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.
15
,
No
.
5
,
Octo
b
er
20
26
,
p
p
.
4
5
2
3
~
4
5
3
9
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
5
.
p
p
4
5
2
3
-
4
5
3
9
4523
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Co
nfidence
-
driv
e
n f
ea
tu
re sele
ctio
n and
ensemb
le cl
a
ss
ificatio
n
using
ResNe
t,
A
le
x
Ne
t,
and
Dense
N
et
for
scene
recog
nition
Su
m
a
t
hi K
rish
na
m
urt
hy
1
,
Sh
et
t
y
P
ra
m
o
d K
um
a
r
1
,
H
a
la
s
ur
Rudra
pp
a
M
a
ha
dev
a
s
w
a
m
y
2
,
Uj
wa
la
B
a
g
epa
lli Sa
t
is
hb
a
b
u
1
1
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
J
a
w
a
h
a
r
l
a
l
N
e
h
r
u
N
e
w
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
V
i
sv
e
sv
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
s
i
t
y
,
B
e
l
a
g
a
v
i
,
I
n
d
i
a
2
Jag
a
d
g
u
r
u
S
r
i
S
h
i
v
a
r
a
t
h
r
e
e
s
h
w
a
r
a
U
n
i
v
e
r
s
i
t
y
,
N
o
i
d
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
Oct
23
,
2
0
2
5
R
ev
is
ed
J
u
l
17
,
2
0
2
6
Acc
ep
ted
Au
g
12
,
2
0
2
6
In
d
o
o
r
sc
e
n
e
re
c
o
g
n
it
io
n
is
a
si
g
n
ifi
c
a
n
t
p
r
o
b
lem
b
e
c
a
u
se
o
f
it
s
stru
c
tu
ra
l
c
o
m
p
lex
it
y
,
il
l
u
m
in
a
ti
o
n
v
a
riati
o
n
s
,
a
n
d
v
isu
a
l
sim
il
a
rit
ies
b
e
t
we
e
n
th
e
v
a
rio
u
s
sc
e
n
e
c
las
se
s
.
Re
c
e
n
t
a
p
p
ro
a
c
h
e
s
u
se
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s
(
CNN
s
)
a
n
d
tran
sfe
r
lea
rn
i
n
g
to
e
x
trac
t
fe
a
tu
re
s
fo
r
c
las
sifica
ti
o
n
.
In
t
h
e
se
m
e
th
o
d
s,
a
ll
t
h
e
fe
a
tu
re
s
we
re
c
o
n
sid
e
re
d
f
o
r
c
las
sifica
ti
o
n
,
wh
ich
lea
d
s
t
o
a
n
in
c
re
a
se
in
c
o
m
p
u
tatio
n
a
l
c
o
m
p
lex
it
y
a
n
d
train
in
g
ti
m
e
.
Th
i
s
lea
d
s
t
o
m
o
re
ti
m
e
c
o
n
su
m
p
t
io
n
f
o
r
p
r
o
c
e
ss
in
g
.
T
o
a
d
d
re
ss
th
e
se
c
h
a
ll
e
n
g
e
s,
th
is
stu
d
y
p
r
o
p
o
se
d
a
h
y
b
rid
in
d
o
o
r
s
c
e
n
e
re
c
o
g
n
it
i
o
n
fra
m
e
wo
rk
t
h
a
t
in
teg
ra
tes
il
lu
m
in
a
ti
o
n
-
a
wa
re
p
re
-
p
r
o
c
e
ss
in
g
,
m
u
lt
i
-
m
o
d
e
l
d
e
e
p
fe
a
tu
re
e
x
trac
ti
o
n
,
a
n
d
c
o
n
fid
e
n
c
e
-
b
a
se
d
fe
a
tu
re
f
u
sio
n
.
Th
e
p
re
-
p
ro
c
e
ss
in
g
tec
h
n
iq
u
e
u
se
s
c
o
n
tras
t
-
li
m
it
e
d
a
d
a
p
ti
v
e
h
ist
o
g
ra
m
e
q
u
a
li
z
a
ti
o
n
(CLAHE)
a
n
d
Re
ti
n
e
x
e
n
h
a
n
c
e
m
e
n
t.
De
e
p
fe
a
tu
re
s
a
re
e
x
trac
ted
u
sin
g
Ale
x
Kriz
h
e
v
s
k
y
n
e
two
r
k
(
Ale
x
Ne
t
)
,
d
e
n
se
ly
c
o
n
n
e
c
ted
c
o
n
v
o
lu
ti
o
n
a
l
n
e
two
r
k
(
De
n
se
Ne
t
)
,
a
n
d
re
sid
u
a
l
n
e
two
rk
1
0
1
lay
e
rs
(
Re
sN
e
t1
0
1
)
to
c
a
p
t
u
re
sp
a
ti
a
l
a
n
d
se
m
a
n
ti
c
in
fo
rm
a
ti
o
n
.
Th
e
fe
a
tu
re
s
e
x
trac
t
e
d
a
re
fe
d
to
c
o
n
f
id
e
n
c
e
-
d
ri
v
e
n
wo
rld
c
u
p
o
p
ti
m
iza
ti
o
n
(CD
-
WCO)
t
o
e
x
trac
t
th
e
sig
n
ifi
c
a
n
t
fe
a
tu
re
s.
Th
e
fi
n
a
l
re
p
re
se
n
tatio
n
is
c
las
sified
u
si
n
g
a
d
e
e
p
li
q
u
i
d
sta
te
m
a
c
h
in
e
(
D
LS
M
)
c
las
sifier
to
e
n
h
a
n
c
e
c
las
s
d
isc
rimin
a
ti
o
n
.
O
u
r
e
x
p
e
rime
n
ts
,
c
o
n
d
u
c
ted
o
n
th
e
NY
U
in
d
o
o
r
sc
e
n
e
d
a
tas
e
t
with
2
0
sc
e
n
e
c
a
teg
o
ries
a
n
d
th
e
M
I
T
In
d
o
o
r
-
1
5
c
las
s
su
b
se
t
re
su
l
ts
i
n
d
ica
te
t
h
e
e
ffe
c
ti
v
e
n
e
ss
o
f
t
h
e
p
ro
p
o
se
d
fr
a
m
e
wo
rk
,
a
c
h
iev
in
g
a
n
o
v
e
ra
ll
c
las
sific
a
ti
o
n
a
c
c
u
ra
c
y
o
f
9
6
.
5
%
a
n
d
9
5
.
2
%
,
re
sp
e
c
ti
v
e
ly
.
K
ey
w
o
r
d
s
:
C
o
n
f
id
en
ce
-
d
r
iv
en
w
o
r
ld
cu
p
o
p
tim
izatio
n
C
o
n
tr
ast
-
lim
ited
ad
ap
tiv
e
h
is
to
g
r
am
eq
u
aliza
tio
n
Dee
p
liq
u
id
s
tate
m
ac
h
in
e
class
if
ier
Featu
r
e
ex
tr
ac
tio
n
Featu
r
e
s
elec
tio
n
R
etin
ex
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
m
ath
i K
r
is
h
n
am
u
r
th
y
Dep
ar
tm
e
n
t
o
f
E
l
ec
t
r
o
n
ics
a
n
d
C
o
m
m
u
n
i
ca
ti
o
n
E
n
g
i
n
ee
r
i
n
g
,
J
awa
h
a
r
l
al
Ne
h
r
u
N
ew
C
o
ll
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
Vis
v
esv
ar
ay
a
T
ec
h
n
o
l
o
g
ical
Un
iv
er
s
ity
B
elag
av
i,
I
n
d
ia
E
m
ail:
s
u
m
ath
ik
@
jn
n
ce
.
ac
.
in
1.
I
NT
RO
D
UCT
I
O
N
Scen
e
r
ec
o
g
n
itio
n
is
a
h
ig
h
-
le
v
el
co
m
p
u
ter
v
is
io
n
task
th
at
in
v
o
lv
es
class
if
y
in
g
an
im
ag
e
o
r
v
id
eo
b
ased
o
n
its
o
v
er
all
en
v
ir
o
n
m
en
t
an
d
co
n
te
x
t,
s
u
ch
as
id
en
t
if
y
in
g
it
as
a
“
b
ea
ch
,
”
“
k
itch
e
n
,
”
o
r
“
s
tr
ee
t”
[
1
]
.
Scen
e
r
ec
o
g
n
itio
n
n
o
t
o
n
ly
co
n
ce
n
tr
ates
o
n
id
en
tify
i
n
g
t
h
e
class
o
f
th
e
im
ag
e,
b
u
t
also
f
o
cu
s
es
o
n
th
e
s
em
an
tic
m
ea
n
in
g
ass
o
ciate
d
with
th
e
p
ar
ticu
lar
s
ce
n
e
[
2
]
.
Scen
e
id
e
n
tific
atio
n
h
as
a
wid
e
r
an
g
e
o
f
ap
p
licatio
n
s
in
r
o
b
o
tics
,
au
to
n
o
m
o
u
s
d
r
iv
i
n
g
,
an
d
m
ed
ical
im
ag
in
g
.
to
ac
h
iev
e
t
h
e
s
ce
n
e
class
if
icatio
n
task
,
th
e
s
tep
s
ar
e
f
ea
tu
r
e
d
etec
tio
n
,
f
ea
tu
r
e
d
escr
ip
tio
n
,
an
d
class
if
icatio
n
[
3
]
.
Scen
e
r
ec
o
g
n
iti
o
n
u
s
in
g
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
an
d
tr
an
s
f
er
lear
n
in
g
-
b
ased
tech
n
i
q
u
es
h
as
s
ig
n
i
f
ican
tly
in
cr
ea
s
ed
th
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.
15
,
No
.
5
,
Octo
b
er
20
26
:
4
5
2
3
-
4
5
3
9
4524
ef
f
icien
cy
o
f
s
ce
n
e
class
if
icatio
n
.
T
h
e
s
ce
n
e
ca
n
b
e
s
tatic
o
r
d
y
n
am
ic.
Dy
n
am
ic
s
ce
n
es
ar
e
r
ef
er
r
ed
t
o
as
v
id
eo
s
eq
u
e
n
ce
s
,
an
d
s
tatic
s
ce
n
es
ar
e
s
till
im
ag
es
[
4
]
.
Static
s
ce
n
e
an
aly
s
is
is
d
ep
en
d
en
t
o
n
s
in
g
le
-
f
r
am
e
in
f
o
r
m
atio
n
.
Static
s
ce
n
e
class
if
icatio
n
is
f
u
r
th
er
d
iv
id
ed
in
to
in
d
o
o
r
an
d
o
u
td
o
o
r
s
c
en
e
class
if
icatio
n
.
Ou
td
o
o
r
s
ce
n
es,
s
u
c
h
as
s
tr
ee
ts
,
f
o
r
ests
,
o
r
b
ea
ch
es
—
g
en
er
ally
co
n
tain
s
im
ilar
o
b
ject
s
ets,
wh
ile
in
d
o
o
r
s
ce
n
es
—
s
u
ch
as
clas
s
r
o
o
m
s
,
b
ed
r
o
o
m
s
,
o
r
liv
in
g
r
o
o
m
s
—
f
ea
tu
r
e
a
v
ar
iety
o
f
o
b
je
cts
with
d
if
f
er
in
g
d
im
en
s
io
n
s
,
s
p
atial
ar
r
an
g
em
e
n
ts
,
an
d
lev
els
o
f
o
cc
lu
s
io
n
.
I
n
d
o
o
r
s
ce
n
e
class
if
icatio
n
p
la
y
s
a
v
e
r
y
i
m
p
o
r
ta
n
t
r
o
le
in
s
ev
er
al
co
m
p
u
ter
ap
p
licatio
n
s
,
in
clu
d
in
g
r
o
b
o
tics
n
av
ig
atio
n
,
h
u
m
an
-
co
m
p
u
te
r
in
ter
ac
tio
n
,
an
d
co
n
ten
t
-
awa
r
e
s
y
s
tem
s
.
Un
lik
e
o
u
td
o
o
r
s
ce
n
e
im
ag
es,
in
d
o
o
r
s
ce
n
es
h
av
e
h
i
g
h
in
tr
a
-
class
v
ar
iab
ilit
y
,
s
tr
o
n
g
illu
m
in
atio
n
ch
an
g
es,
an
d
th
e
p
r
esen
ce
o
f
v
is
u
ally
s
im
ilar
o
b
jects
ac
r
o
s
s
d
if
f
er
en
t
ca
t
eg
o
r
ies.
T
h
e
ab
o
v
e
ch
allen
g
es m
ak
e
in
d
o
o
r
s
ce
n
e
class
if
icatio
n
a
n
o
n
-
tr
iv
ial
task
,
ev
en
with
r
ec
en
t a
d
v
an
ce
s
i
n
d
ee
p
lear
n
in
g
.
E
ar
ly
ap
p
r
o
ac
h
es
to
s
ce
n
e
r
ec
o
g
n
itio
n
r
elied
o
n
h
an
d
cr
af
ted
f
ea
tu
r
es,
s
u
c
h
as
s
ca
le
-
in
v
a
r
ian
t
f
ea
tu
r
e
tr
an
s
f
o
r
m
(
SIFT
)
,
h
is
to
g
r
a
m
o
f
o
r
ie
n
t
ed
g
r
ad
ie
n
ts
(
H
OG)
,
a
n
d
g
l
o
b
al
i
m
a
g
e
d
esc
r
i
p
t
o
r
(
GI
ST)
d
esc
r
i
p
t
o
r
s
[
2
]
.
T
h
e
ab
o
v
e
ap
p
r
o
ac
h
es
ca
p
t
u
r
e
lo
ca
l
tex
tu
r
e
an
d
s
tr
u
ct
u
r
al
in
f
o
r
m
atio
n
;
th
ey
ar
e
h
ig
h
ly
s
en
s
itiv
e
to
illu
m
in
atio
n
v
a
r
iatio
n
s
.
T
h
es
e
h
an
d
c
r
af
ted
f
ea
tu
r
es
r
eq
u
ir
e
m
an
u
al
d
esig
n
a
n
d
lack
t
h
e
r
ep
r
esen
tatio
n
al
ca
p
ac
ity
to
g
en
er
alize
ac
r
o
s
s
co
m
p
lex
i
n
d
o
o
r
s
ce
n
e
im
ag
es,
r
esu
ltin
g
in
lim
ited
class
if
icatio
n
ac
cu
r
ac
y
.
W
ith
th
e
ev
o
lu
tio
n
o
f
d
ee
p
lear
n
in
g
,
s
ce
n
e
r
ec
o
g
n
itio
n
u
s
in
g
C
NN
an
d
tr
an
s
f
er
lear
n
in
g
-
b
ased
tech
n
iq
u
es
h
as
s
ig
n
if
ican
tly
in
cr
ea
s
ed
th
e
ef
f
i
cien
cy
o
f
s
ce
n
e
class
if
icatio
n
.
C
NN
s
b
ec
am
e
th
e
d
o
m
in
an
t
ap
p
r
o
ac
h
f
o
r
s
ce
n
e
class
if
icatio
n
d
u
e
to
th
eir
ab
il
ity
to
lear
n
h
ier
ar
ch
ical
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
.
Ar
ch
itectu
r
es
s
u
ch
as
r
esid
u
al
n
etwo
r
k
(
R
esNet)
,
Alex
Kr
i
zh
ev
s
k
y
n
etwo
r
k
(
Alex
Net)
,
an
d
d
en
s
ely
co
n
n
ec
ted
c
o
n
v
o
lu
tio
n
al
n
etwo
r
k
(
Den
s
eNe
t)
h
av
e
s
h
o
wn
ex
c
ep
tio
n
al
p
er
f
o
r
m
an
ce
i
n
ex
tr
ac
tin
g
f
ea
tu
r
es
f
r
o
m
im
a
g
e
d
ata,
wh
ich
aid
s
in
class
if
y
in
g
s
ce
n
e
im
ag
es
[
5
]
.
Ho
wev
er
,
d
esp
ite
th
e
e
x
ce
p
ti
o
n
al
p
er
f
o
r
m
a
n
ce
f
r
o
m
th
ese
m
o
d
els,
ch
allen
g
es
r
em
ain
in
s
elec
tin
g
d
iv
er
s
if
ie
d
an
d
d
is
tin
g
u
is
h
ed
f
ea
tu
r
es
f
r
o
m
t
h
e
e
n
tire
f
ea
t
u
r
e
s
et
e
x
tr
ac
ted
f
r
o
m
t
h
e
d
ee
p
lear
n
in
g
m
o
d
els an
d
c
o
m
b
in
i
n
g
th
e
d
iv
e
r
s
e
f
ea
tu
r
es f
r
o
m
th
e
m
u
ltip
le
m
o
d
els.
Ho
wev
er
,
s
in
g
le
-
m
o
d
el
C
NN
-
b
ased
ap
p
r
o
ac
h
es
h
av
e
s
ev
er
a
l
lim
itatio
n
s
.
First,
th
ey
o
f
ten
o
v
er
f
it
to
d
o
m
in
an
t
v
is
u
al
p
atter
n
s
,
le
ad
in
g
to
p
o
o
r
g
en
er
aliza
tio
n
ac
r
o
s
s
v
ar
io
u
s
in
d
o
o
r
s
ce
n
es.
Seco
n
d
,
d
ee
p
er
n
etwo
r
k
s
in
c
r
ea
s
e
co
m
p
u
ta
tio
n
al
co
m
p
lex
ity
an
d
tr
ai
n
in
g
tim
e,
m
a
k
in
g
th
em
less
s
u
itab
le
f
o
r
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
en
v
i
r
o
n
m
en
ts
.
R
ec
en
t
ad
v
an
ce
m
en
ts
in
f
ea
tu
r
e
f
u
s
io
n
a
n
d
en
s
em
b
le
lear
n
in
g
tech
n
iq
u
es
h
av
e
ad
d
r
ess
ed
th
ese
d
r
awb
a
ck
s
o
f
co
m
b
in
in
g
f
ea
tu
r
es
f
r
o
m
m
u
ltip
le
d
ee
p
ar
ch
itectu
r
es
.
T
h
e
m
u
lti
-
m
o
d
el
f
u
s
io
n
im
p
r
o
v
es
class
if
icatio
n
ac
cu
r
ac
y
b
ec
au
s
e
o
f
h
ig
h
r
ep
r
esen
tatio
n
al
d
iv
er
s
ity
,
em
p
lo
y
in
g
s
im
p
le
co
n
ca
ten
atio
n
o
r
f
ix
e
d
-
weig
h
t
av
er
ag
in
g
.
Su
ch
s
tatic
f
u
s
io
n
m
ec
h
an
is
m
s
f
ail
to
ad
a
p
t
to
class
-
s
p
ec
if
ic
co
n
f
id
en
ce
v
ar
iatio
n
s
an
d
m
a
y
in
tr
o
d
u
ce
r
ed
u
n
d
an
t
o
r
n
o
is
y
f
ea
tu
r
es,
u
ltima
tely
lim
itin
g
p
er
f
o
r
m
a
n
c
e
g
ain
s
.
T
o
o
v
e
r
co
m
e
th
is
lim
itatio
n
,
i
n
o
u
r
p
ap
e
r
,
we
h
a
v
e
p
r
o
p
o
s
e
d
a
n
o
v
el
m
eth
o
d
o
f
a
h
y
b
r
id
s
ce
n
e
r
ec
o
g
n
itio
n
f
r
am
ewo
r
k
co
m
b
in
in
g
p
r
e
p
r
o
ce
s
s
in
g
tech
n
iq
u
es
s
u
ch
as
R
etin
ex
an
d
c
o
n
tr
ast
-
lim
ited
ad
ap
tiv
e
h
is
to
g
r
am
eq
u
aliza
tio
n
(
C
L
AHE
)
[
6
]
f
o
r
illu
m
in
atio
n
n
o
r
m
aliza
tio
n
with
co
n
f
id
e
n
ce
-
d
r
i
v
en
wo
r
ld
cu
p
o
p
tim
izatio
n
(
C
D
-
W
C
O)
f
o
r
d
iv
er
s
if
ied
f
ea
tu
r
e
s
elec
tio
n
a
n
d
a
d
ee
p
liq
u
id
s
tate
m
ac
h
in
e
(
DL
S
M)
class
if
ier
at
th
e
class
if
icatio
n
s
tag
e,
with
a
s
p
ik
in
g
win
n
er
-
tak
es
-
all
(
SW
T
A)
[
7
]
lay
er
t
h
at
en
h
a
n
ce
s
n
o
n
-
lin
ea
r
d
ec
is
io
n
b
o
u
n
d
ar
ies,
to
class
if
y
th
e
s
ce
n
e
im
ag
es,
r
esu
ltin
g
in
im
p
r
o
v
e
d
illu
m
in
atio
n
r
o
b
u
s
tn
ess
an
d
s
ce
n
e
d
is
cr
im
in
atio
n
p
er
f
o
r
m
a
n
ce
.
2.
L
I
T
E
R
AT
U
RE
R
E
VI
E
W
Ov
er
th
e
y
ea
r
s
,
m
an
y
d
ee
p
lear
n
in
g
a
p
p
r
o
ac
h
es
h
av
e
b
ee
n
in
tr
o
d
u
ce
d
in
s
ce
n
e
r
e
co
g
n
itio
n
.
ar
ch
itectu
r
es
s
u
ch
as
R
esNet,
Alex
Net,
an
d
Den
s
eNe
t
h
av
e
s
h
o
wn
ex
ce
p
tio
n
al
p
er
f
o
r
m
an
ce
in
ex
tr
ac
tin
g
f
ea
tu
r
es
f
r
o
m
i
m
ag
e
d
ata,
wh
ich
aid
s
in
class
if
y
in
g
s
ce
n
e
im
ag
es
.
Ho
we
v
er
,
d
esp
ite
th
e
ex
ce
p
tio
n
al
p
er
f
o
r
m
an
ce
f
r
o
m
th
ese
m
o
d
els,
ch
allen
g
es
r
em
ain
in
s
ele
ctin
g
d
iv
e
r
s
if
ied
an
d
d
is
tin
g
u
is
h
ed
f
ea
tu
r
es
f
r
o
m
th
e
en
tire
f
ea
tu
r
e
s
et
ex
tr
ac
te
d
f
r
o
m
th
e
d
ee
p
lear
n
in
g
m
o
d
els
an
d
co
m
b
in
in
g
th
e
d
iv
er
s
e
f
ea
tu
r
es
f
r
o
m
th
e
m
u
ltip
le
m
o
d
els
[
5
]
.
T
o
a
d
d
r
e
s
s
th
ese
ch
allen
g
es,
in
th
is
r
es
ea
r
ch
wo
r
k
p
r
o
p
o
s
ed
an
atten
tio
n
m
ec
h
a
n
is
m
th
at
ac
ts
as
a
p
o
wer
f
u
l
to
o
l
in
f
o
c
u
s
in
g
o
n
th
e
m
o
s
t
im
p
o
r
ta
n
t
f
ea
tu
r
es
r
ep
r
esen
tin
g
th
e
in
f
o
r
m
ativ
e
d
ata
o
f
th
e
im
ag
es
,
an
d
in
p
ar
allel,
en
s
em
b
le
lear
n
in
g
tech
n
iq
u
es
co
m
b
in
e
th
e
s
tr
en
g
th
s
o
f
m
u
ltip
le
m
o
d
els,
lev
er
ag
in
g
th
eir
d
iv
er
s
e
f
ea
t
u
r
e
e
x
tr
ac
tio
n
ca
p
ab
ilit
ies
to
ac
h
iev
e
m
o
r
e
ef
f
icien
t
p
e
r
f
o
r
m
an
ce
th
a
n
an
y
s
in
g
le
m
o
d
el
alo
n
e.
I
n
th
is
p
r
o
p
o
s
ed
r
e
s
ea
r
ch
wo
r
k
in
tr
o
d
u
ce
d
a
co
n
f
id
en
ce
-
m
o
d
u
lated
wo
r
ld
cu
p
o
p
tim
izatio
n
(
C
M
-
W
C
O)
[
8
]
tech
n
iq
u
e
f
o
r
f
ea
tu
r
e
s
elec
tio
n
an
d
an
en
s
em
b
le
lear
n
in
g
f
r
am
ew
o
r
k
th
at
in
teg
r
ates
R
es
Net,
Alex
Net,
an
d
Den
s
eNe
t f
o
r
r
o
b
u
s
t scen
e
r
ec
o
g
n
itio
n
.
T
h
is
r
esear
ch
wo
r
k
c
o
n
ce
n
tr
a
tes
o
n
co
n
tr
ib
u
tin
g
to
th
e
d
ev
elo
p
m
en
t
o
f
m
o
r
e
ef
f
icien
t
an
d
ac
cu
r
ate
s
ce
n
e
r
ec
o
g
n
itio
n
s
y
s
tem
s
b
y
co
m
b
in
in
g
ad
v
a
n
ce
d
d
ee
p
lear
n
in
g
ar
c
h
itectu
r
es
with
atten
tio
n
-
g
u
i
d
ed
f
ea
tu
r
e
s
elec
tio
n
an
d
en
s
em
b
le
s
tr
ate
g
ies,
p
r
o
v
id
i
n
g
a
way
f
o
r
im
p
r
o
v
ed
s
ce
n
e
r
ec
o
g
n
itio
n
.
Scen
e
class
if
icatio
n
is
a
f
u
n
d
am
e
n
tal
task
with
in
th
e
b
r
o
ad
er
f
ield
o
f
s
ce
n
e
u
n
d
e
r
s
tan
d
in
g
,
wh
er
e
th
e
o
b
jectiv
e
is
to
in
ter
p
r
et
a
g
iv
e
n
s
ce
n
e
an
d
ass
ig
n
it
to
a
s
p
e
cif
ic
ca
teg
o
r
y
.
T
h
e
ea
r
ly
r
ec
o
g
n
itio
n
p
r
o
ce
s
s
was
co
m
p
le
tely
d
ep
en
d
en
t
o
n
p
ix
el
-
b
ased
r
ep
r
esen
tatio
n
s
,
wh
er
e
th
e
r
aw
p
i
x
els
o
f
th
e
s
ce
n
e
wer
e
u
s
ed
f
o
r
class
if
icat
io
n
.
T
h
ese
m
eth
o
d
s
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
C
o
n
fid
en
ce
-
d
r
iven
fea
tu
r
e
s
elec
tio
n
a
n
d
en
s
emb
le
cla
s
s
ifica
tio
n
u
s
in
g
…
(
S
u
ma
th
i Kr
is
h
n
a
mu
r
th
y
)
4525
o
f
ten
f
ac
ed
lim
itatio
n
s
r
elate
d
to
im
ag
e
r
eso
lu
tio
n
an
d
th
e
g
r
an
u
lar
ity
o
f
p
ix
el
in
f
o
r
m
atio
n
.
T
o
o
v
er
c
o
m
e
th
is
lim
itatio
n
,
r
esear
ch
er
s
h
av
e
s
h
if
ted
to
o
b
ject
-
b
ased
class
if
icatio
n
,
wh
ich
u
tili
ze
s
o
b
ject
-
b
ased
in
f
o
r
m
atio
n
to
class
if
y
th
e
s
ce
n
e
im
ag
es.
B
u
t
o
b
ject
-
b
ased
class
if
icatio
n
d
o
es
n
o
t
co
n
ce
n
tr
ate
o
n
th
e
s
em
an
tic
co
n
tex
t
o
f
th
e
s
ce
n
e
im
ag
es
.
T
h
is
ch
allen
g
e
h
as
led
r
esear
ch
er
s
to
m
o
v
e
to
war
d
s
th
e
s
em
an
tic
u
n
d
e
r
s
tan
d
in
g
o
f
s
ce
n
e
im
ag
es.
Scen
e
class
if
icatio
n
in
v
o
lv
in
g
th
e
s
em
an
tic
in
f
o
r
m
atio
n
d
ep
e
n
d
s
o
n
i
d
en
tify
in
g
t
h
e
d
is
tin
ct
f
ea
tu
r
es
o
f
th
e
im
ag
e.
So
m
e
m
eth
o
d
s
,
s
u
ch
as
SIFT
[
9
]
an
d
HOG,
ar
e
b
ased
o
n
h
an
d
c
r
af
ted
f
ea
t
u
r
e
tech
n
iq
u
es
th
at
wer
e
u
s
ed
to
class
if
y
im
ag
es,
b
u
t
as
th
ese
tec
h
n
iq
u
es
r
ely
o
n
m
a
n
u
ally
d
esig
n
e
d
f
ilter
s
,
th
ey
ar
e
p
r
o
n
e
to
m
is
s
in
g
im
p
o
r
tan
t
f
ea
tu
r
e
d
et
ails
,
lim
itin
g
th
eir
e
f
f
ec
tiv
en
e
s
s
.
Ad
d
itio
n
al
f
ea
tu
r
e
e
x
tr
ac
tio
n
tech
n
iq
u
es
h
av
e
b
ee
n
d
ev
elo
p
ed
,
s
u
ch
as
s
p
atial
p
y
r
am
id
m
atch
in
g
an
d
f
is
h
er
v
ec
to
r
co
d
i
n
g
.
On
ce
th
ese
f
ea
tu
r
es
h
av
e
b
ee
n
ex
tr
ac
ted
,
th
e
y
h
a
v
e
b
ee
n
f
e
d
in
to
th
e
m
ac
h
in
e
lear
n
in
g
al
g
o
r
ith
m
s
lik
e
k
-
m
ea
n
s
cl
u
s
ter
in
g
,
s
p
ar
s
e
co
d
in
g
,
an
d
p
r
in
cip
al
co
m
p
o
n
en
t
an
al
y
s
is
(
PC
A)
[
1
0
]
to
p
er
f
o
r
m
th
e
class
if
icatio
n
f
u
n
ctio
n
.
T
h
e
ef
f
icien
cy
o
f
th
es
e
m
eth
o
d
s
is
d
ep
en
d
en
t o
n
h
o
w
th
e
f
ea
tu
r
es a
r
e
f
e
d
in
to
t
h
e
m
o
d
el.
T
h
e
ad
v
en
t
o
f
th
e
d
ee
p
lea
r
n
i
n
g
co
n
ce
p
t
o
f
ex
tr
ac
tin
g
h
ig
h
-
lev
el
f
ea
tu
r
es
in
th
e
g
iv
e
n
d
ataset
h
as
p
av
ed
way
n
ew
er
a
h
as
b
ee
n
p
av
ed
.
Dee
p
lear
n
in
g
m
o
d
els
ex
tr
ac
t
th
e
f
ea
tu
r
es
au
to
m
atica
lly
f
r
o
m
th
e
d
ata.
C
NNs
ar
e
u
s
ed
to
ex
tr
ac
t
th
e
f
ea
tu
r
es
au
to
m
atica
lly
.
Ma
jo
r
d
ev
elo
p
m
e
n
ts
in
th
e
d
o
m
ain
o
f
d
ee
p
lear
n
in
g
h
ap
p
en
e
d
af
ter
th
e
d
ev
elo
p
m
e
n
t
o
f
t
h
e
Alex
Net
n
etwo
r
k
,
wh
ich
is
p
r
e
-
tr
ain
e
d
o
n
th
e
I
m
ag
eNe
t
d
ataset.
Ma
n
y
d
ee
p
lear
n
in
g
m
o
d
els
h
av
e
b
e
en
d
ev
elo
p
e
d
,
wh
ich
in
clu
d
e
v
is
u
al
g
eo
m
etr
y
g
r
o
u
p
n
etwo
r
k
(
VGGN
et)
[
1
1
]
,
Go
o
g
L
eNe
t
[
1
2
]
,
s
q
u
ee
ze
n
et
wo
r
k
(
Sq
u
ee
ze
Net)
,
R
esNet,
an
d
Den
s
eNe
t.
W
ith
th
e
ad
v
en
t
o
f
lar
g
e
im
ag
e
d
atasets
an
d
p
o
wer
f
u
l
p
r
o
ce
s
s
o
r
s
,
m
an
y
r
esear
ch
wo
r
k
s
ar
e
b
ein
g
ca
r
r
ied
o
u
t
with
th
ese
d
ee
p
C
NN
m
o
d
els
in
th
e
ar
ea
o
f
s
ce
n
e
id
en
tific
atio
n
.
Pre
tr
ain
ed
C
NN
m
o
d
els
au
to
m
atica
lly
ex
tr
ac
t
th
e
f
ea
tu
r
e
s
f
r
o
m
th
e
d
ataset,
an
d
th
ese
f
ea
tu
r
es
ar
e
m
illi
o
n
s
in
n
u
m
b
e
r
.
T
h
ese
f
ea
t
u
r
es
ar
e
th
en
g
iv
e
n
as
in
p
u
t
to
th
e
cl
ass
if
ier
lay
er
o
f
th
e
C
NN
m
o
d
el.
Ov
er
th
e
r
ec
en
t
y
ea
r
s
,
th
e
f
ea
tu
r
e
s
elec
tio
n
an
d
en
s
em
b
lin
g
tech
n
iq
u
es
h
av
e
s
ig
n
if
ican
tly
im
p
r
o
v
e
d
in
p
e
r
f
o
r
m
in
g
th
e
s
ce
n
e
r
ec
o
g
n
itio
n
task
.
Mu
c
h
r
esear
ch
h
as
tak
en
p
lace
ex
p
l
o
r
in
g
o
p
tim
izatio
n
alg
o
r
ith
m
s
,
f
ea
tu
r
e
s
elec
tio
n
,
an
d
f
ea
tu
r
e
ag
g
r
eg
atio
n
tec
h
n
iq
u
es
to
im
p
r
o
v
e
th
e
class
if
icatio
n
ac
cu
r
ac
y
an
d
en
s
u
r
e
r
o
b
u
s
tn
ess
ac
r
o
s
s
d
iv
er
s
e
d
atasets
.
Am
o
n
g
o
p
tim
izatio
n
alg
o
r
ith
m
s
,
s
u
ch
as
p
ar
ticle
s
war
m
o
p
tim
izatio
n
(
PS
O)
[
1
3
]
is
a
m
etah
eu
r
is
tic
o
p
tim
izatio
n
alg
o
r
ith
m
p
r
o
p
o
s
ed
b
y
Ken
n
ed
y
an
d
E
b
er
h
ar
t
in
1
9
9
5
.
T
h
is
alg
o
r
ith
m
m
ai
n
ly
aim
s
at
f
in
d
in
g
th
e
o
p
tim
al
v
alu
es
f
o
r
s
p
ec
i
f
ic
p
ar
a
m
eter
s
to
f
u
lf
ill
th
e
d
esig
n
r
eq
u
ir
em
en
ts
with
t
h
e
lo
west
p
o
s
s
ib
le
co
m
p
lex
ity
a
n
d
co
s
t.
T
h
is
al
g
o
r
ith
m
m
ain
ly
c
o
n
ce
n
t
r
ates
o
n
s
o
l
v
in
g
t
h
e
o
p
tim
al
co
m
b
in
atio
n
o
f
f
ea
t
u
r
e
co
m
p
en
s
atio
n
c
o
ef
f
icien
ts
f
o
r
m
u
ltip
le
co
m
p
u
tatio
n
f
ac
to
r
s
s
o
th
at
th
e
r
ec
o
g
n
itio
n
ef
f
icien
cy
o
f
t
h
e
s
y
s
tem
ca
n
b
e
im
p
r
o
v
ed
,
b
u
t
th
e
lim
it
atio
n
o
f
th
is
alg
o
r
ith
m
is
th
at
it
ca
n
b
e
tr
ap
p
ed
in
lo
ca
l
m
in
i
m
a
if
th
e
d
iv
er
s
ity
o
f
th
e
d
ata
is
r
ed
u
ce
d
,
r
e
q
u
ir
i
n
g
tu
n
i
n
g
o
f
p
a
r
am
eter
s
lik
e
i
n
er
tia
weig
h
t
an
d
lear
n
in
g
co
ef
f
icien
ts
.
Gen
etic
alg
o
r
ith
m
(
GA)
is
an
o
p
tim
izatio
n
alg
o
r
ith
m
th
at
was
in
tr
o
d
u
ce
d
b
y
Ho
llan
d
in
th
e
1
9
7
0
s
[
1
4
]
an
d
h
as
s
in
ce
b
ee
n
wid
ely
ap
p
lied
i
n
th
e
d
o
m
ain
s
o
f
en
g
in
ee
r
in
g
,
ar
tific
ia
l
in
tellig
en
ce
,
an
d
im
ag
e
class
if
icatio
n
task
s
.
GA
s
ea
r
ch
es
f
o
r
th
e
n
u
m
b
er
o
f
lay
er
s
,
lear
n
in
g
r
ate,
b
atch
s
ize,
an
d
ac
tiv
atio
n
f
u
n
ctio
n
s
f
o
r
th
e
o
p
tim
al
co
m
b
in
atio
n
t
o
m
ax
im
ize
c
lass
if
icatio
n
p
er
f
o
r
m
an
ce
.
I
t
ca
n
h
an
d
le
co
m
p
lex
a
n
d
lar
g
e
d
atasets
.
I
t
au
to
m
atica
lly
s
elec
ts
th
e
h
ig
h
-
p
er
f
o
r
m
in
g
n
etwo
r
k
ar
ch
it
ec
tu
r
e.
I
ts
s
ea
r
ch
s
p
ac
e
in
cl
u
d
es
b
o
th
n
etwo
r
k
s
tr
u
ctu
r
e
co
n
f
ig
u
r
atio
n
a
n
d
h
y
p
er
p
a
r
am
eter
s
.
B
u
t
th
ese
alg
o
r
ith
m
s
ar
e
c
o
m
p
u
tatio
n
ally
ex
p
en
s
iv
e
f
o
r
la
r
g
e
p
o
p
u
latio
n
s
o
r
d
ee
p
n
etwo
r
k
s
an
d
h
a
v
e
s
lo
wer
c
o
n
v
e
r
g
en
ce
co
m
p
ar
ed
to
g
r
ad
ie
n
t
-
b
ased
m
eth
o
d
s
.
An
t
c
o
lo
n
y
o
p
tim
izatio
n
(
AC
O
)
[
1
5
]
is
a
n
o
p
tim
izatio
n
al
g
o
r
ith
m
in
tr
o
d
u
ce
d
b
y
D
o
r
ig
o
i
n
th
e
ea
r
ly
1
9
9
0
s
[
1
6
]
.
AC
O
was
o
r
ig
in
ally
d
ev
elo
p
ed
f
o
r
co
m
b
in
ato
r
ial
p
r
o
b
lem
s
s
u
ch
as
th
e
tr
av
elin
g
s
alesm
an
p
r
o
b
lem
;
it
h
as
b
ee
n
s
u
cc
ess
f
u
lly
ad
ap
ted
f
o
r
o
p
ti
m
izatio
n
task
s
in
im
ag
e
cla
s
s
i
f
icatio
n
.
AC
O
h
as
b
ee
n
u
s
ed
to
o
p
tim
ize
v
ar
io
u
s
n
eu
r
al
n
etwo
r
k
p
latf
o
r
m
s
,
s
u
c
h
as
th
e
n
u
m
b
er
o
f
h
i
d
d
en
lay
er
s
an
d
lear
n
in
g
r
ates,
an
d
it
i
s
u
s
ed
to
s
elec
t
th
e
m
o
s
t
r
elev
an
t
s
u
b
s
et
o
f
f
ea
t
u
r
e
d
ata
f
r
o
m
h
i
g
h
-
d
i
m
en
s
io
n
al
co
m
p
le
x
d
ata,
wh
ich
h
el
p
s
in
r
e
d
u
cin
g
th
e
co
m
p
u
tatio
n
al
co
m
p
lex
ity
a
n
d
in
cr
ea
s
es
th
e
class
if
icatio
n
ac
cu
r
ac
y
an
d
in
ter
p
r
etab
le
cl
ass
if
icatio
n
m
o
d
els.
B
u
t
th
ese
ar
e
co
m
p
u
tatio
n
ally
in
ten
s
iv
e
alg
o
r
ith
m
s
an
d
m
a
y
n
ee
d
in
teg
r
atio
n
with
l
o
ca
l
s
e
ar
ch
tech
n
iq
u
es
to
im
p
r
o
v
e
e
f
f
icien
cy
,
an
d
c
ar
ef
u
l tu
n
in
g
is
r
eq
u
ir
ed
to
av
o
id
p
r
em
atu
r
e
co
n
v
er
g
e
n
ce
.
T
h
e
r
esear
ch
h
as
b
ee
n
co
n
d
u
cted
b
ased
o
n
m
u
lti
-
m
o
d
al
f
e
atu
r
e
s
elec
tio
n
,
wh
ich
r
ef
er
s
t
o
s
elec
tin
g
f
ea
tu
r
es
f
r
o
m
h
eter
o
g
e
n
eo
u
s
d
ata
m
o
d
alities
s
u
ch
as
im
ag
es
,
d
ep
th
m
ap
s
,
tex
t,
a
n
d
au
d
io
.
I
n
co
n
tr
ast,
m
u
lti
-
m
o
d
el
f
ea
tu
r
e
s
elec
tio
n
f
o
cu
s
es
o
n
lev
er
a
g
in
g
m
u
ltip
le
d
ee
p
lear
n
in
g
ar
ch
itectu
r
es
to
ex
tr
ac
t
an
d
co
m
b
in
e
f
ea
t
u
r
es
f
r
o
m
th
e
s
am
e
d
ata
m
o
d
ality
.
R
ath
er
th
an
in
teg
r
atin
g
d
ata
f
r
o
m
d
if
f
er
en
t
s
o
u
r
ce
s
,
m
u
lti
-
m
o
d
el
tech
n
i
q
u
es
em
p
h
asize
th
e
d
iv
er
s
ity
o
f
r
ep
r
e
s
en
tatio
n
s
lear
n
ed
b
y
m
o
d
el
s
s
u
ch
as
R
e
s
Net,
Alex
Net,
an
d
De
n
s
eNe
t.
T
h
is
r
esear
ch
wo
r
k
is
in
s
p
ir
ed
b
y
t
h
e
m
u
lti
-
m
o
d
al
tech
n
iq
u
es
c
o
n
ce
p
t
o
f
co
m
b
in
in
g
d
iv
er
s
e
in
f
o
r
m
atio
n
s
o
u
r
ce
s
.
Mu
lti
-
m
o
d
el
f
ea
tu
r
e
s
elec
tio
n
is
p
ar
ticu
la
r
ly
s
u
itab
le
f
o
r
s
ce
n
ar
io
s
wh
er
e
a
s
in
g
le
m
o
d
ality
is
av
ailab
le,
b
u
t r
ich
er
f
ea
t
u
r
e
r
e
p
r
esen
tatio
n
s
ar
e
d
esire
d
.
3.
M
E
T
H
O
D
T
h
e
p
r
o
p
o
s
ed
f
lo
w
d
iag
r
am
i
s
d
ep
icted
in
Fig
u
r
e
1
.
T
h
is
r
esear
ch
wo
r
k
h
as
d
esig
n
ed
o
u
r
m
o
d
el
u
s
in
g
m
u
ltip
le
p
r
e
-
tr
ain
ed
m
o
d
els
to
ef
f
icien
tly
r
ec
o
g
n
ize
th
e
d
if
f
e
r
en
t
class
es
o
f
s
ce
n
e
im
ag
es
with
h
ig
h
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.
15
,
No
.
5
,
Octo
b
er
20
26
:
4
5
2
3
-
4
5
3
9
4526
ac
cu
r
ac
y
.
T
h
e
d
etailed
p
r
o
p
o
s
ed
m
eth
o
d
f
o
r
class
if
y
in
g
s
c
en
e
im
ag
es
is
s
h
o
wn
i
n
Fig
u
r
e
2
.
T
h
is
p
r
o
p
o
s
ed
m
eth
o
d
in
tr
o
d
u
ce
d
a
n
ap
p
r
o
ac
h
to
p
er
f
o
r
m
s
ce
n
e
r
ec
o
g
n
itio
n
b
y
co
m
b
in
in
g
th
e
C
L
AHE
an
d
R
etin
ex
illu
m
in
atio
n
tech
n
iq
u
e
[
1
7
]
,
with
th
e
CD
-
WCO
alg
o
r
ith
m
f
o
r
s
elec
tin
g
d
iv
er
s
if
ied
f
ea
tu
r
es.
T
h
is
u
n
iq
u
e
co
m
b
in
atio
n
m
eth
o
d
o
f
th
e
C
L
AHE
an
d
R
etin
ex
illu
m
i
n
atio
n
tech
n
iq
u
e
with
th
e
CD
-
W
C
O
alg
o
r
ith
m
en
h
an
ce
s
class
if
icatio
n
ac
cu
r
ac
y
an
d
ad
d
r
ess
es
v
ar
io
u
s
ch
allen
g
es,
s
u
ch
as
r
ed
u
n
d
a
n
cy
,
o
v
er
f
itti
n
g
,
an
d
th
e
r
em
o
v
al
o
f
i
r
r
elev
an
t
f
ea
tu
r
es
.
Featu
r
e
e
x
tr
ac
tio
n
p
lay
s
a
p
r
o
m
in
en
t
r
o
le
i
n
th
e
im
ag
e
class
if
icatio
n
task
.
I
n
th
e
tr
ad
itio
n
al
m
eth
o
d
,
a
s
in
g
l
e
d
ee
p
lear
n
in
g
m
o
d
el
was
u
s
ed
to
ex
tr
ac
t
th
e
f
ea
tu
r
es
f
r
o
m
th
e
im
ag
es
,
wh
ich
m
ay
b
e
lim
ited
in
its
ca
p
ac
ity
to
ca
p
tu
r
e
d
iv
er
s
e
an
d
co
m
p
le
m
en
tar
y
in
f
o
r
m
atio
n
.
Fig
u
r
e
1
.
Gen
e
r
al
f
lo
w
d
ia
g
r
a
m
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
l
o
g
y
Fig
u
r
e
2
.
Gen
e
r
al
b
lo
c
k
d
iag
r
a
m
o
f
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
with
th
e
f
ea
t
u
r
e
s
elec
tio
n
t
ec
h
n
iq
u
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
C
o
n
fid
en
ce
-
d
r
iven
fea
tu
r
e
s
elec
tio
n
a
n
d
en
s
emb
le
cla
s
s
ifica
tio
n
u
s
in
g
…
(
S
u
ma
th
i Kr
is
h
n
a
mu
r
th
y
)
4527
I
n
t
h
i
s
p
r
o
p
o
s
e
d
w
o
r
k
o
f
m
u
lt
i
-
m
o
d
e
l
f
e
a
t
u
r
e
e
x
tr
a
c
t
io
n
h
a
v
e
co
n
s
i
d
er
ed
e
n
s
e
m
b
l
i
n
g
th
e
f
e
a
tu
r
e
s
e
x
t
r
a
c
te
d
f
r
o
m
m
u
l
t
i
p
l
e
m
o
d
e
l
s
a
n
d
s
e
l
e
c
t
in
g
t
h
e
s
u
p
e
r
io
r
f
e
a
tu
r
e
s
.
T
h
e
f
e
a
tu
r
e
e
x
tr
a
c
ti
o
n
s
u
m
m
a
r
y
t
ab
l
e
f
o
r
t
h
e
m
u
l
t
i
p
l
e
m
o
d
e
l
s
u
s
e
d
i
n
th
i
s
p
r
o
p
o
s
ed
w
o
r
k
i
s
s
h
o
w
n
in
T
a
b
l
e
1
.
T
ab
l
e
1
r
e
p
r
e
s
e
n
t
s
t
h
e
f
e
a
tu
r
e
s
e
x
t
r
a
c
te
d
f
r
o
m
t
h
r
e
e
w
i
d
e
ly
u
s
e
d
d
ee
p
l
e
ar
n
in
g
ar
c
h
i
te
ct
u
r
e
s
w
h
en
a
p
p
l
i
ed
t
o
t
h
e
N
Y
U
i
n
d
o
o
r
s
c
en
e
d
a
t
a
s
e
t
,
co
n
s
i
s
t
i
n
g
o
f
1
,
9
2
0
i
m
a
g
e
s
.
F
ea
t
u
r
e
s
a
r
e
c
o
l
l
ec
te
d
f
r
o
m
t
h
e
l
a
s
t
f
u
l
l
y
co
n
n
ec
t
e
d
l
ay
e
r
(
FC
8
in
A
l
e
x
N
e
t
,
FC
1
0
0
0
in
R
e
s
N
e
t1
0
1
)
o
r
g
l
o
b
a
l
a
v
er
a
g
e
p
o
o
l
in
g
l
a
y
er
(
a
v
g
_
p
o
o
l
in
D
en
s
e
N
e
t2
0
1
)
.
T
h
e
s
e
l
ay
e
r
s
p
r
o
v
i
d
e
co
m
p
a
c
t
,
h
ig
h
-
l
e
v
e
l
s
e
m
a
n
t
ic
r
e
p
r
e
s
e
n
t
a
t
io
n
s
o
f
i
m
ag
e
s
.
T
a
b
le
1
r
e
p
r
e
s
e
n
ts
t
h
e
q
u
an
t
i
t
a
t
iv
e
e
v
i
d
en
c
e
th
a
t
ex
t
r
a
c
t
e
d
f
e
a
tu
r
e
s
c
a
n
b
e
co
m
e
l
a
r
g
e
an
d
r
ed
u
n
d
an
t
.
T
h
er
e
f
o
r
e,
a
p
p
l
y
in
g
a
f
e
a
tu
r
e
s
e
l
e
c
t
i
o
n
m
e
t
h
o
d
i
s
v
er
y
c
r
u
c
i
a
l
.
T
h
e
co
n
ca
ten
atio
n
ap
p
r
o
ac
h
allo
ws
th
e
m
o
d
el
to
p
r
eser
v
e
b
o
th
th
e
lo
w
-
lev
el
d
is
cr
im
in
ativ
e
p
o
wer
o
f
Alex
Net
an
d
th
e
h
ig
h
-
lev
e
l
s
em
an
tic
ab
s
tr
ac
tio
n
s
f
r
o
m
R
esNet
an
d
Den
s
eNe
t
.
T
h
e
e
n
s
em
b
le
ac
ts
as
a
co
n
s
en
s
u
s
m
ec
h
an
is
m
:
th
e
s
tr
en
g
th
s
o
f
o
n
e
m
o
d
el
co
m
p
en
s
ate
f
o
r
th
e
wea
k
n
ess
es
o
f
an
o
th
er
,
r
esu
ltin
g
in
a
m
o
r
e
b
alan
ce
d
an
d
r
o
b
u
s
t
s
ce
n
e
class
if
icatio
n
f
r
am
ewo
r
k
.
T
h
e
R
esNet
m
o
d
el
u
s
es
s
k
ip
co
n
n
ec
tio
n
s
f
o
r
th
e
f
lo
w
o
f
g
r
a
d
ien
ts
an
d
en
a
b
le
s
th
e
lear
n
in
g
o
f
d
ee
p
er
f
ea
t
u
r
es
with
o
u
t
d
eg
r
a
d
atio
n
.
Al
ex
Net
is
o
n
e
o
f
th
e
p
io
n
ee
r
in
g
C
NN
ar
ch
itectu
r
es,
wh
ich
f
ac
ilit
ates
id
en
tify
in
g
p
r
o
m
i
n
en
t
lo
w
-
a
n
d
m
id
-
lev
el
f
ea
tu
r
es
s
u
ch
as
ed
g
es,
s
h
ap
es,
an
d
te
x
tu
r
es.
Den
s
eNe
t
en
h
an
ce
s
f
ea
tu
r
e
r
eu
s
e,
an
d
d
e
n
s
e
co
n
n
ec
tiv
ity
h
elp
s
in
lear
n
in
g
d
etailed
h
ig
h
-
le
v
el
r
ep
r
esen
tatio
n
s
th
at
en
co
d
e
s
p
atial
an
d
s
em
an
tic
r
elatio
n
s
h
ip
s
with
in
th
e
s
ce
n
e.
T
h
is
r
esear
ch
wo
r
k
p
r
o
p
o
s
ed
a
f
ea
tu
r
e
s
elec
tio
n
tech
n
iq
u
e
b
ased
o
n
th
e
C
D
-
W
C
O
tech
n
iq
u
e,
wh
ich
s
elec
ts
th
e
s
u
p
er
io
r
f
ea
tu
r
es
a
n
d
t
h
en
f
ee
d
s
th
e
s
elec
ted
f
ea
tu
r
es
to
th
e
n
ex
t
s
tag
e
o
f
class
if
icatio
n
.
So
m
e
o
f
th
e
co
n
tr
ib
u
tio
n
s
ar
e
s
u
m
m
ar
ized
as f
o
llo
ws:
−
I
m
p
lem
en
tatio
n
o
f
a
h
y
b
r
id
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
e,
wh
ich
in
clu
d
es
C
L
AHE
an
d
R
etin
ex
illu
m
in
atio
n
co
r
r
ec
tio
n
s
,
a
n
d
i
n
teg
r
atin
g
th
ese
p
r
e
-
p
r
o
ce
s
s
in
g
tec
h
n
iq
u
es
with
th
e
C
D
-
W
C
O
r
ed
u
ce
s
n
o
is
e
am
p
lific
atio
n
an
d
en
h
a
n
ce
s
th
e
v
is
ib
ilit
y
o
f
f
in
e
d
etails an
d
t
ex
tu
r
es.
−
Featu
r
e
ex
tr
ac
tio
n
u
s
in
g
m
u
ltip
le
m
o
d
els
en
ab
les
d
iv
er
s
if
ied
lear
n
in
g
in
s
tead
o
f
th
e
ex
tr
ac
tio
n
o
f
f
ea
tu
r
es u
s
in
g
a
s
in
g
le
n
eu
r
al
n
etwo
r
k
m
o
d
el.
−
DL
SM
[
1
8
]
,
u
s
ed
as
a
class
if
ier
in
th
e
f
in
al
s
tag
e
,
p
r
o
v
id
es
s
tate
-
b
ased
class
if
icatio
n
d
ec
i
s
io
n
s
,
lead
in
g
to
im
p
r
o
v
ed
ac
cu
r
ac
y
a
n
d
s
tab
ilit
y
.
T
h
e
s
tep
s
in
v
o
lv
ed
in
o
u
r
p
r
o
p
o
s
ed
wo
r
k
f
o
r
f
ea
tu
r
e
s
elec
tio
n
ar
e
s
u
m
m
ar
ize
d
as f
o
llo
ws:
i)
T
h
e
alg
o
r
ith
m
in
p
u
t is th
e
f
ea
t
u
r
es e
x
tr
ac
ted
f
r
o
m
th
e
m
u
ltip
le
m
o
d
els
.
ii)
T
h
e
co
n
f
id
en
ce
s
co
r
e
f
o
r
ea
ch
s
o
lu
tio
n
is
ca
lcu
lated
.
iii)
Fo
r
ea
ch
o
f
th
e
f
ea
tu
r
e
s
u
b
s
ets
o
b
tain
ed
,
th
e
f
itn
ess
v
alu
e
is
ca
lcu
lated
u
s
in
g
th
e
p
r
e
d
ef
in
ed
f
itn
ess
f
u
n
ctio
n
.
iv
)
T
h
e
f
ea
tu
r
es with
a
h
ig
h
er
v
al
u
e
o
f
c
o
n
f
id
e
n
ce
ar
e
s
elec
ted
as o
p
tim
al
attr
ib
u
tes.
v)
Featu
r
es with
lo
w
co
n
f
id
en
ce
v
alu
es a
r
e
d
is
ca
r
d
ed
.
v
i)
T
h
e
alg
o
r
ith
m
o
u
t
p
u
t is
th
e
in
d
ex
ed
f
ea
t
u
r
e
v
al
u
es u
s
ed
f
o
r
class
if
icatio
n
p
u
r
p
o
s
es
.
T
ab
le
1
.
Featu
r
e
d
im
en
s
io
n
alit
y
o
f
p
r
e
-
tr
ain
e
d
m
o
d
els o
n
th
e
NYU
in
d
o
o
r
s
ce
n
e
d
ataset
M
o
d
e
l
S
i
z
e
o
f
t
h
e
i
n
p
u
t
i
m
a
g
e
F
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
l
a
y
e
r
F
e
a
t
u
r
e
v
e
c
t
o
r
A
l
e
x
N
e
t
2
2
7
×
2
2
7
×
3
FC
8
1
9
2
0
×
4
0
9
6
R
e
sN
e
t
1
0
1
2
2
4
×
2
2
4
×
3
FC
1
0
0
0
1
9
2
0
×
2
0
4
8
D
e
n
seN
e
t
2
0
1
2
2
4
×
2
2
4
×
3
A
vg
p
o
o
l
l
a
y
e
r
1
9
2
0
×
1
9
2
0
3
.
1
.
Da
t
a
s
et
a
nd
pre
-
pro
ce
s
s
ing
I
n
th
is
wo
r
k
,
ex
p
e
r
im
en
tatio
n
is
p
er
f
o
r
m
ed
o
n
th
e
NYU
[
1
9
]
an
d
MI
T
I
n
d
o
o
r
-
1
5
class
s
u
b
s
et
[
3
]
d
atasets
.
T
h
is
wo
r
k
u
s
ed
a
h
y
b
r
id
p
r
e
-
p
r
o
ce
s
s
in
g
tech
n
iq
u
e
,
wh
ich
in
clu
d
es
C
L
AHE
an
d
R
etin
ex
illu
m
in
atio
n
co
r
r
ec
tio
n
s
,
wh
ic
h
im
p
r
o
v
e
th
e
co
n
tr
ast
o
f
th
e
im
ag
es
an
d
n
o
r
m
alize
th
e
lig
h
tin
g
v
a
r
iatio
n
s
in
th
e
im
ag
es.
C
L
AHE
m
ak
es
th
e
s
m
all
d
etails
m
o
r
e
v
is
ib
le
in
th
e
i
m
ag
es.
T
h
e
f
u
s
io
n
o
f
b
o
th
g
iv
es
illu
m
in
atio
n
-
n
o
r
m
alize
d
a
n
d
d
etail
-
en
h
an
ce
d
im
ag
es,
wh
ic
h
p
r
o
v
id
e
r
ic
h
er
an
d
clea
n
er
f
ea
tu
r
es
f
o
r
d
ee
p
lear
n
in
g
m
o
d
els.
C
L
AHE
an
d
R
etin
ex
ar
e
ap
p
lied
s
eq
u
en
tia
lly
.
C
L
AHE
en
s
u
r
es
lo
ca
l
co
n
tr
ast
en
h
an
ce
m
en
t,
wh
ile
R
etin
ex
co
r
r
ec
ts
g
lo
b
al
illu
m
in
atio
n
in
co
n
s
is
ten
cies.
T
h
e
co
m
b
i
n
atio
n
p
r
o
d
u
ce
s
im
ag
es
with
b
alan
ce
d
b
r
ig
h
tn
ess
,
im
p
r
o
v
ed
co
n
tr
ast
,
an
d
r
ed
u
ce
d
lig
h
tin
g
ar
tifa
cts.
T
h
is
h
y
b
r
id
p
r
e
p
r
o
ce
s
s
in
g
p
ip
elin
e
e
n
ab
les
m
o
d
els
to
f
o
c
u
s
o
n
m
ea
n
in
g
f
u
l
s
tr
u
ctu
r
al
an
d
tex
tu
r
al
d
eta
ils
,
wh
ich
is
cr
u
cial
f
o
r
s
ce
n
e
class
if
icatio
n
task
s
wh
er
e
lig
h
tin
g
v
ar
iatio
n
s
ar
e
c
o
m
m
o
n
.
T
h
e
jo
in
t
ap
p
licatio
n
o
f
R
etin
ex
an
d
C
L
AHE
,
f
o
llo
wed
b
y
C
D
-
WCO
,
en
s
u
r
es
b
o
th
g
l
o
b
al
illu
m
i
n
a
tio
n
n
o
r
m
aliza
tio
n
an
d
lo
ca
li
ze
d
co
n
tr
ast
am
p
lific
atio
n
.
R
etin
ex
m
o
d
els
th
e
h
u
m
an
v
is
u
al
s
y
s
tem
to
m
ain
tain
n
atu
r
al
r
e
f
lecta
n
ce
u
n
d
er
v
ar
y
in
g
illu
m
in
atio
n
,
wh
ile
C
L
AHE
ad
ap
tiv
ely
en
h
an
ce
s
m
icr
o
-
co
n
tr
ast
ac
r
o
s
s
s
p
atially
d
is
tin
ct
r
eg
io
n
s
.
Su
b
s
eq
u
en
tly
,
C
D
-
W
C
O
p
er
f
o
r
m
s
ch
an
n
el
-
wis
e
d
u
al
-
weig
h
ted
f
u
s
io
n
to
b
alan
ce
ch
r
o
m
atic
an
d
s
tr
u
ctu
r
al
i
n
f
o
r
m
atio
n
.
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.
15
,
No
.
5
,
Octo
b
er
20
26
:
4
5
2
3
-
4
5
3
9
4528
3
.
2
.
Co
nfidence
-
ba
s
ed
wo
rld c
up
o
ptim
iza
t
io
n
-
ba
s
ed
f
ea
t
ure
s
elec
t
io
n
Featu
r
e
s
elec
tio
n
is
a
v
er
y
im
p
o
r
tan
t
an
d
cr
u
cial
s
tep
in
d
ee
p
lear
n
in
g
.
L
et
=
{
1
,
2
,
3
,
…
,
}
d
en
o
te
th
e
f
ea
tu
r
e
s
et
ex
tr
ac
t
ed
f
r
o
m
m
u
ltip
le
p
r
e
-
tr
ain
e
d
m
o
d
els
(
R
esNet,
Den
s
eNe
t,
an
d
Alex
Net)
.
T
h
e
o
p
tim
izatio
n
p
r
o
ce
s
s
s
ee
k
s
an
o
p
tim
al
s
u
b
s
e
∗
⊆
th
at
m
ax
im
izes
class
if
icatio
n
ac
cu
r
ac
y
wh
ile
m
in
im
izin
g
r
e
d
u
n
d
an
cy
.
T
o
im
p
r
o
v
e
th
e
class
if
icatio
n
ef
f
icien
cy
f
o
r
th
e
e
n
tire
s
et
o
f
ex
tr
ac
ted
f
ea
t
u
r
es
o
b
tain
ed
f
r
o
m
th
e
p
r
e
-
tr
ain
ed
m
o
d
el
,
t
h
e
p
ass
es
th
r
o
u
g
h
th
e
co
n
f
id
en
ce
-
b
ased
weig
h
in
g
m
ec
h
an
is
m
.
Fo
r
ea
ch
o
f
th
e
f
ea
tu
r
es
,
co
n
f
id
e
n
ce
-
b
ased
ca
lcu
lated
,
an
d
it
is
r
ep
r
esen
ted
as
asep
r
esen
ed
y
(
)
.
T
h
e
co
n
f
id
en
ce
s
co
r
e
is
ca
lcu
lated
b
y
(
1
)
.
(
)
=
1
∑
=
1
(
=
ˆ
∣
)
(
1
)
W
h
er
e
is
th
e
to
tal
n
u
m
b
er
o
f
v
alid
atio
n
s
am
p
les
;
is
th
e
g
r
o
u
n
d
tr
u
th
lab
el
o
f
s
am
p
le
;
ˆ
is
th
e
p
r
ed
icted
lab
el
u
s
in
g
f
ea
tu
r
e
;
is
an
in
d
icato
r
f
u
n
ctio
n
th
at
e
q
u
als
1
if
th
e
co
n
d
itio
n
is
tr
u
e
an
d
0
o
th
er
wis
e
;
an
d
(
)
∈
[
0
,
1
]
r
ep
r
esen
ts
th
e
r
eliab
ilit
y
o
f
f
ea
tu
r
e
.
T
h
is
co
n
f
id
en
ce
s
co
r
e
ca
lcu
latio
n
q
u
an
tifie
s
th
e
r
eliab
ilit
y
o
f
ea
ch
f
ea
tu
r
e
to
p
er
f
o
r
m
clas
s
if
icatio
n
.
3
.
3
.
F
i
t
nes
s
f
un
ct
io
n in
t
he
p
ro
po
s
ed
m
et
ho
do
lo
g
y
T
h
e
f
itn
ess
f
u
n
ctio
n
is
th
e
k
ey
cr
iter
io
n
th
at
g
u
id
es
th
e
s
ea
r
c
h
to
war
d
s
o
p
tim
al
s
o
lu
tio
n
s
.
I
n
th
e
ca
s
e
o
f
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
o
f
f
ea
tu
r
e
s
elec
tio
n
,
th
e
f
itn
e
s
s
f
u
n
ctio
n
ev
alu
ates
th
e
q
u
ality
o
f
a
ca
n
d
id
ate
f
ea
tu
r
e
s
u
b
s
et
b
y
b
alan
cin
g
class
if
icatio
n
ac
cu
r
ac
y
,
co
n
f
id
en
ce
,
r
eliab
ilit
y
,
an
d
f
ea
tu
r
e
s
u
b
s
et
co
m
p
ac
tn
ess
.
⊂
b
e
a
ca
n
d
id
ate
f
ea
tu
r
e
s
u
b
s
et,
wh
er
e
F
is
th
e
co
m
p
lete
f
ea
tu
r
e
s
et
ex
tr
ac
ted
f
r
o
m
th
e
en
s
em
b
le
o
f
p
r
e
-
tr
ain
ed
m
o
d
els.
T
h
e
f
itn
es
s
o
f
is
g
iv
e
n
as (
2
)
.
(
)
=
(
)
+
1
∑
∈
(
)
−
⋅
|
|
|
|
(
2
)
W
h
er
e
(
)
is
class
if
icatio
n
ac
cu
r
ac
y
o
b
tain
ed
u
s
in
g
s
u
b
s
et
;
(
)
is
co
n
f
i
d
en
ce
s
co
r
e
o
f
f
ea
t
u
r
e
s
;
|
|
is
n
u
m
b
er
o
f
s
elec
ted
f
ea
tu
r
es
in
s
u
b
s
et
;
|
|
is
to
tal
n
u
m
b
er
o
f
a
v
ailab
le
f
ea
tu
r
es
;
an
d
,
,
is
co
n
tr
o
l p
a
r
am
eter
s
th
at
b
alan
c
e
th
e
co
n
tr
ib
u
tio
n
s
o
f
ac
cu
r
ac
y
,
co
n
f
id
en
ce
,
a
n
d
d
im
en
s
io
n
a
l r
ed
u
ctio
n
.
T
h
e
f
itn
ess
f
u
n
ctio
n
in
co
n
f
i
d
en
ce
-
b
ased
W
C
O
is
a
m
u
lti
-
o
b
jectiv
e
ev
alu
atio
n
m
ea
s
u
r
e
t
h
at
en
s
u
r
es
th
e
f
in
al
s
elec
ted
f
ea
tu
r
e
s
u
b
s
et
is
n
o
t o
n
ly
ac
cu
r
ate
b
u
t
also
r
o
b
u
s
t
an
d
ef
f
icien
t.
T
h
is
ap
p
r
o
ac
h
is
well
-
s
u
ited
f
o
r
h
i
g
h
-
d
im
en
s
io
n
al
f
ea
t
u
r
e
s
p
ac
es
g
en
er
ated
b
y
d
ee
p
e
n
s
em
b
les
lik
e
R
esNet,
Den
s
eNe
t,
an
d
Alex
Net.
B
y
ap
p
ly
in
g
co
n
f
id
e
n
ce
-
b
ased
W
C
O,
th
is
wo
r
k
o
b
tain
s
an
o
p
tim
al
s
u
b
s
et
∗
th
at
co
n
tain
s
o
n
ly
th
e
m
o
s
t
in
f
o
r
m
ativ
e,
r
eliab
le,
an
d
c
o
m
p
ac
t
f
ea
t
u
r
es.
T
h
is
r
ed
u
ce
s
co
m
p
u
tatio
n
al
o
v
e
r
h
ea
d
,
en
h
a
n
ce
s
m
o
d
el
in
ter
p
r
etab
ilit
y
,
a
n
d
im
p
r
o
v
e
s
th
e
g
en
er
aliza
tio
n
ab
ilit
y
o
f
th
e
class
if
ier
ac
r
o
s
s
d
iv
er
s
e
s
ce
n
e
ca
teg
o
r
ies.
Mid
-
lev
el
f
u
s
io
n
was
ch
o
s
en
as
it
o
f
f
er
s
a
b
alan
ce
d
tr
ad
e
-
o
f
f
b
etwe
en
r
ep
r
esen
tatio
n
al
r
ich
n
ess
an
d
co
m
p
u
tatio
n
al
ef
f
icien
cy
.
Un
lik
e
ea
r
ly
f
u
s
io
n
,
wh
ich
o
p
er
ates
d
ir
ec
tly
o
n
r
aw
in
p
u
ts
an
d
m
ay
in
tr
o
d
u
c
e
r
ed
u
n
d
an
cy
,
a
n
d
d
ec
is
io
n
-
le
v
el
f
u
s
io
n
,
wh
ich
c
o
m
b
in
es
in
d
ep
en
d
en
t
class
if
ier
o
u
tp
u
ts
an
d
m
ay
lo
s
e
in
ter
-
f
ea
tu
r
e
c
o
r
r
elatio
n
s
,
m
i
d
-
lev
el
f
u
s
io
n
p
r
eser
v
es
d
is
c
r
im
in
ativ
e
s
em
an
tic
r
elatio
n
s
h
ip
s
wh
ile
en
ab
lin
g
ef
f
ec
tiv
e
d
im
en
s
io
n
ality
r
ed
u
c
tio
n
th
r
o
u
g
h
o
p
tim
izatio
n
.
3
.
4
.
Dee
p liqu
id s
t
a
t
e
m
a
chi
ne
cla
s
s
if
ier
At
th
e
f
in
al
s
tag
e
o
f
th
e
m
eth
o
d
o
lo
g
y
,
th
e
DL
SM
[
2
0
]
,
[
2
1
]
is
u
s
ed
as
th
e
clas
s
if
ier
th
at
m
ap
s
th
e
ex
tr
ac
ted
f
ea
t
u
r
es
in
to
th
e
c
o
r
r
esp
o
n
d
in
g
s
ce
n
e
ca
teg
o
r
ie
s
.
Un
lik
e
co
n
v
en
tio
n
al
class
if
ier
s
(
e.
g
.
,
s
o
f
tm
ax
lay
er
s
o
r
s
u
p
p
o
r
t
v
ec
t
o
r
m
ac
h
in
es
(
SVMs))
,
th
e
DL
SM
lev
er
ag
es
th
e
d
y
n
a
m
ic
tem
p
o
r
al
en
co
d
in
g
p
r
o
p
er
ties
o
f
liq
u
id
s
tates
co
m
b
in
ed
with
th
e
h
ier
ar
ch
ical
lear
n
in
g
a
b
ilit
y
o
f
d
ee
p
n
etwo
r
k
s
,
m
ak
in
g
it
p
ar
ticu
lar
ly
ef
f
ec
tiv
e
f
o
r
co
m
p
lex
s
ce
n
e
d
ata
[
2
1
]
.
T
h
e
p
r
o
p
o
s
ed
DL
SM
ar
ch
itectu
r
e
is
s
h
o
wn
in
Fig
u
r
e
3
.
L
SM
c
o
n
s
is
ts
o
f
th
r
ee
s
p
ec
ial
lay
er
s
:
an
in
p
u
t
lay
er
,
a
r
eser
v
o
ir
o
r
liq
u
i
d
lay
er
,
an
d
a
m
em
o
r
y
-
less
r
ea
d
o
u
t
cir
cu
it.
T
h
e
r
eser
v
o
ir
is
co
n
s
id
er
ed
th
e
g
en
er
ic
p
r
ep
r
o
ce
s
s
o
r
th
at
co
n
s
is
ts
o
f
n
u
m
er
o
u
s
leak
y
in
teg
r
ate
-
an
d
-
f
ir
e
(
L
I
F)
n
eu
r
o
n
s
.
R
ea
d
o
u
t n
e
u
r
o
n
s
ar
e
also
ca
lled
p
r
o
ce
s
s
o
r
s
,
an
d
th
e
y
p
r
o
d
u
ce
th
e
f
in
al
o
u
tp
u
t f
r
o
m
th
e
L
SM.
T
h
e
h
y
p
er
p
a
r
am
eter
s
u
s
ed
f
o
r
th
e
DL
SM
class
if
ier
ar
e
as
s
h
o
wn
b
el
o
w.
T
h
e
s
p
ec
tr
al
r
ad
iu
s
o
f
th
e
r
eser
v
o
ir
was
ch
o
s
en
t
o
b
e
l
ess
th
an
o
n
e
in
o
r
d
er
to
m
ai
n
tain
s
tab
le
r
eser
v
o
ir
d
y
n
a
m
ics
an
d
s
atis
f
y
th
e
ec
h
o
-
s
tate
p
r
o
p
e
r
ty
.
T
h
e
leak
r
ate
was
s
e
lecte
d
to
co
n
tr
o
l
th
e
r
ate
at
wh
ich
p
ast
in
f
o
r
m
a
tio
n
in
f
lu
en
ce
s
th
e
cu
r
r
en
t
n
eu
r
o
n
s
tate,
th
er
e
b
y
allo
win
g
th
e
m
o
d
el
to
r
etain
u
s
ef
u
l
tem
p
o
r
al
in
f
o
r
m
atio
n
with
o
u
t
ca
u
s
in
g
in
s
tab
ilit
y
.
T
o
r
e
d
u
ce
co
m
p
u
t
atio
n
al
co
m
p
lex
ity
,
s
p
ar
s
e
r
eser
v
o
ir
c
o
n
n
ec
ti
v
ity
was
ad
o
p
t
ed
,
wh
ich
lim
its
th
e
n
u
m
b
er
o
f
ac
tiv
e
c
o
n
n
ec
tio
n
s
wh
ile
p
r
eser
v
in
g
s
u
f
f
icien
t
r
e
p
r
esen
tatio
n
al
ca
p
a
b
ilit
y
.
I
n
a
d
d
itio
n
,
t
h
e
SW
T
A
m
ec
h
an
is
m
was
em
p
lo
y
ed
to
s
u
p
p
r
ess
wea
k
n
eu
r
o
n
ac
tiv
ati
o
n
s
an
d
em
p
h
asize
d
o
m
in
a
n
t
r
esp
o
n
s
es,
lead
in
g
to
im
p
r
o
v
ed
class
s
ep
ar
ab
ilit
y
d
u
r
in
g
th
e
r
ea
d
o
u
t stag
e
as sh
o
wn
in
T
ab
le
2
.
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
C
o
n
fid
en
ce
-
d
r
iven
fea
tu
r
e
s
elec
tio
n
a
n
d
en
s
emb
le
cla
s
s
ifica
tio
n
u
s
in
g
…
(
S
u
ma
th
i Kr
is
h
n
a
mu
r
th
y
)
4529
Fig
u
r
e
3
.
Pro
p
o
s
ed
DL
SM
ar
c
h
itectu
r
e
with
co
n
f
id
en
ce
-
g
u
id
ed
atten
tio
n
an
d
SW
T
A
lay
er
s
f
o
r
s
ce
n
e
r
ec
o
g
n
itio
n
T
ab
le
2
.
Hy
p
er
p
ar
a
m
eter
s
ettin
g
s
o
f
th
e
SW
T
A
r
eser
v
o
ir
an
d
r
ea
d
o
u
t la
y
er
s
P
a
r
a
me
t
e
r
D
e
scri
p
t
i
o
n
V
a
l
u
e
R
e
ser
v
o
i
r
s
i
z
e
N
u
mb
e
r
o
f
sp
i
k
i
n
g
n
e
u
r
o
n
s
5
0
0
I
n
p
u
t
sca
l
i
n
g
S
c
a
l
i
n
g
o
f
i
n
p
u
t
f
e
a
t
u
r
e
w
e
i
g
h
t
s
0
.
5
S
p
e
c
t
r
a
l
r
a
d
i
u
s
C
o
n
t
r
o
l
s
r
e
ser
v
o
i
r
s
t
a
b
i
l
i
t
y
0
.
9
Le
a
k
r
a
t
e
N
e
u
r
o
n
s
t
a
t
e
u
p
d
a
t
e
r
a
t
e
0
.
3
S
p
a
r
si
t
y
P
e
r
c
e
n
t
a
g
e
o
f
n
o
n
-
z
e
r
o
r
e
s
e
r
v
o
i
r
c
o
n
n
e
c
t
i
o
n
s
1
0
%
A
c
t
i
v
a
t
i
o
n
f
u
n
c
t
i
o
n
R
e
ser
v
o
i
r
n
e
u
r
o
n
a
c
t
i
v
a
t
i
o
n
t
a
n
h
R
e
a
d
o
u
t
l
a
y
e
r
Li
n
e
a
r
r
e
g
r
e
s
si
o
n
/
s
o
f
t
ma
x
Li
n
e
a
r
Le
a
r
n
i
n
g
r
a
t
e
R
e
a
d
o
u
t
l
e
a
r
n
i
n
g
r
a
t
e
0
.
0
0
1
Tr
a
i
n
i
n
g
e
p
o
c
h
s
N
u
mb
e
r
o
f
t
r
a
i
n
i
n
g
i
t
e
r
a
t
i
o
n
s
1
0
0
R
e
g
u
l
a
r
i
z
a
t
i
o
n
L2
r
e
g
u
l
a
r
i
z
a
t
i
o
n
c
o
e
f
f
i
c
i
e
n
t
1e
-
4
S
W
TA
t
h
r
e
s
h
o
l
d
S
t
a
t
e
se
l
e
c
t
i
o
n
t
h
r
e
sh
o
l
d
0
.
8
DL
SMs
ar
e
ty
p
ically
u
s
ed
f
o
r
tem
p
o
r
al
d
ata,
a
n
d
s
tatic
im
ag
es
ca
n
b
e
tr
an
s
f
o
r
m
ed
in
t
o
p
s
eu
d
o
-
tem
p
o
r
al
s
p
ik
e
s
eq
u
en
ce
s
th
r
o
u
g
h
r
ate
c
o
d
in
g
.
I
n
tr
o
d
u
cin
g
a
SW
T
A
lay
er
en
s
u
r
es
th
at
o
n
ly
th
e
m
o
s
t
ac
tiv
e
n
eu
r
o
n
s
co
n
tr
i
b
u
te
to
t
h
e
liq
u
id
s
tate.
E
ac
h
im
ag
e
f
e
atu
r
e
(
f
r
o
m
th
e
C
NN
en
s
em
b
l
e)
is
f
ir
s
t
co
n
v
er
ted
to
s
p
ik
e
r
ates.
Giv
en
a
f
ea
tu
r
e
v
ec
to
r
=
[
1
,
2
,
…
,
]
,
as in
(
3
)
.
(
)
=
{
1
,
≤
0
,
ℎ
}
(
3
)
W
h
er
e
(
)
is
s
p
ik
e
tr
ain
f
o
r
n
e
u
r
o
n
;
=
−
−
is
n
o
r
m
alize
d
f
ir
in
g
r
ate
;
an
d
is
s
im
u
latio
n
win
d
o
w
(
e.
g
.
,
1
0
0
m
s
)
.
T
h
e
SW
T
A
lay
er
s
e
lects
o
n
ly
th
e
m
o
s
t
d
o
m
in
an
t
s
p
ik
es
p
er
tim
e
s
tep
.
Fo
r
N
in
p
u
t
n
eu
r
o
n
s
,
th
e
m
em
b
r
an
e
p
o
ten
t
ial
is
(
4
)
.
(
)
=
∑
(
)
−
(
4
)
At
ea
ch
tim
e
s
tep
,
o
n
ly
t
h
e
n
e
u
r
o
n
with
t
h
e
m
ax
im
u
m
p
o
ten
tial f
ir
es,
as sh
o
wn
in
(
5
)
.
(
)
=
{
1
,
=
a
r
g
(
)
0
,
ℎ
}
(
5
)
T
h
is
en
f
o
r
ce
s
co
m
p
etitio
n
a
m
o
n
g
n
eu
r
o
n
s
;
o
n
ly
th
e
wi
n
n
er
em
its
a
s
p
ik
e,
im
p
r
o
v
in
g
f
ea
tu
r
e
s
elec
tiv
ity
an
d
s
p
ar
s
ity
.
T
h
e
SW
T
A
—
f
i
lter
ed
s
p
ik
es
en
ter
th
e
liq
u
id
lay
er
o
f
=
1000
L
I
F
n
eu
r
o
n
s
,
with
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.
15
,
No
.
5
,
Octo
b
er
20
26
:
4
5
2
3
-
4
5
3
9
4530
r
an
d
o
m
s
p
ar
s
e
co
n
n
ec
tiv
ity
(
1
0
%).
As
s
h
o
wn
in
Fig
u
r
e
3
,
th
e
o
p
tim
ized
f
ea
tu
r
es
ar
e
p
r
o
ce
s
s
ed
th
r
o
u
g
h
s
p
ik
in
g
r
eser
v
o
ir
lay
e
r
s
with
SW
T
A
co
m
p
etitio
n
,
f
o
llo
wed
b
y
a
co
n
f
id
e
n
ce
-
g
u
id
ed
atten
t
io
n
m
ec
h
an
is
m
an
d
a
r
ea
d
o
u
t
lay
er
f
o
r
f
in
al
s
ce
n
e
class
if
ica
tio
n
ac
r
o
s
s
2
0
ca
t
eg
o
r
ies.
T
h
e
r
eser
v
o
ir
d
y
n
am
ics
tr
an
s
f
o
r
m
s
tatic
f
ea
tu
r
es
in
to
h
i
g
h
-
d
im
e
n
s
io
n
a
l
tem
p
o
r
al
r
e
p
r
esen
tatio
n
s
,
w
h
ich
ar
e
class
if
ied
v
ia
a
s
o
f
t
m
ax
r
ea
d
o
u
t.
T
h
is
in
teg
r
atio
n
lev
er
ag
es
b
o
t
h
th
e
d
iv
er
s
ity
o
f
C
NN
f
ea
tu
r
es
an
d
th
e
tem
p
o
r
al
r
ich
n
ess
o
f
s
p
ik
in
g
r
eser
v
o
ir
s
,
ac
h
iev
in
g
a
f
in
al
ac
cu
r
ac
y
o
f
9
6
.
5
%.
T
h
e
p
r
o
b
ab
ilit
y
o
f
s
y
n
ap
tic
co
n
n
ec
tio
n
s
b
etwe
en
n
eu
r
o
n
s
is
r
elate
d
to
th
e
E
u
clid
ea
n
d
is
tan
ce
b
etwe
en
th
e
n
eu
r
o
n
s
.
T
h
e
s
y
n
a
p
tic
c
o
n
n
ec
tio
n
p
r
o
b
ab
ilit
y
(
,
)
,
f
r
o
m
n
eu
r
o
n
to
,
d
ep
en
d
s
o
n
th
e
E
u
clid
ea
n
d
is
tan
ce
,
(
,
)
,
b
etwe
en
th
em
as sh
o
w
n
(
6
)
.
(
,
)
=
∗
(
−
(
,
)
/
)
2
(
6
)
3
.
5
.
Da
t
a
s
et
T
h
is
p
r
o
p
o
s
ed
w
o
r
k
is
ca
r
r
i
ed
o
u
t
u
s
in
g
th
e
NYU
d
ep
t
h
d
ataset
V2
.
T
h
e
d
ataset
c
o
n
s
is
ts
o
f
2
0
d
if
f
e
r
en
t
s
ce
n
e
ca
teg
o
r
ies
an
d
o
v
er
1
,
0
0
0
o
b
ject
class
es.
Fo
r
th
is
wo
r
k
h
av
e
w
o
r
k
ed
o
n
a
s
u
b
s
et
o
f
1
,
9
2
0
in
d
o
o
r
s
ce
n
e
im
ag
es
w
ith
a
to
tal
s
ize
o
f
7
9
.
5
MB.
T
h
ese
im
ag
es
ex
h
ib
it
v
ar
iatio
n
s
in
s
p
atial
lay
o
u
t,
o
r
ien
tatio
n
,
an
d
s
ce
n
e
co
m
p
o
s
itio
n
an
d
wer
e
s
y
s
tem
atica
ll
y
s
o
r
ted
an
d
r
e
o
r
g
a
n
ized
in
t
o
2
0
d
is
tin
ct
s
ce
n
e
class
es.
T
h
e
d
ataset
co
n
s
i
s
ts
o
f
in
d
o
o
r
s
ce
n
es
s
u
ch
as
b
asem
en
t,
b
ath
r
o
o
m
,
b
ed
r
o
o
m
,
b
o
o
k
s
h
o
p
,
ca
f
e,
class
r
o
o
m
,
co
n
f
e
r
en
ce
r
o
o
m
,
d
in
ette,
d
in
in
g
r
o
o
m
,
ex
e
r
cise
r
o
o
m
,
f
u
r
n
itu
r
e
s
h
o
p
,
h
o
m
e
o
f
f
ice,
h
o
m
e
s
to
r
a
g
e
,
k
itch
en
,
c
o
m
p
u
ter
lab
,
liv
i
n
g
ar
ea
,
p
la
y
r
o
o
m
,
r
ec
ep
tio
n
,
s
tu
d
en
t
lo
u
n
g
e,
an
d
s
tu
d
y
r
o
o
m
,
as
s
h
o
w
n
in
Fig
u
r
e
4
.
T
h
is
p
r
o
p
o
s
ed
wo
r
k
h
as
co
llected
an
d
s
o
r
ted
th
e
d
if
f
er
en
t
im
ag
es
in
ea
ch
class
b
ased
o
n
th
e
d
iv
er
s
e
n
atu
r
e
th
ey
ex
h
ib
it,
a
n
d
th
e
co
u
n
t o
f
im
a
g
es in
ea
ch
class
v
ar
ies ac
co
r
d
in
g
ly
.
Fig
u
r
e
4
.
Sam
p
le
s
ce
n
e
im
ag
e
s
o
f
v
ar
io
u
s
class
es su
ch
as b
asem
en
t,
b
ath
r
o
o
m
,
p
lay
r
o
o
m
,
b
ed
r
o
o
m
,
r
ec
ep
tio
n
,
a
n
d
d
i
n
in
g
r
o
o
m
in
th
e
NYU
d
ataset
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
NS
T
h
is
r
esear
ch
wo
r
k
h
as
ev
al
u
ated
th
e
p
er
f
o
r
m
an
c
e
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
o
n
th
e
NYU
d
ataset
,
wh
ich
in
clu
d
es
d
if
f
er
en
t
class
es
o
f
s
ce
n
e
im
ag
es
.
T
o
en
h
an
ce
th
e
q
u
ality
o
f
th
e
in
p
u
t
s
ce
n
e
im
ag
es
,
th
is
wo
r
k
ha
s
ap
p
lied
th
e
C
L
AHE
an
d
R
etin
ex
illu
m
in
atio
n
co
r
r
ec
tio
n
s
as
a
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
e
to
en
h
an
ce
t
h
e
q
u
ality
o
f
th
e
i
m
ag
es.
I
t
r
e
d
u
ce
s
th
e
b
ac
k
g
r
o
u
n
d
n
o
is
e
,
as
s
h
o
wn
i
n
Fig
u
r
e
5
,
wh
er
e
Fig
u
r
e
5
(
a)
s
h
o
ws
th
e
s
am
p
le
o
r
ig
in
al
in
p
u
t
im
ag
e
,
Fig
u
r
e
5
(
b
)
s
h
o
ws
th
e
C
L
AHE
ap
p
lied
,
an
d
Fig
u
r
e
5
(
c)
s
h
o
ws
th
e
R
etin
ex
ap
p
lied
.
T
h
e
d
ataset
co
n
s
is
ts
o
f
d
if
f
er
en
t
s
ce
n
e
class
es
s
u
ch
as
b
ed
r
o
o
m
,
d
in
in
g
r
o
o
m
,
b
ath
r
o
o
m
,
an
d
s
tu
d
y
r
o
o
m
.
C
L
AHE
aim
s
to
im
p
r
o
v
e
lo
ca
l
co
n
tr
ast,
esp
ec
ially
in
ar
ea
s
th
at
ar
e
to
o
d
ar
k
o
r
to
o
b
r
ig
h
t
,
wh
er
ea
s
f
o
cu
s
o
n
im
p
r
o
v
in
g
t
h
e
d
y
n
am
ic
r
an
g
e
an
d
co
lo
r
c
o
n
s
tan
cy
o
f
an
im
a
g
e,
o
f
ten
m
ak
in
g
d
etails
m
o
r
e
v
is
ib
le
in
b
o
th
b
r
ig
h
t a
n
d
d
ar
k
r
eg
io
n
s
.
T
h
e
in
teg
r
atio
n
o
f
C
L
AHE
an
d
R
etin
ex
p
r
e
p
r
o
ce
s
s
in
g
s
ig
n
i
f
ican
tly
en
h
a
n
ce
s
th
e
s
ce
n
e
r
e
co
g
n
itio
n
p
er
f
o
r
m
an
ce
a
cr
o
s
s
all
b
ac
k
b
o
n
e
n
etwo
r
k
s
.
T
h
e
p
r
o
p
o
s
ed
DL
SM
with
SW
T
A
class
if
ier
ac
h
iev
es
9
6
.
5
%
ac
cu
r
ac
y
,
r
ef
lectin
g
a
2
.
7
%
im
p
r
o
v
em
en
t
as
s
h
o
w
n
in
T
ab
le
3
.
T
h
is
d
em
o
n
s
tr
ates
th
at
illu
m
in
atio
n
n
o
r
m
aliza
tio
n
en
h
an
ce
s
f
ea
tu
r
e
co
n
s
is
ten
cy
,
b
en
e
f
itin
g
b
o
th
C
NN
f
ea
tu
r
e
ex
tr
ac
tio
n
a
n
d
s
p
ik
in
g
n
eu
r
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
C
o
n
fid
en
ce
-
d
r
iven
fea
tu
r
e
s
elec
tio
n
a
n
d
en
s
emb
le
cla
s
s
ifica
tio
n
u
s
in
g
…
(
S
u
ma
th
i Kr
is
h
n
a
mu
r
th
y
)
4531
s
tab
ilit
y
d
u
r
in
g
class
if
icatio
n
.
T
ab
le
3
s
h
o
ws
a
co
m
p
ar
ativ
e
ev
alu
atio
n
o
f
class
if
icatio
n
ac
cu
r
ac
y
o
b
tain
e
d
b
y
d
if
f
er
en
t
d
ee
p
lear
n
in
g
m
o
d
el
s
with
an
d
with
o
u
t
th
e
p
r
o
p
o
s
ed
p
r
e
-
p
r
o
ce
s
s
in
g
s
tr
ateg
y
o
n
th
e
NYU
in
d
o
o
r
s
ce
n
e
d
ataset.
T
ab
le
3
clea
r
ly
d
em
o
n
s
tr
ates
th
e
im
p
ac
t
o
f
illu
m
in
atio
n
-
awa
r
e
p
r
e
-
p
r
o
ce
s
s
in
g
o
n
b
o
th
in
d
iv
id
u
al
C
NN
m
o
d
els an
d
e
n
s
em
b
le
-
b
ased
ap
p
r
o
ac
h
es.
(
a)
(
b
)
(
c)
Fig
u
r
e
5
.
T
h
e
s
am
p
le
in
p
u
t im
ag
e
an
d
t
h
e
im
ag
es a
f
ter
a
p
p
li
ca
tio
n
o
f
(
a)
o
r
ig
in
al
in
p
u
t,
(
b
)
C
L
AHE
,
an
d
(
c)
R
etin
ex
T
ab
le
3
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
b
ef
o
r
e
an
d
af
ter
p
r
ep
r
o
ce
s
s
in
g
f
o
r
th
e
NYU
an
d
MI
T
I
n
d
o
o
r
s
ce
n
e
-
15
d
ataset
M
o
d
e
l
W
i
t
h
o
u
t
p
r
e
p
r
o
c
e
ss
i
n
g
(
%)
[NYU
d
a
t
a
se
t]
W
i
t
h
C
LA
H
E
+
R
e
t
i
n
e
x
(
%)
[NYU
d
a
t
a
se
t
]
W
i
t
h
o
u
t
p
r
e
p
r
o
c
e
ss
i
n
g
(
%)
[
M
I
T
i
n
d
o
o
r
1
5
d
a
t
a
se
t
]
W
i
t
h
C
LA
H
E
+
R
e
t
i
n
e
x
(
%)
[
M
I
T
i
n
d
o
o
r
1
5
d
a
t
a
se
t
]
A
l
e
x
N
e
t
8
9
.
8
0
9
2
.
7
0
8
6
.
4
1
8
9
.
6
1
R
e
sN
e
t
9
2
.
5
2
9
5
.
3
1
8
9
.
2
2
9
2
.
1
2
D
e
n
seN
e
t
9
0
.
6
0
9
3
.
1
1
8
8
.
0
0
9
1
.
0
4
M
u
l
t
i
-
mo
d
e
l
e
n
s
e
m
b
l
e
(
c
o
n
c
a
t
e
n
a
t
i
o
n
)
9
4
.
12
9
6
.
2
2
9
1
.
8
1
9
4
.
6
0
P
r
o
p
o
se
d
D
LS
M
w
i
t
h
S
W
TA
9
3
.
8
0
9
6
.
5
0
9
2
.
4
0
9
5
.
2
0
W
ith
o
u
t
p
r
ep
r
o
ce
s
s
in
g
,
Alex
Net
ac
h
iev
es
an
ac
cu
r
ac
y
o
f
8
9
.
8
%,
in
d
icatin
g
its
s
en
s
itiv
ity
to
illu
m
in
atio
n
v
a
r
iatio
n
s
an
d
b
ac
k
g
r
o
u
n
d
clu
tter
d
u
e
to
i
ts
r
elativ
ely
s
h
allo
w
ar
c
h
itectu
r
e.
R
esNet
an
d
Den
s
eNe
t
s
h
o
w
im
p
r
o
v
ed
b
aselin
e
p
er
f
o
r
m
a
n
ce
(
9
2
.
5
%
an
d
9
0
.
6
%,
r
esp
ec
tiv
ely
)
o
win
g
to
d
ee
p
er
ar
ch
itectu
r
es
an
d
im
p
r
o
v
ed
f
e
atu
r
e
r
eu
s
e;
h
o
wev
er
,
th
e
y
s
till
ex
h
ib
it
p
er
f
o
r
m
an
ce
d
e
g
r
ad
a
tio
n
u
n
d
er
u
n
ev
e
n
lig
h
tin
g
co
n
d
itio
n
s
.
Af
ter
ap
p
l
y
in
g
th
e
h
y
b
r
id
C
L
AHE
+
R
et
in
ex
p
r
ep
r
o
ce
s
s
in
g
,
all
m
o
d
els
s
h
o
w
a
co
n
s
is
ten
t
an
d
n
o
tab
le
im
p
r
o
v
e
m
en
t
in
ac
cu
r
ac
y
.
Alex
Net
im
p
r
o
v
es
to
9
2
.
7
%,
wh
ile
R
esNet
an
d
Den
s
eNe
t
ac
h
iev
e
9
5
.
3
%
an
d
9
3
.
1
%,
r
esp
ec
tiv
ely
,
co
n
f
i
r
m
in
g
t
h
at
p
r
e
-
p
r
o
ce
s
s
in
g
en
h
an
ce
s
f
ea
tu
r
e
q
u
ality
b
y
im
p
r
o
v
in
g
co
n
tr
ast
an
d
illu
m
in
atio
n
co
n
s
is
ten
cy
.
T
h
e
m
u
lti
-
m
o
d
el
en
s
em
b
le
u
s
in
g
f
ea
tu
r
e
co
n
ca
ten
atio
n
f
u
r
th
e
r
im
p
r
o
v
es
p
er
f
o
r
m
a
n
ce
to
9
6
.
2
%,
d
em
o
n
s
tr
atin
g
th
e
b
e
n
ef
it
o
f
co
m
b
in
in
g
co
m
p
lem
e
n
t
ar
y
r
ep
r
esen
tatio
n
s
f
r
o
m
m
u
ltip
le
ar
ch
itectu
r
es.
Ho
wev
er
,
t
h
e
p
r
o
p
o
s
ed
DL
SM
with
ac
h
iev
es
th
e
h
ig
h
est
ac
cu
r
ac
y
o
f
9
6
.
5
%,
o
u
tp
e
r
f
o
r
m
in
g
b
o
th
in
d
iv
id
u
al
m
o
d
els
a
n
d
th
e
c
o
n
v
en
tio
n
al
e
n
s
em
b
le
a
p
p
r
o
ac
h
.
T
h
is
im
p
r
o
v
em
en
t
h
ig
h
lig
h
ts
t
h
e
ef
f
ec
tiv
e
n
ess
o
f
th
e
DL
SM’
s
m
em
o
r
y
-
b
ased
lear
n
in
g
an
d
th
e
SW
T
A
m
ec
h
an
is
m
in
lev
er
a
g
in
g
en
h
an
ce
d
f
ea
t
u
r
e
r
ep
r
esen
tatio
n
s
m
o
r
e
e
f
f
icien
t
ly
th
an
s
tatic
f
u
s
io
n
s
tr
ateg
ies.
Ov
er
all,
th
e
r
esu
lts
in
T
ab
le
4
v
alid
ate
th
at
th
e
p
r
o
p
o
s
ed
p
r
e
-
p
r
o
ce
s
s
in
g
n
o
t
o
n
ly
b
o
o
s
ts
in
d
i
v
id
u
al
C
NN
p
er
f
o
r
m
an
ce
b
u
t
also
s
y
n
er
g
is
tically
en
h
an
ce
s
t
h
e
p
r
o
p
o
s
ed
DL
SM
–
SW
T
A
f
r
am
ewo
r
k
,
lead
i
n
g
to
s
u
p
e
r
io
r
an
d
m
o
r
e
s
tab
le
in
d
o
o
r
s
ce
n
e
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
Fro
m
th
e
r
esu
lt
s
in
T
ab
le
4
,
it
is
ev
id
en
t
th
at
C
D
-
W
C
O
co
n
s
i
s
ten
tly
r
ed
u
ce
s
th
e
n
u
m
b
er
o
f
s
elec
ted
f
ea
tu
r
es
ac
r
o
s
s
all
d
ee
p
lear
n
in
g
m
o
d
els.
C
D
-
W
C
O
elim
in
ates
n
ea
r
ly
o
n
e
-
f
o
u
r
th
o
f
th
e
r
ed
u
n
d
an
t
f
ea
tu
r
es.
T
h
is
r
ed
u
ctio
n
is
attr
ib
u
ted
to
th
e
in
c
o
r
p
o
r
atio
n
o
f
co
n
f
id
e
n
ce
s
co
r
es
b
ef
o
r
e
th
e
o
p
tim
izatio
n
p
r
o
ce
s
s
,
wh
ich
en
a
b
les
th
e
alg
o
r
ith
m
to
d
is
ca
r
d
f
ea
tu
r
es
with
lo
w
r
eliab
ilit
y
.
CD
-
W
C
O
im
p
r
o
v
es
co
m
p
u
tatio
n
al
ef
f
icie
n
cy
b
y
lo
wer
in
g
s
to
r
ag
e
an
d
tr
ain
in
g
co
s
ts
,
wh
ile
s
im
u
ltan
eo
u
s
ly
en
h
an
cin
g
th
e
r
o
b
u
s
tn
ess
o
f
th
e
f
in
al
class
if
icatio
n
s
tag
e.
T
h
e
r
etain
ed
f
e
atu
r
es
ar
e
m
o
r
e
d
is
cr
im
in
ativ
e,
wh
ich
im
p
r
o
v
es
class
if
ier
tr
ain
in
g
s
p
ee
d
,
r
ed
u
c
es
m
em
o
r
y
u
s
ag
e
,
a
n
d
en
h
a
n
c
es
g
en
er
aliza
tio
n
.
T
h
e
b
a
r
g
r
a
p
h
r
ep
r
esen
tin
g
th
e
f
ea
tu
r
e
r
ed
u
ctio
n
is
s
h
o
wn
i
n
Fig
u
r
e
6
.
As
d
ep
icted
i
n
Fig
u
r
e
6
,
ac
r
o
s
s
all
th
r
ee
m
o
d
els
,
th
e
W
C
O
a
lg
o
r
ith
m
r
ed
u
c
es
th
e
n
u
m
b
er
o
f
f
ea
t
u
r
es
ex
tr
ac
ted
f
r
o
m
th
e
m
o
d
els.
Featu
r
e
d
im
en
s
io
n
ality
g
ets
r
ed
u
ce
d
.
T
h
e
C
D
-
W
C
O
f
u
r
th
er
im
p
r
o
v
es
th
e
p
er
f
o
r
m
an
ce
b
y
c
o
n
s
is
ten
tly
r
ed
u
cin
g
th
e
n
u
m
b
e
r
o
f
f
ea
t
u
r
e
s
an
d
s
elec
tin
g
th
e
d
iv
e
r
s
if
ied
an
d
d
is
cr
im
in
ativ
e
f
ea
tu
r
es.
T
h
is
r
e
p
r
esen
ts
th
at
th
e
co
n
f
id
en
ce
-
d
r
iv
en
s
elec
tio
n
en
h
an
ce
s
th
e
ef
f
icien
c
y
o
f
f
ea
tu
r
e
ex
tr
ac
tio
n
,
lead
in
g
to
m
o
r
e
co
m
p
ac
t
r
ep
r
esen
tatio
n
s
th
at
ar
e
co
m
p
u
tati
o
n
ally
less
ex
p
en
s
iv
e
an
d
p
o
t
en
tially
m
o
r
e
r
o
b
u
s
t
f
o
r
class
if
icatio
n
task
s
.
T
h
e
p
r
o
p
o
s
ed
m
et
h
o
d
is
ev
alu
ated
on
d
atasets
co
n
s
is
tin
g
o
f
1
,
9
0
0
im
ag
es
an
d
ac
r
o
s
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.
15
,
No
.
5
,
Octo
b
er
20
26
:
4
5
2
3
-
4
5
3
9
4532
2
0
d
if
f
e
r
en
t
ca
teg
o
r
ies
o
f
s
c
en
e
im
ag
es
in
th
e
d
ataset.
T
h
e
d
ataset
h
as
b
ee
n
s
p
lit
in
t
o
8
0
%
tr
ain
in
g
an
d
2
0
%
test
in
g
r
esu
lts
in
1
,
5
3
6
tr
ain
in
g
im
ag
es
an
d
3
8
4
test
i
n
g
im
ag
es.
T
h
is
wo
r
k
h
as
em
p
lo
y
ed
th
e
Ad
am
o
p
tim
izer
with
a
lear
n
in
g
r
ate
o
f
0
.
0
1
,
a
n
d
t
h
e
tr
ain
in
g
p
r
o
ce
s
s
was
f
o
r
5
0
0
iter
atio
n
s
an
d
f
o
r
a
m
in
i
b
atch
s
ize
o
f
1
2
8
.
T
h
e
p
lo
t
o
f
ac
cu
r
ac
y
an
d
lo
s
s
cu
r
v
es
is
as
s
h
o
wn
in
Fig
u
r
es
7
an
d
8
.
Fig
u
r
e
7
s
h
o
ws
th
at
th
e
m
o
d
el
is
lear
n
in
g
b
etter
to
o
b
tain
an
ac
c
u
r
ac
y
o
f
9
6
.
5
%.
I
n
th
is
s
im
u
latio
n
,
th
e
v
ali
d
a
tio
n
ac
cu
r
ac
y
also
in
cr
ea
s
es
an
d
s
tab
ilizes
at
a
s
lig
h
tly
lo
wer
p
er
ce
n
tag
e
ar
o
u
n
d
9
4
%.
Fig
u
r
e
8
s
h
o
ws
th
at
th
e
m
o
d
el’
s
lo
s
s
is
r
ed
u
ce
d
i
n
th
e
s
u
cc
ess
iv
e
ep
o
c
h
s
.
T
ab
le
4
.
R
ed
u
ctio
n
in
f
ea
t
u
r
e
d
im
en
s
io
n
ality
th
r
o
u
g
h
co
n
f
id
en
ce
-
d
r
iv
e
n
W
C
O
co
m
p
ar
ed
with
s
tan
d
ar
d
W
C
O
ac
r
o
s
s
d
ee
p
lear
n
in
g
m
o
d
els o
n
th
e
NYU
s
ce
n
e
d
ataset
M
o
d
e
l
F
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
l
a
y
e
r
F
e
a
t
u
r
e
s
e
x
t
r
a
c
t
e
d
F
e
a
t
u
r
e
s
e
x
t
r
a
c
t
e
d
u
si
n
g
W
C
O
F
e
a
t
u
r
e
s
e
x
t
r
a
c
t
e
d
u
si
n
g
C
D
-
W
C
O
R
e
d
u
c
t
i
o
n
p
e
r
c
e
n
t
a
g
e
(
%)
A
l
e
x
N
e
t
FC
8
2
3
,
5
7
2
,
4
8
0
1
1
,
6
4
8
,
1
2
0
8
,
7
3
6
,
0
9
0
2
5
.
0
0
R
e
sN
e
t
1
0
1
FC
1
0
0
0
1
1
,
7
8
6
,
2
4
0
5
,
8
8
1
,
6
1
0
4
,
4
1
1
,
2
0
7
2
5
.
0
0
D
e
n
seN
e
t
2
0
1
a
v
g
_
p
o
o
l
1
1
,
0
4
9
,
6
0
0
5
,
5
8
8
,
1
0
5
4
,
1
9
1
,
0
7
9
2
5
.
0
0
Fig
u
r
e
6
.
Featu
r
e
d
im
en
s
io
n
ali
ty
f
o
r
p
r
e
-
tr
ain
e
d
m
o
d
els co
m
p
ar
ed
with
th
o
s
e
r
ed
u
ce
d
u
s
in
g
W
C
O
an
d
CD
-
W
C
O
Fig
u
r
e
7
.
Plo
t o
f
ac
cu
r
ac
y
cu
r
v
es f
o
r
th
e
s
ce
n
e
im
ag
e
NYU
d
ataset
Evaluation Warning : The document was created with Spire.PDF for Python.