
u
s
er
co
n
f
id
e
n
ce
in
au
t
o
m
ated
d
iag
n
o
s
tic
s
y
s
tem
s
,
p
ar
ticu
lar
ly
wh
e
n
d
ec
is
io
n
s
ar
e
m
ad
e
u
n
d
er
tim
e
a
n
d
r
eso
u
r
ce
co
n
s
tr
ain
ts
.
T
o
o
v
e
r
co
m
e
th
ese
lim
itatio
n
s
,
r
esear
ch
er
s
in
co
m
p
u
ter
v
is
io
n
h
a
v
e
in
teg
r
ated
atten
tio
n
m
ec
h
an
is
m
s
in
to
d
ee
p
lear
n
in
g
n
etwo
r
k
s
.
Atten
tio
n
en
ab
les
m
o
d
els
to
l
ea
r
n
th
e
im
p
o
r
tan
ce
o
f
ch
an
n
els
as
well
a
s
th
e
s
p
atial
r
elev
an
ce
o
f
f
ea
tu
r
es.
T
h
is
ca
p
ab
ilit
y
h
elp
s
d
is
tin
g
u
is
h
in
f
o
r
m
ativ
e
f
ea
tu
r
es,
wh
ic
h
im
p
r
o
v
es
m
o
d
e
l
p
er
f
o
r
m
an
ce
[
2
3
]
–
[
2
5
]
.
No
n
eth
eless
,
th
e
m
ajo
r
ity
o
f
ex
is
tin
g
wo
r
k
s
eith
er
u
tili
ze
a
s
in
g
le
atten
tio
n
m
ec
h
an
is
m
o
r
i
n
co
r
p
o
r
ate
a
tten
tio
n
m
o
d
u
les
in
to
s
tan
d
ar
d
C
NN
ar
ch
itectu
r
es
with
o
u
t
ex
p
l
o
r
in
g
t
h
e
in
ter
ac
tio
n
b
etwe
en
d
if
f
e
r
en
t
atten
tio
n
m
ec
h
an
is
m
s
[
2
6
]
–
[
2
9
]
.
T
h
u
s
,
th
ese
m
eth
o
d
s
d
o
n
o
t
s
h
ed
s
ig
n
if
ican
t
lig
h
t o
n
h
o
w
atten
tio
n
a
f
f
ec
ts
f
ea
tu
r
e
s
elec
tio
n
b
eh
av
io
r
a
n
d
d
ec
is
io
n
co
n
s
is
ten
cy
u
n
d
er
r
ea
l
-
f
ield
co
n
d
itio
n
s
.
T
h
e
u
tili
za
tio
n
o
f
atten
tio
n
m
ec
h
an
is
m
s
is
n
o
w
p
r
o
v
in
g
in
cr
ea
s
in
g
ly
p
o
p
u
lar
in
d
ee
p
lear
n
in
g
r
esear
ch
.
Ho
wev
e
r
,
th
er
e
ar
e
n
o
t
m
a
n
y
s
tu
d
ies
th
at
p
r
o
v
i
d
e
a
s
y
s
tem
atic
in
v
esti
g
atio
n
o
f
th
e
im
p
ac
t
o
f
atten
tio
n
m
o
d
u
les
with
d
if
f
er
en
t
s
tr
u
ctu
r
al
p
r
o
p
e
r
ties
.
T
h
is
lim
itatio
n
is
esp
ec
ially
o
b
s
er
v
ed
in
ar
ch
itectu
r
es
with
a
h
ig
h
er
d
eg
r
ee
o
f
s
tr
u
c
tu
r
al
co
m
p
lex
ity
,
s
u
ch
as
R
e
s
NeX
t
-
50
[
3
0
]
,
wh
ic
h
e
n
h
an
c
es
r
ep
r
esen
tatio
n
al
ca
p
ac
ity
th
r
o
u
g
h
ca
r
d
i
n
ality
a
n
d
p
a
r
allel
tr
an
s
f
o
r
m
atio
n
s
[
3
1
]
.
Alth
o
u
g
h
R
esNeXt
-
50
is
ca
p
ab
le
o
f
lear
n
in
g
co
m
p
lex
f
ea
tu
r
es
f
r
o
m
v
ar
iab
le
d
ata,
th
e
e
x
is
tin
g
liter
atu
r
e
d
o
es
n
o
t
p
r
o
v
id
e
e
m
p
ir
ical
a
n
aly
s
es
o
f
h
o
w,
in
p
r
ac
tice,
th
r
ee
-
d
im
en
s
io
n
al
(
c
h
an
n
el
-
b
ased
)
an
d
two
-
d
im
en
s
io
n
al
(
s
p
atial
-
b
ased
)
atten
tio
n
m
o
d
u
les
o
p
e
r
ate
jo
in
tly
with
in
R
esNeX
t
-
50
.
T
h
e
ef
f
ec
tiv
e
co
m
b
i
n
atio
n
o
f
ef
f
icien
t
ch
an
n
el
atten
tio
n
(
E
C
A)
an
d
th
e
co
n
v
o
l
u
tio
n
al
b
lo
c
k
atten
tio
n
m
o
d
u
le
(
C
B
AM
)
with
in
R
esNeXt
-
50
h
as
n
o
t
y
et
b
ee
n
i
n
v
esti
g
ated
f
o
r
th
e
class
if
icatio
n
o
f
p
o
tato
lea
f
d
is
ea
s
es
u
s
in
g
r
ea
l
-
f
ield
im
ag
es
.
T
h
e
ab
s
en
ce
o
f
k
n
o
wled
g
e
r
eg
ar
d
in
g
h
o
w
d
ee
p
lear
n
in
g
m
o
d
els atten
d
to
s
p
atial
ev
id
en
ce
an
d
f
o
r
m
d
ec
is
io
n
s
in
v
is
u
ally
u
n
s
tab
le
ag
r
icu
ltu
r
al
s
ce
n
ar
io
s
h
elp
s
ex
p
lain
th
is
g
ap
.
T
h
is
s
tu
d
y
co
m
b
in
es
E
C
A
an
d
th
e
C
B
AM
in
th
e
R
esNeXt
-
50
m
o
d
el
f
o
r
th
e
class
if
icatio
n
o
f
p
o
tat
o
leaf
d
is
ea
s
es
u
s
in
g
r
ea
l
-
f
ield
i
m
ag
es.
T
h
e
e
f
f
icien
cy
o
f
t
h
e
m
o
d
el
is
ev
alu
ated
q
u
a
n
titativ
ely
,
an
d
th
e
s
p
atial
d
ec
is
io
n
b
eh
av
io
r
is
ex
am
in
e
d
u
s
in
g
g
r
a
d
ien
t
-
weig
h
te
d
class
ac
tiv
atio
n
m
ap
p
in
g
(
Gr
a
d
-
C
AM
)
.
T
h
e
aim
o
f
th
is
r
esear
ch
is
t
o
ascer
tain
wh
eth
er
a
jo
in
t
atten
tio
n
d
esig
n
im
p
r
o
v
es
f
ea
t
u
r
e
-
s
elec
tio
n
s
t
ab
ilit
y
,
clar
if
ies
th
e
s
p
atial
ev
id
en
ce
th
at
en
ter
s
th
e
d
ec
is
io
n
-
m
ak
i
n
g
p
r
o
ce
s
s
,
an
d
en
h
a
n
ce
s
th
e
d
ep
lo
y
ab
ilit
y
o
f
d
ee
p
lea
r
n
in
g
m
o
d
els
in
r
ea
l
ag
r
icu
ltu
r
al
a
p
p
licatio
n
s
.
T
h
is
r
esear
ch
wo
r
k
,
f
r
o
m
a
n
ac
a
d
em
ic
p
e
r
s
p
ec
t
iv
e,
o
f
f
er
s
in
s
ig
h
ts
in
to
th
e
u
n
d
er
s
tan
d
i
n
g
o
f
d
ec
is
io
n
b
eh
av
io
r
in
atten
tio
n
-
au
g
m
en
ted
d
ee
p
lear
n
in
g
m
o
d
el
s
.
Fro
m
a
p
r
ac
tical
p
er
s
p
ec
tiv
e,
it
s
u
p
p
o
r
ts
th
e
d
ev
elo
p
m
en
t
o
f
d
iag
n
o
s
tic
s
y
s
tem
s
th
at
b
etter
r
ef
lect
r
ea
l
-
wo
r
ld
f
a
r
m
in
g
co
n
s
tr
ain
ts
an
d
s
u
s
tain
ab
ilit
y
g
o
als.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
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
7
9
0
-
1
8
0
3
1792
2.
RE
S
E
ARCH
M
E
T
H
O
D
T
h
e
g
o
al
o
f
th
is
s
tu
d
y
was
to
cr
ea
te
an
au
to
m
ate
d
s
y
s
tem
f
o
r
ass
ess
in
g
p
o
tato
leaf
d
is
e
ases
u
s
in
g
d
ee
p
lear
n
in
g
,
f
o
c
u
s
in
g
o
n
p
er
f
o
r
m
an
ce
ev
al
u
atio
n
u
n
d
er
a
co
n
tr
o
lled
en
v
i
r
o
n
m
e
n
t
an
d
a
r
ea
l
-
f
ield
en
v
ir
o
n
m
en
t.
T
h
e
r
esear
ch
p
r
o
ce
s
s
was
co
n
d
u
cted
in
th
r
ee
m
ain
s
tag
es.
T
h
e
C
iR
A
C
O
R
E
p
latf
o
r
m
was
u
s
ed
to
s
elec
t
th
e
in
itial
b
aselin
e
m
o
d
el
in
th
e
f
ir
s
t
p
h
ase.
T
h
e
s
ec
o
n
d
p
h
ase
f
o
cu
s
ed
o
n
m
o
d
el
r
e
f
in
em
en
t
an
d
an
aly
s
is
th
r
o
u
g
h
th
e
in
co
r
p
o
r
atio
n
o
f
atten
tio
n
m
ec
h
an
is
m
s
o
n
th
e
G
o
o
g
le
C
o
lab
p
latf
o
r
m
.
T
h
e
f
in
al
p
h
ase
ev
alu
ated
th
e
p
r
ac
tical
ap
p
lic
ab
ilit
y
o
f
th
e
m
o
d
el
u
s
in
g
3
0
0
im
ag
es
f
r
o
m
a
r
ea
l
-
f
ield
i
m
ag
e
d
ataset.
T
h
e
r
esear
ch
d
esig
n
was
d
e
v
elo
p
ed
to
m
ee
t
th
e
o
b
jectiv
es
o
f
th
e
s
tu
d
y
b
y
ass
ess
in
g
m
o
d
el
s
tab
ilit
y
an
d
g
en
er
aliza
tio
n
a
b
ilit
y
u
n
d
e
r
r
e
alis
tic
ag
r
icu
ltu
r
al
co
n
d
itio
n
s
.
2
.
1
.
Da
t
a
co
llect
io
n
T
h
e
d
ataset
u
s
ed
in
th
is
wo
r
k
co
n
s
is
ted
o
f
a
co
n
tr
o
lled
d
ata
s
et
d
o
wn
lo
ad
ed
f
r
o
m
a
p
u
b
lic
s
o
u
r
ce
as
well
as
a
r
ea
l
-
f
iel
d
im
ag
e
d
ataset.
T
h
e
co
n
tr
o
lled
d
atase
t
u
s
ed
in
th
is
wo
r
k
was
s
o
u
r
ce
d
f
r
o
m
Kag
g
le,
co
m
p
r
is
in
g
a
to
tal
o
f
7
,
1
2
8
p
o
tato
leaf
im
ag
es
class
if
ied
in
to
th
r
ee
class
es,
n
am
ely
ea
r
ly
b
lig
h
t,
late
b
lig
h
t,
an
d
h
ea
lth
y
[
3
2
]
.
T
h
e
d
ataset
was
s
p
lit
in
to
8
0
:2
0
tr
ain
in
g
an
d
test
in
g
s
u
b
s
ets.
T
h
is
d
a
taset
was
s
elec
ted
b
ec
au
s
e
it
is
well
k
n
o
wn
in
th
e
p
lan
t
d
is
ea
s
e
class
if
icatio
n
co
m
m
u
n
ity
,
co
n
tain
s
h
ig
h
-
q
u
al
ity
im
ag
es,
an
d
h
as
r
eliab
le
an
d
clea
r
ly
an
n
o
tated
lab
els.
T
o
f
u
r
th
e
r
test
th
e
p
r
ac
tical
ap
p
licab
ilit
y
o
f
th
e
d
e
v
elo
p
e
d
m
o
d
els,
th
e
tr
ain
e
d
m
o
d
els
wer
e
also
ev
alu
ated
u
s
in
g
a
r
ea
l
-
f
ield
i
m
ag
e
d
ataset
co
n
s
is
tin
g
o
f
3
0
0
im
ag
es
p
r
o
v
id
ed
b
y
ex
p
e
r
ts
f
r
o
m
t
h
e
T
h
ai
Dep
ar
tm
en
t
o
f
Ag
r
icu
ltu
r
e
[
3
3
]
.
T
h
ese
im
ag
es
wer
e
co
llected
f
r
o
m
ac
t
u
al
p
o
tato
f
a
r
m
s
in
T
h
ailan
d
an
d
ex
h
ib
ited
s
u
b
s
tan
tial
v
ar
ia
b
ilit
y
in
lig
h
tin
g
c
o
n
d
itio
n
s
,
b
ac
k
g
r
o
u
n
d
c
o
m
p
lex
ity
,
an
d
leaf
d
ef
o
r
m
atio
n
.
T
h
e
h
i
g
h
v
ar
iab
ilit
y
o
f
th
is
d
a
taset
m
ad
e
it
s
u
itab
le
f
o
r
ass
ess
in
g
th
e
m
o
d
els’
g
e
n
er
aliza
tio
n
ca
p
ab
ilit
y
u
n
d
er
r
ea
l
-
wo
r
ld
co
n
d
itio
n
s
.
T
a
b
le
1
s
u
m
m
ar
izes th
e
d
etailed
ch
a
r
ac
ter
is
tics
o
f
th
e
d
atasets
.
T
ab
le
1
.
Su
m
m
a
r
y
o
f
d
atasets
u
s
ed
in
th
is
s
tu
d
y
D
a
t
a
s
e
t
s
o
u
r
c
e
C
l
a
s
s
N
u
mb
e
r
o
f
i
m
a
g
e
s
P
u
r
p
o
se
P
u
b
l
i
c
d
a
t
a
se
t
(
K
a
g
g
l
e
)
Ea
r
l
y
b
l
i
g
h
t
2
,
4
2
4
Tr
a
i
n
i
n
g
/
t
e
st
i
n
g
La
t
e
b
l
i
g
h
t
2
,
4
2
4
Tr
a
i
n
i
n
g
/
t
e
st
i
n
g
H
e
a
l
t
h
y
2
,
2
8
0
Tr
a
i
n
i
n
g
/
t
e
st
i
n
g
F
i
e
l
d
d
a
t
a
s
e
t
(
D
O
A
,
Th
a
i
l
a
n
d
)
Ea
r
l
y
b
l
i
g
h
t
1
0
0
Ex
t
e
r
n
a
l
v
a
l
i
d
a
t
i
o
n
La
t
e
b
l
i
g
h
t
1
0
0
Ex
t
e
r
n
a
l
v
a
l
i
d
a
t
i
o
n
H
e
a
l
t
h
y
1
0
0
Ex
t
e
r
n
a
l
v
a
l
i
d
a
t
i
o
n
2
.
2
.
Da
t
a
prepa
ra
t
i
o
n
T
h
e
d
ataset
was
p
r
ep
r
o
ce
s
s
ed
b
ef
o
r
e
h
an
d
t
o
en
h
a
n
ce
im
ag
e
q
u
ality
an
d
co
n
s
is
ten
cy
b
e
f
o
r
e
m
o
d
el
tr
ain
in
g
.
E
ac
h
im
ag
e
was
ch
ec
k
ed
to
r
em
o
v
e
th
e
b
lu
r
r
ed
,
d
u
p
licated
,
o
r
in
co
r
r
ec
tly
lab
eled
s
am
p
les
s
o
th
at
th
e
n
o
is
e
in
th
e
s
am
p
le
co
u
l
d
b
e
r
em
o
v
ed
.
T
h
e
q
u
ality
-
co
n
tr
o
l
s
tep
m
ad
e
s
u
r
e
th
at
th
e
d
ataset
was
r
eliab
l
e
an
d
tr
u
s
two
r
th
y
f
o
r
d
e
v
elo
p
i
n
g
d
ee
p
lear
n
in
g
m
o
d
els.
T
h
e
im
p
o
r
tan
ce
o
f
p
r
e
p
r
o
ce
s
s
in
g
f
o
r
im
p
r
o
v
i
n
g
r
o
b
u
s
tn
ess
in
v
is
io
n
-
b
ased
r
e
co
g
n
itio
n
task
s
h
as
b
ee
n
wid
ely
r
ep
o
r
ted
.
I
n
p
r
ac
tical
s
y
s
tem
s
o
p
er
atin
g
with
r
ea
l
ca
p
tu
r
e
d
im
ag
es,
p
r
ep
r
o
c
ess
in
g
s
tep
s
s
u
ch
as
n
o
is
e
s
u
p
p
r
ess
io
n
,
co
n
tr
ast
en
h
a
n
ce
m
en
t,
an
d
s
tr
u
ctu
r
al
r
ef
in
em
en
t
h
a
v
e
b
ee
n
s
h
o
wn
to
s
ig
n
if
ican
tly
im
p
r
o
v
e
r
ec
o
g
n
itio
n
r
eliab
ilit
y
u
n
d
er
v
ar
iab
le
lig
h
tin
g
an
d
b
ac
k
g
r
o
u
n
d
co
n
d
itio
n
s
[
3
4
]
.
T
h
e
au
g
m
e
n
tatio
n
m
eth
o
d
s
in
cl
u
d
ed
th
e
f
o
llo
win
g
.
Ho
r
izo
n
tal
o
r
v
e
r
tical
in
clu
s
io
n
o
f
f
lip
p
e
d
im
ag
es
as
n
ew
s
am
p
les
in
th
e
d
ataset.
Ho
r
izo
n
tal
f
lip
p
in
g
is
a
lef
t
-
to
-
r
i
g
h
t
m
i
r
r
o
r
i
n
g
o
f
an
im
ag
e,
wh
ile
v
er
tical
f
lip
p
in
g
is
a
to
p
-
to
-
b
o
tto
m
r
e
f
lectio
n
o
f
th
e
im
ag
e
,
as
s
h
o
wn
in
Fig
u
r
e
1
.
T
o
tr
ain
th
e
m
o
d
el
to
r
ec
o
g
n
ize
d
is
ea
s
e
f
ea
tu
r
es
f
r
o
m
d
if
f
er
en
t
a
n
g
les,
im
ag
es
wer
e
r
o
tated
at
f
ix
ed
an
g
les
(
9
0
°,
1
8
0
°,
an
d
2
7
0
°)
an
d
also
r
an
d
o
m
an
g
les
b
etwe
en
3
0
°
an
d
+3
0
°,
as
s
h
o
wn
in
Fig
u
r
e
2
.
I
m
ag
e
b
r
ig
h
tn
ess
ad
ju
s
tm
en
t
was
p
er
f
o
r
m
e
d
to
h
e
lp
with
m
o
d
el
s
tab
ilit
y
an
d
cl
ea
r
er
d
is
cr
im
in
atio
n
o
f
th
e
ef
f
ec
ts
o
f
illu
m
in
atio
n
b
etwe
en
s
am
p
les.
T
h
e
b
r
i
g
h
tn
e
s
s
lev
els
o
f
th
e
im
ag
es
wer
e
c
h
an
g
ed
to
m
ak
e
th
e
im
ag
e
lig
h
ter
o
r
d
ar
k
er
,
as in
Fig
u
r
e
3
.
Fig
u
r
e
1
.
E
x
am
p
les o
f
d
ata
au
g
m
en
tatio
n
th
r
o
u
g
h
h
o
r
izo
n
tal
an
d
v
e
r
tical
im
ag
e
f
lip
p
i
n
g
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
A
s
eq
u
en
tia
l a
tten
tio
n
-
en
h
a
n
c
ed
d
ee
p
le
a
r
n
in
g
fr
a
mewo
r
k
fo
r
r
o
b
u
s
t
…
(
Wa
tch
a
r
ko
r
n
Yo
o
ch
o
mb
o
o
n
)
1793
Fig
u
r
e
2
.
E
x
am
p
les o
f
d
ata
au
g
m
en
tatio
n
th
r
o
u
g
h
f
i
x
ed
-
an
g
l
e
an
d
r
a
n
d
o
m
im
ag
e
r
o
tatio
n
Fig
u
r
e
3
.
E
x
am
p
les o
f
d
ata
au
g
m
en
tatio
n
th
r
o
u
g
h
b
r
ig
h
tn
ess
ad
ju
s
tm
en
t u
n
d
er
d
if
f
er
en
t lig
h
tin
g
co
n
d
itio
n
s
2
.
3
B
a
s
eline
m
o
del e
v
a
lua
t
io
n us
ing
CiRA
CO
RE
T
h
e
b
aselin
e
m
o
d
els
wer
e
d
ev
elo
p
ed
u
s
in
g
th
e
C
iR
A
C
OR
E
p
latf
o
r
m
f
o
r
tr
ain
in
g
an
d
ev
alu
atin
g
d
ee
p
lear
n
in
g
ar
ch
itectu
r
es
in
a
s
y
s
tem
atic
an
d
co
n
s
is
ten
t
m
an
n
er
d
u
e
to
its
in
teg
r
ated
en
v
ir
o
n
m
en
t.
Fiv
e
C
NN
ar
ch
itectu
r
es
wer
e
tr
ain
ed
an
d
co
m
p
ar
ed
as
b
aselin
e
ca
n
d
id
ates
in
th
is
s
tu
d
y
,
n
am
ely
Alex
Net,
Mo
b
ileNet
-
V2
,
E
f
f
icien
tNet
-
B
0
,
R
esNet
-
5
0
,
an
d
R
esNeXt
-
5
0
.
E
ac
h
m
o
d
el
was
d
ep
lo
y
ed
u
s
in
g
tr
an
s
f
er
lear
n
in
g
an
d
tr
ain
ed
u
n
d
er
th
e
s
am
e
tr
ain
in
g
s
ettin
g
s
.
T
h
is
s
tr
ateg
y
was
ad
o
p
ted
to
en
s
u
r
e
th
at
p
er
f
o
r
m
an
ce
v
ar
iatio
n
s
wer
e
d
u
e
to
ar
ch
itectu
r
al
d
if
f
er
en
ce
s
r
ath
er
th
an
tr
ain
in
g
co
n
d
itio
n
s
.
Fig
u
r
e
4
illu
s
tr
ates
th
e
wo
r
k
f
lo
w
o
f
th
e
b
aselin
e
m
o
d
el
s
elec
tio
n
p
r
o
ce
s
s
.
Fig
u
r
e
4
.
W
o
r
k
f
lo
w
o
f
b
aselin
e
C
NN
m
o
d
el
tr
ain
in
g
a
n
d
s
elec
tio
n
u
s
in
g
th
e
C
iR
A
C
OR
E
p
latf
o
r
m
2.
3
.1
.
ResNeXt
-
50
a
rc
hite
ct
u
re
R
esNeXt
-
5
0
is
a
d
ee
p
C
NN
b
u
ilt
o
n
th
e
p
r
in
cip
le
o
f
r
esid
u
al
lear
n
in
g
.
I
t
is
ch
ar
ac
ter
ized
b
y
m
u
lti
-
b
r
an
ch
tr
an
s
f
o
r
m
atio
n
s
g
o
v
er
n
ed
b
y
ca
r
d
in
ality
,
wh
ich
en
h
an
ce
r
ep
r
esen
tatio
n
al
ca
p
ac
ity
b
y
in
cr
ea
s
in
g
f
ea
tu
r
e
d
iv
er
s
ity
with
o
u
t
in
cr
ea
s
in
g
th
e
d
ep
th
o
r
wid
th
o
f
th
e
n
etwo
r
k
.
I
n
th
is
s
tu
d
y
,
th
e
in
p
u
t
im
ag
e
was
f
ir
s
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
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
7
9
0
-
1
8
0
3
1794
p
r
o
ce
s
s
ed
th
r
o
u
g
h
co
n
v
o
lu
tio
n
al
lay
er
s
,
f
o
llo
wed
b
y
b
atch
n
o
r
m
aliza
tio
n
an
d
R
eL
U
ac
tiv
atio
n
.
T
h
e
s
p
atial
d
im
en
s
io
n
s
wer
e
th
en
r
ed
u
ce
d
v
ia
m
ax
p
o
o
lin
g
b
ef
o
r
e
en
ter
in
g
th
e
d
ee
p
f
ea
tu
r
e
ex
tr
ac
tio
n
s
tag
es.
T
h
e
b
o
ttlen
ec
k
r
esid
u
al
b
lo
ck
s
wer
e
ar
r
an
g
ed
h
ier
ar
ch
ically
to
p
er
f
o
r
m
f
ea
tu
r
e
ex
tr
ac
tio
n
.
E
ac
h
b
lo
ck
co
n
s
is
ted
o
f
1
×1
,
3
×3
,
an
d
1
×1
co
n
v
o
lu
tio
n
al
lay
er
s
.
I
n
ad
d
itio
n
,
th
e
3
×3
co
n
v
o
lu
tio
n
em
p
lo
y
ed
g
r
o
u
p
ed
co
n
v
o
lu
tio
n
with
a
ca
r
d
in
ality
o
f
3
2
an
d
a
b
ase
wid
th
o
f
4
.
T
h
is
d
esig
n
en
ab
led
g
r
ea
ter
ch
an
n
el
-
wis
e
f
ea
tu
r
e
d
iv
er
s
ity
with
o
u
t
in
cr
ea
s
in
g
co
m
p
u
tatio
n
al
co
s
t.
Af
ter
f
ea
tu
r
e
ex
tr
ac
tio
n
,
th
e
r
esu
ltin
g
f
ea
tu
r
e
m
ap
s
wer
e
p
ass
ed
th
r
o
u
g
h
a
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
lay
er
an
d
a
f
u
lly
co
n
n
ec
ted
lay
er
with
a
So
f
tMa
x
f
u
n
ctio
n
to
class
if
y
im
ag
es
in
to
th
r
ee
ca
teg
o
r
ies:
ea
r
ly
b
lig
h
t,
late
b
lig
h
t,
an
d
h
ea
lth
y
.
T
h
e
o
v
er
all
s
tr
u
ctu
r
e
o
f
th
e
R
esNeXt
-
5
0
ar
ch
itectu
r
e
is
s
h
o
wn
in
Fig
u
r
e
5
.
R
esNeXt
-
5
0
was
s
elec
ted
as
th
e
b
ac
k
b
o
n
e
m
o
d
el
b
ec
au
s
e
it
p
r
o
v
id
es
s
em
an
tically
r
ich
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
an
d
s
er
v
es
as
a
s
u
itab
le
f
o
u
n
d
atio
n
f
o
r
in
teg
r
atin
g
atten
tio
n
m
ec
h
an
is
m
s
.
Fig
u
r
e
5
.
Ar
c
h
itectu
r
e
o
f
th
e
R
esNeX
t
-
50
b
ac
k
b
o
n
e
n
etwo
r
k
2
.
4
.
I
nte
g
ra
t
io
n o
f
a
t
t
ent
io
n
m
ec
ha
nis
m
s
into
t
he
m
o
del
a
rc
hite
ct
ure
T
h
e
d
esig
n
o
f
atten
tio
n
m
ec
h
an
is
m
s
d
r
aws
h
ea
v
ily
f
r
o
m
h
u
m
an
v
is
u
al
p
er
ce
p
tio
n
,
wh
ich
allo
ws
h
u
m
an
s
to
b
e
s
elec
tiv
e
in
r
ec
o
g
n
izin
g
o
b
jects
an
d
ac
tiv
ities
.
T
h
u
s
,
th
e
h
u
m
an
v
is
u
al
p
er
ce
p
tio
n
p
r
o
ce
s
s
ca
n
f
ilter
in
f
o
r
m
atio
n
an
d
f
o
cu
s
o
n
a
s
p
ec
if
ic
s
tim
u
lu
s
wh
ile
ig
n
o
r
in
g
o
th
er
s
.
T
h
er
ef
o
r
e,
atten
tio
n
m
ec
h
an
is
m
s
d
r
aw
in
s
p
ir
atio
n
f
r
o
m
th
e
ab
ilit
y
o
f
h
u
m
an
s
to
f
ilter
ir
r
elev
an
t
in
f
o
r
m
atio
n
to
p
er
f
o
r
m
r
ec
o
g
n
itio
n
task
s
.
Atten
tio
n
m
ec
h
an
is
m
s
h
av
e
en
ab
led
d
ee
p
lear
n
in
g
m
o
d
els
to
em
p
h
asize
u
s
ef
u
l
v
is
u
al
f
ea
tu
r
es
in
m
o
r
e
lo
ca
lized
r
eg
io
n
s
o
f
th
e
s
o
u
r
ce
im
ag
es.
I
n
th
e
f
ield
o
f
co
m
p
u
ter
v
is
io
n
r
esear
ch
,
atten
tio
n
m
ec
h
an
is
m
s
h
av
e
b
ee
n
wid
ely
in
teg
r
ated
with
C
NNs.
I
n
o
r
d
er
to
m
ak
e
ch
an
n
el
-
wis
e
an
d
s
p
atial
-
wis
e
f
ea
tu
r
es
m
o
r
e
d
is
cr
im
in
ativ
e,
f
ea
tu
r
e
weig
h
ts
ar
e
au
to
m
atica
lly
ad
ap
ted
.
T
h
is
s
tu
d
y
in
teg
r
ated
atten
tio
n
m
ec
h
an
is
m
s
in
to
th
e
R
esNeXt
-
5
0
m
o
d
el
to
im
p
r
o
v
e
th
e
m
o
d
el’
s
ca
p
ac
ity
to
lear
n
d
is
ea
s
e
p
atter
n
f
ea
tu
r
es
f
r
o
m
p
o
tato
leaf
im
ag
es.
T
h
e
d
esig
n
in
co
r
p
o
r
ated
atten
tio
n
in
b
o
th
th
e
ch
an
n
el
-
wis
e
an
d
s
p
atial
d
o
m
ain
s
with
in
th
e
b
ac
k
b
o
n
e
n
etwo
r
k
.
T
h
is
en
co
u
r
ag
es
th
e
m
o
d
el
to
f
o
cu
s
o
n
r
eg
io
n
s
co
n
tain
in
g
d
is
ea
s
e
s
y
m
p
to
m
s
,
th
er
eb
y
r
ed
u
cin
g
n
o
is
e
f
r
o
m
b
ac
k
g
r
o
u
n
d
clu
tter
an
d
en
v
ir
o
n
m
en
tal
v
ar
iatio
n
in
f
ield
im
ag
es.
T
h
is
tech
n
iq
u
e
was
esp
ec
ially
u
s
ef
u
l
f
o
r
d
if
f
er
en
tiatin
g
two
v
is
u
ally
s
im
ilar
class
es,
ea
r
ly
b
lig
h
t a
n
d
late
b
lig
h
t,
wh
ich
r
eq
u
ir
e
m
o
r
e
co
n
tex
t
-
s
p
ec
if
ic
f
ea
tu
r
e
s
elec
tio
n
.
2
.
4
.
1
.
I
nte
g
ra
t
io
n o
f
cha
nn
el
-
wis
e
a
t
t
ent
io
n us
ing
ef
f
icien
t
cha
nn
el
a
t
t
ent
io
n
T
o
im
p
r
o
v
e
ch
an
n
el
-
wis
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
,
E
C
A
[
2
4
]
was
in
teg
r
ated
in
to
th
e
R
esNeXt
-
50
ar
ch
itectu
r
e.
E
C
A
is
a
lig
h
tweig
h
t a
tten
tio
n
m
ec
h
an
is
m
th
at
d
ir
ec
tly
lear
n
s
in
ter
-
ch
an
n
el
r
elatio
n
s
h
ip
s
u
s
in
g
a
s
im
p
le
s
tr
u
ctu
r
e
with
a
lo
w
n
u
m
b
er
o
f
p
ar
am
eter
s
.
T
h
is
d
esig
n
allo
ws
th
e
m
o
d
el
to
f
lex
ib
ly
r
ewe
ig
h
t
ch
an
n
el
im
p
o
r
tan
ce
ac
co
r
d
in
g
to
th
e
class
if
icatio
n
task
wh
ile
in
cu
r
r
in
g
m
in
im
al
co
m
p
u
tatio
n
al
co
s
t.
W
ith
in
th
e
E
C
A
m
ec
h
an
is
m
,
a
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
o
p
er
atio
n
is
f
ir
s
t
ap
p
lied
to
th
e
f
ea
tu
r
e
m
ap
s
g
en
er
ated
b
y
th
e
b
ac
k
b
o
n
e
n
etwo
r
k
to
co
n
v
er
t
s
p
atial
in
f
o
r
m
atio
n
in
to
a
ch
an
n
el
-
wis
e
d
escr
ip
to
r
.
T
h
e
r
esu
ltin
g
v
ec
to
r
is
th
en
p
ass
ed
th
r
o
u
g
h
a
o
n
e
-
d
im
en
s
io
n
al
co
n
v
o
lu
tio
n
to
ca
p
tu
r
e
lo
ca
l
cr
o
s
s
-
ch
an
n
el
in
ter
ac
tio
n
s
d
ir
ec
tly
,
with
o
u
t
d
im
en
s
io
n
ality
r
ed
u
ctio
n
o
r
th
e
u
s
e
o
f
f
u
lly
co
n
n
ec
ted
lay
er
s
.
C
h
an
n
el
atten
tio
n
weig
h
ts
ar
e
th
en
g
en
er
ated
u
s
in
g
a
s
ig
m
o
id
ac
tiv
atio
n
f
u
n
ctio
n
an
d
ap
p
lied
to
th
e
o
r
ig
in
al
f
ea
tu
r
e
m
ap
s
th
r
o
u
g
h
elem
en
t
-
wis
e
m
u
ltip
licatio
n
.
T
h
is
ap
p
r
o
ac
h
p
r
eser
v
es
s
em
an
tic
in
f
o
r
m
atio
n
in
th
e
f
ea
tu
r
e
m
ap
s
an
d
av
o
id
s
in
f
o
r
m
atio
n
lo
s
s
th
at
m
ay
r
esu
lt
f
r
o
m
o
v
er
ly
co
m
p
lex
tr
an
s
f
o
r
m
atio
n
s
.
T
h
e
co
n
ce
p
tu
al
ar
ch
itectu
r
e
o
f
th
e
E
C
A
m
o
d
u
le
is
illu
s
tr
ated
in
Fig
u
r
e
6
.
Du
e
to
its
s
im
p
licity
an
d
lo
w
co
m
p
u
tatio
n
al
co
s
t,
E
C
A
was
s
elec
ted
f
o
r
ap
p
licatio
n
to
r
ea
l
-
f
ield
im
ag
e
d
ata
th
at
ar
e
s
tr
o
n
g
ly
af
f
ec
ted
b
y
en
v
ir
o
n
m
en
tal
v
ar
iatio
n
.
Pre
v
io
u
s
s
tu
d
ies
h
av
e
s
h
o
wn
th
at
E
C
A
ca
n
b
e
ef
f
ec
tiv
ely
in
teg
r
ated
in
to
C
NNs
to
en
h
an
ce
d
is
ea
s
e
-
r
elate
d
ch
an
n
el
f
ea
tu
r
e
r
ep
r
esen
tatio
n
.
T
h
ese
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
A
s
eq
u
en
tia
l a
tten
tio
n
-
en
h
a
n
c
ed
d
ee
p
le
a
r
n
in
g
fr
a
mewo
r
k
fo
r
r
o
b
u
s
t
…
(
Wa
tch
a
r
ko
r
n
Yo
o
ch
o
mb
o
o
n
)
1795
p
r
o
p
er
ties
m
ak
e
E
C
A
s
u
itab
le
f
o
r
s
tr
en
g
th
en
in
g
ch
an
n
el
-
wis
e
d
is
cr
im
in
atio
n
in
f
ield
-
b
ased
p
o
tato
leaf
d
is
ea
s
e
class
if
icatio
n
.
Fig
u
r
e
6
.
C
o
n
ce
p
tu
al
ar
ch
itect
u
r
e
o
f
t
h
e
E
C
A
m
o
d
u
le
2
.
4
.
2
.
I
nte
g
ra
t
io
n o
f
cha
nn
el
-
wis
e
a
nd
s
pa
t
ia
l a
t
t
ent
io
n us
ing
t
he
co
nv
o
lutio
na
l blo
ck
a
t
t
ent
io
n m
o
du
le
T
o
b
o
o
s
t
b
o
th
ch
an
n
el
-
wis
e
an
d
s
p
atial
f
ea
tu
r
e
r
ep
r
esen
tatio
n
,
th
e
C
B
AM
[
2
5
]
was
f
u
s
ed
with
th
e
R
esNeXt
-
5
0
ar
ch
itectu
r
e.
T
h
e
d
esig
n
o
f
C
B
AM
is
s
u
ch
th
at
ch
an
n
el
atten
tio
n
is
ap
p
lied
f
ir
s
t,
f
o
llo
wed
b
y
s
p
atial
atten
tio
n
,
to
r
ef
in
e
th
e
f
ea
tu
r
e
m
ap
s
s
eq
u
en
tially
.
T
h
e
u
s
e
o
f
th
is
s
eq
u
en
tial
atten
tio
n
s
tr
ateg
y
allo
wed
th
e
m
o
d
el
to
f
o
cu
s
o
n
in
f
o
r
m
ativ
e
f
ea
tu
r
es
an
d
d
is
r
eg
ar
d
th
e
b
ac
k
g
r
o
u
n
d
in
s
tead
o
f
g
iv
in
g
all
f
ea
tu
r
e
v
ec
to
r
s
eq
u
al
im
p
o
r
tan
ce
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
e
C
B
AM
u
s
ed
in
th
is
s
tu
d
y
f
o
llo
ws
th
e
o
r
ig
in
al
d
esig
n
,
as
s
h
o
wn
in
Fig
u
r
e
7
.
Fig
u
r
e
7
.
Ar
c
h
itectu
r
e
o
f
th
e
C
B
AM
T
h
e
m
o
d
u
le
in
th
e
ch
an
n
el
atten
tio
n
s
tag
e
is
r
esp
o
n
s
ib
le
f
o
r
id
en
tify
in
g
wh
ich
ch
an
n
els
in
th
e
f
ea
tu
r
e
m
ap
ar
e
m
o
s
t
im
p
o
r
tan
t
f
o
r
th
e
class
if
icatio
n
task
.
T
o
ca
p
tu
r
e
co
m
p
lem
en
tar
y
ch
an
n
el
-
wis
e
s
tatis
tics
,
f
ea
tu
r
e
m
ap
s
wer
e
g
lo
b
ally
av
er
ag
e
-
p
o
o
led
an
d
g
lo
b
ally
m
ax
-
p
o
o
led
alo
n
g
th
e
s
p
atial
d
im
en
s
io
n
.
T
h
e
r
esu
ltin
g
d
escr
ip
to
r
s
wer
e
p
ass
ed
th
r
o
u
g
h
a
s
h
ar
ed
m
u
ltil
ay
er
p
er
ce
p
tr
o
n
an
d
th
en
co
m
b
in
ed
,
f
o
llo
wed
b
y
a
s
ig
m
o
id
ac
tiv
atio
n
f
u
n
ctio
n
to
g
en
er
ate
th
e
ch
an
n
el
atten
tio
n
m
ap
.
T
h
e
r
esu
ltin
g
atten
tio
n
m
ap
was a
p
p
lied
to
th
e
b
ase
f
ea
tu
r
e
m
ap
th
r
o
u
g
h
elem
en
t
-
wis
e
m
u
ltip
licatio
n
,
th
er
eb
y
en
h
an
cin
g
ch
an
n
els
r
elev
an
t
to
d
is
ea
s
e
ch
ar
ac
ter
is
tics
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
e
ch
an
n
el
atten
tio
n
co
m
p
o
n
en
t is sh
o
wn
in
Fig
u
r
e
8
.
Fig
u
r
e
8
.
Ar
c
h
itectu
r
e
o
f
th
e
c
h
an
n
el
atten
tio
n
co
m
p
o
n
en
t i
n
th
e
C
B
AM
H
W
C
G
A
P
H
W
C
σ
A
d
ap
t
i
ve
s
e
l
e
c
t
i
on
of
k
e
r
n
e
l
s
i
z
e
1x1xC
1x1xC
I
n
p
u
t
F
e
a
t
u
r
e
C
h
an
n
e
l
A
t
t
e
n
t
i
on
M
od
u
l
e
S
p
at
i
al
A
t
t
e
n
t
i
on
M
od
u
l
e
R
e
f
i
n
e
d
F
e
a
t
u
r
e
C
o
n
v
o
l
u
ti
o
n
a
l
B
l
o
c
k
A
t
te
n
ti
o
n
M
o
d
u
l
e
I
n
p
u
t
F
e
a
t
u
r
e
F
M
ax
P
oo
l
i
n
g
A
vg
P
oo
l
i
n
g
S
h
a
r
e
d
M
L
P
C
h
a
n
n
e
l
A
t
t
e
n
t
i
o
n
M
+
c
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
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
7
9
0
-
1
8
0
3
1796
S
u
b
s
e
q
u
e
n
t
l
y
,
t
h
e
c
h
a
n
n
e
l
-
r
e
f
i
n
e
d
f
e
a
t
u
r
e
m
a
p
w
a
s
p
r
o
c
e
s
s
e
d
u
s
i
n
g
s
p
a
t
i
a
l
a
t
t
e
n
t
i
o
n
t
o
i
d
e
n
t
i
f
y
i
m
p
o
r
t
a
n
t
s
p
a
t
i
a
l
l
o
c
a
t
i
o
n
s
w
i
t
h
i
n
t
h
e
f
e
a
t
u
r
e
m
a
p
.
T
h
e
c
h
a
n
n
e
l
-
r
e
f
i
n
e
d
f
e
a
t
u
r
e
m
a
p
u
n
d
e
r
w
e
n
t
a
v
e
r
a
g
e
p
o
o
l
i
n
g
a
n
d
m
a
x
p
o
o
l
i
n
g
a
l
o
n
g
t
h
e
c
h
a
n
n
e
l
d
i
m
e
n
s
i
o
n
t
o
g
e
n
e
r
a
t
e
a
s
p
a
t
i
a
l
d
e
s
c
r
i
p
t
o
r
.
T
h
i
s
d
e
s
c
r
i
p
t
o
r
w
a
s
t
h
e
n
p
a
s
s
e
d
t
h
r
o
u
g
h
a
c
o
n
v
o
l
u
t
i
o
n
a
l
l
a
y
e
r
,
f
o
l
l
o
w
e
d
b
y
a
s
i
g
m
o
i
d
a
c
t
i
v
a
t
i
o
n
f
u
n
c
t
i
o
n
,
t
o
p
r
o
d
u
c
e
t
h
e
s
p
a
t
i
a
l
a
t
t
e
n
t
i
o
n
m
a
p
.
B
y
a
p
p
l
y
i
n
g
t
h
e
s
p
a
t
i
a
l
a
t
t
e
n
t
i
o
n
m
a
p
t
o
t
h
e
f
e
a
t
u
r
e
m
a
p
t
h
r
o
u
g
h
e
l
e
m
e
n
t
-
w
i
s
e
m
u
l
t
i
p
l
i
c
a
t
i
o
n
,
t
h
e
m
o
d
e
l
c
a
n
e
m
p
h
a
s
i
z
e
r
e
g
i
o
n
s
i
n
d
i
c
a
t
i
v
e
o
f
d
i
s
e
a
s
e
s
y
m
p
t
o
m
s
w
h
i
l
e
s
u
p
p
r
e
s
s
i
n
g
b
a
c
k
g
r
o
u
n
d
a
r
e
a
s
.
Fig
u
r
e
9
illu
s
tr
ates
th
e
s
tr
u
ctu
r
e
o
f
th
e
s
p
atial
atten
tio
n
co
m
p
o
n
en
t.
B
y
ap
p
ly
in
g
ch
an
n
el
atten
tio
n
f
ir
s
t
an
d
s
p
atial
atten
tio
n
s
ec
o
n
d
,
th
e
m
o
d
el
was
ab
le
to
lear
n
n
o
t
o
n
ly
wh
at
f
ea
tu
r
es
wer
e
im
p
o
r
tan
t,
b
u
t
also
wh
er
e
th
ey
wer
e
lo
ca
ted
.
T
h
is
d
esig
n
was
p
ar
ticu
lar
ly
well
-
s
u
ited
f
o
r
f
ield
im
ag
e
class
if
icatio
n
task
s
ch
ar
ac
ter
ized
b
y
co
m
p
lex
b
ac
k
g
r
o
u
n
d
s
an
d
ir
r
eg
u
lar
d
is
tr
ib
u
tio
n
s
o
f
d
is
ea
s
e
s
y
m
p
to
m
s
.
Fig
u
r
e
9
.
Ar
c
h
itectu
r
e
o
f
th
e
s
p
atial
atten
tio
n
co
m
p
o
n
en
t i
n
t
h
e
C
B
AM
2
.
4
.
3
.
S
e
q
u
e
n
t
i
a
l
i
n
t
e
g
r
a
t
i
o
n
o
f
e
f
f
i
c
i
e
n
t
c
h
a
n
n
e
l
a
t
t
e
n
t
i
o
n
a
n
d
s
p
a
t
i
a
l
a
t
t
e
n
t
i
o
n
T
o
en
h
an
ce
th
e
ca
p
ac
ity
o
f
th
e
m
o
d
el
to
lear
n
d
is
ea
s
e
-
r
elate
d
f
ea
tu
r
es
in
b
o
th
ch
an
n
el
-
wis
e
an
d
s
p
atial
d
im
en
s
io
n
s
,
a
s
eq
u
en
tial
in
teg
r
atio
n
o
f
E
C
A
an
d
s
p
atial
atten
tio
n
was
in
co
r
p
o
r
ated
in
to
th
e
R
esNeXt
-
5
0
ar
ch
itectu
r
e
in
th
is
s
tu
d
y
.
T
h
e
atten
tio
n
m
ec
h
an
is
m
s
wer
e
o
r
g
an
ized
f
r
o
m
ch
an
n
el
to
s
p
atial
atten
tio
n
with
in
th
e
b
ac
k
b
o
n
e
f
ea
tu
r
e
ex
tr
ac
tio
n
p
r
o
ce
s
s
.
T
h
is
d
esig
n
aim
ed
to
en
h
an
ce
f
ea
tu
r
e
im
p
o
r
tan
ce
h
ier
ar
ch
y
wh
ile
lim
itin
g
r
ed
u
n
d
an
cy
in
atten
tio
n
weig
h
tin
g
.
I
n
th
e
p
r
o
p
o
s
ed
s
tr
u
ctu
r
e,
th
e
E
C
A
m
ec
h
an
is
m
was
ap
p
lied
at
th
e
ea
r
ly
s
tag
e
o
f
f
ea
tu
r
e
ex
tr
ac
tio
n
in
th
e
R
esNeXt
-
5
0
b
ac
k
b
o
n
e
to
ca
lib
r
ate
ch
an
n
el
-
wis
e
f
ea
tu
r
e
r
esp
o
n
s
es.
B
y
m
o
d
elin
g
in
ter
-
ch
an
n
el
r
elatio
n
s
h
ip
s
d
ir
ec
tly
an
d
m
ain
tain
in
g
lo
w
co
m
p
u
tatio
n
al
co
m
p
lex
ity
,
E
C
A
allo
ws
th
e
m
o
d
el
to
s
elec
tiv
ely
em
p
h
asize
ch
an
n
els
ass
o
ciate
d
with
d
is
ea
s
e
ch
ar
ac
ter
is
tics
.
W
ith
o
u
t
in
tr
o
d
u
cin
g
s
ig
n
if
ican
t
ar
ch
itectu
r
al
o
r
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
.
T
h
e
f
e
a
t
u
r
e
m
a
p
s
p
r
o
d
u
c
e
d
b
y
E
C
A
w
e
r
e
t
h
e
n
f
o
r
w
a
r
d
e
d
t
o
t
h
e
s
p
a
t
i
a
l
a
t
t
e
n
t
i
o
n
c
o
m
p
o
n
e
n
t
o
f
C
B
A
M
t
o
i
d
e
n
t
i
f
y
s
p
a
t
i
a
l
l
o
c
a
t
i
o
n
s
t
h
a
t
w
e
r
e
m
o
s
t
r
e
l
e
v
a
n
t
f
o
r
d
i
s
e
a
s
e
c
l
a
s
s
i
f
i
c
a
t
i
o
n
.
T
o
a
v
o
i
d
r
e
d
u
n
d
a
n
c
y
,
o
n
l
y
t
h
e
s
p
a
t
i
a
l
a
t
t
e
n
t
i
o
n
c
o
m
p
o
n
e
n
t
w
a
s
e
m
p
l
o
y
e
d
a
t
t
h
i
s
s
t
a
g
e
,
a
s
c
h
a
n
n
e
l
-
w
i
s
e
w
e
i
g
h
t
i
n
g
h
a
d
a
l
r
e
a
d
y
b
e
e
n
p
e
r
f
o
r
m
e
d
b
y
E
C
A
.
T
h
r
o
u
g
h
s
p
a
t
i
a
l
r
e
f
i
n
e
m
e
n
t
,
t
h
e
m
o
d
e
l
w
a
s
a
b
l
e
t
o
f
o
c
u
s
o
n
l
e
s
i
o
n
r
e
g
i
o
n
s
w
h
i
l
e
m
i
n
i
m
i
z
i
n
g
t
h
e
i
n
f
l
u
e
n
c
e
o
f
b
a
c
k
g
r
o
u
n
d
n
o
i
s
e
a
n
d
e
n
v
i
r
o
n
m
e
n
t
a
l
v
a
r
i
a
t
i
o
n
p
r
e
s
e
n
t
i
n
r
e
a
l
-
f
i
e
l
d
i
m
a
g
e
s
.
Or
g
an
izin
g
atten
tio
n
m
ec
h
an
is
m
s
in
a
s
eq
u
en
tial
m
an
n
er
en
ab
led
th
e
m
o
d
el
to
r
eg
u
late
lear
n
in
g
o
f
b
o
th
wh
ich
f
ea
tu
r
es
wer
e
im
p
o
r
tan
t
an
d
wh
er
e
th
ey
wer
e
lo
ca
ted
.
T
h
is
d
esig
n
was
p
ar
ticu
lar
ly
s
u
itab
le
f
o
r
an
aly
zin
g
p
o
tato
leaf
im
ag
es
ca
p
tu
r
ed
u
n
d
er
h
ig
h
ly
v
ar
iab
le
an
d
co
m
p
lex
f
ield
co
n
d
itio
n
s
.
T
h
e
p
r
o
p
o
s
ed
E
C
A+
S
p
atial
-
C
B
AM
+Res
NeX
t
-
5
0
ar
ch
itectu
r
e
is
illu
s
tr
ated
in
Fig
u
r
e
1
0
.
2
.
5
.
I
m
ple
m
ent
a
t
io
n e
nv
iro
nm
ent
Du
e
to
th
e
s
tr
u
ctu
r
al
co
n
s
tr
ain
ts
im
p
o
s
ed
b
y
th
e
C
iR
A
C
OR
E
p
latf
o
r
m
th
at
r
estrict
d
ir
ec
t
m
o
d
if
icatio
n
s
o
f
in
ter
n
al
co
m
p
o
n
en
ts
with
in
d
ee
p
n
eu
r
al
n
etwo
r
k
ar
ch
itectu
r
es
(
i.e
.
,
s
p
ec
if
y
in
g
an
d
in
s
er
tin
g
atten
tio
n
m
ec
h
an
is
m
s
at
th
e
lay
er
lev
el)
,
m
o
d
el
d
ev
elo
p
m
en
t
an
d
tr
ain
in
g
wer
e
ca
r
r
ied
o
u
t
u
s
in
g
Py
th
o
n
with
th
e
Py
T
o
r
ch
f
r
am
ewo
r
k
.
T
h
e
r
esear
ch
er
s
o
p
ted
to
u
s
e
Py
T
o
r
ch
f
o
r
its
f
lex
ib
le
ar
ch
itectu
r
e
.
I
t
also
allo
ws
im
p
lem
en
tatio
n
o
f
atten
tio
n
-
r
elate
d
ad
ju
s
tm
en
ts
th
at
alig
n
with
th
e
g
o
als o
f
th
is
r
esear
ch
.
All
ex
p
er
im
en
ts
wer
e
co
n
d
u
cted
o
n
a
co
m
p
u
tin
g
s
y
s
tem
co
n
s
is
tin
g
o
f
an
AM
D
R
y
ze
n
7
7
8
4
0
HS
p
r
o
ce
s
s
o
r
,
1
6
GB
o
f
m
ain
m
em
o
r
y
,
an
d
an
NVI
DI
A
GeFo
r
ce
R
T
X
4
0
5
0
g
r
ap
h
ics
p
r
o
ce
s
s
in
g
u
n
it
(
GPU)
with
6
GB
o
f
d
ed
icate
d
m
em
o
r
y
.
T
h
e
ca
p
ac
ity
o
f
th
is
co
n
f
ig
u
r
atio
n
was
s
u
f
f
icien
t
to
f
ac
ilit
ate
m
o
d
el
tr
ain
in
g
,
ev
alu
atio
n
,
an
d
test
in
g
th
r
o
u
g
h
o
u
t
th
e
ex
p
er
im
en
t.
T
o
en
s
u
r
e
th
at
th
e
tr
ain
in
g
p
ip
elin
e
was
r
u
n
in
a
s
y
s
tem
atic
m
an
n
er
an
d
th
at
th
e
p
r
o
ject
was
r
ep
r
o
d
u
cib
le,
a
co
n
s
is
ten
t
ex
p
er
im
en
tal
en
v
ir
o
n
m
en
t
was
s
elec
ted
.
C
h
a
n
n
e
l
-
r
e
f
i
n
e
d
F
e
a
t
u
r
e
F'
Ma
x
P
o
o
l
,
A
v
g
P
o
o
l
C
o
n
v
l
a
y
e
r
S
p
a
ti
a
l
A
tte
n
ti
o
n
M
s
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
A
s
eq
u
en
tia
l a
tten
tio
n
-
en
h
a
n
c
ed
d
ee
p
le
a
r
n
in
g
fr
a
mewo
r
k
fo
r
r
o
b
u
s
t
…
(
Wa
tch
a
r
ko
r
n
Yo
o
ch
o
mb
o
o
n
)
1797
I
n
o
r
d
er
to
co
m
p
ar
e
th
e
m
o
d
els
in
an
o
b
jectiv
e
an
d
f
air
m
an
n
er
,
all
s
ettin
g
s
wer
e
co
n
tr
o
lled
.
All
m
o
d
els
u
tili
ze
d
th
e
s
am
e
R
esNeXt
-
5
0
ar
ch
itectu
r
e
as
th
eir
b
ac
k
b
o
n
e
an
d
u
s
ed
th
e
s
am
e
p
r
etr
ain
ed
weig
h
ts
,
d
atasets
,
d
ata
p
r
ep
ar
atio
n
m
eth
o
d
s
,
an
d
h
y
p
er
p
ar
am
eter
co
n
f
ig
u
r
atio
n
s
.
Mo
d
el
v
ar
iatio
n
s
ar
o
s
e
s
o
lely
d
u
e
to
d
if
f
er
en
ce
s
in
th
e
d
esig
n
an
d
in
teg
r
atio
n
o
f
atten
tio
n
m
ec
h
an
is
m
s
.
T
h
e
ex
p
er
im
en
t
was
co
n
tr
o
lled
to
cr
ea
te
a
s
ettin
g
th
at
co
u
ld
m
ea
s
u
r
e
th
e
ef
f
ec
t o
f
atten
tio
n
o
n
m
o
d
el
b
eh
av
io
r
an
d
p
er
f
o
r
m
an
ce
.
Fig
u
r
e
10
.
Ar
ch
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
R
esNeXt
-
50
with
s
eq
u
en
tial E
C
A
an
d
s
p
atial
atten
tio
n
in
teg
r
atio
n
2
.
6
.
M
o
del
ev
a
lua
t
i
o
n a
nd
re
a
l
-
f
ield t
esting
T
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
el
was
ev
alu
ated
u
s
in
g
s
tan
d
ar
d
m
etr
ics
ap
p
lied
in
im
ag
e
class
if
icatio
n
,
in
clu
d
in
g
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e.
C
o
n
f
u
s
io
n
m
atr
ix
an
aly
s
is
was
co
n
d
u
cted
to
p
r
o
v
id
e
d
ee
p
er
in
s
ig
h
t
in
to
class
if
icatio
n
b
eh
av
io
r
.
Acc
u
r
ac
y
was
u
s
ed
to
co
m
p
u
te
o
v
er
all
p
r
ed
ictiv
e
co
r
r
ec
tn
ess
,
wh
ile
th
e
r
em
ain
in
g
m
etr
ics
ev
alu
ated
th
e
m
o
d
el’
s
ab
ilit
y
to
d
is
tin
g
u
is
h
b
etwe
en
v
is
u
ally
s
im
ilar
class
es,
wh
ich
is
p
ar
ticu
lar
ly
im
p
o
r
tan
t f
o
r
p
o
tato
leaf
d
is
ea
s
e
class
if
icatio
n
.
Fo
llo
win
g
m
o
d
el
tr
ain
in
g
an
d
s
elec
tio
n
b
ased
o
n
p
er
f
o
r
m
an
ce
o
n
th
e
test
d
ataset,
th
e
R
esNeXt
-
50
m
o
d
el
with
in
teg
r
ated
atten
tio
n
m
ec
h
an
is
m
s
was
ev
alu
ated
o
n
an
ex
ter
n
al
d
ataset
co
n
s
is
tin
g
o
f
3
0
0
r
ea
l
-
f
ield
im
ag
es.
T
h
ese
im
ag
es
wer
e
n
o
t
u
s
ed
d
u
r
in
g
m
o
d
el
tr
ain
in
g
o
r
h
y
p
er
p
ar
am
eter
tu
n
in
g
.
T
h
e
f
ield
im
ag
es
ex
h
ib
ited
n
o
ticea
b
le
v
ar
iatio
n
in
b
r
ig
h
tn
ess
,
b
ac
k
g
r
o
u
n
d
co
m
p
lex
ity
,
leaf
d
ef
o
r
m
atio
n
,
an
d
d
is
ea
s
e
s
y
m
p
to
m
ap
p
ea
r
an
ce
.
T
h
e
g
o
al
o
f
th
is
ev
alu
atio
n
s
tag
e
was
to
ass
ess
th
e
m
o
d
el’
s
ab
ilit
y
to
g
en
er
alize
to
u
n
s
ee
n
d
ata
u
n
d
er
p
r
ac
tical
co
n
d
itio
n
s
.
T
o
en
h
an
ce
ev
alu
atio
n
cr
ed
ib
ilit
y
an
d
an
aly
ze
q
u
alitativ
e
m
o
d
el
d
ec
is
io
n
b
eh
av
io
r
,
Gr
ad
-
C
AM
was
em
p
lo
y
ed
.
Gr
ad
-
C
AM
v
is
u
alize
s
s
p
atial
ac
tiv
atio
n
m
ap
s
in
d
icatin
g
im
ag
e
r
eg
io
n
s
th
at
co
n
tr
ib
u
te
m
o
s
t
s
tr
o
n
g
ly
to
ea
ch
class
if
icatio
n
d
ec
is
io
n
.
T
h
is
an
aly
s
is
was
u
s
ed
to
ex
am
in
e
wh
eth
er
th
e
m
o
d
el
f
o
cu
s
ed
o
n
lesi
o
n
r
eg
io
n
s
co
n
s
is
ten
t w
ith
b
io
lo
g
ical
d
is
ea
s
e
ch
ar
ac
ter
is
tics
,
th
er
eb
y
s
u
p
p
o
r
tin
g
m
o
d
el
in
ter
p
r
etab
ilit
y
an
d
s
tr
en
g
th
en
in
g
co
n
f
id
en
ce
in
th
e
r
eliab
ilit
y
o
f
th
e
au
to
m
ated
d
iag
n
o
s
tic
s
y
s
tem
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
p
r
esen
ts
th
e
o
u
tco
m
es
o
f
th
e
p
er
f
o
r
m
an
ce
ev
alu
atio
n
o
f
th
e
d
ee
p
lear
n
in
g
m
o
d
els
cr
ea
ted
f
o
r
th
e
d
iag
n
o
s
is
o
f
p
o
tato
leaf
d
is
ea
s
e.
T
h
is
s
ec
tio
n
also
d
escr
ib
es
th
e
ex
p
er
im
en
tal
r
esu
lts
to
f
ac
ilit
ate
u
n
d
er
s
tan
d
in
g
o
f
m
o
d
el
b
eh
av
io
r
u
n
d
er
th
e
g
iv
en
ex
p
er
im
en
tal
s
ettin
g
s
.
All
ex
p
er
im
en
ts
wer
e
p
er
f
o
r
m
ed
u
n
d
er
th
e
s
am
e
s
ettin
g
s
to
co
n
tr
o
l
f
o
r
p
o
s
s
ib
le
co
n
f
o
u
n
d
in
g
ef
f
ec
ts
,
th
er
eb
y
allo
win
g
a
f
air
co
m
p
ar
is
o
n
o
f
m
o
d
els b
ased
o
n
g
en
u
in
e
ar
ch
itectu
r
al
d
if
f
er
en
ce
s
r
ath
er
th
an
ex
p
er
im
en
tal
co
n
f
o
u
n
d
in
g
.
224
×
224
×
3
I
m
a
g
e
G
l
o
b
a
l
A
v
e
r
a
g
e
P
o
o
l
i
n
g
F
u
l
l
y
C
o
n
n
e
c
t
e
d
L
a
y
e
r
S
o
f
t
m
a
x
C
h
a
n
n
e
l
-
r
e
f
i
n
e
d
F
e
a
t
u
r
e
F'
M
a
x
P
o
o
l
,
A
v
g
P
o
o
l
C
o
n
v
l
a
y
e
r
S
p
a
ti
a
l
A
tt
e
n
ti
o
n
M
s
G
A
P
σ
A
d
ap
t
i
ve
s
e
l
e
c
t
i
on
of
k
e
r
n
e
l
s
i
z
e
1x1xC
1x1xC
S
p
a
ti
a
l
-
C
B
A
M
M
o
d
u
l
e
EC
A
M
o
d
u
l
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
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
7
9
0
-
1
8
0
3
1798
3
.
1
.
B
a
s
eline
m
o
del c
o
m
pa
riso
n
B
aselin
e
m
o
d
el
co
m
p
ar
is
o
n
was
co
n
d
u
cted
u
n
d
er
id
en
tical
ex
p
er
im
en
tal
co
n
d
itio
n
s
to
en
s
u
r
e
f
air
an
d
u
n
b
iased
ev
alu
atio
n
.
T
o
m
itig
ate
th
e
im
p
ac
t
o
f
co
n
f
o
u
n
d
in
g
v
ar
iab
les
o
n
p
er
f
o
r
m
an
ce
o
u
tco
m
es,
all
m
o
d
els
wer
e
tr
ain
ed
an
d
test
ed
u
s
in
g
th
e
s
am
e
d
ataset,
d
ata
p
r
ep
ar
atio
n
p
r
o
ce
d
u
r
es,
an
d
h
y
p
er
p
ar
am
eter
s
.
I
n
itially
,
f
iv
e
C
NN
ar
ch
itectu
r
es
wer
e
tr
ain
ed
an
d
ev
alu
ated
as
b
aselin
e
m
o
d
els,
n
am
ely
Alex
Net,
Mo
b
ileNet
-
V2
,
E
f
f
icien
tNet
-
B
0
,
R
esNet
-
5
0
,
an
d
R
esNeXt
-
5
0
.
T
h
e
ass
ess
m
en
t
o
f
th
ese
m
o
d
els,
with
o
u
t
atten
tio
n
m
ec
h
an
is
m
s
,
s
er
v
ed
as
a
b
aselin
e
f
o
r
f
u
tu
r
e
p
er
f
o
r
m
an
ce
im
p
r
o
v
em
en
t.
Per
f
o
r
m
an
ce
ev
alu
atio
n
u
tili
ze
d
s
tan
d
ar
d
class
if
icatio
n
m
etr
ics,
in
clu
d
in
g
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
f
o
r
ea
ch
class
,
as
well
as
o
v
er
all
ac
cu
r
ac
y
.
Fro
m
th
e
p
er
f
o
r
m
an
ce
r
esu
lts
p
r
esen
ted
in
T
ab
le
2
,
R
esNeXt
-
5
0
ac
h
iev
ed
th
e
b
est
p
er
f
o
r
m
an
ce
with
an
o
v
er
all
ac
cu
r
ac
y
o
f
9
9
.
3
0
%,
wh
ile
R
esNet
-
5
0
ac
h
iev
ed
an
o
v
er
all
ac
cu
r
ac
y
o
f
9
9
.
1
6
%.
Am
o
n
g
th
e
ev
alu
ated
ar
ch
itectu
r
es,
Alex
Net
ex
h
ib
ited
th
e
lo
west
p
er
f
o
r
m
an
ce
.
T
h
ese
r
esu
lts
in
d
icate
th
at
th
e
en
h
an
ce
d
ca
p
ac
ity
o
f
th
e
R
esNeXt
-
5
0
ar
ch
itectu
r
e,
wh
ich
is
b
ased
o
n
g
r
o
u
p
ed
co
n
v
o
lu
tio
n
an
d
ag
g
r
eg
ated
r
esid
u
al
tr
an
s
f
o
r
m
atio
n
s
,
en
ab
led
m
o
r
e
ef
f
ec
tiv
e
lear
n
in
g
o
f
d
iv
er
s
e
an
d
d
is
cr
im
in
ativ
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
.
T
h
is
ar
ch
itectu
r
al
d
esig
n
p
r
o
v
ed
ad
v
an
tag
eo
u
s
f
o
r
ac
cu
r
ately
class
if
y
in
g
p
o
tato
leaf
d
is
ea
s
es
u
n
d
er
co
n
tr
o
lled
co
n
d
itio
n
s
.
B
ased
o
n
th
ese
f
in
d
in
g
s
,
R
esNeXt
-
5
0
was
s
elec
ted
as
th
e
b
ac
k
b
o
n
e
ar
ch
itectu
r
e
f
o
r
in
teg
r
atin
g
atten
tio
n
m
ec
h
an
is
m
s
an
d
f
u
r
th
er
ev
alu
atio
n
in
s
u
b
s
eq
u
en
t stag
es.
T
ab
le
2
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
o
f
th
e
f
iv
e
d
ee
p
lear
n
i
n
g
m
o
d
els o
n
th
e
test
s
et
M
o
d
e
l
E
a
r
l
y
b
l
i
g
h
t
L
a
t
e
b
l
i
g
h
t
H
e
a
l
t
h
y
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
A
l
e
x
N
e
t
9
3
.
6
0
9
9
.
5
9
9
6
.
4
0
9
9
.
5
3
8
7
.
0
1
9
2
.
8
8
9
3
.
8
3
100
9
6
.
8
1
9
5
.
4
4
M
o
b
i
l
e
N
e
t
V
2
1
0
0
9
1
.
7
5
9
5
.
7
0
9
2
.
3
8
100
9
6
.
0
3
1
0
0
1
0
0
1
0
0
9
7
.
2
0
Ef
f
i
c
i
e
n
t
N
e
t
B
0
1
0
0
9
8
.
3
5
9
9
.
1
6
9
7
.
1
9
100
9
8
.
5
7
9
9
.
7
7
9
7
.
1
5
9
8
.
4
4
9
8
.
9
4
R
e
sN
e
t
-
50
9
7
.
7
8
1
0
0
9
8
.
8
8
9
9
.
7
9
9
7
.
7
3
9
8
.
7
4
1
0
0
9
9
.
7
8
9
9
.
8
9
9
9
.
1
6
R
e
sN
e
X
t
-
50
9
9
.
7
9
9
9
.
1
8
9
9
.
5
2
9
9
.
1
7
9
8
.
7
6
9
9
.
0
0
9
8
.
9
2
1
0
0
9
9
.
4
6
9
9
.
3
0
3
.
2
.
E
x
t
er
na
l
v
a
lid
a
t
io
n us
ing
re
a
l
-
f
ield im
a
g
es
T
o
test
th
e
m
o
d
el’
s
ab
ilit
y
to
g
en
er
alize
to
m
o
r
e
r
ea
lis
tic
s
ce
n
ar
io
s
,
an
ex
ter
n
al
d
ataset
co
n
s
is
tin
g
o
f
3
0
0
r
ea
l
-
f
ield
im
ag
es
was
u
s
ed
f
o
r
ev
alu
atio
n
.
T
h
ese
im
ag
es
wer
e
n
o
t
in
co
r
p
o
r
ated
in
eith
er
tr
ain
in
g
o
r
m
o
d
el
tu
n
in
g
.
T
h
e
d
ataset
was
s
o
u
r
ce
d
f
r
o
m
r
ea
l
p
o
tato
cu
ltiv
atio
n
f
ield
s
an
d
ex
h
ib
ited
s
ig
n
if
ican
t
v
ar
iatio
n
s
in
lig
h
tin
g
,
b
ac
k
g
r
o
u
n
d
co
m
p
lex
ity
,
leaf
d
ef
o
r
m
atio
n
,
an
d
d
is
ea
s
e
s
y
m
p
to
m
ap
p
ea
r
an
ce
.
Fig
u
r
e
1
1
s
h
o
ws
r
ep
r
esen
tativ
e
s
am
p
le
im
ag
es
f
r
o
m
th
e
f
ield
,
wh
er
e
Fig
u
r
e
1
1
(
a)
illu
s
tr
ates
late
b
lig
h
t,
Fig
u
r
e
1
1
(
b
)
s
h
o
ws
ea
r
ly
b
lig
h
t,
an
d
Fig
u
r
e
1
1
(
c)
p
r
esen
ts
h
ea
lth
y
p
o
tato
leav
es.
(
a)
(
b
)
(
c)
Fig
u
r
e
11
.
R
ea
l
-
f
ield
p
o
tato
le
af
im
ag
es f
r
o
m
th
e
T
h
ai
De
p
a
r
tm
en
t
o
f
A
g
r
icu
ltu
r
e
f
o
r
(
a)
l
ate
b
lig
h
t
,
(
b
)
ea
r
ly
b
lig
h
t
,
a
n
d
(
c)
h
ea
lth
y
Af
ter
ch
o
o
s
in
g
R
esNeXt
-
5
0
as
th
e
b
ac
k
b
o
n
e
m
o
d
el
in
th
e
b
aselin
e
s
tu
d
y
,
th
e
im
p
ac
t
o
f
em
p
lo
y
in
g
d
if
f
er
en
t
k
in
d
s
o
f
atten
tio
n
m
ec
h
an
is
m
s
o
n
d
is
ea
s
e
class
if
icatio
n
p
er
f
o
r
m
an
ce
was
ex
am
in
ed
.
T
h
e
test
in
g
in
v
o
lv
ed
f
o
u
r
d
if
f
er
en
t
v
ar
ian
ts
o
f
th
e
m
o
d
el,
in
clu
d
in
g
th
e
b
aselin
e
R
esNeXt
-
5
0
,
E
C
A+
R
esNeXt
-
5
0
,
C
B
AM
+Res
NeX
t
-
5
0
,
an
d
E
C
A+
Sp
atial
-
C
B
AM
+Res
NeX
t
-
5
0
.
As
p
r
esen
ted
in
T
ab
le
3
,
q
u
an
titativ
e
p
er
f
o
r
m
an
ce
co
m
p
ar
is
o
n
s
ar
e
r
ep
o
r
ted
,
an
d
th
e
co
n
f
u
s
io
n
m
atr
ix
o
f
th
e
b
est
-
p
er
f
o
r
m
in
g
m
o
d
el
is
s
h
o
wn
in
Fig
u
r
e
1
2
.
Acc
o
r
d
in
g
to
T
ab
le
3
,
all
atten
tio
n
-
en
h
an
ce
d
m
o
d
els
co
n
s
is
ten
tly
o
u
tp
er
f
o
r
m
ed
th
e
b
aselin
e
R
esNeXt
-
5
0
.
T
h
e
E
C
A
m
o
d
u
le
en
h
an
ce
s
p
er
f
o
r
m
an
ce
b
y
s
tr
en
g
th
en
in
g
f
ea
tu
r
e
r
ep
r
esen
tatio
n
o
n
a
p
er
-
ch
an
n
el
b
asis
,
wh
er
ea
s
C
B
AM
d
o
es
th
is
b
y
jo
in
tly
m
o
d
elin
g
ch
an
n
el
an
d
s
p
atial
atten
tio
n
.
Of
all
th
e
m
o
d
els
s
tu
d
ied
in
th
is
p
ap
er
,
th
e
E
C
A+
S
p
atial
-
C
B
AM
+Res
NeX
t
-
5
0
m
o
d
el
ac
h
iev
ed
th
e
h
ig
h
est
ac
cu
r
ac
y
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
A
s
eq
u
en
tia
l a
tten
tio
n
-
en
h
a
n
c
ed
d
ee
p
le
a
r
n
in
g
fr
a
mewo
r
k
fo
r
r
o
b
u
s
t
…
(
Wa
tch
a
r
ko
r
n
Yo
o
ch
o
mb
o
o
n
)
1799
an
d
F1
-
s
co
r
e.
T
h
is
in
d
icate
s
th
at
th
e
m
o
r
e
ef
f
ec
tiv
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
ca
n
b
e
attr
ib
u
ted
to
th
e
co
m
p
lem
en
tar
y
r
o
les
o
f
atten
tio
n
m
ec
h
an
is
m
s
,
with
ch
an
n
el
atten
tio
n
s
u
p
p
o
r
tin
g
g
lo
b
al
f
ea
tu
r
e
s
elec
tio
n
an
d
s
p
atial
atten
tio
n
f
ac
ilit
atin
g
lesi
o
n
r
eg
io
n
lo
ca
lizatio
n
.
T
h
e
co
n
f
u
s
io
n
m
atr
ix
s
h
o
wn
in
Fig
u
r
e
12
in
d
icate
s
th
at
th
e
m
o
d
el
co
m
p
r
is
in
g
E
C
A+
Sp
atial
-
C
B
AM
+Res
NeX
t
-
5
0
ac
h
iev
es
ex
ce
llen
t
class
if
icatio
n
p
er
f
o
r
m
an
ce
ac
r
o
s
s
all
class
es with
h
ig
h
ac
cu
r
ac
y
.
T
h
e
m
ajo
r
ity
o
f
th
e
m
is
class
if
icatio
n
s
o
cc
u
r
r
ed
b
etwe
en
ea
r
ly
b
lig
h
t
an
d
late
b
lig
h
t,
wh
ich
s
h
ar
e
s
im
ilar
v
is
u
al
s
y
m
p
to
m
s
.
T
h
is
f
in
d
in
g
is
co
n
s
is
ten
t
with
an
ar
ch
itectu
r
al
an
aly
s
is
in
d
icatin
g
th
at
s
eq
u
en
tial
atten
tio
n
in
teg
r
atio
n
en
h
an
ce
d
th
e
m
o
d
el’
s
ab
ilit
y
to
d
is
cr
im
in
ate
b
etwe
en
o
v
er
lap
p
in
g
an
d
co
m
p
lex
d
is
ea
s
e
f
ea
tu
r
es in
ag
r
icu
ltu
r
al
en
v
ir
o
n
m
en
ts
.
T
ab
le
3
.
Qu
a
n
titativ
e
co
m
p
ar
i
s
o
n
o
f
atten
tio
n
-
en
h
a
n
ce
d
m
o
d
els
M
o
d
e
l
A
t
t
e
n
t
i
o
n
t
y
p
e
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
R
e
sN
e
X
t
-
50
N
o
n
e
9
6
.
0
0
9
6
.
0
7
9
6
.
0
0
9
5
.
9
8
EC
A
+
R
e
sN
e
X
t
-
50
C
h
a
n
n
e
l
9
7
.
3
3
9
7
.
3
2
9
7
.
3
3
9
7
.
3
2
C
B
A
M
+
R
e
sN
e
X
t
-
50
C
h
a
n
n
e
l
+
s
p
a
t
i
a
l
9
6
.
3
3
9
6
.
4
4
9
6
.
3
3
9
6
.
3
2
EC
A
+
S
p
a
t
i
a
l
-
C
B
A
M
+
R
e
sN
e
X
t
-
50
O
p
t
i
mi
z
e
d
h
y
b
r
i
d
9
8
.
6
7
9
8
.
6
8
9
8
.
6
7
9
8
.
6
7
Fig
u
r
e
12
.
C
o
n
f
u
s
io
n
m
atr
ix
o
f
th
e
E
C
A+
Sp
atial
-
C
B
A
M+
R
esNeXt
-
5
0
m
o
d
el
ev
al
u
ated
o
n
r
ea
l
-
f
ield
im
a
g
es
3.
3
.
B
eha
v
io
ra
l
a
na
ly
s
is
o
f
t
he
m
o
del us
ing
g
ra
d
-
CAM
I
n
th
is
s
tu
d
y
,
Gr
ad
-
C
AM
was
ap
p
lied
f
o
r
q
u
alitativ
e
in
ter
p
r
etatio
n
in
o
r
d
er
to
g
ain
in
s
ig
h
ts
in
to
th
e
in
ter
n
al
d
ec
is
io
n
-
m
ak
in
g
p
r
o
ce
s
s
es
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
els,
as
well
as
to
an
aly
ze
h
o
w
co
n
s
is
ten
t
m
o
d
el
p
r
ed
ictio
n
s
wer
e
with
lesi
o
n
lo
ca
tio
n
s
o
n
p
o
tato
leav
es.
T
h
e
m
ain
g
o
al
o
f
th
is
an
aly
s
is
was to
v
is
u
alize
s
p
atial
ar
ea
s
th
at
af
f
ec
t
class
if
icatio
n
an
d
to
ch
ec
k
wh
eth
er
th
e
m
o
d
els
ar
e
f
o
cu
s
in
g
o
n
b
io
lo
g
ically
r
elev
an
t
f
ea
tu
r
es
f
o
r
p
o
tato
leaf
d
is
ea
s
es.
An
aly
zin
g
m
o
d
el
in
ter
p
r
etab
ilit
y
an
d
ag
r
ee
m
en
t
f
o
r
au
to
m
ated
p
lan
t
d
is
ea
s
e
d
iag
n
o
s
is
is
v
er
y
im
p
o
r
tan
t.
T
h
e
Gr
ad
-
C
AM
ac
tiv
atio
n
m
ap
s
in
Fig
u
r
e
1
3
ex
h
ib
it th
e
b
eh
av
io
r
o
f
th
e
b
aselin
e
R
esNeXt
-
5
0
m
o
d
el
an
d
th
e
atten
tio
n
-
en
h
an
ce
d
m
o
d
els u
n
d
er
v
ar
io
u
s
co
n
f
ig
u
r
atio
n
s
.
T
h
e
b
a
s
e
l
i
n
e
R
e
s
N
e
X
t
-
5
0
m
o
d
e
l
e
x
h
i
b
i
t
e
d
r
e
l
a
t
i
v
e
l
y
b
l
u
r
r
e
d
a
c
t
i
v
a
t
i
o
n
,
w
i
t
h
n
o
t
a
b
l
e
r
e
s
p
o
n
s
e
s
i
n
b
a
c
k
g
r
o
u
n
d
r
e
g
i
o
n
s
o
u
t
s
i
d
e
t
h
e
l
e
s
i
o
n
a
r
e
a
s
.
T
h
e
b
a
s
e
l
i
n
e
m
o
d
e
l
a
p
p
e
a
r
s
t
o
h
a
v
e
r
e
l
i
e
d
,
i
n
p
a
r
t
,
o
n
v
i
s
u
a
l
c
u
e
s
u
n
r
e
l
a
t
e
d
t
o
t
h
e
d
i
s
e
a
s
e
d
u
r
i
n
g
c
l
a
s
s
i
f
i
c
a
t
i
o
n
.
I
n
c
o
r
p
o
r
a
t
i
n
g
a
s
i
n
g
l
e
a
t
t
e
n
t
i
o
n
m
e
c
h
a
n
i
s
m
,
s
u
c
h
a
s
E
C
A
o
r
t
h
e
C
B
A
M
,
c
o
n
c
e
n
t
r
a
t
e
d
t
h
e
a
c
t
i
v
a
t
i
o
n
m
a
p
s
m
o
r
e
o
n
l
e
s
i
o
n
a
r
e
a
s
.
H
o
w
e
v
e
r
,
s
o
m
e
b
a
c
k
g
r
o
u
n
d
r
e
g
i
o
n
s
s
t
i
l
l
r
e
t
a
i
n
e
d
a
c
t
i
v
a
t
i
o
n
s
,
i
n
d
i
c
a
t
i
n
g
t
h
a
t
i
r
r
e
l
e
v
a
n
t
f
e
a
t
u
r
e
s
w
e
r
e
n
o
t
a
l
w
a
y
s
c
o
m
p
l
e
t
e
l
y
s
u
p
p
r
e
s
s
e
d
.
I
n
co
n
tr
ast,
th
e
E
C
A+
Sp
atial
-
C
B
AM
+Res
NeX
t
-
5
0
m
o
d
el
p
r
o
d
u
ce
d
h
ig
h
ly
f
o
ca
l
an
d
well
-
lo
ca
lized
ac
tiv
atio
n
m
ap
s
th
at
clo
s
ely
m
atch
ed
lesi
o
n
r
eg
io
n
s
,
esp
ec
ially
in
ca
s
es
in
v
o
lv
in
g
s
m
all
o
r
ill
-
d
ef
in
ed
lesi
o
n
s
,
wh
ich
ar
e
co
m
m
o
n
ly
o
b
s
er
v
ed
in
b
o
th
ea
r
ly
b
lig
h
t
an
d
late
b
lig
h
t
u
n
d
er
f
ield
co
n
d
itio
n
s
.
T
h
e
r
esu
lts
r
ev
ea
l
th
at
th
e
m
o
d
el
ca
n
lear
n
to
f
o
cu
s
o
n
b
io
lo
g
ically
m
ea
n
in
g
f
u
l
r
eg
io
n
s
with
in
th
e
im
ag
e
b
y
s
ep
ar
atin
g
atten
tio
n
m
ec
h
an
is
m
s
.
E
C
A
o
p
er
ates
o
n
th
e
ch
an
n
el
d
im
en
s
io
n
to
s
elec
t
s
em
an
tic
f
ea
tu
r
es,
wh
ile
Sp
atial
-
C
B
AM
f
o
cu
s
es o
n
th
e
s
p
atial
d
im
en
s
io
n
f
o
r
lesi
o
n
lo
ca
lizatio
n
T
h
e
q
u
alitativ
e
r
esu
lts
o
b
tain
ed
f
r
o
m
th
e
Gr
ad
-
C
AM
an
aly
s
is
alig
n
with
th
e
q
u
an
titativ
e
p
er
f
o
r
m
an
ce
r
esu
lts
r
ep
o
r
ted
ab
o
v
e
.
I
n
th
ese
r
esu
lts
,
E
C
A+
Sp
atial
-
C
B
AM
+Res
NeX
t
-
5
0
ac
h
iev
ed
th
e
h
ig
h
est
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
T
h
is
an
aly
s
is
d
em
o
n
s
tr
ates
th
at
atten
tio
n
-
b
ased
o
p
tim
izatio
n
n
o
t
o
n
ly
im
p
r
o
v
es
Evaluation Warning : The document was created with Spire.PDF for Python.