I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
,
p
p
.
361
~
373
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
1
.
p
p
3
6
1
-
3
7
3
361
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
M
ulti
-
sca
le
featur
es a
ss
isted knowle
dg
e distillatio
n vi
sio
n
trans
former
for l
a
nd cov
er segm
en
tatio
n and
clas
sifi
ca
tion
Su
j
a
t
a
Arj
un
G
a
ik
wa
d
,
Vij
a
y
a
M
us
a
nd
e
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
J
a
w
a
h
a
r
l
a
l
N
e
h
r
u
E
n
g
i
n
e
e
r
i
n
g
C
o
l
l
e
g
e
,
M
G
M
U
n
i
v
e
r
s
i
t
y
,
C
h
h
a
t
r
a
p
a
t
i
S
a
m
b
h
a
j
i
N
a
g
a
r
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
y
2
6
,
2
0
2
5
R
ev
is
ed
No
v
1
7
,
2
0
2
5
Acc
ep
ted
J
an
1
0
,
2
0
2
6
Th
e
m
o
st
si
g
n
ifi
c
a
n
t
p
ro
b
lem
in
re
m
o
te
se
n
sin
g
i
n
terp
re
tatio
n
is
se
m
a
n
ti
c
se
g
m
e
n
tatio
n
,
wh
ich
a
tt
e
m
p
ts
to
g
iv
e
e
a
c
h
p
ix
e
l
i
n
th
e
ima
g
e
a
p
a
rti
c
u
lar
c
las
s.
Th
is
re
se
a
rc
h
wo
rk
fo
ll
o
w
s
th
e
v
a
rio
u
s
ste
p
s,
s
u
c
h
a
s
p
re
-
p
ro
c
e
ss
in
g
,
se
g
m
e
n
tatio
n
,
a
n
d
c
las
sifica
ti
o
n
.
In
it
ially
,
h
i
g
h
sp
a
ti
a
l
re
so
l
u
ti
o
n
re
m
o
te
se
n
sin
g
ima
g
e
s
(RS
I)
a
re
c
o
ll
e
c
te
d
fro
m
th
e
o
p
e
n
-
so
u
rc
e
d
a
tas
e
t.
In
th
e
p
re
-
p
ro
c
e
ss
in
g
sta
g
e
,
a
n
imp
r
o
v
e
d
g
u
i
d
e
d
fil
ter
(Im
p
-
G
F
)
is
u
se
d
to
re
m
o
v
e
v
a
rio
u
s
n
o
ise
s
fro
m
ima
g
e
s.
Ne
x
t,
th
e
se
g
m
e
n
tati
o
n
is
d
o
n
e
b
y
u
sin
g
a
k
n
o
wle
d
g
e
d
isti
ll
a
ti
o
n
-
b
a
se
d
v
isi
o
n
t
ra
n
sfo
rm
e
r
a
p
p
ro
a
c
h
i
n
teg
ra
t
e
d
with
a
n
a
tro
u
s
s
p
a
ti
a
l
m
u
lt
i
-
sc
a
le
p
y
ra
m
id
a
l
m
o
d
u
le
(KD
-
M
u
ViT
P
y
).
Ba
s
e
d
o
n
th
e
se
g
m
e
n
ted
ima
g
e
,
la
n
d
c
o
v
e
r
c
la
ss
e
s
su
c
h
a
s
v
e
g
e
tati
o
n
,
u
r
b
a
n
a
re
a
s,
fo
re
st,
wa
ter
b
o
d
ies
,
a
n
d
ro
a
d
s
a
re
c
las
sified
.
T
h
e
p
ro
p
o
se
d
m
e
th
o
d
o
u
t
p
e
rfo
rm
e
d
th
e
B
h
u
v
a
n
sa
telli
te
d
a
tas
e
t,
a
c
h
i
e
v
in
g
b
e
tt
e
r
a
c
c
u
ra
c
y
,
p
re
c
isio
n
,
re
c
a
ll
,
F
1
-
sc
o
re
,
Dic
e
sc
o
r
e
,
in
ters
e
c
ti
o
n
o
v
e
r
u
n
i
o
n
(
I
o
U
)
,
a
n
d
Ka
p
p
a
sc
o
re
a
t
v
a
lu
e
s
o
f
9
8
.
0
1
%
,
9
8
.
9
9
%
,
9
7
.
4
9
%
,
9
8
.
2
3
%
,
9
8
.
2
3
%
,
9
6
.
5
5
%
,
a
n
d
9
5
.
9
1
%
,
re
sp
e
c
ti
v
e
ly
.
K
ey
w
o
r
d
s
:
I
m
p
r
o
v
ed
g
u
id
ed
f
ilter
Kn
o
wled
g
e
d
is
till
atio
n
Mu
lti s
ca
le
s
eg
m
en
tatio
n
Py
r
am
id
al
m
o
d
u
le
R
em
o
te
s
en
s
in
g
im
ag
e
Vis
io
n
tr
an
s
f
o
r
m
er
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Su
jata
Ar
ju
n
Gaik
wad
D
e
p
a
r
t
m
e
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
J
a
w
a
h
a
r
l
a
l
N
e
h
r
u
E
n
g
i
n
e
e
r
i
n
g
C
o
l
l
e
g
e
,
M
G
M
U
n
i
v
e
r
s
i
t
y
C
h
h
atr
ap
ati
Sam
b
h
aji
Nag
ar
,
Ma
h
ar
ash
tr
a,
I
n
d
ia
E
m
ail:
s
u
jatag
aik
wad
4
1
4
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
L
a
n
d
c
o
v
e
r
d
e
t
e
ct
i
o
n
a
n
d
cl
a
s
s
i
f
i
c
a
ti
o
n
a
r
e
c
r
i
t
i
ca
l
i
n
e
c
o
n
o
m
i
c
as
s
ess
m
e
n
t
,
r
es
o
u
r
c
e
m
a
n
a
g
e
m
e
n
t
,
a
n
d
c
r
o
p
a
r
e
a
a
n
a
l
y
s
is
.
T
h
is
e
v
a
l
u
a
t
i
o
n
s
t
u
d
y
g
r
e
a
tl
y
i
m
p
r
o
v
e
d
h
o
m
e
l
a
n
d
s
e
c
u
r
i
t
y
a
n
d
n
a
ti
o
n
al
e
c
o
n
o
m
i
c
s
t
a
b
i
li
t
y
[
1
]
,
[
2
]
.
R
e
m
o
t
e
s
e
n
s
i
n
g
is
a
p
o
w
e
r
f
u
l
t
e
c
h
n
o
l
o
g
y
f
o
r
e
a
r
t
h
l
a
n
d
o
b
s
e
r
v
a
t
i
o
n
t
h
a
t
e
m
p
l
o
y
s
s
e
n
s
o
r
s
o
n
s
at
e
ll
i
t
es
o
r
a
i
r
p
l
a
n
e
s
t
o
c
a
p
t
u
r
e
i
m
a
g
es
o
f
t
h
e
p
l
a
n
e
t'
s
s
u
r
f
a
c
e
w
i
t
h
o
u
t
r
e
q
u
i
r
i
n
g
d
i
r
e
c
t
p
h
y
s
ic
a
l
co
n
t
a
c
t
[
3
]
,
[
4
]
.
O
n
e
s
i
g
n
i
f
i
c
a
n
t
a
r
e
a
o
f
r
e
m
o
t
e
s
e
n
s
in
g
i
s
o
p
ti
c
a
l r
e
m
o
t
e
s
e
n
s
i
n
g
,
wh
i
c
h
h
a
s
b
e
e
n
u
s
e
d
f
o
r
a
v
a
r
i
e
ty
o
f
p
u
r
p
o
s
e
s
,
s
u
c
h
a
s
s
u
p
e
r
-
r
e
s
o
l
u
ti
o
n
m
a
p
p
i
n
g
o
f
l
a
n
d
c
o
v
e
r
a
n
d
o
b
j
ec
t
d
et
e
c
tio
n
[
5
]
.
U
s
i
n
g
t
h
e
v
as
t
a
m
o
u
n
t
o
f
r
e
m
o
t
e
s
e
n
s
i
n
g
p
h
o
t
o
s
,
r
e
s
e
a
r
c
h
e
r
s
h
a
v
e
f
o
c
u
s
e
d
o
n
a
u
t
o
m
a
t
i
c
a
ll
y
c
l
as
s
i
f
y
i
n
g
l
a
n
d
c
o
v
e
r
u
s
i
n
g
s
at
e
l
li
t
e
p
i
ct
u
r
e
s
[
6
]
.
T
h
e
ex
am
in
atio
n
o
f
lan
d
co
v
e
r
ag
e
h
as
b
ec
o
m
e
m
o
r
e
v
is
ib
le
in
r
ec
en
t
y
ea
r
s
d
u
e
to
th
e
d
e
v
elo
p
m
en
t
o
f
ar
tific
ial
in
tellig
en
ce
(
AI
)
[
7
]
.
AI
-
p
o
wer
ed
m
ac
h
in
e
lear
n
in
g
(
ML
)
as
well
as
d
ee
p
lear
n
in
g
(
DL
)
m
o
d
els
ar
e
em
p
lo
y
ed
to
id
en
tify
ch
a
n
g
es
in
th
e
la
n
d
[
8
]
.
B
ec
au
s
e
o
f
its
d
ee
p
lay
er
s
,
DL
m
o
d
e
ls
o
u
tp
er
f
o
r
m
ML
m
o
d
els
in
d
etec
tio
n
.
ML
m
o
d
els
u
s
e
s
tatis
tical
alg
o
r
ith
m
s
t
o
d
etec
t
p
atter
n
s
in
d
ata,
b
u
t
th
ey
m
ay
s
tr
u
g
g
le
with
co
m
p
licated
,
m
u
ltid
im
en
s
io
n
al
d
atasets
s
u
ch
as satelli
t
e
im
ag
es
[
9
]
.
DL
m
o
d
els,
wh
ich
u
s
e
m
u
lti
-
l
ay
er
ed
n
eu
r
al
n
etwo
r
k
s
,
p
r
o
v
i
d
e
g
r
ea
ter
d
etec
tio
n
s
k
ills
.
T
h
ese
m
o
d
els
ca
n
au
to
m
atica
lly
e
x
tr
ac
t
d
eta
iled
f
ea
tu
r
es
f
r
o
m
r
aw
d
ata
u
s
in
g
d
ee
p
er
lay
er
s
,
ca
tch
in
g
s
m
all
ch
an
g
es
m
o
r
e
ef
f
ec
tiv
ely
[
1
0
]
.
DL
'
s
ca
p
ac
it
y
to
lear
n
h
ier
ar
ch
ical
r
ep
r
esen
tatio
n
s
allo
ws
it
to
p
er
f
o
r
m
well
in
task
s
lik
e
im
ag
e
s
eg
m
en
tatio
n
,
class
if
icatio
n
,
as we
ll a
s
ch
an
g
e
r
ec
o
g
n
itio
n
[
1
1
]
.
As a
r
esu
lt,
DL
-
b
ased
tech
n
iq
u
es h
av
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
361
-
3
7
3
362
em
er
g
ed
as
th
e
p
r
ef
er
r
e
d
m
eth
o
d
f
o
r
d
etec
tin
g
la
n
d
c
o
v
er
ch
an
g
es,
as
th
ey
o
f
f
er
h
ig
h
er
ac
c
u
r
ac
y
an
d
r
esil
ien
ce
th
an
s
tan
d
ar
d
ML
m
o
d
els
[
1
2
]
.
T
h
is
m
ak
es
t
h
em
esp
ec
ially
v
al
u
ab
le
in
u
r
b
an
p
r
e
p
ar
atio
n
,
d
ef
o
r
estatio
n
m
o
n
ito
r
in
g
,
as w
ell
as a
g
r
icu
ltu
r
al
m
an
a
g
em
en
t
[
1
3
]
.
I
n
iti
all
y
,
l
an
d
c
o
v
e
r
class
i
f
i
ca
t
io
n
s
ar
e
d
o
n
e
u
s
i
n
g
m
u
lti
-
s
ca
l
e
c
o
n
v
o
lu
ti
o
n
al
n
eu
r
al
n
e
tw
o
r
k
s
(
C
NNs
)
.
Ho
w
ev
er
,
is
s
u
es
l
ik
e
in
ad
e
q
u
ate
d
ep
th
,
c
o
n
s
tr
ai
n
e
d
r
e
ce
p
ti
v
e
f
i
el
d
s
,
a
n
d
t
h
e
i
n
c
a
p
ac
it
y
t
o
ac
c
u
r
at
el
y
ca
p
t
u
r
e
lo
n
g
-
r
a
n
g
e
i
n
t
e
r
d
ep
e
n
d
e
n
ce
i
n
i
n
t
r
ic
ate
s
it
u
at
io
n
s
h
i
n
d
e
r
e
d
t
h
e
ir
p
e
r
f
o
r
m
a
n
c
e
[
1
4
]
.
T
h
e
y
t
h
u
s
h
av
e
t
r
o
u
b
l
e
m
a
n
a
g
i
n
g
i
r
r
e
g
u
la
r
p
att
er
n
s
in
lar
g
e
-
s
ca
l
e
r
em
o
t
e
s
e
n
s
i
n
g
d
at
a
a
n
d
f
i
n
e
-
g
r
ai
n
ed
class
if
ica
ti
o
n
.
F
u
r
t
h
e
r
m
o
r
e
,
t
h
e
s
eg
m
e
n
t
ati
o
n
a
n
d
c
lass
i
f
i
ca
ti
o
n
o
f
o
v
e
r
l
ap
p
i
n
g
o
r
u
n
cle
ar
la
n
d
c
o
v
e
r
t
y
p
es
w
er
e
l
ess
a
cc
u
r
at
el
y
a
cc
o
m
p
lis
h
e
d
b
y
t
h
e
ea
r
l
y
m
u
lti
-
s
c
al
e
C
NNs
d
u
e
t
o
t
h
eir
a
b
s
en
ce
o
f
s
t
r
o
n
g
f
ea
t
u
r
e
f
u
s
i
o
n
p
r
o
ce
s
s
es
[
1
5
]
.
Alt
h
o
u
g
h
C
NN
-
b
as
ed
m
u
lti
-
s
ca
l
e
m
o
d
els
h
a
v
e
b
ee
n
wi
d
el
y
u
s
ed
f
o
r
la
n
d
c
o
v
er
s
eg
m
e
n
tati
o
n
,
t
h
e
y
s
till
f
a
ce
s
o
m
e
k
e
y
c
h
a
lle
n
g
es
,
s
u
c
h
as
li
m
ite
d
r
e
ce
p
ti
v
e
f
iel
d
s
,
a
m
b
ig
u
o
u
s
la
n
d
c
o
v
e
r
t
y
p
es
,
i
n
s
u
f
f
ic
ie
n
t
d
e
p
t
h
to
ca
p
t
u
r
e
l
o
n
g
-
r
a
n
g
e
d
e
p
en
d
e
n
ci
es,
a
n
d
w
ea
k
f
ea
t
u
r
e
f
u
s
i
o
n
.
Ma
n
y
e
x
is
ti
n
g
m
o
d
els
als
o
s
u
f
f
e
r
e
d
f
r
o
m
h
i
g
h
co
m
p
u
tati
o
n
a
l
c
o
s
t
,
m
em
o
r
y
in
e
f
f
ic
ie
n
c
y
,
a
n
d
f
ea
t
u
r
e
r
e
d
u
n
d
an
c
y
,
w
h
i
c
h
r
e
d
u
ce
t
h
e
s
ca
l
ab
ilit
y
f
o
r
h
i
g
h
-
r
es
o
l
u
ti
o
n
s
at
elli
te
i
m
a
g
e
r
y
.
T
o
b
r
i
d
g
e
t
h
ese
g
a
p
s
,
t
h
is
s
t
u
d
y
in
tr
o
d
u
ce
s
a
n
o
v
el
k
n
o
wl
e
d
g
e
d
is
ti
lla
ti
o
n
-
b
as
ed
v
is
i
o
n
t
r
a
n
s
f
o
r
m
e
r
a
p
p
r
o
a
c
h
i
n
t
eg
r
ate
d
wi
th
m
u
lt
i
-
s
ca
l
e
p
y
r
am
i
d
al
m
o
d
u
le
(
KD
-
M
u
ViT
P
y
)
m
o
d
el
.
T
h
e
p
r
o
p
o
s
e
d
m
o
d
e
l
i
n
te
g
r
at
es
n
o
i
s
e
r
e
d
u
c
ti
o
n
v
i
a
a
n
i
m
p
r
o
v
e
d
g
u
i
d
e
d
f
ilt
er
(
I
m
p
-
GF)
,
r
es
p
o
n
s
e
-
b
as
ed
k
n
o
wl
ed
g
e
d
is
t
illa
ti
o
n
t
o
t
r
a
n
s
f
e
r
r
o
b
u
s
t
r
ep
r
ese
n
ta
ti
o
n
f
r
o
m
te
ac
h
e
r
t
o
s
tu
d
e
n
t
n
e
tw
o
r
k
s
,
an
d
d
y
n
a
m
ic
m
u
l
ti
-
s
ca
le
p
y
r
am
i
d
al
p
o
o
l
in
g
t
o
ca
p
t
u
r
e
b
o
t
h
g
l
o
b
al
an
d
f
i
n
e
-
g
r
ai
n
ed
s
p
a
tia
l
f
ea
t
u
r
es
.
T
h
is
co
m
b
in
ati
o
n
e
n
h
a
n
c
es
s
eg
m
e
n
t
ati
o
n
ac
cu
r
ac
y
,
r
e
d
u
c
es
co
m
p
u
tat
io
n
al
o
v
er
h
e
a
d
,
a
n
d
d
e
li
v
e
r
s
s
ta
te
-
of
-
t
h
e
-
a
r
t
p
e
r
f
o
r
m
a
n
ce
f
o
r
l
a
n
d
co
v
e
r
c
lass
i
f
i
ca
ti
o
n
.
T
h
e
m
aj
o
r
c
o
n
tr
ib
u
t
io
n
o
f
t
h
is
r
es
ea
r
c
h
wo
r
k
is
g
i
v
e
n
b
el
o
w
:
n
o
is
e
f
r
o
m
th
e
i
n
p
u
t
im
ag
es
is
r
em
o
v
e
d
b
y
u
s
in
g
a
n
I
m
p
-
G
F,
w
h
i
c
h
h
e
lp
s
t
o
e
n
h
a
n
ce
i
m
a
g
e
q
u
alit
y
.
T
o
s
eg
m
e
n
t
t
h
e
p
r
e
-
p
r
o
c
ess
e
d
im
ag
es
an
d
i
d
e
n
ti
f
y
l
a
n
d
c
lass
es
u
s
i
n
g
at
r
o
u
s
s
p
at
ial
KD
-
M
u
V
iTP
y
.
2.
RE
L
AT
E
D
WO
RK
S
Var
io
u
s
p
r
e
v
io
u
s
s
tu
d
ies
h
a
v
e
r
ec
o
m
m
e
n
d
ed
a
n
u
m
b
e
r
o
f
m
eth
o
d
s
b
ased
o
n
m
u
lti
-
s
ca
le
an
d
DL
to
im
p
r
o
v
e
t
h
e
ac
cu
r
ac
y
o
f
ex
tr
a
ctin
g
f
ea
tu
r
es a
n
d
g
r
o
u
p
i
n
g
im
ag
es g
en
er
ated
f
r
o
m
r
em
o
te
s
e
n
s
in
g
,
s
u
ch
as:
‒
C
ar
d
am
a
et
a
l.
[
1
6
]
d
ev
elo
p
ed
a
co
n
s
en
s
u
s
m
u
lti
-
s
ca
le
b
in
ar
y
alter
atio
n
r
ec
o
g
n
itio
n
ap
p
r
o
ac
h
f
o
r
o
b
ject
-
b
ased
f
ea
tu
r
e
ex
tr
ac
tio
n
.
Mu
ltip
le
d
etec
to
r
s
b
ased
o
n
d
if
f
er
e
n
t
s
eg
m
en
tatio
n
a
p
p
r
o
ac
h
es
wer
e
u
tili
ze
d
at
d
if
f
er
en
t
s
ca
les
to
ex
p
lo
it
th
e
v
er
y
h
ig
h
r
eso
lu
tio
n
(
VHR
)
p
ictu
r
es
'
h
ig
h
s
p
atial
r
eso
lu
tio
n
an
d
ca
p
tu
r
e
ch
an
g
es
at
d
if
f
er
e
n
t
g
r
an
u
lar
ity
le
v
els.
T
h
e
ch
an
g
e
v
ec
to
r
an
aly
s
is
-
s
eg
m
en
t
an
y
th
in
g
m
o
d
el
(
C
VA
-
SAM
)
was u
s
ed
o
n
th
e
s
eg
m
en
t le
v
el
r
ath
er
t
h
an
th
e
p
ix
el
lev
el.
‒
W
an
g
et
a
l
.
[
1
7
]
d
ev
elo
p
ed
th
e
p
ar
allel
s
win
(P
-
Swin
)
tr
an
s
f
o
r
m
er
n
etwo
r
k
,
a
t
r
an
s
f
o
r
m
er
n
etwo
r
k
b
ased
o
n
p
ar
allel
win
d
o
ws.
T
h
e
P
-
s
win
tr
an
s
f
o
r
m
e
r
b
lo
c
k
,
wh
ich
co
m
p
r
is
es
th
e
f
ee
d
f
o
r
war
d
n
etwo
r
k
(
FF
N)
an
d
win
d
o
w
-
b
ased
s
elf
-
atten
tio
n
in
ter
ac
tio
n
(
W
SAI
)
,
is
a
cr
itical
co
m
p
o
n
en
t
o
f
P
-
Swin
.
W
SAI
ca
n
d
eter
m
in
e
th
e
lin
k
b
etwe
en
win
d
o
ws
as
well
a
s
th
e
r
elatio
n
s
h
ip
with
in
win
d
o
ws.
As
a
r
esu
lt,
i
t
in
cr
ea
s
es th
e
n
etwo
r
k
'
s
ab
ilit
y
to
co
llect
f
ea
tu
r
e
c
o
n
tex
t
d
ata.
‒
Ma
et
a
l
.
[
1
8
]
d
ev
elo
p
ed
a
n
ew
C
NN
p
ix
el
-
by
-
p
ix
el
class
if
icatio
n
ap
p
r
o
ac
h
with
s
m
aller
s
izes.
T
h
e
ap
p
r
o
ac
h
ad
d
r
ess
es
th
e
p
r
o
b
l
em
o
f
in
ad
eq
u
ate
m
u
lti
-
s
ca
le
lear
n
in
g
f
o
r
class
if
icatio
n
b
y
em
p
lo
y
in
g
m
u
lti
-
s
ca
le
n
etwo
r
k
s
to
r
em
o
v
e
m
u
lti
-
s
ca
le
co
n
tex
tu
al
d
ata
a
t a
f
in
e
-
g
r
ai
n
ed
lev
el.
‒
J
ia
et
a
l
.
[
1
9
]
d
ev
elo
p
e
d
a
m
u
lti
-
atten
tio
n
s
em
an
tic
s
eg
m
en
tatio
n
n
etwo
r
k
f
o
r
r
em
o
te
s
en
s
in
g
im
ag
es
(
R
SI)
.
T
h
e
b
aselin
e
m
o
d
el
is
U
-
Net
,
an
d
th
e
b
ac
k
b
o
n
e
n
etwo
r
k
'
s
ca
p
ac
ity
to
r
em
o
v
e
f
in
e
-
g
r
ain
e
d
s
tr
u
ctu
r
es
is
im
p
r
o
v
e
d
b
y
in
co
r
p
o
r
atin
g
an
o
r
g
a
n
ized
atten
ti
o
n
-
b
ased
r
esid
u
al
n
etwo
r
k
in
t
o
th
e
en
c
o
d
er
.
T
o
in
cr
ea
s
e
n
etwo
r
k
in
f
o
r
m
ati
o
n
ex
t
r
ac
tio
n
,
th
e
d
e
co
d
er
'
s
tr
ad
itio
n
al
u
p
-
s
am
p
lin
g
o
p
e
r
ato
r
was
r
ep
lace
d
b
y
a
co
n
ten
t
-
awa
r
e
r
ea
r
r
an
g
e
m
en
t m
o
d
u
le.
‒
Xu
et
a
l.
[
2
0
]
d
ev
elo
p
ed
a
two
-
b
r
an
ch
ed
s
u
p
er
v
is
ed
s
em
an
t
ic
s
eg
m
en
tatio
n
s
y
s
tem
.
A
n
e
w
s
y
m
m
etr
ic
atten
tio
n
m
o
d
u
le
with
en
h
an
c
ed
s
tr
ip
p
o
o
lin
g
was
d
e
v
elo
p
e
d
.
T
h
e
m
u
ltip
le
lo
n
g
ac
ce
s
s
ib
le
f
ield
s
allo
w
f
o
r
th
e
ac
q
u
is
itio
n
o
f
m
o
r
e
a
n
i
s
o
tr
o
p
ic
co
n
te
x
tu
al
in
f
o
r
m
atio
n
an
d
b
etter
v
is
ib
ilit
y
o
f
ir
r
eg
u
lar
o
b
jects.
‒
Z
h
an
g
et
a
l
.
[
2
1
]
d
ev
elo
p
ed
t
h
e
ex
ten
d
e
d
to
p
o
lo
g
y
p
r
eser
v
in
g
s
eg
m
en
tatio
n
(
E
T
PS
)
m
o
d
el
-
b
ased
m
u
lti
-
s
ca
le
as
well
as
m
u
lti
-
f
ea
tu
r
e
m
eth
o
d
th
at
em
p
lo
y
s
co
n
v
o
l
u
tio
n
al
n
e
u
r
al
s
eg
m
e
n
tatio
n
.
T
h
e
s
u
g
g
ested
m
o
d
el
s
p
lits
th
e
im
ag
es
in
t
o
s
u
p
er
p
ix
els
em
p
lo
y
in
g
th
e
E
T
PS
m
eth
o
d
.
Ho
wev
er
,
th
e
m
u
lti
-
r
eso
lu
tio
n
s
eg
m
en
tatio
n
m
o
d
el
r
etr
iev
e
d
f
ea
tu
r
es a
n
d
m
a
p
p
ed
th
em
to
s
u
p
er
p
ix
els f
o
r
m
u
lti
-
r
ep
r
esen
t
atio
n
.
‒
Sh
i
et
a
l
.
[
2
2
]
d
ev
el
o
p
ed
a
lan
d
c
o
v
er
class
if
icatio
n
s
y
s
tem
b
ased
o
n
m
u
lti
-
s
p
ec
tr
al
lig
h
t
d
etec
tio
n
a
n
d
r
an
g
in
g
(
L
iDAR
)
with
s
p
atial
m
u
lti
-
s
ca
le
as
well
as
s
p
ec
tr
al
f
ea
tu
r
e
s
elec
tio
n
.
I
n
iti
ally
,
k
-
n
ea
r
est
n
eig
h
b
o
r
h
o
o
d
was
em
p
lo
y
e
d
to
c
h
o
o
s
e
n
eig
h
b
o
r
h
o
o
d
p
o
in
ts
f
r
o
m
m
u
lti
-
s
p
ec
tr
al
L
iDAR
d
ata.
Ad
d
itio
n
ally
,
s
p
atial
as we
ll a
s
s
p
ec
tr
al
in
f
o
r
m
atio
n
was r
etr
iev
ed
f
r
o
m
th
e
m
u
lti
-
s
ca
le
n
eig
h
b
o
r
h
o
o
d
.
‒
L
i
et
a
l
.
[
2
3
]
d
ev
elo
p
ed
a
m
u
lti
-
s
ca
le
f
u
lly
co
n
v
o
lu
tio
n
al
n
etwo
r
k
(
MSFC
N)
with
a
m
u
lti
-
s
ca
le
co
n
v
o
l
u
tio
n
al
k
er
n
el
as
well
as
a
ch
an
n
el
atten
tio
n
b
l
o
ck
(
C
AB
)
as
well
a
s
a
g
lo
b
al
p
o
o
li
n
g
m
o
d
u
le
f
o
r
lan
d
co
v
e
r
ca
teg
o
r
izatio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mu
lti
-
s
ca
le
fea
tu
r
es a
s
s
i
s
ted
kn
o
w
led
g
e
d
is
till
a
tio
n
visi
o
n
tr
a
n
s
fo
r
mer fo
r
la
n
d
…
(
S
u
ja
t
a
A
r
ju
n
Ga
ikw
a
d
)
363
‒
Ma
r
tin
s
et
a
l
.
[
2
4
]
s
u
g
g
ested
th
e
m
u
lti
-
s
ca
le
o
b
ject
-
b
ase
d
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etw
o
r
k
s
(
OC
NN)
m
o
d
el.
Ho
we
v
er
,
th
e
s
u
g
g
e
s
ted
m
o
d
el
class
if
ies
th
e
lar
g
e
-
s
ca
le
lan
d
ar
ea
with
a
1
4
5
,
7
5
0
Km
2
r
eso
lu
tio
n
.
Mo
r
eo
v
er
,
th
e
s
u
g
g
ested
m
eth
o
d
co
n
s
is
ts
o
f
t
h
r
ee
p
h
ases
:
im
ag
e
s
eg
m
en
tatio
n
,
s
k
eleto
n
b
ased
alg
o
r
ith
m
f
o
r
o
b
ject
an
a
ly
s
is
,
an
d
th
e
ap
p
licatio
n
o
f
m
u
ltip
le
C
NNs f
o
r
u
ltima
te
clas
s
if
icatio
n
.
‒
C
h
e
n
e
t
a
l
.
[
2
5
]
d
ev
e
lo
p
ed
t
h
e
m
u
l
t
i
-
l
e
v
e
l
f
ea
t
u
r
e
ag
g
r
e
g
a
t
io
n
n
e
t
wo
r
k
(
M
F
A
N
e
t)
m
o
d
e
l
.
T
h
e
s
u
g
g
e
s
t
ed
ap
p
r
o
ac
h
h
ad
im
p
r
o
v
ed
t
wo
p
h
a
s
e
s
,
s
u
c
h
a
s
d
e
ep
f
e
a
t
u
r
e
ex
t
r
ac
t
i
o
n
a
s
we
l
l
as
u
p
-
s
a
m
p
l
in
g
f
e
a
t
u
r
e
f
u
s
i
o
n
.
T
h
e
liter
atu
r
e
s
ea
r
ch
id
e
n
tif
ied
s
ev
er
al
co
n
s
tr
ain
ts
,
in
cl
u
d
in
g
th
e
m
o
d
el'
s
h
ig
h
co
m
p
u
tatio
n
al
co
m
p
lex
ity
,
n
ee
d
f
o
r
m
o
r
e
m
em
o
r
y
to
p
r
o
ce
s
s
,
an
d
h
ig
h
co
m
p
u
tatio
n
al
co
s
t;
also
,
th
e
m
o
d
el
h
as
s
ca
lin
g
co
n
ce
r
n
s
[
1
6
]
–
[
1
8
]
.
I
n
ad
d
itio
n
,
co
n
v
en
tio
n
al
m
o
d
els
r
eq
u
ir
e
m
o
r
e
tim
e
f
o
r
tr
ai
n
in
g
.
Fu
r
t
h
er
m
o
r
e
,
in
cr
ea
s
ed
f
ea
tu
r
e
r
ed
u
n
d
an
c
y
r
en
d
er
s
th
e
m
o
d
el
in
ef
f
icien
t
[
1
9
]
–
[
2
3
]
.
R
eg
ar
d
less
o
f
h
o
w
m
an
y
lay
er
s
ar
e
p
r
esen
t
in
th
e
n
etwo
r
k
,
n
etwo
r
k
d
if
f
icu
lties
m
ay
ar
is
e,
an
d
f
u
tu
r
e
f
u
s
io
n
ap
p
r
o
ac
h
es
m
ay
e
n
h
an
ce
co
m
p
lex
ity
[
2
4
]
,
[
2
5
]
.
T
h
ese
ar
e
th
e
o
v
er
all
lim
itatio
n
s
m
en
tio
n
ed
i
n
th
e
ex
is
tin
g
s
u
r
v
ey
.
T
o
ad
d
r
ess
th
ese
cu
r
r
en
t
ch
allen
g
es,
a
u
n
iq
u
e
k
n
o
wled
g
e
d
is
till
atio
n
-
b
ased
v
is
io
n
tr
an
s
f
o
r
m
er
ap
p
r
o
ac
h
was
p
r
esen
ted
,
wh
ich
is
co
u
p
led
with
an
atr
o
u
s
s
p
atial
m
u
lti
-
s
ca
le
p
y
r
am
id
al
m
o
d
u
le
to
d
iv
id
e
an
d
class
if
y
lan
d
co
v
er
ef
f
ec
tiv
ely
.
T
h
is
s
tu
d
y
f
o
u
n
d
th
at
a
k
n
o
wled
g
e
d
is
till
atio
n
-
b
ased
v
is
io
n
tr
an
s
f
o
r
m
er
m
o
d
el
en
h
an
ce
s
ac
c
u
r
ac
y
wh
ile
lo
w
er
in
g
c
o
m
p
u
tin
g
co
s
ts
.
I
n
ad
d
itio
n
,
th
e
u
n
iq
u
e
m
o
d
u
le
in
c
lu
d
es
a
m
u
lti
-
s
ca
le
p
y
r
am
id
al
m
o
d
u
le
lay
er
th
at
p
r
o
ce
s
s
es
an
d
ex
tr
ac
ts
th
e
m
u
lti
-
s
ca
le
asp
ec
ts
o
f
th
e
in
p
u
t
im
ag
e
wh
ile
also
ass
is
tin
g
in
th
e
ac
q
u
is
itio
n
o
f
h
ig
h
-
lev
el
co
n
tex
tu
al
in
f
o
r
m
atio
n
.
Fu
r
th
e
r
m
o
r
e
,
th
e
m
u
lti
-
s
ca
le
p
y
r
am
id
al
m
o
d
u
le
im
p
r
o
v
es
co
m
p
u
tatio
n
al
ef
f
icien
cy
an
d
is
ap
p
r
o
p
r
iate
f
o
r
r
ea
l
-
tim
e
ap
p
licatio
n
s
.
T
h
e
KD
-
b
ased
v
is
io
n
tr
an
s
f
o
r
m
er
c
o
n
s
is
ten
tly
im
p
r
o
v
es
class
if
icatio
n
p
er
f
o
r
m
an
ce
to
b
o
o
s
t
s
em
an
tic
s
eg
m
en
tatio
n
tr
an
s
f
o
r
m
er
s
,
r
esu
ltin
g
in
h
i
g
h
er
ac
cu
r
ac
y
,
lo
wer
co
m
p
u
tatio
n
al
co
s
t,
an
d
im
p
r
o
v
ed
p
r
e
-
tr
a
in
in
g
n
ee
d
s
.
3.
P
RO
P
O
SE
D
M
E
T
H
O
D
T
h
is
s
tu
d
y
p
r
esen
ted
a
m
u
lti
-
s
ca
le
f
ea
tu
r
e
ass
is
ted
s
em
an
tic
s
eg
m
en
tatio
n
m
eth
o
d
f
o
r
lan
d
s
eg
m
en
tatio
n
an
d
class
if
icatio
n
.
T
h
e
wo
r
k
i
n
g
f
lo
w
o
f
r
esea
r
ch
is
r
ep
r
esen
ted
in
Fig
u
r
e
1
.
T
h
e
ar
ch
itectu
r
e
s
h
o
ws
th
e
f
lo
w
o
f
th
e
s
u
g
g
ested
m
eth
o
d
o
lo
g
y
;
i
n
itially
,
d
at
a
wer
e
ac
q
u
ir
ed
f
r
o
m
th
e
B
h
u
v
an
s
atellite
im
ag
e
d
ataset.
T
h
e
p
r
e
-
p
r
o
ce
s
s
in
g
s
tag
es
u
s
e
th
e
I
m
p
-
GF
m
eth
o
d
,
wh
ich
ef
f
ec
tiv
ely
elim
in
ates
th
e
u
n
d
esire
d
n
o
is
e
an
d
ar
tifa
cts
f
r
o
m
th
e
im
ag
e.
Af
ter
p
r
e
-
p
r
o
ce
s
s
in
g
is
d
o
n
e,
s
eg
m
e
n
tatio
n
is
p
er
f
o
r
m
ed
b
y
u
s
in
g
th
e
KD
-
Mu
ViT
Py
ap
p
r
o
ac
h
.
T
h
er
ef
o
r
e,
t
h
e
s
eg
m
e
n
ted
im
a
g
es
a
r
e
u
s
ed
t
o
class
if
y
th
e
lan
d
co
v
er
in
to
n
u
m
er
o
u
s
k
in
d
s
,
s
u
ch
as r
o
a
d
s
,
u
r
b
an
ar
e
as,
f
o
r
ests
,
wate
r
b
o
d
ies,
as we
ll a
s
v
eg
etatio
n
.
Im
prove
d gui
de
d
fi
l
t
e
r
Im
a
ge
a
c
qui
s
i
t
i
on
know
l
e
dge
di
s
t
i
l
l
a
t
i
on ba
s
e
d vi
s
i
on
t
ra
ns
for
m
e
r a
ppro
a
c
h i
nt
e
gra
t
e
d
w
i
t
h a
t
rous
s
pa
t
i
a
l
m
ul
t
i
s
c
a
l
e
pyra
m
i
da
l
m
odul
e
S
e
gm
e
nt
e
d out
c
om
e
s
Bhuva
n s
a
t
e
l
l
i
t
e
i
m
a
ge
da
t
a
s
e
t
P
re
-
proc
e
s
s
i
ng
S
e
gm
e
nt
a
t
i
on a
nd
c
l
a
s
s
i
fi
e
r m
ode
l
V
e
ge
t
a
t
i
on
U
rba
n a
re
a
s
F
ore
s
t
W
a
t
e
r bodi
e
s
Roa
ds
Cl
a
s
s
i
fi
c
a
t
i
on
Fig
u
r
e
1
.
Ar
c
h
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
m
et
h
o
d
o
lo
g
y
3
.
1
.
P
re
-
pro
ce
s
s
ing
by
us
ing
a
n im
pro
v
ed
g
uid
ed
f
ilte
r
I
n
th
is
wo
r
k
,
im
a
g
es
ar
e
o
b
tai
n
ed
f
r
o
m
a
s
atellite
im
ag
e
d
at
aset;
n
ev
er
th
eless
,
th
e
co
llect
ed
im
ag
es
co
n
tain
s
o
m
e
ar
tifa
cts
an
d
u
n
d
esire
d
n
o
is
es
s
u
ch
as
s
p
ec
k
le
n
o
is
e
an
d
s
alt
an
d
p
ep
p
er
n
o
is
e.
T
h
ese
n
o
is
e
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
361
-
3
7
3
364
wer
e
r
em
o
v
ed
u
s
in
g
th
e
I
m
p
-
GF
m
eth
o
d
.
Ho
wev
er
,
th
is
f
ilter
ca
n
b
e
u
s
ed
to
elim
in
ate
t
h
e
n
o
is
e
f
r
o
m
th
e
s
u
b
-
b
an
d
im
ag
es
[
2
6
]
.
T
h
e
n
o
i
s
e
elim
in
ated
f
r
o
m
th
e
im
ag
e
b
y
u
s
in
g
I
m
p
-
GF is f
o
r
m
u
late
d
in
(
1
)
.
=
+
,
∀
∈
(
1
)
Her
e
d
en
o
tes
th
e
win
d
o
w
a
n
d
lin
ea
r
v
ar
ia
b
les
ar
e
r
ep
r
es
en
ted
as
an
d
.
Nex
t,
th
e
(
2
)
r
em
o
v
es
th
e
u
n
wan
ted
n
o
is
e
to
d
eter
m
i
n
e
t
h
e
lin
ea
r
v
a
r
iab
les.
=
1
−
(
2
)
Her
e,
th
e
i
n
p
u
t
im
ag
e
is
s
p
ec
if
ied
as
,
an
d
a
n
o
is
e
elem
en
t
is
r
ep
r
esen
ted
as
,
h
o
wev
er
,
m
in
im
izin
g
th
e
d
if
f
er
en
ce
b
etwe
en
an
d
is
f
o
r
m
u
lated
in
(
3
)
.
(
,
)
=
∑
(
(
+
−
)
2
+
2
)
∈
(
3
)
W
h
er
e
r
ep
r
esen
t
t
h
e
n
o
r
m
ali
za
tio
n
v
ar
iab
le.
I
m
p
-
GF
in
tr
o
d
u
ce
d
th
e
J
ac
ca
r
d
s
im
ilar
ity
,
wh
ich
is
u
s
ed
f
o
r
id
en
tify
in
g
im
a
g
es
n
ea
r
th
e
e
d
g
es,
an
d
is
f
o
r
m
u
lated
in
(
4
)
an
d
(
5
)
.
Her
e
J
ac
ca
r
d
s
im
ilar
ity
is
r
ep
r
esen
ted
as
,
th
u
s
,
th
e
p
r
o
ce
s
s
o
f
I
m
p
-
GF e
f
f
ec
tiv
ely
elim
in
ates
th
e
n
o
is
e
an
d
en
h
an
ce
s
th
e
ed
g
e
an
d
c
o
n
tr
ast
ac
cu
r
ac
y
o
f
th
e
d
e
g
r
ad
e
d
im
ag
e.
(
,
)
=
∑
(
(
+
−
)
2
+
2
)
∈
+
(
4
)
=
1
−
|
∩
|
|
∪
|
(
5
)
3
.
2
.
Seg
m
ent
a
t
i
o
n a
nd
la
nd
co
v
er
cla
s
s
if
ica
t
io
n by
KD
-
M
uVi
T
P
y
T
h
is
s
tu
d
y
in
tr
o
d
u
ce
d
a
KD
-
Mu
ViT
Py
m
o
d
el
f
o
r
p
ictu
r
e
s
eg
m
en
tatio
n
a
n
d
class
if
icatio
n
.
T
h
is
tech
n
iq
u
e
a
d
d
r
ess
es
th
e
ch
allen
g
es
th
at
ex
is
t
in
th
e
u
s
u
al
ap
p
r
o
ac
h
.
I
n
g
en
er
al,
d
is
till
atio
n
m
eth
o
d
o
lo
g
ies,
k
n
o
wled
g
e
ty
p
e
,
an
d
teac
h
er
s
tu
d
en
t
s
tr
u
ctu
r
e
all
h
av
e
an
im
p
ac
t
o
n
s
tu
d
e
n
t
m
o
d
el
lear
n
i
n
g
.
K
n
o
wled
g
e
ca
n
b
e
ca
teg
o
r
ize
d
as
f
ea
tu
r
e
,
r
esp
o
n
s
e,
o
r
r
elatio
n
s
h
ip
-
b
ased
,
e
m
p
lo
y
in
g
th
e
teac
h
er
m
o
d
el'
s
k
n
o
wled
g
e
s
ets.
Fo
r
p
ictu
r
e
class
if
icatio
n
,
r
esp
o
n
s
e
-
b
ased
k
n
o
wled
g
e
s
er
v
es
as
a
class
if
icatio
n
p
r
o
b
lem
,
o
f
t
en
k
n
o
wn
as
h
ar
d
lab
els.
I
n
class
if
icatio
n
,
r
ea
l
lab
els
ar
e
u
tili
ze
d
as
h
ar
d
lab
el
s
,
an
d
th
e
p
r
o
b
ab
ilit
y
d
is
s
em
in
atio
n
o
f
t
h
e
m
o
d
el
is
f
ed
in
to
th
e
So
f
t
M
a
x
f
u
n
ct
io
n
.
Fu
r
th
er
m
o
r
e,
th
e
r
esu
lt
o
f
th
e
So
f
t
M
a
x
f
u
n
ctio
n
is
im
m
ed
iately
m
atch
e
d
with
th
e
h
ar
d
la
b
el
to
estab
lis
h
th
e
s
p
ec
if
ic
ca
teg
o
r
y
.
3
.
2
.
1.
K
no
wledg
e
dis
t
illa
t
io
n
-
ba
s
ed
v
is
io
n t
ra
ns
f
o
rm
er
T
h
is
s
tu
d
y
u
s
ed
th
e
KD
-
b
a
s
ed
v
is
io
n
tr
an
s
f
o
r
m
er
,
wh
ich
co
n
s
is
ten
tly
im
p
r
o
v
es
cl
ass
if
icatio
n
p
er
f
o
r
m
an
ce
an
d
s
em
an
tic
s
eg
m
en
tatio
n
tr
a
n
s
f
o
r
m
er
s
,
r
esu
ltin
g
in
h
ig
h
e
r
ac
cu
r
ac
y
,
lo
w
er
p
r
o
ce
s
s
in
g
co
s
ts
,
an
d
im
p
r
o
v
ed
p
r
e
-
tr
ain
in
g
n
ee
d
s
[
2
7
]
.
I
n
KD,
k
n
o
wled
g
e
is
tr
an
s
f
er
r
ed
f
r
o
m
teac
h
e
r
to
s
tu
d
en
t
u
s
in
g
a
m
o
d
el
th
at
is
ex
ten
s
iv
ely
u
tili
ze
d
in
co
m
p
u
ter
v
is
io
n
.
Ho
we
v
er
,
th
e
m
ain
p
u
r
p
o
s
e
o
f
th
is
r
esear
c
h
is
to
cr
ea
te
a
K
D
s
tr
u
ctu
r
e
f
o
r
s
eg
m
e
n
tatio
n
u
s
i
n
g
tr
an
s
f
o
r
m
er
-
b
ased
m
o
d
els.
Me
an
wh
ile,
th
is
s
tu
d
y
u
s
ed
a
r
esp
o
n
s
e
-
b
ased
KD
tech
n
iq
u
e
to
t
r
ain
a
teac
h
e
r
a
n
d
s
tu
d
e
n
t
m
o
d
el.
C
o
m
p
ar
e
d
to
p
r
e
v
io
u
s
KD
m
o
d
els,
th
e
r
esp
o
n
s
e
-
b
ased
KD
m
eth
o
d
o
f
f
er
s
t
h
e
ad
v
an
tag
e
s
o
f
m
in
im
izin
g
er
r
o
r
,
d
is
tr
ib
u
tin
g
d
ata
ad
a
p
tiv
ely
b
ased
o
n
d
iv
er
s
e
ty
p
es,
en
h
an
cin
g
m
o
d
el
i
n
ter
p
r
etab
ilit
y
,
an
d
s
ig
n
if
ican
tly
im
p
r
o
v
in
g
m
o
d
el
r
o
b
u
s
tn
ess
.
Fig
u
r
e
2
d
ep
icts
th
e
ar
ch
itectu
r
e
o
f
th
e
k
n
o
wled
g
e
d
is
till
atio
n
-
b
ased
v
is
io
n
tr
an
s
f
o
r
m
er
.
T
h
e
s
o
f
t
lab
el
is
th
e
k
e
y
p
o
in
t
o
f
r
esp
o
n
s
e
-
b
ased
KD;
in
th
is
ca
s
e,
th
e
s
o
f
t
lab
el
h
as
o
b
tain
ed
a
s
m
o
o
th
ed
r
ep
r
esen
tatio
n
o
f
t
h
e
So
f
tMa
x
r
o
le
f
r
o
m
th
e
p
r
o
b
ab
ilit
y
o
f
o
u
tp
u
t
d
is
s
em
in
atio
n
.
Ho
wev
er
,
th
e
s
o
f
ten
in
g
o
u
tp
u
t is ca
lcu
lated
em
p
lo
y
in
g
th
e
teac
h
e
r
tr
ain
in
g
m
o
d
el,
as r
e
p
r
esen
ted
in
(
6
)
.
=
(
)
∑
(
)
(
6
)
Her
e,
t
h
e
s
o
f
t
la
b
e
l
is
d
en
o
t
e
d
as
,
w
h
i
c
h
is
en
g
ag
e
d
to
g
u
i
d
e
t
h
e
s
tu
d
e
n
t
m
o
d
el
as
w
ell
a
s
s
ig
n
i
f
i
es
t
h
e
lik
eli
h
o
o
d
o
f
p
o
s
t
er
io
r
d
is
s
e
m
in
a
ti
o
n
,
w
h
ic
h
is
p
e
r
f
o
r
m
e
d
b
ef
o
r
e
S
o
f
tM
a
x
.
T
h
e
d
en
o
t
es
th
e
v
al
u
e
o
f
t
h
e
class
;
h
e
r
e
is
th
e
p
o
s
t
er
io
r
p
r
o
b
ab
ilit
y
.
M
o
r
eo
v
er
,
o
n
e
s
i
g
n
i
f
ic
a
n
t
r
est
r
ic
ti
o
n
is
th
e
te
m
p
e
r
a
tu
r
e
s
y
s
t
em
,
wh
i
ch
p
e
r
m
i
ts
s
m
o
o
t
h
i
n
g
t
h
e
p
o
s
te
r
i
o
r
p
r
o
b
a
b
il
it
y
d
is
s
e
m
i
n
ati
o
n
.
W
h
e
n
th
e
t
em
p
er
at
u
r
e
i
s
1
,
i
t
is
f
u
n
ct
io
n
a
l
to
t
h
e
So
f
t
Ma
x
m
o
d
el
;
i
f
is
l
ar
g
e
r
t
h
a
n
1
,
th
e
p
r
o
b
ab
ilit
y
d
is
tr
i
b
u
ti
o
n
o
f
t
h
e
o
u
t
p
u
t
b
ec
o
m
es
s
m
o
o
t
h
e
r
a
n
d
s
er
v
es
t
o
p
r
es
er
v
e
s
i
m
il
a
r
i
n
f
o
r
m
at
io
n
,
w
h
i
le
is
i
n
f
i
n
it
e,
it
r
es
em
b
les
a
u
n
if
o
r
m
d
is
s
e
m
i
n
a
ti
o
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mu
lti
-
s
ca
le
fea
tu
r
es a
s
s
i
s
ted
kn
o
w
led
g
e
d
is
till
a
tio
n
visi
o
n
tr
a
n
s
fo
r
mer fo
r
la
n
d
…
(
S
u
ja
t
a
A
r
ju
n
Ga
ikw
a
d
)
365
In
p
u
t
i
m
a
g
e
E
m
b
e
d
d
e
d
p
a
t
c
h
e
s
N
o
rm
M
u
l
t
i
-
h
e
a
d
a
t
t
e
n
t
i
o
n
N
o
rm
M
L
P
D
i
l
a
t
e
d
c
o
n
v
o
l
u
t
i
o
n
a
l
b
l
o
c
k
M
u
l
t
i
-
p
y
ra
m
i
d
a
l
m
o
d
u
l
e
S
o
ft
m
a
x
D
i
s
t
i
l
l
a
t
i
o
n
l
o
s
s
E
m
b
e
d
d
e
d
p
a
t
c
h
e
s
N
o
rm
M
u
l
t
i
-
h
e
a
d
a
t
t
e
n
t
i
o
n
N
o
rm
M
L
P
D
i
l
a
t
e
d
c
o
n
v
o
l
u
t
i
o
n
a
l
b
l
o
c
k
M
u
l
t
i
-
p
y
ra
m
i
d
a
l
m
o
d
u
l
e
S
o
ft
m
a
x
Cro
s
s
e
n
t
ro
p
y
l
o
s
s
T
e
a
c
h
e
r
m
o
d
e
l
S
t
u
d
e
n
t
m
o
d
e
l
P
re
d
i
c
t
e
d
re
s
u
l
t
s
S
o
ft
l
a
b
e
l
s
Fig
u
r
e
2
.
Ar
c
h
itectu
r
e
o
f
k
n
o
wled
g
e
d
is
till
atio
n
-
b
ased
v
is
io
n
tr
a
n
s
f
o
r
m
er
3
.
2
.
2
.
M
ulti
-
s
ca
le
py
ra
m
ida
l m
o
du
le
B
ased
o
n
co
m
p
u
ter
v
is
io
n
,
th
e
m
u
lti
-
s
ca
le
p
y
r
am
id
al
m
o
d
u
le
is
in
tr
o
d
u
ce
d
.
T
h
e
ter
m
m
u
lti
-
s
ca
le
r
ef
er
s
to
s
am
p
lin
g
o
f
f
ea
tu
r
e
s
at
d
if
f
er
e
n
t
s
tag
es.
Ho
we
v
er
,
p
e
r
f
o
r
m
in
g
a
s
p
ec
if
ic
task
r
eq
u
ir
es
d
if
f
er
en
t
f
ea
tu
r
es
at
d
if
f
er
en
t
s
ca
le
co
n
d
itio
n
s
;
th
er
e
f
o
r
e,
th
e
m
u
lti
-
s
ca
le
atten
tio
n
n
etwo
r
k
(
MSANe
t)
m
u
s
t
b
e
in
tr
o
d
u
ce
d
.
T
h
e
m
u
lti
-
s
ca
le
p
y
r
am
id
al
p
o
o
lin
g
m
o
d
u
le
o
n
th
e
en
co
d
er
a
n
d
th
e
atten
tio
n
m
ec
h
an
is
m
o
n
th
e
d
ec
o
d
er
,
wh
ich
p
r
im
ar
ily
r
ef
l
ec
t
th
e
m
u
lti
-
s
ca
le
f
ea
tu
r
e
in
t
o
two
p
o
r
tio
n
s
,
s
u
ch
as
f
ea
tu
r
e
f
u
s
io
n
a
n
d
f
ea
tu
r
e
m
ap
,
ar
e
th
e
f
o
u
n
d
atio
n
o
f
th
e
MSANe
t
n
etwo
r
k
'
s
d
esig
n
.
Fig
u
r
e
3
r
ep
r
esen
ts
th
e
ar
ch
itectu
r
e
o
f
th
e
m
u
lti
-
s
ca
le
p
y
r
am
id
al
m
o
d
u
le
m
o
d
el.
D
RConv
D
RCo
nv
D
RConv
C
onc
a
t
UP
X
G
(
X
)
C
o
n
v
K
x
K
A
r
g
m
a
x
(
*
)
D
RCo
nv
.
.
.
64
X
64
32
X
32
+
UP
P
ool
i
ng
Fig
u
r
e
3
.
Ar
c
h
itectu
r
e
o
f
m
u
lt
i
-
s
ca
le
p
y
r
am
id
al
m
o
d
u
le
T
h
e
u
s
e
o
f
a
m
u
lti
-
s
ca
le
in
p
u
t
f
ea
tu
r
e
m
ap
in
cr
ea
s
es
f
u
s
io
n
s
b
etwe
en
lo
w
-
an
d
h
ig
h
-
r
eso
lu
tio
n
d
ep
th
s
as
well
as
s
h
allo
w
f
ea
tu
r
es,
en
r
ich
in
g
d
etailed
an
d
s
em
an
tic
s
eg
m
en
tatio
n
.
Ho
we
v
er
,
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
is
u
s
ed
i
n
th
e
m
u
lti
-
s
ca
le
p
o
o
lin
g
p
y
r
am
id
al
m
o
d
u
le,
wh
ich
allo
ws
f
o
r
a
b
etter
g
r
asp
o
f
co
n
tex
tu
al
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
361
-
3
7
3
366
in
f
o
r
m
atio
n
.
T
h
e
p
o
o
lin
g
m
o
d
u
le
ad
ap
ts
a
f
o
u
r
-
p
ar
allel
lay
e
r
s
tr
u
ctu
r
e,
alth
o
u
g
h
it
h
as
b
ee
n
ad
ju
s
ted
to
allo
w
f
ea
tu
r
e
m
a
p
s
at
d
if
f
e
r
en
t
s
ca
les.
Dy
n
am
ic
r
eg
i
o
n
-
awa
r
e
co
n
v
o
lu
tio
n
(
DR
C
o
n
v
)
is
ap
p
lied
u
n
d
er
ea
ch
p
o
o
lin
g
lay
er
.
T
h
e
m
u
lti
-
s
ca
le
p
o
o
lin
g
m
o
d
u
le
in
t
r
o
d
u
ce
s
th
e
co
n
ce
p
t
o
f
DR
C
o
n
v
,
wh
ich
is
b
ase
d
o
n
th
e
p
y
r
a
m
id
al
p
o
o
lin
g
m
o
d
u
les.
Py
r
am
id
al
p
o
o
lin
g
m
o
d
u
les
ca
n
u
s
e
co
n
tex
tu
al
in
f
o
r
m
atio
n
at
v
ar
io
u
s
s
ca
les,
an
d
DR
C
o
n
v
im
p
r
o
v
es
th
e
s
p
atial
in
ter
ac
ti
o
n
o
f
c
h
an
g
i
n
g
s
em
an
tic
d
ata
.
T
h
e
DR
C
o
n
v
s
ep
ar
ates
th
e
s
p
atial
d
im
en
s
io
n
s
in
to
n
u
m
er
o
u
s
r
e
g
io
n
s
u
s
in
g
t
h
e
lear
n
a
b
le
g
u
id
ed
m
ask
,
all
o
win
g
it
to
c
o
m
b
in
e
a
s
et
o
f
s
em
an
tically
s
im
ilar
f
ea
tu
r
es
in
to
a
s
in
g
le
r
eg
i
o
n
,
a
s
f
o
r
m
u
late
d
in
(
7
)
.
Similar
ly
,
th
e
m
u
lti
-
s
ca
le
p
y
r
am
id
al
p
o
o
lin
g
m
o
d
u
le
ad
d
ed
to
DR
C
o
n
v
ca
n
ef
f
ec
tiv
ely
s
im
u
late
s
em
an
tic
d
ata
f
r
o
m
m
u
ltip
le
lo
ca
tio
n
s
in
h
ig
h
-
lev
el
s
t
r
u
ctu
r
es,
wh
er
e
th
e
r
eg
io
n
ass
ig
n
m
en
t is d
eter
m
in
ed
b
y
th
e
g
u
id
ed
m
ask
d
ef
in
ed
in
(
8
)
,
h
e
n
ce
im
p
r
o
v
in
g
th
e
s
p
atial
in
ter
ac
tio
n
o
f
s
tr
u
ctu
r
es a
t d
if
f
er
en
t scale
s
.
,
,
=
∑
,
,
=
1
∗
,
(
)
(
,
)
∈
(
7
)
,
=
(
,
0
,
,
1
,
.
.
.
.
,
,
−
1
)
(
8
)
Her
e
d
ef
in
es
th
e
m
ask
=
0
,
.
.
.
,
−
1
,
wh
er
e
d
en
o
tes
th
e
r
eg
i
o
n
s
in
th
e
s
p
atial
d
im
en
s
io
n
,
b
ased
o
n
th
e
in
p
u
t
im
a
g
e
f
ea
t
u
r
e
e
x
tr
ac
tio
n
o
f
th
e
m
ask
is
p
er
f
o
r
m
ed
.
Similar
ly
,
s
em
an
tic
s
tr
u
ct
u
r
es
ar
e
g
r
o
u
p
ed
i
n
to
u
n
if
o
r
m
r
eg
i
o
n
s
.
A
s
ig
n
if
ica
n
t
m
ea
s
u
r
e
o
f
DR
C
o
n
v
,
th
e
m
ask
r
eg
u
lates
th
e
d
is
s
em
in
atio
n
o
f
co
n
v
o
lu
tio
n
al
k
er
n
els
in
th
e
s
p
atial
d
im
en
s
io
n
.
I
n
ea
ch
r
eg
io
n
,
o
n
ly
o
n
e
c
o
n
v
o
lu
ti
o
n
al
k
e
r
n
el
(
∈
[
0
,
−
1
]
)
is
s
h
ar
ed
.
T
h
e
co
n
v
o
lu
tio
n
al
k
e
r
n
els
ar
e
d
ef
in
ed
as
=
[
0
,
.
.
.
.
−
1
]
,
B
t
R
U
co
r
r
esp
o
n
d
in
g
t
o
th
e
r
eg
io
n
.
T
h
e
co
r
r
esp
o
n
d
in
g
co
n
v
o
l
u
tio
n
al
k
er
n
el,
wh
ich
co
n
v
o
lv
es
o
n
t
h
e
f
ea
tu
r
e
m
ap
,
is
alter
ed
ev
e
r
y
tim
e
b
ased
o
n
its
m
ask
.
Acc
o
r
d
in
g
to
th
e
s
h
a
r
ed
ar
ea
,
th
e
co
n
v
o
lu
tio
n
al
k
er
n
el
will
au
to
m
atica
lly
ad
ju
s
t
th
e
s
u
b
s
tan
tial
s
tr
u
ctu
r
es
o
f
th
e
in
p
u
t
im
ag
e.
At
th
e
en
d
o
f
th
e
m
o
d
u
le,
a
f
ea
tu
r
e
m
ap
is
im
p
lem
e
n
ted
u
s
in
g
b
ilin
ea
r
in
ter
p
o
latio
n
.
Fu
r
th
er
,
at
ea
c
h
lev
el,
f
ea
tu
r
es
f
o
llo
w
th
e
s
u
p
er
p
o
s
itio
n
o
p
er
atio
n
,
wh
i
ch
f
in
ally
p
r
o
v
i
d
es
o
u
tp
u
t
to
t
h
e
d
ec
o
d
er
.
T
h
er
ef
o
r
e,
a
m
u
lti
-
s
ca
le
p
y
r
a
m
id
al
m
o
d
u
le
allo
ws
b
etter
co
n
tex
tu
al
d
ata
ac
r
o
s
s
v
ar
io
u
s
r
eg
io
n
s
an
d
im
p
r
o
v
es
its
ab
ilit
y
to
s
eize
g
lo
b
al
d
ata.
T
h
er
e
f
o
r
e,
th
e
m
u
lti
-
s
ca
le
f
ea
tu
r
es
ac
cu
r
ately
id
en
tify
an
d
class
if
y
th
e
lan
d
c
o
v
er
ty
p
es e
f
f
ec
tiv
ely
.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
s
u
g
g
ested
m
eth
o
d
h
as
b
ee
n
r
elate
d
to
v
a
r
io
u
s
p
r
ev
ailin
g
m
eth
o
d
s
lik
e
U
-
Ne
t,
HR
Net,
Dee
p
L
ab
V3
,
R
esNet5
0
+D
ee
p
L
ab
V3
,
a
n
d
Xce
p
ti
o
n
+D
ee
p
L
ab
V3
.
An
al
y
tical
p
er
f
o
r
m
an
ce
o
f
th
is
m
eth
o
d
,
s
u
ch
as
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
al
l,
F
1
-
s
co
r
e,
Dice
s
co
r
e
,
in
ter
s
ec
tio
n
o
v
er
u
n
i
o
n
(
I
o
U)
,
Kap
p
a
s
co
r
e
,
an
d
er
r
o
r
m
etr
ics,
is
ev
alu
ated
a
s
well
as
co
m
p
ar
ed
with
r
ec
en
t
r
esear
ch
wo
r
k
s
.
T
h
e
s
u
g
g
ested
m
eth
o
d
is
im
p
lem
en
ted
in
MA
T
L
AB
R
2
0
2
1
b
,
u
tili
zin
g
W
in
d
o
ws 1
0
OS a
n
d
8
GB
R
AM
.
4
.
1
.
B
hu
v
a
n
s
a
t
ellit
e
im
a
g
e
da
t
a
s
et
des
cr
iptio
n
Fo
r
task
s
in
v
o
lv
in
g
th
e
an
aly
s
is
an
d
s
eg
m
en
tatio
n
o
f
lan
d
co
v
er
,
th
is
d
ataset
is
co
n
s
id
er
ed
a
v
alu
ab
le
r
eso
u
r
ce
[
2
8
]
.
A
s
et
o
f
s
atellite
p
h
o
to
s
with
a
h
ig
h
s
p
atial
r
eso
lu
tio
n
is
p
ar
t
o
f
th
e
d
ataset.
T
h
e
d
ataset
in
clu
d
es
s
atellite
2
D
im
ag
es
o
f
Var
a
n
asi,
a
city
lo
ca
ted
in
th
e
n
o
r
th
e
r
n
p
ar
t
o
f
I
n
d
ia,
i
n
th
e
s
tate
o
f
Uttar
Pra
d
esh
,
o
r
g
an
ized
alter
n
atin
g
f
r
o
m
2
5
.
3
° to
2
5
.
5
° N
latitu
d
e
as we
ll
as 8
3
°
to
8
3
.
2
°
E
lo
n
g
i
tu
d
e.
I
t c
o
n
s
is
ts
o
f
a
s
er
ies
o
f
h
ig
h
-
r
eso
l
u
tio
n
im
ag
es
o
f
th
e
E
ar
th
'
s
s
u
r
f
ac
e.
T
h
ese
im
ag
es
wer
e
ca
p
tu
r
ed
b
y
th
e
I
n
d
ian
r
em
o
te
s
en
s
in
g
s
atellite
(
I
R
S),
p
r
o
ce
s
s
ed
,
an
d
m
ad
e
ac
ce
s
s
ib
le
th
r
o
u
g
h
th
e
B
h
u
v
an
g
e
o
p
latf
o
r
m
,
wh
ich
is
o
p
er
ated
b
y
th
e
I
n
d
ian
s
p
ac
e
r
esear
ch
o
r
g
an
izatio
n
(
I
SR
O)
.
4
.
2
.
E
x
perim
ent
a
l
a
na
ly
s
is
Fig
u
r
e
4
d
ep
icts
an
ex
p
er
im
e
n
tal
in
v
esti
g
atio
n
o
f
t
h
e
s
u
g
g
ested
s
tr
ateg
y
u
s
in
g
th
e
B
h
u
v
an
s
atellite
im
ag
e
d
ataset
.
T
h
e
in
p
u
t
im
ag
e
is
g
iv
en
in
Fig
u
r
e
4
(
a)
.
Fig
u
r
e
4
(
b
)
s
h
o
ws
th
e
p
r
e
-
p
r
o
ce
s
s
ed
im
ag
e
ac
q
u
ir
ed
th
r
o
u
g
h
th
e
ap
p
licatio
n
o
f
an
Imp
-
GF
.
Fig
u
r
e
4
(
c)
s
h
o
ws
th
e
s
eg
m
en
ted
im
ag
e
th
at
th
e
s
u
g
g
ested
KD
-
Mu
ViT
Py
m
o
d
el
o
b
tain
s
.
4
.
3
.
P
er
f
o
r
m
a
nce
a
na
ly
s
is
in bo
t
h t
ra
ini
ng
a
nd
t
esting
T
r
ain
in
g
b
ased
p
er
f
o
r
m
an
ce
a
n
aly
s
is
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
,
ass
ess
ed
in
ter
m
s
o
f
th
e
m
et
r
ics
s
u
ch
as
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
a
s
well
as
F1
-
s
co
r
e
u
s
in
g
th
e
B
h
u
v
an
s
atellite
im
ag
e
d
ata
s
et,
is
d
ep
icted
in
Fig
u
r
e
5
.
T
h
e
p
er
f
o
r
m
an
ce
a
n
aly
s
is
o
f
th
e
r
ec
o
m
m
en
d
ed
m
o
d
el
u
s
in
g
th
e
B
h
u
v
an
s
atellite
im
ag
e
d
ataset
is
s
h
o
wn
in
Fig
u
r
e
6
.
T
h
is
an
aly
s
is
i
s
a
s
s
e
s
s
ed
in
ter
m
s
o
f
m
etr
ics
s
u
ch
as
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
F1
-
s
co
r
e,
Dice
s
co
r
e,
I
o
U,
an
d
Kap
p
a
s
co
r
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mu
lti
-
s
ca
le
fea
tu
r
es a
s
s
i
s
ted
kn
o
w
led
g
e
d
is
till
a
tio
n
visi
o
n
tr
a
n
s
fo
r
mer fo
r
la
n
d
…
(
S
u
ja
t
a
A
r
ju
n
Ga
ikw
a
d
)
367
(
a)
(
b
)
(
c)
Fig
u
r
e
4
.
E
x
p
er
im
e
n
tal
an
aly
s
is
u
s
in
g
th
e
B
h
u
v
an
s
atellite
d
ataset
in
ter
m
s
o
f
(
a)
o
r
ig
in
al
i
m
ag
e
,
(
b
)
p
r
e
-
p
r
o
ce
s
s
ed
im
ag
e
,
an
d
(
c)
s
eg
m
en
ted
o
u
tco
m
es
Fig
u
r
e
5
.
T
r
ain
in
g
b
ased
p
er
f
o
r
m
an
ce
an
aly
s
is
o
f
s
eg
m
e
n
tatio
n
u
s
in
g
B
h
u
v
an
Satellite
d
ata
s
et
T
h
e
s
eg
m
en
tatio
n
ac
cu
r
ac
y
e
v
alu
ated
b
y
e
x
is
tin
g
m
eth
o
d
s
,
s
u
ch
as
U
-
Net
,
HR
Net,
Dee
p
L
ab
V3
,
R
esNet5
0
+D
ee
p
L
ab
V3
,
Xce
p
tin
+D
ee
p
L
ab
V3
,
ac
h
iev
es
v
alu
es
o
f
8
4
.
3
5
7
2
7
%,
8
7
.
1
8
5
3
4
%,
9
3
.
5
0
8
0
9
%,
9
6
.
2
9
5
7
%,
an
d
9
7
.
8
2
0
9
%,
r
e
s
p
ec
tiv
ely
,
as
d
ep
icted
in
Fi
g
u
r
e
6
(
a)
.
Similar
ly
,
t
h
e
p
r
ec
is
io
n
ev
alu
ated
b
y
ex
is
tin
g
m
eth
o
d
s
is
illu
s
tr
ated
in
Fig
u
r
e
6
(
b
)
,
w
h
ich
ac
h
ie
v
e
d
v
alu
es
o
f
8
3
.
9
5
3
2
6
%,
8
7
.
8
2
1
6
5
%,
9
3
.
6
0
9
0
9
%,
9
6
.
8
5
1
2
7
%,
a
n
d
9
7
.
9
9
2
6
%,
r
esp
ec
tiv
ely
.
T
h
e
r
ec
all
w
as
ass
e
s
s
ed
u
s
in
g
ex
is
tin
g
m
eth
o
d
s
,
s
u
ch
as
8
2
.
7
1
0
9
3
%,
8
5
.
8
6
2
2
%,
9
3
.
1
5
4
5
8
%,
9
6
.
0
7
3
5
5
%,
an
d
9
7
.
0
8
3
5
8
%,
r
esp
ec
tiv
ely
.
T
h
e
F1
-
s
co
r
e
ev
alu
ated
b
y
th
e
s
u
g
g
ested
m
eth
o
d
is
9
9
.
2
2
3
9
7
%,
as
d
ep
icted
in
Fig
u
r
es
6
(
c)
an
d
6
(
d
)
.
T
h
e
Dice
s
co
r
e
was
ev
alu
ated
b
y
ex
is
tin
g
m
eth
o
d
s
,
s
u
c
h
as
8
1
.
5
9
%,
8
5
.
9
8
%,
9
2
.
3
1
%,
9
5
.
1
3
%,
an
d
9
6
.
0
1
%,
r
esp
ec
tiv
el
y
.
T
h
e
Dice
s
co
r
e
ass
es
s
ed
b
y
th
e
s
u
g
g
ested
m
eth
o
d
is
9
8
.
2
3
%,
as
s
h
o
wn
in
Fig
u
r
e
6
(
e)
.
T
h
e
I
o
U
ev
a
lu
ated
b
y
ex
is
tin
g
m
eth
o
d
s
is
8
0
.
0
2
%,
8
3
.
9
9
%,
8
8
.
2
5
%,
9
2
.
6
8
%,
a
n
d
9
4
.
9
9
%.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
o
b
tain
ed
a
n
I
o
U
o
f
9
6
.
5
5
%,
wh
ich
is
s
h
o
wn
i
n
F
ig
u
r
e
6
(
f
)
.
T
h
e
Kap
p
a
s
co
r
e
was
ev
alu
ated
b
y
e
x
is
tin
g
m
eth
o
d
s
at
v
alu
es
o
f
7
9
.
2
5
%,
8
2
.
7
5
%,
8
7
.
8
8
%,
9
1
.
3
5
%,
an
d
9
3
.
8
6
%.
T
h
e
s
u
g
g
e
s
ted
m
eth
o
d
o
b
tain
e
d
a
Kap
p
a
s
co
r
e
o
f
9
5
.
9
1
%,
wh
ich
is
d
ep
icted
in
Fig
u
r
e
6
(
g
)
.
T
h
e
s
u
g
g
ested
m
et
h
o
d
p
r
o
d
u
ce
d
th
e
b
est
r
esu
lts
wh
en
co
m
p
ar
ed
to
o
th
e
r
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
361
-
3
7
3
368
tech
n
iq
u
es.
Fig
u
r
e
7
s
h
o
ws
th
e
p
er
f
o
r
m
an
ce
an
aly
s
is
o
f
th
e
r
ec
o
m
m
e
n
d
ed
tech
n
iq
u
e
ac
r
o
s
s
i
n
cl
u
d
i
n
g
m
an
y
class
es;
Fi
g
u
r
e
7
(
a)
f
o
r
u
r
b
a
n
,
Fi
g
u
r
e
7
(
b
)
f
o
r
w
ate
r
,
Fi
g
u
r
e
7
(
c
)
f
o
r
f
o
r
est
,
Fi
g
u
r
e
7
(
d
)
f
o
r
a
g
r
icu
ltu
r
e,
an
d
Fig
u
r
e
7
(
e)
f
o
r
r
o
ad
;
u
s
in
g
th
e
B
h
u
v
an
Satellite im
ag
e
d
ataset.
(
a)
(
b
)
(
c)
(
d
)
(
e)
(f)
(
g
)
Fig
u
r
e
6
.
T
esti
n
g
b
ased
p
er
f
o
r
m
an
ce
an
aly
s
is
o
f
s
eg
m
e
n
tatio
n
u
s
in
g
B
h
u
v
a
n
s
atellite
d
ataset
in
ter
m
s
o
f
(
a)
ac
cu
r
ac
y
,
(
b
)
p
r
ec
is
io
n
,
(
c)
r
ec
all
,
(
d
)
F1
-
s
co
r
e,
(
e
)
Dice
s
co
r
e,
(
f
)
I
o
U,
an
d
(
g
)
Ka
p
p
a
s
co
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mu
lti
-
s
ca
le
fea
tu
r
es a
s
s
i
s
ted
kn
o
w
led
g
e
d
is
till
a
tio
n
visi
o
n
tr
a
n
s
fo
r
mer fo
r
la
n
d
…
(
S
u
ja
t
a
A
r
ju
n
Ga
ikw
a
d
)
369
(
a)
(
b
)
(
c)
(
d
)
(
e)
Fig
u
r
e
7
.
L
a
n
d
c
o
v
er
class
if
icatio
n
ac
cu
r
ac
y
of
(
a)
u
r
b
an
,
(
b
)
wate
r
,
(
c)
f
o
r
est
,
(
d
)
ag
r
ic
u
ltu
r
e
,
an
d
(
e)
r
o
ad
4
.
4
.
Co
m
pa
ra
t
iv
e
a
na
ly
s
is
I
n
th
is
an
aly
s
is
,
th
e
s
u
g
g
ested
m
o
d
el
is
r
elate
d
to
co
n
v
en
tio
n
al
m
o
d
els
s
u
ch
as
U
-
Net
,
HR
Net,
Dee
p
L
ab
V3
,
R
esNet5
0
+D
ee
p
L
ab
V3
,
an
d
Xce
p
tin
+D
ee
p
L
a
b
V3
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
h
as
o
b
tain
ed
a
tr
ain
in
g
p
h
ase
ac
cu
r
ac
y
o
f
9
9
%
with
3
0
0
ep
o
c
h
s
an
d
8
0
%
o
f
th
e
t
r
ain
in
g
d
ata.
T
h
e
s
u
g
g
ested
m
o
d
el
h
as
a
tr
ain
in
g
lo
s
s
v
alu
e
o
f
less
th
an
1
with
3
0
0
ep
o
c
h
s
,
with
8
0
% o
f
th
e
tr
ain
in
g
d
ata.
T
h
e
s
u
g
g
ested
m
o
d
el
ac
h
iev
ed
b
ette
r
o
u
tco
m
es
co
m
p
ar
ed
with
o
th
e
r
ex
is
tin
g
m
o
d
els.
T
h
e
s
u
g
g
es
ted
m
o
d
el
h
as
o
b
tain
e
d
a
test
i
n
g
p
h
ase
ac
cu
r
ac
y
o
f
9
8
%
with
3
0
0
ep
o
ch
s
.
T
h
e
s
u
g
g
ested
m
o
d
el
h
as
a
test
in
g
lo
s
s
v
alu
e
o
f
less
th
an
1
wi
th
3
0
0
ep
o
ch
s
.
T
h
e
s
u
g
g
ested
m
o
d
el
ac
h
iev
e
d
b
et
ter
o
u
tco
m
es
c
o
m
p
a
r
ed
with
o
th
er
ex
is
tin
g
m
eth
o
d
s
.
B
o
th
tr
ain
in
g
as
well
as
test
in
g
ac
cu
r
ac
y
an
d
lo
s
s
ar
e
s
h
o
wn
in
Fig
u
r
e
8
.
Sp
ec
if
ically
,
Fig
u
r
e
8
(
a
)
illu
s
tr
ates
th
e
tr
ain
in
g
ac
c
u
r
ac
y
,
Fig
u
r
e
8
(
b
)
s
h
o
ws
th
e
tr
ain
i
n
g
lo
s
s
,
Fig
u
r
e
8
(
c)
p
r
esen
ts
t
h
e
test
in
g
ac
cu
r
ac
y
,
an
d
Fig
u
r
e
8
(
d
)
d
ep
icts
th
e
test
in
g
lo
s
s
.
Fig
u
r
e
8
s
h
o
ws
th
e
tr
ain
in
g
-
b
ased
p
e
r
f
o
r
m
an
ce
an
aly
s
is
f
o
r
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
as
well
as
th
e
F1
-
s
co
r
e.
W
h
en
co
m
p
a
r
e
d
to
cu
r
r
en
t
m
o
d
els,
th
e
s
u
g
g
ested
m
o
d
el
attain
ed
an
ac
cu
r
ac
y
o
f
9
9
.
0
0
%.
Similar
ly
,
an
aly
s
es
o
f
p
r
ec
is
io
n
,
r
ec
all,
as
well
as
F1
-
s
co
r
e
r
ev
ea
led
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
v
alu
es
o
f
9
9
.
2
8
%,
9
8
.
4
7
%,
a
n
d
9
9
.
2
2
%,
r
esp
ec
tiv
ely
.
T
ab
le
1
s
h
o
ws a
co
m
p
ar
i
s
o
n
o
f
th
e
e
x
is
tin
g
an
d
p
r
o
p
o
s
ed
m
o
d
els.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
361
-
3
7
3
370
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
8
.
T
r
ain
in
g
a
n
d
test
in
g
ac
cu
r
ac
y
as we
ll a
s
lo
s
s
u
s
in
g
B
h
u
v
an
s
atellite
d
ataset
of
(
a)
tr
ain
in
g
ac
cu
r
ac
y
,
(
b
)
tr
ain
in
g
lo
s
s
,
(
c)
test
in
g
ac
cu
r
ac
y
,
a
n
d
(
d
)
test
in
g
lo
s
s
T
ab
le
1
.
T
r
ai
n
in
g
b
ased
co
m
p
ar
ativ
e
an
aly
s
is
o
f
s
eg
m
en
tati
o
n
u
s
in
g
B
h
u
v
an
s
atellite
d
ataset
M
e
t
r
i
c
s
U
-
N
e
t
H
R
N
e
t
D
e
e
p
La
b
V
3
R
e
s
n
e
t
5
0
+
D
e
e
p
La
b
V
3
X
c
e
p
t
i
n
+
D
e
e
p
La
b
V
3
P
r
o
p
o
se
d
m
o
d
e
l
A
c
c
u
r
a
c
y
(
%)
8
4
.
3
5
7
2
7
8
7
.
1
8
5
3
4
9
3
.
5
0
8
0
9
9
6
.
2
9
5
7
6
9
7
.
8
2
0
9
9
9
.
0
0
0
6
4
P
r
e
c
i
s
i
o
n
(
%)
8
3
.
9
5
3
2
6
8
7
.
8
2
1
6
5
9
3
.
6
0
9
0
9
9
6
.
8
5
1
2
7
9
7
.
9
9
2
6
9
9
.
2
8
6
7
2
R
e
c
a
l
l
(
%)
8
2
.
7
1
0
9
3
8
5
.
8
6
2
2
8
9
.
8
4
1
7
9
4
.
9
3
2
2
3
9
6
.
3
3
6
1
6
9
8
.
4
7
4
8
9
F1
-
sc
o
r
e
(
%)
8
2
.
7
1
0
9
3
8
6
.
7
5
1
0
3
9
3
.
1
5
4
5
8
9
6
.
0
7
3
5
5
9
7
.
0
8
3
5
8
9
9
.
2
2
3
9
7
T
esti
n
g
b
ased
co
m
p
ar
ativ
e
a
n
aly
s
is
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
u
s
in
g
B
h
u
v
a
n
s
atellite
im
ag
e
d
ata
is
s
h
o
wn
in
T
ab
le
2
.
T
h
e
s
u
g
g
ested
m
o
d
el
attain
ed
th
e
m
a
x
im
al
ac
cu
r
ac
y
o
f
9
8
.
0
1
%,
wh
ich
is
1
4
.
4
9
%,
1
1
.
6
9
%,
5
.
4
3
%,
2
.
6
7
%,
an
d
1
.
1
6
%
s
u
p
er
io
r
to
th
e
ex
is
tin
g
U
-
Net
,
HR
Net,
Dee
p
L
ab
V3
,
R
esNet5
0
+D
ee
p
L
ab
V3
,
an
d
Xce
p
tin
+D
ee
p
L
ab
V3
m
o
d
els
.
Similar
ly
,
t
h
e
s
u
g
g
ested
m
o
d
el
o
b
tain
ed
th
e
p
r
ec
is
io
n
v
alu
e
o
f
9
8
.
9
9
%,
wh
ich
is
1
5
.
8
7
%,
1
2
.
0
4
%,
6
.
3
1
%,
3
.
1
%,
an
d
1
.
9
7
%
s
u
p
er
i
o
r
to
th
e
tr
ad
itio
n
a
l
m
o
d
els.
T
h
e
s
u
g
g
ested
m
o
d
el
o
b
tain
ed
a
r
ec
all
v
alu
e
o
f
9
7
.
4
9
%,
wh
ich
is
1
5
.
6
%,
1
2
.
4
8
%,
8
.
5
4
%,
3
.
5
%,
an
d
2
.
1
1
%
s
u
p
er
i
o
r
to
th
e
ex
is
tin
g
m
o
d
els.
T
h
e
s
u
g
g
ested
m
o
d
el
o
b
tain
e
d
an
F1
-
s
co
r
e
o
f
9
8
.
2
3
%,
wh
ich
is
1
6
.
3
4
%,
1
2
.
3
4
%,
6
.
0
0
%,
3
.
1
1
%,
an
d
2
.
1
1
%
s
u
p
er
io
r
to
t
h
e
ex
is
tin
g
m
o
d
els.
Ho
wev
er
,
t
h
e
p
r
o
p
o
s
ed
m
o
d
el
o
b
tain
ed
a
Dice
s
co
r
e
v
alu
e
o
f
9
8
.
2
3
%,
wh
ich
is
1
6
.
6
4
%,
1
2
.
2
5
%,
5
.
9
2
%,
3
.
1
%,
an
d
2
.
2
2
%
m
o
r
e
th
an
th
e
ex
is
tin
g
m
o
d
els.
T
h
e
p
r
o
p
o
s
e
d
m
o
d
el
o
b
tain
ed
an
I
o
U
v
al
u
e
o
f
9
6
.
5
5
%,
wh
ich
is
1
6
.
5
3
%,
1
2
.
5
6
%,
8
.
3
%,
3
.
8
7
%,
an
d
1
.
5
6
%
s
u
p
er
i
o
r
to
th
e
co
n
v
en
tio
n
al
m
o
d
el.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
o
b
tain
e
d
a
K
ap
p
a
s
co
r
e
v
alu
e
o
f
9
5
.
9
1
%,
wh
ich
is
1
6
.
6
6
%,
1
3
.
1
6
%,
8
.
0
3
%,
4
.
5
6
%,
an
d
2
.
0
5
% b
etter
th
an
th
e
c
u
r
r
en
t
m
o
d
els.
T
ab
le
2
.
T
esti
n
g
b
ased
co
m
p
a
r
ativ
e
an
aly
s
is
o
f
s
eg
m
en
tatio
n
u
s
in
g
B
h
u
v
a
n
s
atellite
d
ataset
[
2
9
]
M
e
t
r
i
c
s
U
-
N
e
t
H
R
N
e
t
D
e
e
p
La
b
V
3
R
e
s
n
e
t
5
0
+
D
e
e
p
La
b
V
3
X
c
e
p
t
i
n
+
D
e
e
p
La
b
V
3
P
r
o
p
o
se
d
m
o
d
e
l
A
c
c
u
r
a
c
y
(
%)
8
3
.
5
2
8
6
.
3
2
9
2
.
5
8
9
5
.
3
4
9
6
.
8
5
9
8
.
0
1
P
r
e
c
i
s
i
o
n
(
%)
8
3
.
1
2
8
6
.
9
5
9
2
.
6
8
9
5
.
8
9
9
7
.
0
2
9
8
.
9
9
R
e
c
a
l
l
(
%)
8
1
.
8
9
8
5
.
0
1
8
8
.
9
5
9
3
.
9
9
9
5
.
3
8
9
7
.
4
9
F1
-
sc
o
r
e
(
%)
8
1
.
8
9
8
5
.
8
9
9
2
.
2
3
9
5
.
1
2
9
6
.
1
2
9
8
.
2
3
D
i
c
e
sc
o
r
e
(
%)
8
1
.
5
9
8
5
.
9
8
9
2
.
3
1
9
5
.
1
3
9
6
.
0
1
9
8
.
2
3
I
o
U
(
%)
8
0
.
0
2
8
3
.
9
9
8
8
.
2
5
9
2
.
6
8
9
4
.
9
9
9
6
.
5
5
K
a
p
p
a
s
c
o
r
e
(
%)
7
9
.
2
5
8
2
.
7
5
8
7
.
8
8
9
1
.
3
5
9
3
.
8
6
9
5
.
9
1
Evaluation Warning : The document was created with Spire.PDF for Python.