I
nte
rna
t
io
na
l J
o
urna
l o
f
I
nfo
rm
a
t
ics a
nd
Co
m
m
un
ica
t
io
n T
ec
hn
o
lo
g
y
(
I
J
-
I
CT
)
Vo
l.
14
,
No
.
3
,
Dec
em
b
er
20
25
,
p
p
.
9
0
3
~
9
1
3
I
SS
N:
2252
-
8
7
7
6
,
DOI
:
1
0
.
1
1
5
9
1
/iji
ct
.
v14
i
3
.
pp
903
-
9
1
3
903
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ict.
ia
esco
r
e.
co
m
A
h
y
brid app
ro
a
ch using
V
G
G
1
6
-
E
ff
cien
tNe
tV2B
3
-
F
CNets
for
a
ccurate indo
o
r v
s o
utdoo
r and a
ni
ma
ted
v
s na
tural
ima
g
e
cla
ss
ificatio
n
M
eg
ha
na
Deshm
uk
h,
Am
it
G
a
ik
wa
d,
Sn
eha
l K
uche
D
e
p
a
r
t
me
n
t
o
f
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
a
n
d
En
g
i
n
e
e
r
i
n
g
,
G
.
H
.
R
a
i
so
n
i
U
n
i
v
e
r
si
t
y
A
mr
a
v
a
t
i
(
M
a
h
a
r
a
sh
t
r
a
)
,
M
a
h
a
r
a
s
h
t
r
a
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Au
g
2
7
,
2
0
2
4
R
ev
is
ed
Ma
r
5
,
2
0
2
5
Acc
ep
ted
J
u
n
9
,
2
0
2
5
Th
e
p
a
p
e
r
i
n
tro
d
u
c
e
s
a
h
y
b
ri
d
a
p
p
r
o
a
c
h
th
a
t
sy
n
e
rg
ist
ica
ll
y
c
o
m
b
in
e
s
th
e
stre
n
g
th
s
o
f
VG
G
1
6
,
Eff
icie
n
t
Ne
tV2
B3
,
a
n
d
fu
ll
y
c
o
n
n
e
c
ted
n
e
two
rk
s
(F
CNe
ts)
to
a
c
h
iev
e
p
re
c
ise
ima
g
e
c
las
sifica
ti
o
n
.
S
p
e
c
ifi
c
a
ll
y
,
o
u
r
fo
c
u
s
li
e
s
in
d
isc
e
rn
in
g
b
e
twe
e
n
b
a
sic
i
n
d
o
o
r
a
n
d
o
u
t
d
o
o
r
sc
e
n
e
s,
f
u
rth
e
r
e
x
ten
d
e
d
to
d
isti
n
g
u
is
h
b
e
twe
e
n
a
n
ima
ted
a
n
d
n
a
tu
ra
l
ima
g
e
s.
Ou
r
p
r
o
p
o
s
e
d
h
y
b
ri
d
a
rc
h
it
e
c
tu
re
h
a
rn
e
ss
e
s
th
e
u
n
iq
u
e
c
h
a
ra
c
teristics
o
f
e
a
c
h
c
o
m
p
o
n
e
n
t
t
o
sig
n
ifi
c
a
n
t
ly
e
n
h
a
n
c
e
th
e
m
o
d
e
l
’
s
o
v
e
ra
ll
p
e
rfo
rm
a
n
c
e
in
fi
n
e
-
g
ra
i
n
e
d
ima
g
e
ca
teg
o
riza
ti
o
n
.
I
n
o
u
r
m
e
t
h
o
d
o
l
o
g
y
,
we
u
t
il
ize
VG
G
1
6
a
n
d
Eff
icie
n
tNe
tV2
B3
a
s
t
h
e
fe
a
tu
re
e
x
trac
to
rs.
Du
ri
n
g
e
v
a
lu
a
ti
o
n
,
we
e
x
a
m
in
e
d
v
a
rio
u
s
c
las
sifica
ti
o
n
a
l
g
o
rit
h
m
s,
su
c
h
a
s
VG
G
1
6
,
Eff
i
c
ien
tNe
t,
F
e
a
tu
re
_
Ag
g
r
_
Av
g
,
a
n
d
F
e
a
tu
re
_
Ag
g
r
_
m
a
x
,
a
m
o
n
g
o
th
e
rs.
N
o
t
a
b
ly
,
o
u
r
h
y
b
rid
fe
a
tu
re
a
g
g
re
g
a
ti
o
n
a
p
p
ro
a
c
h
d
e
m
o
n
stra
tes
a
re
m
a
rk
a
b
le
imp
ro
v
e
m
e
n
t
o
f
0
.
5
%
in
a
c
c
u
ra
c
y
c
o
m
p
a
re
d
to
e
x
isti
n
g
so
l
u
ti
o
n
s
e
m
p
lo
y
i
n
g
VG
G
1
6
a
n
d
Eff
icie
n
tNe
t
a
s
fe
a
tu
re
e
x
trac
to
rs.
No
ta
b
ly
,
fo
r
i
n
d
o
o
r
v
e
rsu
s
o
u
t
d
o
o
r
ima
g
e
c
las
sifica
ti
o
n
,
f
e
a
tu
re
_
a
g
g
r
_
a
v
g
a
c
h
ie
v
e
s
a
n
a
c
c
u
ra
c
y
o
f
9
8
.
5
1
%
.
S
imilarly
,
w
h
e
n
d
isti
n
g
u
ish
in
g
b
e
twe
e
n
a
n
ima
ted
a
n
d
n
a
tu
ra
l
ima
g
e
s,
F
e
a
tu
re
_
Ag
g
r_
Av
g
a
c
h
ie
v
e
s
a
n
imp
re
ss
iv
e
a
c
c
u
ra
c
y
o
f
9
9
.
2
0
%
.
Ou
r
fin
d
in
g
s d
e
m
o
n
stra
te i
m
p
ro
v
e
d
a
c
c
u
ra
c
y
with
th
e
h
y
b
rid
m
o
d
e
l,
p
ro
v
in
g
it
s
a
d
a
p
ta
b
il
it
y
a
c
ro
ss
d
i
v
e
rse
c
l
a
ss
ifi
c
a
ti
o
n
tas
k
s.
T
h
e
m
o
d
e
l
is
p
ro
m
isin
g
fo
r
a
p
p
li
c
a
ti
o
n
s
li
k
e
a
u
to
m
a
ted
s
u
rv
e
il
lan
c
e
,
c
o
n
ten
t
fil
teri
n
g
,
a
n
d
in
telli
g
e
n
t
v
isu
a
l
re
c
o
g
n
it
io
n
,
with
it
s
ro
b
u
st
n
e
ss
a
n
d
p
re
c
isio
n
m
a
k
in
g
i
t
i
d
e
a
l
fo
r
re
a
l
-
wo
rld
sc
e
n
a
rio
s re
q
u
ir
i
n
g
n
u
a
n
c
e
d
c
a
teg
o
riza
ti
o
n
.
K
ey
w
o
r
d
s
:
C
NN
E
f
f
icien
tNetV2
B
3
FC
Net
Featu
r
e
ag
g
r
e
g
atio
n
Scen
e
class
if
icatio
n
VGG1
6
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
:
Me
g
h
an
a
Desh
m
u
k
h
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
E
n
g
in
ee
r
in
g
,
G.
H.
R
aiso
n
i U
n
iv
er
s
ity
Am
r
av
ati
(
M
ah
ar
ash
tr
a)
Ma
h
ar
ash
tr
a,
I
n
d
ia
E
m
ail:
m
eg
h
an
a
d
esh
m
u
k
h
4
4
0
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
I
n
th
e
e
r
a
o
f
ev
e
r
-
ex
p
a
n
d
in
g
d
ig
ital
co
n
te
n
t,
im
ag
e
class
i
f
icatio
n
p
lay
s
a
p
iv
o
tal
r
o
le
in
v
ar
io
u
s
ap
p
licatio
n
s
r
an
g
i
n
g
f
r
o
m
au
to
m
ated
s
u
r
v
eillan
ce
t
o
co
n
ten
t
r
ec
o
m
m
e
n
d
atio
n
s
y
s
tem
s
.
Acc
u
r
ate
class
if
icatio
n
o
f
im
ag
es
in
to
r
elev
an
t
ca
teg
o
r
ies
is
cr
u
cial
f
o
r
o
p
tim
izin
g
t
h
e
p
er
f
o
r
m
a
n
c
e
o
f
th
ese
s
y
s
tem
s
.
T
h
is
r
esear
ch
p
ap
er
in
tr
o
d
u
ce
s
a
n
o
v
el
h
y
b
r
id
ap
p
r
o
ac
h
th
a
t
co
m
b
in
es
th
e
s
tr
en
g
th
s
o
f
th
r
ee
s
tate
-
of
-
th
e
-
ar
t
d
ee
p
lear
n
in
g
ar
c
h
itectu
r
es:
VGG1
6
,
E
f
f
icien
tNetV2
B
3
,
an
d
FC
Nets.
T
h
e
p
r
o
p
o
s
ed
h
y
b
r
id
m
o
d
el
aim
s
to
i
n
cr
ea
s
e
th
e
ac
cu
r
ac
y
o
f
im
a
g
e
ca
teg
o
r
izatio
n
,
s
p
ec
if
ically
f
o
cu
s
in
g
o
n
d
is
tin
g
u
is
h
in
g
b
etwe
en
in
d
o
o
r
an
d
o
u
td
o
o
r
s
ce
n
es,
as
well
as
d
is
cr
im
in
atin
g
b
etwe
en
an
im
ate
d
an
d
n
atu
r
al
im
ag
es
,
T
h
e
au
th
o
r
s
p
r
o
p
o
s
e
a
d
ee
p
f
u
n
ctio
n
ag
g
r
eg
atio
n
f
r
a
m
ewo
r
k
c
o
m
b
in
i
n
g
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs)
f
o
r
f
u
n
ctio
n
ex
tr
ac
tio
n
an
d
g
r
a
p
h
c
o
n
v
o
lu
tio
n
al
n
etw
o
r
k
s
(
GC
Ns)
to
s
eize
s
p
atial
r
elatio
n
s
h
ip
s
am
o
n
g
f
u
n
ctio
n
s
f
o
r
s
tep
p
ed
f
o
r
war
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
14
,
No
.
3
,
Dec
em
b
er
20
25
:
903
-
9
1
3
904
s
ce
n
e
ty
p
e
o
f
d
ee
p
f
ea
tu
r
e
ag
g
r
e
g
atio
n
f
r
am
ew
o
r
k
d
r
i
v
en
b
y
g
r
ap
h
c
o
n
v
o
lu
tio
n
al
n
etwo
r
k
f
o
r
s
ce
n
e
class
if
icatio
n
in
r
em
o
te
s
en
s
in
g
an
d
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
a
ch
iev
es
o
v
er
all
p
e
r
f
o
r
m
an
ce
i
n
f
ar
f
l
u
n
g
s
en
s
in
g
s
ce
n
e
ty
p
e
d
u
ties
,
o
u
tp
er
f
o
r
m
i
n
g
ex
is
tin
g
m
eth
o
d
s
in
ac
c
u
r
a
cy
o
n
b
en
ch
m
a
r
k
d
atasets
[
1
]
,
[
2
]
.
T
h
e
ch
o
ice
o
f
in
d
o
o
r
Vs
o
u
td
o
o
r
an
d
an
im
ated
Vs
n
atu
r
al
i
m
ag
e
class
if
icatio
n
[
3
]
is
m
o
t
iv
ated
by
th
eir
wid
esp
r
ea
d
ap
p
licatio
n
s
.
R
ec
o
g
n
izin
g
wh
eth
er
an
im
a
g
e
is
tak
en
in
d
o
o
r
s
o
r
o
u
td
o
o
r
s
is
ess
en
tial
f
o
r
task
s
s
u
ch
as
in
tellig
en
t
c
am
er
a
s
y
s
tem
s
,
s
ec
u
r
ity
s
u
r
v
eillan
ce
,
an
d
au
g
m
en
ted
r
ea
lity
ap
p
li
ca
tio
n
s
.
Similar
ly
,
d
if
f
er
en
tiatin
g
b
etwe
en
an
im
ated
an
d
n
atu
r
al
im
ag
es
is
cr
u
cial
f
o
r
co
n
ten
t
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
,
en
ter
tain
m
en
t
p
latf
o
r
m
s
,
an
d
g
am
in
g
in
d
u
s
tr
ies.
T
h
e
h
y
b
r
id
ap
p
r
o
ac
h
p
r
esen
te
d
in
th
is
r
es
ea
r
ch
lev
er
a
g
es
th
e
d
is
tin
ctiv
e
f
ea
tu
r
es
an
d
ca
p
ab
ilit
ies
o
f
ea
ch
co
m
p
o
n
en
t
a
r
c
h
itectu
r
e.
VGG1
6
,
k
n
o
wn
f
o
r
its
s
im
p
lici
ty
an
d
ef
f
ec
tiv
en
ess
,
f
o
r
m
s
th
e
b
ac
k
b
o
n
e
o
f
th
e
m
o
d
el,
p
r
o
v
i
d
in
g
a
s
o
lid
f
o
u
n
d
atio
n
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
.
E
f
f
icien
tNetV2
B
3
,
an
ad
v
an
c
ed
an
d
ef
f
icien
t
ar
c
h
itectu
r
e
d
esig
n
ed
f
o
r
o
p
tim
al
p
er
f
o
r
m
an
ce
,
co
n
tr
ib
u
tes
to
im
p
r
o
v
e
d
f
ea
tu
r
e
r
e
p
r
esen
t
atio
n
an
d
co
m
p
u
tatio
n
al
e
f
f
icien
cy
.
FC
Nets,
f
o
cu
s
in
g
o
n
f
in
e
-
g
r
ain
ed
class
if
icatio
n
,
ar
e
in
co
r
p
o
r
ate
d
to
in
cr
ea
s
e
th
e
m
o
d
el
’
s
ab
ilit
y
to
d
is
cr
im
in
ate
in
-
b
etwe
en
s
u
b
tle
d
if
f
er
en
ce
s
in
im
ag
e
ca
teg
o
r
ies.
T
h
e
s
y
n
er
g
y
b
etwe
en
th
ese
a
r
ch
itectu
r
es
cr
ea
tes
a
r
o
b
u
s
t
an
d
ac
cu
r
ate
im
ag
e
class
if
icatio
n
m
o
d
el
th
at
ad
d
r
ess
es
th
e
ch
allen
g
es
p
o
s
ed
b
y
th
e
co
m
p
le
x
n
atu
r
e
o
f
in
d
o
o
r
Vs
o
u
td
o
o
r
an
d
a
n
im
ated
Vs
n
at
u
r
al
im
ag
e
ca
teg
o
r
izatio
n
.
B
y
c
o
m
b
in
in
g
t
h
e
p
o
wer
o
f
th
ese
ar
c
h
itectu
r
es,
we
aim
to
ac
h
iev
e
a
m
o
d
el
t
h
at
ex
ce
ls
in
g
en
e
r
aliza
tio
n
,
a
d
ap
tab
ilit
y
.
T
h
is
r
esear
ch
c
o
n
tr
ib
u
tes
to
th
e
g
r
o
win
g
f
ield
o
f
im
a
g
e
class
if
icatio
n
b
y
p
r
esen
tin
g
a
h
y
b
r
id
ap
p
r
o
ac
h
th
at
n
o
t
o
n
ly
o
u
t
p
er
f
o
r
m
s
in
d
iv
id
u
al
ar
c
h
itectu
r
es
b
u
t
also
estab
lis
h
es
a
f
o
u
n
d
atio
n
f
o
r
ex
p
lo
r
in
g
s
y
n
e
r
g
ies
b
etwe
en
d
if
f
e
r
en
t
d
ee
p
l
ea
r
n
in
g
m
o
d
els.
T
h
e
r
em
ain
d
er
o
f
th
is
p
ap
er
will
d
elv
e
in
to
th
e
m
eth
o
d
o
lo
g
y
,
ex
p
er
im
en
tal
s
etu
p
,
an
d
r
esu
l
ts
to
v
alid
ate
th
e
ef
f
ec
t
iv
en
ess
o
f
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
m
o
d
el
f
o
r
ac
cu
r
ate
im
ag
e
class
if
icatio
n
in
th
e
s
p
ec
if
ied
d
o
m
ain
s
[
4
]
,
[
5
]
.
C
ateg
o
r
izin
g
i
n
d
o
o
r
v
s
.
o
u
td
o
o
r
s
s
n
ap
s
h
o
ts
en
h
an
ce
s
co
n
s
u
m
e
r
e
n
jo
y
in
p
ac
k
a
g
es
lik
e
p
h
o
to
c
o
n
tr
o
l,
p
e
r
s
o
n
alis
ed
s
u
g
g
esti
o
n
s
,
an
d
s
m
ar
t
d
o
m
esti
c
s
y
s
te
m
s
,
im
p
r
o
v
in
g
alg
o
r
ith
m
s
,
d
ataset
p
leasan
t,
an
d
m
o
d
el
p
er
f
o
r
m
a
n
ce
ac
r
o
s
s
v
ar
io
u
s
d
o
m
ain
n
am
es.
R
ec
en
t
liter
atu
r
e
r
e
v
iews
d
ee
p
lear
n
i
n
g
f
o
r
s
ce
n
e
class
if
icatio
n
an
d
r
e
m
o
te
s
en
s
in
g
.
I
n
r
e
ce
n
t
y
ea
r
s
,
th
er
e
h
av
e
b
ee
n
s
u
b
s
tan
tial
ad
d
itio
n
s
to
th
e
liter
atu
r
e
o
n
th
e
to
p
ic
o
f
u
tili
s
in
g
d
ee
p
lear
n
in
g
m
o
d
els
to
ca
teg
o
r
is
e
p
h
o
to
g
r
ap
h
s
as
eith
er
in
d
o
o
r
s
o
r
o
u
ts
id
e,
o
r
as
an
i
m
ated
o
r
n
at
u
r
al.
T
h
is
liter
atu
r
e
r
ev
iew
co
m
p
iles
an
d
s
u
m
m
ar
is
es
th
e
r
e
s
u
lts
o
f
r
esear
c
h
th
at
h
as
lo
o
k
ed
at
s
ce
n
e
class
if
icatio
n
,
m
o
d
e
l
ar
ch
itectu
r
es,
a
n
d
r
em
o
te
s
en
s
in
g
im
ag
er
y
ap
p
l
icatio
n
s
f
r
o
m
a
v
a
r
iety
o
f
a
n
g
les.
So
u
za
et
a
l.
[
6
]
p
r
o
p
o
s
es
th
e
m
eth
o
d
I
,
e
u
tili
zin
g
b
ig
d
ata
to
d
ev
elo
p
a
h
y
b
r
id
m
o
d
el
f
o
r
s
ce
n
e
class
if
icatio
n
an
d
h
ig
h
lig
h
ted
im
p
o
r
tan
ce
o
f
b
ig
d
ata
an
d
p
r
o
p
o
s
ed
h
y
b
r
id
m
o
d
el
’
s
f
u
n
ctio
n
ality
.
Per
tu
r
b
atio
n
-
s
ee
k
in
g
g
en
e
r
ativ
e
ad
v
er
s
ar
ial
n
e
two
r
k
s
f
o
r
r
em
o
te
s
en
s
in
g
in
tr
o
d
u
ce
d
b
y
C
h
en
g
et
a
l.
[
7
]
in
w
h
ich
h
e
s
ay
s
e
m
p
lo
y
in
g
g
en
e
r
ativ
e
ad
v
er
s
ar
ial
n
etwo
r
k
s
th
a
t
r
eliab
ly
class
if
y
s
ce
n
es
in
r
e
m
o
te
s
en
s
in
g
im
ag
es
o
f
f
er
ed
v
iewp
o
in
ts
o
n
im
p
r
o
v
in
g
r
o
b
u
s
tn
ess
o
f
s
ce
n
e
ca
teg
o
r
izatio
n
m
o
d
els.
Yao
et
a
l.
[
8
]
in
v
e
n
ted
u
tili
zin
g
w
ea
k
s
u
p
er
v
is
io
n
f
o
r
o
b
ject
d
etec
tio
n
in
r
em
o
te
s
en
s
in
g
im
ag
es
wh
ich
h
as
p
r
o
v
id
ed
in
s
ig
h
ts
f
o
r
f
u
tu
r
e
s
tu
d
i
es
o
n
s
im
ilar
m
eth
o
d
s
.
Hu
an
g
et
a
l.
[
9
]
p
r
o
p
o
s
es
t
r
ac
k
ed
ch
an
g
es
in
ec
o
s
y
s
tem
s
er
v
ices
u
s
in
g
d
ee
p
lear
n
in
g
an
d
h
ig
h
-
r
eso
lu
tio
n
r
e
m
o
t
ely
s
en
s
ed
im
ag
er
y
wh
ich
o
u
ts
s
h
o
wed
p
o
ten
tial
wid
er
u
s
es
o
f
p
ictu
r
e
ca
teg
o
r
izatio
n
in
en
v
ir
o
n
m
e
n
tal
tr
ac
k
in
g
u
s
in
g
tr
ac
k
in
g
ch
an
g
es
in
ec
o
s
y
s
tem
s
er
v
ices
u
s
in
g
d
ee
p
le
ar
n
i
n
g
.
T
h
e
i
n
tr
o
d
u
ctio
n
o
f
f
r
am
ewo
r
k
f
o
r
h
ig
h
s
p
atial
r
eso
lu
tio
n
im
ag
e
s
ce
n
e
ca
teg
o
r
izatio
n
u
s
in
g
d
ee
p
s
p
ar
s
e
s
em
an
tic
m
o
d
elin
g
is
d
o
n
e
b
y
Z
h
u
et
a
l.
[
1
0
]
wh
ich
o
u
tp
u
ts
th
e
c
o
n
tr
ib
u
ted
to
u
n
d
e
r
s
tan
d
in
g
ef
f
ec
tiv
e
p
r
o
ce
s
s
in
g
o
f
h
ig
h
-
r
eso
lu
tio
n
im
ag
es.
L
an
d
-
u
s
e
class
if
icatio
n
u
s
in
g
d
ee
p
c
o
n
v
o
lu
tio
n
al
f
ea
tu
r
e
-
b
a
s
ed
ex
tr
em
e
lear
n
in
g
class
if
ier
b
y
au
t
h
o
r
W
en
g
et
a
l.
[
1
1
]
Su
g
g
ested
m
eth
o
d
f
o
r
la
n
d
-
u
s
e
class
if
icatio
n
u
tili
zin
g
d
ee
p
f
ea
tu
r
es
an
d
o
f
f
er
ed
in
s
ig
h
ts
in
to
u
s
in
g
d
ee
p
f
e
atu
r
es
f
o
r
lan
d
-
u
s
e
class
if
icatio
n
.
L
i
et
a
l.
[
1
2
]
in
t
r
o
d
u
ce
s
s
p
atial
-
tem
p
o
r
al
s
u
p
e
r
-
r
eso
lu
tio
n
lan
d
co
v
er
m
ap
p
i
n
g
m
o
d
el
f
o
r
s
u
p
er
-
r
eso
lu
tio
n
lan
d
c
o
v
er
m
ap
p
i
n
g
with
s
p
atial
-
tem
p
o
r
al
c
o
n
s
id
er
atio
n
s
an
d
s
u
g
g
ested
p
o
ten
tial
m
eth
o
d
s
f
o
r
ad
d
r
ess
in
g
tem
p
o
r
al
co
n
ce
r
n
s
.
C
lass
if
icat
io
n
o
f
in
te
r
io
r
s
ce
n
es
u
s
in
g
a
c
o
m
b
in
atio
n
o
f
g
l
o
b
al
an
d
s
em
an
tic
v
ar
iab
les
in
v
en
ted
b
y
Per
eir
a
et
a
l.
[
1
3
]
an
d
co
n
clu
d
es
with
e
m
p
h
asized
s
ig
n
if
ican
ce
o
f
f
ea
tu
r
e
f
u
s
io
n
f
o
r
s
ce
n
e
class
if
icatio
n
.
C
h
en
g
et
a
l.
[
1
4
]
an
d
C
h
en
g
et
a
l.
[
1
5
]
b
o
th
au
th
o
r
s
in
tr
o
d
u
ce
c
o
m
b
in
in
g
v
is
u
al
ter
m
s
with
m
u
lti
-
s
ca
le
co
m
p
leted
l
o
ca
l
b
in
a
r
y
p
atter
n
s
wh
ic
h
p
r
o
v
id
es
p
r
o
v
id
ed
h
is
to
r
ical
co
n
tex
t
f
o
r
s
ce
n
e
class
if
icatio
n
tech
n
iq
u
es.
L
iu
et
a
l.
[
1
6
]
ad
v
o
ca
te
a
r
an
d
o
m
-
s
ca
le
s
tr
etch
ed
co
n
v
o
lu
tio
n
al
n
eu
r
a
l
co
m
m
u
n
ity
(
R
S
-
SC
NN
)
,
wh
ich
ap
p
lies
r
an
d
o
m
-
s
ca
le
s
tr
etch
in
g
to
en
ter
p
ics
f
o
r
r
ein
f
o
r
cin
g
m
u
lti
-
s
ca
le
f
u
n
ctio
n
lear
n
in
g
,
im
p
r
o
v
i
n
g
s
ce
n
e
s
e
m
an
tic
ca
teg
o
r
y
i
n
ex
ce
s
s
iv
esp
atial
d
ec
is
io
n
f
ar
awa
y
s
en
s
in
g
im
ag
er
y
a
n
d
co
n
clu
d
es
T
h
e
R
S
-
SC
NN
v
er
s
io
n
ac
h
ie
v
es
ad
v
a
n
ce
d
class
o
v
er
all
p
er
f
o
r
m
a
n
ce
in
co
m
p
ar
is
o
n
t
o
co
n
v
en
tio
n
al
C
NNs,
d
em
o
n
s
t
r
atin
g
s
tep
p
ed
f
o
r
war
d
ac
cu
r
ac
y
in
co
p
in
g
with
d
iv
er
s
e
s
ce
n
e
s
ca
les
in
h
ig
h
-
d
ec
is
io
n
f
ar
awa
y
s
en
s
in
g
d
ata
s
ets.
A
n
o
v
el
m
etah
eu
r
is
tic
o
p
tim
izer
in
s
p
ir
ed
b
y
b
e
h
av
io
r
o
f
jelly
f
is
h
i
n
o
c
ea
n
b
y
au
th
o
r
C
h
o
u
a
n
d
T
r
u
o
n
g
[
1
7
]
r
ec
o
m
m
e
n
d
ed
t
h
e
J
elly
f
is
h
s
ee
k
(
J
S)
o
p
tim
izer
,
a
n
o
v
el
m
etah
eu
r
is
tic
alg
o
r
ith
m
i
n
s
p
ir
ed
v
ia
jelly
f
is
h
co
n
d
u
ct
in
s
id
e
th
e
o
ce
an
,
s
u
ch
as
p
ass
iv
e
m
o
tio
n
an
d
en
er
g
etic
s
wim
m
in
g
.
T
h
e
alg
o
r
ith
m
m
im
ics jelly
f
is
h
d
y
n
a
m
ics to
d
is
co
v
er
a
n
d
e
x
p
lo
it sear
ch
ar
ea
s
f
o
r
o
p
tim
iza
tio
n
tr
o
u
b
les.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
A
h
yb
r
id
a
p
p
r
o
a
ch
u
s
in
g
V
GG
1
6
-
E
ffcien
tN
etV
2
B
3
-
F
C
N
ets f
o
r
a
cc
u
r
a
te
…
(
Meg
h
a
n
a
Des
h
mu
kh
)
905
Mir
jalili
an
d
L
ewis
[
1
8
]
r
ec
o
m
m
en
d
s
t
h
e
W
h
ale
Op
tim
i
za
tio
n
alg
o
r
ith
m
(
W
OA)
is
a
n
atu
r
e
-
s
tim
u
lated
o
p
tim
izatio
n
a
p
p
r
o
ac
h
th
at
m
im
ics
th
e
b
u
b
b
l
e
-
n
et
lo
o
k
in
g
m
eth
o
d
o
f
h
u
m
p
b
ac
k
wh
ales.
I
t
b
alan
ce
s
ex
p
lo
r
atio
n
a
n
d
ex
p
lo
itatio
n
v
ia
m
o
d
elin
g
en
cir
cl
in
g
an
d
s
p
ir
al
m
o
v
em
en
t
t
o
war
d
s
th
e
f
ir
s
t
-
class
an
s
wer
.
W
OA
is
wid
ely
u
s
ed
f
o
r
s
o
l
v
in
g
co
m
p
licated
o
p
tim
izatio
n
tr
o
u
b
les
in
en
g
i
n
ee
r
in
g
an
d
g
a
d
g
et
g
ettin
g
to
k
n
o
w
.
T
h
e
a
u
th
o
r
W
u
et
a
l.
[
1
9
]
i
n
tr
o
d
u
ce
YOL
O
-
DC
Net,
a
f
lex
ib
le,
lig
h
t
-
weig
h
t
h
u
m
an
d
etec
tio
n
alg
o
r
ith
m
th
at
in
teg
r
ates
s
em
an
tic
-
b
ased
s
tatis
tic
s
with
th
e
YOL
O
(
Yo
u
b
est
ap
p
ea
r
an
ce
o
n
ce
)
s
tr
u
ctu
r
e.
T
h
e
m
o
d
el
s
p
ec
ializes
in
im
p
r
o
v
i
n
g
d
etec
tio
n
ac
cu
r
ac
y
wh
ile
k
ee
p
in
g
p
er
f
o
r
m
an
ce
in
ac
t
u
al
-
tim
e
ev
en
tu
alities
wh
ich
p
r
o
p
o
s
es
YOL
O
-
DC
N
et
d
em
o
n
s
tr
ates
ad
v
an
ce
d
o
v
e
r
all
p
er
f
o
r
m
a
n
ce
in
h
u
m
a
n
d
e
tectio
n
o
b
lig
atio
n
s
,
r
ea
ch
in
g
b
etter
d
etec
tio
n
ac
cu
r
ac
y
an
d
d
ec
r
ea
s
e
co
m
p
u
tatio
n
al
ch
ar
g
es
in
co
m
p
a
r
is
o
n
to
co
n
v
en
tio
n
al
YOL
O
-
p
r
im
ar
ily
b
ased
m
o
d
els,
m
ak
in
g
it
s
u
itab
le
f
o
r
r
ea
l
-
tim
e
p
ac
k
ag
es.
Hao
et
a
l.
[
2
0
]
p
r
o
p
o
s
es
an
en
tr
o
p
y
-
au
g
m
en
ted
n
eu
r
al
co
m
m
u
n
ity
f
o
r
co
r
r
ec
t
h
ig
h
-
d
e
n
s
ity
cr
o
wd
c
o
u
n
tin
g
.
I
t
co
n
tai
n
s
f
ac
ts
en
tr
o
p
y
to
m
an
u
al
th
e
m
o
d
el
in
f
o
cu
s
in
g
o
n
u
n
ce
r
tain
o
r
am
b
i
g
u
o
u
s
r
eg
io
n
s
in
d
e
n
s
e
cr
o
w
d
s
.
T
h
i
s
m
eth
o
d
im
p
r
o
v
es
r
o
b
u
s
tn
ess
an
d
ac
cu
r
ac
y
in
co
m
p
lex
r
ea
l
-
in
ter
n
atio
n
al
cr
o
w
d
an
aly
s
is
ev
en
tu
alities
.
Ma
s
u
d
et
a
l.
[
2
1
]
p
r
esen
t
a
m
eth
o
d
th
e
u
s
e
o
f
p
r
e
-
s
k
illed
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs)
f
o
r
b
r
ea
s
t
m
o
s
t
ca
n
ce
r
s
d
etec
tio
n
in
u
ltra
s
o
u
n
d
p
ics
.
T
h
e
m
eth
o
d
l
ev
er
ag
es
tr
an
s
f
er
g
ettin
g
t
o
k
n
o
w
to
en
h
an
ce
d
iag
n
o
s
tic
ac
cu
r
ac
y
with
co
n
f
in
e
d
class
if
ied
in
f
o
r
m
atio
n
.
E
f
f
ec
ts
d
is
p
lay
h
ig
h
o
v
e
r
all
p
er
f
o
r
m
an
ce
in
class
if
y
in
g
b
e
n
ig
n
an
d
m
alig
n
a
n
t
tu
m
o
r
s
ef
f
icac
io
u
s
ly
.
M
u
ltiro
u
n
d
tr
a
n
s
f
er
lear
n
i
n
g
a
n
d
m
o
d
if
ie
d
g
en
er
ativ
e
a
d
v
er
s
ar
ial
n
etwo
r
k
s
f
o
r
lu
n
g
ca
n
ce
r
d
etec
tio
n
b
y
C
h
u
i
et
a
l.
[
2
2
]
p
r
o
p
o
s
ed
m
o
d
el
ac
h
iev
es
s
iza
b
le
u
p
g
r
ad
es
in
l
u
n
g
ca
n
ce
r
d
etec
tio
n
ac
cu
r
ac
y
,
o
u
tp
er
f
o
r
m
in
g
tr
ad
itio
n
al
tec
h
n
iq
u
es.
T
h
e
m
u
ltiro
u
n
d
tr
a
n
s
f
er
lear
n
in
g
a
n
d
GAN
au
g
m
en
tatio
n
r
esu
lt
in
h
ig
h
er
g
e
n
er
aliza
tio
n
an
d
r
o
b
u
s
tn
ess
,
m
ain
ly
in
co
p
in
g
wit
h
r
estricte
d
o
r
im
b
alan
ce
d
d
at
asets
.
Z
h
u
et
a
l.
[
2
3
]
t
h
is
p
ap
er
p
r
o
p
o
s
es
a
b
ag
-
of
-
v
is
ib
le
-
p
h
r
ases
(
B
o
VW
)
s
ce
n
e
class
if
ier
th
at
co
m
b
in
es
n
ea
r
b
y
an
d
in
ter
n
atio
n
al
f
ea
tu
r
es
f
o
r
class
if
y
in
g
h
ig
h
-
r
eso
lu
tio
n
f
ar
f
lu
n
g
s
en
s
in
g
p
h
o
to
g
r
ap
h
s
.
T
h
e
tec
h
n
iq
u
e
c
o
m
p
lem
en
ts
s
p
atial
f
u
n
ctio
n
r
ep
r
esen
tatio
n
f
o
r
c
o
m
p
licated
s
ce
n
es.
I
t
ac
h
iev
e
s
ef
f
ec
tiv
e
o
v
er
all
p
er
f
o
r
m
a
n
ce
in
lan
d
-
u
s
e
an
d
co
n
cr
ete
s
ce
n
e
r
ec
o
g
n
itio
n
r
e
s
p
o
n
s
ib
ilit
ies
.
Hu
an
g
et
a
l.
[
2
4
]
in
tr
o
d
u
ce
a
m
eth
o
d
u
s
in
g
m
u
lti
-
s
ca
le
f
in
is
h
ed
lo
ca
l
b
in
ar
y
s
ty
les
(
L
B
P)
an
d
f
is
h
er
v
ec
to
r
s
f
o
r
f
ar
f
l
u
n
g
s
e
n
s
in
g
s
ce
n
e
ty
p
e
.
T
h
e
co
m
b
in
atio
n
ca
p
tu
r
es
n
ice
-
g
r
ain
ed
tex
tu
r
e
an
d
s
p
atial
la
y
o
u
t
f
u
n
ctio
n
s
.
I
t
s
u
g
g
ests
h
i
g
h
ac
cu
r
ac
y
in
d
is
cr
im
in
atin
g
d
iv
er
s
e
lan
d
co
v
e
r
ty
p
es
in
s
atellite
tv
f
o
r
p
c
s
n
a
p
s
h
o
ts
.
Mo
h
a
m
ed
et
a
l.
[
2
5
]
t
h
is
p
ap
er
p
r
o
v
i
d
es
a
d
ee
p
g
et
tin
g
to
k
n
o
w
-
b
ased
to
tally
s
em
an
tic
s
eg
m
en
tatio
n
m
eth
o
d
to
class
if
y
in
d
o
o
r
an
d
o
u
td
o
o
r
en
v
ir
o
n
m
en
ts
.
T
h
e
s
y
s
tem
is
d
esig
n
ed
to
h
elp
v
is
u
ally
im
p
air
ed
w
h
ee
lch
air
c
u
s
to
m
er
s
b
y
u
s
in
g
en
a
b
lin
g
r
ea
l
-
tim
e
s
ce
n
e
e
x
p
er
tis
e
,
im
p
r
o
v
in
g
n
av
ig
atio
n
p
r
o
tectio
n
a
n
d
a
u
to
n
o
m
y
.
Alv
es
et
a
l.
[
2
6
]
t
h
e
au
th
o
r
s
s
u
g
g
est
a
s
in
g
u
lar
m
eth
o
d
f
o
r
in
d
o
o
r
/
o
u
td
o
o
r
ty
p
e
of
u
s
er
d
ev
ice
in
ce
ll
n
etwo
r
k
s
.
T
h
eir
v
er
s
io
n
lev
er
ag
es
s
ig
n
an
d
c
o
n
tex
tu
al
in
f
o
r
m
atio
n
to
en
h
an
ce
th
e
ac
cu
r
ac
y
o
f
lo
c
atio
n
-
b
ased
o
f
f
er
in
g
s
in
n
ex
t
-
t
ec
h
n
o
lo
g
y
ce
llu
lar
n
etwo
r
k
s
.
a)
I
m
ag
e
class
if
icatio
n
ar
ch
itectu
r
es
Dee
p
g
ettin
g
to
k
n
o
w
ad
v
an
c
es
p
h
o
to
an
aly
s
is
with
VGG1
6
,
E
f
f
icien
tNetV2
B
3
,
an
d
FC
Nets,
o
p
tim
izin
g
ac
cu
r
ac
y
,
p
er
f
o
r
m
an
ce
,
an
d
b
e
s
t
-
g
r
ain
ed
ty
p
e
in
aid
-
lim
ited
en
v
ir
o
n
m
en
ts
.
b)
I
n
d
o
o
r
v
s
o
u
td
o
o
r
s
ce
n
e
class
if
icatio
n
I
n
d
o
o
r
-
o
u
td
o
o
r
s
ce
n
e
ca
teg
o
r
izatio
n
ad
v
an
ce
d
f
r
o
m
h
an
d
c
r
af
ted
f
ea
tu
r
es
to
d
ee
p
lear
n
i
n
g
m
o
d
els
lik
e
Go
o
g
L
eNe
t,
R
esNet,
an
d
h
y
b
r
id
ap
p
r
o
ac
h
es,
en
h
an
cin
g
ac
cu
r
ac
y
.
c)
An
im
ated
v
s
n
atu
r
al
im
a
g
e
cla
s
s
if
icatio
n
C
las
s
if
y
in
g
an
im
ated
v
e
r
s
u
s
n
atu
r
al
im
ag
es
r
eq
u
ir
es
n
u
an
ce
d
tech
n
iq
u
es.
T
r
ad
itio
n
a
l
m
eth
o
d
s
u
s
ed
tex
tu
r
e
f
ea
tu
r
es,
wh
ile
d
ee
p
l
ea
r
n
in
g
,
in
co
r
p
o
r
atin
g
tem
p
o
r
al
an
d
s
p
atial
in
f
o
r
m
atio
n
,
ac
h
iev
es
s
u
p
er
io
r
ac
cu
r
ac
y
in
r
ec
en
t stu
d
ies.
d)
Hy
b
r
id
ap
p
r
o
ac
h
es in
d
ee
p
lear
n
in
g
Hy
b
r
id
ar
ch
itectu
r
es c
o
m
b
in
e
d
iv
er
s
e
m
o
d
els lik
e
r
esid
u
al
n
etwo
r
k
s
an
d
atten
tio
n
m
ec
h
a
n
i
s
m
s
,
en
h
an
cin
g
p
er
f
o
r
m
an
ce
b
y
in
teg
r
atin
g
s
p
atial
an
d
te
m
p
o
r
al
s
tr
en
g
th
s
f
o
r
im
ag
e
task
s
.
e)
C
h
allen
g
es
an
d
o
p
p
o
r
tu
n
ities
Hy
b
r
id
ap
p
r
o
ac
h
es
lev
er
a
g
e
d
iv
er
s
e
ar
ch
itectu
r
es
f
o
r
i
m
p
r
o
v
e
d
d
is
cr
im
in
atio
n
,
o
p
t
im
izin
g
m
o
d
els,
ex
p
lo
r
in
g
tr
an
s
f
er
lear
n
in
g
,
en
h
an
cin
g
ex
p
lain
ab
ilit
y
,
an
d
ad
v
an
cin
g
r
ea
l
-
tim
e,
a
d
ap
tab
le
im
ag
e
class
if
icatio
n
s
o
lu
tio
n
s
.
2.
P
RO
P
O
SE
D
M
E
T
H
O
D
B
y
u
s
in
g
in
teg
r
atin
g
VGG1
6
,
E
f
f
icien
tNetV2
B
3
,
an
d
FC
Ns,
th
is
h
y
b
r
id
m
eth
o
d
im
p
r
o
v
es
p
ictu
r
e
ca
teg
o
r
y
b
y
way
o
f
lev
er
ag
i
n
g
VGG1
6
’
s
f
u
n
ctio
n
r
ec
o
g
n
itio
n
,
E
f
f
icien
t
NetV2
B
3
’
s
p
atter
n
d
etec
tio
n
,
an
d
FC
Ns
’
s
y
n
th
esis
,
im
p
r
o
v
in
g
ac
cu
r
ac
y
an
d
p
er
f
o
r
m
an
ce
at
th
e
s
am
e
tim
e
as
r
ef
in
in
g
class
m
eth
o
d
s
with
s
u
p
e
r
io
r
a
n
al
y
s
is
an
d
th
r
esh
o
l
d
in
g
.
Fig
u
r
es
1
an
d
2
d
is
p
lay
an
p
h
o
to
g
r
ap
h
ev
alu
atio
n
p
ip
elin
e
with
d
ataset
co
llectio
n
,
p
r
ep
r
o
ce
s
s
in
g
,
f
u
n
ctio
n
ex
tr
ac
tio
n
,
a
g
g
r
e
g
atio
n
,
a
n
d
ev
alu
atio
n
f
o
r
p
r
ed
ictio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
14
,
No
.
3
,
Dec
em
b
er
20
25
:
903
-
9
1
3
906
Fig
u
r
e
1
.
Pro
p
o
s
ed
a
p
p
r
o
ac
h
f
o
r
im
ag
e
class
if
icatio
n
Fig
u
r
e
2
.
Pro
p
o
s
ed
m
et
h
o
d
o
lo
g
y
f
lo
wch
a
r
t f
o
r
im
ag
e
class
if
icatio
n
3.
RE
S
E
ARCH
M
E
T
H
O
DO
L
O
G
Y
3
.
1
.
B
a
s
ic
a
rc
hite
ct
ure
o
f
re
s
ea
rc
h m
et
ho
do
lo
g
y
Her
e
ar
e
th
e
7
s
tep
s
ty
p
ically
in
v
o
lv
ed
i
n
a
p
ip
elin
e
f
o
r
an
i
m
ag
e
class
if
icatio
n
p
r
o
b
lem
a
s
s
h
o
wn
in
Fig
u
r
e
3.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
A
h
yb
r
id
a
p
p
r
o
a
ch
u
s
in
g
V
GG
1
6
-
E
ffcien
tN
etV
2
B
3
-
F
C
N
ets f
o
r
a
cc
u
r
a
te
…
(
Meg
h
a
n
a
Des
h
mu
kh
)
907
a)
Data
co
llectio
n
an
d
p
r
ep
ar
atio
n
:
C
o
llect
a
ap
p
licab
le
im
ag
e
d
ataset,
p
r
ep
ar
e
it
in
to
s
ch
o
o
lin
g
,
v
alid
atio
n
,
an
d
tr
y
in
g
o
u
t
s
ets
,
an
d
p
r
ep
r
o
ce
s
s
p
ictu
r
es
b
y
u
s
in
g
r
esizin
g
an
d
n
o
r
m
alizin
g
f
o
r
u
n
if
o
r
m
ity
an
d
d
ec
r
ea
s
ed
co
m
p
l
ex
ity
.
b)
E
x
p
lo
r
ato
r
y
d
ata
a
n
aly
s
is
(
E
DA)
:
-
T
h
e
d
is
tr
ib
u
tio
n
o
f
class
es
is
o
n
e
d
ataset
ch
ar
ac
ter
is
tic
th
at
y
o
u
s
h
o
u
ld
b
e
f
am
iliar
with
,
im
ag
e
r
eso
lu
tio
n
s
,
an
d
a
n
y
a
n
o
m
alie
s
.
-
Vis
u
alize
s
am
p
le
im
ag
es f
r
o
m
d
if
f
er
en
t c
lass
es to
g
ain
in
s
ig
h
ts
in
to
th
e
d
ata.
c)
Featu
r
e
e
x
tr
ac
tio
n
:
Dee
p
lear
n
in
g
r
o
u
tin
el
y
lear
n
s
f
ea
tu
r
es th
r
u
co
n
v
o
lu
tio
n
al
la
y
er
s
,
n
o
t lik
e
tr
a
d
itio
n
al
m
eth
o
d
s
.
d)
Mo
d
el
s
elec
tio
n
:
Select
a
s
u
itab
le
m
ac
h
in
e
lea
r
n
in
g
m
o
d
el
f
o
r
im
a
g
e
class
i
f
icatio
n
.
C
NNs
ar
e
wid
ely
u
t
ilized
in
d
ee
p
lear
n
in
g
task
s
b
ec
au
s
e
o
f
th
eir
ex
ce
p
tio
n
al
a
b
ilit
y
to
ca
p
tu
r
e
s
p
atial
h
ier
ar
ch
ies an
d
p
atter
n
s
in
im
ag
es.
e)
Mo
d
el
t
r
ain
in
g
:
E
d
u
ca
te
th
e
v
er
s
io
n
,
o
p
tim
ize
h
y
p
er
p
ar
am
eter
s
,
an
d
r
ev
ea
l
p
er
f
o
r
m
an
ce
at
th
e
v
alid
atio
n
s
et
to
s
av
e
y
o
u
o
v
er
f
itti
n
g
.
f)
Mo
d
el
e
v
alu
atio
n
:
E
x
am
in
e
v
er
s
io
n
o
v
er
all
p
er
f
o
r
m
an
ce
th
e
u
s
ag
e
o
f
ac
cu
r
a
cy
,
p
r
ec
is
io
n
,
d
o
n
o
t
f
o
r
g
et
,
F1
-
s
co
r
e,
an
d
v
is
u
alize
ef
f
ec
ts
with
co
n
f
u
s
io
n
m
atr
ices.
g)
Mo
d
el
d
ep
lo
y
m
en
t:
I
n
s
tallatio
n
th
e
v
er
s
io
n
f
o
r
r
ea
l
-
tim
e
p
r
ed
ictio
n
s
,
c
o
m
b
in
e
it
in
to
p
ac
k
a
g
es,
an
d
m
o
n
ito
r
p
e
r
f
o
r
m
a
n
ce
f
o
r
im
p
o
r
tan
t r
etr
ai
n
in
g
.
Fig
u
r
e
3
.
R
esear
ch
m
eth
o
d
o
l
o
g
y
f
o
r
im
ag
e
class
if
icatio
n
3
.
2
.
VG
G
1
6
(
f
ea
t
ure
e
x
t
ra
ct
o
r)
VGG1
6
is
k
n
o
wn
f
o
r
its
d
ee
p
co
n
v
o
lu
tio
n
al
s
tr
u
ct
u
r
e.
T
h
e
ar
ch
itectu
r
e
is
co
m
p
o
s
ed
o
f
m
u
ltip
le
co
n
v
o
l
u
tio
n
al
lay
er
s
,
w
h
ich
ar
e
th
en
f
o
llo
wed
b
y
m
ax
-
p
o
o
lin
g
lay
er
s
,
a
n
d
f
i
n
ally
cu
lm
in
ate
in
f
u
lly
co
n
n
ec
ted
la
y
er
s
.
T
h
e
o
u
tco
m
e
o
f
th
e
f
in
al
f
u
lly
c
o
n
n
ec
te
d
l
ay
er
ac
ts
as a
h
ig
h
-
le
v
el
f
ea
t
u
r
e
r
ep
r
esen
tatio
n
.
3
.
3
.
E
f
f
icient
Net
V2
B
3
(
f
e
a
t
ure
ex
t
ra
ct
o
r
)
E
f
f
icien
tNetV2
B
3
is
a
v
ar
ian
t
o
f
th
e
E
f
f
icien
tNet
ar
ch
itectu
r
e,
d
esig
n
ed
f
o
r
o
p
tim
al
p
e
r
f
o
r
m
a
n
ce
an
d
ef
f
icien
cy
.
I
t
in
clu
d
es
m
u
ltip
le
b
lo
ck
s
with
d
if
f
er
e
n
t
s
p
atial
r
eso
lu
tio
n
s
(
wid
th
,
h
eig
h
t,
an
d
d
ep
t
h
)
.
T
h
e
f
ea
tu
r
e
m
ap
s
f
r
o
m
th
ese
b
lo
ck
s
ca
p
tu
r
e
d
iv
er
s
e
in
f
o
r
m
at
io
n
ab
o
u
t th
e
in
p
u
t im
ag
e.
3
.
4
.
F
CNe
t
s
(
c
la
s
s
if
ier)
Fu
lly
C
o
n
n
ec
ted
Netwo
r
k
s
ar
e
co
m
m
o
n
ly
em
p
lo
y
e
d
f
o
r
f
in
e
-
g
r
ain
ed
class
if
icatio
n
task
s
.
T
h
e
f
u
lly
co
n
n
ec
ted
lay
er
s
u
tili
ze
th
e
f
ea
tu
r
es
ex
tr
ac
ted
b
y
th
e
VGG1
6
an
d
E
f
f
icien
tNetV2
B
3
an
d
m
ap
th
em
to
th
e
o
u
tp
u
t c
lass
es,
p
er
f
o
r
m
in
g
t
h
e
f
in
al
class
if
icatio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
14
,
No
.
3
,
Dec
em
b
er
20
25
:
903
-
9
1
3
908
3
.
5
.
M
a
t
hema
t
ica
l
e
qu
a
t
io
n
s
L
et
’
s
d
en
o
te:
−
X
as th
e
in
p
u
t im
ag
e,
−
VGG(
X)
as th
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
o
b
tain
e
d
f
r
o
m
VGG1
6
,
−
E
f
f
Net(
X)
as th
e
f
ea
tu
r
e
r
ep
r
e
s
en
tatio
n
o
b
tain
ed
f
r
o
m
E
f
f
ici
en
tNetV2
B
3
,
−
FC
Net
s
(
[
VGG(
X)
,
E
f
f
Net(
X)
]
)
as th
e
r
esu
lts
o
b
tain
ed
f
r
o
m
t
h
e
f
u
lly
c
o
n
n
ec
ted
lay
er
s
.
−
T
h
e
f
o
llo
win
g
is
a
m
ath
em
atic
al
ex
p
r
ess
io
n
o
f
t
h
e
h
y
b
r
id
m
o
d
el:
Y=FC
N
ets([V
GG(
X
)
,
E
ffN
et(
X
)
]
)
(
1
)
h
er
e,
Y
r
ep
r
esen
ts
th
e
o
u
tp
u
t
lo
g
its
o
r
p
r
o
b
a
b
ilit
ies
f
o
r
ea
ch
class
.
I
n
teg
r
atin
g
th
e
f
ea
tu
r
es
ex
tr
ac
ted
,
f
u
lly
co
n
n
ec
ted
la
y
er
s
b
y
VGG1
6
a
n
d
E
f
f
icien
tNetV2
B
3
f
o
r
th
e
f
in
al
class
if
icatio
n
.
3
.
6
.
I
nte
g
ra
t
io
n o
f
VG
G
1
6
a
nd
E
f
f
icient
Net
V2
B
3
T
h
e
in
teg
r
atio
n
o
f
th
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
f
r
o
m
VGG1
6
an
d
E
f
f
icien
tNetV2
B
3
ca
n
b
e
d
o
n
e
b
y
co
n
ca
ten
atin
g
t
h
e
f
ea
tu
r
e
v
ec
t
o
r
s
:
A
g
g
r
eg
a
ted
_
F
ea
tu
r
es=Ag
g
r
eg
a
tio
n
(
[
V
GG(
X
)
,
E
fficien
tN
etV
2
B
3
(
X
)
]
)
(
2
)
l
in
e
an
d
ar
ea
c
h
ar
ac
ter
is
tics
ar
e
av
er
ag
ed
u
s
in
g
weig
h
ted
m
e
an
s
tatis
tic
s
.
T
o
g
et
th
e
weig
h
t
ed
m
ea
n
,
u
s
e
th
e
f
o
llo
win
g
f
o
r
m
u
la:
‾
=
∑
=
1
⋅
∑
=
1
(
3
)
w
h
er
e:
N
=
n
u
m
b
er
o
f
o
b
s
er
v
a
tio
n
s
,
x
i =
o
b
s
er
v
atio
n
s
,
W
i =
weig
h
ts
T
h
e
n
e
u
r
al
n
etwo
r
k
u
s
es
R
eL
U
ac
tiv
atio
n
in
h
id
d
en
lay
er
s
an
d
s
ig
m
o
i
d
in
th
e
o
u
tp
u
t
lay
e
r
f
o
r
b
i
n
ar
y
class
if
icatio
n
.
Acc
u
r
ac
y
m
ea
s
u
r
es
m
o
d
el
p
er
f
o
r
m
a
n
ce
,
en
s
u
r
in
g
ef
f
icien
t
tr
ain
in
g
with
ap
p
r
o
p
r
iate
lay
er
s
,
ac
tiv
atio
n
f
u
n
cti
o
n
s
,
an
d
m
etr
i
cs.
3
.
7
.
H
idd
en
l
a
y
er
wit
h Re
L
U
a
ct
iv
a
t
io
n
Su
p
p
o
s
e
we
h
av
e
an
in
p
u
t
v
e
cto
r
x
an
d
weig
h
ts
W
an
d
b
iases
b
f
o
r
th
e
h
id
d
en
lay
er
.
T
h
e
o
u
tp
u
t
o
f
th
e
h
id
d
e
n
lay
er
h
af
ter
a
p
p
ly
i
n
g
th
e
R
eL
U
ac
tiv
atio
n
is
:
h
=R
eLU
(
Wx+b
)
(
4
)
wh
er
e
th
e
R
eL
U
f
u
n
ctio
n
is
d
ef
in
ed
as:
R
eLU(
z
)
=ma
x
(
0
,
z
)
(
5
)
3
.
8
.
O
utput
la
y
er
wit
h si
g
mo
id a
ct
iv
a
t
i
o
n
T
h
e
o
u
tp
u
t
lay
er
o
with
s
ig
m
o
id
ac
tiv
atio
n
,
g
iv
en
t
h
e
o
u
t
p
u
t
f
r
o
m
t
h
e
h
id
d
en
lay
e
r
h
,
w
eig
h
ts
W
o
,
an
d
b
iases
b
o
,
is
: o
=σ
(
W
o
h
+
b
o
)
. W
h
er
e
t
h
e
s
ig
m
o
id
f
u
n
cti
o
n
σ
(
z)
is
d
e
f
in
ed
as:
1
1
+
−
σ(
z
)
(
6
)
3
.
9
.
B
ina
ry
cr
o
s
s
-
ent
ro
py
lo
s
s
Fo
r
class
if
icatio
n
,
th
e
lo
s
s
f
o
r
a
s
in
g
le
tr
ain
in
g
e
x
am
p
le
wit
h
tr
u
e
lab
el
an
d
p
r
ed
icted
o
u
t
p
u
t
ˆ
is
:
=
−
∑
=
1
(
ˆ
)
(
7
)
wh
er
e:
−
As th
e
n
u
m
b
er
o
f
class
es.
−
As
th
e
b
in
ar
y
in
d
icato
r
(
0
o
r
1
)
if
class
lab
el
is
th
e
c
o
r
r
ec
t
class
if
icatio
n
f
o
r
th
e
g
iv
en
tr
ain
in
g
ex
am
p
le.
−
ˆ
I
s
th
e
p
r
ed
icted
p
r
o
b
ab
ilit
y
o
f
th
e
tr
ain
in
g
e
x
am
p
le
b
elo
n
g
in
g
to
class
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
A
h
yb
r
id
a
p
p
r
o
a
ch
u
s
in
g
V
GG
1
6
-
E
ffcien
tN
etV
2
B
3
-
F
C
N
ets f
o
r
a
cc
u
r
a
te
…
(
Meg
h
a
n
a
Des
h
mu
kh
)
909
Fo
r
a
b
atch
o
f
ex
am
p
les,
th
e
a
v
er
ag
e
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
is
:
L=
1
∑
=
1
∑
=
1
(
ˆ
)
(
8
)
wh
er
e:
−
is
th
e
b
in
ar
y
in
d
icato
r
f
o
r
th
e
j
-
th
ex
am
p
le
an
d
i
-
th
class
.
−
ˆ
is
th
e
p
r
ed
icted
p
r
o
b
a
b
ilit
y
f
o
r
th
e
j
-
th
ex
am
p
le
an
d
i
-
th
class
.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
co
m
p
ar
es
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
with
s
tate
-
of
-
th
e
-
ar
t
m
et
h
o
d
s
,
e
m
p
h
asizin
g
ac
cu
r
ac
y
an
d
m
etr
ics
lik
e
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e.
W
h
ile
ac
cu
r
ac
y
o
f
f
er
s
an
o
v
er
all
p
er
f
o
r
m
an
ce
s
n
ap
s
h
o
t,
m
etr
ics
d
e
r
iv
ed
f
r
o
m
th
e
c
o
n
f
u
s
io
n
m
atr
ix
r
ev
ea
l
class
-
s
p
ec
if
ic
p
er
f
o
r
m
an
ce
.
P
r
ec
is
io
n
m
in
im
izes
f
alse
p
o
s
itiv
es,
r
ec
all
d
etec
ts
r
elev
an
t
in
s
tan
ce
s
,
an
d
F1
-
s
co
r
e
b
ala
n
ce
s
b
o
th
,
h
ig
h
lig
h
tin
g
ar
ea
s
f
o
r
im
p
r
o
v
em
e
n
t.
T
ab
le
1
s
h
o
ws
th
e
class
if
icatio
n
ac
cu
r
ac
y
in
p
er
ce
n
tag
e
o
f
p
r
ev
io
u
s
ly
u
s
ed
alg
o
r
ith
m
a
p
p
r
o
ac
h
es.
So
,
th
e
C
NN
en
s
em
b
le
s
h
o
ws
th
e
h
ig
h
est ac
cu
r
ac
y
.
B
ased
o
n
th
e
f
in
d
in
g
s
p
r
esen
ted
in
2
,
i
t is ev
id
en
t th
at
o
u
r
m
eth
o
d
o
l
o
g
y
h
as
y
ield
ed
th
e
h
ig
h
est
ac
cu
r
ac
y
am
o
n
g
t
h
e
ev
alu
ated
ap
p
r
o
ac
h
es,
ac
h
ie
v
in
g
an
im
p
r
ess
iv
e
ac
cu
r
ac
y
r
ate
o
f
9
9
.
2
0
%
th
r
o
u
g
h
th
e
f
ea
tu
r
e
a
g
g
r
eg
atio
n
a
v
er
ag
e
(
a
v
g
)
m
eth
o
d
.
T
ab
le
2
co
m
p
ar
es
ac
cu
r
ac
y
f
o
r
n
atu
r
al
v
s
.
an
im
ated
im
ag
es,
h
ig
h
lig
h
tin
g
o
u
r
ap
p
r
o
ac
h
’
s
s
u
p
er
io
r
p
e
r
f
o
r
m
an
ce
u
s
in
g
th
e
Featu
r
e_
Ag
g
r
_
A
v
g
tech
n
iq
u
e
o
v
er
VGG1
6
,
E
f
f
icien
tNet,
an
d
Featu
r
e_
Ag
g
r
_
Ma
x
.
T
ab
le
1
.
C
lass
if
icatio
n
a
cc
u
r
a
cy
u
s
in
g
v
a
r
io
u
s
ap
p
r
o
ac
h
es
A
p
p
r
o
a
c
h
u
s
e
d
C
l
a
s
si
f
i
c
a
t
i
o
n
a
c
c
u
r
a
c
y
(
%)
D
e
e
p
l
e
a
r
n
i
n
g
7
6
.
8
5
C
N
N
a
n
d
K
-
m
e
a
n
s
7
9
.
2
1
Tr
a
n
sf
e
r
l
e
a
r
n
i
n
g
a
n
d
C
N
N
8
2
.
1
7
C
N
N
e
n
sem
b
l
e
8
5
.
3
6
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
o
f
ac
cu
r
ac
y
with
n
atu
r
al
a
n
d
an
im
ated
im
ag
es
A
l
g
o
r
i
t
h
m
/
M
e
t
h
o
d
o
l
o
g
y
C
l
a
s
si
f
i
c
a
t
i
o
n
a
c
c
u
r
a
c
y
(
%)
I
n
d
o
o
r
v
s
o
u
t
d
o
o
r
A
n
i
m
a
t
e
d
v
s
n
a
t
u
r
a
l
V
G
G
1
6
9
8
.
9
3
9
7
.
8
8
Ef
f
i
c
i
e
n
t
N
e
t
9
7
.
6
0
9
8
.
0
9
F
e
a
t
u
r
e
_
A
g
g
r
_
M
a
x
8
9
.
3
6
9
8
.
3
0
F
e
a
t
u
r
e
_
A
g
g
r
_
A
v
g
(
o
u
r
a
p
p
o
r
a
c
h
)
9
9
.
2
0
9
8
.
5
1
T
h
e
Featu
r
e_
Ag
g
r
_
Av
g
m
et
h
o
d
ex
ce
ls
in
ca
p
tu
r
in
g
s
p
atial
in
f
o
r
m
atio
n
,
lead
in
g
to
ac
cu
r
ate
class
if
icatio
n
s
.
Fig
u
r
es
4
an
d
5
co
m
p
ar
e
alg
o
r
ith
m
p
er
f
o
r
m
an
ce
,
s
h
o
win
g
Featu
r
e
_
Ag
g
r
_
Av
g
’
s
s
u
p
er
io
r
ac
cu
r
ac
y
in
class
if
icatio
n
task
s
.
Fig
u
r
e
4
.
Acc
u
r
ac
y
o
f
n
atu
r
al
v
s
an
im
ated
i
m
ag
es
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
14
,
No
.
3
,
Dec
em
b
er
20
25
:
903
-
9
1
3
910
Featu
r
e_
Ag
g
r
_
A
v
g
o
u
tp
er
f
o
r
m
s
VGG1
6
an
d
E
f
f
icien
tNet,
im
p
r
o
v
in
g
ac
c
u
r
ac
y
b
y
0
.
5
%
ac
r
o
s
s
v
ar
io
u
s
class
if
icatio
n
task
s
.
Fig
u
r
e
5
s
h
o
ws
f
o
r
th
e
task
o
f
in
d
o
o
r
v
er
s
u
s
o
u
td
o
o
r
i
m
ag
e
class
if
icatio
n
,
Featu
r
e_
Ag
g
r
_
A
v
g
ac
h
iev
es
a
n
im
p
r
ess
iv
e
ac
cu
r
ac
y
o
f
9
8
.
5
1
%.
T
h
is
s
ig
n
if
ies
its
ef
f
ec
tiv
en
ess
in
ac
cu
r
at
ely
d
is
tin
g
u
is
h
in
g
b
etwe
en
in
d
o
o
r
an
d
o
u
td
o
o
r
s
ce
n
es.
Similar
ly
,
in
t
h
e
class
if
icatio
n
o
f
an
i
m
ated
v
er
s
u
s
n
atu
r
al
im
ag
es,
Featu
r
e_
Ag
g
r
_
Av
g
e
x
h
ib
its
ex
ce
p
tio
n
al
p
er
f
o
r
m
an
ce
,
ac
h
iev
in
g
an
ac
cu
r
ac
y
r
ate
o
f
9
9
.
2
0
%.
Fig
u
r
e
5
.
Acc
u
r
ac
y
o
f
in
d
o
o
r
vs
o
u
td
o
o
r
i
m
ag
es
T
h
e
b
ar
p
lo
ts
s
er
v
e
as
a
v
is
u
al
co
n
f
ir
m
atio
n
o
f
th
e
s
u
p
er
io
r
ac
cu
r
ac
y
a
ch
iev
ed
b
y
Featu
r
e_
Ag
g
r
_
A
v
g
co
m
p
ar
ed
to
o
th
er
alg
o
r
ith
m
s
.
I
ts
r
esu
lts
d
em
o
n
s
tr
ate
h
o
w
th
e
h
y
b
r
id
f
ea
tu
r
e
ag
g
r
e
g
atio
n
m
eth
o
d
im
p
r
o
v
es
class
if
icatio
n
ac
cu
r
ac
y
,
o
u
tp
e
r
f
o
r
m
in
g
p
o
p
u
lar
f
ea
tu
r
e
ex
tr
ac
to
r
s
lik
e
VGG1
6
an
d
E
f
f
icien
tNet.
E
x
p
lo
r
ato
r
y
d
ata
an
aly
s
is
r
esu
lts
:
Fig
u
r
es
6
an
d
7
s
h
o
w
co
n
v
er
s
io
n
s
f
o
r
n
atu
r
al
v
s
an
i
m
ated
an
d
in
d
o
o
r
v
s
o
u
td
o
o
r
.
C
h
an
n
el
-
wis
e
d
is
tr
ib
u
tio
n
s
(
r
ed
,
g
r
ee
n
,
b
lu
e
)
r
ed
u
ce
d
im
e
n
s
io
n
ality
wh
e
n
c
o
n
v
er
te
d
to
g
r
ay
s
ca
le.
Fig
u
r
e
6
.
C
o
n
v
er
s
io
n
f
o
r
n
at
u
r
al
v
s
an
im
ated
i
m
ag
es
Fig
u
r
e
7
.
C
h
an
n
el
wis
e
d
is
tr
ib
u
tio
n
n
atu
r
al
v
s
an
im
ated
i
m
a
g
es
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
A
h
yb
r
id
a
p
p
r
o
a
ch
u
s
in
g
V
GG
1
6
-
E
ffcien
tN
etV
2
B
3
-
F
C
N
ets f
o
r
a
cc
u
r
a
te
…
(
Meg
h
a
n
a
Des
h
mu
kh
)
911
T
ab
les
3
an
d
4
ev
alu
ate
C
NN,
VGG1
6
,
Alex
Net,
E
f
f
icie
n
tNet,
an
d
h
y
b
r
id
f
ash
io
n
s
o
n
in
d
o
o
r
-
o
u
ts
id
e
an
d
an
im
ated
-
h
er
b
al
s
n
ap
s
h
o
ts
,
co
m
p
a
r
in
g
ac
cu
r
a
cy
,
p
r
ec
is
io
n
,
r
em
em
b
er
,
F1
r
atin
g
,
an
d
g
r
ea
ter
.
T
ab
le
4
is
m
o
d
el
r
esu
lts
f
o
r
an
im
ated
v
s
n
atu
r
al
.
T
ab
le
3
.
T
a
b
u
lar
r
esu
lt a
n
aly
s
is
f
o
r
in
d
o
o
r
v
s
o
u
td
o
o
r
im
a
g
es
S
.
N
o
A
l
g
o
r
i
t
h
m
A
c
c
u
r
a
c
y
Er
r
o
r
r
a
t
e
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
TP
TN
FP
FN
0
C
N
N
9
1
.
4
8
9
8
.
5
1
0
9
1
.
5
3
6
9
1
.
4
8
9
9
1
.
5
0
2
1
9
9
1
4
5
14
18
1
V
G
G
1
6
9
8
.
9
3
6
1
.
0
6
3
9
8
.
9
4
3
9
8
.
9
3
6
9
8
.
9
3
7
2
1
4
1
5
8
1
3
2
A
l
e
x
_
N
e
t
5
7
.
7
1
2
4
2
.
2
8
7
3
3
.
3
0
7
5
7
.
7
1
2
4
2
.
2
3
8
2
1
7
0
1
5
9
0
3
Ef
f
i
c
i
e
n
t
N
e
t
9
7
.
6
0
6
2
.
3
9
3
9
7
.
6
3
1
9
7
.
6
0
6
9
7
.
6
0
0
2
1
5
1
5
2
7
2
4
H
y
b
r
i
d
_
M
o
d
e
l
_
A
v
g
9
9
.
2
0
2
8
0
.
7
9
7
9
9
.
2
0
2
9
9
.
2
0
2
9
9
.
2
0
1
2
1
6
1
5
7
2
1
5
H
y
b
r
i
d
_
M
o
d
e
l
_
M
a
x
8
9
.
3
6
1
1
0
.
6
3
8
8
9
.
5
7
8
8
9
.
3
6
1
8
9
.
2
5
4
2
0
6
1
3
0
29
11
6
H
P
_
H
y
b
r
i
d
_
A
v
g
9
2
.
2
8
7
7
.
7
1
2
9
2
.
3
6
5
9
2
.
2
8
7
9
2
.
2
4
4
2
0
8
1
3
9
20
9
T
ab
le
4
.
T
a
b
u
lar
r
esu
lt a
n
aly
s
is
f
o
r
n
atu
r
al
v
s
an
im
ated
i
m
a
g
es
S
.
N
o
A
l
g
o
r
i
t
h
m
A
c
c
u
r
a
c
y
Er
r
o
r
r
a
t
e
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
TP
TN
FP
FN
0
C
N
N
9
3
.
0
0
8
6
.
9
9
1
9
3
.
0
3
1
9
3
.
0
0
8
9
3
.
0
1
9
3
4
5
94
16
17
1
V
G
G
1
6
9
7
.
8
8
1
2
.
1
1
8
9
7
.
9
0
7
9
7
.
8
8
1
9
7
.
8
5
4
3
5
7
1
0
5
9
1
2
A
l
e
x
_
N
e
t
9
5
.
1
2
7
4
.
8
7
2
9
5
.
0
9
2
9
5
.
1
2
7
9
5
.
0
3
5
3
5
6
93
17
6
3
Ef
f
i
c
i
e
n
t
N
e
t
9
8
.
0
9
3
1
.
9
0
6
7
8
0
9
8
.
1
1
2
9
8
.
0
9
3
9
8
.
0
7
2
3
5
6
1
0
7
8
1
4
H
y
b
r
i
d
_
M
o
d
e
l
_
A
v
g
9
8
.
5
1
6
1
.
4
8
3
9
8
.
5
2
2
9
8
.
5
1
6
9
8
.
5
1
9
3
5
8
1
0
7
3
4
5
H
y
b
r
i
d
_
M
o
d
e
l
_
M
a
x
9
8
.
3
0
5
1
.
6
9
4
9
8
.
3
0
5
9
8
.
3
0
5
9
8
.
3
0
5
3
5
8
1
0
6
4
4
6
H
P
_
H
y
b
r
i
d
_
A
v
g
9
8
.
7
2
8
1
.
2
7
1
9
8
.
7
9
4
9
8
.
7
2
8
9
8
.
7
4
0
3
5
6
1
1
0
0
6
5.
CO
NCLU
SI
O
N
T
h
is
r
esear
ch
p
r
esen
ts
a
h
y
b
r
id
m
o
d
el
co
m
b
in
in
g
VGG1
6
,
E
f
f
icien
tNetV2
B
3
,
an
d
FC
Net
s
f
o
r
ac
cu
r
ate
im
ag
e
class
if
icatio
n
,
ex
ce
llin
g
in
d
is
tin
g
u
is
h
in
g
in
d
o
o
r
/
o
u
td
o
o
r
an
d
an
im
at
ed
/n
atu
r
al
s
ce
n
es.
Ach
iev
in
g
8
5
.
3
6
%
ac
c
u
r
ac
y
i
n
s
ce
n
e
ca
teg
o
r
izatio
n
an
d
o
v
er
9
8
%
in
n
u
an
ce
d
task
s
,
th
e
f
ea
tu
r
e
ag
g
r
eg
atio
n
av
er
ag
e
ap
p
r
o
ac
h
s
tan
d
s
o
u
t,
h
ig
h
lig
h
tin
g
its
p
o
ten
tial f
o
r
f
u
tu
r
e
ad
v
an
ce
m
en
ts
an
d
ap
p
licatio
n
s
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
T
h
is
r
esear
ch
d
id
n
o
t
r
ec
eiv
e
an
y
s
p
ec
if
ic
g
r
an
t
f
r
o
m
f
u
n
d
i
n
g
ag
en
cies
in
th
e
p
u
b
lic,
co
m
m
er
cial
,
o
r
not
-
f
o
r
-
p
r
o
f
it secto
r
s
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
Me
g
h
an
a
Desh
m
u
k
h
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Am
it Ga
ik
wad
✓
✓
✓
✓
✓
✓
✓
Sn
eh
al
Ku
ch
e
✓
✓
✓
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
All th
e
au
th
o
r
s
claim
th
at
th
e
y
h
av
e
n
o
b
attle
o
f
h
o
b
b
y
.
DATA AV
AI
L
AB
I
L
I
T
Y)
C
o
d
e
av
ailab
ilit
y
: T
h
e
co
d
e
m
ig
h
t b
e
to
b
e
h
a
d
u
p
o
n
r
e
q
u
est
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
14
,
No
.
3
,
Dec
em
b
er
20
25
:
903
-
9
1
3
912
RE
F
E
R
E
NC
E
S
[
1
]
D
.
B
h
a
r
d
w
a
j
a
n
d
V
.
T
o
d
w
a
l
,
“
A
n
o
v
e
l
d
e
e
p
l
e
a
r
n
i
n
g
mo
d
e
l
f
o
r
i
n
d
o
o
r
-
o
u
t
d
o
o
r
s
c
e
n
e
c
l
a
ssi
f
i
c
a
t
i
o
n
u
s
i
n
g
VGG
-
16
d
e
e
p
C
N
N
,
”
Wo
r
l
d
J
o
u
rn
a
l
o
f
Re
s
e
a
r
c
h
a
n
d
R
e
v
i
e
w
(
WJ
RR)
,
v
o
l
.
1
3
,
n
o
.
1
,
p
p
.
2
7
–
3
5
,
2
0
2
1
,
d
o
i
:
1
0
.
5
8
1
5
/
i
j
i
s
a
.
2
0
2
1
.
0
1
.
0
7
.
[
2
]
K
.
X
u
,
H
.
H
u
a
n
g
,
P
.
D
e
n
g
,
a
n
d
Y
.
Li
,
“
D
e
e
p
f
e
a
t
u
r
e
a
g
g
r
e
g
a
t
i
o
n
f
r
a
me
w
o
r
k
d
r
i
v
e
n
b
y
g
r
a
p
h
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
t
w
o
r
k
f
o
r
sce
n
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
i
n
r
e
m
o
t
e
s
e
n
s
i
n
g
,
”
I
EE
E
T
r
a
n
s
a
c
t
i
o
n
s
o
n
N
e
u
r
a
l
N
e
t
w
o
r
k
s
a
n
d
L
e
a
r
n
i
n
g
S
y
s
t
e
m
s
,
v
o
l
.
3
3
,
n
o
.
1
0
,
p
p
.
5
7
5
1
–
5
7
6
5
,
O
c
t
.
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
TN
N
LS
.
2
0
2
1
.
3
0
7
1
3
6
9
.
[
3
]
S
.
B
i
a
n
c
o
,
G
.
C
i
o
c
c
a
,
C
.
C
u
s
a
n
o
,
a
n
d
R
.
S
c
h
e
t
t
i
n
i
,
“
I
mp
r
o
v
i
n
g
c
o
l
o
r
c
o
n
s
t
a
n
c
y
u
si
n
g
i
n
d
o
o
r
–
o
u
t
d
o
o
r
i
m
a
g
e
c
l
a
s
si
f
i
c
a
t
i
o
n
,
”
I
EEE
T
r
a
n
s
a
c
t
i
o
n
s
o
n
I
m
a
g
e
Pro
c
e
ss
i
n
g
,
v
o
l
.
1
7
,
n
o
.
1
2
,
p
p
.
2
3
8
1
–
2
3
9
2
,
D
e
c
.
2
0
0
8
,
d
o
i
:
1
0
.
1
1
0
9
/
TI
P
.
2
0
0
8
.
2
0
0
6
6
6
1
.
[
4
]
Y
.
H
u
a
n
g
,
X
.
C
a
o
,
Q
.
W
a
n
g
,
B
.
Z
h
a
n
g
,
X
.
Z
h
e
n
,
a
n
d
X
.
Li
,
“
L
o
n
g
-
s
h
o
r
t
-
t
e
r
m
f
e
a
t
u
r
e
s
f
o
r
d
y
n
a
m
i
c
s
c
e
n
e
c
l
a
ssi
f
i
c
a
t
i
o
n
,
”
I
EEE
T
ra
n
s
a
c
t
i
o
n
s
o
n
C
i
r
c
u
i
t
s
a
n
d
S
y
st
e
m
s
f
o
r
V
i
d
e
o
T
e
c
h
n
o
l
o
g
y
,
v
o
l
.
2
9
,
n
o
.
4
,
p
p
.
1
0
3
8
–
1
0
4
7
,
A
p
r
.
2
0
1
9
,
d
o
i
:
1
0
.
1
1
0
9
/
T
C
S
V
T.
2
0
1
8
.
2
8
2
3
3
6
0
.
[
5
]
X
.
X
u
,
Y
.
C
h
e
n
,
J.
Zh
a
n
g
,
Y
.
C
h
e
n
,
P
.
A
n
a
n
d
h
a
n
,
a
n
d
A
.
M
a
n
i
c
k
a
m,
“
R
ET
R
A
C
T
ED
A
R
TI
C
LE
:
a
n
o
v
e
l
a
p
p
r
o
a
c
h
f
o
r
s
c
e
n
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
f
r
o
m
r
e
m
o
t
e
se
n
si
n
g
i
m
a
g
e
s
u
si
n
g
d
e
e
p
l
e
a
r
n
i
n
g
me
t
h
o
d
s,”
E
u
ro
p
e
a
n
J
o
u
r
n
a
l
o
f
Re
m
o
t
e
S
e
n
s
i
n
g
,
v
o
l
.
5
4
,
n
o
.
su
p
2
,
p
p
.
3
8
3
–
3
9
5
,
M
a
r
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
8
0
/
2
2
7
9
7
2
5
4
.
2
0
2
0
.
1
7
9
0
9
9
5
.
[
6
]
R
.
P
.
P
.
M
.
S
o
u
z
a
,
L
.
J
.
A
.
d
o
s
S
a
n
t
o
s,
G
.
T.
P
.
C
o
i
mb
r
a
,
F
.
A
.
S
i
l
v
a
,
a
n
d
T.
R
.
M
.
B
.
S
i
l
v
a
,
“
A
b
i
g
d
a
t
a
-
d
r
i
v
e
n
h
y
b
r
i
d
so
l
u
t
i
o
n
t
o
t
h
e
i
n
d
o
o
r
-
o
u
t
d
o
o
r
d
e
t
e
c
t
i
o
n
p
r
o
b
l
e
m
,
”
B
i
g
D
a
t
a
R
e
se
a
rc
h
,
v
o
l
.
2
4
,
p
.
1
0
0
1
9
4
,
M
a
y
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
b
d
r
.
2
0
2
1
.
1
0
0
1
9
4
.
[
7
]
G
.
C
h
e
n
g
,
X
.
S
u
n
,
K
.
L
i
,
L.
G
u
o
,
a
n
d
J.
H
a
n
,
“
P
e
r
t
u
r
b
a
t
i
o
n
-
se
e
k
i
n
g
g
e
n
e
r
a
t
i
v
e
a
d
v
e
r
sari
a
l
n
e
t
w
o
r
k
s
:
a
d
e
f
e
n
s
e
f
r
a
mew
o
r
k
f
o
r
r
e
mo
t
e
se
n
s
i
n
g
i
ma
g
e
s
c
e
n
e
c
l
a
s
s
i
f
i
c
a
t
i
o
n
,
”
I
EEE
T
r
a
n
s
a
c
t
i
o
n
s
o
n
G
e
o
s
c
i
e
n
c
e
a
n
d
Re
m
o
t
e
S
e
n
s
i
n
g
,
v
o
l
.
6
0
,
p
p
.
1
–
1
1
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
TG
R
S
.
2
0
2
1
.
3
0
8
1
4
2
1
.
[
8
]
X
.
Y
a
o
,
X
.
F
e
n
g
,
J
.
H
a
n
,
G
.
C
h
e
n
g
,
a
n
d
L
.
G
u
o
,
“
A
u
t
o
ma
t
i
c
w
e
a
k
l
y
s
u
p
e
r
v
i
se
d
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
f
r
o
m
h
i
g
h
sp
a
t
i
a
l
r
e
s
o
l
u
t
i
o
n
r
e
mo
t
e
s
e
n
s
i
n
g
i
m
a
g
e
s
v
i
a
d
y
n
a
mi
c
c
u
r
r
i
c
u
l
u
m
l
e
a
r
n
i
n
g
,
”
I
E
EE
T
r
a
n
sa
c
t
i
o
n
s
o
n
G
e
o
sc
i
e
n
c
e
a
n
d
Re
m
o
t
e
S
e
n
si
n
g
,
v
o
l
.
5
9
,
n
o
.
1
,
p
p
.
6
7
5
–
6
8
5
,
J
a
n
.
2
0
2
1
,
d
o
i
:
1
0
.
1
1
0
9
/
TG
R
S
.
2
0
2
0
.
2
9
9
1
4
0
7
.
[
9
]
X
.
H
u
a
n
g
,
X
.
H
a
n
,
S
.
M
a
,
T.
Li
n
,
a
n
d
J.
G
o
n
g
,
“
M
o
n
i
t
o
r
i
n
g
e
c
o
s
y
st
e
m
s
e
r
v
i
c
e
c
h
a
n
g
e
i
n
t
h
e
C
i
t
y
o
f
S
h
e
n
z
h
e
n
b
y
t
h
e
u
s
e
o
f
h
i
g
h
‐
r
e
so
l
u
t
i
o
n
r
e
m
o
t
e
l
y
se
n
se
d
i
ma
g
e
r
y
a
n
d
d
e
e
p
l
e
a
r
n
i
n
g
,
”
L
a
n
d
D
e
g
ra
d
a
t
i
o
n
&
D
e
v
e
l
o
p
m
e
n
t
,
v
o
l
.
3
0
,
n
o
.
1
2
,
p
p
.
1
4
9
0
–
1
5
0
1
,
Ju
l
.
2
0
1
9
,
d
o
i
:
1
0
.
1
0
0
2
/
l
d
r
.
3
3
3
7
.
[
1
0
]
Q
.
Zh
u
,
Y
.
Zh
o
n
g
,
L
.
Z
h
a
n
g
,
a
n
d
D
.
Li
,
“
A
d
a
p
t
i
v
e
d
e
e
p
s
p
a
r
s
e
s
e
ma
n
t
i
c
mo
d
e
l
i
n
g
f
r
a
mew
o
r
k
f
o
r
h
i
g
h
s
p
a
t
i
a
l
r
e
so
l
u
t
i
o
n
i
m
a
g
e
sce
n
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
I
E
EE
T
ra
n
s
a
c
t
i
o
n
s
o
n
G
e
o
s
c
i
e
n
c
e
a
n
d
Re
m
o
t
e
S
e
n
s
i
n
g
,
v
o
l
.
5
6
,
n
o
.
1
0
,
p
p
.
1
–
1
6
,
2
0
1
8
,
d
o
i
:
1
0
.
1
1
0
9
/
TG
R
S
.
2
0
1
8
.
2
8
3
3
2
9
3
.
[
1
1
]
Q
.
W
e
n
g
,
Z.
M
a
o
,
J.
Li
n
,
a
n
d
W
.
G
u
o
,
“
La
n
d
-
u
se
c
l
a
s
si
f
i
c
a
t
i
o
n
v
i
a
e
x
t
r
e
me
l
e
a
r
n
i
n
g
c
l
a
ss
i
f
i
e
r
b
a
s
e
d
o
n
d
e
e
p
c
o
n
v
o
l
u
t
i
o
n
a
l
f
e
a
t
u
r
e
s,
”
I
EEE
G
e
o
sc
i
e
n
c
e
a
n
d
Re
m
o
t
e
S
e
n
s
i
n
g
L
e
t
t
e
rs
,
v
o
l
.
1
4
,
n
o
.
5
,
p
p
.
7
0
4
–
7
0
8
,
M
a
y
2
0
1
7
,
d
o
i
:
1
0
.
1
1
0
9
/
LG
R
S
.
2
0
1
7
.
2
6
7
2
6
4
3
.
[
1
2
]
X
.
L
i
e
t
a
l
.
,
“
S
p
a
t
i
a
l
–
t
e
mp
o
r
a
l
su
p
e
r
-
r
e
s
o
l
u
t
i
o
n
l
a
n
d
c
o
v
e
r
ma
p
p
i
n
g
w
i
t
h
a
l
o
c
a
l
sp
a
t
i
a
l
–
t
e
m
p
o
r
a
l
d
e
p
e
n
d
e
n
c
e
mo
d
e
l
,
”
I
EEE
T
r
a
n
sa
c
t
i
o
n
s
o
n
G
e
o
s
c
i
e
n
c
e
a
n
d
R
e
m
o
t
e
S
e
n
s
i
n
g
,
v
o
l
.
5
7
,
n
o
.
7
,
p
p
.
4
9
5
1
–
4
9
6
6
,
Ju
l
.
2
0
1
9
,
d
o
i
:
1
0
.
1
1
0
9
/
TG
R
S
.
2
0
1
9
.
2
8
9
4
7
7
3
.
[
1
3
]
R
.
P
e
r
e
i
r
a
,
N
.
G
o
n
c
a
l
v
e
s
,
L.
G
a
r
r
o
t
e
,
T
.
B
a
r
r
o
s,
A
.
Lo
p
e
s,
a
n
d
U
.
J.
N
u
n
e
s,
“
D
e
e
p
-
l
e
a
r
n
i
n
g
b
a
s
e
d
g
l
o
b
a
l
a
n
d
sema
n
t
i
c
f
e
a
t
u
r
e
f
u
si
o
n
f
o
r
i
n
d
o
o
r
sce
n
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
i
n
2
0
2
0
I
E
EE
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
Au
t
o
n
o
m
o
u
s
R
o
b
o
t
S
y
s
t
e
m
s
a
n
d
C
o
m
p
e
t
i
t
i
o
n
s
(
I
C
ARS
C
)
,
A
p
r
.
2
0
2
0
,
p
p
.
6
7
–
7
3
,
d
o
i
:
1
0
.
1
1
0
9
/
I
C
A
R
S
C
4
9
9
2
1
.
2
0
2
0
.
9
0
9
6
0
6
8
.
[
1
4
]
G
.
C
h
e
n
g
,
J.
H
a
n
,
a
n
d
X
.
L
u
,
“
R
e
m
o
t
e
se
n
s
i
n
g
i
ma
g
e
sce
n
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
:
b
e
n
c
h
m
a
r
k
a
n
d
st
a
t
e
o
f
t
h
e
a
r
t
,
”
P
ro
c
e
e
d
i
n
g
s
o
f
t
h
e
I
EEE
,
v
o
l
.
1
0
5
,
n
o
.
1
0
,
p
p
.
1
8
6
5
–
1
8
8
3
,
O
c
t
.
2
0
1
7
,
d
o
i
:
1
0
.
1
1
0
9
/
JP
R
O
C
.
2
0
1
7
.
2
6
7
5
9
9
8
.
[
1
5
]
G
.
C
h
e
n
g
,
X
.
X
i
e
,
J.
H
a
n
,
L.
G
u
o
,
a
n
d
G
.
-
S
.
X
i
a
,
“
R
e
m
o
t
e
se
n
s
i
n
g
i
m
a
g
e
sc
e
n
e
c
l
a
ssi
f
i
c
a
t
i
o
n
m
e
e
t
s
d
e
e
p
l
e
a
r
n
i
n
g
:
c
h
a
l
l
e
n
g
e
s,
met
h
o
d
s,
b
e
n
c
h
m
a
r
k
s
,
a
n
d
o
p
p
o
r
t
u
n
i
t
i
e
s,
”
I
EEE
J
o
u
r
n
a
l
o
f
S
e
l
e
c
t
e
d
T
o
p
i
c
s
i
n
Ap
p
l
i
e
d
E
a
rt
h
O
b
ser
v
a
t
i
o
n
s
a
n
d
Re
m
o
t
e
S
e
n
si
n
g
,
v
o
l
.
1
3
,
p
p
.
3
7
3
5
–
3
7
5
6
,
2
0
2
0
,
d
o
i
:
1
0
.
1
1
0
9
/
JS
TA
R
S
.
2
0
2
0
.
3
0
0
5
4
0
3
.
[
1
6
]
Y
.
L
i
u
,
Y
.
Zh
o
n
g
,
F
.
F
e
i
,
a
n
d
L.
Z
h
a
n
g
,
“
S
c
e
n
e
sem
a
n
t
i
c
c
l
a
ssi
f
i
c
a
t
i
o
n
b
a
s
e
d
o
n
r
a
n
d
o
m
-
s
c
a
l
e
st
r
e
t
c
h
e
d
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
f
o
r
h
i
g
h
-
s
p
a
t
i
a
l
r
e
s
o
l
u
t
i
o
n
r
e
mo
t
e
se
n
s
i
n
g
i
ma
g
e
r
y
,
”
i
n
2
0
1
6
I
EEE
I
n
t
e
r
n
a
t
i
o
n
a
l
G
e
o
s
c
i
e
n
c
e
a
n
d
Re
m
o
t
e
S
e
n
si
n
g
S
y
m
p
o
s
i
u
m
(
I
G
ARS
S
)
,
J
u
l
.
2
0
1
6
,
v
o
l
.
2
0
1
6
-
N
o
v
e
m
,
p
p
.
7
6
3
–
7
6
6
,
d
o
i
:
1
0
.
1
1
0
9
/
I
G
A
R
S
S
.
2
0
1
6
.
7
7
2
9
1
9
2
.
[
1
7
]
J.
-
S
.
C
h
o
u
a
n
d
D
.
-
N
.
Tr
u
o
n
g
,
“
A
n
o
v
e
l
me
t
a
h
e
u
r
i
s
t
i
c
o
p
t
i
mi
z
e
r
i
n
s
p
i
r
e
d
b
y
b
e
h
a
v
i
o
r
o
f
j
e
l
l
y
f
i
sh
i
n
o
c
e
a
n
,
”
A
p
p
l
i
e
d
M
a
t
h
e
m
a
t
i
c
s
a
n
d
C
o
m
p
u
t
a
t
i
o
n
,
v
o
l
.
3
8
9
,
p
.
1
2
5
5
3
5
,
Ja
n
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
a
mc.
2
0
2
0
.
1
2
5
5
3
5
.
[
1
8
]
S
.
M
i
r
j
a
l
i
l
i
a
n
d
A
.
Le
w
i
s
,
“
Th
e
w
h
a
l
e
o
p
t
i
mi
z
a
t
i
o
n
a
l
g
o
r
i
t
h
m
,
”
Ad
v
a
n
c
e
s i
n
En
g
i
n
e
e
ri
n
g
S
o
f
t
w
a
re
,
v
o
l
.
9
5
,
p
p
.
5
1
–
6
7
,
M
a
y
2
0
1
6
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
a
d
v
e
n
g
so
f
t
.
2
0
1
6
.
0
1
.
0
0
8
.
[
1
9
]
Y
.
W
u
,
J.
D
o
n
g
,
a
n
d
J
.
C
h
e
n
,
“
Y
O
L
O
-
D
C
N
e
t
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
rn
a
l
o
n
S
e
m
a
n
t
i
c
W
e
b
a
n
d
I
n
f
o
rm
a
t
i
o
n
S
y
s
t
e
m
s
,
v
o
l
.
2
0
,
n
o
.
1
,
p
p
.
1
–
2
3
,
F
e
b
.
2
0
2
4
,
d
o
i
:
1
0
.
4
0
1
8
/
I
JS
W
I
S
.
3
3
9
0
0
0
.
[
2
0
]
Y
.
H
a
o
,
L.
W
a
n
g
,
Y
.
Li
u
,
a
n
d
J.
F
a
n
,
“
I
n
f
o
r
m
a
t
i
o
n
e
n
t
r
o
p
y
a
u
g
m
e
n
t
e
d
h
i
g
h
d
e
n
s
i
t
y
c
r
o
w
d
c
o
u
n
t
i
n
g
n
e
t
w
o
r
k
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
n
S
e
m
a
n
t
i
c
W
e
b
a
n
d
I
n
f
o
rm
a
t
i
o
n
S
y
st
e
m
s
,
v
o
l
.
1
8
,
n
o
.
1
,
p
p
.
1
–
1
5
,
A
p
r
.
2
0
2
2
,
d
o
i
:
1
0
.
4
0
1
8
/
I
JS
W
I
S
.
2
9
7
1
4
4
.
[
2
1
]
M
.
M
a
s
u
d
e
t
a
l
.
,
“
P
r
e
-
t
r
a
i
n
e
d
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
f
o
r
b
r
e
a
st
c
a
n
c
e
r
d
e
t
e
c
t
i
o
n
u
s
i
n
g
u
l
t
r
a
so
u
n
d
i
m
a
g
e
s,”
AC
M
T
ra
n
s
a
c
t
i
o
n
s
o
n
I
n
t
e
rn
e
t
T
e
c
h
n
o
l
o
g
y
,
v
o
l
.
2
1
,
n
o
.
4
,
p
p
.
1
–
1
7
,
N
o
v
.
2
0
2
1
,
d
o
i
:
1
0
.
1
1
4
5
/
3
4
1
8
3
5
5
.
[
2
2
]
K
.
T.
C
h
u
i
e
t
a
l
.
,
“
M
u
l
t
i
r
o
u
n
d
t
r
a
n
s
f
e
r
l
e
a
r
n
i
n
g
a
n
d
mo
d
i
f
i
e
d
g
e
n
e
r
a
t
i
v
e
a
d
v
e
r
s
a
r
i
a
l
n
e
t
w
o
r
k
f
o
r
l
u
n
g
c
a
n
c
e
r
d
e
t
e
c
t
i
o
n
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
I
n
t
e
l
l
i
g
e
n
t
S
y
st
e
m
s
,
v
o
l
.
2
0
2
3
,
n
o
.
1
,
Ja
n
.
2
0
2
3
,
d
o
i
:
1
0
.
1
1
5
5
/
2
0
2
3
/
6
3
7
6
2
7
5
.
[
2
3
]
Q
.
Z
h
u
,
Y
.
Z
h
o
n
g
,
B
.
Z
h
a
o
,
G
.
-
S
.
X
i
a
,
a
n
d
L.
Z
h
a
n
g
,
“
B
a
g
-
of
-
v
i
s
u
a
l
w
o
r
d
s sc
e
n
e
c
l
a
ssi
f
i
e
r
w
i
t
h
l
o
c
a
l
a
n
d
g
l
o
b
a
l
f
e
a
t
u
r
e
s
f
o
r
h
i
g
h
sp
a
t
i
a
l
r
e
s
o
l
u
t
i
o
n
r
e
mo
t
e
se
n
s
i
n
g
i
ma
g
e
r
y
,
”
i
n
I
E
EE
G
e
o
sc
i
e
n
c
e
a
n
d
Re
m
o
t
e
S
e
n
s
i
n
g
L
e
t
t
e
rs
,
2
0
1
6
,
v
o
l
.
1
3
,
n
o
.
6
,
p
p
.
7
4
7
–
7
5
1
,
d
o
i
:
1
0
.
1
1
0
9
/
LG
R
S
.
2
0
1
6
.
2
5
5
9
0
3
8
.
[
2
4
]
L.
H
u
a
n
g
,
C
.
C
h
e
n
,
W
.
Li
,
a
n
d
Q
.
D
u
,
“
R
e
mo
t
e
s
e
n
si
n
g
i
m
a
g
e
sc
e
n
e
c
l
a
ssi
f
i
c
a
t
i
o
n
u
si
n
g
m
u
l
t
i
-
sc
a
l
e
c
o
m
p
l
e
t
e
d
l
o
c
a
l
b
i
n
a
r
y
p
a
t
t
e
r
n
s
a
n
d
f
i
sh
e
r
v
e
c
t
o
r
s
,
”
Re
m
o
t
e
S
e
n
s
i
n
g
,
v
o
l
.
8
,
n
o
.
6
,
p
.
4
8
3
,
Ju
n
.
2
0
1
6
,
d
o
i
:
1
0
.
3
3
9
0
/
r
s
8
0
6
0
4
8
3
.
[
2
5
]
E.
M
o
h
a
m
e
d
,
K
.
S
i
r
l
a
n
t
z
i
s
a
n
d
G
.
H
o
w
e
l
l
s,
“
I
n
d
o
o
r
/
o
u
t
d
o
o
r
s
e
ma
n
t
i
c
s
e
g
me
n
t
a
t
i
o
n
u
si
n
g
d
e
e
p
l
e
a
r
n
i
n
g
f
o
r
v
i
s
u
a
l
l
y
i
mp
a
i
r
e
d
w
h
e
e
l
c
h
a
i
r
u
sers,
”
i
n
I
EEE
A
c
c
e
s
s,
v
o
l
.
9
,
p
p
.
1
4
7
9
1
4
-
1
4
7
9
3
2
,
2
0
2
1
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
1
.
3
1
2
3
9
5
2
.
[
2
6
]
P
.
A
l
v
e
s
e
t
a
l
.
,
“
A
n
o
v
e
l
a
p
p
r
o
a
c
h
f
o
r
u
ser
e
q
u
i
p
m
e
n
t
i
n
d
o
o
r
/
o
u
t
d
o
o
r
c
l
a
ss
i
f
i
c
a
t
i
o
n
i
n
m
o
b
i
l
e
n
e
t
w
o
r
k
s
,
”
i
n
I
EEE
A
c
c
e
ss
,
v
o
l
.
9
,
p
p
.
1
6
2
6
7
1
-
1
6
2
6
8
6
,
2
0
2
1
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
1
.
3
1
3
0
4
2
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.