I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
15
,
No
.
2
,
A
p
r
il
20
25
,
p
p
.
2
3
0
4
~
2
3
1
3
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
15
i
2
.
pp
2
3
0
4
-
2
3
1
3
2304
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
A cons
trained
co
nv
o
lutiona
l neural network
wi
th
a
t
t
ention
mecha
nism
f
o
r i
ma
g
e ma
nipula
tion detec
tion
K
a
m
a
g
a
t
e
B
ema
n H
a
m
idja
1
,
F
a
t
o
um
a
t
a
Wo
ng
bé
Ro
s
a
lie
T
o
k
pa
2
,
Vincent
M
o
ns
a
n
2
,
So
uley
m
a
ne
O
um
t
a
na
g
a
3
1
La
b
o
r
a
t
o
i
r
e
d
e
s Sc
i
e
n
c
e
s e
t
Te
c
h
n
o
l
o
g
i
e
s
d
e
l
’
I
n
f
o
r
mat
i
o
n
e
t
d
e
l
a
C
o
mm
u
n
i
c
a
t
i
o
n
(
LA
S
TI
C
)
,
Ec
o
l
e
S
u
p
é
r
i
e
u
r
e
A
f
r
i
c
a
i
n
e
d
e
s
Te
c
h
n
o
l
o
g
i
e
s
d
e
l
’
I
n
f
o
r
ma
t
i
o
n
e
t
d
e
l
a
C
o
mm
u
n
i
c
a
t
i
o
n
(
ESA
TI
C
)
,
A
b
i
d
j
a
n
,
C
ô
t
e
d
’
I
v
o
i
r
e
2
La
b
o
r
a
t
o
i
r
e
d
e
M
e
c
a
n
i
q
u
e
e
t
d
’
I
n
f
o
r
mat
i
q
u
e
(
LA
M
I
)
,
U
n
i
v
e
r
si
t
é
F
é
l
i
x
H
o
u
p
h
o
u
ë
t
-
B
o
i
g
n
y
(
U
F
H
B
)
,
A
b
i
d
j
a
n
,
C
ô
t
e
d
’
I
v
o
i
r
e
3
La
b
o
r
a
t
o
i
r
e
d
e
R
e
c
h
e
r
c
h
e
e
n
I
n
f
o
r
m
a
t
i
q
u
e
e
t
Té
l
é
c
o
mm
u
i
c
a
t
i
o
n
(
LA
R
I
T)
,
I
n
st
i
t
u
t
N
a
t
i
o
n
a
l
P
o
l
y
t
e
c
h
n
i
q
u
e
F
é
l
i
x
H
o
u
p
h
o
u
ë
t
-
B
o
i
g
n
y
(
I
N
P
H
B
)
,
Y
a
mo
u
ss
o
u
k
r
o
,
C
ô
t
e
d
’
I
v
o
i
r
e
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
y
1
5
,
2
0
2
4
R
ev
is
ed
Oct
5
,
2
0
2
4
Acc
ep
ted
Oct
2
3
,
2
0
2
4
Th
e
in
f
o
rm
a
ti
o
n
d
isse
m
in
a
ted
b
y
o
n
li
n
e
m
e
d
ia
is
o
ften
p
re
se
n
te
d
in
th
e
fo
rm
o
f
ima
g
e
s,
in
o
r
d
e
r
to
q
u
ick
ly
c
a
p
ti
v
a
te
re
a
d
e
rs
a
n
d
in
c
re
a
se
a
u
d
ie
n
c
e
ra
ti
n
g
s.
H
o
we
v
e
r,
t
h
e
se
ima
g
e
s
c
a
n
b
e
m
a
n
ip
u
late
d
f
o
r
m
a
li
c
io
u
s
p
u
r
p
o
se
s,
su
c
h
a
s
in
f
lu
e
n
c
i
n
g
p
u
b
li
c
o
p
in
i
o
n
,
u
n
d
e
rm
in
in
g
m
e
d
ia
c
re
d
ib
il
i
ty
,
d
isru
p
ti
n
g
d
e
m
o
c
ra
ti
c
p
ro
c
e
ss
e
s
o
r
c
re
a
ti
n
g
c
o
n
fl
ict
with
in
s
o
c
iety
.
Va
rio
u
s
a
p
p
ro
a
c
h
e
s,
wh
e
th
e
r
re
ly
i
n
g
o
n
m
a
n
u
a
ll
y
d
e
v
e
lo
p
e
d
fe
a
tu
re
s
o
r
d
e
e
p
lea
rn
in
g
,
h
a
v
e
b
e
e
n
d
e
v
ise
d
to
d
e
tec
t
fa
lsifi
e
d
ima
g
e
s.
H
o
we
v
e
r,
th
e
y
fre
q
u
e
n
tl
y
p
r
o
v
e
les
s
e
ffe
c
ti
v
e
wh
e
n
c
o
n
fr
o
n
ted
with
wi
d
e
sp
re
a
d
a
n
d
m
u
lt
ip
le
m
a
n
ip
u
lati
o
n
s.
T
o
a
d
d
r
e
ss
th
is
c
h
a
ll
e
n
g
e
,
in
o
u
r
st
u
d
y
,
we
h
a
v
e
d
e
sig
n
e
d
a
m
o
d
e
l
c
o
m
p
risin
g
a
c
o
n
stra
in
e
d
c
o
n
v
o
lu
ti
o
n
lay
e
r
c
o
m
b
in
e
d
wit
h
a
n
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
a
n
d
a
tr
a
n
sfe
r
lea
rn
in
g
Re
sN
e
t5
0
n
e
two
rk
.
T
h
e
se
c
o
m
p
o
n
e
n
ts
a
re
i
n
ten
d
e
d
to
a
u
t
o
m
a
ti
c
a
ll
y
lea
rn
ima
g
e
m
a
n
ip
u
lati
o
n
fe
a
tu
re
s
in
t
h
e
in
it
ial
lay
e
r
a
n
d
e
x
trac
t
sp
a
ti
a
l
fe
a
tu
re
s,
re
sp
e
c
ti
v
e
ly
.
It
m
a
k
e
s
p
o
ss
i
b
le
to
d
e
tec
t
v
a
rio
u
s
fa
lsifi
c
a
ti
o
n
s
wi
th
m
u
c
h
m
o
re
a
c
c
u
ra
c
y
a
n
d
p
re
c
i
sio
n
.
T
h
e
pr
o
p
o
se
d
m
o
d
e
l
h
a
s
b
e
e
n
train
e
d
a
n
d
tes
ted
o
n
re
a
l
d
a
tas
e
ts
so
u
rc
e
d
fro
m
th
e
li
tera
tu
re
,
wh
ic
h
i
n
c
lu
d
e
M
e
d
iaEv
a
l
a
n
d
Ca
sia
.
Th
e
o
b
tain
e
d
re
su
lt
s
in
d
ica
te
t
h
a
t
o
u
r
p
ro
p
o
sa
l
su
r
p
a
ss
e
s
o
th
e
r
m
o
d
e
ls
d
o
c
u
m
e
n
t
e
d
in
th
e
li
tera
tu
re
.
S
p
e
c
ifi
c
a
ll
y
,
we
a
c
h
iev
e
a
n
a
c
c
u
ra
c
y
o
f
8
7
%
a
n
d
a
p
r
e
c
isio
n
o
f
9
3
%
o
n
t
h
e
M
e
d
iaE
v
a
l
d
a
tas
e
t.
I
n
c
o
m
p
a
riso
n
,
t
h
e
p
e
rf
o
rm
a
n
c
e
o
f
m
e
th
o
d
s
fro
m
th
e
li
tera
tu
re
o
n
th
e
sa
m
e
d
a
tas
e
t
d
o
e
s
n
o
t
e
x
c
e
e
d
8
4
%
fo
r
a
c
c
u
ra
c
y
a
n
d
9
0
%
fo
r
p
re
c
isio
n
.
K
ey
w
o
r
d
s
:
Atten
tio
n
m
ec
h
an
is
m
C
o
n
s
tr
ain
ed
co
n
v
o
lu
tio
n
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
I
m
ag
e
m
an
i
p
u
latio
n
T
r
an
s
f
er
lear
n
i
n
g
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
:
Kam
ag
ate
B
em
an
Ham
id
ja
L
ab
o
r
ato
ir
e
d
es Scien
ce
s
et
T
ec
h
n
o
lo
g
ies d
e
l’
I
n
f
o
r
m
atio
n
et
d
e
la
C
o
m
m
u
n
icatio
n
(
L
ASTI
C
)
d
e
l’
E
co
le
Su
p
ér
ieu
r
e
Af
r
icai
n
e
d
es
T
ec
h
n
o
lo
g
ies d
e
l’
I
n
f
o
r
m
atio
n
et
d
e
la
C
o
m
m
u
n
icatio
n
(
E
SATI
C
)
B
d
.
d
e
Ma
r
s
eille
f
ac
e
à
B
er
n
a
b
e
-
Km
4
.
T
r
eich
v
ille
-
1
8
B
P 1
5
0
1
Ab
id
jan
1
8
.
Ab
i
d
jan
-
C
ô
te
d
'
I
v
o
ir
e
E
-
m
ail: b
em
an
.
k
am
ag
ate@es
atic.
ed
u
.
ci
1.
I
NT
RO
D
UCT
I
O
N
I
n
an
ag
e
o
f
in
f
o
r
m
atio
n
tech
n
o
lo
g
y
ad
v
an
ce
s
an
d
r
a
p
id
d
ev
elo
p
m
en
t
o
f
s
o
cial
n
etwo
r
k
s
,
in
f
o
r
m
atio
n
ca
n
b
e
p
r
o
v
id
e
d
b
y
th
e
m
ed
ia
in
th
e
f
o
r
m
o
f
i
m
ag
es
o
r
v
id
eo
to
q
u
ick
ly
ca
p
tiv
ate
r
ea
d
er
s
an
d
in
cr
ea
s
e
au
d
ien
ce
n
u
m
b
e
r
s
.
Vis
u
al
co
n
ten
t,
wh
ich
in
clu
d
es
im
ag
es
o
r
v
id
e
o
s
,
is
a
d
y
n
am
ic
m
ea
n
s
o
f
ex
p
r
ess
io
n
th
at
is
m
o
r
e
s
tr
ik
in
g
th
an
s
im
p
le
tex
t,
s
tim
u
latin
g
th
e
d
is
s
em
in
atio
n
o
f
in
f
o
r
m
at
io
n
.
Vis
u
al
co
n
ten
t
is
also
f
r
eq
u
e
n
tly
u
s
ed
as
e
v
id
en
ce
,
r
ein
f
o
r
cin
g
th
e
cr
ed
ib
ilit
y
o
f
in
f
o
r
m
atio
n
.
Ho
we
v
er
,
in
d
iv
id
u
als
o
r
in
ter
est
g
r
o
u
p
s
d
is
s
em
in
ate
f
alse
v
is
u
al
co
n
ten
t
in
o
r
d
er
to
in
cr
ea
s
e
th
e
v
is
ib
ilit
y
an
d
ac
h
iev
e
th
eir
o
b
jectiv
e,
wh
ich
is
to
m
an
ip
u
late
th
e
o
p
in
io
n
o
f
a
m
ass
o
f
p
eo
p
le,
d
estab
ilize
a
g
o
v
er
n
m
e
n
t,
in
c
r
ea
s
e
th
eir
f
in
an
cial
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
co
n
s
tr
a
in
ed
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
k
w
ith
a
tten
tio
n
mec
h
a
n
is
m
fo
r
(
K
a
ma
g
a
te
B
e
ma
n
Ha
mid
ja
)
2305
in
co
m
e
o
r
in
f
lu
e
n
ce
th
e
r
esu
lts
d
em
o
cr
atic
p
r
o
ce
s
s
[
1
]
T
h
ese
o
b
jectiv
es
h
av
e
n
eg
ativ
e
im
p
ac
ts
,
s
u
ch
as
u
n
d
er
m
i
n
in
g
th
e
cr
ed
ib
ilit
y
o
f
s
o
cial
m
ed
ia,
wea
k
e
n
in
g
d
em
o
cr
atic
p
r
o
ce
s
s
es
an
d
g
en
er
atin
g
c
o
n
f
lict
b
etwe
en
m
em
b
er
s
o
f
a
s
o
ciety
.
I
m
ag
e
f
alsi
f
icatio
n
is
a
s
et
o
f
d
ig
ital
tech
n
iq
u
es
d
esig
n
ed
to
co
n
v
ey
m
is
lead
in
g
in
f
o
r
m
ati
o
n
th
r
o
u
g
h
m
o
d
if
ied
im
ag
es
[
2
]
.
C
o
p
y
a
n
d
p
aste
co
n
s
is
ts
in
r
ep
lacin
g
th
e
co
n
ten
t
with
in
a
r
eg
i
o
n
o
f
an
im
ag
e
with
co
n
ten
t
f
r
o
m
an
o
th
er
im
ag
e
[
3
]
.
E
r
ase
-
f
ill
o
r
i
n
p
ain
tin
g
co
n
s
is
ts
in
f
illi
n
g
a
r
eg
io
n
o
f
an
im
ag
e
with
a
co
m
p
o
s
itio
n
o
f
elem
en
tar
y
p
a
r
ts
o
f
th
e
o
r
ig
in
al
im
a
g
e
o
r
a
n
o
th
er
im
a
g
e
[
4
]
.
Als
o
,
with
t
h
e
ad
v
e
n
t
o
f
m
o
d
er
n
ar
tific
ial
in
tellig
en
ce
tech
n
o
lo
g
y
s
u
ch
as
g
en
er
ativ
e
a
d
v
er
s
a
r
ial
n
etwo
r
k
s
(
GANs)
,
im
ag
es
ca
n
b
e
g
en
er
ated
o
r
f
alsi
f
ied
to
r
esem
b
le
au
th
en
tic
im
ag
es
[
5
]
.
Mo
s
t
r
esear
ch
o
n
f
ak
e
n
ews
d
etec
tio
n
h
as
p
r
ed
o
m
in
an
tly
co
n
ce
n
tr
ated
o
n
tex
tu
al
co
n
te
n
t,
i.e
.
,
n
ews
th
at
r
elies
s
o
lely
o
n
tex
t.
I
n
th
is
co
n
tex
t,
T
o
k
p
a
et
a
l.
[
6
]
in
tr
o
d
u
ce
s
a
m
eth
o
d
t
h
at
co
m
b
in
es
two
n
e
u
r
al
n
etwo
r
k
ar
ch
itectu
r
es,
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs)
an
d
b
id
ir
ec
tio
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
e
m
o
r
y
(
B
iLST
M)
,
to
im
p
r
o
v
e
f
ak
e
n
ews
d
etec
tio
n
ac
cu
r
ac
y
ac
r
o
s
s
v
ar
io
u
s
tex
t
-
b
ased
d
atasets
.
Similar
ly
,
Ajao
et
a
l.
[
7
]
ex
p
lo
r
es
th
e
u
s
e
o
f
a
h
y
b
r
i
d
o
f
C
NNs
an
d
r
ec
u
r
r
e
n
t
n
e
u
r
al
n
etwo
r
k
s
(
R
NNs)
to
id
en
tif
y
an
d
class
if
y
f
ak
e
n
ews
m
ess
ag
es
o
n
T
witter
.
Me
an
wh
ile,
Po
p
at
et
a
l.
[
8
]
p
r
esen
t
a
n
eu
r
al
n
etwo
r
k
m
o
d
el
th
at
in
teg
r
ates
s
ig
n
als
f
r
o
m
ex
ter
n
al
ev
id
e
n
ce
ar
ticles,
co
n
s
id
er
s
th
e
lan
g
u
ag
e
u
s
ed
,
an
d
ass
ess
e
s
th
e
cr
ed
ib
ilit
y
o
f
s
o
u
r
ce
s
.
T
h
eir
m
o
d
el
also
g
en
er
ates
u
s
ef
u
l
f
ea
tu
r
es
th
at
p
r
o
v
id
e
clea
r
ex
p
lan
atio
n
s
f
o
r
u
s
er
s
,
en
s
u
r
in
g
t
r
an
s
p
ar
en
cy
i
n
th
e
p
r
ed
ictio
n
s
.
Fin
ally
,
R
an
i
et
a
l.
[
9
]
p
r
o
p
o
s
e
a
h
y
b
r
id
ap
p
r
o
ac
h
co
m
b
in
in
g
C
NN
an
d
B
iLST
M
with
g
lo
b
al
v
ec
to
r
s
f
o
r
wo
r
d
r
ep
r
esen
tatio
n
(
Glo
Ve)
em
b
e
d
d
in
g
s
t
o
cla
s
s
if
y
twee
ts
as
r
u
m
o
r
s
o
r
n
o
n
-
r
u
m
o
r
s
.
Ho
wev
er
,
th
ese
s
tu
d
ies
o
v
er
lo
o
k
th
e
is
s
u
e
o
f
f
a
k
e
n
ews
in
v
o
lv
in
g
f
alsi
f
ied
im
ag
es
[
1
0
]
.
R
esear
ch
i
n
[
1
1
]
in
d
icate
s
th
at
n
ews
p
o
s
ts
co
n
tain
in
g
im
ag
es
r
ec
eiv
e
m
o
r
e
in
ter
ac
tio
n
c
o
m
p
ar
ed
to
th
o
s
e
with
o
n
ly
tex
t.
As
a
r
esu
lt,
th
er
e
i
s
g
r
o
win
g
i
n
ter
est in
d
etec
tin
g
f
alsi
f
ied
im
ag
es o
n
s
o
cial
m
ed
i
a.
T
h
e
f
alsi
f
ied
im
ag
e
ca
n
b
e
d
etec
ted
u
s
in
g
m
an
u
ally
d
ev
elo
p
ed
f
ea
t
u
r
e
-
b
ased
ap
p
r
o
ac
h
es.
I
t
is
ass
u
m
ed
th
at
m
an
ip
u
lated
o
r
f
ak
e
im
ag
es e
x
h
ib
it
v
is
u
al
an
d
s
tatis
t
ical
d
is
tr
ib
u
tio
n
p
atter
n
s
d
is
tin
ct
f
r
o
m
th
o
s
e
o
f
g
en
u
i
n
e
im
ag
es.
J
in
et
a
l.
[
1
2
]
p
r
o
p
o
s
e
s
ev
er
al
v
is
u
al
an
d
s
tatis
tica
l
f
ea
tu
r
es
to
d
etec
t
f
ak
e
im
ag
es
wh
ich
ar
e
v
is
u
al
clar
ity
,
c
o
n
s
is
ten
cy
s
co
r
e,
v
is
u
al
s
im
ilar
ity
d
is
tr
ib
u
tio
n
h
is
to
g
r
am
,
v
is
u
al
d
iv
er
s
ity
,
clu
s
ter
in
g
s
co
r
e,
n
u
m
b
er
o
f
im
ag
es
in
a
p
o
s
t,
im
a
g
e
s
ize
an
d
im
ag
e
p
o
p
u
lar
ity
.
T
h
ese
f
ea
tu
r
es
ar
e
u
s
ed
as
in
p
u
ts
to
class
if
ier
s
alg
o
r
ith
m
s
s
u
ch
as
d
ec
is
io
n
tr
ee
s
,
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
es
(
SVMs)
,
lo
g
is
tic
r
e
g
r
ess
io
n
an
d
o
th
er
class
if
ier
s
to
class
if
y
an
im
ag
e
as
f
o
r
g
e
d
o
r
n
o
t.
Fo
r
th
ese
ty
p
es
o
f
m
eth
o
d
s
,
t
h
e
r
o
b
u
s
tn
ess
o
f
th
e
f
ea
tu
r
e
v
ec
to
r
s
o
b
tain
ed
is
n
o
t
s
u
f
f
ici
en
t,
as
k
n
o
wled
g
e
o
f
th
e
f
alsi
f
icatio
n
tr
ac
es
in
an
im
ag
e
is
lack
in
g
.
As
a
r
esu
lt,
it is
d
if
f
icu
lt to
u
s
e
th
ese
f
ea
tu
r
es to
d
etec
t f
alse im
ag
es with
ac
ce
p
tab
le
ac
cu
r
ac
y
.
C
ao
et
a
l.
[
1
3
]
p
r
o
v
id
es
an
in
-
d
ep
th
a
n
aly
s
is
o
f
v
is
u
al
c
o
n
t
en
t,
h
ig
h
li
g
h
tin
g
f
u
n
d
am
en
tal
co
n
ce
p
ts
,
im
p
o
r
tan
t
v
is
u
al
f
ea
tu
r
es,
e
f
f
ec
tiv
e
d
etec
tio
n
m
eth
o
d
s
,
an
d
th
e
ch
allen
g
es
en
co
u
n
t
er
ed
in
th
is
f
ield
.
B
er
th
et
[
1
4
]
f
o
cu
s
es
o
n
ar
tific
ial
in
tellig
en
ce
-
b
ased
co
m
p
r
es
s
io
n
to
o
ls
to
d
etec
t
f
o
r
g
ed
im
a
g
es.
I
n
s
tu
d
y
[
1
5
]
,
p
r
in
cip
al
co
m
p
o
n
e
n
t
an
aly
s
i
s
(
PC
A)
i
s
u
s
ed
to
d
etec
t
m
an
ip
u
lated
ar
tifa
cts
in
J
PEG
f
o
r
m
at
im
ag
es.
Vijay
alak
s
h
m
i
et
a
l.
[
1
6
]
in
tr
o
d
u
ce
s
an
a
u
to
en
co
d
er
-
b
ased
m
eth
o
d
f
o
r
i
d
en
tify
in
g
co
p
y
-
p
aste
f
o
r
g
e
r
ies
in
d
ig
ital
im
ag
es.
T
h
is
ap
p
r
o
ac
h
in
clu
d
es
im
ag
e
n
o
r
m
aliza
tio
n
,
r
escalin
g
,
an
d
er
r
o
r
lev
el
an
aly
s
is
(
E
L
A)
to
en
h
an
ce
ac
cu
r
ac
y
an
d
r
e
d
u
c
e
o
v
er
f
itti
n
g
in
th
e
n
etwo
r
k
m
o
d
el.
T
o
f
u
r
th
e
r
im
p
r
o
v
e
p
er
f
o
r
m
a
n
ce
,
im
a
g
e
au
g
m
en
tatio
n
is
ap
p
lied
to
in
cr
ea
s
e
th
e
d
ataset
s
ize.
Ultim
ately
,
th
e
p
r
o
p
o
s
ed
au
to
e
n
co
d
er
-
b
ased
tec
h
n
iq
u
e
ef
f
ec
tiv
ely
class
if
ies
f
o
r
g
ed
i
m
ag
es.
Stu
d
y
[
1
7
]
a
d
d
r
ess
es
th
e
d
etec
tio
n
o
f
cu
t
-
p
aste
m
an
i
p
u
latio
n
s
in
im
ag
es
u
s
in
g
tex
tu
r
e
an
aly
s
is
o
f
s
p
liced
im
ag
es.
Sp
ec
if
ically
,
it
e
x
tr
ac
ts
f
ea
tu
r
es
b
ased
o
n
th
e
lo
ca
l
en
tr
o
p
y
o
f
th
e
m
ed
ian
f
ilter
r
esid
u
al
(
MFR
)
o
f
th
e
m
an
ip
u
lated
im
ag
e,
w
h
ich
h
el
p
s
r
ed
u
ce
n
o
is
e
wh
il
e
p
r
eser
v
in
g
e
d
g
es.
T
h
ese
f
ea
tu
r
es
ar
e
th
en
u
s
ed
to
cr
ea
te
th
e
g
r
o
u
n
d
tr
u
th
m
a
s
k
.
T
h
e
g
o
al
o
f
s
tu
d
y
[
1
8
]
is
to
cr
ea
te
a
p
h
o
to
f
o
r
en
s
ics
alg
o
r
ith
m
ca
p
ab
le
o
f
d
etec
tin
g
all
ty
p
es
o
f
p
h
o
to
m
an
ip
u
latio
n
.
T
o
en
h
an
ce
th
e
er
r
o
r
lev
el
an
aly
s
is
,
th
e
s
tu
d
y
em
p
lo
y
s
v
er
tical
a
n
d
h
o
r
iz
o
n
tal
h
is
to
g
r
a
m
s
o
f
th
e
E
L
A
im
ag
e
to
ac
c
u
r
ately
i
d
e
n
tify
th
e
lo
ca
tio
n
o
f
m
o
d
if
icatio
n
s
.
T
h
r
o
u
g
h
th
e
p
r
ev
io
u
s
ly
cited
wo
r
k
s
i
n
t
h
is
s
ec
tio
n
,
we
o
b
s
er
v
e
th
at
d
if
f
e
r
en
t
t
y
p
es
o
f
r
esid
u
als,
s
u
ch
as
MFR
an
d
E
L
A,
ar
e
u
s
ed
.
T
h
ese
m
eth
o
d
s
ac
h
iev
e
h
ig
h
ac
cu
r
ac
y
f
o
r
s
in
g
le
m
an
ip
u
latio
n
s
b
u
t
s
h
o
w
lo
wer
ac
cu
r
ac
y
f
o
r
m
u
ltip
le
m
an
ip
u
latio
n
s
.
C
o
n
s
e
q
u
en
tly
,
th
eir
p
er
f
o
r
m
an
ce
wh
en
ap
p
lied
to
s
o
cial
m
ed
ia
im
ag
es is
in
s
u
f
f
icien
t,
a
s
th
ese
im
ag
es o
f
ten
u
n
d
er
g
o
m
u
ltip
le
m
an
ip
u
latio
n
s
[
1
9
]
.
I
n
r
ec
en
t
y
ea
r
s
,
ap
p
r
o
ac
h
es
b
ased
o
n
d
ee
p
n
eu
r
al
n
etwo
r
k
s
,
in
p
a
r
ticu
lar
co
n
v
o
lu
tio
n
n
eu
r
a
l
n
etwo
r
k
s
,
h
av
e
em
e
r
g
ed
[
1
4
]
.
Xu
e
et
a
l.
[
1
0
]
p
r
o
p
o
s
es
a
m
u
ltimo
d
al
n
eu
r
al
n
etwo
r
k
c
o
m
p
o
s
ed
o
f
s
ev
er
al
m
o
d
u
les,
in
clu
d
in
g
s
em
an
tic
f
ea
tu
r
e
e
x
tr
ac
tio
n
a
n
d
v
is
u
al
f
alsi
f
icatio
n
.
I
n
t
h
e
f
ir
s
t
m
o
d
u
le,
f
ea
tu
r
es
ar
e
ex
tr
ac
ted
f
r
o
m
th
e
p
r
e
-
tr
ain
e
d
R
esNet5
0
n
eu
r
al
n
etwo
r
k
,
an
d
th
ese
f
ea
tu
r
es
ar
e
p
ass
e
d
o
n
to
th
e
n
eu
r
a
l
n
etwo
r
k
(
B
iGR
U)
f
o
r
s
em
an
tic
f
ea
tu
r
e
ex
tr
ac
tio
n
.
T
h
e
f
ea
tu
r
es
o
f
th
e
f
alsi
f
ied
im
ag
e
ar
e
o
b
tain
ed
b
y
ap
p
ly
in
g
R
esNet5
0
to
th
e
E
L
A
tr
an
s
f
o
r
m
atio
n
o
f
th
e
im
a
g
e
.
I
n
s
tu
d
ies
[
2
0
]
an
d
[
2
1
]
,
to
d
etec
t
m
an
ip
u
latio
n
s
in
im
ag
es,
th
e
au
th
o
r
s
p
r
o
p
o
s
e
a
d
ee
p
lear
n
i
n
g
m
o
d
el.
I
m
a
g
es
g
en
er
ated
b
y
E
L
A
a
r
e
u
s
ed
as
in
p
u
ts
f
o
r
th
e
n
eu
r
al
n
etwo
r
k
m
o
d
el
E
f
f
ici
en
tNetB
0
.
T
h
is
m
o
d
el
is
tr
ain
ed
an
d
test
ed
o
n
Me
d
iaE
v
al
[
2
2
]
.
Sin
g
h
an
d
Sh
ar
m
a
[
2
0
]
ac
h
iev
e
a
n
ac
cu
r
ac
y
o
f
7
9
.
4
7
%
o
n
Me
d
iaE
v
al
an
d
in
[
2
1
]
th
e
y
ac
h
iev
e
a
n
a
cc
u
r
ac
y
o
f
8
1
.
0
7
%
o
n
Me
d
iaE
v
al.
So
m
e
wid
ely
u
s
ed
m
an
ip
u
lated
im
ag
es
m
ay
h
av
e
u
n
d
e
r
g
o
n
e
s
ev
er
al
ty
p
es
o
f
p
r
o
ce
s
s
in
g
,
th
u
s
in
cr
ea
s
in
g
th
e
d
if
f
icu
lty
o
f
c
ap
tu
r
in
g
m
an
ip
u
latio
n
tr
ac
es.
I
n
d
ee
d
,
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
s
u
c
h
as
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
25
:
2
3
0
4
-
2
3
1
3
2306
R
esNet
an
d
E
f
f
icien
tNet
h
av
e
d
em
o
n
s
tr
ated
s
o
m
e
ca
p
a
b
ilit
y
to
d
etec
t
v
ar
i
o
u
s
ty
p
e
s
o
f
m
an
ip
u
latio
n
s
.
Ho
wev
er
,
th
e
y
d
o
s
o
with
less
ef
f
icien
cy
a
n
d
a
r
e
m
o
r
e
ef
f
ec
tiv
e
at
lear
n
in
g
f
ea
tu
r
es
th
at
r
e
p
r
esen
t
th
e
co
n
ten
t
o
f
th
e
im
a
g
e
r
at
h
er
th
a
n
f
ea
tu
r
es
th
at
in
d
icate
th
e
p
r
esen
ce
o
f
m
an
ip
u
latio
n
.
T
h
eir
p
e
r
f
o
r
m
an
ce
,
wh
ich
h
o
v
e
r
s
ar
o
u
n
d
8
0
%
ac
cu
r
ac
y
,
n
ee
d
s
i
m
p
r
o
v
e
m
en
t.
T
h
is
is
wh
y
we
p
r
o
p
o
s
e
in
co
r
p
o
r
ati
n
g
a
c
o
n
s
tr
ain
ed
c
o
n
v
o
lu
tio
n
lay
er
to
au
to
m
atica
lly
lear
n
p
r
ed
ictio
n
er
r
o
r
f
ilter
s
in
th
e
in
i
tial
lay
er
.
T
h
is
ap
p
r
o
ac
h
en
a
b
les
th
e
is
o
latio
n
o
f
m
u
ltip
le
f
alsi
f
icatio
n
ar
tifa
ct
s
in
tr
o
d
u
ce
d
in
to
a
n
im
ag
e
b
y
elim
in
atin
g
ir
r
elev
a
n
t
i
n
f
o
r
m
atio
n
,
th
e
r
eb
y
ef
f
ec
tiv
ely
d
etec
tin
g
m
u
ltip
le
m
an
ip
u
latio
n
s
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
is
p
ap
er
is
as
f
o
llo
ws:
Se
ctio
n
2
o
u
tlin
es
th
e
m
eth
o
d
o
l
o
g
y
em
p
l
o
y
ed
an
d
elu
cid
ates
th
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
s
.
Sectio
n
3
d
elin
ea
tes
th
e
co
n
d
u
cted
ex
p
er
im
en
ts
,
p
r
o
v
i
d
es
an
aly
s
is
,
an
d
in
ter
p
r
ets
th
e
o
b
tain
ed
r
esu
lts
.
Fin
ally
,
th
e
co
n
cl
u
s
io
n
o
f
f
e
r
s
a
s
u
m
m
ar
y
o
f
o
u
r
s
tu
d
y
a
n
d
o
u
tlin
es f
u
tu
r
e
d
ir
ec
tio
n
s
.
2.
M
E
T
H
O
D
T
h
is
s
ec
tio
n
is
s
u
b
d
iv
id
ed
in
t
o
two
s
u
b
s
ec
tio
n
s
,
th
e
f
ir
s
t
o
f
wh
ich
is
th
e
p
r
o
b
lem
f
o
r
m
u
l
atio
n
an
d
f
ea
tu
r
es e
x
tr
ac
tio
n
m
eth
o
do
lo
g
y
,
f
o
llo
wed
b
y
th
e
ex
p
er
im
en
tatio
n
m
eth
o
d
.
2
.
1
.
P
r
o
blem
f
o
rm
ula
t
io
n
a
nd
f
ea
t
ures e
x
t
ra
ct
io
n m
e
t
ho
do
lo
g
y
Fals
if
ied
im
ag
e
d
etec
tio
n
ca
n
b
e
m
o
d
eled
as
a
b
in
ar
y
class
if
icatio
n
p
r
o
b
lem
th
at
in
d
icate
s
wh
eth
er
an
im
ag
e
is
g
en
u
in
e
o
r
f
alsi
f
ied
.
C
o
n
s
id
er
=
{
1
,
…
,
}
⊂
ℝ
x
x
th
e
in
p
u
t
f
ea
tu
r
es,
=
{
1
,
…
,
}
⊂
ℝ
th
e
co
r
r
esp
o
n
d
in
g
lab
els.
T
h
e
p
r
o
b
lem
is
to
f
in
d
a
f
u
n
ctio
n
ℱ
th
at
au
to
m
atica
lly
lear
n
s
to
r
ec
o
g
n
ize
th
e
ch
ar
ac
ter
is
tics
o
f
an
im
ag
e
an
d
to
p
r
e
d
ict
its
tr
u
th
f
u
ln
ess
,
i
.
e
.
ℱ
(
)
=
{
0
1
(
1
)
T
h
e
f
ea
tu
r
e
ex
tr
ac
tio
n
m
o
d
el
co
n
s
is
ts
o
f
two
m
o
d
u
les
as
s
h
o
wn
in
Fig
u
r
e
1
.
On
e
f
o
r
th
e
e
x
tr
ac
tio
n
o
f
f
ea
tu
r
es
f
r
o
m
wea
th
er
i
n
g
tr
ac
es a
n
d
th
e
s
ec
o
n
d
f
o
r
th
e
e
x
tr
ac
tio
n
o
f
s
p
atial
f
ea
tu
r
es f
r
o
m
th
e
im
ag
e.
Fig
u
r
e
1
.
Fals
if
icatio
n
an
d
s
p
a
tial v
is
u
al
f
ea
tu
r
es e
x
tr
ac
tio
n
b
y
B
ay
ar
_
Att
T
h
e
f
ea
tu
r
e
e
x
tr
ac
tio
n
m
o
d
u
l
e
f
o
r
wea
th
er
in
g
tr
ac
es
as
s
h
o
wn
in
Fig
u
r
e
2
is
co
m
p
o
s
ed
with
th
e
co
m
b
in
atio
n
o
f
co
n
s
tr
ain
ed
c
o
n
v
o
lu
ti
o
n
p
r
o
p
o
s
ed
b
y
B
ay
ar
an
d
Stam
m
[
2
3
]
an
d
atten
ti
o
n
m
ec
h
an
is
m
f
o
r
r
elev
an
t
f
ea
tu
r
es
ex
tr
ac
tio
n
b
ased
o
n
atten
tio
n
m
ec
h
an
is
m
ap
p
lied
in
[
2
4
]
.
I
t
is
m
ad
e
u
p
o
f
s
ev
er
al
n
eu
r
al
n
etwo
r
k
lay
er
s
:
a
co
n
s
tr
ain
e
d
co
n
v
o
lu
tio
n
la
y
er
f
o
r
ex
tr
a
ctin
g
lo
w
-
lev
el
tam
p
er
i
n
g
f
e
atu
r
es,
an
atten
tio
n
m
o
d
u
le
f
o
r
ex
tr
ac
tin
g
m
o
r
e
im
p
o
r
tan
t
f
ea
tu
r
es,
a
n
d
a
c
o
n
v
o
lu
tio
n
lay
er
f
o
llo
wed
b
y
a
p
o
o
lin
g
la
y
er
t
o
im
p
r
o
v
e
th
e
g
en
e
r
aliza
tio
n
c
ap
ab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el.
A
C
o
n
v
o
l
u
tio
n
lay
er
is
a
s
et
o
f
f
ilter
s
o
r
m
atr
ices
ap
p
lied
b
y
co
n
v
o
l
u
tio
n
o
p
er
atio
n
o
n
an
o
th
e
r
m
atr
i
x
.
Fil
ter
s
ar
e
f
ea
tu
r
e
ex
tr
ac
to
r
s
an
d
th
e
r
esu
lt
o
f
th
e
co
n
v
o
lu
tio
n
is
ca
lled
a
f
ea
tu
r
e
m
ap
.
L
et
b
e
th
e
in
p
u
t
im
ag
e
o
f
d
im
en
s
io
n
×
×
wh
er
e
is
th
e
len
g
th
o
f
th
e
im
ag
e,
th
e
wid
th
an
d
th
e
n
u
m
b
e
r
o
f
ch
a
n
n
els.
T
h
is
im
ag
e
co
n
v
o
lv
es
with
a
f
ilter
o
f
d
im
en
s
io
n
×
×
with
a
s
tep
o
f
1
p
r
o
d
u
ce
s
th
e
ch
ar
ac
ter
is
tic
m
ap
=
(
∗
)
+
o
f
d
im
e
n
s
io
n
(
−
+
1
)
×
(
−
+
1
)
wh
er
e
*
is
th
e
co
n
v
o
lu
tio
n
o
p
er
atio
n
,
an
ac
tiv
atio
n
f
u
n
cti
o
n
an
d
th
e
b
ias.
Fo
r
f
ilter
s
ap
p
lied
o
n
th
e
im
a
g
e
we
o
b
tain
f
ea
tu
r
es
m
a
p
s
at
th
e
o
u
tp
u
t
o
f
th
e
c
o
n
v
o
lu
tio
n
la
y
er
i.e
.
th
e
o
u
tp
u
t
o
f
t
h
e
co
n
v
o
l
u
tio
n
lay
e
r
is
o
f
d
im
en
s
io
n
(
−
+
1
)
x
(
−
+
1
)
×
.
I
n
o
u
r
p
r
o
p
o
s
al,
in
th
e
B
ay
ar
co
n
s
tr
ain
e
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
co
n
s
tr
a
in
ed
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
k
w
ith
a
tten
tio
n
mec
h
a
n
is
m
fo
r
(
K
a
ma
g
a
te
B
e
ma
n
Ha
mid
ja
)
2307
co
n
v
o
l
u
tio
n
lay
er
,
we
p
r
o
p
o
s
e
to
u
s
e
7
f
ilter
s
o
f
s
ize
5
×
5
o
n
an
in
p
u
t
o
f
150
×
150
×
3
.
T
h
is
lay
er
p
r
o
d
u
ce
s
a
f
ea
tu
r
e
v
ec
to
r
.
Fig
u
r
e
2
.
W
ea
th
er
in
g
tr
ac
es e
x
tr
ac
tio
n
m
o
d
u
le
I
n
s
p
ir
ed
b
y
th
e
atten
tio
n
m
ec
h
an
is
m
o
f
s
tu
d
y
[
2
5
]
wh
ich
a
llo
ws
im
p
o
r
tan
t
f
ea
tu
r
es
to
b
e
s
elec
ted
an
d
ad
ap
ti
v
ely
,
we
ad
d
th
e
att
en
tio
n
m
ec
h
a
n
is
m
m
o
d
u
le
in
Fig
u
r
e
2
.
T
h
is
atten
tio
n
m
o
d
u
l
e
is
f
ir
s
t m
ad
e
u
p
o
f
an
av
er
ag
e
p
o
o
lin
g
(
AP)
lay
er
with
wh
ich
we
o
b
tain
th
e
f
ea
tu
r
e
v
ec
to
r
=
(
)
.
Seco
n
d
ly
it
i
s
m
ad
e
u
p
o
f
a
s
p
atial
f
ea
tu
r
e
ex
tr
ac
tio
n
lay
e
r
f
o
llo
wed
b
y
an
with
wh
ich
we
o
b
tain
t
h
e
f
ea
tu
r
e
v
ec
to
r
=
(
∘
)
.
A
s
u
m
o
f
an
d
is
p
r
o
d
u
ce
d
in
o
r
d
er
t
o
ca
p
t
u
r
e
t
h
e
two
ty
p
es
o
f
ch
ar
ac
ter
is
tics
,
th
en
a
(
)
ac
tiv
atio
n
f
u
n
ctio
n
is
ap
p
lied
to
th
e
s
u
m
o
f
an
d
to
allo
w
th
e
n
etwo
r
k
to
b
e
m
o
r
e
atten
tiv
e
to
th
e
m
o
s
t
im
p
o
r
tan
t
r
eg
io
n
s
.
T
h
e
o
b
tain
ed
f
ea
tu
r
es
ar
e
p
r
o
p
ag
ate
d
o
n
th
e
B
ay
ar
lay
er
in
o
r
d
er
t
o
o
b
tain
th
e
atten
tio
n
f
ea
tu
r
es k
in
(
2
)
.
=
(
⊕
)
⨂
(
2
)
T
o
en
h
an
ce
th
e
m
o
d
el'
s
ab
ili
ty
to
g
en
er
alize
an
d
r
ef
in
e
th
e
f
ea
tu
r
e
s
elec
tio
n
,
we
em
p
lo
y
a
co
n
v
o
lu
tio
n
a
l
o
p
er
atio
n
o
n
t
h
e
p
r
ec
ed
in
g
la
y
er
.
T
h
is
o
p
er
atio
n
u
tili
ze
s
1
6
f
ilter
s
,
ea
ch
with
a
s
ize
o
f
3
×
3
.
Su
b
s
eq
u
en
tly
,
we
ap
p
ly
a
g
lo
b
al
av
e
r
ag
e
p
o
o
lin
g
(
GAP)
lay
er
.
T
h
e
r
esu
ltin
g
f
ea
tu
r
e
v
ec
to
r
is
d
e
n
o
ted
b
y
(
3
)
:
=
(
∘
)
(
3
)
T
h
is
m
o
d
u
le,
r
esp
o
n
s
ib
le
f
o
r
ex
tr
ac
tin
g
f
ea
tu
r
es
f
r
o
m
alter
atio
n
tr
ac
es,
is
d
etailed
in
Al
g
o
r
ith
m
2
,
wh
ich
is
ca
lled
Alg
o
r
ith
m
1
.
Alg
o
r
ith
m
2
s
tan
d
s
f
o
r
a
r
tifa
ct
f
ea
t
u
r
e
s
ex
tr
ac
tio
n
an
d
Alg
o
r
ith
m
1
is
f
o
r
t
h
e
c
o
n
tain
ed
co
n
v
o
l
u
tio
n
lay
er
.
Alg
o
r
ith
m
1
.
C
o
n
s
tr
ain
ed
co
n
v
o
lu
tio
n
la
y
er
Initialize randomly the weights
=
1
While
(
≤
max
_
){
perform a feedforward pass
Update the filter weights using stochastic gradient
descent and backpropagate the errors
For each
filters
Define
(
0
,
0
)
(
1
)
=
0
Normalize
so that
∑
(
,
)
(
1
)
=
1
,
≠
0
Define
(
0
,
0
)
(
1
)
=
−
1
End for
=
+
1
If the training accuracy converges, then Exit
}
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
25
:
2
3
0
4
-
2
3
1
3
2308
Alg
o
r
ith
m
2
.
Ar
tifa
ct
f
ea
tu
r
es
Input:
∈
ℝ
x
x
: an image
Output:
:
Manipulation traces features
Begin
Use
Algorithm
1 to obtain
Feature selection by attention mechanism
followed by Average Pooling:
=
(
)
Select by Convolution and by Average
Pooling":
=
(
∘
)
Element
-
wise summation:
=
⊕
Normalization:
=
(
)
Diffusion on
:
=
⊗
Refine the selection to obtain artifact features
=
(
∘
)
End
As
f
o
r
th
e
s
p
atial
f
ea
tu
r
e
ex
tr
ac
tio
n
as
s
h
o
wn
in
Fig
u
r
e
1
,
th
e
f
ir
s
t
lay
er
s
lear
n
l
o
w
-
lev
e
l
f
ea
tu
r
es
s
u
ch
as
ed
g
es,
co
lo
r
s
,
an
d
a
s
th
e
n
u
m
b
er
o
f
th
ese
lay
er
s
in
cr
ea
s
es,
th
e
f
ea
tu
r
e
lear
n
i
n
g
b
ec
o
m
es
m
o
r
e
ac
cu
r
ate
[
2
6
]
.
Settin
g
u
p
s
u
ch
a
n
etwo
r
k
is
co
s
tly
in
ter
m
s
o
f
co
m
p
u
tin
g
p
o
wer
an
d
th
e
s
ize
o
f
th
e
tr
ain
in
g
d
ata
[
2
7
]
.
An
o
th
e
r
alter
n
ativ
e
is
to
ap
p
ly
tr
an
s
f
er
lear
n
in
g
.
T
r
an
s
f
er
lear
n
i
n
g
is
a
m
ac
h
in
e
lear
n
in
g
tech
n
iq
u
e
th
at
tr
an
s
f
er
s
k
n
o
wled
g
e
ac
q
u
ir
ed
in
o
n
e
o
r
m
o
r
e
s
o
u
r
ce
task
s
in
o
r
d
er
to
u
s
e
it
to
im
p
r
o
v
e
lear
n
in
g
in
a
r
elate
d
tar
g
et
task
[
2
8
]
.
T
h
e
r
e
f
o
r
e,
we
u
s
e
th
e
p
r
e
-
tr
ain
e
d
R
esNet5
0
[
2
7
]
m
o
d
el
to
o
b
tain
th
e
s
p
atial
f
ea
tu
r
e
v
ec
to
r
.
T
h
e
R
esNet5
0
m
o
d
el
is
a
C
NN
m
o
d
el
co
m
p
o
s
ed
o
f
5
0
lay
e
r
s
.
T
h
e
ar
ch
itectu
r
e
o
f
th
e
R
esNet5
0
in
Fig
u
r
e
3
m
o
d
el
u
s
ed
is
th
at
o
f
[
2
7
]
,
ex
ce
p
t
th
at
we
r
em
o
v
e
d
th
e
last
lay
er
co
n
s
is
tin
g
o
f
t
h
e
av
er
ag
e
p
o
o
lin
g
(
AP)
,
th
e
f
u
lly
co
n
n
ec
ted
lay
er
an
d
th
e
class
if
icatio
n
lay
er
b
y
th
e
g
lo
b
al
av
e
r
ag
e
p
o
o
lin
g
(
GAP)
la
y
er
.
T
h
is
ar
ch
itectu
r
e
is
d
escr
ib
ed
as f
o
llo
ws:
Fig
u
r
e
3
.
T
r
an
s
f
er
lear
n
R
esNet5
0
ar
ch
itectu
r
e
L
et
50
=
50
(
)
b
e
th
e
f
ea
tu
r
e
v
e
cto
r
o
b
t
ain
ed
af
ter
ap
p
licatio
n
o
f
R
esNet5
0
.
B
ef
o
r
e
th
e
last
lay
er
wh
ich
is
th
e
cla
s
s
if
icatio
n
lay
er
,
we
f
ir
s
t
p
o
o
l
th
e
p
r
ev
io
u
s
ch
ar
ac
ter
is
tics
an
d
50
.
T
h
en
a
d
en
s
e
la
y
er
is
ad
d
ed
in
o
r
d
er
to
lear
n
th
e
s
h
ar
e
d
f
ea
tu
r
es.
W
e
f
in
ally
o
b
tain
t
h
e
f
o
llo
win
g
f
ea
tu
r
e
v
ec
to
r
:
=
(
(
⊕
50
)
)
wh
er
e
is
th
e
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
ac
tiv
ati
o
n
f
u
n
ctio
n
an
d
th
e
weig
h
ts
o
f
th
e
d
en
s
e
lay
er
.
As
o
u
r
p
r
o
b
lem
is
a
b
in
ar
y
class
if
icatio
n
,
we
u
s
e
th
e
s
ig
m
o
id
f
u
n
ctio
n
f
o
r
th
e
d
is
tr
ib
u
tio
n
o
f
p
r
ed
ictio
n
s
.
At
th
e
o
u
t
p
u
t o
f
o
u
r
ar
ch
i
tectu
r
e,
we
o
b
tain
t
h
e
p
r
e
d
ictio
n
f
u
n
ctio
n
(
4
)
:
ℱ
=
(
)
=
1
1
+
(
4
)
T
o
allo
w
o
u
r
m
o
d
el
to
lear
n
an
d
allo
w
it
to
im
p
r
o
v
e
in
th
e
p
r
ed
ictio
n
o
f
ℱ
,
an
er
r
o
r
f
u
n
ctio
n
is
ca
lcu
lated
an
d
m
in
im
ized
ac
c
o
r
d
in
g
to
p
a
r
am
eter
s
.
I
n
th
is
s
tu
d
y
,
we
ad
o
p
t
th
e
cr
o
s
s
-
en
tr
o
p
y
-
b
ased
er
r
o
r
f
u
n
ctio
n
.
I
t
is
a
f
u
n
ctio
n
th
at
m
ea
s
u
r
es
th
e
d
if
f
er
en
ce
b
et
wee
n
th
e
m
o
d
el'
s
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
an
d
th
e
p
r
ed
icted
d
is
tr
ib
u
tio
n
.
I
t is d
es
cr
ib
ed
as f
o
llo
ws b
y
(
5
)
:
(
ℱ
;
)
=
−
1
∑
[
l
og
(
ℱ
(
)
)
+
(
1
−
)
l
og
(
1
−
ℱ
(
)
)
]
=
1
(
5
)
{
0
,
1
}
,
N
th
e
n
u
m
b
er
o
f
o
b
s
er
v
atio
n
s
,
th
e
in
p
u
t
ch
a
r
ac
ter
is
tics
o
f
th
e
th
im
ag
e
an
d
th
e
class
if
icatio
n
p
ar
am
eter
s
.
T
h
e
p
ar
am
eter
s
ar
e
o
p
tim
ized
b
y
m
in
i
m
izatio
n
o
f
th
e
er
r
o
r
f
u
n
ctio
n
wh
ich
g
i
v
es b
y
(
6
)
:
̂
=
min
(
ℱ
;
)
(
6
)
Im
a
g
e
Glo
b
a
l
Av
e
ra
g
e
P
o
o
l
in
g
Co
n
v
7
×
7
,
6
4
Co
n
v
1
×
1
,
6
4
Co
n
v
3
×
3
,
6
4
Co
n
v
1
×
1
,
2
5
6
Co
n
v
1
×
1
,
5
1
2
Co
n
v
7
×
7
,
5
1
2
Co
n
v
7
×
7
,
2
0
4
8
…………
…….
50
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
co
n
s
tr
a
in
ed
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
k
w
ith
a
tten
tio
n
mec
h
a
n
is
m
fo
r
(
K
a
ma
g
a
te
B
e
ma
n
Ha
mid
ja
)
2309
2
.
2
.
E
x
perim
ent
a
t
io
n m
et
ho
d
Fo
r
im
p
lem
en
tin
g
o
u
r
m
o
d
els,
we
u
s
ed
an
HP
C
o
r
e
i7
co
m
p
u
ter
with
1
6
GB
o
f
m
em
o
r
y
an
d
a
6
4
-
b
it
o
p
er
atin
g
s
y
s
tem
.
W
e
em
p
lo
y
ed
Py
th
o
n
3
an
d
th
e
f
o
ll
o
win
g
lib
r
ar
ies:
Pan
d
as
f
o
r
tr
a
n
s
f
o
r
m
in
g
th
e
d
ataset
in
to
a
D
ata
F
r
am
e,
Nu
m
P
y
f
o
r
m
atr
ix
ca
lcu
latio
n
s
,
Ma
tp
lo
tlib
f
o
r
d
ata
v
is
u
aliza
tio
n
,
Op
en
C
V
f
o
r
p
r
ep
r
o
ce
s
s
in
g
r
aw
im
ag
es,
an
d
Ker
as with
T
en
s
o
r
Flo
w
f
o
r
d
esig
n
in
g
an
d
tr
ain
in
g
d
ee
p
le
ar
n
in
g
m
o
d
els.
R
eg
ar
d
in
g
th
e
d
ataset,
we
u
s
ed
r
o
y
alty
-
f
r
ee
d
atasets
co
m
m
o
n
ly
u
s
ed
in
t
h
e
liter
atu
r
e
f
o
r
ev
alu
atin
g
im
ag
e
m
a
n
ip
u
latio
n
m
o
d
els.
T
h
e
f
ir
s
t
is
Me
d
iaE
v
al
[
2
2
]
,
a
d
ataset
co
llected
f
r
o
m
T
witter
as
p
ar
t
o
f
th
e
au
to
m
atic
d
etec
tio
n
o
f
th
e
m
a
n
ip
u
latio
n
an
d
m
is
u
s
e
o
f
m
u
lt
im
ed
ia
co
n
ten
t
o
n
th
e
we
b
.
T
h
ese
m
an
ip
u
latio
n
s
in
clu
d
e
ass
em
b
lin
g
,
d
eletin
g
,
ad
d
in
g
,
an
d
o
u
t
-
of
-
c
o
n
tex
t im
ag
es.
E
ac
h
en
tr
y
in
th
is
d
ataset
is
ac
co
m
p
an
ied
b
y
tex
tu
al
co
n
te
n
t,
an
im
ag
e
o
r
v
i
d
eo
,
a
n
d
s
o
cial
co
n
te
x
t
in
f
o
r
m
atio
n
.
T
h
e
d
im
e
n
s
io
n
s
o
f
th
is
d
ataset
r
an
g
e
f
r
o
m
100
×
1
0
0
p
ix
els
to
2
7
0
9
×
3
4
0
0
p
ix
els.
T
h
e
s
ec
o
n
d
d
ataset
is
C
ASI
A
[
2
9
]
,
wh
ich
co
n
t
ain
s
7
,
4
9
1
g
en
u
i
n
e
im
ag
es
an
d
5
,
1
2
3
tam
p
er
ed
i
m
ag
es.
T
h
e
f
alsi
f
ied
im
ag
es
in
th
is
d
ataset
ar
e
r
ea
l
im
ag
es
m
an
ip
u
lated
f
ir
s
t
b
y
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es
s
u
c
h
as
cr
o
p
p
i
n
g
,
d
is
to
r
tio
n
,
an
d
r
o
tatio
n
,
th
e
n
b
y
s
titch
in
g
o
p
er
atio
n
s
an
d
p
o
s
t
-
p
r
o
ce
s
s
in
g
o
p
er
atio
n
s
lik
e
b
lu
r
r
in
g
o
n
e
d
g
es
o
r
alter
ed
r
eg
i
o
n
s
.
I
m
ag
e
d
im
en
s
io
n
s
in
th
is
d
ataset
r
an
g
e
f
r
o
m
320
×
2
4
0
p
i
x
els to
8
0
0
×
6
0
0
p
i
x
els.
I
n
th
e
d
ata
p
r
ep
r
o
ce
s
s
in
g
,
we
r
ed
u
ce
d
R
GB
im
ag
es
to
1
5
0
×
1
5
0
p
ix
els.
T
h
e
Op
e
n
C
V,
Pil
lo
w,
an
d
Nu
m
Py
lib
r
ar
ies
wer
e
u
s
ed
to
r
ea
d
an
d
d
ig
itize
th
e
im
ag
es.
W
e
al
s
o
r
em
o
v
ed
d
u
p
licate
im
ag
es.
All
im
ag
es
wer
e
r
esized
to
a
wid
th
o
f
1
5
0
p
ix
els
an
d
a
h
eig
h
t
o
f
1
5
0
p
ix
els.
T
h
e
ex
p
er
im
en
t
was
co
n
d
u
cted
b
y
r
a
n
d
o
m
l
y
d
iv
id
in
g
th
e
d
atasets
with
8
0
%
allo
ca
ted
f
o
r
tr
ain
in
g
d
ata
a
n
d
2
0
%
f
o
r
test
d
ata.
T
h
e
tr
ain
in
g
d
ata
wer
e
u
s
ed
f
o
r
h
y
p
er
p
ar
am
eter
s
ea
r
ch
a
n
d
m
o
d
el
s
elec
tio
n
.
Per
f
o
r
m
an
ce
-
o
p
tim
izin
g
h
y
p
er
p
ar
a
m
e
ter
s
,
s
u
ch
as
th
e
n
u
m
b
er
an
d
s
ize
o
f
f
ilter
s
in
co
n
v
o
l
u
tio
n
lay
er
s
,
d
r
o
p
o
u
t
r
ate,
b
atch
s
ize,
an
d
s
tr
id
e
in
p
o
o
lin
g
lay
er
s
,
wer
e
o
b
tain
ed
u
s
in
g
Gr
id
Sear
ch
f
r
o
m
th
e
Ker
as
lib
r
ar
y
o
n
tr
ai
n
in
g
d
ata
f
r
o
m
[
2
3
]
.
W
e
v
a
r
ied
t
h
e
b
atch
s
ize
in
th
e
s
et
{5
,
1
5
,
3
0
,
3
2
,
4
0
,
5
0
},
th
e
n
u
m
b
er
o
f
f
ilter
s
in
th
e
s
et
{1
,
2
,
3
,
4
,
5
,
6
,
7
,
8
,
9
,
1
0
,
1
6
,
3
2
},
th
e
f
ilter
s
ize
in
th
e
s
et
{3
,
5
,
7
}
,
th
e
d
r
o
p
o
u
t
r
ate
in
th
e
s
et
{0
.
1
,
0
.
2
,
0
.
3
,
0
.
4
,
0
.
5
,
0
.
6
,
0
.
7
,
0
.
8
,
0
.
9
},
t
h
e
p
o
o
lin
g
s
ize
in
th
e
s
et
{2
,
3
,
4
,
5
},
an
d
th
e
p
o
o
lin
g
s
tep
in
th
e
s
et
{1
,
2
,
3
}.
T
h
e
tr
ain
in
g
p
r
o
ce
s
s
was
s
to
p
p
ed
wh
en
th
e
l
o
s
s
f
u
n
ctio
n
c
o
n
v
er
g
ed
,
a
n
d
th
e
Ad
am
o
p
tim
izer
was
u
s
ed
.
T
h
e
p
ar
am
eter
s
o
b
tain
ed
wer
e:
a
b
atch
s
ize
o
f
3
2
,
a
co
n
s
tr
ain
ed
c
o
n
v
o
l
u
tio
n
la
y
er
with
7
f
ilter
s
o
f
s
ize
5
,
p
o
o
li
n
g
o
f
s
ize
3
with
a
s
tep
s
ize
o
f
2
,
an
d
a
s
ec
o
n
d
co
n
v
o
l
u
tio
n
lay
er
with
1
6
f
ilt
er
s
o
f
s
ize
3
.
T
h
e
f
in
al
r
esu
lts
wer
e
o
b
tain
ed
b
y
s
elec
tin
g
t
h
e
m
o
d
el
th
at
g
av
e
th
e
b
est AUC in
cr
o
s
s
-
v
alid
atio
n
with
5
f
o
ld
s
.
T
o
co
n
d
u
ct
a
co
m
p
a
r
ativ
e
ev
alu
atio
n
o
f
t
h
e
p
e
r
f
o
r
m
an
ce
o
f
o
u
r
p
r
o
p
o
s
ed
m
o
d
el,
n
am
ed
B
ay
ar
R
esn
et,
wh
ich
u
tili
ze
s
a
co
n
s
tr
ain
ed
co
n
v
o
l
u
tio
n
la
y
er
alo
n
g
with
an
atten
tio
n
m
ec
h
an
is
m
an
d
a
tr
an
s
f
er
lear
n
in
g
R
esNet5
0
n
etwo
r
k
,
two
ad
d
itio
n
al
m
o
d
els
wer
e
tr
ain
ed
a
n
d
test
ed
o
n
th
e
p
r
e
v
io
u
s
ly
m
en
tio
n
ed
d
atasets
,
u
s
in
g
th
e
s
am
e
p
ar
am
eter
s
estab
lis
h
ed
d
u
r
in
g
th
e
d
ata
p
r
e
p
r
o
ce
s
s
in
g
s
tep
.
T
h
e
f
ir
s
t
is
th
e
Sin
g
h
Z
am
il
m
o
d
el,
as
o
u
tli
n
ed
in
r
ef
er
e
n
ce
s
[
2
1
]
,
[
2
2
]
,
wh
ich
in
teg
r
ates
E
L
A
with
th
e
p
r
e
-
tr
ain
ed
E
f
f
icien
tB
0
m
o
d
el
t
o
id
en
tif
y
m
an
ip
u
lated
im
ag
es.
T
h
e
s
ec
o
n
d
is
B
ay
ar
E
f
f
,
wh
ich
r
ef
er
s
to
th
e
B
ay
a
r
E
f
f
icien
tNet
m
o
d
el,
lik
ely
in
t
eg
r
atin
g
th
e
m
eth
o
d
o
lo
g
ies
o
r
en
h
a
n
ce
m
en
ts
p
r
o
p
o
s
ed
b
y
B
ay
ar
an
d
Stam
m
[
2
3
]
with
in
t
h
e
E
f
f
icien
tNet
ar
ch
itectu
r
e.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
Resul
t
s
T
h
e
r
esu
lts
s
h
o
w
th
at
th
e
p
r
o
p
o
s
ed
B
ay
ar
R
esn
et
m
eth
o
d
is
b
etter
th
an
th
e
o
th
er
s
in
ter
m
s
o
f
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all
an
d
s
p
ec
if
icity
o
n
b
o
th
ca
s
s
ia
a
n
d
Me
d
iev
al
d
ataset
in
T
a
b
le
s
1
an
d
2
.
On
th
e
Me
d
iaE
v
al
d
ataset,
th
e
B
ay
ar
R
esn
et
p
r
o
p
o
s
al
g
iv
es
b
etter
r
esu
lts
f
o
r
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all
an
d
F1
-
s
co
r
e
an
d
ev
en
o
n
s
p
ec
if
icity
.
Sin
g
Z
am
il
m
o
d
el
o
u
t
p
er
f
o
r
m
s
B
ay
ar
R
esn
et
o
n
ly
o
n
s
p
ec
if
icity
wh
en
u
s
in
g
C
ASI
A
Data
s
et.
T
ab
le
1
.
Per
f
o
r
m
an
ce
o
f
v
ar
io
u
s
ap
p
r
o
ac
h
es u
s
in
g
Me
d
iaE
v
al
d
ataset
M
o
d
e
l
s
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
S
p
e
c
i
f
i
c
i
t
y
F1
-
sc
o
r
e
S
i
n
g
Za
mi
l
0
.
7
1
6
0
.
7
0
2
0
.
7
2
2
0
.
71
0
.
7
1
2
B
a
y
a
r
Ef
f
0
.
8
5
1
0
.
9
0
3
0
.
7
7
7
0
.
9
2
1
0
.
8
3
5
B
a
y
a
r
R
e
sn
e
t
0
.
8
7
8
0
.
9
3
5
4
0
.
8
0
5
0
.
9
4
7
0
.
8
6
5
T
ab
le
2
.
Per
f
o
r
m
an
ce
o
f
v
ar
io
u
s
ap
p
r
o
ac
h
es u
s
in
g
C
ASI
A
d
ataset
M
o
d
e
l
s
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
S
p
e
c
i
f
i
c
i
t
y
F1
-
sc
o
r
e
S
i
n
g
Za
mi
l
0
.
6
6
5
0
.
6
0
7
0
.
4
9
4
0
.
7
8
1
0
.
5
4
5
B
a
y
a
r
Ef
f
0
.
6
8
4
0
.
6
0
4
0
.
65
0
.
7
0
8
0
.
6
2
6
B
a
y
a
r
R
e
sn
e
t
0
.
7
0
5
0
.
6
2
3
0
.
6
9
6
0
.
7
1
1
0
.
6
5
7
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
25
:
2
3
0
4
-
2
3
1
3
2310
T
h
e
Me
d
iaE
v
al
d
ataset
is
tr
ai
n
ed
with
a
p
r
o
p
o
r
tio
n
o
f
8
0
%
o
f
th
e
d
ata
o
v
er
3
0
ep
o
c
h
s
.
Alr
ea
d
y
at
th
e
1
0
th
e
p
o
ch
d
u
r
in
g
tr
ain
in
g
,
th
e
B
ay
ar
R
esn
et
an
d
B
ay
ar
E
f
f
m
o
d
els
ac
h
iev
e
o
v
er
9
0
%
h
i
g
h
er
ac
cu
r
ac
y
th
a
n
th
e
o
th
er
m
o
d
els.
T
h
e
tr
ain
in
g
an
d
v
alid
atio
n
ac
cu
r
ac
ies
o
f
th
e
B
ay
ar
R
esn
et
an
d
B
ay
ar
E
f
f
m
o
d
els
ar
e
b
etter
an
d
in
cr
ea
s
e
with
ea
ch
e
p
o
c
h
.
B
u
t
f
in
ally
,
B
ay
ar
R
esn
et
p
er
f
o
m
s
well
th
an
B
ay
ar
E
f
f
.
T
h
is
s
itu
atio
n
is
illu
s
tr
ated
b
y
Fig
u
r
e
4
.
Fig
u
r
e
4
.
T
r
ain
in
g
a
n
d
v
alid
atio
n
ac
cu
r
ac
y
C
o
n
ce
r
n
in
g
th
e
lo
s
s
v
alu
es
in
Fig
u
r
e
5
,
o
v
er
lear
n
in
g
is
o
b
s
er
v
ed
f
o
r
t
h
e
B
ay
ar
m
o
d
el
f
r
o
m
th
e
1
0
th
ep
o
ch
an
d
f
o
r
t
h
e
Sin
g
h
Z
am
il
m
o
d
el
f
r
o
m
th
e
5
th
ep
o
c
h
.
I
n
ad
d
itio
n
,
th
e
v
ali
d
atio
n
an
d
t
r
ain
in
g
lo
s
s
v
alu
es
o
f
th
e
o
t
h
er
two
m
o
d
els
(
B
ay
er
R
esn
et,
B
ay
aE
f
f
)
d
ec
r
ea
s
e
p
r
o
g
r
ess
iv
ely
to
war
d
s
ze
r
o
an
d
s
tab
ilize
f
r
o
m
th
e
15
th
ep
o
ch
.
T
h
e
s
m
allest lo
s
s
v
alu
es a
r
e
o
b
s
er
v
e
d
with
th
e
B
ay
ar
R
esn
et
p
r
o
p
o
s
al.
Fig
u
r
e
5
.
L
o
s
s
cu
r
v
e
d
u
r
in
g
tr
ain
in
g
an
d
v
alid
atio
n
Fig
u
r
e
6
d
is
p
la
y
s
t
h
e
R
OC
c
u
r
v
e
a
ch
ie
v
e
d
o
n
t
h
e
M
ed
iEV
al
d
at
ase
t,
w
h
il
e
F
ig
u
r
e
7
d
ep
icts
t
h
e
c
u
r
v
e
f
o
r
t
h
e
C
AS
I
A
d
at
ase
t.
T
h
e
B
ay
ar
R
esn
et
m
o
d
el
e
x
h
i
b
its
t
h
e
h
i
g
h
est
a
r
ea
u
n
d
e
r
t
h
e
c
u
r
v
e,
ap
p
r
o
x
i
m
a
tel
y
8
7
%
o
n
t
h
e
M
ed
iEV
al
d
atas
et
a
n
d
7
0
%
o
n
t
h
e
C
AS
I
A
d
atas
et.
I
n
co
n
t
r
as
t,
t
h
e
Si
n
g
h
Z
a
m
il
m
o
d
el
h
as
th
e
l
o
w
est
r
es
p
e
cti
v
e
a
r
e
as
u
n
d
e
r
th
e
c
u
r
v
e,
a
p
p
r
o
x
i
m
at
el
y
7
1
%
f
o
r
M
ed
iEV
a
l
an
d
6
3
%
f
o
r
C
ASI
A.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
co
n
s
tr
a
in
ed
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
k
w
ith
a
tten
tio
n
mec
h
a
n
is
m
fo
r
(
K
a
ma
g
a
te
B
e
ma
n
Ha
mid
ja
)
2311
Fig
u
r
e
6
.
R
OC
cu
r
v
e
o
n
Me
d
i
E
VAL
Fig
u
r
e
7
.
R
OC
cu
r
v
e
o
n
C
ASI
A
3
.
2
.
Dis
cu
s
s
io
n
I
n
th
is
s
tu
d
y
,
we
co
n
d
u
cte
d
ex
p
er
im
en
ts
to
co
m
p
ar
e
th
e
p
er
f
o
r
m
an
ce
o
f
o
u
r
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
,
b
ased
o
n
th
e
B
ay
ar
R
esn
et
m
o
d
el,
with
r
ec
e
n
t
ap
p
r
o
ac
h
es
f
r
o
m
th
e
ex
is
tin
g
liter
atu
r
e
,
s
u
c
h
as
B
ay
ar
E
f
f
icien
t
[
2
3
]
a
n
d
Sin
g
h
Z
am
il
[
2
1
]
,
[
2
2
]
.
As
in
d
icate
d
in
th
e
r
es
u
lts
s
ec
tio
n
in
T
ab
les
1
a
n
d
2
,
o
u
r
p
r
o
p
o
s
ed
B
ay
ar
R
esn
et
m
o
d
el
d
em
o
n
s
tr
ated
s
tr
o
n
g
p
e
r
f
o
r
m
an
ce
ac
r
o
s
s
two
d
is
tin
ct
d
atasets
(
Me
d
I
E
VAl
an
d
C
ASI
A)
.
W
e
ac
h
iev
ed
h
ig
h
er
ac
cu
r
ac
y
an
d
p
r
ec
is
io
n
o
n
b
o
th
d
atasets
co
m
p
ar
ed
to
th
e
alter
n
ativ
e
p
r
o
p
o
s
als.
T
h
u
s
,
we
ca
n
ass
er
t
th
at
o
u
r
ap
p
r
o
ac
h
is
well
-
s
u
ited
f
o
r
d
etec
tin
g
f
o
r
g
ed
im
ag
es
an
d
ex
h
ib
its
g
r
ea
ter
p
o
ten
tial
f
o
r
g
en
er
aliza
tio
n
.
An
aly
zin
g
th
e
lo
s
s
cu
r
v
e
in
Fig
u
r
e
5
,
w
e
n
o
ted
a
co
n
s
is
ten
t
d
ec
r
ea
s
e
in
lo
s
s
with
th
e
B
ay
ar
R
esn
et
m
o
d
el,
in
d
icati
n
g
its
r
o
b
u
s
t
lear
n
in
g
ca
p
ac
ity
wh
en
co
m
p
ar
e
d
to
B
ay
a
r
E
f
f
icien
t
[
2
3
]
a
n
d
Sin
g
h
Z
am
il
[
2
1
]
,
[
2
2
]
.
Mo
r
e
o
v
er
,
th
e
l
o
s
s
cu
r
v
e
o
f
B
ay
ar
R
esn
et
d
is
p
lay
ed
s
tab
ilit
y
with
f
ewe
r
f
lu
ctu
atio
n
s
,
s
u
g
g
esti
n
g
an
o
p
tim
al
lear
n
i
n
g
r
ate.
C
o
n
v
er
s
ely
,
th
e
lo
s
s
cu
r
v
e
o
f
t
h
e
B
ay
ar
m
o
d
el,
lack
in
g
th
e
co
n
s
tr
ain
ed
lay
er
f
o
u
n
d
in
B
ay
ar
R
esn
et,
e
x
h
ib
ited
h
ig
h
er
f
lu
ctu
atio
n
s
in
Fig
u
r
e
5
.
T
h
is
s
u
g
g
ests
th
at
th
e
ad
d
itio
n
o
f
th
e
co
n
s
tr
ain
ed
lay
er
in
o
u
r
p
r
o
p
o
s
al
en
h
an
ce
s
its
ef
f
icien
cy
in
i
d
en
tify
in
g
f
alsi
f
icatio
n
ar
tifa
ct
s
.
Up
o
n
ex
am
in
in
g
th
e
r
ec
eiv
er
o
p
er
atin
g
ch
ar
ac
ter
is
tic
(
R
O
C
)
cu
r
v
e
o
n
th
e
Me
d
iEV
AL
d
ataset
in
Fig
u
r
e
6
an
d
th
e
C
ASI
A
d
ataset
in
Fig
u
r
e
7
,
it
b
ec
am
e
ap
p
ar
en
t
th
at
th
e
ar
ea
u
n
d
er
t
h
e
R
OC
cu
r
v
e
o
f
B
ay
ar
R
esn
e
t
s
u
r
p
ass
ed
th
at
o
f
B
ay
ar
,
B
ay
ar
E
f
f
icien
t
[
2
3
]
a
n
d
Sin
g
h
Z
am
il
[
2
1
]
,
[
2
2
]
.
T
h
is
im
p
lies
th
at
B
ay
ar
R
esn
et
y
ield
s
f
ewe
r
f
alse
p
o
s
itiv
es
co
m
p
ar
ed
to
th
e
o
th
er
m
o
d
els.
T
h
is
im
p
r
o
v
em
en
t
is
p
r
ed
o
m
in
a
n
tly
attr
ib
u
te
d
to
th
e
in
clu
s
io
n
o
f
th
e
co
n
s
tr
ain
in
g
lay
e
r
an
d
atten
tio
n
m
ec
h
an
is
m
f
o
r
d
etec
tin
g
im
ag
e
f
alsi
f
icatio
n
f
ea
tu
r
es.
4.
CO
NCLU
SI
O
N
T
o
tack
le
th
e
ch
allen
g
e
o
f
d
e
s
ig
n
in
g
an
d
t
r
ain
in
g
a
d
ee
p
n
eu
r
al
n
etwo
r
k
ca
p
ab
le
o
f
au
to
n
o
m
o
u
s
ly
lear
n
in
g
f
ea
tu
r
es
f
r
o
m
v
ar
io
u
s
m
an
ip
u
latio
n
s
wh
ile
m
in
im
izin
g
f
alse
alar
m
s
,
th
e
B
ay
ar
R
esn
et
m
o
d
el
was
in
tr
o
d
u
ce
d
.
T
h
is
m
o
d
el
in
te
g
r
ates
b
o
th
n
o
is
e
r
esid
u
al
f
ea
tu
r
es
an
d
g
lo
b
al
im
ag
e
f
ea
tu
r
es
t
o
id
en
tify
f
alsi
f
ied
im
ag
es.
I
ts
r
esid
u
e
ex
tr
ac
tio
n
m
o
d
u
le
in
cl
u
d
es
a
co
n
v
o
lu
ti
o
n
al
co
n
s
tr
ain
ed
lay
er
p
air
ed
with
an
atten
tio
n
m
ec
h
an
is
m
,
en
ab
lin
g
au
to
n
o
m
o
u
s
lear
n
in
g
o
f
f
alsi
f
icatio
n
p
a
tter
n
s
.
Me
an
wh
ile,
th
e
f
ea
tu
r
e
ex
tr
ac
tio
n
m
o
d
u
le
f
o
cu
s
es
o
n
ca
p
tu
r
in
g
s
p
atial
f
ea
tu
r
es
ac
r
o
s
s
th
e
en
tire
im
ag
e,
u
tili
zin
g
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
R
esNet5
0
to
ef
f
ec
tiv
ely
ex
t
r
ac
t
g
lo
b
al
f
ea
tu
r
es.
T
h
e
p
r
o
p
o
s
e
d
m
o
d
el
is
e
v
alu
ated
th
r
o
u
g
h
t
r
ain
in
g
an
d
test
in
g
o
n
two
estab
lis
h
ed
d
atasets
,
d
em
o
n
s
tr
atin
g
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
co
m
p
ar
e
d
to
r
ec
en
t
m
o
d
els
in
th
e
liter
atu
r
e.
Fu
tu
r
e
r
esear
ch
d
ir
ec
tio
n
s
m
a
y
ex
p
l
o
r
e
n
ew
p
o
s
s
ib
ilit
ies
b
y
in
co
r
p
o
r
atin
g
m
u
ltimo
d
al
i
n
f
o
r
m
atio
n
,
s
u
c
h
as
tex
t,
v
id
eo
,
an
d
im
ag
e
r
y
,
i
n
to
f
ak
e
n
ews
v
ec
t
o
r
s
,
an
d
in
v
esti
g
atin
g
th
e
ad
ap
tab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
in
s
u
ch
s
ce
n
ar
io
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
2
,
Ap
r
il
20
25
:
2
3
0
4
-
2
3
1
3
2312
RE
F
E
R
E
NC
E
S
[
1
]
M
.
C
a
n
t
a
r
e
l
l
a
,
N
.
F
r
a
c
c
a
r
o
l
i
,
a
n
d
R
.
V
o
l
p
e
,
“
D
o
e
s fa
k
e
n
e
w
s a
f
f
e
c
t
v
o
t
i
n
g
b
e
h
a
v
i
o
u
r
?
,
”
Re
s
e
a
rc
h
P
o
l
i
c
y
,
v
o
l
.
5
2
,
n
o
.
1
,
J
a
n
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
r
e
s
p
o
l
.
2
0
2
2
.
1
0
4
6
2
8
.
[
2
]
L.
Z
h
e
n
g
,
Y
.
Z
h
a
n
g
,
a
n
d
V
.
L
.
L.
T
h
i
n
g
,
“
A
s
u
r
v
e
y
o
n
i
m
a
g
e
t
a
m
p
e
r
i
n
g
a
n
d
i
t
s
d
e
t
e
c
t
i
o
n
i
n
r
e
a
l
-
w
o
r
l
d
p
h
o
t
o
s
,
”
J
o
u
rn
a
l
o
f
V
i
su
a
l
C
o
m
m
u
n
i
c
a
t
i
o
n
a
n
d
I
m
a
g
e
Re
p
res
e
n
t
a
t
i
o
n
,
v
o
l
.
5
8
,
p
p
.
3
8
0
–
3
9
9
,
J
a
n
.
2
0
1
9
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
j
v
c
i
r
.
2
0
1
8
.
1
2
.
0
2
2
.
[
3
]
A
.
D
i
x
i
t
a
n
d
S
.
B
a
g
,
“
A
f
a
s
t
t
e
c
h
n
i
q
u
e
t
o
d
e
t
e
c
t
c
o
p
y
-
m
o
v
e
i
ma
g
e
f
o
r
g
e
r
y
w
i
t
h
r
e
f
l
e
c
t
i
o
n
a
n
d
n
o
n
-
a
f
f
i
n
e
t
r
a
n
sf
o
r
mat
i
o
n
a
t
t
a
c
k
s,
”
Ex
p
e
rt
S
y
st
e
m
s w
i
t
h
Ap
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
1
8
2
,
N
o
v
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
sw
a
.
2
0
2
1
.
1
1
5
2
8
2
.
[
4
]
Y
.
R
a
o
a
n
d
J.
N
i
,
“
A
d
e
e
p
l
e
a
r
n
i
n
g
a
p
p
r
o
a
c
h
t
o
d
e
t
e
c
t
i
o
n
o
f
s
p
l
i
c
i
n
g
a
n
d
c
o
p
y
-
m
o
v
e
f
o
r
g
e
r
i
e
s
i
n
i
ma
g
e
s,”
i
n
2
0
1
6
I
EE
E
I
n
t
e
r
n
a
t
i
o
n
a
l
W
o
rks
h
o
p
o
n
I
n
f
o
rm
a
t
i
o
n
Fo
r
e
n
s
i
c
s
a
n
d
S
e
c
u
r
i
t
y
(
WIFS
)
,
I
EEE,
D
e
c
.
2
0
1
6
,
p
p
.
1
–
6
,
d
o
i
:
1
0
.
1
1
0
9
/
W
I
F
S
.
2
0
1
6
.
7
8
2
3
9
1
1
.
[
5
]
D
.
K
.
S
h
a
r
ma,
B
.
S
i
n
g
h
,
S
.
A
g
a
r
w
a
l
,
L.
G
a
r
g
,
C
.
K
i
m,
a
n
d
K
.
-
H
.
J
u
n
g
,
“
A
su
r
v
e
y
o
f
d
e
t
e
c
t
i
o
n
a
n
d
mi
t
i
g
a
t
i
o
n
f
o
r
f
a
k
e
i
m
a
g
e
s
o
n
so
c
i
a
l
me
d
i
a
p
l
a
t
f
o
r
ms,”
A
p
p
l
i
e
d
S
c
i
e
n
c
e
s
,
v
o
l
.
1
3
,
n
o
.
1
9
,
O
c
t
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
a
p
p
1
3
1
9
1
0
9
8
0
.
[
6
]
F
.
W
.
R
.
To
k
p
a
,
B
.
H
.
K
a
m
a
g
a
t
é
,
V
.
M
o
n
sa
n
,
a
n
d
S
.
O
u
m
t
a
n
a
g
a
,
“
F
a
k
e
n
e
w
s
d
e
t
e
c
t
i
o
n
i
n
so
c
i
a
l
med
i
a
:
H
y
b
r
i
d
d
e
e
p
l
e
a
r
n
i
n
g
a
p
p
r
o
a
c
h
e
s
,
”
J
o
u
rn
a
l
o
f
Ad
v
a
n
c
e
s
i
n
I
n
f
o
r
m
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
v
o
l
.
1
4
,
n
o
.
3
,
p
p
.
6
0
6
–
6
1
5
,
2
0
2
3
,
d
o
i
:
1
0
.
1
2
7
2
0
/
j
a
i
t
.
1
4
.
3
.
6
0
6
-
6
1
5
.
[
7
]
O
.
A
j
a
o
,
D
.
B
h
o
w
mi
k
,
a
n
d
S
.
Za
r
g
a
r
i
,
“
F
a
k
e
n
e
w
s
i
d
e
n
t
i
f
i
c
a
t
i
o
n
o
n
Tw
i
t
t
e
r
w
i
t
h
h
y
b
r
i
d
C
N
N
a
n
d
R
N
N
mo
d
e
l
s,
”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
9
t
h
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
S
o
c
i
a
l
M
e
d
i
a
a
n
d
S
o
c
i
e
t
y
,
i
n
S
M
S
o
c
i
e
t
y
’
1
8
.
A
C
M
,
Ju
l
.
2
0
1
8
,
d
o
i
:
1
0
.
1
1
4
5
/
3
2
1
7
8
0
4
.
3
2
1
7
9
1
7
.
[
8
]
K
.
P
o
p
a
t
,
S
.
M
u
k
h
e
r
j
e
e
,
A
.
Y
a
t
e
s,
a
n
d
G
.
W
e
i
k
u
m,
“
D
e
C
l
a
r
E:
D
e
b
u
n
k
i
n
g
f
a
k
e
n
e
w
s
a
n
d
f
a
l
se
c
l
a
i
ms
u
s
i
n
g
e
v
i
d
e
n
c
e
-
a
w
a
r
e
d
e
e
p
l
e
a
r
n
i
n
g
,
”
i
n
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
2
0
1
8
C
o
n
f
e
re
n
c
e
o
n
Em
p
i
r
i
c
a
l
Me
t
h
o
d
s
i
n
N
a
t
u
r
a
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
i
n
g
,
A
sso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
s
t
i
c
s,
2
0
1
8
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
d
1
8
-
1
0
0
3
.
[
9
]
N
.
R
a
n
i
,
P
.
D
a
s,
a
n
d
A
.
K
.
B
h
a
r
d
w
a
j
,
“
A
h
y
b
r
i
d
d
e
e
p
l
e
a
r
n
i
n
g
m
o
d
e
l
b
a
se
d
o
n
C
N
N
-
B
i
LST
M
f
o
r
r
u
m
o
r
d
e
t
e
c
t
i
o
n
,
”
i
n
2
0
2
1
6
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
C
o
m
m
u
n
i
c
a
t
i
o
n
a
n
d
El
e
c
t
ro
n
i
c
s
S
y
st
e
m
s
(
I
C
C
ES
)
,
I
EEE,
J
u
l
.
2
0
2
1
,
p
p
.
1
4
2
3
–
1
4
2
7
,
d
o
i
:
1
0
.
1
1
0
9
/
i
c
c
e
s5
1
3
5
0
.
2
0
2
1
.
9
4
8
9
2
1
4
.
[
1
0
]
J.
X
u
e
,
Y
.
W
a
n
g
,
Y
.
Ti
a
n
,
Y
.
Li
,
L.
S
h
i
,
a
n
d
L.
W
e
i
,
“
D
e
t
e
c
t
i
n
g
f
a
k
e
n
e
w
s
b
y
e
x
p
l
o
r
i
n
g
t
h
e
c
o
n
s
i
st
e
n
c
y
o
f
mu
l
t
i
m
o
d
a
l
d
a
t
a
,
”
I
n
f
o
rm
a
t
i
o
n
Pr
o
c
e
ssi
n
g
a
n
d
Ma
n
a
g
e
m
e
n
t
,
v
o
l
.
5
8
,
S
e
p
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
p
m.
2
0
2
1
.
1
0
2
6
1
0
.
[
1
1
]
Y
.
Li
a
n
d
Y
.
X
i
e
,
“
I
s
a
p
i
c
t
u
r
e
w
o
r
t
h
a
t
h
o
u
sa
n
d
w
o
r
d
s?
A
n
e
mp
i
r
i
c
a
l
st
u
d
y
o
f
i
m
a
g
e
c
o
n
t
e
n
t
a
n
d
s
o
c
i
a
l
me
d
i
a
e
n
g
a
g
e
m
e
n
t
,
”
J
o
u
rn
a
l
o
f
M
a
rke
t
i
n
g
R
e
se
a
rc
h
,
v
o
l
.
5
7
,
n
o
.
1
,
p
p
.
1
–
1
9
,
N
o
v
.
2
0
1
9
,
d
o
i
:
1
0
.
1
1
7
7
/
0
0
2
2
2
4
3
7
1
9
8
8
1
1
1
3
.
[
1
2
]
Z.
J
i
n
,
J.
C
a
o
,
Y
.
Z
h
a
n
g
,
J.
Z
h
o
u
,
a
n
d
Q
.
Ti
a
n
,
“
N
o
v
e
l
v
i
s
u
a
l
a
n
d
s
t
a
t
i
s
t
i
c
a
l
i
m
a
g
e
f
e
a
t
u
r
e
s
f
o
r
m
i
c
r
o
b
l
o
g
s
n
e
w
s
v
e
r
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
M
u
l
t
i
m
e
d
i
a
,
v
o
l
.
1
9
,
n
o
.
3
,
p
p
.
5
9
8
–
6
0
8
,
M
a
r
.
2
0
1
7
,
d
o
i
:
1
0
.
1
1
0
9
/
t
mm
.
2
0
1
6
.
2
6
1
7
0
7
8
.
[
1
3
]
J.
C
a
o
,
P
.
Q
i
,
Q
.
S
h
e
n
g
,
T.
Y
a
n
g
,
J.
G
u
o
,
a
n
d
J.
L
i
,
“
Ex
p
l
o
r
i
n
g
t
h
e
r
o
l
e
o
f
v
i
s
u
a
l
c
o
n
t
e
n
t
i
n
f
a
k
e
n
e
w
s
d
e
t
e
c
t
i
o
n
,
”
i
n
D
i
si
n
f
o
rm
a
t
i
o
n
,
Mi
s
i
n
f
o
rm
a
t
i
o
n
,
a
n
d
F
a
k
e
N
e
w
s
i
n
S
o
c
i
a
l
M
e
d
i
a
,
S
p
r
i
n
g
e
r
I
n
t
e
r
n
a
t
i
o
n
a
l
P
u
b
l
i
s
h
i
n
g
,
2
0
2
0
,
p
p
.
1
4
1
–
1
6
1
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
030
-
4
2
6
9
9
-
6
_
8
.
[
1
4
]
A
.
B
e
r
t
h
e
t
,
“
D
e
e
p
l
e
a
r
n
i
n
g
me
t
h
o
d
s
a
n
d
a
d
v
a
n
c
e
m
e
n
t
s
i
n
d
i
g
i
t
a
l
i
ma
g
e
f
o
r
e
n
si
c
s,”
P
h
.
D
.
d
i
ss
e
r
t
a
t
i
o
n
,
D
e
p
a
r
t
men
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
Te
l
e
c
o
mm
u
n
i
c
a
t
i
o
n
s
a
n
d
El
e
c
t
r
o
n
i
c
s,
S
o
r
b
o
n
n
e
U
n
i
v
e
r
s
i
t
y
,
2
0
2
2
.
[
1
5
]
N
.
K
r
a
w
e
t
z
,
“
A
p
i
c
t
u
r
e
’
s
w
o
r
t
h
,
”
H
a
c
k
e
r
F
a
c
t
o
r
S
o
l
u
t
i
o
n
s
,
2
0
0
7
.
Acc
ess
ed
:
M
a
y
1
5
,
2
0
2
4
.
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s
:
/
/
w
w
w
.
b
l
a
c
k
h
a
t
.
c
o
m
/
p
r
e
se
n
t
a
t
i
o
n
s/
b
h
-
u
sa
-
0
7
/
K
r
a
w
e
t
z
/
W
h
i
t
e
p
a
p
e
r
/
b
h
-
u
sa
-
07
-
k
r
a
w
e
t
z
-
W
P
.
p
d
f
[
1
6
]
N
.
V
.
S
.
K
.
V
i
j
a
y
a
l
a
k
sh
mi
K
,
J.
S
a
s
i
k
a
l
a
,
a
n
d
C
.
S
h
a
n
mu
g
a
n
a
t
h
a
n
,
“
C
o
p
y
-
p
a
st
e
f
o
r
g
e
r
y
d
e
t
e
c
t
i
o
n
u
s
i
n
g
d
e
e
p
l
e
a
r
n
i
n
g
w
i
t
h
e
r
r
o
r
l
e
v
e
l
a
n
a
l
y
s
i
s
,
”
Mu
l
t
i
m
e
d
i
a
T
o
o
l
s
a
n
d
Ap
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
8
3
,
n
o
.
2
,
p
p
.
3
4
2
5
–
3
4
4
9
,
M
a
y
2
0
2
3
,
d
o
i
:
1
0
.
1
0
0
7
/
s1
1
0
4
2
-
023
-
1
5
5
9
4
-
5.
[
1
7
]
K
.
H
.
R
h
e
e
,
“
D
e
t
e
c
t
i
o
n
o
f
s
p
l
i
c
e
d
i
ma
g
e
f
o
r
e
n
si
c
s
u
si
n
g
t
e
x
t
u
r
e
a
n
a
l
y
si
s
o
f
me
d
i
a
n
f
i
l
t
e
r
r
e
s
i
d
u
a
l
,
”
I
EEE
Ac
c
e
ss
,
v
o
l
.
8
,
p
p
.
1
0
3
3
7
4
–
1
0
3
3
8
4
,
2
0
2
0
,
d
o
i
:
1
0
.
1
1
0
9
/
a
c
c
e
s
s.
2
0
2
0
.
2
9
9
9
3
0
8
.
[
1
8
]
T.
S
.
G
u
n
a
w
a
n
,
S
.
A
.
M
.
H
a
n
a
f
i
a
h
,
M
.
K
a
r
t
i
w
i
,
N
.
I
smai
l
,
N
.
F
.
Za
’
b
a
h
,
a
n
d
A
.
N
.
N
o
r
d
i
n
,
“
D
e
v
e
l
o
p
me
n
t
o
f
p
h
o
t
o
f
o
r
e
n
s
i
c
s
a
l
g
o
r
i
t
h
m
b
y
d
e
t
e
c
t
i
n
g
p
h
o
t
o
s
h
o
p
m
a
n
i
p
u
l
a
t
i
o
n
u
s
i
n
g
e
r
r
o
r
l
e
v
e
l
a
n
a
l
y
si
s,”
I
n
d
o
n
e
s
i
a
n
J
o
u
r
n
a
l
o
f
E
l
e
c
t
r
i
c
a
l
En
g
i
n
e
e
ri
n
g
a
n
d
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
v
o
l
.
7
,
n
o
.
1
,
J
u
l
.
2
0
1
7
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
e
e
c
s.
v
7
.
i
1
.
p
p
1
3
1
-
1
3
7
.
[
1
9
]
X
.
L
i
n
e
t
a
l
.
,
“
I
mag
e
m
a
n
i
p
u
l
a
t
i
o
n
d
e
t
e
c
t
i
o
n
b
y
m
u
l
t
i
p
l
e
t
a
mp
e
r
i
n
g
t
r
a
c
e
s
a
n
d
e
d
g
e
a
r
t
i
f
a
c
t
e
n
h
a
n
c
e
m
e
n
t
,
”
P
a
t
t
e
r
n
Re
c
o
g
n
i
t
i
o
n
,
v
o
l
.
1
3
3
,
J
a
n
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
p
a
t
c
o
g
.
2
0
2
2
.
1
0
9
0
2
6
.
[
2
0
]
B
.
S
i
n
g
h
a
n
d
D
.
K
.
S
h
a
r
ma
,
“
P
r
e
d
i
c
t
i
n
g
i
ma
g
e
c
r
e
d
i
b
i
l
i
t
y
i
n
f
a
k
e
n
e
w
s
o
v
e
r
so
c
i
a
l
m
e
d
i
a
u
s
i
n
g
mu
l
t
i
-
mo
d
a
l
a
p
p
r
o
a
c
h
,
”
N
e
u
r
a
l
C
o
m
p
u
t
i
n
g
a
n
d
Ap
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
3
4
,
n
o
.
2
4
,
p
p
.
2
1
5
0
3
–
2
1
5
1
7
,
M
a
y
2
0
2
1
,
d
o
i
:
1
0
.
1
0
0
7
/
s
0
0
5
2
1
-
021
-
0
6
0
8
6
-
4.
[
2
1
]
Y
.
K
.
Za
mi
l
a
n
d
N
.
M
.
C
h
a
r
k
a
r
i
,
“
C
o
mb
a
t
i
n
g
f
a
k
e
n
e
w
s
o
n
so
c
i
a
l
m
e
d
i
a
:
A
f
u
s
i
o
n
a
p
p
r
o
a
c
h
f
o
r
i
mp
r
o
v
e
d
d
e
t
e
c
t
i
o
n
a
n
d
i
n
t
e
r
p
r
e
t
a
b
i
l
i
t
y
,
”
I
EE
E
A
c
c
e
ss
,
v
o
l
.
1
2
,
p
p
.
2
0
7
4
–
2
0
8
5
,
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
a
c
c
e
ss
.
2
0
2
3
.
3
3
4
2
8
4
3
.
[
2
2
]
C
.
B
o
i
d
i
d
o
u
,
S
.
P
a
p
a
d
o
p
o
u
l
o
s,
M
.
Za
m
p
o
g
l
o
u
,
L
.
A
p
o
st
o
l
i
d
i
s,
O
.
P
a
p
a
d
o
p
o
u
l
o
u
,
a
n
d
Y
.
K
o
m
p
a
t
si
a
r
i
s
,
“
D
e
t
e
c
t
i
o
n
a
n
d
v
i
s
u
a
l
i
z
a
t
i
o
n
o
f
m
i
sl
e
a
d
i
n
g
c
o
n
t
e
n
t
o
n
Tw
i
t
t
e
r
,
”
I
n
t
e
rn
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
Mu
l
t
i
m
e
d
i
a
I
n
f
o
rm
a
t
i
o
n
R
e
t
r
i
e
v
a
l
,
v
o
l
.
7
,
n
o
.
1
,
p
p
.
7
1
–
8
6
,
D
e
c
.
2
0
1
7
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
3
7
3
5
-
0
1
7
-
0
1
4
3
-
x.
[
2
3
]
B
.
B
a
y
a
r
a
n
d
M
.
C
.
S
t
a
m
m,
“
C
o
n
s
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:
A
n
e
w
a
p
p
r
o
a
c
h
t
o
w
a
r
d
s
g
e
n
e
r
a
l
p
u
r
p
o
se
i
ma
g
e
man
i
p
u
l
a
t
i
o
n
d
e
t
e
c
t
i
o
n
,
”
I
EE
E
T
ra
n
s
a
c
t
i
o
n
s
o
n
I
n
f
o
rm
a
t
i
o
n
F
o
re
n
si
c
s
a
n
d
S
e
c
u
r
i
t
y
,
v
o
l
.
1
3
,
n
o
.
1
1
,
p
p
.
2
6
9
1
–
2
7
0
6
,
N
o
v
.
2
0
1
8
,
d
o
i
:
1
0
.
1
1
0
9
/
t
i
f
s
.
2
0
1
8
.
2
8
2
5
9
5
3
.
[
2
4
]
S
.
W
o
o
,
J.
P
a
r
k
,
J.
-
Y
.
Le
e
,
a
n
d
I
.
S
.
K
w
e
o
n
,
“
C
B
A
M
:
C
o
n
v
o
l
u
t
i
o
n
a
l
b
l
o
c
k
a
t
t
e
n
t
i
o
n
mo
d
u
l
e
,
”
i
n
L
e
c
t
u
re
N
o
t
e
s
i
n
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
(
i
n
c
l
u
d
i
n
g
s
u
b
ser
i
e
s
L
e
c
t
u
r
e
N
o
t
e
s
i
n
Ar
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
L
e
c
t
u
re
N
o
t
e
s
i
n
B
i
o
i
n
f
o
rm
a
t
i
c
s)
,
v
o
l
.
1
1
2
1
1
,
S
p
r
i
n
g
e
r
I
n
t
e
r
n
a
t
i
o
n
a
l
P
u
b
l
i
s
h
i
n
g
,
2
0
1
8
,
p
p
.
3
–
1
9
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
030
-
0
1
2
3
4
-
2
_
1
.
[
2
5
]
Q
.
X
u
e
t
a
l
.
,
“
En
h
a
n
c
i
n
g
a
d
a
p
t
i
v
e
h
i
s
t
o
r
y
r
e
ser
v
i
n
g
b
y
s
p
i
k
i
n
g
c
o
n
v
o
l
u
t
i
o
n
a
l
b
l
o
c
k
a
t
t
e
n
t
i
o
n
mo
d
u
l
e
i
n
r
e
c
u
r
r
e
n
t
n
e
u
r
a
l
n
e
t
w
o
r
k
s,”
i
n
T
h
e
C
o
n
f
e
r
e
n
c
e
o
n
N
e
u
ra
l
I
n
f
o
rm
a
t
i
o
n
Pro
c
e
ss
i
n
g
S
y
s
t
e
m
s
,
2
0
2
3
.
[
2
6
]
M
.
D
.
Z
e
i
l
e
r
a
n
d
R
.
F
e
r
g
u
s
,
“
V
i
su
a
l
i
z
i
n
g
a
n
d
U
n
d
e
r
s
t
a
n
d
i
n
g
C
o
n
v
o
l
u
t
i
o
n
a
l
N
e
t
w
o
r
k
s
,
”
i
n
C
o
m
p
u
t
e
r
V
i
si
o
n
–
E
C
C
V
2
0
1
4
,
S
p
r
i
n
g
e
r
I
n
t
e
r
n
a
t
i
o
n
a
l
P
u
b
l
i
sh
i
n
g
,
2
0
1
4
,
p
p
.
8
1
8
–
8
3
3
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
3
1
9
-
1
0
5
9
0
-
1
_
5
3
.
[
2
7
]
K
.
H
e
,
X
.
Zh
a
n
g
,
S
.
R
e
n
,
a
n
d
J.
S
u
n
,
“
D
e
e
p
r
e
s
i
d
u
a
l
l
e
a
r
n
i
n
g
f
o
r
i
ma
g
e
r
e
c
o
g
n
i
t
i
o
n
,
”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
I
EEE
C
o
m
p
u
t
e
r
S
o
c
i
e
t
y
C
o
n
f
e
re
n
c
e
o
n
C
o
m
p
u
t
e
r V
i
si
o
n
a
n
d
P
a
t
t
e
r
n
Re
c
o
g
n
i
t
i
o
n
,
2
0
1
6
,
p
p
.
7
7
0
–
7
7
8
,
d
o
i
:
1
0
.
1
1
0
9
/
C
V
P
R
.
2
0
1
6
.
9
0
.
[
2
8
]
Y
.
M
a
,
S
.
C
h
e
n
,
S
.
Er
m
o
n
,
a
n
d
D
.
B
.
Lo
b
e
l
l
,
“
Tr
a
n
sf
e
r
l
e
a
r
n
i
n
g
i
n
e
n
v
i
r
o
n
m
e
n
t
a
l
r
e
mo
t
e
se
n
si
n
g
,
”
R
e
m
o
t
e
S
e
n
s
i
n
g
o
f
En
v
i
r
o
n
m
e
n
t
,
v
o
l
.
3
0
1
,
F
e
b
.
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
r
s
e
.
2
0
2
3
.
1
1
3
9
2
4
.
[
2
9
]
J.
G
u
,
Y
.
X
u
,
J.
S
u
n
,
a
n
d
W
.
L
i
u
,
“
I
mag
e
t
a
mp
e
r
i
n
g
d
e
t
e
c
t
i
o
n
b
a
s
e
d
o
n
f
e
a
t
u
r
e
c
o
n
s
i
st
e
n
c
y
a
t
t
e
n
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
f
o
rm
a
t
i
o
n
a
n
d
C
o
m
p
u
t
e
r
S
e
c
u
ri
t
y
,
v
o
l
.
2
3
,
n
o
.
1
,
p
p
.
1
–
1
5
,
2
0
2
4
,
d
o
i
:
1
0
.
1
5
0
4
/
i
j
i
c
s
.
2
0
2
4
.
1
3
6
7
0
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
co
n
s
tr
a
in
ed
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
k
w
ith
a
tten
tio
n
mec
h
a
n
is
m
fo
r
(
K
a
ma
g
a
te
B
e
ma
n
Ha
mid
ja
)
2313
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
K
a
m
a
g
a
te
Be
m
a
n
H
a
m
id
ja
r
e
c
e
iv
e
d
h
is
m
a
ste
r
d
e
g
re
e
in
c
o
m
p
u
ter
sc
ien
c
e
2
0
1
4
at
Un
it
é
d
e
F
o
rm
a
ti
o
n
e
t
d
e
Re
c
h
e
rc
h
e
e
n
S
c
ien
c
e
s
F
o
n
d
e
m
e
n
tal
e
s
e
t
Ap
p
li
q
u
é
s
(UFR
-
S
F
A)
,
Un
iv
e
rsité
Na
n
g
u
i
Ab
r
o
g
o
u
a
,
I
v
o
ry
Co
a
st.
He
re
c
e
iv
e
d
P
h
D
d
e
g
re
e
in
th
e
sa
m
e
field
a
t
In
stit
u
t
Na
ti
o
n
a
l
P
o
l
y
tec
h
n
i
q
u
e
Ho
u
p
h
o
u
ë
t
-
Bo
ig
n
y
d
e
Ya
m
o
u
ss
o
u
k
ro
(IN
P
HB)
in
2
0
1
8
.
F
ro
m
th
a
t
d
a
te
to
n
o
w
h
e
a
ss
istan
t
P
ro
fe
ss
o
r
a
t
Eco
le
S
u
p
e
rieu
re
Afric
a
in
d
e
s
TIC
(ES
ATIC)
Ab
id
jan
,
Treic
h
v
il
le.
He
field
s
o
f
re
se
a
rc
h
in
c
l
u
d
e
n
e
two
rk
o
p
ti
m
iza
ti
o
n
,
I
o
T,
c
y
b
e
rse
c
u
rit
y
a
n
d
a
p
p
li
e
d
a
rti
ficia
l
in
telli
g
e
n
t
.
He
is
a
m
e
m
b
e
r
o
f
tea
m
In
fo
rm
a
ti
q
u
e
e
t
S
e
c
u
rit
é
d
e
s
s
y
ste
m
e
s
Nu
m
e
rq
ie
(I2
S
N)
o
f
La
b
o
ra
t
o
ire
d
e
s
S
c
ien
c
e
s
e
t
Tec
h
n
o
l
o
g
ies
d
e
l’In
f
o
rm
a
ti
o
n
e
t
d
e
la
C
o
m
m
u
n
ica
ti
o
n
(LAS
TIC)
o
f
ES
ATIC.
H
e
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
b
e
m
a
n
.
k
a
m
a
g
a
te@
e
sa
ti
c
.
e
d
u
.
c
i
.
Fa
to
u
m
a
ta
Wo
n
g
b
é
Ro
sa
li
e
T
o
k
p
a
re
c
e
iv
e
d
h
is
m
a
ste
r
d
e
g
r
e
e
in
d
a
tab
a
se
a
n
d
so
ftwa
re
e
n
g
in
e
e
rin
g
a
t
Un
it
é
d
e
F
o
rm
a
ti
o
n
e
t
d
e
Re
c
h
e
rc
h
e
M
a
th
é
m
a
ti
q
u
e
s
e
t
In
fo
rm
a
ti
q
u
e
(UFR
-
M
I)
a
t
Un
iv
e
rsité F
é
li
x
H
o
u
p
h
o
u
ë
t
-
B
o
i
g
n
y
(U
F
HB) Cô
te d
’
Iv
o
ire,
i
n
2
0
1
9
.
Cu
rre
n
tl
y
,
s
h
e
is
a
P
h
.
D.
st
u
d
e
n
t
i
n
c
o
m
p
u
ter
s
c
ien
c
e
a
t
(UFHB).
S
h
e
is
a
m
e
m
b
e
r
o
f
th
e
M
a
th
e
m
a
ti
c
s
a
n
d
Co
m
p
u
ter
S
c
ien
c
e
Lab
o
ra
to
r
y
(
LAM
I)
a
t
th
is
Un
i
v
e
rsity
,
a
lso
a
m
e
m
b
e
r
o
f
th
e
Artifi
c
ial
In
telli
g
e
n
c
e
a
n
d
Da
tab
a
se
M
o
d
e
ll
in
g
tea
m
o
f
Lab
o
ra
to
i
re
d
e
Re
c
h
e
rc
h
e
e
n
In
fo
rm
a
ti
q
u
e
e
t
Télé
c
o
m
u
n
ica
ti
o
n
(LARIT
)
o
f
In
stit
u
t
Na
ti
o
n
a
l
P
o
ly
tec
h
n
iq
u
e
F
é
li
x
Ho
u
p
h
o
u
ë
t
-
Bo
i
g
n
y
(INP
-
HB),
Ya
m
o
u
ss
o
u
k
r
o
.
His
r
e
se
a
rc
h
fo
c
u
se
s
o
n
m
a
th
e
m
a
ti
c
a
l
m
o
d
e
ll
in
g
m
a
c
h
i
n
e
lea
rn
i
n
g
a
n
d
d
a
ta
v
e
ra
c
it
y
.
S
h
e
c
a
n
b
e
c
o
n
t
a
c
ted
a
t
e
m
a
il
:
to
k
p
a
fa
to
u
@
g
m
a
il
.
c
o
m
.
Vin
c
e
n
t
Mo
n
sa
n
o
b
tain
e
d
h
is
P
h
.
D
.
i
n
c
o
m
m
o
n
sc
ien
c
e
s
a
n
d
t
e
c
h
n
iq
u
e
s
fro
m
th
e
Un
iv
e
rsity
o
f
Ro
u
e
n
in
F
ra
n
c
e
i
n
1
9
9
4
.
D
u
rin
g
h
is
d
o
c
to
ra
te,
h
e
wo
rk
e
d
o
n
sp
e
c
tral
e
stim
a
ti
o
n
in
th
e
p
e
rio
d
ica
ll
y
c
o
rre
late
d
p
ro
c
e
ss
e
s.
He
is
c
u
rre
n
tl
y
a
lec
tu
re
r
in
sta
ti
stic
a
t
Un
it
é
d
e
F
o
rm
a
ti
o
n
e
t
d
e
Re
c
h
e
rc
h
e
M
a
th
é
m
a
ti
q
u
e
s
e
t
In
f
o
rm
a
ti
q
u
e
(U
F
R
-
M
I)
a
t
Un
i
v
e
rsité
F
é
l
ix
Ho
u
p
h
o
u
ë
t
-
Bo
i
g
n
y
(U
F
HB),
C
ô
t
e
d
’Iv
o
ire.
His
re
se
a
rc
h
i
n
tere
sts
in
c
l
u
d
e
F
o
u
rier
c
o
e
fficie
n
t,
P
e
rio
d
ic
c
o
rre
latio
n
,
n
o
n
p
a
ra
m
e
tri
c
e
stim
a
ti
o
n
,
a
n
d
M
a
t
h
e
m
a
ti
c
a
l
sta
ti
stics
.
He
c
u
rre
n
tl
y
h
o
ld
s
th
e
p
o
siti
o
n
o
f
v
ice
-
p
re
si
d
e
n
t
a
t
Un
iv
e
rsité
F
e
li
x
H
o
u
p
h
o
u
ë
t
-
B
o
ig
n
y
,
Ab
i
d
jan
,
Iv
o
ry
C
o
a
st.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
v
m
o
n
sa
n
@y
a
h
o
o
.
fr
.
S
o
u
ley
m
a
n
e
O
u
m
t
a
n
a
g
a
is
a
c
u
rre
n
t
d
irec
to
r
o
f
Un
i
té
d
M
i
x
te
d
e
Re
c
h
e
rc
h
e
e
n
M
a
th
é
m
a
ti
q
u
e
s
e
t
S
c
ien
c
e
s
d
u
Nu
m
e
riq
u
e
s
(
UMRI
M
S
N)
a
t
I
n
ti
tu
t
Na
ti
o
n
a
l
P
o
l
y
tec
h
n
iq
u
e
Ho
u
p
h
o
u
ë
t
-
Bo
i
g
n
y
,
Ya
m
o
u
ss
o
u
k
ro
(
INPHB)
,
Iv
o
ry
Co
st.
He
is als
o
a
m
e
m
b
e
r
o
f
th
e
WACREN
b
o
a
rd
a
n
d
m
e
m
b
e
r
o
f
S
c
ien
ti
fi
Co
u
n
c
if
o
f
F
ON
TS
I
(
F
u
n
d
f
o
r
S
c
ien
c
e
,
Tec
h
o
n
o
l
o
g
y
a
n
d
In
n
o
v
a
ti
o
n
).
He
p
lay
e
d
a
k
e
y
ro
le
in
th
e
e
sta
b
li
sh
m
e
n
t
o
f
In
ter
n
e
t
d
o
m
a
in
o
f
I
v
o
r
y
C
o
st.
He
re
c
e
iv
e
d
h
is
P
h
.
D.
in
c
o
m
p
u
ter
s
c
ien
c
e
a
t
Un
iv
e
rsité
P
a
u
l
S
a
b
a
ti
e
r
(To
u
lo
u
se
,
F
ra
n
c
e
)
i
n
1
9
9
5
.
He
is
fu
ll
p
ro
fe
ss
o
r
in
c
o
m
p
u
ter
sc
ien
c
e
a
t
INP
-
HB
sin
c
e
2
0
0
7
.
His
re
se
a
rc
h
in
tere
sts
in
c
lu
d
e
n
e
two
rk
,
Io
T,
c
y
b
e
rse
c
u
rit
y
a
n
d
a
p
p
li
e
d
a
rti
ficia
l
in
telli
g
e
n
t.
H
e
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
o
u
m
tan
a
@g
m
a
il
.
c
o
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.