Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
10
,
No.
2
,
A
pr
il
2020, p
p. 18
33
~
1841
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
2
.
pp1833
-
18
41
1833
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Optimi
zation
of
deep l
earning f
eatur
es for
age
-
inv
ar
i
ant
face r
ecogniti
on
Ama
l
A
.
Mou
stafa,
Ah
me
d
El
na
kib,
N
ih
al
F.
F.
A
ree
d
El
e
ct
roni
cs
and
Com
m
unic
at
ions E
ngineeri
ng
Depa
rtment
,
Fa
cult
y
of
Engi
n
ee
rin
g,
Mansoura
Unive
rsit
y
,
Mansour
a,
Eg
y
p
t
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Oct
1
, 2
019
Re
vised
Oct
18
,
20
19
Accepte
d
Oct
3
1
, 20
19
Thi
s
pape
r
pre
s
ent
s
a
m
et
hodo
log
y
for
Age
-
I
nvar
ia
n
t
Face
Rec
ognition
(AIF
R),
base
d
o
n
the
op
ti
m
ization
of
de
ep
l
ea
r
ning
fe
at
ure
s
.
T
he
propose
d
m
et
hod
ext
ra
ct
s
dee
p
learni
ng
fe
at
ure
s
using
tra
n
sfer
dee
p
learni
n
g,
ext
r
ac
t
e
d
from
the
unpro
ce
ss
ed
fa
ce
images.
To
opt
imize
the
ext
r
ac
t
e
d
fe
a
ture
s,
a
Gene
tic
Algori
thm
(GA
)
proc
edur
e
is
designe
d
in
orde
r
to
select
the
m
ost
re
le
v
ant
f
ea
tu
re
s
to
the
prob
lem
of
ide
nti
f
y
i
n
g
a
per
son
b
ase
d
on
his
/he
r
fa
cial
images
o
ver
diff
ere
nt
ag
e
s
.
For
class
ifi
c
at
ion
,
K
-
Nea
r
est
Neighbor
(KN
N)
cl
assifiers
with
diff
erent
dista
n
ce
m
etrics
ar
e
inve
sti
gat
ed
,
i
.
e
.
,
Corre
lat
ion,
E
ucl
idian,
Cosi
ne,
and
Man
hat
t
an
dist
ance
m
et
ri
cs.
Expe
riment
al
r
e
sults
using
a
Manha
ttan
distance
KN
N
cl
assi
fie
r
a
chiev
e
the
best
R
ank
-
1
re
cognition
r
at
e
s
of
86.
2%
and
9
6%
on
the
st
and
ard
FG
NET
and
MO
RP
H
dat
ase
ts,
r
espe
c
ti
vely
.
Com
par
ed
to
the
sta
t
e
-
of
-
the
-
art
m
et
hods,
our
pr
oposed
m
et
hod
nee
ds
no
pre
proc
essing
stage
s.
I
n
addi
ti
on
,
the
expe
rimen
ts
show
it
s pri
vi
le
g
e
over
oth
er
r
el
a
t
ed
m
et
hods
.
Ke
yw
or
d
s
:
AI
FR
Deep t
ra
ns
fe
r
l
earn
i
ng
Gen
et
ic
Algori
thm
Copyright
©
202
0
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Ah
m
ed
Elna
ki
b,
Ele
ct
ro
nics
and C
omm
un
ic
ation
s
Enginee
ri
ng D
e
pa
rtm
ent
,
Faculty
of E
ngineerin
g
,
Ma
nsoura
U
niv
er
sit
y,
El
-
G
om
ho
ria S
T, Man
soura
3551
6,
Da
kah
li
a
, E
gypt.
Em
a
il
:
nak
ib@
m
ans.
ed
u.
e
g
1.
INTROD
U
CTION
Id
e
ntific
at
ion
of
a
pe
rson
ba
s
ed
on
his
face
i
m
ages
ov
er
diff
e
ren
t
ages
is
m
and
at
ory
for
secur
it
y
and
forensi
c
ap
plica
ti
on
s
,
e
.g.,
i
de
ntific
at
ion
of
crim
inals
and
m
issi
ng
pe
rs
ons.
H
ow
e
ver,
t
his
pro
blem
is
ver
y
chall
eng
i
ng
due
to
the
rap
i
d
cha
ng
e
in
f
ace
i
m
ages
with
age,
espec
ia
ll
y
wh
en
im
ages
from
diff
ere
nt
age
-
sta
ges
are
con
si
der
e
d,
e
.g
.
,
ne
wbo
rn,
toddler,
te
ena
ge
,
and
a
dult
face
i
m
ages.
In
add
it
io
n,
the
facial
changes
a
re
ve
ry
sp
eci
fic
f
or
eac
h
per
s
on
ba
sed
o
n
his
ge
nes,
li
fest
yl
e.
Fo
r
t
hese
reasons
,
de
ve
lop
in
g
a reli
able, acc
ura
te
m
et
ho
d for AIFR is
n
ece
ssary
In
t
he
li
te
rature,
m
any
resear
ch
gr
oups
ha
ve
pr
ese
nte
d
di
fferent
m
et
ho
ds
to
ide
ntify
pers
on
base
d
on
facial
age
i
m
ages
.
T
hese
m
et
ho
ds
ca
n
be
cat
eg
or
iz
e
d
into
th
ree
cat
egories:
(i)
gen
e
rati
ve
m
et
ho
ds,
(ii)
discrim
inati
ve
m
et
ho
ds
,
a
nd
(iii
)
dee
p
l
earn
i
ng
m
et
hod
s.
Ge
ner
at
ive
m
et
ho
ds
c
ons
truct
a
pe
rs
on
a
l
face
m
od
el
based
on
the
c
ollec
te
d
facial
-
age
im
a
ges.
For
exam
ple,
s
hap
e
a
nd intensit
y
feature
s
are u
sed
by Lan
it
is
et
al
.
[
1
]
to
bu
i
ld
a
3D
ag
e
m
od
el
,
wh
ic
h
ac
hieves
68.5
%
r
ecognit
ion
acc
ur
acy
on
a
pri
vate
agin
g
dat
a.
Po
se
correct
ion
is
a
dd
e
d,
by
Par
k
at
.
al
[
2
]
,
to
a
3D
s
ha
pe
-
a
nd
te
xtu
re
-
ba
sed
f
ace
m
od
el
to
achieve
a
n
accu
racy
of
37.4%
an
d
79.
8%
on
t
he
tw
o
popula
r
on
li
ne
sta
nd
a
rd
dat
abases
for
A
IFR
,
i.e.,
F
GNE
T
[
3
]
a
nd
M
O
RPH
-
II
[
4
]
database
s,
res
pecti
vely
.
The
m
ai
n
dra
wb
ac
k
of
t
hes
e
m
et
ho
ds
is
the
need
f
or
unreali
sti
c
pa
r
a
m
et
ri
c
assum
ption
s
[
5
].
On
the
oth
e
r
ha
nd,
discrim
in
at
ive
m
et
ho
ds
do
not
r
el
y
on
face
m
od
el
ing.
They
e
xtract
direct
facial
featur
e
s
for
cl
assifi
cat
ion
,
s
uc
h
as
G
ra
dient
O
rientat
ion
P
yram
id
(
GO
P
),
Scal
e
-
I
nvaria
nt
F
eat
ur
e
T
ra
ns
f
or
m
(S
I
FT)
[
6
]
,
a
nd
Mult
i
-
Scal
e
Local
Bi
nar
y
Patt
ern
(MLB
P)
[
7,
8]
.
For
exam
ple,
Ling
et
al
.
[
9
]
us
e
d
G
OP
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
2
,
A
pr
i
l 202
0
:
1833
-
1841
1834
featur
e
s
to
des
cribe
faces
.
Th
en,
the
e
xtract
ed
featu
res
we
re
input
to
a
S
upport
Vect
or
Ma
chine
(SV
M)
for
per
s
onal
ide
ntific
at
ion
.
Li
et
al
.
[1
0
]
us
e
d
S
IF
T
a
nd
MLB
P
feat
ur
es
as
l
ocal
de
scri
pto
r
s
f
or
face
rec
ogniti
on
achievin
g
rec
ogniti
on
rates
of
47.5
%
an
d
83.
9%
for
FGN
ET
an
d
M
ORP
H
-
II
dataset
s,
r
especti
vely
.
G
ong
et
al
[
11
]
a
pp
li
ed
a
Ma
xim
u
m
Entr
op
y
Feat
ure
Descr
i
pto
r
(
MEFD)
that
e
ncode
d
th
e
m
i
cro
st
ru
ct
ur
e
of
facial
i
m
ages
into
a
s
et
of
discrete
c
od
e
s
in
te
rm
s
of
m
axi
m
u
m
entropy.
F
or
m
atch
in
g
face
s,
t
he
y
fu
rther
ap
pl
ie
d
an
Id
e
ntit
y
Fact
or
A
naly
sis
(IF
A)
t
o
ac
hieve
rec
ogniti
on
r
at
es
of
76.
2%
for
FGNET
dataset
an
d
92
.3
%
f
or
MORPH
-
I
I
da
ta
set
.
Li
et
a
l.
[
12
]
a
pp
li
e
d
a
M
od
i
fied
Hidde
n
Fact
or
A
naly
sis
(
MHFA)
t
o
a
set
of
discrim
inati
ve
featur
e
s,
ac
hie
ving
reco
gnit
i
on
rates
of
72.
8%
an
d
87.
9%
on
F
GNET
a
nd
M
ORPH
-
II
da
ta
set
s,
resp
ect
ively
.
A
lt
ho
ug
h
these
m
et
ho
ds
ac
hie
ved
s
uccess
i
n
face
rec
ogniti
on,
thei
r
ge
ner
al
iz
at
ion
to
rec
ognize
age
-
i
nv
a
riant
f
aces i
s sti
ll
u
nd
er investi
gatio
n [
5
].
To
in
vestigat
e
m
or
e
gen
eral
f
eat
u
res
f
or
A
I
FR
,
the
thir
d
c
at
egory
of
fac
e
recogn
it
io
n
m
et
ho
ds
us
e
deep
le
ar
ning.
These
m
et
ho
ds
us
ually
us
e
diff
e
re
nt
Conv
olu
ti
onal
Ne
ural
Network
(CNN)
arc
hitec
tu
res
to
extract
com
pact
face
featur
e
s.
CNN
a
rc
hitec
tures,
c
om
po
se
d
of
co
nvol
ution
al
la
ye
rs,
pool
in
g
la
ye
rs,
a
nd
fu
ll
y
connecte
d
la
ye
rs,
ha
d
achieve
d
a
rem
a
rk
a
ble
su
ccess
in
AI
FR
.
Y.
Li
et
al
.
[
13
]
us
e
d
a
CNN
arc
hitec
ture
com
po
sed
of
s
even
la
ye
rs
tha
t
can
le
ar
n
fac
e
featu
res,
dist
ance
betwee
n
f
aces,
a
nd
discr
i
m
inati
ng
thres
ho
l
ds
si
m
ultaneou
sly
to
achi
ev
e
an
accuracy
of
93.6
%
on
MORP
H
-
II
dataset
.
P
arkhi
et
al
.
[
14
]
,
a
group
in
O
xfo
r
d
Un
i
ver
sit
y,
na
m
el
y,
Visu
al
Geo
m
et
ry
Gro
up
(
VGG
),
ha
ve
bu
il
t
a
de
ep
le
ar
ning
a
r
chite
ct
ur
e
f
or
face
recog
niti
on
na
m
ed
as
VG
G
-
Face
CNN
m
od
el
.
The
m
od
el
has
been
trai
ne
d
us
in
g
m
or
e
than
2.6
m
illi
on
face
i
m
ages
achiev
ing
hi
gh
face
recog
niti
on
rat
e.
Lat
er
on,
th
is
m
od
el
has
been
rep
eat
e
dly
inv
oked
by
oth
e
r
researc
h
te
am
s,
us
i
ng
tra
ns
f
er
le
ar
ning,
for
age
est
im
a
tio
n
[
1
5
,
1
6
]
a
nd
f
or
A
IF
R
[1
7
].
ElK
hiya
r
i
an
d
Wechsl
er
[
1
7
]
us
e
d
this
m
odel
with
an
E
nsem
ble
of
S
ubs
pace
D
isc
rim
i
nan
t
(E
SD)
c
la
ssifie
rs
t
o
achi
eve
an
accuracy
of
80.
6%
for
F
G
NET
dataset
a
nd
92.
2%
f
or
MORP
H
-
II
da
ta
set
.
In
this
pa
per,
we
e
m
plo
ye
d
the VG
G
-
face
CNN
m
od
el
,
wi
th two m
ajo
r
c
on
t
rib
ution
s:
The
pro
posed
m
e
tho
d
i
nvolve
s
no
pr
e
proce
ssing
ste
ps,
i.e
.,
im
ages
are
only
resized
to
t
he
sta
nd
a
rd
siz
e
of
VGG
-
m
od
el
inp
ut
without
any
f
or
m
of
r
egistrat
ion.
Re
gistrati
on
ste
ps,
usual
ly
us
ed
in
the
li
te
rature
[
13
,
1
7
-
1
9
]
,
m
ay
no
t
only
in
vo
l
ve
al
ig
nm
e
nt
er
rors
[
20
,
2
1
]
,
bu
t
al
s
o
m
ake
the
m
et
ho
d
le
ss
r
obus
t,
ti
m
e
consum
ing
, a
nd se
ns
it
ive to
t
he
s
uccess
of t
he registrat
io
n st
ep.
The
e
xtracted
f
eat
ur
es
us
in
g
t
he
tra
ns
fe
r
le
ar
ning VG
G
-
m
od
el
are fu
rthe
r op
ti
m
iz
ed
us
in
g
a s
ophisti
cat
ed
GA
in
or
der
to
sel
ect
the
m
os
t
releva
nt
fe
at
ur
es
to
A
IFR
.
In
ad
diti
on,
GA
pe
rfor
m
s
featur
e
sel
ect
ion
,
wh
ic
h
si
gn
ific
a
ntly
r
ed
uces t
he
f
eat
ure s
pace
and im
pr
oves
the r
ec
ogniti
on
sp
ee
d.
This
pa
pe
r
is
orga
nized
as
fo
l
lows
:
Sect
io
n
2
il
lustrate
s
th
e
detai
ls
of
the
propose
d
m
eth
od
ology
for
face
recog
ni
ti
on
.
Sect
io
n
3
outl
ines
the
re
s
u
lt
s
with
detai
le
d
discuss
i
ons,
a
nd
fi
nally
,
sect
ion
4
co
nc
lud
es
the
pa
pe
r
a
nd the
fu
t
ur
e
w
ork.
2.
METHO
DS
The
pro
pos
e
d
m
et
ho
d
f
or
AIFR
con
sist
s
of
two
m
ai
n
ste
p
s,
as
show
n
in
Figure
1.
Fi
rs
t,
the
VGG
-
face
m
od
el
is
use
d
t
o
e
xtract
c
om
pact
face
fe
at
ur
es
from
of
the
in
pu
t
ra
w
i
m
ages.
T
he
act
ivati
on
s
of
the
fu
ll
y
connecte
d
la
ye
r
FC6
,
of
le
ngth
=409
6,
t
he
first
f
ully
connecte
d
la
ye
r,
i
s
us
e
d
as
t
he
fa
ce
feat
ur
e
s
for
the
pro
po
se
d
s
yst
e
m
.
Secon
d,
featur
e
sel
ect
ion
is
ap
plied
by
the
ai
d
of
a
n
offli
ne
sm
art
Gen
et
ic
Algor
it
h
m
(GA)
op
ti
m
iz
a
ti
on
.
Usi
ng
a
trai
ning
datab
ase,
the
GA
i
te
rati
vely
evo
lves
to
the
bes
t
chr
om
os
om
e
that
represe
nts
the
m
os
t
rev
ena
nt
FC6
f
eat
ur
e
s
of
le
ngth
<
fo
r
t
he
prob
le
m
of
AI
FR
.
Fin
al
ly
,
a
cl
assifi
cat
ion
ste
p
is
a
pp
li
e
d
to
pe
rfo
rm
the
face
ide
ntific
at
ion
.
The
det
ai
ls
of
eac
h
of
t
hese
ste
ps
a
re
il
lustra
te
d
in
this sect
ion.
2.1. Fe
at
ure
e
xt
r
act
i
on
The
pro
po
se
d
m
et
ho
d
a
ppli
es
trans
fer
le
ar
ni
ng
[22,
23]
us
ing
t
he
VGG
-
Face
m
od
el
[14]
to
e
xtract
c
o
m
pact
face
f
eat
ur
es
for
fac
e
recog
niti
on
.
The
m
od
el
is
com
po
sed
of
12
conv
olu
ti
onal
la
ye
rs
and
t
hr
e
e
fu
ll
y
connecte
d
la
ye
rs
trai
ne
d
us
i
ng
2.6
m
illi
on
f
ace
i
m
ages
colle
ct
ed
from
2,
622
pe
rs
ons
[14].
Since
the
ut
il
iz
e
d
age
in
var
ia
nt
f
ace
databases
,
i.e.,
F
GNET
and
M
ORP
H
-
II
,
a
re
of
m
uch
s
m
al
le
r
siz
es,
trans
fer
le
ar
ni
ng
is
pr
e
ferred
to
avo
i
d
over
-
fitt
ing
[22,
23
]
.
Applyi
ng
tra
nsfer
le
ar
ning,
the
trai
ned
conv
olu
ti
onal
la
ye
rs
’
par
am
et
ers
are
tran
sfe
rr
e
d
t
o
the
pro
posed
m
od
el
,
w
he
re
as
the
f
ully
connecte
d
la
ye
r
s
are
trai
ne
d
us
in
g
the
util
iz
ed
ag
e
inv
aria
nt
fac
e
databases
,
i.e.,
FGNET
a
nd
MORP
H
dat
aset
s.
In
a
dd
it
ion,
trans
fer
le
arn
i
ng
involves
re
pla
ci
ng
the
cl
ass
ific
at
ion
la
ye
r
(t
he
la
st
fu
ll
y
con
necte
d
la
ye
r
)
with
th
e
pr
o
pose
d
syst
e
m
cl
assifi
er
[
24]
;
in
this
m
anuscript,
a
K
N
N
cl
assifi
er
is
us
e
d.
In
the
pro
posed
syst
em
,
the
act
ivatio
ns
of
the
fir
st
f
ully
co
nnect
ed
lay
er,
FC6
,
of le
ng
t
h
=4
096
a
re
us
e
d
as
the c
om
pact face fe
at
ur
e
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Op
ti
mize
d
Dee
p
Le
arnin
g
A
ge
-
I
nvari
ant
Fa
ce Reco
gnit
ion (
A. M
ou
st
afa
)
1835
Figure
1. Dia
gra
m
o
f
t
he pr
opos
e
d
m
et
h
od
f
or AIFR.
I
nput
r
a
w
im
age is
proces
sed
b
y
t
he
V
GG
-
Face
m
od
el
.
The
act
ivati
ons
of the
full
y co
nn
ect
e
d
la
ye
r,
FC6,
of len
gth
=409
6,
is
u
se
d as a feat
ur
e
d
es
cripto
r.
Feat
ur
e
sel
ect
ion
is a
ppli
ed of
fline
usi
ng
a
G
A, w
hic
h
sel
ect
s the
m
os
t rele
van
t
fea
tures f
or
A
IF
R
of len
gth
<
.
Finall
y, cl
assif
ic
at
ion
is
pe
r
form
ed
for
face
i
den
ti
ficat
io
n
2.2.
Gene
tic
alg
orit
hm
To
s
el
ect
the
m
os
t
relevan
t
featur
e
s
to
the
pro
blem
of
A
IF
R,
a
sm
art
GA
is
trai
ned
offli
ne,
us
i
ng
the
act
ivati
on
s
of
the
f
ully
connecte
d
la
y
er
FC6
of
siz
e
=409
6.
T
he
propose
d
GA
blo
c
k
dia
gr
a
m
is
il
lustrate
d
i
n
F
igure
2.
In
it
ia
ll
y,
a
ra
ndom
po
pu
la
ti
on
of
siz
e
is
gen
e
rated
from
lo
gical
ve
ct
or
s
of
le
ngt
h
=409
6,
w
her
e
is
the
t
otal
nu
m
ber
of
im
age
s
(total
data
si
ze)
.
A
l
og
ic
one
i
nd
ic
at
es
t
ha
t
the
featur
e
is
sel
ect
ed,
w
herea
s
a
l
og
ic
ze
ro
in
dicat
es
th
at
the
featu
re
is
not
sel
ect
ed
.
T
he
c
os
t
f
unct
ion
(
)
of
ea
c
h
chrom
os
om
e i
n
the
po
pu
la
ti
on is esti
m
at
ed
as the i
nv
e
rse of
the
recog
niti
on
rate
(
)
as
fo
l
lows
:
Figure
2. Dia
gra
m
o
f
t
he pr
opos
e
d
sm
art GA optim
iz
at
ion
pe
rfor
m
ed
offli
ne
at the trai
ning
ph
a
se
of
the
pr
opos
e
d
m
et
ho
d. T
he
in
it
ia
l po
pula
ti
on is ra
ndom
ly
g
ener
at
e
d
f
r
om
l
og
ic
vecto
rs o
f
leng
t
h
=409
6,
wh
e
re a
lo
gic one in
dicat
es th
at
the f
eat
ur
e
is sele
ct
ed
a
nd
a
log
ic
ze
ro in
dicat
es that t
he f
eat
ur
e is
not
the
b
est
chrom
os
om
e o
f
the
selec
te
d m
os
t rele
van
t f
eat
ur
es
t
o AIF
R of le
ngth
M
<
N
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
2
,
A
pr
i
l 202
0
:
1833
-
1841
1836
The
GA
r
uns
thr
ough
a
num
ber
of
gen
e
rati
on
s
N
G
to
m
ini
m
iz
e
th
e
cost
functi
on,
i.e.,
m
axi
m
iz
e
the
thr
ough
perform
ing
G
A
ope
rati
ons
(select
ion
,
cr
oss
over
a
nd
m
uta
ti
on
).
Th
e
chrom
os
ome
of
the
m
ini
m
u
m
(m
axi
m
u
m
)
represe
nts
the
sel
ect
ed
feat
ures
of
siz
e
<
as
s
how
n
i
n
Figure
2.
The
pro
po
se
d
GA
us
es
t
he
f
ollow
i
ng
set
ti
ng
:
t
ourn
am
ent
sel
ect
ion
[
25]
,
with
a
to
urnam
ent
siz
e
of
t
w
o,
blen
ding
cr
os
s
ov
e
r
[
26]
,
an
d
adap
ti
ve
feasi
bl
e
m
utati
on
[27].
The
best
po
pu
la
ti
on
siz
e
a
nd
best
num
ber
of
gen
e
rati
ons
for
the
prop
os
e
d GA are
in
vestigat
ed
t
hro
ugho
ut
ex
pe
rim
enta
ti
on
s.
2.3.
Cl
as
sific
ati
on
The
pro
po
s
ed
m
e
tho
d
in
ve
sti
gates
us
i
ng
diff
e
re
nt
di
sta
nce
m
et
ric
s
(E
uclidia
n,
Ma
nh
at
ta
n,
Correl
at
ion
,
a
nd
Cosine
)
f
or
the
K
-
Nea
rest
Neig
hbor
(
K
N
N)
[
2
8
]
cl
assifi
er
to
perform
f
ace
recogn
it
io
n
us
in
g
the
sel
ect
ed
G
A
feat
ur
es
of
l
eng
t
h
<
.
T
he
Ra
nk
-
1
K
NN
c
la
ssifie
r
is
sel
e
ct
ed
in
th
e
pr
opose
d
syst
em
fo
r
three
reas
on
s:
(
i
)
it
s si
m
plici
ty and effic
ie
nt ti
m
e p
erfo
rm
ance and
c
om
pu
ta
ti
on
s
, (
ii
)
t
o
in
ve
sti
gate the
qual
ity
of
t
he
sel
ect
ed
G
A
feat
ur
es
by
app
ly
in
g
a
si
m
ple
cl
assifi
er
,
an
d
(
iii
)
K
N
N
cl
assifi
er
do
es
not
nee
d
t
ra
ining;
since
it
perform
s
reco
gnit
ion
by
sim
ply
sel
ect
ing
the
database
im
age
of
feature
vecto
r
that
ha
s
t
he
m
ini
m
u
m
distance
to
the
te
st
im
age
fea
ture
vect
or
.
L
et
be
le
ngth
of
t
he
G
A
featu
re
vecto
r.
Let
be
the
i
nd
e
x
of
a
featur
e
point
in
a
vecto
r
;
*
+
.
T
he
f
ollow
i
ng
di
sta
nce
m
et
rics
are
i
nv
est
ig
at
ed
f
or
the pr
opos
e
d
s
yst
e
m
:
Eucli
dian
d
ist
a
nce
:
(
)
√
∑
(
)
(1)
Ma
nh
at
ta
n (Ci
ty
blo
ck
) dist
anc
e
:
(
)
∑
|
|
(2)
Correl
at
ion
dis
ta
nce
:
(
)
∑
(
̅
)
(
̅
)
√
∑
(
̅
)
∑
(
̅
)
(3)
Cosine
distanc
e
:
(
)
∑
(
)
√
∑
(
)
∑
(
)
(4)
3.
E
X
PERI
MEN
TAL RES
UL
TS A
ND DIS
CUSSIO
NS
This
sect
io
n
de
scribes
i
n
detai
ls
the
util
iz
ed
a
ging
data
ba
ses,
e
xp
e
rim
en
ta
l
set
up
,
e
xpe
rim
ental
resu
lt
s,
as w
el
l as t
he
c
om
par
at
ive r
es
ults to
oth
e
r
sta
te
-
of
-
the
-
art m
et
hods
.
3.1.
D
atase
ts
Tw
o
well
-
known
a
ging
databases
a
re
us
e
d
to
te
st
the
pe
rfor
m
ance
of
t
he
pr
opos
e
d
m
et
ho
d,
nam
ely,
FGNET
[
3
]
a
nd
MORP
H
-
I
I
(av
ai
la
ble
onli
ne
f
or
academ
ic
researc
h)
[
4
]
.
Table
1
des
cribes
t
he
deta
il
s
of
these
data.
As
sho
wn
in
the
Table,
MOR
PH
c
onta
ins
m
or
e
than
55
thousa
nd
im
ag
es
f
ro
m
13
th
ou
s
a
nd
su
bject
s,
ra
nging
f
ro
m
16
t
o
77
ye
ars
old.
On
the
oth
e
r
ha
nd,
F
G
NET
c
on
ta
in
s
10
02
im
ages
from
82
perso
ns,
rangin
g
f
ro
m
0
to
69
ye
ars
old
.
Sam
ples
fo
r
a
s
pan
of
one
sub
j
ect
agin
g
im
ages
for
each
dataset
are
exem
plifie
d
in
Figure
3
a
nd
Fig
ure
4.
T
he
bi
om
et
ric
recog
niti
on
of
faces
in
t
hes
e
database
s
is
ver
y
chall
eng
i
ng
du
e
to
the
li
m
i
ted
nu
m
ber
of
i
m
ages
pe
r
s
ubj
ect
acr
os
s
t
he
dif
fer
e
nt
age
s
,
m
aking
the
t
r
ai
nin
g
process
tr
ai
ning
ha
r
der.
I
n
a
dd
it
io
n,
t
he
im
ages
are
of
diff
e
re
nt
poses
,
blurri
ng,
il
lu
m
ina
ti
on
,
a
nd
ca
m
era
distances
as
show
n
in
Fig
ur
e
3
an
d
Fig
ur
e
4
.
More
over
,
at
lowe
r
age
s,
i.
e
.,
bet
wee
n
0
and
2
as
in
F
GN
E
T
,
the
pe
rs
on
al
f
ace
at
tribu
te
s
are
sti
ll
dev
el
op
i
ng
an
d
it
is
dif
ficult
to
re
la
te
the
i
m
age
to
the
c
orres
pondin
g
su
bject
,
ev
e
n v
isuall
y,
see
the
f
irst
row
i
n
Fi
gure
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Op
ti
mize
d
Dee
p
Le
arnin
g
A
ge
-
I
nvari
ant
Fa
ce Reco
gnit
ion (
A. M
ou
st
afa
)
1837
Table
1.
FGNE
T and M
ORPH
(alb
um
-
II
)
d
at
abase
desc
ription
Data
#
Su
b
jects
#
I
m
ag
es
#
I
m
ag
es/Su
b
j.
Ag
e Ran
g
e
Ag
e Gap
FGNE
T
82
1002
5
-
13
0
-
69
0
-
45
MORPH
1
3
,61
7
5
5
,13
4
2
-
5
16
-
77
0
-
5
Figure
3.
A
sa
m
ple o
f
a
n i
m
a
ge
s
pa
n of
one
su
bject
over a
ge
for
FGNET
dat
abase.
T
he
subj
ect
a
ge
is
pro
vid
e
d
belo
w
eac
h
im
age
Figure
4
.
A
sa
m
ple o
f
a
n
im
a
ge
s
pa
n
of
one
su
bject
over a
ge
for
M
ORP
H data
base
.
The
s
ubj
ect
ag
e is
pro
vid
e
d b
el
ow
eac
h
im
age
3.2. Ex
peri
ment
al setup
The
ra
w
i
m
ages,
as
sh
ow
n
in
Fig
ur
e
3
a
nd
Fig
ur
e
4
,
ar
e
inp
ut
to
the
pro
po
se
d
m
et
h
od
without
an
y
pr
e
processi
ng
ste
ps
.
T
he
pro
po
s
ed
syst
em
process
es
the
r
aw
i
m
ages
us
ing
the
VGG
-
F
ace
m
od
el
fo
r
featur
e
s
extracti
on,
o
ff
l
ine
G
A
f
or
fea
ture
sel
ect
io
n,
and
K
N
N
f
or
c
la
ssific
at
ion
as
show
n
Fig
ur
e
1
.
T
o
sel
ect
th
e
best
par
am
et
ers
of
the
pro
p
os
e
d
GA,
i.e.
,
di
ff
e
r
ent
val
ues
for
the
num
ber
of
ge
ner
at
io
n
s
and
the
popula
ti
on
siz
e
are
inv
e
sti
gated,
i.e
.,
=6
0,
80,
a
nd
100
and
=30
a
nd
40
,
res
pecti
vel
y
.
In
orde
r
to
t
rain
the
G
A
,
data
is
div
i
ded
into
trai
n
an
d
te
st
dataset
s
usi
ng
Lea
ve
-
One
-
Per
son
-
O
ut
(
LOPO)
sc
hem
e.
F
or
a
datab
ase
of
siz
e
,
L
OPO
r
un
s
a
num
ber
of
experim
ent
s,
on
eac
h
us
e
s
i
m
ages
for
trai
ning
a
nd
on
ly
on
e
im
ag
e
for
te
sti
ng
.
T
he
LOP
O
sc
he
m
e,
us
ed
f
or
A
IF
R
thr
oughou
t
literat
ur
e
(e.
g.,
in
[1
0
,
1
9
,
2
9
]
),
is
us
ually
us
e
d
in
f
oren
sic
ap
plica
ti
on
s
a
nd
m
issed
per
s
on
ide
ntific
at
ion
,
w
he
re
one
im
age
of
t
he
pe
rs
on
at
a
certai
n
a
ge
is
avail
able
an
d
it
is require
d
t
o
r
et
rieve
the c
orr
ect
p
ers
on ide
nt
it
y fr
om
the ag
in
g database
.
K
N
N
cl
assifi
e
r
[
28]
is use
d for
m
atch
in
g wit
h
.
3.3. Res
ults
of VGG
-
f
ace m
od
el
The
act
ivat
ions
of
the
f
ully
connecte
d
la
ye
r
,
FC6,
of
the
VGG
-
Face
m
od
el
are
us
e
d
as
the
featur
es
for
rec
ogniti
on,
ac
hievi
ng
t
he
be
st
recog
niti
on
rates
of
80.
7%
a
nd
92.8%
on
F
G
NET
a
nd
M
O
R
PH
databases
,
re
spe
ct
ively
,
as
show
n
i
n
Ta
ble
2
.
The
vis
ual
re
su
lt
s,
re
pres
ent
ed
in
Fig
ur
e
5,
an
d
the
quanti
ta
ti
ve
resu
lt
s,
r
e
porte
d
in
Ta
ble 2, in
dicat
e that FC
6 ca
n re
pr
ese
nt
a good i
niti
al
f
eat
ur
e
vecto
r f
or AIFR.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
2
,
A
pr
i
l 202
0
:
1833
-
1841
1838
Table
2
.
C
om
par
iso
n betwee
n usi
ng all
VGG
-
m
od
el
’s
FC
6 f
eat
ur
es
(
=409
6)
a
nd the
optim
iz
ed
G
A
featu
r
e
vecto
r of
le
ng
t
h
in term
s
of the Ra
nk
=
1
f
or FGNET
and
MORPH
-
I
I data
set
s w
it
h
dif
fe
ren
t
K
NN
distance m
et
rics (i.e.
, Ma
nh
at
t
an,
Eucli
dea
n,
Correl
at
ion
,
a
nd C
os
ine
)
us
in
g pop
ulati
on
siz
e of
=30
a
nd
nu
m
ber
of g
e
ne
rati
on
s
of
=6
0
FGNE
T
MORPH
VGG,
FC6
GA
VGG,
FC6
GA
Distan
ce
N
RR%
M
RR%
N
RR
M
RR%
Euclid
ean
4096
8
0
.7
2022
8
5
.7
4096
9
2
.8
2010
9
5
.7
Co
rr
elatio
n
4096
8
0
.7
2022
8
4
.0
4096
9
2
.8
2085
9
4
.0
Co
sin
e
4096
8
0
.7
2022
8
5
.5
4096
9
2
.8
2049
9
5
.4
Manh
attan
4096
8
0
.7
2021
8
5
.8
4096
9
2
.8
2005
9
6
.0
(a)
(b)
Figure
5
.
V
is
ua
l sy
stem
r
eco
gn
it
io
n for
(a)
t
hr
ee
sam
ples f
r
om
FG
NET
and
(b)
th
ree sam
ples fr
om
MORPH
(alb
um
-
II
) data
set
s u
si
ng LOP
O
sc
hem
e and
KNN
cl
assi
fier
w
it
h
Ma
nhat
ta
n dist
ance.
F
or
each
dataset
,
the f
ir
st col
umn is
t
he
te
st i
m
age.
C
olu
m
ns
2
a
nd 3 re
pr
e
se
nt v
is
ual r
e
su
lt
s of
us
in
g
i
nd
i
vidual FC
6
la
ye
r
a
nd
of u
si
ng
FC
6
l
ay
er af
te
r
b
ei
ng
op
ti
m
iz
ed
by
the
GA (
with
=30 an
d
=6
0)
,
resp
ect
ively
.
A
larg
e
r
e
d
rectan
gu
la
r
f
ra
m
e ind
ic
at
es
fa
il
ed
rec
ogniti
on, w
her
eas
a
s
m
al
l gr
een
rect
angular
fram
e i
nd
ic
at
es c
orrec
t
recog
niti
on
3.4.
Res
ults
of G
A
o
p
timi
z
at
ion
of
VGG
-
f
ace m
od
el
To
im
pr
ov
e
t
he
accuracy
of
the
V
GG
-
Face
m
od
el
,
the
pr
opos
e
d
G
A
is
us
ed
offli
ne
f
or
featur
e
sel
ect
ion
.
T
o
sel
ect
the
best
distance
for
the
K
NN
cl
ass
ifie
r,
rec
ogniti
on
r
at
es
are
cal
culat
ed
us
in
g
eac
h
distance,
ke
epi
ng
the
G
A
pa
r
a
m
et
ers
fixe
d
(
=30
a
nd
=60).
As
repo
rted
in Tab
le
2
,
the
G
A
op
ti
m
iz
ation,
us
in
g
a
Ma
nh
a
tt
an
dista
nce
KNN
cl
assifi
er
,
is
a
ble
t
o
im
pro
ve
t
he
f
or
FGNET
datase
t
from
=80.7%
us
in
g
t
he
w
hole
40
96
FC
6
featu
res
t
o
=85.8%
usi
ng
t
he
op
ti
m
al
featur
e
set
of
l
eng
t
h
=
2021
.
Fo
r
MOR
PH
-
I
I
dataset
,
G
A
optim
iz
at
ion
suc
cessf
ully
i
m
pr
oves
the
fro
m
=92.2
%
to
=96.0%
us
in
g
an
op
ti
m
al
featur
e
set
of
le
ngth
=2
005
.
I
n
a
dd
it
io
nal
to
th
e
quantit
at
ive
r
esults,
visu
al
r
esults,
represe
nted
in
Fig
ur
e
5,
s
uppo
rt
the
ef
fici
ency
of
the
propose
d
G
A
optim
iz
at
ion
to
im
pr
ov
e
the
pe
rfor
m
ance.
To
furth
e
r
inv
est
igate
the
best
par
am
et
er
set
ti
ng
of
the
GA,
dif
fer
e
nt
nu
m
ber
s
of
po
pu
la
ti
on
siz
es
of
=30
a
nd
40
an
d
diff
e
re
nt
num
ber
of
ge
ner
at
io
ns
of
=6
0,
80,
and
10
0
are
in
vestigat
ed
us
i
n
g
the
best
K
N
N
distance
m
etr
ic
,
i.e.,
t
he
Ma
nhat
ta
n
dista
nce.
As
s
how
n
i
n
Figure
6,
val
ues
a
re
sli
gh
tl
y
var
yi
ng
f
or
t
he
FGNET
dataset
us
in
g
dif
fer
e
nt
pa
ram
et
ers,
be
tween
85%
a
nd
86.
2%
,
with
t
he
best
set
t
ing
=60
a
nd
=40,
ac
hievi
ng
=86.2%
,
with
a
featu
re
vecto
r
of
le
ng
t
h
=2014.
I
n
c
ontrast
,
cha
ng
i
ng
the
par
am
et
ers
for
MORPH
do
not
change
the
an
d
is
ke
pt
fixe
d
at
96
.
0%
,
us
i
ng
a
fixe
d
feat
ur
e
vecto
r
of
le
ng
t
h
=
2005
.
These
re
su
lt
s
su
pp
or
t
the
fac
t
that
the
FGN
ET
dataset
is
m
or
e
chall
eng
i
ng
t
han
the
M
ORPH
dataset
,
and
t
her
e
f
or
e
require
s
m
or
e
effor
t
t
o
reac
h
t
he
op
ti
m
al
per
f
or
m
an
ce.
Fi
gure
7
s
umm
arizes
the
im
pr
ov
em
ent
o
f
the
pro
pose
d
GA
op
ti
m
iz
ation
on the
achie
ve
d
value
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Op
ti
mize
d
Dee
p
Le
arnin
g
A
ge
-
I
nvari
ant
Fa
ce Reco
gnit
ion (
A. M
ou
st
afa
)
1839
Figure
6.
Com
par
is
on b
et
wee
n
f
or
FGNET
dataset
u
si
ng
di
ff
ere
nt
GA p
a
ram
et
er s
et
ti
ng
(
=60
,
80,
or
100
a
nd
=30
or
40), an
d
a
Man
hatta
n dist
ance
K
NN cla
ssifie
r
Figure
7
.
F
or
F
GN
E
T, t
he
Ra
nk
-
1
i
m
pr
oves
f
r
om
8
0.7%
(
usi
ng all
FC6 fe
at
ur
es,
=
4096) t
o
86%
after
GA feat
ur
e
sel
ect
ion
, a
chie
ve
d usin
g
a
Ma
nhat
ta
n
distance
K
NN
cl
assifi
e
r
. Fo
r
MOR
PH, the Ra
nk
-
1
i
m
pr
oves
from
92.8%
(u
si
ng
al
l FC6 f
eat
ur
e
s,
=
4096)
t
o 96
% after
GA
fea
ture
sel
ect
io
n,
achie
ve
d usin
g
a
Ma
nhat
ta
n dist
ance
KNN
cl
assifi
er
3.5. C
ompari
ng
wi
th st
ate
-
of
-
the
-
art al
gori
th
ms
To
highli
ght
the
co
ntributi
on
of
the
pr
opos
e
d
m
et
ho
d,
Table
3
com
par
es
the
perf
orm
ance
of
the
pro
po
se
d
m
et
ho
d
with
t
he
sta
te
-
of
-
t
he
-
art
te
c
hniq
ues
,
in
te
rm
s
of
t
he
on
tw
o
st
and
a
r
d
bio
m
etr
ic
dataset
s,
i.e
.
F
GN
E
T
a
nd
M
ORPH.
As
re
porte
d
in
the
Ta
ble,
th
e
propos
ed
GA
opti
m
izati
on
s
hows
a
n
abili
t
y
to
ac
hieve
a
s
up
e
rio
r
perf
orm
ance,
e
vid
e
nc
ed
by
the
hi
gh
e
st
.
No
te
t
ha
t
so
m
e
of
t
he
sta
te
-
of
-
t
he
-
art
al
gorithm
s can’
t be c
om
par
ed
to
the pr
opos
e
d
m
et
ho
d
since
they ei
ther
use
a subset of the
d
at
a (
no
t t
he whole
dataset
,
as
the
propose
d
m
et
hod
do
with
the
FG
N
ET
dataset
)
an
d/or
us
in
g
di
ff
e
r
ent
trai
ni
ng
schem
e
(e.
g.
,
5
-
fo
l
d
cr
os
s
-
validat
io
n
instea
d
of
le
av
e
-
one
-
out
as
the
pro
posed
m
et
hod
does)
.
T
he
el
apse
d
tim
es
for
the
pr
opos
e
d
m
et
ho
d
f
or
fe
at
ur
e
e
xtracti
on
a
nd
cl
assifi
c
at
ion
a
re
0.7
92
sec
an
d
0.0
139
sec,
res
pe
ct
ively
.
Ther
e
f
or
e,
the
ov
e
rall
tim
e o
f
the co
m
plete
prop
os
ed
fram
e
work is
0.806
s
ec.
Table
3.
C
om
par
at
ive re
su
lt
s i
n
te
rm
s
of the
Ra
nk
-
1
betwee
n
the
pr
opos
e
d m
et
ho
d
a
nd t
he
stat
e
-
of
the
-
art m
et
h
od
s for
A
IF
R
on
FGNET a
nd M
PROP
-
II
datas
et
s
Metho
d
p
reproces
sin
g
FGNE
T
MORPH
Gen
erative
Mod
el
[
2
8
]
YES
3
7
.4%
7
9
.8%
Discri
m
in
ativ
e M
o
d
el [
9
]
YES
4
7
.5%
8
3
.9%
Hid
d
en
Factor Ana
ly
sis
[
5
]
YES
69%
8
6
.8%
Maxi
m
u
m
E
n
trop
y
m
o
d
el [
11
]
YES
7
6
.2%
9
2
.3%
Deep Jo
in
t L
earnin
g
app
roach
[
13
]
YES
NA
9
3
.6%
FR acros
s ti
m
e
lap
se
u
sin
g
CNN
[
17
]
YES
8
0
.6%
9
2
.2%
Mod
if
ied
Hidd
en
Facto
r
An
al
y
sis
[
12
]
YES
7
6
.8%
88%
AIFR
-
CN
N
m
o
d
el
[
19
]
YES
7
6
.5%
9
2
.5%
Prop
o
sed
Sy
ste
m
NO
8
6
.2%
9
6
.0%
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
2
,
A
pr
i
l 202
0
:
1833
-
1841
1840
3.6.
Li
mi
tatio
ns
As
e
xp
ect
e
d,
exp
e
ri
m
ents
on
MOR
PH
dat
abase
ac
hieve
sign
ific
a
ntly
be
tt
er
res
ults
th
an
F
G
NET.
This
is
du
e
to
the
fact
that
t
he
FGNET
is
bein
g
m
or
e
chall
eng
in
g
sinc
e
it
includes
hi
gh
e
r
age
ga
ps
than
MORPH
(m
axi
m
u
m
age
ga
p
betwee
n
im
ages
of
MORP
H
s
ubj
ect
s
is
4
ye
ar)
.
Fig
ur
e
8
s
hows
s
om
e
sam
ple
s
of
fail
ed
rec
o
gniz
ed
im
ages
us
in
g
the
pro
pose
d
m
et
ho
d.
Figure
8
(
b)
i
nd
ic
at
e
s
that
i
m
age
pro
per
ti
es
li
ke
il
lu
m
inati
on
an
d
distance
fro
m
the
cam
era
m
ay
deg
ra
de
t
he
resu
lt
of
M
ORPH
da
ta
set
.
Ex
per
im
ental
resu
lt
on
FGN
ET
datase
t
shows
a
ver
y
high
sim
il
arity
a
m
on
g
c
hild
ren
in
a
ge
ga
p
(
0
-
3
ye
ar
s
old),
s
o
m
os
t
inc
orrect
m
at
ching
occ
urs
in
this
a
ge
gap
,
as
sho
wn
in
Fi
gure
8
(
a).
Howe
ver,
i
n
the
te
ena
ge
gap
the
face
f
eat
ur
es
app
ea
r
to
be
unsta
ble
in
this
sta
ge
an
d
this
is
cl
early
ob
vio
us
in
th
e
m
is
m
at
ching
im
ag
es.
More
sta
bil
it
y
in
featur
e
s
occur
s
in a
du
lt
sta
ge
t
hat
le
ads
to
im
pro
ving
the
rec
ogniti
on
rates
.
(a)
(b)
Figure
8
.
V
is
ua
l sy
stem
o
f
fai
le
d
im
age r
et
rieval (a
)
th
ree
s
a
m
ples o
f
F
G
NET
a
nd
(b)
th
ree
sam
ples f
r
om
MORPH
(
al
bu
m
-
II
)
dataset
s.
The first
row
is
the test
im
age.
The
se
co
nd
row
s
hows
Fail
ed
Re
trie
ve
d (F
R)
i
m
ages u
sin
g o
ur syst
em
. Th
e last
r
ow is
the
gro
und
tr
uth (
GT),
dete
rm
ined
as
the
recog
nized pe
rson
i
m
age
of the
near
e
st a
ge
4.
CONCL
US
I
O
N
A
deep
le
a
rn
i
ng
m
et
ho
d
for
AI
FR
is
pro
po
sed.
The
pro
po
sed
m
et
ho
d
i
nvol
ves
no
re
gis
trat
ion
ste
ps
,
wh
e
re
ra
w
fac
e
data
are
i
nput
directl
y
to
a
VGG
-
Face
dee
p
le
ar
ning
m
odel
to
extract
fa
ce
featu
res,
a
voidin
g
tim
e
-
con
su
m
ing
,
un
reli
able
pr
e
processi
ng
ste
ps
.
T
he
ext
racted
face
fe
at
ur
es
are
further
op
ti
m
iz
ed
us
i
ng
a
sm
art
GA
that
pe
rfor
m
s
featur
e
sel
ect
ion
a
nd
re
duct
ion
,
res
ulti
ng
on
t
he
m
os
t
relevan
t
feat
ures
t
o
the
pro
blem
of
AIFR.
In
ad
di
ti
on
,
t
he
propo
sed
GA
opti
m
i
zat
ion
s
hows
a
n
a
bili
ty
to
i
m
pro
ve
the
rec
og
niti
on
accuracy
f
ro
m
80
.
7%
a
nd
92
.8
%
to
86.2%
and
96%
on
two
publicl
y
avail
able
databa
ses,
i.e.,
F
G
N
ET
an
d
MORPH
,
res
pe
ct
ively
.
Co
m
par
i
ng
to
the
st
at
e
-
of
-
the
-
art
m
et
ho
d,
t
he
propose
d
m
et
ho
d
has
show
n
pr
om
isi
ng
resu
lt
s
.
REFERE
NCE
S
[1]
A.
La
nitis
,
C.
J.
Tay
lor,
and
T
.
F.
Coote
s,
“
Towa
rd
aut
om
at
i
c
sim
ula
ti
on
of
agi
n
g
eff
ec
ts
on
fa
c
e
images,
”
IEEE
Tr
ans.
Pat
te
rn
A
nal.
Ma
ch. I
nt
el
l
.
,
vol. 24, no. 4,
pp.
442
–
455
,
20
02.
[2]
U.
Park
and
A.
K.
Jain,
“
Face
m
at
chi
ng
and
re
tri
ev
al
using
soft
biometri
c
s,
”
IEE
E
Tr
ans.
Inf.
Forensic
s
Sec
u
r.
,
vol.
5
,
no
.
3
,
pp
.
406
–
415,
2010
.
[3]
The
Fac
e
and
Gesture
Rec
ognition
Network
(FG
-
NET)
Aging
Data
b
ase
,
ht
tps:/
/ww
w
-
prima.
inrialp
es.
f
r/FGnet/
html/benchm
ark
s.ht
m
l
[
ac
c
essed
Dec
ember
2019]
.
[4]
K.
Ric
an
ek
an
d
T.
T
esa
fa
y
e
,
“
Mor
ph:
A
l
ongit
udinal
ima
ge
databa
se
of
norm
al
adul
t
age
-
progre
ss
ion
,
”
in
Au
tomati
c
Fa
ce
and
Gesture Rec
ogn
it
ion
,
7
th
Inte
rnational
Co
nfe
renc
e
on
FG
R
,
pp
.
341
–
345
,
2006.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Op
ti
mize
d
Dee
p
Le
arnin
g
A
ge
-
I
nvari
ant
Fa
ce Reco
gnit
ion (
A. M
ou
st
afa
)
1841
[5]
D.
Gong,
Z
.
Li,
D.
Li
n
,
J.
Liu,
and
X.
T
ang
,
“
Hidden
f
ac
to
r
anal
y
s
is
for
a
ge
inv
ari
an
t
fa
c
e
re
cogni
t
ion,”
in
Proc
ee
dings
o
f
th
e
I
EE
E
Int
ernati
onal
Con
fe
re
nce
on
Compute
r V
ision
,
pp.
287
2
–
2879,
2013
.
[6]
D.
G.
Lowe
,
“
Distinc
ti
v
e
image
fe
at
ur
es
fro
m
sca
le
-
inv
aria
nt
ke
y
poin
ts,
”
I
nt.
J.
Comput.
Vi
s.
,
vol
.
60,
n
o.
2,
pp.
91
–
110
,
200
4.
[7]
T.
Oj
al
a
,
M.
P
ietika
in
en,
and
T.
Mae
npaa,
“
Mult
ire
soluti
on
gra
y
-
sca
le
and
ro
ta
t
io
n
inva
ri
ant
te
x
tu
re
c
la
ss
ifi
c
at
ion
with
lo
ca
l
bin
ar
y
patter
ns,
”
I
EE
E
Tr
ans.
Pat
te
rn
A
nal.
Ma
ch. I
nt
el
l
.
,
vol. 24, no. 7,
pp.
971
–
987
,
20
02.
[8]
S.
R.
Konda,
V.
Kum
ar,
and
V.
Krishna,
“
Face
re
co
gni
ti
on
using
m
ult
i
re
gion
prom
ine
nt
lbp
re
pre
senta
ti
on
,
”
Int
.
J
.
El
e
ct
r.
Comput.
Eng.
,
vol
.
6
,
no
.
6,
p
.
2781
,
2016
.
[9]
H.
Li
ng,
S.
Soatt
o
,
N.
Ramanat
han
,
and
D.
W
.
Jac
obs,
“Fac
e
ver
if
icati
o
n
ac
ross
age
progre
ss
ion
using
discri
m
ina
tive
m
et
hods,”
IE
EE Tr
ans.
Inf. F
orens
ic
s Secur.
,
vol
.
5
,
no
.
1
,
pp
.
82
–
9
1,
2010
.
[10]
Z.
Li
,
U.
Park
,
and
A.
K
.
Jain
,
“
A
discri
m
ina
tive
m
odel
for
a
ge
inv
ari
an
t
f
ace
re
cogni
t
ion,”
IEE
E
Tr
ans.
Inf
.
forensic
s Secur.
,
vol. 6, no. 3, pp. 1028
–
1037,
201
1.
[11]
D.
Gong,
Z
.
L
i,
D.
Ta
o
,
J.
Li
u,
and
X
.
Li,
“
A
m
axi
m
um
ent
rop
y
f
eatur
e
desc
r
ipt
or
fo
r
age
inva
r
ia
nt
fa
ce
re
cogni
t
io
n,
”
in
Proce
ed
ings
of
the
I
E
EE
conf
ere
n
ce
on
comput
er
vi
sion
and
pattern
rec
ogni
ti
on
,
pp.
5289
–
5297,
2015.
[12]
H.
Li,
H.
Zou,
a
nd
H.
Hu,
“
Modifie
d
h
idde
n
f
act
or
ana
l
y
sis
for
c
ross
-
age
fa
c
e
re
cogni
ti
on
,
”
IEEE
Signal
pro
ce
s
s.
Lett
.
,
vol
.
24,
no
.
4
,
pp
.
465
–
469
,
2017.
[13]
Y.
Li
,
G.
W
ang
,
L.
Li
n,
and
H.
Chang,
“
A
de
ep
joi
nt
l
ea
rn
in
g
appr
oac
h
for
ag
e
invarian
t
fa
ce
ver
ifi
c
a
ti
on
,
”
in
Chin
ese
Conf
ere
nce on
Comp
ute
r V
ision
CCF
,
pp
.
296
–
305
,
2
015.
[14]
O.
M.
Parkhi,
A.
Veda
ldi,
A.
Zi
ss
erman,
and
othe
rs,
“
Dee
p
f
ac
e
r
ec
ogn
it
ion
.
,
”
in
Proceedi
n
gs
of
the
Bri
ti
sh
Mac
hine V
ision
Confe
renc
e
(
BM
VC)
,
Swansea
,
v
ol.
1
,
no
.
3
,
p
.
6
.
[15]
Z.
Qawaqne
h
,
A.
A.
Mall
ouh,
an
d
B.
D.
Barka
na
,
“
Dee
p
Convolut
iona
l
Neur
al
Network
for
Age
Esti
m
at
ion
base
d
on
VG
G
-
Face
M
odel
,
”
arX
iv
Pre
pr.
arXiv
1709
.
0
1664
,
2017
.
[16]
A.
Anand,
R.
Donida
La
b
at
i
,
A.
Genove
se,
E.
Munoz
Bal
le
st
er
,
V.
Piuri,
and
F.
Scott
i
,
“
Age
es
ti
m
at
ion
base
d
o
n
fa
ce i
m
age
s
and pre
-
traine
d
conv
olut
ional
n
eur
al
net
works
,
”
in
CI
SDA
,
2017
.
[17]
H.
El
Khi
y
a
ri
a
nd
H.
W
ec
hsler
,
“
Face
re
cogni
t
i
on
ac
ross
ti
m
e
l
apse
using
conv
olut
ional
neu
ra
l
net
works
,
”
J.
In
f.
Sec
ur.
,
vol
.
7
,
no
.
03
,
pp
.
1028
–
1
037,
2016
.
[18]
C.
Xu,
Q.
Li
u
,
and
M.
Ye,
“
Age
inva
ri
ant
f
ace
re
cogn
it
ion
a
nd
re
tri
ev
al
b
y
coupl
ed
au
to
-
en
code
r
ne
tworks,”
Neurocomputi
ng
,
vol
.
222
,
pp
.
62
–
71,
2017
.
[19]
M.
Nim
bar
te
an
d
K.
Bho
y
ar
,
“
Age
Inva
ria
n
t
Face
R
ec
ogni
ti
on
Us
ing
Convolut
iona
l
Neur
al
N
et
work,”
In
t.
J.
El
e
ct
r.
Comput.
Eng.
,
vol
.
8
,
no
.
4,
2018
.
[20]
A.
Dutta
,
M.
Günther
,
L
.
El
Shafe
y
,
S.
Marc
el,
R.
Veldhui
s,
and
L.
Spree
uwers,
“
Im
pac
t
of
e
y
e
d
et
e
ct
ion
err
or
o
n
fa
ce recogni
t
ion performance,”
I
ET bi
ometri
cs
,
v
ol.
4
,
no
.
3
,
pp
.
1
37
–
150,
2015
.
[21]
H.
Moham
m
adz
ade
,
A.
S
a
yy
a
fa
n,
and
B.
Ghoj
ogh,
“
Pixel
-
le
v
e
l
al
ignment
of
f
ac
i
al
images
for
high
ac
cur
a
c
y
re
cognition
usin
g
ense
m
ble of
p
a
tc
hes,”
JOSA
A
,
vol.
35
,
no
.
7
,
pp
.
1149
–
1159
,
20
18.
[22]
A.
S.
Ra
za
vi
an,
H.
Aziz
pour
,
J.
Sulli
van,
and
S
.
Carl
ss
on,
“
CNN
fe
atures
off
-
th
e
-
shelf:
an
astoun
ding
base
l
ine
for
re
cognition,”
in
Proce
edi
ngs
of
the
IEE
E
con
f
ere
nce
on
comp
ute
r
vi
sion
and
patt
ern
rec
ognit
ion
workshops
,
pp.
806
–
813
,
20
14.
[23]
X.
Yin,
X.
Yu,
K.
Sohn,
X.
Li
u
,
and
M.
Ch
andr
ake
r,
“
Fea
ture
T
ra
nsfer
Learni
ng
for
Dee
p
Fac
e
Rec
ognition
wit
h
Long
-
Tail
Da
ta,
”
arXiv
Prepr
.
a
r
Xiv1803.
09014,
2018.
[24]
C.
-
L.
Zha
ng
,
H.
Zha
ng
,
X.
-
S.
W
ei
,
and
J
.
W
u,
“
Dee
p
bimoda
l
re
gre
ss
ion
for
appa
re
n
t
per
son
al
ity
anal
y
s
is,
”
i
n
Eur
opean
Conf
e
renc
e
on
Compu
te
r V
ision
,
pp.
3
11
–
324,
2016
.
[25]
B.
L.
Mil
le
r
,
D.
E.
Goldber
g
,
a
nd
othe
rs,
“
Gene
tic
al
gori
thms
,
tourna
m
ent
sel
e
ct
ion
,
and
the
e
ffe
ct
s
of
noise
,
”
Complex
S
yst
.
,
v
ol.
9
,
no
.
3
,
pp
.
1
93
–
212,
1995
.
[26]
A.
E.
E
ibe
n
and
C.
A.
Schippers
,
“
On
ev
olut
ion
a
r
y
expl
ora
ti
on
a
nd
expl
oitat
ion,
”
Fundam.
Infor
matic
ae
,
vol
.
35
,
no.
1
–
4
,
pp
.
35
–
50,
1998
.
[27]
D.
Thiere
ns,
“
Adapti
v
e
m
uta
t
ion
ra
t
e
con
trol
sch
emes
in
gen
et
i
c
al
gorit
hm
s,”
in
Proce
ed
ings
of
t
he
2002
Congre
ss
on
Ev
o
lut
ionary
Computati
on.
C
EC’02
(
Cat.
No
.
02TH
8600
)
,
vol. 1, pp. 980
–
985,
2002.
[28]
T.
Cover
and
P.
Hart
,
“
Nea
r
est
n
ei
ghbor
pa
tt
e
rn
cl
assifi
ca
t
ion,”
I
EE
E
Tr
ans.
Inf
.
the
ory
,
vo
l.
13
,
no.
1,
pp
.
21
–
27
,
1967.
[29]
U.
Park,
Y.
To
ng,
and
A.
K.
J
ai
n,
“
Age
-
inva
ri
ant
f
ace
r
ec
ogni
ti
on,
”
IE
EE
Tr
ans.
Pattern
Ana
l.
Ma
ch.
Int
el
l
.
,
vol.
32
,
no
.
5
,
pp
.
947
–
954
,
2010
.
Evaluation Warning : The document was created with Spire.PDF for Python.