I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
5
,
O
c
to
be
r
2025
, pp.
3634
~
3646
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
5
.pp
3634
-
3646
3634
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
C
om
p
ar
at
i
ve
an
al
ysi
s of
ge
n
d
e
r
c
l
ass
i
f
i
c
at
i
on
m
e
t
h
od
s u
s
i
n
g
c
on
vol
u
t
i
on
al
n
e
u
r
al
n
e
t
w
or
k
s
P
an
c
a D
e
w
i
P
am
u
n
gk
as
ar
i
1
, I
lh
an
A
li
m
A
s
f
an
d
im
a
2
, A
c
h
m
ad
P
r
at
am
a R
if
ai
3
, N
gu
ye
n
H
u
u
T
h
o
4
1
I
nf
or
m
a
t
i
on S
ys
t
e
m
s
S
t
udy P
r
og
r
a
m
,
F
a
c
ul
t
y of
C
om
put
e
r
I
n
f
or
m
a
t
i
on
T
e
c
hnol
ogy, U
ni
ve
r
s
i
t
a
s
N
a
s
i
on
a
l
, J
a
ka
r
t
a
, I
ndone
s
i
a
2
I
nf
or
m
a
t
i
c
s
S
t
udy P
r
og
r
a
m
, F
a
c
ul
t
y of
C
om
put
e
r
I
n
f
or
m
a
t
i
on
T
e
c
hnol
ogy, U
ni
ve
r
s
i
t
a
s
N
a
s
i
on
a
l
, J
a
ka
r
t
a
, I
ndone
s
i
a
3
D
e
pa
r
t
m
e
nt
of
M
e
c
ha
ni
c
a
l
a
nd I
ndus
t
r
i
a
l
E
ngi
ne
e
r
i
ng,
F
a
c
ul
t
y of
E
ngi
ne
e
r
i
ng,
U
ni
ve
r
s
i
t
a
s
G
a
dj
a
h
M
a
da
, Y
ogya
k
a
r
t
a
, I
ndone
s
i
a
4
F
a
c
ul
t
y of
E
ngi
ne
e
r
i
ng a
nd T
e
c
hnol
ogy, N
guye
n T
a
t
T
ha
nh U
ni
ve
r
s
i
t
y, H
o C
h
i
M
i
nh C
i
t
y, V
i
e
t
na
m
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
O
c
t
16
,
2024
R
e
vi
s
e
d
J
un
19
,
2025
A
c
c
e
pt
e
d
J
ul
10
,
2025
Gender
classification
has
become
an
important
application
in
the
fields
of
system
automation
and
artificial
intelligence,
having
important
impli
cations
across
various
fields.
The
main
challenge
in
this
classification
t
ask
is
variation
in
illumination
that
affects
the
quality
of
facial
images.
This
study
presents
method
for
identifying
genders
with
convolut
ional
neural
ne
tworks
(CNNs).
To
address
this
issue,
various
preprocessing
methods
are
a
pplied,
including
self
quotient
image
(SQI),
locally
tuned
inverse
sine
no
nline
ar
(LTISN),
histogram
equalization
(HE)
,
difference
of
gaussian
(Do
G),
and
gamma
intensity
correction
(GIC),
to
stabilize
the
effects
of
illumi
nation
variations
before
the
images
are
processed
by
CNN.
The
CNN
archi
tecture
used
consists
of
5
convolut
ional
blocks
and
2
fully
connected
blocks,
which
have
proven
effective
in
image
recognition.
The
results
of
study
sho
w
that
model traine
d with DoG me
thod achie
ved acc
uracy
of 91.07%,
makin
g it the
best
preprocessing
technique
compared
to
other
methods
such
as
S
QI
and
HE,
which
achieved
accuracy
of
90.39%
and
88.76%,
respectively.
These
findings
demonstrate
that
application
of
SQI
in
CNN
can
improve
ac
curacy
of
gender
classification
on
facial
images,
providing
better
performance
than
previous
methods
.
These
findings
are
expected
to
serve
as
foundati
on
for
further
development
s
in
fa
cial
image
cl
assificatio
n
and
its
applicati
ons
in
various fields.
K
e
y
w
o
r
d
s
:
C
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
F
a
c
e
r
e
c
ogni
ti
on
G
e
nde
r
c
la
s
s
if
ic
a
ti
on
I
m
a
ge
pr
oc
e
s
s
in
g
P
r
e
pr
oc
e
s
s
in
g m
e
th
od
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
A
c
hm
a
d P
r
a
ta
m
a
R
if
a
i
D
e
pa
r
tm
e
nt
of
M
e
c
ha
ni
c
a
l
a
nd I
ndus
tr
ia
l
E
ngi
ne
e
r
in
g, F
a
c
ul
ty
of
E
ngi
ne
e
r
in
g, U
ni
ve
r
s
it
a
s
G
a
dj
a
h M
a
da
S
t.
G
r
a
f
ik
a
N
o. 2, Yogya
ka
r
ta
55284
, I
ndone
s
ia
E
m
a
il
:
a
c
hm
a
d.p.r
if
a
i@ugm.a
c
.i
d
1.
I
N
T
R
O
D
U
C
T
I
O
N
T
he
hum
a
n
f
a
c
e
is
th
e
f
ir
s
t
in
di
c
a
to
r
us
e
d
to
id
e
nt
if
y
a
pe
r
s
o
n’
s
ge
nde
r
in
e
ve
r
yda
y
li
f
e
.
G
e
nde
r
r
e
c
ogni
ti
on
th
r
ough
f
a
c
ia
l
a
na
ly
s
is
ha
s
va
r
io
us
im
por
ta
nt
a
ppl
ic
a
ti
ons
in
th
e
f
ie
ld
of
s
e
c
ur
it
y,
s
uc
h
a
s
ve
r
if
yi
ng
ge
nde
r
dur
in
g
r
e
gi
s
tr
a
ti
on
or
s
e
c
ur
it
y
s
ur
ve
il
la
nc
e
in
s
e
ns
it
iv
e
a
r
e
a
s
[
1]
.
I
n
m
a
r
ke
ti
ng,
da
ta
a
bout
th
e
ge
nde
r
of
c
us
to
m
e
r
s
c
a
n
a
ls
o
b
e
us
e
d
to
ta
il
or
m
a
r
ke
ti
ng
o
r
c
om
m
e
r
c
ia
l
[
2]
,
s
of
twa
r
e
[
3]
,
or
s
e
r
vi
c
e
s
to
m
or
e
pr
e
c
is
e
ly
ta
r
ge
t
th
e
r
ig
ht
a
udi
e
nc
e
.
A
ddi
ti
ona
ll
y,
it
c
a
n
be
ut
il
iz
e
d
in
va
r
io
us
ty
pe
s
of
da
ta
a
na
ly
s
i
s
to
id
e
nt
if
y
us
e
f
ul
tr
e
nds
in
f
ie
ld
s
s
uc
h
a
s
s
o
c
io
lo
gy,
ps
yc
hol
ogy
[
4]
,
or
e
c
onomi
c
s
. T
he
te
c
hnol
ogy
us
e
d
f
or
th
is
ta
s
k
m
us
t
be
a
bl
e
to
c
a
pt
ur
e
a
nd
a
na
ly
z
e
f
a
c
ia
l
f
e
a
tu
r
e
s
,
w
hi
c
h
c
a
n
th
e
n
be
us
e
d
to
c
la
s
s
if
y
w
he
th
e
r
a
pe
r
s
on
is
m
a
le
or
f
e
m
a
le
.
G
e
nde
r
c
la
s
s
if
ic
a
ti
on
m
e
th
ods
e
m
pl
oy
di
s
ti
nc
t
c
ha
r
a
c
te
r
is
ti
c
s
,
s
uc
h
a
s
pa
tt
e
r
n
r
e
c
ogni
ti
on
m
e
th
ods
,
to
di
f
f
e
r
e
nt
ia
te
be
twe
e
n
m
a
s
c
ul
in
e
a
nd
f
e
m
in
in
e
a
tt
r
i
but
e
s
[
5]
.
M
e
a
nw
hi
le
,
m
e
th
ods
f
or
id
e
nt
i
f
yi
ng
ge
nde
r
c
a
n be
br
oa
dl
y c
a
te
gor
iz
e
d i
nt
o f
e
a
tu
r
e
-
ba
s
e
d
[
6]
a
nd a
p
pe
a
r
a
nc
e
-
ba
s
e
d a
ppr
oa
c
he
s
[
7]
, [
8]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
C
om
par
at
iv
e
analy
s
is
of
ge
nde
r
c
la
s
s
if
ic
at
io
n m
e
th
ods
u
s
in
g c
o
nv
ol
ut
io
nal
…
(
P
anc
a D
e
w
i
P
am
ungk
as
ar
i
)
3635
C
la
s
s
if
yi
ng
obj
e
c
t
s
ba
s
e
d
on
im
a
g
e
da
ta
c
a
n
b
e
a
c
hi
e
v
e
d
us
in
g
a
dva
nc
e
d
m
a
c
hi
ne
te
c
hnol
ogi
e
s
. T
he
te
c
hnol
ogy be
hi
nd t
he
a
bi
li
ty
of
m
a
c
hi
ne
s
t
o r
e
c
ogni
z
e
a
pe
r
s
on’
s
f
a
c
e
i
s
m
a
c
hi
ne
l
e
a
r
ni
ng. M
a
c
hi
ne
le
a
r
ni
ng
is
a
br
a
nc
h
of
a
r
ti
f
ic
ia
l
in
te
ll
ig
e
nc
e
,
s
pe
c
if
ic
a
ll
y
f
oc
us
e
d
on
how
c
om
put
e
r
s
c
a
n
le
a
r
n
f
r
om
da
ta
to
im
pr
ove
th
e
ir
in
te
ll
ig
e
nc
e
[
9]
.
M
a
c
hi
ne
le
a
r
ni
ng
pl
a
y
s
a
c
r
uc
ia
l
r
ol
e
in
c
la
s
s
if
yi
ng
a
nd
d
e
te
c
ti
ng
obj
e
c
ts
.
O
ne
of
th
e
e
vol
vi
ng
f
ie
ld
s
is
de
e
p
le
a
r
ni
ng
(
D
L
)
.
DL
is
a
b
r
a
nc
h
of
m
a
c
h
in
e
le
a
r
ni
ng
c
lo
s
e
ly
r
e
la
te
d
to
a
r
ti
f
ic
ia
l
ne
ur
a
l
ne
twor
ks
(
A
N
N
)
.
U
nl
ik
e
tr
a
di
ti
ona
l
A
N
N
s
,
D
L
in
vol
ve
s
de
e
pe
r
a
r
c
hi
te
c
tu
r
e
c
om
pos
e
d
of
m
ul
ti
pl
e
la
ye
r
s
,
e
na
bl
in
g t
he
m
ode
l
to
l
e
a
r
n hi
e
r
a
r
c
hi
c
a
l
r
e
pr
e
s
e
nt
a
ti
ons
of
da
t
a
th
r
ough s
uc
c
e
s
s
iv
e
c
onvolut
io
ns
[
10]
.
V
a
r
io
us
a
lg
or
it
hm
s
e
na
bl
e
m
a
c
hi
ne
s
to
c
la
s
s
if
y
im
a
ge
da
ta
,
w
it
h
th
e
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
C
N
N
)
be
in
g
one
of
th
e
m
os
t
w
id
e
ly
us
e
d
m
e
th
ods
[
11]
,
[
12
]
.
C
N
N
is
a
not
he
r
e
f
f
e
c
ti
ve
te
c
hni
que
in
th
e
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
pr
oc
e
s
s
.
C
N
N
is
a
ty
pe
of
DL
a
lg
or
it
hm
th
a
t
pr
oc
e
s
s
es
in
put
im
a
ge
s
by
a
ppl
yi
ng
f
il
te
r
s
or
ke
r
ne
ls
to
id
e
nt
if
y
a
nd
e
xt
r
a
c
t
im
por
ta
nt
f
e
a
tu
r
e
s
[
13]
.
T
hi
s
te
c
hni
que
c
a
n
e
f
f
e
c
ti
ve
ly
e
xt
r
a
c
t
hi
e
r
a
r
c
hi
c
a
l
f
e
a
tu
r
e
s
f
r
om
im
a
ge
s
,
e
na
bl
in
g
m
or
e
a
c
c
ur
a
te
a
nd
c
om
pl
e
x
pa
t
te
r
n
r
e
c
ogni
ti
on.
T
he
s
uc
c
e
s
s
of
C
N
N
in
ta
s
k
s
s
uc
h
a
s
obj
e
c
t
r
e
c
ogni
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
m
ot
iv
a
te
s
th
is
s
t
udy
to
us
e
C
N
N
f
or
ge
nde
r
c
la
s
s
if
ic
a
ti
on
on
f
a
c
ia
l
im
a
ge
s
.
T
he
r
e
a
r
e
m
a
ny
f
a
c
to
r
s
th
a
t
in
f
lu
e
nc
e
th
e
pr
oc
e
s
s
of
f
a
c
e
r
e
c
ogni
ti
on,
a
s
m
e
nt
io
ne
d
in
[
14]
,
w
hi
c
h
s
ta
t
e
s
t
ha
t
di
f
f
e
r
e
nc
e
s
in
il
lu
m
in
a
ti
on
i
nt
e
n
s
it
y
-
h
ig
he
r
or
lo
w
e
r
in
s
om
e
a
r
e
a
s
-
w
il
l
a
f
f
e
c
t
t
he
f
a
c
e
r
e
c
ogni
ti
on pr
o
c
e
s
s
, w
he
r
e
t
he
s
e
m
e
th
od
s
di
d n
ot
pr
ovi
d
e
good
r
e
s
ul
t
s
.
A
s
m
e
nt
io
n
e
d i
n
[
15]
,
w
hi
c
h
s
ta
t
e
s
t
ha
t
be
c
a
us
e
th
e
il
lu
m
i
na
ti
o
n
in
te
ns
it
y
of
th
e
e
n
vi
r
onm
e
nt
di
f
f
e
r
s
,
th
e
f
a
c
e
r
e
c
ogni
ti
on
r
a
t
e
i
s
of
te
n
lo
w
.
I
n
ge
n
de
r
c
la
s
s
if
i
c
a
ti
on
on
f
a
c
ia
l
im
a
g
e
s
,
im
a
g
e
pr
o
c
e
s
s
i
ng
is
c
lo
s
e
ly
r
e
la
te
d.
T
he
r
e
a
r
e
m
a
ny
m
e
th
od
s
th
a
t
c
a
n
b
e
us
e
d
to
e
nha
nc
e
im
a
g
e
s
.
R
a
g
ha
v
a
n
a
n
d
A
hm
a
di
[
16]
pr
ov
e
d
th
a
t
a
ddi
ng
pr
e
pr
oc
e
s
s
in
g
te
c
hni
qu
e
s
im
pr
ove
s
a
c
c
ur
a
c
y
in
f
a
c
e
r
e
c
ogni
ti
o
n.
T
he
be
s
t
-
p
e
r
f
or
m
in
g
t
e
c
h
ni
qu
e
in
th
e
s
tu
dy
of
R
a
gh
a
va
n
a
n
d
A
hm
a
di
[
16]
w
a
s
S
Q
I
, w
hi
c
h i
m
pr
ove
d a
c
c
ur
a
c
y by 2.6%
c
om
pa
r
e
d t
o r
e
s
ul
ts
be
f
or
e
us
in
g S
Q
I
.
T
he
ba
c
kgr
ound
of
th
is
s
tu
dy
be
gi
ns
w
it
h
th
e
unde
r
s
ta
ndi
ng
th
a
t
ge
nde
r
c
la
s
s
if
ic
a
ti
on
on
f
a
c
ia
l
im
a
ge
s
i
s
a
n i
m
por
ta
nt
t
opi
c
i
n va
r
io
us
f
ie
ld
s
, a
s
i
t
pl
a
ys
a
c
r
uc
i
a
l
r
ol
e
i
n e
nha
nc
in
g t
he
a
c
c
ur
a
c
y of
bi
om
e
tr
ic
id
e
nt
if
ic
a
ti
on,
im
pr
ovi
ng
us
e
r
-
c
e
nt
r
ic
s
e
r
vi
c
e
s
,
a
nd
e
n
a
bl
in
g
a
dva
nc
e
d
a
na
ly
ti
c
s
in
s
e
c
to
r
s
s
uc
h
a
s
s
ur
ve
il
la
nc
e
,
he
a
lt
hc
a
r
e
,
a
nd
di
gi
ta
l
m
a
r
ke
ti
ng.
M
us
ta
f
a
a
nd
M
e
e
ha
n
[
17]
s
uc
c
e
s
s
f
ul
ly
a
ppl
ie
d
C
N
N
w
it
h
a
n
a
c
c
ur
a
c
y
of
85%
.
I
n
a
not
he
r
s
tu
dy,
I
s
la
m
e
t
al
.
[
18]
de
ve
lo
pe
d
a
f
a
c
e
r
e
c
ogni
ti
on
m
ode
l
u
s
in
g
tr
a
ns
f
e
r
le
a
r
ni
ng
vi
a
P
a
r
e
to
f
r
ont
ie
r
C
N
N
.
B
e
nka
ddour
[
19]
a
ls
o
pr
opo
s
e
d
C
N
N
m
ode
ls
f
or
ge
nde
r
c
la
s
s
if
ic
a
ti
on
a
nd
a
ge
e
s
ti
m
a
ti
on.
B
e
f
or
e
C
N
N
,
s
e
ve
r
a
l
s
tu
di
e
s
pr
opos
e
d
va
r
io
us
m
e
th
ods
f
or
th
e
s
e
pr
obl
e
m
s
,
s
uc
h
a
s
lo
c
a
l
bi
na
r
y
pa
tt
e
r
ns
[
20]
,
bi
o
-
in
s
pi
r
e
d
f
e
a
tu
r
e
s
[
21]
,
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
[
22]
,
a
nd
hi
e
r
a
r
c
hi
c
a
l
c
la
s
s
ifi
e
r
[
23]
.
H
ow
e
ve
r
,
s
uc
h
m
e
th
ods
a
r
e
r
e
la
ti
ve
ly
out
of
f
a
vor
in
r
e
c
e
nt
ye
a
r
s
due
to
th
e
s
upe
r
io
r
it
y
of
C
N
N
in
im
a
ge
c
la
s
s
if
ic
a
ti
on,
a
s
in
di
c
a
te
d
by
[
24]
,
[
25]
.
T
he
r
e
f
or
e
,
th
is
s
tu
d
y
opt
s
to
de
ve
lo
p
C
N
N
m
ode
ls
w
it
h
va
r
io
us
pr
e
pr
oc
e
s
s
in
g m
e
th
ods
f
or
th
e
ge
nde
r
c
la
s
s
if
ic
a
ti
on pr
obl
e
m
.
T
he
c
ont
r
ib
ut
io
n
of
a
ppl
yi
ng
va
r
io
us
pr
e
pr
oc
e
s
s
in
g
m
e
th
ods
—
s
e
lf
quot
ie
nt
im
a
ge
(
S
Q
I
)
,
l
oc
a
ll
y
tu
ne
d
in
ve
r
s
e
s
in
e
nonl
in
e
a
r
(
L
T
I
S
N
)
,
hi
s
to
gr
a
m
e
qua
li
z
a
ti
o
n
(
H
E
)
,
di
f
f
e
r
e
nc
e
of
G
a
us
s
ia
n
(
D
oG
)
,
a
n
d
ga
m
m
a
in
te
ns
it
y
c
or
r
e
c
ti
on
(
G
I
C
)
is
to
e
nha
nc
e
th
e
qua
li
ty
a
nd
c
ons
is
te
n
c
y
of
f
a
c
ia
l
im
a
ge
s
,
w
hi
c
h
i
s
c
r
uc
ia
l
f
or
im
pr
ovi
ng
ge
nde
r
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y.
E
a
c
h
m
e
th
od
a
ddr
e
s
s
e
s
s
pe
c
if
ic
c
ha
ll
e
nge
s
in
im
a
ge
pr
e
pr
oc
e
s
s
in
g:
S
Q
I
r
e
duc
e
s
th
e
e
f
f
e
c
ts
of
va
r
yi
ng
li
ght
in
g
c
ondi
ti
ons
,
HE
e
nha
nc
e
s
im
a
ge
c
ont
r
a
s
t,
L
T
I
S
N
m
in
im
iz
e
s
non
-
uni
f
or
m
il
lu
m
in
a
ti
on,
G
I
C
a
dj
us
ts
in
te
ns
it
y
f
or
be
tt
e
r
f
e
a
tu
r
e
vi
s
ib
il
it
y,
a
nd
D
oG
e
m
pha
s
iz
e
s
e
dge
s
a
nd
te
xt
ur
e
s
by
r
e
duc
in
g
noi
s
e
w
hi
le
r
e
ta
in
in
g
e
s
s
e
nt
ia
l
de
ta
il
s
.
B
y
s
y
s
te
m
a
ti
c
a
ll
y
e
va
lu
a
ti
ng
th
e
pe
r
f
or
m
a
nc
e
of
th
e
s
e
m
e
th
ods
,
th
is
s
tu
dy
a
im
s
to
id
e
nt
if
y
th
e
m
os
t
e
f
f
e
c
ti
ve
a
ppr
oa
c
h
or
c
om
bi
na
ti
on
of
te
c
hni
que
s
th
a
t
c
a
n s
ig
ni
f
ic
a
nt
ly
im
pr
ove
th
e
r
obus
tn
e
s
s
a
nd
r
e
li
a
bi
li
ty
of
C
N
N
-
ba
s
e
d
ge
nd
e
r
c
la
s
s
if
ic
a
ti
on mode
ls
.
2.
P
R
E
P
R
O
C
E
S
S
I
N
G
M
E
T
H
O
D
P
r
e
pr
oc
e
s
s
in
g
te
c
hni
que
s
a
r
e
us
e
d
to
e
nh
a
nc
e
th
e
qua
li
ty
of
im
a
ge
s
s
o
th
a
t
th
e
in
f
or
m
a
ti
on
c
ont
a
in
e
d
in
th
e
im
a
g
e
s
i
s
e
a
s
ie
r
to
e
xt
r
a
c
t
a
nd
a
n
a
ly
z
e
.
T
he
s
e
te
c
hni
que
s
in
c
lu
de
r
e
s
iz
in
g,
nor
m
a
li
z
a
ti
on,
noi
s
e
r
e
duc
ti
on,
a
nd
c
ont
r
a
s
t
a
dj
us
tm
e
nt
,
w
hi
c
h
a
im
to
s
ta
n
da
r
di
z
e
im
a
ge
da
ta
a
nd
m
in
im
iz
e
va
r
ia
ti
ons
c
a
us
e
d
by
li
ght
in
g
c
ondi
ti
ons
,
b
a
c
kgr
ound
c
lu
tt
e
r
,
or
im
a
ge
r
e
s
ol
ut
io
n.
E
f
f
e
c
ti
ve
pr
e
pr
oc
e
s
s
in
g
is
c
r
uc
ia
l
f
or
im
pr
ovi
ng
th
e
a
c
c
ur
a
c
y
a
nd
r
obus
tn
e
s
s
of
im
a
ge
c
l
a
s
s
if
ic
a
t
io
n
a
nd
r
e
c
ogni
ti
on
s
ys
te
m
s
.
P
r
e
pr
oc
e
s
s
in
g
f
oc
us
e
s
on r
e
duc
in
g nois
e
, i
ll
um
in
a
ti
on va
r
ia
ti
ons
, a
nd l
ow
c
on
tr
a
s
t
in
i
m
a
ge
s
.
2.1. S
e
lf
q
u
ot
ie
n
t
i
m
age
S
Q
I
is
a
n
il
lu
m
in
a
ti
on
-
in
va
r
ia
nt
a
lg
or
it
h
m
de
s
ig
ne
d
to
a
ddr
e
s
s
va
r
ia
ti
ons
in
li
ght
in
g
a
nd
s
ha
dow
s
.
I
t
is
c
om
put
e
d
by
ta
ki
ng
th
e
r
a
ti
o
be
tw
e
e
n
th
e
or
ig
in
a
l
im
a
ge
in
te
ns
it
y
a
nd
a
s
m
oot
he
d
ve
r
s
io
n
of
th
e
s
a
m
e
im
a
ge
, a
s
s
how
n i
n (
1)
.
(
,
)
=
(
,
)
(
,
)
=
(
,
)
(
,
)
∗
(
,
)
(
1)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
5
,
O
c
to
be
r
20
25
:
3634
-
3646
3636
W
he
r
e
(
,
)
is
th
e
f
a
c
e
im
a
ge
a
nd
(
,
)
is
th
e
s
m
oot
he
d
ve
r
s
io
n
of
th
e
im
a
ge
,
a
nd
*
is
th
e
c
onvolut
io
n
ope
r
a
ti
on.
is
th
e
ke
r
ne
l
f
or
s
m
oot
hi
ng
th
e
im
a
ge
,
w
hi
c
h
in
th
is
c
a
s
e
is
a
G
a
us
s
ia
n
f
il
te
r
,
a
nd
is
th
e
r
e
s
ul
t
of
t
he
S
Q
I
c
a
lc
ul
a
ti
on
[
16]
.
2.2. Hi
s
t
ogr
am
e
q
u
al
iz
at
io
n
A
n
im
a
ge
hi
s
to
gr
a
m
is
a
gr
a
phi
c
a
l
r
e
pr
e
s
e
nt
a
ti
on
th
a
t
s
how
s
how
pi
xe
l
in
te
ns
it
y
va
lu
e
s
a
r
e
di
s
tr
ib
ut
e
d
a
c
r
os
s
a
n
e
nt
ir
e
im
a
ge
or
a
s
e
le
c
te
d
r
e
gi
on
of
it
.
H
E
is
a
n
im
a
ge
e
nha
nc
e
m
e
nt
m
e
th
od
w
he
r
e
th
e
pi
xe
l
hi
s
to
gr
a
m
of
th
e
im
a
ge
be
c
om
e
s
m
or
e
s
pr
e
a
d
out
a
nd
uni
f
or
m
.
S
in
c
e
th
e
hi
s
to
gr
a
m
r
e
pr
e
s
e
nt
s
th
e
pr
oba
bi
li
ty
of
pi
xe
ls
w
it
h c
e
r
ta
in
gr
a
y l
e
ve
ls
, t
he
f
or
m
ul
a
f
or
c
a
lc
ul
a
ti
ng
HE
is
us
e
d
in
(
2)
.
=
(
−
1
)
∑
(
)
=
0
(
2)
W
he
r
e
th
e
gr
a
y
le
ve
l
kkk
i
s
nor
m
a
li
z
e
d
a
g
a
in
s
t
th
e
hi
ghe
s
t
gr
a
y
le
ve
l
(
L
-
1)
.
T
he
va
lu
e
=
0
r
e
pr
e
s
e
nt
s
bl
a
c
k, a
nd
=
1
r
e
pr
e
s
e
nt
s
w
hi
te
on a
de
f
in
e
d gr
a
ys
c
a
le
[
26]
.
2.3. L
oc
al
ly
t
u
n
e
d
i
n
ve
r
s
e
s
in
e
n
on
-
li
n
e
ar
L
T
I
S
N
is
a
nonl
in
e
a
r
a
pp
r
oa
c
h
th
a
t
ope
r
a
te
s
on
e
a
c
h
pi
xe
l
.
T
he
c
or
r
e
c
te
d
in
te
ns
it
y
va
lu
e
is
c
a
lc
ul
a
te
d
by
a
ppl
yi
ng
a
n
in
ve
r
s
e
s
in
e
f
unc
ti
on
.
T
hi
s
f
unc
ti
on
us
e
s
a
dj
us
ta
bl
e
p
a
r
a
m
e
te
r
s
ba
s
e
d
on
th
e
s
ur
r
ounding pi
xe
l
va
lu
e
s
, a
s
gi
ve
n i
n (
3)
[
16]
.
ℎ
(
,
)
=
2
−
1
(
(
,
)
2
)
(
3)
2.4. Gam
m
a i
n
t
e
n
s
it
y c
or
r
e
c
t
io
n
G
r
a
y
in
te
ns
it
y
c
or
r
e
c
ti
on
(
G
I
C
)
is
a
nonl
in
e
a
r
gr
a
y
-
le
ve
l
tr
a
n
s
f
or
m
a
ti
on
te
c
hni
que
th
a
t
a
dj
us
ts
th
e
im
a
ge
'
s
gr
a
y
le
v
e
ls
by
r
e
pl
a
c
in
g
e
a
c
h
or
ig
in
a
l
gr
a
y
va
lu
e
w
it
h
a
c
or
r
e
s
ponding
G
I
C
-
a
dj
us
te
d
gr
a
y
le
ve
l,
a
s
de
s
c
r
ib
e
d i
n (
4)
.
(
,
)
=
(
,
)
1
/
(
4)
F
or
a
ga
m
m
a
v
a
lu
e
le
s
s
th
a
n
1.0,
th
e
im
a
g
e
w
il
l
b
e
c
om
e
da
r
ke
r
,
a
nd
f
or
a
g
a
m
m
a
va
lu
e
gr
e
a
te
r
th
a
n
1.0,
th
e
im
a
ge
w
il
l
be
c
om
e
br
ig
ht
e
r
. W
he
n t
he
ga
m
m
a
va
lu
e
i
s
1.0, no
e
f
f
e
c
t
is
pr
oduc
e
d
[
16]
.
2.5. Dif
f
e
r
e
n
c
e
o
f
G
au
s
s
ia
n
D
oG
is
a
g
r
a
ys
c
a
le
im
a
ge
e
nha
nc
e
m
e
nt
a
lg
or
it
hm
th
a
t
in
vol
ve
s
s
ubt
r
a
c
ti
ng
th
e
s
m
oot
he
d
ve
r
s
io
n
of
th
e
or
ig
in
a
l
im
a
ge
f
r
om
a
not
he
r
ve
r
s
io
n
of
th
e
or
ig
in
a
l
im
a
ge
t
ha
t
is
not
a
s
s
m
oot
hl
y
f
il
te
r
e
d.
T
he
s
m
oot
he
d
im
a
ge
s
a
r
e
obt
a
in
e
d
by
c
onvolvi
ng
th
e
gr
a
ys
c
a
le
im
a
ge
w
it
h
a
G
a
us
s
ia
n
f
il
te
r
ke
r
ne
l
w
it
h
di
f
f
e
r
e
nt
s
ta
nda
r
d
de
vi
a
ti
ons
, a
s
gi
ve
n i
n (
5)
[
16]
.
(
,
)
=
1
2
1
2
−
2
+
2
2
1
2
−
1
2
2
2
−
2
+
2
2
2
2
(
5)
w
he
r
e
1
a
nd
2
th
e
s
e
a
r
e
t
he
w
id
th
s
of
t
he
G
a
us
s
ia
n f
il
te
r
ke
r
ne
l.
3.
S
Y
S
T
E
M
D
E
S
I
G
N
T
hi
s
s
ys
te
m
c
ons
i
s
ts
of
s
e
ve
r
a
l
s
ta
ge
s
to
c
l
a
s
s
if
y
g
e
nde
r
us
in
g
th
e
C
N
N
m
e
th
od:
c
ol
le
c
ti
ng
d
a
ta
s
e
t
s
f
r
om
r
e
li
a
bl
e
s
our
c
e
s
,
pr
e
pr
oc
e
s
s
in
g
th
e
da
ta
s
e
ts
to
im
pr
ove
i
m
a
ge
qua
li
ty
a
nd
c
ons
is
te
nc
y,
a
nd
tr
a
in
in
g
th
e
da
ta
us
in
g t
he
C
N
N
m
od
e
l
to
r
e
c
ogni
z
e
a
nd pr
e
di
c
t
ge
nd
e
r
w
it
h hi
gh a
c
c
ur
a
c
y.
3.1. Dat
as
e
t
T
hi
s
s
tu
dy
u
s
e
s
a
d
a
ta
s
e
t
c
on
s
is
ti
ng
of
f
a
c
ia
l
im
a
ge
s
of
bot
h
m
e
n
a
nd
w
om
e
n.
T
he
r
e
s
e
a
r
c
h
ut
il
iz
e
s
two
ty
pe
s
of
da
ta
:
s
e
c
onda
r
y
da
ta
obt
a
in
e
d
f
r
om
K
a
ggl
e
a
nd
pr
im
a
r
y
da
ta
,
w
hi
c
h
w
a
s
pe
r
s
ona
ll
y
r
e
que
s
te
d
f
r
om
t
he
s
our
c
e
s
w
it
h pe
r
m
is
s
io
n. T
he
da
ta
s
e
t
w
il
l
be
di
vi
de
d i
nt
o t
hr
e
e
pa
r
ts
:
K
a
ggl
e
t
r
a
in
in
g da
ta
f
o
r
m
ode
l
le
a
r
ni
ng,
K
a
ggl
e
te
s
t
da
ta
f
or
in
it
ia
l
pe
r
f
or
m
a
nc
e
e
va
lu
a
ti
on,
a
nd
pr
im
a
r
y
te
s
t
da
ta
f
or
f
in
a
l
va
li
da
ti
on
to
a
s
s
e
s
s
t
he
m
ode
l'
s
a
c
c
ur
a
c
y a
nd r
obus
tn
e
s
s
on e
xt
e
r
na
l
da
ta
.
T
he
da
ta
s
e
t
c
ons
is
t
s
of
bot
h t
r
a
in
in
g a
nd
te
s
t
da
ta
s
our
c
e
d f
r
om
K
a
ggl
e
a
nd pr
im
a
r
y da
ta
s
pe
c
if
ic
a
ll
y
c
ol
le
c
te
d
f
or
th
is
r
e
s
e
a
r
c
h
th
a
t
a
s
s
e
e
n
in
F
ig
ur
e
1.
T
he
tr
a
in
in
g
da
ta
s
e
t
in
c
lu
de
s
10,582
im
a
ge
s
-
5,291
m
a
le
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
C
om
par
at
iv
e
analy
s
is
of
ge
nde
r
c
la
s
s
if
ic
at
io
n m
e
th
ods
u
s
in
g c
o
nv
ol
ut
io
nal
…
(
P
anc
a D
e
w
i
P
am
ungk
as
ar
i
)
3637
a
nd
5,291
f
e
m
a
le
-
f
e
a
tu
r
in
g
f
a
c
ia
l
im
a
ge
s
th
a
t
a
r
e
not
in
f
lu
e
nc
e
d
by
il
lu
m
in
a
ti
on
c
ondi
ti
ons
.
T
he
te
s
t
da
ta
s
e
t
c
ont
a
in
s
2,646
im
a
ge
s
-
1,323
m
a
le
a
nd
1,323
f
e
m
a
l
e
-
w
it
h
f
a
c
ia
l
im
a
ge
s
a
f
f
e
c
te
d
by
il
lu
m
in
a
ti
on
f
r
om
va
r
io
us
a
ngl
e
s
.
W
he
r
e
a
s
th
e
pr
im
a
r
y
da
ta
s
e
t
a
l
s
o
in
c
lu
de
s
95
im
a
ge
s
-
4
5
m
a
le
a
nd
50
f
e
m
a
le
-
de
pi
c
ti
ng
f
a
c
i
a
l
im
a
ge
s
unde
r
di
f
f
e
r
e
nt
i
ll
um
in
a
ti
on c
ondi
ti
ons
f
r
om
m
ul
ti
pl
e
di
r
e
c
ti
ons
.
F
ig
ur
e
1. E
xa
m
pl
e
of
t
r
a
in
in
g da
ta
T
he
t
e
s
t
da
ta
f
e
a
tu
r
e
s
va
r
yi
ng i
ll
um
in
a
ti
on c
ondi
ti
ons
w
it
h
th
e
goa
l
of
e
va
lu
a
ti
ng t
he
pe
r
f
or
m
a
nc
e
of
th
e
pr
e
pr
oc
e
s
s
in
g
m
e
th
ods
.
I
t
a
ll
ow
s
us
to
de
te
r
m
in
e
w
hi
c
h
m
e
th
od
is
m
os
t
e
f
f
e
c
ti
ve
in
ha
ndl
in
g
di
f
f
e
r
e
nt
il
lu
m
in
a
ti
on
c
ondi
ti
ons
.
T
he
di
s
tr
ib
ut
io
n
of
f
a
c
ia
l
im
a
ge
s
in
th
e
da
ta
s
e
t
a
c
r
os
s
di
f
f
e
r
e
nt
c
a
te
gor
ie
s
is
di
vi
de
d
in
to
th
r
e
e
pa
r
ts
:
tr
a
in
in
g
da
ta
,
te
s
t
da
ta
,
a
nd
pr
im
a
r
y
da
ta
,
a
s
s
e
e
n
in
T
a
bl
e
1.
T
he
te
s
t
da
ta
f
e
a
tu
r
e
s
va
r
yi
ng
il
lu
m
in
a
ti
on
c
ondi
ti
ons
w
it
h
th
e
go
a
l
of
e
va
lu
a
ti
ng
th
e
pe
r
f
or
m
a
nc
e
of
th
e
pr
e
pr
oc
e
s
s
in
g
m
e
th
ods
.
I
t
a
ll
ow
s
us
t
o de
te
r
m
in
e
w
hi
c
h m
e
th
od i
s
m
os
t
e
f
f
e
c
ti
ve
i
n ha
ndl
in
g di
f
f
e
r
e
nt
i
ll
um
in
a
ti
on c
ondi
ti
ons
. T
he
di
s
tr
ib
ut
io
n
of
f
a
c
ia
l
im
a
ge
s
in
th
e
da
ta
s
e
t
a
c
r
os
s
di
f
f
e
r
e
nt
c
a
te
gor
ie
s
is
di
vi
de
d
in
to
th
r
e
e
pa
r
ts
:
5
,
291
im
a
ge
s
of
th
e
tr
a
in
in
g
da
ta
,
1
,
323
im
a
ge
s
of
th
e
te
s
t
da
ta
,
a
nd
t
he
pr
im
a
r
y
da
ta
c
ons
i
s
ts
of
50
im
a
ge
s
a
nd
45
im
a
ge
s
of
th
e
f
e
m
a
le
s
pr
im
a
r
y
a
nd
m
a
le
s
,
r
e
s
pe
c
ti
ve
ly
.
A
ll
m
ode
ls
a
r
e
tr
a
in
e
d
a
nd
e
va
lu
a
te
d
us
in
g
th
e
s
a
m
e
tr
a
in
in
g,
te
s
ti
ng,
a
nd
pr
im
a
r
y
da
ta
to
e
ns
ur
e
f
a
ir
c
om
pa
r
is
on
of
th
e
m
od
e
ls
’
pe
r
f
or
m
a
nc
e
.
I
n
a
ddi
ti
on,
he
r
e
,
th
e
te
s
ti
ng
a
nd
pr
im
a
r
y
da
ta
f
or
e
va
lu
a
ti
on
a
r
e
dr
a
w
n
f
r
om
di
f
f
e
r
e
nt
s
ubj
e
c
t
s
to
a
s
s
e
s
s
th
e
g
e
ne
r
a
li
z
a
ti
on
of
th
e
tr
a
in
e
d
m
ode
ls
i
n r
e
c
ogni
z
in
g ne
w
da
ta
.
A
f
t
e
r
w
a
r
d
, da
t
a
a
ugm
e
n
ta
ti
o
n
i
s
p
e
r
f
or
m
e
d
t
o a
r
t
if
i
c
i
a
ll
y
i
nc
r
e
a
s
e
t
he
s
i
z
e
a
nd
di
ve
r
s
it
y
of
a
d
a
t
a
s
e
t
b
y
a
pp
ly
i
ng
va
r
io
us
tr
a
n
s
f
or
m
a
t
io
ns
to
t
h
e
e
xi
s
ti
ng
da
ta
.
D
a
t
a
a
ug
m
e
n
ta
ti
o
n
i
s
e
s
pe
c
i
a
l
ly
v
a
l
ua
bl
e
w
he
n
t
he
r
e
i
s
li
m
it
e
d
d
a
t
a
a
va
il
a
bl
e
.
G
e
n
e
r
a
ti
ng
a
ugm
e
n
te
d
s
a
m
p
le
s
h
e
lp
s
c
r
e
a
t
e
a
la
r
ge
r
a
nd
m
or
e
d
iv
e
r
s
e
d
a
t
a
s
e
t
,
w
h
ic
h
i
s
c
r
it
ic
a
l
f
or
tr
a
i
ni
n
g
DL
m
o
de
ls
t
ha
t
ty
p
ic
a
l
ly
r
e
qui
r
e
l
a
r
g
e
d
a
t
a
s
e
t
s
.
D
a
ta
a
ug
m
e
nt
a
ti
on
th
a
t
i
m
pl
e
m
e
nt
e
d
ge
o
m
e
tr
i
c
tr
a
n
s
f
or
m
a
t
io
n
s
(
r
ot
a
ti
o
n,
f
l
ip
p
in
g,
s
c
a
li
ng
,
c
r
o
ppi
n
g)
,
a
nd
p
hot
om
e
tr
i
c
a
dj
u
s
t
m
e
nt
s
(
br
i
ght
n
e
s
s
,
c
on
tr
a
s
t,
no
i
s
e
a
ddi
ti
o
n)
.
T
h
e
s
e
t
e
c
hn
iq
u
e
s
e
nh
a
n
c
e
v
a
r
i
a
ti
on,
r
e
d
uc
i
ng
o
ve
r
f
it
t
in
g
a
nd
i
m
pr
ovi
ng
ge
n
e
r
a
li
z
a
ti
on.
D
a
t
a
di
s
tr
ib
u
ti
o
n
i
s
c
a
r
e
f
ul
ly
m
a
in
t
a
i
ne
d
t
o
p
r
e
v
e
nt
c
la
s
s
im
b
a
l
a
n
c
e
;
ty
p
ic
a
l
ly
,
e
a
c
h
c
l
a
s
s
r
e
c
e
i
ve
s
a
n
e
q
ua
l
pr
opo
r
ti
on of
a
ugm
e
n
te
d s
a
m
pl
e
s
. E
n
s
ur
in
g
a
d
iv
e
r
s
e
y
e
t
ba
la
nc
e
d
d
a
ta
s
e
t
a
ll
ow
s
th
e
m
o
de
l
to
le
a
r
n
in
v
a
r
i
a
n
t
f
a
c
ia
l
f
e
a
t
ur
e
s
,
e
nh
a
nc
i
ng
r
e
c
o
gni
ti
o
n
a
c
c
ur
a
c
y
a
c
r
o
s
s
d
if
f
e
r
e
nt
l
ig
ht
i
ng
c
o
ndi
ti
o
ns
,
a
n
gl
e
s
,
a
nd
oc
c
lu
s
io
n
s
i
n
r
e
a
l
-
w
or
ld
a
p
pl
i
c
a
ti
on
s
.
T
h
e
num
be
r
of
i
m
a
g
e
s
a
f
t
e
r
d
a
ta
a
ug
m
e
nt
a
ti
on
is
pr
e
s
e
nt
e
d
i
n
T
a
bl
e
2
.
T
a
bl
e
1. D
a
ta
di
s
tr
ib
ut
io
n
C
a
t
e
gor
y
T
r
a
i
ni
ng da
t
a
T
e
s
t
da
t
a
P
r
i
m
a
r
y da
t
a
F
e
m
a
l
e
5,291
1,323
50
M
a
l
e
5,291
1,323
45
T
ot
a
l
10,582
2,646
90
T
a
bl
e
2. D
a
ta
di
s
tr
ib
ut
io
n a
f
te
r
a
ugm
e
nt
a
ti
on
C
a
t
e
gor
y
T
r
a
i
ni
ng
da
t
a
T
e
s
t
da
t
a
P
r
i
m
a
r
y
da
t
a
F
e
m
a
l
e
8
,
142
1
,
985
835
M
a
l
e
8
,
142
1
,
985
825
T
ot
a
l
16
,
284
3
,
970
1
,
660
3.2. Ap
p
ly
p
r
e
p
r
oc
e
s
s
in
g m
e
t
h
od
s
t
o i
m
age
s
I
n
th
e
pr
e
p
r
oc
e
s
s
in
g
s
ta
ge
,
im
a
ge
s
a
r
e
r
e
s
iz
e
d
to
96
×
9
6
pi
xe
ls
to
e
ns
u
r
e
un
if
or
m
in
p
ut
di
m
e
ns
i
ons
,
w
h
ic
h
is
c
r
u
c
ia
l
f
o
r
c
o
ns
is
te
n
t
m
ode
l
pe
r
f
o
r
m
a
nc
e
.
A
f
te
r
r
e
s
iz
i
ng,
no
r
m
a
l
iz
a
ti
o
n
is
a
p
pl
ie
d
to
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
5
,
O
c
to
be
r
20
25
:
3634
-
3646
3638
s
c
a
le
th
e
p
ix
e
l
v
a
l
ue
s
to
a
[
0,
1
]
r
a
n
ge
,
s
ta
nda
r
d
iz
in
g
t
he
da
t
a
a
n
d
e
nha
nc
i
ng
th
e
le
a
r
ni
ng
p
r
oc
e
s
s
.
I
l
lu
m
i
na
t
io
n
c
o
nd
it
io
ns
c
a
n
s
ig
n
i
f
ic
a
nt
ly
i
m
pa
c
t
im
a
ge
qua
li
t
y,
s
o
ha
nd
li
ng
il
lu
m
i
na
ti
o
n
e
f
f
e
c
ti
ve
l
y
d
ur
in
g
pr
e
pr
oc
e
s
s
in
g
is
vi
ta
l.
V
a
r
io
us
m
e
t
hods
,
i
nc
lu
d
in
g
g
r
a
ys
c
a
l
e
c
onve
r
s
io
n
,
S
Q
I
,
H
E
,
L
T
S
I
N
,
G
I
C
,
a
nd
D
o
G
,
a
r
e
e
m
p
lo
ye
d
to
a
dd
r
e
s
s
th
e
s
e
va
r
ia
ti
ons
.
T
he
e
f
f
e
c
ts
o
f
t
he
s
e
i
ll
um
in
a
ti
on
a
dj
us
t
m
e
n
t
t
e
c
h
ni
que
s
on
pr
e
pr
oc
e
s
s
in
g
a
r
e
il
lu
s
t
r
a
te
d
in
F
ig
ur
e
2
,
w
he
r
e
di
f
f
e
r
e
n
t
m
e
t
hods
de
m
o
ns
t
r
a
te
th
e
i
r
e
f
f
e
c
t
iv
e
ne
s
s
in
c
r
e
a
ti
ng
m
o
r
e
c
o
ns
is
te
n
t
a
nd
r
e
l
ia
b
le
da
t
a
f
o
r
m
o
de
l
t
r
a
in
i
ng.
T
he
S
Q
I
a
n
d
D
o
G
m
e
th
ods
e
f
f
e
c
ti
ve
ly
r
e
d
uc
e
il
lu
m
in
a
t
io
n
c
on
di
ti
ons
.
H
ow
e
ve
r
,
th
e
D
o
G
m
e
t
ho
d
p
r
i
m
a
r
i
ly
s
h
a
r
pe
ns
e
dge
s
,
w
h
ic
h
c
a
n
le
a
d
to
a
lo
s
s
o
f
in
f
o
r
m
a
t
io
n
.
H
E
a
nd
G
I
C
a
ls
o
he
lp
in
m
i
ni
m
iz
in
g
i
ll
u
m
in
a
t
io
n
va
r
ia
ti
ons
,
b
ut
bot
h
m
e
t
hods
te
nd
to
da
r
ke
n
a
r
e
a
s
th
a
t
a
r
e
no
t
a
f
f
e
c
te
d
by
l
ig
ht
.
F
ig
ur
e
2. C
om
pa
r
is
on of
pr
e
pr
oc
e
s
s
in
g
m
e
th
ods
3.3. Con
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
or
k
s
A
C
N
N
is
a
ty
pe
of
ANN
de
s
ig
n
e
d
f
or
im
a
ge
r
e
c
ogni
ti
on.
C
N
N
s
a
ut
om
a
ti
c
a
ll
y
le
a
r
n
a
nd
e
xt
r
a
c
t
f
e
a
tu
r
e
s
f
r
om
in
put
im
a
ge
s
a
nd
in
te
gr
a
te
th
e
s
e
f
e
a
tu
r
e
s
w
it
h
a
c
la
s
s
if
ic
a
ti
on
m
e
c
ha
ni
s
m
.
O
ne
of
th
e
k
e
y
a
dva
nt
a
ge
s
of
C
N
N
c
la
s
s
if
ie
r
s
is
th
e
ir
r
e
la
ti
ve
ly
s
im
pl
e
a
r
c
hi
te
c
tu
r
e
c
om
pa
r
e
d
to
ot
he
r
m
e
th
ods
,
w
it
h
a
c
le
a
r
s
e
que
nc
e
of
la
ye
r
s
th
a
t
tr
a
ns
f
or
m
in
put
da
ta
in
to
out
put
pr
e
di
c
ti
ons
[
27]
.
T
ypi
c
a
ll
y,
C
N
N
s
ope
r
a
te
in
two
m
a
in
s
ta
ge
s
:
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on.
T
he
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
pha
s
e
de
r
iv
e
s
m
e
a
ni
ngf
ul
f
e
a
tu
r
e
s
f
r
om
th
e
or
ig
in
a
l
da
ta
s
e
t,
w
hi
c
h
he
lp
s
r
e
duc
e
c
om
put
a
ti
ona
l
r
e
s
our
c
e
s
w
hi
le
pr
e
s
e
r
vi
ng
c
r
it
ic
a
l
in
f
or
m
a
ti
on
[
28]
.
I
n
th
e
c
la
s
s
if
ic
a
ti
on
pha
s
e
,
f
ul
ly
c
onne
c
t
e
d
la
ye
r
s
a
c
t
a
s
t
he
c
la
s
s
if
ie
r
,
a
s
s
ig
ni
ng
pr
oba
bi
li
ti
e
s
to
pr
e
di
c
t
th
e
obj
e
c
t
pr
e
s
e
nt
i
n t
he
i
m
a
ge
ba
s
e
d on the
e
xt
r
a
c
te
d f
e
a
tu
r
e
s
[
29]
.
T
he
C
N
N
m
ode
l
a
r
c
hi
te
c
tu
r
e
,
c
ons
is
t
s
of
5
c
onvolut
io
na
l
bl
oc
ks
a
nd
2
f
ul
ly
c
onne
c
te
d
la
ye
r
s
.
T
he
in
put
la
ye
r
ha
s
a
s
iz
e
of
96
×
96
pi
xe
ls
w
it
h
a
c
ol
or
c
ha
nne
l
o
f
1,
r
e
s
ul
ti
ng
in
a
n
in
pu
t
la
ye
r
c
om
pos
it
io
n
of
(
96,
96,
1)
.
T
he
im
a
ge
th
e
n
unde
r
goe
s
f
e
a
tu
r
e
le
a
r
ni
ng
th
r
ough
c
onvolut
io
na
l
ope
r
a
ti
ons
pe
r
f
or
m
e
d
by
th
e
5
c
onvolut
io
na
l
bl
oc
ks
.
S
ubs
e
que
nt
ly
,
th
e
im
a
ge
w
il
l
be
f
la
tt
e
n
e
d,
c
onve
r
ti
ng
th
e
r
e
s
ul
ts
of
f
e
a
tu
r
e
le
a
r
ni
ng
in
to
a
ve
c
to
r
.
T
he
f
in
a
l
s
ta
ge
in
vol
ve
s
c
la
s
s
if
ic
a
ti
on
us
in
g
th
e
f
ul
ly
c
onne
c
te
d
la
ye
r
s
.
F
ig
ur
e
3
il
lu
s
tr
a
te
s
th
e
m
ode
l
a
r
c
hi
te
c
tu
r
e
us
e
d i
n t
hi
s
s
tu
dy.
A
s
s
how
n
in
F
ig
ur
e
3,
th
e
r
e
a
r
e
5
c
onvol
ut
io
n
a
l
bl
oc
ks
a
nd
2
f
ul
ly
c
o
nne
c
te
d
b
lo
c
k
s
.
T
he
c
onvol
ut
io
n
a
l
l
a
ye
r
i
s
t
he
e
a
r
li
e
s
t
la
y
e
r
i
n a
C
N
N
. T
h
e
pa
r
a
m
e
te
r
s
in
th
e
c
on
vol
ut
io
na
l
la
ye
r
a
r
e
d
e
te
r
m
in
e
d
by
th
e
num
be
r
of
ke
r
ne
ls
us
e
d
in
t
he
c
onv
ol
ut
io
n
a
l
o
pe
r
a
ti
on. T
h
e
ope
r
a
ti
on
i
s
pe
r
f
or
m
e
d
on
th
e
in
put
to
pr
od
uc
e
th
e
ne
ur
on
s
out
p
ut
[
t
o
e
nha
nc
e
-
30]
.
T
hi
s
i
s
f
ol
lo
w
e
d
b
y
poo
li
ng
to
e
xt
r
a
c
t
f
e
a
tu
r
e
s
f
r
om
th
e
in
p
ut
im
a
ge
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
C
om
par
at
iv
e
analy
s
is
of
ge
nde
r
c
la
s
s
if
ic
at
io
n m
e
th
ods
u
s
in
g c
o
nv
ol
ut
io
nal
…
(
P
anc
a D
e
w
i
P
am
ungk
as
ar
i
)
3639
s
e
qu
e
nt
i
a
ll
y.
T
he
p
ool
in
g
la
y
e
r
,
a
ls
o
kno
w
n
a
s
s
ub
s
a
m
pl
in
g
or
dow
n
s
a
m
pl
i
ng,
r
e
duc
e
s
th
e
s
pa
ti
a
l
di
m
e
n
s
io
n
s
of
e
a
c
h
f
e
a
tu
r
e
m
a
p w
hi
le
pr
e
s
e
r
vi
n
g
th
e
m
o
s
t
im
p
or
ta
nt
in
f
or
m
a
ti
on
[
30]
–
[
3
3]
.
T
he
pur
p
os
e
of
us
in
g
pool
in
g
is
to
r
e
du
c
e
th
e
num
be
r
of
f
e
a
tu
r
e
s
f
r
om
th
e
c
on
vol
ut
io
na
l
la
y
e
r
’
s
out
p
ut
or
f
e
a
t
ur
e
m
a
p,
th
e
r
e
by
s
p
e
e
di
ng
u
p
th
e
m
ode
l’
s
tr
a
in
i
ng
pr
oc
e
s
s
.
A
f
te
r
p
a
s
s
in
g
th
r
o
ugh
a
ll
th
e
la
y
e
r
s
,
th
e
da
t
a
is
f
or
w
a
r
de
d
to
th
e
f
ul
ly
c
onne
c
te
d/
de
ns
e
l
a
ye
r
.
I
n
th
i
s
l
a
y
e
r
,
e
a
c
h
ne
ur
on
i
s
l
in
ke
d
to
e
v
e
r
y
a
c
ti
v
a
ti
on
f
r
om
th
e
pr
e
c
e
di
ng
la
y
e
r
.
W
h
e
n
th
e
s
e
f
ul
ly
c
onn
e
c
t
e
d l
a
ye
r
s
a
r
e
c
om
bi
ne
d
w
it
h a
S
of
t
M
a
x
f
unc
ti
on, t
he
y f
or
m
a
m
ul
t
i
-
la
y
e
r
pe
r
c
e
pt
r
on (
M
L
P
)
,
w
hi
c
h
s
e
r
ve
s
a
s
t
h
e
c
la
s
s
if
i
e
r
i
n
th
e
ne
tw
or
k
[
14]
.
F
ig
ur
e
3. C
N
N
m
ode
l
a
r
c
hi
te
c
tu
r
e
I
n
th
e
C
N
N
a
r
c
hi
te
c
tu
r
e
,
e
a
c
h
la
y
e
r
in
c
lu
de
s
th
e
s
c
a
le
d
e
xpo
ne
nt
ia
l
li
ne
a
r
uni
t
(
S
e
L
U
)
a
c
ti
va
ti
on
f
unc
ti
on
to
e
na
bl
e
hi
gh
-
le
ve
l
a
bs
tr
a
c
t
r
e
pr
e
s
e
nt
a
ti
ons
.
S
e
L
U
ha
s
th
e
pr
ope
r
ty
of
s
e
lf
-
nor
m
a
li
z
a
ti
on,
w
hi
c
h
a
ut
om
a
ti
c
a
ll
y
dr
iv
e
s
th
e
out
put
to
w
a
r
d
a
m
e
a
n
of
z
e
r
o
a
nd
va
r
yi
ng
uni
ts
[
34]
–
[
36
]
.
T
hi
s
he
lp
s
m
a
in
ta
in
gr
a
di
e
nt
s
ta
bi
li
ty
dur
in
g t
r
a
in
in
g a
nd a
c
c
e
le
r
a
te
s
c
onv
e
r
ge
nc
e
.
T
he
S
e
L
U
e
qua
ti
on c
a
n be
s
e
e
n i
n (
6)
.
(
)
=
{
≥
0
(
−
1
)
<
0
(
6)
A
ddi
ti
ona
ll
y,
e
a
c
h
c
onvolut
io
na
l
la
ye
r
in
c
or
por
a
te
s
ba
tc
h
nor
m
a
li
z
a
ti
on
a
nd
dr
opout.
B
a
tc
h
nor
m
a
li
z
a
ti
on
he
lp
s
m
in
im
iz
e
in
te
r
na
l
c
ova
r
ia
te
s
hi
f
t
by
nor
m
a
li
z
in
g
th
e
out
put
s
of
th
e
pr
e
vi
ous
a
c
ti
va
ti
on
la
ye
r
—
th
is
is
done
by
s
ubt
r
a
c
ti
ng
th
e
ba
tc
h
m
e
a
n
a
nd
di
vi
di
ng
by
th
e
ba
tc
h
s
ta
nda
r
d
de
vi
a
ti
on,
w
hi
c
h
im
pr
ove
s
th
e
ne
twor
k’
s
s
ta
bi
li
ty
a
nd s
pe
e
d
s
up
tr
a
in
in
g
[
37]
.
T
he
dr
opout
te
c
hni
que
is
in
tr
oduc
e
d
to
m
it
ig
a
te
ove
r
f
it
ti
ng
by
te
m
por
a
r
il
y
de
a
c
ti
va
ti
ng
a
s
ubs
e
t
of
ne
ur
ons
du
r
in
g
tr
a
in
in
g,
ba
s
e
d
on
a
s
pe
c
if
ie
d
pr
oba
bi
li
ty
(
ℎ
0
<
<
1
)
[
38]
.
T
he
e
xpe
r
im
e
nt
a
l
c
ondi
ti
ons
a
r
e
out
li
ne
d
in
T
a
bl
e
3.
T
he
ta
bl
e
s
how
s
th
a
t
th
e
ba
tc
h
nor
m
a
li
z
a
ti
on
s
iz
e
i
s
s
e
t
to
64,
e
ns
ur
in
g
s
ta
bl
e
tr
a
in
in
g.
O
pt
im
iz
a
ti
on
i
s
pe
r
f
or
m
e
d
us
in
g
th
e
A
da
m
opt
im
iz
e
r
w
it
h
a
le
a
r
ni
ng
r
a
te
of
0.001,
pr
ovi
di
ng
e
f
f
ic
ie
nt
c
onve
r
ge
nc
e
.
T
he
out
put
la
ye
r
e
m
pl
oys
th
e
S
of
tM
a
x
f
unc
ti
on,
s
ui
ta
bl
e
f
or
c
la
s
s
if
ic
a
ti
on
ta
s
ks
,
w
hi
le
th
e
hi
dde
n
l
a
ye
r
s
u
s
e
S
e
L
U
f
or
s
e
lf
-
nor
m
a
li
z
a
ti
on.
A
ddi
ti
ona
ll
y,
2×
2
pool
in
g
is
a
ppl
ie
d
to
r
e
duc
e
s
pa
ti
a
l
di
m
e
ns
io
ns
,
a
nd
in
pu
t
da
ta
is
r
e
s
c
a
le
d
by
1/
255,
no
r
m
a
li
z
in
g
pi
xe
l
va
lu
e
s
t
o a
[
0,
1]
r
a
nge
f
or
c
ons
is
te
nt
a
nd e
f
f
e
c
ti
ve
t
r
a
in
in
g.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
5
,
O
c
to
be
r
20
25
:
3634
-
3646
3640
T
a
bl
e
3. M
od
e
l
s
pe
c
if
ic
a
ti
on
s
S
pe
c
i
f
i
c
a
t
i
on
V
a
l
ue
O
pt
i
m
i
z
e
r
A
da
m
L
e
a
r
ni
ng
r
a
t
e
0.0001
B
a
t
c
h
s
i
z
e
64
E
poc
h
40
A
c
t
i
va
t
i
on
f
unc
t
i
on
(
hi
dde
n
)
S
e
L
U
A
c
t
i
va
t
i
on
f
unc
t
i
on
(
out
put
)
S
of
t
M
a
x
P
ool
i
ng
2
×
2
R
e
s
c
a
l
e
1/
255
4.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
A
t
th
is
s
t
a
ge
,
th
e
tr
a
in
in
g
of
th
e
m
ode
l
w
il
l
be
c
onduc
te
d
u
s
i
ng
s
e
c
ond
a
r
y
da
ta
,
w
it
h
th
e
tr
a
in
in
g
da
ta
10,582
s
a
m
pl
e
s
a
nd
th
e
te
s
t
d
a
ta
2,098
s
a
m
pl
e
s
w
it
h
pr
e
pr
oc
e
s
s
in
g
m
e
th
od.
T
he
m
ode
l
w
il
l
be
tr
a
in
e
d
a
c
c
or
di
ng
to
th
e
s
pe
c
if
ic
a
ti
ons
m
ode
l.
D
ur
in
g
tr
a
in
in
g,
th
e
m
ode
l
w
il
l
e
va
lu
a
te
th
e
te
s
t
da
t
a
a
nd
pr
oduc
e
a
c
c
ur
a
c
y
m
e
tr
ic
s
a
t
e
a
c
h
e
poc
h.
O
nc
e
a
ll
m
ode
ls
a
r
e
tr
a
in
e
d,
th
e
e
va
lu
a
ti
on
w
il
l
be
pe
r
f
or
m
e
d
on
pr
im
a
r
y
da
ta
t
o de
te
r
m
in
e
t
he
be
s
t
pr
e
pr
oc
e
s
s
in
g m
e
th
od f
or
ge
nde
r
c
la
s
s
if
ic
a
ti
on.
4
.1.
T
r
ai
n
in
g r
e
s
u
lt
s
T
he
C
N
N
m
ode
l
s
a
r
e
tr
a
in
e
d
in
40
e
poc
h
s
.
T
he
obt
a
in
e
d
a
c
c
ur
a
c
y
a
nd
lo
s
s
of
e
a
c
h
e
poc
h
dur
in
g
tr
a
in
in
g
pr
ogr
e
s
s
of
th
e
m
ode
ls
a
r
e
s
how
n
in
F
ig
ur
e
4.
F
or
th
e
gr
a
ys
c
a
l
e
m
ode
l,
th
e
a
c
c
ur
a
c
y
gr
a
ph
in
F
ig
ur
e
4(
a
)
s
ho
w
s
a
c
o
n
s
i
s
te
n
t
u
pw
a
r
d
t
r
e
n
d
w
it
h
m
in
or
r
ip
p
le
s
,
i
nd
i
c
a
ti
n
g
p
e
r
io
d
s
o
f
s
l
ow
e
r
g
a
i
n
s
.
T
h
e
l
o
s
s
g
r
a
p
h
s
i
m
i
la
r
l
y
r
e
f
l
e
c
t
s
a
g
e
n
e
r
a
l
d
e
c
li
n
e
w
it
h
o
c
c
a
s
i
on
a
l
f
l
u
c
t
u
a
t
io
n
s
,
s
u
gg
e
s
ti
ng
t
h
e
m
od
e
l’
s
o
ng
o
in
g
a
d
j
u
s
t
m
e
nt
s
i
n
m
i
ni
m
i
z
i
ng
e
r
r
or
s
.
T
he
S
Q
I
t
e
s
t
d
a
ta
r
e
s
ul
t
s
i
n a
s
m
o
ot
h
e
r
c
ur
v
e
c
om
p
a
r
e
d t
o
t
h
e
gr
a
y
s
c
a
l
e
m
od
e
l
a
s
v
i
s
u
a
li
z
e
d
i
n
F
i
g
ur
e
4(
b)
,
in
di
c
a
ti
ng
t
h
a
t
t
h
e
m
od
e
l
w
a
s
tr
a
i
n
e
d
a
d
e
qu
a
t
e
l
y
to
ha
n
dl
e
v
a
r
i
a
t
io
n
s
i
n
il
lu
m
i
n
a
t
io
n.
M
e
a
nw
h
il
e
,
th
e
r
e
s
ul
t
s
f
r
o
m
t
h
e
H
E
m
o
de
ls
i
n F
ig
ur
e
4
(
c
)
r
e
v
e
a
l
a
w
a
vy
a
nd
u
ne
ve
n c
ur
v
e
,
s
ug
ge
s
t
in
g
th
a
t
th
e
m
od
e
l
w
a
s
n
ot
e
f
f
e
c
ti
v
e
l
y t
r
a
in
e
d
to
h
a
ndl
e
v
a
r
i
a
t
io
n
s
i
n
il
lu
m
in
a
ti
on
.
T
hi
s
i
r
r
e
gul
a
r
i
ty
i
ndi
c
a
t
e
s
t
h
a
t
t
h
e
m
od
e
l
s
tr
ug
gl
e
s
w
i
th
ge
n
e
r
a
li
z
in
g
a
c
r
os
s
di
f
f
e
r
e
nt
il
lu
m
i
n
a
ti
on
c
ondi
ti
o
n
s
.
D
e
s
pi
t
e
t
h
e
s
e
is
s
ue
s
,
th
e
m
od
e
l
w
a
s
e
v
a
lu
a
t
e
d
w
it
h
pr
im
a
r
y
da
t
a
a
nd
a
c
hi
e
v
e
d
a
n
a
c
c
ur
a
c
y
of
88
.7
6
%
.
W
h
il
e
t
hi
s
a
c
c
ur
a
c
y
i
s
r
e
l
a
t
iv
e
ly
s
tr
ong
,
th
e
ob
s
e
r
ve
d
c
ur
v
e
c
h
a
r
a
c
t
e
r
i
s
ti
c
s
hi
g
hl
i
ght
li
m
it
a
ti
on
s
i
n
th
e
m
od
e
l
’
s
tr
a
i
ni
n
g,
poi
nt
i
ng
t
o
p
ot
e
nt
ia
l
im
p
r
ov
e
m
e
n
ts
ne
e
d
e
d
i
n
ha
ndl
in
g
di
v
e
r
s
e
il
lu
m
in
a
t
io
n
s
c
e
na
r
io
s
to
e
n
ha
nc
e
ov
e
r
a
ll
p
e
r
f
or
m
a
nc
e
a
nd
c
on
s
i
s
te
nc
y.
F
ig
ur
e
4(
d)
il
lu
s
tr
a
te
s
th
e
L
T
I
S
N
m
ode
l,
w
hi
c
h
di
s
pl
a
ys
a
r
e
la
ti
ve
ly
s
ta
bl
e
a
c
c
ur
a
c
y
tr
e
nd,
th
ough
m
in
or
f
lu
c
tu
a
ti
ons
a
r
e
pr
e
s
e
nt
in
la
te
r
e
poc
hs
.
T
hi
s
in
di
c
a
t
e
s
th
a
t
th
e
m
od
e
l
w
a
s
m
od
e
r
a
te
ly
s
uc
c
e
s
s
f
ul
in
a
da
pt
in
g
to
va
r
ia
ti
ons
in
il
lu
m
in
a
ti
on, s
how
in
g be
tt
e
r
c
ons
is
te
nc
y c
om
p
a
r
e
d t
o t
he
H
E
m
ode
l
but
s
ti
ll
l
e
a
vi
ng r
oom
f
or
i
m
pr
ove
m
e
nt
.
S
im
il
a
r
to
th
e
H
E
m
ode
ls
,
th
e
te
s
t
da
ta
c
ur
ve
of
G
I
C
m
ode
l
pr
oduc
e
s
a
ja
gge
d
a
nd
une
ve
n
c
ur
ve
,
a
s
s
how
n
in
F
ig
ur
e
4
(
e
)
,
w
it
h
a
c
c
ur
a
c
y
e
xhi
bi
ti
ng
a
n
uns
ta
bl
e
d
e
c
li
ne
a
f
te
r
e
poc
h
25.
T
hi
s
s
ugge
s
ts
th
a
t
th
e
m
ode
l
ha
s
not
be
e
n
a
de
qua
te
ly
tr
a
in
e
d
to
m
a
na
ge
va
r
ia
ti
ons
i
n
il
lu
m
in
a
ti
on.
F
in
a
ll
y,
th
e
te
s
t
da
ta
c
ur
ve
of
D
oG
m
ode
l
s
how
n
in
F
ig
ur
e
4(
f
)
pr
oduc
e
s
a
r
e
la
ti
ve
ly
s
m
oot
h
c
ur
ve
,
w
it
h
a
m
in
or
a
c
c
ur
a
c
y
dr
op
ob
s
e
r
ve
d
a
f
te
r
e
poc
h
34.
T
hi
s
s
ugge
s
ts
th
a
t
th
e
m
ode
l
is
a
de
qua
te
ly
t
r
a
in
e
d
to
m
a
na
ge
va
r
ia
ti
ons
in
il
lu
m
in
a
ti
on
.
E
va
lu
a
te
d
w
it
h
pr
im
a
r
y
da
ta
,
th
e
m
ode
l
a
c
hi
e
ve
d a
n
a
c
c
ur
a
c
y
o
f
91.07%
.
T
hi
s
pe
r
f
or
m
a
nc
e
in
di
c
a
te
s
th
a
t
th
e
m
ode
l
is
e
f
f
e
c
ti
ve
ly
ha
ndl
in
g
th
e
c
ha
ll
e
nge
s
pos
e
d
by
di
f
f
e
r
e
nt
il
lu
m
in
a
ti
on
c
ondi
ti
ons
,
de
m
ons
tr
a
ti
ng
it
s
r
obus
tn
e
s
s
a
nd
r
e
li
a
bi
li
ty
in
ge
nde
r
c
la
s
s
if
ic
a
ti
on
ta
s
k
s
.
T
he
s
m
oot
h
c
ur
ve
a
nd
hi
gh
a
c
c
ur
a
c
y
unde
r
s
c
or
e
th
e
m
ode
l’
s
pr
of
ic
ie
nc
y i
n a
da
pt
in
g t
o va
r
ia
ti
ons
a
nd e
ns
ur
in
g c
ons
is
te
nt
r
e
s
ul
ts
.
4
.2.
T
e
s
t
r
e
s
u
lt
s
T
he
te
s
t
r
e
s
ul
ts
a
r
e
f
ur
th
e
r
pr
e
s
e
nt
e
d
in
th
e
c
onf
us
io
n
m
a
tr
ix
i
n
F
ig
ur
e
5.
T
he
gr
a
ys
c
a
le
m
ode
l
f
o
r
ge
nde
r
c
la
s
s
if
ic
a
ti
on
de
m
on
s
tr
a
te
s
a
n
ove
r
a
ll
a
c
c
ur
a
c
y
of
87%
.
A
s
s
how
n
in
F
ig
ur
e
5(
a
)
,
th
e
m
ode
l
s
how
s
hi
gh
pr
e
c
is
io
n
f
or
f
e
m
a
le
s
(
97%
)
,
but
a
lo
w
e
r
r
e
c
a
ll
(
76%
)
,
in
di
c
a
ti
ng
a
te
nde
nc
y
to
m
is
c
la
s
s
if
y
f
e
m
a
le
s
a
s
m
a
le
s
.
I
n
c
ont
r
a
s
t,
m
a
le
c
la
s
s
if
ic
a
ti
ons
e
xhi
bi
t
a
r
e
c
a
ll
of
9
7%
but
a
lo
w
e
r
pr
e
c
is
io
n
(
80%
)
,
s
ugge
s
ti
ng
oc
c
a
s
io
na
l
m
is
c
la
s
s
if
ic
a
ti
on
of
m
a
le
s
a
s
f
e
m
a
le
s
.
T
he
F
1
-
s
c
or
e
s
f
or
f
e
m
a
le
s
a
nd
m
a
le
s
a
r
e
0.85
a
nd
0.88,
r
e
s
pe
c
ti
ve
ly
,
r
e
f
le
c
ti
ng
a
good
ba
la
n
c
e
be
tw
e
e
n
pr
e
c
i
s
io
n
a
nd
r
e
c
a
ll
.
W
hi
le
th
e
m
ode
l
pe
r
f
or
m
s
w
e
ll
ov
e
r
a
ll
,
im
pr
ovi
ng r
e
c
a
ll
f
or
f
e
m
a
le
c
la
s
s
if
ic
a
ti
ons
c
oul
d e
nha
nc
e
i
ts
e
f
f
e
c
ti
ve
ne
s
s
.
T
he
S
Q
I
m
ode
l
a
c
hi
e
ve
s
a
n
a
c
c
ur
a
c
y
of
90%
f
or
ge
nde
r
c
la
s
s
if
ic
a
ti
on
w
it
h
th
e
c
onf
us
io
n
m
a
tr
i
x
s
how
n
in
F
ig
ur
e
5(
b)
.
P
r
e
c
is
io
n
a
nd
r
e
c
a
ll
f
or
bot
h
f
e
m
a
le
a
nd
m
a
le
c
la
s
s
e
s
a
r
e
ba
la
n
c
e
d,
w
it
h
va
lu
e
s
of
0.91
a
nd
0.89
f
or
f
e
m
a
le
s
,
a
nd
0.90
a
nd
0.91
f
or
m
a
le
s
,
r
e
s
p
e
c
ti
v
e
ly
.
T
he
F
1
-
s
c
or
e
s
f
or
bot
h
c
la
s
s
e
s
a
r
e
0.90.
T
he
s
e
r
e
s
ul
ts
in
di
c
a
te
th
a
t
th
e
S
Q
I
pr
e
pr
oc
e
s
s
in
g
te
c
hni
que
e
f
f
e
c
ti
ve
ly
e
nha
nc
e
s
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
unde
r
va
r
yi
ng
il
lu
m
in
a
ti
on
c
ondi
ti
ons
.
S
im
il
a
r
ly
,
th
e
H
E
m
o
de
l
a
c
hi
e
ve
s
a
n
a
c
c
ur
a
c
y
of
89%
f
or
ge
nde
r
c
la
s
s
if
ic
a
ti
on.
T
he
pr
e
c
is
io
n
a
nd
r
e
c
a
ll
f
or
th
e
f
e
m
a
le
c
la
s
s
a
r
e
0.88
a
nd
0.90,
w
hi
le
f
or
th
e
m
a
le
c
la
s
s
th
e
y
a
r
e
0.90
a
nd
0.87,
r
e
s
pe
c
ti
ve
ly
.
B
ot
h
c
la
s
s
e
s
ha
ve
a
n
F
1
-
s
c
or
e
of
0.89.
T
he
s
e
r
e
s
ul
ts
in
di
c
a
te
th
a
t
th
e
H
E
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
C
om
par
at
iv
e
analy
s
is
of
ge
nde
r
c
la
s
s
if
ic
at
io
n m
e
th
ods
u
s
in
g c
o
nv
ol
ut
io
nal
…
(
P
anc
a D
e
w
i
P
am
ungk
as
ar
i
)
3641
pr
e
pr
oc
e
s
s
in
g
te
c
hni
que
a
ls
o
pr
ovi
de
s
ba
la
nc
e
d
pe
r
f
or
m
a
nc
e
a
s
vi
s
ua
li
z
e
d
in
F
ig
ur
e
5(
c
)
,
th
ough
s
li
ght
ly
lo
w
e
r
t
ha
n t
he
S
Q
I
m
e
th
od, with e
f
f
e
c
ti
ve
ha
ndl
in
g of
i
ll
um
in
a
t
io
n va
r
ia
ti
ons
.
T
he
L
T
I
S
N
m
ode
l
a
c
hi
e
ve
s
a
n
a
c
c
ur
a
c
y
of
82%
.
F
ig
ur
e
5(
d)
s
how
s
hi
gh
pr
e
c
is
io
n
f
or
f
e
m
a
le
s
(
0.97)
but
lo
w
e
r
r
e
c
a
ll
(
0.66)
,
w
hi
le
m
a
l
e
s
h
a
ve
a
lo
w
e
r
pr
e
c
is
io
n
(
0.74)
but
a
hi
gh
r
e
c
a
ll
(
0.98)
.
T
he
F1
-
s
c
or
e
s
a
r
e
0.78
f
or
f
e
m
a
le
s
a
nd
0.84
f
or
m
a
le
s
.
T
h
e
s
e
r
e
s
ul
ts
in
di
c
a
te
a
n
im
ba
la
nc
e
in
th
e
m
ode
l’
s
pe
r
f
or
m
a
nc
e
,
pa
r
ti
c
ul
a
r
ly
in
m
is
c
la
s
s
if
yi
ng
a
s
ig
ni
f
ic
a
nt
num
be
r
of
f
e
m
a
le
s
a
s
m
a
le
s
,
w
hi
c
h
a
f
f
e
c
ts
ove
r
a
ll
a
c
c
ur
a
c
y.
T
he
G
I
C
m
ode
l
a
c
hi
e
ve
s
a
n
a
c
c
ur
a
c
y
of
86%
.
T
he
c
onf
us
io
n
m
a
tr
ix
of
G
I
C
m
ode
l
in
F
ig
ur
e
5(
e
)
s
how
s
s
tr
ong
pr
e
c
is
io
n
f
or
f
e
m
a
le
s
(
0.92)
but
lo
w
e
r
r
e
c
a
ll
(
0.8
0)
,
w
hi
le
m
a
le
s
ha
ve
a
lo
w
e
r
pr
e
c
is
io
n
(
0.82)
but
a
hi
ghe
r
r
e
c
a
ll
(
0.93)
.
T
he
F
1
-
s
c
or
e
s
a
r
e
0.85
f
o
r
f
e
m
a
le
s
a
nd
0.87
f
or
m
a
le
s
.
T
he
s
e
r
e
s
ul
ts
in
di
c
a
te
ba
la
nc
e
d ove
r
a
ll
pe
r
f
or
m
a
nc
e
, w
it
h a
s
li
ght
ly
hi
ghe
r
a
bi
li
ty
t
o c
or
r
e
c
tl
y c
la
s
s
if
y m
a
le
s
.
T
he
D
oG
m
ode
l
a
c
hi
e
ve
s
a
n a
c
c
ur
a
c
y of
91%
.
F
ig
ur
e
5(
f
)
s
ho
w
s
t
ha
t
D
oG
m
ode
l
ha
s
hi
gh pr
e
c
is
io
n
(
0.87)
a
nd
r
e
c
a
ll
(
0.96)
f
or
f
e
m
a
le
s
,
w
hi
le
m
a
le
s
c
la
s
s
d
e
m
ons
tr
a
te
hi
ghe
r
pr
e
c
is
io
n
(
0.96)
but
lo
w
e
r
r
e
c
a
ll
(
0.86)
.
T
he
F
1
-
s
c
or
e
s
a
r
e
0.92
f
or
f
e
m
a
le
s
a
nd
0.91
f
or
m
a
le
s
.
T
h
e
s
e
r
e
s
ul
ts
in
di
c
a
t
e
s
tr
ong
ove
r
a
ll
pe
r
f
or
m
a
nc
e
,
w
it
h
th
e
m
ode
l
s
how
in
g
b
e
tt
e
r
r
e
c
a
ll
f
or
f
e
m
a
l
e
s
a
nd
hi
ghe
r
pr
e
c
is
io
n
f
or
m
a
le
s
,
le
a
di
ng
to
ba
la
nc
e
d c
la
s
s
if
ic
a
ti
on outc
om
e
s
.
(
a
)
(
b)
(
c
)
(
d)
(
e
)
(f)
F
ig
ur
e
4. T
r
a
in
in
g pr
ogr
e
s
s
of
(
a
)
gr
a
ys
c
a
le
m
ode
l,
(
b)
S
Q
I
m
o
de
l,
(
c
)
H
E
m
ode
l,
(
d)
L
T
I
S
N
m
ode
l,
(
e
)
G
I
C
m
ode
l,
a
nd (
f
)
D
o
G
m
ode
l
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
5
,
O
c
to
be
r
20
25
:
3634
-
3646
3642
(
a
)
(
b)
(
c
)
(
d)
(
e
)
(f)
F
ig
ur
e
5. C
onf
us
io
n m
a
tr
ix
of
a
ll
m
ode
ls
:
(
a
)
gr
a
ys
c
a
le
m
ode
l,
(
b)
S
Q
I
m
ode
l,
(
c
)
H
E
m
ode
l,
(
d)
L
T
I
S
N
m
ode
l,
(
e
)
G
I
C
m
ode
l,
a
nd (
f
)
D
o
G
m
ode
l
4.3. M
od
e
l
c
om
p
ar
is
on
T
a
bl
e
4
is
pr
ovi
de
d
to
s
im
pl
if
y
th
e
c
om
pa
r
is
on
of
pe
r
f
or
m
a
nc
e
m
e
tr
ic
s
f
r
om
va
r
io
us
pr
e
pr
oc
e
s
s
in
g
m
e
th
od
a
ppl
ie
d,
a
ll
ow
in
g
a
c
le
a
r
e
r
e
v
a
lu
a
ti
on
of
th
e
ir
im
pa
c
t
on
th
e
m
ode
l'
s
a
c
c
ur
a
c
y
a
nd
e
f
f
ic
ie
nc
y.
T
hi
s
c
om
pa
r
is
on
he
lp
s
id
e
nt
if
y
w
hi
c
h
m
e
th
od
de
li
ve
r
s
th
e
m
os
t
opt
im
a
l
r
e
s
ul
t
f
o
r
th
e
gi
ve
n
da
ta
s
e
t.
T
he
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
C
om
par
at
iv
e
analy
s
is
of
ge
nde
r
c
la
s
s
if
ic
at
io
n m
e
th
ods
u
s
in
g c
o
nv
ol
ut
io
nal
…
(
P
anc
a D
e
w
i
P
am
ungk
as
ar
i
)
3643
c
om
pa
r
is
on
is
ba
s
e
d
on
ke
y
pe
r
f
or
m
a
nc
e
m
e
tr
ic
s
s
u
c
h
a
s
a
c
c
ur
a
c
y,
pr
e
c
is
io
n,
r
e
c
a
ll
,
a
nd
F
1
-
s
c
or
e
f
or
bot
h
f
e
m
a
le
a
nd
m
a
le
c
la
s
s
if
ic
a
ti
ons
.
A
c
c
ur
a
c
y
is
a
c
r
it
ic
a
l
m
e
tr
ic
,
r
e
f
le
c
ti
ng
th
e
ove
r
a
ll
c
or
r
e
c
tn
e
s
s
of
th
e
m
ode
l
.
T
he
D
oG
m
ode
l
a
c
hi
e
ve
d
th
e
hi
ghe
s
t
a
c
c
ur
a
c
y
of
91%
,
c
lo
s
e
ly
f
ol
lo
w
e
d
by
S
Q
I
w
it
h
90%
,
a
nd
H
E
w
it
h
89%
.
T
he
s
e
r
e
s
ul
ts
s
ugge
s
t
th
a
t
th
e
D
oG
a
nd
S
Q
I
m
e
th
ods
a
r
e
m
or
e
e
f
f
e
c
ti
ve
in
im
pr
ovi
ng
th
e
m
ode
l'
s
a
bi
li
ty
to
c
or
r
e
c
tl
y
c
la
s
s
if
y
bot
h
ge
nd
e
r
s
a
c
r
os
s
va
r
io
us
il
lu
m
i
na
ti
on
c
ondi
ti
ons
.
O
n
th
e
lo
w
e
r
e
nd,
L
T
I
S
N
pr
oduc
e
d
a
n
a
c
c
ur
a
c
y
of
82%
,
in
di
c
a
ti
ng
le
s
s
r
e
li
a
bl
e
ov
e
r
a
ll
pe
r
f
or
m
a
nc
e
,
e
s
pe
c
ia
ll
y
in
c
or
r
e
c
tl
y
id
e
nt
if
yi
ng
f
e
m
a
le
i
m
a
ge
s
.
T
a
bl
e
4. M
od
e
l
pe
r
f
or
m
a
nc
e
m
e
tr
ic
s
M
ode
l
A
c
c
ur
a
c
y
(%)
L
os
s
P
r
e
c
i
s
i
on
(
F
e
m
a
l
e
)
R
e
c
a
l
l
(
F
e
m
a
l
e
)
F
1 s
c
or
e
(
F
e
m
a
l
e
)
P
r
e
c
i
s
i
on
(
M
a
l
e
)
R
e
c
a
l
l
(
M
a
l
e
)
F
1 s
c
or
e
(
M
a
l
e
)
G
r
a
ys
c
a
l
e
86.84
0.3088
0.97
0.76
0.85
0.80
0.97
0.88
S
Q
I
90.39
0.1085
0.91
0.89
0.90
0.90
0.91
0.90
HE
88.76
0.3330
0.88
0.90
0.89
0.90
0.87
0.89
L
T
I
S
N
81.56
0.1937
0.97
0.66
0.78
0.74
0.98
0.84
G
I
C
86.26
0.3830
0.92
0.80
0.85
0.82
0.93
0.87
D
oG
91.07
0.2504
0.87
0.96
0.92
0.96
0.86
0.91
P
r
e
c
is
i
on
m
e
a
s
u
r
e
s
t
he
pe
r
c
e
nt
a
g
e
o
f
c
or
r
e
c
tl
y
i
de
nt
i
f
ie
d
in
s
t
a
nc
e
s
o
u
t
of
a
l
l
i
ns
ta
nc
e
s
p
r
e
di
c
t
e
d
a
s
a
s
pe
c
if
i
c
ge
nd
e
r
.
T
he
L
T
I
S
N
m
o
de
l
de
m
on
s
t
r
a
t
e
d
th
e
hi
g
he
s
t
p
r
e
c
is
io
n
f
o
r
th
e
f
e
m
a
l
e
c
l
a
s
s
(
0
.
97
)
,
b
u
t
th
is
is
o
f
f
s
e
t
by
l
ow
e
r
r
e
c
a
l
l
va
lu
e
s
,
i
nd
i
c
a
ti
ng
t
ha
t
w
h
il
e
i
t
i
s
h
i
gh
ly
a
c
c
u
r
a
t
e
i
n
th
e
p
r
e
d
ic
t
io
ns
i
t
m
a
ke
s
,
i
t
f
a
i
ls
t
o
d
e
te
c
t
m
a
ny
t
r
u
e
p
os
it
i
ve
s
.
T
h
e
D
o
G
m
o
de
l,
w
hi
le
ha
vi
n
g
a
s
li
g
ht
l
y
l
ow
e
r
p
r
e
c
i
s
i
on
f
o
r
f
e
m
a
le
s
(
0.
87
)
,
m
a
i
nt
a
i
ns
a
m
uc
h
m
or
e
ba
la
nc
e
d
p
e
r
f
o
r
m
a
nc
e
w
i
t
h
hi
g
h
p
r
e
c
is
io
n
f
o
r
m
a
l
e
s
(
0.
96
)
,
i
n
d
ic
a
t
in
g
f
e
w
e
r
f
a
ls
e
po
s
i
t
iv
e
s
a
c
r
os
s
b
ot
h
c
la
s
s
e
s
.
S
Q
I
a
n
d
H
E
a
ls
o
s
h
ow
b
a
l
a
n
c
e
d
p
r
e
c
is
i
on
a
c
r
os
s
b
o
t
h
c
l
a
s
s
e
s
,
w
i
th
v
a
lu
e
s
a
r
o
un
d
0
.9
0
,
de
m
ons
t
r
a
t
in
g
r
e
li
a
bl
e
pe
r
f
o
r
m
a
n
c
e
i
n
c
o
r
r
e
c
t
ly
id
e
n
t
i
f
y
in
g
b
o
t
h
ge
nd
e
r
s
.
R
e
c
a
l
l,
w
h
ic
h
r
e
f
le
c
ts
t
he
a
bi
li
ty
o
f
th
e
m
o
de
l
t
o
c
or
r
e
c
tl
y
i
de
nt
i
f
y
a
l
l
in
s
t
a
nc
e
s
o
f
a
pa
r
t
ic
u
la
r
c
la
s
s
,
is
w
he
r
e
t
he
D
o
G
m
ode
l
s
ta
nds
o
ut
,
e
s
pe
c
i
a
l
ly
f
or
f
e
m
a
le
s
,
w
it
h
a
r
e
c
a
ll
o
f
0.9
6.
T
hi
s
h
ig
h
r
e
c
a
ll
f
o
r
f
e
m
a
le
s
m
e
a
ns
th
e
D
o
G
m
od
e
l
is
hi
gh
ly
e
f
f
e
c
ti
v
e
i
n
id
e
n
ti
f
yi
ng
a
lm
os
t
a
ll
f
e
m
a
le
im
a
ge
s
,
w
i
th
m
in
im
a
l
f
a
ls
e
ne
ga
t
iv
e
s
.
I
n
c
o
nt
r
a
s
t,
L
T
I
S
N
,
de
s
p
i
te
i
ts
h
ig
h
pr
e
c
is
io
n
f
or
f
e
m
a
l
e
s
,
ha
s
a
s
ig
ni
f
ic
a
n
tl
y
lo
w
e
r
r
e
c
a
ll
(
0.6
6
)
,
m
e
a
n
in
g
i
t
f
a
i
ls
t
o
c
o
r
r
e
c
tl
y
c
la
s
s
i
f
y
a
s
u
bs
ta
n
ti
a
l
po
r
t
io
n
o
f
f
e
m
a
le
im
a
ge
s
.
T
he
G
I
C
m
ode
l
e
xhi
bi
ts
a
m
o
r
e
ba
la
nc
e
d
r
e
c
a
ll
be
twe
e
n
bo
th
ge
n
de
r
s
,
w
it
h
va
lu
e
s
o
f
0.
80
f
o
r
f
e
m
a
le
s
a
nd
0.9
3
f
o
r
m
a
le
s
,
r
e
f
le
c
ti
ng
a
r
e
li
a
bl
e
a
b
il
it
y
to
c
a
p
tu
r
e
m
os
t
in
s
ta
n
c
e
s
o
f
e
a
c
h
ge
nde
r
,
th
ou
gh
s
l
ig
ht
ly
bi
a
s
e
d
t
ow
a
r
d
m
a
le
c
l
a
s
s
i
f
ic
a
t
io
ns
.
T
he
F
1
-
s
c
o
r
e
ba
la
nc
e
s
pr
e
c
is
io
n
a
n
d
r
e
c
a
l
l,
o
f
f
e
r
in
g
a
m
or
e
c
o
m
p
r
e
he
ns
iv
e
v
ie
w
o
f
m
o
de
l
pe
r
f
o
r
m
a
nc
e
.
D
o
G
a
nd
S
Q
I
a
g
a
in
de
m
ons
t
r
a
te
s
upe
r
io
r
r
e
s
ul
ts
,
w
it
h
F
1
-
s
c
o
r
e
s
o
f
0
.9
2
f
o
r
f
e
m
a
le
s
a
nd
0.9
1
f
o
r
m
a
le
s
in
th
e
D
o
G
m
o
de
l
a
nd
ba
l
a
nc
e
d
0.9
0
s
c
or
e
s
f
or
b
ot
h
g
e
nde
r
s
i
n
th
e
S
Q
I
m
ode
l.
T
he
s
e
s
c
or
e
s
h
ig
hl
ig
h
t
th
e
i
r
a
bi
li
ty
to
m
a
i
nt
a
in
a
s
tr
on
g
tr
a
de
-
o
f
f
b
e
twe
e
n
p
r
e
c
is
i
on
a
n
d
r
e
c
a
ll
,
m
a
k
in
g
t
he
m
th
e
m
os
t
ba
la
nc
e
d
m
ode
ls
in
t
hi
s
c
o
m
pa
r
is
on
.
I
n
c
o
nt
r
a
s
t
,
L
T
I
S
N
,
de
s
pi
te
it
s
h
ig
h
pr
e
c
is
io
n
f
o
r
f
e
m
a
le
s
,
s
uf
f
e
r
s
f
r
om
a
lo
w
e
r
F
1
-
s
c
o
r
e
(
0.7
8
f
o
r
f
e
m
a
l
e
s
)
du
e
to
it
s
w
e
a
k
r
e
c
a
ll
, l
i
m
i
ti
n
g
it
s
ov
e
r
a
l
l
e
f
f
e
c
t
iv
e
ne
s
s
.
E
a
c
h
m
ode
l
pr
e
s
e
nt
s
a
di
f
f
e
r
e
nt
s
e
t
o
f
s
tr
e
ng
th
s
a
nd
w
e
a
kne
s
s
e
s
.
D
o
G
e
m
e
r
ge
s
a
s
t
he
be
s
t
ov
e
r
a
ll
m
ode
l,
a
c
hi
e
v
in
g
hi
gh
a
c
c
ur
a
c
y
,
ba
l
a
nc
e
d
p
r
e
c
is
io
n,
a
nd
r
e
c
a
ll
,
a
l
on
g
w
i
th
s
tr
on
g
F
1
-
s
c
or
e
s
f
or
bot
h
ge
nd
e
r
s
.
T
h
is
b
a
la
nc
e
s
ugg
e
s
ts
th
a
t
D
o
G
is
hi
gh
ly
e
f
f
e
c
ti
ve
in
r
e
a
l
-
w
or
ld
s
c
e
na
r
io
s
w
he
r
e
il
lu
m
i
na
t
io
n
c
ond
it
io
ns
c
a
n
va
r
y
s
ig
n
i
f
ic
a
nt
ly
.
S
Q
I
a
ls
o
pe
r
f
or
m
s
w
e
l
l
a
c
r
os
s
a
ll
m
e
t
r
ic
s
,
m
a
ki
n
g
i
t
a
no
th
e
r
s
t
r
o
ng
c
a
nd
id
a
te
f
o
r
r
e
l
ia
b
le
ge
n
de
r
c
la
s
s
i
f
ic
a
t
io
n.
L
T
I
S
N
,
w
h
il
e
e
xc
e
ll
in
g
in
p
r
e
c
is
io
n
f
o
r
f
e
m
a
le
s
,
s
t
r
u
ggl
e
s
w
it
h
r
e
c
a
ll
,
le
a
di
n
g
t
o
m
a
n
y
m
is
c
la
s
s
if
ic
a
ti
ons
i
n
th
is
c
la
s
s
.
I
ts
h
ig
h
pr
e
c
is
io
n
but
lo
w
r
e
c
a
ll
f
o
r
f
e
m
a
le
s
s
ugge
s
ts
th
a
t
w
hi
le
i
t
m
a
ke
s
a
c
c
ur
a
te
p
r
e
d
ic
t
io
ns
f
o
r
th
e
i
m
a
g
e
s
i
t
c
la
s
s
i
f
ie
s
a
s
f
e
m
a
le
,
i
t
f
a
il
s
to
i
de
n
ti
f
y
a
s
ig
n
i
f
ic
a
nt
po
r
ti
on
o
f
a
c
tu
a
l
f
e
m
a
le
i
m
a
g
e
s
.
T
hi
s
im
ba
l
a
nc
e
m
a
ke
s
i
t
le
s
s
s
u
it
a
bl
e
f
o
r
a
pp
li
c
a
t
io
ns
w
h
e
r
e
i
t
is
c
r
i
ti
c
a
l
to
c
o
r
r
e
c
tl
y
c
la
s
s
i
f
y
a
l
l
f
e
m
a
le
i
ns
ta
nc
e
s
.
H
E
a
nd
G
I
C
o
f
f
e
r
m
o
r
e
ba
la
nc
e
d
pe
r
f
o
r
m
a
nc
e
t
ha
n
L
T
I
S
N
,
bu
t
th
e
i
r
a
c
c
u
r
a
c
y
a
nd
F
1
-
s
c
o
r
e
s
,
w
h
il
e
c
om
pe
t
it
iv
e
,
f
a
ll
s
l
ig
h
tl
y
s
ho
r
t
o
f
t
he
to
p
-
pe
r
f
o
r
m
in
g
m
ode
ls
,
D
o
G
a
nd
S
Q
I
.
T
he
s
e
m
od
e
ls
m
ig
ht
b
e
c
ons
id
e
r
e
d
r
e
l
ia
b
le
b
ut
l
e
s
s
r
obus
t
w
h
e
n
de
a
l
in
g
w
it
h
hi
g
hl
y v
a
r
ie
d
o
r
c
ha
ll
e
ng
in
g
i
m
a
ge
da
ta
s
e
ts
.
I
n
c
o
nc
l
us
i
on,
th
e
D
o
G
a
nd
S
Q
I
m
od
e
ls
de
m
o
ns
t
r
a
t
e
s
upe
r
i
or
pe
r
f
o
r
m
a
nc
e
a
c
r
os
s
ke
y
m
e
tr
ic
s
,
pa
r
ti
c
ul
a
r
ly
in
ba
l
a
nc
in
g
p
r
e
c
is
io
n,
r
e
c
a
ll
,
a
nd
F
1
-
s
c
o
r
e
s
f
o
r
bot
h
g
e
nde
r
s
.
D
o
G
s
l
ig
ht
ly
e
dge
s
ou
t
S
Q
I
w
it
h
it
s
hi
ghe
r
r
e
c
a
ll
f
o
r
f
e
m
a
le
s
a
nd
ov
e
r
a
l
l
a
c
c
ur
a
c
y
.
T
he
s
e
m
o
de
ls
a
r
e
t
hus
w
e
l
l
-
s
u
it
e
d
f
o
r
a
pp
li
c
a
ti
ons
th
a
t
r
e
q
ui
r
e
h
ig
h
a
c
c
u
r
a
c
y
in
ge
nde
r
c
la
s
s
i
f
ic
a
ti
on
u
nde
r
va
r
y
in
g
il
lu
m
in
a
t
io
n
c
ond
it
io
ns
.
W
h
il
e
L
T
I
S
N
a
nd
G
I
C
e
xh
ib
it
s
tr
ong
pe
r
f
or
m
a
n
c
e
in
s
p
e
c
i
f
i
c
a
r
e
a
s
,
th
e
y
a
r
e
le
s
s
r
e
l
ia
bl
e
d
ue
to
im
ba
la
nc
e
s
be
twe
e
n
pr
e
c
is
i
on
a
nd
r
e
c
a
l
l,
m
a
k
in
g
th
e
m
le
s
s
e
f
f
e
c
t
iv
e
f
or
ge
ne
r
a
l
us
e
.
F
ut
ur
e
i
m
p
r
ov
e
m
e
nt
s
m
i
gh
t
f
oc
us
on
Evaluation Warning : The document was created with Spire.PDF for Python.