I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
,
p
p
.
888
~
900
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
1
.
p
p
8
8
8
-
9
0
0
888
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Bridg
ing
hybrid
deep learning
det
ection a
nd lig
h
tw
eig
ht
ha
ndcra
ft
ed
f
ea
t
u
res for robus
t
sin
g
le sa
mple face
re
co
g
nition
F
a
uli
nd
a
E
ly
Na
s
t
it
i
1
,
So
pin
g
i
1,
2
,
Dedy
H
a
riy
a
di
3
,
Sri Su
m
a
rlind
a
1,
4
1
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
S
y
s
t
e
m,
F
a
c
u
l
t
y
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
U
n
i
v
e
r
si
t
a
s
D
u
t
a
B
a
n
g
sa
S
u
r
a
k
a
r
t
a
,
S
u
r
a
k
a
r
t
a
,
I
n
d
o
n
e
si
a
2
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
c
s
E
n
g
i
n
e
e
r
i
n
g
,
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
,
U
n
i
v
e
r
si
t
a
s Has
a
n
u
d
d
i
n
,
M
a
k
a
ss
a
r
,
I
n
d
o
n
e
s
i
a
3
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
E
n
g
i
n
e
e
r
i
n
g
a
n
d
I
n
f
o
r
mat
i
o
n
T
e
c
h
n
o
l
o
g
y
,
U
n
i
v
e
r
s
i
t
a
s J
e
n
d
e
r
a
l
A
c
h
m
a
d
Y
a
n
i
,
S
l
e
ma
n
,
I
n
d
o
n
e
si
a
4
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
,
U
n
i
v
e
r
si
t
i
K
u
a
l
a
L
u
m
p
u
r
,
K
u
a
l
a
Lu
m
p
u
r
,
M
a
l
a
y
si
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
n
1
6
,
2
0
2
5
R
ev
is
ed
Dec
3
1
,
2
0
2
5
Acc
ep
ted
J
an
2
2
,
2
0
2
6
S
in
g
le
sa
m
p
le
fa
c
e
r
e
c
o
g
n
it
i
o
n
(S
S
F
R)
re
m
a
in
s
a
c
h
a
ll
e
n
g
in
g
tas
k
d
u
e
to
t
h
e
li
m
it
a
ti
o
n
o
f
h
a
v
in
g
o
n
l
y
o
n
e
re
fe
re
n
c
e
ima
g
e
p
e
r
id
e
n
ti
t
y
,
wh
ic
h
re
d
u
c
e
s
e
m
b
e
d
d
in
g
d
iv
e
rsit
y
a
n
d
d
e
c
re
a
se
s
ro
b
u
st
n
e
ss
u
n
d
e
r
v
a
riatio
n
s
o
f
p
o
se
,
e
x
p
re
ss
io
n
,
a
n
d
il
lu
m
i
n
a
ti
o
n
.
T
h
i
s
stu
d
y
p
ro
p
o
se
d
a
h
y
b
ri
d
fra
m
e
wo
rk
t
h
a
t
in
teg
ra
tes
d
e
e
p
lea
rn
i
n
g
-
b
a
se
d
d
e
tec
ti
o
n
th
r
o
u
g
h
a
n
c
h
o
r
b
o
x
o
p
t
imiz
a
ti
o
n
a
n
d
n
o
n
-
m
a
x
imu
m
s
u
p
p
re
ss
io
n
(
NMS
)
wit
h
li
g
h
twe
ig
h
t
h
a
n
d
c
ra
fted
fe
a
tu
re
e
x
trac
ti
o
n
u
si
n
g
lo
c
a
l
b
i
n
a
ry
p
a
t
tern
(LBP
)
.
T
h
e
d
e
tec
ti
o
n
sta
g
e
lev
e
ra
g
e
s
d
e
e
p
lea
rn
in
g
t
o
e
n
s
u
re
ro
b
u
s
t
fa
c
e
lo
c
a
li
sa
ti
o
n
,
w
h
il
e
LBP
m
a
in
tain
s
c
o
m
p
u
tati
o
n
a
l
e
fficie
n
c
y
u
n
d
e
r
li
m
it
e
d
-
sa
m
p
le
c
o
n
d
it
i
o
n
s.
Th
e
train
in
g
p
ro
c
e
ss
sh
o
we
d
a
c
c
u
ra
c
y
imp
r
o
v
e
m
e
n
t
fro
m
4
7
.
5
%
a
t
t
h
e
in
i
ti
a
l
e
p
o
c
h
to
9
8
.
0
%
a
t
e
p
o
c
h
7
2
,
wh
i
le
tes
ti
n
g
a
c
c
u
ra
c
y
sta
b
il
ize
d
a
t
8
5
-
8
8
%
wit
h
th
e
b
e
st
v
a
lu
e
o
f
8
7
.
9
%
.
Ev
a
l
u
a
ti
o
n
o
n
4
8
n
e
w
fa
c
ial
ima
g
e
s
a
c
h
iev
e
d
8
9
.
6
%
a
c
c
u
ra
c
y
,
9
5
.
3
%
p
re
c
isio
n
,
9
1
.
1
%
re
c
a
ll
,
9
3
.
1
%
F
1
-
sc
o
re
,
a
n
d
0
.
9
4
a
re
a
u
n
d
e
r
th
e
re
c
e
iv
e
r
o
p
e
ra
ti
n
g
c
h
a
ra
c
teristic
c
u
rv
e
(
AU
C
ROC
)
.
R
e
a
l
-
wo
rld
imp
lem
e
n
tatio
n
o
n
A
n
d
r
o
id
a
n
d
iOS
-
b
a
se
d
a
tt
e
n
d
a
n
c
e
a
p
p
l
ica
ti
o
n
s
fu
r
th
e
r
v
a
li
d
a
ted
t
h
e
m
o
d
e
l,
re
a
c
h
in
g
8
8
.
4
6
%
a
c
c
u
ra
c
y
a
c
ro
ss
5
2
tes
ts
u
n
d
e
r
5
0
-
4
0
0
lu
x
il
lu
m
i
n
a
ti
o
n
.
T
h
e
fin
d
i
n
g
s
p
ro
v
e
d
t
h
a
t
th
e
p
ro
p
o
se
d
h
y
b
ri
d
d
e
sig
n
p
ro
v
id
e
s
imp
r
o
v
e
d
a
c
c
u
ra
c
y
a
n
d
sta
b
il
it
y
c
o
m
p
a
re
d
wit
h
p
re
v
i
o
u
s
a
p
p
ro
a
c
h
e
s.
K
ey
w
o
r
d
s
:
An
ch
o
r
b
o
x
B
io
m
etr
ic
L
o
ca
l b
in
ar
y
p
atter
n
s
No
n
-
m
ax
im
u
m
s
u
p
p
r
ess
io
n
Sin
g
le
s
am
p
le
f
ac
e
r
ec
o
g
n
itio
n
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
:
Fau
lin
d
a
E
ly
Nastit
i
Dep
ar
tm
en
t o
f
I
n
f
o
r
m
atio
n
Sy
s
tem
,
Facu
lty
o
f
C
o
m
p
u
te
r
Scien
ce
,
Un
iv
er
s
itas
Du
ta
B
an
g
s
a
Su
r
ak
ar
ta
Km
.
2
0
Ki
Ma
n
g
u
n
Sar
k
o
r
o
R
o
ad
,
Su
r
a
k
ar
ta,
C
en
tr
al
J
av
a,
I
n
d
o
n
esia
E
m
ail: f
au
lin
d
a_
ely
@
u
d
b
.
ac
.
i
d
1.
I
NT
RO
D
UCT
I
O
N
Face
r
ec
o
g
n
itio
n
h
as
em
er
g
ed
as
o
n
e
o
f
th
e
cr
itical
tech
n
o
lo
g
ies
in
d
ig
ital
s
ec
u
r
ity
an
d
au
th
en
ticatio
n
s
y
s
tem
s
[
1
]
,
with
v
ar
io
u
s
ap
p
licatio
n
s
in
p
u
b
lic
s
u
r
v
eillan
ce
[
2
]
,
[
3
]
,
ac
c
ess
co
n
tr
o
l
[
4
]
,
an
d
id
en
tity
v
er
if
icatio
n
[
5
]
,
[
6
]
.
I
t
s
ad
v
an
tag
es
in
p
r
o
v
id
in
g
ef
f
i
cien
t
s
ec
u
r
ity
an
d
u
s
er
co
n
v
e
n
ien
ce
h
av
e
m
a
d
e
it
a
p
r
ef
er
r
ed
c
h
o
ice
ac
r
o
s
s
m
u
l
tip
le
s
ec
to
r
s
,
in
clu
d
in
g
g
o
v
er
n
m
en
t,
b
u
s
in
ess
,
an
d
th
e
in
ter
n
et
o
f
t
h
in
g
s
(
I
o
T
)
[
7
]
–
[
9
]
.
I
n
cr
ea
s
in
g
l
y
,
th
ese
s
y
s
tem
s
ar
e
ad
o
p
ted
in
em
b
ed
d
ed
d
e
v
ices,
s
u
ch
as
m
o
b
ile
d
ev
ices
[
1
0
]
,
[
1
1
]
,
wh
ich
r
eq
u
i
r
e
r
ea
l
-
tim
e
i
d
e
n
tific
atio
n
with
h
ig
h
ac
cu
r
a
cy
[
1
2
]
,
[
1
3
]
.
Ho
wev
er
,
i
m
p
lem
en
tin
g
f
ac
ial
r
ec
o
g
n
itio
n
in
em
b
ed
d
e
d
d
ev
i
ce
s
p
r
esen
ts
s
ig
n
if
ican
t
ch
alle
n
g
es,
p
ar
ticu
lar
ly
in
th
e
co
n
te
x
t
o
f
s
in
g
le
s
am
p
le
f
ac
e
r
ec
o
g
n
itio
n
(
SS
FR
)
,
wh
er
e
o
n
l
y
a
s
in
g
le
f
ac
ial
im
ag
e
i
s
av
ailab
le
f
o
r
tr
ain
in
g
p
u
r
p
o
s
es
[
1
4
]
–
[
1
6
]
.
T
h
is
co
n
d
itio
n
c
o
m
m
o
n
ly
ar
is
es
in
s
ec
u
r
ity
s
y
s
tem
s
p
er
m
itti
n
g
o
n
ly
a
o
n
e
-
tim
e
r
eg
is
tr
atio
n
,
s
u
ch
as
elec
tr
o
n
ic
id
en
tity
ca
r
d
s
(
e
-
I
D)
,
p
ass
p
o
r
t
s
,
o
r
u
s
er
v
e
r
if
icatio
n
in
m
o
b
il
e
b
an
k
in
g
ap
p
licatio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
B
r
id
g
in
g
h
y
b
r
id
d
ee
p
lea
r
n
i
n
g
d
etec
tio
n
a
n
d
lig
h
tw
eig
h
t
h
a
n
d
cra
fted
fea
tu
r
es fo
r
… (
F
a
u
li
n
d
a
E
ly
N
a
s
titi
)
889
Alth
o
u
g
h
d
ee
p
lear
n
in
g
m
eth
o
d
s
h
av
e
d
em
o
n
s
tr
ated
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
in
f
ac
ial
r
ec
o
g
n
i
tio
n
task
s
[
1
7
]
–
[
1
9
]
,
th
ei
r
ef
f
ec
tiv
e
n
ess
d
ec
r
ea
s
es
s
ig
n
if
ican
tly
in
o
n
b
o
ar
d
a
p
p
licatio
n
s
u
n
d
er
SS
FR
co
n
d
itio
n
s
.
I
n
s
u
ch
ca
s
es,
th
e
m
o
d
els
s
u
f
f
er
f
r
o
m
p
o
o
r
g
e
n
er
aliza
tio
n
ac
r
o
s
s
v
a
r
iatio
n
s
in
lig
h
tin
g
,
p
o
s
e,
an
d
ex
p
r
ess
io
n
m
ak
i
n
g
SS
F
R
o
n
e
o
f
th
e
m
o
s
t
ch
allen
g
in
g
s
ce
n
ar
io
s
in
m
o
b
ile
-
b
ased
f
ac
e
r
ec
o
g
n
itio
n
[
2
0
]
,
[
2
1
]
.
SS
F
R
was
r
eg
ar
d
ed
as
an
ex
tr
e
m
e
ca
s
e
o
f
f
ew
-
s
h
o
t
lear
n
in
g
,
as
ea
ch
id
en
tity
wa
s
r
ep
r
esen
ted
b
y
o
n
l
y
a
s
in
g
le
f
ac
ial
im
ag
e
[
2
2
]
.
As
a
ch
allen
g
e
with
in
th
e
lo
w
-
d
ata
r
e
g
im
e,
SS
FR
r
eq
u
i
r
ed
s
p
ec
ialized
s
tr
ateg
ies
to
en
ab
le
m
o
d
els
to
g
en
er
alize
ef
f
ec
tiv
el
y
d
esp
ite
th
e
s
ca
r
city
o
f
d
ata
[
2
3
]
,
[
2
4
]
.
Ap
p
r
o
ac
h
es
s
u
ch
as
m
etr
ic
l
ea
r
n
in
g
,
t
h
e
u
s
e
o
f
lo
ca
l
f
ea
tu
r
es,
atten
tio
n
m
ec
h
an
is
m
s
,
an
d
ev
e
n
m
eta
-
lea
r
n
i
n
g
o
r
g
e
n
er
ativ
e
a
u
g
m
e
n
tatio
n
h
av
e
f
r
e
q
u
en
tly
b
ee
n
e
x
p
lo
r
e
d
.
Ho
wev
er
,
th
i
s
s
tu
d
y
p
r
o
p
o
s
ed
a
lig
h
ter
alter
n
ativ
e
b
y
f
o
cu
s
in
g
o
n
t
h
e
o
p
tim
izatio
n
o
f
d
etec
tio
n
an
d
f
ea
tu
r
e
e
x
tr
ac
tio
n
p
r
o
ce
s
s
es.
A
p
leth
o
r
a
o
f
s
tu
d
ies
h
av
e
p
r
e
v
io
u
s
ly
ex
a
m
in
ed
t
h
e
ef
f
icac
y
o
f
d
iv
er
s
e
f
ac
ial
r
ec
o
g
n
itio
n
m
o
d
els
in
ad
d
r
ess
in
g
th
e
SS
F
R
ch
all
en
g
e.
Ho
wev
er
,
th
e
attain
e
d
ac
cu
r
ac
ies
h
av
e
b
ee
n
d
ee
m
ed
in
ad
eq
u
ate.
C
o
m
p
ar
ativ
e
s
tu
d
ies
ev
alu
atin
g
m
o
d
els
f
o
r
SS
FR
h
av
e
r
ep
o
r
ted
ac
cu
r
ac
ies
o
f
p
r
in
ci
p
al
c
o
m
p
o
n
en
t
an
aly
s
is
(
PC
A
)
(
1
0
%),
lin
ea
r
d
is
cr
im
in
an
t
an
aly
s
is
(
L
DA
)
(
2
7
.
6
9
%
)
,
k
er
n
el
p
r
i
n
cip
al
co
m
p
o
n
en
t
an
aly
s
is
(
KPC
A
)
(
1
4
.
6
2
%),
k
e
r
n
el
f
is
h
er
an
aly
s
is
(
KFA
)
(
2
5
.
3
8
%),
r
eg
u
lar
i
ze
d
s
u
p
er
v
is
ed
L
DA
(
R
SLDA
)
(
5
7
.
4
6
%),
l
o
ca
lly
r
o
b
u
s
t
p
atter
n
p
r
o
p
a
g
atio
n
with
g
lo
b
al
r
eg
u
lar
izatio
n
r
e
d
u
ctio
n
(
L
R
PP
-
GR
R
)
(
5
7
.
4
3
%),
d
ee
p
k
-
n
ea
r
est
n
eig
h
b
o
r
s
(
KNN
)
(
3
7
.
6
9
%),
an
d
Dlib
f
ac
e
r
ec
o
g
n
itio
n
lib
r
ar
y
(
DL
I
B
)
d
ee
p
lear
n
i
n
g
(
6
3
.
2
8
%)
[
2
5
]
.
E
v
e
n
with
th
e
in
tr
o
d
u
ctio
n
o
f
m
o
r
e
ad
v
an
ce
d
m
o
d
els
s
u
ch
as,
Ar
cFac
e
attain
ed
5
2
%
ac
cu
r
ac
y
[
2
6
]
,
co
n
d
itio
n
al
g
en
er
ativ
e
ad
v
er
s
ar
ial
n
etwo
r
k
(
C
GAN
)
ac
h
iev
e
d
7
6
%
[
2
7
]
,
Mo
b
ileFace
Net
[
2
8
]
,
Gh
o
s
t
Face
Net
[
2
9
]
,
[
3
0
]
,
b
o
th
ap
p
r
o
ac
h
es
wer
e
n
o
t
o
r
i
g
in
ally
d
esig
n
ed
f
o
r
SS
FR
,
th
er
eb
y
lim
itin
g
th
eir
co
n
tr
ib
u
ti
o
n
s
to
m
u
lti
-
s
am
p
le
s
ce
n
ar
io
s
.
R
en
d
er
in
g
t
h
em
in
a
d
eq
u
ate
f
o
r
s
in
g
le
-
s
am
p
le
f
ac
i
al
r
ec
o
g
n
itio
n
task
s
.
T
h
er
ef
o
r
e
,
s
tr
en
g
th
en
in
g
th
e
s
tag
es o
f
f
ac
e
d
etec
tio
n
a
n
d
e
x
tr
ac
tio
n
b
ec
am
e
ess
en
tial p
r
io
r
to
th
e
em
b
e
d
d
in
g
p
r
o
ce
s
s
.
T
o
e
n
h
a
n
ce
S
S
F
R
p
e
r
f
o
r
m
a
n
c
e
,
p
r
ec
i
s
e
f
a
c
e
d
e
te
c
t
i
o
n
was
c
o
n
s
i
d
e
r
e
d
e
s
s
e
n
t
i
a
l
p
r
i
o
r
to
f
e
a
t
u
r
e
e
x
t
r
a
c
t
io
n
.
I
n
m
o
d
er
n
f
a
c
i
a
l
r
e
c
o
g
n
i
t
io
n
p
i
p
e
l
i
n
e
s
,
a
n
ch
o
r
b
o
x
-
b
a
s
e
d
ap
p
r
o
ac
h
e
s
w
er
e
a
d
o
p
t
e
d
to
en
a
b
le
m
u
l
t
i
-
s
c
a
le
a
n
d
m
u
l
t
i
-
a
s
p
e
c
t
r
a
t
i
o
p
r
e
d
ic
t
i
o
n
s
,
w
h
i
l
e
n
o
n
-
m
a
x
im
u
m
s
u
p
p
r
e
s
s
i
o
n
(
N
MS
)
w
a
s
a
p
p
l
i
ed
t
o
e
l
i
m
i
n
a
t
e
o
v
e
r
l
ap
p
in
g
p
r
e
d
ict
i
o
n
s
,
r
e
t
a
in
i
n
g
o
n
l
y
t
h
e
m
o
s
t
p
r
o
b
ab
l
e
c
an
d
id
a
t
e
s
[
3
1
]
.
T
h
e
o
u
t
p
u
t
s
f
r
o
m
d
e
t
e
c
t
io
n
r
em
a
i
n
ed
a
s
lo
c
a
l
iz
a
t
i
o
n
r
e
s
u
l
t
s
an
d
r
e
q
u
i
r
ed
f
u
r
t
h
e
r
p
r
o
c
e
s
s
i
n
g
t
h
r
o
u
g
h
f
ea
t
u
r
e
ex
t
r
a
c
t
io
n
to
r
e
p
r
e
s
e
n
t
f
ac
i
a
l
d
a
t
a
n
u
m
e
r
i
ca
l
l
y
.
A
t
t
h
i
s
s
t
a
g
e,
lo
c
a
l
b
in
ar
y
p
a
t
te
r
n
s
(
L
B
P
)
w
er
e
w
i
d
e
ly
em
p
lo
y
ed
d
u
e
t
o
t
h
e
i
r
s
im
p
l
i
c
i
ty
,
s
p
e
ed
,
an
d
r
o
b
u
s
t
n
e
s
s
ag
a
in
s
t
i
l
l
u
m
in
a
t
io
n
v
a
r
ia
t
i
o
n
s
[
3
2
]
.
S
tu
d
i
e
s
r
e
p
o
r
t
e
d
t
h
a
t
L
B
P
a
c
h
i
ev
e
d
ac
c
u
r
a
c
y
r
a
t
e
s
o
f
7
6
%
w
i
t
h
e
l
l
i
p
t
i
c
a
l
m
a
s
k
s
,
8
4
.
1
%
w
i
t
h
r
e
c
t
an
g
u
l
a
r
m
a
s
k
s
,
an
d
7
8
%
in
s
t
a
n
d
a
r
d
i
m
p
l
e
m
e
n
t
a
t
io
n
s
[
3
3
]
.
H
o
wev
e
r
,
L
B
P
w
a
s
s
t
i
l
l
l
i
m
i
t
e
d
u
n
d
er
e
x
tr
e
m
e
p
o
s
e
v
a
r
i
a
t
io
n
s
,
i
n
d
i
ca
t
i
n
g
t
h
a
t
i
t
s
i
n
t
e
g
r
a
t
i
o
n
w
i
th
m
o
d
e
r
n
a
n
ch
o
r
-
N
M
S
-
b
a
s
e
d
d
e
te
c
t
i
o
n
wa
s
v
i
e
w
ed
a
s
a
p
r
o
m
i
s
in
g
s
t
r
a
t
e
g
y
to
im
p
r
o
v
e
S
S
F
R
a
c
c
u
r
a
cy
.
P
r
e
v
i
o
u
s
s
t
u
d
i
e
s
h
a
d
p
r
im
a
r
il
y
f
o
c
u
s
ed
o
n
f
e
a
tu
r
e
e
m
b
e
d
d
i
n
g
o
r
d
a
t
a
au
g
m
e
n
t
a
ti
o
n
,
w
h
i
l
e
t
h
e
o
p
t
i
m
iz
a
t
i
o
n
o
f
a
n
ch
o
r
-
N
M
S
-
b
a
s
e
d
d
e
t
e
c
t
i
o
n
u
n
d
e
r
S
S
F
R
s
c
e
n
a
r
io
s
h
ad
n
o
t
b
e
en
ex
t
e
n
s
i
v
e
ly
ex
p
lo
r
ed
.
F
u
r
t
h
er
m
o
r
e,
a
l
th
o
u
g
h
L
B
P
h
ad
b
e
e
n
p
r
o
v
e
n
e
f
f
i
c
i
en
t
a
n
d
r
o
b
u
s
t
ag
a
i
n
s
t
i
l
l
u
m
in
a
t
i
o
n
c
h
an
g
e
s
,
i
t
s
a
p
p
l
i
ca
t
i
o
n
h
a
d
r
a
r
e
ly
b
e
en
s
t
r
a
t
e
g
i
ca
l
l
y
i
n
te
g
r
a
t
e
d
w
i
th
m
o
d
e
r
n
d
e
t
e
c
t
io
n
t
e
ch
n
iq
u
e
s
.
E
v
a
l
u
a
t
io
n
s
u
n
d
er
r
e
a
l
m
o
b
i
l
e
d
e
v
i
c
e
co
n
d
i
t
i
o
n
s
a
l
s
o
r
e
m
a
in
e
d
l
im
i
t
e
d
,
d
e
s
p
i
te
th
e
h
ig
h
d
em
an
d
f
o
r
b
io
m
etr
i
c
a
p
p
l
i
c
a
t
i
o
n
s
o
n
s
u
c
h
p
l
a
t
f
o
r
m
s
.
T
h
i
s
s
t
u
d
y
p
r
o
p
o
s
e
d
a
h
y
b
r
i
d
f
r
a
m
e
w
o
r
k
t
h
a
t
c
o
m
b
i
n
e
d
a
n
c
h
o
r
-
N
M
S
f
o
r
d
e
t
e
c
t
i
o
n
p
r
e
p
r
o
c
e
s
s
i
n
g
w
i
t
h
L
B
P
f
o
r
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
.
A
n
c
h
o
r
-
N
M
S
w
a
s
e
m
p
l
o
y
e
d
t
o
e
n
s
u
r
e
p
r
e
c
i
s
e
f
a
c
e
l
o
c
a
l
i
z
a
t
i
o
n
b
y
s
u
p
p
r
e
s
s
i
n
g
d
u
p
l
i
c
a
t
e
s
,
w
h
i
l
e
L
B
P
w
a
s
u
t
i
l
i
z
e
d
t
o
p
r
o
v
i
d
e
l
i
g
h
t
w
e
i
g
h
t
t
e
x
t
u
r
e
r
e
p
r
e
s
e
n
t
a
t
i
o
n
r
e
s
i
s
t
a
n
t
t
o
i
l
l
u
m
i
n
a
t
i
o
n
v
a
r
i
a
t
i
o
n
s
.
T
h
i
s
i
n
t
e
g
r
a
t
i
o
n
w
a
s
e
x
p
e
c
t
e
d
t
o
y
i
e
l
d
a
m
o
r
e
a
c
c
u
r
a
t
e
,
r
o
b
u
s
t
,
a
n
d
e
f
f
i
c
i
e
n
t
S
S
F
R
s
y
s
t
e
m
s
u
i
t
a
b
l
e
f
o
r
m
o
b
i
l
e
d
e
v
i
c
e
i
m
p
l
e
m
e
n
t
a
t
i
o
n
.
A
d
d
i
t
i
o
n
a
l
l
y
,
t
h
e
s
t
u
d
y
a
i
m
e
d
t
o
e
v
a
l
u
a
t
e
t
h
e
s
y
s
t
e
m
’
s
r
o
b
u
s
t
n
e
s
s
a
g
a
i
n
s
t
v
a
r
i
a
t
i
o
n
s
i
n
l
i
g
h
t
i
n
g
,
p
o
s
e
,
a
n
d
f
a
c
i
a
l
a
t
t
r
i
b
u
t
e
s
,
w
h
i
l
e
a
l
s
o
d
e
m
o
n
s
t
r
a
t
i
n
g
i
t
s
i
m
p
l
e
m
e
n
t
a
t
i
o
n
f
e
a
s
i
b
i
l
i
t
y
t
h
r
o
u
g
h
c
o
s
i
n
e
s
i
m
i
l
a
r
i
t
y
-
b
a
s
e
d
e
v
a
l
u
a
t
i
o
n
a
n
d
r
e
a
l
-
w
o
r
l
d
m
o
b
i
l
e
a
p
p
l
i
c
a
t
i
o
n
s
c
e
n
a
r
i
o
s
.
2.
RE
S
E
ARCH
M
E
T
H
O
D
2
.
1
.
M
a
t
er
ia
ls
T
h
e
u
s
e
o
f
a
SS
F
R
f
o
r
ea
ch
s
u
b
ject
h
as
b
ee
n
id
en
tifie
d
as
a
f
u
n
d
am
e
n
tal
lim
itatio
n
,
as
i
t
n
o
t
o
n
ly
in
cr
ea
s
es
th
e
lik
elih
o
o
d
o
f
f
alse
p
o
s
itiv
es
b
u
t
al
s
o
co
n
s
tr
ain
s
th
e
m
o
d
el’
s
ab
ilit
y
t
o
ac
h
iev
e
r
eliab
le
g
en
er
aliza
tio
n
.
T
o
o
v
er
c
o
m
e
t
h
is
co
n
d
itio
n
,
an
e
x
p
er
im
e
n
tal
d
esig
n
was
f
o
r
m
u
lated
as
s
h
o
w
n
in
Fig
u
r
e
1
.
S
o
th
at
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
alg
o
r
ith
m
co
u
l
d
b
e
test
ed
d
ir
ec
tly
in
a
lim
ited
d
ata
s
it
u
atio
n
,
with
r
esu
lts
m
ea
s
u
r
ed
th
r
o
u
g
h
ac
cu
r
ac
y
p
e
r
ce
n
tag
e,
d
etec
tio
n
er
r
o
r
r
ates,
an
d
co
m
p
u
tatio
n
al
e
f
f
icien
cy
.
T
o
en
s
u
r
e
t
h
at
th
e
m
o
d
el
h
as
a
h
ig
h
le
v
el
o
f
ac
c
u
r
ac
y
in
ac
co
r
d
an
ce
with
th
e
SS
FR
s
ce
n
ar
io
,
f
ac
e
d
etec
tio
n
ex
p
er
im
en
ts
wer
e
co
n
d
u
cte
d
in
a
Py
th
o
n
-
b
ase
d
v
ir
tu
al
la
b
o
r
ato
r
y
e
n
v
ir
o
n
m
en
t.
T
h
e
r
esear
ch
d
ataset
was
co
llected
d
ir
ec
tly
b
y
r
esear
ch
er
s
at
Un
iv
er
s
itas
Du
ta
B
an
g
s
a
Su
r
ak
ar
ta
.
T
h
e
d
ataset
co
n
s
is
ts
o
f
2
3
9
s
u
b
jects
allo
ca
ted
f
o
r
d
at
a
tr
ain
in
g
1
9
1
s
u
b
jects
an
d
d
ata
test
in
g
4
8
s
u
b
jects,
with
o
n
ly
o
n
e
f
ac
e
im
a
g
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
888
-
9
0
0
890
p
er
s
u
b
ject
in
ac
co
r
d
a
n
ce
with
th
e
SS
F
R
p
r
in
cip
le.
E
ac
h
im
ag
e
was
p
r
o
ce
s
s
ed
th
r
o
u
g
h
a
p
r
e
-
p
r
o
ce
s
s
in
g
s
tag
e
an
d
d
etec
ted
u
s
in
g
y
o
u
o
n
ly
lo
o
k
o
n
c
e
v
er
s
io
n
5
(
YOL
Ov
5
)
,
wh
ich
h
ad
b
ee
n
o
p
tim
is
ed
with
an
ch
o
r
-
NM
S.
I
n
ad
d
itio
n
,
4
8
n
ew
s
u
b
jects
f
r
o
m
Un
iv
er
s
itas
Du
ta
B
an
g
s
a
Su
r
ak
ar
ta
s
tu
d
en
ts
wer
e
also
co
llected
an
d
u
s
ed
e
x
clu
s
iv
ely
in
th
e
c
o
s
in
e
s
im
ilar
ity
-
b
ased
ev
alu
atio
n
s
tag
e.
W
ith
th
is
ap
p
r
o
ac
h
,
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
o
p
tim
is
ed
m
o
d
el
was
n
o
t
o
n
l
y
test
ed
f
o
r
d
etec
tio
n
ca
p
ab
ilit
ies
b
u
t
also
f
o
r
f
a
ce
v
er
if
icatio
n
an
d
id
en
tific
atio
n
p
r
o
ce
s
s
es.
T
h
is
s
tr
ateg
y
was
d
esig
n
ed
to
r
ep
r
esen
t
a
r
ea
lis
tic
s
ce
n
ar
io
o
f
S
SF
R
u
s
e
o
n
m
o
b
ile
d
ev
ices,
wh
er
e
tr
ain
in
g
d
ata
is
lim
ited
,
b
u
t
th
e
m
o
d
el
is
s
ti
ll
r
eq
u
ir
ed
to
b
e
a
b
le
to
g
en
e
r
alis
e
id
en
titi
es
th
at
h
av
e
n
o
t b
ee
n
i
n
v
o
lv
e
d
b
ef
o
r
e
.
All
p
r
o
ce
d
u
r
es
in
th
is
s
tu
d
y
wer
e
co
n
d
u
cted
in
c
o
m
p
l
ian
ce
with
r
ec
o
g
n
ize
d
eth
ic
al
r
esear
ch
s
tan
d
ar
d
s
.
T
h
e
f
ac
ial
im
ag
es
em
p
lo
y
ed
f
o
r
ex
p
e
r
im
en
tatio
n
wer
e
co
llected
with
in
f
o
r
m
ed
co
n
s
en
t
f
r
o
m
th
e
p
ar
ticip
an
ts
an
d
wer
e
u
s
ed
ex
clu
s
iv
ely
f
o
r
ac
ad
em
ic
a
n
d
s
cien
tific
p
u
r
p
o
s
es.
T
h
r
o
u
g
h
o
u
t
th
e
p
r
o
ce
s
s
o
f
d
at
a
co
llectio
n
an
d
a
n
aly
s
is
,
co
n
f
i
d
en
tiality
was
m
ain
tain
e
d
,
p
o
ten
tial
r
is
k
s
wer
e
m
in
i
m
ized
,
an
d
th
e
r
ig
h
ts
o
f
p
ar
ticip
an
ts
wer
e
f
u
lly
r
esp
ec
ted
in
lin
e
with
eth
ical
r
esear
c
h
p
r
in
cip
les.
2
.
2
.
M
et
ho
ds
T
h
e
p
r
o
ce
s
s
b
eg
in
s
as
s
h
o
wn
in
Fig
u
r
e
1
with
f
ac
ia
l
im
ag
es
th
at
ar
e
p
r
o
ce
s
s
ed
th
r
o
u
g
h
p
r
e
-
p
r
o
ce
s
s
in
g
an
d
d
etec
tio
n
s
tag
es
u
s
in
g
YOL
Ov
5
,
wh
ich
is
o
p
tim
is
ed
u
s
in
g
an
c
h
o
r
-
NM
S.
T
h
e
d
etec
tio
n
r
esu
lts
ar
e
th
en
ex
tr
ac
ted
u
s
i
n
g
R
GB
co
lo
r
m
ap
p
in
g
co
m
b
in
e
with
L
B
P
an
d
p
r
o
jecte
d
in
to
v
ec
to
r
s
p
ac
e,
wh
ich
is
s
u
b
s
eq
u
en
tly
u
s
ed
in
th
e
test
in
g
p
r
o
ce
s
s
with
b
o
th
tr
ain
in
g
an
d
test
d
ata.
Nex
t,
a
n
ev
alu
atio
n
b
ased
o
n
co
s
in
e
s
im
ilar
ity
an
d
o
n
b
o
a
r
d
m
o
b
ile
is
p
er
f
o
r
m
ed
.
Fig
u
r
e
1
.
T
h
e
p
ip
elin
e
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
m
o
d
elin
g
Ph
ase
o
n
e
aim
ed
to
m
in
im
ize
b
ias
an
d
en
h
an
ce
g
e
n
er
aliza
tio
n
,
all
im
ag
es
wer
e
co
llected
th
r
o
u
g
h
a
r
an
d
o
m
ize
d
s
am
p
lin
g
p
r
o
ce
s
s
co
v
er
in
g
t
h
r
ee
f
ac
ial
co
n
d
itio
n
s
,
as sh
o
wn
in
Fig
u
r
e
2
.
Fig
u
r
e
2
(
a)
f
r
o
n
tal
an
g
le
ey
ewe
ar
,
Fig
u
r
e
2
(
b
)
h
ijab
s
lig
h
t
an
g
le,
a
n
d
Fig
u
r
e
2
(
c
)
f
r
o
n
tal
an
g
le
an
g
le
with
o
u
t
ad
d
itio
n
al
attr
ib
u
tes.
T
h
e
r
an
d
o
m
s
elec
tio
n
en
s
u
r
ed
th
at
ea
ch
s
u
b
ject
h
ad
an
eq
u
al
p
r
o
b
ab
ilit
y
o
f
r
ep
r
esen
tatio
n
ac
r
o
s
s
d
if
f
er
en
t
p
o
s
e
v
ar
iatio
n
s
,
p
r
e
v
en
tin
g
o
v
e
r
f
i
ttin
g
to
war
d
s
p
ec
if
ic
f
ac
ial
o
r
ien
tatio
n
s
o
r
attr
ib
u
tes.
T
h
is
ap
p
r
o
ac
h
also
s
im
u
lates
r
ea
lis
tic
d
ep
lo
y
m
en
t
s
ce
n
ar
io
s
o
f
SS
FR
,
wh
er
e
s
y
s
tem
u
s
er
s
m
ay
ap
p
ea
r
with
d
iv
er
s
e
v
is
u
al
co
n
d
itio
n
s
th
at
ca
n
n
o
t b
e
p
r
e
d
eter
m
in
ed
d
u
r
in
g
m
o
d
el
tr
ain
i
n
g
.
(
a)
(
b
)
(
c)
Fig
u
r
e
2
.
Data
s
et
f
ac
e
co
n
d
itio
n
s
am
p
le
of
(
a)
f
r
o
n
tal
a
n
g
le
ey
ewe
ar
,
(
b
)
h
ijab
s
lig
h
t a
n
g
le
,
an
d
(
c
)
f
r
o
n
tal
an
g
le
an
g
le
with
o
u
t a
d
d
itio
n
al
attr
ib
u
tes
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
B
r
id
g
in
g
h
y
b
r
id
d
ee
p
lea
r
n
i
n
g
d
etec
tio
n
a
n
d
lig
h
tw
eig
h
t
h
a
n
d
cra
fted
fea
tu
r
es fo
r
… (
F
a
u
li
n
d
a
E
ly
N
a
s
titi
)
891
Ph
ase
two
,
f
o
cu
s
ed
o
n
p
r
e
p
r
o
ce
s
s
in
g
o
p
tim
izatio
n
to
en
s
u
r
e
ac
cu
r
ate
f
ac
e
d
etec
tio
n
an
d
co
n
s
is
ten
t
ex
tr
ac
tio
n
.
As
illu
s
tr
ated
in
Fig
u
r
e
3
,
th
e
p
r
o
ce
s
s
b
eg
an
w
ith
th
e
ap
p
licatio
n
o
f
a
n
ch
o
r
b
o
x
es
to
g
en
e
r
ate
m
u
ltip
le
b
o
u
n
d
in
g
b
o
x
p
r
o
p
o
s
als
w
ith
v
ar
y
in
g
s
izes
an
d
asp
ec
t
r
atio
s
ac
co
r
d
in
g
to
th
e
s
h
ap
e
o
f
th
e
f
ac
e.
T
h
ese
p
r
o
p
o
s
als
wer
e
s
u
b
s
eq
u
en
tly
f
ilter
ed
u
s
in
g
NM
S,
wh
ich
r
em
o
v
ed
o
v
er
lap
p
i
n
g
o
r
lo
w
-
co
n
f
id
e
n
ce
b
o
u
n
d
in
g
b
o
x
es,
leav
in
g
o
n
ly
th
e
m
o
s
t r
ep
r
esen
tativ
e
d
etec
tio
n
f
o
r
ea
c
h
f
ac
e.
Af
ter
o
p
tim
izatio
n
,
th
e
d
etec
te
d
f
ac
ial
r
eg
io
n
s
wer
e
cr
o
p
p
e
d
f
r
o
m
th
e
o
r
ig
i
n
al
im
ag
es
an
d
p
r
ep
ar
e
d
f
o
r
t
h
e
f
o
llo
w
in
g
s
tep
s
,
s
u
ch
as
n
o
r
m
alis
atio
n
an
d
f
e
atu
r
e
e
x
tr
ac
tio
n
.
T
h
is
r
ef
in
em
e
n
t
en
s
u
r
ed
t
h
at
o
n
l
y
th
e
r
elev
an
t
f
ac
e
r
eg
io
n
s
wer
e
p
r
o
ce
s
s
ed
,
r
ed
u
cin
g
b
ac
k
g
r
o
u
n
d
n
o
is
e
an
d
im
p
r
o
v
in
g
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
.
Fig
u
r
e
3
.
I
m
ag
e
p
r
ep
r
o
ce
s
s
in
g
o
p
tim
izatio
n
s
tep
s
T
o
ca
r
r
y
o
u
t
th
is
d
etec
tio
n
p
r
o
ce
s
s
,
YOL
Ov
5
was
em
p
l
o
y
ed
e
x
clu
s
iv
ely
as
t
h
e
f
ac
e
d
etec
tio
n
m
o
d
u
le.
W
h
ile
Fig
u
r
e
3
illu
s
tr
ates
th
e
s
eq
u
en
tial
p
r
ep
r
o
ce
s
s
in
g
s
tep
s
,
Fig
u
r
e
4
p
r
esen
ts
th
e
YOL
Ov
5
ar
ch
itectu
r
e
th
at
en
a
b
led
th
is
s
tag
e.
T
h
e
a
r
ch
itectu
r
e
co
n
s
is
ted
o
f
th
r
ee
s
eq
u
en
tial
c
o
m
p
o
n
en
ts
:
a
b
ac
k
b
o
n
e
f
o
r
h
ier
ar
ch
ical
f
ea
tu
r
e
ex
tr
ac
tio
n
,
a
n
ec
k
(
p
at
h
ag
g
r
eg
atio
n
n
etwo
r
k
(
PANet
)
)
f
o
r
ag
g
r
eg
atin
g
m
u
lti
-
s
ca
le
f
ea
tu
r
es,
an
d
a
h
e
ad
f
o
r
g
e
n
er
atin
g
th
e
f
i
n
al
b
o
u
n
d
in
g
b
o
x
p
r
ed
ictio
n
s
.
I
n
th
is
s
tu
d
y
,
YOL
Ov
5
was
o
p
tim
is
ed
with
an
c
h
o
r
-
N
MS
to
ac
h
iev
e
r
o
b
u
s
t
f
ac
e
lo
c
alis
atio
n
with
o
u
t
ex
ten
d
in
g
its
f
u
n
ctio
n
ality
to
th
e
r
ec
o
g
n
itio
n
s
tag
e.
Fig
u
r
e
4
.
YOL
Ov
5
s
tep
s
P
h
a
s
e
th
r
e
e
in
v
o
l
v
ed
t
h
e
tr
a
n
s
f
o
r
m
a
t
i
o
n
o
f
cr
o
p
p
ed
R
G
B
f
a
c
e
i
m
ag
e
s
in
t
o
d
i
s
c
r
i
m
i
n
a
t
i
v
e
n
u
m
e
r
i
c
a
l
r
e
p
r
e
s
e
n
t
a
t
io
n
s
.
T
o
a
c
h
i
e
v
e
t
h
i
s
,
L
B
P
w
e
r
e
ap
p
l
ie
d
t
o
th
e
g
r
a
y
s
c
a
l
e
-
co
n
v
e
r
t
ed
i
m
a
g
e
s
t
o
ex
t
r
ac
t
t
e
x
t
u
r
e
-
b
a
s
e
d
d
e
s
cr
i
p
to
r
s
.
T
h
i
s
m
e
th
o
d
co
m
p
a
r
e
s
th
e
in
t
e
n
s
i
t
y
o
f
e
a
ch
ce
n
tr
a
l
p
i
x
e
l
w
i
t
h
th
a
t
o
f
i
t
s
n
e
i
g
h
b
o
r
in
g
p
ix
e
l
s
,
p
r
o
d
u
c
in
g
a
r
o
b
u
s
t
lo
c
a
l
t
e
x
tu
r
e
s
i
g
n
a
t
u
r
e
.
T
h
e
b
i
n
ar
y
co
d
e
s
g
en
e
r
a
t
ed
f
r
o
m
t
h
i
s
p
r
o
c
e
s
s
w
e
r
e
co
n
v
e
r
t
ed
i
n
to
n
u
m
e
r
i
c
a
l
ar
r
ay
s
a
n
d
s
u
b
s
e
q
u
e
n
t
ly
n
o
r
m
a
l
i
s
e
d
u
s
i
n
g
th
e
L
2
m
e
th
o
d
t
o
f
o
r
m
u
n
i
t
-
le
n
g
t
h
f
e
a
tu
r
e
v
e
c
to
r
s
.
T
h
e
d
e
c
i
s
i
o
n
t
o
u
s
e
L
B
P
wa
s
b
a
s
e
d
o
n
it
s
c
o
m
p
u
ta
t
i
o
n
a
l
e
f
f
i
c
i
en
c
y
a
n
d
s
u
i
t
a
b
i
l
i
t
y
f
o
r
m
o
b
i
l
e
d
ep
l
o
y
m
e
n
t
.
Un
l
i
k
e
co
n
v
o
lu
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
wo
r
k
(
C
N
N
)
-
b
a
s
e
d
f
e
a
tu
r
e
ex
t
r
a
c
t
o
r
s
,
w
h
i
c
h
t
y
p
i
c
a
l
l
y
d
e
m
an
d
h
i
g
h
p
r
o
c
e
s
s
i
n
g
p
o
w
er
an
d
m
e
m
o
r
y
,
L
B
P
p
r
o
d
u
c
e
d
l
i
g
h
t
w
e
ig
h
t
d
e
s
c
r
i
p
to
r
s
w
i
t
h
l
o
w
c
o
m
p
l
ex
i
t
y
,
th
e
r
e
b
y
r
ed
u
c
i
n
g
i
n
f
e
r
en
c
e
t
i
m
e
w
h
i
l
e
p
r
e
s
e
r
v
i
n
g
d
i
s
cr
i
m
in
a
t
i
v
e
ca
p
a
b
i
l
i
ty
.
T
h
i
s
m
a
d
e
L
B
P
p
a
r
t
i
cu
l
a
r
ly
r
e
l
ev
an
t
i
n
t
h
e
c
o
n
t
ex
t
o
f
S
S
F
R
,
w
h
e
r
e
b
o
t
h
e
f
f
i
c
i
e
n
c
y
an
d
g
en
e
r
a
l
i
s
a
t
i
o
n
w
er
e
cr
i
t
i
c
al
u
n
d
e
r
r
e
s
o
u
r
c
e
-
c
o
n
s
t
r
a
in
e
d
en
v
i
r
o
n
m
e
n
t
s
.
T
h
e
n
o
r
m
a
l
i
s
ed
f
e
a
t
u
r
e
v
e
c
t
o
r
s
t
h
en
s
e
r
v
e
d
as
t
h
e
f
o
u
n
d
a
t
io
n
f
o
r
t
h
e
t
r
a
i
n
in
g
a
n
d
t
e
s
t
i
n
g
s
t
ag
e
s
.
Ph
ase
f
o
u
r
o
u
tlin
ed
th
e
tr
ain
i
n
g
an
d
test
in
g
p
r
o
ce
d
u
r
e
o
f
th
e
r
ec
o
g
n
itio
n
m
o
d
el
u
s
in
g
L
B
P
f
ea
tu
r
e
v
ec
to
r
s
.
Fro
m
2
3
9
s
u
b
jects,
1
9
1
wer
e
allo
ca
ted
f
o
r
t
r
ain
in
g
an
d
4
8
f
o
r
test
in
g
,
with
o
n
e
i
m
ag
e
p
er
s
u
b
ject
to
f
o
llo
w
th
e
SS
F
R
p
r
in
cip
le.
T
r
ain
in
g
was
co
n
d
u
cted
in
a
Py
th
o
n
-
b
ased
en
v
ir
o
n
m
en
t
f
o
r
u
p
to
7
2
ep
o
c
h
s
.
Per
f
o
r
m
an
ce
was
m
o
n
it
o
r
ed
u
s
in
g
f
o
u
r
m
etr
ics:
tr
ain
in
g
l
o
s
s
,
v
alid
atio
n
lo
s
s
,
tr
ain
in
g
a
cc
u
r
ac
y
,
an
d
test
in
g
ac
cu
r
ac
y
.
An
ea
r
ly
s
to
p
p
in
g
m
ec
h
an
is
m
was
ap
p
lied
o
n
ce
v
alid
atio
n
p
er
f
o
r
m
an
ce
s
t
ab
ilis
ed
to
p
r
ev
en
t
o
v
er
f
itti
n
g
.
L
o
g
s
an
d
p
lo
ts
w
er
e
g
en
er
ate
d
th
r
o
u
g
h
o
u
t
th
e
p
r
o
ce
s
s
to
p
r
o
v
id
e
d
ata
f
o
r
an
aly
s
is
in
th
e
r
esu
lts
an
d
d
is
cu
s
s
io
n
s
ec
tio
n
.
I
n
ad
d
itio
n
,
to
v
alid
ate
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
p
r
o
p
o
s
ed
o
p
tim
izatio
n
,
an
ab
latio
n
s
tu
d
y
was
co
n
d
u
cte
d
with
th
r
ee
co
n
f
ig
u
r
atio
n
s
:
i)
o
r
ien
ted
f
ea
tu
r
es
f
r
o
m
ac
ce
ler
ated
s
eg
m
en
t
test
an
d
r
o
tated
b
i
n
ar
y
r
o
b
u
s
t
in
d
e
p
en
d
e
n
t
elem
en
tar
y
f
ea
tu
r
es
(
OR
B
)
as
a
b
aseli
n
e
n
o
n
-
f
ac
e
d
escr
ip
to
r
,
ii)
M
o
b
ileFace
Net
with
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
888
-
9
0
0
892
Ar
cFac
e
lo
s
s
,
an
d
iii)
th
e
h
y
b
r
id
an
ch
o
r
-
NM
S
an
d
L
B
P.
T
h
is
allo
wed
th
e
co
n
tr
ib
u
tio
n
o
f
ea
ch
co
m
p
o
n
en
t
to
b
e
ex
am
in
ed
s
ep
ar
ately
.
A
c
o
m
p
ar
ativ
e
an
aly
s
is
ag
ain
s
t
p
r
io
r
ap
p
r
o
ac
h
es
was
also
in
clu
d
ed
to
h
ig
h
lig
h
t
p
er
f
o
r
m
an
ce
g
ai
n
s
u
n
d
e
r
th
e
S
SF
R
s
ce
n
ar
io
.
Ph
ase
f
iv
e
in
v
o
lv
ed
ev
al
u
atin
g
th
e
r
ec
o
g
n
itio
n
m
o
d
el
u
s
in
g
a
co
s
in
e
s
im
ilar
ity
ap
p
r
o
ac
h
.
I
n
th
is
s
tag
e,
4
8
ad
d
itio
n
al
s
u
b
jects,
n
o
t
in
clu
d
ed
in
th
e
tr
ain
in
g
o
r
test
in
g
s
ets,
wer
e
co
llected
to
ass
ess
th
e
s
y
s
tem
’
s
v
er
if
icatio
n
an
d
id
en
tific
atio
n
ca
p
ab
ilit
ies
u
n
d
er
SS
F
R
co
n
d
itio
n
s
.
C
o
s
in
e
s
im
i
lar
ity
was
ap
p
lied
to
co
m
p
ar
e
th
e
f
ea
tu
r
e
v
ec
t
o
r
s
o
f
q
u
er
y
i
m
ag
es
with
th
o
s
e
s
to
r
ed
in
th
e
r
ef
er
en
ce
d
atab
ase.
A
s
im
ilar
ity
s
co
r
e
clo
s
e
to
1
in
d
icate
d
a
m
atch
,
wh
ile
lo
we
r
v
alu
es
in
d
icate
d
d
is
s
im
ilar
it
y
.
T
h
is
p
r
o
ce
d
u
r
e
en
ab
le
d
th
e
s
y
s
tem
to
b
e
test
ed
o
n
u
n
s
ee
n
id
en
titi
es,
p
r
o
v
id
in
g
a
r
ea
lis
tic
ev
alu
atio
n
o
f
r
o
b
u
s
tn
ess
,
g
en
er
alis
atio
n
,
an
d
f
ea
s
ib
ilit
y
f
o
r
m
o
b
ile
-
b
ased
f
ac
e
r
ec
o
g
n
itio
n
.
Ph
ase
s
ix
,
as
th
e
f
in
al
ev
al
u
atio
n
,
was
co
n
d
u
cte
d
in
a
n
in
teg
r
ated
m
an
n
er
th
r
o
u
g
h
a
m
o
b
ile
ap
p
licatio
n
d
e
v
elo
p
ed
f
o
r
th
e
An
d
r
o
id
a
n
d
iOS
p
latf
o
r
m
s
.
T
h
e
test
in
g
s
ce
n
ar
io
f
o
llo
wed
S
SF
R
,
wh
er
eb
y
o
n
e
im
ag
e
p
er
id
e
n
tity
was
u
s
ed
as
an
en
r
o
lm
e
n
t
tem
p
late.
Sin
ce
SS
FR
d
o
es
n
o
t
p
r
o
v
i
d
e
a
d
d
itio
n
al
im
ag
es
to
ad
ap
t
to
ch
an
g
es
in
illu
m
in
a
tio
n
,
lig
h
tin
g
v
ar
iatio
n
s
wer
e
id
en
tifie
d
as
th
e
m
ain
co
n
f
o
u
n
d
in
g
f
ac
to
r
th
at
n
ee
d
ed
t
o
b
e
test
ed
f
o
r
r
esil
ien
ce
.
T
h
e
test
d
ata
co
n
s
is
ted
o
f
4
8
im
ag
es
o
f
s
tu
d
en
ts
f
r
o
m
Un
iv
er
s
itas
Du
ta
B
an
g
s
a
Su
r
ak
ar
ta
wh
o
wer
e
n
o
t
in
v
o
l
v
ed
in
th
e
p
r
ev
io
u
s
m
o
d
el
tr
ain
in
g
o
r
test
in
g
p
r
o
ce
s
s
.
E
ac
h
im
ag
e
was
ev
alu
ated
in
two
illu
m
in
atio
n
r
an
g
es:
i)
n
o
r
m
al
≈
3
5
0
-
4
0
0
lu
x
an
d
ii)
d
im
≈
5
0
-
1
0
0
lu
x
.
I
llu
m
in
atio
n
v
al
u
es
wer
e
m
ea
s
u
r
ed
u
s
in
g
a
lu
x
m
eter
o
n
th
e
f
ac
e
ar
ea
to
e
n
s
u
r
e
r
an
g
e
s
u
itab
ilit
y
.
T
h
e
ac
cu
r
ac
y
ev
alu
atio
n
p
r
o
ce
d
u
r
e
f
o
r
th
e
o
n
b
o
a
r
d
m
o
b
ile
f
ac
e
r
ec
o
g
n
itio
n
s
y
s
tem
,
as
d
ef
in
ed
in
(
1
)
,
is
s
im
p
ly
th
e
r
atio
o
f
co
r
r
ec
t
p
r
ed
ictio
n
s
to
th
e
to
tal
n
u
m
b
er
o
f
s
am
p
les u
n
d
er
ea
ch
illu
m
in
atio
n
co
n
d
itio
n
.
=
×
100%
(
1
)
W
h
er
e
T
r
ep
r
esen
ts
th
e
n
u
m
b
er
o
f
co
r
r
ec
t
p
r
ed
ictio
n
s
an
d
N
r
ep
r
esen
ts
th
e
n
u
m
b
er
o
f
s
am
p
les
in
a
g
iv
en
test
co
n
d
itio
n
.
I
n
t
h
e
id
en
tific
atio
n
s
ch
em
e
(
1
:N
)
,
a
p
r
ed
ictio
n
is
co
n
s
id
er
ed
co
r
r
ec
t
if
th
e
f
a
ce
s
ca
n
n
ed
b
y
th
e
m
o
b
ile
ap
p
licatio
n
m
atch
es th
e
co
r
r
ec
t la
b
el;
in
t
h
e
v
er
if
icat
io
n
s
ch
em
e
(
1
:1
)
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
I
m
a
g
e
prepro
ce
s
ing
o
pt
im
iza
t
io
n
T
h
e
an
c
h
o
r
b
ox
i
n
itializatio
n
p
r
o
ce
s
s
is
an
im
p
o
r
tan
t
in
itial
s
tep
in
th
e
SS
FR
b
ased
f
ac
ial
d
etec
tio
n
s
y
s
tem
.
T
h
e
m
o
d
el
m
u
s
t
as
ce
r
tain
th
e
s
tar
tin
g
p
o
s
itio
n
an
d
d
im
e
n
s
io
n
s
o
f
th
e
b
o
u
n
d
in
g
b
o
x
p
r
io
r
to
ad
d
itio
n
al
r
eg
r
ess
io
n
an
d
f
ilter
in
g
p
r
o
ce
s
s
es.
T
h
e
b
o
u
n
d
in
g
b
o
x
r
eg
r
ess
io
n
f
o
r
m
u
la
u
s
ed
(
2
)
to
(
5
)
[
3
4
]
.
=
(
)
+
(
2
)
=
(
)
+
(
3
)
=
(
4
)
ℎ
=
ℎ
ℎ
(
5
)
I
n
th
is
co
n
tex
t,
an
d
in
d
icate
s
th
e
co
o
r
d
i
n
ates o
f
th
e
p
r
ed
icte
d
ce
n
ter
,
wh
ile
an
d
ℎ
s
p
ec
if
ie
s
th
e
d
im
en
s
io
n
s
o
f
th
e
p
r
ed
ict
ed
b
o
u
n
d
in
g
b
o
x
.
Ad
d
itio
n
all
y
,
,
,
,
an
d
ℎ
ar
e
th
e
p
ar
a
m
eter
s
p
r
ed
icted
b
y
th
e
m
o
d
el.
T
h
e
s
y
m
b
o
l
σ
r
ep
r
esen
ts
th
e
s
ig
m
o
id
ac
tiv
at
io
n
f
u
n
ctio
n
,
wh
ich
en
s
u
r
es
t
h
at
th
e
c
o
o
r
d
in
ates
s
tay
with
in
a
v
alid
r
an
g
e.
T
h
i
s
r
e
s
e
a
r
c
h
e
m
p
l
o
y
s
a
s
c
a
l
e
o
f
1
t
o
e
n
s
u
r
e
t
h
a
t
t
h
e
i
n
i
t
i
a
l
b
o
u
n
d
i
n
g
b
o
x
s
i
z
e
r
e
m
a
i
n
s
p
r
o
p
o
r
t
i
o
n
a
l
t
o
t
h
e
f
e
a
t
u
r
e
m
a
p
.
T
h
i
s
a
p
p
r
o
a
c
h
w
a
s
s
e
l
e
c
t
e
d
b
e
c
a
u
s
e
t
h
e
f
a
c
i
a
l
i
m
a
g
e
s
i
n
t
h
e
d
a
t
a
s
e
t
a
r
e
o
f
u
n
i
f
o
r
m
s
i
z
e
,
w
h
i
c
h
a
l
l
o
w
s
t
h
e
m
o
d
e
l
t
o
o
p
e
r
a
t
e
m
o
r
e
s
t
a
b
l
y
a
n
d
e
f
f
i
c
i
e
n
t
l
y
w
i
t
h
o
u
t
r
e
q
u
i
r
i
n
g
s
i
g
n
i
f
i
c
a
n
t
a
d
j
u
s
t
m
e
n
t
s
t
o
t
h
e
b
o
u
n
d
i
n
g
b
o
x
s
i
z
e
.
T
h
e
a
n
c
h
o
r
b
o
x
u
s
e
d
i
n
t
h
i
s
e
x
p
e
r
i
m
e
n
t
m
a
i
n
t
a
i
n
s
a
1
:
1
r
a
t
i
o
,
w
i
t
h
a
n
i
n
i
t
i
a
l
s
i
z
e
o
f
6
0
×
6
0
p
i
x
e
l
s
o
n
t
h
e
f
e
a
t
u
r
e
m
a
p
.
A
f
t
e
r
i
n
i
t
i
a
l
i
z
a
t
i
o
n
,
t
h
e
m
o
d
e
l
a
d
j
u
s
t
s
t
h
e
p
o
s
i
t
i
o
n
a
n
d
s
i
z
e
o
f
t
h
e
b
o
u
n
d
i
n
g
b
o
x
t
h
r
o
u
g
h
c
o
o
r
d
i
n
a
t
e
t
r
a
n
s
f
o
r
m
a
t
i
o
n
-
b
a
s
e
d
r
e
g
r
e
s
s
i
o
n
.
T
h
i
s
t
r
a
n
s
f
o
r
m
a
t
i
o
n
u
t
i
l
i
z
e
s
s
i
g
m
o
i
d
a
n
d
e
x
p
o
n
e
n
t
i
a
l
f
u
n
c
t
i
o
n
s
t
o
e
n
s
u
r
e
t
h
a
t
t
h
e
b
o
u
n
d
i
n
g
b
o
x
v
a
l
u
e
s
s
t
a
y
w
i
t
h
i
n
a
v
a
l
i
d
r
a
n
g
e
.
T
h
e
r
e
s
u
l
t
s
i
n
d
i
c
a
t
e
t
h
a
t
t
h
e
b
o
u
n
d
i
n
g
b
o
x
,
f
o
l
l
o
w
i
n
g
r
e
g
r
e
s
s
i
o
n
,
h
a
s
c
e
n
t
e
r
c
o
o
r
d
i
n
a
t
e
s
o
f
(
1
0
1
.
0
5
,
1
0
0
.
9
8
)
.
I
t
s
d
i
m
e
n
s
i
o
n
s
h
a
v
e
c
h
a
n
g
e
d
t
o
1
4
6
.
5
7
×
1
0
8
.
5
8
p
i
x
e
l
s
,
w
h
i
c
h
s
t
i
l
l
c
l
o
s
e
l
y
a
p
p
r
o
x
i
m
a
t
e
s
t
h
e
i
n
i
t
i
a
l
1
:
1
r
a
t
i
o
.
Fig
u
r
e
5
s
h
o
ws an
ex
p
er
im
en
t
r
esu
lt o
f
th
r
ee
r
ec
tan
g
u
lar
a
n
c
h
o
r
b
o
x
es th
at
ar
e
in
ten
tio
n
ally
ce
n
ter
ed
r
ig
h
t
o
n
th
e
s
u
b
ject’
s
f
ac
e
to
i
llu
s
tr
ate
th
e
co
n
ce
p
t
o
f
a
n
ch
o
r
p
lace
m
en
t
b
ef
o
r
e
th
e
in
f
er
e
n
ce
p
r
o
ce
s
s
b
eg
in
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
B
r
id
g
in
g
h
y
b
r
id
d
ee
p
lea
r
n
i
n
g
d
etec
tio
n
a
n
d
lig
h
tw
eig
h
t
h
a
n
d
cra
fted
fea
tu
r
es fo
r
… (
F
a
u
li
n
d
a
E
ly
N
a
s
titi
)
893
T
h
e
b
o
x
es
ar
e
o
f
d
if
f
e
r
en
t
s
ca
les:
s
m
all,
m
ed
iu
m
,
an
d
la
r
g
e
with
a
1
:1
r
atio
,
s
o
th
at
ea
ch
s
ca
le
ca
n
r
ep
r
esen
t
th
e
f
ac
e
at
d
if
f
er
e
n
t
ca
m
er
a
d
i
s
tan
ce
s
.
T
h
e
s
m
allest
an
ch
o
r
(
tu
r
q
u
o
is
e)
c
o
v
er
s
th
e
n
ar
r
o
we
s
t
ar
ea
o
f
th
e
f
ac
e,
th
e
m
ed
iu
m
an
ch
o
r
(
y
ello
w)
o
v
er
lap
s
with
th
e
g
r
o
u
n
d
-
tr
u
t
h
,
wh
ile
th
e
lar
g
est
an
ch
o
r
(
r
ed
)
s
lig
h
tly
e
x
ten
d
s
b
ey
o
n
d
th
e
c
o
n
to
u
r
o
f
th
e
f
ac
e
.
Fig
u
r
e
5
.
An
c
h
o
r
b
o
x
T
h
e
3
an
ch
o
r
b
o
x
es g
en
er
ate
d
ar
e
th
en
p
r
o
ce
s
s
ed
with
th
e
NM
S
m
eth
o
d
.
I
n
NM
S,
o
n
ly
o
n
e
b
o
x
with
th
e
h
ig
h
est
co
n
f
id
e
n
ce
s
co
r
e
an
d
in
ter
s
ec
tio
n
o
v
e
r
u
n
io
n
(
I
o
U
)
in
(
6
)
[
3
5
]
is
s
elec
te
d
af
ter
th
e
n
etwo
r
k
r
eg
r
ess
es
ea
ch
an
ch
o
r
an
d
ca
lcu
lates
th
e
co
n
f
id
en
ce
s
co
r
e
.
T
h
is
m
eth
o
d
also
r
em
o
v
es
o
v
er
lap
p
i
n
g
b
o
x
es
ab
o
v
e
a
ce
r
tain
lev
el
.
T
h
er
ef
o
r
e,
ev
en
th
o
u
g
h
th
er
e
ar
e
th
r
e
e
in
itial
an
ch
o
r
s
,
th
e
d
etec
to
r
u
ltima
tely
p
r
o
d
u
ce
s
o
n
ly
o
n
e
co
r
r
ec
t f
ac
e
b
o
x
.
=
(
6
)
Af
ter
th
e
NM
S
p
r
o
ce
s
s
is
co
m
p
lete,
th
e
n
ex
t
s
tep
in
v
o
l
v
es
ex
tr
ac
tin
g
th
e
b
o
u
n
d
i
n
g
b
o
x
co
o
r
d
in
ates
an
d
p
er
f
o
r
m
in
g
im
ag
e
cr
o
p
p
in
g
to
is
o
late
th
e
d
etec
ted
f
ac
ial
r
eg
io
n
s
.
T
h
e
r
esu
lts
o
f
th
is
p
r
o
ce
s
s
,
as
s
h
o
wn
in
Fig
u
r
e
6
,
illu
s
tr
ate
th
e
o
p
tim
i
ze
d
o
u
tp
u
ts
o
f
th
e
p
r
e
p
r
o
ce
s
s
in
g
s
tag
e.
E
ac
h
cr
o
p
p
ed
im
a
g
e
r
ep
r
esen
ts
a
f
ac
e
th
at
h
as
b
ee
n
co
n
s
is
ten
tly
lo
c
alize
d
,
r
eg
a
r
d
less
o
f
v
ar
iatio
n
s
in
o
r
ien
tatio
n
o
r
v
is
u
al
attr
i
b
u
tes.
Sp
ec
if
ically
,
Fig
u
r
e
6
(
a)
d
em
o
n
s
tr
ates
th
e
s
u
cc
ess
f
u
l
lo
ca
lizatio
n
o
f
a
f
ac
e
with
ey
ewe
ar
,
wh
ile
Fig
u
r
e
6
(
b
)
s
h
o
ws
th
e
r
esu
lt
o
n
a
s
u
b
ject
wea
r
in
g
a
h
ijab
.
Fig
u
r
e
6
(
c)
s
er
v
es
as
a
clea
r
ex
am
p
le
o
f
th
e
o
p
tim
ized
o
u
tp
u
t
f
o
r
a
f
r
o
n
tal
an
g
le
with
o
u
t a
d
d
itio
n
al
attr
ib
u
tes.
(
a)
(
b
)
(
c)
Fig
u
r
e
6
.
E
x
am
p
le
o
f
im
ag
e
p
r
ep
r
o
ce
s
s
in
g
o
p
tim
izatio
n
r
esu
l
ts
o
f
(
a)
f
r
o
n
tal
an
g
le
ey
ewe
ar
,
(
b
)
h
ija
b
s
lig
h
t
an
g
le,
an
d
(
c)
f
r
o
n
tal
an
g
le
an
g
le
with
o
u
t a
d
d
itio
n
al
attr
ib
u
t
es
3
.
2
.
I
ma
g
e
t
ra
ns
f
o
rma
t
io
n
T
h
e
r
esu
lts
o
f
th
e
im
ag
e
tr
a
n
s
f
o
r
m
atio
n
p
r
o
ce
s
s
f
o
llo
win
g
th
e
f
ac
e
d
etec
tio
n
an
d
cr
o
p
p
in
g
s
tag
es
ar
e
illu
s
tr
ated
in
Fig
u
r
e
7
,
wh
i
ch
d
is
p
lay
s
R
GB
co
lo
r
m
ap
p
in
g
a
n
d
th
e
f
i
v
e
-
lan
d
m
a
r
k
p
o
in
t
f
o
r
m
at.
As
d
em
o
n
s
tr
ated
in
Fig
u
r
e
s
7
(
a)
-
7
(
c)
,
th
e
im
ag
e
tr
an
s
f
o
r
m
atio
n
p
r
o
ce
s
s
s
u
cc
ess
f
u
lly
f
o
r
m
s
a
p
r
o
p
er
b
o
u
n
d
i
n
g
b
o
x
a
r
o
u
n
d
th
e
f
ac
e,
r
eg
ar
d
less
o
f
v
ar
iatio
n
s
s
u
ch
as
ey
ew
ea
r
Fig
u
r
e
7
(
a)
s
h
o
ws
th
e
f
r
o
n
tal
an
g
le
e
y
ewe
ar
;
Fig
u
r
e
7
(
b
)
s
h
o
ws th
e
h
ijab
s
lig
h
t a
n
g
le
; a
n
d
Fig
u
r
e
7
(
c)
s
h
o
ws th
e
f
r
o
n
tal
an
g
le
w
ith
o
u
t a
d
d
itio
n
al
attr
ib
u
tes.
Fu
r
th
er
m
o
r
e
,
th
e
s
y
s
tem
ac
cu
r
ately
d
etec
ts
f
iv
e
m
ain
lan
d
m
ar
k
s
(
b
o
t
h
ey
es,
th
e
tip
o
f
th
e
n
o
s
e,
an
d
b
o
th
co
r
n
er
s
o
f
th
e
m
o
u
th
)
.
T
h
ese
lan
d
m
ar
k
s
ar
e
v
is
u
alize
d
as
r
ed
d
o
ts
with
n
u
m
er
ic
co
o
r
d
in
ates,
s
er
v
in
g
as
ess
en
tial r
ef
er
en
ce
s
f
o
r
s
p
atial
n
o
r
m
aliza
tio
n
a
n
d
th
e
p
r
ec
is
e
ar
r
an
g
e
m
en
t o
f
f
ac
ial
f
ea
tu
r
es
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
888
-
9
0
0
894
T
h
e
b
lu
e
an
d
p
u
r
p
le
c
o
lo
r
s
th
at
ap
p
ea
r
in
th
e
im
ag
e
a
r
e
th
e
r
esu
lt
o
f
th
e
tr
a
n
s
f
o
r
m
atio
n
p
r
o
ce
s
s
to
in
ten
s
if
y
an
d
em
p
h
asize
im
p
o
r
tan
t
f
ea
t
u
r
e
ar
ea
s
.
T
h
e
m
a
in
p
u
r
p
o
s
e
o
f
th
is
co
n
v
er
s
io
n
is
to
r
ed
u
ce
th
e
d
im
en
s
io
n
o
f
th
e
d
ata
(
elim
in
ate
ir
r
elev
an
t
co
lo
r
in
f
o
r
m
atio
n
)
an
d
s
h
ar
p
en
th
e
lo
ca
l
tex
tu
r
e
p
atter
n
th
at
will
b
e
u
s
ed
in
th
e
f
e
atu
r
e
e
x
tr
ac
tio
n
p
r
o
ce
s
s
.
Af
ter
f
ac
ial
la
n
d
m
ar
k
s
ar
e
d
etec
ted
an
d
th
e
i
n
ten
s
ity
co
n
v
e
r
s
io
n
is
co
m
p
leted
f
r
o
m
th
e
o
r
i
g
in
al
R
GB
im
ag
e,
f
ac
ial
f
ea
tu
r
es
ar
e
ex
tr
ac
ted
u
s
in
g
th
e
L
B
P
m
eth
o
d
.
T
h
is
p
r
o
ce
s
s
p
r
o
d
u
ce
s
a
f
ix
ed
-
d
im
en
s
io
n
al
n
u
m
er
ic
f
ea
t
u
r
e
ar
r
ay
,
w
h
ich
d
escr
ib
es
th
e
m
icr
o
-
tex
tu
r
e
p
a
tter
n
o
f
t
h
e
f
ac
e
i
n
th
e
f
o
r
m
o
f
d
ata
th
at
ca
n
b
e
p
r
o
ce
s
s
ed
m
ath
em
atica
lly
.
T
h
is
ar
r
ay
is
th
e
n
in
p
u
t
f
o
r
th
e
n
o
r
m
aliza
tio
n
an
d
class
if
icatio
n
s
tag
es
in
th
e
n
ex
t
p
r
o
ce
s
s
.
Vis
u
ally
,
it
ca
n
b
e
o
b
s
er
v
ed
th
at
th
e
s
y
s
tem
wo
r
k
s
co
n
s
is
ten
tly
o
n
v
ar
io
u
s
f
ac
ial
co
n
d
itio
n
s
,
in
cl
u
d
in
g
s
u
b
jects
wea
r
in
g
g
lass
es,
h
ijab
s
,
an
d
f
r
o
n
tal
p
o
s
es.
T
h
is
s
h
o
ws
th
at
th
e
tr
an
s
f
o
r
m
atio
n
s
tag
e
h
as
s
u
cc
e
ed
ed
in
f
o
r
m
in
g
an
i
n
itial
n
u
m
er
ic
r
ep
r
esen
tatio
n
th
at
is
r
ea
d
y
to
b
e
u
s
ed
in
t
h
e
f
ac
ial
r
ec
o
g
n
itio
n
p
r
o
ce
s
s
.
(
a)
(
b
)
(
c)
Fig
u
r
e
7
.
R
GB
co
lo
r
m
ap
p
in
g
an
d
f
iv
e
la
n
d
m
ar
k
p
o
in
ts
f
o
r
m
ats
of
(
a)
f
r
o
n
tal
an
g
le
ey
ewe
a
r
,
(
b
)
h
ijab
s
lig
h
t
an
g
le,
an
d
(
c)
f
r
o
n
tal
an
g
le
wi
th
o
u
t a
d
d
itio
n
al
attr
ib
u
tes
3
.
3
.
F
a
ce
re
co
g
nitio
n pre
dict
io
n m
o
delin
g
T
h
e
n
ee
d
f
o
r
SS
FR
o
n
m
o
b
ile
d
ev
ices
r
eq
u
ir
es
a
s
o
lu
tio
n
th
at
ca
n
m
ain
tain
d
etec
tio
n
s
p
ee
d
an
d
ac
cu
r
ac
y
in
lim
ited
r
eso
u
r
ce
c
o
n
d
itio
n
s
.
I
n
th
is
s
tu
d
y
,
an
c
h
o
r
-
NM
S
co
m
b
in
e
d
with
L
B
P
ar
e
u
s
ed
as
a
m
o
d
er
n
ap
p
r
o
ac
h
th
at
b
alan
ce
s
co
m
p
u
tatio
n
al
ef
f
icien
c
y
with
f
ac
e
d
etec
tio
n
a
cc
u
r
ac
y
.
T
h
is
s
tr
a
teg
y
is
d
esig
n
e
d
to
en
s
u
r
e
th
at
th
e
m
o
d
el
r
em
ain
s
lig
h
tweig
h
t,
ac
cu
r
ate,
an
d
r
es
is
tan
t
to
o
v
e
r
f
itti
n
g
.
T
o
test
th
e
ef
f
ec
tiv
en
ess
o
f
th
is
d
esig
n
,
th
e
tr
ain
in
g
an
d
t
esti
n
g
r
esu
lts
ar
e
p
r
esen
ted
th
r
o
u
g
h
a
co
m
p
ar
is
o
n
o
f
ep
o
ch
s
an
d
lo
s
s
ac
cu
r
ac
y
u
p
to
7
2
ep
o
ch
s
f
r
o
m
1
9
1
s
u
b
jects
d
ata
tr
ain
in
g
an
d
d
ata
tes
tin
g
4
8
s
u
b
jects.
T
a
b
le
1
s
u
m
m
ar
is
es
th
e
tr
ain
in
g
lo
s
s
,
v
alid
atio
n
lo
s
s
,
tr
ain
in
g
a
cc
u
r
ac
y
,
a
n
d
test
in
g
ac
cu
r
ac
y
v
alu
es a
t e
ac
h
ep
o
c
h
.
T
ab
le
1
.
E
p
o
ch
v
s
.
lo
s
s
tr
ain
in
g
-
test
in
g
lo
g
Ep
o
c
h
Tr
a
i
n
_
l
o
s
s
T
e
st
_
l
o
ss
Tr
a
i
n
_
a
c
c
Te
st
_
a
c
c
1
1
.
7
4
6
8
1
.
9
2
3
5
0
.
4
7
5
0
.
4
1
3
4
15
1
.
0
1
9
1
.
1
5
0
7
0
.
6
9
3
7
0
.
6
2
2
1
30
0
.
6
1
0
7
0
.
6
1
8
2
0
.
8
4
5
8
0
.
7
5
3
3
50
0
.
2
6
6
6
0
.
2
7
1
9
0
.
9
1
9
1
0
.
8
3
1
51
0
.
2
9
7
7
0
.
3
3
4
6
0
.
9
3
6
0
.
8
5
6
1
52
0
.
2
7
3
3
0
.
2
4
4
4
0
.
9
4
0
8
0
.
8
3
3
6
53
0
.
2
5
7
6
0
.
2
9
8
8
0
.
9
3
0
7
0
.
8
4
3
5
54
0
.
2
7
3
9
0
.
3
4
2
1
0
.
9
4
3
6
0
.
8
5
55
0
.
2
7
3
1
0
.
2
4
2
4
0
.
9
4
5
0
.
8
6
2
4
56
0
.
2
6
2
2
0
.
2
5
1
5
0
.
9
3
5
2
0
.
8
4
1
2
57
0
.
2
1
8
3
0
.
2
6
8
2
0
.
9
5
2
4
0
.
8
6
3
8
58
0
.
2
2
0
6
0
.
2
4
7
3
0
.
9
5
6
5
0
.
8
5
6
2
59
0
.
2
2
5
5
0
.
2
1
3
6
0
.
9
6
3
8
0
.
8
5
60
0
.
2
3
0
7
0
.
2
6
0
2
0
.
9
6
5
4
0
.
8
6
3
1
…
…
…
…
…
70
0
.
1
3
4
9
0
.
2
9
7
1
0
.
9
7
6
8
0
.
8
5
6
4
71
0
.
1
4
9
8
0
.
2
1
1
3
0
.
9
6
8
3
0
.
8
7
1
8
72
0
.
1
6
8
3
0
.
2
6
6
9
0
.
9
8
0
2
0
.
8
7
8
6
T
ab
le
1
s
h
o
ws
th
e
d
ev
elo
p
m
e
n
t
o
f
lo
s
s
an
d
ac
cu
r
ac
y
in
tr
ai
n
in
g
an
d
test
in
g
d
ata
u
p
to
7
2
ep
o
ch
s
.
I
n
th
e
ea
r
ly
s
tag
es
(
ep
o
c
h
1
)
,
th
e
tr
ain
l
o
s
s
v
alu
e
was
s
till
h
i
g
h
at
1
.
7
4
6
8
with
a
tr
ain
in
g
ac
cu
r
ac
y
o
f
0
.
4
7
5
,
wh
ile
th
e
v
alid
atio
n
lo
s
s
was
r
ec
o
r
d
e
d
at
1
.
9
2
3
5
with
a
test
in
g
ac
cu
r
ac
y
o
f
0
.
4
1
3
4
.
T
h
is
is
n
o
r
m
al
b
ec
au
s
e
th
e
m
o
d
el
is
s
till
in
th
e
in
itial
ad
ap
tatio
n
p
h
ase
to
th
e
d
ata.
As
th
e
n
u
m
b
er
o
f
ep
o
c
h
s
in
cr
ea
s
es,
th
e
lo
s
s
v
alu
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
B
r
id
g
in
g
h
y
b
r
id
d
ee
p
lea
r
n
i
n
g
d
etec
tio
n
a
n
d
lig
h
tw
eig
h
t
h
a
n
d
cra
fted
fea
tu
r
es fo
r
… (
F
a
u
li
n
d
a
E
ly
N
a
s
titi
)
895
d
ec
r
ea
s
es
co
n
s
is
ten
tly
an
d
th
e
ac
cu
r
ac
y
in
cr
ea
s
es.
Fo
r
ex
am
p
le,
at
ep
o
ch
3
0
,
th
e
tr
ain
l
o
s
s
h
as
d
r
o
p
p
ed
to
0
.
6
1
0
7
with
an
a
cc
u
r
ac
y
o
f
0
.
8
4
5
8
,
wh
ile
th
e
test
ac
cu
r
ac
y
h
as in
cr
ea
s
ed
to
0
.
7
5
3
3
.
I
n
th
e
m
id
d
le
s
tag
e
(
ep
o
ch
s
5
0
-
6
0
)
,
it
ca
n
b
e
s
ee
n
t
h
a
t
th
e
tr
ain
in
g
l
o
s
s
is
g
ettin
g
s
m
aller
(
0
.
2
6
6
6
-
0
.
2
3
0
7
)
with
a
tr
ain
i
n
g
ac
cu
r
ac
y
ab
o
v
e
0
.
9
2
.
T
h
e
test
in
g
ac
cu
r
ac
y
also
in
cr
ea
s
es
s
ig
n
if
ican
tly
to
ar
o
u
n
d
0
.
8
3
-
0
.
8
6
,
in
d
icatin
g
th
at
th
e
m
o
d
el
n
o
t
o
n
ly
le
ar
n
s
f
r
o
m
th
e
tr
ain
in
g
d
ata
b
u
t
is
also
ab
le
to
g
en
er
alis
e
to
th
e
test
d
ata.
E
ar
ly
s
to
p
p
in
g
was
ap
p
lied
at
ep
o
ch
5
8
,
w
h
er
e
th
e
t
r
ain
in
g
ac
cu
r
ac
y
r
ea
ch
e
d
0
.
9
5
6
5
a
n
d
th
e
test
in
g
ac
cu
r
ac
y
r
ea
c
h
ed
0
.
8
5
6
2
.
T
h
is
d
e
cisi
o
n
was
im
p
o
r
ta
n
t
to
av
o
i
d
o
v
er
f
itti
n
g
,
ev
e
n
th
o
u
g
h
th
e
ac
cu
r
ac
y
v
al
u
e
co
n
tin
u
ed
to
in
c
r
ea
s
e
in
th
e
f
o
llo
win
g
ep
o
ch
s
.
Alth
o
u
g
h
ea
r
l
y
s
to
p
p
in
g
was
ap
p
lied
at
ep
o
ch
5
8
to
p
r
ev
en
t
o
v
er
f
itti
n
g
,
s
u
b
s
eq
u
en
t
ev
alu
atio
n
u
p
to
ep
o
ch
7
1
d
em
o
n
s
tr
ated
th
at
m
o
d
el
p
er
f
o
r
m
a
n
ce
r
em
ain
e
d
s
tab
le,
with
a
m
ar
g
in
al
ac
cu
r
ac
y
im
p
r
o
v
em
en
t
to
0
.
8
7
1
8
.
T
h
is
in
d
icate
s
th
at
th
e
m
o
d
el
h
ad
r
ea
c
h
ed
co
n
v
e
r
g
en
ce
wh
ile
m
ain
tain
in
g
g
e
n
er
aliza
tio
n
with
o
u
t
s
ig
n
s
o
f
o
v
er
f
itti
n
g
.
T
h
e
n
u
m
er
ical
d
ata
i
n
T
a
b
le
1
s
h
o
ws
a
p
o
s
itiv
e
tr
e
n
d
in
t
h
e
tr
ain
i
n
g
p
r
o
ce
s
s
,
b
u
t
th
e
co
n
v
er
g
en
ce
p
atter
n
an
d
in
d
i
ca
tio
n
s
o
f
o
v
er
f
itti
n
g
will
b
e
m
o
r
e
clea
r
l
y
v
is
ib
le
in
th
e
g
r
ap
h
s
.
T
h
er
ef
o
r
e,
Fig
u
r
es 8
an
d
9
ar
e
p
r
esen
ted
t
o
s
h
o
w
th
e
d
y
n
am
ics o
f
ac
cu
r
ac
y
an
d
l
o
s
s
o
v
er
7
2
ep
o
c
h
s
.
Fig
u
r
e
8
.
Acc
u
r
ac
y
v
s
.
e
p
o
ch
Fig
u
r
e
9
.
L
o
s
s
v
s
.
ep
o
ch
T
h
e
r
esu
lts
o
f
m
o
d
el
tr
ain
in
g
an
d
test
in
g
ar
e
s
h
o
wn
in
Fig
u
r
es
8
an
d
9
.
T
h
e
ac
cu
r
a
cy
g
r
ap
h
in
Fig
u
r
e
8
s
h
o
ws
th
at
tr
ain
in
g
ac
cu
r
ac
y
in
c
r
ea
s
ed
co
n
s
is
ten
tly
to
n
ea
r
l
y
1
.
0
,
wh
ile
tes
tin
g
ac
cu
r
ac
y
also
s
h
o
wed
a
s
tead
y
u
p
war
d
tr
en
d
to
ar
o
u
n
d
0
.
9
.
T
h
e
d
if
f
er
en
ce
b
etwe
en
t
h
e
two
cu
r
v
es
in
d
icate
s
o
v
er
f
itti
n
g
,
alth
o
u
g
h
it
is
s
till
with
in
ac
ce
p
tab
le
lim
its
.
T
h
e
ea
r
ly
s
to
p
p
in
g
m
ec
h
an
is
m
ap
p
lied
at
ep
o
ch
5
8
was
a
b
le
to
p
r
ev
en
t
o
v
er
f
itti
n
g
,
alth
o
u
g
h
t
h
e
b
est v
alid
atio
n
ac
c
u
r
ac
y
wa
s
r
ec
o
r
d
ed
at
e
p
o
ch
7
1
.
T
h
e
lo
s
s
g
r
a
p
h
in
Fig
u
r
e
9
s
h
o
ws
a
s
im
ilar
p
atter
n
,
w
h
er
e
b
o
th
tr
ain
in
g
lo
s
s
an
d
v
alid
atio
n
lo
s
s
d
ec
r
ea
s
ed
co
n
s
is
ten
tly
th
r
o
u
g
h
o
u
t
th
e
tr
ai
n
in
g
p
r
o
ce
s
s
.
Af
te
r
ep
o
ch
5
0
,
a
s
m
all
g
ap
b
eg
an
to
ap
p
ea
r
b
etwe
en
th
e
two
cu
r
v
es,
in
d
icatin
g
th
e
ea
r
ly
s
ig
n
s
o
f
o
v
er
f
itti
n
g
.
Ho
wev
er
,
th
e
lo
s
s
v
alu
e
at
th
e
e
n
d
o
f
tr
ai
n
in
g
was
r
elativ
ely
s
tab
le,
in
d
icatin
g
th
at
th
e
m
o
d
el
h
ad
r
ea
ch
ed
a
co
n
v
er
g
e
d
s
tate.
T
h
is
co
n
v
er
g
en
ce
in
d
icate
s
th
at
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
1
,
Feb
r
u
ar
y
2
0
2
6
:
888
-
9
0
0
896
m
o
d
el
p
ar
am
eter
s
h
av
e
a
d
ap
t
ed
o
p
tim
ally
to
th
e
tr
ai
n
in
g
d
ata
with
o
u
t
lo
s
in
g
th
eir
g
en
er
alis
atio
n
ab
ilit
y
o
n
th
e
test
d
ata.
T
o
ev
alu
ate
th
e
r
o
b
u
s
tn
ess
an
d
g
e
n
er
aliza
tio
n
ca
p
ab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
SS
FR
f
r
am
ewo
r
k
,
a
co
m
p
ar
ativ
e
ex
p
er
im
en
t
was
co
n
d
u
cte
d
ag
ain
s
t
two
b
aseli
n
e
m
o
d
els,
OR
B
an
d
Mo
b
ileFace
Net.
T
h
e
s
am
e
4
8
s
u
b
ject
test
s
et
u
s
ed
f
o
r
m
o
d
el
e
v
alu
atio
n
was
em
p
lo
y
ed
to
en
s
u
r
e
f
air
b
en
c
h
m
ar
k
in
g
.
As
s
h
o
wn
in
T
ab
le
2
,
OR
B
,
u
tili
ze
d
as
a
b
aselin
e
n
o
n
-
f
ac
e
d
escr
ip
to
r
,
ac
h
iev
ed
o
n
ly
4
5
.
2
%
ac
c
u
r
ac
y
,
r
ef
lectin
g
its
lim
itatio
n
s
in
h
an
d
lin
g
p
o
s
e
v
ar
iatio
n
s
an
d
o
cc
lu
s
io
n
s
.
Mo
b
ileFace
Net,
o
p
tim
ized
th
r
o
u
g
h
d
ata
au
g
m
en
tatio
n
an
d
a
d
ap
tiv
e
t
h
r
esh
o
ld
tu
n
in
g
,
ac
h
ie
v
ed
8
2
.
4
%
ac
cu
r
ac
y
.
Desp
ite
b
ein
g
tr
ain
ed
o
n
d
iv
er
s
e
f
ac
ial
d
ata,
its
p
er
f
o
r
m
an
ce
d
ec
r
ea
s
ed
in
SS
FR
s
ett
in
g
s
d
u
e
to
th
e
lack
o
f
m
u
ltip
le
r
ef
e
r
en
ce
s
am
p
les
p
er
id
en
tity
,
wh
ich
co
n
s
tr
ain
ed
em
b
ed
d
in
g
d
iv
e
r
s
ity
.
I
n
co
n
tr
ast,
th
e
p
r
o
p
o
s
ed
h
y
b
r
id
m
o
d
el
i
n
teg
r
atin
g
an
ch
o
r
-
b
ased
d
etec
tio
n
,
NM
S,
an
d
L
B
P
ac
h
iev
ed
a
co
n
v
er
g
ed
test
in
g
ac
cu
r
ac
y
o
f
8
5
.
6
%
,
wh
ich
was
s
elec
ted
as
th
e
o
f
f
icial
ev
alu
atio
n
v
al
u
e
f
o
r
c
o
m
p
ar
ativ
e
an
aly
s
is
.
T
h
e
u
s
e
o
f
th
is
co
n
v
er
g
ed
v
alu
e,
r
ath
er
th
a
n
th
e
p
e
ak
ac
cu
r
ac
y
o
f
8
7
.
1
8
%
ac
h
iev
ed
at
e
p
o
ch
7
1
,
was
in
ten
d
ed
to
e
n
s
u
r
e
t
h
e
v
ali
d
ity
an
d
r
e
p
r
o
d
u
cib
ilit
y
o
f
r
esu
lts
wh
ile
p
r
ev
e
n
tin
g
o
v
er
f
itti
n
g
b
ias.
T
h
e
co
n
v
er
g
en
ce
-
b
ased
ev
alu
atio
n
r
ef
lects
a
b
alan
ce
b
etwe
en
ac
cu
r
ac
y
an
d
m
o
d
el
s
tab
ilit
y
,
p
r
o
v
id
i
n
g
a
r
eliab
le
m
ea
s
u
r
e
o
f
p
er
f
o
r
m
an
ce
in
S
SF
R
.
Acc
o
r
d
in
g
ly
,
th
e
3
.
2
%
im
p
r
o
v
em
e
n
t
o
v
er
Mo
b
ileFace
Net
in
d
icate
s
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
el
r
em
ain
s
c
o
m
p
etitiv
e
an
d
ca
n
o
u
tp
er
f
o
r
m
co
n
v
en
tio
n
al
d
ee
p
lear
n
i
n
g
ar
ch
itectu
r
es
u
n
d
er
SS
F
R
,
wh
er
e
f
ea
tu
r
e
d
i
v
er
s
ity
an
d
g
e
n
er
aliza
tio
n
ar
e
ty
p
ical
ly
lim
ited
.
T
ab
le
2
.
C
o
m
p
a
r
ativ
e
ex
p
er
im
en
tal
m
o
d
els
M
o
d
e
l
t
e
s
t
i
n
g
A
c
c
u
r
a
c
y
(
%)
O
R
B
4
5
.
2
M
o
b
i
l
e
F
a
c
e
N
e
t
8
2
.
4
P
r
o
p
o
se
d
m
o
d
e
l
(
a
n
c
h
o
r
-
N
M
S
-
LB
P
)
8
5
.
6
3
.
4
.
F
a
ce
re
co
g
nitio
n c
o
s
ine sim
ila
rit
y
ev
a
lua
t
i
o
n
T
h
e
e
v
a
l
u
a
t
i
o
n
o
f
t
h
e
f
a
c
e
r
e
c
o
g
n
i
t
i
o
n
m
o
d
e
l
i
n
t
h
i
s
s
t
u
d
y
w
a
s
c
a
r
r
i
e
d
o
u
t
u
s
i
n
g
4
8
n
e
w
f
a
c
i
a
l
i
m
a
g
e
s
t
h
a
t
i
n
c
l
u
d
e
d
v
a
r
i
a
t
i
o
n
s
i
n
l
i
g
h
t
i
n
g
,
a
n
g
l
e
s
,
a
n
d
e
x
p
r
e
s
s
i
o
n
s
.
T
o
e
v
a
l
u
a
t
e
t
h
e
m
o
d
e
l
p
e
r
f
o
r
m
a
n
c
e
,
c
o
s
i
n
e
s
i
m
i
l
a
r
i
t
y
w
a
s
u
s
e
d
a
s
t
h
e
m
a
i
n
m
e
t
r
i
c
,
w
i
t
h
a
m
i
n
i
m
u
m
s
i
m
i
l
a
r
i
t
y
l
e
v
e
l
o
f
3
0
%
a
s
t
h
e
a
c
c
e
p
t
a
n
c
e
t
h
r
e
s
h
o
l
d
.
T
h
e
t
h
r
e
s
h
o
l
d
o
f
3
0
%
w
a
s
c
h
o
s
e
n
b
a
s
e
d
o
n
r
e
s
e
a
r
c
h
r
e
s
u
l
t
s
s
h
o
w
i
n
g
t
h
a
t
t
h
i
s
l
e
v
e
l
o
f
s
i
m
i
l
a
r
i
t
y
i
s
s
t
i
l
l
s
u
f
f
i
c
i
e
n
t
t
o
u
n
i
q
u
e
l
y
d
i
s
t
i
n
g
u
i
s
h
i
n
d
i
v
i
d
u
a
l
s
w
h
i
l
e
s
t
i
l
l
a
l
l
o
w
i
n
g
f
o
r
n
a
t
u
r
a
l
v
a
r
i
a
t
i
o
n
s
i
n
f
a
c
i
a
l
e
x
p
r
e
s
s
i
o
n
a
n
d
l
i
g
h
t
i
n
g
.
I
n
a
d
d
i
t
i
o
n
,
t
h
i
s
v
a
l
u
e
h
a
s
a
l
s
o
b
e
e
n
t
e
s
t
e
d
i
n
v
a
r
i
o
u
s
f
a
c
e
r
e
c
o
g
n
i
t
i
o
n
s
c
e
n
a
r
i
o
s
,
w
h
e
r
e
t
h
e
m
o
d
e
l
p
e
r
f
o
r
m
a
n
c
e
r
e
m
a
i
n
s
o
p
t
i
m
a
l
w
i
t
h
o
u
t
s
i
g
n
i
f
i
c
a
n
t
l
y
i
n
c
r
e
a
s
i
n
g
t
h
e
n
u
m
b
e
r
o
f
f
a
l
s
e
p
o
s
i
t
i
v
e
s
.
T
h
e
ev
alu
atio
n
was
co
n
d
u
cte
d
u
s
in
g
th
e
co
s
in
e
s
im
ilar
ity
ap
p
r
o
ac
h
a
n
d
p
r
o
d
u
ce
d
a
p
e
r
f
o
r
m
a
n
ce
d
is
tr
ib
u
tio
n
as sh
o
wn
in
Fig
u
r
e
1
0
.
Fig
u
r
e
1
0
(
a
)
s
h
o
ws th
e
co
n
f
u
s
io
n
m
atr
ix
.
Ou
t o
f
th
e
4
8
im
ag
es,
4
1
im
ag
es
wer
e
co
r
r
ec
tly
r
ec
o
g
n
is
ed
(
tr
u
e
p
o
s
itiv
e
(
T
P)
)
,
wh
ile
2
im
ag
es
wer
e
m
is
id
en
tifie
d
as
o
th
er
id
en
titi
es
(
f
alse
p
o
s
itiv
e
(
FP
)
)
,
an
d
4
im
ag
es
wer
e
n
o
t
r
ec
o
g
n
is
ed
(
f
alse
n
eg
ativ
e
(
FN)
)
.
On
e
o
th
er
im
ag
e
was
r
ec
o
r
d
ed
as
tr
u
e
n
eg
ativ
e
(
T
N)
.
T
h
ese
r
esu
lts
p
r
o
v
id
e
an
o
v
er
all
ac
c
u
r
ac
y
o
f
8
9
.
6
%,
wh
ic
h
s
h
o
ws
th
e
c
o
n
s
is
ten
cy
o
f
th
e
m
o
d
el
in
d
ea
lin
g
with
v
ar
i
atio
n
s
in
test
co
n
d
itio
n
s
.
F
u
r
t
h
er
m
o
r
e,
d
i
s
c
r
i
m
in
a
t
io
n
e
v
a
l
u
a
t
i
o
n
w
a
s
p
er
f
o
r
m
e
d
u
s
in
g
t
h
e
r
e
c
e
i
v
e
r
o
p
e
r
a
t
in
g
ch
ar
a
c
t
e
r
i
s
t
i
c
(
R
O
C
)
c
u
r
v
e
a
s
s
h
o
w
n
in
F
i
g
u
r
e
1
0
(
b
)
,
w
i
t
h
a
n
ar
e
a
u
n
d
e
r
c
u
r
v
e
(
A
U
C
)
v
a
lu
e
o
f
0
.
9
4
.
T
h
i
s
v
a
lu
e
i
n
d
i
c
a
t
e
s
t
h
a
t
th
e
m
o
d
e
l
h
a
s
a
f
a
i
r
l
y
s
t
r
o
n
g
ab
i
l
i
t
y
t
o
d
i
s
t
i
n
g
u
i
s
h
b
e
tw
e
e
n
d
i
f
f
e
r
en
t
id
en
t
i
t
i
e
s
,
a
l
t
h
o
u
g
h
th
er
e
i
s
s
t
i
l
l
a
r
e
a
s
o
n
a
b
l
e
m
ar
g
in
o
f
er
r
o
r
.
I
n
a
d
d
i
t
i
o
n
,
th
e
ev
a
l
u
a
t
i
o
n
r
e
s
u
l
t
s
c
an
b
e
s
u
m
m
ar
i
s
e
d
i
n
t
o
p
er
f
o
r
m
an
c
e
m
e
t
r
i
c
s
.
T
h
e
p
r
e
c
i
s
i
o
n
r
a
t
e
wa
s
9
5
.
3
%
,
c
a
l
cu
l
a
t
e
d
f
r
o
m
t
h
e
r
a
t
io
o
f
co
r
r
e
c
t
ly
r
e
c
o
g
n
i
s
ed
f
ac
e
s
(
T
P
)
to
a
l
l
p
o
s
i
t
i
v
e
p
r
e
d
ic
t
i
o
n
s
(
T
P
+
F
P
)
.
T
h
e
r
e
c
a
l
l
v
a
lu
e
wa
s
9
1
.
1
%
,
d
er
i
v
e
d
f
r
o
m
t
h
e
r
a
t
io
o
f
c
o
r
r
e
c
t
ly
r
e
c
o
g
n
i
s
ed
f
a
c
e
s
t
o
a
l
l
ac
t
u
a
l
f
a
c
e
d
a
t
a
(
T
P
+
F
N
)
.
M
ea
n
wh
i
l
e
,
t
h
e
F
1
-
s
c
o
r
e
v
a
l
u
e
r
e
a
ch
e
d
9
3
.
1
%
a
s
a
r
e
s
u
l
t
o
f
t
h
e
h
a
r
m
o
n
i
s
a
t
io
n
b
e
t
w
e
en
p
r
e
c
is
i
o
n
a
n
d
r
e
ca
l
l
.
T
h
e
s
e
t
h
r
e
e
m
e
t
r
i
c
s
s
h
o
w
th
a
t
t
h
e
m
o
d
e
l
i
s
n
o
t
o
n
l
y
a
c
cu
r
a
te
i
n
r
e
c
o
g
n
i
s
i
n
g
f
a
c
e
s
c
o
r
r
e
c
t
ly
,
b
u
t
a
l
s
o
c
ap
ab
l
e
o
f
m
a
i
n
ta
in
i
n
g
a
b
a
l
an
c
e
b
e
t
w
ee
n
p
r
e
d
ic
t
i
o
n
a
c
cu
r
a
cy
an
d
d
e
t
e
c
t
io
n
c
o
m
p
l
e
t
en
e
s
s
.
T
h
e
ev
alu
atio
n
r
esu
lts
d
em
o
n
s
tr
ate
th
at
th
e
p
r
o
p
o
s
ed
m
o
d
e
l
is
r
o
b
u
s
t
an
d
ca
p
a
b
le
o
f
m
ain
tain
in
g
a
b
alan
ce
d
tr
ad
e
-
o
f
f
b
etwe
en
p
r
ec
is
io
n
an
d
r
ec
all.
C
o
m
p
ar
e
d
with
p
r
io
r
s
tu
d
ies,
s
u
c
h
as
Ar
cFac
e
(
5
2
%)
[
2
6
]
,
an
d
C
GAN
(
7
6
%)
[
2
7
]
th
e
p
r
o
p
o
s
ed
o
p
tim
izatio
n
p
r
o
v
ed
to
b
e
m
o
r
e
ac
cu
r
ate,
ef
f
icien
t,
an
d
s
tab
le.
T
h
e
p
er
f
o
r
m
an
ce
also
s
u
r
p
ass
ed
th
e
s
tu
d
y
b
y
Din
g
et
a
l
.
[
3
6
]
,
wh
ich
r
ep
o
r
ted
8
7
.
3
%
ac
cu
r
ac
y
o
f
SS
FR
u
s
in
g
u
n
if
o
r
m
g
e
n
er
ic
r
e
p
r
esen
tatio
n
an
d
d
ee
p
f
ea
tu
r
e
o
n
lab
eled
f
ac
es
in
th
e
wild
(
L
FW
)
.
T
h
es
e
f
in
d
in
g
s
co
n
f
ir
m
th
at
th
e
in
teg
r
atio
n
o
f
an
ch
o
r
b
o
x
,
NM
S,
an
d
L
B
P
is
ef
f
ec
tiv
e
an
d
r
o
b
u
s
t
in
th
e
co
n
tex
t
o
f
SS
FR
,
in
co
n
tr
ast
to
m
o
s
t p
r
ev
i
o
u
s
wo
r
k
s
th
at
r
e
lied
o
n
m
u
lti
-
s
am
p
le
co
n
f
i
g
u
r
atio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
B
r
id
g
in
g
h
y
b
r
id
d
ee
p
lea
r
n
i
n
g
d
etec
tio
n
a
n
d
lig
h
tw
eig
h
t
h
a
n
d
cra
fted
fea
tu
r
es fo
r
… (
F
a
u
li
n
d
a
E
ly
N
a
s
titi
)
897
(
a)
(
b
)
Fig
u
r
e
1
0
.
C
o
s
in
e
s
im
ilar
ity
ev
alu
atio
n
o
f
(
a)
c
o
n
f
u
s
io
n
m
at
r
ix
an
d
(
b
)
R
OC
cu
r
v
e
3
.
5
.
I
llu
m
ina
t
io
n v
a
ria
t
io
n e
v
a
lua
t
io
n
T
h
e
s
y
s
t
e
m
e
v
a
l
u
a
t
i
o
n
w
a
s
c
o
n
d
u
c
t
e
d
i
n
a
n
i
n
t
e
g
r
a
t
e
d
m
a
n
n
e
r
t
h
r
o
u
g
h
a
m
o
b
i
l
e
a
p
p
l
i
c
a
t
i
o
n
d
e
v
e
l
o
p
e
d
f
o
r
t
h
e
A
n
d
r
o
i
d
a
n
d
i
O
S
p
l
a
t
f
o
r
m
s
.
S
i
n
c
e
f
a
c
i
a
l
r
e
c
o
g
n
i
t
i
o
n
w
a
s
a
p
p
l
i
e
d
i
n
a
S
S
F
R
s
c
e
n
a
r
i
o
,
l
i
g
h
t
i
n
g
v
a
r
i
a
t
i
o
n
s
e
m
e
r
g
e
d
a
s
t
h
e
m
a
i
n
o
b
s
t
a
c
l
e
;
t
h
e
s
y
s
t
e
m
d
i
d
n
o
t
h
a
v
e
a
d
d
i
t
i
o
n
a
l
i
m
a
g
e
s
t
o
a
d
a
p
t
t
o
c
h
a
n
g
e
s
i
n
i
l
l
u
m
i
n
a
t
i
o
n
.
T
h
e
r
e
f
o
r
e
,
t
h
e
e
v
a
l
u
a
t
i
o
n
w
a
s
c
o
n
d
u
c
t
e
d
b
y
m
e
a
s
u
r
i
n
g
t
h
e
m
o
d
e
l
'
s
r
e
s
i
l
i
e
n
c
e
t
o
t
h
e
s
e
o
b
s
t
a
c
l
e
s
.
T
h
e
a
p
p
l
i
c
a
t
i
o
n
w
a
s
t
e
s
t
e
d
o
n
4
8
i
m
a
g
e
s
o
f
s
t
u
d
e
n
t
s
f
r
o
m
U
n
i
v
e
r
s
i
t
a
s
D
u
t
a
B
a
n
g
s
a
S
u
r
a
k
a
r
t
a
,
w
h
i
c
h
w
e
r
e
n
o
t
u
s
e
d
i
n
t
h
e
t
r
a
i
n
i
n
g
o
r
t
e
s
t
i
n
g
p
r
o
c
e
s
s
.
T
h
e
e
v
a
l
u
a
t
i
o
n
r
e
s
u
l
t
s
T
a
b
l
e
3
s
h
o
w
t
h
a
t
u
n
d
e
r
n
o
r
m
a
l
l
i
g
h
t
i
n
g
c
o
n
d
i
t
i
o
n
s
o
f
a
r
o
u
n
d
3
5
0
-
4
0
0
l
u
x
,
t
h
e
m
o
d
e
l
s
u
c
c
e
s
s
f
u
l
l
y
r
e
c
o
g
n
i
s
e
d
4
3
o
u
t
o
f
4
8
f
a
c
e
s
w
i
t
h
a
n
a
c
c
u
r
a
c
y
o
f
8
9
.
5
8
%
.
M
e
a
n
w
h
i
l
e
,
u
n
d
e
r
l
o
w
l
i
g
h
t
i
n
g
c
o
n
d
i
t
i
o
n
s
o
f
5
0
-
1
0
0
l
u
x
,
t
h
e
s
y
s
t
e
m
w
a
s
s
t
i
l
l
a
b
l
e
t
o
i
d
e
n
t
i
f
y
4
1
f
a
c
e
s
w
i
t
h
a
n
a
c
c
u
r
a
c
y
o
f
8
5
.
4
2
%
.
T
h
e
l
i
m
i
t
e
d
d
i
f
f
e
r
e
n
c
e
b
e
t
w
e
e
n
t
h
e
s
e
t
w
o
v
a
l
u
e
s
i
n
d
i
c
a
t
e
s
t
h
a
t
t
h
e
p
r
o
p
o
s
e
d
a
p
p
r
o
a
c
h
i
s
q
u
i
t
e
t
o
l
e
r
a
n
t
t
o
i
l
l
u
m
i
n
a
t
i
o
n
v
a
r
i
a
t
i
o
n
s
,
e
v
e
n
t
h
o
u
g
h
i
t
i
s
e
q
u
i
p
p
e
d
w
i
t
h
o
n
l
y
o
n
e
s
a
m
p
l
e
p
e
r
i
d
e
n
t
i
t
y
.
W
i
t
h
i
n
f
e
r
e
n
c
e
l
a
t
e
n
c
y
i
n
t
h
e
r
a
n
g
e
o
f
t
e
n
s
o
f
m
i
l
l
i
s
e
c
o
n
d
s
o
n
m
o
b
i
l
e
d
e
v
i
c
e
s
,
t
h
i
s
s
y
s
t
e
m
n
o
t
o
n
l
y
m
e
e
t
s
S
S
F
R
p
e
r
f
o
r
m
a
n
c
e
r
e
q
u
i
r
e
m
e
n
t
s
b
u
t
i
s
a
l
s
o
s
u
i
t
a
b
l
e
f
o
r
i
m
p
l
e
m
e
n
t
a
t
i
o
n
o
n
m
o
b
i
l
e
p
l
a
t
f
o
r
m
s
w
i
t
h
l
i
m
i
t
e
d
p
o
w
e
r
a
n
d
c
o
m
p
u
t
i
n
g
r
e
s
o
u
r
c
e
s
.
T
ab
le
3
.
Mo
b
ile
f
ac
e
r
ec
o
g
n
iti
o
n
lu
x
ev
alu
atio
n
Lu
x
N
u
mb
e
r
o
f
f
a
c
i
a
l
t
e
st
e
d
C
o
r
r
e
c
t
p
r
e
d
i
c
t
i
o
n
A
c
c
u
r
a
c
y
%
3
5
0
-
400
48
43
8
9
.
5
8
50
-
1
0
0
48
41
8
5
.
4
2
4.
CO
NCLU
SI
O
N
T
h
e
h
y
b
r
id
m
o
d
el
f
o
r
SS
FR
in
teg
r
atin
g
an
ch
o
r
,
NM
S
,
an
d
L
B
P
im
p
r
o
v
ed
tr
ain
in
g
ac
cu
r
ac
y
f
r
o
m
4
7
.
5
-
9
8
.
0
%
at
ep
o
ch
7
2
,
with
test
in
g
ac
cu
r
ac
y
r
em
ain
i
n
g
s
tab
le
in
th
e
r
a
n
g
e
o
f
8
5
-
8
8
%
(
b
est
at
8
7
.
9
%).
E
v
alu
atio
n
o
n
4
8
n
ew
f
ac
ia
l
im
ag
es
ac
h
iev
ed
8
9
.
6
%
ac
cu
r
ac
y
,
9
5
.
3
%
p
r
ec
is
io
n
,
9
1
.
1
%
r
ec
all,
9
3
.
1
%
F1
-
s
co
r
e,
a
n
d
0
.
9
4
AUC
R
OC
.
Pra
ctica
l
d
ep
lo
y
m
en
t
v
ia
An
d
r
o
id
an
d
iOS
-
b
ased
atten
d
an
ce
a
p
p
licatio
n
s
co
n
f
ir
m
e
d
th
e
tr
an
s
f
er
ab
ilit
y
o
f
th
e
m
o
d
el
with
lev
els
b
etwe
en
5
0
lu
x
a
n
d
4
0
0
lu
x
,
it
p
r
o
d
u
ce
s
an
ac
cu
r
ac
y
o
f
8
8
.
4
6
%.
T
h
ese
f
in
d
in
g
s
co
n
f
i
r
m
th
at
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
is
n
o
t
o
n
l
y
ac
cu
r
ate
b
u
t
also
m
o
r
e
r
o
b
u
s
t
th
a
n
class
ical
b
aselin
es
an
d
p
r
e
v
io
u
s
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
e
s
,
wh
ile
r
em
ain
i
n
g
co
m
p
u
tatio
n
ally
e
f
f
icien
t
a
n
d
s
u
itab
le
f
o
r
r
ea
l
-
tim
e
b
io
m
etr
i
c
au
th
en
ticatio
n
in
p
u
b
lic
s
ec
to
r
,
ed
u
ca
tio
n
,
an
d
I
o
T
ec
o
s
y
s
tem
s
.
Fu
tu
r
e
s
tu
d
ies
s
h
o
u
ld
ex
p
a
n
d
d
ataset
d
iv
er
s
i
ty
an
d
in
clu
d
e
b
r
o
a
d
er
d
em
o
g
r
ap
h
ic
v
alid
atio
n
to
m
itig
ate
p
o
ten
tial
b
ias
an
d
f
air
n
ess
is
s
u
es
in
SS
F
R
.
I
n
co
r
p
o
r
atin
g
f
ac
e
alig
n
m
e
n
t,
ad
ap
tiv
e
th
r
esh
o
ld
in
g
,
an
d
p
r
iv
ac
y
-
p
r
eser
v
i
n
g
tech
n
iq
u
es
s
u
ch
as
o
n
-
d
ev
ice
in
f
er
en
ce
o
r
f
e
d
er
ated
lear
n
i
n
g
m
ay
f
u
r
th
er
en
h
an
ce
r
o
b
u
s
tn
ess
an
d
s
ec
u
r
ity
.
Ad
d
itio
n
ally
,
lar
g
e
-
s
ca
le
r
ea
l
-
tim
e
m
u
lti
-
u
s
er
ev
alu
atio
n
s
ar
e
r
ec
o
m
m
en
d
ed
to
ass
ess
s
c
alab
ilit
y
in
m
o
b
ile
an
d
I
o
T
en
v
ir
o
n
m
e
n
ts
.
ACK
NO
WL
E
DG
M
E
N
T
S
G
r
a
t
i
t
u
d
e
i
s
f
ir
s
t
ex
p
r
e
s
s
e
d
t
o
t
h
e
M
in
i
s
t
r
y
o
f
H
i
g
h
er
E
d
u
c
a
t
io
n
,
S
c
ie
n
c
e,
a
n
d
T
e
c
h
n
o
l
o
g
y
,
R
e
p
u
b
l
i
c
o
f
I
n
d
o
n
e
s
i
a
,
a
s
th
i
s
p
ap
er
r
e
p
r
e
s
e
n
t
s
o
n
e
o
f
t
h
e
o
u
t
p
u
t
s
o
f
t
h
e
Fu
n
d
a
m
en
t
a
l
R
es
e
a
r
c
h
G
r
a
n
t
.
T
h
e
a
u
t
h
o
r
s
a
l
s
o
ex
t
e
n
d
a
p
p
r
e
c
ia
t
i
o
n
to
U
n
i
v
e
r
s
i
t
a
s
D
u
t
a
B
a
n
g
s
a
S
u
r
ak
ar
t
a
,
f
o
r
p
r
o
v
id
i
n
g
a
d
m
in
i
s
t
r
a
t
iv
e
Evaluation Warning : The document was created with Spire.PDF for Python.