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
.
3
,
J
u
n
e
20
25
,
p
p
.
2951
~
2
9
6
4
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
15
i
3
.
pp
2
9
5
1
-
2
9
6
4
2951
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Enha
ncing
traini
ng
perf
o
rma
nce
f
o
r sma
ll mo
dels u
sing
da
ta
-
centri
c appr
o
a
ches
Reda
A.
E
l
-
K
ho
ribi
1
,
E
i
d E
m
a
ry
1
,
2
,
Am
r
E
s
s
a
m
H
a
s
s
a
n
1
1
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
Te
c
h
n
o
l
o
g
y
,
F
a
c
u
l
t
y
o
f
C
o
m
p
u
t
e
r
s a
n
d
A
r
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
U
n
i
v
e
r
s
i
t
y
o
f
C
a
i
r
o
,
C
a
i
r
o
,
E
g
y
p
t
2
F
a
c
u
l
t
y
o
f
C
o
mp
u
t
i
n
g
S
t
u
d
i
e
s
,
A
r
a
b
O
p
e
n
U
n
i
v
e
r
si
t
y
,
C
a
i
r
o
,
E
g
y
p
t
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
l 1
5
,
2
0
2
4
R
ev
is
ed
Dec
1
9
,
2
0
2
4
Acc
ep
ted
J
an
1
6
,
2
0
2
5
In
th
is
wo
r
k
,
we
p
ro
p
o
se
a
n
e
w
sy
ste
m
to
imp
r
o
v
e
th
e
p
e
rfo
rm
a
n
c
e
o
f
c
las
sifica
ti
o
n
m
o
d
e
ls
b
y
a
p
p
l
y
in
g
d
a
ta
-
c
e
n
tri
c
p
ri
n
c
ip
les
.
Th
e
sy
ste
m
o
p
ti
m
ize
s
d
a
tas
e
ts
b
y
re
m
o
v
i
n
g
p
o
o
r
-
q
u
a
li
ty
sa
m
p
les
a
n
d
g
e
n
e
ra
ti
n
g
h
ig
h
-
q
u
a
li
ty
sy
n
th
e
ti
c
d
a
ta.
We
tes
ted
th
e
s
y
ste
m
o
n
v
a
rio
u
s
c
la
ss
ifi
c
a
ti
o
n
m
o
d
e
ls
a
n
d
d
a
tas
e
ts,
m
e
a
su
rin
g
it
s
p
e
rf
o
rm
a
n
c
e
with
a
c
c
u
ra
c
y
,
p
re
c
isio
n
,
re
c
a
ll
,
a
n
d
F
1
-
sc
o
re
.
Th
e
re
su
lt
s
sh
o
we
d
sig
n
ifi
c
a
n
t
imp
r
o
v
e
m
e
n
ts
in
c
las
sifica
ti
o
n
p
e
rfo
rm
a
n
c
e
,
h
ig
h
li
g
h
ti
n
g
t
h
e
e
ffe
c
ti
v
e
n
e
ss
o
f
th
is
d
a
ta
-
c
e
n
tri
c
a
p
p
ro
a
c
h
.
Wh
il
e
t
h
e
sc
a
l
a
b
il
it
y
to
larg
e
-
sc
a
le
d
a
tas
e
ts
is
stil
l
a
n
o
p
e
n
q
u
e
stio
n
,
it
o
ffe
rs
g
re
a
t
p
o
ten
ti
a
l
fo
r
f
u
tu
re
re
se
a
rc
h
.
T
h
is
a
p
p
r
o
a
c
h
c
o
u
l
d
b
e
v
a
lu
a
b
le
i
n
c
rit
ica
l
a
re
a
s
li
k
e
h
e
a
lt
h
c
a
re
,
fin
a
n
c
e
,
a
n
d
a
u
t
o
n
o
m
o
u
s
sy
ste
m
s,
wh
e
re
h
ig
h
-
q
u
a
l
it
y
d
a
ta
is
c
ru
c
ial.
F
u
t
u
re
wo
rk
c
o
u
l
d
e
x
p
lo
re
a
d
v
a
n
c
e
d
d
a
ta
a
u
g
m
e
n
tatio
n
,
a
d
a
p
ti
n
g
th
e
s
y
st
e
m
fo
r
d
iffere
n
t
d
a
ta
t
y
p
e
s
li
k
e
tex
t
a
n
d
ti
m
e
-
se
ries
,
a
n
d
e
x
ten
d
in
g
it
t
o
se
m
i
-
su
p
e
rv
ise
d
a
n
d
u
n
s
u
p
e
rv
ise
d
lea
rn
in
g
.
Ou
r
fin
d
i
n
g
s
e
m
p
h
a
siz
e
th
e
imp
o
rtan
c
e
o
f
d
a
ta
q
u
a
li
t
y
in
a
c
h
iev
in
g
b
e
tt
e
r
m
o
d
e
l
p
e
rfo
rm
a
n
c
e
,
o
ften
o
v
e
rl
o
o
k
e
d
i
n
fa
v
o
r
o
f
m
o
d
e
l
a
rc
h
it
e
c
tu
re
.
By
a
d
v
a
n
c
in
g
d
a
ta
-
c
e
n
tri
c
a
rti
ficia
l
i
n
telli
g
e
n
c
e
(
AI
)
,
th
is
wo
r
k
o
ffe
rs a
p
ra
c
ti
c
a
l
fra
m
e
wo
rk
fo
r
re
se
a
rc
h
e
rs
a
n
d
p
ra
c
ti
ti
o
n
e
rs
to
o
p
ti
m
ize
d
a
t
a
se
ts
a
n
d
imp
ro
v
e
m
a
c
h
i
n
e
lea
rn
in
g
sy
ste
m
s.
K
ey
w
o
r
d
s
:
C
o
m
p
u
ter
v
is
io
n
Data
-
ce
n
tr
ic
Dee
p
lear
n
in
g
Gen
er
ativ
e
ad
v
e
r
s
ar
ial
n
etwo
r
k
Mo
d
el
-
ce
n
tr
ic
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
:
Am
r
E
s
s
am
Hass
an
D
e
p
a
r
t
m
e
n
t
o
f
I
n
f
o
r
m
a
t
i
o
n
T
ec
h
n
o
l
o
g
y
,
F
a
c
u
l
t
y
o
f
C
o
m
p
u
t
e
r
s
a
n
d
A
r
t
i
f
i
c
i
al
I
n
t
e
l
li
g
e
n
c
e
,
Un
i
v
e
r
s
i
t
y
o
f
C
a
i
r
o
C
air
o
,
E
g
y
p
t
E
m
ail:
am
r
.
ess
am
.
h
ass
an
@
o
u
tlo
o
k
.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Dee
p
lear
n
in
g
h
as
r
ev
o
lu
tio
n
i
ze
d
ar
tific
ial
in
tellig
en
ce
(
AI
)
,
d
r
iv
in
g
ad
v
an
ce
m
en
ts
in
d
o
m
ain
s
s
u
ch
as
m
ed
ical
d
iag
n
o
s
is
,
au
to
n
o
m
o
u
s
s
y
s
tem
s
,
an
d
lar
g
e
-
s
ca
le
d
ec
is
io
n
-
m
ak
in
g
[
1
]
.
T
h
ese
d
ev
elo
p
m
en
ts
h
av
e
s
ig
n
if
ican
tly
im
p
r
o
v
e
d
ef
f
icie
n
cy
,
ac
cu
r
ac
y
,
an
d
in
n
o
v
atio
n
ac
r
o
s
s
v
ar
io
u
s
in
d
u
s
tr
ies.
Ho
wev
er
,
ch
allen
g
es
r
em
ain
,
p
ar
ticu
lar
l
y
r
eg
ar
d
in
g
th
e
av
ailab
ilit
y
an
d
q
u
ality
o
f
d
atasets
cr
u
cial
f
o
r
tr
ain
in
g
ef
f
ec
tiv
e
m
o
d
els
[
2
]
–
[
4
]
.
T
h
e
s
ca
r
city
o
f
r
el
iab
le
lab
eled
d
ata
an
d
th
e
h
ig
h
ac
q
u
is
itio
n
co
s
ts
co
n
tin
u
e
to
h
in
d
er
A
I
d
ev
elo
p
m
e
n
t.
Fo
r
e
x
am
p
le,
m
ed
ical
im
ag
in
g
d
atasets
o
f
ten
co
n
tain
m
is
lab
eled
s
am
p
les,
lo
w
-
r
eso
lu
tio
n
im
ag
es,
o
r
in
s
u
f
f
icien
t
r
ep
r
e
s
en
tatio
n
s
o
f
r
ar
e
d
is
ea
s
es,
all
o
f
wh
ic
h
d
e
g
r
ad
e
m
o
d
el
p
er
f
o
r
m
an
ce
[
5
]
.
Alth
o
u
g
h
d
ata
au
g
m
e
n
tatio
n
tech
n
iq
u
es,
s
u
ch
as
f
lip
p
i
n
g
an
d
n
o
is
e
in
jectio
n
,
ar
e
co
m
m
o
n
ly
ap
p
lied
,
th
e
y
o
f
ten
in
tr
o
d
u
ce
b
iases
o
r
f
ail
to
f
u
lly
ad
d
r
ess
d
ataset
d
ef
icien
cies
[
6
]
,
[
7
]
.
As
m
o
d
els
b
ec
o
m
e
m
o
r
e
co
m
p
lex
,
th
eir
p
er
f
o
r
m
an
ce
in
c
r
ea
s
in
g
ly
d
ep
en
d
s
o
n
h
ig
h
-
q
u
ality
d
atasets
r
ath
er
th
an
s
o
lely
o
n
ar
ch
itectu
r
a
l
in
n
o
v
atio
n
s
[
3
]
,
[
4
]
.
T
h
is
s
h
if
t
h
as
g
iv
en
r
is
e
to
d
ata
-
ce
n
tr
ic
AI
,
wh
ich
f
o
cu
s
es
o
n
o
p
tim
izin
g
d
atasets
r
ath
er
th
an
ex
cl
u
s
iv
ely
en
h
an
cin
g
m
o
d
el
ar
c
h
itectu
r
es
[
2
]
,
[
3
]
.
Data
-
ce
n
tr
ic
m
et
h
o
d
o
lo
g
ies
p
r
io
r
itize
r
ef
in
in
g
d
atasets
b
y
ad
d
r
ess
in
g
lab
el
n
o
is
e,
class
im
b
alan
ce
,
an
d
ir
r
el
ev
an
t d
ata
p
o
in
ts
[
4
]
–
[
8
]
.
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
.
3
,
J
u
n
e
20
25
:
2
9
5
1
-
2
9
6
4
2952
An
d
r
ew
Ng
'
s
d
ata
-
ce
n
tr
ic
AI
c
o
m
p
etitio
n
u
n
d
er
s
co
r
ed
th
e
im
p
o
r
tan
ce
o
f
d
ataset
q
u
ality
in
ac
h
iev
in
g
s
u
p
er
io
r
m
o
d
el
p
er
f
o
r
m
an
ce
,
ev
en
with
s
m
aller
d
atasets
[
1
]
.
B
y
s
y
s
tem
atica
lly
en
h
an
cin
g
d
ata
q
u
ality
,
r
esear
ch
er
s
h
av
e
d
em
o
n
s
tr
ated
n
o
ta
b
le
im
p
r
o
v
em
en
ts
in
m
o
d
el
r
o
b
u
s
tn
ess
,
g
en
e
r
aliza
tio
n
,
an
d
ef
f
icien
cy
ac
r
o
s
s
d
iv
er
s
e
a
p
p
licatio
n
s
[
2
]
–
[
7
]
.
T
h
is
co
m
p
etitio
n
h
i
g
h
li
g
h
ted
t
h
e
o
f
ten
-
o
v
er
lo
o
k
ed
r
o
le
o
f
h
ig
h
-
q
u
ality
d
ata
in
d
r
iv
i
n
g
p
e
r
f
o
r
m
an
ce
g
ain
s
,
ev
en
wh
en
u
s
in
g
s
im
p
le
r
o
r
s
m
aller
m
o
d
els,
r
ein
f
o
r
cin
g
th
e
s
h
if
t
to
war
d
d
ata
-
ce
n
tr
ic
AI
p
r
ac
tices.
B
u
ild
in
g
u
p
o
n
th
ese
in
s
ig
h
ts
,
th
is
s
tu
d
y
in
tr
o
d
u
ce
s
a
d
at
a
-
ce
n
tr
ic
alg
o
r
ith
m
d
esig
n
ed
to
im
p
r
o
v
e
d
atasets
f
o
r
tr
ain
in
g
r
o
b
u
s
t
d
ee
p
n
e
u
r
al
n
etwo
r
k
s
(
DNNs)
.
T
h
e
alg
o
r
ith
m
em
p
lo
y
s
tech
n
iq
u
es
to
i
d
en
tify
an
d
r
e
m
o
v
e
n
o
is
y
o
r
m
is
lab
eled
s
a
m
p
les,
wh
ich
ca
n
d
is
to
r
t
m
o
d
el
lear
n
in
g
an
d
r
ed
u
ce
p
er
f
o
r
m
an
ce
.
Ad
d
itio
n
ally
,
it
in
co
r
p
o
r
ates
h
ig
h
-
q
u
ality
s
y
n
th
etic
d
ata
g
en
er
ated
u
s
in
g
g
en
er
ativ
e
m
o
d
els
s
u
ch
as
g
en
er
ativ
e
ad
v
e
r
s
ar
ial
n
etwo
r
k
s
(
GANs)
[
9
]
.
T
h
ese
GAN
-
g
en
er
ated
s
am
p
les
ad
d
r
ess
is
s
u
es
lik
e
cla
s
s
im
b
alan
ce
an
d
d
ataset
s
p
ar
s
ity
,
p
r
o
v
id
in
g
th
e
m
o
d
el
with
a
m
o
r
e
co
m
p
r
eh
en
s
iv
e
an
d
r
ep
r
esen
tativ
e
tr
ain
in
g
s
et.
T
h
is
ap
p
r
o
ac
h
p
r
o
v
es
p
ar
t
icu
lar
ly
v
alu
ab
le
i
n
h
ig
h
-
s
tak
e
s
d
o
m
ain
s
lik
e
h
ea
lth
ca
r
e
a
n
d
f
in
an
ce
,
wh
er
e
th
e
ac
cu
r
ac
y
an
d
r
eliab
ilit
y
o
f
p
r
ed
ictiv
e
m
o
d
els
ar
e
cr
itical
f
o
r
i
n
f
o
r
m
ed
d
ec
is
io
n
-
m
a
k
in
g
[
4
]
–
[
8
]
,
[
1
0
]
.
B
y
d
em
o
n
s
tr
atin
g
th
at
p
r
io
r
itizin
g
d
ataset
q
u
ality
ca
n
s
ig
n
if
ican
tly
im
p
r
o
v
e
p
er
f
o
r
m
a
n
ce
,
th
is
r
esear
ch
co
n
tr
ib
u
tes
to
th
e
g
r
o
win
g
b
o
d
y
o
f
ev
id
en
ce
s
u
p
p
o
r
tin
g
d
ata
-
ce
n
tr
ic
AI
as
a
co
m
p
lem
en
t
ar
y
m
eth
o
d
o
lo
g
y
to
m
o
d
el
-
ce
n
tr
ic
a
p
p
r
o
ac
h
es.
T
h
e
f
in
d
in
g
s
s
u
g
g
est
th
at
an
e
f
f
e
ctiv
e
b
alan
ce
b
etwe
en
th
ese
two
p
ar
ad
ig
m
s
lead
s
to
b
etter
p
er
f
o
r
m
an
ce
an
d
m
o
r
e
s
tab
le
o
u
tco
m
es.
T
h
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
,
th
er
ef
o
r
e,
em
p
h
asizes
th
e
im
p
o
r
tan
ce
o
f
in
teg
r
atin
g
b
o
t
h
p
er
s
p
ec
tiv
es
to
ad
v
a
n
ce
m
a
ch
in
e
lear
n
in
g
s
y
s
tem
s
,
en
s
u
r
in
g
th
at
m
o
d
els
ca
n
lear
n
f
r
o
m
clea
n
er
,
m
o
r
e
m
ea
n
in
g
f
u
l
d
ata
wh
ile
le
v
er
ag
in
g
o
p
tim
ized
ar
c
h
itectu
r
es
f
o
r
e
n
h
an
ce
d
p
er
f
o
r
m
an
ce
.
2.
RE
L
AT
E
D
WO
RK
S
Sig
n
i
f
ic
a
n
t
p
r
io
r
r
ese
ar
c
h
h
as
a
d
v
an
ce
d
d
at
a
-
ce
n
t
r
i
c
A
I
b
y
a
d
d
r
ess
in
g
d
a
tase
t
q
u
al
ity
is
s
u
es.
Stu
d
y
[
1
]
d
em
o
n
s
t
r
at
e
d
t
h
a
t v
a
r
ia
ti
o
n
s
i
n
d
a
tase
t
s
i
ze
,
l
a
b
el
in
g
q
u
al
it
y
,
a
n
d
t
r
a
in
-
t
est
s
p
l
its
s
u
b
s
ta
n
ti
all
y
im
p
ac
t
m
o
d
el
p
e
r
f
o
r
m
an
ce
,
u
n
d
er
s
co
r
in
g
t
h
e
i
m
p
o
r
ta
n
c
e
o
f
d
at
a
-
ce
n
t
r
ic
m
e
th
o
d
o
l
o
g
ies
.
Si
m
il
ar
l
y
,
Sa
m
b
asi
v
a
n
e
t
a
l.
[
2
]
e
m
p
h
asiz
ed
t
h
e
r
i
p
p
le
e
f
f
e
cts
o
f
la
b
el
in
g
e
r
r
o
r
s
,
a
d
v
o
c
ati
n
g
f
o
r
s
y
s
te
m
at
ic
d
at
a
q
u
al
it
y
i
m
p
r
o
v
e
m
e
n
ts
.
No
r
th
cu
tt
et
a
l.
[
1
1
]
id
en
tifie
d
p
er
v
asiv
e
lab
elin
g
er
r
o
r
s
in
wid
ely
u
s
ed
d
atasets
,
r
ein
f
o
r
cin
g
th
e
n
ee
d
f
o
r
d
ata
-
ce
n
tr
ic
s
tr
ateg
ies.
Po
ly
zo
tis
an
d
Z
ah
ar
ia
[
4
]
p
r
o
p
o
s
ed
en
d
-
to
-
en
d
v
er
s
io
n
tr
ac
k
in
g
an
d
ac
tio
n
a
b
le
m
o
n
ito
r
in
g
f
o
r
m
a
n
ag
in
g
d
y
n
am
ic.
d
atasets
in
p
r
o
d
u
ctio
n
s
y
s
tem
s
.
Ham
id
[
1
2
]
h
i
g
h
lig
h
t
ed
th
e
a
p
p
licab
ilit
y
o
f
d
ata
-
ce
n
tr
ic
AI
in
I
n
d
u
s
tr
y
4
.
0
b
y
en
h
a
n
cin
g
d
ata
q
u
ality
f
o
r
r
o
b
u
s
t in
d
u
s
tr
ial
au
to
m
atio
n
.
I
n
n
atu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
(
NL
P),
Xu
et
a
l.
[
3
]
in
tr
o
d
u
ce
d
th
e
d
ata
C
L
UE
b
en
ch
m
ar
k
,
s
h
o
win
g
s
im
p
le
y
et
ef
f
ec
tiv
e
s
tr
ateg
ies
f
o
r
im
p
r
o
v
in
g
d
ata
q
u
ality
.
Seed
at
et
a
l.
[
1
3
]
p
r
o
p
o
s
ed
DC
-
ch
ec
k
,
a
ch
ec
k
lis
t
f
o
r
s
y
s
tem
atica
lly
ev
alu
atin
g
d
ata
-
ce
n
tr
ic
r
eliab
ilit
y
.
Seed
at
et
a
l.
[
1
4
]
also
em
p
h
asized
d
ataset
tr
an
s
p
ar
en
cy
with
d
atash
ee
ts
f
o
r
d
atasets
,
a
f
o
u
n
d
atio
n
al
to
o
l i
n
h
ig
h
-
s
tak
es a
p
p
licatio
n
s
.
Mo
t
am
e
d
i
et
a
l
.
[
6
]
d
em
o
n
s
t
r
a
ted
d
at
a
q
u
ali
ty
e
n
h
a
n
c
em
e
n
ts
wit
h
GA
N
-
g
e
n
er
ate
d
s
am
p
l
es,
ac
h
ie
v
i
n
g
im
p
r
o
v
e
d
a
cc
u
r
a
cy
wh
ile
r
ed
u
ci
n
g
d
a
tase
t
s
iz
e.
Ma
et
a
l
.
[
1
5
]
r
e
v
i
ewe
d
d
ata
-
c
en
tr
ic
A
I
’
s
r
o
le
in
a
d
d
r
ess
in
g
lab
eli
n
g
e
r
r
o
r
s
a
n
d
cl
ass
i
m
b
al
a
n
ce
s
.
S
h
a
n
k
a
r
an
d
E
v
a
n
s
[
1
6
]
i
d
en
ti
f
ie
d
p
it
f
alls
i
n
d
atas
et
c
o
n
s
t
r
u
ct
io
n
,
p
r
o
p
o
s
i
n
g
b
est
p
r
ac
tic
es
f
o
r
r
el
iab
ilit
y
a
n
d
f
a
ir
n
ess
.
Z
h
a
et
a
l.
[
5
]
in
tr
o
d
u
ce
d
a
f
r
am
ewo
r
k
in
te
g
r
atin
g
s
tatis
tica
l
m
ea
s
u
r
es
to
en
s
u
r
e
d
ataset
r
eliab
ilit
y
.
T
h
ese
s
tu
d
ies
h
ig
h
lig
h
t
th
e
t
r
an
s
f
o
r
m
ativ
e
p
o
ten
tial
o
f
d
a
ta
-
ce
n
tr
ic
AI
in
ad
d
r
ess
in
g
d
ata
q
u
ality
is
s
u
es,
f
o
r
m
in
g
.
T
h
e
f
o
u
n
d
atio
n
f
o
r
th
e
d
ataset
o
p
tim
izatio
n
a
p
p
r
o
a
ch
p
r
esen
ted
i
n
th
is
wo
r
k
.
3.
M
E
T
H
O
D
I
n
m
ac
h
in
e
lear
n
in
g
,
class
if
icatio
n
r
ef
e
r
s
to
a
p
r
ed
ictiv
e
m
o
d
elin
g
p
r
o
b
lem
wh
er
e
a
clas
s
lab
el
is
p
r
ed
icted
f
o
r
a
g
iv
en
in
p
u
t
d
at
a
ex
am
p
le.
T
o
co
n
s
tr
u
ct
s
u
ch
a
p
r
ed
ictiv
e
m
o
d
el,
we
d
ef
i
n
e
lo
s
s
f
u
n
ctio
n
s
,
s
et
h
y
p
er
p
ar
am
eter
s
f
o
r
th
e
m
o
d
e
l,
an
d
,
g
iv
en
a
tr
ain
in
g
d
ataset
,
o
p
tim
ize
th
e
m
o
d
el
p
ar
a
m
eter
s
to
m
in
im
ize
th
e
lo
s
s
f
u
n
ctio
n
,
as e
x
p
r
ess
ed
in
(
1
)
.
,
,
lo
ss
(
(
,
)
)
(
1
)
T
h
e
tar
g
eted
lo
s
s
f
u
n
ctio
n
,
d
e
n
o
ted
as
lo
s
s
,
p
lay
s
a
cr
u
cial
r
o
le
in
m
o
d
el
o
p
tim
izatio
n
,
wh
er
e
f
is
th
e
m
ap
p
i
n
g
m
o
d
el,
φ
r
ep
r
esen
ts
th
e
h
y
p
er
p
ar
am
eter
s
,
θ
th
e
m
o
d
el
p
ar
am
eter
s
,
a
n
d
γ
th
e
g
i
v
en
d
ataset.
R
ec
en
t
ad
v
an
ce
m
e
n
ts
in
d
ee
p
lear
n
in
g
h
av
e
estab
lis
h
ed
it
as
a
p
o
w
er
f
u
l
to
o
l
f
o
r
lear
n
in
g
h
ier
ar
c
h
ical
r
ep
r
esen
tatio
n
s
f
r
o
m
lar
g
e
d
atasets
,
d
r
iv
in
g
b
r
ea
k
th
r
o
u
g
h
s
in
d
o
m
ai
n
s
lik
e
im
ag
e
r
ec
o
g
n
itio
n
,
n
atu
r
al
la
n
g
u
ag
e
p
r
o
ce
s
s
in
g
,
an
d
au
to
n
o
m
o
u
s
s
y
s
tem
s
[
1
]
–
[
8
]
,
[
1
0
]
–
[
1
7
]
.
T
o
o
p
tim
ize
p
er
f
o
r
m
a
n
ce
,
v
ar
io
u
s
lo
s
s
f
u
n
ctio
n
s
h
av
e
b
ee
n
d
ev
elo
p
e
d
b
ased
o
n
task
r
eq
u
ir
em
en
ts
,
s
u
ch
as
m
ea
n
s
q
u
ar
ed
er
r
o
r
f
o
r
r
eg
r
ess
io
n
an
d
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
f
o
r
class
if
icatio
n
[
1
8
]
,
[
1
9
]
.
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
E
n
h
a
n
cin
g
tr
a
in
in
g
p
erfo
r
ma
n
ce
fo
r
s
ma
ll mo
d
els u
s
in
g
d
a
t
a
-
ce
n
tr
ic
a
p
p
r
o
a
ch
es
(
R
ed
a
A
.
E
l
-
K
h
o
r
ib
i
)
2953
3
.
1
.
Reg
re
s
s
io
n
lo
s
s
es
R
eg
r
ess
io
n
lo
s
s
es q
u
an
tify
d
is
cr
ep
an
cies b
etwe
en
p
r
ed
icted
an
d
ac
tu
al
co
n
tin
u
o
u
s
v
alu
es:
a.
Me
an
s
q
u
ar
ed
e
r
r
o
r
(
MSE
)
: Pen
alize
s
lar
g
e
er
r
o
r
s
m
o
r
e
h
ea
v
ily
,
m
ak
in
g
it sen
s
itiv
e
to
o
u
t
lier
s
[
2
0
]
.
M
S
E
=
1
∑
(
−
̂
)
2
=
1
(
2
)
b.
Me
an
ab
s
o
lu
te
er
r
o
r
(
L
1
)
: M
e
asu
r
es a
b
s
o
lu
te
d
if
f
er
e
n
ce
s
,
less
s
en
s
i
tiv
e
to
o
u
tlier
s
[
1
]
.
L1
=
1
∑
|
−
̂
|
=
1
(
3
)
c.
Hu
b
er
l
o
s
s
:
C
o
m
b
in
es
MSE
f
o
r
s
m
all
er
r
o
r
s
an
d
L
1
f
o
r
lar
g
e
er
r
o
r
s
,
o
f
f
er
in
g
r
o
b
u
s
tn
ess
[
2
1
]
.
=
{
1
2
(
−
̂
)
2
,
|
−
̂
|
≤
,
|
−
̂
|
,
ℎ
.
(
4
)
3
.
2
.
Cla
s
s
if
ica
t
io
n
lo
s
s
es
C
las
s
if
icatio
n
lo
s
s
es
ev
alu
ate
m
o
d
el
p
e
r
f
o
r
m
an
ce
in
class
if
icatio
n
task
s
b
y
m
ea
s
u
r
in
g
d
is
cr
ep
an
cies
b
etwe
en
p
r
ed
icted
an
d
ac
t
u
al
class
lab
els:
a.
B
in
ar
y
cr
o
s
s
-
en
tr
o
p
y
(
B
C
E
)
:
Used
f
o
r
b
in
ar
y
class
if
icatio
n
b
y
m
ea
s
u
r
in
g
th
e
d
if
f
e
r
en
ce
b
etwe
en
p
r
o
b
a
b
ilit
y
d
is
tr
ib
u
tio
n
s
[
2
0
]
.
BCE
=
−
∑
[
l
og
(
̂
)
+
(
1
−
)
l
og
(
1
−
̂
)
]
=
1
(
5
)
b.
C
ate
g
o
r
i
ca
l
c
r
o
s
s
-
e
n
t
r
o
p
y
(
C
C
E
)
:
E
x
te
n
d
s
B
C
E
t
o
m
u
l
ti
-
cl
as
s
c
lass
i
f
i
ca
t
io
n
b
y
s
u
m
m
in
g
t
h
e
cr
o
s
s
-
e
n
t
r
o
p
y
f
o
r
al
l
class
es
[
2
0
]
.
CCE
=
−
∑
∑
=
1
l
og
(
̂
)
=
1
(
6
)
c.
Fo
ca
l
lo
s
s
(
FL)
:
Ad
d
r
ess
es
cla
s
s
im
b
alan
ce
b
y
d
o
w
n
-
weig
h
ti
n
g
ea
s
y
ex
am
p
les an
d
em
p
h
as
izin
g
h
ar
d
o
n
es
[
2
2
]
.
FL
=
−
(
1
−
̂
)
l
og
(
̂
)
(
7
)
3
.
3
.
O
t
her
lo
s
s
es
Sp
ec
ialized
lo
s
s
f
u
n
ctio
n
s
ad
d
r
ess
u
n
iq
u
e
ch
allen
g
es in
tr
ain
in
g
m
o
d
els:
a.
Ku
llb
ac
k
-
L
eib
ler
d
iv
er
g
e
n
ce
(
KL
)
:
Me
asu
r
es
th
e
d
iv
er
g
en
ce
b
etwe
en
two
p
r
o
b
a
b
ilit
y
d
is
tr
ib
u
tio
n
s
,
co
m
m
o
n
l
y
u
s
ed
in
p
r
o
b
ab
ilis
tic
m
o
d
els
[
1
7
]
.
KL
=
∑
=
1
l
og
̂
(
8
)
b.
Hin
g
e
lo
s
s
:
Used
in
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
es (
SVMs)
to
m
a
x
im
ize
class
m
ar
g
in
s
[
2
3
]
.
Hinge
=
ma
x
(
0
,
1
−
̂
)
(
9
)
c.
T
o
tal
v
ar
iatio
n
(
T
V)
:
Pro
m
o
te
s
s
m
o
o
th
n
ess
in
im
ag
es o
r
s
ig
n
als,
r
ed
u
cin
g
n
o
is
e
[
7
]
.
TV
=
∑
(
|
|
+
1
,
−
,
|
|
+
|
|
,
+
1
−
,
|
|
)
,
(
1
0
)
3
.
4
.
H
y
perpa
ra
m
e
t
er
s
Hy
p
er
p
ar
a
m
eter
s
ar
e
p
r
ed
e
f
in
ed
p
ar
a
m
eter
s
cr
itical
to
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
m
ac
h
in
e
lear
n
i
n
g
m
o
d
els.
T
h
eir
s
elec
tio
n
s
ig
n
if
ican
tly
i
m
p
ac
ts
m
o
d
el
tr
ain
in
g
an
d
o
u
t
co
m
es.
a.
L
ea
r
n
in
g
r
ate
(
η
)
:
Dete
r
m
in
es
th
e
s
tep
s
ize
f
o
r
g
r
ad
ien
t
d
escen
t.
Sm
aller
η
en
s
u
r
es
co
n
v
er
g
en
ce
b
u
t
s
lo
ws
tr
ain
in
g
,
wh
ile
lar
g
er
η
m
ay
c
au
s
e
in
s
tab
ilit
y
[
1
7
]
.
≔
−
∇
(
)
(
1
1
)
b.
B
atch
s
ize:
Sp
ec
if
ies
th
e
n
u
m
b
er
o
f
tr
ain
in
g
ex
am
p
les p
er
iter
atio
n
.
Sm
aller
b
atch
es r
ed
u
c
e
o
v
er
f
itti
n
g
b
u
t
in
cr
ea
s
e
g
r
ad
ien
t
n
o
is
e
[
1
7
]
–
[
2
5
]
.
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
.
3
,
J
u
n
e
20
25
:
2
9
5
1
-
2
9
6
4
2954
c.
Dep
th
a
n
d
wi
d
t
h
:
T
h
e
n
u
m
b
e
r
o
f
l
ay
e
r
s
an
d
n
e
u
r
o
n
s
p
e
r
la
y
er
im
p
ac
t
t
h
e
m
o
d
e
l’
s
r
e
p
r
ese
n
t
a
tio
n
a
l
ca
p
a
cit
y
.
d.
E
x
ce
s
s
iv
e
lay
er
s
o
r
n
e
u
r
o
n
s
m
ay
lead
to
o
v
er
f
itti
n
g
an
d
i
n
ef
f
icien
cies
[
1
7
]
–
[
2
5
]
,
[
2
6
]
[
2
7
]
.
e.
Activ
atio
n
f
u
n
ctio
n
s
:
Ad
d
n
o
n
-
lin
ea
r
ity
to
n
etwo
r
k
s
,
en
ab
lin
g
co
m
p
lex
p
atter
n
lear
n
in
g
.
C
o
m
m
o
n
ch
o
ices
in
clu
d
e
R
eL
U,
s
ig
m
o
id
,
a
n
d
ta
n
h
[
1
7
]
–
[
2
5
]
,
[
2
6
]
–
[
3
0
]
.
f.
R
eg
u
lar
izatio
n
p
ar
a
m
eter
s
:
Pre
v
en
t
o
v
er
f
itti
n
g
b
y
p
en
alizi
n
g
lar
g
e
weig
h
ts
.
L
1
p
r
o
m
o
t
es
s
p
ar
s
ity
;
L
2
r
ed
u
ce
s
v
ar
ian
ce
[
2
8
]
–
[
3
2
]
.
3
.
5
.
Da
t
a
s
et
o
pti
m
iza
t
io
n
Op
tim
izin
g
th
e
d
ataset
is
a
cr
itical
asp
ec
t
o
f
th
is
wo
r
k
.
Op
er
atio
n
s
s
u
ch
as
p
r
u
n
i
n
g
tr
ain
in
g
p
o
in
ts
,
in
s
er
tin
g
n
ewly
g
en
er
ated
d
at
a
p
o
in
ts
,
an
d
weig
h
tin
g
s
p
ec
i
f
ic
p
o
in
ts
in
t
h
e
lo
s
s
f
u
n
ctio
n
ar
e
p
er
f
o
r
m
ed
t
o
en
h
an
ce
d
ataset
q
u
ality
.
T
h
es
e
o
p
tim
izatio
n
s
aim
to
im
p
r
o
v
e
th
e
r
ep
r
esen
tativ
en
ess
,
b
alan
ce
,
an
d
clea
n
lin
ess
o
f
th
e
tr
ain
i
n
g
d
ataset,
lead
in
g
to
m
o
r
e
r
o
b
u
s
t a
n
d
ac
cu
r
ate
m
ac
h
in
e
lear
n
in
g
m
o
d
els.
B
ef
o
r
e
im
p
lem
en
tin
g
d
ataset
o
p
tim
izatio
n
,
d
iag
n
o
s
in
g
p
o
ten
tial
is
s
u
es
with
in
th
e
d
ataset
is
n
ec
ess
ar
y
to
id
en
tify
a
p
p
r
o
p
r
iate
r
em
ed
ies.
A
s
tr
aig
h
tf
o
r
war
d
ap
p
r
o
ac
h
to
d
iag
n
o
s
in
g
d
ataset
p
r
o
b
lem
s
in
v
o
lv
es
tr
ain
in
g
a
b
asic
m
o
d
el
o
n
th
e
d
ataset
an
d
an
aly
zin
g
its
p
er
f
o
r
m
an
ce
to
d
etec
t
n
o
is
y
o
r
m
is
lab
eled
s
am
p
les.
On
ce
id
en
tifie
d
,
th
es
e
p
r
o
b
lem
atic
s
am
p
les ca
n
eith
er
b
e
c
o
r
r
ec
ted
o
r
r
e
m
o
v
e
d
.
T
h
e
d
ataset
o
p
tim
izatio
n
p
r
o
c
ess
in
clu
d
es two
p
r
im
ar
y
o
p
er
atio
n
s
:
a.
Sam
p
le
p
r
u
n
i
n
g
:
R
em
o
v
in
g
n
o
is
y
o
r
m
is
lab
eled
s
am
p
les f
r
o
m
th
e
d
ataset.
b.
Sam
p
le
in
s
er
tio
n
:
Au
g
m
e
n
tin
g
th
e
d
ataset
with
h
i
g
h
-
q
u
ality
s
y
n
th
etic
s
am
p
les
to
a
d
d
r
ess
class
im
b
alan
ce
o
r
u
n
d
er
-
r
e
p
r
esen
ted
f
ea
tu
r
e
.
W
e
b
eg
in
b
y
tr
ain
in
g
a
b
aselin
e
m
o
d
el
o
n
th
e
o
r
i
g
in
al
d
a
taset,
an
d
b
ased
o
n
its
o
u
tp
u
ts
,
we
ap
p
ly
th
ese
en
h
an
ce
m
e
n
t o
p
er
ati
o
n
s
.
3
.
5
.
1
.
Sa
m
ple
prun
ing
Sam
p
l
e
p
r
u
n
in
g
in
v
o
l
v
es
id
en
t
if
y
i
n
g
a
n
d
r
em
o
v
i
n
g
n
o
is
y
o
r
p
r
o
b
l
em
ati
c
s
a
m
p
les
f
r
o
m
t
h
e
d
at
ase
t
,
as
s
h
o
wn
in
Fi
g
u
r
e
1
,
T
h
ese
is
s
u
es
m
a
y
i
n
cl
u
d
e
m
is
l
ab
ele
d
e
x
a
m
p
les
,
a
m
b
ig
u
o
u
s
d
ata
p
o
i
n
ts
,
o
r
s
a
m
p
les
wi
th
co
n
f
li
cti
n
g
f
e
at
u
r
es.
T
h
is
p
r
o
c
ess
t
y
p
ic
all
y
r
eli
es o
n
t
ec
h
n
i
q
u
es
s
u
c
h
as
co
n
f
i
d
e
n
t
le
ar
n
i
n
g
,
wh
i
ch
es
tim
at
es
t
h
e
jo
i
n
t
d
is
tr
ib
u
t
io
n
o
f
n
o
is
y
a
n
d
cle
an
la
b
els
t
o
d
e
tec
t
a
n
d
a
d
d
r
ess
e
r
r
o
r
s
s
y
s
t
em
ati
ca
l
ly
.
T
h
e
s
te
p
s
i
n
v
o
lv
e
d
i
n
s
am
p
le
p
r
u
n
in
g
ar
e
as
f
o
ll
o
w
s
:
i)
co
lle
ct
o
r
s
el
ec
t
a
r
e
p
r
es
en
t
ati
v
e
d
atas
et;
ii
)
t
r
ai
n
a
b
a
s
eli
n
e
m
o
d
el
o
n
th
e
d
at
ase
t
t
o
est
a
b
lis
h
a
p
er
f
o
r
m
an
ce
b
e
n
c
h
m
a
r
k
;
ii
i)
u
s
e
c
r
o
s
s
-
v
ali
d
ati
o
n
o
n
t
h
e
t
r
ai
n
i
n
g
d
a
ta
t
o
i
d
e
n
t
if
y
n
o
is
y
s
am
p
les
;
i
v
)
ap
p
l
y
te
c
h
n
iq
u
es
s
u
ch
as
la
b
el
er
r
o
r
s
a
n
d
co
n
f
id
e
n
t
l
ea
r
n
in
g
t
o
d
e
tec
t
m
is
l
a
b
el
e
d
o
r
a
m
b
i
g
u
o
u
s
d
at
a
p
o
i
n
ts
;
a
n
d
v
)
r
em
o
v
e
o
r
r
ewe
i
g
h
t
t
h
e
p
r
o
b
le
m
a
tic
s
a
m
p
l
es t
o
o
b
t
ai
n
a
cle
a
n
d
a
tas
et
f
o
r
tr
a
in
in
g
.
3
.
5
.
2
.
Sa
m
ple ins
er
t
io
n
Fo
llo
win
g
th
e
p
r
u
n
in
g
o
f
n
o
is
y
s
am
p
les,
as
s
h
o
wn
in
Fig
u
r
e
2
,
n
ew
s
am
p
les
ar
e
g
en
er
ate
d
to
en
r
ich
th
e
d
ataset.
T
h
is
p
r
o
ce
s
s
in
v
o
l
v
es
id
en
tify
i
n
g
wea
k
class
es,
ad
d
r
ess
in
g
n
am
in
g
er
r
o
r
s
,
a
n
d
tr
ain
in
g
a
GAN
to
g
en
er
ate
n
ew
in
s
tan
ce
s
f
o
r
th
e
u
n
d
er
r
ep
r
esen
ted
class
.
T
h
is
s
tep
en
s
u
r
es
b
etter
class
r
ep
r
e
s
en
tatio
n
an
d
h
el
p
s
th
e
m
o
d
el
lear
n
m
o
r
e
ef
f
ec
tiv
ely
f
r
o
m
th
e
a
v
ailab
le
d
ata.
T
h
e
s
tep
s
f
o
r
im
p
lem
en
tin
g
s
a
m
p
le
in
s
er
tio
n
a
r
e
as
f
o
llo
ws:
i)
C
o
llect
o
r
s
elec
t
a
r
ep
r
esen
tativ
e
d
ataset
to
s
er
v
e
as
th
e
f
o
u
n
d
atio
n
f
o
r
tr
ai
n
in
g
;
ii)
T
r
ain
t
h
e
d
ataset
o
n
a
b
aselin
e
m
o
d
el
an
d
ca
lcu
late
th
e
in
itial
a
cc
u
r
ac
y
t
o
estab
lis
h
p
e
r
f
o
r
m
an
ce
b
e
n
ch
m
a
r
k
s
;
iii)
I
d
en
tify
th
e
wea
k
class
b
y
an
aly
zi
n
g
t
h
e
ac
cu
r
ac
y
r
esu
lts
an
d
d
eter
m
in
in
g
wh
ich
class
es
u
n
d
er
p
er
f
o
r
m
;
iv
)
T
r
ain
a
GAN
m
o
d
el
s
p
e
cif
ically
f
o
r
th
e
wea
k
class
to
g
en
er
ate
h
ig
h
-
q
u
ality
s
y
n
th
etic
in
s
tan
ce
s
th
a
t
im
p
r
o
v
e
class
b
alan
ce
an
d
o
v
er
all
m
o
d
el
p
er
f
o
r
m
a
n
ce
[
3
2
]
;
an
d
v
)
Ad
d
th
e
g
en
er
ate
d
in
s
tan
ce
s
to
th
e
ex
is
tin
g
d
ataset
to
im
p
r
o
v
e
r
e
p
r
esen
tatio
n
an
d
class
if
icatio
n
ac
cu
r
ac
y
.
Af
ter
n
o
is
y
s
am
p
les
ar
e
p
r
u
n
ed
in
Ph
ase
1
,
th
e
r
em
ai
n
in
g
d
ataset
is
en
r
ich
ed
b
y
g
e
n
er
atin
g
n
ew
s
y
n
th
etic
s
am
p
les.
W
ea
k
class
es
ar
e
id
en
tifie
d
an
d
ad
d
r
ess
ed
b
y
g
e
n
er
atin
g
in
s
tan
ce
s
u
s
in
g
clea
n
d
ata
to
tr
ain
a
GAN
m
o
d
el.
T
h
ese
n
ew
s
a
m
p
les en
h
an
ce
th
e
d
ataset'
s
r
e
p
r
esen
tatio
n
an
d
class
if
icatio
n
ac
cu
r
ac
y
.
T
h
e
GAN
m
o
d
el'
s
o
p
tim
izatio
n
p
r
o
b
lem
is
d
ef
in
ed
as
[
3
2
]
.
min
ma
x
(
,
)
=
∼
dt
(
)
[
l
og
(
)
]
+
∼
(
)
[
l
og
(
1
−
(
(
)
)
)
]
(
1
2
)
T
h
e
g
e
n
er
ato
r
'
s
lo
s
s
f
u
n
ctio
n
m
in
im
izes
th
e
m
ax
i
m
u
m
v
alu
e
o
f
th
e
d
is
cr
im
in
ato
r
'
s
v
alu
e
f
u
n
ctio
n
b
y
g
en
er
atin
g
s
am
p
les lik
ely
to
b
e
m
is
class
if
ied
b
y
th
e
d
is
cr
im
in
ato
r
[
3
2
]
.
min
ma
x
(
,
)
=
∼
dt
(
)
[
l
og
(
)
]
+
∼
(
)
[
l
og
(
1
−
(
(
)
)
)
]
(
1
3
)
T
h
e
d
is
cr
im
in
ato
r
'
s
lo
s
s
f
u
n
ctio
n
m
ax
im
izes
its
v
alu
e
f
u
n
ctio
n
wh
ile
m
in
im
izin
g
th
e
g
en
er
ato
r
'
s
v
alu
e
f
u
n
ctio
n
,
en
ab
lin
g
it to
co
r
r
ec
t
ly
class
if
y
b
o
th
r
ea
l a
n
d
g
e
n
er
ated
s
am
p
les.
T
h
e
r
esu
ltin
g
e
n
h
an
ce
d
d
ataset
is
:
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
E
n
h
a
n
cin
g
tr
a
in
in
g
p
erfo
r
ma
n
ce
fo
r
s
ma
ll mo
d
els u
s
in
g
d
a
t
a
-
ce
n
tr
ic
a
p
p
r
o
a
ch
es
(
R
ed
a
A
.
E
l
-
K
h
o
r
ib
i
)
2955
̂
=
s
e
lec
ted
∪
ge
ne
r
a
ted
(
1
4
)
Her
e,
s
e
lec
ted
r
ep
r
esen
ts
th
e
p
r
u
n
ed
d
ataset,
an
d
ge
ne
r
a
ted
r
ep
r
esen
ts
n
ewly
g
en
er
ated
s
am
p
les.
T
h
e
o
p
tim
izatio
n
en
s
u
r
es th
at
a
m
o
d
el
f
tr
ain
ed
o
n
̂
o
u
tp
er
f
o
r
m
s
t
h
e
o
r
ig
in
al
d
ataset
.
Fig
u
r
e
1
.
Sam
p
le
p
r
u
n
in
g
s
u
b
s
y
s
tem
Fig
u
r
e
2
.
Sam
p
le
i
n
s
er
tio
n
wo
r
k
f
lo
w
3
.
6
.
E
x
perim
ent
a
l
s
et
up
W
e
v
alid
ated
th
e
p
r
o
p
o
s
ed
d
ata
-
ce
n
tr
ic
m
eth
o
d
s
u
s
in
g
t
h
e
C
I
FAR
-
1
0
d
ataset
[
9
]
,
wh
ic
h
in
clu
d
es
6
0
,
0
0
0
3
2
×
3
2
co
l
o
r
im
ag
es
ac
r
o
s
s
1
0
class
es
(
5
0
,
0
0
0
f
o
r
tr
ain
in
g
an
d
1
0
,
0
0
0
f
o
r
test
in
g
)
.
Pre
p
r
o
ce
s
s
in
g
ad
d
r
ess
ed
lab
el
n
o
is
e
an
d
lo
w
-
r
eso
lu
tio
n
s
am
p
les v
ia
p
r
u
n
in
g
an
d
GAN
-
b
ased
a
u
g
m
e
n
tatio
n
.
a.
B
aselin
e
m
o
d
els
:
−
R
esNet
-
18
[
3
3
]
:
An
18
-
lay
e
r
C
NN
lev
er
ag
in
g
r
esid
u
al
co
n
n
ec
tio
n
s
f
o
r
g
r
ad
ie
n
t
f
lo
w,
p
r
e
-
tr
ain
ed
o
n
I
m
ag
eNe
t.
−
I
n
ce
p
tio
n
V3
[
3
4
]
:
A
d
ee
p
C
NN
with
f
ac
to
r
ized
co
n
v
o
lu
ti
o
n
s
an
d
au
x
iliar
y
class
if
ier
s
f
o
r
im
p
r
o
v
ed
ac
cu
r
ac
y
.
b.
E
v
alu
atio
n
m
etr
ics:
Per
f
o
r
m
a
n
ce
was
ass
e
s
s
ed
u
s
in
g
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
F1
-
s
co
r
e,
an
d
s
tan
d
ar
d
d
ev
iatio
n
.
c.
T
r
ai
n
i
n
g
co
n
f
i
g
u
r
ati
o
n
:
T
r
a
in
i
n
g
em
p
l
o
y
e
d
t
h
e
A
d
am
o
p
ti
m
iz
er
[
3
0
]
wi
th
a
le
ar
n
i
n
g
r
a
te
o
f
1
×
10
−
4
b
at
c
h
s
ize
o
f
1
2
8
,
5
0
ep
o
c
h
s
,
an
d
s
t
an
d
ar
d
a
u
g
m
en
tat
io
n
s
(
r
a
n
d
o
m
c
r
o
p
p
in
g
,
f
l
ip
p
i
n
g
,
n
o
r
m
al
iz
ati
o
n
)
.
E
x
p
e
r
i
m
e
n
ts
w
er
e
p
e
r
f
o
r
m
e
d
o
n
a
n
N
VI
D
I
A
T
esl
a
V
1
0
0
GP
U
(
3
2
GB
)
.
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
.
3
,
J
u
n
e
20
25
:
2
9
5
1
-
2
9
6
4
2956
d.
Pro
p
o
s
ed
m
eth
o
d
s
etu
p
:
−
Sam
p
le
p
r
u
n
in
g
:
co
n
f
id
en
t
l
ea
r
n
in
g
[
3
5
]
i
d
en
tifie
d
an
d
r
em
o
v
ed
n
o
is
y
s
am
p
les
b
ased
o
n
lab
el
u
n
ce
r
tain
ty
.
−
Sam
p
le
in
s
er
tio
n
:
d
GAN
[
3
2
]
g
e
n
er
ated
s
y
n
th
etic
s
am
p
l
es
f
o
r
u
n
d
e
r
r
ep
r
esen
ted
class
es
u
s
in
g
a
s
tan
d
ar
d
GAN
ar
ch
itectu
r
e
with
a
1
0
0
-
d
im
en
s
io
n
al
late
n
t
v
ec
to
r
.
T
r
ain
in
g
co
n
s
is
te
d
o
f
1
0
,
0
0
0
iter
atio
n
s
,
with
v
is
u
al
in
s
p
ec
tio
n
o
f
g
en
er
ated
s
am
p
les f
o
r
q
u
ality
ass
u
r
an
ce
.
e.
B
aselin
e
co
m
p
ar
is
o
n
:
−
Or
ig
in
al
d
ataset:
b
aselin
e
m
o
d
els tr
ain
ed
o
n
u
n
alter
ed
C
I
FAR
-
10.
−
Pru
n
ed
d
ataset:
m
o
d
els tr
ain
ed
o
n
h
ig
h
-
q
u
ality
s
am
p
les p
o
s
t
-
p
r
u
n
i
n
g
.
−
E
n
h
an
ce
d
d
ataset:
m
o
d
els tr
ai
n
ed
o
n
p
r
u
n
ed
s
am
p
les au
g
m
e
n
ted
with
GAN
-
g
en
er
ated
in
s
tan
ce
s
.
4.
RE
SU
L
T
S
AND
D
I
SCU
SS
I
O
N
I
n
th
is
s
ec
tio
n
,
we
p
r
esen
t
th
e
o
u
tco
m
es
o
f
o
u
r
ex
p
er
im
en
ts
an
d
an
al
y
s
es
aim
ed
at
e
n
h
an
cin
g
t
h
e
tr
ain
in
g
p
er
f
o
r
m
an
ce
o
f
s
m
all
m
o
d
els
u
s
in
g
d
ata
-
ce
n
tr
ic
ap
p
r
o
ac
h
es
[
1
]
,
[
2
]
.
T
h
e
f
o
c
u
s
is
o
n
two
d
ee
p
class
if
icatio
n
m
o
d
els,
I
n
ce
p
ti
o
n
V3
a
n
d
R
esNet1
8
,
wh
ich
s
er
v
e
as
o
u
r
b
aselin
e
m
o
d
els.
W
e
d
etail
th
e
p
er
f
o
r
m
an
ce
im
p
r
o
v
e
m
en
ts
a
ch
iev
ed
th
r
o
u
g
h
v
a
r
io
u
s
s
tag
es
o
f
d
ata
o
p
tim
izatio
n
,
i
n
clu
d
in
g
p
r
u
n
in
g
a
n
d
au
g
m
en
tatio
n
with
g
en
er
ated
s
am
p
les.
T
h
e
ev
alu
atio
n
cr
iter
ia
u
s
ed
f
o
r
ass
ess
in
g
m
o
d
el
p
er
f
o
r
m
a
n
ce
in
clu
d
e
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
F1
-
s
co
r
e,
an
d
s
tan
d
ar
d
d
ev
iatio
n
[
3
]
.
4
.
1
.
Ado
pte
d
mo
dels
Gen
er
ally
,
an
y
m
a
p
p
in
g
m
o
d
el
f
ca
n
b
e
ad
o
p
te
d
as
a
b
aselin
e
f
o
r
test
in
g
o
u
r
p
r
o
p
o
s
als
f
o
r
d
ata
-
ce
n
tr
ic
en
h
an
ce
m
e
n
t.
I
n
th
is
s
tu
d
y
,
we
u
s
ed
two
m
o
d
els
as
b
aselin
es:
I
n
ce
p
tio
n
V3
an
d
R
esNet1
8
,
wh
ich
ar
e
r
ec
en
t e
x
am
p
les o
f
d
ee
p
class
i
f
icatio
n
m
o
d
els
[
3
3
]
,
[
3
4
]
.
a.
I
n
ce
p
tio
n
V3
:
T
h
is
im
a
g
e
r
e
co
g
n
itio
n
m
o
d
el
h
as
d
em
o
n
s
tr
ated
ac
cu
r
ac
y
g
r
ea
ter
th
a
n
7
8
.
1
%
o
n
t
h
e
I
m
ag
eNe
t
d
ataset.
I
t
in
teg
r
ate
s
s
y
m
m
etr
ic
an
d
asy
m
m
etr
ic
b
u
ild
in
g
b
lo
ck
s
s
u
ch
as
co
n
v
o
lu
tio
n
s
,
av
er
ag
e
p
o
o
lin
g
,
m
ax
p
o
o
lin
g
,
co
n
ca
ten
atio
n
s
,
d
r
o
p
o
u
ts
,
an
d
f
u
ll
y
co
n
n
ec
te
d
lay
er
s
.
B
atch
n
o
r
m
aliza
tio
n
is
ap
p
lied
e
x
ten
s
iv
ely
th
r
o
u
g
h
o
u
t
th
e
m
o
d
el,
e
n
h
an
ci
n
g
its
r
o
b
u
s
tn
ess
,
wh
ile
lo
s
s
co
m
p
u
tati
o
n
is
p
e
r
f
o
r
m
ed
u
s
in
g
So
f
tm
ax
[
2
7
]
.
b.
R
esNet1
8
:
A
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
t
h
at
is
1
8
lay
er
s
d
ee
p
,
R
esNet
-
1
8
h
as
b
ee
n
p
r
e
-
tr
ain
ed
o
n
o
v
e
r
a
m
illi
o
n
im
ag
es
f
r
o
m
th
e
I
m
a
g
eNe
t
d
atab
ase.
I
t
ac
h
ie
v
es
class
if
icatio
n
ac
r
o
s
s
1
,
0
0
0
ca
te
g
o
r
ies,
in
clu
d
in
g
v
ar
io
u
s
o
b
jects a
n
d
an
im
als,
u
tili
zin
g
r
esid
u
al
lear
n
in
g
to
en
h
an
ce
g
r
a
d
ien
t f
lo
w
t
h
r
o
u
g
h
it
s
lay
er
s
[
2
8
]
.
4
.
2
.
Da
t
a
s
et
W
e
ad
o
p
ted
th
e
p
u
b
lic
C
I
FAR
-
1
0
d
ataset
to
v
alid
ate
th
e
o
p
tim
izatio
n
m
eth
o
d
s
ap
p
lied
to
th
ese
b
aselin
e
m
o
d
els.
T
h
e
C
I
FAR
-
1
0
d
ataset
co
n
s
is
ts
o
f
6
0
,
0
0
0
3
2
×
3
2
co
lo
r
im
a
g
es
d
iv
i
d
e
d
in
to
1
0
m
u
tu
ally
ex
clu
s
iv
e
class
es,
ea
ch
co
n
tai
n
in
g
6
,
0
0
0
im
ag
es
[
9
]
.
T
h
e
d
ataset
is
f
u
r
th
er
s
p
lit
in
to
5
0
,
0
0
0
tr
ain
in
g
im
ag
es
an
d
1
0
,
0
0
0
test
im
ag
es.
T
h
e
d
ataset's
class
e
s
ar
e
d
esig
n
ed
to
b
e
en
tire
ly
d
is
tin
ct.
Fo
r
ex
am
p
le,
th
e
“a
u
to
m
o
b
ile”
class
in
clu
d
es
s
ed
an
s
an
d
SUVs
,
wh
ile
th
e
“t
r
u
ck
”
class
ex
clu
s
iv
ely
co
v
er
s
lar
g
e
tr
u
ck
s
,
with
n
o
o
v
er
lap
[
1
5
]
.
Fig
u
r
e
3
h
ig
h
lig
h
ts
co
m
m
o
n
lab
el
er
r
o
r
s
f
o
u
n
d
in
th
e
d
ataset,
wh
ich
c
an
ad
v
er
s
ely
af
f
ec
t
m
o
d
el
tr
ain
in
g
.
T
h
ese
er
r
o
r
s
u
n
d
er
s
co
r
e
t
h
e
im
p
o
r
tan
ce
o
f
d
ata
clea
n
in
g
to
en
s
u
r
e
a
cc
u
r
ate
an
d
r
elia
b
le
lear
n
in
g
o
u
tc
o
m
es
[
8
]
.
Ad
d
i
tio
n
ally
,
s
ev
er
al
im
ag
es
in
th
e
d
ataset
s
u
f
f
er
f
r
o
m
p
o
o
r
r
eso
lu
tio
n
,
f
u
r
th
e
r
h
am
p
er
in
g
ef
f
ec
tiv
e
lear
n
i
n
g
b
y
th
e
m
o
d
els.
L
o
w
-
r
eso
lu
tio
n
im
ag
es o
f
ten
lack
th
e
n
ec
ess
ar
y
d
etail
f
o
r
f
ea
t
u
r
e
ex
tr
ac
tio
n
,
r
ed
u
cin
g
th
e
m
o
d
e
l's
ab
ilit
y
to
d
is
tin
g
u
is
h
b
etwe
en
s
im
ilar
class
e
s
.
Fig
u
r
e
4
illu
s
tr
ates
s
u
ch
lo
w
-
r
eso
lu
tio
n
im
ag
es,
h
i
g
h
lig
h
tin
g
th
e
im
p
o
r
tan
ce
o
f
o
p
tim
izin
g
d
ataset
q
u
ality
th
r
o
u
g
h
r
eso
l
u
tio
n
en
h
an
ce
m
en
t
tech
n
iq
u
es to
ac
h
ie
v
e
b
etter
p
er
f
o
r
m
a
n
ce
[
1
1
]
.
Fig
u
r
e
3
.
E
x
am
p
les o
f
lab
el
er
r
o
r
s
am
p
les.
T
h
ese
ty
p
es o
f
er
r
o
r
s
em
p
h
asize
th
e
n
ec
ess
ity
o
f
d
ata
clea
n
in
g
to
en
s
u
r
e
ac
cu
r
ate
m
o
d
el
tr
ain
i
n
g
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
E
n
h
a
n
cin
g
tr
a
in
in
g
p
erfo
r
ma
n
ce
fo
r
s
ma
ll mo
d
els u
s
in
g
d
a
t
a
-
ce
n
tr
ic
a
p
p
r
o
a
ch
es
(
R
ed
a
A
.
E
l
-
K
h
o
r
ib
i
)
2957
Fig
u
r
e
4
.
E
x
am
p
les o
f
lo
w
-
r
es
o
lu
tio
n
im
ag
es in
t
h
e
d
ataset.
Ad
d
r
ess
in
g
th
ese
is
s
u
es th
r
o
u
g
h
d
ata
clea
n
i
n
g
an
d
en
h
an
ce
m
en
t is cr
u
cial
f
o
r
im
p
r
o
v
i
n
g
m
o
d
el
p
e
r
f
o
r
m
an
c
e
4
.
3
.
E
v
a
lua
t
i
o
n
cr
it
er
ia
T
o
ass
ess
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
els,
s
ev
er
al
s
tan
d
ar
d
class
if
icatio
n
m
etr
ics
wer
e
u
s
ed
.
T
h
ese
m
etr
ics ca
p
tu
r
e
d
if
f
er
e
n
t a
s
p
ec
ts
o
f
m
o
d
el
ac
cu
r
ac
y
,
r
o
b
u
s
tn
ess
,
an
d
r
eliab
ilit
y
,
alig
n
in
g
w
ith
b
est p
r
ac
tices in
m
ac
h
in
e
lear
n
in
g
ev
al
u
atio
n
[
2
0
]
–
[
2
8
]
.
a.
Acc
u
r
ac
y
:
T
h
is
m
etr
ic
m
ea
s
u
r
es
th
e
p
r
o
p
o
r
tio
n
o
f
co
r
r
ec
tly
class
if
ied
in
s
tan
ce
s
o
u
t
o
f
th
e
to
tal
n
u
m
b
e
r
o
f
in
s
tan
ce
s
.
I
t is f
o
r
m
ally
ex
p
r
ess
ed
as
(
1
5
)
[
1
6
]
:
=
+
+
+
+
(
1
5
)
b.
Pre
cisi
o
n
:
Pre
cisi
o
n
f
o
cu
s
es o
n
th
e
ac
cu
r
ac
y
o
f
th
e
p
o
s
itiv
e
p
r
ed
ictio
n
s
.
I
t is d
ef
i
n
ed
as
(
1
6
)
[
1
6
]
:
=
+
(
1
6
)
c.
R
ec
all:
R
ec
all
,
also
k
n
o
wn
as
s
en
s
itiv
ity
,
ev
alu
ates
th
e
m
o
d
el'
s
ab
ilit
y
to
co
r
r
ec
tly
id
en
t
if
y
all
r
elev
an
t
in
s
tan
ce
s
.
I
t is ca
lcu
lated
as
(
1
7
)
[
1
6
]
:
=
+
(
1
7
)
d.
F1
-
s
co
r
e
:
T
h
e
F1
-
s
co
r
e
co
m
b
i
n
es
p
r
ec
is
io
n
an
d
r
ec
all
in
to
a
s
in
g
le
h
ar
m
o
n
ic
m
ea
n
,
p
r
o
v
i
d
in
g
a
b
alan
ce
d
m
etr
ic
f
o
r
e
v
alu
atin
g
p
er
f
o
r
m
an
ce
o
n
d
atasets
with
class
im
b
alan
ce
s
[
1
6
]
:
1
=
2
⋅
⋅
+
(
1
8
)
e.
Sig
n
if
ican
ce
test
(t
-
test
)
:
A
t
-
test
was
em
p
lo
y
ed
to
co
m
p
ar
e
p
er
f
o
r
m
an
ce
m
etr
ics
b
e
f
o
r
e
a
n
d
af
ter
ap
p
ly
in
g
t
h
e
p
r
o
p
o
s
ed
d
ata
o
p
tim
izatio
n
tech
n
iq
u
es.
T
h
e
t
est
ass
e
s
s
e
s
wh
eth
er
th
e
o
b
s
er
v
ed
d
if
f
er
en
ce
s
ar
e
s
tatis
tically
s
ig
n
if
ican
t.
I
t i
s
ca
lcu
lated
u
s
in
g
:
,
=
−
√
2
(
1
+
1
)
,
wh
er
e
(
)
an
d
(
)
r
ep
r
esen
t
th
e
t
wo
s
y
s
tem
s
u
n
d
er
co
m
p
ar
is
o
n
,
(
)
d
en
o
tes
th
e
p
er
f
o
r
m
an
ce
m
etr
ic,
(
)
is
th
e
s
am
p
le
s
ize,
an
d
(
2
)
is
th
e
p
o
o
led
v
ar
ian
ce
[
1
3
]
–
[
1
9
]
.
f.
Stan
d
ar
d
d
e
v
iatio
n
:
T
h
is
m
et
r
ic
m
ea
s
u
r
es
th
e
v
ar
ia
b
ilit
y
o
f
th
e
p
e
r
f
o
r
m
an
ce
m
et
r
ics
ac
r
o
s
s
d
if
f
er
en
t
class
es,
h
ig
h
lig
h
tin
g
d
is
cr
ep
an
cies
o
r
in
co
n
s
is
ten
cies
in
class
if
icatio
n
r
esu
lts
.
R
ed
u
cin
g
s
tan
d
ar
d
d
ev
iatio
n
is
cr
u
cial
f
o
r
en
s
u
r
in
g
b
alan
ce
d
p
er
f
o
r
m
an
ce
ac
r
o
s
s
all
class
e
s
[
2
]
–
[
4
]
.
4
.
4
.
ResNet
1
8
perf
o
rma
nce
ev
a
lua
t
io
n
T
h
e
im
p
ac
t
o
f
d
ata
clea
n
in
g
an
d
o
p
tim
izatio
n
was
an
aly
ze
d
u
s
in
g
th
e
R
esNet1
8
an
d
I
n
ce
p
tio
n
V3
m
o
d
els as th
e
b
aselin
e
ar
ch
itectu
r
es.
Fig
u
r
e
5
p
r
o
v
id
es a
co
m
p
r
eh
en
s
iv
e
o
v
er
v
iew
o
f
th
e
p
r
o
b
lem
atic
s
am
p
les
with
in
th
e
d
ataset,
in
clu
d
i
n
g
lab
el
er
r
o
r
s
a
n
d
lo
w
-
r
eso
lu
tio
n
im
ag
es.
T
h
ese
is
s
u
es
wer
e
ad
d
r
ess
ed
th
r
o
u
g
h
r
ig
o
r
o
u
s
d
ata
p
r
e
p
r
o
ce
s
s
in
g
[
4
]
–
[
8
]
,
[
1
0
]
–
[
1
7
]
,
[
1
8
]
–
[
2
0
]
.
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
.
3
,
J
u
n
e
20
25
:
2
9
5
1
-
2
9
6
4
2958
Fig
u
r
e
5
.
E
x
am
p
les o
f
lab
el
er
r
o
r
s
an
d
l
o
w
-
r
eso
lu
tio
n
im
ag
e
s
in
th
e
d
ataset.
Ad
d
r
ess
in
g
th
ese
is
s
u
es wa
s
cr
itical
f
o
r
o
p
tim
izin
g
m
o
d
el
p
er
f
o
r
m
a
n
ce
[
4
]
–
[
8
]
,
[
1
0
]
–
[
1
7
]
,
[
1
8
]
–
[
2
1
]
T
h
e
p
er
f
o
r
m
an
ce
m
etr
ics,
in
clu
d
in
g
p
r
ec
is
io
n
,
r
ec
all,
F1
-
s
co
r
e,
an
d
ac
cu
r
ac
y
,
wer
e
m
ea
s
u
r
ed
at
d
if
f
er
en
t
s
tag
es
o
f
th
e
o
p
tim
i
za
tio
n
p
r
o
ce
s
s
.
T
ab
les
1
,
2,
3
,
an
d
4
s
u
m
m
ar
ize
th
e
r
esu
lts
f
o
r
R
esNet1
8
.
T
h
e
f
ir
s
t
co
lu
m
n
r
ep
r
esen
ts
th
e
m
o
d
el'
s
b
asel
in
e
p
er
f
o
r
m
an
ce
o
n
th
e
o
r
ig
in
al
d
ataset,
wh
ile
th
e
s
ec
o
n
d
co
lu
m
n
r
ef
lects
th
e
p
er
f
o
r
m
an
ce
af
ter
d
ata
clea
n
in
g
.
T
h
e
f
in
al
co
lu
m
n
d
em
o
n
s
tr
ates
th
e
en
h
a
n
ce
m
en
ts
ac
h
iev
ed
b
y
in
co
r
p
o
r
atin
g
g
en
er
ated
s
am
p
les u
s
in
g
a
GAN
[
2
8
]
,
[
2
9
]
.
T
ab
le
1
h
ig
h
lig
h
ts
th
e
im
p
r
o
v
em
en
ts
in
p
r
ec
is
io
n
o
b
s
er
v
ed
af
ter
ea
ch
s
tag
e
o
f
d
ataset
o
p
tim
izatio
n
.
W
h
ile
p
r
u
n
in
g
in
itially
r
ed
u
c
es
th
e
n
u
m
b
er
o
f
tr
ai
n
in
g
s
am
p
les,
it
h
elp
s
elim
in
ate
n
o
is
y
d
ata,
r
esu
ltin
g
i
n
b
etter
p
r
ec
is
io
n
f
o
r
s
o
m
e
class
es.
I
n
co
r
p
o
r
atin
g
g
en
er
ate
d
s
am
p
les f
u
r
th
er
en
h
a
n
ce
s
p
r
ec
is
io
n
,
p
ar
ticu
lar
ly
f
o
r
u
n
d
er
r
ep
r
esen
ted
o
r
m
is
class
if
ied
class
es,
s
u
ch
as “
ca
t”
an
d
“d
o
g
”
[
2
8
]
,
[
2
9
]
.
T
ab
le
1
.
Pre
cisi
o
n
co
m
p
ar
is
o
n
(
R
esNet1
8
)
.
T
h
is
tab
le
co
m
p
a
r
es th
e
p
r
ec
is
io
n
o
f
R
esNet1
8
ac
r
o
s
s
d
if
f
er
en
t
s
tag
es: b
aselin
e,
af
ter
p
r
u
n
i
n
g
,
an
d
af
ter
in
co
r
p
o
r
atin
g
g
en
e
r
ated
s
am
p
les
C
l
a
s
s
B
a
se
l
i
n
e
A
f
t
e
r
P
r
u
n
i
n
g
W
i
t
h
G
A
N
A
i
r
p
l
a
n
e
0
.
8
6
9
0
.
8
8
7
0
.
8
2
2
A
u
t
o
mo
b
i
l
e
0
.
9
2
9
0
.
9
8
4
0
.
9
1
9
B
i
r
d
0
.
9
4
2
0
.
9
5
0
0
.
8
5
6
Cat
0
.
7
8
6
0
.
6
5
9
0
.
8
4
3
D
e
e
r
0
.
8
1
9
0
.
9
5
4
0
.
8
7
7
D
o
g
0
.
8
7
0
0
.
6
2
7
0
.
8
7
6
F
r
o
g
0
.
8
7
7
0
.
8
6
1
0
.
8
9
9
H
o
r
se
0
.
9
1
7
0
.
8
4
3
0
.
9
6
3
S
h
i
p
0
.
9
2
3
0
.
9
4
1
0
.
9
7
4
Tr
u
c
k
0
.
9
7
9
0
.
7
4
7
0
.
9
5
5
A
v
e
r
a
g
e
0
.
8
9
1
0
.
8
4
5
0
.
8
9
8
S
t
d
D
e
v
0
.
0
5
9
0
.
1
2
7
0
.
0
5
3
T
ab
le
2
d
em
o
n
s
tr
ates
h
o
w
r
ec
all
im
p
r
o
v
es
af
ter
d
ata
au
g
m
en
tatio
n
with
GAN.
W
h
il
e
p
r
u
n
in
g
o
cc
asio
n
ally
r
e
d
u
ce
s
r
e
ca
ll
d
u
e
to
a
s
m
aller
d
ataset
s
ize,
th
e
ad
d
itio
n
o
f
g
en
er
ated
s
am
p
le
s
h
elp
s
r
ec
o
v
er
an
d
ev
en
e
n
h
an
ce
r
ec
all
b
y
b
alan
cin
g
class
r
ep
r
esen
tatio
n
a
n
d
r
eso
lv
in
g
an
n
o
tatio
n
in
co
n
s
is
ten
cies
[
2
1
]
–
[
3
0
]
.
T
h
is
p
r
o
ce
s
s
en
s
u
r
es
th
at
t
h
e
m
o
d
el
ca
n
b
etter
i
d
en
tif
y
r
elev
an
t
p
atter
n
s
,
p
ar
tic
u
lar
ly
in
p
r
ev
io
u
s
ly
u
n
d
er
r
ep
r
esen
ted
class
es.
T
ab
le
3
f
u
r
th
e
r
h
ig
h
lig
h
ts
th
e
b
alan
ce
d
im
p
r
o
v
em
e
n
ts
in
th
e
F1
-
s
co
r
e
ac
h
iev
ed
th
r
o
u
g
h
th
e
p
r
o
p
o
s
ed
o
p
tim
izatio
n
tech
n
iq
u
es.
T
h
e
a
d
d
itio
n
o
f
h
ig
h
-
q
u
ality
GAN
-
g
en
er
ated
s
am
p
les
m
itig
ates
th
e
tr
ad
e
-
o
f
f
b
etwe
en
p
r
ec
is
io
n
an
d
r
ec
all
b
y
p
r
o
v
id
in
g
m
o
r
e
d
iv
er
s
e
a
n
d
r
ep
r
esen
tativ
e
tr
ain
in
g
d
ata.
As
a
r
esu
lt,
th
e
m
o
d
el
d
em
o
n
s
tr
ates
co
n
s
is
ten
t
p
er
f
o
r
m
an
ce
g
ain
s
ac
r
o
s
s
class
es,
r
ein
f
o
r
ci
n
g
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
d
ata
-
ce
n
tr
ic
ap
p
r
o
ac
h
in
im
p
r
o
v
i
n
g
m
o
d
el
r
eliab
ilit
y
[
9
]
,
[
3
2
]
.
T
ab
le
4
s
u
m
m
ar
izes
th
e
ac
c
u
r
ac
y
im
p
r
o
v
em
en
ts
o
f
R
esNet1
8
ac
r
o
s
s
d
if
f
er
e
n
t
s
tag
es
o
f
d
ataset
o
p
tim
izatio
n
.
W
h
ile
p
r
u
n
in
g
r
ed
u
ce
s
ac
cu
r
ac
y
in
s
o
m
e
clas
s
es
d
u
e
to
d
ata
lo
s
s
,
th
e
i
n
co
r
p
o
r
atio
n
o
f
GAN
-
g
en
er
ated
s
am
p
les
r
esto
r
es
an
d
en
h
an
ce
s
o
v
er
all
ac
cu
r
ac
y
b
y
b
alan
cin
g
class
r
ep
r
esen
tatio
n
an
d
im
p
r
o
v
i
n
g
m
o
d
el
r
o
b
u
s
tn
ess
[
2
9
]
–
[
3
1
]
.
As
s
h
o
wn
in
Fig
u
r
e
6
,
th
e
GAN
-
g
en
er
ated
s
am
p
les
ex
h
ib
i
t
h
ig
h
f
id
elity
an
d
v
ar
iety
,
ef
f
ec
tiv
el
y
ad
d
r
ess
in
g
is
s
u
es
lik
e
clas
s
im
b
alan
ce
an
d
p
o
o
r
-
q
u
ality
o
r
i
g
in
al
d
ata.
T
h
ese
h
ig
h
-
q
u
ality
s
am
p
les
en
h
an
ce
m
o
d
el
g
e
n
e
r
aliza
tio
n
an
d
im
p
r
o
v
e
class
-
s
p
ec
if
ic
ac
cu
r
ac
y
b
y
p
r
o
v
i
d
in
g
m
o
r
e
d
iv
er
s
e
an
d
r
ep
r
esen
tativ
e
tr
ain
in
g
d
ata
[
9
]
,
[
3
2
]
.
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
E
n
h
a
n
cin
g
tr
a
in
in
g
p
erfo
r
ma
n
ce
fo
r
s
ma
ll mo
d
els u
s
in
g
d
a
t
a
-
ce
n
tr
ic
a
p
p
r
o
a
ch
es
(
R
ed
a
A
.
E
l
-
K
h
o
r
ib
i
)
2959
T
ab
le
2
.
R
ec
all
co
m
p
a
r
is
o
n
(
R
esNet1
8
)
.
T
h
is
tab
le
co
m
p
ar
es th
e
r
ec
all
o
f
R
esNet1
8
at
ea
ch
s
tag
e
o
f
o
p
tim
izatio
n
: b
aselin
e,
p
r
u
n
in
g
,
an
d
d
ata
au
g
m
en
tatio
n
with
GAN
C
l
a
s
s
B
a
se
l
i
n
e
A
f
t
e
r
P
r
u
n
i
n
g
W
i
t
h
G
A
N
A
i
r
p
l
a
n
e
0
.
9
2
9
0
.
8
4
3
0
.
9
4
5
A
u
t
o
mo
b
i
l
e
0
.
9
6
2
0
.
8
1
2
0
.
9
5
9
B
i
r
d
0
.
7
2
7
0
.
5
7
3
0
.
8
7
3
Cat
0
.
8
5
0
0
.
7
3
7
0
.
7
8
9
D
e
e
r
0
.
9
3
5
0
.
7
1
9
0
.
9
1
4
D
o
g
0
.
7
8
8
0
.
8
8
4
0
.
8
5
6
F
r
o
g
0
.
9
5
3
0
.
9
2
9
0
.
9
5
7
H
o
r
se
0
.
9
3
3
0
.
8
8
4
0
.
9
0
6
S
h
i
p
0
.
9
4
5
0
.
8
5
0
0
.
8
6
0
Tr
u
c
k
0
.
8
4
7
0
.
9
7
8
0
.
9
0
1
A
v
e
r
a
g
e
0
.
8
8
7
0
.
8
2
1
0
.
8
9
6
S
t
d
D
e
v
0
.
0
8
0
0
.
1
1
8
0
.
0
5
3
T
ab
le
3
.
F1
-
s
co
r
e
co
m
p
ar
is
o
n
(
(
R
esNet1
8
)
.
T
h
is
tab
le
illu
s
tr
ates th
e
co
m
b
in
ed
e
f
f
ec
ts
o
f
p
r
ec
is
io
n
an
d
r
ec
all
im
p
r
o
v
em
e
n
ts
o
n
th
e
F1
-
s
co
r
e
at
d
if
f
er
en
t
o
p
tim
izatio
n
s
tag
e
s
C
l
a
s
s
B
a
se
l
i
n
e
A
f
t
e
r
P
r
u
n
i
n
g
W
i
t
h
G
A
N
A
i
r
p
l
a
n
e
0
.
8
9
8
0
.
8
6
5
0
.
8
7
9
A
u
t
o
mo
b
i
l
e
0
.
9
4
5
0
.
8
9
0
.
9
3
9
B
i
r
d
0
.
8
2
1
0
.
7
1
5
0
.
8
6
4
Cat
0
.
8
1
7
0
.
6
9
6
0
.
8
1
5
D
e
e
r
0
.
8
7
3
0
.
8
2
0
.
8
9
5
D
o
g
0
.
8
2
7
0
.
7
3
4
0
.
8
6
6
F
r
o
g
0
.
9
1
3
0
.
8
9
4
0
.
9
2
7
H
o
r
se
0
.
9
2
5
0
.
8
6
3
0
.
9
3
4
S
h
i
p
0
.
9
3
4
0
.
8
9
3
0
.
9
1
3
Tr
u
c
k
0
.
9
0
8
0
.
8
4
7
0
.
9
2
7
A
v
e
r
a
g
e
0
.
8
8
6
0
.
8
2
2
0
.
8
9
6
0
.
0
4
S
t
d
D
e
v
0
.
0
4
9
0
.
0
7
7
T
ab
le
4
.
Acc
u
r
ac
y
co
m
p
ar
is
o
n
(
R
esNet1
8
)
.
T
h
is
tab
le
h
ig
h
li
g
h
ts
th
e
ac
cu
r
ac
y
o
f
R
esNet1
8
at
d
if
f
er
en
t
o
p
tim
izatio
n
s
tag
es: b
aselin
e,
af
ter
p
r
u
n
in
g
,
an
d
a
f
ter
in
co
r
p
o
r
atin
g
GAN
-
g
en
e
r
ated
s
am
p
l
es
C
l
a
s
s
B
a
se
l
i
n
e
A
f
t
e
r
P
r
u
n
i
n
g
W
i
t
h
GAN
A
i
r
p
l
a
n
e
0
.
9
0
5
0
.
8
2
4
0
.
9
6
3
A
u
t
o
mo
b
i
l
e
0
.
8
9
8
0
.
7
8
7
0
.
9
4
B
i
r
d
0
.
8
7
0
.
6
0
7
0
.
8
4
3
Cat
0
.
6
4
1
0
.
7
1
5
0
.
8
2
D
e
e
r
0
.
8
9
2
0
.
8
0
9
0
.
9
1
8
D
o
g
0
.
8
4
2
0
.
8
4
8
0
.
8
4
7
F
r
o
g
0
.
9
2
9
0
.
8
7
8
0
.
9
2
5
H
o
r
se
0
.
8
7
3
0
.
9
5
2
0
.
9
1
9
S
h
i
p
0
.
9
6
5
0
.
7
5
0
.
9
5
2
Tr
u
c
k
0
.
9
5
1
0
.
9
8
6
0
.
9
0
3
A
v
e
r
a
g
e
0
.
8
7
7
0
.
8
1
6
0
.
9
0
3
S
t
d
D
e
v
0
.
0
8
6
0
.
1
0
6
0
.
0
4
7
Fig
u
r
e
1
.
E
x
am
p
les o
f
GAN
-
g
en
er
ated
s
am
p
les u
s
ed
f
o
r
au
g
m
en
tin
g
th
e
d
ataset.
T
h
ese
h
ig
h
-
q
u
ality
s
am
p
les
en
h
an
ce
th
e
d
iv
er
s
ity
an
d
r
ep
r
esen
tatio
n
o
f
u
n
d
er
r
ep
r
esen
ted
class
es,
s
ig
n
if
ican
tly
co
n
tr
i
b
u
tin
g
to
im
p
r
o
v
ed
class
if
icatio
n
p
er
f
o
r
m
an
ce
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
.
3
,
J
u
n
e
20
25
:
2
9
5
1
-
2
9
6
4
2960
T
h
e
p
e
r
f
o
r
m
an
ce
o
f
R
esNet1
8
d
em
o
n
s
tr
ates
th
at
th
e
p
r
o
p
o
s
e
d
d
ata
-
ce
n
tr
ic
a
p
p
r
o
a
ch
,
wh
ic
h
in
v
o
lv
es
p
r
u
n
in
g
an
d
a
u
g
m
e
n
tatio
n
,
co
n
s
is
ten
tly
im
p
r
o
v
es
k
ey
e
v
alu
atio
n
m
etr
ics.
T
h
ese
f
in
d
in
g
s
v
alid
ate
th
e
ef
f
ec
tiv
en
ess
o
f
p
r
i
o
r
itizin
g
d
ataset
q
u
ality
o
v
er
m
o
d
el
co
m
p
le
x
ity
.
P
ar
ticu
la
r
ly
w
h
en
e
n
h
an
ci
n
g
t
h
e
p
er
f
o
r
m
an
ce
o
f
s
m
aller
m
o
d
els
[
2
]
.
4
.
5
.
I
ncept
io
nV3
perf
o
rma
nce
ev
a
lua
t
io
n
W
e
r
ep
licated
th
e
d
ataset
o
p
t
im
izatio
n
ex
p
er
im
en
ts
u
s
in
g
t
h
e
I
n
ce
p
tio
n
V
3
ar
ch
itectu
r
e
t
o
ev
alu
ate
th
e
g
en
er
aliza
b
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
ac
r
o
s
s
d
if
f
e
r
e
n
t
m
o
d
els.
T
a
b
les
5
to
8
p
r
esen
t
th
e
r
esu
lts
f
o
r
p
r
ec
is
io
n
,
r
ec
all,
F1
-
s
co
r
e,
a
n
d
ac
cu
r
ac
y
,
d
em
o
n
s
tr
atin
g
co
n
s
is
ten
t
p
er
f
o
r
m
an
ce
im
p
r
o
v
em
en
ts
s
im
ilar
to
th
o
s
e
o
b
s
er
v
e
d
with
R
esNet1
8
.
T
ab
le
5
in
d
icate
s
a
s
im
ilar
tr
en
d
in
p
r
ec
is
io
n
im
p
r
o
v
em
e
n
t f
o
r
I
n
ce
p
tio
n
V3
as
s
ee
n
in
R
esNet1
8
,
co
n
f
ir
m
in
g
th
e
ef
f
ec
tiv
e
n
ess
o
f
th
e
d
ata
-
c
en
tr
ic
ap
p
r
o
ac
h
ac
r
o
s
s
m
u
ltip
le
ar
ch
itectu
r
es
[
9
]
,
[
2
0
]
–
[
2
6
]
,
[
2
8
]
,
[
2
9
]
.
As
p
r
esen
ted
in
T
ab
le
6
th
e
r
ec
all
v
a
lu
es
s
h
o
w
co
n
s
is
ten
t
im
p
r
o
v
em
en
ts
with
GAN
-
au
g
m
en
ted
s
am
p
les,
p
ar
ticu
la
r
ly
f
o
r
u
n
d
er
r
e
p
r
esen
ted
class
es
lik
e
b
ir
d
an
d
ca
t.
T
h
ese
r
esu
lts
em
p
h
asize
th
e
ab
ilit
y
o
f
th
e
o
p
tim
ized
d
ataset
to
im
p
r
o
v
e
m
o
d
el
s
en
s
itiv
ity
to
r
elev
an
t in
s
tan
ce
s
[
3
0
]
,
[
3
1
]
.
T
ab
le
7
d
em
o
n
s
tr
ates
th
e
co
m
b
in
ed
im
p
r
o
v
e
m
en
ts
in
p
r
e
cisi
o
n
an
d
r
ec
all
th
r
o
u
g
h
th
e
F1
-
s
co
r
e
m
etr
ic.
GAN
-
b
ased
au
g
m
en
ta
tio
n
s
ig
n
if
ican
tly
en
h
a
n
ce
s
F1
-
s
co
r
es,
p
ar
ticu
la
r
ly
f
o
r
clas
s
es
lik
e
"d
o
g
"
a
n
d
"c
at"
th
at
p
r
ev
io
u
s
ly
s
u
f
f
er
e
d
f
r
o
m
lo
w
r
ec
all
o
r
p
r
ec
is
io
n
.
T
h
is
im
p
r
o
v
em
e
n
t
r
ein
f
o
r
ce
s
t
h
e
ef
f
ec
tiv
en
ess
o
f
b
alan
ce
d
d
ata
a
u
g
m
e
n
tatio
n
s
tr
ateg
ies in
im
p
r
o
v
in
g
m
o
d
el
p
er
f
o
r
m
a
n
ce
[
3
0
]
–
[
3
2
]
.
As
s
h
o
wn
in
T
ab
le
8
,
th
e
o
v
er
all
ac
cu
r
ac
y
o
f
I
n
ce
p
tio
n
V3
im
p
r
o
v
es
s
ig
n
if
ican
tly
af
ter
GAN
au
g
m
en
tatio
n
.
T
h
is
im
p
r
o
v
e
m
en
t
v
alid
ates
th
e
g
e
n
er
aliza
b
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
d
ata
-
c
en
tr
ic
o
p
tim
izatio
n
ap
p
r
o
ac
h
ac
r
o
s
s
d
if
f
er
en
t
m
o
d
el
ar
ch
itectu
r
es.
T
h
e
co
n
s
i
s
ten
t
ac
cu
r
ac
y
g
ain
s
d
em
o
n
s
tr
ate
th
e
m
eth
o
d
's
ef
f
ec
tiv
en
ess
b
ey
o
n
d
R
esNet1
8
,
h
ig
h
li
g
h
tin
g
its
p
o
te
n
tial f
o
r
b
r
o
ad
e
r
ap
p
licatio
n
s
in
d
ee
p
l
ea
r
n
in
g
[
2
8
]
,
[
2
9
]
.
T
ab
le
5
.
Pre
cisi
o
n
co
m
p
ar
is
o
n
(
I
n
ce
p
tio
n
V3
)
.
T
h
is
tab
le
co
m
p
ar
es th
e
p
r
ec
is
io
n
o
f
I
n
ce
p
t
io
n
V3
at
d
if
f
er
en
t stag
es: b
aselin
e,
af
ter
p
r
u
n
in
g
,
a
n
d
with
GAN
-
g
e
n
er
ated
s
am
p
les
C
l
a
s
s
B
a
se
l
i
n
e
A
f
t
e
r
P
r
u
n
i
n
g
W
i
t
h
G
A
N
A
i
r
p
l
a
n
e
0
.
9
0
9
0
.
7
7
9
0
.
8
9
1
A
u
t
o
mo
b
i
l
e
0
.
9
4
9
0
.
9
4
2
0
.
9
5
6
B
i
r
d
0
.
9
1
3
0
.
7
4
0
.
9
0
7
Cat
0
.
9
2
5
0
.
8
6
2
0
.
8
5
9
D
e
e
r
0
.
9
2
6
0
.
9
5
0
.
9
6
3
D
o
g
0
.
8
5
9
0
.
9
2
9
0
.
8
7
5
F
r
o
g
0
.
8
8
6
0
.
9
3
7
0
.
9
4
4
H
o
r
se
0
.
8
7
6
0
.
9
0
9
0
.
9
5
9
S
h
i
p
0
.
9
5
5
0
.
9
7
4
0
.
9
1
6
Tr
u
c
k
0
.
8
9
4
0
.
9
5
6
0
.
9
1
8
A
v
e
r
a
g
e
0
.
9
0
9
0
.
8
9
8
0
.
9
1
9
S
t
d
D
e
v
0
.
0
3
1
0
.
0
8
0
.
0
3
6
T
ab
le
6
.
R
ec
all
co
m
p
a
r
is
o
n
(
I
n
ce
p
tio
n
V3
)
.
T
h
is
tab
le
h
ig
h
lig
h
ts
th
e
r
ec
all
v
alu
es o
f
I
n
ce
p
tio
n
V3
at
d
if
f
er
en
t
o
p
tim
izatio
n
s
tag
e
s
: b
aselin
e,
af
ter
p
r
u
n
in
g
,
a
n
d
with
GAN
-
g
en
er
ated
s
am
p
les
C
l
a
s
s
B
a
se
l
i
n
e
A
f
t
e
r
P
r
u
n
i
n
g
W
i
t
h
G
A
N
A
i
r
p
l
a
n
e
0
.
9
3
1
0
.
9
6
3
0
.
9
3
5
A
u
t
o
mo
b
i
l
e
0
.
9
6
1
0
.
9
7
6
0
.
9
7
3
B
i
r
d
0
.
8
7
0
0
.
9
2
1
0
.
8
6
6
Cat
0
.
7
3
7
0
.
7
8
2
0
.
8
3
0
D
e
e
r
0
.
8
9
4
0
.
8
8
3
0
.
8
8
3
D
o
g
0
.
8
5
7
0
.
7
8
1
0
.
8
7
5
F
r
o
g
0
.
9
6
7
0
.
9
4
0
0
.
9
5
6
H
o
r
se
0
.
9
7
6
0
.
9
4
5
0
.
9
5
3
S
h
i
p
0
.
9
2
0
0
.
7
8
6
0
.
9
6
5
Tr
u
c
k
0
.
9
6
6
0
.
9
1
4
0
.
9
4
9
A
v
e
r
a
g
e
0
.
9
0
8
0
.
8
8
9
0
.
9
1
8
S
t
d
D
e
v
0
.
0
7
3
0
.
0
7
8
0
.
0
5
0
4
.
6
.
P
er
f
o
r
m
a
nce
a
na
ly
s
is
o
f
t
he
m
o
d
T
h
e
im
p
ac
t
o
f
d
ata
-
ce
n
tr
ic
o
p
tim
izatio
n
was
an
aly
ze
d
ac
r
o
s
s
all
clas
s
es,
f
o
cu
s
in
g
o
n
th
e
b
aselin
e,
p
r
u
n
e
d
,
an
d
GAN
-
au
g
m
e
n
ted
s
tag
es.
T
h
is
ev
alu
atio
n
h
ig
h
lig
h
ts
p
er
f
o
r
m
a
n
ce
d
if
f
er
e
n
ce
s
r
esu
ltin
g
f
r
o
m
d
ata
clea
n
in
g
a
n
d
a
u
g
m
e
n
tatio
n
.
Fig
u
r
es
3
a
n
d
5
,
alo
n
g
wit
h
tab
les
in
s
ec
tio
n
s
1
th
r
o
u
g
h
8
,
p
r
esen
t
th
e
p
er
f
o
r
m
an
ce
m
etr
ics
f
o
r
R
esNet1
8
an
d
I
n
ce
p
tio
n
V
3
.
T
h
e
r
esu
lts
d
em
o
n
s
tr
ate
th
e
co
n
s
is
ten
t
im
p
r
o
v
e
m
en
ts
ac
h
iev
ed
th
r
o
u
g
h
th
e
p
r
o
p
o
s
ed
o
p
tim
izatio
n
a
p
p
r
o
ac
h
.
Evaluation Warning : The document was created with Spire.PDF for Python.