I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
2
,
Apr
il
2025
,
pp.
1
056
~
1
066
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
2
.
pp
1
0
56
-
1
066
1056
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
E
n
h
an
c
in
g f
a
c
ia
l
r
e
c
ogn
itio
n
ac
c
u
r
ac
y
t
h
r
ou
g
h
f
e
at
u
r
e
e
xt
r
a
c
t
io
n
s a
n
d
a
r
t
ifi
c
ia
l
n
e
u
r
al
n
e
t
w
o
r
k
s
Adh
i
Ku
s
n
ad
i
,
I
vr
an
z
a
Z
u
h
d
i
P
a
n
e
,
F
e
n
in
a
Adl
in
e
T
win
c
e
T
ob
in
g
D
e
pa
r
tm
e
nt
of
I
nf
or
ma
ti
c
s
, F
a
c
ul
ty
of
E
ngi
ne
e
r
in
g
a
nd
I
n
f
or
ma
tc
s
, U
ni
ve
r
s
it
a
s
M
ul
ti
me
di
a
N
us
a
nt
a
r
a
, B
a
nt
e
n, I
ndone
s
ia
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
De
c
28
,
2023
R
e
vis
e
d
Oc
t
26
,
2024
Ac
c
e
pted
Nov
14
,
2024
Faci
al
rec
o
g
n
i
t
i
o
n
i
s
a
b
i
o
met
r
i
c
s
y
s
t
em
u
s
ed
t
o
i
d
en
t
i
f
y
i
n
d
i
v
i
d
u
a
l
s
t
h
r
o
u
g
h
faces
.
A
l
t
h
o
u
g
h
t
h
i
s
t
ech
n
o
l
o
g
y
h
a
s
man
y
ad
v
a
n
t
a
g
es
,
i
t
s
t
i
l
l
faces
s
ev
era
l
ch
al
l
en
g
es
.
O
n
e
o
f
t
h
e
mai
n
ch
a
l
l
e
n
g
e
s
i
s
t
h
at
t
h
e
l
ev
e
l
o
f
accu
racy
h
as
y
et
t
o
reach
i
t
s
max
i
m
u
m
p
o
t
e
n
t
i
al
.
T
h
i
s
res
earch
ai
m
s
t
o
i
mp
r
o
v
e
faci
al
reco
g
n
i
t
i
o
n
p
erfo
rma
n
ce
b
y
ap
p
l
y
i
n
g
t
h
e
d
i
s
cre
t
e
c
o
s
i
n
e
t
ran
s
fo
rm
(D
CT
)
a
n
d
G
au
s
s
i
a
n
mi
x
t
u
re
m
o
d
e
l
(G
MM),
w
h
i
ch
a
re
t
h
en
t
rai
n
ed
w
i
t
h
b
ac
k
w
ar
d
p
r
o
p
a
g
at
i
o
n
o
f
erro
rs
(
b
ack
p
r
o
p
a
g
at
i
o
n
)
an
d
co
n
v
o
l
u
t
i
o
n
a
l
n
eu
ra
l
n
e
t
w
o
rk
s
(CN
N
).
T
h
e
res
earch
res
u
l
t
s
s
h
o
w
l
o
w
D
CT
an
d
G
MM
feat
u
re
ex
t
ract
i
o
n
accu
racy
w
i
t
h
b
ack
p
ro
p
ag
a
t
i
o
n
o
f
4
.
8
8
%
.
H
o
w
ev
er,
t
h
e
co
m
b
i
n
at
i
o
n
o
f
D
CT
,
G
MM,
an
d
CN
N
fea
t
u
re
e
x
t
ra
ct
i
o
n
p
r
o
d
u
ces
a
n
acc
u
racy
o
f
u
p
t
o
9
8
.
2
%
an
d
a
t
rai
n
i
n
g
t
i
me
o
f
3
6
0
s
eco
n
d
s
o
n
t
h
e
O
l
i
v
et
t
i
Res
earc
h
L
ab
o
rat
o
ry
(
O
RL
)
d
at
as
et
,
an
accu
racy
o
f
9
8
.
9
%
an
d
a
t
rai
n
i
n
g
t
i
me
o
f
1
2
1
0
s
ec
o
n
d
s
o
n
t
h
e
Y
a
l
e
d
a
t
as
e
t
,
an
d
1
0
0
%
accu
racy
an
d
t
rai
n
i
n
g
t
i
me
1
7
4
9
s
ec
o
n
d
s
o
n
t
h
e
J
a
p
an
e
s
e
femal
e
faci
al
ex
p
re
s
s
i
o
n
(
J
A
FF
E
)
d
a
t
as
e
t
.
T
h
i
s
i
mp
r
o
v
em
en
t
i
s
d
u
e
t
o
t
h
e
co
mb
i
n
a
t
i
o
n
o
f
D
C
T
,
G
MM,
an
d
CN
N
's
a
b
i
l
i
t
y
t
o
rem
o
v
e
n
o
i
s
e
a
n
d
s
t
u
d
y
i
mag
e
s
accu
rat
el
y
.
T
h
i
s
res
earch
i
s
ex
p
ec
t
ed
t
o
s
i
g
n
i
f
i
can
t
l
y
co
n
t
r
i
b
u
t
e
t
o
o
v
erc
o
mi
n
g
accu
rac
y
ch
al
l
en
g
es
an
d
i
n
crea
s
i
n
g
t
h
e
fl
ex
i
b
i
l
i
t
y
o
f
faci
al
reco
g
n
i
t
i
o
n
s
y
s
t
ems
i
n
v
ar
i
o
u
s
p
rac
t
i
ca
l
s
i
t
u
a
t
i
o
n
s
,
as
w
el
l
as
t
h
e
p
o
t
e
n
t
i
al
t
o
i
mp
r
o
v
e
s
ec
u
ri
t
y
a
n
d
rel
i
ab
i
l
i
t
y
i
n
s
ecu
r
i
t
y
an
d
b
i
o
me
t
ri
cs
.
K
e
y
w
o
r
d
s
:
B
a
c
kpr
opa
ga
ti
on
C
onvolut
ional
ne
ur
a
l
ne
twor
k
Dis
c
r
e
te
c
os
ine
tr
a
ns
f
or
m
F
a
c
e
r
e
c
ognit
ion
F
e
a
tur
e
e
xtr
a
c
ti
on
Ga
us
s
ian
mi
xtur
e
model
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
I
vr
a
ns
a
Z
uhdi
P
a
ne
De
pa
r
tm
e
nt
of
I
nf
or
mat
ics
,
F
a
c
ult
y
o
f
E
nginee
r
ing
a
nd
I
n
f
or
mat
i
c
s
,
Unive
r
s
it
a
s
M
ult
im
e
dia
Nus
a
ntar
a
S
c
ientia
B
ouleva
r
d
Ga
ding
S
t
.
,
T
a
nge
r
a
ng
R
e
ge
nc
y
,
B
a
nten
15810
,
I
ndone
s
ia
E
mail:
iv
r
a
ns
a
.
z
uhdi@l
e
c
tur
e
r
.
umn
.
a
c
.
id
1.
I
NT
RODU
C
T
I
ON
F
a
c
ial
r
e
c
ognit
ion
is
a
n
e
s
s
e
nti
a
l
s
ys
tem
in
the
digi
tal
wor
ld
that
is
us
e
d
to
identi
f
y
a
pe
r
s
on
f
r
om
digi
tal
im
a
ge
s
[
1
]
.
T
his
s
ys
tem
is
a
ppli
e
d
a
s
a
s
olu
ti
on
in
va
r
ious
f
ields
s
uc
h
a
s
s
e
c
ur
it
y
,
biom
e
t
r
ics
,
r
oboti
c
s
,
im
a
ge
s
e
a
r
c
h,
a
nd
im
a
ge
a
nd
video
indexing
[
2
]
–
[
5
]
.
As
tec
hnology
de
ve
lops
,
r
e
c
ognit
ion
pr
ovides
s
i
gnif
ica
nt
a
dva
ntage
s
in
va
r
ious
c
ontexts
[
6]
.
One
of
it
s
s
up
e
r
ior
f
e
a
tur
e
s
is
it
s
s
oli
d
s
e
c
ur
it
y
be
c
a
us
e
thi
s
tec
hnology
of
f
e
r
s
a
s
a
f
e
a
nd
c
onve
nient
wa
y
of
a
uthentica
ti
on
[
7]
,
r
e
duc
ing
de
pe
nde
nc
e
on
pa
s
s
wor
ds
a
n
d
c
onv
e
nti
ona
l
a
c
c
e
s
s
c
a
r
ds
.
F
a
c
ial
r
e
c
ognit
ion
tec
hnology
is
wide
ly
us
e
d
in
s
e
c
ur
it
y
to
identif
y
indi
viduals
a
nd
c
ontr
ol
a
c
c
e
s
s
to
r
e
s
tr
icte
d
a
r
e
a
s
.
F
or
e
xa
mpl
e
,
many
a
i
r
por
ts
us
e
r
e
mot
e
r
e
c
ognit
ion
s
ys
tems
to
c
he
c
k
pa
s
s
e
nge
r
s
a
n
d
e
ns
ur
e
that
they
a
r
e
pe
o
ple
r
e
gis
ter
e
d
in
the
a
ir
li
ne
's
da
taba
s
e
,
whic
h
c
a
n
he
lp
im
pr
ove
s
a
f
e
ty
a
nd
e
f
f
ici
e
nc
y.
I
n
biom
e
tr
ics
,
f
a
c
ial
r
e
c
ognit
ion
is
us
e
d
in
identi
f
ica
ti
on
a
nd
ve
r
if
ica
ti
on
s
ys
tems
,
s
uc
h
a
s
us
ing
the
f
a
c
e
to
unlock
a
s
mar
tphone
[
8]
–
[
10]
.
T
his
s
e
c
ur
it
y
r
e
li
e
s
he
a
vil
y
on
the
s
ys
tem's
a
bil
it
y
to
r
e
c
ognize
f
a
c
e
s
a
c
c
ur
a
tely,
ther
e
by
incr
e
a
s
ing
e
f
f
icie
nc
y
a
nd
making
it
e
a
s
ier
f
or
us
e
r
s
to
ope
n
their
s
mar
tphones
withou
t
e
nter
ing
a
p
a
s
s
wor
d.
De
s
pit
e
it
s
many
a
dva
ntage
s
,
f
a
c
ial
r
e
c
ognit
ion
tec
hnology
f
a
c
e
s
s
e
ve
r
a
l
c
ha
ll
e
nge
s
that
mus
t
be
ov
e
r
c
ome.
One
of
the
main
c
ha
ll
e
nge
s
is
the
leve
l
of
a
c
c
ur
a
c
y
that
ha
s
ye
t
to
r
e
a
c
h
it
s
maximum
po
tential
[
11]
.
F
a
c
tor
s
s
uc
h
a
s
li
ghti
ng
c
ondit
ions
,
f
a
c
ial
a
ngles
,
a
nd
va
r
ying
us
e
r
de
mogr
a
phics
c
a
n
inf
luenc
e
the
c
ons
is
tenc
y
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
E
nhanc
ing
facia
l
r
e
c
ognit
ion
ac
c
ur
ac
y
thr
ough
featur
e
e
x
tr
ac
ti
ons
and
ar
ti
fi
c
ial
ne
ur
al
…
(
A
dhi
K
us
nadi
)
1057
a
c
c
ur
a
c
y
of
identif
ying
indi
viduals
[
12
]
,
[
13]
.
T
hi
s
r
a
is
e
s
c
onc
e
r
ns
r
e
ga
r
ding
mi
s
identif
ica
ti
on
a
nd
potential
bias
in
thi
s
tec
hnology,
s
o
r
e
s
e
a
r
c
h
a
nd
de
ve
lop
ment
mus
t
be
c
onduc
ted
to
e
ns
ur
e
c
ons
is
tent
a
nd
r
e
li
a
ble
pe
r
f
or
manc
e
a
c
r
os
s
a
ll
us
e
r
gr
oups
.
Va
r
ious
m
e
thods
ha
ve
be
e
n
pr
opos
e
d
to
im
pr
ove
f
a
c
e
r
e
c
ognit
ion
a
c
c
ur
a
c
y,
including
f
e
a
tur
e
e
xtr
a
c
ti
on
[
14
]
.
T
he
m
a
in
goa
l
of
f
e
a
tur
e
e
xtr
a
c
ti
on
is
to
e
xtr
a
c
t
e
s
s
e
nti
a
l
f
e
a
tur
e
s
f
r
om
f
a
c
ial
im
a
ge
s
to
r
e
duc
e
nois
e
dur
ing
c
las
s
if
ic
a
ti
on
a
nd
incr
e
a
s
e
a
c
c
ur
a
c
y
[
15]
.
T
o
ove
r
c
ome
thi
s
tec
hnologi
c
a
l
c
ha
ll
e
nge
,
pr
e
v
ious
r
e
s
e
a
r
c
h
ha
s
pr
opos
e
d
s
e
ve
r
a
l
a
lgor
it
hms
,
including
dis
c
r
e
te
c
os
ine
t
r
a
ns
f
or
m
(
DC
T
)
[
16]
,
gr
a
y
leve
l
c
o
-
oc
c
ur
r
e
nc
e
matr
ix
(
GL
C
M
)
[
17
]
,
a
nd
Ga
u
s
s
ian
mi
xtur
e
model
(
GM
M
)
[
18
]
.
P
r
e
vious
r
e
s
e
a
r
c
h
us
ing
GL
C
M
a
nd
ba
c
kw
a
r
d
pr
opa
ga
ti
on
o
f
e
r
r
o
r
s
(
b
a
c
kpr
opa
ga
ti
on
)
s
howe
d
89%
a
c
c
ur
a
c
y
with
a
d
is
tanc
e
of
1
pixel
[
19]
.
T
he
r
e
s
ult
s
of
c
onvolut
ion
a
l
ne
ur
a
l
ne
twor
ks
(
C
NN
)
r
e
s
e
a
r
c
h
with
the
Ale
xNe
t
a
r
c
hit
e
c
tur
e
pr
ovide
a
n
a
c
c
ur
a
c
y
of
98.
5%
[
20]
.
T
he
r
e
s
e
a
r
c
h
r
e
s
ult
s
us
ing
DC
T
ha
ve
a
n
a
c
c
ur
a
c
y
of
95%
[
21]
.
T
he
r
e
s
e
a
r
c
h
r
e
s
ult
s
us
ing
low
-
f
r
e
que
nc
y
DC
T
da
t
a
f
or
f
a
c
e
a
nd
pa
lm
r
e
c
ognit
ion
p
r
oduc
e
d
a
n
a
c
c
ur
a
c
y
of
95%
.
T
he
s
e
s
tudi
e
s
s
how
s
igni
f
ica
nt
leve
ls
of
f
a
c
ial
r
e
c
ognit
ion
ac
c
ur
a
c
y,
but
ther
e
is
s
ti
ll
r
oom
f
or
im
p
r
ove
ment,
e
s
pe
c
ially
in
de
a
li
ng
with
f
a
c
ial
va
r
iations
invol
ving
c
ha
nge
s
in
pos
it
ion
a
nd
o
r
ienta
ti
on.
A
c
ompr
e
he
ns
ive
li
ter
a
tur
e
r
e
view
wa
s
c
onduc
t
e
d
that
c
a
r
e
f
ull
y
e
xplo
r
e
s
the
methodology
a
nd
theor
e
ti
c
a
l
f
ounda
ti
ons
r
e
late
d
to
f
a
c
e
r
e
c
ognit
ion
,
with
a
pa
r
ti
c
ular
f
oc
us
on
s
e
ve
r
a
l
vit
a
l
a
ppr
oa
c
he
s
,
i
nc
ludi
ng
DC
T
[
22]
,
GM
M
[
23]
,
b
a
c
kpr
opa
ga
ti
on
,
a
nd
C
NN
[
24]
.
An
in
-
de
pth
a
na
lys
is
is
c
onduc
ted
to
unde
r
s
tand
the
a
dva
ntage
s
,
we
a
kne
s
s
e
s
,
a
nd
late
s
t
de
ve
lopm
e
nts
i
n
e
a
c
h
method
or
theor
y
dis
c
us
s
e
d.
S
our
c
e
s
o
f
inf
or
mation
take
n
include
pr
e
vious
s
c
ientif
ic
jou
r
na
ls
,
a
c
a
de
mi
c
thes
e
s
,
e
s
s
e
nt
ial
a
r
ti
c
les
,
a
nd
r
e
leva
nt
digi
tal
r
e
s
our
c
e
s
.
S
our
c
e
s
e
lec
ti
on
is
ba
s
e
d
on
s
tr
ict
c
r
it
e
r
ia
to
e
ns
ur
e
the
va
li
dit
y
a
nd
r
e
leva
nc
e
of
the
inf
or
mation
pr
e
s
e
nted.
T
he
li
ter
a
tu
r
e
r
e
view
a
ls
o
c
ove
r
s
the
late
s
t
li
ter
a
tur
e
in
thi
s
f
ield,
e
ns
ur
ing
that
the
knowle
dge
pr
e
s
e
nted
r
e
mains
r
e
leva
nt
a
nd
up
-
to
-
da
te.
T
his
r
e
s
e
a
r
c
h
a
im
s
to
ove
r
c
ome
thes
e
obs
tac
les
by
c
ombi
ning
DC
T
a
nd
GM
M
f
e
a
tur
e
a
c
c
ur
a
c
y
tec
hniques
.
T
his
r
e
s
e
a
r
c
h
will
a
ls
o
e
va
luate
the
potential
of
a
r
ti
f
icia
l
ne
ur
a
l
ne
twor
k
(
AN
N)
a
lgor
it
h
ms
s
uc
h
a
s
b
a
c
kpr
opa
g
a
ti
on
a
nd
C
NN
,
whic
h
ha
ve
be
e
n
pr
ove
n
e
f
f
e
c
ti
ve
in
objec
t
r
e
c
ognit
ion.
T
he
s
e
a
lgor
it
hms
will
be
int
e
gr
a
ted
with
f
e
a
tur
e
e
xtr
a
c
ti
on
to
inc
r
e
a
s
e
f
a
c
ial
r
e
c
ognit
ion
a
c
c
ur
a
c
y,
e
s
pe
c
ially
f
or
f
a
c
ial
va
r
iat
ions
that
include
f
a
c
ial
pos
it
ion
a
nd
or
ienta
ti
on
c
ha
nge
s
.
T
h
is
pr
oc
e
s
s
will
invol
ve
a
c
a
r
e
f
ul
t
r
a
ini
ng
s
tage
to
e
ns
ur
e
the
int
e
gr
a
ted
a
lgor
it
h
ms
c
a
n
r
e
c
ognize
f
a
c
ial
va
r
iat
ions
a
c
c
ur
a
tely,
including
f
a
c
ial
pos
it
ion
a
nd
or
i
e
ntation
c
ha
nge
s
.
How
e
ve
r
,
it
is
e
s
s
e
nti
a
l
to
note
that
c
omb
ini
ng
thes
e
a
lgor
it
hms
c
a
n
a
l
s
o
incr
e
a
s
e
the
c
omp
utational
c
ompl
e
xit
y
of
the
s
ys
tem,
whic
h
c
a
n
a
f
f
e
c
t
pr
oc
e
s
s
ing
ti
me.
B
y
c
ombi
ning
the
DC
T
a
nd
GM
M
f
e
a
tur
e
e
xtr
a
c
ti
on
methods
with
the
AN
N
a
lgor
it
h
m,
thi
s
r
e
s
e
a
r
c
h
c
a
n
s
igni
f
ica
ntl
y
c
ont
r
ibut
e
to
the
de
ve
lopm
e
nt
o
f
f
a
c
ial
r
e
c
ognit
ion
te
c
hnology.
T
he
r
e
s
ult
s
of
thi
s
r
e
s
e
a
r
c
h
a
r
e
e
xpe
c
ted
to
incr
e
a
s
e
the
a
c
c
ur
a
c
y
of
f
a
c
ial
r
e
c
ognit
ion
s
igni
f
ica
ntl
y.
T
hus
,
thi
s
r
e
s
e
a
r
c
h
ope
ns
up
ne
w
oppor
tuni
ti
e
s
f
o
r
de
ve
lopi
ng
mor
e
s
ophis
ti
c
a
ted
f
a
c
ial
r
e
c
ognit
ion
tec
hnology
a
nd
c
a
n
p
r
ovide
mor
e
e
f
f
e
c
ti
ve
s
olut
ions
in
va
r
ious
c
ontexts
.
2.
M
E
T
HO
D
T
he
method
us
e
d
in
thi
s
r
e
s
e
a
r
c
h
include
s
s
tage
s
,
a
s
de
tailed
in
F
igur
e
1
.
T
his
r
e
s
e
a
r
c
h
dif
f
e
r
s
f
r
om
pr
e
vious
r
e
s
e
a
r
c
h
[
20]
in
that
it
doe
s
not
r
e
move
th
e
im
a
ge
ba
c
kgr
ound.
S
e
lec
ti
on
be
c
a
us
e
it
is
r
e
quir
e
d
a
t
the
DC
T
f
e
a
tur
e
e
xtr
a
c
ti
on
s
tage
.
DC
T
inher
e
ntl
y
f
oc
us
e
s
on
a
nd
mi
ti
ga
tes
high
-
f
r
e
qu
e
nc
y
da
ta,
e
f
f
e
c
ti
ve
ly
mi
nim
izing
the
in
f
luenc
e
of
ba
c
kgr
ound
c
ompone
nts
,
s
o
e
xpli
c
it
ba
c
kgr
ound
r
e
moval
is
unne
c
e
s
s
a
r
y
.
T
his
r
e
s
e
a
r
c
h
methodology
a
ppr
oa
c
h
us
e
s
f
e
a
tur
e
e
xtr
a
c
ti
on
f
r
om
im
a
ge
s
us
ing
DC
T
a
t
low
f
r
e
que
nc
ies
s
o
that
it
ha
s
the
potential
to
ha
v
e
mo
r
e
inf
o
r
mation
that
c
a
n
be
us
e
d
to
identif
y
f
e
a
tur
e
s
i
n
im
a
ge
s
[
25]
.
Ne
xt,
the
GM
M
a
lgo
r
it
hm
obtains
f
a
c
ial
im
a
ge
textur
e
inf
o
r
mation,
whic
h
c
a
n
be
us
e
d
a
s
a
n
ident
if
ica
ti
on
f
e
a
tur
e
[
26
]
.
Af
ter
f
e
a
tur
e
e
xtr
a
c
ti
on,
f
a
c
ial
da
ta
is
r
e
c
ognize
d
us
ing
AN
N
a
lgor
it
hms
,
na
mely
b
a
c
kpr
o
pa
ga
ti
on
a
nd
C
NN
.
B
a
c
kpr
opa
ga
ti
on
a
lgor
it
hms
lea
r
n
quickly
by
c
omput
ing
s
yna
pti
c
upda
tes
us
ing
f
e
e
dba
c
k
c
onne
c
ti
ons
to
s
e
nd
e
r
r
or
s
ignals
[
27]
.
C
NN
wa
s
c
h
os
e
n
a
s
a
c
las
s
if
ica
ti
on
method
be
c
a
us
e
of
it
s
c
omp
a
ti
bil
it
y
with
im
a
ge
da
ta,
whe
r
e
C
NN
c
a
n
indepe
nde
ntl
y
lea
r
n
a
nd
e
xtr
a
c
t
f
e
a
tur
e
s
f
r
om
a
n
im
a
ge
[
28
]
.
I
n
a
dd
it
ion,
the
f
e
a
tur
e
s
e
xtr
a
c
ted
by
DC
T
a
nd
GM
M
a
r
e
c
ombi
ne
d
to
im
pr
ove
the
a
c
c
ur
a
c
y
of
f
a
c
e
r
e
c
ognit
ion
in
th
e
f
a
c
e
of
va
r
iations
,
s
uc
h
a
s
c
ha
nge
s
in
f
a
c
ial
pos
it
ion
a
nd
or
ienta
ti
on.
T
he
r
e
s
ult
s
of
the
tr
a
ined
AN
N
model
will
be
tes
ted,
a
nd
it
s
a
c
c
ur
a
c
y
will
be
c
a
lcula
ted.
2
.
1.
I
m
age
p
r
e
p
r
oc
e
s
s
in
g
F
igur
e
2
il
lus
tr
a
tes
a
s
a
mpl
e
of
s
ome
of
the
da
tas
e
ts
us
e
d.
T
he
f
a
c
ial
da
tas
e
t
us
e
d
in
thi
s
r
e
s
e
a
r
c
h
is
the
Olivetti
R
e
s
e
a
r
c
h
L
a
bor
a
tor
y
(
OR
L
)
da
tas
e
t
[
29]
,
whic
h
c
ons
is
ts
of
410
f
a
c
ial
i
mage
s
f
r
om
41
dif
f
e
r
e
nt
pe
ople,
a
nd
e
a
c
h
pe
r
s
on
ha
s
10
f
a
c
ial
im
a
ge
s
,
a
n
e
xa
mpl
e
of
whic
h
c
a
n
be
s
e
e
n
in
F
igur
e
2(
a
)
.
E
a
c
h
im
a
ge
is
80×
70
pixels
in
s
ize
a
nd
is
in
J
P
G
f
o
r
mat.
T
he
s
e
c
ond
da
ta
s
e
t
is
the
Ya
le
da
tas
e
t.
T
his
da
ta
ha
s
1
65
f
a
c
ial
im
a
ge
s
f
r
om
15
di
f
f
e
r
e
nt
pe
ople
with
dif
f
e
r
e
nt
f
a
c
ial
im
a
ge
s
;
a
n
e
xa
mpl
e
c
a
n
be
s
e
e
n
in
F
igur
e
2(
b)
;
e
a
c
h
i
mage
is
320×
243
pixels
a
nd
is
in
GI
F
f
o
r
mat.
T
he
thi
r
d
da
tas
e
t
is
the
J
a
pa
ne
s
e
f
e
male
f
a
c
ial
e
xpr
e
s
s
ion
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
2
,
Apr
il
20
25
:
1
056
-
1
066
1058
(
J
AFF
E
)
da
tas
e
t,
whic
h
c
ontains
213
f
a
c
ial
im
a
ge
s
with
10
J
a
pa
ne
s
e
f
e
male
f
a
c
e
s
,
a
n
e
xa
mpl
e
of
whic
h
c
a
n
be
s
e
e
n
in
F
igur
e
2(
c
)
.
T
he
s
ize
of
e
a
c
h
i
mage
is
256×
256
pixels
in
T
I
F
F
f
or
mat
.
T
he
s
e
thr
e
e
da
tas
e
ts
we
r
e
c
hos
e
n
be
c
a
us
e
they
ha
ve
a
va
r
iety
o
f
s
ubjec
ts
,
s
o
t
he
y
ha
ve
s
uf
f
icie
nt
r
e
s
our
c
e
s
to
t
r
a
in
the
model
we
l
l.
B
e
f
or
e
the
da
ta
is
us
e
d,
it
is
pr
oc
e
s
s
e
d
to
im
pr
ove
it
s
s
ui
tabili
ty
to
the
model
a
nd
f
e
a
t
ur
e
e
xtr
a
c
ti
on.
T
he
i
mage
is
c
onve
r
ted
f
r
om
r
e
d,
g
r
e
e
n,
blue
(
R
GB
)
c
olor
[
29
]
to
gr
a
ys
c
a
le
[
30]
.
I
n
thi
s
pr
oc
e
s
s
,
the
int
e
ns
it
y
of
the
gr
a
y
c
olor
is
maintaine
d
s
o
that
the
im
a
ge
s
ti
ll
c
ontains
e
s
s
e
nti
a
l
inf
or
mation.
F
igur
e
1.
R
e
s
e
a
r
c
h
m
e
thod
(
a
)
(
b)
(
c
)
F
igur
e
2.
S
a
mpl
e
o
f
(
a
)
OR
L
d
a
tas
e
t,
(
b
)
Ya
le
d
a
ta
s
e
t,
a
nd
(
c
)
J
AFF
E
d
a
tas
e
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
E
nhanc
ing
facia
l
r
e
c
ognit
ion
ac
c
ur
ac
y
thr
ough
featur
e
e
x
tr
ac
ti
ons
and
ar
ti
fi
c
ial
ne
ur
al
…
(
A
dhi
K
us
nadi
)
1059
M
e
thod
de
ve
lopm
e
nt
r
e
quir
e
s
that
the
da
ta
be
pr
oc
e
s
s
e
d
f
ir
s
t
by
c
onve
r
ti
ng
R
GB
c
olor
s
to
gr
a
y
s
c
a
le
[
30]
,
[
31]
.
T
he
b
r
ight
ne
s
s
leve
l
r
e
pr
e
s
e
nts
the
pixel
int
e
ns
it
y
va
lue
in
a
gr
a
ys
c
a
le
im
a
ge
,
mea
s
ur
e
d
on
a
gr
a
ys
c
a
le
f
r
om
0
(
blac
k)
to
255
(
white)
.
T
he
goa
l
of
thi
s
s
tage
is
to
s
im
pli
f
y
the
a
na
lys
is
a
s
it
r
e
d
uc
e
s
the
c
ompl
e
xit
y
of
the
da
ta
f
r
o
m
thr
e
e
c
olo
r
c
ha
nne
ls
t
o
one
c
olor
c
ha
nne
l
a
nd
r
e
tai
ns
e
s
s
e
nti
a
l
inf
or
mati
on
a
bout
the
br
ight
ne
s
s
leve
ls
r
e
quir
e
d
f
or
f
a
c
ial
r
e
c
ognit
ion
.
2
.
2.
F
e
a
t
u
r
e
e
xt
r
ac
t
ion
2
.
2.
1
.
L
ow
-
f
r
e
q
u
e
n
c
y
d
is
c
r
e
t
e
c
os
in
e
t
r
an
s
f
or
m
f
e
at
u
r
e
e
xt
r
ac
t
ion
L
ow
-
f
r
e
que
nc
y
DC
T
[
31]
is
a
tec
hnique
u
s
e
d
in
f
e
a
tur
e
e
xt
r
a
c
ti
on,
us
ua
ll
y
a
ppli
e
d
in
s
ignal
pr
oc
e
s
s
ing
tas
ks
s
uc
h
a
s
im
a
ge
a
nd
a
udio
a
na
lys
is
;
a
vis
ua
li
z
a
ti
on
o
f
the
DC
T
c
oe
f
f
icie
nt
matr
ix
c
a
n
be
s
e
e
n
in
F
igur
e
3
.
I
t
invol
ve
s
c
onve
r
ti
ng
da
ta
int
o
a
ne
w
r
e
pr
e
s
e
ntation
that
c
ombi
ne
s
c
os
ine
f
unc
ti
ons
with
va
r
ying
f
r
e
que
nc
ies
.
I
n
thi
s
c
ontext,
“
low
f
r
e
que
nc
y”
c
a
pt
ur
e
s
s
low
a
nd
s
igni
f
ica
nt
da
ta
va
r
iations
while
e
li
mi
na
ti
ng
f
a
s
t
f
luctua
ti
ons
[
32]
.
T
his
is
e
s
pe
c
ially
us
e
f
ul
in
tas
ks
th
a
t
e
mphas
iz
e
ba
s
ic
s
tr
uc
tur
e
s
or
f
unda
menta
l
c
ha
r
a
c
ter
is
ti
c
s
.
T
o
us
e
low
-
f
r
e
que
nc
y
DC
T
f
or
f
e
a
t
ur
e
e
xtr
a
c
ti
on
,
da
ta,
s
uc
h
a
s
a
n
i
mage
,
is
div
ided
in
t
o
blocks
,
a
nd
DC
T
is
a
ppli
e
d
to
e
a
c
h
block
.
T
he
r
e
s
ult
ing
c
o
e
f
f
icie
nts
,
whic
h
e
mphas
ize
low
-
f
r
e
que
nc
y
inf
o
r
m
a
ti
on,
a
r
e
s
e
lec
ted
a
nd
c
ombi
ne
d
in
to
a
f
e
a
tu
r
e
ve
c
tor
.
T
his
c
ompac
t
r
e
pr
e
s
e
ntation
p
r
e
s
e
r
ve
s
im
por
tant
f
e
a
tur
e
s
while
r
e
duc
ing
dim
e
ns
ions
,
making
i
t
us
e
f
ul
f
o
r
tas
ks
s
uc
h
a
s
im
a
ge
c
ompr
e
s
s
ion,
pa
tt
e
r
n
r
e
c
ognit
ion,
a
nd
da
ta
a
na
lys
is
.
At
thi
s
s
tage
,
the
pr
e
vious
ly
p
r
oc
e
s
s
e
d
da
tas
e
t
is
e
xtr
a
c
ted
us
ing
DC
T
to
pr
oduc
e
c
oe
f
f
icie
nts
with
thr
e
e
types
of
f
r
e
que
nc
ies
.
T
he
f
r
e
que
nc
y
that
wil
l
be
us
e
d
is
low
be
c
a
us
e
it
is
a
t
thi
s
f
r
e
que
nc
y
th
a
t
f
a
c
ial
f
e
a
tur
e
s
a
r
e
s
tor
e
d.
L
ow
c
oe
f
f
icie
nts
,
only
8
×
8
pi
xe
ls
in
s
ize
,
a
r
e
s
e
lec
ted
a
ga
in
a
t
the
top
lef
t
of
t
he
DC
T
matr
ix
c
oe
f
f
icie
nt
im
a
ge
[
33
]
.
F
igur
e
3.
DC
T
c
oe
f
icie
nt
matr
ix
[
34
]
B
e
s
ide
s
,
low
f
r
e
que
nc
ies
a
r
e
s
e
lec
ted
ba
s
e
d
o
n
r
e
s
e
a
r
c
h
[
35]
.
T
his
r
e
s
e
a
r
c
h
tes
ted
va
r
ious
c
ombi
na
ti
ons
of
DC
T
low
-
f
r
e
que
nc
y
pe
r
c
e
ntage
s
on
the
de
tec
tor
f
e
a
tur
e
s
a
c
c
ur
a
c
y
leve
l.
(
,
)
=
2
(
)
(
)
∑
−
1
=
0
∑
−
1
=
0
(
,
)
[
(
2
+
1
)
2
]
[
(
2
+
1
)
2
]
(
1)
W
he
r
e
(
,
)
is
the
DC
T
va
lue
in
f
r
e
que
nc
y
c
oor
dinate
s
(
,
)
, f
(
,
)
i
s
the
pixel
va
lue
in
s
pa
ti
a
l
c
oor
dinate
s
(
,
)
,
N
is
the
DC
T
b
lock
s
ize
,
a
nd
(
)
is
a
c
os
ine
f
unc
ti
on
r
e
late
d
to
f
r
e
que
nc
y
(
)
.
2
.
2.
2
.
Gau
s
s
ian
m
at
r
ix
m
od
e
l
f
e
at
u
r
e
s
GM
M
[
36]
is
a
pr
oba
bil
is
ti
c
model
that
a
na
lyze
s
da
ta
with
ove
r
lapping
Ga
us
s
ian
c
omponents
[
37]
.
T
his
model
c
a
n
be
us
e
d
f
o
r
da
ta
c
lus
ter
ing
a
nd
c
a
n
a
ls
o
be
us
e
d
to
id
e
nti
f
y
the
unde
r
lyi
ng
dis
tr
ibut
i
on
of
the
da
ta
[
38]
.
T
he
ba
s
ic
f
or
mul
a
f
or
GM
M
is
a
s
(
2)
[
3
9]
.
(
X
|
Θ
)
=
∑
=
1
∙
(
X
|
µ
,
∑
)
(
2)
W
he
r
e
P
(
X|Θ
)
is
pr
oba
bil
it
y
of
da
ta
X
given
pa
r
a
mete
r
Θ
in
GM
M
,
K
is
number
of
Ga
us
s
ian
c
omp
one
nts
in
GM
M
,
is
the
we
ight
f
or
e
a
c
h
Ga
us
s
ian
c
omponent,
whic
h
indi
c
a
tes
the
pr
opor
ti
on
or
pr
oba
bil
it
y
of
oc
c
ur
r
e
nc
e
of
that
c
omponent,
a
nd
N
(
X|
µ
,
∑
)
is
the
G
a
us
s
ian
de
ns
it
y
f
unc
ti
on
f
or
c
omponent
k
with
me
a
n
µ
a
nd
c
ova
r
ianc
e
matr
ix
∑
.
T
he
main
objec
ti
ve
of
GM
M
is
to
f
ind
the
op
ti
mal
Θ
pa
r
a
mete
r
s
that
give
the
highes
t
p
r
oba
bil
it
y
f
or
the
pr
ovided
da
ta.
T
o
de
ter
mi
ne
the
model
pa
r
a
mete
r
s
,
GM
M
us
e
s
the
e
xpe
c
tation
maximi
z
a
ti
on
(
E
M
)
a
lgor
it
hm,
whe
r
e
in
the
e
xpe
c
tation
(
E
)
s
tage
,
the
e
xpe
c
ted
va
lue
of
e
a
c
h
Ga
us
s
ian
c
omponent
in
the
mi
xtur
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
2
,
Apr
il
20
25
:
1
056
-
1
066
1060
is
c
a
lcula
ted,
a
nd
in
the
maxim
iza
ti
on
(
M
)
s
tage
,
the
model
pa
r
a
mete
r
s
a
r
e
r
e
c
a
lcula
ted
us
ing
that
pr
e
dicte
d
va
lue
[
40]
.
T
he
it
e
r
a
ti
on
pr
oc
e
s
s
c
onti
nue
s
unti
l
c
o
nve
r
ge
nc
e
oc
c
ur
s
whe
n
the
GM
M
pa
r
a
mete
r
s
a
r
e
s
ta
ble,
or
the
dif
f
e
r
e
nc
e
be
twe
e
n
s
uc
c
e
s
s
ive
it
e
r
a
ti
ons
be
c
omes
mi
nim
a
l.
T
he
E
M
p
r
oc
e
s
s
in
GM
M
invol
ve
s
t
wo
s
tage
s
[
41]
:
i)
E
s
tage
is
e
s
ti
mate
s
the
pos
ter
io
r
pr
oba
bil
it
y
of
e
a
c
h
Ga
us
s
ian
c
omponent
(
gr
oup)
f
or
e
a
c
h
da
ta
point
.
T
he
f
o
r
mul
a
f
or
c
a
lcula
ti
ng
the
pos
ter
ior
pr
oba
bil
it
y
(
r
e
s
pons
ibi
li
ty)
f
or
e
a
c
h
Ga
us
s
ian
c
omponent
a
nd
ii
)
M
s
tage
is
u
s
e
s
the
pos
ter
ior
pr
oba
bil
it
ies
e
s
ti
mate
d
in
the
E
s
tage
to
upda
te
the
GM
M
pa
r
a
mete
r
s
,
i
nc
ludi
ng
the
we
ight
s
,
mea
n,
a
nd
c
ova
r
ianc
e
matr
ix
.
2
.
3.
T
r
ain
in
g
d
at
a
2
.
3.
1
.
Dat
a
s
p
li
t
t
i
n
g
Af
ter
e
xtr
a
c
ti
on
,
the
da
ta
is
divi
de
d
in
to
t
r
a
ini
ng,
v
a
li
da
ti
on,
a
nd
tes
ti
ng
.
T
r
a
ini
ng
da
ta
is
us
e
d
to
t
r
a
in
f
a
c
ial
r
e
c
ognit
ion
a
lgor
it
hms
s
o
that
they
c
a
n
und
e
r
s
tand
the
da
ta
f
or
c
las
s
if
ica
ti
on
pur
pos
e
s
.
Va
li
da
ti
on
da
ta
is
us
e
d
to
e
va
luate
model
pe
r
f
o
r
manc
e
dur
ing
the
tr
a
ini
ng
pr
oc
e
s
s
but
is
no
t
us
e
d
to
tr
a
in
the
model
it
s
e
lf
.
Da
ta
tes
ti
ng
is
the
f
inal
s
tage
to
tes
t
model
pe
r
f
or
manc
e
on
da
ta
that
ha
s
ne
ve
r
be
e
n
s
e
e
n
be
f
or
e
.
Da
ta
dis
t
r
ibut
ion
with
a
pr
opor
ti
on
of
60
%
t
r
a
ini
ng
da
ta,
20%
va
li
d
a
ti
on
da
ta,
a
nd
20%
tes
ti
ng
da
ta.
T
he
divi
s
ion
is
a
s
tr
a
tegic
a
ppr
oa
c
h
in
mac
hine
lea
r
ning
a
nd
da
ta
s
c
ienc
e
a
im
e
d
a
t
opti
mi
z
ing
the
model
de
ve
lopm
e
nt
pr
oc
e
s
s
.
T
his
s
pe
c
if
ic
dis
tr
ibut
ion
r
e
f
lec
ts
a
ba
lanc
e
d
a
ppr
oa
c
h,
e
ns
ur
ing
s
uf
f
icie
nt
da
ta
f
or
t
r
a
ini
ng
while
a
ll
oc
a
ti
ng
a
mpl
e
r
e
s
our
c
e
s
f
or
both
model
tuni
ng
a
nd
unbias
e
d
e
va
l
ua
ti
on
[
42]
,
[
43]
.
2
.
3.
2
.
Dat
a
p
r
oc
e
s
s
in
g
m
e
t
h
od
s
wi
t
h
d
is
c
r
e
t
e
c
os
in
e
t
r
an
s
f
or
m
an
d
gau
s
s
ian
m
a
t
r
ix
m
o
d
e
l
T
he
method
c
ombi
ne
s
the
DC
T
tr
a
ns
f
or
mati
on
wit
h
the
GM
M
model
to
pr
oduc
e
a
r
e
pr
e
s
e
ntation
of
da
ta
f
e
a
tur
e
s
with
a
f
oc
us
on
low
f
r
e
que
nc
ies
us
ing
DC
T
a
nd
then
a
pplyi
ng
the
GM
M
model
f
or
f
ur
the
r
a
na
lys
is
a
nd
da
ta
c
las
s
if
ica
ti
on.
T
he
pr
oc
e
s
s
be
gins
by
c
ha
nging
the
da
ta
int
o
a
one
-
dim
e
ns
ional
(
1D)
f
o
r
m
t
hr
ough
a
r
e
s
ha
ping
pr
oc
e
s
s
,
a
ll
owing
f
ur
ther
p
r
oc
e
s
s
ing
us
i
ng
the
DC
T
tr
a
ns
f
or
mation
.
He
r
e
is
the
ps
e
udoc
od
e
f
or
the
c
ombi
na
ti
on:
//
F
unc
ti
on
to
e
xtr
a
c
t
low
-
f
r
e
que
nc
y
c
omponents
of
DC
T
Function performDCTLowFrequency(inputSignal, lowFrequencyThreshol
d):
//
Apply
DC
T
to
the
input
s
ignal
transformedSignal=DCTAlgorithm(inputSignal)
//
E
xt
r
a
c
t
low
-
f
r
e
que
nc
y
c
omponents
ba
s
e
d
on
the
s
pe
c
if
ied
thr
e
s
hold
lowFrequencyComponents=extractLowFrequency (transformedSignal, lowFrequencyThreshold)
return
lowFrequencyComponents
//
F
unc
ti
on
to
ini
ti
a
li
z
e
a
nd
tr
a
in
a
Ga
us
s
ian
M
ixt
ur
e
M
ode
l
(
GM
M
)
function trainGMM(data, numberOfComponents):
//
I
nit
ialize
a
GM
M
with
the
s
pe
c
if
ied
numbe
r
of
c
omponents
gmm=InitializeGMM(numberOfComponents)
//
T
r
a
in
the
GM
M
on
the
pr
ovided
da
ta
gmm.fit(data)
return gmm
//
M
a
in
pr
oc
e
s
s
ing
f
unc
ti
on
to
c
ombi
ne
DC
T
(
low
f
r
e
que
nc
y)
a
nd
GM
M
function processSignal(inputSignal):
//
S
tep
1:
Apply
DC
T
to
the
input
s
ignal
a
nd
e
xtr
a
c
t
low
-
f
r
e
que
nc
y
c
omponents
//
De
f
ine
a
thr
e
s
hold
to
identif
y
low
-
f
r
e
que
nc
y
c
omponents
lowFrequencyThreshold=defineThreshold()
lowFrequencyDCTOutput=performDCTLowFrequency (inputSignal, lowFrequencyThreshold)
//
Optional:
F
ur
ther
f
e
a
tur
e
e
xt
r
a
c
ti
on
or
s
e
lec
ti
on
f
r
om
the
low
-
f
r
e
que
nc
y
DC
T
output
features=extr
actFeatures(lowFrequencyDCTOutput)
//
S
tep
2:
T
r
a
in
a
GM
M
on
the
low
-
f
r
e
que
nc
y
DC
T
//
S
e
lec
t
the
nu
mber
o
f
GM
M
c
omponents
ba
s
e
d
on
a
ppli
c
a
ti
on
-
s
pe
c
if
ic
c
r
it
e
r
ia
numberOfGMMComponents=
selectNumberOfComponents()
gmmModel=trainGMM (features, numberOfGMMComp
onents)
return gmmModel
Af
ter
pr
oc
e
s
s
ing
the
da
ta
th
r
ough
a
c
ombi
na
ti
on
of
DC
T
a
nd
GM
M
,
the
da
ta
is
input
to
the
AN
N.
2
.
4.
F
ac
ial
r
e
c
ogn
it
ion
ac
c
u
r
ac
y
2
.
4.
1.
B
ac
k
p
r
op
agat
ion
B
a
c
kpr
opa
ga
ti
on
[
44]
,
a
vit
a
l
t
r
a
ini
ng
tec
hnique
in
the
c
ontext
of
ANN
us
e
d
in
va
r
ious
a
ppli
c
a
ti
ons
,
including
f
a
c
ial
r
e
c
ognit
ion.
B
a
c
kpr
opa
ga
ti
on
wa
s
c
hos
e
n
in
thi
s
r
e
s
e
a
r
c
h
be
c
a
u
s
e
of
it
s
c
r
it
ica
l
a
bil
it
y
to
tr
a
in
ANN
,
e
s
pe
c
ially
f
or
c
ompl
e
x
tas
ks
s
uc
h
a
s
f
a
c
e
r
e
c
ognit
ion.
B
a
c
kpr
opa
ga
ti
on
a
ll
ows
the
ne
twor
k
t
o
upda
te
we
ight
s
a
nd
bias
e
s
ba
s
e
d
on
pr
e
diction
e
r
r
or
s
,
e
na
bli
n
g
e
r
r
or
c
o
r
r
e
c
ti
on
a
nd
pe
r
f
o
r
manc
e
im
pr
ove
m
e
nts
ove
r
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
E
nhanc
ing
facia
l
r
e
c
ognit
ion
ac
c
ur
ac
y
thr
ough
featur
e
e
x
tr
ac
ti
ons
and
ar
ti
fi
c
ial
ne
ur
al
…
(
A
dhi
K
us
nadi
)
1061
ti
me.
B
a
c
kpr
opa
ga
ti
on
ha
s
3
types
of
laye
r
s
,
na
me
ly
i
)
input
laye
r
is
a
pa
r
t
c
ons
is
ti
ng
of
unit
s
whe
r
e
the
unit
s
s
tar
t
f
r
om
1
to
n
;
ii
)
a
hidden
laye
r
is
a
laye
r
that
c
ons
is
ts
of
a
t
lea
s
t
one
laye
r
,
whe
r
e
e
a
c
h
laye
r
c
o
ns
is
ts
of
s
e
ve
r
a
l
unit
s
;
a
nd
ii
i)
the
output
laye
r
is
e
a
c
h
ne
u
r
o
n
unit
in
the
input
laye
r
c
onne
c
ted
to
a
l
l
uni
ts
in
th
e
hidden
laye
r
be
low
it
.
Vic
e
ve
r
s
a
,
e
ve
r
y
unit
in
the
hidden
laye
r
is
c
onne
c
ted
to
a
ll
unit
s
in
the
ou
tput
laye
r.
I
n
F
igu
r
e
4
,
the
b
a
c
kpr
opa
ga
ti
on
a
r
c
hit
e
c
tur
e
is
pr
e
s
e
nted,
il
lus
tr
a
ti
ng
the
s
tr
uc
tur
e
a
nd
r
e
lations
hips
a
mong
the
thr
e
e
types
of
laye
r
s
.
T
he
s
e
laye
r
s
c
ons
is
t
of
the
input
laye
r
,
the
hidden
laye
r
,
a
nd
the
ou
tp
ut
laye
r
.
E
a
c
h
laye
r
p
lays
a
s
pe
c
if
ic
r
ole
i
n
the
ba
c
kpr
opa
ga
ti
on
pr
oc
e
s
s
by
a
djus
ti
ng
the
we
ight
s
ba
s
e
d
on
the
c
a
lcula
ted
e
r
r
or
,
e
na
bli
ng
the
model
to
lea
r
n
mo
r
e
a
c
c
ur
a
tely.
F
igur
e
4.
Ar
c
hit
e
c
tur
e
of
b
a
c
kpr
opa
ga
ti
on
[
45]
B
a
c
kpr
opa
ga
ti
on
a
lgor
it
hms
a
r
e
ke
y
in
im
p
r
oving
ne
twor
k
pe
r
f
o
r
manc
e
f
or
c
ompl
e
x
tas
ks
[
46]
.
T
his
a
lgor
it
hm
wor
ks
be
c
a
us
e
a
ne
ur
a
l
ne
twor
k
c
a
n
i
mpr
ove
by
unde
r
s
tanding
a
nd
c
or
r
e
c
ti
n
g
pr
e
dicti
on
e
r
r
o
r
s
dur
ing
tr
a
ini
ng
.
T
he
t
r
a
ini
ng
be
gins
with
ini
ti
a
li
z
i
ng
the
we
ight
s
a
nd
bias
e
s
f
or
e
a
c
h
ne
ur
on
in
the
ne
twor
k.
Ne
xt,
t
r
a
ini
ng
da
ta
in
the
f
or
m
o
f
f
a
c
ial
im
a
ge
s
or
f
a
c
ial
e
xa
mpl
e
s
is
pr
e
s
e
nted
to
the
ne
twor
k
.
T
his
da
ta
f
lows
thr
ough
the
n
e
twor
k
in
a
s
e
r
ies
of
s
teps
c
a
ll
e
d
f
e
e
df
or
wa
r
d,
whe
r
e
e
a
c
h
ne
ur
on
pe
r
f
o
r
ms
c
a
lcula
ti
ons
ba
s
e
d
on
the
we
ight
s
a
nd
input
s
ignals
it
r
e
c
e
ives
.
At
the
e
nd
of
the
f
e
e
df
or
wa
r
d
pr
oc
e
s
s
,
the
ne
twor
k
pr
oduc
e
s
pr
e
dictions
of
f
e
a
tur
e
s
or
c
ha
r
a
c
ter
is
ti
c
s
of
the
e
xtr
a
c
ted
f
a
c
e
s
.
Ne
xt,
a
c
ompar
is
on
is
made
be
twe
e
n
the
pr
e
dicte
d
r
e
s
ult
s
a
nd
the
c
o
r
r
e
c
t
labe
ls
,
r
e
pr
e
s
e
nti
ng
the
pe
r
s
on's
identit
y
in
the
i
mage
.
T
he
pr
e
diction
e
r
r
o
r
is
mea
s
ur
e
d
a
s
a
n
e
r
r
or
;
the
ne
xt
s
tep
is
r
e
tur
ning
(
b
a
c
kpr
opa
ga
ti
ng)
thi
s
e
r
r
or
thr
ough
the
n
e
twor
k
.
T
his
i
nvolves
c
a
lcula
ti
ng
the
e
r
r
or
gr
a
dient
a
ga
ins
t
the
we
ight
s
a
nd
bias
e
s
in
e
a
c
h
ne
ur
on.
T
he
we
ight
s
a
nd
bias
a
r
e
upda
ted
by
s
ubtr
a
c
ti
ng
th
e
e
r
r
or
gr
a
dient
f
r
om
the
c
ur
r
e
nt
we
ight
s
a
nd
bias
,
a
nd
thi
s
pr
oc
e
s
s
is
r
e
pe
a
ted
r
e
pe
a
tedly
f
or
e
a
c
h
tr
a
i
n
ing
e
xa
mpl
e
in
the
da
tas
e
t.
T
he
b
a
c
kpr
opa
ga
ti
on
a
lgor
it
hm
tr
ies
to
f
ind
a
s
e
t
of
we
ight
s
that
opti
m
ize
s
the
ne
twor
k's
a
bil
it
y
to
r
e
c
ognize
f
a
c
e
s
with
high
a
c
c
ur
a
c
y.
T
his
c
a
n
take
ti
me
a
nd
many
f
a
c
tor
s
,
s
uc
h
a
s
the
lea
r
ni
ng
r
a
te,
the
number
o
f
ne
u
r
ons
in
th
e
hidden
laye
r
,
a
nd
the
number
of
it
e
r
a
ti
ons
r
e
quir
e
d
.
T
his
it
e
r
a
ti
ve
p
r
oc
e
s
s
gr
a
dua
ll
y
im
pr
ove
s
the
ne
ur
a
l
ne
twor
k's
a
bil
it
y
to
r
e
c
ognize
pa
tt
e
r
ns
a
nd
f
e
a
tur
e
s
on
f
a
c
e
s
unti
l
it
f
inally
r
e
a
c
he
s
a
s
uf
f
icie
nt
leve
l
of
a
c
c
ur
a
c
y.
T
h
e
r
e
f
or
e
,
b
a
c
kpr
opa
ga
ti
on
is
a
c
r
i
ti
c
a
l
f
ounda
ti
on
in
de
ve
lopi
ng
s
ophis
ti
c
a
ted
a
nd
e
f
f
icie
nt
ANN
in
va
r
ious
a
ppli
c
a
ti
ons
,
including
f
a
c
ial
r
e
c
ognit
ion
.
W
it
h
a
de
e
p
unde
r
s
tanding
of
thes
e
a
lgor
it
hms
,
de
ve
loper
s
a
nd
r
e
s
e
a
r
c
he
r
s
c
a
n
a
c
hieve
opti
mal
r
e
s
ult
s
in
c
ompl
e
x
f
a
c
ial
r
e
c
ognit
i
on
tas
ks
.
2
.
4.
2
.
Convol
u
t
ion
al
n
e
u
r
al
n
e
t
wor
k
m
o
d
e
l
T
he
C
NN
model
is
a
de
e
p
lea
r
ning
a
r
c
hit
e
c
tur
e
de
s
igned
to
tac
kle
im
a
ge
a
nd
im
a
ge
pr
oc
e
s
s
ing
ta
s
ks
[
47]
.
C
NN
wa
s
c
hos
e
n
in
thi
s
r
e
s
e
a
r
c
h
be
c
a
us
e
of
it
s
e
xc
e
ll
e
nt
a
bil
it
y
to
ha
ndle
im
a
ge
p
r
oc
e
s
s
in
g
tas
ks
,
including
f
a
c
e
r
e
c
ognit
ion.
C
NN
s
a
r
e
s
pe
c
if
ica
ll
y
de
s
igned
to
e
xtr
a
c
t
hier
a
r
c
hica
l
f
e
a
tur
e
s
f
r
om
im
a
ge
da
ta,
e
na
bli
ng
a
de
e
pe
r
unde
r
s
tanding
of
vis
ua
l
s
tr
uc
tur
e
s
a
nd
pa
tt
e
r
ns
.
C
NN
c
ons
is
ts
of
s
e
ve
r
a
l
laye
r
s
,
including
c
onvolut
ional
laye
r
s
that
hie
r
a
r
c
hica
ll
y
e
xtr
a
c
t
e
s
s
e
nti
a
l
f
e
a
tur
e
s
f
r
om
im
a
ge
s
,
r
e
c
ti
f
ied
li
ne
a
r
unit
(
R
e
L
U)
a
c
ti
va
ti
on
laye
r
s
to
int
r
oduc
e
non
-
li
ne
a
r
it
y,
pooli
n
g
laye
r
s
that
r
e
duc
e
da
ta
di
mens
ions
,
a
nd
f
ull
y
c
o
nne
c
ted
laye
r
s
that
play
a
r
ole
in
de
c
is
ion
making.
C
NN
is
tr
a
ined
us
ing
mac
hine
lea
r
ning
a
lgor
it
h
ms
li
ke
ba
c
kpr
opa
ga
ti
on
to
opti
mi
z
e
pe
r
f
o
r
manc
e
in
tas
ks
s
uc
h
a
s
im
a
ge
c
las
s
if
ica
ti
on.
W
i
th
it
s
a
bil
it
y
to
a
uto
matica
ll
y
e
xtr
a
c
t
f
e
a
tur
e
s
f
r
om
im
a
ge
da
ta
,
C
NN
ha
s
domi
na
ted
many
i
mage
pr
oc
e
s
s
ing
a
ppli
c
a
ti
ons
.
I
t
is
a
c
or
ne
r
s
tone
in
de
ve
lopi
ng
tec
hnologi
e
s
li
ke
objec
t
r
e
c
ognit
ion,
a
utonom
ous
ve
hicle
s
,
a
nd
medic
a
l
i
mage
a
na
lys
is
.
T
he
L
e
Ne
t
model,
a
ls
o
known
a
s
L
e
Ne
t
-
5,
wa
s
e
mpl
oye
d
in
thi
s
r
e
s
e
a
r
c
h.
I
t
r
e
p
r
e
s
e
nts
one
of
the
e
a
r
ly
mi
les
tones
in
de
ve
lopi
ng
C
NN
s
[
48]
,
[
49
]
.
De
s
igned
by
L
e
C
un
e
t
al.
[
50]
in
1998
,
L
e
Ne
t
wa
s
ini
ti
a
ll
y
c
r
e
a
ted
f
or
ha
ndwr
it
ten
c
ha
r
a
c
ter
r
e
c
ogn
it
ion
tas
ks
.
T
his
model
c
ons
is
ts
of
c
onvolut
ional
laye
r
s
tha
t
uti
li
z
e
f
il
ter
s
to
e
xtr
a
c
t
f
e
a
tur
e
s
f
r
om
input
i
mage
s
,
f
oll
owe
d
by
pooli
ng
laye
r
s
that
r
e
duc
e
da
ta
di
mens
ions
.
S
ubs
e
que
ntl
y,
two
f
ul
ly
c
onne
c
ted
laye
r
s
pr
oc
e
s
s
t
he
s
e
f
e
a
tur
e
s
a
nd
ge
ne
r
a
te
pr
e
dic
ti
ons
.
L
e
Ne
t
int
r
o
duc
e
d
the
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
2
,
Apr
il
20
25
:
1
056
-
1
066
1062
c
onc
e
pt
of
c
onvolut
ional
laye
r
s
,
whic
h
ha
s
now
be
c
ome
the
c
or
e
of
moder
n
C
NN
a
r
c
hit
e
c
tu
r
e
s
.
Althou
gh
ther
e
a
r
e
now
mor
e
e
xtens
ive
a
nd
c
ompl
e
x
C
NN
a
r
c
hit
e
c
tur
e
s
,
L
e
Ne
t
r
e
mains
a
s
igni
f
ica
nt
landma
r
k
in
de
e
p
lea
r
ning
a
nd
im
a
ge
pr
o
c
e
s
s
ing
his
tor
y,
pa
ving
the
wa
y
f
or
f
ur
the
r
innovations
in
thi
s
f
ield.
I
n
F
igur
e
5,
you
c
a
n
obs
e
r
ve
the
vis
ua
l
r
e
pr
e
s
e
ntation
of
the
'
mode
l
C
NN
L
e
Ne
t'
.
T
his
diagr
a
m
il
lus
tr
a
tes
the
a
r
c
hit
e
c
tur
e
of
L
e
Ne
t,
s
howc
a
s
ing
the
a
r
r
a
nge
ment
of
c
onvolut
io
na
l
laye
r
s
,
pooli
ng
laye
r
s
,
a
nd
f
ull
y
c
onne
c
ted
laye
r
s
.
F
igur
e
5.
M
ode
l
C
NN
L
e
Ne
t
[
51]
2
.
4.
3
.
T
e
s
t
in
g
an
d
e
valu
at
ion
T
e
s
ti
ng
a
nd
e
va
luating
thi
s
f
a
c
ial
r
e
c
ognit
ion
m
ode
l
us
e
s
s
e
ve
r
a
l
ke
y
metr
ics
to
mea
s
ur
e
model
pe
r
f
or
manc
e
[
52]
.
F
i
r
s
t,
the
a
c
c
ur
a
c
y
a
nd
los
s
dur
ing
tr
a
ini
ng
a
nd
tes
ti
ng
will
be
c
a
lcula
ted.
Ac
c
ur
a
c
y
s
hows
how
f
a
r
the
model
r
e
c
ognize
s
f
a
c
e
s
c
or
r
e
c
tl
y
[
53]
,
while
los
s
mea
s
ur
e
s
how
w
e
ll
the
model
mi
nim
iz
e
s
e
r
r
or
s
[
54]
.
Additi
ona
ll
y
,
e
va
luations
we
r
e
pe
r
f
or
med
us
ing
c
las
s
if
ica
ti
on
r
e
por
ts
a
nd
c
onf
us
ion
matr
ice
s
t
o
a
s
s
e
s
s
the
model's
f
a
c
e
r
e
c
o
gnit
ion
pe
r
f
or
manc
e
[
55]
,
inc
ludi
ng
a
c
c
ur
a
c
y,
los
s
,
r
e
c
a
ll
,
pr
e
c
is
ion,
a
nd
F
1
s
c
or
e
.
Ne
xt,
to
mea
s
ur
e
the
leve
l
o
f
de
ter
mi
na
ti
on
of
the
mod
e
l,
the
c
or
r
e
lation
c
oe
f
f
icie
nt
is
us
e
d,
whic
h
mea
s
ur
e
s
the
c
los
e
ne
s
s
of
the
r
e
lations
hip
be
twe
e
n
the
indepe
nde
nt
va
r
iable
(
f
e
a
tur
e
e
xtr
a
c
ti
on
da
ta)
a
nd
the
d
e
pe
nde
nt
va
r
iable
(
f
a
c
e
r
e
c
ognit
ion
a
c
c
ur
a
c
y
leve
l)
to
pr
ovi
de
a
n
unde
r
s
tanding
of
the
e
xtent
to
whic
h
the
m
ode
l
c
a
n
dif
f
e
r
e
nti
a
te
be
twe
e
n
di
f
f
e
r
e
nt
a
nd
s
im
il
a
r
f
a
c
e
s
in
the
da
tas
e
t
[
56]
.
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
I
n
thi
s
s
e
c
ti
on,
e
xpe
r
im
e
ntal
a
na
lys
is
is
c
a
r
r
ied
out
us
ing
the
OR
L
da
tas
e
t
c
on
s
is
ti
ng
of
410
im
a
ge
s
f
r
om
41
pe
ople.
S
o,
e
a
c
h
f
a
c
e
ha
s
ten
im
a
ge
s
wi
t
h
dif
f
e
r
e
nt
f
a
c
ial
e
xp
r
e
s
s
ions
a
nd
a
ngles
.
All
i
mage
s
we
r
e
gr
a
ys
c
a
led
be
f
or
e
the
DC
T
t
r
a
ns
f
or
mation.
E
xtr
a
c
ti
ng
low
-
f
r
e
que
nc
y
da
ta
f
r
om
DC
T
is
c
a
r
r
ied
out
b
y
taking
the
8×
8
i
mage
a
t
the
top
lef
t.
Af
ter
e
xtr
a
c
ti
ng
the
low
-
f
r
e
que
nc
y
DC
T
da
ta,
GM
M
is
a
ppli
e
d
to
e
a
c
h
da
ta,
pr
oduc
ing
a
GM
M
matr
ix
f
or
e
a
c
h
da
ta.
3
.
1.
I
m
p
lem
e
n
t
at
ion
of
f
e
at
u
r
e
e
xt
r
ac
t
ion
a
n
d
b
ac
k
p
r
op
agat
ion
At
thi
s
s
tage
,
the
b
a
c
kpr
opa
ga
ti
on
method
is
a
ppl
ied
to
t
r
a
in
a
f
a
c
ial
r
e
c
ognit
ion
model
us
ing
da
ta
e
xtr
a
c
ted
thr
ough
DC
T
a
nd
GM
M
.
T
his
method
plays
a
c
r
it
ica
l
r
ole
in
a
djus
ti
ng
the
model's
we
ight
s
to
mi
nim
ize
e
r
r
o
r
a
nd
im
pr
ove
pe
r
f
or
manc
e
ove
r
ti
m
e
.
As
s
h
own
in
T
a
ble
1
,
the
r
e
s
ult
s
indi
c
a
te
that
the
tr
a
ini
ng
a
c
c
ur
a
c
y
a
c
hieve
d
with
b
a
c
kpr
opa
ga
ti
on
a
nd
DC
T
-
GM
M
f
e
a
tur
e
e
xtr
a
c
ti
on
r
e
mains
r
e
latively
low,
s
u
gge
s
ti
ng
that
f
ur
ther
opti
mi
z
a
ti
on
or
a
lt
e
r
na
ti
ve
a
ppr
oa
c
he
s
may
be
r
e
quir
e
d
.
T
a
ble
1.
R
e
s
ult
of
da
ta
tr
a
ini
ng
tr
ial
wi
t
h
DC
T
a
nd
GM
M
f
e
a
tur
e
e
xtr
a
c
ti
on
L
e
a
r
ni
ng
r
a
te
H
id
de
n
node
A
c
c
ur
a
c
y
(%)
E
poc
h
T
r
a
in
in
g
ti
me
(
s
)
50
4.88
402
1.79
0.01
150
1.22
320
3.89
350
3.66
290
8.29
50
3.66
1643
16.7
0.001
150
2.44
1270
12.56
350
1.22
1096
13.73
50
4.88
616
8.73
0.005
150
1.22
473
8.8
350
2.44
424
7.95
50
4.88
134
2.11
0.2
150
0.00
126
1.65
350
3.66
123
1.26
50
4.88
124
1.25
0.8
150
1.22
121
2.02
350
4.88
120
3.02
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
E
nhanc
ing
facia
l
r
e
c
ognit
ion
ac
c
ur
ac
y
thr
ough
featur
e
e
x
tr
ac
ti
ons
and
ar
ti
fi
c
ial
ne
ur
al
…
(
A
dhi
K
us
nadi
)
1063
F
r
om
the
r
e
s
ult
s
in
T
a
ble
1,
the
highes
t
a
c
c
ur
a
c
y
va
lue
is
f
ound
in
the
lea
r
ning
r
a
te
pa
r
a
mete
r
,
na
mely
0.
8,
a
nd
hidden
node
s
,
na
mely
50
,
with
a
n
a
c
c
ur
a
c
y
va
lue
of
4
.
88%
.
I
n
thi
s
e
xpe
r
im
e
n
t,
GM
M
e
xtr
a
c
ti
on
us
e
s
thr
e
e
n
-
c
omponent
pa
r
a
mete
r
s
a
nd
a
r
a
ndom
s
t
a
te
1.
B
a
s
e
d
on
thes
e
r
e
s
ult
s
,
the
s
maller
the
lea
r
ning
r
a
te
pa
r
a
mete
r
va
lue,
the
f
a
s
ter
the
tr
a
ini
n
g
da
ta
tr
a
ini
ng
ti
me
be
c
a
us
e
,
ba
s
e
d
on
the
tr
ials
c
a
r
r
ied
out,
the
lea
r
ning
r
a
te
va
lue
of
0
.
001
ha
s
t
he
longes
t
tr
a
ini
ng
ti
me
wa
s
16.
7
s
e
c
onds
on
50
hidden
node
s
a
nd
with
the
s
a
me
number
of
hidden
node
s
with
a
lea
r
ning
r
a
te
va
lue
of
0
.
8
the
f
a
s
tes
t
tr
a
ini
ng
ti
me
wa
s
1.
25
s
e
c
onds
.
3
.
2.
I
m
p
lem
e
n
t
at
ion
of
f
e
at
u
r
e
e
xt
r
ac
t
ion
a
n
d
c
on
volu
t
ion
al
n
e
u
r
al
n
e
t
wor
k
S
e
ve
r
a
l
s
tudi
e
s
ha
ve
s
hown
that
a
lea
r
ning
r
a
te
o
f
0.
0001
yielde
d
the
be
s
t
pe
r
f
or
manc
e
,
a
nd
tes
ti
ng
wa
s
c
onduc
ted
onc
e
the
da
ta
s
e
t
wa
s
pr
e
pa
r
e
d.
T
he
tes
ti
ng
pr
oc
e
s
s
wa
s
c
a
r
r
ied
out
to
e
va
luate
the
model’
s
a
c
c
ur
a
c
y
a
nd
r
obus
tnes
s
a
c
r
os
s
dif
f
e
r
e
nt
da
tas
e
ts
.
T
he
r
e
s
ult
s
a
r
e
pr
e
s
e
nted
in
T
a
ble
2
f
o
r
models
us
i
ng
only
C
NN
,
a
nd
in
T
a
ble
3
f
or
models
uti
li
z
ing
f
e
a
tur
e
e
xtr
a
c
ti
on
methods
with
DC
T
,
GM
M
,
a
nd
C
NN
.
T
a
ble
2.
T
e
s
t
r
e
s
ult
s
us
ing
C
NN
without
f
e
a
tur
e
e
xtr
a
c
ti
on
D
a
ta
s
e
t
A
c
c
ur
a
c
y
(%)
T
r
a
in
in
g
ti
me
(
s
)
O
R
L
97.2
372.59
Y
a
le
97.9
1330.48
J
A
F
F
E
99.2
2017.49
T
a
ble
3.
T
e
s
t
r
e
s
ult
s
us
ing
DC
T
,
GM
M
,
a
nd
C
NN
D
a
ta
s
e
t
A
c
c
ur
a
c
y
(%)
T
r
a
in
in
g
ti
me
(
s
)
O
R
L
98.2
360.59
Y
a
le
98.9
1210.8
J
A
F
F
E
100
1749.49
F
r
o
m
t
he
r
e
s
u
l
ts
o
f
t
he
e
xpe
r
i
men
ts
t
ha
t
ha
ve
be
e
n
c
a
r
r
ied
ou
t
,
i
t
c
a
n
be
s
e
e
n
in
T
a
b
le
3
tha
t
t
he
a
dd
i
ti
on
o
f
th
e
DC
T
a
nd
GM
M
m
e
t
ho
ds
w
it
h
a
l
e
a
r
n
i
ng
r
a
te
o
f
0
.
0
00
1
a
n
d
GM
M
pa
r
a
me
te
r
s
(
n
_c
o
m
pon
e
n
t
1
0
,
r
a
ndom_s
tate
300)
pr
oduc
e
s
the
be
s
t
a
c
c
ur
a
c
y
c
o
mpar
e
d
to
jus
t
us
ing
c
las
s
if
ica
ti
on
f
r
om
C
NN
a
s
s
hown
in
T
a
ble
2.
T
his
e
xpe
r
i
ment
pr
oduc
e
s
a
n
a
c
c
ur
a
c
y
o
f
98.
2
%
with
a
t
r
a
ini
ng
t
im
e
of
360
s
e
c
onds
on
t
he
OR
L
da
tas
e
t,
98.
9%
a
c
c
ur
a
c
y
with
a
tr
a
ini
ng
ti
me
of
12
10
s
e
c
onds
on
the
Ya
le
da
ta
s
e
t,
a
nd
100%
a
c
c
ur
a
c
y
with
a
tr
a
ini
ng
ti
me
o
f
1749
s
e
c
onds
on
the
J
AFF
E
da
tas
e
t.
B
e
f
or
e
looki
ng
f
o
r
the
c
oe
f
f
icie
nt
of
de
ter
m
inati
on
va
lue,
look
f
or
the
c
or
r
e
lation
c
oe
f
f
icie
nt
va
lue
or
r
with
a
r
e
s
ult
of
0
.
989
in
the
model
c
r
e
a
ted.
T
he
n,
the
va
l
ue
of
the
c
oe
f
f
icie
nt
of
de
ter
mi
na
ti
on
is
c
a
lcula
ted
by
incr
e
a
s
ing
the
c
or
r
e
lation
c
oe
f
f
icie
nt
to
the
powe
r
of
t
wo
or
r
2
,
whic
h
is
then
mul
ti
pli
e
d
by
100%
to
obtain
the
pe
r
c
e
ntage
.
T
he
c
oe
f
f
icie
n
t
o
f
de
te
r
mi
na
ti
on
va
lue
ob
tai
ne
d
wa
s
97.
9%
,
wh
ich
mea
ns
th
e
mo
de
l
c
r
e
a
te
d
ha
s
a
s
t
r
on
g
c
or
r
e
la
ti
o
n
a
nd
c
a
n
s
how
tha
t
the
e
f
f
ic
ienc
y
of
th
e
metho
d
us
e
d
inf
luenc
e
s
the
f
a
c
ial
r
e
c
og
nit
ion
v
a
lue
b
y
9
7
.
9
%
,
a
nd
the
r
e
ma
inde
r
is
i
nf
luenc
e
d
by
ot
he
r
f
a
c
t
or
s
by
2.
3%
.
T
his
r
e
s
e
a
r
c
h's
incr
e
a
s
e
in
a
c
c
ur
a
c
y
a
nd
tes
ti
ng
ti
me
wa
s
c
a
u
s
e
d
by
a
dding
f
e
a
tur
e
e
xtr
a
c
ti
on
methods
,
na
mely
DC
T
a
nd
dif
f
e
r
e
nt
c
las
s
if
ier
s
(
C
NN
)
.
App
lyi
ng
low
-
f
r
e
que
nc
y
DC
T
he
lps
e
li
mi
na
te
no
is
e
a
t
medium
a
nd
high
f
r
e
que
nc
ies
s
uc
h
a
s
ba
c
kgr
ound
,
s
kin,
a
nd
ha
ir
.
B
e
c
a
us
e
thi
s
r
e
s
e
a
r
c
h
f
oc
us
e
s
on
low
-
f
r
e
que
nc
y
f
e
a
tur
e
s
s
uc
h
a
s
the
no
s
e
,
mout
h,
a
nd
e
ye
s
.
T
he
r
ole
of
C
NN
a
s
a
c
las
s
if
ica
ti
on
method
a
ls
o
c
ontr
ibut
e
s
to
incr
e
a
s
ing
a
c
c
ur
a
c
y
be
c
a
us
e
the
c
onvolut
ion
meth
od
us
e
d
by
C
NN
he
lps
the
model
lea
r
n
im
a
ge
s
s
o
that
the
r
e
s
ult
ing
model
c
a
n
c
las
s
if
y
im
a
ge
s
much
mor
e
a
c
c
ur
a
tely.
T
h
is
r
e
s
e
a
r
c
h
s
hows
s
igni
f
ica
nt
s
uc
c
e
s
s
c
ompar
e
d
to
pr
e
vious
r
e
s
e
a
r
c
h
with
OR
L
da
ta
us
ing
the
GL
C
M
a
nd
b
a
c
kpr
opa
ga
ti
on
method
s
,
whic
h
obtaine
d
a
c
c
ur
a
c
y
r
e
s
ult
s
of
89%
.
T
his
r
e
s
e
a
r
c
h
c
ombi
ne
s
thr
e
e
met
hods
,
na
mely
DC
T
,
GM
M
,
a
nd
C
NN
,
whic
h
a
c
hi
e
ve
d
a
n
a
c
c
ur
a
c
y
leve
l
of
98.
2
%
with
a
s
igni
f
ica
nt
incr
e
a
s
e
in
a
c
c
ur
a
c
y
due
to
the
c
ombi
na
ti
on
of
f
e
a
tur
e
e
x
tr
a
c
ti
on,
e
s
pe
c
ially
DC
T
,
whic
h
us
e
s
low
f
r
e
que
nc
ies
.
I
n
te
r
ms
of
t
r
a
ini
ng
ti
me
,
thi
s
r
e
s
e
a
r
c
h
r
e
a
c
he
d
360.
59
s
e
c
onds
,
lowe
r
than
pr
e
vious
r
e
s
e
a
r
c
h,
whic
h
took
3.
53
s
e
c
onds
.
3
.
3.
Dis
c
u
s
s
ion
T
his
r
e
s
e
a
r
c
h
s
hows
s
igni
f
ica
nt
im
pr
ove
ments
c
o
mpar
e
d
to
p
r
e
vious
r
e
s
e
a
r
c
h
[
49
]
,
whic
h
us
e
d
a
c
ombi
ne
d
GL
C
M
a
nd
ne
ur
a
l
ne
twor
ks
method
with
a
n
a
c
c
ur
a
c
y
of
89%
,
tr
a
ini
ng
ti
me
o
f
3
.
53
s
e
c
onds
,
pr
e
c
is
ion
va
lue
of
0.
85,
r
e
c
a
ll
va
lue
of
0
.
86,
a
nd
f
1
s
c
or
e
of
85%
.
T
his
r
e
s
e
a
r
c
h
a
im
s
to
f
il
l
the
ga
ps
in
pr
e
vious
r
e
s
e
a
r
c
h
by
a
pplyi
ng
va
r
ious
f
e
a
tur
e
e
xtr
a
c
ti
on
te
c
hniques
,
s
uc
h
a
s
DC
T
a
nd
GM
M
,
a
nd
uti
li
z
ing
C
NN
to
im
pr
ove
f
a
c
ial
r
e
c
ognit
ion
a
c
c
ur
a
c
y.
T
he
r
e
s
ult
s
of
thi
s
s
tudy
s
how
va
r
iations
in
a
c
c
ur
a
c
y
de
p
e
nding
on
pa
r
a
mete
r
s
s
uc
h
a
s
lea
r
ning
r
a
te
a
nd
number
o
f
hid
de
n
node
s
.
T
he
highes
t
a
c
c
ur
a
c
y
va
lue,
a
lt
hough
r
e
latively
low,
is
4.
88
%
,
a
c
hieve
d
with
a
c
ombi
na
ti
on
of
a
lea
r
ning
r
a
te
o
f
0.
8
a
nd
50
hidden
node
s
u
s
ing
the
b
a
c
kpr
opa
ga
ti
on
method.
De
s
pit
e
the
low
a
c
c
ur
a
c
y,
the
r
e
latively
f
a
s
t
tr
a
ini
ng
ti
me
is
a
t
r
a
de
-
of
f
.
T
he
n,
the
im
pleme
ntation
o
f
DC
T
a
nd
GM
M
f
e
a
tur
e
e
xtr
a
c
ti
on,
whic
h
wa
s
p
r
oc
e
s
s
e
d
us
ing
a
C
NN
,
s
howe
d
s
i
gnif
ica
nt
a
c
c
ur
a
c
y
r
e
s
ult
s
,
na
mely
a
n
a
c
c
ur
a
c
y
o
f
98
.
2%
a
nd
a
t
r
a
ini
ng
ti
me
o
f
360
s
e
c
onds
on
the
OR
L
da
tas
e
t,
a
n
a
c
c
ur
a
c
y
of
98.
9%
a
nd
a
tr
a
ini
ng
ti
me
of
1210
s
e
c
onds
on
the
da
tas
e
t.
Ya
le,
100%
a
c
c
ur
a
c
y
a
nd
174
9
s
e
c
onds
tr
a
ini
ng
ti
me
on
the
J
AFF
E
da
tas
e
t.
T
he
tr
a
ini
ng
t
im
e
is
longer
than
the
b
a
c
kpr
opa
ga
ti
on
method,
e
s
pe
c
ially
on
the
J
AFF
E
da
tas
e
t
,
but
the
r
e
s
ult
s
s
how
that
th
e
c
ombi
na
ti
on
of
DC
T
,
GM
M
,
a
nd
C
NN
p
r
ovides
s
upe
r
ior
pe
r
f
or
manc
e
.
T
he
t
r
a
ini
ng
ti
me
f
or
the
b
a
c
kpr
opa
g
a
ti
on
method
is
r
e
latively
f
a
s
t,
e
ve
n
though
the
a
c
c
ur
a
c
y
is
low.
At
the
s
a
me
ti
me,
the
c
ombi
na
ti
on
o
f
DC
T
,
GM
M
,
a
nd
C
NN
r
e
quir
e
s
longer
tr
a
ini
ng
ti
me,
e
s
pe
c
ially
on
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
2
,
Apr
il
20
25
:
1
056
-
1
066
1064
the
J
AFF
E
da
tas
e
t,
but
pr
oduc
e
s
high
a
c
c
ur
a
c
y
.
S
o,
the
r
e
s
ult
s
obtaine
d
f
r
om
the
c
ombi
na
ti
on
of
DC
T
,
GM
M
,
a
nd
C
NN
s
how
that
the
be
ne
f
it
s
r
e
c
e
ived
f
r
om
thi
s
method
a
r
e
h
igher
than
the
incr
e
a
s
e
in
tr
a
ini
ng
ti
m
e
.
I
n
c
onc
lus
ion,
thi
s
r
e
s
e
a
r
c
h
ove
r
c
a
me
the
li
mi
tatio
ns
of
p
r
e
vious
r
e
s
e
a
r
c
h
by
a
pplyi
ng
va
r
i
ous
f
e
a
tur
e
e
xtr
a
c
ti
on
tec
hniques
a
nd
c
las
s
if
ier
s
.
Although
tr
a
i
ning
ti
me
c
a
n
be
a
li
mi
ti
ng
f
a
c
tor
,
the
r
e
s
ult
s
obtai
ne
d
f
r
om
the
c
ombi
na
ti
on
of
DC
T
,
GM
M
,
a
nd
C
NN
s
how
a
s
igni
f
ica
nt
incr
e
a
s
e
in
a
c
c
ur
a
c
y.
W
it
h
a
c
oe
f
f
i
c
ient
of
de
ter
mi
na
ti
on
of
97.
9
%
,
thi
s
r
e
s
e
a
r
c
h
s
igni
f
ica
ntl
y
c
ontr
ibut
e
s
to
unde
r
s
tanding
the
f
a
c
tor
s
inf
luenc
ing
f
a
c
ial
r
e
c
ognit
ion
r
e
s
ult
s
.
I
t
is
hope
d
that
thi
s
r
e
s
e
a
r
c
h
c
a
n
be
c
ome
a
r
e
f
e
r
e
nc
e
in
the
de
ve
lopm
e
nt
of
f
a
c
ial
r
e
c
ognit
ion
tec
hnology
in
the
f
u
tur
e
a
nd
c
a
n
ove
r
c
ome
s
e
ve
r
a
l
obs
tac
le
s
f
a
c
e
d
in
pr
e
vious
r
e
s
e
a
r
c
h.
4.
CONC
L
USI
ON
T
his
r
e
s
e
a
r
c
h
e
xplor
e
s
DC
T
a
nd
GM
M
f
e
a
tur
e
e
xtr
a
c
ti
on
to
im
pr
ove
f
a
c
ial
r
e
c
ognit
ion
a
c
c
ur
a
c
y,
c
ombi
ne
d
with
b
a
c
kpr
opa
ga
ti
on
a
nd
C
NN
t
r
a
ini
n
g
methods
.
T
he
tes
t
r
e
s
ult
s
s
how
that
the
b
a
c
kpr
opa
ga
ti
on
method
with
DC
T
a
nd
GM
M
f
e
a
tur
e
e
xtr
a
c
ti
on
pr
ovides
a
li
mi
ted
a
c
c
ur
a
c
y
of
4.
88%
but
with
the
a
dva
ntage
of
a
r
e
latively
f
a
s
t
tr
a
ini
ng
ti
me
of
1
.
25
s
e
c
onds
.
On
the
othe
r
ha
nd,
c
omb
ini
ng
DC
T
,
GM
M
,
a
nd
C
NN
s
igni
f
ica
ntl
y
im
pr
ove
s
the
a
c
c
ur
a
c
y
r
a
te,
r
e
a
c
hing
9
8.
2,
98.
9
,
a
nd
100
%
f
or
the
OR
L
,
Ya
le
,
a
nd
J
AFF
E
da
tas
e
ts
,
r
e
s
pe
c
ti
ve
ly.
Although
it
r
e
quir
e
s
mor
e
e
xtende
d
tr
a
ini
ng,
thi
s
c
ombi
na
ti
on
pr
ovides
s
upe
r
ior
r
e
s
ult
s
a
nd
s
hows
e
xc
e
ll
e
nt
potential
f
or
de
ve
lopi
ng
f
a
c
ial
r
e
c
ognit
io
n
tec
hnology.
Ana
lys
is
of
the
c
oe
f
f
icie
nt
o
f
de
ter
m
ination
of
97
.
9%
c
onf
i
r
ms
that
the
e
f
f
icie
nc
y
o
f
the
metho
d
us
e
d
g
r
e
a
tl
y
inf
luenc
e
s
the
f
a
c
ial
r
e
c
ognit
ion
r
e
s
ult
s
,
with
other
f
a
c
tor
s
c
ont
r
ibut
ing
a
r
ound
2
.
3%
.
T
h
is
c
onc
lus
ion
highl
ight
s
the
s
tr
e
ngth
of
the
de
ve
loped
model
in
ha
ndli
ng
va
r
iations
in
f
a
c
ial
pos
it
ion
a
nd
or
ienta
ti
o
n
a
nd
im
pr
ove
s
ove
r
a
ll
a
c
c
ur
a
c
y.
C
ompar
is
on
with
pr
e
vious
r
e
s
e
a
r
c
h
s
hows
a
pos
it
ive
e
volut
ion
in
thi
s
tec
hno
logy,
a
nd
the
de
ve
lopm
e
nt
of
ne
w
methods
,
e
s
pe
c
ially
the
c
ombi
na
ti
on
of
DC
T
,
GM
M
,
a
nd
C
NN
,
ope
ns
the
door
to
f
u
r
ther
a
dva
nc
e
s
in
f
a
c
ial
r
e
c
ognit
ion.
T
h
e
r
e
f
or
e
,
thi
s
r
e
s
e
a
r
c
h
make
s
a
va
luable
c
ontr
ibut
ion
to
the
de
ve
lopm
e
nt
of
f
a
c
ial
r
e
c
ognit
ion
tec
hnology
,
w
it
h
wide
a
ppli
c
a
ti
on
potential
in
va
r
ious
s
e
c
tor
s
,
e
s
pe
c
iall
y
in
im
pr
oving
the
s
e
c
ur
it
y
a
nd
r
e
li
a
bil
it
y
o
f
in
divi
dua
l
iden
ti
f
ica
ti
on.
T
hus
,
thi
s
innovative
c
ombi
na
ti
on
ope
ns
up
a
ne
w
di
r
e
c
ti
on
in
im
p
r
oving
f
a
c
ial
r
e
c
ognit
ion
a
c
c
ur
a
c
y
a
nd
pos
it
ively
im
pa
c
ts
pe
r
s
ona
l
identif
ica
ti
on
tec
hnology's
s
e
c
ur
it
y
de
ve
lopm
e
nt.
AC
KNOWL
E
DGE
M
E
NT
S
Our
a
ppr
e
c
iation
goe
s
to
Ke
mendikbud
-
R
is
tek
R
e
publi
c
of
I
ndone
s
ia
with
number
of
c
ontr
a
c
t
073/E
5/P
G.
02
.
00.
P
L
/2023
,
f
o
r
thei
r
f
inanc
ial
s
uppor
t,
a
nd
to
Unive
r
s
it
a
s
M
ult
im
e
dia
Nus
a
ntar
a
f
o
r
p
r
ovidi
ng
ne
c
e
s
s
a
r
y
r
e
s
our
c
e
s
.
RE
F
E
RE
NC
E
S
[
1]
S
.
M
.
B
a
h
a
nd
F
.
M
in
g,
“
A
n
im
pr
ove
d
f
a
c
e
r
e
c
ogni
ti
on
a
lg
or
it
hm
a
nd
it
s
a
ppl
ic
a
ti
on
in
a
tt
e
nda
nc
e
ma
na
ge
me
nt
s
ys
te
m,
”
A
r
r
ay
,
vol
. 5, 2020, doi:
10.1016/j
.a
r
r
a
y.2019.100014.
[
2]
R
.
V
.
P
e
tr
e
s
c
u,
“
F
a
c
e
r
e
c
ogni
ti
on
a
s
a
bi
ome
tr
ic
a
ppl
ic
a
ti
on,”
SSR
N
E
le
c
tr
oni
c
J
ou
r
nal
,
vol
.
3,
pp.
237
–
257,
2019,
doi
:
10.2139/s
s
r
n.3417325.
[
3]
M
.
T
a
s
ki
r
a
n,
N
.
K
a
hr
a
ma
n,
a
nd
C
.
E
.
E
r
de
m,
“
F
a
c
e
r
e
c
ogni
ti
on:
P
a
s
t,
pr
e
s
e
nt
a
nd
f
ut
ur
e
(
a
r
e
vi
e
w
)
,”
D
ig
it
al
Si
gnal
P
r
oc
e
s
s
in
g
,
vol
. 106, 2020, doi:
10.1016/j
.ds
p.2020.102809.
[
4]
S
.
K
a
r
ni
la
,
S
.
I
r
ia
nt
o,
a
nd
R
.
K
ur
ni
a
w
a
n,
“
F
a
c
e
r
e
c
ogni
ti
on
us
in
g
c
ont
e
nt
b
a
s
e
d
im
a
g
e
r
e
tr
ie
va
l
f
or
in
te
ll
ig
e
nt
s
e
c
ur
it
y,”
I
nt
e
r
nat
io
nal
J
our
nal
of
A
dv
anc
e
d E
ngi
ne
e
r
in
g R
e
s
e
a
r
c
h and S
c
ie
nc
e
, vol
. 6, no. 1, pp. 91
–
98, 2019, doi:
10.22161/i
ja
e
r
s
.6.1.1
3.
[
5]
Z
.
W
a
ng,
X
.
Z
ha
ng,
P
.
Y
u,
W
.
D
ua
n,
D
.
Z
hu,
a
nd
N
.
C
a
o,
“
A
ne
w
f
a
c
e
r
e
c
ogni
ti
o
n
me
th
od
f
or
in
te
ll
ig
e
nt
s
e
c
ur
it
y,”
A
ppl
ie
d
Sc
ie
nc
e
s
, vol
. 10, no. 3, 2020, doi
:
10.3390/a
pp10030852.
[
6]
T
.
S
a
a
r
ik
ko,
U
.
H
.
W
e
s
te
r
gr
e
n,
a
nd
T
.
B
lo
mqui
s
t,
“
D
ig
it
a
l
tr
a
ns
f
or
ma
ti
on:
F
iv
e
r
e
c
omm
e
nda
ti
ons
f
or
th
e
di
g
it
a
ll
y
c
ons
c
io
us
f
i
r
m,”
B
us
in
e
s
s
H
o
r
iz
ons
, v
ol
. 63, no. 6, pp. 825
–
839, 2020, doi:
10.1
016/
j.
bus
hor
.2020.07.005.
[
7]
A
. A
nw
a
r
a
nd A
. R
a
yc
ho
w
dhur
y, “
M
a
s
ke
d f
a
c
e
r
e
c
ogni
ti
on f
or
s
e
c
ur
e
a
ut
he
nt
ic
a
ti
on,
”
ar
X
iv
-
C
om
put
e
r
S
c
ie
nc
e
, pp. 1
–
8, 2020
.
[
8]
M
.
S
mi
th
a
nd
S
.
M
il
le
r
,
“
T
he
e
th
ic
a
l
a
ppl
ic
a
ti
on
of
bi
ome
tr
ic
f
a
c
ia
l
r
e
c
ogni
ti
on
te
c
hnol
ogy,”
A
I
&
SO
C
I
E
T
Y
,
vo
l.
37,
no.
1,
pp.
167
–
175, 2022, doi:
10.1007/s
00146
-
021
-
01199
-
9.
[
9]
V
.
W
a
ti
,
K
.
K
us
r
in
i,
H
.
A
l
F
a
tt
a
,
a
nd
N
.
K
a
poor
,
“
S
e
c
ur
it
y
of
f
a
c
ia
l
bi
ome
tr
ic
a
ut
he
nt
ic
a
ti
on
f
o
r
a
tt
e
nda
nc
e
s
ys
te
m,”
M
ul
ti
m
e
di
a
T
ool
s
and A
ppl
ic
at
io
ns
, vol
. 80, no. 15, pp. 23625
–
23646, 2021, doi:
10.1007/s
11042
-
020
-
10246
-
4.
[
10]
K
.
H
.
T
e
oh,
R
. C
.
I
s
ma
il
,
S
.
Z
.
M
.
N
a
z
ir
i,
R
.
H
us
s
in
,
M
.
N
. M
.
I
s
a
,
a
nd
M
.
B
a
s
ir
,
“
F
a
c
e
r
e
c
ogni
ti
on
a
nd
id
e
nt
if
ic
a
ti
on
us
in
g
de
e
p
le
a
r
ni
ng a
ppr
oa
c
h,”
J
our
nal
of
P
hy
s
ic
s
:
C
onf
e
r
e
nc
e
Se
r
ie
s
, vol
.
1755, no. 1, 2021, doi
:
10.1088/1742
-
6596/1755/
1/
012006.
[
11]
Y
.
K
or
tl
i,
M
.
J
r
id
i,
A
.
A
l
F
a
lo
u,
a
nd
M
.
A
tr
i,
“
F
a
c
e
r
e
c
o
gni
ti
on
s
ys
te
ms
:
a
s
ur
ve
y,”
Se
n
s
or
s
,
vol
.
20,
no.
2,
2020,
doi
:
10.3390/s
20020342.
[
12]
C
.
M
.
C
ook,
J
.
J
.
H
ow
a
r
d,
Y
.
B
.
S
ir
ot
in
,
J
.
L
.
T
ip
to
n,
a
nd
A
.
R
.
V
e
mur
y,
“
D
e
mogr
a
phi
c
e
f
f
e
c
t
s
in
f
a
c
ia
l
r
e
c
ogni
ti
on
a
nd
t
he
ir
de
pe
nde
nc
e
on i
ma
ge
a
c
qui
s
it
io
n:
a
n e
va
lu
a
ti
on of
e
le
ve
n c
om
me
r
c
ia
l
s
ys
te
ms
,”
I
E
E
E
T
r
ans
ac
ti
ons
on B
io
m
e
tr
ic
s
, B
e
hav
io
r
, and
I
de
nt
it
y
Sc
ie
nc
e
,
vol
. 1, no. 1, pp. 32
–
41, 2019, doi:
10.1109/T
B
I
O
M
.2019.2897801.
[
13]
K
.
R
a
ju
,
B
.
C
.
R
a
o
,
K
.
S
a
i
kum
a
r
,
a
n
d
N
.
L
.
P
r
a
t
a
p,
“
A
n
opt
im
a
l
h
ybr
id
s
ol
ut
io
n
to
lo
c
a
l
a
nd
gl
o
ba
l
f
a
c
i
a
l
r
e
c
ogni
t
io
n
th
r
o
ugh
m
a
c
h
in
e
le
a
r
ni
ng,
”
i
n
I
nt
e
ll
ig
e
nt
S
y
s
te
m
s
R
e
fe
r
e
nc
e
L
i
br
a
r
y
,
S
pr
i
nge
r
, C
h
a
m, 2
022,
pp.
203
–
226
, do
i
:
10
.100
7/
97
8
-
3
-
030
-
7665
3
-
5_1
1.
[
14]
L
.
C
.
N
gugi
,
M
.
A
be
lwa
h
a
b,
a
nd
M
.
A
bo
-
Z
a
hha
d,
“
R
e
c
e
nt
a
d
va
nc
e
s
in
im
a
ge
pr
oc
e
s
s
in
g
te
c
hni
que
s
f
or
a
ut
oma
te
d
le
a
f
pe
s
t
a
nd
di
s
e
a
s
e
r
e
c
ogni
ti
on
–
a
r
e
vi
e
w
,”
I
nf
or
m
at
io
n P
r
oc
e
s
s
in
g i
n
A
gr
i
c
ul
tu
r
e
, vol
. 8, no. 1, pp. 27
–
51, 2021.
[
15]
K
. A
dna
n a
nd
R
. A
kb
a
r
, “
A
n
a
na
ly
ti
c
a
l
s
tu
dy of
i
nf
or
ma
ti
on e
x
tr
a
c
ti
on f
r
om uns
tr
uc
tu
r
e
d a
nd mul
ti
di
me
ns
io
na
l
bi
g d
a
ta
,”
J
ou
r
nal
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
E
nhanc
ing
facia
l
r
e
c
ognit
ion
ac
c
ur
ac
y
thr
ough
featur
e
e
x
tr
ac
ti
ons
and
ar
ti
fi
c
ial
ne
ur
al
…
(
A
dhi
K
us
nadi
)
1065
of
B
ig
D
at
a
, vol
. 6, no. 1, 2019, doi:
10.1186/s
40537
-
019
-
0254
-
8.
[
16]
V
. P
. V
is
hw
a
ka
r
ma
a
nd T
.
G
oe
l,
“
A
n e
f
f
ic
ie
nt
hybr
id
D
W
T
-
f
u
z
z
y f
il
te
r
i
n
D
C
T
doma
in
ba
s
e
d i
ll
umi
na
ti
on no
r
ma
li
z
a
ti
on f
o
r
f
a
c
e
r
e
c
ogni
ti
on,”
M
ul
ti
m
e
di
a T
ool
s
and A
ppl
ic
at
io
ns
, vol
. 78, no. 1
1, pp. 15213
–
15233, 2019, doi:
10.10
07/
s
11042
-
018
-
6837
-
0.
[
17]
M
.
M
.
O
gha
z
,
M
.
A
. M
a
a
r
of
,
M
.
F
.
R
oh
a
ni
,
A
.
Z
a
in
a
l,
a
nd
S
. Z
.
M
.
S
ha
id
,
“
A
n
opt
im
iz
e
d s
ki
n
te
xt
ur
e
mode
l
us
in
g
gr
a
y
-
le
ve
l
c
o
-
oc
c
ur
r
e
nc
e
ma
tr
ix
,”
N
e
ur
al
C
om
put
in
g and A
ppl
ic
at
io
ns
, vol
. 31, no. 6, pp. 1835
–
1853, 2019, doi:
10.1007
/s
00521
-
017
-
3164
-
8.
[
18]
J
.
X
ie
e
t
al
.
,
“
D
S
-
U
I
:
D
ua
l
-
s
upe
r
vi
s
e
d
mi
xt
ur
e
of
ga
us
s
ia
n
mi
xt
ur
e
mode
ls
f
or
unc
e
r
ta
in
ty
in
f
e
r
e
nc
e
in
im
a
ge
r
e
c
ogni
ti
on,”
I
E
E
E
T
r
ans
ac
ti
ons
on I
m
age
P
r
oc
e
s
s
in
g
, vol
. 30, pp. 9208
–
9219, 20
21, doi:
10.1109/T
I
P
.2021.3123555.
[
19]
V
e
r
a
,
A
.
K
us
na
di
,
I
.
Z
.
P
a
ne
,
M
.
V
.
O
ve
r
be
e
k,
a
nd
S
.
G
.
P
r
a
s
e
ty
a
,
“
F
a
c
e
r
e
c
ogni
ti
on
a
c
c
ur
a
c
y
im
pr
ovi
ng
us
in
g
gr
a
y
le
ve
l
c
o
-
oc
c
ur
r
e
nc
e
ma
tr
ix
s
e
le
c
ti
on
f
e
a
tu
r
e
a
lg
or
it
hm,”
in
2023
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
Sm
ar
t
C
om
put
in
g
and
A
ppl
ic
at
io
n
(
I
C
S
C
A
)
,
2023, pp. 1
–
6
, doi
:
10.1109/I
C
S
C
A
57840.2023.10087414.
[
20]
S
.
K
ha
n,
M
.
H
.
J
a
v
e
d,
E
.
A
hme
d,
S
.
A
.
A
.
S
ha
h,
a
nd
S
.
U
.
A
li
,
“
F
a
c
ia
l
r
e
c
ogni
ti
on
us
in
g
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
a
nd
im
pl
e
me
nt
a
ti
on
on
s
ma
r
t
gl
a
s
s
e
s
,”
in
2019
I
nt
e
r
nat
io
nal
C
on
fe
r
e
nc
e
on
I
nf
or
m
at
io
n
S
c
ie
nc
e
and
C
om
m
uni
c
at
io
n
T
e
c
hnol
ogy
(
I
C
I
SC
T
)
, 2019, pp. 1
–
6
, doi
:
10.1109/C
I
S
C
T
.2019.8777442.
[
21]
S
.
H
s
ia
,
S
.
W
a
ng,
a
nd
C
.
C
he
n,
“
F
a
s
t
s
e
a
r
c
h
r
e
a
l‐
ti
me
f
a
c
e
r
e
c
ogni
ti
on
ba
s
e
d
on
D
C
T
c
oe
f
f
ic
ie
nt
s
di
s
tr
ib
ut
io
n,”
I
E
T
I
m
age
P
r
oc
e
s
s
in
g
, vol
. 14, no. 3, pp. 570
–
57
5, 2020, doi:
10.1049/i
e
t
-
ip
r
.2018.6175.
[
22]
C
.
S
c
r
ib
a
no,
G
.
F
r
a
nc
hi
ni
,
M
.
P
r
a
to
,
a
nd
M
.
B
e
r
to
gna
,
“
D
C
T
-
f
or
me
r
:
e
f
f
ic
ie
nt
s
e
lf
-
a
tt
e
nt
io
n
w
it
h
di
s
c
r
e
te
c
os
in
e
tr
a
ns
f
or
m,”
J
our
nal
of
Sc
ie
nt
if
ic
C
om
put
in
g
, vol
. 94, no. 3, 2023, doi
:
10.1007/s
10915
-
023
-
02125
-
5.
[
23]
A
.
S
in
gha
l,
P
.
S
in
gh,
B
.
L
a
ll
,
a
nd
S
.
D
.
J
o
s
hi
,
“
M
od
e
li
ng
a
nd
p
r
e
di
c
ti
on
of
C
O
V
I
D
-
19
pa
nde
mi
c
u
s
in
g
G
a
us
s
ia
n
mi
xt
ur
e
mod
e
l,
”
C
haos
, Sol
it
ons
& F
r
ac
ta
ls
, vol
. 138, 2020, doi:
10.1016/j
.c
ha
o
s
.2020.110023.
[
24]
L
.
C
he
n,
S
.
L
i,
Q
.
B
a
i,
J
.
Y
a
ng,
S
.
J
ia
n
g,
a
nd
Y
.
M
ia
o,
“
R
e
vi
e
w
of
im
a
ge
c
la
s
s
if
ic
a
ti
on
a
lg
or
it
hms
ba
s
e
d
on
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
,”
R
e
m
ot
e
Se
ns
in
g
, vol
. 13, no. 22, 2021, doi
:
10.3390/
r
s
13224712.
[
25]
S
.
P
.
J
a
ip
r
a
ka
s
h,
M
.
B
.
D
e
s
a
i,
C
.
S
.
P
r
a
ka
s
h,
V
.
H
.
M
i
s
tr
y,
a
nd
K
.
L
.
R
a
da
di
ya
,
“
L
ow
di
me
ns
io
na
l
D
C
T
a
nd
D
W
T
f
e
a
tu
r
e
ba
s
e
d
mode
l
f
or
de
te
c
ti
on
of
im
a
ge
s
pl
ic
in
g
a
nd
c
opy
-
move
f
o
r
ge
r
y,”
M
ul
ti
m
e
di
a
T
ool
s
and
A
ppl
ic
at
io
ns
,
vol
.
79,
no.
39
–
40,
pp.
299
77
–
30005, 2020, doi:
10.1007/s
11042
-
020
-
09415
-
2.
[
26]
S
.
M
is
r
a
a
nd
R
.
H
.
L
a
s
k
a
r
,
“
I
nt
e
gr
a
te
d
f
e
a
tu
r
e
s
a
nd
G
M
M
ba
s
e
d
ha
nd
de
te
c
to
r
a
pp
li
e
d
to
c
ha
r
a
c
t
e
r
r
e
c
og
ni
ti
o
n
s
y
s
te
m
u
nde
r
pr
a
c
t
ic
a
l
c
ondi
ti
on
s
,
”
M
ul
ti
m
e
di
a T
o
ol
s
an
d A
ppl
i
c
at
i
on
s
, v
ol
. 7
8, n
o. 2
4,
pp.
3492
7
–
34
961,
201
9, d
oi
:
1
0.10
07/
s
110
42
-
01
9
-
0810
5
-
y.
[
27]
T
.
P
.
L
il
li
c
r
a
p,
A
.
S
a
nt
or
o,
L
.
M
a
r
r
is
,
C
.
J
.
A
ke
r
ma
n,
a
nd
G
.
H
in
to
n,
“
B
a
c
kpr
opa
ga
ti
on
a
nd
th
e
br
a
in
,”
N
at
ur
e
R
e
v
i
e
w
s
N
e
ur
os
c
ie
nc
e
, vol
. 21, no. 6, pp. 335
–
346, 2020, doi:
10.1038/s
41583
-
020
-
0277
-
3.
[
28]
M
.
A
.
H
o
s
s
a
in
a
nd
M
.
S
.
A
.
S
a
ji
b,
“
C
la
s
s
if
ic
a
ti
on
of
im
a
ge
us
in
g
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
C
N
N
)
,”
G
lo
bal
J
our
na
l
of
C
om
put
e
r
Sc
ie
n
c
e
and T
e
c
hnol
ogy
, vol
. 19, no. 2, pp. 13
–
18, 2
019, doi:
10.34257/gj
c
s
td
vol
19i
s
2pg13.
[
29]
M
.
T
a
va
r
e
s
,
“
T
h
e
O
R
L
da
ta
ba
s
e
f
or
tr
a
in
in
g
a
nd
te
s
ti
ng,”
K
aggl
e
.
2020.
A
c
c
e
s
s
e
d:
N
ov.
04,
2024.
[
O
nl
in
e
]
.
A
va
il
a
bl
e
:
ht
tp
s
:
//
w
w
w
.ka
ggl
e
.c
om/
da
ta
s
e
ts
/t
a
va
r
e
z
/t
he
-
or
l
-
da
ta
ba
s
e
-
f
or
-
t
r
a
in
in
g
-
a
nd
-
te
s
ti
ng
[
30]
W
.
W
a
ng,
X
.
W
u,
X
.
Y
u
a
n,
a
nd
Z
.
G
a
o,
“
A
n
e
xpe
r
im
e
nt
-
ba
s
e
d
r
e
vi
e
w
of
lo
w
-
li
ght
im
a
ge
e
nha
nc
e
m
e
nt
me
th
ods
,”
I
E
E
E
A
c
c
e
s
s
,
vol
. 8, pp. 87884
–
87917, 2020, doi:
10.1109/AC
C
E
S
S
.2020.29
92749.
[
31]
W
.
A
.
M
us
ta
f
a
e
t
al
.
,
“
I
ma
ge
e
nha
nc
e
me
nt
ba
s
e
d
on
di
s
c
r
e
te
c
os
in
e
t
r
a
ns
f
or
ms
(
D
C
T
)
a
nd
di
s
c
r
e
te
w
a
ve
le
t
tr
a
ns
f
or
m
(
D
W
T
)
:
a
r
e
vi
e
w
,”
I
O
P
C
onf
e
r
e
nc
e
Se
r
ie
s
:
M
at
e
r
ia
ls
Sc
ie
nc
e
and
E
ngi
ne
e
r
in
g
,
vol
.
557,
no.
1,
2019,
do
i:
10.1088/1757
-
899X/557
/1
/0
12
027.
[
32]
M
.
D
.
D
e
e
pa
k,
P
.
K
a
r
th
ik
,
S
.
S
.
K
uma
r
,
a
nd
N
.
A
.
D
e
e
p
a
k,
“
C
ompa
r
a
ti
v
e
s
tu
dy
of
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
us
in
g
di
f
f
e
r
e
nt
tr
a
ns
f
or
m
te
c
hni
que
s
in
f
r
e
que
nc
y
doma
in
,”
in
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
A
ut
om
at
io
n,
Si
gnal
P
r
oc
e
s
s
in
g,
I
ns
t
r
um
e
nt
at
io
n
and
C
ont
r
ol
,
2021, pp. 2835
–
2846
, doi
:
10.1007/978
-
981
-
15
-
8221
-
9_265.
[
33]
W
.
M
.
A
la
lu
os
i
,
“
R
e
c
ogni
ti
on
of
huma
n
f
a
c
ia
l
e
xpr
e
s
s
io
n
s
us
in
g
D
C
T
-
D
W
T
a
nd
a
r
ti
f
ic
ia
l
ne
ur
a
l
ne
twor
k,”
I
r
aqi
J
ou
r
nal
of
S
c
ie
nc
e
,
vol
. 62, no. 6, pp. 2090
–
2098, 2021, doi:
10.24996/i
js
.2021.62.6.34.
[
34]
E
.
G
.
W
a
hyuni,
L
.
M
.
F
.
F
a
uz
a
n,
F
.
A
br
iy
a
ni
,
N
.
F
.
M
uc
hl
is
,
a
nd
M
.
U
lf
a
,
“
R
a
in
f
a
ll
pr
e
di
c
ti
on
w
it
h
ba
c
kpr
opa
ga
ti
on
me
th
o
d,”
J
our
nal
of
P
hy
s
ic
s
:
C
onf
e
r
e
nc
e
Se
r
ie
s
, vol
. 983, 2018, doi:
10.1088/1742
-
6596/983/
1/
012059.
[
35]
A
.
K
us
na
di
,
L
.
N
a
th
a
ni
a
,
I
.
Z
.
P
a
ne
,
a
nd
M
.
V
.
O
ve
r
be
e
k,
“
F
a
c
e
de
te
c
ti
on
k
e
ypoi
nt
s
us
in
g
D
c
t
A
nd
C
la
he
,
”
T
ur
k
is
h
J
our
n
al
of
C
om
put
e
r
and
M
at
he
m
at
ic
s
E
duc
at
io
n (
T
U
R
C
O
M
A
T
)
, vol
. 12, no. 11, pp. 4365
–
4372,
2021, doi:
10.17762/t
ur
c
oma
t.
v12i
11.6568.
[
36]
H
.
W
a
n,
H
.
W
a
ng,
B
.
S
c
ot
ne
y,
a
nd
J
.
L
iu
,
“
A
nove
l
ga
us
s
ia
n
mi
xt
ur
e
mode
l
f
or
c
la
s
s
if
ic
a
ti
on,”
in
2019
I
E
E
E
I
n
te
r
nat
io
nal
C
onf
e
r
e
nc
e
on Sy
s
te
m
s
, M
an and C
y
be
r
ne
ti
c
s
(
SM
C
)
, 2019, pp. 3298
–
3303
, doi
:
10.1109/S
M
C
.2019.8914215.
[
37]
L
.
J
ia
o,
T
.
D
e
nœ
ux,
Z
.
L
iu
,
a
nd
Q
. P
a
n,
“
E
G
M
M
:
A
n
e
vi
de
nt
ia
l
ve
r
s
io
n
of
th
e
ga
us
s
ia
n
mi
xt
ur
e
mode
l
f
or
c
lu
s
te
r
in
g,”
A
ppl
ie
d Soft
C
om
put
i
ng
, vol
. 129, 2022, doi:
10.1016/j
.a
s
oc
.2022.109619.
[
38]
E
.
P
a
te
l
a
nd
D
.
S
.
K
us
hw
a
ha
,
“
C
lu
s
te
r
in
g
c
lo
ud
w
or
kl
oa
ds
:
k
-
me
a
ns
vs
ga
us
s
ia
n
mi
xt
ur
e
mode
l,
”
P
r
oc
e
di
a
C
om
put
e
r
Sc
i
e
nc
e
,
vol
. 171, pp. 158
–
167, 2020, doi:
10.1016/j
.pr
oc
s
.2020.04.017.
[
39]
S
.
C
a
o,
Z
.
H
u,
X
.
L
uo,
a
nd H
.
W
a
ng,
“
R
e
s
e
a
r
c
h
on
f
a
ul
t
di
a
gn
os
is
te
c
hnol
ogy
of
c
e
nt
r
if
uga
l
pump
bl
a
de
c
r
a
c
k
ba
s
e
d
on
P
C
A
a
nd
G
M
M
,”
M
e
as
ur
e
m
e
nt
, vol
. 173, 2021, doi:
10.1016/j
.me
a
s
ur
e
me
nt
.2020.108558.
[
40]
J
.
Q
ia
o
e
t
al
.
,
“
D
a
ta
on
M
R
I
br
a
in
le
s
io
n
s
e
gme
nt
a
ti
on
us
in
g
k
-
me
a
ns
a
nd
ga
us
s
ia
n
mi
xt
ur
e
mode
l
-
e
xpe
c
ta
ti
on
ma
xi
mi
z
a
ti
on,”
D
at
a
in
B
r
ie
f
, vol
. 27, 2019, doi:
10.1016/j
.di
b.2019.104628.
[
41]
W
.
J
a
nna
h a
nd
D
.
R
.
S
.
S
a
put
r
o,
“
P
a
r
a
me
te
r
e
s
ti
ma
ti
on
of
ga
us
s
ia
n
mi
xt
ur
e
mode
ls
(
G
M
M
)
w
it
h
e
xpe
c
ta
ti
on
ma
xi
mi
z
a
ti
on
(
E
M
)
a
lg
or
it
hm,”
in
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
of
M
at
he
m
at
ic
s
and
M
at
he
m
at
ic
s
E
duc
at
io
n
(
I
-
C
M
M
E
)
2021
,
2022
,
doi
:
10.1063/5.0117119.
[
42]
I
. G
ood
f
e
ll
ow
, Y
. B
e
ngi
o, a
nd A
. C
our
vi
ll
e
,
D
e
e
p l
e
a
r
ni
ng
. C
a
mbr
id
ge
, M
a
s
s
a
c
hus
e
tt
s
:
M
I
T
P
r
e
s
s
, 2016.
[
43]
M
. K
uhn a
nd K
. J
ohn
s
o
n,
A
ppl
ie
d pr
e
di
c
ti
v
e
m
ode
li
ng
. N
e
w
Y
or
k:
S
pr
in
ge
r
, 2013
, doi
:
10.1007
/9
78
-
1
-
4614
-
6849
-
3.
[
44]
T
.
P
.
L
il
li
c
r
a
p
a
nd
A
.
S
a
nt
or
o,
“
B
a
c
kpr
opa
ga
ti
on
th
r
ough
ti
me
a
nd
th
e
br
a
in
,”
C
ur
r
e
nt
O
pi
ni
on
in
N
e
ur
obi
ol
ogy
,
vol
.
55,
pp. 82
–
89, 2019, doi:
10.1016/j
.c
onb.201
9.01.011.
[
45]
S
. A
na
m, “
R
a
in
f
a
ll
pr
e
di
c
ti
on us
in
g ba
c
kpr
opa
ga
ti
on a
lg
or
it
hm
opt
im
iz
e
d by B
r
oyde
n
-
F
le
tc
he
r
-
G
ol
d
f
a
r
b
-
S
ha
nno a
lg
or
i
th
m,”
I
O
P
C
onf
e
r
e
nc
e
Se
r
ie
s
:
M
at
e
r
ia
ls
Sc
ie
n
c
e
and E
ngi
ne
e
r
in
g
, vol
. 56
7, no. 1, 2019, doi:
10.1088/1757
-
899X/567/1/
012008.
[
46]
P
.
R
.
V
la
c
ha
s
e
t
al
.
,
“
B
a
c
kpr
opa
g
a
ti
on
a
lg
or
it
hms
a
nd
r
e
s
e
r
v
oi
r
c
omput
in
g
in
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
ks
f
or
th
e
f
or
e
c
a
s
ti
n
g
of
c
ompl
e
x s
pa
ti
ot
e
mpor
a
l
dyna
mi
c
s
,”
N
e
ur
al
N
e
tw
o
r
k
s
, vol
. 126,
pp. 191
–
217, 2020, doi:
10.1016/
j.
ne
une
t.
2020.02.016.
[
47]
L
. A
lz
uba
id
i
e
t
al
.
, “
R
e
vi
e
w
of
de
e
p l
e
a
r
ni
ng:
c
onc
e
pt
s
, C
N
N
a
r
c
hi
te
c
tu
r
e
s
, c
ha
ll
e
nge
s
, a
ppl
ic
a
ti
ons
, f
ut
ur
e
di
r
e
c
ti
ons
,”
J
our
na
l
of
B
ig
D
at
a
, vol
. 8, no. 1, 2021, doi:
10.1186/s
40537
-
021
-
00444
-
8.
[
48]
A
.
K
ha
n,
A
.
S
oha
il
,
U
.
Z
a
ho
or
a
,
a
nd
A
.
S
.
Q
ur
e
s
hi
,
“
A
s
ur
ve
y of
th
e
r
e
c
e
nt
a
r
c
hi
te
c
tu
r
e
s
of
de
e
p c
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
,”
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
R
e
v
ie
w
, vol
. 53, no. 8, pp. 5455
–
5516, 2020, doi:
10.1007/s
10462
-
020
-
09825
-
6.
[
49]
M
. K
r
ic
he
n, “
C
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
:
A
s
ur
ve
y,”
C
om
p
ut
e
r
s
, vol
. 12, no. 8, 2023, doi
:
10.3390/c
omput
e
r
s
12080151.
[
50]
Y
.
L
e
C
un
,
L
.
B
ot
to
u,
Y
.
B
e
ngi
o,
a
nd
P
.
H
a
f
f
ne
r
,
“
G
r
a
di
e
nt
-
ba
s
e
d
le
a
r
ni
ng
a
ppl
ie
d
to
do
c
ume
nt
r
e
c
ogni
ti
on,”
P
r
oc
e
e
di
ngs
of
th
e
I
E
E
E
, vol
. 86, no. 11, pp. 2278
–
2324, 1998, doi:
10.
1109/5.726
791.
[
51]
T
.
S
a
e
e
d,
M
.
S
uf
ia
n,
M
.
A
li
,
a
nd
A
.
U
.
R
e
hma
n,
“
C
onvolut
io
n
a
l
ne
ur
a
l
ne
twor
k
ba
s
e
d
c
a
r
e
e
r
r
e
c
omm
e
nde
r
s
y
s
te
m
f
or
P
a
ki
s
t
a
ni
e
ngi
ne
e
r
in
g
s
tu
de
nt
s
,”
in
2021
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
I
nnov
at
iv
e
C
om
put
in
g
(
I
C
I
C
)
,
2021,
pp.
1
–
10
,
doi
:
Evaluation Warning : The document was created with Spire.PDF for Python.