I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
A
I
)
V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024, pp.
4147
~
4157
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v13.i4.pp
4147
-
4157
4147
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
E
n
h
a
n
c
i
n
g r
oad
i
m
age
c
l
ar
i
t
y w
i
t
h
r
e
si
d
u
al
n
e
u
r
al
n
e
t
w
or
k
d
e
h
az
i
n
g m
o
d
e
l
A
e
r
u
n
M
ar
t
in
1
, M
oh
d
N
az
e
r
i
K
am
ar
u
d
d
in
1
,
Z
am
an
i
M
d
S
an
i
2
, H
ad
h
r
am
i
A
b
d
u
l
G
h
an
i
3
1
F
a
c
ul
t
y of
E
ngi
ne
e
r
i
ng a
nd T
e
c
hnol
ogy, M
ul
t
i
m
e
di
a
U
ni
ve
r
s
i
t
y, M
e
l
a
ka
, M
a
l
a
ys
i
a
2
F
a
c
ul
t
y of
T
e
c
hnol
ogy a
nd E
l
e
c
t
r
i
c
a
l
E
ngi
ne
e
r
i
ng, U
ni
ve
r
s
i
t
i
T
e
kni
ka
l
M
a
l
a
ys
i
a
M
e
l
a
ka
,
M
e
l
a
ka
, M
a
l
a
ys
i
a
3
F
a
c
ul
t
y of
D
a
t
a
S
c
i
e
nc
e
a
nd C
om
put
i
ng, U
ni
ve
r
s
i
t
i
M
a
l
a
ys
i
a
K
e
l
a
nt
a
n, P
e
ngk
a
l
a
n C
he
pa
,
M
a
l
a
ys
i
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
D
e
c
8
,
2023
R
e
vi
s
e
d
A
pr
6
,
2024
A
c
c
e
pt
e
d
A
pr
17
,
2024
Lane
markers,
or
road
markers,
are
the
painted
lines
on
a
roadway
that
separate
different
lanes
of
traffic.
Lane
markers
guide
drivers
and
ensure
orderly
vehicle
flow.
They
are
essential
for
advanced
driver
assi
stance
systems
(ADAS),
providing
refere
nce
po
ints
for
vehicle
positioning
on
the
road.
These
markers
enable
ADAS
to
give
warnings,
assistanc
e,
and
automati
on
features
that
enhance
driver
safety
and
convenience.
Ho
wever,
unpredictable
illumination,
such
as
a
foggy
environment,
can
suppress
marker
visi
bility,
impacting
ADAS'
s
performance.
Deep
learning
-
based
meth
ods
are
well
-
known
for
their
superiority
in
handling
various
haze
patterns.
Thi
s
paper
presents
a
r
esidual
network
(ResNet)
-
based
deep
learning
model
to
i
mprove
road
image
clarity
impacted
by
fog.
The
residual
neural
network
dehaze
model
(RNN
-
D)
utilises
a
joint
loss
function
to
produce
haze
-
free
images
with
improved
lighting
conditions
and
enhance
d
details.
The
model
was
t
rai
ned,
validated,
and
fine
-
tuned
using
hazy
and
corresponding
non
-
hazy
data
sets
to
ensure
that
the
model
is
quantitatively
sup
erior
in
the
peak
signal
-
to
-
noise
ratio
(PSNR)
and
the
structural
similarity
index
measure
(SSIM).
R
NN
-
D
achieved
an
average
PSNR
of
27.98
and
SSIM
of
0.8
on
multiple
open
-
sourced
datasets.
The
proposed
algorithm'
s
superior
performanc
e
and
v
isually
appealing
results
make
it
a
powerful
tool
for
real
-
world
image
de
hazing
applicati
ons.
K
e
y
w
o
r
ds
:
A
dva
nc
e
d
dr
iv
e
r
a
s
s
is
ta
nc
e
s
ys
te
m
D
e
e
p
le
a
r
ni
ng
D
e
ha
z
in
g a
lg
or
it
hm
I
m
a
ge
e
nha
nc
e
m
e
nt
L
a
ne
m
a
r
ke
r
s
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
M
ohd Na
z
e
r
i
K
a
m
a
r
uddi
n
F
a
c
ul
ty
of
E
ngi
ne
e
r
in
g a
nd T
e
c
hnol
ogy, M
ul
ti
m
e
di
a
U
ni
ve
r
s
it
y
(
M
e
la
ka
C
a
m
pu
s
)
S
t.
A
ye
r
K
e
r
oh L
a
m
a
, M
e
la
ka
-
75450, M
a
la
y
s
ia
E
m
a
il
:
na
z
e
r
i.
ka
m
a
r
uddi
n@
m
m
u.e
du.my
1.
I
N
T
R
O
D
U
C
T
I
O
N
N
a
tu
r
a
l
a
tm
os
phe
r
ic
phe
nom
e
na
s
uc
h
a
s
f
og, ha
z
e
,
a
nd s
m
og
i
m
pa
c
t
th
e
qu
a
li
ty
a
nd vi
s
ib
il
it
y of
r
oa
d
im
a
ge
f
e
a
tu
r
e
s
,
e
s
pe
c
ia
ll
y
on
th
e
la
ne
m
a
r
ke
r
in
f
or
m
a
ti
on.
M
a
r
ke
r
in
f
or
m
a
ti
on
is
c
r
it
ic
a
l
f
or
a
c
c
ur
a
te
a
nd
pr
e
c
is
e
c
l
a
s
s
if
ic
a
ti
on
of
la
ne
m
a
r
ke
r
s
to
im
pr
ove
th
e
p
e
r
f
or
m
a
nc
e
of
th
e
a
dva
nc
e
d
dr
iv
e
r
a
s
s
i
s
ta
nc
e
s
y
s
te
m
(
A
D
A
S
)
[
1]
.
A
s
a
r
e
s
ul
t,
de
ve
lo
pi
ng
e
f
f
e
c
ti
ve
a
nd
r
obu
s
t
de
ha
z
in
g
te
c
hni
que
s
ha
s
be
c
om
e
c
r
it
ic
a
l
to
e
nha
nc
in
g
im
a
ge
qua
li
ty
in
r
e
a
l
-
w
or
ld
s
c
e
n
a
r
io
s
.
T
r
a
di
ti
ona
l
de
ha
z
in
g
m
e
th
ods
li
ke
c
ont
r
a
s
t
s
tr
e
tc
hi
ng
[
2]
,
gui
de
d
f
il
te
r
in
g
[
3]
,
a
nd
da
r
k
c
ha
nne
l
pr
io
r
s
(
D
C
P
)
[
4]
a
r
e
pi
one
e
r
i
ng
a
ppr
oa
c
he
s
to
w
a
r
ds
im
a
ge
d
e
ha
z
in
g
th
a
t
a
c
c
e
le
r
a
te
th
e
d
e
ve
lo
pm
e
nt
of
de
ha
z
in
g
m
e
th
ods
.
H
ow
e
ve
r
,
t
he
s
e
m
e
th
ods
a
r
e
r
ig
id
a
nd
on
ly
a
ppl
ic
a
bl
e
to
c
e
r
ta
in
le
ve
ls
of
f
og,
w
hi
c
h
a
r
e
not
pr
a
c
ti
c
a
l
f
or
r
e
a
l
-
w
or
ld
s
c
e
na
r
io
s
w
it
h
di
ve
r
s
e
ha
z
e
p
a
tt
e
r
ns
.
T
h
e
s
e
li
m
it
a
ti
ons
ha
ve
pr
om
pt
e
d
th
e
e
xpl
or
a
ti
on
of
de
e
p
le
a
r
ni
ng
-
ba
s
e
d
a
ppr
oa
c
he
s
,
w
hi
c
h
ha
v
e
s
how
n
s
ig
ni
f
ic
a
nt
im
pr
ove
m
e
nt
s
in
im
a
ge
de
ha
z
in
g
[
5]
,
[
6]
.
R
e
s
id
ua
l
ne
ur
a
l
ne
two
r
ks
(
R
N
N
)
a
r
e
one
of
th
e
de
e
p
le
a
r
ni
ng
m
ode
ls
popula
r
f
or
m
a
c
hi
ne
vi
s
io
n
a
ppl
ic
a
ti
ons
s
uc
h
a
s
s
e
gm
e
nt
a
ti
on, obje
c
t
id
e
nt
if
ic
a
ti
on,
a
nd
im
a
ge
c
la
s
s
if
ic
a
ti
on.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024:
4147
-
4157
4148
T
he
ir
a
bi
li
ty
to
le
a
r
n
c
om
pl
e
x
f
e
a
tu
r
e
s
w
it
h
r
a
pi
d
tr
a
in
in
g
th
r
ough
s
ki
p
c
onne
c
ti
on
m
a
ke
s
th
e
m
one
of
th
e
c
om
pe
ti
ti
ve
m
e
th
odol
ogi
e
s
in
r
e
c
e
nt
de
ha
z
in
g
te
c
hni
que
s
[
7]
.
L
e
ve
r
a
gi
ng
th
e
c
om
put
a
ti
ona
l
e
f
f
ic
ie
nc
y
of
r
e
s
id
ua
l
ne
twor
k
(
R
e
s
N
e
t)
,
th
is
pa
pe
r
pr
opos
e
s
a
R
e
s
N
e
t
-
ba
s
e
d
de
ha
z
in
g
m
e
th
odol
ogy
us
in
g
a
tm
os
phe
r
ic
s
c
a
tt
e
r
in
g
e
s
ti
m
a
ti
on
w
it
hout
e
xpl
ic
it
tr
a
ns
m
is
s
io
n
m
a
p
a
nd
a
t
m
os
phe
r
ic
li
ght
e
s
ti
m
a
ti
on,
w
hi
c
h
ha
s
be
e
n
a
pr
om
in
e
nt
m
e
th
odol
ogy
in
pa
s
t
s
tu
di
e
s
.
T
hi
s
a
ppr
oa
c
h
de
m
ons
tr
a
te
s
a
s
ig
ni
f
ic
a
nt
im
pr
ove
m
e
nt
in
pe
a
k
s
ig
na
l
-
to
-
noi
s
e
r
a
ti
o
(
P
S
N
R
)
a
nd
s
tr
uc
tu
r
a
l
s
im
il
a
r
it
y
in
de
x
m
e
a
s
ur
e
(
S
S
I
M
)
c
om
pa
r
e
d
to
s
ta
te
-
of
-
th
e
-
a
r
t
de
ha
z
in
g
te
c
hni
que
s
on
va
r
io
us
h
a
z
e
p
a
tt
e
r
ns
w
hi
le
m
a
in
ta
in
in
g
th
e
f
in
e
de
ta
il
s
of
th
e
or
ig
in
a
l
s
c
e
ne
.
T
he
m
a
in
c
ont
r
ib
ut
io
ns
of
t
hi
s
s
tu
dy
a
r
e
s
um
m
a
r
i
s
e
d a
s
f
ol
lo
w
s
:
i)
nove
l
R
e
s
N
e
t
-
ba
s
e
d de
h
a
z
in
g m
e
th
odol
ogy u
s
in
g t
he
a
tm
os
phe
r
ic
s
c
a
tt
e
r
in
g
m
ode
l
(
A
S
M
)
w
it
hout
t
he
ne
e
d
to
c
om
p
ut
e
th
e
tr
a
ns
m
is
s
io
n
m
a
p
a
nd
a
tm
os
phe
r
ic
li
ght
s
e
pa
r
a
te
ly
;
a
nd
ii
)
v
a
li
da
ti
on
a
nd
r
e
f
in
e
m
e
nt
of
th
e
lo
s
s
f
unc
ti
on
f
or
e
nh
a
nc
in
g
th
e
P
S
N
R
a
nd
S
S
I
M
of
de
h
a
z
e
d
im
a
ge
s
.
T
he
r
e
m
a
in
de
r
of
th
is
pa
pe
r
is
o
r
ga
ni
s
e
d
a
s
f
ol
lo
w
s
.
I
n
s
e
c
ti
on
2,
r
e
la
te
d
w
or
k
on
im
a
ge
de
ha
z
in
g
te
c
hni
que
s
is
r
e
vi
e
w
e
d.
S
e
c
ti
on
3
pr
e
s
e
nt
s
th
e
pr
opos
e
d
R
e
s
N
e
t
-
ba
s
e
d
de
ha
z
in
g
a
lg
or
it
hm
.
S
e
c
ti
on
4
de
ta
il
s
th
e
e
xpe
r
im
e
nt
a
l
s
e
tu
p
a
nd
e
va
lu
a
ti
on
of
th
e
m
e
th
od
.
F
in
a
ll
y,
s
e
c
ti
on
5
c
on
c
lu
de
s
th
e
pa
p
e
r
a
nd
di
s
c
u
s
s
e
s
pot
e
nt
ia
l
f
ut
ur
e
w
or
k.
2.
R
E
L
A
T
E
D
WORK
T
he
r
e
a
r
e
two
pr
im
a
r
y
a
ppr
oa
c
he
s
to
im
a
ge
de
ha
z
in
g:
c
la
s
s
ic
a
l
a
nd
d
e
e
p
-
le
a
r
ni
ng
-
ba
s
e
d.
C
la
s
s
ic
a
l
m
e
th
ods
,
s
uc
h
a
s
th
e
D
C
P
[
4]
,
de
te
c
t
lo
w
-
in
te
ns
it
y
pi
xe
ls
in
ha
z
e
-
f
r
e
e
a
r
e
a
s
a
nd
e
m
pl
oy
a
ha
z
e
im
a
gi
ng
m
ode
l
to
e
s
ti
m
a
te
th
e
th
ic
kn
e
s
s
of
th
e
ha
z
e
a
nd
r
e
c
ove
r
hi
gh
-
qua
li
ty
ha
z
e
-
f
r
e
e
im
a
ge
s
.
H
ow
e
ve
r
,
one
li
m
it
a
ti
on
of
D
C
P
is
th
e
pr
e
s
e
nc
e
of
ha
lo
a
r
te
f
a
c
ts
a
r
ound
hi
gh
-
c
ont
r
a
s
t
e
d
ge
s
in
de
ha
z
e
d
im
a
ge
s
,
w
hi
c
h
c
a
n
ne
ga
ti
ve
ly
a
f
f
e
c
t
ove
r
a
ll
vi
s
ua
l
qua
li
ty
.
T
o
a
ddr
e
s
s
th
is
is
s
u
e
,
a
gui
de
d
f
il
te
r
[
3
]
w
a
s
pr
opos
e
d
th
a
t
a
ppl
ie
s
lo
c
a
l
li
ne
a
r
m
ode
ls
to
f
il
te
r
a
n
im
a
ge
ba
s
e
d
on
th
e
c
ont
e
nt
of
th
e
gui
da
nc
e
i
m
a
ge
,
w
he
th
e
r
it
is
th
e
in
put
im
a
g
e
it
s
e
lf
or
a
di
f
f
e
r
e
nt
im
a
ge
.
T
he
gui
de
d
f
il
te
r
e
f
f
e
c
ti
ve
ly
pe
r
f
or
m
s
e
dge
-
pr
e
s
e
r
vi
ng
s
m
oot
hi
ng,
pa
r
ti
c
ul
a
r
ly
ne
a
r
th
e
e
dge
s
,
th
e
r
e
by
m
it
i
ga
ti
ng
ha
lo
a
r
te
f
a
c
ts
.
A
ddi
ti
ona
ll
y,
c
ont
r
a
s
t
s
tr
e
tc
hi
ng
[
8
]
a
nd
hi
s
to
gr
a
m
e
qua
li
s
a
ti
on
[
9
]
ha
ve
be
e
n
pr
opos
e
d
f
or
th
e
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
of
im
a
ge
s
due
to
ha
z
e
r
e
m
ova
l
by
D
C
P
.
C
ont
r
a
s
t
s
tr
e
tc
hi
ng
a
im
s
to
e
nha
nc
e
c
ont
r
a
s
t
by
e
xt
e
ndi
ng
th
e
r
a
nge
of
in
te
ns
i
ty
va
lu
e
s
.
O
n t
he
ot
he
r
ha
nd,
hi
s
to
gr
a
m
e
qua
li
s
a
ti
on
f
oc
us
e
s
on
r
e
s
ha
pi
ng
in
te
ns
it
y
va
lu
e
di
s
tr
ib
ut
io
ns
f
or
a
m
or
e
uni
f
or
m
s
pr
e
a
d.
H
ow
e
ve
r
,
it
is
e
s
s
e
nt
ia
l
to
a
c
knowl
e
dge
th
a
t
a
ppl
yi
ng
hi
s
to
gr
a
m
e
qua
li
s
a
ti
on
c
a
n
in
tr
oduc
e
th
e
r
is
k
of
e
xa
c
e
r
ba
ti
ng
noi
s
e
a
nd
a
r
te
f
a
c
ts
,
m
a
in
ly
w
he
n
de
a
li
ng
w
it
h
f
oggy
im
a
ge
s
s
ubj
e
c
te
d
to
il
lu
m
in
a
ti
on.
M
e
ng
e
t
al
.
[1
0
]
pr
e
s
e
nt
e
d
a
m
e
th
odol
ogy
th
a
t
e
f
f
ic
ie
nt
l
y
c
om
bi
ne
s
bounda
r
y
c
on
s
tr
a
in
ts
a
nd
c
ont
e
xt
ua
l
r
e
gul
a
r
is
a
ti
on
to
a
ddr
e
s
s
th
e
s
e
c
ha
ll
e
ng
e
s
.
T
hi
s
a
ppr
oa
c
h
e
f
f
e
c
ti
ve
ly
ha
r
ne
s
s
e
s
th
e
s
yn
e
r
gy
be
twe
e
n
bounda
r
y
in
f
or
m
a
ti
on
a
nd
c
ont
e
xt
ua
l
c
ohe
r
e
nc
e
to
a
c
hi
e
ve
im
pr
ove
d
r
e
s
ul
ts
i
n i
m
a
ge
de
h
a
z
in
g.
L
i
a
nd Z
he
ng
[1
1
]
di
s
c
us
s
e
d a
n
e
dge
-
ba
s
e
d a
ppr
oa
c
h t
o
s
in
gl
e
-
im
a
ge
ha
z
e
r
e
m
ova
l
us
in
g
a
da
pt
iv
e
c
ont
r
a
s
t
e
nha
nc
e
m
e
nt
.
T
hi
s
m
e
th
od
ut
il
is
e
s
a
n
e
dge
-
pr
e
s
e
r
vi
ng
f
il
te
r
to
e
nha
nc
e
th
e
im
a
ge
’
s
c
ont
r
a
s
t
w
hi
le
pr
e
s
e
r
vi
ng
e
dge
in
f
or
m
a
ti
on.
T
he
e
nha
n
c
e
m
e
nt
s
t
r
e
ngt
h
is
a
da
pt
iv
e
ly
a
d
ju
s
te
d
ba
s
e
d
on
th
e
lo
c
a
l
c
ont
r
a
s
t
le
ve
l
to
a
c
hi
e
ve
im
pr
ove
d
r
e
s
ul
ts
.
D
e
s
pi
te
th
e
s
e
e
f
f
or
ts
,
tr
a
di
ti
ona
l
te
c
hni
que
s
of
te
n
s
tr
uggl
e
to
c
ons
is
te
nt
ly
ge
ne
r
a
te
hi
gh
-
qua
li
ty
ha
z
e
-
f
r
e
e
im
a
ge
s
due
to
t
he
ir
li
m
it
a
ti
ons
in
de
a
li
ng
w
it
h
di
ve
r
s
e
a
nd
c
ha
ll
e
ngi
ng ha
z
e
pa
tt
e
r
n
s
.
R
e
c
e
nt
a
dva
nc
e
s
in
d
e
e
p
-
le
a
r
ni
ng
-
ba
s
e
d
a
ppr
oa
c
he
s
ha
v
e
le
d
t
o
s
ig
ni
f
ic
a
nt
im
pr
ove
m
e
nt
s
in
im
a
ge
de
ha
z
in
g,
m
a
ki
ng
th
e
m
a
pow
e
r
f
ul
a
lt
e
r
na
ti
ve
f
or
va
r
io
us
a
ppl
ic
a
ti
ons
.
D
e
ha
z
e
N
e
t
[
5]
is
a
gr
oundbr
e
a
ki
ng
de
e
p l
e
a
r
ni
ng
-
ba
s
e
d
a
ppr
oa
c
h t
ha
t
e
s
ti
m
a
te
s
a
tm
os
phe
r
ic
li
ght
a
nd ha
z
e
t
r
a
ns
m
is
s
io
n m
a
p
s
f
r
om
ha
z
y i
m
a
ge
s
.
H
ow
e
ve
r
, a
pot
e
nt
ia
l
di
s
a
dva
nt
a
g
e
of
D
e
ha
z
e
N
e
t
i
s
t
he
ne
e
d f
or
a
r
e
f
in
e
m
e
nt
pr
oc
e
s
s
t
o e
nha
nc
e
t
he
a
c
c
ur
a
c
y
of
th
e
tr
a
ns
m
is
s
io
n
m
a
p,
w
hi
c
h
c
oul
d
a
dd
c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
.
B
ui
ld
in
g
on
th
e
s
u
c
c
e
s
s
of
D
e
ha
z
e
N
e
t,
R
e
n
e
t
al
.
[1
2
]
pr
opos
e
d
a
m
ul
ti
-
s
c
a
le
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
C
N
N
)
th
a
t
c
a
pt
ur
e
s
f
e
a
tu
r
e
s
a
t
va
r
io
us
le
ve
ls
of
de
ta
il
.
T
hi
s
s
tr
a
te
gy
e
f
f
e
c
ti
ve
ly
c
a
pt
ur
e
s
lo
c
a
l
a
nd
gl
oba
l
c
ont
e
xt
ua
l
in
f
or
m
a
ti
on,
bu
t
it
m
a
y
s
ti
ll
s
uf
f
e
r
f
r
om
s
om
e
lo
s
s
of
f
in
e
de
ta
il
s
due
to
th
e
f
ix
e
d
r
e
c
e
pt
iv
e
f
i
e
ld
s
iz
e
s
of
c
onvolut
io
na
l
f
il
te
r
s
.
D
e
noi
s
in
g
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
(
D
nC
N
N
)
-
e
nha
nc
e
[1
3
]
ut
il
is
e
s
r
e
s
id
ua
l
le
a
r
ni
ng
a
nd
ba
tc
h
nor
m
a
li
s
a
ti
on
to
e
nha
nc
e
im
a
ge
opt
im
is
a
ti
on.
W
hi
le
th
is
m
e
th
od
a
c
hi
e
ve
s
s
t
a
te
-
of
-
th
e
-
a
r
t
r
e
s
ul
ts
f
or
va
r
io
us
im
a
ge
r
e
s
to
r
a
ti
on
ta
s
ks
,
it
m
ig
ht
r
e
qui
r
e
s
ubs
ta
nt
ia
l
c
om
put
a
ti
ona
l
r
e
s
our
c
e
s
f
or
tr
a
in
in
g
due
to
its
de
e
p
ne
twor
k
a
r
c
hi
te
c
tu
r
e
.
A
ll
-
in
-
one
de
ha
z
in
g
ne
twor
k
(
A
O
D
-
N
e
t)
[
6]
is
a
c
om
pr
e
he
ns
iv
e
de
ha
z
in
g
ne
twor
k
th
a
t
di
r
e
c
tl
y
le
a
r
ns
th
e
m
a
ppi
ng
f
r
om
ha
z
y
to
h
a
z
e
-
f
r
e
e
im
a
ge
s
in
a
s
in
gl
e
s
t
e
p. T
he
a
d
va
nt
a
ge
of
A
O
D
-
N
e
t
is
it
s
a
bi
li
ty
to
e
li
m
in
a
te
th
e
e
xpl
ic
it
c
a
l
c
ul
a
ti
on
of
th
e
tr
a
ns
m
is
s
io
n
m
a
p
a
nd
a
tm
os
p
he
r
ic
li
ght
,
s
im
pl
if
yi
n
g
th
e
de
ha
z
in
g
pr
oc
e
s
s
.
H
ow
e
ve
r
,
a
li
m
it
a
ti
on
of
A
O
D
-
N
e
t
m
ig
h
t
be
it
s
r
e
li
a
nc
e
on
la
r
ge
-
s
c
a
le
da
ta
s
e
ts
f
or
tr
a
in
in
g,
w
hi
c
h
c
oul
d
be
c
ha
ll
e
ngi
ng
to
a
c
qui
r
e
in
s
om
e
c
a
s
e
s
.
G
r
id
D
e
ha
z
e
N
e
t
[
1
4
]
le
ve
r
a
ge
s
a
gr
id
s
tr
uc
tu
r
e
a
nd
m
ul
ti
-
s
c
a
le
da
ta
f
or
e
nha
nc
e
d
de
ha
z
in
g
pe
r
f
or
m
a
nc
e
.
W
hi
le
th
is
a
ppr
oa
c
h
e
f
f
e
c
ti
ve
ly
ha
ndl
e
s
c
om
pl
e
x
ha
z
e
pa
tt
e
r
ns
,
it
m
a
y
s
ti
ll
s
tr
uggl
e
w
it
h
th
ic
k
or
de
ns
e
s
c
e
ne
s
.
G
a
te
d
f
us
io
n
ne
twor
k
(
G
F
N
)
[
1
5
]
a
nd
its
a
da
pt
iv
e
f
us
io
n
of
m
u
lt
i
-
s
c
a
le
f
e
a
tu
r
e
s
th
r
ough
ga
ti
ng
m
e
c
ha
ni
s
m
s
e
f
f
e
c
ti
ve
ly
e
m
pha
s
is
e
in
f
or
m
a
ti
ve
f
e
a
tu
r
e
s
f
o
r
im
pr
ove
d
de
ha
z
in
g
pe
r
f
or
m
a
nc
e
.
H
ow
e
ve
r
,
one
pot
e
nt
ia
l
li
m
it
a
ti
on
o
f
th
is
m
e
th
o
d
is
th
a
t
th
e
ga
ti
ng
m
e
c
ha
ni
s
m
m
ig
ht
r
e
qui
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
E
nhanc
in
g r
oad image
c
la
r
it
y
w
it
h r
e
s
id
ual
ne
u
r
al
ne
tw
or
k
de
h
az
in
g m
ode
l
(
A
e
r
un M
ar
ti
n)
4149
c
a
r
e
f
ul
tu
ni
ng
a
nd
m
ig
ht
not
ge
ne
r
a
li
s
e
opt
im
a
ll
y
a
c
r
o
s
s
a
ll
ha
z
y
s
c
e
ne
s
.
K
im
e
t
al
.
[1
6
]
pr
e
s
e
nt
e
d a
s
p
a
ti
a
ll
y
a
da
pt
iv
e
a
tm
os
phe
r
ic
poi
nt
-
s
pr
e
a
d
f
unc
ti
on
(
A
P
S
F
)
te
c
hni
que
f
or
de
ha
z
in
g
r
oa
d
s
c
e
ne
im
a
ge
s
.
W
hi
le
th
is
a
ppr
oa
c
h
is
pr
a
c
ti
c
a
l
f
or
r
oa
d
s
c
e
ne
s
,
it
m
ig
ht
f
a
c
e
c
ha
ll
e
nge
s
in
de
a
li
ng
w
it
h
m
or
e
di
ve
r
s
e
a
nd
c
om
pl
e
x
na
tu
r
a
l
s
c
e
ne
s
.
T
h
e
a
dva
nt
a
ge
s
of
th
e
ne
w
m
e
th
od
s
,
s
u
c
h
a
s
A
O
D
-
N
e
t
a
nd
G
F
N
,
ove
r
pa
s
t
m
e
th
ods
li
ke
D
e
ha
z
e
N
e
t
a
nd
m
ul
ti
-
s
c
a
le
C
N
N
li
e
in
th
e
ir
a
bi
li
ty
to
ha
ndl
e
ha
z
e
r
e
m
ova
l
in
a
s
in
gl
e
s
te
p
w
i
th
out
e
xpl
ic
it
ly
e
s
ti
m
a
ti
ng
tr
a
ns
m
is
s
io
n
m
a
p
s
a
nd
a
tm
o
s
phe
r
ic
li
ght
.
T
hi
s
s
tr
e
a
m
li
ne
d
a
ppr
oa
c
h
s
im
pl
if
ie
s
th
e
de
h
a
z
in
g
pr
oc
e
s
s
a
nd
r
e
duc
e
s
c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
.
A
ddi
ti
ona
ll
y,
th
e
s
e
ne
w
e
r
m
e
th
ods
in
c
or
por
a
te
a
da
pt
iv
e
f
e
a
tu
r
e
f
us
io
n
a
nd
ga
ti
ng
m
e
c
ha
ni
s
m
s
,
e
n
a
bl
in
g
th
e
m
to
c
a
pt
ur
e
in
f
or
m
a
ti
ve
f
e
a
tu
r
e
s
m
or
e
e
f
f
e
c
ti
ve
ly
a
nd
im
pr
ove
de
ha
z
in
g
pe
r
f
or
m
a
nc
e
.
I
n
th
e
s
tr
e
a
m
li
ne
of
th
e
C
N
N
ne
twor
k,
M
e
ng
e
t
a
l.
[1
7
]
pr
opos
e
d
a
two
-
s
tr
e
a
m
C
N
N
or
s
in
gl
e
-
im
a
ge
de
ha
z
in
g.
T
hi
s
ne
twor
k
a
r
c
hi
te
c
tu
r
e
c
om
pr
is
e
s
a
s
p
a
ti
a
l
in
f
o
r
m
a
ti
on
f
e
a
tu
r
e
s
tr
e
a
m
a
nd
a
hi
gh
-
le
ve
l
s
e
m
a
nt
ic
f
e
a
tu
r
e
s
tr
e
a
m
.
T
he
s
pa
ti
a
l
i
nf
or
m
a
t
io
n f
e
a
t
ur
e
s
tr
e
a
m
pr
e
s
e
r
v
e
s
i
n
tr
i
c
a
te
d
e
ta
il
s
of
t
h
e
de
ha
z
e
d i
m
a
ge
,
w
hi
l
e
t
he
hi
gh
-
l
e
v
e
l
s
e
m
a
nt
ic
f
e
a
tu
r
e
s
tr
e
a
m
c
a
p
tu
r
e
s
m
u
lt
i
-
s
c
a
le
s
tr
uc
tu
r
a
l
f
e
a
tu
r
e
s
.
A
n
a
ux
il
i
a
r
y
m
od
ul
e
f
o
c
u
s
e
d
on
s
p
a
ti
a
l
in
f
or
m
a
ti
on
i
s
in
t
e
gr
a
t
e
d
b
e
tw
e
e
n
t
h
e
f
e
a
t
ur
e
s
tr
e
a
m
s
.
T
h
is
m
o
dul
e
e
m
p
lo
ys
a
n
a
tt
e
nt
io
n
m
e
c
h
a
n
i
s
m
t
o
uni
f
y
di
v
e
r
s
e
in
f
or
m
a
ti
on
t
yp
e
s
,
pr
ogr
e
s
s
iv
e
ly
r
e
s
to
r
in
g
im
a
g
e
c
la
r
it
y
w
it
h
s
e
m
a
n
ti
c
in
f
or
m
a
ti
on.
A
p
a
r
a
ll
e
l
r
e
s
id
ua
l
m
odu
le
e
nh
a
n
c
e
s
ha
z
e
d
is
c
r
i
m
in
a
t
io
n
b
y
d
e
h
a
z
in
g
f
e
a
tu
r
e
d
if
f
e
r
e
nc
e
s
a
c
r
o
s
s
s
ta
ge
s
. T
h
e
a
d
va
nc
e
m
e
n
ts
in
de
e
p
l
e
a
r
ni
ng
-
b
a
s
e
d
te
c
h
ni
qu
e
s
h
a
v
e
s
ho
w
n
gr
e
a
t
pr
om
i
s
e
i
n
a
d
dr
e
s
s
in
g
c
om
pl
e
x
ha
z
e
pa
tt
e
r
n
s
a
n
d
s
i
gni
f
ic
a
n
tl
y
im
pr
o
vi
n
g
i
m
a
ge
d
e
h
a
z
in
g
r
e
s
ul
t
s
.
H
o
w
e
ve
r
,
i
t
i
s
e
s
s
e
nt
i
a
l
to
c
on
s
i
de
r
e
a
c
h
m
e
th
od’
s
s
pe
c
if
ic
r
e
q
ui
r
e
m
e
nt
s
a
n
d
li
m
i
ta
ti
o
n
s
w
he
n
a
p
pl
y
in
g t
he
m
i
n d
if
f
e
r
e
nt
r
e
a
l
-
w
or
ld
s
c
e
n
a
r
io
s
.
R
e
s
N
e
t
[
7]
ha
ve
ga
in
e
d
popula
r
it
y
in
va
r
io
us
c
om
put
e
r
vi
s
io
n
ta
s
ks
,
in
c
lu
di
ng
im
a
ge
de
ha
z
in
g,
du
e
to
th
e
ir
e
f
f
e
c
ti
ve
de
e
p
f
e
a
tu
r
e
le
a
r
ni
ng
c
a
pa
bi
li
ti
e
s
a
nd
a
bi
li
t
y
to
m
it
ig
a
te
th
e
va
ni
s
hi
ng
gr
a
di
e
nt
pr
obl
e
m
.
Z
hou
e
t
al
.
[1
8
]
pr
opos
e
d
a
R
e
s
N
e
t
-
ba
s
e
d
m
ode
l
c
a
ll
e
d
a
m
ul
ti
-
s
c
a
le
r
e
s
id
ua
l
n
e
twor
k
(
M
S
R
N
)
f
or
s
i
ngl
e
-
im
a
g
e
de
ha
z
in
g,
w
hi
c
h
i
nt
e
gr
a
te
s
m
ul
ti
-
s
c
a
le
f
e
a
tu
r
e
s
f
r
om
di
f
f
e
r
e
n
t
n
e
twor
k
la
ye
r
s
to
c
a
pt
ur
e
h
a
z
e
d
is
tr
ib
ut
io
n
a
nd
s
c
e
ne
de
t
a
il
s
e
f
f
e
c
ti
v
e
ly
.
T
he
m
od
e
l
ou
tp
e
r
f
or
m
e
d
pr
e
vi
ou
s
s
t
a
te
-
of
-
th
e
-
a
r
t
m
e
t
hod
s
in
bot
h
qu
a
nt
it
a
ti
ve
a
nd
qua
li
t
a
ti
ve
e
v
a
lu
a
ti
on
s
.
Y
i
e
t
al
.
[
19
]
pr
e
s
e
nt
e
d
a
d
ua
l
r
e
s
i
dua
l
ne
twor
k
f
or
s
in
gl
e
-
im
a
ge
d
e
h
a
z
in
g
th
a
t
c
o
m
bi
ne
s
s
e
p
a
r
a
t
e
r
e
s
id
ua
l
ne
tw
or
ks
t
o
e
s
ti
m
a
t
e
tr
a
n
s
m
i
s
s
io
n
m
a
ps
a
nd
a
t
m
os
ph
e
r
ic
li
g
ht
.
B
y
ut
il
i
s
in
g
a
jo
in
t
lo
s
s
f
unc
t
io
n
c
ons
id
e
r
in
g
bo
th
tr
a
n
s
m
is
s
io
n
a
nd
a
tm
o
s
ph
e
r
ic
li
ght
e
s
ti
m
a
ti
ons
,
D
R
-
R
e
s
N
e
t
a
c
h
ie
v
e
s
im
pr
ov
e
d
de
h
a
z
i
ng
pe
r
f
or
m
a
n
c
e
.
R
e
s
N
e
t
-
ba
s
e
d
m
o
de
l
kno
w
n
a
s
th
e
c
a
s
c
a
de
d
r
e
s
id
ua
l
de
ns
e
ne
tw
or
k
(
C
R
D
N
)
[
2
0
]
e
m
pl
oys
a
c
a
s
c
a
de
d
s
tr
u
c
tu
r
e
t
o
r
e
f
in
e
th
e
de
ha
z
e
d
r
e
s
ul
t
it
e
r
a
ti
ve
l
y.
T
he
s
e
R
e
s
N
e
t
-
b
a
s
e
d
a
ppr
o
a
c
h
e
s
s
ho
w
c
a
s
e
th
e
ve
r
s
a
ti
li
ty
a
nd
e
f
f
e
c
ti
ve
ne
s
s
of
r
e
s
id
ua
l
le
a
r
ni
ng
in
im
a
ge
d
e
h
a
z
i
ng
ta
s
k
s
,
a
da
pt
i
ng
to
di
f
f
e
r
e
nt
ha
z
e
pa
tt
e
r
ns
a
nd
e
f
f
e
c
ti
v
e
ly
r
e
m
ovi
ng
ha
z
e
w
hi
le
pr
e
s
e
r
vi
ng
s
c
e
ne
de
t
a
il
s
.
I
n
th
e
r
e
a
lm
of
s
in
gl
e
-
im
a
ge
d
e
h
a
z
in
g
,
a
pr
om
in
e
nt
c
ha
ll
e
ng
e
i
n
lo
w
-
le
ve
l
c
om
p
ut
e
r
vi
s
i
on, r
e
c
e
nt
d
e
ve
l
opm
e
nt
s
ha
ve
i
nt
r
o
duc
e
d i
n
nov
a
ti
ve
a
p
pr
oa
c
he
s
t
o
a
d
dr
e
s
s
th
e
c
om
pl
e
xi
t
ie
s
of
th
e
pr
o
bl
e
m
.
P
yr
a
m
id
a
tt
e
n
ti
on
r
e
s
id
u
a
l
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
n
e
twor
k
(
P
A
R
F
E
-
N
e
t)
[2
1
]
m
ode
l
c
om
bi
n
e
s
a
f
e
a
tu
r
e
pyr
a
m
i
d
ne
tw
or
k
to
c
a
pt
ur
e
di
ve
r
s
e
im
a
ge
f
e
a
tu
r
e
s
a
c
r
os
s
di
f
f
e
r
e
nt
s
c
a
l
e
s
a
nd
a
n
a
tt
e
nt
i
on
-
ba
s
e
d r
e
s
bl
o
c
k t
ha
t
s
tr
a
te
gi
c
a
ll
y e
nh
a
n
c
e
s
c
r
u
c
ia
l
a
r
e
a
s
, i
m
pr
ovi
ng o
ut
put
i
m
a
g
e
qua
l
it
y.
R
e
s
e
a
r
c
h by
J
ia
n
g
e
t
al
.
[
2
2]
is
a
n
e
nd
-
to
-
e
nd
tr
a
in
a
bl
e
,
de
ns
e
ly
c
onn
e
c
t
e
d
r
e
s
id
ua
l
s
p
a
ti
a
l
a
nd
c
h
a
nne
l
a
tt
e
nt
io
n
n
e
twor
k
th
a
t
e
li
m
in
a
te
s
th
e
n
e
e
d
f
or
e
x
pl
ic
it
e
s
ti
m
a
ti
on
of
a
tm
os
p
he
r
ic
pa
r
a
m
e
te
r
s
a
nd
e
m
pl
oy
s
a
uni
que
r
e
s
id
ua
l
a
tt
e
nt
io
n
m
odul
e
t
o
dyn
a
m
ic
a
ll
y
a
dj
u
s
t
f
e
a
tu
r
e
w
e
ig
ht
s
, e
f
f
e
c
t
iv
e
l
y
f
oc
u
s
in
g
on
r
e
g
io
n
s
of
im
por
t
a
nc
e
.
W
a
ng
e
t
al
.
[2
3
]
a
ddr
e
s
s
e
s
dom
a
i
n
s
h
if
t
is
s
ue
s
b
e
twe
e
n
s
ynt
h
e
ti
c
a
nd
r
e
a
l
-
w
or
l
d
ha
z
y
im
a
ge
s
th
r
o
ugh
th
e
twof
ol
d
m
ul
ti
-
s
c
a
l
e
ge
ne
r
a
ti
ve
a
d
ve
r
s
a
r
ia
l
ne
t
w
or
k
(
T
M
S
-
G
A
N
)
f
r
a
m
e
w
or
k.
T
hi
s
d
ua
l
-
s
ta
g
e
tr
a
in
in
g
s
c
he
m
e
c
om
pr
i
s
e
s
a
ha
z
e
-
ge
n
e
r
a
ti
o
n
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
i
a
l
n
e
twor
k
(
H
gG
A
N
)
a
nd
a
ha
z
e
-
r
e
m
ova
l
ge
n
e
r
a
ti
v
e
a
d
ve
r
s
a
r
ia
l
ne
t
w
or
k
(
H
r
G
A
N
)
w
or
ki
ng
c
oop
e
r
a
ti
ve
ly
to
g
e
ne
r
a
te
a
nd
r
e
m
ov
e
ha
z
e
,
a
id
e
d
by
s
tr
uc
tu
r
a
l
a
dv
a
nc
e
m
e
nt
s
li
k
e
a
ha
z
e
r
e
s
id
u
a
l
m
a
p
a
n
d
a
m
ul
t
i
-
a
tt
e
nt
io
n
pr
ogr
e
s
s
iv
e
f
u
s
io
n
m
odu
le
(
M
A
P
F
M
)
f
or
e
f
f
e
c
ti
ve
f
e
a
t
ur
e
f
us
i
on.
T
h
e
s
e
in
nova
t
iv
e
a
ppr
oa
c
he
s
c
ol
le
c
ti
v
e
ly
pus
h
th
e
b
ound
a
r
ie
s
of
s
in
gl
e
-
im
a
ge
de
h
a
z
i
ng,
s
ho
w
c
a
s
i
ng
pr
o
m
is
i
ng
r
e
s
ul
ts
a
nd a
dva
nc
e
m
e
nt
s
ov
e
r
e
xi
s
t
in
g t
e
c
h
ni
que
s
.
W
it
h
th
e
in
c
r
e
a
s
in
g
de
ha
z
in
g
m
e
th
ods
in
te
gr
a
te
d
w
it
h
r
obus
t
ba
s
e
ne
twor
ks
s
uc
h
a
s
C
N
N
a
nd
R
e
s
N
e
t,
th
e
e
nd
-
to
-
e
nd
a
r
c
hi
te
c
tu
r
e
w
it
h
in
c
r
e
a
s
e
d
a
c
c
ur
a
c
y
a
nd
le
s
s
c
om
put
a
ti
on
s
ti
ll
ha
s
a
hi
gh
pot
e
nt
ia
l
to
ta
c
kl
e
va
r
io
us
ha
z
e
pa
tt
e
r
ns
.
T
hi
s
te
c
hni
que
is
m
or
e
pr
a
c
ti
c
a
l
in
r
e
a
l
-
w
or
ld
s
c
e
na
r
io
s
,
r
e
qui
r
in
g
m
in
im
a
l
ti
m
e
to
pr
oc
e
s
s
th
e
ha
z
y
im
a
ge
a
nd
out
put
hi
gh
-
qua
li
ty
de
ha
z
e
im
a
ge
s
to
yi
e
ld
be
tt
e
r
r
e
s
ul
ts
a
nd
c
ont
r
ib
ut
e
to
a
ut
onomous
ve
hi
c
le
s
a
nd A
D
A
S
’
s
s
a
f
e
ty
a
nd e
f
f
ic
ie
nc
y.
3.
M
E
T
H
O
D
H
a
z
y
im
a
ge
s
of
te
n
s
uf
f
e
r
c
ol
our
de
vi
a
ti
on
a
nd
lo
w
c
ont
r
a
s
t
due
to
s
e
le
c
ti
ve
a
bs
or
pt
io
n
a
nd
li
ght
s
c
a
tt
e
r
in
g,
w
hos
e
de
gr
a
d
a
ti
on
is
g
e
ne
r
a
ll
y
de
s
c
r
ib
e
d
by
a
n
A
S
M
.
I
n
(
1)
r
e
pr
e
s
e
nt
s
th
e
m
a
th
e
m
a
ti
c
a
l
e
xpr
e
s
s
io
n
f
or
t
he
A
S
M
.
(
)
=
(
)
(
)
+
(
1
−
(
)
)
(
1)
W
he
r
e
, I
(
x)
r
e
pr
e
s
e
nt
s
t
he
obs
e
r
ve
d ha
z
y i
m
a
ge
,
(
)
r
e
pr
e
s
e
nt
s
t
he
s
c
e
ne
r
a
di
a
nc
e
or
“
c
le
a
n i
m
a
ge
”
r
e
qui
r
e
d
to
r
e
c
ove
r
,
r
e
pr
e
s
e
nt
s
th
e
gl
oba
l
a
tm
os
phe
r
ic
li
ght
,
a
nd
(
)
r
e
pr
e
s
e
nt
s
th
e
tr
a
ns
m
is
s
io
n
m
a
p.
T
he
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024:
4147
-
4157
4150
tr
a
ns
m
is
s
io
n m
a
p pl
a
ys
a
c
r
uc
ia
l
r
ol
e
i
n
A
S
M
.
I
t
de
te
r
m
in
e
s
ho
w
m
uc
h l
ig
ht
i
s
a
tt
e
nua
te
d a
nd s
c
a
tt
e
r
e
d w
he
n
it
t
r
a
ve
ls
t
hr
ough a
ha
z
y m
e
di
um
. I
n e
qua
ti
on f
or
m
, t
he
t
r
a
ns
m
is
s
io
n m
a
p c
a
n b
e
e
xpr
e
s
s
e
d a
s
(
2)
:
(
)
=
−
(
)
(
2)
W
he
r
e
r
e
pr
e
s
e
nt
s
t
h
e
s
c
a
tt
e
r
in
g c
oe
f
f
ic
ie
nt
of
t
he
a
tm
os
phe
r
e
,
a
nd
(
)
r
e
pr
e
s
e
nt
s
t
he
di
s
ta
nc
e
b
e
twe
e
n t
he
obj
e
c
t
a
nd
th
e
c
a
m
e
r
a
.
T
he
A
S
M
i
s
a
f
unda
m
e
nt
a
l
c
on
c
e
pt
u
s
e
d
to
unde
r
s
ta
nd
ha
z
y
im
a
ge
f
or
m
a
ti
on
a
nd
i
s
c
om
m
onl
y
e
m
pl
oye
d
in
im
a
ge
r
e
s
to
r
a
ti
on
a
nd
e
nha
nc
e
m
e
nt
.
T
he
c
le
a
n
im
a
ge
(
)
c
a
n
be
obt
a
in
e
d
f
r
om
th
e
ha
z
y i
m
a
ge
(
)
us
in
g t
he
t
r
a
ns
m
is
s
io
n m
a
p
(
)
a
nd a
tm
os
phe
r
ic
l
ig
ht
a
c
c
or
di
ng t
o (
3)
[
6]
:
(
)
=
1
(
)
(
)
−
1
(
)
+
(
3)
T
he
a
lg
or
it
hm
of
f
e
r
s
a
s
ig
ni
f
ic
a
nt
a
dva
nt
a
g
e
by mi
ni
m
is
in
g t
he
r
e
c
ons
tr
uc
ti
on e
r
r
or
(
)
di
r
e
c
tl
y on
th
e
c
le
a
n
im
a
ge
(
)
,
w
it
hout
th
e
ne
e
d
f
or
s
e
pa
r
a
te
e
s
ti
m
a
ti
on
of
th
e
tr
a
ns
m
is
s
io
n
m
a
p
t(
x)
a
nd
a
tm
os
phe
r
ic
li
ght
.
B
y
c
om
bi
ni
ng
th
e
e
s
ti
m
a
ti
on
of
(
)
a
nd
,
th
e
y
e
f
f
e
c
ti
ve
ly
r
e
d
uc
e
th
e
r
e
c
on
s
tr
uc
ti
on
e
r
r
or
a
t
th
e
pi
xe
l
le
ve
l
of
t
he
i
m
a
ge
. T
hi
s
w
a
s
a
c
c
om
pl
is
he
d u
s
in
g
(
4)
:
(
)
=
(
)
(
)
+
(
)
+
(
4)
w
he
r
e
(
)
=
1
(
)
(
(
)
−
)
+
(
−
)
(
)
−
1
(
5)
r
e
pr
e
s
e
nt
s
a
c
ons
ta
nt
bi
a
s
t
e
r
m
i
n t
he
a
lg
or
it
hm
.
T
hi
s
t
e
r
m
pl
a
ys
a
r
ol
e
i
n a
dj
us
ti
ng t
he
r
e
c
ons
tr
uc
ti
on e
r
r
or
,
c
ont
r
ib
ut
in
g
to
th
e
ove
r
a
ll
a
c
c
ur
a
c
y
of
th
e
de
ha
z
in
g
pr
oc
e
s
s
.
T
he
jo
in
t
e
s
ti
m
a
ti
on
pr
oc
e
s
s
c
om
pe
n
s
a
te
s
f
or
in
a
c
c
ur
a
c
ie
s
i
n ot
he
r
hype
r
pa
r
a
m
e
te
r
s
, s
uc
h a
s
ga
m
m
a
c
or
r
e
c
ti
on, e
nha
nc
in
g ove
r
a
ll
i
m
a
ge
qua
li
ty
.
T
he
R
e
s
N
e
t
a
r
c
hi
te
c
tu
r
e
w
a
s
c
ho
s
e
n
due
to
it
s
a
bi
li
ty
to
tr
a
in
de
e
p
ne
twor
ks
by
a
ll
e
vi
a
ti
ng
th
e
va
ni
s
hi
ng
gr
a
di
e
nt
pr
obl
e
m
.
T
he
ne
twor
k
is
tr
a
in
e
d
u
s
in
g
th
e
A
da
m
opt
im
is
e
r
f
or
te
n
e
poc
hs
,
w
it
h
a
le
a
r
ni
ng
r
a
te
1e
-
4
. T
he
ba
tc
h s
iz
e
i
s
s
e
t
to
8. T
he
A
da
m
opt
im
is
e
r
, s
hor
t
f
or
a
da
pt
iv
e
m
om
e
nt
e
s
ti
m
a
ti
on
, ha
s
be
c
om
e
a
pr
e
f
e
r
r
e
d c
hoi
c
e
i
n de
e
p l
e
a
r
ni
ng due
t
o i
t
is
di
s
ti
nc
t
a
dva
nt
a
ge
s
. A
t
it
i
s
c
or
e
, A
da
m
s
e
a
m
le
s
s
ly
c
om
bi
ne
s
t
he
s
tr
e
ngt
hs
of
two
popula
r
e
xt
e
n
s
io
ns
of
s
to
c
ha
s
ti
c
gr
a
di
e
nt
de
s
c
e
nt
:
m
om
e
nt
um
a
nd
r
oot
m
e
a
n
s
qua
r
e
pr
opa
ga
ti
on
(
R
M
S
pr
op
)
.
W
hi
le
m
om
e
nt
um
le
ve
r
a
ge
s
pa
s
t
gr
a
di
e
nt
s
to
s
m
oot
h
out
th
e
upda
te
p
r
oc
e
s
s
,
R
M
S
pr
op
dyna
m
ic
a
ll
y
a
dj
us
ts
th
e
le
a
r
ni
ng
r
a
te
ba
s
e
d
on
r
e
c
e
nt
gr
a
di
e
nt
m
a
gni
tu
de
s
.
T
hi
s
bl
e
nd
r
e
s
ul
ts
in
pe
r
-
pa
r
a
m
e
te
r
a
da
pt
iv
e
l
e
a
r
ni
ng r
a
te
s
, m
a
ki
ng A
d
a
m
pa
r
ti
c
ul
a
r
l
y e
f
f
e
c
ti
ve
f
or
pr
obl
e
m
s
w
it
h
s
pa
r
s
e
gr
a
di
e
nt
s
,
w
hi
c
h
is
s
ta
nda
r
d
in
c
om
put
e
r
vi
s
io
n.
A
not
he
r
pi
vot
a
l
f
e
a
tu
r
e
of
A
da
m
is
it
s
bi
a
s
-
c
or
r
e
c
ti
on
m
e
c
ha
ni
s
m
,
w
hi
c
h
c
or
r
e
c
ts
t
he
i
ni
ti
a
l
z
e
r
o bi
a
s
i
n i
t
s
m
ovi
ng a
ve
r
a
ge
s
, e
n
s
ur
in
g a
m
or
e
a
c
c
ur
a
te
c
onve
r
ge
nc
e
.
T
he
R
e
s
N
e
t
a
r
c
hi
te
c
tu
r
e
is
ba
s
e
d
on
it
s
uni
que
r
e
s
id
ua
l
bl
oc
k
s
tr
uc
tu
r
e
.
E
a
c
h
bl
oc
k
c
ons
is
ts
of
two
c
onvolut
io
na
l
la
ye
r
s
,
f
ol
lo
w
e
d
by
ba
tc
h
nor
m
a
li
s
a
ti
on
a
nd
r
e
c
ti
f
ie
d
li
ne
a
r
uni
t
(
R
e
L
U
)
a
c
ti
va
ti
on.
T
he
f
ir
s
t
la
ye
r
us
e
s
a
1×
1 ke
r
ne
l
w
it
h s
tr
id
e
s
, w
hi
le
t
h
e
s
e
c
ond la
ye
r
us
e
s
a
3x3 ke
r
ne
l
w
it
h a
s
tr
id
e
of
1. T
he
out
put
i
s
a
ugm
e
nt
e
d
by
a
n
in
te
gr
a
l
s
hor
tc
ut
c
onne
c
ti
on,
w
hi
c
h
in
c
lu
de
s
a
1
×
1
c
onvolut
io
na
l
la
ye
r
a
li
gne
d
w
it
h
th
e
bl
oc
k’
s
f
il
te
r
c
ount
a
nd
b
a
tc
h
nor
m
a
li
s
a
ti
on.
T
hi
s
de
s
ig
n
a
ll
ow
s
th
e
n
e
twor
k
to
le
a
r
n
id
e
nt
it
y
f
unc
ti
ons
,
pr
e
ve
nt
in
g pe
r
f
or
m
a
nc
e
de
gr
a
da
ti
on w
he
n a
ddi
ng l
a
ye
r
s
. T
he
n
e
twor
k s
ta
r
ts
w
it
h a
7×
7 ke
r
ne
l,
s
tr
id
e
2, a
nd a
2D
c
onvolut
io
na
l
la
ye
r
,
f
ol
lo
w
e
d
by
ba
tc
h
nor
m
a
li
s
a
ti
on
a
nd
R
e
L
U
a
c
ti
va
ti
on.
A
3×
3
ke
r
ne
l
f
ol
lo
w
s
th
is
,
s
tr
id
e
two
m
a
x
pool
in
g
.
S
ubs
e
que
nt
s
ta
ge
s
c
ons
i
s
t
of
m
ul
ti
pl
e
r
e
s
id
ua
l
bl
oc
ks
,
in
c
r
e
a
s
in
g
th
e
f
il
te
r
c
ount
w
hi
le
r
e
duc
in
g
s
pa
ti
a
l
di
m
e
ns
io
ns
.
A
f
te
r
e
a
c
h
s
ta
ge
,
th
e
f
il
te
r
c
ount
double
s
w
hi
le
th
e
di
m
e
ns
io
ns
ha
lv
e
.
T
he
ne
twor
k
r
e
ve
r
s
e
s
th
e
s
pa
ti
a
l
r
e
duc
ti
on
us
in
g
tr
a
ns
pos
e
c
onvol
ut
io
na
l
la
ye
r
s
f
or
ups
a
m
pl
in
g,
f
ol
lo
w
e
d
by
ba
tc
h
nor
m
a
li
s
a
ti
on
a
nd
R
e
L
U
a
c
ti
va
ti
on.
T
he
num
be
r
of
f
il
te
r
s
in
th
e
s
e
la
ye
r
s
de
c
r
e
a
s
e
s
s
im
il
a
r
ly
to
th
e
in
it
ia
l
s
ta
ge
s
.
T
h
e
f
in
a
l
la
ye
r
is
a
3
-
f
i
lt
e
r
c
onvolut
io
na
l
la
ye
r
th
a
t
m
a
tc
he
s
th
e
in
put
im
a
ge
c
ha
nne
ls
w
it
h
ba
tc
h
nor
m
a
li
s
a
ti
on.
A
s
pe
c
if
ic
s
te
p
f
or
ha
z
e
r
e
m
ova
l
c
om
bi
ne
s
th
e
o
ut
put
of
th
e
f
in
a
l
la
ye
r
w
it
h
th
e
or
ig
in
a
l
in
put
,
s
ubt
r
a
c
ti
ng t
he
e
s
ti
m
a
te
d ha
z
e
f
r
om
t
he
ha
z
y i
m
a
ge
t
o r
e
s
to
r
e
c
l
a
r
it
y.
T
he
obj
e
c
ti
ve
f
unc
ti
on
f
or
tr
a
in
in
g
th
e
ne
twor
k
is
th
e
m
e
a
n
s
qua
r
e
d
e
r
r
or
(
M
S
E
)
be
twe
e
n
th
e
de
ha
z
e
d
out
put
i
m
a
ge
a
nd t
he
c
or
r
e
s
ponding gr
ound
-
tr
ut
h c
le
a
r
i
m
a
ge
.
T
he
out
put
i
m
a
ge
s
a
r
e
r
e
s
iz
e
d
t
o m
a
tc
h
t
he
s
i
z
e
of
th
e
gr
ound
t
r
ut
h
im
a
ge
be
f
or
e
c
om
put
in
g
th
e
M
S
E
.
T
he
A
da
m
opt
im
is
a
ti
on
a
lg
or
it
hm
is
us
e
d
to
m
in
im
is
e
th
e
lo
s
s
.
M
S
E
is
c
om
m
onl
y
us
e
d
f
or
r
e
gr
e
s
s
io
n
ta
s
k
s
due
to
it
i
s
s
im
pl
ic
it
y,
e
m
pha
s
is
on
m
or
e
s
ig
ni
f
ic
a
nt
e
r
r
or
s
,
a
nd
s
m
oot
h
gr
a
di
e
nt
pr
ope
r
ti
e
s
,
w
hi
c
h
m
a
ke
it
s
ui
ta
bl
e
f
or
opt
im
is
a
ti
on.
I
n
a
ppl
ic
a
ti
ons
li
ke
im
a
ge
h
a
z
e
r
e
m
ova
l,
th
e
goa
l
i
s
to
m
in
im
is
e
th
e
de
vi
a
ti
on
b
e
twe
e
n
th
e
de
ha
z
e
d
out
put
a
nd
th
e
pr
e
c
i
s
e
r
e
f
e
r
e
nc
e
im
a
ge
.
M
S
E
na
tu
r
a
ll
y
qua
nt
if
ie
s
th
is
d
e
vi
a
ti
on
by
c
a
lc
ul
a
ti
ng
th
e
a
ve
r
a
ge
s
qua
r
e
d
di
f
f
e
r
e
nc
e
be
twe
e
n
th
e
two.
I
t
i
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
E
nhanc
in
g r
oad image
c
la
r
it
y
w
it
h r
e
s
id
ual
ne
u
r
al
ne
tw
or
k
de
h
az
in
g m
ode
l
(
A
e
r
un M
ar
ti
n)
4151
pe
na
li
s
a
ti
on
of
m
or
e
s
ig
ni
f
ic
a
nt
e
r
r
or
s
e
ns
ur
e
s
th
a
t
th
e
m
ode
l
a
im
s
f
or
pr
e
c
is
io
n
in
h
a
z
e
r
e
m
ova
l,
w
hi
c
h
is
c
r
uc
ia
l
f
or
i
m
a
ge
c
la
r
it
y. T
he
M
S
E
e
qua
ti
on i
s
a
s
(
6)
:
=
∑
(
=
1
−
̂
)
2
(
6)
T
o
a
ddr
e
s
s
th
e
is
s
ue
of
e
xpl
odi
ng
gr
a
di
e
nt
s
of
te
n
e
nc
ount
e
r
e
d
in
de
e
p
ne
twor
ks
,
th
e
pa
pe
r
im
pl
e
m
e
nt
e
d
gr
a
di
e
nt
c
li
ppi
ng
by
li
m
it
in
g
th
e
L
2
nor
m
of
th
e
gr
a
di
e
nt
s
to
0.1
a
f
te
r
c
om
put
a
ti
on.
A
f
te
r
tr
a
in
in
g
f
or
te
n
e
poc
hs
,
th
e
r
e
s
id
ua
l
ne
ur
a
l
ne
twor
k
de
h
a
z
e
(
R
N
N
-
D
)
pr
oc
e
s
s
e
d
ha
z
y
im
a
ge
s
a
nd
ge
ne
r
a
te
d
r
e
s
ul
ts
f
r
om
it
s
f
in
a
l
la
ye
r
.
T
he
ne
twor
k
c
ons
is
te
d
of
101
la
ye
r
s
w
it
h
8,911,
369
pa
r
a
m
e
te
r
s
a
nd
w
a
s
c
us
to
m
is
e
d
to
ha
ndl
e
s
pe
c
if
ic
ha
z
e
c
h
a
r
a
c
te
r
is
ti
c
s
.
U
nl
ik
e
g
e
ne
r
ic
a
r
c
hi
te
c
tu
r
e
s
,
th
e
ta
il
or
e
d
la
ye
r
s
e
f
f
e
c
ti
ve
l
y
a
ddr
e
s
s
e
d
th
e
im
pa
c
t
of
ha
z
e
on
c
ont
r
a
s
t,
c
ol
our
,
a
nd
d
e
ta
il
.
B
y
ut
il
is
in
g
r
e
s
id
ua
l
bl
oc
k
s
,
th
e
ne
twor
k
le
a
r
ne
d
in
tr
ic
a
te
r
e
pr
e
s
e
nt
a
ti
ons
w
hi
le
pr
e
s
e
r
vi
ng
im
por
ta
nt
in
put
in
f
o
r
m
a
ti
on
th
r
ough
s
ki
p
c
onn
e
c
ti
ons
,
w
hi
c
h
a
id
e
d
in
r
e
c
ove
r
in
g
f
in
e
de
ta
il
s
.
U
ps
a
m
pl
i
ng
us
in
g
tr
a
ns
pos
e
c
onvolut
io
na
l
la
y
e
r
s
r
e
ta
in
e
d
c
r
uc
ia
l
s
pa
ti
a
l
in
f
or
m
a
ti
on
ne
c
e
s
s
a
r
y
to
r
e
s
to
r
e
or
ig
in
a
l
te
xt
ur
e
s
a
nd de
ta
il
s
obs
c
ur
e
d by h
a
z
e
.
T
he
a
r
c
hi
te
c
tu
r
e
of
t
he
R
N
N
-
D
i
s
s
how
n i
n F
ig
ur
e
1.
F
ig
ur
e
1. R
e
s
N
e
t
-
ba
s
e
d d
e
ha
z
in
g a
r
c
hi
te
c
tu
r
e
w
it
h k
-
e
s
ti
m
a
ti
on f
or
i
m
pr
ove
d i
m
a
ge
c
la
r
it
y
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024:
4147
-
4157
4152
4.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
A
c
c
e
s
s
to
hi
gh
-
qua
li
ty
tr
a
in
in
g
da
ta
is
c
r
uc
i
a
l
f
or
e
f
f
e
c
ti
ve
tr
a
in
in
g
m
ode
ls
a
nd
a
c
hi
e
vi
ng
r
obus
t
a
nd
a
c
c
ur
a
te
de
h
a
z
in
g
r
e
s
ul
t
s
[
2
2
]
.
T
hi
s
s
tu
dy
ut
il
is
e
s
th
e
N
Y
U
2
d
a
ta
s
e
t
to
ov
e
r
c
om
e
th
is
c
ha
ll
e
nge
f
or
tr
a
in
in
g
pur
pos
e
s
.
T
he
N
Y
U
2
da
ta
s
e
t
pr
ovi
de
s
pr
e
c
is
e
tr
a
in
in
g
d
a
ta
,
a
ll
o
w
in
g
th
e
m
ode
l
to
le
a
r
n
e
f
f
e
c
ti
ve
ly
a
nd
pr
oduc
e
e
nha
nc
e
d
r
e
s
ul
ts
in
im
a
g
e
de
ha
z
in
g
ta
s
k
s
[
19
]
,
[
24
]
.
T
he
tr
a
in
in
g
s
e
t
c
on
s
is
ts
of
27,256
im
a
ge
s
f
r
om
th
e
N
Y
U
2
da
ta
s
e
t,
w
hi
le
th
e
te
s
t
s
e
t
in
c
lu
de
s
3,170
non
-
ove
r
la
ppi
ng
im
a
ge
s
.
T
h
e
s
e
d
a
ta
s
e
t
s
c
ove
r
di
ve
r
s
e
in
door
s
c
e
n
e
s
w
it
h va
r
yi
ng l
e
ve
ls
of
ha
z
e
. H
ow
e
ve
r
, i
t
i
s
e
s
s
e
nt
i
a
l
to
not
e
t
ha
t
th
e
N
Y
U
2 d
a
ta
s
e
t
pr
im
a
r
il
y f
oc
us
e
s
on i
ndoor
s
c
e
ne
s
.
W
he
n
th
e
s
e
m
ode
l
s
a
r
e
te
s
te
d
on
out
doo
r
da
ta
s
e
t
s
,
s
uc
h
a
s
th
os
e
c
ont
a
in
in
g
r
oa
ds
,
a
c
c
ur
a
c
y
m
a
y
de
c
r
e
a
s
e
due
to
th
e
dom
a
in
s
hi
f
t
be
twe
e
n
in
door
a
nd
out
doo
r
i
m
a
ge
s
.
O
ut
door
im
a
ge
s
of
te
n
e
xhi
bi
t
di
f
f
e
r
e
nt
li
ght
in
g
c
ondi
ti
ons
a
nd
a
tm
o
s
phe
r
ic
e
f
f
e
c
ts
c
om
pa
r
e
d
to
th
e
in
do
or
da
ta
s
e
t,
im
pa
c
ti
ng
th
e
m
o
de
l'
s
pe
r
f
or
m
a
nc
e
.
T
he
tr
a
in
in
g
s
e
ts
w
e
r
e
e
xpa
nd
e
d
to
in
c
lu
de
th
e
R
E
S
I
D
E
da
ta
s
e
t
to
a
ddr
e
s
s
th
is
li
m
it
a
ti
on
a
nd
im
pr
ove
th
e
m
ode
l’
s
a
da
pt
a
bi
li
ty
to
a
br
oa
de
r
r
a
nge
of
e
nvi
r
onm
e
nt
s
[
2
5
]
.
T
he
R
E
S
I
D
E
da
ta
s
e
t
pr
ovi
de
s
a
m
or
e
di
v
e
r
s
e
s
e
t
of
im
a
ge
s
f
or
t
r
a
in
in
g,
s
pe
c
if
ic
a
ll
y
th
e
out
door
tr
a
in
in
g
s
e
t
(
O
T
S
)
,
w
hi
c
h
c
ons
is
ts
of
72,135
ha
z
y
a
nd
c
or
r
e
s
ponding
ha
z
e
-
f
r
e
e
im
a
ge
s
.
I
t
w
a
s
c
om
p
a
r
e
d
to
e
ig
ht
s
t
a
te
-
of
-
th
e
-
a
r
t
m
e
th
ods
us
in
g
49
c
le
a
n
im
a
ge
s
a
ugm
e
nt
e
d
w
it
h
s
ynt
he
ti
c
ha
z
e
to
e
va
lu
a
te
th
e
m
e
th
od’
s
p
e
r
f
or
m
a
nc
e
.
T
o
e
ns
ur
e
a
f
a
ir
c
om
pa
r
is
on,
1,400
tr
a
in
in
g
im
a
ge
s
f
r
om
th
e
N
Y
U
2
da
ta
s
e
t
w
e
r
e
e
xc
lu
de
d.
T
he
e
va
lu
a
ti
on
ut
il
is
e
d
two
s
t
a
nda
r
d
m
e
tr
ic
s
:
S
S
I
M
a
nd P
S
N
R
[
2
6
].
=
10
10
(
7)
(
,
)
=
(
2
+
1
)
(
2
,
+
2
)
(
2
+
2
+
1
)
(
2
+
2
+
2
)
(
8)
D
ur
in
g
th
e
e
va
lu
a
ti
on
,
t
hr
e
e
pr
io
r
-
b
a
s
e
d
m
e
th
od
s
[
4]
,
[
1
0
]
,
[
27
]
a
n
d
f
iv
e
le
a
r
ni
ng
-
b
a
s
e
d
m
e
th
ods
[
5]
,
[
6]
,
[1
2
]
,
[
15
]
,
[
2
8]
w
e
r
e
c
on
s
id
e
r
e
d
f
or
c
om
pa
r
i
s
on
.
T
he
f
ir
s
t
ba
s
e
li
n
e
a
ppr
oa
c
h,
D
C
P
,
e
s
ti
m
a
te
s
a
ha
z
e
-
f
r
e
e
im
a
ge
us
i
ng
th
e
lo
c
a
l
m
i
ni
m
um
s
ta
ti
s
ti
c
s
in
th
e
in
put
im
a
ge
’
s
ha
z
e
-
r
e
le
va
nt
c
h
a
nn
e
l.
B
ound
a
r
y
-
c
o
ns
tr
a
in
e
d
c
ont
e
xt
r
e
gul
a
r
is
a
ti
on
(
B
C
C
R
)
c
om
bi
ne
s
c
o
nt
e
x
tu
a
l
in
f
or
m
a
ti
on
a
nd
g
r
a
di
e
nt
s
to
e
s
ti
m
a
te
th
e
tr
a
n
s
m
is
s
io
n
m
a
p.
T
he
nonl
o
c
a
l
im
a
g
e
de
ha
z
in
g
(
N
L
D
)
a
lg
or
it
hm
c
o
ns
id
e
r
s
nonl
oc
a
l
s
e
lf
-
s
im
il
a
r
it
y
a
nd
te
xt
ur
e
de
t
a
il
s
dur
in
g
de
ha
z
in
g. A
m
ong the
l
e
a
r
ni
n
g
-
ba
s
e
d
m
e
th
ods
,
c
ol
or
a
tt
e
nua
ti
on
pr
io
r
(
C
A
P
)
e
m
pl
oy
s
a
li
n
e
a
r
m
ode
l
b
a
s
e
d
o
n
t
he
a
s
s
um
p
ti
on
of
c
ol
our
c
o
ns
i
s
t
e
nc
y
a
c
r
o
s
s
di
f
f
e
r
e
nt
r
e
gi
on
s
.
D
e
ha
z
e
N
e
t
a
nd
m
ul
ti
-
s
c
a
le
c
o
nvol
ut
io
na
l
n
e
ur
a
l
ne
twor
k
(
M
S
C
N
N
)
a
r
e
C
N
N
-
b
a
s
e
d
m
e
th
od
s
th
a
t
l
e
a
r
n
t
he
m
a
ppi
ng
be
t
w
e
e
n
h
a
z
y
i
nput
s
a
nd
th
e
ir
tr
a
n
s
m
i
s
s
io
n.
AOD
-
N
e
t
in
t
e
gr
a
t
e
s
tr
a
n
s
m
i
s
s
i
on
a
nd
a
t
m
os
phe
r
i
c
li
gh
t
in
to
a
n
e
w
va
r
i
a
bl
e
f
or
ha
z
e
-
f
r
e
e
im
a
ge
e
s
ti
m
a
ti
o
n.
G
F
N
us
e
s
a
f
us
i
on
-
ba
s
e
d
s
tr
a
t
e
gy
th
a
t
in
c
or
por
a
te
s
w
hi
t
e
b
a
l
a
nc
e
(
W
B
)
,
c
on
tr
a
s
t
e
nha
nc
e
m
e
nt
(
C
E
)
,
a
n
d
g
a
m
m
a
c
or
r
e
c
ti
on
(
G
C
)
to
de
r
iv
e
c
onf
id
e
nc
e
m
a
p
s
.
T
he
s
e
m
a
p
s
a
r
e
th
e
n
us
e
d
to
b
le
nd
in
f
or
m
a
t
io
n
a
nd
pr
e
s
e
r
ve
r
e
gi
o
ns
w
it
h
im
pr
ove
d
vi
s
ib
il
it
y,
r
e
s
ul
ti
ng
in
th
e
f
in
a
l
d
e
h
a
z
e
d
im
a
ge
.
A
c
om
p
a
r
a
ti
v
e
a
n
a
ly
s
is
of
va
r
io
u
s
de
h
a
z
i
ng
te
c
hni
que
s
,
i
nc
lu
d
in
g
pr
io
r
-
b
a
s
e
d
m
e
th
od
s
a
nd
l
e
a
r
ni
n
g
-
ba
s
e
d
m
e
th
od
s
,
w
a
s
c
on
duc
t
e
d,
a
nd
c
om
p
a
r
e
d
w
it
h
R
N
N
-
D
.
T
he
r
e
s
ul
t
s
s
h
ow
e
d
th
a
t
pr
io
r
-
ba
s
e
d
m
e
th
od
s
t
e
nd
to
ov
e
r
e
s
t
im
a
t
e
th
e
ha
z
e
th
ic
kne
s
s
,
l
e
a
di
ng
t
o
d
a
r
ke
r
,
de
ha
z
e
d
i
m
a
g
e
s
w
it
h
c
ol
our
di
s
to
r
ti
o
ns
in
c
e
r
ta
i
n r
e
gi
ons
. A
lt
ho
ugh C
A
P
, D
e
ha
z
e
N
e
t,
M
S
C
N
N
, a
n
d A
O
D
-
N
e
t
pr
oduc
e
d
r
e
s
ul
t
s
c
lo
s
e
r
to
th
e
gr
ound
tr
ut
h,
r
e
s
id
ua
l
h
a
z
e
r
e
m
a
i
ne
d
i
n
th
e
d
e
ha
z
e
d
im
a
ge
s
.
O
n
th
e
ot
he
r
ha
nd
,
th
e
R
N
N
-
D
a
ppr
oa
c
h
a
nd
G
F
N
pr
od
uc
e
d
d
e
ha
z
e
d
im
a
ge
s
t
ha
t
c
lo
s
e
ly
r
e
s
e
m
b
le
d
th
e
gr
ou
nd
tr
ut
h
a
nd
de
m
on
s
tr
a
t
e
d
s
up
e
r
io
r
pe
r
f
or
m
a
n
c
e
.
T
h
e
qu
a
li
t
a
ti
ve
a
n
a
ly
s
is
is
s
how
n i
n
F
ig
ur
e
2.
B
a
s
e
d
on
th
e
qua
nt
it
a
ti
ve
a
na
ly
s
is
in
T
a
bl
e
1,
R
N
N
-
D
out
pe
r
f
or
m
e
d
th
e
o
th
e
r
m
e
th
ods
in
P
S
N
R
va
lu
e
s
, i
ndi
c
a
ti
ng highe
r
i
m
a
ge
qua
li
ty
.
T
he
m
e
th
od e
f
f
e
c
ti
ve
ly
r
e
duc
e
s
ha
z
e
, r
e
s
ul
ti
ng i
n a
hi
ghe
r
P
S
N
R
a
nd
th
e
pr
oduc
ti
on
of
de
ha
z
e
d
im
a
ge
s
th
a
t
c
lo
s
e
ly
a
li
gn
w
it
h
th
e
gr
ound
tr
ut
h.
H
ow
e
ve
r
,
it
is
im
por
ta
nt
t
o
not
e
th
a
t
th
e
a
ppr
oa
c
h
m
a
y
not
c
ons
is
te
nt
ly
out
pe
r
f
or
m
ot
he
r
m
e
th
ods
in
t
e
r
m
s
of
S
S
I
M
f
or
a
ll
f
og
le
ve
ls
.
T
hi
s
c
oul
d
be
due
to
li
m
it
a
ti
ons
in
c
a
pt
ur
in
g
s
tr
uc
tu
r
a
l
in
f
or
m
a
ti
on.
S
S
I
M
m
e
a
s
ur
e
s
s
tr
uc
tu
r
a
l
s
im
il
a
r
it
y
a
nd
c
ons
id
e
r
s
lu
m
in
a
nc
e
, c
ont
r
a
s
t,
a
nd s
tr
uc
tu
r
e
. R
N
N
-
D
m
a
y be
le
s
s
e
f
f
e
c
ti
ve
i
n pr
e
s
e
r
vi
ng s
tr
uc
tu
r
a
l
i
nf
or
m
a
ti
on,
l
e
a
di
ng
to
lo
w
e
r
S
S
I
M
va
lu
e
s
a
t
c
e
r
ta
in
f
og
le
ve
ls
.
N
e
ve
r
th
e
le
s
s
,
th
e
a
ppr
oa
c
h
ge
ne
r
a
te
s
vi
s
u
a
ll
y
a
ppe
a
li
ng
de
ha
z
e
d
im
a
ge
s
w
it
h
f
e
w
e
r
c
o
lo
ur
di
s
to
r
ti
ons
a
nd
de
m
ons
tr
a
te
s
s
upe
r
io
r
im
a
ge
qua
li
ty
ba
s
e
d
on
P
S
N
R
va
lu
e
s
.
T
hi
s
e
f
f
e
c
ti
ve
ly
r
e
duc
e
s
ha
z
e
, pr
oduc
in
g hi
gh
-
qua
li
ty
de
ha
z
e
d i
m
a
ge
s
t
ha
t
c
lo
s
e
ly
m
a
tc
h t
h
e
gr
ound tr
ut
h.
T
he
pe
r
f
or
m
a
nc
e
of
R
N
N
-
D
w
a
s
f
ur
th
e
r
e
va
lu
a
te
d
on
th
e
R
E
S
I
D
E
da
ta
s
e
t,
w
hi
c
h
in
c
lu
de
s
in
door
a
nd
out
door
im
a
ge
s
w
it
h f
our
s
ta
te
-
of
-
th
e
-
a
r
t
im
a
ge
de
ha
z
in
g
a
lg
or
i
th
m
s
(
D
C
P
,
A
O
D
-
N
e
t,
D
e
ha
z
e
N
e
t,
a
nd
G
F
N
)
.
F
ig
ur
e
3
vi
s
ua
ll
y
c
om
pa
r
e
s
th
e
a
lg
or
it
hm
s
f
or
in
door
a
nd
out
door
s
c
e
ne
s
.
D
C
P
e
xhi
bi
ts
s
ig
ni
f
ic
a
nt
c
ol
our
di
s
to
r
ti
on
a
nd
lo
s
s
of
de
ta
il
,
w
hi
le
A
O
D
-
N
e
t
f
a
il
s
to
obl
it
e
r
a
te
ha
z
e
a
nd
pr
oduc
e
s
lo
w
-
br
ig
ht
ne
s
s
im
a
ge
s
.
D
e
ha
z
e
N
e
t
t
e
nds
to
ove
r
c
om
pe
n
s
a
te
f
or
br
ig
ht
ne
s
s
r
e
la
ti
ve
to
th
e
gr
ound
tr
ut
h.
I
n
c
ont
r
a
s
t,
R
N
N
-
D
de
m
ons
tr
a
te
s
r
e
m
a
r
ka
bl
e
a
c
c
ur
a
c
y
in
r
e
c
ove
r
in
g
de
ta
il
s
,
pa
r
ti
c
ul
a
r
ly
in
r
ow
4,
w
he
r
e
it
s
uc
c
e
s
s
f
ul
ly
r
e
s
to
r
e
s
th
e
to
w
e
r
s
in
th
e
de
pt
h
of
th
e
im
a
ge
.
T
he
r
e
s
ul
ts
obt
a
in
e
d
by
th
e
ot
he
r
a
lg
or
it
hm
s
a
r
e
uns
a
ti
s
f
a
c
to
r
y
in
c
om
pa
r
is
on,
hi
ghl
ig
ht
in
g
th
e
s
upe
r
io
r
pe
r
f
or
m
a
nc
e
of
R
N
N
-
D
in
te
r
m
s
of
r
e
a
li
s
ti
c
im
a
ge
de
ta
il
s
a
nd
c
ol
our
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
E
nhanc
in
g r
oad image
c
la
r
it
y
w
it
h r
e
s
id
ual
ne
u
r
al
ne
tw
or
k
de
h
az
in
g m
ode
l
(
A
e
r
un M
ar
ti
n)
4153
f
id
e
li
ty
.
T
a
bl
e
2
s
um
m
a
r
is
e
s
th
e
P
S
N
R
a
nd
S
S
I
M
f
or
in
door
a
nd
out
door
s
c
e
ne
s
.
R
N
N
-
D
d
e
m
ons
tr
a
te
s
s
upe
r
io
r
pe
r
f
or
m
a
nc
e
,
a
c
hi
e
vi
ng
s
ig
ni
f
ic
a
nt
ly
hi
ghe
r
P
S
N
R
va
lu
e
s
th
a
n
th
e
ot
he
r
m
e
th
ods
.
S
pe
c
if
ic
a
ll
y,
it
a
c
hi
e
ve
s
P
S
N
R
va
lu
e
s
of
28.87
a
nd
28.39
f
or
in
door
a
n
d
out
door
s
c
e
ne
s
,
r
e
s
pe
c
ti
ve
ly
,
s
ur
pa
s
s
in
g
th
e
r
e
s
ul
t
s
of
th
e
ot
he
r
a
lg
or
it
hm
s
.
I
n
c
onc
lu
s
io
n,
R
N
N
-
D
out
pe
r
f
or
m
s
s
ta
te
-
of
-
th
e
-
a
r
t
de
ha
z
in
g
a
lg
or
it
hm
s
qua
nt
it
a
ti
ve
ly
a
nd
qua
li
ta
ti
ve
ly
, r
e
s
to
r
in
g c
le
a
r
a
nd na
tu
r
a
l
im
a
ge
s
f
r
om
ha
z
y s
c
e
n
e
s
.
(
a
)
(
b)
(
c
)
(
d)
(
e
)
(f)
(
g)
(
h)
(
i)
(
j)
(
k)
F
ig
ur
e
2. T
he
de
ha
z
e
d r
e
s
ul
ts
on t
he
s
ynt
h
e
ti
c
da
ta
s
e
t
(
a
)
ha
z
y i
nput
, (
b)
D
C
P
[
4]
,
(
c
)
B
C
C
R
[
10]
,
(
d)
N
L
D
[
27]
,
(
e
)
C
A
P
[
28]
,
(
f
)
M
S
C
N
N
[
12]
,
(
g)
D
e
ha
z
e
N
e
t
[
5]
, (
h)
A
O
D
-
N
e
t
[
6]
,
(
i)
G
F
N
[
15
]
,
(
j)
R
N
N
-
D
, a
nd (
k)
gr
ound t
r
ut
hs
T
a
bl
e
1. A
ve
r
a
ge
P
S
N
R
a
nd S
S
I
M
va
lu
e
s
of
de
ha
z
e
d r
e
s
ul
ts
on
N
Y
U
2 da
ta
s
e
t
E
va
l
ua
t
i
on
P
S
N
R
/
S
S
I
M
P
a
pe
r
s
D
C
P
[
4]
BCCR
[
10]
N
L
D
[
27]
C
A
P
[
28]
M
S
C
N
N
[
12]
D
e
ha
z
e
N
e
t
[
5]
AOD
-
N
e
t
[
6]
G
F
N
(
G
)
[
15]
P
r
opos
e
d
L
i
ght
18.74/
0.77
17.72/
0.76
18.61/
0.71
21.92/
0.83
22.34/
0.82
24.87/
0.84
22.64/
0.85
24.78/
0.85
27.97/
0.85
M
e
di
um
18.68/
0.77
17.54/
0.75
18.47/
0.70
21.40/
0.82
21.21/
0.80
23.37/
0.83
21.33/
0.84
23.68/
0.84
27.96/
0.78
H
e
a
vy
18.67/
0.77
17.43/
0.75
18.21/
0.70
20.21/
0.80
20.51/
0.79
21.98/
0.82
20.24/
0.81
22.32/
0.83
27.98/
0.80
R
a
ndom
18.58/
0.77
17.35/
0.75
18.28/
0.71
19.99/
0.78
20.01/
0.78
20.97/
0.80
19.36/
0.78
22.41/
0.81
28.00/
0.77
(
a
)
(
b)
(
c
)
(
d)
(
e
)
(f)
(
g)
F
ig
ur
e
3. I
ndoor
a
nd outdoor
r
e
s
ul
ts
i
n S
O
T
S
:
(
a
)
ha
z
y i
nput
,
(
b)
D
C
P
[
4]
,
(
c
)
A
O
D
-
N
e
t
[
6
]
,
(
d)
D
e
ha
z
e
N
e
t
[
5]
,
(
e
)
G
F
N
[
16]
,
(
f
)
R
N
N
-
D
, a
nd
(
g)
gr
ound t
r
ut
h
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024:
4147
-
4157
4154
T
a
bl
e
2. Q
ua
nt
it
a
ti
ve
c
om
p
a
r
is
ons
of
S
O
T
S
f
or
di
f
f
e
r
e
nt
m
e
th
ods
M
e
t
hods
I
ndoor
O
ut
door
P
S
N
R
S
S
I
M
P
S
N
R
S
S
I
M
D
C
P
[
4]
16.62
0.82
19.13
0.81
AOD
-
N
e
t
[
6]
19.06
0.85
20.29
0.88
D
e
ha
z
e
N
e
t
[
5]
21.14
0.85
22.46
0.85
G
F
N
[
15]
22.30
0.88
21.55
0.84
R
N
N
-
D
28.87
0.83
28.39
0.82
S
in
c
e
th
e
de
ve
lo
pe
d
m
ode
l
pe
r
f
or
m
e
d
w
e
ll
in
th
e
in
door
a
nd
ou
td
oor
da
ta
s
e
ts
,
it
w
a
s
e
va
lu
a
te
d
f
or
it
s
r
obus
tn
e
s
s
i
n
r
oa
d da
ta
s
e
ts
f
r
om
t
he
F
r
id
a
a
nd F
r
id
a
2 D
a
ta
s
e
ts
.
T
he
s
e
da
ta
s
e
ts
c
ont
a
in
a
di
ve
r
s
e
r
a
nge
of
r
oa
d
im
a
ge
s
w
it
h
di
f
f
e
r
e
nt
ha
z
e
le
v
e
ls
,
a
ll
ow
in
g
f
or
a
c
om
pr
e
he
n
s
iv
e
a
s
s
e
s
s
m
e
nt
of
th
e
m
e
th
od'
s
e
f
f
e
c
ti
ve
ne
s
s
und
e
r
va
r
io
us
c
ondi
ti
ons
a
nd
a
c
om
pa
r
is
on
w
it
h
tr
a
di
ti
ona
l
a
lg
or
it
hm
s
.
T
he
e
va
lu
a
ti
on
in
c
lu
de
d
qua
li
ta
ti
ve
a
nd
qua
nt
it
a
ti
ve
a
na
ly
s
is
us
in
g
th
e
S
S
I
M
a
nd
P
S
N
R
,
f
ol
lo
w
in
g
a
s
im
il
a
r
a
ppr
oa
c
h
in
a
s
s
e
s
s
in
g
th
e
R
E
S
I
D
E
a
nd
N
Y
U
2
da
ta
s
e
t
s
to
e
ns
ur
e
c
ons
is
te
n
c
y
a
nd
c
om
pa
r
a
bi
li
ty
.
Q
ua
li
ta
ti
ve
hi
ghl
ig
ht
e
d
th
e
di
f
f
e
r
e
nc
e
s
be
twe
e
n
th
e
pr
opos
e
d
m
e
th
od
a
nd
pr
e
vi
ous
a
ppr
oa
c
he
s
,
f
oc
us
in
g
on
th
e
r
e
a
li
s
ti
c
r
e
pr
e
s
e
nt
a
ti
on
of
im
a
ge
de
ta
il
s
a
nd
c
ol
our
a
c
c
ur
a
c
y
in
r
oa
d
im
a
ge
s
.
F
ig
ur
e
4
s
how
s
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
th
e
pr
opos
e
d
a
lg
or
it
hm
c
om
pa
r
e
d
to
e
a
r
li
e
r
a
ppr
oa
c
he
s
on
s
ynt
he
ti
c
im
a
ge
s
f
r
om
th
e
F
R
I
D
A
a
nd
F
R
I
D
A
2
da
ta
s
e
t
s
.
A
m
ong
th
e
c
onve
nt
io
na
l
m
e
th
od
s
,
th
e
D
C
P
de
m
ons
tr
a
te
d
th
e
be
s
t
pe
r
f
or
m
a
nc
e
in
r
e
m
ovi
ng
th
e
f
our
ty
pe
s
of
ha
z
e
.
T
he
pr
opo
s
e
d
m
e
th
od
a
c
hi
e
v
e
d
r
e
s
ul
ts
c
om
pa
r
a
bl
e
to
th
e
D
C
P
'
s
,
but
s
om
e
r
oa
d
c
ol
our
di
s
to
r
ti
o
n
w
a
s
obs
e
r
ve
d
in
s
ynt
he
ti
c
im
a
ge
s
.
I
n
th
e
c
a
s
e
of
he
te
r
oge
ne
ous
ha
z
e
,
th
e
pr
opos
e
d
m
e
th
od,
a
lo
ng
w
it
h
th
e
A
P
S
F
a
nd
D
C
P
,
e
f
f
e
c
ti
ve
ly
de
ha
z
e
d
th
e
im
a
ge
s
a
nd
r
e
ve
a
le
d
bui
ld
in
gs
w
it
h
m
or
e
c
la
r
it
y
th
a
n
ot
he
r
m
e
th
ods
.
T
he
pa
tc
h
m
a
p
a
nd
hom
ogr
a
phy
-
ba
s
e
d
lo
c
a
l
de
ha
z
in
g
(
P
M
H
L
D
)
m
e
th
od
pe
r
f
o
r
m
e
d
w
e
ll
ove
r
a
ll
but
e
nc
ount
e
r
e
d
di
f
f
ic
ul
ti
e
s
in
de
ns
e
ha
z
e
a
r
e
a
s
w
it
h
c
lo
udy
e
nvi
r
o
nm
e
nt
s
.
T
he
C
A
P
a
lg
or
it
hm
c
a
us
e
d
a
de
c
r
e
a
s
e
in
s
a
tu
r
a
ti
on
in
a
c
hr
om
a
ti
c
a
r
e
a
s
w
it
h
obj
e
c
ts
s
uc
h
a
s
c
a
r
s
a
nd
r
oa
ds
.
T
he
A
P
S
F
m
e
th
od, l
ik
e
t
he
pr
opos
e
d t
e
c
hni
q
ue
, s
uc
c
e
s
s
f
ul
ly
r
e
m
ove
d a
ll
f
our
t
ype
s
of
ha
z
e
w
it
hout
c
a
us
in
g a
ny c
ol
our
c
ha
nge
i
n obje
c
ts
w
it
hi
n t
he
r
oa
d s
c
e
ne
. T
he
qua
nt
it
a
ti
ve
m
e
a
s
ur
e
m
e
nt
s
pr
e
s
e
nt
e
d
in
T
a
bl
e
3
pr
ovi
de
qua
nt
it
a
ti
ve
e
vi
de
nc
e
of
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
t
he
R
N
N
-
D
m
e
th
od
c
om
pa
r
e
d
to
ot
h
e
r
m
e
th
ods
.
R
N
N
-
D
a
c
hi
e
ve
d
a
P
S
N
R
s
c
or
e
of
27.799
dB
,
a
n
S
S
I
M
s
c
or
e
of
0.584
on
th
e
F
R
I
D
A
da
ta
s
e
t,
a
P
S
N
R
s
c
or
e
of
28.412 dB
, a
nd a
n S
S
I
M
s
c
or
e
o
f
0.73 on the FR
I
D
A
2 da
ta
s
e
t.
(
a
)
(
b)
(
c
)
(
d)
(
e
)
(f)
(
g)
(
h)
(
i)
(
j)
(
k)
F
ig
ur
e
4. Q
ua
li
ta
ti
ve
c
om
pa
r
is
on of
t
he
F
R
I
D
A
a
nd F
R
I
D
A
2 da
ta
s
e
ts
:
(
a
)
ha
z
y i
nput
, (
b)
D
C
P
[
4]
,
(
c
)
C
A
P
[
29]
, (
d)
D
E
F
A
D
E
[
30]
, (
e
)
D
e
ha
z
e
N
e
t
[
5]
, (
f
)
A
O
D
-
N
e
t
[
6]
, (
g)
A
M
E
F
[
31
]
, (
h)
P
M
H
L
D
[
32]
,
(
i)
A
P
S
F
[
33]
, (
j
)
R
N
N
-
D
, a
nd (
k)
gr
ound t
r
ut
hs
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
E
nhanc
in
g r
oad image
c
la
r
it
y
w
it
h r
e
s
id
ual
ne
u
r
al
ne
tw
or
k
de
h
az
in
g m
ode
l
(
A
e
r
un M
ar
ti
n)
4155
T
a
bl
e
3. Q
ua
nt
it
a
ti
ve
c
om
p
a
r
is
ons
on F
R
I
D
A
a
nd F
R
I
D
A
2 da
t
a
s
e
ts
D
a
t
a
s
e
t
M
e
t
r
i
c
s
H
a
z
y
D
C
P
[
4]
C
A
P
[
29]
D
E
F
A
D
E
[
30]
D
e
ha
z
e
N
e
t
[
5]
AOD
-
N
e
t
[
6]
A
M
E
F
[
31]
P
M
H
L
D
[
32]
A
S
M
F
[
33]
R
N
N
-
D
F
R
I
D
A
(
76
s
a
m
pl
e
s
)
S
S
I
M
0.616
0.6
0.522
0.566
0.614
0.629
0.68
0.563
0.693
0.584
P
S
N
R
11.167
13.316
12.494
13.226
12.926
13.229
1.237
12.697
13.875
27.799
F
R
I
D
A
2
(
264
s
a
m
pl
e
s
)
S
S
I
M
0.737
0.642
0.334
0.756
0.662
0.44
0.787
0.656
0.803
0.73
P
S
N
R
10.472
11.638
8.826
12.048
11.8
9.749
12.713
11.695
12.549
28.412
T
he
a
va
il
a
bi
li
ty
of
th
e
or
ig
in
a
l
im
a
ge
s
in
th
e
s
ynt
he
ti
c
da
ta
s
e
t
a
ll
ow
s
f
or
m
or
e
a
c
c
ur
a
te
c
om
p
a
r
is
ons
of
t
he
ha
z
e
r
e
m
ova
l
a
c
hi
e
ve
d by e
a
c
h a
lg
or
it
hm
.
H
ow
e
ve
r
, a
na
ly
s
in
g t
he
a
lg
or
it
hm
us
in
g r
e
a
l
r
oa
d da
ta
s
e
ts
i
s
c
r
it
ic
a
l
to
i
de
nt
if
y
it
s
e
f
f
e
c
ti
ve
ne
s
s
. T
hus
, t
he
a
lg
or
it
hm
w
a
s
f
ur
th
e
r
e
va
lu
a
te
d on the
c
it
ys
c
a
pe
r
oa
d
da
ta
s
e
t
to
a
s
s
e
s
s
i
ts
r
obus
tn
e
s
s
i
n de
ha
z
in
g
a
c
tu
a
l
r
oa
d i
m
a
ge
s
w
it
h r
oa
d
m
a
r
ke
r
s
. T
a
bl
e
4
pr
e
s
e
nt
s
a
qua
li
ta
ti
ve
a
na
ly
s
is
of
th
e
m
ode
l'
s
pe
r
f
or
m
a
nc
e
on
s
e
le
c
te
d
im
a
g
e
s
f
r
om
th
e
da
t
a
s
e
t
s
in
c
e
th
e
r
e
a
r
e
no
ha
z
e
-
f
r
e
e
im
a
ge
s
f
or
qua
nt
it
a
ti
ve
a
na
ly
s
is
.
T
he
de
ha
z
e
d
im
a
ge
s
us
in
g
th
e
R
N
N
-
D
a
l
g
or
it
hm
s
ig
ni
f
ic
a
nt
ly
i
m
pr
ove
th
e
vi
s
ib
il
i
ty
o
f
r
oa
d
m
a
r
ke
r
s
.
T
hi
s
e
nha
nc
e
m
e
nt
is
a
tt
r
ib
ut
e
d
to
th
e
in
c
r
e
a
s
e
d
c
ont
r
a
s
t
di
f
f
e
r
e
nt
ia
ti
on
be
twe
e
n
th
e
r
oa
d
s
ur
f
a
c
e
a
nd
th
e
m
a
r
ke
r
s
.
T
he
r
e
duc
e
d
c
ont
r
a
s
t
be
twe
e
n
th
e
s
e
e
le
m
e
nt
s
c
a
n
le
a
d
to
c
la
s
s
if
ic
a
ti
on
e
r
r
or
s
in
f
oggy
c
ondi
ti
ons
.
H
ow
e
ve
r
,
th
e
R
N
N
-
D
de
ha
z
in
g
te
c
hni
que
e
nha
n
c
e
s
th
e
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
c
a
pa
bi
li
ti
e
s
s
pe
c
if
ic
to
r
oa
d m
a
r
ke
r
s
, r
e
duc
in
g t
he
l
ik
e
li
hood of
m
is
c
la
s
s
if
ic
a
ti
on.
T
a
bl
e
4.
Q
ua
li
ta
ti
ve
a
n
a
ly
s
is
of
R
N
N
-
D
e
f
f
e
c
ti
ve
ne
s
s
on
c
it
ys
c
a
pe
r
e
a
l
r
oa
d
da
ta
s
e
t
I
m
a
ge
no
A
c
t
ua
l
i
m
a
ge
D
e
ha
z
e
d
i
m
a
ge
s
I
m
a
ge
no
A
c
t
ua
l
i
m
a
ge
D
e
ha
z
e
d
i
m
a
ge
s
1
7
2
8
3
9
4
10
5
11
6
12
5.
C
O
N
C
L
U
S
I
O
N
I
n
c
onc
lu
s
io
n,
th
e
pr
opos
e
d
R
e
s
N
e
t
-
ba
s
e
d
de
ha
z
in
g
a
lg
or
it
hm
,
R
N
N
-
D
,
ha
s
de
m
ons
tr
a
te
d
e
xc
e
ll
e
nt
pe
r
f
or
m
a
nc
e
on
va
r
io
us
da
ta
s
e
ts
,
in
c
lu
di
ng
N
Y
U
2,
R
E
S
I
D
E
,
F
R
I
D
A
,
a
nd
F
R
I
D
A
2.
I
t
out
pe
r
f
or
m
e
d
e
xi
s
ti
ng
s
ta
te
-
of
-
th
e
-
a
r
t
de
ha
z
in
g
a
lg
or
it
hm
s
in
t
e
r
m
s
of
obj
e
c
ti
ve
m
e
tr
ic
s
s
u
c
h
a
s
P
S
N
R
a
nd
S
S
I
M
,
a
s
w
e
ll
a
s
s
ubj
e
c
ti
ve
e
va
lu
a
ti
ons
ba
s
e
d
on
vi
s
u
a
l
qua
li
ty
.
T
he
a
lg
or
it
hm
e
f
f
e
c
ti
ve
ly
r
e
c
ove
r
s
c
le
a
r
a
nd
na
tu
r
a
l
im
a
ge
s
f
r
om
ha
z
y
s
c
e
ne
s
,
e
ve
n
und
e
r
c
ha
ll
e
ngi
ng
c
ondi
ti
ons
w
it
h
va
r
yi
ng
de
gr
e
e
s
a
nd
ty
pe
s
of
ha
z
e
. T
hi
s
m
a
ke
s
it
a
pr
om
is
in
g
s
ol
ut
io
n
f
or
e
nha
nc
in
g
im
a
ge
qua
li
ty
in
a
dve
r
s
e
w
e
a
th
e
r
s
it
ua
ti
ons
.
T
he
qua
nt
it
a
ti
ve
a
na
ly
s
is
s
how
e
d
th
a
t
th
e
pr
opos
e
d
a
lg
or
it
hm
out
pe
r
f
or
m
e
d
e
xi
s
ti
ng
te
c
hni
que
s
on
m
ul
ti
pl
e
da
ta
s
e
ts
,
in
c
lu
di
ng
S
O
T
S
:
in
door
,
out
door
,
N
Y
U
2,
F
R
I
D
A
,
a
nd
F
R
I
D
A
2.
A
ddi
ti
ona
ll
y,
th
e
va
li
da
ti
on
on
r
e
a
l
im
a
ge
s
de
m
ons
tr
a
te
d
hi
gh
c
ont
r
a
s
t
be
twe
e
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024:
4147
-
4157
4156
r
oa
d
m
a
r
ke
r
s
a
nd
th
e
r
oa
d s
ur
f
a
c
e
,
w
hi
c
h
e
nh
a
nc
e
s
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
f
or
r
oa
d
m
a
r
ke
r
c
la
s
s
if
ic
a
ti
on. T
hi
s
ha
s
s
ig
ni
f
ic
a
nt
im
pl
ic
a
ti
ons
f
or
a
ppl
ic
a
ti
ons
s
uc
h
a
s
a
ut
onomou
s
d
r
iv
in
g,
w
he
r
e
im
a
ge
qua
li
ty
a
nd
a
c
c
ur
a
c
y
a
r
e
c
r
uc
ia
l
f
or
s
a
f
e
ty
a
nd
r
e
li
a
bi
li
ty
.
T
he
pr
opos
e
d
a
lg
or
it
hm
’
s
s
upe
r
io
r
pe
r
f
or
m
a
nc
e
a
nd
vi
s
ua
ll
y
a
ppe
a
li
ng
r
e
s
ul
ts
m
a
ke
it
a
pow
e
r
f
ul
to
ol
f
or
im
a
ge
-
de
ha
z
in
g
a
ppl
ic
a
ti
ons
.
F
ut
ur
e
w
or
k
c
oul
d
di
s
c
us
s
th
e
t
e
c
hni
que
f
or
c
la
s
s
if
yi
ng
r
oa
d
m
a
r
ke
r
s
e
xt
r
a
c
te
d
f
r
om
de
ha
z
e
d
im
a
ge
s
,
f
ur
th
e
r
e
nha
nc
in
g
th
e
c
la
s
s
if
ic
a
ti
on
a
lg
or
it
hm
a
nd
r
e
duc
in
g e
r
r
or
s
c
a
us
e
d by f
oggy il
lu
m
in
a
ti
on.
A
C
K
N
O
WL
E
D
G
E
M
E
N
T
S
T
he
a
ut
hor
s
e
xpr
e
s
s
th
e
ir
gr
a
ti
tu
de
to
th
e
M
a
la
ys
ia
n
gov
e
r
nm
e
nt
f
or
s
pons
or
in
g
th
is
pr
oj
e
c
t
a
nd
th
os
e
w
ho
ha
ve
c
ont
r
ib
ut
e
d
to
c
om
pl
e
ti
ng
th
is
pa
p
e
r
di
r
e
c
tl
y
or
in
di
r
e
c
tl
y.
T
he
f
undi
ng
of
th
is
r
e
s
e
a
r
c
h
is
pr
ovi
de
d
by a
gr
a
nt
f
r
om
t
he
M
a
la
ys
ia
n M
in
is
tr
y of
H
ig
he
r
E
duc
a
ti
on (
F
R
G
S
/1
/2
019/
T
K
04/
M
M
U
/0
2/
2)
.
R
E
F
E
R
E
N
C
E
S
[
1]
H
. A
.
G
ha
ni
e
t
al
.
, “
A
dva
nc
e
s
i
n l
a
ne
m
a
r
ki
ng de
t
e
c
t
i
on a
l
gor
i
t
hm
s
f
or
a
l
l
-
w
e
a
t
he
r
c
ondi
t
i
ons
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
E
l
e
c
t
r
i
c
al
and C
om
put
e
r
E
ngi
ne
e
r
i
ng (
I
J
E
C
E
)
, vol
. 11, no. 4, pp. 3365
–
3373, A
ug. 2021, doi
:
10.11591/
i
j
e
c
e
.v11i
4.pp3365
-
3373.
[
2]
J.
-
P
.
T
a
r
e
l
a
nd
N
.
H
a
ut
i
e
r
e
,
“
F
a
s
t
vi
s
i
bi
l
i
t
y
r
e
s
t
or
a
t
i
on
f
r
om
a
s
i
ngl
e
c
ol
or
o
r
gr
a
y
l
e
ve
l
i
m
a
ge
,”
i
n
2009
I
E
E
E
12
th
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
i
s
i
on
, I
E
E
E
,
S
e
p. 2009, pp. 2201
–
2208
, doi
:
10.110
9/
I
C
C
V
.2009.5459251.
[
3]
K
.
H
e
,
J
.
S
un,
a
nd
X
.
T
a
ng,
“
G
ui
de
d
i
m
a
g
e
f
i
l
t
e
r
i
ng,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
P
at
t
e
r
n
A
nal
y
s
i
s
and
M
ac
hi
ne
I
nt
e
l
l
i
ge
nc
e
,
vol
.
35
,
no. 6, pp. 1397
–
1409, J
un. 2013, doi
:
10.1109/
T
P
A
M
I
.2012.213.
[
4]
H
.
K
a
i
m
i
ng,
S
.
J
i
a
n,
a
nd
T
.
X
i
a
oou,
“
S
i
ngl
e
i
m
a
ge
ha
z
e
r
e
m
ova
l
us
i
ng
d
a
r
k
c
ha
nne
l
pr
i
or
,”
i
n
2009
I
E
E
E
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
i
s
i
on and P
at
t
e
r
n R
e
c
ogni
t
i
on
, I
E
E
E
,
J
un. 2009, pp. 1956
–
1963
, doi
:
10.1109
/
C
V
P
R
.2009.5206515.
[
5]
B
. C
a
i
, X
. X
u, K
. J
i
a
, C
. Q
i
ng, a
nd D
. T
a
o, “
D
e
ha
z
e
N
e
t
:
a
n e
nd
-
to
-
e
nd s
ys
t
e
m
f
or
s
i
ngl
e
i
m
a
ge
ha
z
e
r
e
m
ova
l
,”
I
E
E
E
T
r
ans
ac
t
i
ons
on I
m
age
P
r
oc
e
s
s
i
ng
, vol
. 25, no. 11, pp. 5187
–
5198, N
ov. 2016, doi
:
10.1109/
T
I
P
.2016.2598681.
[
6]
B
. L
i
, X
. P
e
ng, Z
.
W
a
ng, J
.
X
u, a
nd D
.
F
e
ng, “
A
O
D
-
N
e
t
:
a
l
l
-
in
-
one
de
ha
z
i
ng ne
t
w
or
k,”
i
n
2017
I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
i
s
i
on (
I
C
C
V
)
, I
E
E
E
, O
c
t
. 2017, pp. 4780
–
4788
, doi
:
10.1109/
I
C
C
V
.2017.511.
[
7]
S
. Z
ha
ng a
nd F
. H
e
, “
D
R
C
D
N
:
l
e
a
r
ni
ng de
e
p r
e
s
i
dua
l
c
onvol
ut
i
ona
l
de
ha
z
i
ng
ne
t
w
or
ks
,”
T
he
V
i
s
ual
C
om
put
e
r
, vol
. 36,
no. 9, pp
.
1797
–
1808, S
e
p. 2020, doi
:
10.1007/
s
00371
-
019
-
01774
-
8.
[
8]
E
r
w
i
n,
“
I
m
pr
ovi
ng
r
e
t
i
na
l
i
m
a
ge
qua
l
i
t
y
us
i
ng
t
he
c
ont
r
a
s
t
s
t
r
e
t
c
hi
ng,
hi
s
t
ogr
a
m
e
qua
l
i
z
a
t
i
on,
a
nd
C
L
A
H
E
m
e
t
hod
s
w
i
t
h
m
e
di
a
n
f
i
l
t
e
r
s
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
I
m
age
,
G
r
aphi
c
s
and
Si
gnal
P
r
oc
e
s
s
i
n
g
,
vol
.
12,
no.
2,
pp.
30
–
41,
A
p
r
.
2020,
doi
:
10.5815/
i
j
i
gs
p.2020.02.04.
[
9]
Z
. M
. S
a
ni
, L
. J
. C
hua
n,
T
. A
. I
z
z
udi
n, H
. A
. G
ha
ni
,
a
nd A
. M
a
r
t
i
n, “
U
r
ba
n R
oa
d m
a
r
ke
r
c
l
a
s
s
i
f
i
c
a
t
i
on us
i
ng hi
s
t
ogr
a
m
of
or
i
e
nt
e
d
gr
a
di
e
nt
a
nd
l
oc
a
l
bi
na
r
y
pa
t
t
e
r
n
w
i
t
h
a
r
t
i
f
i
c
i
a
l
ne
ur
a
l
ne
t
w
or
k,”
i
n
P
r
oc
e
e
di
ngs
of
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
E
m
e
r
gi
n
g
T
e
c
hnol
ogi
e
s
and I
nt
e
l
l
i
ge
nt
Sy
s
t
e
m
s
:
I
C
E
T
I
S 2021
, 2022, pp. 126
–
135
, doi
:
10.1007/
978
-
3
-
030
-
82616
-
1_12.
[
10]
G
.
M
e
ng,
Y
.
W
a
ng,
J
.
D
ua
n,
S
.
X
i
a
ng,
a
nd
C
.
P
a
n,
“
E
f
f
i
c
i
e
nt
i
m
a
ge
de
ha
z
i
ng
w
i
t
h
bounda
r
y
c
ons
t
r
a
i
nt
a
nd
c
ont
e
xt
ua
l
r
e
gul
a
r
i
z
a
t
i
on,”
i
n
T
he
I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
i
s
i
on
, 20
13, pp. 617
–
624
, doi
:
10.1109/
I
C
C
V
.2013.82.
[
11]
Z
. L
i
a
nd J
. Z
he
ng, “
E
dge
-
pr
e
s
e
r
vi
ng de
c
om
pos
i
t
i
on
-
ba
s
e
d
s
i
ngl
e
i
m
a
ge
ha
z
e
r
e
m
ova
l
,”
I
E
E
E
T
r
ans
ac
t
i
ons
on I
m
age
P
r
oc
e
s
s
i
ng
,
vol
. 24, no. 12, pp. 5432
–
5441, D
e
c
. 2015, doi
:
10.1109/
T
I
P
.2015.2482903.
[
12]
W
.
R
e
n,
S
.
L
i
u,
H
.
Z
ha
ng,
J
.
P
a
n,
X
.
C
a
o,
a
nd M
.
-
H
.
Y
a
ng,
“
S
i
ngl
e
i
m
a
ge
de
h
a
z
i
ng
vi
a
m
ul
t
i
-
s
c
a
l
e
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
,”
i
n
C
om
put
e
r
V
i
s
i
on
-
E
C
C
V
2016:
14
th
E
ur
ope
an
C
onf
e
r
e
nc
e
,
A
m
s
t
e
r
dam
,
T
he
N
e
t
he
r
l
ands
,
O
c
t
obe
r
11
-
14,
2016,
P
r
oc
e
e
di
ng
s
,
P
ar
t
I
I
14
, 2016, pp. 154
–
169
, doi
:
10.1007/
978
-
3
-
319
-
46475
-
6_10.
[
13]
M
.
M
.
H
o,
J
.
Z
hou,
a
nd
G
.
H
e
,
“
R
R
-
D
nC
N
N
v2.0:
e
nh
a
nc
e
d
r
e
s
t
or
a
t
i
on
-
r
e
c
ons
t
r
uc
t
i
on
de
e
p
ne
ur
a
l
ne
t
w
or
k
f
or
dow
n
-
s
a
m
pl
i
ng
-
ba
s
e
d vi
de
o c
odi
ng,”
I
E
E
E
T
r
ans
ac
t
i
ons
on I
m
age
P
r
oc
e
s
s
i
ng
, vol
. 30, pp. 170
2
–
1715, 2021, doi
:
10.1109/
T
I
P
.2020.3046872.
[
14]
X
. L
i
u, Y
. M
a
, Z
. S
hi
, a
nd
J
. C
he
n, “
G
r
i
dD
e
ha
z
e
N
e
t
:
a
t
t
e
nt
i
on
-
ba
s
e
d m
ul
t
i
-
s
c
a
l
e
ne
t
w
or
k f
or
i
m
a
ge
de
ha
z
i
ng,”
i
n
2019 I
E
E
E
/
C
V
F
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
i
s
i
on (
I
C
C
V
)
, I
E
E
E
, O
c
t
. 2019, pp. 73
13
–
7322
, doi
:
10.1109/
I
C
C
V
.2019.00741.
[
15]
W
. R
e
n
e
t
al
.
, “
G
a
t
e
d f
us
i
on ne
t
w
or
k f
or
s
i
ngl
e
i
m
a
ge
de
ha
z
i
ng,”
i
n
2018 I
E
E
E
/
C
V
F
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
i
s
i
on and P
at
t
e
r
n
R
e
c
ogni
t
i
on
, I
E
E
E
,
J
un. 2018, pp. 3253
–
3261
, doi
:
10.1109/
C
V
P
R
.2018.00343.
[
16]
M
.
K
i
m
,
S
.
H
ong,
H
.
L
e
e
,
a
nd
M
.
G
.
K
a
ng,
“
S
i
ngl
e
i
m
a
ge
de
ha
z
i
ng
of
r
oa
d
s
c
e
ne
s
us
i
ng
s
pa
t
i
a
l
l
y
a
da
pt
i
ve
a
t
m
os
phe
r
i
c
poi
nt
s
pr
e
a
d
f
unc
t
i
on,”
I
E
E
E
A
c
c
e
s
s
, vol
. 9, pp. 76135
–
76152, 2021, doi
:
10.1109/
A
C
C
E
S
S
.2021.3082175.
[
17]
J
.
M
e
ng,
Y
.
L
i
,
H
.
L
i
a
ng,
a
nd
Y
.
M
a
,
“
S
i
ngl
e
i
m
a
ge
de
ha
z
i
ng
ba
s
e
d
on
t
w
o
-
s
t
r
e
a
m
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k,”
J
our
nal
o
f
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
and T
e
c
hnol
ogy
, J
un. 2022, doi
:
10.37965/
j
a
i
t
.2022.0110
.
[
1
8]
J
.
Z
hou
,
C
.
T
.
L
e
o
ng,
a
n
d
C
.
L
i
,
“
M
u
l
t
i
-
s
c
a
l
e
a
n
d
a
t
t
e
nt
i
o
n
r
e
s
i
d
ua
l
ne
t
w
o
r
k
f
o
r
s
i
ng
l
e
i
m
a
g
e
de
ha
z
i
ng
,”
i
n
2
02
1
6
th
I
nt
e
r
na
t
i
o
na
l
C
onf
e
r
e
nc
e
on
I
nt
e
l
l
i
g
e
nt
C
om
p
ut
i
n
g
and
S
i
g
na
l
P
r
o
c
e
s
s
i
ng
(
I
C
S
P
)
,
A
p
r
.
2
021
,
p
p.
48
3
–
4
87
,
do
i
:
1
0.1
10
9/
I
C
S
P
518
82
.20
21
.94
08
801
.
[
19]
W
.
Y
i
,
L
.
D
ong,
M
.
L
i
u,
Y
.
Z
ha
o,
M
.
H
ui
,
a
nd
L
.
K
ong,
“
D
C
N
e
t
:
dua
l
-
c
a
s
c
a
de
ne
t
w
or
k
f
or
s
i
ngl
e
i
m
a
g
e
de
ha
z
i
ng,”
N
e
u
r
al
C
om
put
i
ng and A
ppl
i
c
at
i
ons
, vol
. 34, no. 19, pp. 16771
–
16783, O
c
t
. 2022, doi
:
10.1007/
s
00521
-
022
-
07319
-
w.
[
20]
Y
.
Y
a
ng,
C
.
H
ou,
H
.
H
ua
ng,
Z
.
Z
ha
ng,
a
nd
G
.
X
i
e
,
“
C
a
s
c
a
de
d
de
e
p
r
e
s
i
d
ua
l
l
e
a
r
ni
ng
ne
t
w
or
k
f
o
r
s
i
ngl
e
i
m
a
ge
de
ha
z
i
ng,”
M
ul
t
i
m
e
di
a Sy
s
t
e
m
s
, vol
. 29, no. 4, pp. 2037
–
2048, A
ug. 2023, doi
:
10.1007/
s
0
0530
-
023
-
01087
-
w.
[
21]
J
. Z
hou, C
. T
. L
e
ong, a
nd
C
.
L
i
, “
M
ul
t
i
-
s
c
a
l
e
a
nd
a
t
t
e
nt
i
on r
e
s
i
dua
l
ne
t
w
or
k f
o
r
s
i
ngl
e
i
m
a
ge
de
ha
z
i
ng,”
i
n
2021 6t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
I
nt
e
l
l
i
ge
nt
C
om
put
i
ng
and
Si
gnal
P
r
oc
e
s
s
i
ng
(
I
C
SP
)
,
I
E
E
E
,
A
pr
.
2021,
pp.
483
–
487
,
doi
:
10.1109/
I
C
S
P
51882.2021.9408801.
[
22]
X
.
J
i
a
ng,
C
.
Z
ha
o,
M
.
Z
hu,
Z
.
H
a
o,
a
nd
W
.
G
a
o,
“
R
e
s
i
dua
l
s
pa
t
i
a
l
a
nd
c
ha
n
ne
l
a
t
t
e
nt
i
on
ne
t
w
or
ks
f
or
s
i
ngl
e
i
m
a
ge
de
ha
z
i
ng,”
Se
ns
or
s
, vol
. 21, no. 23, p. 7922, N
ov. 2021, doi
:
10.3390/
s
21237922.
[
23]
P
. W
a
ng, H
. Z
hu, H
. H
ua
ng, H
.
Z
ha
ng, a
nd N
. W
a
ng, “
T
M
S
-
G
A
N
:
a
t
w
of
o
l
d
m
ul
t
i
-
s
c
a
l
e
ge
ne
r
a
t
i
ve
a
dve
r
s
a
r
i
a
l
ne
t
w
or
k f
or
s
i
ngl
e
i
m
a
ge
de
ha
z
i
ng,”
I
E
E
E
T
r
ans
ac
t
i
ons
on C
i
r
c
ui
t
s
and
S
y
s
t
e
m
s
f
or
V
i
de
o
T
e
c
hno
l
ogy
,
vol
.
32,
no.
5,
pp.
2760
–
2772,
M
a
y
2022,
doi
:
10.1109/
T
C
S
V
T
.2021.3097713.
[
24]
C
. Y
. J
e
ong, K
. M
oon, a
nd M
. K
i
m
,
“
A
n e
nd
-
to
-
e
nd de
e
p l
e
a
r
n
i
ng a
ppr
oa
c
h f
o
r
r
e
a
l
-
t
i
m
e
s
i
ngl
e
i
m
a
ge
de
ha
z
i
ng,”
J
our
nal
of
R
e
al
-
T
i
m
e
I
m
age
P
r
oc
e
s
s
i
ng
, vol
. 20, no. 1, p. 12, F
e
b. 2023, doi
:
10.1007/
s
11554
-
0
23
-
01270
-
2.
[
25]
B
. L
i
e
t
al
.
, “
B
e
nc
hm
a
r
ki
ng s
i
ngl
e
-
i
m
a
ge
de
ha
z
i
ng a
nd be
yond,
”
I
E
E
E
T
r
ans
ac
t
i
ons
on I
m
age
P
r
oc
e
s
s
i
ng
, vol
. 28, no. 1, pp. 492
–
505, J
a
n. 2019, doi
:
10.1109/
T
I
P
.2018.2867951.
Evaluation Warning : The document was created with Spire.PDF for Python.