I
nte
rna
t
io
na
l J
o
urna
l o
f
I
nfo
rm
a
t
ics a
nd
Co
m
m
un
ica
t
io
n T
ec
hn
o
lo
g
y
(
I
J
-
I
CT
)
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
,
p
p
.
9
4
4
~
9
5
4
I
SS
N:
2252
-
8
7
7
6
,
DOI
:
1
0
.
1
1
5
9
1
/iji
ct
.
v15
i
3
.
pp
944
-
9
5
4
944
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ict.
ia
esco
r
e.
co
m
Lig
htweight para
llel f
ee
dba
ck ne
tw
o
rk bas
ed on C
R
L with
po
licy
t
ra
nsfer
and enha
nceme
nt
for ima
g
e super
-
res
o
lution
S
V
R
M
a
nim
a
la
,
T
K
a
v
it
ha
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
mm
u
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
M
V
S
R
En
g
i
n
e
e
r
i
n
g
C
o
l
l
e
g
e
,
H
y
d
e
r
a
b
a
d
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Au
g
1
,
2
0
2
5
R
ev
is
ed
Ap
r
6
,
2
0
2
6
Acc
ep
ted
Ap
r
1
7
,
2
0
2
6
Im
a
g
e
su
p
e
r
-
re
so
lu
ti
o
n
(S
R)
is
e
ss
e
n
ti
a
l
in
a
p
p
li
c
a
ti
o
n
s
su
c
h
a
s
su
r
v
e
il
lan
c
e
,
m
e
d
ica
l
ima
g
in
g
,
a
n
d
re
m
o
te
s
e
n
sin
g
,
b
u
t
e
x
isti
n
g
d
e
e
p
lea
rn
in
g
(DL)
m
o
d
e
ls
o
ften
re
q
u
ire
h
ig
h
c
o
m
p
u
tatio
n
a
l
re
so
u
rc
e
s
a
n
d
stru
g
g
le
to
re
c
o
v
e
r
fin
e
d
e
tails
in
li
g
h
twe
ig
h
t
a
rc
h
it
e
c
tu
re
s.
Alth
o
u
g
h
fe
e
d
b
a
c
k
a
n
d
a
tt
e
n
ti
o
n
-
b
a
se
d
m
e
th
o
d
s
h
a
v
e
sh
o
wn
im
p
ro
v
e
m
e
n
ts,
th
e
y
stil
l
lac
k
a
n
e
ffe
c
ti
v
e
c
o
m
b
in
a
ti
o
n
o
f
e
fficie
n
t
fe
a
tu
re
re
fin
e
m
e
n
t,
e
d
g
e
e
n
h
a
n
c
e
m
e
n
t,
a
n
d
lo
w
p
a
ra
m
e
ter co
m
p
lex
it
y
.
To
a
d
d
re
ss
th
is g
a
p
,
we
p
r
o
p
o
se
a
li
g
h
twe
ig
h
t
p
a
ra
ll
e
l
fe
e
d
b
a
c
k
n
e
tw
o
rk
(LP
F
N)
th
a
t
c
o
m
b
in
e
s
t
h
re
e
k
e
y
c
o
m
p
o
n
e
n
ts:
a
fe
e
d
b
a
c
k
b
lo
c
k
f
o
r
re
p
e
a
ted
fe
a
tu
re
re
fi
n
e
m
e
n
t,
a
d
isp
e
rsio
n
-
a
wa
re
a
tt
e
n
ti
o
n
re
sid
u
a
l
b
lo
c
k
(DA
RB)
fo
r
h
i
g
h
l
ig
h
ti
n
g
imp
o
rtan
t
sp
a
ti
a
l
a
n
d
c
h
a
n
n
e
l
d
e
tails,
a
n
d
Ed
g
e
Ne
t
fo
r
e
d
g
e
sh
a
rp
e
n
in
g
fo
r
sh
a
rp
e
r
b
o
u
n
d
a
ries
.
Th
e
se
c
o
m
p
o
n
e
n
ts
a
re
su
p
p
o
rted
b
y
c
u
rric
u
lu
m
re
in
f
o
rc
e
m
e
n
t
lea
rn
in
g
(CRL),
a
n
a
d
a
p
ti
v
e
train
in
g
stra
teg
y
t
h
a
t
g
ra
d
u
a
ll
y
imp
r
o
v
e
s
th
e
m
o
d
e
l
’
s
lea
rn
in
g
b
e
h
a
v
io
r.
In
ste
a
d
o
f
re
ly
in
g
o
n
a
fi
x
e
d
lo
ss
f
u
n
c
ti
o
n
,
LP
F
N
u
se
s
a
d
y
n
a
m
ica
ll
y
lea
rn
e
d
g
lo
b
a
l
fe
e
d
b
a
c
k
lo
ss
to
re
fi
n
e
re
c
o
n
stru
c
ti
o
n
q
u
a
li
t
y
a
t
e
a
c
h
sta
g
e
.
Ex
p
e
rime
n
ts
o
n
DIV
2
K
a
n
d
F
l
ick
r2
K
s
h
o
w
t
h
a
t
LP
F
N
a
c
h
iev
e
s
h
ig
h
e
r
P
S
NR
a
n
d
S
S
IM
sc
o
re
s
wh
il
e
k
e
e
p
in
g
t
h
e
m
o
d
e
l
li
g
h
twe
ig
h
t
a
n
d
e
fficie
n
t
.
Th
is
st
u
d
y
e
m
p
h
a
siz
e
s
a
n
e
ffe
c
ti
v
e
li
g
h
tw
e
ig
h
t
fe
e
d
b
a
c
k
fra
m
e
wo
rk
,
a
n
e
n
h
a
n
c
e
d
a
tt
e
n
ti
o
n
a
n
d
e
d
g
e
-
re
fin
e
m
e
n
t
m
e
c
h
a
n
ism
,
a
n
d
a
n
a
d
a
p
ti
v
e
lea
rn
in
g
stra
teg
y
th
a
t
imp
ro
v
e
s
b
o
th
a
c
c
u
ra
c
y
a
n
d
sta
b
i
li
ty
u
n
d
e
r
d
iffere
n
t
d
e
g
ra
d
a
ti
o
n
c
o
n
d
i
ti
o
n
s.
K
ey
w
o
r
d
s
:
C
u
r
r
icu
lu
m
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
E
d
g
eNe
t
Feed
b
ac
k
n
etwo
r
k
Hig
h
-
r
eso
lu
tio
n
im
a
g
e
r
ec
o
n
s
tr
u
ctio
n
I
m
ag
e
s
u
p
er
-
r
eso
lu
tio
n
L
ig
h
tweig
h
t p
ar
allel
f
ee
d
b
ac
k
n
etwo
r
k
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
:
S
V
R
Ma
n
im
ala
Dep
ar
tm
en
t o
f
E
lectr
o
n
ics an
d
C
o
m
m
u
n
icatio
n
E
n
g
in
ee
r
i
n
g
,
MV
SR
E
n
g
in
ee
r
in
g
C
o
lleg
e
Hy
d
er
ab
ad
,
T
elan
g
a
n
a
,
5
0
1
5
1
0
,
I
n
d
ia
E
m
ail:
s
r
ik
u
r
m
am
.
m
a
n
im
ala@g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
I
m
ag
e
s
u
p
er
-
r
eso
lu
tio
n
(
SR
)
is
a
p
iv
o
tal
task
in
v
is
u
al
p
r
o
ce
s
s
in
g
,
aim
ed
at
r
ec
o
n
s
tr
u
c
tin
g
h
ig
h
-
r
eso
lu
tio
n
(
HR
)
im
ag
es
f
r
o
m
lo
w
-
r
eso
lu
tio
n
(
L
R
)
in
p
u
ts
.
I
n
en
co
d
er
–
d
ec
o
d
er
m
o
d
els,
th
e
en
co
d
er
e
x
tr
ac
ts
m
u
lti
-
s
ca
le
f
ea
tu
r
es
f
r
o
m
d
o
wn
s
am
p
led
L
R
in
p
u
ts
,
wh
ile
th
e
d
ec
o
d
er
u
p
s
am
p
les
an
d
r
ef
i
n
es
th
ese
f
ea
tu
r
es
to
g
en
er
ate
HR
o
u
tp
u
ts
[
1
]
.
HR
im
ag
es a
r
e
v
ital in
ar
ea
s
lik
e
s
u
r
v
eillan
ce
,
m
ed
ical
im
ag
i
n
g
,
a
n
d
r
em
o
te
s
en
s
in
g
,
wh
er
e
h
ig
h
clar
ity
s
u
p
p
o
r
ts
ac
cu
r
ate
an
aly
s
is
an
d
d
ec
is
io
n
-
m
ak
i
n
g
[
2
]
.
Mo
s
t
SR
m
e
th
o
d
s
r
eq
u
ir
e
h
i
g
h
co
m
p
u
tatio
n
al
co
s
ts
,
lim
itin
g
th
eir
ap
p
licatio
n
in
d
e
v
ices
wit
h
lim
ited
r
eso
u
r
ce
s
[
3
]
.
Desp
ite
ad
v
an
ce
m
en
ts
i
n
im
ag
in
g
tech
n
o
l
o
g
ies,
ca
p
tu
r
i
n
g
an
d
en
h
a
n
cin
g
HR
im
ag
es
r
em
ain
s
ch
allen
g
in
g
d
u
e
to
:
i)
th
e
r
ig
id
ity
an
d
co
s
t
-
p
r
o
h
ib
itiv
e
n
atu
r
e
o
f
r
ea
l
-
wo
r
ld
ap
p
licatio
n
s
,
an
d
ii)
th
e
p
r
io
r
itizatio
n
o
f
ca
p
tu
r
in
g
n
ew
HR
im
ag
es
o
v
er
en
h
an
cin
g
ex
is
tin
g
L
R
o
n
es.
E
ar
ly
m
ac
h
in
e
lear
n
in
g
-
b
ased
SR
m
eth
o
d
s
u
s
ed
h
an
d
c
S
r
a
f
ted
f
ea
tu
r
es
an
d
in
ter
p
o
latio
n
,
w
h
ich
wer
e
ef
f
icien
t b
u
t f
ailed
to
r
ec
o
v
e
r
f
in
e
d
etails an
d
lack
ed
s
ca
lab
ili
ty
.
Dee
p
lear
n
in
g
(
DL
)
,
esp
ec
i
ally
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs
)
,
o
v
er
ca
m
e
th
ese
lim
its
b
y
lear
n
in
g
h
ier
a
r
ch
ical
f
ea
tu
r
es
d
ir
ec
tly
f
r
o
m
d
ata,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Lig
h
tw
eig
h
t p
a
r
a
llel fe
ed
b
a
ck
n
etw
o
r
k
b
a
s
ed
o
n
C
R
L
w
ith
p
o
licy
… (
S
V
R
Ma
n
ima
la
)
945
ac
h
iev
in
g
s
u
p
er
io
r
SR
p
er
f
o
r
m
an
ce
[
4
]
.
Mo
s
t
C
NN
-
b
ased
m
eth
o
d
s
em
p
h
asize
ad
v
an
ce
d
ar
ch
itectu
r
es,
s
u
ch
as
r
esid
u
al
lear
n
in
g
an
d
d
en
s
e
co
n
n
ec
tio
n
s
[
5
]
.
Ho
wev
er
,
d
ee
p
n
etwo
r
k
s
b
r
in
g
n
ew
ch
allen
g
es,
in
clu
d
in
g
h
ig
h
c
o
m
p
u
tatio
n
al
co
m
p
lex
it
y
,
s
ig
n
if
ican
t
m
em
o
r
y
r
eq
u
ir
e
m
en
ts
,
an
d
d
if
f
ic
u
lties
in
tr
ai
n
in
g
d
u
e
t
o
is
s
u
es
lik
e
g
r
ad
ien
t v
a
n
is
h
in
g
an
d
e
x
p
lo
s
io
n
as n
etwo
r
k
d
e
p
th
in
cr
ea
s
es [
6
]
.
As n
etwo
r
k
d
ep
th
in
cr
ea
s
es,
is
s
u
e
s
s
u
ch
as
g
r
ad
ien
t
v
an
is
h
in
g
a
n
d
e
x
p
lo
s
io
n
b
ec
o
m
e
p
r
o
m
in
en
t,
co
m
p
licatin
g
th
e
co
n
v
er
g
en
ce
d
u
r
in
g
t
r
ain
in
g
.
T
o
en
h
an
ce
r
eso
lu
tio
n
p
er
f
o
r
m
an
ce
,
r
esear
ch
er
s
h
a
v
e
d
ev
el
o
p
ed
d
ee
p
er
n
etwo
r
k
s
lev
er
ag
i
n
g
r
esid
u
al
lear
n
i
n
g
tech
n
iq
u
es
[
7
]
.
T
r
a
d
itio
n
al
DL
m
eth
o
d
s
p
r
im
ar
ily
ad
o
p
t
a
f
ee
d
-
f
o
r
war
d
in
f
o
r
m
atio
n
f
l
o
w.
T
h
is
ap
p
r
o
ac
h
r
estricts
f
ee
d
b
ac
k
f
r
o
m
later
lay
er
s
to
ea
r
lier
o
n
es,
h
i
n
d
er
i
n
g
d
y
n
am
ic
a
d
ju
s
tm
en
t
o
f
th
e
in
p
u
t
at
p
r
ev
i
o
u
s
lay
er
s
[
8
]
.
T
o
s
o
lv
e
th
ese
is
s
u
es,
lig
h
tweig
h
t
SR
m
eth
o
d
s
h
av
e
b
ee
n
p
r
o
p
o
s
ed
.
Ho
wev
e
r
,
th
ey
o
f
ten
f
ail
t
o
f
u
lly
ex
p
l
o
it th
e
r
ep
r
esen
tatio
n
al
ca
p
ab
ilit
ies o
f
C
NNs.
R
ec
e
n
t
s
t
u
d
ies
h
i
g
h
li
g
h
t
t
h
e
v
a
lu
e
o
f
f
ee
d
b
a
ck
m
e
ch
a
n
is
m
s
f
o
r
s
tr
en
g
t
h
en
in
g
l
o
w
-
l
ev
el
f
ea
t
u
r
es
u
s
in
g
h
i
g
h
-
l
e
v
el
c
u
es
i
n
v
is
i
o
n
tas
k
s
[
9
]
.
D
esp
ite
t
h
e
ir
e
f
f
e
cti
v
e
n
es
s
,
m
a
n
y
SR
m
et
h
o
d
s
r
el
y
o
n
a
r
c
h
it
ec
t
u
r
es
wi
th
a
lar
g
e
n
u
m
b
e
r
o
f
p
a
r
am
ete
r
s
,
li
m
iti
n
g
t
h
e
ir
u
s
e
i
n
r
es
o
u
r
ce
-
c
o
n
s
tr
ai
n
e
d
en
v
i
r
o
n
m
en
ts
.
I
n
a
d
d
iti
o
n
,
t
h
e
n
e
ed
t
o
m
o
d
el
l
o
n
g
s
e
q
u
e
n
c
e
d
e
p
e
n
d
e
n
ci
es
o
f
t
e
n
r
es
u
l
t
i
n
la
r
g
e
n
u
m
b
e
r
s
o
f
h
i
d
d
e
n
s
tat
es,
ca
u
s
i
n
g
ch
an
n
el
r
ed
u
n
d
a
n
c
y
an
d
h
i
n
d
er
in
g
t
h
e
le
ar
n
i
n
g
o
f
ess
e
n
ti
al
SR
i
m
a
g
e
r
e
p
r
e
s
en
t
ati
o
n
s
[
1
0
]
.
T
h
is
h
as
in
cr
ea
s
e
d
i
n
t
er
est
i
n
lig
h
t
wei
g
h
t
SR
n
e
tw
o
r
k
s
[
1
1
]
,
w
h
e
r
e
f
ee
d
b
ac
k
p
r
o
v
es
es
p
e
cial
ly
u
s
ef
u
l
.
R
e
ce
n
t
w
o
r
k
o
n
th
e
d
ee
p
r
es
id
u
al
ch
an
n
el
att
e
n
ti
o
n
n
etw
o
r
k
(
R
C
AN)
[
1
2
]
f
o
c
u
s
es
o
n
o
p
ti
m
i
zin
g
t
h
e
l
_
1
l
o
s
s
b
et
we
e
n
th
e
g
e
n
e
r
a
te
d
HR
im
ag
e
an
d
th
e
g
r
o
u
n
d
t
r
u
t
h
.
Al
th
o
u
g
h
i
t
ac
h
ie
v
es
h
i
g
h
PS
NR
,
it
o
f
ten
p
r
o
d
u
ce
s
o
v
er
-
s
m
o
o
t
h
ed
r
esu
l
ts
w
it
h
b
l
u
r
r
ed
ed
g
es
.
Mo
r
eo
v
e
r
,
s
t
u
d
i
es
[
1
3
]
s
h
o
w
t
h
at
in
cr
ea
s
in
g
n
etw
o
r
k
d
e
p
t
h
d
o
es
n
o
t
a
lwa
y
s
im
p
r
o
v
e
p
e
r
f
o
r
m
a
n
ce
a
n
d
ca
n
e
v
e
n
le
ad
t
o
d
e
g
r
a
d
at
io
n
.
T
o
e
n
h
an
ce
p
er
ce
p
t
u
al
q
u
a
lit
y
,
th
e
s
u
p
e
r
-
r
es
o
l
u
t
io
n
g
en
er
ati
v
e
a
d
v
e
r
s
a
r
ia
l
n
et
wo
r
k
(
SR
GAN
)
[
1
4
]
i
n
t
r
o
d
u
c
ed
a
p
e
r
ce
p
t
u
al
l
o
s
s
c
o
m
b
in
i
n
g
a
d
v
e
r
s
a
r
ia
l
l
o
s
s
wit
h
c
o
n
t
en
t lo
s
s
d
e
r
i
v
ed
f
r
o
m
th
e
VGG
n
e
tw
o
r
k
,
b
u
t
th
is
im
p
r
o
v
e
m
e
n
t
in
v
is
u
al
q
u
ali
ty
co
m
es
a
t
t
h
e
co
s
t
o
f
l
o
w
er
PS
N
R
c
o
m
p
ar
e
d
t
o
o
th
er
SR
m
et
h
o
d
s
.
M
o
r
eo
v
er
,
e
x
is
t
in
g
f
ee
d
b
ac
k
-
b
ase
d
SR
m
et
h
o
d
s
r
e
ly
o
n
h
ea
v
y
ar
c
h
it
ec
t
u
r
es,
a
n
d
th
e
y
d
o
n
o
t
ex
p
li
cit
ly
a
d
d
r
ess
t
h
e
c
h
al
le
n
g
e
o
f
a
c
h
ie
v
i
n
g
ef
f
i
ci
en
t
h
i
g
h
-
f
r
eq
u
e
n
cy
d
et
ail
r
ec
o
v
e
r
y
in
li
g
h
twei
g
h
t
n
etw
o
r
k
s
.
W
h
ile
ea
r
lier
s
tu
d
ies
h
av
e
i
n
v
esti
g
ated
lig
h
tweig
h
t
co
m
p
o
n
en
ts
,
r
esid
u
al
lear
n
in
g
m
o
d
u
les,
an
d
f
ee
d
b
ac
k
m
ec
h
an
is
m
s
in
d
i
v
id
u
ally
,
th
e
y
h
a
v
e
n
o
t
ex
p
licitly
ex
am
in
ed
h
o
w
th
ese
elem
en
ts
b
eh
av
e
wh
e
n
co
m
b
in
ed
with
in
a
u
n
if
ie
d
lig
h
tweig
h
t
f
r
am
ewo
r
k
esp
ec
iall
y
r
eg
ar
d
in
g
h
ig
h
-
f
r
eq
u
e
n
cy
d
etail
r
ec
o
v
er
y
an
d
tr
ain
in
g
s
tab
ilit
y
.
T
o
ad
d
r
ess
th
is
g
ap
,
th
is
s
tu
d
y
in
v
esti
g
ate
s
th
e
ef
f
ec
ts
o
f
in
teg
r
atin
g
a
l
ig
h
tweig
h
t
p
ar
allel
f
ee
d
b
ac
k
m
ec
h
a
n
is
m
,
d
is
p
er
s
io
n
-
awa
r
e
atten
tio
n
,
a
n
d
cu
r
r
icu
lu
m
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
(
C
R
L
)
wi
th
in
a
s
in
g
le
SR
m
o
d
el.
T
h
er
ef
o
r
e,
t
h
is
ar
ticle
p
r
o
p
o
s
es
a
n
o
v
el
lig
h
tweig
h
t
p
ar
allel
f
ee
d
b
ac
k
n
etwo
r
k
(
L
PF
N)
to
d
eliv
er
h
ig
h
-
q
u
ality
SR
wh
ile
m
ain
tain
in
g
co
m
p
u
tatio
n
a
l
ef
f
icien
cy
.
I
n
itially
,
th
e
FB
is
in
tr
o
d
u
ce
d
to
iter
ativ
ely
r
ef
in
e
lo
w
-
lev
el
f
ea
tu
r
es
u
s
in
g
h
ig
h
-
lev
el
f
ee
d
b
ac
k
,
th
er
eb
y
r
ed
u
cin
g
g
r
ad
i
en
t
v
an
is
h
in
g
an
d
im
p
r
o
v
in
g
co
n
v
er
g
en
ce
.
Seco
n
d
,
DARB
i
s
in
teg
r
ated
to
s
el
ec
tiv
ely
f
o
cu
s
o
n
in
f
o
r
m
ativ
e
s
p
atial
an
d
ch
an
n
el
f
ea
tu
r
es,
en
h
a
n
cin
g
ed
g
e
s
h
ar
p
n
ess
an
d
f
in
e
-
tex
tu
r
e
r
ec
o
n
s
tr
u
ctio
n
.
T
h
en
,
E
d
g
eNe
t
is
em
p
lo
y
ed
as
a
lig
h
tweig
h
t
ed
g
e
-
e
n
h
an
ce
m
e
n
t
m
o
d
u
le
th
at
s
h
ar
p
e
n
s
co
n
to
u
r
s
with
o
u
t
ad
d
in
g
h
e
av
y
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
.
Fin
ally
,
a
g
lo
b
al
f
ee
d
b
ac
k
lo
s
s
,
o
p
tim
ized
th
r
o
u
g
h
C
R
L
with
p
o
licy
tr
a
n
s
f
er
a
n
d
e
n
h
a
n
ce
m
en
t,
is
u
s
ed
to
d
y
n
am
ically
ad
j
u
s
t
th
e
lo
s
s
f
u
n
ctio
n
d
u
r
in
g
tr
ain
i
n
g
.
T
h
is
ad
ap
tiv
e
lo
s
s
lear
n
in
g
en
s
u
r
es
s
tab
le
co
n
v
er
g
e
n
ce
,
im
p
r
o
v
e
d
s
tr
u
ctu
r
al
c
o
n
s
is
ten
cy
,
an
d
r
o
b
u
s
tn
ess
u
n
d
er
v
ar
y
in
g
d
eg
r
a
d
atio
n
lev
els.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
o
f
th
is
wo
r
k
is
d
etailed
in
b
elo
w:
−
A
n
o
v
el
L
PF
N
is
in
tr
o
d
u
ce
d
,
in
teg
r
atin
g
p
ar
allel
f
ee
d
b
ac
k
a
n
d
r
esid
u
al
lear
n
in
g
to
iter
ativ
ely
r
ef
in
e
lo
w
-
lev
el
f
ea
tu
r
es
u
s
in
g
h
ig
h
-
le
v
el
in
f
o
r
m
atio
n
,
en
ab
lin
g
h
i
g
h
-
q
u
ality
SR
with
m
in
im
al
co
m
p
u
tatio
n
al
o
v
er
h
ea
d
.
−
An
en
h
an
ce
d
f
ea
tu
r
e
r
ef
in
e
m
e
n
t
m
ec
h
an
is
m
is
d
ev
elo
p
ed
t
h
r
o
u
g
h
DARB
an
d
E
d
g
eNe
t,
w
h
ich
s
elec
tiv
ely
em
p
h
asize
im
p
o
r
tan
t
s
p
atial
a
n
d
ch
a
n
n
el
f
ea
tu
r
es
an
d
ten
d
t
o
p
r
o
d
u
ce
a
h
ig
h
er
p
r
o
p
o
r
tio
n
o
f
s
h
a
r
p
e
d
g
es
an
d
clea
r
tex
tu
r
e
d
etails in
lig
h
tweig
h
t e
n
v
ir
o
n
m
en
ts
.
−
A
d
y
n
am
ic
lear
n
in
g
s
tr
ateg
y
is
in
co
r
p
o
r
ated
u
s
in
g
g
lo
b
al
f
e
ed
b
ac
k
lo
s
s
o
p
tim
ized
v
ia
C
R
L
,
allo
win
g
th
e
m
o
d
el
to
ad
a
p
tiv
ely
ad
ju
s
t
th
e
lear
n
in
g
p
r
o
ce
s
s
,
ac
c
eler
ate
co
n
v
er
g
e
n
ce
,
a
n
d
m
ain
tain
s
tab
le
r
ec
o
n
s
tr
u
ctio
n
q
u
ality
u
n
d
er
v
ar
y
in
g
d
eg
r
a
d
atio
n
co
n
d
itio
n
s
.
2.
RE
L
AT
E
D
WO
RK
S
2
.
1
.
T
ra
ditio
na
l SR
m
et
ho
ds
I
m
ag
e
SR
p
r
esen
ts
a
co
m
p
le
x
in
v
e
r
s
e
p
r
o
b
lem
,
as
e
x
tr
ac
ti
n
g
h
i
g
h
-
f
r
eq
u
e
n
cy
d
etails
f
r
o
m
lo
w
L
R
im
ag
es
is
in
h
er
en
tly
d
if
f
icu
lt.
Ho
wev
er
,
m
er
ely
in
cr
ea
s
in
g
th
e
n
etwo
r
k
’
s
d
ep
th
d
o
es
n
o
t
y
ield
s
ig
n
if
ican
t
p
er
f
o
r
m
an
ce
g
ain
s
an
d
in
cu
r
s
s
u
b
s
tan
tial
co
m
p
u
tatio
n
al
co
s
ts
.
T
o
ad
d
r
ess
th
is
,
Fan
g
et
a
l.
[
1
5
]
d
ev
elo
p
e
d
s
o
f
t
-
ed
g
e
ass
is
ted
n
etwo
r
k
(
SeaNe
t)
,
wh
ich
lev
er
a
g
es
s
o
f
t
-
ed
g
e
f
ea
tu
r
es
with
C
NNs
to
r
ef
i
n
e
SR
an
d
p
r
o
d
u
ce
h
ig
h
-
q
u
a
lity
im
ag
es.
Ho
wev
e
r
,
a
co
n
tin
u
o
u
s
ch
allen
g
e
lies
in
th
e
co
m
p
lex
task
o
f
r
ec
o
v
er
in
g
f
in
e
tex
tu
r
e
d
etails
in
th
e
r
ec
o
n
s
tr
u
cted
im
ag
es.
T
o
ad
d
r
ess
th
is
ch
allen
g
e,
Z
h
ao
et
a
l.
[
1
6
]
p
r
o
p
o
s
ed
en
h
an
ce
d
lap
lacia
n
p
y
r
am
id
g
en
er
ativ
e
a
d
v
er
s
ar
ial
n
etwo
r
k
(
E
L
SR
GAN)
,
wh
ich
u
s
es
a
L
ap
lacia
n
p
y
r
am
id
to
ca
p
tu
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
944
-
9
5
4
946
h
ig
h
-
f
r
eq
u
e
n
cy
d
etails
an
d
em
p
lo
y
s
r
esid
u
al
-
in
-
r
esid
u
al
d
en
s
e
b
lo
ck
s
(
R
R
DB
)
to
o
v
er
co
m
e
g
r
a
d
ien
t
v
an
is
h
in
g
.
No
n
eth
eless
,
th
ese
ap
p
r
o
ac
h
es
m
ay
en
co
u
n
ter
d
is
to
r
tio
n
s
wh
en
p
r
o
ce
s
s
in
g
s
m
all
im
ag
es
th
at
co
n
tain
s
ig
n
if
i
ca
n
t
in
ter
f
er
e
n
ce
.
T
o
ad
d
r
ess
th
ese
is
s
u
es
i
n
SR
r
ec
o
n
s
tr
u
ctio
n
,
Ko
n
g
e
t
a
l.
[
1
7
]
p
r
o
p
o
s
ed
en
h
an
ce
d
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
m
o
d
el
(
FISR
C
N)
,
an
en
h
a
n
ce
d
C
NN
f
o
r
co
lo
r
an
d
tex
tu
r
e
r
esto
r
atio
n
.
I
t
u
s
es
a
So
b
el
f
ilt
er
f
o
r
ed
g
es,
m
ed
ia
n
f
ilter
s
f
o
r
n
o
is
e
r
e
d
u
ctio
n
,
an
d
p
i
x
el
s
h
u
f
f
lin
g
f
o
r
ef
f
icie
n
t
u
p
s
am
p
lin
g
,
c
o
n
v
e
r
tin
g
L
R
im
ag
es to
HR
wh
ile
lev
er
ag
in
g
h
ig
h
-
d
im
en
s
io
n
al
f
ea
tu
r
es.
2
.
2
.
L
ig
htw
eig
ht
ima
g
e
SR
m
et
ho
ds
L
ig
h
tweig
h
t
SR
r
esear
ch
i
n
cr
e
asin
g
ly
aim
s
to
b
alan
ce
r
ec
o
n
s
tr
u
ctio
n
q
u
ality
with
l
o
w
co
m
p
u
tatio
n
al
co
s
t.
L
i
et
a
l.
[
1
8
]
i
n
tr
o
d
u
ce
d
Fo
u
r
ier
SR
,
wh
ich
r
ed
u
ce
s
c
o
m
p
lex
ity
u
s
in
g
f
r
e
q
u
en
c
y
-
d
o
m
ain
o
p
e
r
atio
n
s
b
u
t
s
tr
u
g
g
les
to
ca
p
tu
r
e
f
in
e
l
o
ca
l
tex
tu
r
es.
L
i
et
a
l.
[
1
9
]
p
r
o
p
o
s
ed
lig
h
tweig
h
t
cr
o
s
s
-
r
ec
ep
tiv
e
f
o
cu
s
ed
in
f
er
e
n
ce
n
etwo
r
k
(
C
FIN
)
,
b
len
d
i
n
g
C
NN
an
d
T
r
an
s
f
o
r
m
e
r
f
ea
tu
r
e
s
f
o
r
im
p
r
o
v
ed
co
n
te
x
tu
al
m
o
d
elin
g
,
th
o
u
g
h
its
m
u
ltip
le
m
o
d
u
les
in
cr
ea
s
e
ar
c
h
itectu
r
al
o
v
er
h
ea
d
.
Gao
et
a
l.
[
2
0
]
d
ev
elo
p
e
d
FDI
W
N
with
f
ea
tu
r
e
d
is
till
atio
n
an
d
weig
h
tin
g
f
o
r
ef
f
icien
t
r
eu
s
e,
b
u
t
its
n
u
m
e
r
o
u
s
d
is
till
atio
n
b
lo
c
k
s
m
ak
e
o
p
tim
izatio
n
m
o
r
e
d
if
f
icu
lt.
Xu
e
et
a
l.
[
2
1
]
d
esig
n
ed
m
u
lti
-
p
ath
f
ee
d
b
ac
k
f
u
s
io
n
n
etwo
r
k
(
MFFN)
to
en
h
an
ce
iter
ativ
e
r
ef
in
em
en
t
u
s
in
g
f
u
s
io
n
-
atten
tio
n
f
ee
d
b
ac
k
b
lo
ck
s
,
y
et
its
m
u
lti
-
p
ath
s
tr
u
ct
u
r
e
r
aises
m
em
o
r
y
u
s
ag
e
a
n
d
s
lo
ws
in
f
er
en
ce
.
L
i
et
a
l
.
[
2
2
]
p
r
o
p
o
s
ed
a
li
g
h
tweig
h
t
ad
ap
tiv
e
weig
h
ted
s
u
p
er
-
r
eso
lu
tio
n
n
etwo
r
k
(
L
W
-
AW
SR
N)
th
a
t
in
teg
r
ate
s
f
u
s
io
n
lo
ca
l
f
u
s
io
n
b
lo
ck
(
L
FB
)
with
ad
ap
tiv
e
m
u
lti
-
s
ca
le,
b
u
t
it
s
till
ad
d
s
e
x
tr
a
co
m
p
o
n
en
ts
t
o
m
an
ag
e
s
ca
le
r
ed
u
n
d
an
c
y
.
Ho
wev
er
,
th
e
u
s
e
o
f
m
u
ltip
le
s
ca
le
b
r
an
ch
es
s
till
in
t
r
o
d
u
ce
s
p
ar
a
m
eter
r
ed
u
n
d
an
cy
,
lim
itin
g
its
s
u
itab
ilit
y
f
o
r
h
ig
h
ly
r
eso
u
r
ce
-
c
o
n
s
tr
ain
ed
d
ep
l
o
y
m
en
ts
.
Ho
wev
er
,
MFFN,
an
d
L
FB
-
b
ased
n
etwo
r
k
s
im
p
r
o
v
e
r
ec
o
n
s
tr
u
ctio
n
th
r
o
u
g
h
iter
ativ
e
o
r
m
u
lti
-
p
ath
r
e
f
in
em
en
t,
b
u
t e
ac
h
s
u
f
f
er
s
f
r
o
m
d
r
awb
ac
k
s
th
at
lim
it
lig
h
tweig
h
t
d
ep
lo
y
m
en
t.
I
n
co
n
tr
ast,
th
e
p
r
o
p
o
s
ed
L
PF
N
in
tr
o
d
u
ce
s
a
s
tr
ea
m
lin
ed
p
ar
all
el
f
ee
d
b
ac
k
m
ec
h
a
n
is
m
th
at
r
eu
s
es f
ea
tu
r
es m
o
r
e
ef
f
ici
en
tly
.
2
.
3
.
Reinf
o
rc
e
m
ent
lea
rning
-
ba
s
ed
SR m
et
ho
ds
I
n
th
is
s
tu
d
y
,
C
h
en
et
a
l.
[
2
3
]
p
r
o
p
o
s
ed
s
p
atial
-
tem
p
o
r
al
h
ier
ar
ch
ical
r
ein
f
o
r
ce
m
en
t
lear
n
in
g
(
STAR
-
R
L
)
f
o
r
p
ath
o
lo
g
y
im
a
g
e
SR
.
I
t m
o
d
els S
R
as a
n
M
DP,
u
s
in
g
a
h
ier
ar
ch
ical
p
atch
-
lev
el
r
ec
o
v
er
y
with
a
s
p
atial
m
an
ag
er
to
tar
g
et
d
eg
r
ad
ed
p
atch
es
an
d
a
tem
p
o
r
al
m
an
ag
er
to
d
ec
i
d
e
ea
r
ly
s
to
p
p
in
g
,
im
p
r
o
v
in
g
r
ec
o
n
s
tr
u
ctio
n
ef
f
icien
c
y
an
d
av
o
id
in
g
o
v
e
r
-
p
r
o
ce
s
s
in
g
.
I
n
th
e
co
n
tex
t
o
f
s
in
g
le
im
ag
e
SR
,
B
o
u
f
f
ar
d
et
a
l.
[
2
4
]
p
r
o
p
o
s
ed
a
m
u
lti
-
ag
e
n
t
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
(
MRL
)
ap
p
r
o
ac
h
f
o
r
s
in
g
le
-
im
ag
e
SR
,
wh
er
e
ag
en
ts
ad
ju
s
t
p
ix
el
in
ten
s
ities
u
s
in
g
l
o
ca
l
en
h
an
ce
m
e
n
t
o
p
er
ato
r
s
in
a
co
n
ten
t
-
awa
r
e
m
a
n
n
er
.
T
h
is
m
eth
o
d
im
p
r
o
v
es
r
eso
lu
tio
n
with
o
u
t
th
e
c
o
m
p
le
x
ity
o
f
GAN
-
b
ased
m
o
d
el.
Al
tin
k
ay
a
an
d
B
ar
ak
li
[
2
5
]
p
r
o
p
o
s
ed
(
DR
L
-
SR
F
R
)
,
a
d
ee
p
r
ein
f
o
r
ce
m
en
t
lear
n
in
g
b
a
s
ed
SR
o
f
f
ac
e
r
eg
io
n
s
.
I
t
u
s
es
v
is
u
al
atten
tio
n
an
d
R
R
DB
s
f
o
r
iter
ativ
e,
p
atch
-
b
ased
d
etail
en
h
an
ce
m
en
t,
b
u
t
its
u
s
e
o
f
DR
L
an
d
d
en
s
e
r
esid
u
al
b
lo
ck
s
in
c
r
e
ases
co
m
p
u
tatio
n
al
co
m
p
lex
ity
c
o
m
p
ar
e
d
to
tr
a
d
itio
n
al
m
eth
o
d
s
.
3.
P
RO
P
O
SE
D
M
E
T
H
O
DO
L
O
G
Y
T
h
e
L
PF
N
m
o
d
el
f
o
llo
ws
a
clea
r
en
d
-
to
-
en
d
p
ip
elin
e
wh
er
e
ea
ch
co
m
p
o
n
e
n
t
r
ef
in
es
t
h
e
f
ea
tu
r
es
p
ass
ed
f
r
o
m
th
e
p
r
ev
io
u
s
s
tag
e.
First,
th
e
LR
im
ag
e
is
p
r
o
ce
s
s
ed
b
y
th
e
f
ee
d
b
ac
k
b
lo
c
k
,
wh
ich
ex
tr
ac
ts
in
itial
f
ea
tu
r
es
wh
ile
also
r
ec
eiv
in
g
h
ig
h
-
lev
el
f
ee
d
b
ac
k
f
r
o
m
later
lay
er
s
to
co
r
r
ec
t
ea
r
ly
f
ea
tu
r
e
er
r
o
r
s
.
T
h
e
r
ef
i
n
ed
f
ee
d
b
ac
k
b
lo
c
k
o
u
tp
u
t
is
th
en
s
en
t
to
th
e
d
is
p
er
s
io
n
-
awa
r
e
atten
tio
n
r
esid
u
al
b
lo
ck
(
DARB
)
,
wh
er
e
s
p
atia
l
an
d
ch
an
n
el
atten
tio
n
s
elec
tiv
ely
en
h
an
ce
s
im
p
o
r
tan
t
tex
t
u
r
es
an
d
ed
g
es.
Nex
t,
t
h
e
en
h
a
n
c
ed
f
ea
tu
r
es
m
o
v
e
in
to
E
d
g
eNe
t,
a
lig
h
tweig
h
t
m
o
d
u
le
th
at
s
h
ar
p
e
n
s
b
o
u
n
d
ar
ies
an
d
f
in
e
s
tr
u
ctu
r
es
b
ef
o
r
e
u
p
s
am
p
lin
g
.
Du
r
in
g
tr
ain
in
g
,
th
e
C
R
L
m
o
d
u
le
s
u
p
er
v
is
es
all
s
tag
es
th
r
o
u
g
h
a
g
lo
b
al
f
ee
d
b
ac
k
lo
s
s
,
wh
ich
a
d
j
u
s
ts
au
to
m
atica
lly
b
ased
o
n
th
e
lear
n
in
g
p
r
o
g
r
ess
.
T
h
is
allo
ws
f
ee
d
b
ac
k
b
lo
ck
,
DARB
,
an
d
E
d
g
eNe
t
to
o
p
e
r
ate
in
o
n
e
u
n
if
ie
d
f
ee
d
b
ac
k
p
ath
way
,
e
n
ab
lin
g
s
tab
le
tr
ain
in
g
,
f
aster
co
n
v
er
g
en
ce
,
an
d
clea
r
e
r
r
ec
o
n
s
tr
u
ctio
n
.
3
.
1
.
L
ig
htw
eig
ht
pa
ra
llel f
ee
db
a
ck
net
wo
rk
(
L
P
F
N)
L
PF
N
is
d
esig
n
ed
f
o
r
h
ig
h
-
q
u
ality
SR
o
n
d
ev
ices
with
li
m
ited
co
m
p
u
tin
g
p
o
wer
.
I
t
t
ak
es
a
LR
im
ag
e
as
in
p
u
t
an
d
im
p
r
o
v
es
it
s
tep
-
by
-
s
tep
u
s
in
g
a
p
ar
al
lel
f
ee
d
b
ac
k
m
ec
h
an
is
m
t
h
at
r
eu
s
es
an
d
r
e
f
in
es
f
ea
tu
r
es,
h
elp
in
g
th
e
n
etwo
r
k
co
r
r
ec
t
ea
r
lier
f
ea
tu
r
e
e
r
r
o
r
s
f
o
r
clea
r
er
tex
tu
r
es
an
d
s
h
ar
p
er
ed
g
es.
Fig
u
r
e
1
p
r
esen
ts
th
e
o
v
er
all
ar
ch
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
L
PF
N
m
o
d
el,
wh
ic
h
p
er
f
o
r
m
s
SR
th
r
o
u
g
h
a
p
r
o
g
r
ess
iv
e
f
ea
tu
r
e
-
r
ef
in
e
m
en
t
p
ip
elin
e.
T
h
e
L
R
im
ag
e
f
ir
s
t
u
n
d
er
g
o
es
s
h
allo
w
f
ea
tu
r
e
ex
tr
ac
tio
n
,
an
d
th
e
r
esu
ltin
g
f
ea
tu
r
es
a
r
e
p
ass
ed
i
n
to
th
e
f
ee
d
b
ac
k
b
lo
c
k
s
.
T
h
e
s
e
b
lo
ck
s
iter
ativ
ely
r
e
f
in
e
lo
w
-
lev
el
f
ea
tu
r
es
b
y
u
s
in
g
f
ee
d
b
ac
k
s
ig
n
als
f
r
o
m
d
ee
p
er
lay
er
s
,
h
elp
i
n
g
th
e
m
o
d
el
co
r
r
ec
t
ea
r
l
y
f
ea
tu
r
e
er
r
o
r
s
.
T
h
e
r
e
f
in
ed
f
ea
tu
r
es
ar
e
th
en
s
en
t
to
th
e
DARB
,
wh
er
e
ch
an
n
el
an
d
s
p
ati
al
atten
tio
n
e
m
p
h
asize
im
p
o
r
t
an
t
tex
tu
r
es,
ed
g
es,
an
d
s
tr
u
ctu
r
al
in
f
o
r
m
atio
n
.
T
h
e
en
h
a
n
ce
d
f
ea
tu
r
es
ar
e
f
u
r
th
er
p
r
o
ce
s
s
ed
b
y
E
d
g
e
Net,
a
lig
h
tweig
h
t
ed
g
e
-
s
h
ar
p
e
n
in
g
u
n
it
d
esig
n
ed
to
s
tr
en
g
th
en
b
o
u
n
d
ar
y
d
et
ails
b
e
f
o
r
e
u
p
s
a
m
p
l
i
n
g
.
A
l
l
im
p
r
o
v
e
d
f
e
a
t
u
r
e
s
a
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Lig
h
tw
eig
h
t p
a
r
a
llel fe
ed
b
a
ck
n
etw
o
r
k
b
a
s
ed
o
n
C
R
L
w
ith
p
o
licy
… (
S
V
R
Ma
n
ima
la
)
947
f
i
n
a
l
l
y
f
u
s
e
d
a
n
d
u
p
s
a
m
p
l
e
d
t
o
g
e
n
e
r
a
t
e
t
h
e
H
R
o
u
t
p
u
t
.
D
u
r
i
n
g
t
r
a
i
n
i
n
g
,
t
h
e
C
R
L
m
o
d
u
l
e
s
u
p
e
r
v
i
s
e
s
t
h
e
e
n
t
i
r
e
a
r
c
h
i
t
e
c
t
u
r
e
u
s
i
n
g
a
g
l
o
b
a
l
f
e
ed
b
a
c
k
l
o
s
s
,
e
n
s
u
r
i
n
g
t
h
a
t
f
ee
d
b
ac
k
b
lo
c
k
,
D
A
R
B
,
a
n
d
E
d
g
eN
e
t
w
o
r
k
s
m
o
o
t
h
l
y
t
o
g
e
t
h
e
r
d
u
r
i
n
g
t
r
a
i
n
i
n
g
t
o
p
r
o
d
u
c
e
c
l
e
a
r
e
r
d
et
a
il
s
a
n
d
m
o
r
e
a
c
c
u
r
a
t
e
S
R
o
u
t
c
o
m
e
.
T
h
e
o
u
tc
o
m
e
o
f
t
h
e
p
a
r
t
o
n
s
h
a
l
l
o
w
f
ea
t
u
r
e
e
x
t
r
a
ct
i
o
n
is
i
n
d
i
c
a
t
e
d
as
l
s
,
w
h
i
c
h
c
a
n
b
e
c
a
l
c
u
l
a
t
e
d
as
f
o
l
l
o
ws
,
l
s
=
F
c
(
lR
)
(
1
)
T
o
ex
tr
ac
t
s
h
allo
w
l
o
w
R
f
ea
t
u
r
es,
F
c
co
n
s
is
t
o
f
(
3
,
1
2
8
)
a
n
d
(
1
2
8
,
3
2
)
co
n
v
o
lu
tio
n
s
as
in
(
1
)
.
I
n
th
e
f
ee
d
b
ac
k
p
ar
t
s
et,
n
is
co
n
s
id
er
ed
as
s
ev
er
al
f
ee
d
b
ac
k
s
f
r
o
m
1
to
N
an
d
t
is
co
n
s
id
er
ed
as
s
ev
er
al
iter
atio
n
s
f
r
o
m
1
to
T
.
Hen
ce
th
e
o
u
t
p
u
t
o
f
th
e
n
-
th
t
er
m
o
f
th
e
f
ee
d
b
a
ck
n
etwo
r
k
in
t
-
th
iter
atio
n
is
d
ef
in
ed
as,
l
n
t
.
I
n
th
e
f
ir
s
t
iter
atio
n
,
t
h
e
f
ir
s
t
f
ee
d
b
ac
k
in
p
u
t
an
d
o
th
er
o
u
tp
u
ts
co
n
tain
o
n
ly
s
h
allo
w
f
ea
t
u
r
es
l
s
,
wh
ile
r
esid
u
al
lear
n
in
g
allo
ws
s
u
b
s
eq
u
e
n
t
f
e
ed
b
ac
k
o
u
tp
u
ts
to
i
n
clu
d
e
r
esu
lts
f
r
o
m
p
r
e
v
io
u
s
n
etwo
r
k
s
l
1
1
…
l
n
−
1
1
as
d
escr
ib
e
d
in
(
2
)
,
l
n
1
=
{
F
f
eed
b
ack
(
l
s
)
n
=
1
F
f
eed
b
ack
(
l
s
,
l
1
1
…
.
l
n
−
1
1
)
n
≥
2
}
(
2
)
w
h
er
e
o
p
er
atio
n
s
o
f
f
ee
d
b
ac
k
b
lo
ck
ar
e
co
n
s
id
er
e
d
as
F
f
eed
b
ack
.
I
n
th
e
r
ec
o
n
s
tr
u
ctio
n
s
tag
e,
th
e
co
n
ca
ten
ated
f
ee
d
b
ac
k
b
lo
ck
o
u
tp
u
ts
ar
e
u
p
s
am
p
led
v
ia
d
ec
o
n
v
o
l
u
tio
n
al
l
ay
er
s
,
I
n
th
e
g
l
o
b
al
f
ee
d
b
ac
k
s
tag
e,
t
h
e
d
o
wn
s
am
p
lin
g
o
p
er
ato
r
g
e
n
er
ates
lR
′
in
(
3
)
,
wh
ich
is
co
m
p
a
r
ed
with
th
e
o
r
ig
in
al
lR
to
s
u
p
er
v
is
e
an
d
g
u
id
e
th
e
tr
ai
n
in
g
o
f
th
e
LP
F
N
.
lR
′
=
F
d
o
w
n
s
am
p
l
e
(
Supe
r
−
R
e
s
ol
ution
)
,
(
3
)
Fig
u
r
e
1
.
Pro
p
o
s
ed
L
PF
N
f
r
a
m
ewo
r
k
3.
2
.
F
ee
db
a
c
k
blo
ck
Af
ter
th
e
L
PF
N
m
o
d
u
le
g
e
n
er
ates
th
e
r
ef
in
e
d
f
ea
tu
r
e
o
u
tp
u
t
,
it
is
s
en
t
b
ac
k
in
to
th
e
f
ee
d
b
ac
k
b
lo
c
k
to
b
eg
in
th
e
n
ex
t
r
ef
in
em
e
n
t
cy
cle,
wh
er
e
lo
w
-
lev
el
f
ea
tu
r
es
ar
e
iter
ativ
ely
im
p
r
o
v
e
d
u
s
in
g
h
ig
h
-
lev
el
in
f
o
r
m
atio
n
,
r
esu
ltin
g
in
en
h
a
n
ce
d
ed
g
e
s
h
ar
p
n
ess
an
d
tex
tu
r
e
r
ec
o
n
s
tr
u
ctio
n
.
T
h
r
o
u
g
h
r
ep
ea
ted
u
p
s
am
p
lin
g
an
d
d
o
wn
s
am
p
lin
g
o
f
HR
an
d
L
R
f
ea
tu
r
e
f
lo
ws
d
ef
in
ed
as
h
1
,
h
2
,
h
3
,
h
4
an
d
l
in
,
l
1
,
l
2
,
l
3
,
l
4
,
l
o
ut
r
esp
ec
tiv
ely
,
it
en
h
an
ce
s
co
n
v
er
g
e
n
ce
,
s
tab
ilit
y
,
an
d
ef
f
icien
t
f
ea
tu
r
e
r
eu
s
e,
b
o
o
s
tin
g
r
ec
o
n
s
tr
u
ctio
n
p
er
f
o
r
m
an
ce
wh
ile
k
ee
p
in
g
c
o
m
p
u
tatio
n
al
co
s
t
lo
w,
m
ak
in
g
it
s
u
itab
le
f
o
r
r
eso
u
r
ce
-
c
o
n
s
tr
ain
ed
en
v
ir
o
n
m
e
n
ts
.
I
n
th
e
f
ee
d
b
ac
k
b
lo
ck
,
s
e
t th
e
n
u
m
b
e
r
o
f
p
r
o
je
ctio
n
s
as
g
=
4
.
T
h
e
f
ee
d
b
ac
k
p
r
o
ce
s
s
is
d
ef
in
ed
as f
o
llo
ws,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
944
-
9
5
4
948
l
g
=
{
F
D
A
RB
(
l
in
)
,
g
=
1
F
D
A
RB
(
F
d
o
w
n
s
am
p
l
e
(
h
g
−
1
)
)
,
2
≤
g
≤
4
}
,
(
4
)
h
g
=
F
D
A
RB
(
F
up
s
a
m
p
l
e
(
l
g
)
)
1
≤
g
≤
4
,
(
5
)
l
o
ut
=
F
D
A
RB
(
F
d
o
w
n
s
a
m
p
l
e
(
h
4
)
)
,
(
6
)
I
n
(
4
)
-
(
6
)
,
F
D
A
RB
is
co
n
s
id
er
ed
as
a
b
asic
b
lo
ck
o
f
DA
RB
in
L
R
an
d
HR
f
ea
tu
r
e
f
lo
ws
an
d
F
u
p
s
am
p
l
e
,
F
d
o
w
n
s
am
p
l
e
d
en
o
tes
th
e
p
r
o
ce
s
s
es
f
o
r
u
p
s
am
p
lin
g
u
s
in
g
d
ec
o
n
v
o
lu
tio
n
a
n
d
d
o
wn
s
am
p
lin
g
t
h
r
o
u
g
h
co
n
v
o
l
u
tio
n
,
r
esp
ec
tiv
el
y
.
3.
3
.
Dis
persio
n
-
a
wa
re
a
t
t
ent
io
n r
esid
ua
l blo
ck
(
DARB)
T
h
e
f
ea
tu
r
es
r
ef
in
ed
b
y
th
e
f
ee
d
b
ac
k
b
lo
ck
ar
e
th
e
n
p
ass
e
d
to
th
e
DARB
m
o
d
u
le,
to
e
n
h
an
ce
t
h
e
tex
tu
r
es
an
d
e
d
g
e
r
eg
io
n
s
.
T
h
e
DARB
,
s
h
o
wn
in
Fig
u
r
e
2
,
to
im
p
r
o
v
e
th
e
e
f
f
icien
cy
o
f
t
h
e
f
ee
d
b
ac
k
b
l
o
ck
.
DAR
B
en
h
an
ce
s
f
ea
tu
r
e
r
ef
in
em
en
t
in
L
PF
N
b
y
in
teg
r
atin
g
d
is
p
er
s
io
n
-
awa
r
e
ch
an
n
el
atten
tio
n
(
DACA)
an
d
s
p
atial
atten
tio
n
(
DA
SA)
wit
h
in
a
r
esid
u
al
s
tr
u
ctu
r
e.
I
t
u
s
es
s
tan
d
ar
d
d
ev
iatio
n
(
SD)
alo
n
g
s
id
e
av
er
ag
e
an
d
m
ax
p
o
o
lin
g
to
ca
p
tu
r
e
r
ic
h
r
ep
r
esen
tatio
n
s
,
with
ch
an
n
el
atten
tio
n
em
p
h
asizin
g
im
p
o
r
tan
t
ch
an
n
els
an
d
s
p
atial
atten
tio
n
r
ef
in
in
g
p
ix
el
-
lev
el
d
etails.
W
h
en
co
m
b
in
e
d
with
f
ee
d
b
ac
k
b
lo
ck
s
,
DARB
en
ab
les
iter
ativ
e
en
h
an
ce
m
e
n
t o
f
e
d
g
e
s
an
d
tex
tu
r
es,
im
p
r
o
v
in
g
SR
q
u
ality
w
ith
lo
w
co
m
p
u
tatio
n
al
co
s
t.
Fig
u
r
e
2
.
Dis
p
er
s
io
n
-
awa
r
e
atten
tio
n
r
esid
u
al
b
lo
ck
(
DARB
)
3
.
3
.
1
.
Dis
persio
n
-
a
wa
re
cha
nn
el
a
t
t
ent
io
n
(
DACA
)
T
o
en
r
ich
c
h
an
n
el
r
e
p
r
esen
tatio
n
,
SD,
av
er
ag
e
p
o
o
lin
g
,
a
n
d
m
ax
p
o
o
lin
g
d
escr
ip
to
r
s
ar
e
co
m
b
in
ed
.
Av
er
ag
e
p
o
o
lin
g
s
u
m
m
ar
izes
o
v
er
all
ch
an
n
el
in
f
o
r
m
atio
n
,
wh
ile
m
ax
p
o
o
lin
g
h
ig
h
lig
h
ts
p
r
o
m
in
e
n
t
f
ea
tu
r
es.
SD a
d
d
s
a
n
o
v
el
p
er
s
p
ec
tiv
e
b
y
ca
p
tu
r
in
g
p
ix
el
d
is
p
er
s
io
n
a
n
d
s
tr
u
ctu
r
al
b
o
u
n
d
ar
ies.
T
o
g
e
th
er
,
th
ey
p
r
o
v
id
e
a
r
ich
er
co
n
tex
t,
en
a
b
lin
g
DACA m
ec
h
an
is
m
to
s
ig
n
if
ican
tly
b
o
o
s
t th
e
d
is
cr
im
in
ativ
e
p
o
we
r
o
f
f
ea
t
u
r
e
m
a
p
s
.
A
1
D
ch
a
n
n
el
atten
tio
n
m
a
p
R
C
×
1
×
1
is
g
en
er
ated
u
s
in
g
an
ML
P
an
d
s
ig
m
o
i
d
n
o
r
m
aliza
tio
n
,
t
h
en
ap
p
lied
v
ia
elem
e
n
t
-
wis
e
m
u
ltip
licatio
n
to
th
e
in
p
u
t f
ea
tu
r
e
m
ap
f
in
.
T
h
e
f
in
al
DACA
o
u
tp
u
t
is
d
ef
in
ed
in
(
7
)
.
f
d
ca
=
f
in
∗
σ
F
m
i
p
(
V
c
)
.
(
7
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Lig
h
tw
eig
h
t p
a
r
a
llel fe
ed
b
a
ck
n
etw
o
r
k
b
a
s
ed
o
n
C
R
L
w
ith
p
o
licy
… (
S
V
R
Ma
n
ima
la
)
949
3
.
3
.
2
.
Dis
persio
n
-
a
wa
re
s
pa
t
ia
l a
t
t
ent
io
n
(
DASA)
Similar
to
th
e
DACA,
DASA
en
h
a
n
ce
s
s
p
atial
f
ea
tu
r
e
r
e
f
in
em
en
t
b
y
co
m
b
i
n
in
g
th
o
s
e
d
escr
ip
to
r
s
.
W
h
ile
av
er
ag
e
an
d
m
a
x
p
o
o
l
in
g
ca
p
tu
r
e
o
v
er
all
an
d
p
r
o
m
in
en
t
p
ix
el
v
alu
es,
s
tan
d
a
r
d
d
ev
iatio
n
h
ig
h
lig
h
ts
p
ix
el
d
is
p
er
s
io
n
ac
r
o
s
s
ch
an
n
e
ls
,
allo
win
g
th
e
n
etwo
r
k
to
f
o
c
u
s
o
n
cr
u
cial
s
p
atial
in
f
o
r
m
ati
o
n
.
A
2
D
s
p
atial
atten
tio
n
m
ap
R
1
×
h
×
W
is
p
r
o
d
u
ce
d
u
s
in
g
a
7
×7
co
n
v
o
lu
tio
n
an
d
s
ig
m
o
id
ac
tiv
ati
o
n
,
th
en
ap
p
lied
to
th
e
in
p
u
t
f
d
ca
v
ia
e
lem
en
t
-
wis
e
m
u
ltip
licatio
n
[
2
5
]
.
T
h
e
f
in
al
DASA
o
u
tp
u
t is
g
iv
en
in
(
8
)
:
f
d
s
a
=
f
d
ca
∗
σ
(
f
7
×
7
(
V
i
,
j
)
)
.
(
8
)
Fin
ally
,
f
o
ut
,
as th
e
o
u
tp
u
t o
f
DARB
,
ca
n
b
e
o
b
tain
ed
b
y
u
s
in
g
(
9
)
:
f
o
ut
=
f
d
ca
+
f
in
.
(
9
)
3.
4
.
E
dg
eNe
t
E
d
g
eNe
t,
in
teg
r
ated
in
to
th
e
L
PF
N
f
r
am
ewo
r
k
,
en
h
an
ce
s
e
d
g
e
clar
ity
an
d
tex
tu
r
e
s
h
ar
p
n
ess
in
H
R
im
ag
es
with
lo
w
co
m
p
u
tatio
n
al
co
s
t.
I
n
s
p
ir
ed
b
y
r
ich
e
r
c
o
n
v
o
lu
tio
n
al
f
ea
tu
r
es
(
R
C
F)
b
u
t simp
lifie
d
f
r
o
m
f
iv
e
s
tag
es
to
th
r
ee
an
d
u
s
in
g
d
ep
th
-
wis
e
s
ep
ar
ab
le
co
n
v
o
lu
tio
n
s
,
it
r
ed
u
ce
s
p
ar
am
eter
s
wh
ile
p
r
eser
v
in
g
p
er
f
o
r
m
an
ce
.
B
y
f
o
cu
s
in
g
o
n
ed
g
e
-
s
p
ec
if
ic
f
ea
tu
r
es,
E
d
g
eN
et
p
r
ev
en
ts
b
l
u
r
r
in
g
,
im
p
r
o
v
es
SR
v
is
u
al
q
u
ality
,
an
d
r
em
ain
s
ef
f
icien
t f
o
r
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
en
v
ir
o
n
m
en
ts
.
Fig
u
r
e
3
illu
s
tr
ates
E
d
g
e
Net,
a
lig
h
tweig
h
t
ad
ap
tatio
n
o
f
t
h
e
R
C
F
ed
g
e
d
etec
tio
n
m
o
d
el.
E
d
g
eNe
t
r
ed
u
ce
s
th
e
o
r
ig
in
al
f
iv
e
-
s
tag
e
VGG1
6
d
esig
n
to
t
h
r
ee
s
tag
e
s
an
d
r
ep
lace
s
3
×
3
co
n
v
o
lu
tio
n
s
with
d
ep
th
-
wis
e
s
ep
ar
ab
le
co
n
v
o
l
u
tio
n
s
(
‘
Dis
co
n
v
’
)
f
o
r
ef
f
icien
cy
.
I
t
u
s
es
d
ec
o
n
v
o
l
u
tio
n
(
‘
Dec
o
n
v
’
)
an
d
p
o
o
lin
g
lay
er
s
f
o
r
f
ea
tu
r
e
r
e
co
n
s
tr
u
ctio
n
an
d
r
ed
u
ctio
n
,
with
k
er
n
el
n
o
tatio
n
“k
×
k
−
”
in
d
icatin
g
s
ize
a
n
d
f
ilter
c
o
u
n
t.
T
h
is
s
tr
ea
m
lin
ed
ar
ch
itectu
r
e
ef
f
ic
ien
tly
ex
tr
ac
ts
an
d
r
ec
o
n
s
tr
u
cts
ed
g
e
f
ea
tu
r
es
wh
ile
r
ed
u
cin
g
co
m
p
u
tatio
n
al
co
s
t,
m
ak
in
g
it
ef
f
ec
tiv
e
f
o
r
ed
g
e
en
h
a
n
ce
m
en
t in
SR
.
Fig
u
r
e
3
.
T
h
e
f
r
a
m
ewo
r
k
o
f
e
d
g
e
en
h
an
ce
d
n
etwo
r
k
(
E
d
g
e
Net)
3
.
4
.
1
.
F
us
io
n
T
h
e
f
u
s
io
n
m
ec
h
a
n
is
m
co
m
b
i
n
es
th
e
s
u
p
er
-
r
eso
lv
ed
im
ag
e
I
SR
an
d
ed
g
e
-
e
n
h
an
ce
d
im
ag
e
I
Ed
g
e
t
o
in
teg
r
ate
tex
tu
r
e
an
d
ed
g
e
d
e
tails
,
en
h
an
cin
g
cla
r
ity
an
d
s
h
ar
p
n
ess
.
T
h
e
f
u
s
io
n
m
ec
h
an
is
m
co
m
b
in
es
th
e
s
u
p
er
-
r
eso
lv
ed
im
a
g
e
(
I
SR
)
an
d
ed
g
e
-
en
h
a
n
ce
d
im
ag
e
(
I
Ed
g
e
)
th
r
o
u
g
h
co
n
ca
ten
atio
n
,
f
o
llo
we
d
b
y
a
1
×1
co
n
v
o
l
u
tio
n
t
o
r
e
d
u
ce
d
im
en
s
io
n
ality
an
d
g
e
n
er
ate
th
e
f
in
a
l
o
u
tp
u
t.
T
h
is
p
r
o
ce
s
s
in
teg
r
a
tes
en
h
an
ce
d
ed
g
e
d
etails
with
tex
tu
r
e
in
f
o
r
m
atio
n
,
p
r
o
d
u
cin
g
s
h
ar
p
er
an
d
clea
r
er
r
ec
o
n
s
tr
u
ctio
n
s
.
Ad
d
itio
n
a
lly
,
th
e
m
ec
h
an
is
m
lev
er
ag
es
in
f
o
r
m
atio
n
f
r
o
m
p
r
o
jectio
n
g
r
o
u
p
s
to
r
ef
in
e
LR
f
ea
tu
r
es
f
o
r
s
u
b
s
eq
u
en
t
iter
atio
n
s
,
en
ab
lin
g
ef
f
ec
tiv
e
f
ee
d
b
ac
k
g
en
er
atio
n
an
d
im
p
r
o
v
ed
r
ec
o
n
s
tr
u
ctio
n
q
u
ality
.
3.
5
.
G
l
o
ba
l
f
ee
db
a
ck
l
o
s
s
f
un
ct
io
n
L
PF
N
u
s
es
a
g
lo
b
al
f
ee
d
b
ac
k
m
ec
h
an
is
m
to
r
ela
y
SR
d
eg
r
ad
atio
n
i
n
f
o
r
m
atio
n
to
L
R
im
ag
es,
im
p
r
o
v
in
g
L
R
-
HR
m
ap
p
in
g
.
I
t
ad
ap
tiv
ely
b
alan
ce
s
im
ag
e
q
u
ality
an
d
ef
f
icien
c
y
u
s
in
g
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
,
with
two
l
o
s
s
es:
th
e
p
r
im
ar
y
SR
-
HR
r
eg
r
ess
io
n
lo
s
s
an
d
a
f
ee
d
b
ac
k
-
r
e
g
r
ess
io
n
lo
s
s
f
o
r
L
R
an
d
m
o
d
if
ied
LR
’
,
d
ef
in
ed
as
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
944
-
9
5
4
950
l
oss
=
l
1
(
SR
,
hR
)
+
θ
l
1
(
LR
′
,
LR
)
,
(
1
0
)
wh
er
e
θ
,
co
n
tr
o
ls
th
e
f
ee
d
b
ac
k
lo
s
s
weig
h
t
i
n
(
1
0
)
.
Feed
b
a
c
k
is
ap
p
lied
at
th
e
n
etwo
r
k
’
s
en
d
,
o
p
tim
ized
v
ia
cu
r
r
icu
lu
m
r
ein
f
o
r
ce
m
en
t
le
ar
n
in
g
,
a
n
d
g
en
er
ated
with
lig
h
tweig
h
t
co
n
v
o
lu
tio
n
al
d
o
wn
s
am
p
lin
g
a
n
d
ch
an
n
el
-
tr
a
n
s
f
o
r
m
lay
e
r
s
f
o
r
m
in
im
al
p
ar
am
eter
o
v
er
h
ea
d
.
3.
6
.
CRL wit
h
po
licy
t
ra
ns
f
er
a
nd
enha
ncem
ent
T
h
e
p
r
o
p
o
s
ed
C
R
L
-
b
ased
m
eth
o
d
u
s
es
r
ein
f
o
r
ce
m
e
n
t
lea
r
n
in
g
t
o
ad
ap
tiv
el
y
u
p
d
ate
w
eig
h
ts
an
d
o
p
tim
ize
g
lo
b
al
f
ee
d
b
ac
k
lo
s
s
f
o
r
im
p
r
o
v
ed
SR
p
er
f
o
r
m
an
c
e.
Usi
n
g
th
e
l
1
lo
s
s
,
th
e
L
PF
N
ad
ap
ts
to
v
ar
y
in
g
im
ag
e
co
m
p
le
x
ities
b
y
ar
r
an
g
in
g
HR
tar
g
ets
s
eq
u
en
tially
o
v
er
T
iter
atio
n
s
,
f
o
llo
win
g
a
c
u
r
r
icu
lu
m
s
tr
ateg
y
.
T
h
e
C
R
L
f
r
am
ewo
r
k
p
r
o
g
r
es
s
iv
ely
r
ef
in
es
im
ag
e
f
ea
tu
r
es
th
r
o
u
g
h
a
n
e
u
r
al
p
o
licy
t
h
at
g
en
er
ates
o
p
tim
al
f
ea
tu
r
e
r
ef
i
n
em
en
t
s
eq
u
e
n
ce
s
,
m
ax
im
izin
g
a
r
ewa
r
d
b
ased
o
n
p
ix
el
-
lev
el
ac
c
u
r
ac
y
a
n
d
p
er
ce
p
tu
al
q
u
ality
f
o
r
e
f
f
ec
tiv
e
SR
en
h
an
ce
m
e
n
t.
I
n
(
1
1
)
f
o
r
m
u
lates
th
e
p
o
licy
o
p
ti
m
izatio
n
o
b
jectiv
e
as
m
ax
im
iz
in
g
th
e
o
v
er
all
SR
r
ewa
r
d
to
ac
h
iev
e
th
e
m
o
s
t e
f
f
ec
tiv
e
f
ea
tu
r
e
en
h
an
ce
m
e
n
t stra
teg
y
.
ma
x
θ
R
(
ξ
∗
(
z
(
θ
)
)
)
s
.
t
.
ξ
∗
(
z
(
θ
)
)
=
f
MPC
(
z
(
θ
)
)
(
1
1
)
T
h
e
r
ewa
r
d
R
is
ca
lcu
lated
wit
h
r
esp
ec
t
to
n
etwo
r
k
p
ar
am
et
er
s
θ
u
s
in
g
th
e
ch
ain
r
u
le
f
o
r
g
r
ad
ien
t
co
m
p
u
tatio
n
.
dR
d
θ
=
∂
r
∂
z
∂
z
∂
θ
.
(
1
2
)
Her
e,
∂
z
∂
θ
is
au
to
m
atica
lly
o
b
tain
ed
th
r
o
u
g
h
b
ac
k
p
r
o
p
a
g
atio
n
,
wh
ile
dR
d
θ
ac
co
u
n
ts
f
o
r
th
e
SR
-
s
p
ec
if
ic
r
ewa
r
d
g
r
a
d
ien
ts
ass
o
ciate
d
with
r
ec
o
n
s
tr
u
ctio
n
q
u
ality
.
T
h
e
o
n
-
p
o
licy
R
L
f
r
a
m
ewo
r
k
d
ir
ec
tly
o
p
tim
izes
n
etwo
r
k
p
ar
am
eter
s
u
s
in
g
SR
-
r
elev
an
t
r
ewa
r
d
s
,
e
n
ab
lin
g
s
tab
le
an
d
ef
f
icien
t
lear
n
in
g
.
T
h
e
C
R
L
f
r
am
ewo
r
k
m
itig
ates
th
is
b
y
p
r
o
g
r
ess
iv
ely
s
h
ap
in
g
th
e
lear
n
in
g
p
r
o
ce
s
s
th
r
o
u
g
h
cu
r
r
icu
l
u
m
-
b
ased
p
o
l
icy
tr
an
s
f
er
,
wh
ich
ac
ce
ler
ates
ex
p
lo
r
atio
n
,
r
ed
u
c
es
in
s
tab
ilit
y
,
an
d
en
h
an
ce
s
f
ea
tu
r
e
r
ef
in
em
e
n
t
ac
r
o
s
s
d
if
f
er
en
t
SR
d
if
f
icu
lty
lev
els.
T
h
r
ee
cu
r
r
icu
l
u
m
m
o
d
es
=
{
c
i
}
,
i
ϵ
{
1
,
2
,
3
}
ar
e
in
tr
o
d
u
ce
d
t
o
s
y
s
tem
atica
lly
tr
an
s
f
er
an
d
r
ef
i
n
e
th
e
lear
n
ed
p
o
licy
.
−
C
u
r
r
icu
lu
m
1
: Rewar
d
s
h
a
p
in
g
f
o
r
t
r
an
s
f
er
ab
le
SR
f
ea
tu
r
e
r
ef
in
em
en
t
T
h
is
s
tag
e
f
o
cu
s
es
o
n
s
im
p
le
LR
im
ag
es
with
m
ild
d
e
g
r
ad
a
tio
n
,
allo
win
g
th
e
m
o
d
el
to
le
ar
n
b
asic
f
ea
tu
r
e
e
n
h
an
ce
m
e
n
t
s
u
ch
a
s
ed
g
e
s
h
a
r
p
en
in
g
a
n
d
s
tr
u
c
tu
r
e
p
r
eser
v
atio
n
.
T
h
e
r
ewa
r
d
f
u
n
ctio
n
g
u
i
d
es
co
n
v
er
g
en
ce
o
f
ess
en
tial SR
p
ar
am
eter
s
,
h
elp
in
g
th
e
n
etwo
r
k
d
ev
elo
p
a
tr
an
s
f
er
a
b
le
en
h
an
ce
m
en
t stra
teg
y
f
o
r
f
u
tu
r
e,
m
o
r
e
c
o
m
p
lex
s
tag
es
.
−
C
u
r
r
icu
lu
m
2
: L
ea
r
n
in
g
s
u
p
er
r
eso
lu
tio
n
Po
licy
u
n
d
er
m
o
d
e
r
ate
d
eg
r
ad
atio
n
co
n
d
itio
n
s
I
n
th
is
s
tag
e,
t
h
e
p
r
e
-
tr
ain
e
d
p
o
licy
f
r
o
m
C
u
r
r
icu
lu
m
1
is
r
ef
in
e
d
o
n
im
a
g
es
with
m
o
d
er
ate
d
eg
r
ad
atio
n
,
in
clu
d
in
g
n
o
is
e,
b
lu
r
,
an
d
c
o
m
p
r
ess
io
n
ar
tifa
ct
s
.
T
h
e
cu
r
r
icu
lu
m
en
h
an
ce
s
th
e
m
o
d
el
’
s
ab
ilit
y
to
g
en
er
alize
u
n
d
er
m
o
r
e
ch
allen
g
in
g
SR
co
n
d
itio
n
s
,
im
p
r
o
v
in
g
b
o
th
s
tr
u
ct
u
r
al
an
d
tex
tu
r
al
r
esto
r
atio
n
.
−
C
u
r
r
icu
lu
m
3
: E
n
h
an
ci
n
g
s
u
p
e
r
-
r
eso
lu
tio
n
Po
licy
u
n
d
er
co
m
p
lex
d
eg
r
ad
atio
n
T
h
e
f
in
al
s
tag
e
tar
g
ets
s
ev
er
e
ly
d
eg
r
a
d
ed
im
a
g
es
with
h
ea
v
y
b
lu
r
an
d
f
in
e
-
d
etail
lo
s
s
.
T
h
e
r
ewa
r
d
f
u
n
ct
io
n
em
p
h
asizes
p
ix
el
-
lev
el
p
er
ce
p
tu
al
q
u
ality
an
d
p
e
n
alize
s
v
is
u
al
ar
tifa
cts.
Star
ti
n
g
f
r
o
m
th
e
p
o
licy
lear
n
ed
in
C
u
r
r
icu
lu
m
2
,
th
e
n
etwo
r
k
p
r
o
g
r
ess
iv
ely
r
ef
in
e
s
its
r
ec
o
n
s
tr
u
ctio
n
ca
p
ab
ilit
y
,
co
n
v
er
g
in
g
to
an
o
p
tim
al
SR
p
o
licy
th
at
h
an
d
le
s
d
iv
er
s
e
d
eg
r
ad
a
tio
n
s
wh
ile
p
r
eser
v
in
g
v
is
u
al
f
i
d
elity
.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
o
u
tlin
es
th
e
im
p
lem
en
tatio
n
an
d
e
v
alu
atio
n
d
etails
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
.
E
v
alu
atio
n
is
co
n
d
u
cted
u
s
in
g
s
tan
d
ar
d
m
etr
ics,
in
clu
d
in
g
PS
NR
an
d
SS
I
M,
wh
ich
co
llectiv
ely
as
s
ess
th
e
ac
cu
r
ac
y
,
v
is
u
al
q
u
ality
,
an
d
s
tr
u
ctu
r
al
f
i
d
elity
o
f
th
e
s
u
p
er
-
r
eso
l
v
ed
im
ag
es.
4
.
1.
I
m
ple
m
ent
a
t
io
n
det
a
ils
T
h
e
p
r
o
p
o
s
e
d
ar
ch
ite
ct
u
r
e
u
s
es
PR
eL
U
a
cti
v
a
ti
o
n
s
t
o
p
r
o
ce
s
s
4
0
×
4
0
LR
R
GB
p
at
ch
es
p
a
ir
ed
wit
h
HR
im
a
g
es
.
I
m
p
le
m
e
n
t
ed
i
n
P
y
T
o
r
ch
o
n
a
n
NV
I
D
I
A
2
0
7
0
T
i
GP
U,
i
t is
t
r
ai
n
e
d
wit
h
a
b
a
tc
h
s
i
z
e
o
f
1
6
u
s
i
n
g
l
1
l
o
s
s
an
d
t
h
e
A
d
am
o
p
ti
m
iz
e
r
f
o
r
1
,
0
0
0
e
p
o
c
h
s
.
T
o
m
a
in
tai
n
a
lig
h
t
wei
g
h
t
d
esi
g
n
,
th
e
m
o
d
el
e
m
p
l
o
y
s
2
s
h
a
r
e
d
-
p
a
r
a
m
e
te
r
f
ee
d
b
ac
k
s
w
h
i
le
r
etai
n
i
n
g
c
o
m
p
eti
ti
v
e
r
ec
o
n
s
t
r
u
ct
io
n
ca
p
a
b
il
it
y
.
Du
r
i
n
g
t
r
ai
n
i
n
g
,
c
o
n
v
o
lu
ti
o
n
al
lay
e
r
s
u
s
e
6
4
f
ilt
e
r
s
,
w
it
h
k
er
n
el
s
iz
e
a
n
d
s
tr
i
d
e
a
d
j
u
s
t
ed
b
y
u
p
s
a
m
p
li
n
g
s
ca
le
2
×
,
k
=
6
a
n
d
s
=
2
;
f
o
r
s
ca
l
es
o
f
3
×
a
n
d
4
×,
k
=
3
a
n
d
s
=
1
.
T
h
i
s
a
d
a
p
t
iv
e
c
o
n
f
i
g
u
r
at
io
n
is
p
a
r
tic
u
la
r
l
y
i
m
p
o
r
t
an
t
f
o
r
l
ig
h
t
we
ig
h
t
SR
m
o
d
els
,
h
el
p
i
n
g
a
v
o
id
t
h
e
p
e
r
f
o
r
m
a
n
c
e
d
e
g
r
ad
ati
o
n
s
o
m
et
im
es
r
e
p
o
r
t
ed
b
y
d
ee
p
er
C
NN
-
b
ase
d
SR
a
p
p
r
o
a
c
h
es
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Lig
h
tw
eig
h
t p
a
r
a
llel fe
ed
b
a
ck
n
etw
o
r
k
b
a
s
ed
o
n
C
R
L
w
ith
p
o
licy
… (
S
V
R
Ma
n
ima
la
)
951
4
.
2.
Da
t
a
s
et
des
cr
iptio
n
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
is
ev
al
u
ated
o
n
two
b
en
ch
m
ar
k
d
a
tasets
co
m
m
o
n
ly
u
s
ed
f
o
r
i
m
ag
e
SR
:
DI
V2
K
an
d
Fli
ck
r
2
K.
DI
V2
K
co
n
s
is
ts
o
f
1
,
0
0
0
HR
R
GB
im
ag
es
with
d
i
v
er
s
e
co
n
ten
t,
s
p
lit
in
to
8
0
0
tr
ain
in
g
,
1
0
0
v
alid
atio
n
,
an
d
1
0
0
test
im
ag
es.
LR
co
u
n
ter
p
ar
ts
ar
e
g
en
er
at
ed
u
s
in
g
d
eg
r
a
d
atio
n
s
s
u
ch
as
b
icu
b
ic
d
o
wn
s
am
p
lin
g
,
m
o
tio
n
b
lu
r
,
p
o
is
s
o
n
n
o
is
e,
an
d
p
ix
el
s
h
if
tin
g
,
with
tr
ain
in
g
im
ag
e
s
d
o
wn
s
am
p
led
at
f
ac
to
r
s
o
f
0
.
6
–
0
.
9
an
d
r
o
tated
to
cr
ea
te
1
6
,
0
0
0
HR
-
L
R
im
ag
e
p
air
s
.
T
h
is
d
ataset
s
etu
p
en
s
u
r
es
co
m
p
ar
ab
ilit
y
with
m
an
y
s
tate
-
of
-
th
e
-
ar
t
m
eth
o
d
s
th
at
also
r
ely
o
n
DI
V2
K
’
s
s
tan
d
ar
d
ized
d
e
g
r
ad
atio
n
s
ettin
g
s
.
Fli
ck
r
2
K
co
n
tain
s
2
,
6
5
0
HR
im
ag
es
s
o
u
r
ce
d
f
r
o
m
Fli
ck
r
,
f
ea
tu
r
in
g
v
ar
ied
co
n
ten
t
an
d
q
u
ality
,
in
clu
d
i
n
g
lan
d
s
ca
p
es,
p
o
r
t
r
aits
,
an
d
s
t
ill
-
life
p
h
o
to
g
r
ap
h
y
,
p
r
o
v
id
i
n
g
ad
d
itio
n
al
d
i
v
er
s
ity
an
d
ch
allen
g
es
f
o
r
SR
ev
alu
atio
n
.
T
ab
le
1
s
h
o
ws
th
at
o
u
r
p
r
o
p
o
s
ed
L
PF
N
ac
h
iev
es
o
p
tim
al
×4
SR
p
er
f
o
r
m
an
ce
o
n
Set5
,
Set1
4
,
B
1
0
0
,
Ur
b
an
1
0
0
,
an
d
Ma
n
g
a1
0
9
wh
en
th
e
f
ee
d
b
ac
k
-
r
e
g
r
ess
io
n
lo
s
s
weig
h
t
(
θ)
is
0
.
1
,
with
PS
N
R
an
d
SS
I
M
v
alu
es
h
ig
h
er
t
h
an
f
o
r
o
th
er
θ
s
ettin
g
s
.
I
n
c
r
ea
s
in
g
θ
b
ey
o
n
d
0
.
1
lead
s
to
l
o
wer
PS
NR
an
d
SS
I
M,
in
d
icatin
g
d
im
in
is
h
ed
r
ec
o
n
s
tr
u
ctio
n
q
u
ality
d
u
e
to
ex
ce
s
s
iv
e
f
ee
d
b
ac
k
weig
h
tin
g
.
T
h
er
e
f
o
r
e,
th
e
id
ea
l
b
alan
ce
f
o
r
ac
h
iev
in
g
h
i
g
h
-
q
u
ality
im
ag
e
SR
i
s
f
o
u
n
d
at
θ
=0
.
1
.
T
h
is
tr
en
d
alig
n
s
with
o
b
s
er
v
ati
o
n
s
in
ea
r
lier
SR
liter
atu
r
e,
wh
er
e
o
v
er
ly
s
tr
o
n
g
f
ee
d
b
ac
k
o
r
atten
tio
n
weig
h
tin
g
h
as
b
ee
n
li
n
k
e
d
to
in
s
tab
ilit
y
an
d
o
v
er
s
h
o
o
tin
g
d
u
r
in
g
o
p
tim
izatio
n
.
T
h
ese
f
in
d
in
g
s
s
u
g
g
es
t
th
at
L
PF
N
b
en
ef
its
f
r
o
m
m
o
d
er
ate
f
ee
d
b
ac
k
r
ein
f
o
r
ce
m
e
n
t w
ith
o
u
t a
d
v
er
s
e
ly
af
f
ec
tin
g
s
tr
u
ct
u
r
al
f
id
elity
.
T
ab
le
1
.
C
o
m
p
a
r
ativ
e
an
aly
s
is
o
f
f
ee
d
b
ac
k
-
r
eg
r
ess
io
n
lo
s
s
weig
h
ts
in
L
PF
N
ac
r
o
s
s
v
ar
io
u
s
d
atasets
W
e
i
g
h
t
S
c
a
l
e
P
a
r
a
me
t
e
r
s
S
e
t
5
S
e
t
1
4
B
1
0
0
U
r
b
a
n
1
0
0
M
a
n
g
a
l
1
0
9
P
S
N
R
/
S
S
I
M
P
S
N
R
/
S
S
I
M
P
S
N
R
/
S
S
I
M
P
S
N
R
/
S
S
I
M
P
S
N
R
/
S
S
I
M
=
0
4
6
4
9
K
3
5
.
5
4
/
0
.
9
3
5
0
3
1
.
4
2
/
0
.
8
6
5
2
3
0
.
2
0
/
0
.
8
1
4
2
2
9
.
1
0
/
0
.
8
6
1
4
3
4
.
5
4
/
0
.
9
2
8
1
=
0
.
01
3
5
.
5
5
/
0
.
9
3
5
4
3
1
.
4
3
/
0
.
8
6
5
4
3
0
.
2
2
/
0
.
8
1
4
8
2
9
.
1
2
/
0
.
8
6
2
2
3
4
.
5
5
/
0
.
9
2
8
3
=
0
.
1
3
5
.
5
6
/
0
.
9
3
5
6
3
1
.
5
0
/
0
.
8
6
6
3
3
0
.
2
5
/
0
.
8
1
5
3
2
9
.
1
5
/
0
.
8
6
3
1
3
4
.
5
6
/
0
.
9
2
8
5
=
0
.
2
3
5
.
5
4
/
0
.
9
3
5
3
3
1
.
4
7
/
0
.
8
6
6
0
3
0
.
2
4
/
0
.
8
1
5
0
2
9
.
1
4
/
0
.
8
6
2
5
3
4
.
5
4
/
0
.
9
2
8
4
=
1
3
4
.
4
6
/
0
.
9
3
4
2
3
1
.
3
6
/
0
.
8
6
2
8
3
0
.
2
0
/
0
.
8
1
4
2
2
9
.
1
0
/
0
.
8
6
0
8
3
4
.
4
6
/
0
.
9
2
7
5
Fig
u
r
e
4
illu
s
tr
ates
th
e
p
r
o
g
r
e
s
s
io
n
o
f
r
ewa
r
d
o
p
tim
izatio
n
ac
r
o
s
s
v
ar
io
u
s
R
L
m
o
d
els
in
clu
d
in
g
th
e
p
r
o
p
o
s
ed
C
R
L
m
eth
o
d
,
wh
en
ap
p
lied
to
cu
r
r
icu
lu
m
-
b
ased
r
ein
f
o
r
ce
m
e
n
t
lear
n
in
g
f
o
r
im
a
g
e
SR
.
T
h
e
r
ewa
r
d
in
cr
ea
s
es
o
v
er
iter
atio
n
s
,
with
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
u
t
p
e
r
f
o
r
m
in
g
o
th
er
m
o
d
els
in
b
o
th
co
n
v
er
g
e
n
ce
r
ate
an
d
f
in
al
r
ewa
r
d
lev
el,
h
ig
h
lig
h
tin
g
its
ef
f
icien
c
y
.
T
h
e
im
p
r
o
v
e
m
en
t
h
ig
h
lig
h
ts
th
e
ef
f
ec
tiv
e
n
ess
o
f
C
R
L
in
th
e
p
r
o
p
o
s
ed
m
o
d
el,
wh
ich
p
r
o
g
r
e
s
s
iv
ely
ad
ju
s
ts
ta
s
k
d
if
f
icu
lty
to
en
h
an
ce
lear
n
in
g
,
o
p
tim
ize
p
o
lic
ies ef
f
icien
tly
,
an
d
ac
h
iev
e
h
ig
h
e
r
r
ewa
r
d
s
,
esp
ec
ially
in
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
im
ag
e
p
r
o
ce
s
s
in
g
task
s
.
C
o
m
p
ar
ed
with
ex
is
tin
g
R
L
ap
p
r
o
ac
h
es
s
u
ch
as
M
R
L
,
D
R
L
-
SR
FR
,
STA
R
-
R
L
,
an
d
DR
L
,
o
u
r
r
esu
lts
d
em
o
n
s
tr
ate
th
at
ac
h
iev
in
g
h
ig
h
er
r
ewa
r
d
v
alu
es
d
o
es
n
o
t
co
m
e
at
th
e
ex
p
e
n
s
e
o
f
v
is
u
al
q
u
ality
.
T
h
is
co
n
tr
asts
with
s
ev
er
al
RL
–
b
ased
SR
m
o
d
els,
wh
er
e
ag
g
r
ess
iv
ely
m
ax
im
izin
g
r
ewa
r
d
s
ig
n
als
o
f
ten
lead
s
to
s
o
f
t,
t
ex
tu
r
e
-
d
im
i
n
is
h
ed
,
o
r
o
v
e
r
ly
s
m
o
o
t
h
o
u
t
p
u
ts
.
Fig
u
r
e
4
.
C
o
m
p
a
r
is
o
n
o
f
C
R
L
m
o
d
els in
r
ewa
r
d
o
p
tim
izatio
n
Fig
u
r
e
5
illu
s
tr
ates
th
e
q
u
alit
ativ
e
co
m
p
ar
is
o
n
o
f
E
d
g
eNe
t
with
SeaNe
t,
DASR
Ne
t,
Me
m
Net,
an
d
SR
Den
s
eNe
t
o
n
th
e
DI
V2
K
an
d
Fli
ck
r
2
K
d
atasets
.
T
h
e
p
r
o
p
o
s
ed
E
d
g
eNe
t
p
r
o
v
id
es
v
is
ib
ly
clea
n
er
an
d
s
h
ar
p
er
r
ec
o
n
s
tr
u
ctio
n
s
,
esp
ec
ially
in
ch
allen
g
i
n
g
r
eg
io
n
s
s
u
ch
as
tex
t
b
o
u
n
d
ar
ies,
o
b
ject
co
n
to
u
r
s
,
an
d
f
in
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
15
,
No
.
3
,
Sep
tem
b
er
20
26
:
944
-
9
5
4
952
tex
tu
r
e
p
atter
n
s
.
W
h
ile
SeaN
et
an
d
DASR
Net
o
f
ten
g
en
er
ate
s
lig
h
tly
b
lu
r
r
ed
ed
g
es
an
d
m
is
s
f
in
e
d
etails,
Me
m
Net
an
d
SR
Den
s
eNe
t
te
n
d
to
lo
s
e
tex
tu
r
es
an
d
p
r
o
d
u
ce
s
o
f
t
o
u
tp
u
ts
.
I
n
c
o
n
tr
as
t,
E
d
g
eNe
t
k
ee
p
s
th
in
lin
es,
ch
ar
ac
ter
s
tr
o
k
es,
a
n
d
o
b
ject
b
o
u
n
d
ar
ies
m
u
c
h
s
h
ar
p
er
,
with
f
ewe
r
ar
tifa
cts
an
d
less
b
lu
r
r
in
g
.
T
h
e
en
h
an
ce
d
clar
ity
in
c
h
ar
ac
ter
s
tr
o
k
es,
b
o
u
n
d
ar
y
tr
an
s
itio
n
s
,
an
d
h
ig
h
-
f
r
eq
u
en
cy
tex
tu
r
e
s
d
em
o
n
s
tr
ates
th
e
ef
f
ec
tiv
en
ess
o
f
E
d
g
eN
et
’
s
ed
g
e
ex
tr
ac
tio
n
an
d
r
ef
in
em
e
n
t
m
ec
h
an
is
m
.
T
h
is
s
h
o
ws
th
at
E
d
g
eNe
t
p
r
eser
v
es
im
p
o
r
tan
t
s
tr
u
ct
u
r
al
d
etails
m
o
r
e
ef
f
ec
tiv
ely
th
an
th
e
o
t
h
er
m
eth
o
d
s
in
d
if
f
icu
lt
r
eg
i
o
n
s
,
m
ak
in
g
it
m
o
r
e
r
eliab
le
f
o
r
h
ig
h
-
d
etail
SR
tas
k
s
.
DIV
2
K
F
li
c
k
r2
K
S
e
a
N
e
t
D
A
S
R
N
e
t
P
r
o
p
o
se
d
M
e
mN
e
t
S
R
D
e
n
seNe
t
P
r
o
p
o
se
d
Fig
u
r
e
5
.
Vis
u
al
co
m
p
ar
is
o
n
s
f
o
r
E
d
g
eNe
t
Fig
u
r
e
6
p
r
esen
ts
a
co
m
p
ar
a
tiv
e
an
aly
s
is
o
f
im
ag
e
SR
te
ch
n
iq
u
es
o
n
s
elec
ted
p
atch
es
f
r
o
m
th
e
Fli
ck
r
2
K
an
d
DI
V2
K
d
ataset
s
,
f
o
cu
s
in
g
o
n
th
e
ef
f
ec
tiv
en
ess
o
f
ea
ch
m
eth
o
d
in
r
ec
o
n
s
tr
u
ctin
g
f
in
e
im
ag
e
d
etails.
T
ec
h
n
iq
u
es
s
u
ch
as
B
icu
b
ic,
FISR
C
N,
C
C
FN,
L
W
-
AW
S
R
N,
an
d
E
MA
SR
N
p
r
o
v
id
e
g
r
ad
u
al
im
p
r
o
v
em
e
n
ts
,
y
et
o
f
ten
f
ail
to
r
ec
o
v
er
f
in
e
d
etails
o
r
m
ain
tain
s
tr
u
ctu
r
al
co
n
s
is
te
n
cy
u
n
d
e
r
lim
ited
co
m
p
u
tat
io
n
al
b
u
d
g
ets.
Ho
we
v
er
,
MFFN,
r
ely
o
n
d
ee
p
f
u
s
io
n
b
r
a
n
ch
es
b
u
t
lac
k
d
y
n
am
ic
lo
s
s
ad
ap
tatio
n
.
I
n
co
n
tr
ast
th
e
L
PF
N
f
r
am
ewo
r
k
d
em
o
n
s
tr
ates
clea
r
er
tex
tu
r
es
an
d
s
h
ar
p
er
e
d
g
es,
an
d
im
p
r
o
v
e
th
e
r
ec
o
n
s
tr
u
ctio
n
q
u
ality
with
o
u
t
in
cr
ea
s
in
g
m
o
d
el
s
ize.
T
h
e
s
e
s
h
o
ws
th
at
h
ig
h
-
f
r
e
q
u
en
c
y
r
ec
o
v
e
r
y
b
en
ef
its
m
o
r
e
f
r
o
m
ef
f
ec
tiv
e
f
ea
tu
r
e
r
eu
s
e
an
d
s
tab
le
o
p
tim
izatio
n
th
an
f
r
o
m
d
ee
p
er
ar
ch
ite
ctu
r
es.
Ov
er
all,
th
e
f
in
d
in
g
s
h
i
g
h
lig
h
t
t
h
at
lig
h
tw
eig
h
t
SR
m
o
d
els
ca
n
d
eliv
e
r
h
ig
h
v
is
u
al
q
u
ality
wh
e
n
s
u
p
p
o
r
ted
b
y
ef
f
icien
t
f
ee
d
b
ac
k
m
ec
h
an
is
m
s
an
d
a
d
a
p
tiv
e
lear
n
in
g
s
tr
ateg
ies,
o
f
f
er
i
n
g
a
p
r
o
m
is
in
g
d
i
r
ec
tio
n
in
S
R
ap
p
licatio
n
s
.
HR
B
i
c
u
b
i
c
F
I
S
R
C
N
M
F
F
N
LPF
N
(
o
u
r
s)
DIV
2
K
HR
B
i
c
u
b
i
c
F
I
S
R
C
N
M
F
F
N
LPF
N
(
o
u
r
s)
F
li
c
k
r
2
K
Fig
u
r
e
6
.
Vis
u
al
co
m
p
ar
is
o
n
s
f
o
r
i
m
ag
e
SR
tech
n
iq
u
es o
n
Fl
ick
r
2
K
an
d
DI
V2
K
d
atasets
5.
DIS
CU
SS
I
O
N
B
ey
o
n
d
v
is
u
al
q
u
ality
im
p
r
o
v
em
en
ts
,
L
PF
N
h
as
s
tr
o
n
g
r
ele
v
an
ce
to
in
f
o
r
m
ati
o
n
an
d
co
m
m
u
n
icatio
n
tech
n
o
lo
g
y
(
I
C
T
)
ap
p
licatio
n
s
.
I
n
co
m
m
u
n
icatio
n
s
y
s
tem
s
,
it
ca
n
b
e
d
ir
ec
tly
in
teg
r
ated
in
t
o
im
ag
e
an
d
v
id
e
o
tr
an
s
m
is
s
io
n
p
ip
elin
es,
im
p
r
o
v
in
g
t
h
e
q
u
ality
o
f
tr
a
n
s
m
itted
LR
im
ag
es
a
n
d
v
id
e
o
s
,
r
ed
u
cin
g
b
an
d
wid
th
r
eq
u
ir
em
e
n
ts
wh
ile
p
r
eser
v
in
g
cr
itical
v
is
u
al
d
etails.
I
n
tele
m
ed
icin
e,
L
PF
N
ca
n
ass
is
t
d
o
cto
r
s
b
y
e
n
h
an
cin
g
d
iag
n
o
s
tic
im
ag
es
ca
p
tu
r
ed
u
n
d
er
l
o
w
b
an
d
wid
th
o
r
n
o
i
s
y
en
v
ir
o
n
m
en
ts
,
en
ab
lin
g
m
o
r
e
r
eliab
le
r
e
m
o
te
co
n
s
u
ltatio
n
s
.
Similar
ly
,
in
n
e
two
r
k
-
b
ased
s
u
r
v
eillan
ce
,
th
e
m
o
d
el
ca
n
r
ec
o
n
s
tr
u
ct
clea
r
er
f
ac
es,
o
b
ject
s
,
an
d
s
ce
n
e
d
etails
f
r
o
m
LR
ca
m
e
r
a
f
ee
d
s
,
im
p
r
o
v
i
n
g
r
ec
o
g
n
it
io
n
ac
cu
r
ac
y
an
d
s
itu
atio
n
al
awa
r
en
ess
.
T
h
ese
ap
p
licatio
n
-
o
r
ien
ted
in
teg
r
atio
n
s
s
tr
en
g
th
en
its
p
r
ac
tical
r
elev
an
ce
an
d
d
em
o
n
s
tr
ate
h
o
w
L
PF
N
ca
n
f
u
n
ctio
n
as
an
ad
a
p
tab
le
e
n
h
an
ce
m
e
n
t
m
o
d
u
le
ac
r
o
s
s
v
ar
i
o
u
s
I
C
T
p
latf
o
r
m
s
.
T
h
ese
in
teg
r
atio
n
p
o
s
s
ib
ilit
ie
s
h
ig
h
lig
h
t
th
e
p
r
ac
tical
v
alu
e
o
f
L
PF
N
in
I
C
T
en
v
ir
o
n
m
e
n
ts
th
at
d
em
an
d
ef
f
icien
t,
r
eliab
le,
a
n
d
h
ig
h
-
q
u
ality
im
ag
e
en
h
an
ce
m
e
n
t u
n
d
er
co
n
s
tr
ain
ed
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
Lig
h
tw
eig
h
t p
a
r
a
llel fe
ed
b
a
ck
n
etw
o
r
k
b
a
s
ed
o
n
C
R
L
w
ith
p
o
licy
… (
S
V
R
Ma
n
ima
la
)
953
6.
CO
NCLU
SI
O
N
T
h
is
wo
r
k
aim
e
d
to
im
p
r
o
v
e
lig
h
tweig
h
t
SR
m
eth
o
d
s
b
y
d
ev
elo
p
in
g
an
ef
f
icien
t
DL
f
r
am
ewo
r
k
ca
p
ab
le
o
f
p
r
o
d
u
cin
g
h
ig
h
-
q
u
ality
r
ec
o
n
s
tr
u
ctio
n
s
with
lo
wer
co
m
p
u
tatio
n
al
co
s
t
th
an
ex
is
tin
g
s
y
s
tem
s
.
R
ec
en
t
ev
alu
atio
n
s
in
d
icate
th
at
ac
h
iev
in
g
h
ig
h
-
q
u
ality
SR
i
n
l
ig
h
tweig
h
t
s
ettin
g
s
r
em
ain
s
ch
allen
g
in
g
d
u
e
t
o
lo
s
s
o
f
tex
tu
r
e
a
n
d
u
n
s
tab
le
r
e
co
n
s
tr
u
ctio
n
.
B
y
in
te
g
r
atin
g
L
PF
N
with
f
ee
d
b
ac
k
r
e
f
in
em
en
t
th
r
o
u
g
h
f
ee
d
b
ac
k
b
lo
ck
s
,
s
elec
tiv
e
f
ea
tu
r
e
en
h
an
ce
m
en
t
u
s
in
g
DARB
,
an
d
ed
g
e
s
h
ar
p
en
in
g
with
E
d
g
eN
et,
s
u
p
p
o
r
ted
b
y
a
CRL
-
d
r
iv
en
g
lo
b
al
f
ee
d
b
ac
k
lo
s
s
f
o
r
s
tab
le
o
p
tim
izatio
n
.
T
h
e
r
esu
lts
o
b
tain
ed
co
n
s
is
ten
tly
s
h
o
w
th
at
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
im
p
r
o
v
e
s
s
tr
u
ctu
r
al
co
n
s
is
ten
cy
an
d
en
h
an
ce
s
ed
g
e
an
d
tex
tu
r
e
clar
ity
ev
en
u
n
d
e
r
s
tr
ict
co
m
p
u
tatio
n
al
lim
its
.
T
h
ese
im
p
r
o
v
em
e
n
ts
h
ig
h
li
g
h
t
t
h
at
th
e
p
er
f
o
r
m
an
ce
g
ai
n
is
p
r
im
ar
ily
d
r
iv
en
b
y
ef
f
ec
tiv
e
f
ea
tu
r
e
r
ef
i
n
em
en
t
a
n
d
s
tab
le
o
p
tim
izatio
n
r
ath
er
t
h
an
in
cr
ea
s
ed
m
o
d
el
s
ize
o
r
d
ee
p
er
ar
ch
itectu
r
es.
Ov
er
all,
th
e
f
in
d
i
n
g
s
co
n
f
ir
m
th
at
lig
h
tweig
h
t
SR
ca
n
a
ch
iev
e
s
tr
o
n
g
v
is
u
al
q
u
ality
wh
en
s
u
p
p
o
r
ted
b
y
ef
f
icien
t
f
ee
d
b
ac
k
m
e
ch
an
is
m
s
an
d
ad
ap
tiv
e
lear
n
in
g
p
r
o
ce
s
s
es
in
r
ea
l
-
wo
r
ld
SR
ap
p
licati
o
n
s
.
Fu
tu
r
e
s
tu
d
ies
m
ay
e
x
p
lo
r
e
ex
ten
d
in
g
L
PF
N
to
h
an
d
le
r
ea
l
-
w
o
r
ld
d
eg
r
ad
a
tio
n
s
an
d
v
id
e
o
SR
,
with
f
ea
s
ib
le
ap
p
r
o
ac
h
es
f
o
r
in
teg
r
atin
g
t
r
an
s
f
o
r
m
e
r
-
b
ased
m
o
d
u
les,
as
well
as
in
v
esti
g
atin
g
d
ep
lo
y
m
en
t
o
n
m
o
b
il
e/I
o
T
d
e
v
ices
an
d
ap
p
licatio
n
s
in
m
ed
ical
a
n
d
s
a
tellite im
ag
in
g
.
ACK
NO
WL
E
DG
M
E
N
T
S
All
lis
ted
au
th
o
r
s
h
av
e
m
ad
e
s
u
b
s
tan
tial
co
n
tr
ib
u
tio
n
s
to
th
e
r
esear
ch
,
h
av
e
r
ea
d
an
d
ap
p
r
o
v
ed
th
e
m
an
u
s
cr
ip
t,
an
d
ag
r
ee
wit
h
its
co
n
ten
t a
n
d
s
u
b
m
is
s
io
n
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
T
h
is
r
esear
ch
d
id
n
o
t
r
ec
eiv
e
an
y
s
p
ec
if
ic
g
r
an
t
f
r
o
m
f
u
n
d
i
n
g
ag
en
cies
in
th
e
p
u
b
lic,
co
m
m
er
cial,
o
r
not
-
f
o
r
-
p
r
o
f
it secto
r
s
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
S
rik
u
rm
a
m
V.
R
.
Ma
n
im
ala
✓
✓
✓
✓
✓
✓
✓
Tad
ik
o
n
d
a
Kav
ith
a
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
T
h
e
au
th
o
r
s
d
ec
lar
e
th
at
th
e
r
e
ar
e
n
o
c
o
n
f
licts
o
f
in
ter
est ass
o
ciate
d
with
th
is
m
an
u
s
cr
ip
t.
DATA AV
AI
L
AB
I
L
I
T
Y
T
h
e
d
ata
s
u
p
p
o
r
tin
g
th
is
s
tu
d
y
ar
e
av
ailab
le
u
p
o
n
r
ea
s
o
n
ab
le
r
eq
u
est.
RE
F
E
R
E
NC
E
S
[
1
]
W
.
L
i
e
t
a
l
.
,
“
Ef
f
i
c
i
e
n
t
f
a
c
e
s
u
p
e
r
-
r
e
s
o
l
u
t
i
o
n
v
i
a
w
a
v
e
l
e
t
-
b
a
s
e
d
f
e
a
t
u
r
e
e
n
h
a
n
c
e
me
n
t
n
e
t
w
o
r
k
,
”
i
n
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
3
2
n
d
A
C
M
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
Mu
l
t
i
m
e
d
i
a
,
O
c
t
.
2
0
2
4
,
p
p
.
4
5
1
5
–
4
5
2
3
,
d
o
i
:
1
0
.
1
1
4
5
/
3
6
6
4
6
4
7
.
3
6
8
1
0
8
8
.
[
2
]
D
.
C
.
L
e
p
c
h
a
,
B
.
G
o
y
a
l
,
A
.
D
o
g
r
a
,
a
n
d
V
.
G
o
y
a
l
,
“
I
mag
e
su
p
e
r
-
r
e
s
o
l
u
t
i
o
n
:
a
c
o
m
p
r
e
h
e
n
si
v
e
r
e
v
i
e
w
,
r
e
c
e
n
t
t
r
e
n
d
s,
c
h
a
l
l
e
n
g
e
s
a
n
d
a
p
p
l
i
c
a
t
i
o
n
s,
”
I
n
f
o
rm
a
t
i
o
n
F
u
si
o
n
,
v
o
l
.
9
1
,
p
p
.
2
3
0
–
2
6
0
,
M
a
r
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
n
f
f
u
s.
2
0
2
2
.
1
0
.
0
0
7
.
[
3
]
W
.
Li
,
H
.
G
u
o
,
Y
.
H
o
u
,
G
.
G
a
o
,
a
n
d
Z
.
M
a
,
“
D
u
a
l
-
d
o
mai
n
m
o
d
u
l
a
t
i
o
n
n
e
t
w
o
r
k
f
o
r
l
i
g
h
t
w
e
i
g
h
t
i
ma
g
e
s
u
p
e
r
-
r
e
so
l
u
t
i
o
n
,
”
I
EEE
T
r
a
n
s
a
c
t
i
o
n
s
o
n
M
u
l
t
i
m
e
d
i
a
,
p
p
.
1
–
1
1
,
2
0
2
6
,
d
o
i
:
1
0
.
1
1
0
9
/
t
mm
.
2
0
2
6
.
3
6
6
0
1
4
9
.
[
4
]
H
.
Y
a
n
,
Z.
W
a
n
g
,
Z.
X
u
,
Z.
W
a
n
g
,
Z.
W
u
,
a
n
d
R
.
L
y
u
,
“
R
e
sea
r
c
h
o
n
i
m
a
g
e
su
p
e
r
-
r
e
s
o
l
u
t
i
o
n
r
e
c
o
n
s
t
r
u
c
t
i
o
n
m
e
c
h
a
n
i
sm
b
a
se
d
o
n
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
,
”
i
n
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
Art
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
A
u
t
o
m
a
t
i
o
n
a
n
d
H
i
g
h
Pe
r
f
o
rm
a
n
c
e
C
o
m
p
u
t
i
n
g
,
2
0
2
4
,
p
p
.
1
4
2
–
1
4
6
,
d
o
i
:
1
0
.
1
1
4
5
/
3
6
9
0
9
3
1
.
3
6
9
0
9
5
6
.
[
5
]
J.
L
i
a
n
g
,
J.
C
a
o
,
G
.
S
u
n
,
K
.
Z
h
a
n
g
,
L.
V
a
n
G
o
o
l
,
a
n
d
R
.
T
i
mo
f
t
e
,
“
S
w
i
n
I
R
:
i
m
a
g
e
r
e
st
o
r
a
t
i
o
n
u
si
n
g
sw
i
n
t
r
a
n
sf
o
r
mer,
”
i
n
2
0
2
1
I
EEE
/
C
V
F
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
C
o
m
p
u
t
e
r
Vi
s
i
o
n
Wo
r
k
sh
o
p
s
(
I
C
C
VW
)
,
O
c
t
.
2
0
2
1
,
p
p
.
1
8
3
3
–
1
8
4
4
,
d
o
i
:
1
0
.
1
1
0
9
/
i
c
c
v
w
5
4
1
2
0
.
2
0
2
1
.
0
0
2
1
0
.
Evaluation Warning : The document was created with Spire.PDF for Python.