I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
d
van
c
e
s
i
n
A
p
p
li
e
d
S
c
ie
n
c
e
s
(
I
JA
A
S
)
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
, pp.
1340
~
1349
I
S
S
N
:
2252
-
8814
,
D
O
I
:
10.11591/
ij
a
a
s
.
v14.
i
4
.
pp
1340
-
1349
1340
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
aas
.i
ae
s
c
or
e
.c
om
G
e
n
e
r
at
i
ve
ad
ve
r
sar
i
al
n
e
t
w
or
k
f
o
r
i
n
t
e
l
l
i
ge
n
t
h
az
e
r
e
m
oval
f
r
om
h
i
gh
q
u
al
i
t
y i
m
age
s
A
li
A
b
d
u
la
z
e
e
z
M
oh
am
m
e
d
B
aq
e
r
Q
az
z
az
1
, H
ayf
aa T
. H
u
s
s
e
in
1
, S
h
r
oou
q
J. A
l
-
j
a
n
ab
i
1
,
Y
ou
s
if
M
u
d
h
af
ar
1,
2
1
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
, F
a
c
ul
t
y of
E
duc
a
t
i
on,
U
ni
ve
r
s
i
t
y of
K
uf
a
,
N
a
j
a
f
,
I
r
a
q
2
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
T
e
c
hni
que
s
E
ngi
ne
e
r
i
ng
,
C
ol
l
e
ge
of
T
e
c
hni
c
a
l
E
ngi
n
e
e
r
i
ng
, T
he
I
s
l
a
m
i
c
U
ni
ve
r
s
i
t
y
,
N
a
j
a
f
,
I
r
a
q
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
J
ul
23
,
2025
R
e
vi
s
e
d
O
c
t
05
,
2025
A
c
c
e
pt
e
d
N
ov 07
,
2025
Suspend
ed
atmosph
eric
partic
ulates
like
haze
,
mist,
and
fog
grea
tly
d
egra
de
captured
images,
creating
considerab
le
challenges
for
computer
vision
applicati
ons
operating
in
safety
-
sensitive
areas
such
as
autonomous
d
riving,
surveillance,
and
remote
sensing.
In
this
paper,
we
treat
the
im
portant
challenge
of
single
-
image
haze
removal
by
proposing
a
novel
and
robust
conditi
onal
generative
adversarial
network
(cGAN)
-
based
framewor
k.
The
proposal
utilizes
a
U
-
Net
-
based
generator
with
self
-
attenti
on
and
skip
-
conn
ections
to
preserve
spatial
fidelit
y,
and
a
PatchGAN
discrim
in
ator
to
enforce local
realism
. At
the heart
of
our cont
ributio
n
is a
careful
ly
we
ighted
multi
-
component
loss
function
that
applies
reconstruct
ion,
pe
rceptual,
edge,
structural
similarity
(SSIM),
and
advers
arial
losses
to
optimize
pixe
l
-
level
accuracy
and
perceptual
quality
.
We
trained
and
evaluated
our
propo
sal
on
the
large
-
scale
real
-
world
LMHaze
dataset.
Exp
erimental
results
demonstrate
state
-
of
-
the
-
art
performance
with
a
peak
signal
-
to
-
noise
r
atio
(PSNR)
of
33.42
dB
and
SSIM
of
0.9590.
Our
qualitative
and
comp
arative
analyses
further
support
our
claims
by
assessin
g
our
proposed
model'
s
capacity
to
recover
clear
and
artifact
-
free
images
from
hazy
images
-
outperforming
the
existing
methods
on
this
challenging
real
-
world
benchmark.
K
e
y
w
o
r
d
s
:
C
ondi
ti
ona
l
G
A
N
D
e
e
p l
e
a
r
ni
ng
G
e
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
I
m
a
ge
de
ha
z
in
g
P
a
tc
h G
A
N
P
e
r
c
e
pt
ua
l
lo
s
s
U
-
N
e
t
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
:
A
li
A
bdul
a
z
e
e
z
M
oha
m
m
e
d B
a
q
e
r
Q
a
z
z
a
z
D
e
pa
r
tm
e
nt
of
C
om
put
e
r
, F
a
c
ul
ty
of
E
duc
a
ti
on, Unive
r
s
it
y
of
K
uf
a
N
a
ja
f
, I
r
a
q
E
m
a
il
:
a
li
a
.qa
z
z
a
z
@
uokuf
a
.e
du.i
q
1.
I
N
T
R
O
D
U
C
T
I
O
N
H
a
z
e
i
s
a
f
r
e
que
nt
w
e
a
th
e
r
phe
nom
e
non
th
a
t
r
e
duc
e
s
vi
s
ib
il
it
y
by
s
c
a
tt
e
r
in
g
a
nd
a
bs
or
bi
ng
li
ght
,
w
hi
c
h
le
a
ds
to
lo
w
c
ont
r
a
s
t,
c
ol
or
di
s
to
r
ti
on,
a
nd
lo
s
s
of
f
in
e
de
ta
il
in
phot
ogr
a
phi
c
im
a
ge
s
.
T
h
e
s
e
de
gr
a
da
ti
ons
ha
ve
c
ons
id
e
r
a
bl
e
im
pl
ic
a
ti
ons
f
or
c
om
put
e
r
vi
s
io
n
a
ppl
ic
a
ti
ons
dow
n
th
e
li
ne
—
s
uc
h
a
s
a
ut
onomous
dr
iv
in
g,
a
e
r
ia
l
s
ur
ve
il
la
nc
e
,
a
nd
r
e
m
ot
e
s
e
ns
in
g
—
th
a
t
de
pe
nd
on
c
le
a
r
vi
s
ib
il
it
y
of
a
s
c
e
ne
f
or
r
e
li
a
bl
e
ope
r
a
ti
on
[
1]
–
[
5]
.
T
he
f
o
r
m
a
ti
on
of
a
ha
z
y
im
a
ge
is
c
om
m
onl
y
m
ode
le
d
by
th
e
a
tm
os
phe
r
ic
s
c
a
tt
e
r
in
g m
ode
l
(
A
S
M
)
:
=
.
+
(
1
−
)
.
(
1)
(
1)
W
he
r
e
r
e
pr
e
s
e
nt
s
or
ig
in
a
l
c
ol
or
,
obs
e
r
ve
d c
ol
or
,
p
pos
it
io
n
of
pi
xe
l,
is
th
e
a
m
bi
e
nt
li
ght
,
a
nd
is
th
e
tr
a
ns
m
is
s
io
n of
t
he
l
ig
ht
r
e
f
le
c
te
d by the
obj
e
c
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
dv A
ppl
S
c
i
I
S
S
N
:
2252
-
8814
G
e
ne
r
at
iv
e
ad
v
e
r
s
ar
ia
l
ne
tw
o
r
k
f
or
i
nt
e
ll
ig
e
nt
ha
z
e
r
e
m
ov
al
…
(
A
li
A
bdu
la
z
e
e
z
M
ohamm
e
d B
aqe
r
Q
a
z
z
a
z
)
1341
=
−
(
2)
W
he
r
e
r
e
pr
e
s
e
nt
s
th
e
a
tt
e
nua
ti
on
c
oe
f
f
ic
ie
nt
de
te
r
m
in
e
d
by
th
e
w
e
a
th
e
r
c
ondi
ti
on,
a
nd
is
th
e
s
c
e
ne
de
pt
h
f
r
om
th
e
c
a
m
e
r
a
[
6]
.
T
r
a
di
ti
ona
l
pr
io
r
-
ba
s
e
d
te
c
hni
que
s
s
uc
h
a
s
D
C
P
,
C
A
P
,
a
nd
f
us
io
n
-
ba
s
e
d
pr
io
r
s
a
tt
e
m
pt
to
e
s
ti
m
a
te
tr
a
ns
m
is
s
io
n
a
nd
a
tm
os
ph
e
r
ic
li
ght
th
r
ou
gh
ha
ndc
r
a
f
te
d
a
s
s
um
pt
io
ns
[
7]
–
[
9]
,
but
f
a
il
unde
r
de
ns
e
or
s
pa
ti
a
ll
y
non
-
hom
oge
ne
ou
s
ha
z
e
.
R
e
c
e
nt
de
e
p
le
a
r
ni
ng
m
e
th
ods
le
ve
r
a
ge
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
(
C
N
N
)
a
nd
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
(
G
A
N
)
to
le
a
r
n
a
di
r
e
c
t
m
a
ppi
ng
f
r
om
ha
z
y
to
c
le
a
r
im
a
ge
s
[
10]
,
[
11
]
,
w
it
h
a
tt
e
nt
io
n
a
nd
tr
a
ns
f
or
m
e
r
-
ba
s
e
d
va
r
ia
nt
s
f
ur
th
e
r
im
pr
ov
in
g
gl
oba
l
c
ont
e
x
t
m
ode
li
ng
[
12]
,
[
13]
.
I
n
pa
r
a
ll
e
l,
di
f
f
u
s
io
n
-
b
a
s
e
d
g
e
ne
r
a
ti
ve
m
ode
l
s
a
nd
p
e
r
c
e
pt
ua
l
-
r
e
gul
a
r
iz
e
d
G
A
N
s
ha
ve
a
ls
o
im
pr
ov
e
d
r
e
s
t
or
a
ti
o
n
f
id
e
li
ty
i
n
r
e
a
l
-
w
or
ld
c
ondi
ti
o
ns
[
14]
–
[
16]
.
G
A
N
o
pt
im
iz
a
ti
on
h
a
s
e
vol
ve
d
th
r
oug
h
r
e
gul
a
r
i
z
a
ti
o
n,
c
ondi
ti
oni
ng,
a
nd
a
r
c
hi
te
c
tu
r
a
l
va
r
ia
nt
s
s
u
c
h
a
s
W
G
A
N
-
G
P
a
nd
P
ix
2P
i
x
-
s
ty
l
e
c
ondi
ti
ona
l
ge
ne
r
a
ti
ve
a
dv
e
r
s
a
r
ia
l
n
e
twor
k
(
c
G
A
N
s
)
,
c
om
m
o
nl
y
c
a
te
g
or
iz
e
d
in
th
e
li
te
r
a
tu
r
e
a
s
s
ho
w
n
in
F
ig
ur
e
1
[
17]
.
A
ddi
ti
ona
ll
y,
th
e
r
e
is
gr
ow
in
g
in
te
r
e
s
t
in
li
ght
w
e
ig
ht
a
nd
h
a
r
dw
a
r
e
-
e
f
f
ic
ie
nt
de
s
ig
ns
f
or
e
m
be
dde
d
pl
a
tf
or
m
s
,
e
na
bl
in
g r
e
a
l
-
ti
m
e
ha
z
e
r
e
m
ova
l
on e
dge
de
vi
c
e
s
[
18]
.
A
m
o
ng
th
e
s
e
a
pp
r
o
a
c
he
s
,
c
G
A
N
s
r
e
m
a
in
h
ig
hl
y
e
f
f
e
c
t
iv
e
f
o
r
de
ha
z
i
ng
be
c
a
us
e
t
he
y
e
n
f
o
r
c
e
b
ot
h
pi
xe
l
-
le
ve
l
a
c
c
u
r
a
c
y
a
n
d
pe
r
c
e
p
tu
a
l
r
e
a
li
s
m
in
a
n
i
m
a
ge
-
to
-
im
a
ge
t
r
a
ns
la
t
io
n
s
e
tt
in
g
[
1
9]
.
H
ow
e
ve
r
,
e
xi
s
ti
n
g
m
e
th
ods
o
f
te
n
s
t
r
ugg
le
t
o
p
r
e
s
e
r
ve
te
x
tu
r
e
s
ha
r
p
n
e
s
s
a
n
d
f
in
e
s
t
r
uc
tu
r
a
l
de
ta
i
l
in
r
e
a
l
-
w
o
r
l
d
hi
g
h
-
r
e
s
ol
ut
io
n
ha
z
e
.
T
o
a
dd
r
e
s
s
th
e
s
e
li
m
i
ta
ti
ons
,
th
is
w
or
k
p
r
op
os
e
s
a
n
e
n
ha
nc
e
d
c
G
A
N
f
r
a
m
e
w
o
r
k
in
t
e
g
r
a
t
in
g
a
n
a
t
te
n
ti
on
-
a
u
gm
e
nt
e
d
U
-
N
e
t
g
e
ne
r
a
t
o
r
,
a
P
a
tc
h
G
A
N
d
is
c
r
i
m
i
na
t
or
,
a
n
d
a
c
a
r
e
f
ul
ly
b
a
la
nc
e
d
m
ul
ti
-
c
om
pon
e
n
t
lo
s
s
f
un
c
t
io
n
.
T
he
m
a
in
c
on
tr
ib
ut
io
ns
of
t
hi
s
w
o
r
k
a
r
e
:
i
)
a
n
a
t
te
nt
io
n
-
gui
de
d
U
-
N
e
t
ge
ne
r
a
t
o
r
to
r
e
c
ov
e
r
b
ot
h
g
lo
ba
l
s
t
r
u
c
t
ur
e
a
n
d
f
in
e
te
xt
u
r
e
s
;
i
i
)
a
w
e
ig
ht
e
d
m
u
lt
i
-
lo
s
s
c
o
m
b
in
a
ti
on
(
r
e
c
o
ns
t
r
uc
ti
on
,
pe
r
c
e
p
tu
a
l,
e
d
ge
,
s
tr
uc
t
u
r
a
l
s
i
m
i
la
r
i
ty
(
S
S
I
M
)
,
a
n
d
a
dv
e
r
s
a
r
ia
l
)
t
o
im
p
r
ov
e
s
tr
uc
t
u
r
a
l
a
nd
pe
r
c
e
pt
ua
l
f
id
e
li
ty
;
a
nd
ii
i)
s
ta
t
e
-
of
-
th
e
-
a
r
t
pe
r
f
or
m
a
nc
e
on
t
he
r
e
a
l
-
w
or
ld
L
M
H
a
z
e
da
ta
s
e
t
,
s
u
r
pa
s
s
in
g
pr
io
r
G
A
N
a
nd
t
r
a
ns
f
or
m
e
r
-
ba
s
e
d a
pp
r
oa
c
he
s
in
pe
a
k s
ig
n
a
l
-
to
-
noi
s
e
r
a
t
io
(
P
S
N
R
)
a
nd
S
S
I
M
.
F
ig
ur
e
1
.
C
la
s
s
if
ic
a
ti
on of
G
A
N
[
15]
2.
R
E
L
A
T
E
D
WORK
R
e
c
e
nt
a
dva
n
c
e
s
in
im
a
ge
de
h
a
z
in
g
s
pa
n
th
r
e
e
m
a
jo
r
di
r
e
c
ti
on
s
:
pr
io
r
-
ba
s
e
d
m
e
th
ods
,
de
e
p
le
a
r
ni
ng
m
e
th
ods
,
a
nd
a
dve
r
s
a
r
ia
l
ge
ne
r
a
ti
ve
a
ppr
oa
c
he
s
.
E
a
r
li
e
r
w
or
ks
s
uc
h
a
s
da
r
k
c
ha
nne
l
pr
io
r
a
nd
f
us
io
n
-
ba
s
e
d
m
ode
ls
[
6]
–
[
9]
e
s
ti
m
a
te
a
tm
os
phe
r
ic
li
ght
a
nd
tr
a
ns
m
is
s
io
n
m
a
ps
vi
a
ha
ndc
r
a
f
te
d
a
s
s
um
pt
io
ns
,
but
ty
pi
c
a
ll
y
de
gr
a
de
i
n de
ns
e
or
s
pa
ti
a
ll
y va
r
yi
ng ha
z
e
.
W
it
h t
he
e
m
e
r
g
e
nc
e
of
de
e
p l
e
a
r
n
in
g,
C
N
N
-
a
nd e
nc
od
e
r
–
d
e
c
o
de
r
-
ba
s
e
d
a
r
c
hi
t
e
c
tu
r
e
s
i
m
pr
ov
e
d r
ob
us
t
ne
s
s
by l
e
a
r
ni
n
g di
r
e
c
t
m
a
p
pi
ngs
f
r
om
ha
z
y t
o ha
z
e
-
f
r
e
e
i
m
a
ge
s
[
10]
,
[
11]
.
G
A
N
-
ba
s
e
d
a
ppr
oa
c
he
s
f
ur
th
e
r
e
nha
nc
e
d
pe
r
c
e
pt
ua
l
r
e
a
li
s
m
by
e
nf
or
c
in
g
di
s
tr
ib
ut
io
n
c
ons
is
te
nc
y
be
twe
e
n
r
e
s
to
r
e
d
a
nd
gr
ound
-
tr
ut
h
im
a
ge
s
.
Z
hu
e
t
al
.
[
20]
in
tr
oduc
e
d
D
e
ha
z
e
G
A
N
,
r
e
f
or
m
ul
a
ti
ng
th
e
A
S
M
w
it
hi
n
a
ge
ne
r
a
ti
ve
f
r
a
m
e
w
or
k
a
nd
de
m
ons
tr
a
ti
ng
im
pr
ove
d
pe
r
f
or
m
a
nc
e
on
bot
h
in
door
a
nd
out
door
s
c
e
ne
s
.
Z
ha
ng
a
n
d
P
a
te
l
[
2
1]
pr
opo
s
e
d
d
e
n
s
e
ly
c
onn
e
c
t
e
d
pyr
a
m
id
de
ha
z
in
g
n
e
twor
k
(
D
C
P
D
N
)
,
w
hi
c
h
jo
in
tl
y
e
s
ti
m
a
te
s
tr
a
n
s
m
is
s
io
n
a
nd
de
h
a
z
e
d
o
ut
put
w
it
h
a
pyr
a
m
i
d
-
ba
s
e
d
d
e
s
i
gn
b
ut
s
tr
ugg
le
s
w
i
th
s
c
e
ne
s
c
ont
a
in
in
g
br
ig
ht
/whi
t
e
o
bj
e
c
ts
.
F
u
e
t
al
.
[
22]
u
ti
li
z
e
d
a
G
A
N
b
a
s
e
d
on
di
s
c
r
e
t
e
w
a
v
e
le
t
tr
a
n
s
f
or
m
f
or
i
m
a
g
e
r
e
s
to
r
a
ti
on
w
it
h
d
e
n
s
e
ha
z
e
,
w
hi
c
h
w
a
s
s
how
n
to
be
tt
e
r
pr
e
s
e
r
ve
hi
gh
-
f
r
e
q
ue
nc
y
te
xt
ur
e
.
L
ia
ng
e
t
al
.
[
13]
in
ve
s
ti
ga
te
d
a
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8814
I
nt
J
A
dv A
ppl
S
c
i
,
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
:
1340
-
1349
1342
tr
a
ns
f
or
m
e
r
b
a
s
e
d
r
e
c
ur
s
io
n s
tr
a
t
e
gy
f
or
im
a
ge
r
e
s
to
r
a
ti
on,
de
m
ons
tr
a
ti
ng
th
a
t
a
tt
e
nt
io
n c
a
n
e
f
f
e
c
ti
v
e
ly
c
a
pt
ur
e
lo
ng
r
a
ng
e
b
e
h
a
vi
or
w
hi
l
e
us
i
ng
f
e
w
e
r
p
a
r
a
m
e
te
r
s
.
M
e
a
n
w
hi
le
,
F
a
n
g
e
t
al
.
[
23]
pr
opo
s
e
d
a
dua
l
c
ol
or
-
s
pa
c
e
gui
de
d
m
od
e
l
t
ha
t
be
tt
e
r
pr
e
s
e
r
ve
d
s
tr
uc
tu
r
a
l
pr
ope
r
ti
e
s
of
s
c
e
n
e
s
in
r
e
a
l
-
w
or
ld
h
a
z
e
.
A
t
t
he
s
a
m
e
ti
m
e
,
A
lh
a
d
e
e
th
y
e
t
al
.
[
24]
a
ppl
i
e
d
p
e
r
c
e
pt
u
a
l
a
nd c
yc
l
e
-
c
o
ns
i
s
te
nc
y
c
ons
tr
a
in
t
s
t
o
in
d
oor
s
c
e
ne
s
, but
w
e
r
e
not
a
bl
e
to
qua
nt
it
a
ti
ve
ly
be
n
c
hm
a
r
k t
h
e
ir
r
e
s
ul
t
s
.
A
br
oa
d
e
r
c
o
m
pa
r
i
s
on
of
th
e
s
e
m
e
th
ods
is
pr
e
s
e
nt
e
d
in
T
a
bl
e
1,
hi
ghl
ig
ht
in
g
da
t
a
s
e
t
ty
pe
s
,
m
od
e
l
s
tr
a
t
e
gi
e
s
,
p
e
r
f
or
m
a
n
c
e
m
e
tr
ic
s
,
s
tr
e
n
gt
h
s
,
a
n
d
li
m
it
a
ti
on
s
.
A
s
s
how
n,
c
G
A
N
-
ba
s
e
d
f
or
m
ul
a
ti
ons
c
on
s
i
s
te
nt
ly
out
pe
r
f
or
m
tr
a
di
ti
ona
l
pr
i
or
s
in
pe
r
c
e
pt
ua
l
qu
a
li
ty
but
of
te
n s
tr
u
ggl
e
w
it
h
te
xt
ur
e
c
ons
is
t
e
nc
y a
nd
c
ol
or
f
id
e
li
ty
unde
r
c
om
pl
e
x
r
e
a
l
-
w
or
ld
h
a
z
e
.
T
he
s
e
ga
p
s
m
ot
iv
a
t
e
th
e
pr
e
s
e
nt
w
or
k
,
w
hi
c
h
in
t
e
gr
a
t
e
s
a
tt
e
nt
io
n
gui
d
a
nc
e
a
nd a
m
ul
ti
-
c
om
pon
e
nt
l
os
s
t
o
i
m
pr
ov
e
s
tr
uc
tu
r
a
l
r
e
c
ov
e
r
y a
nd r
e
a
li
s
m
on
hi
gh
-
q
ua
li
ty
r
e
a
l
-
w
or
ld
d
a
ta
s
e
ts
.
T
a
bl
e
1
.
T
he
s
um
m
a
r
y of
t
he
r
e
la
te
d w
or
k
R
e
f
.#
D
a
t
a
s
e
t
M
e
t
hodol
ogy
M
e
t
r
i
c
s
A
dva
nt
a
ge
s
L
i
m
i
t
a
t
i
ons
[
20]
S
U
N
-
R
G
B
D
,
NYU
-
D
e
pt
h,
a
nd C
O
C
O
C
G
A
N
I
ndoor
(
P
S
N
R
=22.15,
S
S
I
M
=0.8727)
A
nd out
door
(
P
S
N
R
=24.94,
S
S
I
M
=0.9169)
E
a
r
l
i
e
r
c
ondi
t
i
ona
l
G
A
N
w
i
t
h good
r
e
s
ul
t
s
P
r
oduc
e
d a
r
t
i
f
a
c
t
s
i
n
de
ns
e
ha
z
e
[
21]
NYU
-
de
pt
h2
R
E
S
I
D
E
C
yc
l
e
G
A
N
i
n
D
C
P
D
N
S
S
I
M
~
0.9
-
0.965
T
hi
s
m
ode
l
i
s
e
nha
nc
e
d us
i
ng a
ne
w
l
os
s
f
unc
t
i
on
f
or
e
dge
-
pr
e
s
e
r
vi
ng
D
C
P
f
a
i
l
e
d i
n i
m
a
ge
s
w
i
t
h w
hi
t
e
obj
e
c
t
s
[
22]
Nh
-
H
a
z
e
N
h
-
H
a
z
e
2 D
e
ns
e
-
H
a
z
e
A
di
s
c
r
e
t
e
w
a
ve
l
e
t
t
r
a
ns
f
or
m
GAN
P
S
N
R
~
21.99
S
S
I
M
~
0.856
G
ood a
r
c
hi
t
e
c
t
ur
e
c
ons
i
s
t
s
of
t
w
o
ge
ne
r
a
t
or
s
T
r
a
i
ni
ng t
w
o ge
ne
r
a
t
or
s
a
nd di
s
c
r
i
m
i
na
t
i
on i
s
di
f
f
i
c
ul
t
[
13]
R
a
i
n800,
R
a
i
n100L
,
R
a
i
n100H
,
S
now
100K
de
r
a
i
ni
ng a
r
e
c
ur
s
i
ve
t
r
a
ns
f
or
m
e
r
(
D
R
T
)
P
S
N
R
=(
27.02, 37.61,
29.47, a
nd 28.04
-
32.15)
S
S
I
M
=(
0.847, 0.948,
a
nd 0.846)
U
s
e
d one
t
r
a
ns
f
or
m
e
r
a
nd
r
e
pe
a
t
e
d i
t
on
di
f
f
e
r
e
nt
s
a
m
pl
e
s
R
e
c
ur
s
i
ve
c
a
n l
e
a
d t
o
c
um
ul
a
t
i
ve
e
r
r
or
s
i
f
not
de
s
i
gne
d pe
r
f
e
c
t
l
y
[
19]
F
R
I
D
A
F
o
c
u
s
,
f
l
e
x
,
a
nd
e
nt
r
o
py
f
a
de
c
o
m
p
o
ne
n
t
b
l
oc
ks
w
i
t
h
a
n
a
t
t
e
n
t
i
on
m
e
c
h
a
n
i
s
m
P
S
N
R
~
25.4700
-
31.8100
S
S
I
M
~
0.8028
-
0.9573
T
h
e
c
a
p
a
c
i
t
y
t
o
i
m
p
r
o
v
e
i
m
a
g
e
s
ha
r
p
n
e
s
s
a
nd
i
t
s
f
e
a
t
u
r
e
s
A
m
pl
i
f
i
e
d c
om
put
a
t
i
ona
l
c
om
pl
e
xi
t
y
[
23]
R
W
2A
H
a
nd
R
e
a
l
-
w
or
l
d
s
m
oke
G
ui
de
d
de
ha
z
i
ng
ne
t
w
or
k
(
S
G
D
N
)
P
S
N
R
~
(
22.26 a
nd
23.41)
S
S
I
M
~
(
0.668
a
nd 0.790)
S
u
pe
r
i
o
r
pe
r
f
o
r
m
a
n
c
e
o
n
r
e
a
l
-
w
o
r
l
d s
m
ok
e
a
n
d
ha
z
e
T
h
e
i
n
he
r
e
n
t
d
i
f
f
i
c
u
l
t
y
o
f
a
c
h
i
e
v
i
n
g
p
e
r
f
e
c
t
p
i
x
e
l
-
w
i
s
e
a
l
i
g
n
m
e
nt
i
n
t
h
e
r
e
a
l
w
o
r
l
d
[
24]
NYU
d
e
pt
h
GAN
N
um
e
r
i
c
a
l
s
c
or
e
s
not
f
ound
U
s
e
pe
r
c
e
pt
ua
l
a
nd
c
yc
l
e
-
c
ons
i
s
t
e
nc
y
l
os
s
e
s
A
bs
e
nc
e
of
qua
nt
i
t
a
t
i
ve
da
t
a
3.
M
E
T
H
O
D
T
hi
s
w
or
k
pr
opos
e
s
a
c
a
r
e
f
ul
ly
e
ngi
ne
e
r
e
d
P
ix
2P
ix
G
A
N
,
in
te
gr
a
ti
ng
U
-
N
e
t
w
it
h
e
nha
nc
e
d
s
ki
p
c
onne
c
ti
ons
a
nd
a
P
a
tc
hG
A
N
di
s
c
r
im
in
a
to
r
,
a
im
in
g
to
ba
la
nc
e
lo
c
a
l
de
ta
il
pr
e
s
e
r
va
ti
on
w
it
h
gl
oba
l
c
ons
is
te
nc
y.
T
he
pr
opos
e
d
G
A
N
c
ons
is
t
s
of
a
U
-
N
e
t
-
ba
s
e
d
ge
ne
r
a
to
r
a
nd
a
P
a
tc
hG
A
N
di
s
c
r
im
in
a
to
r
.
R
e
li
e
s
e
s
s
e
nt
ia
ll
y
on
di
s
c
ove
r
in
g
a
nd
s
a
vi
ng
th
e
pa
tt
e
r
ns
of
th
e
ha
z
e
pr
e
s
e
nt
e
d
in
th
e
in
put
im
a
ge
s
,
a
s
il
lu
s
tr
a
te
d
in
F
ig
ur
e
2.
W
he
r
e
a
s
c
onve
nt
io
na
l
m
ode
ls
(
e
.g.,
D
a
r
k
C
ha
nne
l
P
r
io
r
)
r
e
ly
on
phys
ic
a
l
pr
io
r
s
a
nd
de
e
p
le
a
r
ni
ng
m
ode
ls
(
e
.g.,
C
yc
le
G
A
N
)
r
e
qui
r
e
pa
ir
e
d
da
ta
,
our
m
ode
l
ove
r
c
om
e
s
li
m
it
a
ti
ons
s
uc
h
a
s
a
r
ti
f
a
c
ts
in
d
e
ns
e
ha
z
e
a
nd
c
os
tl
y
c
om
put
a
ti
ons
w
it
h
a
hybr
id
G
A
N
m
ode
l.
T
he
m
a
in
c
ont
r
ib
ut
io
ns
a
r
e
:
i)
a
U
-
N
e
t
ge
ne
r
a
to
r
w
it
h
s
e
lf
-
at
te
nt
io
n
f
or
gl
oba
l
ha
z
e
r
e
m
ova
l;
ii
)
a
P
a
tc
hG
A
N
di
s
c
r
im
in
a
to
r
f
or
pr
e
s
e
r
ve
d
lo
c
a
l
de
ta
il
s
;
a
nd
ii
i)
a
w
e
ig
ht
e
d l
os
s
f
unc
ti
on f
or
pi
xe
l
-
w
is
e
a
nd pe
r
c
e
pt
ua
l
qua
li
ty
.
3.1. P
r
op
os
e
d
p
ix
2p
ix
G
A
N
n
e
t
w
or
k
T
he
m
od
e
l
a
dh
e
r
e
s
t
o
th
e
tr
a
di
ti
ona
l
pi
x
2pi
x
f
r
a
m
e
w
or
k
in
w
h
ic
h
a
U
-
N
e
t
g
e
ne
r
a
to
r
p
e
r
f
or
m
s
h
a
z
e
r
e
m
ova
l
a
nd
a
P
a
t
c
hG
A
N
di
s
c
r
im
i
na
to
r
pr
om
ot
e
s
lo
c
a
l
r
e
a
li
s
m
.
T
h
e
U
-
N
e
t
e
m
pl
oy
s
s
tr
id
e
d
c
onv
ol
ut
io
n
s
to
dow
n
s
a
m
pl
e
th
e
im
a
ge
,
u
ti
li
z
e
s
s
ki
p
c
on
ne
c
ti
on
s
to
r
e
c
ov
e
r
s
p
a
ti
a
l
d
e
ta
i
l,
a
nd
a
p
pl
ie
s
a
bot
tl
e
ne
c
k
a
t
te
nt
i
on
la
ye
r
f
or
gl
oba
l
c
o
nt
e
xt
.
T
he
ge
n
e
r
a
to
r
g
e
n
e
r
a
te
s
a
r
e
s
to
r
e
d ha
z
e
-
f
r
e
e
i
m
a
ge
c
ondi
ti
one
d
on t
h
e
h
a
z
y
i
nput
. T
he
di
s
c
r
im
in
a
to
r
a
s
s
e
s
s
e
s
th
e
ge
ne
r
a
t
or
im
a
g
e
s
pa
t
c
hw
i
s
e
(
r
a
th
e
r
th
a
n
gl
o
ba
ll
y)
,
a
ll
o
w
in
g
it
to
id
e
nt
if
y
lo
c
a
l
in
c
on
s
is
te
n
c
ie
s
a
nd
in
c
e
nt
iv
iz
in
g
th
e
g
e
ne
r
a
to
r
to
pr
odu
c
e
s
h
a
r
pe
r
a
nd
m
or
e
r
e
a
l
is
ti
c
r
e
c
o
ns
tr
u
c
ti
on
s
.
T
he
ove
r
a
ll
a
r
c
hi
t
e
c
t
ur
e
of
t
h
e
di
s
c
r
im
in
a
to
r
i
s
s
ho
w
n i
n F
ig
ur
e
3.
T
he
opt
im
i
z
a
ti
on u
s
e
s
a
c
G
A
N
l
os
s
c
o
ns
i
s
ti
n
g of
s
e
v
e
r
a
l
c
o
m
pon
e
nt
s
:
t
he
a
dv
e
r
s
a
r
ia
l
obj
e
c
t
iv
e
(
3)
e
n
c
our
a
g
e
s
r
e
a
li
s
m
,
th
e
L
1
r
e
c
on
s
tr
u
c
ti
on
lo
s
s
(
4)
m
a
in
ta
in
s
pi
xe
l
-
w
i
s
e
a
c
c
ur
a
c
y
,
a
nd
t
he
f
in
a
l
w
e
ig
ht
e
d
lo
s
s
f
or
m
u
la
ti
o
n
jo
in
tl
y
e
nf
or
c
e
s
bo
th
s
tr
uc
tu
r
a
l
0.
f
id
e
li
ty
a
nd
pe
r
c
e
pt
ua
l
r
e
a
l
is
m
, e
n
s
ur
in
g t
ha
t
th
e
ge
ne
r
a
t
or
pr
e
s
e
r
v
e
s
f
in
e
te
x
tu
r
e
w
hi
le
pr
e
ve
nt
in
g o
ve
r
-
s
m
oot
hi
ng.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
dv A
ppl
S
c
i
I
S
S
N
:
2252
-
8814
G
e
ne
r
at
iv
e
ad
v
e
r
s
ar
ia
l
ne
tw
o
r
k
f
or
i
nt
e
ll
ig
e
nt
ha
z
e
r
e
m
ov
al
…
(
A
li
A
bdu
la
z
e
e
z
M
ohamm
e
d B
aqe
r
Q
a
z
z
a
z
)
1343
(
,
)
=
{
,
}
[
(
,
)
]
+
{
}
[
l
o
g
(
1
−
(
,
(
)
)
)
]
(
5)
=
[
|
|
−
(
)
|
|
−
1
]
(
6)
(
)
=
1
.
5
×
RL
+
PL
+
EL
+
2
.
5
×
S
S
IM
+
0
.
1
×
AL
(
7)
F
ig
ur
e
2
.
B
lo
c
k di
a
gr
a
m
of
t
he
pr
opos
e
d s
ys
te
m
F
ig
ur
e
3
.
T
he
di
s
c
r
im
in
a
to
r
a
r
c
hi
te
c
tu
r
e
3.2.
D
at
as
e
t
F
ig
ur
e
4
s
how
s
th
e
L
M
H
a
z
e
da
ta
s
e
t
c
ont
a
in
s
a
c
ol
le
c
ti
on
of
1
,
115
pa
ir
e
d
r
e
a
l
-
w
or
ld
ha
z
y
.
F
ig
ur
e
4(
a
)
a
nd
c
le
a
r
im
a
ge
s
c
a
pt
ur
e
d
a
nd
F
ig
ur
e
4(
b)
in
di
f
f
e
r
e
nt
a
tm
os
phe
r
ic
c
ont
e
xt
s
a
nd
li
ght
in
g
c
ondi
ti
ons
.
T
hi
s
da
ta
s
e
t
w
a
s
s
e
le
c
t
e
d
be
c
a
us
e
th
e
c
ol
le
c
ti
on
is
r
e
pr
e
s
e
nt
a
ti
ve
of
a
c
tu
a
l
ha
z
e
c
ha
r
a
c
te
r
is
ti
c
s
of
th
e
im
a
ge
s
in
s
t
e
a
d
of
s
ynt
he
ti
c
a
ll
y
ove
r
la
yi
ng
it
f
or
ge
ne
r
a
li
z
a
t
io
n
to
r
e
a
l
de
pl
oym
e
nt
s
c
e
n
a
r
io
s
.
T
h
e
da
ta
s
e
t
is
di
vi
de
d i
nt
o 80%
f
or
t
r
a
in
in
g a
nd 20%
f
or
t
e
s
ti
ng t
o e
ns
ur
e
a
f
a
ir
e
va
lu
a
ti
on.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8814
I
nt
J
A
dv A
ppl
S
c
i
,
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
:
1340
-
1349
1344
(
a
)
(
b)
F
ig
ur
e
4
.
I
m
a
ge
s
f
r
om
t
he
L
M
H
a
z
e
da
ta
s
e
t
(
a
)
ha
z
e
i
m
a
ge
a
nd
(
b)
c
le
a
r
i
m
a
ge
3
.3.
P
r
e
p
r
oc
e
s
s
in
g an
d
im
p
le
m
e
n
t
at
io
n
d
e
t
ai
ls
T
he
L
M
H
a
z
e
da
ta
s
e
t
c
ont
a
in
s
1
,
115
r
e
a
l
-
w
or
ld
ha
z
y/
c
le
a
r
im
a
ge
pa
ir
s
.
I
m
a
ge
s
a
r
e
r
e
s
iz
e
d
to
256×
256
a
nd
nor
m
a
li
z
e
d
to
[
-
1,
1
]
.
T
he
da
ta
s
e
t
is
s
pl
it
in
to
80%
f
or
tr
a
in
in
g
a
nd
20
%
f
or
te
s
ti
ng
.
T
he
m
ode
l
w
a
s
im
pl
e
m
e
nt
e
d
in
P
yT
or
c
h
a
nd
tr
a
in
e
d
on
a
n
N
V
I
D
I
A
R
T
X
3090
G
P
U
us
in
g
A
da
m
w
it
h
a
ba
tc
h
s
iz
e
of
8
f
or
75
e
poc
hs
.
T
he
ge
ne
r
a
to
r
a
nd
di
s
c
r
im
in
a
to
r
a
r
e
upda
te
d
a
lt
e
r
na
te
ly
w
it
h
le
a
r
ni
ng
r
a
te
s
of
0.0002
a
nd
0.00008, r
e
s
pe
c
ti
ve
ly
.
4.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
T
hi
s
s
e
c
ti
on
f
oc
us
e
s
on
th
e
r
e
s
ul
ts
a
nd
di
s
c
us
s
io
n.
I
t
c
ove
r
s
th
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
a
nd
pe
r
f
or
m
a
nc
e
m
e
tr
ic
s
. A
ddi
ti
ona
ll
y, i
t
pr
ovi
de
s
a
de
ta
il
e
d a
na
ly
s
is
of
a
ll
pr
e
vi
ous
s
e
c
ti
ons
.
4.1.
E
xp
e
r
im
e
n
t
al
r
e
s
u
lt
s
T
he
pr
opos
e
d
s
ys
te
m
w
il
l
be
a
ppl
ie
d
to
m
a
ny
im
a
g
e
s
f
r
om
th
e
pr
opos
e
d
da
ta
s
e
t.
I
t
out
pe
r
f
or
m
e
d
pr
e
vi
ous
s
ta
te
-
of
-
th
e
-
a
r
t
m
e
th
ods
w
it
h
a
P
S
N
R
of
33.42
a
nd
S
S
I
M
of
0.9590
on
th
e
te
s
ti
ng
s
e
t.
T
he
r
e
s
ul
ts
a
r
e
s
how
n
in
F
ig
ur
e
5
to
in
di
c
a
te
th
e
pe
r
f
or
m
a
nc
e
of
th
e
pr
opo
s
e
d s
ys
te
m
in
a
s
ubj
e
c
ti
ve
m
a
nne
r
.
F
ig
ur
e
5(
a
)
s
how
s
t
he
ha
z
e
i
m
a
ge
,
F
ig
ur
e
5
(
b)
s
how
s
t
he
de
h
a
z
e
i
m
a
ge
,
a
n
d
F
ig
ur
e
5
(
c
)
s
how
s
t
he
c
le
a
r
i
m
a
ge
.
(
a
)
(
b)
(
c
)
F
ig
ur
e
5
.
R
e
s
ul
ts
of
t
he
pr
opos
e
d s
y
s
te
m
(
a
)
ha
z
e
i
m
a
ge
,
(
b)
de
ha
z
e
i
m
a
ge
, a
nd
(
c
)
c
le
a
r
i
m
a
ge
4.2.
P
e
r
f
or
m
an
c
e
m
e
t
r
ic
s
T
h
e
pr
op
o
s
e
d
s
y
s
t
e
m
r
e
a
c
h
e
d
t
he
f
ol
lo
w
in
g
e
f
f
i
c
i
e
n
c
y
w
he
n
a
p
pl
ie
d
to
t
h
e
im
a
g
e
s
of
t
h
e
d
a
t
a
s
e
t
,
w
h
e
r
e
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
dv A
ppl
S
c
i
I
S
S
N
:
2252
-
8814
G
e
ne
r
at
iv
e
ad
v
e
r
s
ar
ia
l
ne
tw
o
r
k
f
or
i
nt
e
ll
ig
e
nt
ha
z
e
r
e
m
ov
al
…
(
A
li
A
bdu
la
z
e
e
z
M
ohamm
e
d B
aqe
r
Q
a
z
z
a
z
)
1345
i)
T
r
a
in
in
g
s
e
t:
w
he
n
pe
r
f
or
m
in
g
th
e
tr
a
in
in
g
pr
oc
e
s
s
,
th
e
be
s
t
r
e
s
ul
ts
of
th
e
pr
opos
e
d
s
y
s
te
m
a
r
e
P
S
N
R
=
29.3 a
nd S
S
I
M
=
0.9656. T
h
e
s
e
r
e
s
ul
ts
w
il
l
be
r
e
a
c
he
d i
n
e
poc
h (
75)
a
s
i
ll
us
tr
a
te
d i
n F
ig
ur
e
6
.
ii)
T
e
s
ti
ng
s
e
t:
w
he
n
pe
r
f
or
m
in
g
th
e
te
s
ti
ng
pr
oc
e
s
s
,
th
e
a
ve
r
a
ge
r
e
s
ul
ts
of
th
e
pr
opos
e
d
s
ys
t
e
m
a
r
e
P
S
N
R
=
33.42 a
nd S
S
I
M
=
0.9590.
A
s
s
how
n
in
F
ig
ur
e
6,
bot
h
P
S
N
R
a
nd
S
S
I
M
in
c
r
e
a
s
e
s
te
a
di
ly
w
it
h
tr
a
in
in
g,
c
onf
ir
m
in
g
s
ta
bl
e
c
onve
r
ge
nc
e
.
T
he
P
S
N
R
r
is
e
s
f
r
om
22.31
a
t
e
po
c
h
5
to
29.
3
by
e
poc
h
75,
w
hi
le
S
S
I
M
im
pr
ove
s
f
r
om
0.7831
to
0.9656
ove
r
th
e
s
a
m
e
pe
r
io
d.
T
he
ge
ne
r
a
to
r
lo
s
s
de
c
r
e
a
s
e
s
c
ons
is
t
e
nt
ly
,
a
nd
th
e
di
s
c
r
im
in
a
to
r
r
e
m
a
in
s
s
ta
bl
e
,
in
di
c
a
ti
ng
a
ba
la
nc
e
d
a
dve
r
s
a
r
ia
l
pr
oc
e
s
s
w
it
hout
c
ol
la
ps
e
.
T
he
r
e
s
ul
ti
ng
de
h
a
z
e
d
im
a
ge
s
e
xhi
bi
t
s
ha
r
p
te
xt
ur
e
s
a
nd
na
tu
r
a
l
c
ol
or
s
w
it
h
no
ha
lo
in
g
or
ov
e
r
s
m
oot
hi
ng,
de
m
ons
tr
a
ti
ng
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
t
he
m
ul
ti
-
lo
s
s
opt
im
iz
a
ti
on s
tr
a
te
gy.
F
ig
ur
e
6
.
R
e
s
ul
ts
i
n t
he
t
r
a
in
in
g pr
oc
e
s
s
of
t
he
pr
opos
e
d s
ys
t
e
m
4.3.
A
b
la
t
io
n
s
t
u
d
y
W
he
n
pr
o
pos
in
g
ot
h
e
r
va
lu
e
s
of
th
e
n
e
twor
k
p
a
r
a
m
e
te
r
s
,
a
s
f
o
ll
ow
s
:
t
ot
a
l
e
poc
hs
=
75
,
ba
tc
h
s
i
z
e
=
4,
le
a
r
ni
ng
r
a
te
o
f
th
e
ge
ne
r
a
to
r
=
0.0001,
le
a
r
ni
ng
r
a
te
of
th
e
di
s
c
r
im
in
a
to
r
=
0.00005,
di
s
c
r
im
in
a
to
r
upda
te
f
r
e
que
nc
y=
2,
a
nd
lo
s
s
w
e
ig
ht
s
=
{
r
e
c
on
s
tr
uc
ti
on=
12.0,
pe
r
c
e
pt
ua
l=
0.5,
e
dg
e
=
0.5,
s
s
im
=
2.0,
a
nd
a
dve
r
s
a
r
ia
l=
0.01}
.
F
ig
ur
e
7
s
how
s
t
he
be
s
t
r
e
s
ul
ts
of
th
e
s
y
s
te
m
w
il
l
not
be
a
s
good
a
s
th
e
pr
e
vi
ous
,
w
he
r
e
P
S
N
R
=
20.71 dB
, a
s
s
how
n i
n F
ig
ur
e
7(
a
)
a
nd S
S
I
M
=
0.7943, a
s
s
how
n
in
F
ig
ur
e
7
(
b)
. W
it
h a
n uns
ta
bl
e
s
ta
te
f
or
th
e
va
lu
e
s
of
th
e
e
xpl
a
in
e
d
m
e
tr
ic
s
.
A
s
w
e
ll
a
s
w
he
n
a
ppl
y
in
g
a
n
e
nha
nc
e
d
te
c
hni
que
a
s
a
pr
e
pr
oc
e
s
s
in
g
te
c
hni
que
,
th
e
r
e
s
ul
ts
of
th
e
p
r
opos
e
d
s
ys
te
m
w
il
l
be
w
or
s
e
th
a
n
th
e
or
ig
in
a
l
r
e
s
ul
ts
,
a
s
f
ol
lo
w
s
:
to
ta
l
e
poc
hs
=
75,
ba
tc
h
s
iz
e
=
4,
le
a
r
ni
ng
r
a
te
of
th
e
ge
ne
r
a
to
r
=
0.00
01,
le
a
r
ni
ng
r
a
te
of
th
e
di
s
c
r
im
in
a
to
r
=
0.0001,
di
s
c
r
im
in
a
to
r
upda
te
f
r
e
que
nc
y=
41,
a
nd
lo
s
s
w
e
ig
ht
s
=
{
r
e
c
ons
tr
uc
ti
on=
1.0,
s
s
im
=
1.0,
pe
r
c
e
pt
ua
l
=
0.5,
e
dge
=
1.5,
a
nd
a
dv
e
r
s
a
r
ia
l=
0.05}
a
s
w
e
ll
a
s
th
e
s
y
s
te
m
a
ppl
yi
n
g
C
L
A
H
E
m
e
th
od
to
th
e
in
put
im
a
ge
s
b
e
f
or
e
tr
a
in
in
g
a
s
a
pr
e
pr
oc
e
s
s
in
g
m
e
th
od.
F
ig
ur
e
8
s
how
s
th
e
be
s
t
r
e
s
ul
ts
of
th
e
s
ys
te
m
w
il
l
not
be
a
s
good
a
s
th
e
pr
e
vi
ous
,
w
he
r
e
P
S
N
R
=
20.40
,
a
s
s
how
n
in
F
ig
ur
e
8(
a
)
a
nd
S
S
I
M
=
0.7123
,
a
s
s
ho
w
n
in
F
ig
ur
e
8(
b)
,
a
nd
th
e
va
lu
e
s
of
th
e
m
de
c
r
e
a
s
e
d
a
s
th
e
num
be
r
of
e
poc
hs
in
c
r
e
a
s
e
d
a
f
te
r
e
poc
h
(
15
)
.
T
he
a
bl
a
ti
on
s
tu
dy
c
onf
ir
m
e
d
th
e
pe
r
f
e
c
t
s
e
le
c
ti
on of
hype
r
pa
r
a
m
e
te
r
c
hoi
c
e
s
a
nd l
os
s
w
e
ig
ht
s
.
(
a
)
(
b)
F
ig
ur
e
7
.
R
e
s
ul
ts
of
t
he
a
bl
a
ti
on s
ta
te
1 (
a
)
P
S
N
R
f
or
a
bl
a
ti
on s
t
a
te
a
nd
(
b)
S
S
I
M
f
or
a
bl
a
ti
on s
ta
te
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8814
I
nt
J
A
dv A
ppl
S
c
i
,
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
:
1340
-
1349
1346
(
a
)
(
b)
F
ig
ur
e
8
.
R
e
s
ul
ts
of
t
he
a
bl
a
ti
on s
ta
te
(
a
)
P
S
N
R
f
or
a
bl
a
ti
on s
ta
t
e
(
b)
S
S
I
M
f
or
a
bl
a
ti
on s
ta
te
4.4.
C
om
p
ar
is
on
w
it
h
ot
h
e
r
s
ys
t
e
m
s
T
h
e
r
e
s
ul
t
s
of
th
e
pr
op
o
s
e
d
s
y
s
t
e
m
a
r
e
g
oo
d
w
he
n
c
om
p
a
r
e
d w
it
h
ot
h
e
r
m
e
t
ho
ds
f
or
d
e
h
a
z
in
g
im
a
g
e
s
,
a
s
s
how
n
in
T
a
bl
e
2
a
n
d
F
ig
ur
e
9.
A
s
s
h
ow
n
in
T
a
bl
e
2,
th
e
pr
o
po
s
e
d
m
e
th
od
o
ut
p
e
r
f
o
r
m
s
e
x
i
s
ti
ng
t
e
c
hni
qu
e
s
,
a
c
hi
e
vi
ng
a
n 1
1.
43
-
m
a
r
k hi
gh
e
r
P
S
N
R
t
ha
n D
C
P
D
N
[
21]
a
nd
23
%
be
t
te
r
S
S
I
M
t
h
a
n S
G
D
N
[
2
3]
.
V
i
s
u
a
l
r
e
s
ul
t
s
in
F
i
gur
e
4
d
e
m
on
s
tr
a
t
e
e
f
f
e
c
t
iv
e
h
a
z
e
r
e
m
o
va
l
in
c
o
m
pl
e
x
s
c
e
ne
s
(
e
.
g.,
ur
b
a
n
a
r
e
a
s
w
it
h
d
e
n
s
e
f
og)
.
L
im
it
a
ti
ons
in
c
lu
de
m
ode
r
a
te
c
om
put
a
ti
ona
l
ove
r
he
a
d
(
~
15%
s
lo
w
e
r
th
a
n
S
G
D
N
due
to
th
e
s
e
lf
-
a
tt
e
nt
io
n
m
odul
e
)
.
I
t
is
im
por
ta
nt
to
r
e
a
li
z
e
th
a
t
c
om
pa
r
is
ons
of
di
f
f
e
r
e
nt
m
e
tr
ic
s
f
r
om
va
r
yi
ng
da
ta
s
e
ts
c
a
n
be
de
c
e
pt
iv
e
.
T
he
r
e
f
or
e
,
f
or
a
n
e
qui
ta
bl
e
c
om
p
a
r
is
on,
w
e
w
oul
d
in
ve
s
ti
ga
te
a
ll
c
om
pe
ti
ng
m
e
th
ods
on
our
L
M
H
a
z
e
da
ta
s
e
t,
but
due
to
th
e
r
e
s
ul
ts
a
s
r
e
por
te
d,
w
e
c
a
n
s
e
e
a
m
a
r
ke
d
im
pr
ove
m
e
nt
w
it
h
our
m
e
th
od
ove
r
ot
he
r
m
e
th
ods
s
uc
h
a
s
D
C
P
,
A
O
D
-
N
e
t,
a
nd
F
F
A
-
N
e
t
on
va
r
yi
ng
da
ta
s
e
ts
,
in
c
lu
di
ng
a
di
f
f
ic
ul
t
r
e
a
l
-
w
or
ld
be
nc
hm
a
r
k
da
ta
s
e
t.
O
ur
qua
li
ta
ti
ve
a
n
a
ly
s
e
s
in
f
or
m
us
th
a
t
th
e
r
e
is
be
tt
e
r
te
xt
ur
e
r
e
c
ov
e
r
y
a
nd
f
e
w
e
r
h
a
lo
a
r
ti
f
a
c
ts
, pr
obl
e
m
s
e
xpe
r
ie
nc
e
d i
n pr
io
r
-
ba
s
e
d a
nd e
a
r
li
e
r
c
om
pa
r
e
d t
o our
l
a
te
r
de
e
p
-
le
a
r
ni
ng mode
ls
.
T
a
bl
e
2
.
R
e
s
ul
ts
c
om
p
a
r
is
on
R
e
f
.#
P
S
N
R
↑
S
S
I
M
↑
D
a
t
a
s
e
t
[
13]
A
vg.=30,858
A
vg.=0.8803
R
a
i
n800, R
a
i
n100L
,
R
a
i
n100H
, a
nd S
now
100K
[
19]
A
vg.=28.64
A
vg.=0.88005
F
R
I
D
A
[
20]
24.94
0.9169
S
U
N
-
R
G
B
D
, N
Y
U
-
D
e
pt
h, a
nd C
O
C
O
[
22]
21.99
0.856
Nh
-
H
a
z
e
, N
h
-
H
a
z
e
2, a
nd D
e
n
s
e
-
H
a
z
e
[
23]
A
vg.=22.835
A
vg.=0.729
RW
2
A
H
a
nd R
e
a
l
-
w
or
l
d S
m
oke
our
33.42
0.9590
L
M
H
a
z
e
F
ig
ur
e
9
.
C
om
pa
r
is
on of
P
S
N
R
a
nd S
S
I
M
de
ha
z
in
g m
e
th
ods
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
dv A
ppl
S
c
i
I
S
S
N
:
2252
-
8814
G
e
ne
r
at
iv
e
ad
v
e
r
s
ar
ia
l
ne
tw
o
r
k
f
or
i
nt
e
ll
ig
e
nt
ha
z
e
r
e
m
ov
al
…
(
A
li
A
bdu
la
z
e
e
z
M
ohamm
e
d B
aqe
r
Q
a
z
z
a
z
)
1347
4.5.
S
t
r
e
n
gt
h
s
an
d
li
m
it
at
io
n
s
of
t
h
e
p
r
op
os
e
d
s
ys
t
e
m
T
he
pr
opos
e
d
s
ys
te
m
de
m
on
s
tr
a
te
s
s
e
ve
r
a
l
s
tr
e
ngt
hs
,
in
c
lu
di
ng
th
e
us
e
of
a
r
e
a
l‐
w
or
ld
L
M
H
a
z
e
da
ta
s
e
t
.
A
c
a
r
e
f
ul
ly
de
s
ig
ne
d
m
ul
ti
-
lo
s
s
f
unc
ti
on
th
a
t
e
ns
ur
e
s
hi
gh
f
id
e
li
ty
in
r
e
c
ons
tr
uc
ti
on,
a
nd
s
tr
ong
c
om
pa
r
a
ti
ve
pe
r
f
or
m
a
nc
e
m
e
tr
ic
s
th
a
t
v
a
li
da
te
it
s
e
f
f
e
c
ti
ve
ne
s
s
.
H
ow
e
ve
r
,
it
a
ls
o
ha
s
c
e
r
ta
in
li
m
it
a
ti
ons
,
s
uc
h
a
s
th
e
a
bs
e
nc
e
of
a
bl
a
ti
on
s
tu
di
e
s
to
e
va
lu
a
te
th
e
in
di
vi
dua
l
im
pa
c
t
of
e
a
c
h
lo
s
s
c
om
pone
nt
,
li
m
it
e
d
e
xpl
or
a
ti
on
of
uns
upe
r
vi
s
e
d
or
hybr
id
le
a
r
ni
ng
s
tr
a
te
gi
e
s
,
a
nd
onl
y
a
m
ode
s
t
le
ve
l
of
th
e
or
e
ti
c
a
l
ju
s
ti
f
ic
a
ti
on
f
or
s
om
e
of
t
he
a
r
c
hi
te
c
tu
r
a
l
de
s
ig
n c
hoi
c
e
s
.
4.5.
F
u
t
u
r
e
d
ir
e
c
t
io
n
s
i
n
c
lu
d
e
T
he
a
ddi
ti
on
of
m
or
e
e
xt
r
e
m
e
ha
z
e
c
ondi
ti
ons
to
th
e
da
ta
s
e
t'
s
r
a
nge
of
c
ondi
ti
ons
a
ll
ow
s
f
or
th
e
e
xpl
or
a
ti
on
of
tr
a
ns
f
or
m
e
r
-
ba
s
e
d
a
r
c
hi
te
c
tu
r
e
s
th
a
t
c
a
n
le
a
r
n
g
lo
ba
l
c
ont
e
xt
.
A
not
he
r
a
ppr
oa
c
h
w
oul
d
b
e
to
a
ppl
y
th
e
m
ode
l
on
vi
de
o
s
e
que
nc
e
s
to
a
c
hi
e
ve
te
m
por
a
l
c
ons
is
te
nc
y
in
d
e
ha
z
in
g,
a
l
s
o
ut
il
iz
in
g
s
e
lf
-
s
upe
r
vi
s
e
d
le
a
r
ni
ng
m
e
th
ods
.
T
hi
s
h
a
s
be
e
n
a
f
f
e
c
te
d
by
[
25]
ba
s
e
d
on
F
P
G
A
-
ba
s
e
d
im
pl
e
m
e
nt
a
ti
on,
opt
im
iz
in
g
th
e
m
ode
l
f
or
r
e
a
l
-
ti
m
e
de
pl
oym
e
nt
w
it
hi
n
e
m
be
dde
d
s
ys
te
m
s
or
e
dge
de
vi
c
e
s
to
r
e
du
c
e
la
te
nc
y
f
or
a
ppl
ic
a
ti
ons
in
pr
a
c
ti
c
e
.
V
e
r
if
yi
ng
th
e
im
pl
e
m
e
nt
a
ti
on
f
e
a
s
ib
il
it
y
of
th
e
pr
opos
e
d
m
od
e
l
in
a
lo
w
-
c
os
t
e
dge
c
om
put
in
g
pl
a
tf
or
m
,
s
uc
h
a
s
th
e
R
a
s
pbe
r
r
y
P
i.
A
s
a
im
e
d
by
[
26]
th
is
c
oul
d
a
c
t
a
s
a
a
ppr
oa
c
h
to
v
a
li
da
te
pr
io
r
w
or
k t
he
e
f
f
e
c
t
of
l
ig
h
twe
ig
ht
a
nd r
e
s
our
c
e
c
ons
e
r
vi
ng m
ode
ls
i
n r
e
s
our
c
e
c
on
s
tr
a
in
e
d e
nvi
r
onm
e
nt
.
5.
C
O
N
C
L
U
S
I
O
N
I
n
th
is
s
tu
dy,
w
e
in
tr
oduc
e
a
de
ha
z
in
g
m
ode
l
b
a
s
e
d
on
G
A
N
s
th
a
t
u
s
e
s
a
U
-
N
e
t
ge
ne
r
a
to
r
w
it
h
s
e
lf
-
a
tt
e
nt
io
n
a
nd
a
m
ul
ti
-
c
om
pone
nt
lo
s
s
to
im
pr
ove
s
tr
uc
tu
r
a
l
a
nd
pe
r
c
e
pt
ua
l
r
e
c
ons
tr
uc
ti
on.
O
n
th
e
r
e
a
l
-
w
or
ld
L
M
H
a
z
e
da
t
a
s
e
t,
th
e
m
ode
l
a
c
hi
e
v
e
s
a
P
S
N
R
of
3
3.42
dB
a
nd
S
S
I
M
of
0.9590,
out
pe
r
f
or
m
in
g
ot
he
r
a
ppr
oa
c
he
s
bot
h
qua
nt
it
a
ti
ve
ly
a
nd
qua
li
ta
ti
ve
ly
.
T
he
s
e
r
e
s
ul
ts
s
ugge
s
t
th
a
t
s
e
lf
-
a
tt
e
nt
io
n
c
om
bi
ne
d
w
it
h
a
ba
la
nc
e
d
lo
s
s
f
or
m
ul
a
ti
on
le
a
ds
to
s
upe
r
io
r
ha
z
e
r
e
m
ova
l,
e
s
pe
c
ia
ll
y
in
c
om
pl
e
x
te
xt
ur
a
l
s
c
e
ne
s
of
va
r
yi
ng
ha
z
e
de
ns
it
y.
T
he
pr
opos
e
d
s
ys
t
e
m
pr
ovi
de
s
a
r
obus
t
a
nd
e
f
f
ic
ie
nt
s
ol
ut
io
n
f
or
r
e
a
l
-
w
or
ld
de
ha
z
in
g
a
ppl
ic
a
ti
ons
, e
s
p
e
c
ia
ll
y i
n s
a
f
e
ty
-
c
r
it
ic
a
l
dom
a
in
s
w
h
e
r
e
vi
s
ua
l
c
la
r
it
y i
s
e
s
s
e
nt
ia
l.
A
C
K
N
O
WL
E
D
G
M
E
N
T
S
T
he
a
ut
hor
s
a
r
e
gr
a
te
f
ul
f
or
t
he
i
ns
ti
tu
ti
ona
l
r
e
s
our
c
e
s
t
ha
t
f
a
c
il
i
ta
te
d t
hi
s
r
e
s
e
a
r
c
h
.
F
U
N
D
I
N
G
I
N
F
O
R
M
A
T
I
O
N
A
ut
hor
s
s
ta
te
no f
undi
ng i
nvol
ve
d.
A
U
T
H
O
R
C
O
N
T
R
I
B
U
T
I
O
N
S
S
T
A
T
E
M
E
N
T
T
hi
s
jo
ur
na
l
us
e
s
th
e
C
ont
r
ib
ut
or
R
ol
e
s
T
a
xonomy
(
C
R
e
di
T
)
to
r
e
c
ogni
z
e
in
di
vi
dua
l
a
ut
hor
c
ont
r
ib
ut
io
ns
, r
e
duc
e
a
ut
hor
s
hi
p di
s
put
e
s
,
a
nd f
a
c
il
it
a
te
c
ol
la
bo
r
a
ti
on.
N
am
e
o
f
A
u
t
h
or
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
A
li
A
bdul
a
z
e
e
z
M
oha
m
m
e
d B
a
q
e
r
Q
a
z
z
a
z
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
H
a
yf
a
a
T
. H
u
s
s
e
in
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
S
hr
oouq J
. A
l
-
ja
na
bi
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Y
ou
s
i
f
S
a
m
e
r
M
u
dh
a
f
a
r
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
C
:
C
onc
e
pt
ua
l
i
z
a
t
i
on
M
:
M
e
t
hodol
ogy
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
da
t
i
on
Fo
:
Fo
r
m
a
l
a
na
l
ys
i
s
I
:
I
nve
s
t
i
ga
t
i
on
R
:
R
e
s
our
c
e
s
D
:
D
a
t
a
C
ur
a
t
i
on
O
:
W
r
i
t
i
ng
-
O
r
i
gi
na
l
D
r
a
f
t
E
:
W
r
i
t
i
ng
-
R
e
vi
e
w
&
E
di
t
i
ng
Vi
:
Vi
s
ua
l
i
z
a
t
i
on
Su
:
Su
pe
r
vi
s
i
on
P
:
P
r
oj
e
c
t
a
dm
i
ni
s
t
r
a
t
i
on
Fu
:
Fu
ndi
ng a
c
qui
s
i
t
i
on
C
O
N
F
L
I
C
T
O
F
I
N
T
E
R
E
S
T
S
T
A
T
E
M
E
N
T
A
ut
hor
s
s
ta
te
no c
onf
li
c
t
of
i
nt
e
r
e
s
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8814
I
nt
J
A
dv A
ppl
S
c
i
,
V
ol
.
14
, N
o.
4
,
D
e
c
e
m
be
r
20
25
:
1340
-
1349
1348
D
A
T
A
A
V
A
I
L
A
B
I
L
I
T
Y
D
a
ta
a
va
il
a
bi
li
ty
is
not
a
ppl
ic
a
bl
e
to
th
is
pa
pe
r
a
s
no
ne
w
da
ta
w
e
r
e
c
r
e
a
te
d
or
a
na
ly
z
e
d
in
th
is
s
tu
dy.
R
E
F
E
R
E
N
C
E
S
[
1]
C
.
O
.
A
nc
ut
i
e
t
al
.
,
“
N
T
I
R
E
2024
de
ns
e
a
nd
non
-
hom
oge
ne
ou
s
de
ha
z
i
ng
c
ha
l
l
e
nge
r
e
por
t
,”
i
n
2024
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
W
or
k
s
hops
(
C
V
P
R
W
)
,
I
E
E
E
,
J
un.
2024,
pp.
6453
–
6468
,
doi
:
10.1109/
C
V
P
R
W
63382.2024.00646.
[
2]
G
.
Y
.
L
e
e
,
J
.
C
he
n,
T
.
D
a
m
,
M
.
M
.
F
e
r
da
us
,
D
.
P
.
P
oe
na
r
,
a
nd
V
.
N
.
D
uong
,
“
D
e
ha
z
i
ng
r
e
m
ot
e
s
e
ns
i
ng
a
nd
U
A
V
i
m
a
ge
r
y:
a
r
e
vi
e
w
of
de
e
p l
e
a
r
ni
ng, pr
i
or
-
ba
s
e
d, a
nd hybr
i
d a
ppr
oa
c
he
s
,”
ar
X
i
v
:
2405.07520
, M
a
y 2024
.
[
3]
H
.
H
.
I
.
A
l
hus
s
e
i
n
a
nd
A
.
A
.
M
.
Q
a
z
z
a
z
,
“
L
i
c
e
ns
e
pl
a
t
e
de
t
e
c
t
i
on
a
nd
r
e
c
ogni
t
i
on
us
i
ng
f
a
s
t
e
r
R
C
N
N
,”
i
n
C
y
be
r
I
nt
e
l
l
i
ge
nc
e
and
I
nf
or
m
at
i
on R
e
t
r
i
e
v
al
, S
pr
i
nge
r
, S
i
nga
por
e
, 2025, pp. 173
–
186
, doi
:
10.1007/
9
78
-
981
-
97
-
7603
-
0_17.
[
4]
A
.
A
.
M
.
B
.
Q
a
z
z
a
z
a
nd
Y
.
S
.
M
udha
f
a
r
,
“
G
e
ne
r
a
t
i
ng
de
t
e
c
t
i
on
l
a
be
l
s
f
r
om
c
l
a
s
s
-
l
e
ve
l
e
xpl
a
na
t
i
ons
f
or
de
e
p
l
e
a
r
ni
ng
-
ba
s
e
d
e
ye
di
s
e
a
s
e
di
a
gnos
i
s
,
”
J
our
nal
of
I
nnov
at
i
v
e
I
m
age
P
r
oc
e
s
s
i
ng
,
vol
.
7
,
no.
4,
pp.
1229
–
1246,
D
e
c
.
2025,
doi
:
10.36548/
j
i
i
p.2025.4.008.
[
5]
A
.
A
.
M
.
B
.
Q
a
z
z
a
z
a
nd
N
.
E
.
K
a
dhi
m
,
“
W
a
t
e
r
m
a
r
k
ba
s
e
d
on
s
i
ngul
a
r
v
a
l
ue
d
e
c
om
pos
i
t
i
on
,”
B
aghdad
Sc
i
e
nc
e
J
ou
r
nal
,
vol
.
20
,
no. 5, F
e
b. 2023, doi
:
10.21123/
bs
j
.2023.7168.
[
6]
P
.
C
hi
t
hr
a
,
A
.
R
.
K
a
r
t
hi
ka
,
S
.
E
.
G
.
Q
ue
e
n
,
a
nd
D
.
R
a
m
a
l
i
nga
m
,
“
H
i
gh
s
pe
e
d
i
m
a
ge
de
ha
z
i
ng
m
e
t
hod
ba
s
e
d
on
l
i
ne
a
r
t
r
a
ns
f
or
m
a
t
i
on
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
A
dv
anc
e
R
e
s
e
ar
c
h i
n Sc
i
e
n
c
e
and E
ngi
ne
e
r
i
ng
, vol
. 7, no. 2, 2018.
[
7]
I
.
A
da
k,
P
.
N
i
s
ha
d,
a
nd
P
.
Y
a
da
v,
“
D
e
-
s
m
oki
ng/
de
-
ha
z
i
ng
a
l
gor
i
t
hm
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
R
e
s
e
a
r
c
h
P
ubl
i
c
at
i
on
and
R
e
v
i
e
w
s
, vol
. 5, no. 5, pp. 9243
–
9246, 2024.
[
8]
C
.
J
e
ni
s
ha
a
nd
C
.
S
.
J
oi
c
e
,
“
A
na
l
y
s
i
s
of
r
e
c
e
nt
t
r
e
nds
i
n
s
i
ngl
e
i
m
a
ge
de
ha
z
i
ng
t
e
c
hni
que
s
,”
i
n
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
an
d
C
om
m
uni
c
at
i
on T
e
c
hnol
ogi
e
s
, S
of
t
C
om
put
i
ng R
e
s
e
a
r
c
h S
oc
i
e
t
y, 2022, pp. 10
7
–
126
, doi
:
10.52458/
978
-
81
-
955020
-
5
-
9
-
11.
[
9]
T
.
A
.
A
l
-
A
s
a
di
a
nd
A
.
A
.
A
.
M
.
B
a
qe
r
,
“
F
us
i
on
f
or
m
ul
t
i
pl
e
l
i
ght
s
our
c
e
s
i
n
t
e
xt
ur
e
m
a
ppi
ng
obj
e
c
t
,”
J
ou
r
nal
of
T
e
l
e
c
om
m
uni
c
at
i
on, E
l
e
c
t
r
oni
c
and C
om
put
e
r
E
ngi
ne
e
r
i
ng
, vol
. 9, no. 2
–
11, p
p. 7
–
12, 2017.
[
10]
B
.
L
i
,
Y
.
G
ou,
J
.
Z
.
L
i
u,
H
.
Z
hu, J
.
T
. Z
hou,
a
nd
X
.
P
e
ng,
“
Z
e
r
o
-
s
hot
i
m
a
ge
de
ha
z
i
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
. 29, pp. 8457
–
8466, 2020, doi
:
10.1109/
T
I
P
.2020.3016134.
[
11]
X
.
Q
i
n,
Z
.
W
a
ng,
Y
.
B
a
i
,
X
.
X
i
e
,
a
nd
H
.
J
i
a
,
“
F
F
A
-
N
e
t
:
f
e
a
t
ur
e
f
us
i
on
a
t
t
e
nt
i
o
n
ne
t
w
or
k
f
or
s
i
ngl
e
i
m
a
ge
de
ha
z
i
ng
,”
T
he
T
hi
r
t
y
-
F
our
t
h A
A
A
I
C
onf
e
r
e
nc
e
on A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
(
A
A
A
I
-
20)
,
2020, pp.
11908
-
11915
.
[
12]
C
.
O
.
A
nc
ut
i
,
C
.
A
nc
ut
i
,
R
.
T
i
m
of
t
e
,
a
nd
C
.
D
e
V
l
e
e
s
c
houw
e
r
,
“
O
-
H
A
Z
E
:
a
de
ha
z
i
ng
be
nc
hm
a
r
k
w
i
t
h
r
e
a
l
ha
z
y
a
nd
ha
z
e
-
f
r
e
e
out
door
i
m
a
ge
s
,”
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
W
or
k
s
hop
s
(
C
V
P
R
W
)
,
I
E
E
E
,
J
un.
2018, pp. 867
–
8678
, doi
:
10.1109/
C
V
P
R
W
.2018.00119.
[
13]
Y
.
L
i
a
ng,
S
.
A
nw
a
r
,
a
nd
Y
.
L
i
u,
“
D
R
T
:
a
l
i
ght
w
e
i
ght
s
i
ngl
e
i
m
a
ge
de
r
a
i
ni
ng
r
e
c
ur
s
i
ve
t
r
a
ns
f
or
m
e
r
,”
i
n
2022
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
W
or
k
s
hop
s
(
C
V
P
R
W
)
,
I
E
E
E
,
J
un.
2022,
pp.
588
–
597
,
doi
:
10.1109/
C
V
P
R
W
56347.2022.00074.
[
14]
D
.
E
ngi
n,
A
.
G
e
nc
,
a
nd
H
.
K
.
E
ke
ne
l
,
“
C
yc
l
e
-
de
ha
z
e
:
e
nh
a
nc
e
d
C
yc
l
e
G
A
N
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
W
or
k
s
hop
s
(
C
V
P
R
W
)
,
I
E
E
E
,
J
un.
2018,
pp.
938
–
9388
,
doi
:
10.1109/
C
V
P
R
W
.2018.00127.
[
15]
M
.
S
.
I
r
a
j
i
,
J
.
T
a
nha
,
M
.
-
A
.
B
a
l
a
f
a
r
,
a
nd
M
.
-
R
.
F
e
i
z
i
-
D
e
r
a
khs
hi
,
“
A
nove
l
i
nt
e
r
pol
a
t
i
on
c
ons
i
s
t
e
nc
y
f
or
ba
d
ge
ne
r
a
t
i
ve
a
dve
r
s
a
r
i
a
l
ne
t
w
or
ks
(
I
C
-
B
G
A
N
)
,”
M
ul
t
i
m
e
di
a
T
ool
s
and
A
ppl
i
c
at
i
ons
,
vol
.
83,
no.
38,
p
p.
86161
–
86205,
O
c
t
.
2024,
doi
:
10.1007/
s
11042
-
024
-
20333
-
5.
[
16]
M
.
S
.
I
r
a
j
i
,
“
S
e
m
i
-
s
upe
r
vi
s
e
d
ge
ne
r
a
t
i
ve
a
dve
r
s
a
r
i
a
l
ne
t
w
or
ks
f
o
r
i
m
ba
l
a
nc
e
d
s
ki
n
l
e
s
i
on
di
a
gnos
i
s
w
i
t
h
a
n
unbi
a
s
e
d
ge
ne
r
a
t
or
a
nd
i
nf
or
m
a
t
i
ve
i
m
a
ge
s
,”
E
ngi
ne
e
r
i
ng
A
ppl
i
c
at
i
ons
of
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
,
vol
.
159,
N
ov.
2025,
doi
:
10.1016/
j
.e
nga
ppa
i
.2025.111643.
[
17]
P
.
S
a
l
e
hi
,
A
.
C
ha
l
e
c
ha
l
e
,
a
nd
M
.
T
a
ghi
z
a
d
e
h,
“
G
e
n
e
r
a
t
i
ve
a
dve
r
s
a
r
i
a
l
ne
t
w
or
ks
(
G
A
N
s
)
:
a
n
ove
r
vi
e
w
of
t
he
or
e
t
i
c
a
l
m
ode
l
,
e
va
l
ua
t
i
on m
e
t
r
i
c
s
, a
nd r
e
c
e
nt
de
ve
l
opm
e
nt
s
,”
ar
X
i
v
:
2005.13178
, M
a
y 2020
.
[
18]
R
.
T
i
w
a
r
i
,
B
.
G
oya
l
,
a
nd
A
.
D
ogr
a
,
“
D
e
ha
z
i
ng
m
e
c
ha
ni
s
m
us
i
ng
a
ut
o
-
e
nc
ode
r
w
i
t
h
i
nt
e
ns
i
t
y
a
t
t
e
nt
i
on
s
ys
t
e
m
,”
J
our
nal
of
C
om
put
e
r
Sc
i
e
n
c
e
, vol
. 20, no. 12, pp. 1805
–
1817, D
e
c
. 2024, doi
:
10.3844/
j
c
s
s
p.2024.1805.1817.
[
19]
S
.
A
kt
e
r
,
“
G
e
ne
r
a
t
i
ve
A
I
:
a
P
i
x2pi
x
-
GAN
-
ba
s
e
d
m
a
c
hi
ne
l
e
a
r
ni
ng
a
ppr
oa
c
h
f
or
r
obus
t
a
nd
e
f
f
i
c
i
e
nt
l
ung
s
e
gm
e
nt
a
t
i
on
,
”
ar
X
i
v
:
2412.10826
, D
e
c
. 2024
.
[
20]
H
.
Z
hu,
X
.
P
e
ng,
V
.
C
ha
ndr
a
s
e
kha
r
,
L
.
L
i
,
a
nd
J
.
-
H
.
L
i
m
,
“
D
e
ha
z
e
G
A
N
:
w
he
n
i
m
a
ge
de
ha
z
i
ng
m
e
e
t
s
di
f
f
e
r
e
nt
i
a
l
pr
ogr
a
m
m
i
ng,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
27t
h
I
nt
e
r
nat
i
onal
J
oi
nt
C
onf
e
r
e
nc
e
on
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
(
I
J
C
A
I
-
18)
,
A
A
A
I
P
r
e
s
s
,
2018,
pp. 1234
–
1240.
[
21]
H
.
Z
ha
ng
a
nd
V
.
M
.
P
a
t
e
l
,
“
D
e
n
s
e
l
y
c
onn
e
c
t
e
d
pyr
a
m
i
d
de
h
a
z
i
ng
ne
t
w
or
k
,”
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
o
n
and P
at
t
e
r
n R
e
c
ogni
t
i
on
, I
E
E
E
,
J
un. 2018, pp. 3194
–
3203
, doi
:
10.1109/
C
V
P
R
.2018.00337.
[
22]
M
.
F
u,
H
.
L
i
u,
Y
.
Y
u,
J
.
C
he
n,
a
nd
K
.
W
a
ng,
“
D
W
-
G
A
N
:
A
di
s
c
r
e
t
e
w
a
ve
l
e
t
t
r
a
ns
f
or
m
G
A
N
f
or
nonhom
oge
ne
ous
de
ha
z
i
ng
,”
i
n
2021
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
W
or
k
s
hop
s
(
C
V
P
R
W
)
,
I
E
E
E
,
J
un.
2021,
pp.
203
–
212
,
doi
:
10.1109/
C
V
P
R
W
53098.2021.00029.
[
23]
W
.
F
a
ng,
J
.
F
a
n,
Y
.
Z
he
ng,
J
.
W
e
ng,
Y
.
T
a
i
,
a
nd
J
.
L
i
,
“
G
ui
de
d
r
e
a
l
i
m
a
ge
de
ha
z
i
ng
us
i
ng
Y
C
bC
r
c
ol
or
s
pa
c
e
,”
P
r
oc
e
e
di
ngs
of
t
he
A
A
A
I
C
onf
e
r
e
nc
e
on A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
, vol
. 39, no. 3, pp. 2906
–
2914, A
pr
. 2025, doi
:
10.1609/
a
a
a
i
.v39i
3.32297.
[
24]
N
.
F
.
A
.
A
l
ha
de
e
t
hy,
A
.
M
.
Z
e
ki
,
a
nd
A
.
S
ha
h,
“
I
m
a
ge
de
-
ha
z
i
ng
us
i
ng
de
e
p
l
e
a
r
ni
ng
a
ppr
oa
c
h
,”
i
n
4t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on C
om
m
uni
c
at
i
on E
ngi
ne
e
r
i
ng and C
om
put
e
r
Sc
i
e
nc
e
, 2022
, doi
:
10.24086/
c
o
c
os
2022/
pa
pe
r
.577.
[
25]
S
.
H
.
A
bdul
na
bi
,
Y
.
S
.
M
udha
f
a
r
,
A
.
A
.
K
a
dhi
m
,
M
.
B
.
M
a
hdi
,
a
nd
H
.
H
.
S
oj
a
r
,
“
N
e
ur
a
l
ne
t
w
or
k
-
ba
s
e
d
s
ys
t
e
m
i
de
nt
i
f
i
c
a
t
i
on:
a
c
om
pr
e
he
ns
i
ve
F
P
G
A
de
s
i
gn
a
nd
i
m
pl
e
m
e
nt
a
t
i
on
,”
i
n
2024
I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
and
M
e
c
hat
r
oni
c
s
Sy
s
t
e
m
s
(
A
I
M
S)
, I
E
E
E
, F
e
b. 2024, pp. 1
–
7
, doi
:
10.1109
/
A
I
M
S
6
1812.2024.10512531.
[
26]
A
.
M
.
A
.
A
l
-
m
uqa
r
m
,
Y
.
M
udha
f
a
r
,
A
.
M
.
S
ha
ki
r
,
M
.
K
a
z
e
m
,
R
.
A
.
-
Y
a
hi
ya
,
a
nd
B
.
S
.
A
.
Z
a
h
r
a
,
“
L
ow
-
c
os
t
s
m
a
r
t
l
e
a
r
ni
ng
w
i
t
h
m
oodl
e
-
ba
s
e
d
R
a
s
pbe
r
r
y
P
i
4
f
or
uni
ve
r
s
i
t
y
s
t
ude
nt
s
,”
i
n
2023
6t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
E
ngi
ne
e
r
i
ng
T
e
c
hnol
ogy
and
i
t
s
A
ppl
i
c
at
i
ons
(
I
I
C
E
T
A
)
,
I
E
E
E
,
J
ul
. 2023, pp. 603
–
608
, doi
:
10.1109/
I
I
C
E
T
A
57
613.2023.10351266.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
dv A
ppl
S
c
i
I
S
S
N
:
2252
-
8814
G
e
ne
r
at
iv
e
ad
v
e
r
s
ar
ia
l
ne
tw
o
r
k
f
or
i
nt
e
ll
ig
e
nt
ha
z
e
r
e
m
ov
al
…
(
A
li
A
bdu
la
z
e
e
z
M
ohamm
e
d B
aqe
r
Q
a
z
z
a
z
)
1349
B
I
O
G
R
A
P
H
I
E
S
O
F
A
U
T
H
O
R
S
Ali
Abdulazeez
Mohammed
Baqer
Qazzaz
received
the
M.Sc.
and
Ph.D.
degrees
in
Computer
Science
from
Babylon
University,
Iraq,
in
2012
and
2018,
respectively.
He
worked
as
a
Lecturer
at
the
Univ
ersity
of
Kufa,
College
of
Ed
ucation,
Department
of
Computer
Science
.
His
resea
rch
interes
ts
include
image
proces
sing,
computer
vision,
information
security,
deep
learning,
artificial
intell
igence
,
and
da
ta
mining.
He
can
be
contacted
at
email
: alia.qaz
zaz@uokufa.edu.iq
.
Hayfaa
T.
Hussein
received
the
Ph.D.
degr
ee
with
the
Intelligen
t
Sensing
and
Communicat
ions (ISC)
Resear
ch Grou
p, School
of Engine
ering, N
ewcas
tle Unive
rsity, (U
.K.).
M.Sc.
degrees
in
Computer
Science
from
the
University
of
Babylon
,
Iraq.
She
worked
as
a
l
ecturer
at
the
Faculty
of
Educatio
n,
Kufa
Universi
ty.
Her
resea
rch
interest
s
are
facial
expressio
n
recogniti
on,
image
processin
g,
machine
learning,
deep
learning,
and
artificial
intelligence
. She can be contacted a
t
email
: hayfaa
.abogalal@uokufa.edu.iq
.
Shroouq
J.
Al
-
janabi
received
B.Sc.
and
M.Sc.
degrees
in
Co
mputer
Science
from
Babylon
University,
Iraq,
in
1995
and
2000,
respectively.
She
completed
her
Ph.D.
in
Computer
Science
at
the
Informa
tics
Institute
for
Postgradua
te
Stu
dies,
Baghdad,
Iraq,
in
2007.
She
is
a
faculty
member
at
the
University
of
Kufa,
College
of
E
ducation,
Department
of
Computer
Science
.
Her
resea
rch
interes
ts
include
informat
ion
se
curity,
AI,
and
image
processing. She can be contac
ted at
email
: shroouqj.alja
nabi@
uokufa.ed
u.iq
.
Yousif
Samer
Mudhafar
earned
his
B.Sc.
in
Computer
Techniqu
es
Engineering
from
the
Islamic
University
in
Najaf
in
2018.
He
completed
his
M.S
c.
in
Computer
Science
Engineering
at
the
University
of
Debrecen
in
2022,
graduating
with
honors
and
receiving
the
Outstanding
Student
certificate.
He
works
at
the
Univ
ersity
of
Kuf
a
,
Faculty
of
Education,
Department
of
Computer
Science.
His
research
interests
include
c
om
puter
networks,
IoT,
and
AI. He can be contacted at
email
: yousif.mudha
far@iunajaf.edu.iq.
Evaluation Warning : The document was created with Spire.PDF for Python.