I
A
ES
I
n
te
r
n
at
i
on
al
Jou
r
n
al
of
A
r
ti
fi
c
i
al
I
n
te
l
l
i
ge
n
c
e
(I
J
-
AI
)
V
ol
.
14
,
N
o
.
5,
O
c
t
obe
r
2
025
,
pp
.
4074
~
4089
IS
S
N
:
2252
-
8938
,
D
O
I
:
10.
1
1591
/
i
j
a
i
.
v
14
.i
5
.
p
p
4074
-
4
089
4074
Jou
r
n
al
h
o
m
e
pa
ge
:
ht
t
p:
/
/
i
j
ai
.
i
ae
s
c
or
e
.
c
om
Fa
c
i
a
l
p
a
r
a
l
y
s
i
s
i
m
a
g
e
a
n
a
l
y
si
s
f
o
r
s
t
r
o
k
e
d
e
t
e
c
t
i
o
n
u
si
n
g
d
e
e
p
e
n
s
e
mb
l
e
t
r
a
n
sf
e
r
l
e
a
r
n
i
n
g
a
n
d
o
p
t
i
mi
z
a
t
i
o
n
K
i
r
u
th
i
ga
S
u
b
r
aman
i
yan
1
,
C
h
i
n
n
as
amy
A
n
b
u
an
an
th
1
,
D
h
i
l
i
p
K
u
ma
r
V
e
n
k
ate
s
an
2
1
D
e
p
a
rt
m
e
n
t
of
Co
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
ri
n
g
,
F
a
c
u
l
t
y
of
E
n
g
i
n
e
e
ri
n
g
a
n
d
T
e
c
h
n
o
l
o
g
y
,
A
n
n
a
m
a
l
a
i
U
n
i
v
e
rs
i
t
y
,
Ch
i
d
a
m
b
a
ra
m
,
In
d
i
a
2
D
e
p
a
rt
m
e
n
t
of
Co
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
ri
n
g
,
S
c
h
o
o
l
of
Co
m
p
u
t
i
n
g
,
V
e
l
T
e
c
h
.
Ra
n
g
a
ra
j
a
n
D
r.
S
a
g
u
n
t
h
a
l
a
R&
D
In
s
t
i
t
u
t
e
of
S
c
i
e
n
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
Ch
e
n
n
a
i
,
In
d
i
a
A
r
ti
c
l
e
I
n
fo
A
BS
TR
A
C
T
Ar
t
i
c
l
e
h
i
s
t
or
y
:
Re
c
e
i
v
e
d
J
a
n
6,
20
25
Re
vi
s
e
d
J
u
l
1,
2025
A
c
c
e
pt
e
d
J
u
l
1
3,
2025
F
a
c
i
a
l
pa
r
a
l
ys
i
s
(
F
P
)
w
e
a
k
e
ns
f
a
c
i
a
l
m
u
s
c
l
e
s
,
l
e
a
d
i
n
g
to
a
s
y
m
m
e
t
r
i
c
f
a
c
i
a
l
a
c
t
i
on
s
a
n
d
c
o
m
p
l
i
c
a
t
i
ng
s
t
r
oke
di
a
gno
s
i
s
.
M
a
c
h
i
ne
l
e
a
r
n
i
ng
(
M
L
)
a
nd
de
e
p
l
e
a
r
n
i
ng
(
D
L
)
s
ys
t
e
m
s
ha
ve
be
e
n
e
xp
l
or
e
d
f
o
r
d
i
a
gno
s
i
n
g
F
P
,
bu
t
t
he
e
f
f
e
c
t
i
ve
ne
s
s
of
t
he
s
e
m
e
t
hod
s
is
hi
nde
r
e
d
by
t
he
l
i
m
i
t
e
d
s
i
z
e
a
nd
di
ve
r
s
i
t
y
of
a
va
i
l
a
b
l
e
da
t
a
s
e
t
s
.
T
h
i
s
s
t
udy
p
r
opo
s
e
s
a
no
ve
l
d
e
e
p
e
n
s
e
m
b
l
e
t
r
a
ns
f
e
r
l
e
a
r
n
i
ng
m
e
t
ho
d
f
o
r
a
c
c
u
r
a
t
e
s
t
r
o
ke
d
i
a
g
nos
i
s
u
s
i
ng
f
a
c
i
a
l
pa
r
a
l
ys
i
s
i
m
a
g
i
ng
(
D
E
T
L
M
-
A
S
D
F
P
I
)
.
T
h
e
m
e
t
ho
d
l
e
ve
r
a
g
e
s
pr
e
-
t
r
a
i
n
e
d
m
o
de
l
s
to
r
e
duc
e
c
om
pu
t
a
t
i
o
n
c
o
s
t
s
on
e
dg
e
de
vi
c
e
s
.
T
h
e
f
r
a
m
e
w
or
k
i
n
c
l
u
de
s
d
a
t
a
a
c
qu
i
s
i
t
i
on
,
p
r
e
p
a
r
a
t
i
on
,
a
n
d
pr
e
-
p
r
o
c
e
s
s
i
ng,
w
i
t
h
i
m
a
ge
r
e
s
c
a
l
i
ng
to
s
t
a
nda
r
d
i
z
e
i
np
ut
d
i
m
e
n
s
i
ons
.
F
e
a
t
ur
e
e
x
t
r
a
c
t
i
on
is
pe
r
f
o
r
m
e
d
u
s
i
ng
a
de
e
p
c
a
ps
ul
e
ne
t
w
or
k
(
D
C
a
ps
N
e
t
)
to
c
a
pt
ur
e
c
o
m
p
l
e
x
f
e
a
t
u
r
e
s
.
F
or
s
t
r
o
ke
de
t
e
c
t
i
o
n,
an
e
ns
e
m
bl
e
t
r
a
n
s
f
e
r
l
e
a
r
ni
ng
m
ode
l
i
nt
e
gr
a
t
e
s
t
hr
e
e
c
l
a
s
s
i
f
i
e
r
s:
ga
t
e
d
r
e
c
u
r
r
e
n
t
un
i
t
(
G
R
U
)
,
d
e
e
p
c
onv
ol
ut
i
on
a
l
n
e
u
r
a
l
ne
t
w
o
r
k
(
D
C
N
N
)
,
a
nd
s
t
a
c
k
e
d
s
pa
r
s
e
a
ut
o
-
e
nc
o
de
r
(
S
S
A
E
)
.
T
he
h
i
ppo
pot
a
m
us
op
t
i
m
i
z
a
t
i
on
a
l
go
r
i
t
h
m
(
H
O
A
)
is
a
pp
l
i
e
d
to
f
i
n
e
-
t
un
e
m
ode
l
p
a
r
a
m
e
t
e
r
s
.
T
he
m
e
t
hod
w
a
s
va
l
i
da
t
e
d
u
s
i
ng
t
w
o
be
nc
h
m
a
r
k
da
t
a
s
e
t
s
,
M
a
s
s
a
c
hu
s
e
t
t
s
e
ye
a
nd
e
a
r
i
nf
i
r
m
a
r
y
(
M
E
E
I
)
a
nd
Y
ouT
ub
e
f
a
c
i
a
l
p
a
l
s
y
(
Y
F
P
)
,
a
c
h
i
e
v
i
ng
an
a
c
c
ur
a
c
y
of
97
.
06
%
,
ou
t
pe
r
f
or
m
i
ng
r
e
c
e
n
t
a
pp
r
o
a
c
h
e
s
.
T
hi
s
r
e
s
e
a
r
c
h
d
e
m
ons
t
r
a
t
e
s
t
he
e
f
f
e
c
t
i
ve
ne
s
s
of
t
he
D
E
T
L
M
-
A
S
D
F
P
I
m
e
t
h
od
in
a
c
c
u
r
a
t
e
l
y
d
i
a
gno
s
i
n
g
s
t
r
oke
s
f
r
om
FP
i
m
a
ge
s
w
h
i
l
e
a
dd
r
e
s
s
i
ng
c
ha
l
l
e
ng
e
s
r
e
l
a
t
e
d
to
da
t
a
s
e
t
l
i
m
i
t
a
t
i
on
s
a
nd
c
o
m
p
ut
a
t
i
on
a
l
e
f
f
i
c
i
e
n
c
y
.
Ke
y
w
or
d
s
:
D
a
t
a
pr
e
-
pro
c
e
s
s
i
ng
F
a
c
i
a
l
pa
ra
l
ys
i
s
i
m
a
ge
s
H
i
ppopo
t
a
m
us
o
pt
i
m
i
z
a
t
i
on
S
t
roke
de
t
e
c
t
i
on
T
ra
ns
f
e
r
l
e
a
rn
i
ng
T
hi
s
is
an
op
e
n
ac
c
e
s
s
ar
t
i
c
l
e
u
nde
r
t
he
CC
BY
-
SA
l
i
c
e
n
s
e
.
Cor
r
e
s
pon
di
n
g
Au
t
h
or
:
K
i
rut
h
i
ga
S
ubr
a
m
a
ni
ya
n
D
e
pa
r
t
m
e
nt
of
Co
m
pu
t
e
r
S
c
i
e
n
c
e
a
nd
E
ng
i
ne
e
ri
n
g
,
F
a
c
ul
t
y
of
E
n
gi
n
e
e
r
i
ng
a
n
d
T
e
c
hno
l
ogy
A
nna
m
a
l
a
i
U
ni
v
e
rs
i
t
y
Chi
da
m
b
a
ra
m
,
T
a
m
i
l
na
du,
Ind
i
a
E
m
a
i
l
:
ki
ru
t
hi
ga
r
e
s
e
a
rc
hs
c
hol
a
r
@
gm
a
i
l
.
c
o
m
1.
I
N
TR
O
D
U
C
TI
O
N
A
s
t
ro
k
e
,
t
h
e
s
e
c
on
d
p
a
r
a
m
o
un
t
r
e
a
s
o
n
for
d
e
m
i
s
e
g
l
ob
a
l
l
y
a
n
d
a
pr
i
m
a
ry
r
e
a
s
o
n
is
a
d
i
s
a
b
i
l
i
t
y
,
c
a
n
be
a
n
e
u
ro
l
og
i
c
a
l
d
e
f
i
c
i
e
n
c
y
t
h
a
t
ou
t
c
o
m
e
s
m
os
t
l
y
fr
o
m
s
e
v
e
r
e
f
o
c
a
l
d
a
m
a
g
e
of
t
h
e
c
e
n
t
r
a
l
n
e
r
vo
us
s
ys
t
e
m
(C
N
S
)
ow
i
n
g
to
v
a
s
c
u
l
a
r
r
e
a
s
ons
s
u
c
h
as
i
nt
r
a
c
e
r
e
br
a
l
h
a
e
m
o
rr
h
a
g
e
,
s
u
ba
r
a
c
h
n
oi
d
h
a
e
m
o
rr
h
a
g
e
,
a
n
d
c
e
r
e
br
a
l
i
n
f
a
r
c
t
i
on
[1
]
.
I
t
’
s
a
m
e
d
i
c
a
l
c
r
i
s
i
s
a
nd
h
e
n
c
e
n
e
e
ds
i
ns
t
a
nt
t
r
e
a
t
m
e
nt
.
To
a
c
c
ur
a
t
e
l
y
d
e
t
e
c
t
s
us
p
e
c
t
e
d
s
t
r
ok
e
p
a
t
i
e
n
t
s
,
F
a
c
i
a
l
p
a
r
a
l
ys
i
s
(
FP
)
h
a
s
b
e
e
n
e
x
t
e
ns
i
v
e
l
y
a
p
p
l
i
e
d
f
or
e
m
e
r
g
e
n
c
y
m
e
d
i
c
a
l
t
r
e
a
t
m
e
nt
.
FP
d
a
m
a
g
e
s
f
a
c
i
a
l
a
c
t
i
o
ns
du
e
to
n
e
r
v
e
p
a
t
h
o
l
o
gy
[
2]
.
It
c
a
us
e
s
d
a
m
a
g
e
to
t
h
e
f
un
c
t
i
o
n
of
i
n
t
e
n
t
i
on
a
l
f
a
c
i
a
l
m
u
s
c
l
e
s
s
t
r
e
ng
t
h
e
n
e
d
t
h
r
ou
gh
t
h
e
f
a
c
i
a
l
n
e
r
v
e
,
r
e
s
u
l
t
i
n
g
in
f
a
c
i
a
l
a
n
o
m
a
l
i
e
s
[3
]
.
P
e
r
i
ph
e
r
a
l
FP
c
or
r
e
s
p
o
nds
to
ne
r
v
e
d
i
s
o
rd
e
r
i
ns
i
d
e
t
he
po
n
t
i
n
e
of
t
h
e
b
r
a
i
n
s
t
e
m
.
It
a
f
f
e
c
t
s
t
h
e
m
u
s
c
l
e
s
of
t
h
e
f
a
c
e
in
t
h
e
u
p
pe
r
,
m
i
dd
l
e
,
a
n
d
l
ow
e
r
a
r
e
a
s
of
t
h
e
f
a
c
e
on
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
A
rt
i
f
I
nt
e
l
l
IS
S
N
:
2252
-
8938
F
ac
i
al
par
a
l
y
s
i
s
i
m
ag
e
anal
y
s
i
s
f
or
s
t
r
ok
e
d
e
t
e
c
t
i
o
n
us
i
ng
de
e
p
e
ns
e
m
b
l
e
…
(
Ki
r
u
t
hi
ga
Su
br
am
ani
y
an
)
4075
on
e
s
i
d
e
.
C
e
n
t
r
a
l
FP
(
b
e
c
a
u
s
e
of
s
t
r
o
ke
)
b
e
c
o
m
e
s
n
e
r
v
e
f
a
i
l
u
r
e
w
i
t
h
i
n
t
he
m
o
t
o
r
c
or
t
i
c
a
l
r
e
g
i
on
s
;
t
h
e
l
o
w
e
r
p
a
r
t
of
one
s
i
de
of
t
h
e
f
a
c
e
can
be
i
nf
l
ue
n
c
e
d
[4]
.
M
os
t
of
t
h
e
FP
p
a
t
i
e
n
t
s
b
e
c
o
m
e
i
n
f
e
c
t
e
d
w
i
t
h
p
e
r
i
ph
e
r
a
l
F
P
.
T
h
e
n
,
it
h
a
s
an
e
m
o
t
i
on
a
l
i
m
p
a
c
t
on
m
os
t
f
a
c
i
a
l
m
u
s
c
l
e
s
of
t
h
e
f
a
c
e
on
one
s
i
d
e
;
i
t
'
s
p
ro
b
l
e
m
a
t
i
c
fo
r
t
h
e
p
a
t
i
e
n
t
to
c
a
rr
y
ou
t
t
h
e
r
e
g
u
l
a
r
a
c
t
i
o
ns
of
t
h
e
m
o
u
t
h
,
e
y
e
b
r
ow
s
,
a
nd
e
y
e
s
[
5]
.
D
e
t
e
c
t
i
n
g
t
h
e
fa
c
i
a
l
pa
l
s
y
pr
o
c
e
s
s
is
s
i
gn
i
fi
c
a
n
t
in
m
e
a
s
ur
i
ng
t
h
e
d
i
f
f
i
c
u
l
t
y
of
t
h
e
m
us
c
l
e
f
a
i
l
ur
e
a
nd
fa
c
i
a
l
ne
r
v
e
,
r
e
c
o
r
di
n
g
p
hys
i
c
a
l
de
v
e
l
o
p
m
e
nt
s
a
f
t
e
r
t
r
e
a
t
m
e
nt
,
a
nd
o
bs
e
r
v
i
n
g
t
h
e
p
a
t
i
e
n
t
[
6]
.
Co
m
p
ut
e
r
i
z
e
d
a
ut
o
m
a
t
e
d
FP
d
e
t
e
c
t
i
on
is
s
i
g
n
i
f
i
c
a
nt
in
d
e
v
e
l
op
i
ng
s
ys
t
e
m
a
t
i
z
e
d
d
e
v
i
c
e
s
for
t
r
e
a
t
m
e
n
t
,
m
o
n
i
t
or
i
n
g
,
a
nd
m
e
di
c
a
l
v
a
l
u
a
t
i
o
n
,
a
n
d
m
e
d
i
c
a
l
p
r
i
c
e
s
ha
v
e
d
e
c
r
e
a
s
e
d
ov
e
r
t
h
e
a
dd
i
t
i
o
n
of
a
u
t
o
m
a
t
e
d
m
e
t
h
o
ds
[7
]
.
M
or
e
o
ve
r
,
c
o
m
p
u
t
e
r
i
z
e
d
m
e
t
h
od
s
a
r
e
p
re
d
i
c
t
e
d
to
o
f
f
e
r
c
o
nv
e
n
i
e
n
t
d
e
v
i
c
e
s
,
s
ho
rt
l
y
,
f
or
p
a
t
i
e
nt
m
on
i
t
or
i
ng
in
t
h
e
h
o
us
e
.
M
a
i
n
l
y
in
c
o
m
pu
t
e
r
v
i
s
i
o
n
(
CV
)
,
t
h
e
s
t
ud
y
of
fa
c
i
a
l
m
o
v
e
m
e
n
t
s
h
a
s
i
n
s
p
i
r
e
d
s
e
v
e
r
a
l
i
n
v
e
s
t
i
g
a
t
i
ons
on
a
u
t
o
m
a
t
i
c
fa
c
i
a
l
n
e
rv
e
fu
n
c
t
i
o
n
v
a
l
u
a
t
i
on
from
t
h
e
b
i
ol
o
gi
c
a
l
v
i
s
u
a
l
c
a
p
t
ur
e
of
t
h
e
f
a
c
e
[8
]
.
R
e
c
e
n
t
l
y
,
w
i
t
h
t
h
e
a
s
s
i
s
t
a
n
c
e
of
A
I
,
i
n
v
e
s
t
i
g
a
t
or
s
h
a
v
e
s
ug
g
e
s
t
e
d
g
r
a
d
u
a
l
l
y
m
or
e
p
re
c
i
s
e
d
e
t
e
c
t
i
o
n
a
n
d
a
s
s
e
s
s
m
e
n
t
m
o
d
e
l
s
f
or
C
N
S
c
on
d
i
t
i
on
s
.
In
t
he
c
ur
r
e
nt
i
nv
e
s
t
i
g
a
t
i
o
n
,
d
e
e
p
l
e
a
rn
i
ng
(
DL
)
a
n
d
m
a
c
h
i
n
e
l
e
a
rn
i
ng
(
ML
)
m
o
d
e
l
s
a
r
e
pr
e
s
e
n
t
e
d
fo
r
FP
d
e
t
e
c
t
i
o
n
a
n
d
p
r
e
d
i
c
t
i
o
n
of
t
h
e
s
e
v
e
r
i
t
y
[9
]
.
N
e
v
e
r
t
h
e
l
e
s
s
,
ML
m
o
d
e
l
s
h
a
v
e
r
e
s
t
ri
c
t
i
ons
as
t
h
e
y
d
e
p
e
nd
on
f
a
c
i
a
l
l
a
n
d
m
a
r
k
m
e
t
h
ods
a
nd
p
h
ys
i
c
a
l
fa
c
e
pa
r
a
l
y
s
i
s
a
r
e
a
e
x
t
r
a
c
t
i
on
m
o
d
e
l
s
to
g
e
t
s
pa
t
i
a
l
d
a
t
a
.
M
e
a
n
w
h
i
l
e
,
DL
t
e
c
h
n
i
q
u
e
s
n
o
r
m
a
l
l
y
u
t
i
l
i
z
e
CN
N
s
to
re
m
o
v
e
in
-
d
e
p
t
h
f
e
a
t
ur
e
s
fr
o
m
f
a
c
i
a
l
f
e
a
t
u
r
e
s
[
10
]
.
T
h
e
y
i
d
e
n
t
i
f
y
s
l
i
gh
t
m
od
i
fi
c
a
t
i
ons
a
nd
d
e
t
e
c
t
p
a
t
t
e
r
ns
from
f
a
c
i
a
l
i
m
a
g
e
s
t
h
a
t
a
s
s
i
s
t
in
i
m
pr
o
v
e
d
r
e
c
o
gn
i
t
i
o
n
of
FP.
T
hi
s
s
t
u
d
y
p
r
e
s
e
n
t
s
a
n
e
w
d
e
e
p
e
ns
e
m
b
l
e
t
r
a
ns
fe
r
l
e
a
r
n
i
n
g
m
e
t
h
o
d
f
or
a
c
c
ur
a
t
e
s
t
ro
k
e
di
a
g
nos
i
s
us
i
ng
fa
c
i
a
l
p
a
r
a
l
ys
i
s
i
m
a
g
i
n
g
(D
E
T
L
M
-
A
S
D
F
P
I
)
m
e
t
h
od
.
T
he
D
E
T
L
M
-
A
S
D
F
P
I
m
e
t
h
od
a
i
m
s
to
c
l
a
s
s
i
fy
s
t
ro
k
e
s
t
h
a
t
e
x
i
s
t
in
FP
i
m
a
g
e
s
p
r
of
i
c
i
e
n
t
l
y
.
T
h
i
s
s
t
ud
y
f
o
c
u
s
e
s
on
an
a
d
v
a
nc
e
d
di
a
gn
os
t
i
c
f
r
a
m
e
w
or
k
i
n
t
e
gr
a
t
i
n
g
d
a
t
a
a
c
q
u
i
s
i
t
i
o
n
,
p
r
e
p
a
r
a
t
i
o
n
,
a
n
d
pre
-
pr
o
c
e
s
s
i
ng
of
FP
i
m
a
g
e
s
.
T
he
p
re
-
pr
o
c
e
s
s
i
n
g
ph
a
s
e
i
n
c
l
u
de
s
r
e
s
c
a
l
i
n
g
t
h
e
i
m
a
g
e
s
to
s
t
a
n
d
a
r
di
z
e
i
np
u
t
d
i
m
e
ns
i
on
s
.
A
l
s
o
,
t
h
e
d
e
e
p
c
a
p
s
u
l
e
n
e
t
w
or
k
(
D
C
a
ps
N
e
t
)
m
e
t
ho
d
is
us
e
d
f
or
f
e
a
t
ur
e
e
xt
r
a
c
t
i
o
n
to
l
e
a
r
n
c
o
m
p
l
e
x
f
e
a
t
ur
e
s
f
ro
m
t
h
e
pr
e
-
p
r
o
c
e
s
s
e
d
d
a
t
a
.
F
or
t
h
e
s
t
r
ok
e
d
e
t
e
c
t
i
o
n
pr
o
c
e
s
s
,
an
e
ns
e
m
b
l
e
t
r
a
ns
f
e
r
l
e
a
rn
i
ng
(
T
L
)
m
o
de
l
u
t
i
l
i
z
e
s
t
hr
e
e
c
l
a
s
s
i
f
i
e
r
s
s
u
c
h
as
g
a
t
e
d
r
e
c
ur
r
e
n
t
u
n
i
t
(G
R
U
)
,
d
e
e
p
c
on
v
ol
u
t
i
o
n
a
l
n
e
u
r
a
l
ne
t
w
o
rk
(
D
C
N
N
)
,
a
nd
s
t
a
c
k
e
d
s
p
a
r
s
e
a
ut
o
-
e
n
c
od
e
r
(S
S
A
E
)
.
E
v
e
n
t
u
a
l
l
y
,
t
h
e
h
i
pp
op
o
t
a
m
u
s
op
t
i
m
i
z
a
t
i
o
n
a
l
g
or
i
t
h
m
(H
O
A
)
is
u
t
i
l
i
z
e
d
to
o
pt
i
m
i
z
e
p
a
r
a
m
e
t
e
r
t
u
n
i
n
g
fo
r
t
h
e
t
h
re
e
e
ns
e
m
b
l
e
t
e
c
h
n
i
q
u
e
s
.
To
i
m
pr
ov
e
t
h
e
pr
e
di
c
t
i
on
r
e
s
u
l
t
s
of
t
h
e
D
E
T
L
M
-
A
S
D
F
P
I
m
e
t
h
o
do
l
og
y
,
a
s
e
qu
e
n
c
e
of
e
xp
e
ri
m
e
n
t
s
is
i
m
p
l
e
m
e
n
t
e
d
on
t
w
o
b
e
n
c
h
m
a
r
k
d
a
t
a
s
e
t
s
s
u
c
h
as
M
a
s
s
a
c
h
us
e
t
t
s
e
y
e
a
nd
e
a
r
i
nf
i
r
m
a
r
y
(
M
E
E
I
)
a
n
d
Y
ou
T
ub
e
f
a
c
i
a
l
p
a
l
s
y
(
Y
F
P
)
.
T
h
e
m
a
j
o
r
c
o
n
t
r
i
bu
t
i
on
of
t
h
e
D
E
T
L
M
-
A
S
D
F
P
I
m
e
t
h
od
o
l
o
g
y
is
l
i
s
t
e
d
a
s
f
ol
l
ow
s
:
−
T
he
D
E
T
L
M
-
A
S
D
F
P
I
m
od
e
l
i
n
t
e
gr
a
t
e
s
a
r
ob
us
t
pre
-
p
ro
c
e
s
s
i
ng
s
t
e
p
to
r
e
s
c
a
l
e
m
e
d
i
c
a
l
i
m
a
g
e
s
,
c
on
fi
r
m
i
n
g
un
i
fo
r
m
i
n
p
ut
s
i
z
e
s
for
t
h
e
s
u
bs
e
q
u
e
n
t
f
e
a
t
u
r
e
e
xt
r
a
c
t
i
o
n
a
n
d
m
od
e
l
t
r
a
i
n
i
n
g
s
t
a
ge
s
.
T
h
i
s
no
r
m
a
l
i
z
a
t
i
o
n
a
s
s
i
s
t
s
in
m
a
i
n
t
a
i
ni
n
g
c
o
ns
i
s
t
e
nc
y
a
c
ro
s
s
di
v
e
r
s
e
i
m
a
g
e
f
or
m
a
t
s
a
n
d
i
m
p
r
ov
e
s
t
h
e
p
e
r
fo
r
m
a
n
c
e
of
t
h
e
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
p
ro
c
e
s
s
.
S
t
a
n
da
r
di
z
i
n
g
i
m
a
g
e
d
i
m
e
ns
i
on
s
e
n
a
b
l
e
s
t
h
e
m
e
t
h
od
to
l
e
a
r
n
p
a
t
t
e
r
ns
m
or
e
e
ff
e
c
t
u
a
l
l
y
,
i
m
p
ro
v
i
n
g
s
t
ro
k
e
d
e
t
e
c
t
i
o
n
a
c
c
ur
a
c
y
.
−
T
he
D
E
T
L
M
-
A
S
D
F
P
I
m
e
t
h
o
d
u
t
i
l
i
z
e
s
D
C
a
ps
N
e
t
to
c
a
pt
u
re
s
p
a
t
i
a
l
h
i
e
r
a
rc
h
i
e
s
a
n
d
p
a
r
t
-
w
h
ol
e
re
l
a
t
i
ons
h
i
p
s
in
m
e
di
c
a
l
i
m
a
g
e
s
.
T
h
i
s
a
l
l
ow
s
fo
r
m
or
e
pr
e
c
i
s
e
a
nd
ro
b
us
t
f
e
a
t
u
r
e
e
xt
r
a
c
t
i
o
n
,
w
hi
c
h
is
s
i
gn
i
fi
c
a
n
t
f
or
d
e
t
e
c
t
i
n
g
s
t
r
ok
e
-
r
e
l
a
t
e
d
p
a
t
t
e
rn
s
.
By
e
m
p
l
o
y
i
n
g
D
C
a
ps
N
e
t
,
t
h
e
a
pp
r
oa
c
h
e
n
h
a
n
c
e
s
t
h
e
a
c
c
u
r
a
c
y
a
nd
r
e
l
i
a
b
i
l
i
t
y
of
s
t
ro
k
e
d
e
t
e
c
t
i
o
n
fr
o
m
c
o
m
p
l
e
x
m
e
d
i
c
a
l
i
m
a
g
i
n
g
da
t
a
.
−
T
he
D
E
T
L
M
-
A
S
D
F
P
I
m
o
d
e
l
i
n
c
or
p
or
a
t
e
s
t
hr
e
e
c
l
a
s
s
i
fi
e
rs
,
G
RU
,
D
C
N
N
,
a
nd
S
S
A
E
,
i
n
t
o
an
e
n
s
e
m
b
l
e
TL
m
od
e
l
to
e
nh
a
n
c
e
s
t
ro
k
e
d
e
t
e
c
t
i
o
n
.
T
h
i
s
a
pp
ro
a
c
h
u
t
i
l
i
z
e
s
t
h
e
m
e
ri
t
s
of
e
v
e
r
y
c
l
a
s
s
i
f
i
e
r
,
i
nt
e
gr
a
t
i
ng
s
e
qu
e
n
t
i
a
l
,
s
p
a
t
i
a
l
,
a
nd
a
bs
t
r
a
c
t
f
e
a
t
ur
e
l
e
a
rn
i
ng
for
m
or
e
a
c
c
u
ra
t
e
o
u
t
c
o
m
e
s
.
T
h
e
m
o
d
e
l
be
n
e
f
i
t
s
f
ro
m
pre
-
t
r
a
i
n
e
d
kn
ow
l
e
dg
e
u
t
i
l
i
z
i
ng
TL,
i
m
pr
ov
i
ng
g
e
n
e
r
a
l
i
z
a
t
i
on
a
nd
p
e
r
fo
r
m
a
n
c
e
on
m
e
d
i
c
a
l
d
a
t
a
s
e
t
s
.
−
T
he
D
E
T
L
M
-
A
S
D
F
P
I
m
e
t
ho
d
ol
o
gi
e
s
ut
i
l
i
z
e
HOA
f
or
o
pt
i
m
a
l
h
yp
e
rp
a
r
a
m
e
t
e
r
t
un
i
ng
,
c
o
nf
i
r
m
i
n
g
i
m
p
r
ov
e
d
m
od
e
l
p
e
rf
or
m
a
n
c
e
in
s
t
r
ok
e
d
e
t
e
c
t
i
o
n
.
HOA
a
s
s
i
s
t
s
in
a
v
o
i
d
i
ng
p
r
e
m
a
t
u
r
e
c
on
v
e
r
g
e
n
c
e
by
e
f
f
e
c
t
i
v
e
l
y
e
xp
l
or
i
ng
t
he
s
o
l
u
t
i
on
s
p
a
c
e
a
nd
f
i
n
e
-
t
u
n
i
n
g
m
o
d
e
l
p
a
r
a
m
e
t
e
rs
.
T
h
i
s
e
nh
a
n
c
e
s
t
h
e
t
e
c
h
ni
q
u
e
'
s
a
c
c
u
r
a
c
y
a
nd
r
ob
us
t
n
e
s
s
,
r
e
s
u
l
t
i
n
g
in
m
or
e
r
e
l
i
a
b
l
e
s
t
r
ok
e
d
e
t
e
c
t
i
o
n
r
e
s
u
l
t
s
.
−
T
he
D
E
T
L
M
-
A
S
D
F
P
I
a
p
pr
o
a
c
h
i
n
c
or
po
r
a
t
e
s
e
ns
e
m
b
l
e
TL
w
i
t
h
t
h
r
e
e
d
i
s
t
i
n
c
t
c
l
a
s
s
i
f
i
e
rs
,
G
RU
,
D
CN
N
,
a
nd
S
S
A
E
,
to
u
t
i
l
i
z
e
t
h
e
i
r
c
o
m
p
l
e
m
e
n
t
a
r
y
m
e
ri
t
s
in
s
t
r
o
k
e
d
e
t
e
c
t
i
o
n
.
T
hi
s
i
n
t
e
g
r
a
t
i
o
n
e
n
h
a
n
c
e
s
t
h
e
t
e
c
h
n
i
q
u
e
'
s
c
a
p
a
b
i
l
i
t
y
to
c
a
p
t
ur
e
s
e
q
u
e
n
t
i
a
l
,
s
p
a
t
i
a
l
,
a
n
d
a
bs
t
r
a
c
t
f
e
a
t
u
re
s
from
d
i
v
e
r
s
e
d
a
t
a
s
o
ur
c
e
s
.
M
o
re
o
v
e
r
,
us
i
n
g
t
h
e
H
O
A
for
p
a
r
a
m
e
t
e
r
t
un
i
ng
i
m
pr
ov
e
s
t
h
e
m
od
e
l
'
s
e
ff
i
c
i
e
n
c
y
by
a
v
oi
d
i
n
g
pr
e
m
a
t
ur
e
c
on
v
e
r
g
e
n
c
e
a
n
d
op
t
i
m
i
z
i
n
g
p
e
r
f
or
m
a
n
c
e
.
T
h
i
s
i
n
t
e
gr
a
t
i
on
of
m
od
e
l
s
r
e
s
u
l
t
s
in
a
m
o
r
e
ro
b
us
t
a
nd
a
d
a
p
t
i
v
e
m
od
e
l
fo
r
p
r
e
c
i
s
e
s
t
r
ok
e
d
e
t
e
c
t
i
on
a
c
ro
s
s
v
a
r
i
e
d
d
a
t
a
s
e
t
s
.
2.
R
ELA
TED
WO
R
K
S
Ou
et
al
.
[
11]
i
n
t
rod
uc
e
d
an
i
nno
va
t
i
v
e
m
u
l
t
i
m
o
da
l
DL
t
e
c
h
ni
que
ba
s
e
d
on
t
he
f
a
c
e
a
r
m
s
pe
e
c
h
t
e
s
t
(F
A
S
T
)
for
e
v
a
l
u
a
t
i
ng
s
us
p
i
c
i
ous
s
t
r
oke
pa
t
i
e
nt
s
s
how
i
ng
s
ym
pt
om
s
l
i
k
e
s
p
e
e
c
h
d
i
s
orde
rs
,
l
i
m
b
w
e
a
k
ne
s
s
,
a
nd
f
a
c
i
a
l
p
a
ra
l
ys
i
s
in
s
e
ve
r
e
s
c
e
n
e
ri
e
s
.
Ba
s
e
d
on
t
he
F
A
S
T
,
t
he
a
u
t
hor
ga
t
he
r
e
d
a
d
a
t
a
s
e
t
c
ont
a
i
n
i
ng
a
ud
i
o
a
nd
vi
de
o
r
e
c
o
rdi
n
gs
of
i
n
t
e
ns
i
ve
c
a
r
e
uni
t
pa
t
i
e
nt
s
e
na
c
t
i
ng
s
p
e
c
i
fi
e
d
s
pe
e
c
h
t
e
s
t
s
,
l
i
m
b
m
ov
e
m
e
n
t
s
,
a
nd
fa
c
i
a
l
e
xpr
e
s
s
i
ons
.
T
h
e
a
u
t
hor
e
qua
t
e
d
t
he
d
e
v
e
l
op
e
d
DL
m
e
t
hod
to
proc
e
s
s
m
ul
t
i
m
od
a
l
d
a
t
a
s
e
t
s
w
i
t
h
s
i
x
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8938
Int
J
A
rt
i
f
I
nt
e
l
l
,
V
ol
.
1
4,
N
o.
5
,
O
c
t
o
be
r
2025
:
4074
-
4
089
4076
pre
c
e
di
ng
m
e
t
h
ods
,
w
hi
c
h
a
t
t
a
i
n
e
d
gr
e
a
t
m
o
ve
m
e
n
t
c
l
a
s
s
i
fi
e
r
e
xe
c
ut
i
on
,
c
on
t
a
i
ni
ng
t
h
e
m
u
l
t
i
s
c
a
l
e
v
i
s
i
on
t
ra
ns
for
m
e
r
(
M
V
i
T
)
,
i
n
fl
a
t
e
d
3D
(
I
3D
)
,
S
l
ow
F
a
s
t
,
e
xt
e
nd
e
d
3D
(
X
3D
)
,
a
n
d
t
e
m
por
a
l
pyr
a
m
i
d
n
e
t
w
ork
(
T
P
N
)
.
U
m
i
rz
a
kov
a
et
al
.
[12]
i
nt
rod
uc
e
a
DL
-
b
a
s
e
d
t
e
c
hn
i
que
for
di
a
gnos
i
n
g
FP
i
l
l
ne
s
s
e
s
l
i
ke
B
e
l
l
'
s
pa
l
s
y
a
nd
s
t
rok
e
.
T
hi
s
m
e
t
hod
us
e
s
m
ul
t
i
-
t
a
s
k
n
e
t
w
orks
,
i
n
c
orpo
r
a
t
i
ng
fa
c
i
a
l
a
s
y
m
m
e
t
ry,
f
a
c
e
a
n
a
l
ys
i
s
,
a
nd
t
yp
e
i
m
pro
ve
m
e
n
t
to
i
de
n
t
i
f
y
t
he
c
os
t
s
a
c
c
o
m
pa
nyi
n
g
c
onve
nt
i
on
a
l
di
a
gnos
i
ng
t
e
c
hni
que
s
l
i
k
e
c
o
m
put
e
d
t
om
o
gra
p
hy
(
CT
)
a
n
d
m
a
g
ne
t
i
c
r
e
s
ona
nc
e
i
m
a
gi
n
g
(M
R
I)
s
c
a
nn
e
d
i
m
a
g
e
s
.
S
pa
t
i
a
l
di
ff
e
r
e
nc
e
s
w
e
re
t
a
c
k
l
e
d
by
a
de
p
t
h
-
map
e
s
t
i
m
a
t
i
ng
m
o
dul
e
,
w
h
i
c
h
i
nfl
u
e
n
c
e
s
a
c
a
s
e
-
s
pe
c
i
f
i
e
d
k
e
rn
e
l
t
e
c
hni
qu
e
.
Chow
dhu
ry
et
al
.
[13]
p
re
s
e
nt
an
a
ut
o
m
a
t
i
c
s
ys
t
e
m
to
i
d
e
nt
i
fy
t
h
e
s
t
rok
e
from
pre
-
pro
c
e
s
s
e
d
d
a
t
a
ut
i
l
i
z
i
ng
CN
N
a
nd
ot
h
e
r
DL
m
e
t
hods
.
T
h
e
pre
s
e
nt
e
d
a
p
proa
c
h
is
pr
i
m
a
ri
l
y
to
de
t
e
r
m
i
n
e
t
he
s
t
rok
e
d
p
e
rs
on'
s
f
a
c
e
fro
m
t
h
e
e
xpr
e
s
s
i
ons
or
norm
a
l
fa
c
e
.
F
or
c
l
a
s
s
i
f
i
c
a
t
i
on
,
t
h
e
a
u
t
hor
g
a
ve
pr
e
-
pro
c
e
s
s
e
d
s
t
rok
e
i
m
a
g
e
s
for
t
r
a
i
n
i
ng
,
s
e
rv
e
d
t
h
e
m
i
nt
o
di
ffe
r
e
n
t
d
e
e
p
s
t
r
uc
t
ure
s
,
a
nd
ul
t
i
m
a
t
e
l
y,
b
a
s
e
d
on
t
h
e
c
a
t
e
g
ori
z
e
d
e
xpr
e
s
s
i
ons
,
t
he
a
ut
ho
r
c
a
t
e
go
ri
z
e
d
s
t
rok
e
a
nd
norm
a
l
p
a
t
i
e
n
t
s
.
Ca
i
et
al
.
[14]
pre
s
e
nt
a
n
e
w
m
u
l
t
i
m
o
da
l
DL
m
e
t
hod
,
D
e
e
p
S
t
roke
,
to
a
t
t
a
i
n
c
o
m
put
e
r
-
a
i
d
e
d
s
t
r
oke
i
nc
i
de
n
c
e
e
v
a
l
u
a
t
i
on
by
di
a
gnos
i
ng
m
ode
l
s
of
t
ri
v
i
a
l
fa
c
i
a
l
m
us
c
l
e
d
i
s
c
oo
rdi
n
a
t
i
on
a
nd
s
p
e
e
c
h
i
nc
a
pa
b
i
l
i
t
y
for
pa
t
i
e
nt
s
w
i
t
h
s
t
roke
i
m
pre
s
s
i
on
in
a
c
ri
t
i
c
a
l
s
i
t
ua
t
i
on
.
T
h
e
pre
s
e
nt
e
d
D
e
e
pS
t
rok
e
e
n
dure
s
1
-
m
i
nut
e
f
a
c
i
a
l
a
udi
o
a
nd
vi
d
e
o
d
a
t
a
fre
e
l
y
a
c
c
e
s
s
i
bl
e
in
s
t
rok
e
t
r
i
a
g
e
s
for
l
oc
a
l
FP
r
e
c
og
ni
t
i
on
a
nd
g
l
ob
a
l
s
p
e
e
c
h
i
l
l
n
e
s
s
e
xa
m
i
n
a
t
i
on
.
TL
h
a
s
b
e
e
n
i
m
p
l
e
m
e
nt
e
d
to
de
c
re
a
s
e
fa
c
e
-
a
t
t
ri
bu
t
e
b
i
a
s
e
s
a
n
d
e
nh
a
nc
e
ge
n
e
ra
l
i
z
a
bi
l
i
t
y
.
E
l
h
a
na
s
h
i
et
al
.
[
15]
propos
e
a
n
e
w
t
e
c
hni
q
ue
to
c
on
fron
t
t
he
s
e
c
r
uc
i
a
l
r
e
qui
r
e
m
e
n
t
s
by
pr
e
s
e
nt
i
ng
a
re
a
l
-
t
i
m
e
s
t
roke
re
c
ogn
i
t
i
on
m
e
t
hod
t
h
a
t
d
e
p
e
nds
on
DL
us
i
ng
fe
d
e
ra
t
e
d
le
a
rn
i
ng
(F
L
)
to
i
m
p
rove
pr
e
c
i
s
i
on
a
nd
c
onfi
de
n
t
i
a
l
i
t
y
m
a
i
nt
e
na
nc
e
.
T
h
e
m
a
i
n
a
i
m
of
t
h
i
s
c
a
s
e
is
to
a
d
va
n
c
e
an
e
ffe
c
t
i
ve
a
nd
pre
c
i
s
e
m
e
t
h
od
of
di
s
c
ri
m
i
na
t
i
ng
b
e
t
w
e
e
n
no
n
-
s
t
rok
e
a
n
d
s
t
rok
e
pa
t
i
e
nt
s
at
p
re
s
e
nt
,
e
n
a
bl
i
ng
m
e
di
c
a
l
e
xp
e
r
t
s
to
m
a
ke
i
nt
e
l
l
e
c
t
u
a
l
d
e
c
i
s
i
ons
.
P
hi
e
nph
a
n
i
c
h
et
al
.
[1
6]
pr
opos
e
t
h
e
us
a
ge
of
a
f
a
c
i
a
l
i
m
a
ge
d
a
t
a
s
e
t
e
nc
o
m
p
a
s
s
i
ng
s
m
i
l
i
ng
a
nd
n
e
ut
r
a
l
f
a
c
e
s
to
i
de
n
t
i
f
y
fa
c
i
a
l
f
a
ul
t
s
t
h
a
t
can
be
a
us
u
a
l
s
y
m
pt
om
of
s
t
rok
e
.
T
he
fa
c
i
a
l
i
m
a
ge
da
t
a
s
e
t
c
on
t
a
i
ns
f
a
c
e
i
m
a
ge
s
of
s
t
ro
ke
p
a
t
i
e
n
t
s
a
nd
no
rm
a
l
s
ubj
e
c
t
s
.
T
h
i
s
a
dd
e
d
da
t
a
s
e
t
e
nc
o
m
p
a
s
s
e
s
a
s
e
t
of
s
m
i
l
i
ng
a
nd
n
e
ut
r
a
l
f
a
c
i
a
l
i
m
a
ge
s
c
r
e
a
t
e
d
fr
om
fr
e
e
l
y
a
va
i
l
a
bl
e
d
a
t
a
s
e
t
s
t
ha
t
a
re
a
ugm
e
nt
e
d
to
d
e
ve
l
op
t
w
o
furt
h
e
r
s
m
i
l
i
ng
i
m
a
g
e
s
at
e
i
gh
t
a
g
e
grou
ps
.
G
o
m
e
s
et
al
.
[17]
pre
s
e
nt
e
d
to
c
o
nc
e
nt
r
a
t
e
on
a
na
l
y
z
i
ng
t
hi
s
a
s
ym
m
e
t
ry
u
t
i
l
i
z
i
ng
a
DL
t
e
c
hn
i
que
w
i
t
hout
t
rus
t
i
ng
m
a
nu
a
l
c
o
m
put
a
t
i
ons
,
prop
os
i
ng
t
h
e
f
a
c
i
a
l
poi
nt
gra
p
hs
(F
P
G
)
m
e
t
hod
,
a
ne
w
t
e
c
hni
que
t
h
a
t
s
urp
a
s
s
e
s
in
s
t
udy
i
ng
ge
o
m
e
t
ri
c
a
l
d
a
t
a
a
nd
e
f
fi
c
i
e
nt
l
y
m
a
n
a
gi
n
g
di
ffe
r
e
n
c
e
s
a
ft
e
r
t
he
pos
s
i
b
i
l
i
t
y
of
ha
ndc
r
a
f
t
e
d
c
o
m
pu
t
a
t
i
ons
.
FPG
e
na
b
l
e
s
t
he
t
e
c
hn
i
qu
e
to
e
f
fi
c
i
e
nt
l
y
i
de
n
t
i
fy
t
he
fa
c
i
a
l
di
s
a
bi
l
i
t
y
c
a
us
e
d
by
a
s
t
rok
e
by
u
t
i
l
i
z
i
ng
v
i
de
o
da
t
a
.
3.
TH
E
P
R
O
P
O
S
ED
M
O
D
EL
T
hi
s
w
o
rk
p
re
s
e
n
t
s
a
n
e
w
D
E
T
L
M
-
ASDFPI
t
e
c
hn
i
qu
e
.
T
h
e
m
a
j
or
i
nt
e
nt
i
on
of
t
he
D
E
T
L
M
-
ASDFPI
t
e
c
hni
q
ue
is
to
i
de
n
t
i
fy
s
t
ro
ke
s
in
FP
i
m
a
g
e
s
prof
i
c
i
e
n
t
l
y.
T
h
e
D
E
T
L
M
-
ASDFPI
t
e
c
h
ni
qu
e
h
a
s
d
a
t
a
pre
pa
ra
t
i
on
a
nd
pre
-
proc
e
s
s
i
ng
,
a
fe
a
t
ur
e
e
x
t
ra
c
t
or
,
e
ns
e
m
bl
e
c
l
a
s
s
i
f
i
c
a
t
i
on
pro
c
e
s
s
e
s
,
a
nd
hyp
e
rpa
ra
m
e
t
e
r
t
uni
n
g
to
a
c
c
om
p
l
i
s
h
t
ha
t
.
F
i
g
ur
e
1
por
t
ra
ys
t
he
c
om
pl
e
t
e
proc
e
s
s
of
t
he
D
E
T
L
M
-
ASDFPI
m
od
e
l
.
S
t
ro
ke
di
a
gn
os
i
s
is
c
o
m
pl
i
c
a
t
e
d
by
F
P
,
w
h
i
c
h
a
ffe
c
t
s
f
a
c
e
m
us
c
l
e
s
a
nd
p
rodu
c
e
s
a
s
y
m
m
e
t
ry.
Ins
uf
fi
c
i
e
nt
d
a
t
a
s
e
t
s
re
s
t
ri
c
t
di
a
gnos
t
i
c
s
us
i
ng
ML
a
n
d
D
L
.
A
ne
w
de
e
p
e
ns
e
m
b
l
e
t
r
a
ns
f
e
r
l
e
a
rni
n
g
a
ppro
a
c
h
for
s
t
rok
e
d
i
a
g
nos
i
s
us
i
ng
f
a
c
i
a
l
p
a
r
a
l
ys
i
s
i
m
a
g
i
ng
is
p
ropos
e
d
in
t
hi
s
pa
p
e
r.
P
re
-
t
ra
i
ne
d
m
ode
l
s
l
ow
e
r
e
dg
e
de
v
i
c
e
c
o
m
put
a
t
i
ona
l
e
xpe
ns
e
s
.
T
he
fr
a
m
e
w
ork
i
nc
l
ud
e
s
d
a
t
a
ga
t
he
ri
ng
,
pi
c
t
u
r
e
r
e
s
c
a
l
i
ng
,
a
nd
d
e
e
p
c
a
ps
u
l
e
n
e
t
w
o
rk
fe
a
t
ur
e
e
xt
r
a
c
t
i
on
.
A
s
t
roke
d
e
t
e
c
t
i
on
e
ns
e
m
bl
e
t
r
a
ns
fe
r
l
e
a
rni
ng
m
ode
l
us
e
s
G
RU
,
CN
N
,
a
nd
s
t
a
c
k
e
d
s
pa
rs
e
a
ut
o
-
e
nc
o
de
r
c
l
a
s
s
i
f
i
e
rs
.
H
i
p
popo
t
a
m
us
op
t
i
m
i
z
a
t
i
on
op
t
i
m
i
z
e
s
pa
r
a
m
e
t
e
rs
.
On
M
E
E
I
a
nd
YFP
be
nc
h
m
a
rk
da
t
a
s
e
t
s
,
D
E
T
L
M
-
ASDFPI
out
p
e
rfor
m
s
c
u
rre
n
t
t
e
c
hni
q
ue
s
w
i
t
h
9
7.
0
6%
a
c
c
ura
c
y
.
T
hi
s
m
e
t
hod
a
ddr
e
s
s
e
s
da
t
a
s
e
t
a
nd
c
o
m
put
a
t
i
on
a
l
i
s
s
ue
s
w
hi
l
e
i
m
prov
i
ng
s
t
rok
e
di
a
gnos
i
s
us
i
ng
FP
pi
c
t
ur
e
s
.
3.
1
.
D
ata
p
r
e
p
a
r
at
i
on
an
d
p
r
e
-
p
r
o
c
e
s
s
i
n
g
D
a
t
a
pre
-
pro
c
e
s
s
i
ng
is
v
i
t
a
l
in
t
he
c
o
m
bi
n
e
d
da
t
a
s
e
t
to
i
n
c
re
a
s
e
qu
a
l
i
t
y
by
e
x
e
c
ut
i
ng
i
m
a
ge
pre
-
proc
e
s
s
i
ng
a
n
d
a
ug
m
e
nt
i
ng
t
a
s
ks
[1
8]
.
T
he
s
e
c
ro
ppe
d
i
m
a
ge
s
in
t
he
o
rga
n
i
z
e
d
d
a
t
a
s
e
t
w
e
re
r
e
s
c
a
l
e
d
to
norm
a
l
d
i
m
e
ns
i
ons
of
224
×
224
fo
r
i
npu
t
to
t
h
e
TL
m
e
t
h
od.
A
ddi
t
i
ona
l
l
y
,
t
he
s
e
a
r
e
re
s
i
z
e
d
w
i
t
hi
n
t
he
i
nt
e
rv
a
l
(0
–
1)
by
s
e
p
a
ra
t
i
ng
e
a
c
h
pi
xe
l
v
a
l
ue
by
22
5.
R
e
s
c
a
l
i
n
g
ha
s
be
e
n
a
ppl
i
e
d
to
nor
m
a
l
i
z
e
t
he
i
m
a
ge
d
a
t
a
in
t
h
e
norm
a
l
va
r
i
e
t
y
of
(0
–
1)
.
Ce
rt
a
i
n
c
l
a
s
s
e
s
ge
t
m
or
e
i
m
a
g
e
s
t
h
a
n
ot
h
e
rs
,
w
hi
l
e
s
o
m
e
i
m
a
g
e
s
w
i
t
h
i
n
t
h
e
da
t
a
s
e
t
s
a
re
no
i
s
y
a
nd
c
h
a
r
a
c
t
e
r
i
z
e
i
ns
uff
i
c
i
e
n
t
d
a
t
a
a
bo
ut
FP.
In
i
t
i
a
l
l
y,
no
i
s
e
i
m
a
ge
s
w
e
r
e
e
xt
r
a
c
t
e
d
f
rom
t
he
da
t
a
s
e
t
.
A
dde
d
i
m
a
ge
s
ha
ve
b
e
e
n
m
a
de
w
i
t
h
d
a
t
a
a
ug
m
e
nt
a
t
i
o
n
m
ode
l
s
to
a
t
t
a
i
n
t
h
e
offs
e
t
be
t
w
e
e
n
e
a
c
h
c
l
a
s
s
.
Conve
n
t
i
o
na
l
i
m
a
g
e
-
a
ugm
e
nt
i
ng
m
od
e
l
s
l
i
ke
rot
a
t
i
on,
z
oo
m
i
ng
,
f
l
i
pp
i
ng
,
a
n
d
s
he
a
ri
ng
a
r
e
a
pp
l
i
e
d
for
da
t
a
a
ugm
e
n
t
a
t
i
on
.
T
he
g
oa
l
is
to
a
c
hi
e
v
e
s
i
m
i
l
a
r
i
m
a
ge
c
oun
t
s
in
e
ve
ry
FP
c
l
a
s
s
to
i
m
pr
ove
t
he
t
r
a
i
n
i
ng
d
a
t
a
di
ve
rs
i
t
i
e
s
a
n
d
e
nh
a
n
c
e
t
he
g
e
n
e
ra
l
i
z
a
b
i
l
i
t
y
of
t
h
e
a
l
gor
i
t
h
m
.
T
hi
s
a
ugm
e
n
t
a
t
i
on
p
ha
s
e
is
c
r
i
t
i
c
a
l
in
m
a
ki
n
g
a
hi
ghe
r
-
qu
a
l
i
t
y
d
a
t
a
s
e
t
to
t
r
a
i
n
a
pre
c
i
s
e
FP
c
l
a
s
s
i
fi
c
a
t
i
o
n
m
e
t
hod
.
E
i
t
h
e
r
i
m
a
g
e
a
ugm
e
nt
a
t
i
on
or
p
re
-
proc
e
s
s
i
ng
ope
r
a
t
i
ons
a
r
e
c
a
rr
i
e
d
ou
t
w
i
t
h
t
he
K
e
r
a
s
l
i
br
a
ry
I
m
a
g
e
D
a
t
a
G
e
ne
r
a
t
o
r
c
l
a
s
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
A
rt
i
f
I
nt
e
l
l
IS
S
N
:
2252
-
8938
F
ac
i
al
par
a
l
y
s
i
s
i
m
ag
e
anal
y
s
i
s
f
or
s
t
r
ok
e
d
e
t
e
c
t
i
o
n
us
i
ng
de
e
p
e
ns
e
m
b
l
e
…
(
Ki
r
u
t
hi
ga
Su
br
am
ani
y
an
)
4077
F
i
g
ure
1.
O
v
e
ra
l
l
pro
c
e
s
s
of
D
E
T
L
M
-
ASDFPI
m
e
t
hodo
l
ogy
3.
2
.
D
C
ap
s
N
e
t
-
b
as
e
d
fe
a
tu
r
e
e
xtr
ac
t
i
on
Be
s
i
d
e
s
,
t
h
e
D
C
a
ps
N
e
t
m
ode
l
is
e
m
p
l
oye
d
for
t
h
e
fe
a
t
ur
e
e
xt
r
a
c
t
i
on
pro
c
e
s
s
to
l
e
a
rn
c
o
m
pl
e
x
fe
a
t
ure
s
fro
m
t
he
pre
-
proc
e
s
s
e
d
da
t
a
[19]
.
T
hi
s
m
od
e
l
gi
v
e
s
a
s
i
gn
i
fi
c
a
nt
a
dv
a
nt
a
g
e
in
f
e
a
t
ure
e
x
t
ra
c
t
i
on
t
a
s
ks
be
c
a
us
e
it
can
c
a
pt
ur
e
s
pa
t
i
a
l
h
i
e
r
a
r
c
hi
e
s
a
nd
pre
s
e
rve
p
a
r
t
-
w
hol
e
r
e
l
a
t
i
ons
hi
ps
in
d
a
t
a
.
U
n
l
i
k
e
c
onve
nt
i
ona
l
CN
N
s
,
w
h
i
c
h
fa
c
e
d
i
ffi
c
ul
t
y
ha
n
dl
i
ng
vi
e
w
poi
n
t
di
s
c
r
e
p
a
nc
i
e
s
a
nd
pos
e
c
h
a
ng
e
s
,
D
Ca
ps
N
e
t
ut
i
l
i
z
e
s
c
a
ps
ul
e
s
groups
of
n
e
uro
ns
t
h
a
t
e
n
c
ode
fe
a
t
ur
e
s
'
pre
s
e
nc
e
a
nd
pos
e
i
n
form
a
t
i
on.
T
h
i
s
a
l
l
ow
s
t
he
ne
t
w
ork
to
r
e
c
og
ni
z
e
obj
e
c
t
s
or
p
a
t
t
e
rns
m
or
e
r
obus
t
l
y,
e
v
e
n
in
o
c
c
l
us
i
o
ns
or
t
ra
n
s
form
a
t
i
o
ns
.
A
dd
i
t
i
on
a
l
l
y,
D
Ca
ps
N
e
t
’s
d
yna
m
i
c
rout
i
n
g
m
e
c
h
a
n
i
s
m
b
e
t
w
e
e
n
c
a
ps
u
l
e
s
e
na
b
l
e
s
e
ffe
c
t
u
a
l
d
a
t
a
pro
pa
g
a
t
i
on,
re
s
ul
t
i
n
g
in
m
or
e
a
c
c
ur
a
t
e
a
n
d
c
om
p
a
c
t
re
p
re
s
e
nt
a
t
i
o
ns
.
By
c
hoos
i
ng
D
C
a
ps
N
e
t
o
ve
r
c
onv
e
nt
i
ona
l
m
e
t
ho
ds
,
t
h
e
m
e
t
ho
dol
og
y
c
a
n
i
m
p
rove
ge
ne
r
a
l
i
z
a
t
i
on
,
m
i
t
i
g
a
t
e
t
he
re
qu
i
r
e
m
e
nt
for
e
x
t
e
ns
i
ve
da
t
a
a
ugm
e
nt
a
t
i
on,
a
n
d
i
m
pr
ove
t
h
e
i
nt
e
rpr
e
t
a
bi
l
i
t
y
of
l
e
a
rne
d
fe
a
t
ure
s
.
T
hi
s
m
a
ke
s
it
pa
r
t
i
c
ul
a
rl
y
e
ffe
c
t
u
a
l
for
c
o
m
pl
e
x
t
a
s
ks
s
u
c
h
as
i
m
a
ge
c
l
a
s
s
i
f
i
c
a
t
i
on
,
w
h
e
re
s
pa
t
i
a
l
re
l
a
t
i
o
ns
hi
ps
pl
a
y
a
s
i
gni
fi
c
a
n
t
r
ol
e
.
F
i
g
ur
e
2
i
l
l
us
t
ra
t
e
s
t
he
ov
e
r
a
l
l
s
t
ru
c
t
ure
of
t
h
e
D
Ca
ps
N
e
t
m
od
e
l
.
T
he
D
Ca
ps
N
e
t
is
ut
i
l
i
z
e
d
to
pr
oc
e
s
s
t
h
e
m
e
rge
d
v
a
ri
a
t
i
on
a
l
i
m
a
g
e
.
A
c
c
ord
i
ng
to
t
h
e
c
ons
i
s
t
e
n
t
s
a
m
pl
e
s
,
t
he
y
can
l
a
s
t
l
y
g
a
i
n
an
e
xp
e
rt
n
e
t
w
or
k.
T
he
D
C
a
ps
N
e
t
for
a
ny
va
r
i
a
t
i
on
r
e
c
ogni
t
i
o
n
is
a
c
hi
e
ve
d
in
t
he
s
u
bs
e
qu
e
nt
pha
s
e
s
:
i
)
C
hoos
e
dua
l
×
s
a
m
p
l
e
s
,
re
l
a
t
e
t
h
e
m
d
i
re
c
t
l
y
i
nt
o
a
×
2
di
m
e
ns
i
on
,
a
nd
ut
i
l
i
z
e
t
h
e
s
e
f
or
t
h
e
ne
t
w
ork
i
nput
;
i
i
)
S
e
t
t
h
e
i
np
ut
w
i
t
hi
n
t
h
e
Conv
l
a
y
e
r,
w
i
t
h
s
e
ve
r
a
l
c
onv
ol
u
t
i
on
a
l
k
e
rn
e
l
s
,
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8938
Int
J
A
rt
i
f
I
nt
e
l
l
,
V
ol
.
1
4,
N
o.
5
,
O
c
t
o
be
r
2025
:
4074
-
4
089
4078
to
re
m
ove
v
a
r
i
ous
e
a
s
y
f
e
a
t
ur
e
i
nfor
m
a
t
i
on.
T
he
s
e
m
a
i
n
c
a
p
s
l
a
ye
r
a
ddi
t
i
o
na
l
l
y
c
hoos
e
s
t
h
e
re
m
ov
e
d
f
e
a
t
ure
da
t
a
a
n
d
m
e
rg
e
s
t
h
e
f
e
a
t
ure
da
t
a
i
nt
o
ve
c
t
ors
.
T
hi
s
di
g
i
t
c
a
p
s
l
a
y
e
r
s
t
a
nda
rdi
z
e
s
t
hi
s
v
e
c
t
or
a
nd
c
a
t
e
g
ori
z
e
s
it
i
nt
o
a
gro
up
of
ve
c
t
ors
;
i
i
i
)
R
e
s
ha
p
i
ng
t
hi
s
ve
c
t
or
w
i
t
h
i
n
t
he
1D
ve
c
t
ors
a
nd
m
od
i
fyi
n
g
it
to
i
n
c
l
ud
e
v
a
ri
o
us
i
m
a
ge
bl
oc
ks
of
a
s
e
l
e
c
t
e
d
di
m
e
ns
i
on.
F
or
m
e
r
l
y
,
t
he
y
a
re
i
nput
t
e
d
w
i
t
hi
n
t
he
s
ys
t
e
m
in
t
he
s
a
m
e
w
a
y
as
e
a
r
l
i
e
r
;
a
nd
i
v
)
Ca
l
c
u
l
a
t
e
t
h
e
L
2
nor
m
f
rom
t
he
ve
c
t
o
rs
to
c
l
a
s
s
i
fy
a
nd
g
e
t
t
he
l
a
s
t
c
l
a
s
s
i
fi
e
r
out
c
o
m
e
s
.
F
i
g
ure
2.
O
v
e
ra
l
l
s
t
ru
c
t
u
re
of
D
Ca
ps
N
e
t
m
e
t
ho
dol
o
gy
T
he
ne
u
rons
in
e
ve
r
y
l
a
y
e
r
w
e
re
s
e
pa
ra
t
e
d
i
n
t
o
c
l
us
t
e
rs
,
s
uc
h
as
c
a
ps
u
l
e
s
.
T
he
c
on
ve
n
t
i
o
na
l
n
e
uron
out
pu
t
is
r
e
s
ha
p
e
d
a
nd
i
m
prov
e
d
l
i
ke
a
ve
c
t
or
.
T
he
pa
t
h
c
o
ns
i
s
t
e
n
c
i
e
s
m
o
de
l
m
a
i
nt
a
i
ns
t
he
e
nt
i
t
y'
s
pos
i
t
i
on
i
nfor
m
a
t
i
on
a
nd
a
ddi
t
i
o
na
l
i
nfor
m
a
t
i
on.
T
he
t
ra
i
n
i
ng
of
t
he
ne
t
w
ork
h
a
s
be
e
n
c
a
l
c
ul
a
t
e
d
by
e
m
p
l
oy
i
ng
di
ffe
r
e
n
t
fu
nc
t
i
ons
c
o
nve
n
t
i
o
na
l
l
y
by
ut
i
l
i
z
i
ng
t
he
t
a
s
k
(1)
.
=
ma
x
(
0
,
+
−
|
|
|
|
)
2
+
(
1
−
)
ma
x
(
0
,
|
|
|
|
−
−
)
2
(1)
On
t
he
o
t
he
r
ha
n
d,
s
i
g
ni
f
i
e
s
a
k
i
nd
of
c
l
a
s
s
i
fi
c
a
t
i
o
n,
a
nd
de
not
e
s
a
p
a
ra
m
e
t
e
r.
If
oc
c
urs
,
is
1,
or
0
.
+
=
0
.
9
,
−
=
0
.
1
re
p
re
s
e
nt
s
m
i
s
s
i
ng
c
a
s
e
s
in
t
he
c
urre
nt
c
l
a
s
s
i
f
i
c
a
t
i
on
c
on
di
t
i
ons
.
is
n
a
m
e
d
t
h
e
m
a
rg
i
n
a
l
l
os
s
.
In
Ca
ps
N
e
t
,
t
he
r
e
c
o
ns
t
ru
c
t
i
on
l
os
s
is
i
n
t
e
gra
t
e
d
w
i
t
h
t
h
e
m
a
rgi
n
l
os
s
to
i
m
pro
ve
c
l
a
s
s
i
f
i
c
a
t
i
on
a
c
c
ura
c
y
a
nd
l
e
a
rn
e
d
re
pr
e
s
e
n
t
a
t
i
o
ns
'
qua
l
i
t
y
.
T
h
e
l
os
s
func
t
i
on
e
nf
orc
e
s
m
a
rg
i
n
-
b
a
s
e
d
re
g
ul
a
ri
z
a
t
i
on
,
prom
o
t
i
ng
i
n
t
r
a
-
c
l
a
s
s
c
om
pa
c
t
n
e
s
s
a
n
d
i
nt
e
r
-
c
l
a
s
s
s
e
pa
r
a
t
i
on,
w
i
t
h
t
h
e
hype
rpa
r
a
m
e
t
e
r
b
a
l
a
nc
i
ng
t
h
e
s
e
t
e
r
m
s
for
e
nha
nc
e
d
m
od
e
l
pe
rfor
m
a
n
c
e
in
t
a
s
ks
l
i
k
e
m
e
t
r
i
c
l
e
a
rni
n
g.
3.
3
.
En
s
e
mb
l
e
c
l
as
s
i
fi
c
at
i
on
p
r
oc
e
s
s
An
e
ns
e
m
b
l
e
TL
m
ode
l
i
nvo
l
ve
s
t
hr
e
e
c
l
a
s
s
i
f
i
e
rs
f
or
t
he
s
t
ro
ke
d
e
t
e
c
t
i
on
pro
c
e
s
s
:
G
RU
,
D
CN
N
,
a
nd
S
S
A
E
.
An
e
ns
e
m
bl
e
TL
m
o
de
l
i
n
t
e
gr
a
t
i
ng
G
RU
,
D
CN
N
,
a
nd
SSAE
pr
e
s
e
n
t
s
a
ro
bus
t
a
ppro
a
c
h
to
s
t
ro
ke
de
t
e
c
t
i
on
by
ut
i
l
i
z
i
ng
t
he
m
e
ri
t
s
of
e
ve
ry
m
ode
l
t
ype
.
G
RU
is
h
i
gh
l
y
e
ffi
c
i
e
nt
for
c
a
pt
uri
n
g
s
e
q
ue
n
t
i
a
l
de
pe
nde
n
c
i
e
s
in
t
i
m
e
-
s
e
r
i
e
s
da
t
a
,
s
uc
h
as
E
E
G
s
i
gn
a
l
s
or
pa
t
i
e
nt
h
i
s
t
or
y,
a
l
l
ow
i
n
g
it
to
d
e
t
e
c
t
t
e
m
pora
l
pa
t
t
e
rns
c
ru
c
i
a
l
for
s
t
rok
e
d
e
t
e
c
t
i
on
.
D
CN
N
ou
t
pe
rf
orm
s
in
e
xt
r
a
c
t
i
ng
s
pa
t
i
a
l
fe
a
t
ur
e
s
fr
om
i
m
a
g
e
s
,
s
uc
h
as
M
RI
or
CT
s
c
a
ns
,
m
a
ki
n
g
it
i
d
e
a
l
for
de
t
e
c
t
i
ng
s
t
roke
-
re
l
a
t
e
d
a
no
m
a
l
i
e
s
in
m
e
di
c
a
l
i
m
a
gi
ng
.
On
t
he
c
o
nt
r
a
ry,
SSAE
is
prof
i
c
i
e
n
t
in
l
e
a
r
ni
ng
s
p
a
rs
e
,
h
i
gh
-
l
e
v
e
l
r
e
pr
e
s
e
nt
a
t
i
ons
fr
om
uns
t
ruc
t
ur
e
d
da
t
a
,
w
h
i
c
h
a
s
s
i
s
t
s
in
de
t
e
c
t
i
ng
s
ubt
l
e
f
e
a
t
ure
s
a
nd
p
a
t
t
e
rns
t
h
a
t
m
a
y
not
be
e
a
s
i
l
y
di
s
c
e
rni
b
l
e
.
T
h
e
e
ns
e
m
bl
e
a
ppr
oa
c
h
i
m
p
rove
s
t
h
e
m
ode
l
'
s
pe
r
form
a
n
c
e
by
i
n
c
orpor
a
t
i
ng
t
h
e
c
om
pl
e
m
e
nt
a
r
y
m
e
ri
t
s
of
t
h
e
s
e
di
ve
rs
e
m
ode
l
s
,
e
n
ha
n
c
i
n
g
a
c
c
ura
c
y
,
robus
t
ne
s
s
,
a
nd
g
e
ne
r
a
l
i
z
a
t
i
on
r
e
l
a
t
e
d
to
s
i
n
gl
e
-
m
ode
l
m
e
t
h
ods
.
T
hi
s
i
nt
e
gra
t
i
o
n
c
onf
i
r
m
s
t
ha
t
di
ve
rs
e
f
a
c
e
t
s
of
t
h
e
s
t
rok
e
d
e
t
e
c
t
i
o
n
pr
oc
e
s
s
,
s
u
c
h
as
t
e
m
p
ora
l
,
s
p
a
t
i
a
l
,
a
nd
a
bs
t
r
a
c
t
fe
a
t
ur
e
e
x
t
ra
c
t
i
on,
a
re
opt
i
m
a
l
l
y
a
ddre
s
s
e
d.
3.
3
.
1.
G
R
U
c
l
as
s
i
f
i
e
r
G
RU
s
a
r
e
an
r
e
c
urre
n
t
ne
ura
l
ne
t
w
ork
(
RN
N
)
s
t
r
uc
t
ure
a
ppl
i
e
d
to
pr
oc
e
s
s
s
e
que
n
t
i
a
l
d
a
t
a
[20]
.
G
RU
s
w
e
re
i
ni
t
i
a
l
l
y
pr
e
s
e
n
t
e
d
as
an
e
a
s
i
e
r
s
ubs
t
i
t
ut
i
on
fo
r
t
h
e
m
ore
c
o
m
pos
i
t
e
l
on
g
s
hort
-
t
e
rm
m
e
m
o
ry
(
L
S
T
M
)
s
ys
t
e
m
.
G
RU
s
w
e
re
i
n
t
e
n
de
d
to
t
a
c
kl
e
t
he
p
robl
e
m
of
va
n
i
s
hi
ng
gra
di
e
nt
s
t
ha
t
m
a
y
oc
c
ur
in
nor
m
a
l
RN
N
s
.
If
t
he
l
os
s
f
unc
t
i
o
n
gra
d
i
e
n
t
be
c
o
m
e
s
i
n
a
d
e
qua
t
e
,
u
pgra
d
e
t
h
e
ne
t
w
ork
pa
ra
m
e
t
e
rs
e
ffi
c
i
e
nt
l
y.
T
he
G
RU
s
t
ru
c
t
ure
c
ont
a
i
ns
g
a
t
e
m
e
c
ha
n
i
s
m
s
c
o
nt
ro
l
l
i
ng
t
he
s
ys
t
e
m
'
s
i
nf
orm
a
t
i
on
f
l
ow
.
T
he
g
a
t
e
,
us
ua
l
l
y
an
upda
t
e
g
a
t
e
a
nd
a
re
s
e
t
g
a
t
e
,
c
ont
r
ol
s
w
ha
t
a
m
oun
t
of
t
h
e
pr
e
c
e
di
ng
c
ond
i
t
i
on
h
a
s
be
e
n
m
a
i
n
t
a
i
ne
d
a
nd
w
h
a
t
a
m
ou
nt
of
or
i
gi
na
l
i
nfor
m
a
t
i
o
n
is
i
n
c
orpor
a
t
e
d
w
i
t
hi
n
t
he
pre
s
e
n
t
c
on
di
t
i
on
.
T
h
e
re
s
e
t
g
a
t
e
c
on
t
rol
s
w
h
i
c
h
c
om
p
one
n
t
s
of
t
he
pri
o
r
c
on
di
t
i
on
m
us
t
be
f
orgo
t
t
e
n.
In
c
o
nt
ra
s
t
,
t
h
e
upd
a
t
e
g
a
t
e
re
g
ul
a
t
e
s
t
h
e
a
m
oun
t
of
t
he
ori
gi
n
a
l
i
n
put
t
h
a
t
m
us
t
be
i
nc
l
ude
d
in
t
h
e
pre
s
e
nt
c
ond
i
t
i
on
.
T
he
c
r
i
t
i
c
a
l
f
e
a
t
ur
e
of
G
RU
s
is
t
h
e
i
r
c
a
pa
bi
l
i
t
y
to
d
i
s
c
a
r
d
or
m
a
i
n
t
a
i
n
d
a
t
a
fro
m
t
he
pr
e
c
e
di
n
g
t
i
m
e
ph
a
s
e
,
w
hi
c
h
m
a
k
e
s
t
h
e
m
e
ff
i
c
i
e
nt
f
or
d
e
m
o
ns
t
ra
t
i
n
g
l
onge
r
-
t
e
rm
d
e
p
e
nde
nc
i
e
s
in
s
e
que
nt
i
a
l
d
a
t
a
.
G
RU
s
w
e
re
e
x
pos
e
d
to
s
urp
a
s
s
ot
h
e
r
RN
N
t
e
c
hn
i
qu
e
s
t
hrou
gh
va
ri
o
us
t
a
s
ks
s
u
c
h
as
s
pe
e
c
h
r
e
c
og
ni
t
i
on
,
m
a
c
h
i
ne
t
r
a
ns
l
a
t
i
o
n,
a
nd
l
a
ngu
a
g
e
m
od
e
l
l
i
ng
.
T
h
e
y
w
e
r
e
u
t
i
l
i
z
e
d
in
di
ffe
r
e
n
t
a
ppl
i
c
a
t
i
ons
in
s
i
gn
a
l
pro
c
e
s
s
i
ng
a
nd
CV
,
i
n
c
l
u
di
ng
a
no
m
a
l
y
de
t
e
c
t
i
on
,
i
m
a
ge
c
a
p
t
i
on
i
ng
,
a
nd
m
us
i
c
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
A
rt
i
f
I
nt
e
l
l
IS
S
N
:
2252
-
8938
F
ac
i
al
par
a
l
y
s
i
s
i
m
ag
e
anal
y
s
i
s
f
or
s
t
r
ok
e
d
e
t
e
c
t
i
o
n
us
i
ng
de
e
p
e
ns
e
m
b
l
e
…
(
Ki
r
u
t
hi
ga
Su
br
am
ani
y
an
)
4079
ge
ne
r
a
t
i
on
.
G
RU
s
a
r
e
ve
rs
a
t
i
l
e
a
nd
pow
e
r
ful
de
vi
c
e
s
for
de
m
ons
t
r
a
t
i
ng
s
e
qu
e
nt
i
a
l
d
a
t
a
,
a
nd
t
h
e
i
r
e
f
fi
c
i
e
nc
y
a
nd
s
i
m
pl
i
c
i
t
y
m
a
k
e
t
h
e
m
a
s
t
a
nd
a
rd
s
e
l
e
c
t
i
on
in
t
he
DL
a
r
e
a
.
F
i
g
u
re
3
s
how
s
t
h
e
i
nfr
a
s
t
ru
c
t
u
re
of
t
he
G
RU
m
o
de
l
.
F
i
g
ure
3.
G
RU
a
r
c
h
i
t
e
c
t
ure
3.
3
.
2.
DCNN
C
l
as
s
i
fi
e
r
T
he
D
CN
N
ne
t
w
ork’s
fi
na
l
l
a
ye
r
is
S
oft
M
a
x,
us
i
ng
s
i
m
i
l
a
r
ne
t
w
ork
nod
e
c
o
unt
s
for
t
h
e
nu
m
be
r
of
t
ra
i
ni
ng
da
t
a
c
l
a
s
s
i
f
i
c
a
t
i
ons
[21
]
.
A
f
t
e
rw
a
rds
,
t
he
t
ra
i
ni
ng
w
a
s
fi
n
i
s
he
d
,
a
nd
t
he
fi
n
a
l
l
a
ye
r
ou
t
put
of
t
he
ne
t
w
ork
w
a
s
c
h
os
e
n
to
re
m
ov
e
t
h
e
i
m
a
ge
f
e
a
t
ur
e
s
.
T
ypi
c
a
l
l
y
,
t
h
e
gr
e
a
t
e
r
t
he
ne
t
w
ork
a
rc
h
i
t
e
c
t
ure
,
t
he
m
ul
t
i
p
l
e
n
e
t
w
or
k
l
a
ye
rs
,
a
n
d
t
h
e
a
d
di
t
i
on
a
l
p
a
r
a
m
e
t
e
rs
to
be
s
t
udi
e
d,
t
he
l
a
r
ge
r
t
he
ove
r
fi
t
t
i
n
g
prob
a
bi
l
i
t
y
a
n
d
t
he
s
l
ow
e
s
t
t
ra
i
ni
ng
s
pe
e
d
.
H
ow
e
ve
r
,
t
h
e
t
ra
i
ni
ng
m
i
gh
t
n
ot
fu
nc
t
i
on
w
he
n
t
h
e
ne
t
w
ork
c
on
t
a
i
ns
m
i
n
i
m
a
l
l
a
y
e
rs
.
In
g
e
ne
ra
l
,
t
hi
s
w
ork
a
pp
l
i
e
s
a
ne
t
w
ork
f
ra
m
e
w
or
k
w
i
t
h
t
hr
e
e
c
onv
ol
u
t
i
on
l
a
ye
rs
.
F
i
g
ure
4
d
e
p
i
c
t
s
t
h
e
s
t
ruc
t
ure
of
t
h
e
D
CN
N
m
ode
l
.
F
i
g
ure
4
.
S
t
ruc
t
ure
of
D
CN
N
m
od
e
l
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8938
Int
J
A
rt
i
f
I
nt
e
l
l
,
V
ol
.
1
4,
N
o.
5
,
O
c
t
o
be
r
2025
:
4074
-
4
089
4080
T
he
D
CN
N
w
i
t
h
t
hre
e
c
o
nvol
u
t
i
o
na
l
l
a
ye
rs
h
a
s
be
e
n
a
p
pl
i
e
d.
Be
s
i
de
s
t
he
c
onvo
l
ut
i
on
a
l
l
a
y
e
r,
d
ua
l
ful
l
y
c
on
ne
c
t
e
d
(F
C)
l
a
ye
rs
e
xi
s
t
,
a
l
l
of
w
hi
c
h
can
be
a
c
c
o
m
pa
ni
e
d
by
t
h
e
p
ool
i
ng
l
a
y
e
r
(s
ub
-
s
a
m
pl
i
ng
l
a
y
e
r)
.
T
he
i
nput
i
m
a
ge
is
an
i
m
a
g
e
of
a
s
o
l
i
t
a
ry
t
a
rg
e
t
of
80
×
10
0
d
i
m
e
ns
i
ons
.
T
h
e
pri
m
a
r
y
D
CN
N
l
a
y
e
r
re
pre
s
e
nt
s
t
he
c
on
vol
u
t
i
o
na
l
l
a
y
e
r
r
e
l
a
t
e
d
to
t
h
e
i
npu
t
a
nd
is
r
e
pr
e
s
e
n
t
e
d
by
c
on
vol
u
t
i
o
na
l
l
a
y
e
r
.
T
he
i
np
ut
of
a
s
o
l
i
t
a
ry
t
a
rg
e
t
e
d
i
m
a
ge
can
be
c
onvo
l
ut
e
d
w
i
t
h
t
w
e
nt
y
-
four
f
i
l
t
e
r
s
to
pre
s
e
nt
t
w
e
n
t
y
-
fou
r
fe
a
t
ur
e
m
a
pp
i
ng
w
i
t
h
G
a
us
s
i
a
n
f
i
l
t
e
rs
.
T
h
i
s
ne
t
w
ork
u
t
i
l
i
z
e
s
du
a
l
t
ype
s
of
c
on
vol
ut
i
on
a
l
k
e
rn
e
l
s
of
v
a
ri
o
us
di
m
e
ns
i
ons
.
T
he
di
m
e
ns
i
ons
of
t
h
e
g
re
a
t
e
r
r
e
c
t
a
n
gl
e
bo
x
w
i
t
h
i
n
t
he
f
i
gur
e
a
r
e
5
×
5.
A
dd
i
t
i
ona
l
l
y
,
t
h
e
d
i
m
e
ns
i
ons
of
t
he
l
e
s
s
e
r
re
c
t
a
ng
l
e
b
ox
c
or
re
s
pon
d
to
t
he
d
i
m
e
ns
i
ons
of
t
he
c
onv
ol
u
t
i
ona
l
ke
r
ne
l
s
of
3
×
3,
a
n
d
t
h
e
di
m
e
ns
i
ons
of
t
h
e
c
onvo
l
ut
i
ona
l
ke
rn
e
l
of
t
h
e
s
e
l
a
ye
rs
a
r
e
5
×
5.
T
h
e
c
onv
ol
ut
i
ona
l
pha
s
e
d
i
m
e
ns
i
o
ns
a
r
e
one
.
F
or
t
he
c
onvo
l
ut
i
ona
l
fun
c
t
i
on
to
o
pe
r
a
t
e
by
t
h
e
pi
x
e
l
s
by
t
he
i
m
a
g
e
e
dg
e
,
t
he
y
f
i
l
l
e
d
in
t
w
o
-
p
i
x
e
l
w
i
d
t
hs
(pa
ds
)
on
e
ve
ry
i
m
a
ge
s
i
d
e
.
T
he
n
,
c
onvo
l
ut
i
on
a
l
l
a
ye
r
c
on
t
a
i
ns
t
w
e
nt
y
-
four
c
on
vol
u
t
i
o
na
l
s
a
m
pl
e
s
,
a
nd
e
v
e
ry
c
onvo
l
ut
i
on
ke
rn
e
l
ne
e
ds
5
×
6
p
a
r
a
m
e
t
e
rs
.
O
n
e
of
t
he
pa
r
a
m
e
t
e
rs
is
26
×
2
8,
so
t
h
e
t
o
t
a
l
pa
ra
m
e
t
e
r
c
oun
t
for
c
onv
ol
u
t
i
on
a
l
l
a
ye
r
is
26
×
28
.
T
he
S
oft
M
a
x
c
l
a
s
s
i
fi
e
r
ha
s
be
e
n
u
t
i
l
i
z
e
d
w
i
t
hi
n
t
h
e
o
ut
put
l
a
ye
r.
U
nl
i
ke
l
og
i
s
t
i
c
re
gr
e
s
s
i
on
,
w
h
i
c
h
c
a
n
ha
ndl
e
no
n
-
l
i
ne
a
r
b
i
na
ry
c
l
a
s
s
i
f
i
c
a
t
i
on
i
s
s
ue
s
,
S
oft
M
a
x
can
ha
nd
l
e
m
u
l
t
i
pl
e
c
l
a
s
s
i
f
i
c
a
t
i
on
d
i
ff
i
c
u
l
t
i
e
s
.
A
s
i
ngl
e
i
ns
t
a
n
c
e
can
re
l
a
t
e
to
a
c
l
a
s
s
;
c
l
a
s
s
e
s
a
r
e
e
q
ua
l
l
y
e
xc
l
u
s
i
ve
.
T
h
e
nod
e
c
ount
s
a
r
e
s
i
m
i
l
a
r
for
t
h
e
nu
m
be
r
of
tr
a
i
ni
ng
d
a
t
a
c
l
a
s
s
e
s
.
T
he
a
c
t
i
va
t
i
o
n
f
unc
t
i
ons
of
e
v
e
ry
FC
l
a
ye
r
a
nd
c
onvo
l
u
t
i
on
a
l
l
a
y
e
r
a
r
e
r
e
c
t
i
f
i
e
d
l
i
n
e
a
r
un
i
t
(
R
e
LU
)
.
T
he
f
unc
t
i
on
of
R
e
L
U
d
e
no
t
e
s
a
non
-
l
i
n
e
a
r
a
c
t
i
v
a
t
i
on
fu
nc
t
i
on,
a
nd
t
h
e
fu
nc
t
i
on
of
R
e
L
U
is
s
i
gni
f
i
c
a
nt
l
y
l
e
s
s
m
a
t
he
m
a
t
i
c
a
l
l
y
t
h
a
n
a
no
t
he
r
fu
nc
t
i
on
.
U
t
i
l
i
z
i
ng
t
h
e
f
unc
t
i
on
of
R
e
L
U
w
i
l
l
c
a
us
e
nu
m
e
rous
n
e
t
w
ork
p
oi
n
t
s
to
out
pu
t
z
e
ro
,
so
it
c
on
t
a
i
ns
a
pa
r
t
i
c
ul
a
r
s
e
l
f
-
c
ons
c
i
ous
n
e
s
s
r
e
s
ul
t
i
ng
in
ove
rfi
t
t
i
ng.
3.
3
.
3.
SSAE
c
l
as
s
i
f
i
e
r
An
SAE
is
a
s
t
a
nda
r
d
3
-
l
a
ye
r
A
N
N
,
w
hi
c
h
c
a
n
ut
i
l
i
z
e
uns
upe
rvi
s
e
d
l
e
a
rn
i
ng
to
a
c
qui
r
e
a
c
om
p
re
s
s
e
d
i
np
ut
d
a
t
a
c
od
e
[22
]
.
T
h
e
c
om
p
re
s
s
e
d
c
od
e
r
e
c
o
gni
z
e
s
t
h
e
d
i
m
e
ns
i
on
d
e
c
r
e
a
s
e
of
t
h
e
r
a
w
i
n
put
.
S
i
gni
f
i
c
a
nt
l
y,
t
h
e
AE
is
a
gr
e
a
t
f
e
a
t
ure
e
x
t
ra
c
t
i
on
f
or
t
he
DL
ne
t
w
o
rk.
F
i
g
ur
e
5
de
m
ons
t
r
a
t
e
s
t
he
a
rc
hi
t
e
c
t
ur
e
of
t
h
e
SSAE
m
od
e
l
.
F
i
g
ure
5.
A
r
c
hi
t
e
c
t
ur
e
of
t
h
e
SSAE
a
ppro
a
c
h
In
g
e
ne
r
a
l
,
an
SAE
is
bui
l
t
fro
m
e
nc
odi
n
g
a
n
d
d
e
c
o
di
ng
.
:
=
(
+
)
(2)
:
′
=
′
(
+
′
)
(3)
H
e
re
,
,
,
a
nd
’
,
’
,
'
de
not
e
t
h
e
w
e
i
ght
,
b
i
a
s
v
e
c
t
or
,
a
nd
a
c
t
i
va
t
i
o
n
fu
nc
t
i
on
of
t
h
e
e
n
c
od
e
r
a
nd
d
e
c
o
de
r
proc
e
s
s
.
∈
m
e
a
ns
an
i
np
ut
.
∈
re
f
e
rs
to
c
om
pre
s
s
e
d
c
od
e
.
′
∈
d
e
not
e
s
t
he
r
e
c
o
ns
t
ru
c
t
i
on
ve
c
t
or.
T
h
e
SAE
c
a
n
c
onve
r
t
an
i
npu
t
i
n
t
o
t
h
e
l
a
t
e
nt
v
a
ri
a
bl
e
a
nd
r
e
bu
i
l
d
′
ov
e
r
t
he
d
e
c
od
i
ng
.
T
he
r
e
for
e
,
t
h
e
obj
e
c
t
i
v
e
of
AE
is
to
foc
us
a
nd
’
,
w
hi
c
h
p
e
r
m
i
t
s
an
ou
t
put
of
de
c
od
i
ng
to
i
m
prov
e
t
he
n
e
w
i
npu
t
.
F
or
an
a
s
s
u
m
e
d
p
a
i
r
of
da
t
a
,
t
h
e
re
c
ons
t
ruc
t
i
on
e
rror
of
m
a
t
he
m
a
t
i
c
a
l
for
m
ul
a
t
i
on
as
m
e
n
t
i
on
e
d
i
n
(4)
.
(
,
′
)
=
‖
−
′
‖
2
=
‖
−
′
[
′
(
+
)
+
′
]
‖
2
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
A
rt
i
f
I
nt
e
l
l
IS
S
N
:
2252
-
8938
F
ac
i
al
par
a
l
y
s
i
s
i
m
ag
e
anal
y
s
i
s
f
or
s
t
r
ok
e
d
e
t
e
c
t
i
o
n
us
i
ng
de
e
p
e
ns
e
m
b
l
e
…
(
Ki
r
u
t
hi
ga
Su
br
am
ani
y
an
)
4081
T
he
s
t
a
c
k
e
d
s
e
l
f‐
e
nc
ode
r
s
ys
t
e
m
c
on
t
a
i
ns
n
um
e
rous
AE
l
a
y
e
rs
,
w
hi
c
h
e
m
pl
o
y
an
o
ut
pu
t
of
pr
e
vi
ous
AE
l
a
ye
rs
as
an
i
n
put
f
or
t
he
c
ons
e
que
n
t
on
e
.
I
d
e
not
e
t
he
n
um
be
r
of
s
a
m
p
l
e
s
f
or
a
c
o
m
pl
e
t
e
s
e
t
of
t
r
a
i
n
i
ng
s
a
m
pl
e
s
,
(
)
.
T
he
l
os
s
fun
c
t
i
ons
f
or
S
A
E
a
re
i
n
(5)
.
(
,
)
=
1
∑
[
=
1
(
)
(5)
H
e
re
,
de
not
e
s
a
s
u
m
of
s
a
m
pl
e
s
,
s
i
gni
f
i
e
s
t
he
e
qu
i
va
l
e
n
t
l
a
be
l
of
.
T
h
e
r
e
gul
a
ri
z
a
t
i
o
n
of
w
e
i
gh
t
a
t
t
e
nu
a
t
i
on
is
t
h
e
2
nd
t
e
r
m
e
m
pl
oye
d
for
pr
e
v
e
nt
i
ng
ov
e
r
-
fi
t
t
i
ng
of
t
h
e
m
e
t
hod
,
a
nd
re
fe
rs
to
w
e
i
gh
t
a
t
t
e
nu
a
t
i
on
co
-
e
ffi
c
i
e
nt
a
nd
is
ut
i
l
i
z
e
d
to
r
e
gul
a
t
e
t
he
r
e
l
a
t
i
v
e
w
e
i
ght
of
1
st
a
nd
2
nd
t
e
rm
s
.
a
nd
i
nd
i
c
a
t
e
t
he
t
o
t
a
l
a
m
ou
nt
of
n
e
urons
in
ℎ
ℎ
l
a
y
e
r,
a
nd
re
f
e
rs
to
t
he
c
on
ne
c
t
i
on
w
e
i
ght
a
m
ong
t
he
ne
ur
ons
in
dua
l
l
a
ye
rs
.
T
he
f
e
a
t
ure
'
s
s
pa
rs
i
t
y
is
a
t
t
a
i
ne
d
by
i
nc
l
ud
i
ng
w
ords
in
(
6),
a
n
d
t
h
e
c
o
m
pl
e
t
e
fe
a
t
u
re
e
xt
r
a
c
t
or
proc
e
dure
is
e
nha
n
c
e
d.
N
e
x
t
,
t
he
l
os
s
f
unc
t
i
o
n
of
s
t
a
nd
a
rd
s
t
a
c
k
e
d
S
A
E
(S
S
A
E
s
)
h
a
s
b
e
e
n
(7
)
.
(
,
)
=
(
,
)
+
∑
2
=
1
(
□
̂
j
)
(6)
(
̂
j
)
=
l
o
g
̂
j
+
(
1
−
)
l
o
g
1
−
1
−
̂
j
(7)
H
e
re
,
t
h
e
s
y
m
bo
l
s
c
orr
e
s
pond
i
ngl
y
re
pr
e
s
e
n
t
t
h
e
c
ons
t
a
n
t
of
s
pa
rs
e
n
e
s
s
a
nd
di
v
e
rg
e
nc
e
.
2
de
no
t
e
s
t
he
a
m
oun
t
of
h
i
dd
e
n
l
a
ye
r
ne
urons
.
̂
re
f
e
rs
to
t
h
e
v
a
l
u
e
of
m
e
a
n
a
c
t
i
va
t
i
o
n.
3.
4
.
H
O
A
-
b
as
e
d
p
ar
a
me
t
e
r
op
ti
mi
z
e
r
E
ve
n
t
u
a
l
l
y,
t
he
HOA
is
e
m
pl
oye
d
for
t
he
o
pt
i
m
a
l
p
a
ra
m
e
t
e
r
t
uni
n
g
of
t
he
t
hr
e
e
e
ns
e
m
bl
e
t
e
c
hni
qu
e
s
[23]
.
T
h
e
HOA
is
a
nov
e
l
m
e
t
a
he
uri
s
t
i
c
i
ns
p
i
re
d
by
t
h
e
b
e
h
a
vi
our
of
h
i
ppop
ot
a
m
us
e
s
in
t
h
e
i
r
n
a
t
ur
a
l
ha
b
i
t
a
t
.
It
is
pa
r
t
i
c
ul
a
rl
y
e
ffe
c
t
i
ve
in
pa
r
a
m
e
t
e
r
op
t
i
m
i
z
a
t
i
on
du
e
to
its
un
i
que
b
a
l
a
n
c
e
be
t
w
e
e
n
e
x
pl
or
a
t
i
on
a
nd
e
xpl
o
i
t
a
t
i
on
.
T
h
e
a
l
g
ori
t
hm
'
s
a
b
i
l
i
t
y
to
m
i
m
i
c
t
h
e
dyna
m
i
c
m
ove
m
e
nt
p
a
t
t
e
rns
of
hi
ppos
-
a
l
t
e
rn
a
t
i
ng
be
t
w
e
e
n
gra
z
i
ng
in
op
e
n
fi
e
l
ds
a
nd
d
i
vi
n
g
u
nde
rw
a
t
e
r
-
m
a
k
e
s
it
hi
g
hl
y
s
ui
t
e
d
for
n
a
vi
g
a
t
i
ng
c
o
m
p
l
e
x
,
hi
gh
-
di
m
e
ns
i
on
a
l
s
e
a
rc
h
s
pa
c
e
s
w
i
t
h
l
oc
a
l
a
nd
gl
o
ba
l
op
t
i
m
a
.
Com
p
a
re
d
to
t
r
a
d
i
t
i
ona
l
t
e
c
hni
que
s
l
i
k
e
ge
n
e
t
i
c
a
l
gor
i
t
h
m
s
or
pa
r
t
i
c
l
e
s
w
a
r
m
opt
i
m
i
z
a
t
i
on
(P
S
O
)
,
HOA
s
how
s
be
t
t
e
r
robus
t
ne
s
s
in
a
vo
i
di
ng
p
re
m
a
t
ure
c
onv
e
rge
nc
e
a
n
d
c
a
n
m
or
e
e
ffe
c
t
i
ve
l
y
h
a
ndl
e
m
ul
t
i
m
od
a
l
,
no
i
s
y,
a
nd
unc
e
rt
a
i
n
obj
e
c
t
i
v
e
f
unc
t
i
ons
.
I
t
s
s
i
m
pl
i
c
i
t
y
a
nd
re
l
a
t
i
v
e
l
y
f
e
w
e
r
c
ont
r
ol
pa
ra
m
e
t
e
rs
m
a
k
e
i
m
pl
e
m
e
n
t
i
n
g
a
nd
t
u
ni
ng
fo
r
d
i
v
e
rs
e
op
t
i
m
i
z
a
t
i
on
t
a
s
ks
e
a
s
i
e
r.
T
he
r
e
for
e
,
HOA
off
e
rs
an
a
p
pe
a
l
i
ng
a
l
t
e
rn
a
t
i
v
e
w
he
n
t
a
c
kl
i
ng
i
nt
r
i
c
a
t
e
a
nd
h
i
gh
-
d
i
m
e
ns
i
o
na
l
pa
ra
m
e
t
e
r
op
t
i
m
i
z
a
t
i
on
pr
obl
e
m
s
.
F
i
g
ure
6
s
how
s
t
he
s
t
e
ps
i
nvol
v
e
d
in
t
h
e
HOA
m
e
t
ho
dol
o
gy.
F
i
g
ure
6.
S
t
e
ps
i
nvo
l
v
e
d
in
t
h
e
HOA
a
ppr
oa
c
h
T
he
HOA
m
o
de
l
is
be
t
w
e
e
n
t
he
m
os
t
of
t
e
n
a
p
pl
i
e
d
a
n
d
a
dv
a
n
c
e
d
bi
o
-
i
ns
p
i
re
d
m
e
t
a
he
uri
s
t
i
c
opt
i
m
i
z
a
t
i
on
a
l
gori
t
h
m
s
for
s
ol
vi
ng
c
o
m
pos
i
t
e
opt
i
m
i
z
a
t
i
on
i
s
s
ue
s
.
T
hi
s
m
e
t
hod
ha
s
b
e
e
n
s
t
i
m
ul
a
t
e
d
by
t
h
e
i
r
he
rds
'
d
e
fe
n
c
e
m
e
c
ha
n
i
s
m
s
a
nd
s
oc
i
a
l
be
h
a
vi
our.
T
he
HOA
m
ode
l
c
on
t
a
i
ns
3
s
t
a
ge
s
,
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8938
Int
J
A
rt
i
f
I
nt
e
l
l
,
V
ol
.
1
4,
N
o.
5
,
O
c
t
o
be
r
2025
:
4074
-
4
089
4082
i)
P
ha
s
e
1:
hi
p
popot
a
m
us
e
s
upda
t
e
d
l
oc
a
t
i
on
i
n
t
h
e
pond
or
ri
v
e
r
T
hi
s
s
t
a
ge
t
a
rge
t
s
di
s
c
ove
r
i
ng
t
he
s
e
a
rc
h
s
p
a
c
e
c
orre
s
po
ndi
ng
to
how
h
i
ppos
m
ov
e
n
e
a
r
t
he
i
r
s
urroundi
n
gs
,
t
h
e
w
a
t
e
r.
H
e
nc
e
,
t
he
l
o
c
a
t
i
on
a
nd
m
ov
e
m
e
n
t
of
i
ndi
v
i
du
a
l
m
a
l
e
s
(
ℎ
)
,
f
e
m
a
l
e
s
,
a
nd
t
h
e
dom
i
na
n
t
hi
pp
o
(
ℎ
)
c
on
t
ro
l
t
h
e
e
xp
l
or
a
t
i
on
pro
c
e
s
s
.
In
(14)
,
t
h
e
ℎ
d
e
m
ons
t
ra
t
e
s
t
he
i
de
a
l
s
ol
ut
i
on
to
t
ra
v
e
l
l
i
ng
a
n
ot
h
e
r
e
nt
i
t
y
to
i
t
s
a
re
a
.
T
h
e
s
e
di
s
t
a
nc
e
s
a
re
,
in
s
e
q
ue
n
c
e
,
a
f
unc
t
i
on
of
not
j
us
t
t
he
dom
i
n
a
nt
hi
ppo
bu
t
a
l
s
o
a
r
bi
t
r
a
ry
v
e
c
t
ors
(
)
.
(
)
,
a
n
d
(
1
)
a
nd
an
i
n
t
e
g
e
r
(
1
)
s
i
gni
fy
i
ng
i
n
t
ri
ns
i
c
v
a
ri
a
bi
l
i
t
i
e
s
w
i
t
ne
s
s
e
d
in
e
xp
l
or
a
t
i
on.
ℎ
:
ℎ
=
+
1
.
(
ℎ
−
1
)
(8)
W
he
n
a
fe
m
a
l
e
or
m
a
l
e
hi
ppopo
t
a
m
us
'
s
l
oc
a
t
i
on
l
e
a
ds
to
a
l
a
r
ge
r
va
l
ue
of
t
he
ob
j
e
c
t
i
ve
func
t
i
o
n
t
ha
n
t
h
e
hi
ppo'
s
pre
s
e
nt
dom
i
na
nc
e
,
t
he
do
m
i
n
a
nt
l
o
c
a
t
i
on
is
s
u
bs
t
i
t
u
t
e
d
w
i
t
h
t
h
a
t
i
nd
i
vi
d
ua
l
l
o
c
a
t
i
on
.
T
he
s
e
m
e
c
ha
n
i
s
m
s
e
ns
ur
e
t
h
a
t
t
h
e
e
xp
l
ora
t
i
o
n
pr
oc
e
s
s
c
o
ns
t
a
nt
l
y
s
e
a
rc
h
e
s
fo
r
i
m
p
rove
d
s
o
l
ut
i
ons
.
ii)
P
ha
s
e
2:
t
h
e
d
e
f
e
nc
e
a
c
t
of
h
i
ppo
pot
a
m
us
a
g
a
i
ns
t
pre
d
a
t
o
rs
(
e
xpl
o
ra
t
i
on)
T
he
de
f
e
nc
e
m
e
t
hod
is
a
c
qui
r
e
d
if
t
h
e
r
e
is
s
o
m
e
t
hi
ng
to
de
fe
n
d
a
g
a
i
ns
t
;
f
or
e
xa
m
pl
e
,
t
h
e
fl
o
c
k
i
de
n
t
i
f
i
e
s
c
ro
c
odi
l
e
s
f
rom
t
h
e
N
i
l
e
.
Ins
t
a
n
t
de
fe
n
c
e
fro
m
t
h
e
hun
t
e
r
ha
s
b
e
e
n
a
s
s
u
m
e
d
,
t
oge
t
h
e
r
w
i
t
h
l
oud
voc
a
l
i
z
a
t
i
on
.
In
(9)
r
e
pr
e
s
e
nt
s
t
h
e
fa
s
t
t
ry
ne
a
r
t
he
da
ng
e
r.
T
h
e
a
rbi
t
ra
ry
pre
da
t
or'
s
m
ov
e
m
e
nt
(
∖
:
)
c
a
n
be
re
pre
s
e
nt
e
d
by
a
ra
n
dom
l
y
ge
n
e
ra
t
e
d
ve
c
t
or
(
⃗
8
)
ra
n
ge
b
e
t
w
e
e
n
(0
-
1)
i
ns
i
de
t
h
e
l
ow
e
r
(
)
a
nd
uppe
r
(
)
bounds
of
t
h
e
d
e
c
i
s
i
on
va
r
i
a
b
l
e
s
at
(
ℎ
)
.
M
ore
ov
e
r,
i
n
(1
0)
re
pre
s
e
nt
e
d
t
he
m
odi
f
i
c
a
t
i
on
in
t
h
e
di
s
t
a
n
c
e
be
t
w
e
e
n
hi
pp
os
a
nd
pre
d
a
t
o
rs
a
f
t
e
r
t
h
e
s
e
l
f
-
d
e
f
e
nc
e
m
e
c
h
a
ni
s
m
.
∖
∷
∖
:
∖
:
∖
≔
+
⃗
8
⋅
(
−
)
,
=
1
,
2
,
…
,
(9)
⃗
⃗
⃗
=
|
∖
:
∖
:
−
∖
|
(10)
iii)
P
ha
s
e
3:
hi
p
popot
a
m
us
e
s
c
a
pe
d
fro
m
t
h
e
pr
e
d
a
t
or
T
he
hi
pp
os
m
i
ght
e
s
c
a
p
e
by
ove
r
c
o
m
i
ng
pr
e
da
t
or
a
s
s
a
ul
t
s
or
c
ond
i
t
i
ons
in
w
hi
c
h
t
he
y
c
a
n'
t
i
nt
e
ns
i
fy
an
a
d
e
qu
a
t
e
de
fe
n
c
e
to
s
a
fe
r
r
e
gi
ons
.
A
c
c
ur
a
t
e
l
y,
t
he
s
e
s
t
o
c
ha
s
t
i
c
re
fu
gi
a
a
r
e
de
m
ons
t
ra
t
e
d
to
pr
e
t
e
nd
t
he
ra
ndo
m
ne
s
s
of
e
s
c
a
pi
ng
pa
t
hs
.
M
e
a
nw
h
i
l
e
,
an
ori
g
i
n
a
l
pos
i
t
i
on
of
fe
rs
an
i
m
prov
e
d
obj
e
c
t
i
v
e
fu
nc
t
i
on
va
l
ue
,
re
pre
s
e
nt
i
ng
a
s
upe
ri
or
s
ol
u
t
i
o
n.
T
h
e
s
e
c
oul
d
d
e
s
c
ri
be
how
a
hi
ppo
e
s
c
a
p
e
s
,
re
p
re
s
e
nt
i
ng
an
e
ff
e
c
t
i
v
e
e
s
c
a
pe
.
T
he
HO
m
od
e
l
is
an
a
dva
n
c
e
d
o
pt
i
m
i
z
a
t
i
o
n
a
l
go
ri
t
hm
i
ns
p
i
re
d
by
t
he
l
i
f
e
h
a
bi
t
s
a
nd
be
ha
v
i
ors
of
hi
ppos
,
s
u
c
h
as
t
he
i
r
de
f
e
ns
i
v
e
m
e
c
h
a
ni
s
m
s
a
nd
grou
p
dyn
a
m
i
c
s
.
It
e
ff
i
c
i
e
n
t
l
y
a
ddre
s
s
e
s
c
o
m
p
l
e
x
opt
i
m
i
z
a
t
i
on
prob
l
e
m
s
by
s
i
m
ul
a
t
i
ng
t
he
s
e
b
i
ol
o
gi
c
a
l
s
t
ra
t
e
gi
e
s
,
i
m
prov
i
ng
b
ot
h
t
h
e
rob
us
t
n
e
s
s
a
nd
a
da
p
t
a
b
i
l
i
t
y
of
t
he
s
e
a
rc
h
p
roc
e
s
s
.
T
h
rough
i
t
e
r
a
t
i
ve
s
t
a
g
e
s
,
t
h
e
m
od
e
l
r
e
fi
n
e
s
its
s
ol
u
t
i
o
ns
,
m
a
ki
ng
it
pa
rt
i
c
u
l
a
r
l
y
e
ff
e
c
t
i
v
e
in
t
a
c
kl
i
ng
c
h
a
l
l
e
n
gi
ng
op
t
i
m
i
z
a
t
i
on
t
a
s
ks
.
I
t
s
a
bi
l
i
t
y
to
b
a
l
a
n
c
e
e
xp
l
or
a
t
i
on
a
nd
e
xpl
o
i
t
a
t
i
on
e
nh
a
nc
e
s
p
e
rfor
m
a
nc
e
a
c
r
os
s
va
r
i
ous
dom
a
i
n
s
.
T
he
ne
x
t
is
t
he
fl
ow
c
ha
r
t
a
nd
ps
e
ud
oc
od
e
i
l
l
us
t
ra
t
i
ng
t
he
H
OA
m
od
e
l
a
s
s
h
ow
n
i
n
A
l
gor
i
t
h
m
1
.
A
l
gori
t
hm
1
.
P
s
e
udoc
ode
of
HOA
S
t
a
rt
D
e
s
c
ri
b
e
t
h
e
o
pt
i
m
i
z
a
t
i
o
n
p
robl
e
m
a
va
i
l
a
bl
e
.
E
s
t
a
b
l
i
s
h
i
ng
t
h
e
m
a
xi
m
a
l
i
t
e
ra
t
i
on
c
ount
s
re
pre
s
e
nt
e
d
as
“
i
t
”
a
nd
de
t
e
r
m
i
n
i
ng
t
h
e
num
be
r
of
hi
p
popot
a
m
us
e
s
m
e
n
t
i
on
e
d
as
“
N
”.
G
e
ne
r
a
t
i
ng
t
h
e
i
ni
t
i
a
l
hi
p
popot
a
m
us
po
pul
a
t
i
on
a
nd
e
va
l
ua
t
i
n
g
t
h
e
m
a
i
n
fun
c
t
i
ons
of
t
hi
s
pri
m
a
r
y
g
roup
.
F
or
=
1
,
it
U
pda
t
e
d
t
h
e
l
oc
a
t
i
on
of
t
h
e
do
m
i
na
n
t
h
i
ppos
by
t
h
e
s
t
a
nd
a
rd
of
m
a
i
n
fu
nc
t
i
on
v
a
l
u
e
c
ond
i
t
i
on
S
t
a
ge
1:
upd
a
t
e
d
t
he
l
oc
a
t
i
on
of
t
he
h
i
ppos
w
i
t
h
i
n
t
h
e
p
ond
or
ri
v
e
r
F
or
=
1
:
/
2
Com
pu
t
e
a
nov
e
l
l
oc
a
t
i
on
f
or
ℎ
h
i
ppos
U
pda
t
e
d
t
h
e
l
oc
a
t
i
on
for
ℎ
hi
ppos
E
nd
for
S
t
a
ge
2:
hi
p
popot
a
m
us
d
e
fe
ns
e
a
g
a
i
ns
t
pr
e
d
a
t
ors
F
or
=
1
+
/
2
:
M
a
ke
a
r
bi
t
ra
ry
l
oc
a
t
i
ons
fo
r
pr
e
d
a
t
or
Com
pu
t
e
a
nov
e
l
l
oc
a
t
i
on
f
or
ℎ
h
i
ppos
U
pda
t
e
d
t
h
e
l
oc
a
t
i
on
for
ℎ
hi
ppos
E
nd
for
S
t
a
ge
3:
hi
p
popot
a
m
us
e
s
c
a
pe
d
fro
m
t
h
e
pr
e
d
a
t
or
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
A
rt
i
f
I
nt
e
l
l
IS
S
N
:
2252
-
8938
F
ac
i
al
par
a
l
y
s
i
s
i
m
ag
e
anal
y
s
i
s
f
or
s
t
r
ok
e
d
e
t
e
c
t
i
o
n
us
i
ng
de
e
p
e
ns
e
m
b
l
e
…
(
Ki
r
u
t
hi
ga
Su
br
am
ani
y
an
)
4083
Com
pu
t
e
n
e
w
bo
unda
ri
e
s
of
t
h
e
va
r
i
a
b
l
e
'
s
d
e
c
i
s
i
on
F
or
=
1
:
Com
pu
t
e
t
h
e
nov
e
l
l
o
c
a
t
i
on
for
ℎ
hi
pp
os
U
pda
t
e
d
t
h
e
l
oc
a
t
i
on
for
ℎ
hi
ppos
E
nd
for
S
a
vi
ng
t
h
e
opt
i
m
um
c
a
ndi
da
t
e
pe
r
form
a
n
c
e
foun
d
so
fa
r
E
nd
for
O
ut
put
t
h
e
opt
i
m
um
s
ol
u
t
i
ons
for
t
h
e
pri
m
a
r
y
f
unc
t
i
on
E
N
D
T
he
HOA
r
a
i
s
e
s
a
f
i
t
n
e
s
s
fu
nc
t
i
on
(F
F
)
to
ge
t
s
upe
r
i
or
c
l
a
s
s
i
fi
c
a
t
i
on
pe
rf
orm
a
nc
e
s
.
It
i
de
n
t
i
f
i
e
s
a
pos
i
t
i
ve
i
nt
e
ge
r
to
s
i
gn
i
fy
t
he
g
re
a
t
e
r
i
m
pl
e
m
e
n
t
a
t
i
on
of
t
he
c
a
n
di
d
a
t
e
ou
t
c
o
m
e
s
.
In
t
hi
s
w
or
k,
t
h
e
re
duc
t
i
on
of
t
h
e
c
l
a
s
s
i
f
i
e
r
e
r
ror
ra
t
e
c
a
n
be
e
x
a
m
i
n
e
d
as
t
he
FF,
as
i
nd
i
c
a
t
e
d
in
(11)
.
(
)
=
(
)
=
.
.
×
100
(11)
4.
R
ES
U
LT
A
N
A
LY
S
I
S
AND
DISCUSSIO
N
T
he
D
E
T
L
M
-
A
S
D
F
P
I
t
e
c
h
n
i
q
u
e
'
s
a
n
a
l
y
s
i
s
is
s
t
ud
i
e
d
us
i
n
g
t
w
o
d
a
t
a
s
e
t
s
:
t
h
e
M
E
E
I
d
a
t
a
s
e
t
[
2
4]
a
n
d
t
h
e
YFP
d
a
t
a
s
e
t
[2
5]
,
[2
6
]
,
wh
ich
a
r
e
c
o
m
b
i
n
e
d
to
c
r
e
a
t
e
a
b
a
l
a
nc
e
d
d
a
t
a
s
e
t
,
as
r
e
p
r
e
s
e
n
t
e
d
in
T
a
b
l
e
1.
F
i
g
ur
e
7
s
how
s
pe
rf
orm
a
n
c
e
e
v
a
l
u
a
t
i
on
of
t
h
e
D
E
T
L
M
-
ASDFPI
M
e
t
hod
on
t
h
e
t
e
s
t
d
a
t
a
b
a
s
e
.
F
i
g
ure
s
7
(
a
)
a
nd
7
(
b
)
di
s
pl
a
ys
t
he
c
onf
us
i
on
m
a
t
ri
c
e
s
w
i
t
h
a
c
c
ura
t
e
c
l
a
s
s
i
fi
c
a
t
i
on
a
nd
re
c
ogn
i
t
i
on
of
a
l
l
6
c
l
a
s
s
l
a
be
l
s
on
a
70
:
30
T
RA
S
T
/
T
E
S
S
T
.
F
i
g
ur
e
7
(
c
)
s
how
s
t
he
s
t
u
dy
of
pr
e
c
i
s
i
on
-
re
c
a
l
l
c
urv
e
,
re
pr
e
s
e
n
t
i
n
g
gre
a
t
e
r
pe
rfor
m
a
n
c
e
s
ove
ra
l
l
of
6
c
l
a
s
s
l
a
be
l
s
.
E
v
e
nt
u
a
l
l
y
,
F
i
g
u
re
7
(
d
)
d
e
m
ons
t
r
a
t
e
s
t
he
s
t
udy
of
re
c
e
i
v
e
r
op
e
ra
t
i
n
g
c
ha
r
a
c
t
e
r
i
s
t
i
c
(
RO
C
)
,
s
i
gn
i
fyi
ng
e
ff
i
c
i
e
n
t
v
a
l
u
e
s
w
i
t
h
b
e
t
t
e
r
va
l
ue
s
of
RO
C
for
di
ff
e
r
e
nt
c
l
a
s
s
e
s
.
T
a
b
l
e
2
a
nd
F
i
g
ur
e
8
s
i
g
ni
fy
t
h
e
c
l
a
s
s
i
fi
e
r
s
t
udy
of
t
h
e
D
E
T
L
M
-
ASDFPI
m
e
t
hodo
l
ogy
und
e
r
70%T
RA
S
T
a
nd
30
%T
E
S
S
T
.
T
h
e
f
i
ndi
ngs
s
t
a
t
e
d
t
h
a
t
t
h
e
D
E
T
L
M
-
ASDFPI
m
e
t
hodo
l
ogy
a
ppr
opri
a
t
e
l
y
i
de
n
t
i
f
i
e
d
t
h
e
s
a
m
pl
e
s
.
W
i
t
h
70
%T
RA
S
T
,
t
h
e
D
E
T
L
M
-
A
S
D
F
P
I
a
pproa
c
h
p
rovi
de
s
a
ve
r
a
ge
,
,
,
,
a
nd
of
9
7.
06
%,
91
.
20%
,
9
1.
2
0%,
98
.
24%
,
a
nd
91.
17
%,
r
e
s
pe
c
t
i
ve
l
y
.
M
e
a
nw
h
i
l
e
,
w
i
t
h
30
%T
E
S
S
T
,
t
h
e
D
E
T
L
M
-
ASDFPI
t
e
c
hn
i
que
off
e
rs
a
v
e
ra
ge
,
,
,
,
a
nd
of
96
.
93%
,
90
.
82%
,
90.
8
3%,
98
.
15
%,
a
n
d
9
0.
8
1%,
c
orre
s
p
o
ndi
ng
l
y.
In
F
i
g
ur
e
9,
t
h
e
t
ra
i
ni
ng
(T
RA
A
C)
a
nd
v
a
l
i
da
t
i
on
(V
L
A
A
C)
va
l
ue
s
of
t
h
e
D
E
T
L
M
-
ASDFPI
t
e
c
hni
que
a
re
d
i
s
pl
a
ye
d
.
T
h
e
ra
t
e
of
is
e
s
t
i
m
a
t
e
d
for
0
-
100
e
po
c
h
c
ount
s
.
T
h
e
fi
gu
re
unde
r
l
i
n
e
d
t
h
a
t
t
h
e
v
a
l
u
e
s
of
T
RA
A
C
a
nd
V
L
A
A
C
d
i
s
pl
a
y
an
i
n
c
re
a
s
i
n
g
t
re
n
d,
w
hi
c
h
i
nfor
m
e
d
t
he
c
a
p
a
bi
l
i
t
y
of
t
he
D
E
T
L
M
-
ASDFPI
t
e
c
hn
i
qu
e
w
i
t
h
e
nha
n
c
e
d
e
xe
c
u
t
i
on
ov
e
r
v
a
r
i
ous
i
t
e
ra
t
i
o
ns
.
M
ore
ove
r
,
t
h
e
T
RA
A
C
a
nd
V
L
A
A
C
s
t
a
y
n
e
a
re
r
ov
e
r
t
h
e
e
po
c
hs
,
w
hi
c
h
s
how
s
m
i
n
i
m
u
m
ove
r
fi
t
t
i
n
g
a
nd
di
s
p
l
a
ys
gr
e
a
t
e
r
e
xe
c
ut
i
on
of
t
h
e
D
E
T
L
M
-
ASDFPI
m
od
e
l
,
p
rom
i
s
i
ng
c
ons
t
a
nt
pre
di
c
t
i
o
n
on
unno
t
i
c
e
d
s
a
m
pl
e
s
.
In
F
i
g
u
re
10
,
t
he
T
RA
l
os
s
(T
RA
L
S
)
a
nd
V
L
A
l
os
s
(V
L
A
L
S
)
gra
p
h
of
t
he
D
E
T
L
M
-
ASDFPI
a
ppro
a
c
h
is
e
xhi
b
i
t
e
d
.
T
h
e
r
a
t
e
of
l
os
s
is
e
s
t
i
m
a
t
e
d
for
0
-
100
e
po
c
hs
.
T
he
r
a
t
e
of
T
RA
L
S
a
nd
V
L
A
L
S
e
xhi
b
i
t
s
a
r
e
duc
i
ng
t
re
nd,
i
nfo
rm
i
ng
t
he
c
a
pa
b
i
l
i
t
y
of
t
h
e
D
E
T
L
M
-
ASDFPI
a
ppr
oa
c
h
to
ba
l
a
n
c
e
a
t
ra
de
-
of
f
be
t
w
e
e
n
da
t
a
f
i
t
t
i
ng
a
n
d
ge
ne
r
a
l
i
z
a
t
i
on.
M
ore
ove
r
,
t
he
c
on
s
t
a
nt
re
d
uc
t
i
on
in
l
os
s
ra
t
e
gua
r
a
n
t
e
e
s
s
upe
r
i
or
pe
rfor
m
a
nc
e
s
of
t
h
e
D
E
T
L
M
-
ASDFPI
t
e
c
hni
que
a
n
d
f
i
ne
-
t
u
ni
ng
of
t
h
e
pr
e
di
c
t
i
on
va
l
u
e
s
ov
e
r
t
i
m
e
.
T
a
b
l
e
3
pr
e
s
e
n
t
s
a
c
om
p
a
ri
s
o
n
s
t
u
dy
of
t
h
e
D
E
T
L
M
-
ASDFPI
m
e
t
hod
w
i
t
h
c
urr
e
nt
t
e
c
hn
i
qu
e
s
[18
]
,
[27]
.
T
h
e
V
G
G
16
N
e
t
m
ode
l
a
t
t
a
i
ns
93
.
10%
,
w
hi
l
e
P
H
CN
N
-
L
S
T
M
e
nh
a
nc
e
s
94
.
80
%.
T
P
CN
N
e
xhi
b
i
t
s
hi
gh
97.
09%
but
l
ow
e
r
89.
91%
,
a
nd
R
e
s
N
e
t
50
r
e
a
c
h
e
s
95
.
80%
but
w
i
t
h
l
e
s
s
e
r
78
.
79
%.
r
a
ndo
m
for
e
s
t
(RF
)
a
nd
s
uppo
rt
v
e
c
t
or
m
a
c
hi
n
e
(S
V
M
)
c
l
a
s
s
i
fi
e
rs
gi
v
e
an
i
m
p
rove
d
a
c
c
ur
a
c
y
of
93
.
65
%
a
nd
9
4.
87
%,
r
e
s
pe
c
t
i
ve
l
y
,
w
i
t
h
v
a
ry
i
ng
a
nd
.
T
h
e
e
ns
e
m
b
l
e
c
l
a
s
s
i
fi
e
r
a
c
hi
e
ve
s
96.
13
%
,
a
n
d
t
h
e
D
E
T
L
M
-
ASDFPI
m
od
e
l
out
pe
rfo
rm
s
a
l
l
,
w
i
t
h
97
.
06%
,
91.
20%
,
91.
20
%
,
a
nd
98
.
24%
.
T
a
b
l
e
1
.
D
e
t
a
i
l
s
of
da
t
a
b
a
s
e
S
e
v
e
ri
t
y
c
l
a
s
s
e
s
N
u
m
b
e
r
of
i
m
a
g
e
s
N
o
rm
a
l
500
N
e
a
r
n
o
rm
a
l
500
M
i
l
d
500
M
o
d
e
ra
t
e
500
S
e
v
e
re
500
Co
m
p
l
e
t
e
500
T
o
t
a
l
i
m
a
g
e
s
3
,
000
Evaluation Warning : The document was created with Spire.PDF for Python.