I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
E
lec
t
r
ical
an
d
Com
p
u
t
e
r
E
n
gin
e
e
r
in
g
(
I
JE
CE
)
Vol.
1
4
,
No.
5
,
Oc
tober
20
2
4
,
pp
.
5390
~
5399
I
S
S
N:
2088
-
8708
,
DO
I
:
10
.
11591/i
jec
e
.
v
1
4
i
5
.
pp
5
390
-
5399
5390
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
e
c
e
.
iaes
c
or
e
.
c
om
D
e
e
p
n
e
u
r
al
n
e
t
w
o
r
k
s f
o
r
r
e
m
ovi
n
g c
lo
u
d
s an
d
n
e
b
u
la
e
f
r
om
sat
e
ll
ite
i
m
ag
e
s
Nat
alya
Gl
az
yr
in
a
1
,
Raik
h
an
M
u
r
at
k
h
a
n
2
,
S
e
r
i
k
E
s
lyam
ov
3
,
Gu
ld
e
n
M
u
r
z
ab
e
k
ova
4
,
Nurgul
Az
iyeva
5
,
B
ak
h
yt
gu
l
Rys
b
e
k
k
yz
y
6
,
Ain
u
r
Oryn
b
aye
va
7
,
N
az
ira
B
ak
t
iyarova
7
1
D
e
pa
r
tm
e
nt
of
C
omput
e
r
a
nd S
of
twa
r
e
E
ngi
ne
e
r
in
g, F
a
c
ul
ty
of
I
nf
or
ma
ti
on T
e
c
hnol
ogy, L
.N
.
G
umi
ly
ov E
u
r
a
s
ia
n N
a
ti
ona
l
U
ni
ve
r
s
it
y,
A
s
ta
na
, R
e
publ
i
c
of
K
a
z
a
khs
t
a
n
2
D
e
pa
r
tm
e
nt
of
A
ppl
ie
d M
a
th
e
ma
ti
c
s
a
nd I
nf
or
ma
ti
c
a
,
K
a
r
a
ga
n
da
B
uke
to
v U
ni
ve
r
s
it
y, K
a
r
a
ga
nd
a
, R
e
publ
ic
of
K
a
z
a
khs
ta
n
3
D
e
pa
r
tm
e
nt
of
R
a
di
o E
ngi
ne
e
r
in
g, E
le
c
tr
oni
c
s
, a
nd T
e
l
e
c
omm
uni
c
a
ti
ons
, L
.N
. G
umi
ly
ov E
u
r
a
s
ia
n N
a
ti
ona
l
U
ni
ve
r
s
it
y, A
s
ta
n
a
,
R
e
publ
ic
of
K
a
z
a
kh
s
ta
n
4
D
e
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ie
nc
e
s
, S
. S
e
if
ul
li
n K
a
z
a
kh A
gr
ot
e
c
hni
c
a
l
R
e
s
e
a
r
c
h U
ni
ve
r
s
it
y, A
s
ta
na
, R
e
publ
ic
of
K
a
z
a
kh
s
ta
n
5
D
e
pa
r
tm
e
nt
of
I
nf
or
ma
ti
on S
e
c
ur
it
y, F
a
c
ul
ty
of
I
nf
or
ma
ti
on T
e
c
hnol
ogy, L
.N
. G
umi
ly
ov
E
ur
a
s
ia
n N
a
ti
ona
l
U
ni
ve
r
s
it
y, A
s
ta
n
a
,
R
e
publ
ic
of
K
a
z
a
kh
s
ta
n
6
D
e
pa
r
tm
e
nt
of
I
nf
or
ma
ti
on T
e
c
hnol
ogy a
nd S
e
c
ur
it
y, I
nnova
ti
ve
T
e
c
hnol
ogi
e
s
, K
a
r
a
ga
nd
a
T
e
c
hni
c
a
l
U
ni
ve
r
s
it
y na
me
d
a
f
te
r
A
. S
a
gi
nov, Ka
r
a
ga
nda
, R
e
publ
ic
of
K
a
z
a
khs
ta
n
7
D
e
pa
r
tm
e
nt
of
B
io
s
ta
ti
s
ti
c
s
, B
io
in
f
or
ma
ti
c
s
a
nd I
nf
or
ma
ti
on T
e
c
hnol
ogy, As
ta
na
M
e
di
c
a
l
U
ni
ve
r
s
it
y,
A
s
ta
na
,
R
e
publ
ic
of
K
a
z
a
khs
ta
n
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
Apr
30,
2024
R
e
vis
e
d
J
un
8,
2024
Ac
c
e
pted
J
un
16,
2024
T
h
i
s
re
s
ea
r
c
h
p
a
p
e
r
d
e
l
v
e
s
i
n
t
o
c
o
n
t
em
p
o
ra
r
y
me
t
h
o
d
o
l
o
g
i
e
s
f
o
r
er
a
d
i
c
a
t
i
n
g
c
l
o
u
d
s
a
n
d
n
e
b
u
l
ae
f
r
o
m
s
p
a
ce
i
m
a
g
e
s
u
t
i
l
i
z
i
n
g
a
d
v
a
n
c
e
d
d
ee
p
l
e
ar
n
i
n
g
t
e
c
h
n
o
l
o
g
i
e
s
s
u
c
h
a
s
c
o
n
d
i
t
i
o
n
a
l
g
e
n
e
ra
t
i
v
e
a
d
v
e
r
s
a
r
i
a
l
n
e
t
w
o
r
k
s
(
c
o
n
d
i
t
i
o
n
a
l
G
A
N
),
c
y
c
l
i
c
g
e
n
er
a
t
i
v
e
a
d
v
er
s
ar
i
a
l
n
e
t
w
o
r
k
s
(
C
y
c
l
e
G
A
N
),
a
n
d
s
p
ac
e
-
a
t
t
e
n
t
i
o
n
g
e
n
e
ra
t
i
v
e
a
d
v
e
r
s
a
r
i
a
l
n
e
t
w
o
r
k
s
(
s
p
ac
e
-
a
t
t
e
n
t
i
o
n
G
A
N
).
C
l
o
u
d
c
o
v
e
r
p
r
e
s
e
n
t
s
a
s
i
g
n
i
f
i
c
a
n
t
o
b
s
t
a
c
l
e
i
n
re
m
o
t
e
s
e
n
s
i
n
g
,
i
mp
e
d
i
n
g
a
cc
u
ra
t
e
d
a
t
a
a
n
a
l
y
s
i
s
a
cr
o
s
s
v
ar
i
o
u
s
d
o
m
a
i
n
s
i
n
c
l
u
d
i
n
g
e
n
v
i
r
o
n
m
en
t
a
l
m
o
n
i
t
o
r
i
n
g
a
n
d
n
a
t
u
r
a
l
r
e
s
o
u
r
ce
ma
n
a
g
em
e
n
t
.
T
h
e
p
r
o
p
o
s
e
d
t
e
c
h
n
i
q
u
e
s
o
f
f
er
n
o
v
e
l
s
o
l
u
t
i
o
n
s
b
y
l
e
v
e
ra
g
i
n
g
s
p
a
t
i
a
l
a
t
t
e
n
t
i
o
n
me
c
h
a
n
i
s
m
s
t
o
i
d
e
n
t
i
fy
a
n
d
s
u
b
s
e
q
u
e
n
t
l
y
e
l
i
m
i
n
a
t
e
c
l
o
u
d
s
fr
o
m
i
ma
g
e
s
,
t
h
u
s
u
n
c
o
v
er
i
n
g
p
r
e
v
i
o
u
s
l
y
c
o
n
ce
a
l
e
d
i
n
f
o
r
ma
t
i
o
n
a
n
d
e
n
h
a
n
c
i
n
g
t
h
e
q
u
a
l
i
t
y
o
f
s
p
a
ce
d
a
t
a.
T
h
e
s
t
u
d
y
em
p
h
a
s
i
ze
s
t
h
e
n
e
c
e
s
s
i
t
y
f
o
r
f
u
r
t
h
e
r
r
e
s
e
ar
c
h
a
i
m
e
d
a
t
r
ef
i
n
i
n
g
c
l
o
u
d
re
m
o
v
a
l
a
l
g
o
r
i
t
h
m
s
t
o
a
c
c
o
m
m
o
d
a
t
e
d
i
v
e
r
s
e
d
e
t
ec
t
i
o
n
c
o
n
d
i
t
i
o
n
s
a
n
d
e
n
h
a
n
c
i
n
g
t
h
e
o
v
er
a
l
l
e
f
f
i
c
i
e
n
c
y
o
f
d
e
e
p
l
ea
r
n
i
n
g
i
n
s
a
t
e
l
l
i
t
e
i
m
a
g
e
p
r
o
c
e
s
s
i
n
g
.
B
y
h
i
g
h
l
i
g
h
t
i
n
g
p
o
t
e
n
t
i
a
l
b
e
n
e
f
i
t
s
a
n
d
a
d
v
o
ca
t
i
n
g
f
o
r
o
n
g
o
i
n
g
e
x
p
l
o
r
a
t
i
o
n
,
t
h
e
p
a
p
e
r
u
n
d
e
r
s
c
o
r
e
s
t
h
e
i
m
p
o
r
t
a
n
c
e
o
f
a
d
v
a
n
c
i
n
g
c
l
o
u
d
r
em
o
v
a
l
t
e
c
h
n
i
q
u
e
s
t
o
i
m
p
r
o
v
e
d
a
t
a
q
u
a
l
i
t
y
a
n
d
u
n
l
o
c
k
n
e
w
a
p
p
l
i
c
a
t
i
o
n
s
i
n
E
ar
t
h
re
m
o
t
e
s
e
n
s
i
n
g
.
I
n
c
o
n
c
l
u
s
i
o
n
,
t
h
e
p
r
o
p
o
s
e
d
a
p
p
r
o
ac
h
e
s
h
o
l
d
p
r
o
m
i
s
e
i
n
a
d
d
r
e
s
s
i
n
g
t
h
e
p
e
r
s
i
s
t
e
n
t
c
h
a
l
l
e
n
g
e
o
f
c
l
o
u
d
c
o
v
er
i
n
s
p
a
ce
i
m
a
g
er
y
,
p
a
v
i
n
g
t
h
e
w
a
y
f
o
r
m
o
r
e
ac
c
u
ra
t
e
d
a
t
a
a
n
a
l
y
s
i
s
a
n
d
f
u
t
u
r
e
a
d
v
a
n
ce
me
n
t
s
i
n
r
em
o
t
e
s
e
n
s
i
n
g
t
e
c
h
n
o
l
o
g
i
e
s
.
K
e
y
w
o
r
d
s
:
C
loud
r
e
moval
C
on
di
ti
on
a
l
ge
ne
r
a
ti
ve
ne
tw
or
ks
C
yc
li
c
ge
ne
r
a
ti
ve
ne
twor
ks
De
e
p
lea
r
ning
tec
hnologi
e
s
S
pa
c
e
-
a
tt
e
nti
on
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ial
ne
twor
k
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
R
a
ikhan
M
ur
a
tkhan
De
pa
r
tm
e
nt
of
Applied
mathe
matics
a
nd
I
n
f
or
mati
c
a
of
Ka
r
a
ga
nda
B
uke
tov
Unive
r
s
it
y,
Ka
r
a
ga
nda
,
R
e
publi
c
of
Ka
z
a
khs
tan
100000
Ka
r
a
ga
nda
,
R
e
publi
c
of
Ka
z
a
khs
tan
E
mail:
r
a
ykha
n.
mur
a
tkhan@
mail.
r
u
1.
I
NT
RODU
C
T
I
ON
I
n
the
moder
n
e
r
a
of
s
pa
c
e
e
xplor
a
ti
on
a
nd
r
e
mot
e
s
e
ns
ing
[
1]
,
a
dva
nc
e
d
s
a
telli
te
im
a
ge
a
na
lys
is
methods
[
2]
–
[
4
]
f
a
c
e
c
ha
ll
e
nge
s
pr
e
s
e
nted
by
c
lo
ud
c
ove
r
a
ge
,
whic
h
c
a
n
s
igni
f
ica
ntl
y
l
im
it
the
a
v
a
il
a
bil
it
y
a
nd
c
ompl
e
tene
s
s
of
da
ta
[
5]
.
T
his
pr
oblem
lea
d
s
to
dif
f
iculti
e
s
in
a
na
lyzing
a
nd
int
e
r
pr
e
ti
ng
inf
or
mation,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
De
e
p
ne
ur
al
ne
tw
or
k
s
for
r
e
mov
ing
c
louds
and
ne
bulae
fr
om
s
atelli
te
image
s
(
N
atal
y
a
Glaz
y
r
ina
)
5391
ha
ving
a
s
igni
f
ica
nt
im
pa
c
t
on
a
r
e
a
s
s
uc
h
a
s
e
nvir
onmenta
l
moni
tor
ing
[
6
]
,
a
gr
icultu
r
a
l
planning,
a
n
d
na
tur
a
l
r
e
s
our
c
e
mana
ge
ment
[
7]
,
[
8]
.
T
he
a
bil
it
y
to
a
c
c
ur
a
tely
e
xtr
a
c
t
inf
o
r
mation
f
r
om
s
a
telli
te
i
mage
s
[
9]
obs
c
ur
e
d
by
c
louds
a
nd
f
og
be
c
omes
a
ke
y
f
a
c
tor
in
i
mpr
oving
the
qua
li
ty
of
thes
e
da
ta
a
nd,
ther
e
f
or
e
,
be
ing
e
f
f
e
c
ti
ve
ly
us
e
d
in
va
r
ious
a
ppli
c
a
ti
ons
[
10
]
.
I
n
thi
s
wo
r
k,
de
e
p
lea
r
ning
methods
s
uc
h
a
s
c
on
dit
ional
ge
ne
r
a
ti
ve
ne
twor
ks
(
GC
N)
[
11]
–
[
13]
,
c
yc
li
c
ge
ne
r
a
ti
ve
ne
twor
ks
(
C
yc
le
GA
N)
[
14]
–
[
16]
,
a
n
d
s
pa
c
e
-
a
tt
e
nti
on
GA
N
(
S
pA
GA
N)
[
17]
r
e
pr
e
s
e
nt
r
e
volut
i
ona
r
y
s
olut
ions
to
ove
r
c
ome
thes
e
li
mi
tations
.
T
he
pur
pos
e
of
thi
s
r
e
s
e
a
r
c
h
wor
k
is
to
r
e
view
a
nd
c
ompar
e
a
ppr
oa
c
he
s
ba
s
e
d
on
c
GA
N,
c
yc
li
c
GA
N,
a
nd
S
pA
GA
N
in
r
e
movi
ng
c
louds
a
nd
ne
bulae
f
r
om
s
pa
c
e
im
a
ge
s
.
W
e
will
dis
c
us
s
their
a
pp
li
c
a
ti
on,
e
f
f
e
c
ti
ve
ne
s
s
,
a
nd
pr
os
pe
c
ts
f
or
r
e
c
ove
r
ing
hidden
inf
or
mation
,
whic
h
c
a
n
lea
d
to
a
s
igni
f
ica
nt
im
p
r
ove
ment
in
da
ta
qua
li
ty
a
nd
e
xpa
nde
d
pos
s
ibi
li
ti
e
s
f
or
s
c
ientif
ic
a
nd
pr
a
c
ti
c
a
l
a
ppli
c
a
ti
ons
in
the
f
ield
of
e
a
r
t
h
r
e
mot
e
s
e
ns
ing
[
18]
.
T
e
c
hnologi
c
a
l
pr
og
r
e
s
s
in
the
f
ield
of
de
e
p
lea
r
ning
[
19
]
,
r
e
pr
e
s
e
nted
by
c
GA
N,
C
yc
le
GA
N,
a
nd
S
pA
GA
N,
ope
ns
ne
w
pe
r
s
pe
c
ti
ve
s
in
s
p
a
c
e
im
a
ge
pr
oc
e
s
s
ing
[
20]
.
Us
ing
a
r
ti
f
icia
l
int
e
ll
ig
e
nc
e
a
nd
ge
ne
r
a
ti
ve
a
lgor
it
hms
,
thes
e
methods
c
a
nnot
only
de
tec
t
c
loud
c
ove
r
ings
but
a
ls
o
a
c
ti
ve
ly
e
li
mi
na
t
e
them,
r
e
s
tor
ing
hidden
inf
o
r
mation
unde
r
ne
a
th
them.
T
he
unique
c
ha
r
a
c
ter
is
ti
c
s
of
e
a
c
h
o
f
the
p
r
e
s
e
nted
a
ppr
oa
c
he
s
c
r
e
a
te
the
oppor
tuni
ty
to
s
e
lec
t
the
be
s
t
tool
de
pe
nding
on
the
tas
k's
s
pe
c
if
ic
r
e
quir
e
ment
s
a
nd
the
da
ta's
c
ha
r
a
c
ter
is
ti
c
s
.
How
e
ve
r
,
thi
s
pa
pe
r
a
ls
o
c
a
ll
s
f
or
f
u
r
ther
r
e
s
e
a
r
c
h
a
nd
im
pr
ove
ment
of
c
loud
r
e
moval
tec
hniques
us
ing
de
e
p
lea
r
ning.
T
he
de
ve
lopm
e
nt
of
a
lgo
r
it
hms
a
da
pted
to
di
f
f
e
r
e
nt
c
loud
types
a
n
d
s
ur
ve
y
c
ondit
ions
[
21]
a
nd
im
pr
oving
the
ove
r
a
ll
pe
r
f
or
manc
e
of
ne
twor
ks
be
c
ome
im
por
tant
di
r
e
c
ti
ons
f
or
f
utur
e
r
e
s
e
a
r
c
h.
E
f
f
icie
nt
a
nd
a
c
c
ur
a
te
methods
f
o
r
r
e
mo
ving
c
louds
f
r
om
s
a
telli
te
im
a
ge
s
c
a
n
s
igni
f
ica
ntl
y
incr
e
a
s
e
the
a
va
il
a
bil
it
y
a
nd
va
lue
of
thes
e
da
ta,
us
he
r
ing
in
a
ne
w
e
r
a
in
s
c
ientif
ic
r
e
s
e
a
r
c
h
a
nd
pr
a
c
ti
c
a
l
us
e
of
E
a
r
th
r
e
mot
e
s
e
ns
ing.
Z
ha
ng
e
t
al
.
[
22]
dis
c
us
s
the
a
ppli
c
a
ti
on
of
a
r
ti
f
icia
l
int
e
ll
igenc
e
(
AI
)
in
a
s
tr
onomi
c
a
l
da
ta
pr
oc
e
s
s
ing.
T
he
a
uthor
s
highl
ight
that
AI
is
be
c
omi
ng
c
e
ntr
a
l
to
many
a
r
e
a
s
o
f
a
s
tr
onomi
c
a
l
r
e
s
e
a
r
c
h
due
to
a
dva
nc
e
s
in
da
ta
pr
oc
e
s
s
ing.
T
he
a
r
ti
c
le
pr
ovides
a
s
y
s
tema
ti
c
r
e
view
of
r
e
pr
e
s
e
ntative
e
xa
mpl
e
s
of
the
a
ppli
c
a
ti
on
of
a
r
ti
f
icia
l
int
e
ll
igenc
e
tec
hnologi
e
s
i
n
a
s
tr
onomi
c
a
l
da
ta
p
r
oc
e
s
s
ing.
I
n
pa
r
ti
c
ular
,
s
he
de
s
c
r
ibes
the
us
e
of
AI
to
identif
y
puls
a
r
c
a
ndidate
s
,
de
tec
t
f
a
s
t
r
a
dio
e
mi
s
s
ions
,
de
tec
t
gr
a
vit
a
ti
ona
l
wa
ve
s
,
c
las
s
if
y
s
pe
c
tr
a
,
a
nd
r
e
duc
e
int
e
r
f
e
r
e
nc
e
f
r
om
r
a
dio
f
r
e
que
nc
y
int
e
r
f
e
r
e
nc
e
.
I
n
a
ddit
ion,
the
a
r
ti
c
le
e
xa
mi
ne
s
pos
s
ibl
e
f
utur
e
a
ppli
c
a
ti
ons
o
f
A
I
in
a
s
tr
onomi
c
a
l
r
e
s
e
a
r
c
h
to
p
r
ovide
pe
r
s
pe
c
ti
ve
s
f
o
r
the
de
ve
lop
ment
of
a
s
tr
onomi
c
a
l
r
e
s
e
a
r
c
h
in
the
e
r
a
of
a
r
ti
f
icia
l
int
e
ll
i
ge
nc
e
.
Kulka
r
ni
a
nd
M
ur
a
la
[
23]
inves
ti
ga
ted
the
pr
oblem
of
a
ir
bor
ne
im
a
ge
qua
li
ty
de
gr
a
da
ti
on
due
to
a
tm
os
phe
r
ic
ha
z
e
,
whic
h
s
igni
f
ica
ntl
y
a
f
f
e
c
ts
the
a
c
c
ur
a
c
y
of
high
-
leve
l
a
ppli
c
a
ti
ons
s
uc
h
a
s
mi
li
tar
y
s
ur
ve
il
lanc
e
a
nd
e
a
r
thquake
a
s
s
e
s
s
ment.
T
he
a
uthor
s
pr
opos
e
a
nove
l
method
us
ing
de
f
or
mable
mu
lt
i
-
he
a
d
a
tt
e
nti
on
with
dis
plac
e
ment
e
xtr
a
c
ti
on
with
a
s
pa
ti
a
ll
y
a
tt
e
nti
ve
a
ppr
oa
c
h
to
im
pr
ove
the
qua
li
ty
o
f
a
e
r
ial
im
a
ge
s
a
f
ter
ha
z
e
r
e
moval.
E
xpe
r
im
e
ntal
r
e
s
ult
s
on
s
ynth
e
ti
c
a
nd
r
e
a
l
da
ta,
a
s
we
ll
a
s
a
n
e
xtens
ive
a
blative
s
tudy,
de
mons
tr
a
te
the
s
upe
r
ior
it
y
of
the
pr
opos
e
d
meth
od
ove
r
e
xis
ti
ng
a
ppr
oa
c
he
s
f
or
ha
z
e
r
e
moval
in
a
e
r
ial
im
a
ge
s
.
C
he
ng
e
t
al.
[
24]
p
r
e
s
e
nt
a
ne
w
s
ys
tem
f
or
i
mage
de
nois
ing
c
a
ll
e
d
nois
e
ba
s
is
ne
twor
k
(
NB
Ne
t)
.
Unlike
pr
e
vious
wor
ks
,
they
p
r
opos
e
a
n
a
ppr
oa
c
h
to
s
olve
thi
s
c
ompl
e
x
pr
oblem
f
r
o
m
a
ne
w
pe
r
s
pe
c
ti
ve
:
nois
e
r
e
duc
ti
on
thr
ough
im
a
ge
-
a
da
pti
ve
pr
ojec
ti
on
.
T
he
ba
s
ic
idea
is
to
tr
a
in
a
ne
twor
k
c
a
pa
ble
of
s
e
pa
r
a
ti
ng
s
ignal
a
nd
nois
e
by
lea
r
ning
a
s
e
t
of
r
e
c
ons
tr
uc
ti
o
n
ba
s
e
s
in
the
f
e
a
tur
e
s
pa
c
e
.
Ne
xt,
r
e
movi
ng
nois
e
f
r
om
the
im
a
ge
is
a
c
hieve
d
by
s
e
lec
ti
ng
the
a
ppr
opr
iate
ba
s
is
of
the
s
ignal
s
ubs
pa
c
e
a
nd
pr
ojec
ti
ng
the
input
da
ta
int
o
thi
s
s
ubs
pa
c
e
.
An
im
por
tant
a
s
pe
c
t
is
the
pr
opos
e
d
s
pa
ti
a
ll
y
s
e
lf
-
a
tt
e
nti
on
(
S
S
A)
modul
e
,
whic
h
e
xpli
c
it
ly
lea
r
ns
to
ge
ne
r
a
te
ba
s
e
s
a
nd
pr
ojec
t
int
o
a
s
ub
s
pa
c
e
.
S
S
A
int
e
gr
a
tes
with
NB
Ne
t,
s
tr
uc
tu
r
e
d
a
s
UN
e
t,
de
s
igned
f
or
e
nd
-
to
-
e
nd
im
a
ge
de
nois
ing.
E
va
luat
ion
r
e
s
ult
s
on
s
tanda
r
d
da
tas
e
ts
including
s
tanda
r
d
im
a
ge
de
nois
ing
da
tas
e
t
(
S
I
DD
)
a
nd
Da
r
ms
tadt
nois
e
d
a
tas
e
t
(
DN
D)
s
how
that
NB
Ne
t
a
c
hieve
s
s
tate
-
of
-
the
-
a
r
t
pe
a
k
s
ignal
-
to
-
nois
e
r
a
ti
o
(
P
S
NR
)
a
nd
s
tr
uc
tur
a
l
s
im
il
a
r
it
y
index
(
S
S
I
M
)
pe
r
f
or
manc
e
a
t
a
s
igni
f
ica
nt
ly
lowe
r
c
omput
a
ti
ona
l
c
os
t.
W
a
ldner
a
nd
Dia
kogiannis
[
25]
pr
opos
e
s
a
method
to
a
utom
a
ti
c
a
ll
y
e
xtr
a
c
t
f
ield
bounda
r
ies
f
r
om
s
a
telli
te
im
a
ge
s
f
or
digi
tal
a
gr
i
c
ult
ur
e
.
T
he
de
e
p
c
onvolut
ional
ne
ur
a
l
ne
two
r
k
r
e
s
idual
UN
e
t
-
a
is
us
e
d
to
s
olve
the
pr
oblem
of
im
a
ge
s
e
gmenta
ti
on,
highl
ight
ing
f
ield
a
r
e
a
s
,
bounda
r
ies
,
a
nd
the
dis
tanc
e
to
the
ne
a
r
e
s
t
bounda
r
y.
T
he
p
r
oblem
is
f
or
mul
a
ted
a
s
mul
ti
-
tas
k
s
e
mantic
s
e
gmenta
ti
on.
T
he
model
is
tr
a
ined
on
a
s
ingl
e
s
e
nti
ne
l
-
2
c
ompos
it
e
im
a
ge
,
pr
ovidi
ng
a
c
c
ur
a
te
mapping
of
f
ield
bounda
r
ies
a
nd
r
e
c
ognit
ion
of
ind
ivi
dua
l
f
ields
.
E
xpe
r
i
ments
s
how
that
the
model
s
uc
c
e
s
s
f
ull
y
ge
ne
r
a
li
z
e
s
to
dif
f
e
r
e
nt
r
e
s
olut
ions
,
s
e
ns
or
s
,
a
nd
pe
r
iods
without
ove
r
f
it
ti
ng,
a
nd
a
ve
r
a
ging
pr
e
dictions
a
c
r
os
s
mul
ti
ple
im
a
ge
s
he
lps
r
e
duc
e
tempor
a
l
va
r
iations
in
a
c
c
ur
a
c
y.
T
he
a
ppr
oa
c
h
of
thi
s
wo
r
k
is
e
xpe
c
ted
to
s
im
pli
f
y
the
e
xt
r
a
c
ti
on
of
f
ield
bounda
r
ies
on
a
s
c
a
le.
T
his
s
tudy
e
xplor
e
s
a
dva
nc
e
d
methods
f
or
c
loud
a
nd
ha
z
e
r
e
moval
f
r
o
m
s
pa
c
e
im
a
ge
s
us
ing
de
e
p
lea
r
ning
tec
hnologi
e
s
s
uc
h
a
s
C
GN
,
C
yc
leG
AN
,
a
nd
S
pA
GA
N,
whic
h
a
pply
s
pa
ti
a
l
a
tt
e
nti
on
mec
ha
nis
ms
to
identif
y
a
nd
r
e
move
c
louds
.
E
xpe
r
i
ments
on
the
R
I
C
E
1
a
nd
R
I
C
E
2
da
tas
e
ts
s
how
that
S
pA
GA
N
pe
r
f
or
ms
be
tt
e
r
in
P
S
NR
a
nd
S
S
I
M
metr
ics
,
e
s
pe
c
ially
in
de
ns
e
c
loud
c
ondit
ions
.
T
he
s
tudy
highl
ight
s
the
ne
e
d
to
f
ur
ther
im
pr
ove
c
loud
r
e
moval
a
lgor
it
hms
to
e
nh
a
nc
e
the
qua
li
ty
of
s
pa
c
e
da
ta.
I
n
c
onc
lus
ion,
th
e
r
e
s
ult
s
c
onf
ir
m
that
us
ing
GA
N
f
or
c
loud
r
e
moval
s
igni
f
ica
ntl
y
i
mpr
ove
s
the
qua
li
ty
o
f
g
r
oun
d
objec
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5390
-
5399
5392
r
e
c
ons
tr
uc
ti
on.
T
he
R
I
C
E
1
a
nd
R
I
C
E
2
da
tas
e
ts
a
r
e
va
luable
r
e
s
our
c
e
s
f
or
s
tudyi
ng
a
tm
os
phe
r
ic
phe
nomena
,
c
onf
ir
mi
ng
the
high
pe
r
f
or
manc
e
of
S
pA
GA
N
i
n
va
r
ious
a
tm
os
phe
r
ic
c
ondit
ions
a
nd
pr
ovidi
ng
d
ir
e
c
ti
ons
f
or
f
utur
e
r
e
s
e
a
r
c
h
in
the
f
ield
o
f
c
loud
r
e
moval
us
i
ng
de
e
p
lea
r
ning.
2.
M
E
T
HO
D
S
pA
GA
N
is
us
e
d
to
ge
ne
r
a
te
a
nd
t
r
a
ns
f
or
m
s
pa
ti
a
l
da
ta,
s
uc
h
a
s
high
-
r
e
s
olut
ion
im
a
ge
s
f
r
om
low
-
r
e
s
olut
ion
im
a
ge
s
,
tr
a
ns
late
im
a
ge
s
be
twe
e
n
dif
f
e
r
e
nt
domains
(
f
or
e
xa
mpl
e
,
s
a
telli
te
im
a
ge
r
y
int
o
maps
)
,
a
s
we
ll
a
s
im
pr
oving
the
qua
li
ty
a
nd
diver
s
it
y
of
s
pa
ti
a
l
da
ta,
c
r
e
a
te
de
tailed
maps
f
r
om
c
oa
r
s
e
da
ta,
modeling
ur
ba
n
gr
owth
a
nd
pr
e
dicting
e
nvir
on
menta
l
pa
tt
e
r
ns
.
S
pA
GA
N
is
a
ls
o
us
e
d
in
medic
a
l
im
a
ging
to
im
pr
ove
the
qua
li
ty
of
magne
ti
c
r
e
s
ona
nc
e
im
a
ging
(
M
R
I
)
or
c
omput
e
r
ize
d
tom
ogr
a
phy
(
C
T
)
i
mage
s
,
a
s
we
ll
a
s
ge
ne
r
a
te
r
e
a
li
s
ti
c
medic
a
l
im
a
ge
s
f
or
tr
a
ini
ng
mac
hine
lea
r
ning
models
,
a
s
we
ll
a
s
in
other
a
r
e
a
s
whe
r
e
s
pa
ti
a
l
da
ta
ge
ne
r
a
ti
on
a
nd
tr
a
ns
f
or
mation
a
r
e
r
e
quir
e
d.
Als
o,
S
pA
GA
N
is
a
n
innovative
a
ppr
oa
c
h
in
the
f
ield
of
r
e
mot
e
s
e
ns
ing
im
a
ge
pr
oc
e
s
s
ing
with
a
s
pe
c
ial
f
oc
us
on
the
tas
k
of
c
loud
r
e
moval.
T
his
te
c
hnology
dif
f
e
r
s
f
r
om
tr
a
dit
ional
methods
due
to
i
ts
a
bil
it
y
to
a
na
lyze
a
nd
pr
oc
e
s
s
inf
o
r
mation
in
de
tail
a
t
t
he
pixel
leve
l,
us
ing
s
ophis
ti
c
a
ted
a
tt
e
nti
on
mec
ha
nis
ms
to
identif
y
a
nd
e
li
mi
na
te
c
loud
c
ove
r
without
los
ing
va
luable
inf
or
mation
a
bout
the
unde
r
lyi
ng
s
ur
f
a
c
e
.
T
he
a
r
c
hit
e
c
tur
e
of
S
pA
GA
N
is
ba
s
e
d
on
a
s
tr
uc
tur
e
c
a
ll
e
d
s
pa
ti
a
ll
y
a
tt
e
nti
ve
ne
twor
k
(
S
P
AN
e
t)
,
whic
h
e
f
f
ic
iently
a
na
lyze
s
im
a
ge
s
f
or
c
loud
c
ove
r
a
nd
r
e
moves
them
while
pr
e
s
e
r
ving
im
por
tant
de
tails
a
nd
c
ha
r
a
c
ter
is
ti
c
s
of
the
e
a
r
th's
s
ur
f
a
c
e
.
S
pA
GA
N
a
ls
o
include
s
s
e
ve
r
a
l
ke
y
c
omponents
:
a.
C
onvolut
ional
laye
r
s
us
e
d
to
ini
ti
a
ll
y
e
xtr
a
c
t
f
e
a
tur
e
maps
f
r
om
input
i
mage
s
.
b.
R
e
s
idual
blocks
us
e
d
f
or
de
e
p
pr
oc
e
s
s
ing
of
f
e
a
tur
e
s
without
los
s
of
in
f
or
mation
thr
oughout
the
ne
tw
or
k.
c.
S
pa
ti
a
l
a
tt
e
nti
on
blocks
(
S
AB
s
)
a
r
e
the
he
a
r
t
o
f
S
pA
GA
Ns
,
c
ons
is
ti
ng
of
s
pa
ti
a
l
a
tt
e
nti
ona
l
r
e
s
idual
blocks
(
S
AR
B
s
)
a
nd
s
pa
ti
a
l
a
tt
e
nti
on
modul
e
s
(
S
AM
s
)
that
ope
r
a
te
in
pa
r
a
ll
e
l
.
T
he
s
e
unit
s
c
a
n
dyna
mi
c
a
ll
y
f
oc
us
on
a
r
e
a
s
with
c
louds
,
de
te
r
mi
ne
their
bounda
r
ies
a
nd
c
ha
r
a
c
ter
is
ti
c
s
,
a
nd
then
guide
the
im
a
ge
r
e
s
tor
a
ti
on
pr
oc
e
s
s
in
s
uc
h
a
wa
y
a
s
t
o
e
ns
ur
e
the
mos
t
a
c
c
ur
a
te
c
loud
r
e
moval.
T
he
s
pa
ti
a
l
a
tt
e
nti
on
e
ngine
in
S
pA
GA
N
a
im
s
to
e
mul
a
te
h
uman
vis
ua
l
a
tt
e
nti
on,
whic
h
is
c
a
pa
ble
o
f
r
e
c
ognizing
a
nd
f
oc
us
ing
on
ke
y
e
leme
nts
of
v
is
ua
l
pe
r
c
e
pti
on.
I
n
c
loud
r
e
moval
r
e
s
e
a
r
c
h,
thi
s
mea
ns
that
the
ne
t
wor
k
c
a
n
identif
y
c
loudy
a
r
e
a
s
in
a
n
im
a
ge
a
nd
give
t
he
m
mor
e
e
mphas
is
dur
ing
p
r
oc
e
s
s
ing.
T
his
a
ppr
oa
c
h
a
ll
ows
not
only
to
a
c
c
ur
a
te
r
e
moval
of
c
louds
b
ut
a
ls
o
the
r
e
s
tor
a
ti
on
of
inf
o
r
mation
unde
r
ne
a
th
them
with
a
high
de
gr
e
e
o
f
de
tail,
whic
h
is
e
xt
r
e
mely
i
mpor
tant
f
or
the
a
c
c
ur
a
c
y
a
nd
us
e
f
ulnes
s
of
the
r
e
s
ult
ing
da
ta.
T
he
a
ppli
c
a
ti
on
of
S
pA
GA
N
in
the
f
ield
of
r
e
mot
e
s
e
ns
ing
ope
ns
up
ne
w
pos
s
ibi
li
ti
e
s
f
o
r
the
a
na
lys
is
of
the
E
a
r
th's
s
ur
f
a
c
e
a
nd
a
tm
os
phe
r
e
.
I
mpr
ove
d
c
loud
-
f
r
e
e
im
a
ge
qua
li
ty
he
lps
be
tt
e
r
m
onit
or
e
nvir
onmenta
l
c
ha
nge
,
mana
ge
na
tu
r
a
l
r
e
s
our
c
e
s
,
ur
ba
n
a
nd
a
gr
icultu
r
a
l
planning
,
a
nd
e
nha
nc
e
r
e
s
e
a
r
c
h
c
a
pa
bil
it
ies
in
c
li
mate
,
e
c
ology
,
a
nd
mete
or
o
logy.
F
unda
menta
l
dif
f
e
r
e
nc
e
s
a
nd
innovations
o
f
S
pA
GA
N
a
r
e
:
a.
Unlike
s
tanda
r
d
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ial
ne
twor
ks
,
S
pA
GA
N
int
r
oduc
e
s
a
s
ophis
ti
c
a
ted
s
pa
ti
a
l
a
tt
e
nti
on
s
ys
tem
(
S
P
AN
e
t)
that
a
ll
ows
the
model
to
f
oc
us
on
ke
y
r
e
gions
of
the
im
a
ge
f
or
mor
e
a
c
c
ur
a
te
c
loud
r
e
moval.
T
his
r
e
pr
e
s
e
nts
a
s
igni
f
ica
nt
im
pr
ove
men
t
in
mac
hine
vis
ua
l
pe
r
c
e
pti
on
by
e
mul
a
ti
ng
the
hu
man
a
tt
e
nti
on
mec
ha
nis
m.
b.
De
e
p
pr
oc
e
s
s
ing
us
ing
s
pa
ti
a
l
-
a
tt
e
nti
ona
l
blocks
.
E
a
c
h
s
pa
ti
a
l
a
tt
e
nti
on
block
in
S
P
AN
e
t
c
ontains
a
unique
c
onf
igur
a
ti
on
o
f
s
pa
ti
a
l
a
tt
e
nti
ona
l
r
e
s
idual
blocks
(
S
AR
B
s
)
a
nd
s
pa
ti
a
l
a
tt
e
nti
on
modul
e
s
(
S
AM
s
)
that
pr
ovide
f
ine
-
gr
a
ined
pr
oc
e
s
s
ing
of
e
a
c
h
a
s
pe
c
t
of
the
input
da
ta.
T
his
a
ll
ows
S
pA
GA
N
to
a
da
pt
to
c
ompl
e
x
c
loud
pa
tt
e
r
ns
a
nd
e
f
f
e
c
ti
ve
ly
r
e
c
ons
tr
u
c
t
inf
or
mation
a
bout
the
e
a
r
th's
s
ur
f
a
c
e
a
t
a
lev
e
l
of
de
tail
ne
ve
r
s
e
e
n
be
f
or
e
.
c.
One
of
the
ke
y
a
dva
ntage
s
of
S
pA
GA
N
is
the
a
bil
it
y
of
the
model
to
dyna
mi
c
a
ll
y
a
da
pt
to
dif
f
e
r
e
nt
t
ype
s
a
nd
de
ns
it
ies
of
c
loud
c
ove
r
a
ge
,
whic
h
is
a
c
hieve
d
thr
ough
the
us
e
o
f
a
tt
e
nti
on
maps
.
T
he
s
e
maps
a
ll
ow
the
model
to
de
ter
mi
ne
the
a
mount
of
a
tt
e
nti
on
that
s
hould
be
given
to
e
a
c
h
pixel,
ther
e
by
pr
ovidi
ng
f
lexibil
it
y
a
nd
a
c
c
ur
a
c
y
whe
n
pr
oc
e
s
s
ing
im
a
ge
s
with
va
r
ied
c
loud
c
ove
r
a
ge
.
d.
S
pA
GA
N
us
e
s
a
mul
ti
dim
e
ns
ional
a
ppr
oa
c
h
to
im
a
ge
pr
oc
e
s
s
ing,
whic
h
include
s
not
only
the
r
e
moval
of
c
louds
but
a
ls
o
a
de
e
p
a
na
lys
is
of
the
unde
r
lyi
ng
s
ur
f
a
c
e
s
to
r
e
s
tor
e
de
tailed
inf
o
r
mation,
whic
h
is
made
pos
s
ibl
e
thanks
to
the
c
ompl
e
x
s
tr
uc
tur
e
of
the
n
e
twor
k,
including
many
laye
r
s
a
nd
blocks
with
a
high
de
gr
e
e
of
s
pe
c
ializa
ti
on.
F
igur
e
1
(
a
)
s
hows
the
s
tr
uc
tur
e
of
the
s
pa
ti
a
l
a
tt
e
nt
ional
ne
twor
k
(
S
P
AN
e
t)
a
s
a
whole
.
I
n
pa
r
ti
c
ular
,
F
igur
e
1
(
b)
c
lea
r
ly
de
mons
tr
a
tes
the
s
pa
ti
a
l
a
tt
e
nt
i
on
block
(
S
AB
)
,
whic
h
c
ons
is
ts
of
thr
e
e
s
pa
ti
a
l
a
tt
e
nti
ona
l
r
e
s
idual
blocks
(
S
AR
B
)
blocks
a
nd
one
s
pa
ti
a
l
a
tt
e
nti
ona
l
modul
e
(
S
AM
)
modul
e
.
F
igur
e
1(
c
)
de
tai
ls
one
of
the
S
AR
B
blocks
,
whic
h
is
a
c
ti
ve
ly
a
ppli
e
d
to
e
f
f
e
c
ti
ve
ly
r
e
move
c
louds
us
ing
ne
ga
ti
ve
r
e
s
iduals
g
uided
by
a
n
a
tt
e
nti
on
map.
T
his
block
is
a
n
im
po
r
tant
c
om
pone
nt
in
the
pr
oc
e
s
s
of
da
ta
p
r
oc
e
s
s
ing
a
nd
inf
or
mation
r
e
c
ove
r
y
unde
r
the
c
loud.
Als
o,
F
igu
r
e
1(
d)
pr
e
s
e
nts
the
s
pa
ti
a
l
a
tt
e
nti
on
modul
e
(
S
AM
)
,
whic
h
is
de
s
c
r
ibed
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
De
e
p
ne
ur
al
ne
tw
or
k
s
for
r
e
mov
ing
c
louds
and
ne
bulae
fr
om
s
atelli
te
image
s
(
N
atal
y
a
Glaz
y
r
ina
)
5393
a
s
a
two
-
c
ir
c
uit
,
f
our
-
wa
y
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
tw
or
k.
T
his
modu
le
plays
a
ke
y
r
o
le
in
pr
ovidi
ng
a
c
c
ur
a
te
s
pa
ti
a
l
a
tt
e
nti
on
whe
n
r
e
movi
ng
c
louds
,
whic
h
s
ig
nif
ica
ntl
y
im
pr
ove
s
the
e
f
f
icie
nc
y
o
f
the
a
lgor
it
hm
.
T
his
s
e
t
of
im
a
ge
s
a
nd
blocks
,
including
S
AB
,
S
AR
B
,
a
nd
S
AM
,
r
e
pr
e
s
e
nts
a
n
innovative
a
ppr
oa
c
h
in
the
f
ield
of
s
pa
ti
a
l
a
tt
e
nti
on
a
nd
c
loud
r
e
moval
f
r
om
s
a
telli
te
i
mage
s
.
De
c
oding
of
s
hor
tene
d
wo
r
ds
s
uc
h
a
s
S
AB
,
S
AR
B
,
a
nd
S
AM
indi
c
a
te
the
us
e
of
s
pa
ti
a
l
a
tt
e
nti
ona
l
u
nit
s
,
s
pa
ti
a
l
a
tt
e
nti
ona
l
r
e
s
idual
unit
s
,
a
nd
s
pa
ti
a
l
a
tt
e
nti
on
unit
s
,
r
e
s
pe
c
ti
ve
ly.
T
he
s
e
c
omponents
a
r
e
a
n
int
e
gr
a
l
pa
r
t
of
the
de
ve
loped
s
ys
tem
a
im
e
d
a
t
im
p
r
o
ving
the
qua
li
ty
of
s
pa
c
e
da
ta
p
r
oc
e
s
s
ing
us
ing
de
e
p
lea
r
ning
methods
.
(
a
)
(
b)
(
c
)
(
d)
F
igur
e
1.
T
he
a
r
c
hit
e
c
tur
e
o
f
the
S
pA
GA
N
metho
d:
(
a
)
S
P
AN
e
t,
(
b)
s
pa
ti
a
l
a
tt
e
nti
ve
block
,
(
c
)
S
AR
B
,
a
nd
(
d)
S
AM
c
GA
N
(
c
ondit
ional
GA
N)
C
ondit
ional
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ial
ne
twor
ks
(
c
GA
N)
r
e
pr
e
s
e
nt
a
s
igni
f
ica
nt
de
ve
lopm
e
nt
in
the
f
ield
of
ge
ne
r
a
ti
ve
modeling
.
T
he
s
e
ne
twor
ks
,
pr
op
os
e
d
by
M
ir
z
a
a
nd
Os
inder
o
in
2014
[
26
]
,
e
x
tend
the
c
a
pa
bil
it
ies
of
tr
a
dit
ional
GA
Ns
by
int
r
oduc
ing
a
n
e
leme
nt
of
c
ondit
ionalit
y,
a
ll
owing
da
ta
to
be
g
e
ne
r
a
ted
ba
s
e
d
on
s
pe
c
if
ic
inf
or
mation.
I
ntegr
a
ti
ng
c
ondit
i
ona
li
ty
int
o
c
GA
Ns
im
pr
ove
s
their
a
ppli
c
a
bil
it
y
i
n
va
r
ious
tas
ks
s
uc
h
a
s
im
a
ge
-
to
-
im
a
ge
tr
a
ns
lation,
s
tyl
e
tr
a
ns
f
e
r
,
a
nd
da
ta
s
ynthes
is
.
A
s
tanda
r
d
GA
N
c
ons
is
ts
of
a
ge
ne
r
a
tor
(
G)
a
nd
a
dis
c
r
im
inator
(
D)
pa
r
ti
c
ipat
ing
in
a
mi
ni
max
ga
me,
whe
r
e
the
ge
ne
r
a
tor
s
t
r
ives
to
ge
ne
r
a
te
r
e
a
li
s
ti
c
da
ta
to
f
ool
the
d
is
c
r
im
inator
,
while
the
dis
c
r
im
inator
s
tr
ives
to
c
or
r
e
c
tl
y
dis
ti
nguis
h
be
twe
e
n
r
e
a
l
a
nd
ge
ne
r
a
ted
s
a
mpl
e
s
.
T
he
objec
t
of
the
dis
c
r
im
inato
r
is
to
maximi
z
e
the
pr
oba
bil
it
y
of
c
or
r
e
c
tl
y
c
las
s
if
ying
the
da
ta,
while
the
goa
l
of
the
ge
ne
r
a
tor
is
to
mi
nim
ize
thi
s
pr
oba
bil
it
y.
M
a
thema
ti
c
a
ll
y,
thi
s
a
dve
r
s
a
r
ial
tr
a
ini
ng
is
e
xpr
e
s
s
e
d
by
the
f
oll
owi
ng
mi
nim
a
x
ga
me
(
1)
:
(
,
)
=
~
(
)
[
(
)
]
+
~
(
)
[
(
1
−
(
(
)
)
)
]
(
1)
I
n
p
u
t
C
O
N
V
R
e
L
U
3
x
R
B
3
x
S
A
B
2
x
R
B
C
O
N
V
O
u
t
p
u
t
S
A
M
S
A
R
B
S
A
R
B
S
A
R
B
C
O
N
V
R
e
L
U
C
O
N
V
R
e
L
U
C
O
N
V
R
e
L
U
F
_
1
i
n
p
u
t
f
e
a
t
u
r
e
m
a
p
C
O
N
V
C
O
N
V
R
e
L
U
C
O
N
V
R
e
L
U
F
_
u
F
_
l
F
_
d
F
_
r
W
_
u
W
_
l
W
_
d
W
_
r
x
x
x
x
C
O
N
V
F
_
u
F
_
l
F
_
d
F
_
r
W
_
u
W
_
l
W
_
d
W
_
r
x
x
x
x
C
O
N
V
C
O
N
V
R
e
L
U
C
O
N
V
S
i
g
m
o
i
d
O
u
t
p
u
t
A
t
t
e
n
t
i
o
n
M
a
p
C
O
N
V
R
e
L
U
W
u
,
l
,
d
,
r
(
s
h
a
r
e
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5390
-
5399
5394
whe
r
e
~
(
)
[
(
)
]
is
the
e
xpe
c
ted
va
lue
o
f
the
logar
it
hm
o
f
the
pr
oba
bil
it
y
that
the
dis
c
r
im
inato
r
c
las
s
if
ies
the
r
e
a
l
da
ta
a
nd
~
(
)
r
e
pr
e
s
e
nts
the
dis
tr
ibut
ion
of
the
r
e
a
l
da
ta.
~
(
)
[
(
1
−
(
(
)
)
)
]
is
the
e
xpe
c
ted
va
lue
of
the
logar
i
thm
o
f
the
pr
ob
a
bil
it
y
that
the
dis
c
r
im
inator
c
las
s
if
ies
f
a
ke
da
ta
ge
ne
r
a
ted
by
the
ge
ne
r
a
tor
.
(
)
r
e
pr
e
s
e
nts
the
dis
tr
ibut
ion
of
r
a
ndom
nois
e
,
(
)
is
the
ge
ne
r
a
ted
s
a
mpl
e
.
He
r
e
(
)
r
e
pr
e
s
e
nts
the
dis
tr
ibut
ion
of
r
e
a
l
da
ta,
is
a
r
a
ndom
nois
e
ve
c
tor
s
e
lec
ted
f
r
om
(
)
,
a
nd
(
)
.
I
n
th
is
c
GA
N
wor
k,
the
t
r
a
dit
ional
GA
N
a
r
c
hit
e
c
tur
e
is
e
xtende
d
to
include
c
ondit
io
na
l
inf
or
mation.
T
he
ge
ne
r
a
tor
now
take
s
a
s
input
bot
h
r
a
ndom
nois
e
z
a
nd
c
ondit
ional
inf
or
mation
,
ge
ne
r
a
ti
ng
s
a
mpl
e
s
(
,
)
.
S
im
il
a
r
ly,
the
dis
c
r
im
inator
c
ons
ider
s
both
da
ta
x
a
nd
c
ondit
ional
inf
o
r
mation
y
whe
n
making
c
las
s
if
ica
ti
on.
T
he
los
s
f
unc
ti
on
objec
t
f
or
c
GA
N
is
modi
f
ied
a
c
c
or
dingl
y
(
2)
.
(
,
)
=
~
(
)
[
(
|
)
]
+
~
(
)
[
(
1
−
(
(
|
)
)
)
]
(
2)
I
n
thi
s
f
or
mul
a
ti
on
,
(
|
)
a
nd
(
|
)
indi
c
a
te
that
the
dis
c
r
i
mi
na
tor
a
nd
ge
ne
r
a
tor
a
r
e
c
ondit
ional
on
y
,
a
ll
owing
the
ge
ne
r
a
ti
on
of
da
ta
with
c
e
r
tain
c
ha
r
a
c
ter
is
ti
c
s
.
I
ntr
oduc
ing
c
ondit
ional
inf
o
r
mation
g
iv
e
s
c
GA
N
the
a
bil
it
y
to
c
ontr
o
l
a
nd
manipulate
the
ge
ne
r
a
ted
r
e
s
ult
s
.
F
or
e
xa
mpl
e
,
in
im
a
ge
s
ynthes
is
tas
ks
,
the
c
ondit
ional
inf
or
mation
y
may
r
e
pr
e
s
e
nt
a
c
las
s
labe
l
or
c
e
r
tain
a
tt
r
ibut
e
s
,
in
f
luenc
ing
the
ge
ne
r
a
te
d
im
a
ge
a
c
c
or
dingl
y.
c
GA
Ns
ha
ve
f
ound
wide
a
ppli
c
a
ti
on
in
va
r
ious
f
ields
.
I
n
im
a
ge
-
to
-
im
a
ge
tr
a
ns
lation
tas
ks
,
s
uc
h
a
s
c
onve
r
ti
ng
s
a
telli
te
i
mage
s
int
o
maps
or
c
onve
r
t
ing
blac
k
-
a
nd
-
white
photogr
a
phs
to
c
olo
r
,
c
GA
Ns
e
xc
e
l
in
pr
oduc
ing
r
e
a
li
s
ti
c
a
nd
c
ontextua
ll
y
r
e
leva
nt
r
e
s
ult
s
.
T
he
c
ondit
ional
na
tur
e
of
thes
e
ne
twor
ks
make
s
them
ve
r
s
a
ti
le
tool
s
in
the
f
ields
of
ge
ne
r
a
ti
ve
modeling
a
nd
a
r
ti
f
icia
l
int
e
ll
igenc
e
.
T
he
C
yc
leG
AN
a
r
c
hit
e
c
tur
e
c
ons
is
ts
of
two
(
⇢
)
a
nd
(
⇢
)
a
nd
two
dis
c
r
im
inator
s
(
и
)
.
T
he
f
ir
s
t
ge
ne
r
a
tor
c
onve
r
ts
im
a
ge
s
f
r
om
domai
n
A
to
domain
B
,
a
nd
the
s
e
c
ond
ge
ne
r
a
tor
pe
r
f
or
ms
the
r
e
ve
r
s
e
c
onve
r
s
ion.
Dis
c
r
im
inator
s
e
va
luate
how
r
e
a
l
the
ge
ne
r
a
ted
im
a
ge
s
,
the
or
igi
na
l
im
a
ge
s
,
a
nd
their
looped
ve
r
s
ions
look
.
C
yc
leG
AN
los
s
f
unc
ti
ons
i
nc
lude
los
s
e
s
f
or
dis
c
r
im
inator
s
a
nd
ge
ne
r
a
tor
s
.
T
he
main
c
omponents
of
los
s
f
unc
ti
ons
include
(
3
)
-
(
5)
.
(
⇢
,
,
,
)
=
~
(
)
[
log
(
)
]
+
~
(
)
[
log
(
1
−
(
⇢
(
)
)
)
]
(
3)
whe
r
e
(
)
a
nd
(
)
a
r
e
the
dis
tr
ibut
ions
of
im
a
ge
s
in
dom
a
ins
A
a
nd
B
,
r
e
s
pe
c
ti
ve
ly.
(
⇢
,
⇢
,
,
)
=
~
(
)
[
|
|
⇢
(
⇢
(
)
)
−
|
|
1
]
+
~
(
)
[
|
|
G
⇢
(
G
⇢
(
b
)
)
−
b
|
|
1
]
(
4)
whe
r
e
|
|
∙
|
|
1
–
no
r
mal
L
1.
(
⇢
,
⇢
)
=
~
(
)
[
|
|
⇢
−
|
|
1
]
+
~
(
)
[
|
|
G
⇢
(
b
)
−
b
|
|
1
]
(
5)
C
yc
leG
AN
’
s
it
e
r
a
ti
ve
a
ppr
oa
c
h
f
a
c
il
it
a
tes
e
f
f
ici
e
nt
lea
r
ning
without
the
ne
e
d
f
or
e
xpli
c
it
pa
ir
wis
e
da
ta
be
twe
e
n
dif
f
e
r
e
nt
domains
.
T
his
model
is
e
xtens
ively
a
ppli
e
d
in
tas
ks
s
uc
h
a
s
s
tyl
e
tr
a
ns
f
e
r
a
nd
im
a
ge
tr
a
ns
f
or
mation,
de
mons
tr
a
ti
ng
it
s
ve
r
s
a
ti
li
ty.
T
he
s
tudy
e
mphas
ize
s
the
ne
c
e
s
s
it
y
f
or
c
onti
nue
d
opti
mi
z
a
ti
on
of
models
a
nd
e
xpe
r
im
e
nts
to
e
nha
nc
e
ove
r
a
ll
pe
r
f
or
manc
e
,
pa
r
ti
c
ular
ly
in
c
ha
ll
e
nging
c
ondit
ions
li
ke
high
c
loud
c
ove
r
.
c
GA
N,
S
pA
GA
N
,
a
nd
C
yc
leG
AN
dif
f
e
r
in
their
a
ppr
oa
c
he
s
to
f
r
a
me
-
s
pa
nning
c
loud
r
e
moval.
c
GA
N
us
e
s
c
ondit
ional
da
ta
to
ge
ne
r
a
te
im
a
ge
s
that
matc
h
given
c
ondit
ions
,
s
uc
h
a
s
c
ur
r
e
nt
we
a
ther
c
ondit
ions
.
S
pA
GA
N
e
mpl
oys
e
xter
na
l
a
tt
e
nti
on
mec
ha
nis
ms
to
be
tt
e
r
identi
f
y
a
nd
e
xploi
t
c
loud
r
e
gions
,
a
c
hieving
high
P
S
NR
a
nd
S
S
I
M
.
C
yc
leG
AN
doe
s
not
r
e
qui
r
e
pa
ir
e
d
im
a
ge
s
a
nd
us
e
s
two
ge
ne
r
a
tor
s
a
nd
two
dis
c
r
im
inator
s
to
tr
a
ns
f
or
m
im
a
ge
s
f
r
om
a
c
loudy
s
tate
to
a
non
-
c
loud
s
tate
a
nd
vice
ve
r
s
a
,
whic
h
i
s
us
e
f
ul
whe
n
ther
e
a
r
e
no
c
lea
r
c
or
r
e
s
ponde
nc
e
s
be
twe
e
n
c
loudy
a
nd
non
-
c
loudy
im
a
ge
s
.
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
T
he
R
I
C
E
da
tas
e
t
us
e
d
in
our
e
xpe
r
im
e
nts
is
pub
li
c
ly
a
va
il
a
ble
a
nd
c
a
n
be
f
ound
a
t
the
f
oll
owing
li
nk
htt
ps
:
//
pan
.
baidu.
c
om/s
/1h6SF
W
Snz
H7G
QJ
oM
2Ux
O_ng
.
F
o
r
the
input
da
ta
us
e
d
in
ou
r
models
,
pr
e
pr
oc
e
s
s
ing
include
d
im
a
ge
nor
maliza
ti
on
to
e
n
s
ur
e
c
ons
is
tenc
y
of
pixel
int
e
ns
it
y
va
lues
a
c
r
os
s
da
tas
e
ts
.
S
pe
c
if
ica
ll
y,
e
a
c
h
im
a
ge
wa
s
nor
malize
d
by
s
ubtr
a
c
ti
ng
the
mea
n
a
nd
divi
ding
by
the
s
tanda
r
d
de
v
iation
of
the
pixel
va
lues
c
a
lcula
ted
a
c
r
os
s
the
e
nti
r
e
da
ta
s
e
t.
T
his
s
tep
wa
s
c
r
it
ica
l
to
s
tanda
r
dizing
input
da
ta
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
De
e
p
ne
ur
al
ne
tw
or
k
s
for
r
e
mov
ing
c
louds
and
ne
bulae
fr
om
s
atelli
te
image
s
(
N
atal
y
a
Glaz
y
r
ina
)
5395
im
pr
oving
the
pe
r
f
or
manc
e
o
f
ou
r
de
e
p
-
lea
r
ning
models
by
br
ing
ing
da
ta
to
a
c
omm
on
s
c
a
le.
T
he
models
we
r
e
tr
a
ined
on
a
s
ys
tem
e
quipped
with
a
n
NV
I
DI
A
T
e
s
la
V100
GPU
,
whic
h
s
igni
f
ica
ntl
y
a
c
c
e
ler
a
ted
the
tr
a
ini
ng
pr
oc
e
s
s
.
T
r
a
ini
ng
e
a
c
h
model
took
a
bout
48
hour
s
on
a
ve
r
a
ge
.
T
his
s
e
tup
a
ll
owe
d
us
to
e
f
f
icie
ntl
y
ha
ndle
the
e
xtens
ive
c
omput
a
ti
on
r
e
quir
e
d
f
or
de
e
p
ge
ne
r
a
ti
ve
models
a
nd
mana
ge
the
lar
ge
volum
e
s
of
da
ta
invol
ve
d
in
our
c
loud
a
nd
ne
bula
r
e
moval
tas
ks
.
T
he
da
tas
e
t
is
divi
de
d
int
o
two
s
ubs
e
ts
:
R
I
C
E
1,
whic
h
include
s
thi
n
c
loud
im
a
ge
s
,
a
nd
R
I
C
E
2
,
whic
h
f
oc
us
e
s
on
de
ns
e
c
loud
im
a
ge
s
.
E
a
c
h
s
ubs
e
t
wa
s
de
s
igned
to
a
ddr
e
s
s
di
f
f
e
r
e
nt
c
loud
de
tec
ti
on
a
nd
r
e
moval
c
ha
ll
e
nge
s
,
pr
ovidi
ng
a
c
ompr
e
he
ns
iv
e
r
e
s
our
c
e
f
or
e
va
luating
our
models
unde
r
dif
f
e
r
e
nt
a
tm
os
phe
r
ic
c
ondit
ions
.
T
he
R
I
C
E
da
tas
e
t,
whic
h
c
ons
is
ts
of
two
s
ubs
e
ts
:
R
I
C
E
1
a
nd
R
I
C
E
2,
plays
a
n
im
por
tant
r
ole
in
the
r
e
s
e
a
r
c
h
of
c
loud
r
e
moval
f
r
om
s
a
telli
te
im
a
ge
s
.
B
oth
s
ubs
e
ts
r
e
pr
e
s
e
nt
unique
da
tas
e
ts
c
oll
e
c
ted
us
ing
dif
f
e
r
e
nt
tec
hnologi
e
s
,
de
s
igned
to
s
tudy
a
tm
os
phe
r
ic
phe
nomena
s
uc
h
a
s
c
louds
unde
r
a
va
r
iety
of
li
ght
ing
a
nd
de
ns
it
y
c
ondit
ions
.
R
I
C
E
1
is
a
n
innovative
da
tas
e
t
c
ontaining
500
high
-
r
e
s
olut
ion
im
a
ge
pa
ir
s
a
c
quir
e
d
us
ing
the
Google
E
a
r
th
platf
o
r
m.
E
a
c
h
pa
ir
include
s
a
n
im
a
ge
of
the
s
a
me
a
r
e
a
with
a
nd
without
c
louds
.
T
he
f
oc
us
is
on
im
a
ge
s
of
s
ubtl
e
c
louds
s
uc
h
a
s
ne
bulae
,
making
R
I
C
E
1
a
va
luable
r
e
s
our
c
e
f
or
r
e
s
e
a
r
c
h
r
e
late
d
to
s
ubtl
e
a
tm
os
phe
r
ic
phe
nom
e
na
a
nd
thei
r
i
mpac
t
on
the
vis
ibi
li
ty
of
the
e
a
r
th's
s
ur
f
a
c
e
.
R
I
C
E
2,
in
tur
n
,
is
c
ompi
led
f
r
om
da
ta
f
r
om
L
a
nds
a
t
8
OL
I
/T
I
R
S
s
e
ns
or
s
,
known
f
o
r
their
a
bil
it
y
to
pr
ovide
high
-
qua
li
ty
im
a
ge
s
of
the
E
a
r
th.
A
s
pe
c
ialize
d
f
oc
us
is
on
de
ns
e
c
loud
im
a
ge
r
y,
making
R
I
C
E
2
a
n
im
por
tant
tool
f
or
r
e
s
e
a
r
c
h
int
o
c
louds
a
nd
their
im
pa
c
t
on
c
li
mate
a
nd
we
a
ther
.
E
xpe
r
im
e
nts
c
onduc
ted
with
va
r
ious
de
e
p
lea
r
nin
g
models
on
both
s
ubs
e
ts
r
e
ve
a
led
the
r
e
s
ult
s
that
on
the
R
I
C
E
1
da
ta
s
e
t
with
thi
n
c
louds
,
c
ondit
ional
GA
N
a
c
hieve
d
a
P
S
NR
of
27
.
427
dB
a
nd
S
S
I
M
of
0
.
905,
C
yc
leG
AN
s
howe
d
a
P
S
NR
of
26
.
930
dB
a
nd
S
S
I
M
of
0
.
887,
a
nd
S
pA
GA
N
s
howe
d
the
be
s
t
r
e
s
ult
s
with
a
P
S
NR
of
33
.
232
dB
a
nd
S
S
I
M
0
.
963,
high
li
ghti
ng
it
s
outs
tanding
pe
r
f
or
manc
e
in
thi
n
c
loud
c
o
ndit
ions
.
T
he
s
e
r
e
s
ult
s
highl
ight
the
im
por
tanc
e
of
c
hoos
ing
a
n
a
ppr
opr
iate
model
f
or
a
pa
r
ti
c
ular
c
loud
t
ype
a
nd
s
uppor
t
the
a
ppli
c
a
ti
on
of
S
pA
GA
N
in
thi
n
c
loud
c
ondit
ions
,
whe
r
e
thi
s
model
e
xhibi
ts
the
be
s
t
pe
r
f
or
manc
e
a
s
s
hown
in
T
a
ble
1
.
I
n
the
e
va
luation
of
the
R
I
C
E
2
da
tas
e
t
unde
r
de
ns
e
c
loud
c
ondit
ions
,
c
ondit
ional
GA
N
de
mons
tr
a
ted
a
P
S
NR
of
26
.
354
dB
a
nd
S
S
I
M
of
0.
824
.
C
onve
r
s
e
ly,
C
yc
leG
AN
e
xhibi
ted
s
li
ghtl
y
lowe
r
pe
r
f
or
ma
nc
e
with
a
P
S
NR
of
24.
79
dB
a
nd
S
S
I
M
of
0.
783
.
R
e
mar
ka
bly,
S
pA
GA
N
outper
f
or
med
both
models
,
a
c
hieving
the
highes
t
r
e
s
ult
s
with
a
P
S
NR
o
f
29
.
432
dB
a
nd
S
S
I
M
of
0.
912
,
highl
ight
ing
the
e
f
f
ica
c
y
o
f
it
s
a
tt
e
nti
on
mec
ha
nis
m,
e
s
pe
c
ially
in
c
ha
ll
e
nging
s
c
e
na
r
ios
with
de
ns
e
c
loud
c
ove
r
a
s
s
hown
in
T
a
ble
2
.
T
a
ble
1.
M
ode
l
a
c
c
ur
a
c
y
r
e
s
ult
s
with
thi
n
c
louds
M
ode
l
Q
ua
nt
it
a
ti
ve
me
tr
ic
s
P
N
S
R
S
S
I
M
c
G
A
N
27.427
0.905
C
yc
le
G
A
N
26.930
0.887
S
pA
G
A
N
33.232
0.963
T
a
ble
2.
Ac
c
ur
a
c
y
r
e
s
ult
s
f
or
de
ns
e
c
loud
models
M
ode
l
Q
ua
nt
it
a
ti
ve
me
tr
ic
s
P
N
S
R
S
S
I
M
c
G
A
N
26.354
0.824
C
yc
le
G
A
N
24.79
0.783
S
pA
G
A
N
29.432
0.912
T
he
s
e
r
e
s
ult
s
highl
ight
the
im
po
r
tanc
e
of
c
a
r
e
f
u
l
model
s
e
lec
ti
on
whe
n
pr
oc
e
s
s
ing
im
a
ge
s
with
va
r
ying
de
gr
e
e
s
of
c
loud
c
ove
r
a
nd
di
f
f
iculty
in
c
loud
r
e
moval
.
S
pA
GA
N,
r
e
lyi
ng
on
s
pa
ti
a
l
a
tt
e
nti
on
mec
ha
nis
ms
,
ha
s
de
mons
tr
a
ted
s
upe
r
ior
s
kil
l
in
r
e
c
ove
r
ing
gr
ound
tr
uth
objec
ts
f
r
o
m
both
thi
n
a
n
d
de
ns
e
c
loud
im
a
ge
s
.
T
he
S
pA
GA
N
model
pe
r
f
o
r
ms
be
s
t
in
mos
t
c
ondit
ions
,
howe
ve
r
,
we
e
mpha
s
ize
the
im
por
tanc
e
of
c
a
r
e
f
ul
model
s
e
lec
ti
on
whe
n
pr
oc
e
s
s
ing
im
a
ge
s
with
va
r
ying
de
gr
e
e
s
of
c
loudi
ne
s
s
f
or
the
f
oll
owing
r
e
a
s
ons
:
im
a
ge
s
c
a
n
va
r
y
gr
e
a
tl
y
in
c
loudi
ne
s
s
a
nd
other
we
a
ther
c
ondit
ions
,
a
nd
in
s
om
e
s
pe
c
if
ic
c
ondit
ions
other
models
may
pe
r
f
or
m
be
tt
e
r
;
uni
que
pr
ojec
t
r
e
quir
e
ments
may
r
e
qui
r
e
s
pe
c
ialize
d
models
opti
mi
z
e
d
f
or
s
pe
c
if
ic
c
ondit
ions
;
the
us
e
of
a
va
r
i
e
ty
of
models
im
pr
ove
s
the
r
e
li
a
bil
it
y
a
nd
s
tabili
ty
of
da
ta
a
na
lys
is
;
mul
ti
tas
king
a
nd
the
f
lexibil
it
y
of
s
ome
tas
ks
r
e
quir
e
c
hoos
ing
the
mos
t
s
uit
a
ble
model
f
or
e
a
c
h
s
pe
c
if
ic
s
it
ua
ti
on.
I
n
a
ddit
ion
,
other
models
a
r
e
a
l
s
o
a
ppli
c
a
ble
in
the
tas
k
o
f
r
e
movi
ng
c
louds
a
nd
ne
bulae
f
r
om
i
mage
s
,
s
o
we
c
ompar
e
d
them
with
the
c
om
monl
y
us
e
d
S
pA
GA
N
method.
T
hus
,
while
S
pA
GA
N
did
s
how
e
xc
e
ll
e
nt
r
e
s
ult
s
,
it
is
im
por
tant
to
c
ons
ider
that
dif
f
e
r
e
nt
models
may
be
mor
e
e
f
f
e
c
ti
ve
in
c
e
r
tain
c
ondit
ions
,
e
ns
ur
ing
c
ompr
e
he
ns
ive
a
nd
r
e
li
a
ble
da
ta
a
na
lys
is
a
nd
s
uc
c
e
s
s
f
ul
c
ompl
e
ti
on
of
the
c
ompl
e
x
tas
ks
f
a
c
ing
our
inf
o
r
mation
s
ys
tem.
T
he
s
e
f
indi
ngs
high
li
ght
the
e
f
f
e
c
ti
ve
ne
s
s
of
S
p
A
GA
N
in
a
va
r
iety
of
a
tm
os
phe
r
ic
c
ondit
ions
,
s
uppor
ti
ng
it
s
r
ole
a
s
a
powe
r
f
ul
s
a
telli
te
im
a
ge
pr
oc
e
s
s
ing
tool
in
the
f
ields
of
mete
or
ology
,
c
li
matol
ogy,
a
nd
r
e
mot
e
s
e
ns
ing.
T
he
dif
f
e
r
e
nc
e
s
in
model
pe
r
f
o
r
manc
e
on
the
R
I
C
E
1
a
nd
R
I
C
E
2
tes
t
da
tas
e
ts
c
a
n
be
e
xplaine
d
a
s
f
oll
ows
.
On
the
R
I
C
E
1
da
tas
e
t,
c
o
ndit
ional
GA
N
ge
ne
r
a
tes
im
a
ge
s
while
pr
e
s
e
r
ving
s
pa
ti
a
l
c
onti
nuit
y,
but
the
r
e
s
ult
s
may
ha
ve
f
uz
z
y
a
r
e
a
s
a
nd
gr
ound
de
tail
may
not
be
f
ull
y
r
e
c
ove
r
e
d,
whic
h
e
xplains
the
r
e
latively
low
P
S
NR
a
nd
S
S
I
M
s
c
or
e
s
.
W
hil
e
,
C
yc
le
GA
N
s
howe
d
the
wor
s
t
r
e
s
ult
s
a
mong
the
models
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5390
-
5399
5396
whic
h
may
be
due
to
the
lac
k
of
pa
ir
wis
e
d
is
c
r
im
i
na
ti
ve
inf
or
mation,
whic
h
is
e
s
pe
c
ially
im
po
r
tant
f
or
high
-
r
e
s
olut
ion
r
e
mot
e
s
e
ns
ing
a
nd
c
loud
r
e
moval.
T
hi
s
r
e
s
ult
s
in
a
los
s
of
s
pa
ti
a
l
c
onti
nui
ty
a
nd
a
n
ina
bil
it
y
to
r
e
c
ons
tr
uc
t
de
tails
of
gr
ound
objec
ts
.
Als
o,
S
pA
GA
N
s
hows
the
be
s
t
r
e
s
ult
s
be
c
a
u
s
e
it
e
f
f
e
c
ti
ve
ly
us
e
s
a
tt
e
nti
on
mec
ha
nis
ms
to
de
tec
t
c
loudy
a
r
e
a
s
a
nd
i
mpr
ove
the
c
loud
r
e
moval
pr
oc
e
s
s
,
pr
e
s
e
r
ving
mo
r
e
de
tail
a
nd
c
ons
is
tenc
y
of
the
im
a
ge
,
a
nd
making
it
vis
ua
l
ly
c
los
e
r
to
the
tr
ue
c
loudy
im
a
ge
a
s
s
hown
F
ig
u
r
e
2.
On
the
R
I
C
E
2
c
ondit
ional
da
tas
e
t,
GA
N
f
a
c
e
s
t
he
pr
oblem
o
f
los
ing
inf
o
r
mation
a
bout
gr
ound
objec
ts
unde
r
de
ns
e
c
louds
,
whic
h
make
s
r
e
c
ons
tr
uc
ti
on
dif
f
icult
a
nd
r
e
s
ult
s
in
lowe
r
qua
li
ty
s
c
or
e
s
.
C
yc
leG
AN
r
e
moves
the
white
r
e
gion
of
de
ns
e
c
louds
,
but
due
to
the
c
ompl
e
te
los
s
of
inf
or
mati
on
a
bout
gr
ound
objec
ts
unde
r
de
ns
e
c
louds
,
it
r
e
quir
e
s
tr
a
i
ning
on
a
lar
ge
a
mount
of
s
im
il
a
r
da
ta
to
r
e
c
ons
tr
uc
t
them.
S
pA
GA
N
a
ls
o
pe
r
f
or
ms
be
tt
e
r
in
thi
s
c
a
s
e
,
thank
s
to
it
s
a
bil
i
ty
to
lea
r
n
f
r
om
lar
ge
a
mount
s
of
da
ta
a
nd
the
e
f
f
e
c
ti
ve
us
e
of
the
a
tt
e
nti
on
mec
ha
nis
m
f
or
de
te
c
ti
ng
c
loud
a
r
e
a
s
,
whic
h
c
a
n
s
igni
f
ica
ntl
y
im
pr
o
ve
c
loud
r
e
moval
pe
r
f
o
r
manc
e
,
e
ve
n
whe
n
in
f
or
mation
a
bo
ut
gr
ound
ob
jec
ts
is
c
ompl
e
tely
los
t
a
s
s
hown
F
igu
r
e
3.
F
igur
e
2.
S
pA
GA
N
r
e
s
ult
f
o
r
ne
bula
r
e
moval
on
R
I
C
E
1
da
tas
e
t
F
igur
e
3.
S
pA
GA
N
r
e
s
ult
s
f
or
de
ns
e
c
loud
r
e
mov
a
l
on
the
R
I
C
E
2
da
tas
e
t
S
pA
GA
N
e
xc
e
ll
e
d
in
both
e
va
luations
due
to
it
s
gr
oundbr
e
a
king
inco
r
por
a
ti
on
of
s
pa
ti
a
l
a
t
tention
mec
ha
nis
ms
,
mi
r
r
or
ing
the
human
vis
ua
l
s
ys
tem's
a
bil
it
y
to
r
e
c
ognize
a
nd
c
onc
e
ntr
a
te
on
c
loudy
a
r
e
a
s
.
T
his
innovative
a
ppr
oa
c
h
e
na
bles
the
pr
e
c
is
e
identif
i
c
a
ti
on
of
c
loudy
r
e
gions
a
nd
e
nha
nc
e
s
the
r
e
s
to
r
a
ti
on
o
f
inf
or
mation
r
e
late
d
to
gr
ound
objec
ts
.
C
ons
e
que
n
tl
y,
S
pA
GA
N
p
r
ove
s
e
f
f
e
c
ti
ve
in
ha
ndli
ng
va
r
io
us
c
loud
s
c
e
na
r
ios
,
including
bo
th
thi
n
c
louds
a
nd
de
ns
e
c
loud
s
it
ua
ti
ons
,
lea
ding
to
the
ge
ne
r
a
ti
on
o
f
hig
h
-
qua
li
ty
r
e
c
ons
tr
uc
ted
im
a
ge
s
.
I
n
ou
r
r
e
s
e
a
r
c
h,
we
us
e
d
both
ou
r
de
ve
lopm
e
nts
a
nd
s
tanda
r
d
li
br
a
r
ies
:
GA
N
-
ba
s
e
d
a
lgor
it
hms
,
including
c
GA
N
a
nd
C
yc
leG
AN
,
we
r
e
im
pleme
nted
us
ing
the
popular
P
yT
or
c
h
li
br
a
r
y,
whic
h
made
it
pos
s
ibl
e
to
take
a
dva
ntage
of
it
s
c
onve
nienc
e
f
or
ope
r
a
ti
ons
with
tens
or
s
a
nd
a
utom
a
ti
c
dif
f
e
r
e
nti
a
ti
on,
e
s
pe
c
ially
r
e
ga
r
ding
im
p
leme
ntation
mec
ha
nis
ms
of
c
ondit
ional
ge
ne
r
a
ti
on
a
nd
c
yc
li
c
t
r
a
ns
f
or
mati
ons
.
T
he
de
ve
lopm
e
nt
of
S
pA
GA
N
include
d
the
c
r
e
a
ti
on
of
a
unique
s
pa
ti
a
l
a
tt
e
nti
on
a
r
c
hit
e
c
tur
e
a
nd
the
de
ve
lopm
e
nt
of
c
us
tom
a
tt
e
nti
on
modul
e
s
buil
t
f
r
om
s
c
r
a
tch
by
our
r
e
s
e
a
r
c
h
tea
m,
including
de
tailed
modeling
a
nd
tuni
ng
of
s
pa
ti
a
l
a
tt
e
nti
on
unit
s
f
or
e
f
f
icie
nt
c
loud
c
ove
r
identi
f
ica
ti
on
a
nd
r
e
mo
va
l.
T
his
c
ombi
ne
d
a
ppr
oa
c
h
opti
mi
z
e
d
the
lea
r
ning
pr
oc
e
s
s
a
nd
incr
e
a
s
e
d
the
a
c
c
ur
a
c
y
of
the
r
e
s
ult
s
,
a
nd
we
hope
that
the
inf
or
mation
pr
ovided
will
a
s
s
is
t
r
e
s
e
a
r
c
he
r
s
a
nd
pr
a
c
ti
ti
one
r
s
in
r
e
pr
oduc
ing
o
r
de
ve
lo
ping
the
pr
opos
e
d
methods
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
De
e
p
ne
ur
al
ne
tw
or
k
s
for
r
e
mov
ing
c
louds
and
ne
bulae
fr
om
s
atelli
te
image
s
(
N
atal
y
a
Glaz
y
r
ina
)
5397
Our
e
xpe
r
im
e
nts
s
ometim
e
s
e
nc
ounter
e
d
pr
oblem
s
s
uc
h
a
s
ove
r
f
it
ti
ng,
e
s
pe
c
ially
whe
n
tr
a
ini
ng
on
highl
y
homogene
ous
c
loud
s
tr
uc
tur
e
s
in
R
I
C
E
2.
Additi
ona
ll
y,
while
ou
r
models
we
r
e
s
uc
c
e
s
s
f
ul
a
t
r
e
movi
ng
c
louds
,
they
s
ometim
e
s
a
lt
e
r
e
d
c
loud
-
f
r
e
e
a
r
e
a
s
,
s
ubtl
y
a
f
f
e
c
ti
ng
the
na
tur
a
l
a
ppe
a
r
a
nc
e
o
f
the
lands
c
a
pe
.
R
e
c
ognizing
thes
e
li
mi
tations
is
e
s
s
e
nti
a
l
f
or
f
utu
r
e
im
pr
ove
ments
a
nd
r
e
a
li
s
ti
c
s
a
telli
te
im
a
ge
r
e
c
ons
tr
uc
ti
on.
F
utur
e
r
e
s
e
a
r
c
h
will
f
oc
us
on
a
pplyi
ng
our
c
loud
r
e
moval
tec
hniques
to
othe
r
f
ields
,
s
uc
h
a
s
u
r
ba
n
planning
a
nd
a
gr
icultur
a
l
mon
it
or
ing
,
whe
r
e
a
c
c
ur
a
te
E
a
r
th
obs
e
r
va
ti
on
da
ta
is
c
r
i
ti
c
a
l.
M
or
e
ove
r
,
e
xten
ding
our
methodologi
e
s
to
other
types
of
a
tm
os
phe
r
ic
oc
c
l
us
ions
s
uc
h
a
s
s
moke
a
nd
ha
z
e
c
ould
f
u
r
ther
de
mons
tr
a
te
the
ge
ne
r
a
li
ty
a
nd
i
mpac
t
of
our
wo
r
k
be
yond
the
R
I
C
E
da
tas
e
t.
T
he
r
e
s
ult
s
obtaine
d
in
our
s
tudy
a
r
e
r
e
a
li
s
ti
c
a
n
d
c
or
r
e
c
t.
I
f
pa
r
t
of
the
im
a
ge
is
not
c
ove
r
e
d
by
c
louds
or
ne
bula,
it
s
hould
be
le
f
t
unc
ha
nge
d.
C
ha
nging
only
thos
e
pa
r
ts
of
the
i
mage
that
r
e
quir
e
c
or
r
e
c
ti
on
pr
e
s
e
r
ve
s
the
or
igi
na
l
inf
o
r
mation
a
nd
i
mage
qua
li
ty.
T
his
is
e
s
pe
c
ially
im
por
tant
to
e
ns
ur
e
the
a
c
c
ur
a
c
y
a
nd
r
e
li
a
bil
it
y
of
the
r
e
s
ult
s
s
ince
c
ha
nge
s
in
a
r
e
a
s
that
do
not
r
e
quir
e
c
or
r
e
c
ti
on
c
a
n
lea
d
to
dis
tor
ti
on
a
n
d
los
s
of
im
por
tant
de
tails
.
W
e
a
ls
o
c
ompar
e
d
va
r
ious
models
c
omm
only
us
e
d
f
or
c
loud
a
nd
ne
bula
r
e
moval
tas
ks
,
s
uc
h
a
s
S
pA
GA
N
,
with
other
a
pp
r
oa
c
he
s
to
c
onf
i
r
m
our
hypothes
is
that
c
a
r
e
f
ul
model
s
e
lec
ti
on
is
n
e
c
e
s
s
a
r
y
de
pe
nding
on
the
c
ondit
ions
.
T
hus
,
ou
r
r
e
s
ult
s
highl
ight
the
im
por
tanc
e
of
pr
ope
r
model
s
e
lec
ti
on
f
or
pr
oc
e
s
s
ing
im
a
ge
s
unde
r
dif
f
e
r
e
nt
c
loud
c
ondit
ions
to
e
ns
ur
e
high
a
c
c
ur
a
c
y
a
nd
r
e
li
a
bil
i
ty
of
da
ta
a
na
l
ys
is
.
4.
CONC
L
USI
ON
T
he
s
tudy
de
mons
t
r
a
tes
tha
t
us
ing
GA
N
to
r
e
mo
ve
c
l
ouds
f
r
om
s
a
tel
li
t
e
i
mage
s
c
a
n
s
igni
f
ic
a
ntl
y
im
p
r
ove
the
qu
a
li
ty
of
g
r
ou
nd
objec
t
r
e
c
o
ns
tr
uc
ti
on.
S
pA
G
AN
,
t
ha
nks
to
i
ts
s
pa
t
ial
a
t
tent
ion
mec
ha
nis
ms
,
s
tands
o
ut
a
mon
g
o
the
r
m
ode
ls
,
pr
o
vid
ing
hi
gh
a
c
c
ur
a
c
y
a
nd
de
t
a
il
.
T
he
R
I
C
E
da
tas
e
t,
wh
ich
incl
ude
s
s
ubs
e
ts
R
I
C
E
1
a
nd
R
I
C
E
2
,
is
a
va
luabl
e
r
e
s
ou
r
c
e
f
o
r
s
t
u
dies
on
c
l
oud
r
e
mova
l
f
r
om
s
a
tel
li
te
i
mage
s
.
B
o
th
s
ubs
e
ts
,
c
oll
e
c
t
e
d
us
ing
d
if
f
e
r
e
nt
tec
hnol
ogies
,
pr
ovi
de
u
nique
da
ta
f
o
r
s
tudy
ing
a
t
mos
phe
r
ic
phe
nome
na
unde
r
a
va
r
ie
ty
o
f
l
igh
ti
ng
c
ond
it
i
ons
a
nd
c
lou
d
de
ns
i
ti
e
s
.
R
I
C
E
1
,
w
it
h
it
s
i
nnova
ti
ve
a
pp
r
oa
c
h
f
oc
us
ing
on
t
h
in
c
lo
uds
s
uc
h
a
s
ne
bu
lae
,
p
r
ov
ides
a
v
a
luab
le
r
e
s
e
a
r
c
h
to
ol
f
o
r
a
na
lyz
ing
s
ubt
le
a
tm
os
phe
r
ic
phe
n
omena
a
nd
thei
r
im
pa
c
t
on
the
vis
i
bil
it
y
of
the
e
a
r
th's
s
u
r
f
a
c
e
.
R
I
C
E
2
,
ba
s
e
d
on
L
a
n
ds
a
t
8
OL
I
/T
I
R
S
da
ta
,
w
it
h
a
f
oc
us
o
n
de
ns
e
c
louds
,
pr
ovi
de
s
im
po
r
ta
nt
in
puts
f
or
s
tu
dies
of
c
lo
uds
a
nd
the
ir
i
nf
luenc
e
on
c
li
ma
te
a
nd
we
a
the
r
.
E
xpe
r
im
e
nts
with
dif
f
e
r
e
nt
de
e
p
lea
r
ning
models
o
n
both
s
ubs
e
ts
de
mons
tr
a
ted
that
unde
r
thi
n
c
loud
c
ondit
ions
,
S
pA
GA
N
s
hows
the
be
s
t
pe
r
f
or
ma
nc
e
with
high
P
S
NR
a
nd
S
S
I
M
va
lues
.
E
ve
n
in
mor
e
c
ha
ll
e
nging
e
nvir
onments
with
de
ns
e
c
louds
,
S
pA
GA
N
pr
ove
s
it
s
e
f
f
e
c
ti
ve
ne
s
s
,
de
li
ve
r
ing
ou
ts
tanding
r
e
s
ult
s
.
T
hus
,
the
r
e
s
e
a
r
c
h
r
e
s
ult
s
on
the
R
I
C
E
da
t
a
s
e
t
highl
ight
the
im
por
tanc
e
o
f
model
s
e
lec
ti
on
d
e
pe
nding
on
c
loud
c
ha
r
a
c
ter
is
ti
c
s
,
s
uppor
t
ing
the
a
ppli
c
a
ti
o
n
of
S
pA
GA
N
in
di
f
f
e
r
e
nt
c
loud
s
c
e
na
r
ios
a
nd
c
onf
ir
mi
ng
it
s
high
e
f
f
icie
nc
y
in
s
a
telli
te
im
a
ge
pr
oc
e
s
s
ing
u
nde
r
a
va
r
iety
of
a
tm
os
phe
r
ic
c
ondit
ions
.
T
he
r
e
s
ult
s
pa
ve
the
wa
y
f
or
f
utu
r
e
r
e
s
e
a
r
c
h
in
c
loud
r
e
moval
us
ing
de
e
p
lea
r
ning,
s
ugge
s
ti
ng
the
e
xplor
a
ti
on
of
ne
w
a
r
c
hit
e
c
tur
e
s
a
nd
tec
hniques
to
f
ur
ther
im
p
r
ove
a
c
c
ur
a
c
y
a
nd
e
f
f
icie
nc
y.
RE
F
E
RE
NC
E
S
[
1]
Y
.
Z
ha
ng,
J
. J
ia
ng,
a
nd
G
.
Z
ha
ng,
“
C
ompr
e
s
s
io
n
of
r
e
mot
e
ly
s
e
ns
e
d
a
s
tr
onomi
c
a
l
im
a
ge
us
in
g
w
a
ve
le
t
-
ba
s
e
d
c
ompr
e
s
s
e
d s
e
n
s
in
g
in
de
e
p s
pa
c
e
e
xpl
or
a
ti
on,”
R
e
m
ot
e
Se
ns
in
g
, vol
. 13, no. 2, pp.
1
–
16, J
a
n. 2021, doi:
10.3390/r
s
13020288.
[
2]
G
.
A
.
F
ot
s
o
K
a
mga
,
L
.
B
it
jo
ka
,
T
.
A
k
r
a
m,
A
.
M
e
ngue
M
bom
,
S
.
R
a
me
e
z
N
a
qvi
,
a
nd
Y
.
B
our
oub
i,
“
A
dva
nc
e
me
nt
s
in
s
a
te
ll
it
e
im
a
ge
c
la
s
s
if
ic
a
ti
on :
me
th
odol
ogi
e
s
,
te
c
hni
qu
e
s
,
a
ppr
oa
c
h
e
s
a
nd
a
ppl
ic
a
ti
ons
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
R
e
m
ot
e
Se
ns
in
g
,
vol
.
42,
no. 20, pp. 7662
–
7722, S
e
p. 2021, doi:
10.1080/01431161.
2021.1954261.
[
3]
A
.
A
s
oka
n,
J
.
A
ni
th
a
,
M
.
C
io
ba
nu,
A
.
G
a
bor
,
A
.
N
a
a
ji
,
a
nd
D
.
J
.
H
e
ma
nt
h,
“
I
ma
ge
pr
oc
e
s
s
in
g
te
c
hni
que
s
f
or
a
na
ly
s
is
of
s
a
te
ll
i
te
im
a
ge
s
f
or
hi
s
to
r
ic
a
l
ma
ps
c
la
s
s
if
ic
a
ti
on
—
a
n ove
r
vi
e
w
,”
A
ppl
ie
d Sc
ie
nc
e
s
, vol
. 10, no. 12,
J
un. 2020, doi:
10.3390/app101242
07.
[
4]
G
.
R
a
jm
oha
n,
C
.
V
.
C
hi
nna
ppa
n,
A
.
D
.
J
ohn
W
il
li
a
m,
S
.
C
h
a
ndr
a
kr
is
ha
n
B
a
la
kr
is
hn
a
n,
B
.
A
na
nd
M
ut
hu,
a
nd
G
.
M
a
noga
r
a
n,
“
R
e
va
mpi
ng
la
nd
c
ove
r
a
ge
a
na
ly
s
is
us
in
g
a
e
r
ia
l
s
a
te
ll
it
e
i
ma
ge
ma
ppi
ng,”
T
r
ans
ac
ti
ons
on
E
m
e
r
gi
ng
T
e
le
c
om
m
uni
c
a
ti
ons
T
e
c
hnol
ogi
e
s
, vol
. 32, no. 7, M
a
r
. 2021, doi:
10.1002/ett.3927.
[
5]
K
.
D
a
ur
e
nbe
kov
e
t
al
.
,
“
N
oi
s
y
im
a
ge
e
nha
nc
e
me
nt
s
us
in
g
d
e
e
p
le
a
r
ni
ng
te
c
hni
que
s
,”
I
nt
e
r
nat
io
nal
J
ou
r
nal
of
E
le
c
tr
ic
al
and
C
om
put
e
r
E
ngi
ne
e
r
in
g
, vol
. 14, no. 1, pp. 811
–
818, F
e
b. 2024,
doi
:
10.11591/i
je
c
e
.v14i1.pp811
-
818.
[
6]
M
.
Y
e
s
s
e
no
va
e
t
a
l.
,
“
A
p
pl
ic
a
t
io
n
of
i
nf
o
r
m
a
t
iv
e
t
e
x
tu
r
a
l
L
a
w
’
s
m
a
s
k
s
me
th
od
s
f
or
pr
oc
e
s
s
i
ng
s
p
a
c
e
im
a
g
e
s
,”
I
nt
e
r
n
at
io
nal
J
o
u
r
n
al
o
f
E
l
e
c
t
r
i
c
a
l
an
d
C
om
pu
te
r
E
n
gi
ne
e
r
i
ng
,
vo
l.
1
3,
n
o.
4
,
pp
.
45
57
–
4
56
6,
A
ug
.
20
23
,
do
i:
10
.1
15
91
/i
je
c
e
.
v1
3i
4.
pp
45
57
-
45
66
.
[
7]
M
.
Y
e
s
s
e
nov
a
e
t
al
.
,
“
F
e
a
tu
r
e
s
of
gr
ow
th
of
a
gr
ic
ul
tu
r
a
l
c
r
ops
a
nd
f
a
c
to
r
s
n
e
ga
ti
ve
ly
a
f
f
e
c
ti
ng
th
e
ir
gr
ow
th
,”
I
ndone
s
ia
n
J
ou
r
nal
of
E
le
c
tr
ic
al
E
ngi
ne
e
r
in
g and C
om
put
e
r
S
c
ie
nc
e
, vol
. 30, no. 1,
pp. 625
–
632, Apr
. 2023, doi:
10.11591/i
je
e
c
s
.v30.i1.pp625
-
632.
[
8]
A
.
K
.
R
a
nga
r
a
ja
n,
R
.
P
ur
us
hot
ha
ma
n,
M
.
P
r
a
bha
ka
r
,
a
nd
C
.
S
z
c
z
e
pa
ńs
ki
,
“
C
r
op
id
e
nt
if
ic
a
ti
on
a
nd
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on
u
s
in
g
tr
a
di
ti
ona
l
ma
c
hi
ne
le
a
r
ni
ng
a
nd
de
e
p
le
a
r
ni
ng
a
ppr
oa
c
h
e
s
,
”
J
our
nal
of
E
ngi
ne
e
r
in
g
R
e
s
e
a
r
c
h
(
K
uw
ai
t)
,
vol
.
11,
no.
1,
pp. 228
–
252, Nov. 2023, doi:
10.36909/j
e
r
.11941.
[
9]
J
. T
us
s
upov
e
t
al
.
, “
A
n i
nt
r
oduc
ti
on t
o ma
c
hi
ne
a
nd de
e
p l
e
a
r
ni
ng me
th
ods
f
or
c
lo
ud ma
s
ki
ng a
ppl
ic
a
ti
ons
,”
I
E
E
E
A
c
c
e
s
s
, vol
. 12,
pp. 19902
–
19910, 2024, doi:
10.1109/AC
C
E
S
S
.2024.3361046.
[
10]
N
.
U
z
a
kkyz
y
e
t
al
.
,
“
I
ma
ge
noi
s
e
r
e
duc
ti
on
by
de
e
p
l
e
a
r
ni
ng
me
th
ods
,”
I
nt
e
r
nat
io
nal
J
ou
r
nal
of
E
le
c
tr
ic
al
and
C
om
p
ut
e
r
E
ngi
ne
e
r
in
g
, vol
. 13, no. 6, pp. 6855
–
6861, 2023, doi:
10.11591/i
je
c
e
.v13i6.pp6855
-
6861.
[
11]
X
.
D
in
g,
Y
.
W
a
ng,
Z
.
X
u,
W
.
J
.
W
e
lc
h,
a
nd
Z
.
J
.
W
a
ng,
“
C
c
G
A
N
:
c
ont
in
uous
c
ondi
ti
ona
l
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
ks
f
or
im
a
ge
ge
ne
r
a
ti
on,”
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on L
e
a
r
ni
ng R
e
pr
e
s
e
nt
at
io
ns
,
2021.
[
12]
H
. Y
a
n, Q
. Z
he
ng, a
nd L
.
Z
e
ng, “
C
ondi
ti
ona
l
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
ks
f
or
gr
oundwa
te
r
c
ont
a
mi
na
ti
on c
ha
r
a
c
te
r
iz
a
ti
on
a
n
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5390
-
5399
5398
s
our
c
e
i
de
nt
if
ic
a
ti
on,”
J
our
nal
of
H
y
d
r
ol
ogy
, vol
. 632, M
a
r
. 2024, doi:
10.1016/j
.j
hydr
ol
.2024.130900.
[
13]
Z
.
Z
uo
e
t
al
.
,
“
S
ta
ti
s
ti
c
s
e
nha
nc
e
me
nt
g
e
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
ks
f
or
di
ve
r
s
e
c
ondi
ti
ona
l
im
a
ge
s
ynt
he
s
is
,”
I
E
E
E
T
r
ans
ac
ti
ons
on C
i
r
c
ui
ts
and Sy
s
te
m
s
f
or
V
id
e
o T
e
c
hnol
ogy
, pp
. 1
–
1, 2023, doi:
10.1109/T
C
S
V
T
.2023.3348471.
[
14]
J
.
Y
a
ng
,
K
.
W
a
ng
,
F
.
L
ua
n,
Y
.
Y
i
n,
a
nd
H
.
Z
h
a
n
g,
“
P
r
e
C
a
C
y
c
le
G
A
N
:
p
e
r
c
e
pt
ua
l
c
a
p
s
ul
e
c
y
c
li
c
ge
n
e
r
a
ti
ve
a
d
ve
r
s
a
r
i
a
l
n
e
t
w
or
k
f
or
i
nd
u
s
tr
i
a
l
d
e
f
e
c
ti
ve
s
a
mp
l
e
a
u
gm
e
n
t
a
ti
on
,
”
E
le
c
t
r
on
ic
s
(
Sw
it
z
e
r
l
an
d)
,
vo
l.
1
2,
n
o.
1
6,
A
ug
.
20
23
,
do
i:
10
.3
39
0/
e
l
e
c
t
r
o
ni
c
s
1
21
63
475
.
[
15]
F
.
L
ul
e
c
i
,
F
. N
e
c
a
t
i
C
a
t
ba
s
, a
n
d O
.
A
v
c
i
, “
C
y
c
l
e
G
A
N
f
or
u
nd
a
ma
g
e
d
-
to
-
d
a
m
a
ge
d do
ma
in
tr
a
ns
l
a
ti
on
f
o
r
s
t
r
u
c
t
ur
a
l
h
e
a
lt
h mo
ni
t
or
i
ng
a
n
d
da
m
a
g
e
d
e
t
e
c
ti
on
,
”
M
e
c
h
an
ic
al
S
y
s
t
e
m
s
an
d
Si
g
na
l
P
r
o
c
e
s
s
in
g
,
vol
.
19
7,
A
u
g.
2
02
3,
d
oi
:
1
0.
10
16
/j
.ym
s
s
p.
20
23
.1
10
37
0.
[
16]
X
.
L
iu
,
T
.
Z
ha
ng,
a
nd
M
.
L
iu
,
“
U
D
F
-
G
A
N
:
uns
upe
r
vi
s
e
d
de
ns
e
opt
ic
a
l
-
f
lo
w
e
s
ti
ma
ti
on
us
in
g
c
yc
le
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
ks
,”
K
now
le
dg
e
-
B
as
e
d Sy
s
te
m
s
, vol
. 271, J
ul
. 2023, doi:
10.1016/j
.knos
ys
.2023.110568.
[
17]
Y
.
C
he
n,
T
.
H
umm
e
l,
A
.
S
.
K
oe
pke
,
a
nd
Z
.
A
ka
ta
,
“
W
he
r
e
a
nd
w
he
n:
s
pa
c
e
-
ti
me
a
tt
e
nt
io
n
f
or
a
udi
o
-
vi
s
ua
l
e
xpl
a
na
ti
o
ns
,”
ar
X
iv
:
2105.01517
, M
a
y 2021.
[
18]
A
.
A
nz
a
lo
ne
,
A
.
P
a
gl
ia
r
o,
a
nd
A
.
T
ut
one
,
“
A
n
in
tr
oduc
ti
on
to
ma
c
hi
ne
a
nd
de
e
p
l
e
a
r
ni
ng
me
th
ods
f
or
c
lo
ud
m
a
s
ki
ng
a
ppl
ic
a
ti
ons
,”
A
ppl
ie
d Sc
ie
nc
e
s
, vol
. 14, no. 7, M
a
r
. 2024, doi:
10.3390/a
pp14072887.
[
19]
A
.
N
a
iz
a
ga
r
a
ye
va
e
t
al
.
,
“
D
e
te
c
ti
on
of
he
a
r
t
pa
th
ol
ogy
us
in
g
de
e
p
le
a
r
ni
ng
me
th
ods
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
E
le
c
tr
ic
al
and
C
om
put
e
r
E
ngi
ne
e
r
in
g
, vol
. 13, no. 6, pp. 6673
–
6680, D
e
c
. 202
3, doi:
10.11591/i
je
c
e
.v13i6.pp6673
-
6680.
[
20]
X
. W
a
ng, W
. W
a
ng, Y. C
a
o, C
. S
he
n, a
nd T
. H
ua
ng, “
I
ma
ge
s
s
pe
a
k i
n i
ma
ge
s
:
a
ge
ne
r
a
li
s
t
pa
in
te
r
f
or
i
n
-
c
ont
e
xt
vi
s
ua
l
le
a
r
ni
ng,”
in
P
r
oc
e
e
di
ngs
of
th
e
I
E
E
E
C
om
put
e
r
Soc
ie
ty
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
is
io
n
and
P
at
te
r
n
R
e
c
ogni
ti
on
,
J
un.
2023,
vol
.
20
23
-
J
une
, pp. 6830
–
6839, doi:
10.1109/C
V
P
R
52729.2023.00660.
[
21]
J
. L
iu
, W
. Y
in
, H
. W
a
ng, Y. C
H
e
n, J
.
-
J
. S
onke
, a
nd E
. G
a
vve
s
, “
D
yna
mi
c
pr
ot
ot
ype
a
da
pt
a
ti
on w
it
h di
s
ti
ll
a
ti
on f
or
f
e
w
-
s
hot
poi
nt
c
lo
ud s
e
gme
nt
a
ti
on,”
2024 I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on 3D V
is
io
n (
3D
V
)
,
M
a
r
. 2024, doi:
10.1109/3dv62453.
2024.00045.
[
22]
H
.
Z
ha
ng,
J
.
W
a
ng,
Y
.
Z
ha
ng,
X
.
D
u,
H
.
W
u,
a
nd
T
.
Z
ha
ng,
“
R
e
vi
e
w
of
a
r
ti
f
ic
ia
l
in
te
ll
ig
e
nc
e
a
ppl
ic
a
ti
ons
in
a
s
tr
onomi
c
a
l
da
ta
pr
oc
e
s
s
in
g,”
A
s
tr
ono
m
ic
al
T
e
c
hni
que
s
and I
n
s
tr
um
e
nt
, vol
. 1, n
o. 1, pp. 1
–
15, 2024, doi:
10.61977/ati202400
1.
[
23]
A
.
K
ul
ka
r
ni
a
nd
S
.
M
ur
a
la
,
“
A
e
r
ia
l
im
a
ge
de
ha
z
in
g
w
it
h
a
tt
e
nt
iv
e
de
f
or
ma
bl
e
tr
a
ns
f
or
me
r
s
,”
in
P
r
oc
e
e
di
ngs
-
2023
I
E
E
E
W
i
nt
e
r
C
onf
e
r
e
nc
e
on A
ppl
ic
at
io
ns
of
C
om
put
e
r
V
is
io
n, W
A
C
V
2023
, J
a
n. 2023, pp. 6294
–
6303, doi:
10.1109/WA
C
V
56688.2023.00
624.
[
24]
S
.
C
he
ng,
Y
.
W
a
ng,
H
.
H
ua
ng,
D
.
L
iu
,
H
.
F
a
n,
a
nd
S
.
L
iu
,
“
N
B
N
e
t:
noi
s
e
ba
s
i
s
le
a
r
ni
ng
f
or
im
a
ge
de
noi
s
in
g
w
it
h
s
ubs
pa
c
e
pr
oj
e
c
ti
on,”
in
P
r
oc
e
e
di
ng
s
of
th
e
I
E
E
E
C
om
put
e
r
Soc
ie
t
y
C
o
nf
e
r
e
nc
e
on
C
om
put
e
r
V
is
io
n
and
P
at
te
r
n
R
e
c
ogni
ti
on
,
J
un.
2
021,
pp. 4894
–
4904, doi:
10.1109/C
V
P
R
46437.2021.00486.
[
25]
F
.
W
a
ld
ne
r
a
nd
F
.
I
.
D
ia
kogi
a
nni
s
,
“
D
e
e
p
le
a
r
ni
ng
on
e
dge
:
e
xt
r
a
c
ti
ng
f
ie
ld
bounda
r
ie
s
f
r
om
s
a
te
ll
it
e
im
a
ge
s
w
it
h a
c
onvolut
i
o
na
l
ne
ur
a
l
ne
twor
k,”
R
e
m
ot
e
Se
ns
in
g of
E
nv
ir
onm
e
nt
, vol
. 245, Au
g. 2020, doi:
10.1016/j
.r
s
e
.2020.111741.
[
26]
M
. M
ir
z
a
a
nd S
. O
s
in
d
e
r
o, “
C
ondi
ti
ona
l
ge
ne
r
a
ti
ve
a
dv
e
r
s
a
r
ia
l
N
e
ts
,”
ar
X
iv
:
1411.1784
, N
ov. 2014.
B
I
OG
RA
P
HI
E
S
OF
AU
T
HO
RS
N
a
t
a
l
y
a
Gl
a
zy
r
i
na
g
rad
u
at
e
d
fro
m
t
h
e
In
s
t
i
t
u
t
e
o
f
A
u
t
o
ma
t
i
o
n
,
T
el
ec
o
mmu
n
i
ca
t
i
o
n
s
an
d
In
f
o
rmat
i
o
n
T
ech
n
o
l
o
g
y
o
f
O
ms
k
St
a
t
e
T
ran
s
p
o
rt
U
n
i
v
ers
i
t
y
i
n
2
0
0
3
.
Sh
e
d
efen
d
ed
h
er
d
o
c
t
o
ra
l
t
h
e
s
i
s
o
n
co
m
p
u
t
er
t
e
ch
n
o
l
o
g
y
an
d
s
o
ft
w
are
i
n
2
0
1
5
.
Fro
m
2
0
1
6
t
o
t
h
e
p
res
e
n
t
t
i
me,
s
h
e
i
s
an
a
s
s
o
ci
a
t
e
p
ro
fe
s
s
o
r
o
f
t
h
e
D
e
p
art
me
n
t
o
f
C
o
mp
u
t
er
an
d
So
ft
w
are
E
n
g
i
n
eer
i
n
g
at
L
.
N
.
G
u
mi
l
y
o
v
E
u
ras
i
an
N
a
t
i
o
n
al
U
n
i
v
er
s
i
t
y
.
Sh
e
h
as
au
t
h
o
red
mo
re
t
h
a
n
5
0
p
a
p
ers
.
H
er
res
earc
h
i
n
t
eres
t
s
i
n
cl
u
d
e
mat
h
emat
i
cal
an
d
co
mp
u
t
er
mo
d
el
i
n
g
,
art
i
fi
ci
a
l
i
n
t
el
l
i
g
en
ce,
au
t
o
ma
t
i
o
n
o
f
t
ec
h
n
o
l
o
g
i
ca
l
p
ro
ce
s
s
e
s
.
Sh
e
can
b
e
co
n
t
act
e
d
at
emai
l
:
g
l
az
y
ri
n
a_
n
s
_
1
@
en
u
.
k
z
.
R
a
i
kh
a
n
M
ura
tkh
a
n
i
n
2
0
0
2
h
e
g
ra
d
u
a
t
ed
fro
m
t
h
e
K
arag
a
n
d
a
S
t
at
e
U
n
i
v
ers
i
t
y
n
amed
aft
er
E
.
A
.
Bu
k
et
o
v
w
i
t
h
a
d
e
g
ree
i
n
a
p
p
l
i
e
d
m
at
h
em
a
t
i
c
s
.
In
2
0
1
9
h
e
d
efe
n
d
e
d
h
er
d
i
s
s
er
t
at
i
o
n
i
n
t
h
e
s
p
eci
a
l
t
y
"
6
D
0
6
0
2
0
0
–
co
m
p
u
t
er
s
c
i
e
n
ces
"
an
d
recei
v
ed
a
P
h
.
D
.
Cu
rre
n
t
l
y
,
h
e
i
s
as
s
o
c
i
at
e
p
r
o
fes
s
o
r
a
t
t
h
e
D
e
p
art
me
n
t
o
f
A
p
p
l
i
ed
Mat
h
emat
i
cs
an
d
In
fo
rma
t
i
ca
o
f
K
arag
a
n
d
a
Bu
k
et
o
v
U
n
i
v
er
s
i
t
y
.
H
e
i
s
t
h
e
au
t
h
o
r
o
f
mo
r
e
t
h
an
7
0
s
c
i
e
n
t
i
fi
c
p
a
p
ers
,
i
n
cl
u
d
i
n
g
1
mo
n
o
g
ra
p
h
,
8
ar
t
i
c
l
es
i
n
t
h
e
Sco
p
u
s
d
a
t
ab
a
s
e.
Sci
e
n
t
i
f
i
c
i
n
t
ere
s
t
s
-
i
ma
g
e
p
r
o
ces
s
i
n
g
,
p
at
t
ern
reco
g
n
i
t
i
o
n
t
h
eo
r
y
,
d
at
a
mi
n
i
n
g
,
i
n
f
o
rmat
i
o
n
s
ecu
r
i
t
y
.
H
e
can
b
e
co
n
t
act
e
d
at
emai
l
:
ray
k
h
an
.
mu
ra
t
k
h
an
@
mai
l
.
ru
.
Seri
k
Es
l
y
a
m
o
v
g
rad
u
at
e
d
fro
m
K
ara
g
an
d
a
St
a
t
e
U
n
i
v
er
s
i
t
y
n
amed
af
t
er
E
.
Bu
k
e
t
o
v
i
n
1
9
8
7
w
i
t
h
a
d
e
g
ree
i
n
p
h
y
s
i
cs
.
I
n
1
9
9
3
,
h
e
d
efen
d
e
d
h
i
s
P
h
D
t
h
es
i
s
i
n
t
h
e
s
p
eci
a
l
t
y
"
0
5
.
2
5
.
0
5
-
In
fo
rmat
i
o
n
s
y
s
t
em
s
an
d
p
ro
ce
s
s
e
s
"
at
t
h
e
V
.
G
l
u
s
h
k
o
v
In
s
t
i
t
u
t
e
o
f
Cy
b
ern
e
t
i
cs
(K
i
e
v
)
an
d
rece
i
v
e
d
t
h
e
d
e
g
ree
o
f
Ca
n
d
i
d
a
t
e
o
f
T
ec
h
n
i
cal
Sci
e
n
ces
.
H
e
b
e
g
an
h
i
s
career
i
n
1
9
8
7
a
s
a
l
ect
u
rer
at
t
h
e
D
e
p
art
me
n
t
o
f
C
o
mp
u
t
er
Sc
i
en
ce
an
d
C
o
mp
u
t
er
Sy
s
t
em
s
o
f
K
o
s
t
a
n
a
y
St
at
e
U
n
i
v
er
s
i
t
y
.
C
u
rren
t
l
y
,
h
e
i
s
a
n
ac
t
i
n
g
p
ro
fe
s
s
o
r
at
t
h
e
D
ep
ar
t
men
t
o
f
Rad
i
o
E
n
g
i
n
eeri
n
g
,
E
l
ec
t
ro
n
i
c
s
an
d
T
e
l
eco
mm
u
n
i
cat
i
o
n
s
o
f
t
h
e
L
.
N
.
G
u
mi
l
y
o
v
E
u
ra
s
i
a
n
N
at
i
o
n
al
U
n
i
v
ers
i
t
y
,
a
non
-
p
r
o
fi
t
J
o
i
n
t
-
St
o
ck
Co
m
p
an
y
.
H
e
i
s
t
h
e
a
u
t
h
o
r
o
f
mo
re
t
h
a
n
1
2
0
s
c
i
en
t
i
f
i
c
p
a
p
ers
,
i
n
c
l
u
d
i
n
g
3
mo
n
o
g
rap
h
s
,
2
art
i
cl
e
s
i
n
t
h
e
Sco
p
u
s
d
at
a
b
as
e.
H
i
s
r
es
earc
h
i
n
t
eres
t
s
are
i
n
f
o
rmat
i
o
n
s
y
s
t
ems
,
art
i
fi
c
i
al
i
n
t
el
l
i
g
en
ce,
r
o
b
o
t
i
cs
,
n
eu
ral
n
et
w
o
r
k
s
,
mach
i
n
e
l
earn
i
n
g
.
H
e
can
b
e
co
n
t
ac
t
ed
at
ema
i
l
:
es
l
y
am
o
v
@
g
ma
i
l
.
co
m.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
De
e
p
ne
ur
al
ne
tw
or
k
s
for
r
e
mov
ing
c
louds
and
ne
bulae
fr
om
s
atelli
te
image
s
(
N
atal
y
a
Glaz
y
r
ina
)
5399
Gul
den
M
urza
beko
v
a
g
rad
u
at
e
d
fro
m
t
h
e
Facu
l
t
y
o
f
A
p
p
l
i
e
d
Mat
h
ema
t
i
c
s
-
Co
n
t
r
o
l
Pro
c
es
s
es
o
f
Sai
n
t
-
Pet
e
rs
b
u
r
g
St
at
e
U
n
i
v
ers
i
t
y
i
n
1
9
9
4
,
w
h
ere
s
h
e
al
s
o
s
u
cce
s
s
f
u
l
l
y
d
efen
d
ed
h
er
d
o
c
t
o
r
al
t
h
es
i
s
i
n
1
9
9
7
o
n
d
i
s
c
ret
e
mat
h
e
mat
i
c
s
an
d
ma
t
h
ema
t
i
ca
l
cy
b
ern
e
t
i
c
s
.
Si
n
ce
1
9
9
8
s
h
e
h
a
s
b
ee
n
a
n
as
s
o
c
i
at
e
p
r
o
fes
s
o
r
i
n
t
h
e
D
ep
ar
t
men
t
o
f
In
f
o
rmat
i
cs
,
h
ead
o
f
t
h
e
In
fo
rma
t
i
o
n
a
n
d
C
o
mmu
n
i
ca
t
i
o
n
T
ech
n
o
l
o
g
i
e
s
D
e
p
art
men
t
d
u
r
i
n
g
2
0
0
3
-
2
0
2
2
,
an
d
rece
n
t
l
y
s
h
e
i
s
a
n
as
s
o
c
i
a
t
e
p
r
o
fes
s
o
r
o
f
t
h
e
C
o
mp
u
t
er
Sc
i
e
n
ce
D
e
p
art
me
n
t
a
t
Se
i
fu
l
l
i
n
K
azak
h
A
g
r
o
t
ec
h
n
i
cal
U
n
i
v
ers
i
t
y
.
Sh
e
h
as
au
t
h
o
red
mo
re
t
h
a
n
1
0
0
p
ap
er
s
.
H
er
res
earch
i
n
t
ere
s
t
s
i
n
c
l
u
d
e
n
u
mer
i
cal
me
t
h
o
d
s
o
f
n
o
n
-
s
m
o
o
t
h
a
n
al
y
s
i
s
an
d
n
o
n
d
i
ffere
n
t
i
ab
l
e
o
p
t
i
m
i
zat
i
o
n
,
mat
h
ema
t
i
ca
l
mo
d
e
l
i
n
g
,
art
i
f
i
ci
a
l
i
n
t
el
l
i
g
en
c
e,
an
d
mach
i
n
e
l
earn
i
n
g
.
Sh
e
can
b
e
co
n
t
ac
t
ed
a
t
emai
l
g
.
m
u
rzab
e
k
o
v
a@
k
azat
u
.
k
z
.
N
urg
ul
A
z
i
y
e
v
a
recei
v
ed
h
er
ma
s
t
er’
s
d
e
g
ree
i
n
2
0
0
9
i
n
c
o
mp
u
t
er
s
ci
e
n
ce
fr
o
m
L
.
N
.
G
u
mi
l
y
o
v
E
u
ra
s
i
a
n
N
at
i
o
n
al
U
n
i
v
er
s
i
t
y
,
K
azak
h
s
t
an
.
Cu
rren
t
l
y
,
s
h
e
i
s
a
s
e
n
i
o
r
l
ect
u
rer
o
f
t
h
e
D
ep
ar
t
men
t
o
f
In
f
o
rmat
i
o
n
Secu
ri
t
y
L
.
N
.
G
u
mi
l
y
o
v
E
u
ra
s
i
a
n
N
a
t
i
o
n
a
l
U
n
i
v
er
s
i
t
y
.
H
e
r
res
earch
i
n
t
ere
s
t
s
i
n
cl
u
d
e
s
i
g
n
a
l
p
ro
c
es
s
i
n
g
t
ec
h
n
o
l
o
g
y
o
f
G
ro
u
n
d
Pen
et
r
at
i
n
g
Rad
ar
(G
PR),
art
i
f
i
ci
a
l
i
n
t
el
l
i
g
en
ce
a
n
d
d
a
t
a
mi
n
i
n
g
.
Sh
e
ca
n
b
e
co
n
t
ac
t
ed
at
ema
i
l
n
u
rg
u
l
_
azi
y
ev
a@
ma
i
l
.
ru
.
Ba
khy
tg
u
l
R
y
s
bekky
zy
i
s
a
D
o
ct
o
r
o
f
Ph
i
l
o
s
o
p
h
y
w
i
t
h
a
d
e
g
ree
i
n
mat
h
ema
t
i
ca
l
mo
d
e
l
i
n
g
a
t
L
.
G
u
mi
l
y
o
v
E
u
ra
s
i
a
n
N
a
t
i
o
n
a
l
U
n
i
v
er
s
i
t
y
,
Rep
u
b
l
i
c
o
f
K
azak
h
s
t
a
n
.
Sh
e
w
o
r
k
s
a
t
t
h
e
D
e
p
ar
t
men
t
o
f
In
f
o
rmat
i
o
n
T
ec
h
n
o
l
o
g
y
an
d
Secu
r
i
t
y
o
f
K
arag
a
n
d
a
T
ech
n
i
c
al
U
n
i
v
ers
i
t
y
n
amed
af
t
er
A
.
Sa
g
i
n
o
v
.
H
er
i
n
t
eres
t
s
i
n
c
l
u
d
e
ma
t
h
ema
t
i
cal
m
o
d
e
l
i
n
g
,
b
i
o
med
i
cal
s
i
g
n
a
l
s
,
a
n
d
s
p
ec
t
ral
an
a
l
y
s
i
s
.
Sh
e
can
b
e
c
o
n
t
act
e
d
at
emai
l
:
b
ak
h
y
t
g
u
l
z@
mai
l
.
ru
.
A
i
n
ur
Ory
nba
y
ev
a
g
r
a
d
u
a
t
e
d
f
r
o
m
A
l
m
a
t
y
S
t
a
t
e
U
n
i
v
e
r
s
i
t
y
n
a
me
d
af
t
er
A
b
a
i
i
n
2
0
0
0
,
m
a
j
o
r
i
n
g
i
n
p
h
y
s
i
c
s
a
n
d
c
o
m
p
u
t
e
r
s
c
i
e
n
ce
.
I
n
2
0
1
5
s
h
e
g
r
a
d
u
a
t
e
d
w
i
t
h
a
m
a
s
t
e
r’
s
d
e
g
r
ee
f
r
o
m
t
h
e
K
a
za
k
h
U
n
i
v
e
r
s
i
t
y
o
f
E
c
o
n
o
m
i
c
s
,
F
i
n
a
n
c
e
a
n
d
I
n
t
e
r
n
a
t
i
o
n
a
l
T
r
a
d
e,
ma
j
o
r
i
n
g
i
n
i
n
f
o
r
ma
t
i
o
n
s
y
s
t
em
s
.
S
h
e
w
o
r
k
s
a
s
a
s
e
n
i
o
r
l
ec
t
u
re
r
a
t
t
h
e
D
e
p
a
r
t
m
e
n
t
o
f
B
i
o
s
t
a
t
i
s
t
i
c
s
,
B
i
o
i
n
f
o
rm
a
t
i
c
s
a
n
d
I
n
f
o
r
ma
t
i
o
n
T
e
c
h
n
o
l
o
g
y
a
t
A
s
t
a
n
a
Me
d
i
c
a
l
U
n
i
v
er
s
i
t
y
.
H
er
r
e
s
e
ar
c
h
i
n
t
er
e
s
t
s
i
n
c
l
u
d
e
b
i
o
s
t
a
t
i
s
t
i
c
s
,
ma
c
h
i
n
e
l
ea
r
n
i
n
g
i
n
m
e
d
i
c
i
n
e,
ar
t
i
f
i
c
i
a
l
i
n
t
e
l
l
i
g
e
n
ce,
a
n
d
i
n
f
o
rm
a
t
i
o
n
s
y
s
t
e
m
s
.
S
h
e
c
a
n
b
e
c
o
n
t
ac
t
e
d
a
t
e
m
a
i
l
:
a
i
n
u
r
_
t
a
s
@
ma
i
l
.
ru
.
a
n
d
t
a
s
t
a
n
o
v
a
7
2
8
8
@
g
m
a
i
l
.
c
o
m
.
N
a
z
i
ra
Ba
kti
y
a
r
o
v
a
g
rad
u
at
e
d
fro
m
A
rk
al
y
k
St
a
t
e
In
s
t
i
t
u
t
e
n
amed
aft
er
Ib
ra
y
A
l
t
y
n
s
ar
i
n
i
n
1
9
9
0
w
i
t
h
a
d
eg
ree
i
n
co
m
p
u
t
er
s
ci
e
n
ce
an
d
mat
h
emat
i
cs
.
W
o
r
k
s
as
a
s
en
i
o
r
l
ect
u
rer
at
t
h
e
D
ep
ar
t
men
t
o
f
Bi
o
s
t
at
i
s
t
i
c
s
,
Bi
o
i
n
f
o
rma
t
i
c
s
an
d
In
f
o
rmat
i
o
n
T
ec
h
n
o
l
o
g
i
es
a
t
A
s
t
an
a
Med
i
cal
U
n
i
v
er
s
i
t
y
.
Sh
e
can
b
e
co
n
t
ac
t
ed
at
ema
i
l
:
b
_
n
azi
ra
_
a@
mai
l
.
r
u
.
Evaluation Warning : The document was created with Spire.PDF for Python.