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.
9
, No
.
5
,
Octo
ber
201
9
, pp.
4069~
40
76
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v9
i
5
.
pp4069
-
40
76
4069
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Biologi
cal l
andm
ar
k
v
s
quasi
-
lan
dm
ar
ks
for 3D fa
ce re
cogn
ition
and gend
er
cl
assification
Ha
w
ra
a
H.
A
bb
as
1
,
Amm
ar
A
.
A
l
ta
mee
mi
2
,
Ham
eed
R.
Far
han
3
1,
3
Depa
rt
m
ent of
Elec
tr
ical and
E
le
c
troni
c
Eng
ineeri
ng,
Univer
si
t
y
of
Kerb
al
a
,
Ir
aq
2
Depa
rt
m
ent
of
Biom
edi
cal
Eng
i
nee
ring
,
Univ
ersity
of
Kerba
la,
Ir
aq
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
a
n
2
9
, 2
01
9
Re
vised
A
pr
13
, 2
01
9
Accepte
d
Apr
25
, 201
9
F
ac
e
rec
ogn
it
io
n
and
gende
r
class
ifi
cation
ar
e
vit
al
top
ic
s
in
t
he
fie
ld
o
f
computer
gra
ph
ic
and
pa
ttern
rec
ogni
ti
on.
W
e
uti
lize
d
id
ea
s
from
two
growing
ide
as
in
computer
vision
,
which
ar
e
biol
o
gic
a
l
la
ndm
ark
s
and
quasi
-
la
ndm
ark
s
(de
n
se
m
esh)
to
pr
opose
a
novel
appr
oac
h
to
co
m
par
e
the
ir
per
form
anc
e
in
f
ac
e
re
cognition
and
gende
r
c
la
ss
ifi
c
at
ion
.
The
ex
per
imental
work
is
conducte
d
on
FR
RGv2
dat
ase
t
and
a
c
quire
d
98%
and
94%
fac
e
rec
ogni
ti
on
a
cc
u
rac
i
es
using
the
quasi
and
bio
log
ic
a
l
la
ndm
ark
s
r
espe
ctive
l
y
.
The
gende
r
c
la
s
sific
a
ti
on
accuracie
s
are
92%
for
quasi
-
la
ndm
ar
ks
and
90%
for
biol
og
ic
a
l la
ndm
ark
s.
Ke
yw
or
d
s
:
Bi
olo
gical
lan
dm
ark
s
Den
s
e s
urface
Face rec
ogniti
on
FRR
Gv2
datas
et
Gende
r
cl
assifi
cat
ion
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights
reserv
ed
.
Corres
pond
in
g
Aut
h
or
:
Hawraa
H.
A
bbas,
Dep
a
rtm
ent o
f
Ele
ct
ric an
d
El
ect
ronic
Engi
n
e
eri
ng
,
Un
i
ver
sit
y o
f Ker
bala,
Kerba
la
, I
ra
q.
Em
a
il
: haw
raa.
h@u
ok
e
rb
al
a.
e
du.iq
1.
INTROD
U
CTION
Nowa
days,
ou
r
su
r
r
oundin
g
env
i
ronm
ent
beca
m
e
m
or
e
tech
nical
ly
.
The
dev
ic
es
are
ge
tt
ing
to
be
sm
arter
ever
y
day
.
I
n
the
w
orl
d
of
te
ch
no
l
ogy
,
on
e
of
the
vital
te
chn
ol
og
ic
al
featur
es
is
a
r
ti
fici
al
intel
l
igence
.
Face
rec
ogniti
on,
as
a
bio
m
et
ric
auth
entic
at
ion
te
c
hn
i
que
is
a
n
esse
nt
ia
l
app
li
cat
ion
sco
pe
of
a
rtific
ia
l
intel
li
gen
ce
[
1
,
2
]
.
Its
m
ai
n
su
per
i
or
it
y
inh
eri
te
d
in
high
sp
e
ed
thr
ough
pu
t
wh
e
n
the
use
r
do
e
s
not
sp
en
d
tim
e
in
data
c
ollec
tio
n
proces
s
li
ke
oth
e
r
bi
om
et
ri
c
te
chn
i
qu
es
s
uch
a
s
sp
e
ake
r
recog
niti
on
[
3
]
,
fing
e
r
pr
int
[
4
]
an
d
iris
[
5
]
.
For
ex
a
m
ple,
fac
e
reco
gnit
ion
syst
e
m
can
be
instal
le
d
in
a
pu
bli
c
place
wh
e
re
f
ac
es
of
pe
ople
who
pass
can
be
rec
ognized
a
nd
pi
cking
out
the
wan
te
d
pe
rs
ons
[
6
]
.
T
he
m
a
in d
isa
dvanta
ge of
this
2
D
s
of
t
war
e is
the
sensiti
vity
in
li
gh
ti
ng
co
ndit
ion
s
,
w
hich
consi
der
e
d
as
po
s
es
an
d
occl
us
io
ns
.
I
n
the
l
ast
few
ye
ars
,
d
ue
t
o
the
rob
us
tness
of
3
-
D
te
c
hn
i
qu
e
s
in
il
lu
m
i
nation
var
ia
nc
es
and
po
s
e
al
te
rati
on
,
us
in
g
3D
facial
surf
aces
in
face r
ec
ogniti
on
has b
ec
om
e a
cr
ucial
r
esea
rc
h fiel
d.
Si
m
ultaneo
us
ly
,
gende
r
cl
ass
ific
at
ion
is
an
i
m
po
rtant
pro
bl
e
m
that
can
be
a
valua
ble
pri
m
ary
ste
p
into
fac
e
rec
ogniti
on.
Ge
nd
e
r
is
a
s
of
t
biom
et
ric
trai
ts
.
It
can
com
bin
e
oth
e
r
bio
m
et
ric
trai
ts
su
ch
as
ey
es
colo
rs,
s
kin
c
olo
r,
an
d
s
o
on,
to
di
ff
e
re
ntiat
e
between
hu
m
ans
pee
r
s.
Ma
ny
prese
nt
w
ork
s
on
gende
r
cl
assifi
cat
ion
ha
ve
us
e
d
2
-
D
f
ace
i
m
ages
f
or
so
lvi
ng
ge
nde
r
cl
assifi
cat
ion
pro
blem
,
bu
t
3
-
D
f
ace
s
urfa
ce
has
no
t
s
o
far
bee
n
us
e
d
wi
dely
fo
r
ge
nd
e
r
cl
assifi
cat
ion
.
D
iverse
ps
yc
holog
y
en
co
ur
a
ge
s
to
us
e
3
-
D
sh
a
pe
inf
or
m
at
ion
for ge
nd
e
r
cl
assifi
cat
ion
[
7
,
8
]
.
The
dev
el
op
m
ent
of
a
pr
act
ic
al
and
easy
intel
li
gen
t
syst
e
m
fo
r
a
uto
m
at
ic
face
recogn
it
io
n
a
nd
gende
r
cl
assifi
cat
ion
is
an
im
portant
s
ubj
ect
.
In
t
his
resea
r
ch
,
the
e
ff
ect
i
ve
ness
of
us
in
g
bio
lo
gical
la
nd
m
ark
(certai
n
ver
ti
ce
s
of
face
m
esh)
an
d
quasi
-
la
ndm
ark
s
(r
e
gula
rised
face
m
esh
ve
rtic
es)
on
fa
ce
rec
ogniti
on
an
d
gende
r
cl
assifi
cat
ion
syst
em
p
er
form
ance
ha
s b
ee
n
i
nv
est
i
gated
.
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.
9
, N
o.
5
,
Oct
ober
20
19
:
4069
-
4076
4070
This
resea
rch
m
ai
n
no
velty
is
te
sti
ng
the
pe
rfor
m
ance
of
bio
lo
gical
la
ndm
ark
s
in
face
recog
niti
on
and g
e
nder
classi
ficat
ion
.
S
pe
ci
fical
ly
, th
e m
ai
n
co
ntri
bu
ti
ons
of this
pa
per incl
ude:
-
It test
s the
pe
rfor
m
ance of Bi
ologica
l l
an
dma
rk for
f
ace
rec
ogniti
on and
ge
nd
e
r
cl
assifi
c
at
ion
.
-
It test
s the
pe
rfor
m
ance of
quasi
-
la
nd
m
ark
s
for face
r
ec
ogni
ti
on
and
ge
nd
er classi
ficat
io
n.
-
It
com
par
es
the
Bi
olo
gical
la
nd
m
ark
an
d
quasi
-
la
ndm
ark
s
reli
abili
ty
fo
r
gender
cl
assif
ic
at
ion
an
d
fa
ce
recog
niti
on
.
-
It
proposes
a
si
m
ple
autom
a
ti
c
face
reco
gnit
ion
a
nd
gender
cl
assifi
cat
ion
syst
em
us
i
ng
la
nd
m
ark
s
a
nd
pr
i
ncipal c
om
po
ne
nt a
naly
sis (PCA
).
The
rest
of
this
pap
e
r
is
orga
ni
sed
as
fo
ll
ows
;
In
Sect
io
n
2,
a
br
ie
f
r
evie
w
of
the
a
ppr
oac
hes
that
are
relat
ed
to
the
pro
po
se
d
m
et
ho
d
is
int
rod
uc
ed.
Sect
io
n
3
disp
la
ys
the
pr
oce
dure
of
t
he
propose
d
m
et
hod.
In
Sect
io
n
4,
the
ex
per
im
ental
resu
lt
s
are
intr
oduce
d.
Th
e
discuss
i
on
f
or
the
acqu
ire
d
r
esults
is
pr
ese
nte
d
in
Sect
ion
5.
Fin
al
ly
,
Sect
ion
6
con
cl
udes
the
m
et
ho
do
l
og
y
and
im
po
rtant
achievem
ents
of
the
propose
d
w
ork,
and it
inclu
des suggesti
ons
for f
utu
re
r
esea
rc
h
in
vestigat
io
n.
2.
RELATE
D
W
ORK
Bi
olo
gical
la
ndm
ark
s
are
sta
nd
a
r
d
refe
re
nc
e
sign
ific
at
ive
po
i
nts
on
the
f
ace,
su
c
h
as
th
e
no
se
ti
p
,
inn
e
r
an
d
ou
te
r
cor
ne
rs
of
the
ey
es
,
corners
of
the
m
ou
th,
a
nd
the
chin
po
i
nt
[
9
,
10
]
.
Lan
dm
ark
local
isa
ti
on
is
us
ua
ll
y
local
is
ed
m
anu
al
ly
,
bu
t
num
ero
us
researc
h
has
be
en
achiev
ed
on
a
uto
m
at
ic
accurate
local
i
sat
ion.
Fo
r
e
xam
ple, Ru
iz
et al
.
[
11
]
pr
ese
nted
an
a
lgorit
hm
f
or
aut
om
a
ti
c local
is
at
ion
of b
iol
ogic
a
l l
and
m
ark
s o
n 3D
faces
by
usi
ng
the A
ct
ive
S
ha
pe
M
od
el
(AS
M)
to
l
ocali
se
a set o
f 14 faci
al
b
iolo
gical
la
nd
m
ark
s.
A
series
of
di
fferent
m
easur
e
m
ents
su
c
h
as
the
Eucli
de
an
distance
a
nd
th
e
an
gle
betwee
n
bi
ologica
l
la
nd
m
ark
s
are
us
e
d
su
cce
ssf
ul
ly
in
find
in
g
t
he
di
ff
e
ren
ce
s
a
m
on
g
i
nd
i
viduals
faces
.
Th
e
accuracy
of
these
m
easur
em
ents
dep
e
nds
dr
am
at
ic
al
ly
on
la
ndm
ark
s
local
is
at
ion
a
nd
sel
e
ct
ion
.
M
orecr
oft
[
12
]
pe
rform
ed
a
researc
h
to
e
va
luate
the
us
i
ng
of
bio
l
og
ic
al
la
nd
m
ark
s
for
3D
face
s
ha
pe
a
naly
sis.
Aynec
hi
et
al
.
[
13
]
c
arr
i
e
d
ou
t
an
al
gorithm
rep
resen
ti
ng
a
face
s
ha
pe
by
a
vecto
r
of
distances
m
easur
ed
bet
ween
pairs
of
facial
la
nd
m
ark
s.
T
he
y
sel
ect
ed
the
la
nd
m
ark
s
by
an
ex
hau
sti
ve
searc
h
ov
e
r
possible
l
an
dm
ark
s
com
bin
at
ions
wh
ic
h
offe
red
the
best
recogn
it
io
n
rates.
Their
re
searc
h
resu
lt
s
rec
ord
ed
27
facial
la
nd
m
ark
s
as
es
sentia
l
la
nd
m
ark
s
for
facial
co
m
par
ison an
d
i
den
ti
fi
cat
ion
.
Su
r
face
re
gu
la
risat
ion
is
the
process
of
fin
di
ng
t
he
ge
om
etr
ic
al
trans
f
or
m
a
ti
on
betwee
n
two
or
m
or
e
su
r
faces
t
hat
al
ign
th
e
surfac
e
s
[
14
]
.
It
re
gu
l
arises
the
nu
m
ber
of
ve
rtic
es
by
fin
ding
de
ns
e
c
orrespo
ndences
in
the
3D
m
e
sh
es,
t
hese
re
gu
la
rised
ver
ti
ces
are
cal
le
d
qu
asi
-
la
ndm
a
rk
s
.
The
quasi
la
nd
m
ark
s
ar
e
us
ed
dr
am
at
ic
ally i
n
com
par
in
g
the
d
if
fer
e
nces i
n faci
al
m
or
phol
og
ie
s
for
va
rio
us
po
pula
ti
on
s
and se
xes [
15
,
16
].
The
cl
ear
se
xual
dim
or
phism
of
the
hu
m
an
face
m
akes
face
ge
nder
cl
assifi
cat
ion
consi
der
a
bly
eff
ic
ie
nt
an
d
fa
st
op
erati
on
[
7
]
.
In
s
pite
of,
th
e
abun
dan
t
ge
nder
cl
assifi
cat
ion
r
esearc
h,
th
e
ta
sk
of
dev
el
op
i
ng
easy
to
i
m
ple
m
ent,
autom
ated
an
d
r
obus
t
f
ace
gender
cl
a
ssific
at
ion
is
sti
ll
a
re
m
ote
ob
j
ect
ive.
To
ov
erco
m
e
the
lim
i
ta
ti
on
it
e
m
s
of
2
-
D
i
m
age
base
m
eth
ods,
t
hr
ee
-
dim
ension
al
(
3
-
D)
in
f
or
m
at
ion
of
hu
m
an
face
s
is
us
ed
to
pe
rfor
m
f
ac
e g
e
nd
e
r
cl
assi
ficat
ion
.
F
or
e
xam
ple, in O’
T
oole
et al
.
[
8
]
pr
incipal
co
m
ponen
ts a
naly
sis
(P
CA
)
was
ap
plied
se
par
at
el
y
to
the
3
-
D
m
esh
and
gr
ay
le
vel
i
m
age
pix
el
s
of
3
-
D
hum
an
face
i
m
age
.
The
resu
lt
s
dem
on
strat
ed
t
he
th
ree
-
dim
ension
al
face
m
e
sh
pro
vid
e
d
m
or
e
preci
se
se
x
cl
assifi
cat
ion
resu
lt
s
t
han
th
e
gray
le
vel
i
m
age
pix
el
s.
Y
ua
n
et
al
.
[
1
7
]
us
e
d
the
feat
ur
es
from
four
face
reg
i
on
s
(up
pe
r
re
gion
of
the
face
,
the lo
wer re
gion
of the
face, t
he nose
a
nd th
e left
eye
)
for gen
de
r
cl
assifi
cat
ion
pur
po
se
.
A
gen
e
ral
re
vi
ew
on
3D
fa
ce
recog
niti
on
m
et
hods
is
i
ntrodu
ce
d
i
n
[
1
8
]
.
The
diff
ic
ulty
in
3D
face
recog
niti
on
al
gorithm
s’
i
m
p
lem
entat
ion
is
handling
oc
cl
us
io
ns
,
m
iss
ing
fa
ce
par
t
s
,
an
d
ex
pr
es
sion.
The
e
xisti
ng
3D
face
rec
ogni
ti
on
m
et
ho
ds
c
an
be
cat
eg
or
i
sed
int
o
tw
o
c
at
egories
acco
r
ding
to
the
fea
ture
s
us
e
d
in
recog
ni
ti
on
:
holi
sti
c
m
et
ho
ds
a
nd
r
egio
nally
appr
oach
e
s.
I
n
holi
sti
c
m
et
ho
ds
,
t
he
feat
ur
es
fro
m
the
whole
is
us
e
d
for
recog
niti
on.
D
rira
et
al
.
[
1
9
]
im
ple
m
ent
ed
a
m
et
hod
t
hat
use
s
c
urvat
ur
es
of
t
he
3D
facial
su
r
faces.
T
he
m
et
ho
d
us
ed
t
he
ra
dial
curve
s
of
facial
surf
aces
arou
nd
th
e
no
se
ti
p
f
or
r
ecognit
ion.
Mo
rpha
ble
m
od
el
-
based
a
ppr
oac
hes
were
introduc
ed
i
n
[
20
,
2
1
]
,
th
is
m
od
el
us
e
d
eff
ect
ively
for
face
rec
ogni
ti
o
n.
In
l
ocal
-
reg
i
on
base
d
a
ppr
oa
ches,
l
ocal
in
f
or
m
at
ion
is
ex
tract
ed
f
r
om
s
ever
al
face
re
gions.
For
e
xa
m
ple,
Li
et
al
.
[
22
]
pro
posed
a
r
obust
local
facial
descr
i
ptor
co
nt
ai
ns
f
our
ty
pes
of
ge
om
et
ric
featur
e
s.
T
hen,
this
descr
i
ptor is
use
d for
recog
niti
on
pur
pose
.
In
sp
it
e
of
the
accuracy
of
th
ese
m
et
ho
ds
in
gender
cl
assif
ic
at
ion
and
fac
e
reco
gnit
io
n
,
they
su
ff
e
r
from
the
fo
ll
ow
in
g
pro
blem
s
:
com
plexity
,
a
nd
sl
ow
feature
extracti
on
an
d
rec
ogniti
on
processes
.
The
refore
,
pro
po
se
d
a
sim
ple
an
d
acc
urat
e
face
regnit
io
n
m
et
ho
d
is
e
s
sentia
l.
I
n
t
his
pa
per
,
bi
ologica
l
la
nd
m
ark
a
nd
the
qu
a
si
-
la
ndm
ark
s
a
re
e
xtrac
te
d
an
d
us
e
d
as
feat
ur
e
de
scri
ptor
for
3D
face
rec
ogniti
on
a
nd
ge
nd
e
r
cl
assifi
cat
ion
.
This
resea
rch
r
epr
ese
nts
the
f
irst
researc
h
te
sts
the
eff
ic
ie
nc
y
of
bio
l
og
ic
al
la
nd
m
ark
and
the
qu
a
si
-
la
ndm
ark
s
i
n
face
r
ec
ogniti
on a
nd ge
nd
e
r
cl
assifi
cat
ion
.
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
Biolo
gical la
ndm
ar
k
vs
quas
i
-
la
ndm
ar
ks f
or 3
D
face
rec
og
niti
on
and ge
nder
..
.
(
Ha
wr
aa
H.
Ab
bas)
4071
3.
PROP
OSE
D ALGO
RTHI
M
In
this
sect
io
n,
we
desc
ribe
a
pro
po
se
d
a
ppr
oach
of
face
re
cogniti
on
a
nd
gende
r
cl
assifi
cat
ion
us
i
ng
the
bi
ologica
l
la
nd
m
ark
or
the
qu
a
si
-
la
nd
m
ark
s.
Fig
ur
es
1
a
nd
2
disp
l
ay
blo
c
k
diag
r
a
m
s
of
t
he
propose
d
appr
oach,
w
hich
are
e
xp
la
i
ned
belo
w.
F
igure
1
diag
r
a
m
illustrates
the
pro
po
s
ed
appro
a
c
h
for
face
recog
niti
on
,
w
hile
F
igure
2
il
lustrate
s
the
gende
r
cl
assifi
cat
ion
ap
proac
h.
Fi
gure
3
s
hows
t
he
la
nd
m
ark
s
extracti
on
blo
c
k
c
on
te
nts.
Figure
1.
The
pro
po
se
d
a
ppr
oach f
or f
ace
re
cogniti
on
Figure
2
.
The
pro
po
se
d
a
ppr
oach f
or g
e
nde
r
cl
assifi
cat
io
n
Figure
3
.
The
land
m
ark
s e
xtra
ct
ion
3.1.
Dataset
We
util
iz
e
d
th
e
FRGC
v2
dat
abase
.
T
his
dat
a
con
ta
in
4,007
3D
face
sca
ns
of
466
s
ubj
ec
ts
.
The
data
wer
e
colle
ct
ed
us
ing
a
Mi
no
l
ta
91
0
ty
pe
la
ser
scan
ner
.
T
he
i
m
ages
reco
r
d
dif
fer
e
nt
facial
exp
ressi
ons
(e.
g.,
happine
ss
a
nd
surp
rise
)
.
M
aur
e
r
et
al
.
m
anu
al
ly
cat
egorised
the
F
RGC
v2
i
nto
three
cat
eg
ori
es
[
23
]
dep
e
ndin
g
on
t
he
facial
ex
pre
ssion.
T
hey
cl
a
ssifie
d
796
im
ages
a
s
sm
al
l
e
x
pressi
on
s
a
nd
74
2
im
ages
as
la
rg
e
expressi
on
s
,
wh
il
e
2469
i
m
ages
cl
assifi
ed
as
the
ne
utral
ex
pr
es
sion.
Fig
ur
e
4
FRR
Gv2
3D
face
i
m
age ex
am
ple
.
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.
9
, N
o.
5
,
Oct
ober
20
19
:
4069
-
4076
4072
Figure
4
.
FRR
Gv2 3
D
fa
ce i
m
age ex
am
ple
Fo
r
face
recog
niti
on
e
xperim
ent,
the
i
nput
to
the
al
gorith
m
is
two
set
s
of
im
ages
galle
ry
an
d
pro
be
faces.
Faces
i
n
the
galle
ry
set
re
pr
ese
nt
faci
al
3D
im
ages
known
to
t
he
s
yst
e
m
,
wh
il
e
f
aces
in
t
he
pro
be
set
represe
nt
unkn
own
facial
3D
i
m
ages
pr
ese
nted
to
th
e
syst
e
m
fo
r
th
e
re
cogniti
on.
I
n
this
w
ork,
face
s
with
neu
t
ral
ex
pr
es
s
ion
a
re u
sed
.
A
ccordin
g
to
the
Jack
-
knife
ru
l
e
80%
of
t
he
f
aces
ra
ndom
ly
sel
ect
ed
f
or
tra
inin
g
wh
il
e
the
rest
3D
fac
es
are
use
d
f
or
pr
opose
d
al
gorithm
testing
.
T
her
e
for
e,
1975
i
m
ages
are
us
ed
f
or
trai
ning
and
494
3D
fa
ce
i
m
ages
are
us
e
d
f
or
syst
e
m
te
sti
ng
.
For
gende
r
cl
assifi
cat
ion
a
ppro
ac
h
,
466
s
ubj
ect
s
i
m
ages
are
us
e
d
(
202
f
e
m
al
e
and
264
m
al
e),
al
so
accor
ding
to
Jac
k
-
knife
te
ch
nique
a
total
of
372
im
ages
were
us
e
d
for
trai
ning a
nd
9
4 i
m
ages were
us
ed
for t
es
ti
ng
.
3.2.
Preproces
sing
This
sta
ge
c
onsist
s
of
thre
e
proces
ses:
Clip
t
he
f
aci
al
reg
i
on,
Re
m
ov
e d
at
a
sp
i
kes
(
noise
)
an
d
Fil
l
the
ho
le
s
.
-
Denoisi
ng
: T
he
und
esi
rab
le
e
lem
ents o
f
t
he sc
ans were
r
e
m
ov
ed
m
anu
al
ly
f
ollo
wed
by
autom
at
ic
r
e
m
ov
a
l
of
t
he
la
r
ge
spi
ke,
wh
ic
h
ca
n
natu
rall
y
aris
e
in
sca
nnin
g.
The
im
ages
w
ere
f
urt
her
sm
oo
t
hed
usi
ng
t
he
Laplaci
an
filt
er
avail
able
in
Ra
pid
f
orm
,
wh
ic
h
rem
ov
es
li
tt
le
sp
ikes.
La
placi
an
sm
oo
thin
g
is
a
co
m
m
on
proce
dure
(e.
g., see
[
2
4
]
).
-
Ho
le
fill
ing
:
L
arg
e
or
sm
all
ho
le
s
m
ay
app
ear
in
3D
face
s
cans.
All
hole
s
wer
e
sp
eci
fie
d
and
fill
ed
us
in
g
Ra
pid
f
orm
p
rogr
am
too
ls.
3.3.
Landm
arks E
xt
r
act
i
on
Bi
olo
gical
la
ndm
ark
s
and
quasi
la
nd
m
arks
are
us
e
d
as
facial
featur
e
s
fo
r
face
rec
ogniti
on
a
nd
gende
r
cl
assifi
cat
ion
.
Bi
ol
ogic
al
la
nd
m
ark
s
are
usual
ly
us
e
d
by
physi
ci
an
s
an
d
sci
entist
s
,
they
are
m
ean
ingful
po
i
nts
that
are
def
i
ned
as
sta
ndar
d
points
on
the
face,
su
c
h
as
the
nasio
n,
inn
er
a
nd
oute
r
canthi
of
th
e
ey
es,
enter
of the
up
per li
p,
ce
nter
of the l
ow
e
r
li
p, m
ou
th
c
orners
[
2
5
]
.
In
t
his
w
ork,
21
facial
la
ndm
ark
s
re
pr
ese
nte
d
by
a
total
of
63
(
x,
y,
a
nd
z)
co
ordinate
values
were
m
anu
al
ly
identifie
d
an
d
rec
orded
for
each
3D
facial
i
m
age.
Af
te
r
that,
a
Gen
e
rali
sed
P
r
ocrustes
Al
gor
it
h
m
(G
P
A
)
is
i
m
plem
ented
to
carry
out
the
r
otati
on,
translat
io
n
an
d
scal
ing
of
the
faces
la
ndm
ark
s
.
O
r
din
a
ry
Proc
ru
ste
s
a
naly
sis
is
a
t
ype
of
sta
ti
sti
cal
sh
ap
e
analy
sis
that
al
ign
s
faces
by
m
ini
m
izing
th
e
Pr
oc
r
us
te
s
distance m
et
ric b
et
ween t
heir
lan
dm
ark
s
[
2
6
]
. T
he
al
go
rithm
st
eps
a
re th
e
foll
ow
i
ng
:
-
Choose a
r
e
fere
nce
face lan
dm
ark
s am
on
g t
he
trai
ning set
instances.
-
Align
al
l
ot
her
faces
la
nd
m
ark
s
on
the
cu
r
r
ent
re
fer
e
nce
by
m
ini
m
izing
t
he
P
r
ocrustes
distance
betwe
e
n
the insta
nces.
-
Determ
ine the
m
ean f
ace
of t
he
trai
ning set
faces
3D lan
dm
ark
s.
-
Ca
lc
ulate
Pr
oc
ru
ste
s
distance
betwee
n
the
m
ean f
ac
e an
d
t
he
r
efe
rence
f
ac
e. If this
distan
ce value
is ab
ove
the s
pec
ifie
d t
hresh
old, set t
he
r
efe
re
nce face
to m
ean f
ace a
nd go t
o
ste
p 2.
Figure
5
il
lustr
at
es the b
iolo
gi
cal
f
ace l
and
m
ark
s t
hat u
se
d
in this r
esea
rc
h.
Figure
6
s
how
s the ef
fec
t
of
a
pply
ing
G
PA
on
t
he
la
nd
m
ark
s
co
ordi
na
te
s,
the
re
d
points
is
the
la
ndm
ark
s
co
ordi
nates
be
f
or
e
a
pp
ly
in
g
the
G
PA,
wh
il
e
the
blu
e
poin
t
after
a
pply
ing
this
al
gort
him
.
Qu
asi
la
nd
m
ark
s
a
re
ac
quired
by
s
uperi
m
po
sin
g
al
l
i
m
ages
on
a
ref
ere
nce
fac
e
and
t
hen
regularisi
ng
al
l
the
i
m
ages
to
en
su
re
t
hat
the
re
su
lt
ing
m
eshes
have
exactl
y t
he n
um
ber
o
f refe
re
nce
face
ver
ti
c
es. Speci
fical
ly
, th
e
f
ollow
i
ng
was
im
ple
m
ent
ed
:
-
Ri
gid
re
gistrati
on.
T
he
idea
w
as
to
use
one
of
dataset
faces
as
facial
te
m
pl
at
e
to
cr
op
t
he
facial
i
m
ages
of
interest
.
T
he
t
e
m
plate
was
bro
ught
int
o
the
vicinit
y
of
t
he
ta
rg
et
s
urfac
e,
by
re
posit
io
ning
the
floati
ng
su
r
face,
u
si
ng
it
erati
ve
cl
os
est
point
(I
CP
)
al
gorithm
[
2
7
]
.
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
Biolo
gical la
ndm
ar
k
vs
quas
i
-
la
ndm
ar
ks f
or 3
D
face
rec
og
niti
on
and ge
nder
..
.
(
Ha
wr
aa
H.
Ab
bas)
4073
-
Non
-
rigi
d
reg
i
strat
ion
.
S
urfa
ces
of
dif
fer
e
nt
subj
ect
s
are
m
at
ched
to
ea
ch
oth
e
r
us
in
g
a
non
-
rigi
d
T
hin
Plat
e Sp
li
ne
(
T
PS)
[
2
8
]
.
-
Re
gu
la
risat
io
n.
An
it
erati
ve
proce
dure
wa
s
us
e
d
to
assi
gn
each
floati
ng
point
of
t
he
tem
plate
to
a
corres
pondin
g
po
i
nt
of
the
ta
rg
et
im
a
ge
by
app
ly
in
g
the
di
sta
nce
weig
ht
ed
k
-
near
est
ne
ighbor
r
ule
[
2
9
]
.
As
a
res
ult, all
faces
wer
e
cr
oppe
d
a
nd c
onve
rted
t
o un
if
or
m
ly
d
ist
ribu
te
d m
eshes
.
Figure
5. The
bio
lo
gical
f
ace
la
nd
m
ark
s
Figure
6
.
Th
e
GPA ef
fect
on
la
nd
m
ark
s
c
oord
i
nates
3.4.
Pri
ncipl
e Co
mpo
ne
nt A
na
l
ys
is (
PC
A)
PCA
is
a
s
ucc
essfu
l
m
et
ho
d
pro
gr
e
ssed
f
or
2D
or
3D
fa
ce
rec
ogniti
on.
I
n
this
al
gori
thm
,
ei
gen
deco
m
po
sit
io
n
of
the
2D
m
at
rix
of
facial
i
m
ages
feat
ur
es
is
p
e
rfor
m
ed
to
obta
in
a
set
of
or
th
ogonal
basis
vecto
rs
.
T
hese
vector
s
m
axim
i
s
e
the
var
ia
nce
of
facial
im
age
data
set
.
Ther
e
f
or
e,
t
he
PCA
is
us
e
d
to
re
du
c
e
the
dim
ension
al
it
y
of
da
ta
set
and
re
pr
ese
nts
face
i
m
ages
i
nto
their
pri
nc
ipal
com
ponen
ts
.
Fo
r
face
reco
gnit
ion
,
t
he
to
p
ei
genvecto
rs
are
kn
own
as
Eigen
faces.
T
hese
ei
ge
nvect
or
s
acc
ount
the
m
os
t
var
ia
ti
on
in
t
he
data
[
30
]
.
Al
l
galle
ry
an
d
pro
be
facial
i
m
a
ges
a
re
pr
oj
ect
ed
on
t
o
the
pri
nciple
s
pace
a
nd
the
ei
genvecto
rs
is
u
se
d
as
traini
ng a
nd test
ing el
e
m
ents
in f
a
ce
recog
niti
on
a
nd
gende
r
cl
assi
ficat
ion
syst
em
.
3.5.
Ma
c
hine Le
ar
ning tech
niqu
es
Au
t
om
atic
face
rec
ogniti
on
and
ge
nder
cl
assifi
cat
ion
a
re
patte
rn
rec
ogni
ti
on
pr
ob
le
m
s
,
w
hich
ar
e
no
t
easy
to
so
lve
pro
blem
s
becau
se
their
non
-
li
nea
rity
.
Especial
ly
,
wh
e
n
face
rec
ogniti
on
a
nd
gend
e
r
cl
assifi
cat
ion
a
pp
li
ed
in
high
dim
ension
al
and
onli
ne
dat
aset
.
The
refo
r
e,
us
i
ng
accu
r
at
e
m
achine
le
arn
i
ng
al
gorithm
s ar
e essentia
l f
or
fa
ce rec
ogniti
on
and g
e
nder
classi
ficat
ion
pro
blem
s.
The
K
-
N
earest
Neig
hbour
m
et
hod
(
K
NN)
is
a
super
vised
le
arn
i
ng
al
gorithm
.
In
this
al
gorithm
,
the
te
ste
d
insta
nce
is
cl
assifi
ed
a
ccordin
g
t
o
K
-
near
est
nei
ghbor
cat
eg
or
ie
s.
This
m
et
ho
d
is
us
e
d
m
ark
ed
ly
in
nu
m
erous
com
pu
te
r
visio
n
a
nd
im
age
proc
essing
a
ppli
cat
ion
s
.
S
up
port
Vecto
r
Ma
chi
ne
(SVM)
a
re
al
so
popula
r
s
uper
vised
le
ar
ning
al
gorithm
and
it
is
us
e
d
f
or
so
l
ving
dif
fer
e
nt
m
achine
le
arn
i
ng
pr
oble
m
s.
SV
M
cl
assifi
er
trie
s
to
fi
nd
t
he
opti
m
al
res
ults
by
m
axi
m
iz
ing
the
m
argi
n
betw
een
t
he
trai
ning
an
d
t
est
ing
instances
[
30
]
.
Th
us
,
in
t
his
researc
h,
these
two
m
achine
le
arn
in
g
al
gori
thm
s
are
us
e
d
for
f
ace
re
c
og
niti
on
and g
e
nder
classi
ficat
ion
.
4.
RESU
LT
S
In
t
he
f
ollow
i
ng,
we
presen
t
the
i
m
ple
m
e
nted
e
xp
e
rim
e
nts
an
d
their
r
esults
.
W
e
al
s
o
pe
rfo
rm
a
com
par
at
ive
re
view
with
the
s
ta
te
-
of
-
the
-
art
relat
ed
resea
rc
h
res
ults
.
Th
e
FRGCv
2
dataset
hav
e
4007
3D
face
i
m
ages
,
60%
of
the
face
s
ha
ve
neu
t
ral
ex
pr
e
ssion
.
I
n
ad
diti
on
to
that
the
s
ubj
ect
s
’
ge
nd
e
r
in
this
datase
t
are
div
ide
d
i
nto
57%
m
a
le
and
4
3%
fem
al
e.
The
face
rec
ogniti
on
ta
s
k
is
m
ult
ic
la
ss
cl
assifi
cat
ion
,
t
her
e
fore
On
e
-
Ag
ai
ns
t
-
All
(1AA)
te
chn
i
qu
e
s
are
us
ed
.
In
a
1AA
appr
oa
ch
the
cl
assifi
cat
ion
res
ult
is
determ
ined
for
each
cl
ass
by
discrim
inati
ng
that
cl
ass
agains
t
the
rem
ai
nin
g
(N
-
1)
cl
asses.
To
acce
ss
the
per
f
or
m
ance
of
the
recog
niti
on
a
nd
ge
nder
cl
ass
ific
at
ion
three
m
easur
es
are
us
ed
acc
ur
a
c
y,
sensiti
vity
,
an
d
s
pecifici
ty
[
3
2
]
.
Table
1
rec
ords
the
rec
ogni
ti
on
accu
racy,
sensiti
vi
ty
,
a
nd
s
pecifici
ty
of
the
both
ty
pes
of
la
nd
m
ark
in
g
(b
i
ologica
l
la
nd
m
ark
and
qu
asi
-
la
nd
m
ark
s
)
,
wh
il
e
Table
2
reco
r
ds
th
e
gender
cl
as
sific
at
ion
accu
racy,
sensiti
vity
, and
sp
eci
fici
ty
of
t
he
sam
e ty
pes
of lan
dm
ark
s.
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.
9
, N
o.
5
,
Oct
ober
20
19
:
4069
-
4076
4074
Table
1.
T
he
perf
or
m
ance of
face
rec
ogniti
on syst
em
Land
m
ark t
y
p
e
Clas
sif
icatio
n
algor
h
ti
m
KNN
SVM
Bio
lo
g
ical land
m
a
rks
Accurac
y
8
8
.4%
94%
sen
sitiv
ity
78%
90%
sp
ecif
icity
9
6
.5%
99%
Qu
asi lan
d
m
arks
Accurac
y
90%
98%
sen
sitiv
ity
80%
98%
sp
ecif
icity
96%
99%
Table
2
. T
he
perf
or
m
ance of
gende
r
cl
assifi
cat
ion
syst
em
Land
m
ark t
y
p
e
Clas
sif
icatio
n
algor
h
ti
m
KNN
SVM
Bio
lo
g
ical and
m
a
r
k
s
Accurac
y
85%
90%
sen
sitiv
ity
80%
88%
sp
ecif
icity
82%
87%
Qu
asi lan
d
m
arks
Accurac
y
87%
92%
sen
sitiv
ity
84%
90%
sp
ecif
icity
88%
91%
5.
DISCU
SIO
N
The
e
xp
e
rim
ental
wo
r
k
i
n
the p
rese
nt
stu
dy
a
i
m
s
to
determ
i
ne
wh
ic
h
facial
la
nd
m
ark
s w
e
re
the
m
os
t
eff
ect
ive
feat
ur
es
i
n
a
uto
m
at
ic
face
recogn
it
io
n
a
nd
ge
nd
e
r
cl
assifi
ca
ti
on
.
F
or
face
recog
niti
on
,
t
he
best
perform
ance
was
achieve
d
w
hen
the
qu
a
si
-
la
nd
m
ark
s
are
us
e
d,
as
il
lustrate
d
in
Table
1
,
but
the
bi
ol
og
ic
a
l
la
nd
m
ark
s
acc
ur
acy
in
face
recog
niti
on
was
no
t
far
fro
m
the
qu
a
si
-
la
nd
m
ark
s
accuracy.
The
sam
e
ob
s
er
vation
is
reache
d
from
a
naly
sing
t
he
ge
nd
e
r
cl
assifi
c
at
ion
re
su
lt
s
in
Table
2.
T
hat
m
eans
the
bio
l
og
ic
al
la
nd
m
ark
s
ca
n
be
us
e
d
as
featur
es
for
accu
rate
an
d
high
-
perform
ance
face
r
ec
ogniti
on
a
nd
gende
r
cl
assifi
cat
ion
s
yst
e
m
.
Table
1
an
d
Table
2
resu
lt
s
show
the
perform
ance
of
S
VM
cl
assifi
er
is
bette
r
than
KNN
cl
assifi
er
in
face
rec
ogni
ti
on
an
d
ge
nd
er
cl
assifi
cat
ion
.
As
sho
wn
i
n
Table
3,
t
he
pr
opos
e
d
a
pp
ro
a
ch
f
or
face
rec
ogniti
on
outpe
rfor
m
s
the
sta
te
-
of
-
the
-
art
w
ork,
w
hich
co
nducte
d
on
the
sam
e
dataset
.
Table
4
r
ecords
gende
r
cl
assifi
cat
ion
acc
ur
ac
ie
s
com
par
ed
to
earli
er
m
et
hods
im
ple
m
e
nted
on
FR
G
CV2
an
d
FRGCV
1
dataset
s
Table
3.
C
om
par
iso
n wit
h
sta
t
e
of the
art m
eth
ods
on FR
GC
V2
Metho
d
s
Accurac
y
F.
Dan
i
y
al
[
3
3
]
9
7
.2%
Kak
ad
iaris
[
3
4
]
9
7
.2%
D.
Hu
an
g
[
3
5
]
97%
Co
o
k
[
3
6
]
9
2
.9%
Table
4.
Stat
e o
f
the a
rt g
e
nd
er classi
ficat
io
n resu
lt
s
Metho
d
s
Dataset
Accurac
y
Laho
u
cin
e
[
3
7
]
FRGC
V2
86%
J.W
u
[
3
8
]
FRGC
V1
6
9
.7%
X.
Lu,
H
[
3
9
]
FRGC
V1
8
5
.4%
6.
CONCL
US
I
O
N
This
pa
pe
r
pr
opose
s
a
no
vel
c
om
par
ison
b
et
ween
t
he
qu
asi
and
bi
ologica
l
la
nd
m
ark
s
f
or
an
ef
fecti
ve
gende
r
cl
assifi
cat
ion
an
d
fac
e
reco
gnit
io
n
syst
e
m
.
Twen
ty
-
one
bio
l
og
ic
al
la
nd
m
ark
s
can
be
us
e
d
as
f
eat
ur
es
for
high
-
pe
rform
ance
gender
cl
assifi
cat
ion
and
fac
e
recog
ni
ti
on
syst
e
m
.
We
can
noti
ce
that
the
per
f
or
m
ance
resu
lt
s
of
our
3D
face
rec
ogni
ti
on
a
nd
ge
nd
e
r
cl
assifi
cat
ion
syst
e
m
,
ou
tpe
r
form
the
reo
r
de
red
res
ults
of
m
any
oth
e
r
pu
blishe
d
res
earc
hes
on
FR
GCV
2
da
ta
set
.
In
ad
diti
on
t
o
that
our
syst
e
m
is
easier
to
im
ple
m
e
nt
tha
n
oth
e
r
sta
te
s
of
the
art
face
rec
ogniti
on
an
d
ge
nd
e
r
cl
assifi
c
at
ion
syst
e
m
s.
In
fu
t
ur
e
,
w
e
a
re
pla
nn
i
ng
to
exten
d
our
a
ppro
ac
h
t
o
pe
rs
on
et
hn
i
c
cl
assifi
cat
ion.
I
n
this
resea
r
ch
21
la
ndm
ark
s
are
us
e
d,
te
sti
ng
the
perf
orm
ance
of
diff
e
re
nt
num
ber
s
of
bio
l
ogic
al
la
nd
m
arks
in
face
rec
ogniti
on
a
nd
gender
cl
assifi
cat
ion,
w
hich
will
a
lso
be
consi
der
e
d
i
n f
uture
work.
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
Biolo
gical la
ndm
ar
k
vs
quas
i
-
la
ndm
ar
ks f
or 3
D
face
rec
og
niti
on
and ge
nder
..
.
(
Ha
wr
aa
H.
Ab
bas)
4075
REFERE
NCE
S
[1]
R.
Jafri
and
H.
R
.
Arabni
a,
“
A surve
y
of
face
r
ec
o
gnit
ion te
chni
qu
es,
”
J
IPS
,
vo
l. 5, pp. 41
-
68,
2009
.
[2]
B.
F.
Klar
e
,
et
al
.
,
“
Face
re
cog
nit
ion
per
form
a
nce
:
Rol
e
of
de
m
ogra
phic
infor
m
at
ion,
”
I
EE
E
Tr
ansacti
ons
on
Information
For
ensic
s and
S
ec
ur
it
y
,
vol
.
7
,
pp
.
17
89
-
1801,
2012
.
[3]
Z.
Saquib,
e
t
al.
,
“
A
survey
on
aut
om
at
ic
spea
ker
rec
ogni
ti
on
s
y
stems
,
”
in
Sig
nal
Proce
ss
ing
and
Mult
imedi
a
,
Ed
:
Springer
,
pp
.
134
-
145
,
2010
.
[4]
D.
Malt
on
i,
et al
.
,
“
Handbook
of fingerpri
n
t
r
ec
og
nit
ion
,”
Spring
er
Scie
n
ce &
Busi
ness Medi
a, 200
9.
[5]
K.
W
.
Bow
y
er
a
nd
M.
J.
Burge
,
“
Handbook
of
ir
is r
ec
ogn
it
ion
,”
Springer,
2016
.
[6]
L.
Dorm
ehl
,
“
Faci
al
r
ec
ogni
ti
on:
is
the
t
ec
hnolog
y
ta
k
i
ng
awa
y
y
our
ide
nt
ity?
”
20
14
.
Avai
la
bl
e:
htt
ps://
ww
w.t
h
e
guar
dia
n
.
com/t
e
chnol
og
y
/2014/m
a
y
/04/
f
acia
l
-
r
e
cogni
ti
on
-
techno
log
y
-
ide
nt
ity
-
te
s
co
-
et
h
ic
a
l
-
issues
.
[7]
H.
Abbas,
et
al
.
,
“
A
3D
m
orpho
m
et
ric
per
spe
cti
ve
for
facial
gen
der
anal
y
s
is
and
cl
assifi
ca
t
ion
usi
ng
ge
odesic
p
at
h
cur
vat
ur
e
f
ea
tur
e
s,
”
Computat
ion
al
V
isual
Me
d
ia
,
vol. 4, pp. 17
-
3
2,
Mar
2018.
[8]
A.
J.
O
'
too
le,
e
t
al.
,
“
Sex
class
ifi
cation
is
be
tt
er
with
thr
ee
-
dime
nsional
h
ea
d
str
uct
ure
tha
n
wi
th
image
intensit
y
informati
on,
”
Per
ce
pti
on
,
vol
.
26
,
pp
.
75
-
84
,
199
7.
[9]
R.
Enc
iso,
e
t
al.
,
“T
hre
e
-
d
imen
sional
hea
d
an
t
hropom
et
ric
anal
y
sis,
”
in
Me
di
c
al
Imaging
2003:
Vi
sualizati
on
,
Image
-
Guided P
roce
dures,
and
Display
,
pp
.
590
-
598
,
2003
.
[10]
P.
Ham
m
ond,
et
al.
,
“
3D
an
aly
si
s
of
facia
l
m
orpholog
y
,
”
Ame
ric
an
journal
of
medic
a
l
gen
et
i
cs
Part
A
,
vol
.
126
,
pp.
3
39
-
348
,
20
04.
[11]
M.
C.
Ru
iz
and
J.
Il
li
ngworth
.
“
Autom
at
ic
la
nd
m
ark
ing
of
fa
ces
in
3D
-
ALF<sup>3D</
sup>
,”
IET
Confe
ren
c
e
Proce
ed
ings
,
pp
.
41
-
46
,
2008
.
Available:
h
tt
ps:
//
di
git
al
-
li
br
ar
y
.
th
eiet.
org
/c
ont
ent
/
co
nfe
ren
c
es/10.
10
49/c
p_20080280
[12]
L.
More
cro
ft
,
“
A Sta
ti
st
i
cal
Ap
proa
ch to
Fa
ci
a
l Ident
if
ic
a
ti
on
,
”
Univer
sit
y
of
Sheffield, 2009.
[13]
N.
A
y
ne
chi
,
et
al
.
,
“
Acc
ura
c
y
a
nd
pre
ci
sion
of
a
3D
ant
hropom
et
ri
c
facia
l
ana
l
y
sis
with
and
w
it
hout
l
andmark
la
be
li
ng
b
efo
re
i
m
age
a
cqui
siti
on
,
”
The
Ang
le Or
thodont
ist
,
vol. 8
1,
pp
.
24
5
-
252
,
2011.
[14]
M.
L
.
Z
el
di
tc
h
,
e
t
a
l.
,
“
Geom
et
ri
c
m
orphom
et
ric
s
for
biol
og
ists:
a
primer
,”
Aca
d
e
m
ic
Press
,
2012.
[15]
T
.
Gor,
et
al.,
“
Thre
e
-
d
imensional
compari
son
of
fac
ial
m
orpholog
y
in
whit
e
popula
ti
ons
in
Budape
st,
Hung
ar
y
,
and
Hous
ton, T
e
xas
,”
Ame
rican
Journal
of
Or
thodont
ic
s and
Den
tof
aci
a
l
Or
thopedics
,
vo
l.
137
,
pp
.
424
-
432
,
2010
.
[16]
C
.
H
.
Kau
,
et
a
l.
,
“
Us
e
of
3
-
di
m
ensiona
l
surfa
ce
a
cqui
sit
ion
t
o
stud
y
f
ac
i
al
m
orpholog
y
in
five
popul
at
ions
,”
Ame
rican
Journal
of
Or
thodont
i
cs
and
Dent
ofaci
al
Or
thope
di
cs
,
vol.
137
,
pp
.
S5
6
-
e1,
2010
.
[17]
Y.
Hu,
et
al.
,
“
A
fusion
-
base
d
m
et
hod
for
3D
fac
i
al
g
end
er
c
la
ss
ifi
c
ation
,
”
in
Computer
and
automati
o
n
engi
ne
ering
(
ICCAE
)
,
2010
The
2nd
Inte
rnat
iona
l
Conf
ere
nce on
,
pp.
369
-
372
,
20
10
.
[18]
K.
W
.
Bowy
er,
et
al.
,
“
A
surve
y
of
appr
oac
hes
a
nd
cha
llenges
in
3D
and
m
ult
i
-
m
odal
3D+
2D
f
ac
e
re
cognition
,”
Computer
vi
sion
and
image
unde
rs
tandi
ng
,
vol
.
1
01,
pp
.
1
-
15
,
20
06.
[19]
H.
Drira
,
et
al
.
,
“
Pos
e
and
expr
ession
-
inva
ria
n
t
3d
fac
e
re
cogni
t
i
o
n
using
el
asti
c
r
adi
a
l
cur
ves,
”
in
Brit
ish
machi
ne
vi
sion
conf
ere
n
c
e
,
pp
.
1
-
11
,
2010
.
[20]
V
.
Bla
n
z
and
T.
Vett
er
,
“
Face
r
e
cogni
ti
on
base
d
o
n
fit
t
ing
a
3D
m
orpha
ble
m
ode
l,
”
IEEE
Tr
ansacti
ons
on
pat
te
r
n
analy
sis and
ma
chi
ne
intelligen
c
e
,
vo
l. 25, pp. 10
63
-
1074,
2003
.
[21]
P.
Pa
y
san
,
et
al
.
,
“
A
3D
fac
e
m
o
del
for
po
se
and
il
lumination
inv
ari
an
t
face
re
cog
nit
ion,
”
in
Adv
a
nce
d
v
ide
o
and
signal
based
sur
ve
i
ll
anc
e, 2009.
AV
SS'09.
Sixth
I
EE
E
Inte
rnat
ion
al
Conf
ere
nce o
n
,
pp
.
296
-
301
,
2009
.
[22]
H.
Li
,
et
a
l.
,
“
Learni
ng
weighted
sparse
rep
rese
nt
at
ion
of
en
code
d
fac
i
al
norm
al
in
f
orm
at
ion
for
ex
pre
ss
ion
-
robust
3D fa
c
e
re
cognit
ion,
”
in
Bi
ome
tri
cs
(
IJCB
)
,
2011
Inte
rnational
Jo
i
nt
Conf
ere
nce o
n
,
pp
.
1
-
7
,
2011
.
[23]
P.
J.
Phill
ips,
et
a
l.
,
“
Overvi
e
w
of
the
fa
ce
rec
ogni
ti
on
gra
nd
cha
l
le
nge
,
”
in
Computer
vis
ion
and
patt
er
n
rec
ognit
ion
,
200
5.
CV
PR 2005. I
EE
E
compute
r
society
conference
on
,
pp.
947
-
95
4
,
2005
.
[24]
O.
Sorkine,
e
t
al
.
,
“
La
pl
ac
i
an
surf
ac
e
ed
it
ing
,
”
in
Proceedi
n
gs
of
the
2004
Eurographics/A
CM
SIGG
RA
P
H
sympos
ium
on
Geomet
ry
proc
essing
,
pp
.
175
-
184
,
2004
.
[25]
J.
Djordje
v
ic,
e
t
al
.
,
“
Gene
t
ic
an
d
envi
ronm
ental
cont
ribu
ti
ons
to
facia
l
m
orphologi
cal
var
ia
t
ion:
a
3D
population
-
base
d
twin
stud
y,
”
Pl
oS
one
,
vol
.
11,
p
p
.
e016225
0,
2016
.
[26]
J.
C.
Gow
er,
“
G
ene
ra
li
z
ed
pro
cr
ustes
anal
y
s
is,
”
Psyc
hometri
ka
,
vol.
40
,
pp
.
33
-
5
1,
1975
.
[27]
B.
B.
Am
or,
et
al
.
,
“
3D
Face
rec
ognit
i
on
b
y
IC
P
-
base
d
shape
m
at
chi
ng,
”
LIRI
S
Lab,
Lyon
Re
search
Cent
er
f
or
Images and
Int
ellige
nt
Informatio
n
Syste
ms
,
UMR
,
vol
.
5205
,
2005
.
[28]
G.
Donato
and
S.
Bel
ongie,
“
Approxim
at
e
thi
n
pla
te
spl
ine
m
appi
ngs,
”
in
Eu
ropean
conf
ere
n
ce
on
comput
er
vi
sion
,
pp.
21
-
31
,
2002
.
[29]
S.
A.
Dudani
,
“
The
dista
n
ce
-
weight
ed
k
-
n
ea
r
est
-
nei
ghbor
rul
e,
”
IEEE
Tr
ansacti
ons
on
S
yst
ems,
Man,
and
Cybe
rnet
ic
s
,
pp.
325
-
327,
1976
.
[30]
K.
Sett
,
et
a
l.
,
“
Face
rec
ogni
tion
using
fusion
of
spati
al
and
te
m
pora
l
fe
at
ur
e
s,
”
in
2018
Em
erging
Tr
ends
in
El
e
ct
ronic
Dev
ices and
Computat
ional
Te
chni
que
s (
EDCT
)
,
pp.
1
-
4
,
2018
.
[31]
A.
G
.
Ba
y
din
,
e
t
al.
,
“
Autom
at
i
c
diffe
r
entiation
in
m
ac
hine
lea
rning:
A
surve
y,
”
Journal
of
M
achi
ne
Learning
Re
search
,
vol
.
1
8,
pp
.
1
-
43
,
201
8.
[32]
W
.
Zhu,
et
a
l.
,
“
Sensi
ti
vity
,
spec
ifi
cit
y
,
ac
cur
acy
,
associa
te
d
conf
ide
nc
e
interva
l
a
nd
ROC
ana
l
y
si
s
with
pra
ct
i
ca
l
SA
S
implementa
ti
ons,
”
NESUG
proce
edi
ngs:
h
e
alt
h
care
and
lif
e
scie
n
ce
s,
Balti
mor
e,
Maryland
,
vol.
19
,
p
p.
67,
2010.
[33]
F.
Dani
y
a
l,
et
al.
,
“
Com
pac
t
sign
at
ure
s
for
3D
fa
c
e
re
cogni
t
ion
un
der
var
y
i
ng
expr
essions,
”
in
Ad
v
ance
d
Vi
deo
and
Signal
Based
Su
rve
il
lan
ce,
2009
.
AV
SS'09
.
S
ix
th
I
EE
E
Inte
rnat
ion
al
Conf
ere
nce o
n
,
pp
.
302
-
307
,
2009
.
[34]
I.
A.
Kaka
di
ari
s
,
e
t
al
.
,
“
Thr
ee
-
dimensional
f
ace
re
cogni
t
ion
in
the
pre
sen
ce
o
f
facia
l
expr
essions:
An
annotated
def
orm
abl
e
m
o
del
appr
o
ac
h
,
”
IEE
E
Tr
ansac
ti
ons
on
pattern
analy
sis
and
machine
in
tell
ige
nc
e
,
vo
l.
29
,
pp.
640
-
649
,
20
07.
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.
9
, N
o.
5
,
Oct
ober
20
19
:
4069
-
4076
4076
[35]
D.
Huang,
et
al.
,
“
A
novel
g
eo
m
et
ric
f
ac
i
al
rep
rese
ntation
b
ase
d
on
m
ult
i
-
sc
al
e
extende
d
loc
a
l
bina
r
y
p
at
t
ern
s,
”
in
Aut
omat
ic
Fa
ce
&
Gesture
Re
cognition
and
Workshops
(
FG
2011)
,
2011
IE
EE
Int
ernati
ona
l
Confe
ren
ce
on
,
pp.
1
-
7
,
2011
.
[36]
J.
A.
Cook
,
et
al.
,
“
3D
face
rec
ogn
it
ion
using
log
-
g
abor
t
emplat
es,
”
Proc
.
Bri
ti
sh
Ma
chine
V
ision
Conf
.
Edi
nborugh,
S
co
tl
and
,
2006.
[37]
L.
Ba
ll
ih
i,
et
al.
,
“
Boosting
3
-
D
-
geometri
c
featur
es
for
eff
i
cient
f
ac
e
rec
ogn
it
ion
and
gender
class
ifi
c
at
ion
,
”
I
EEE
Tr
ansacti
ons on Inform
ati
on
For
ensic
s and
S
ec
ur
it
y
,
vol
.
7
,
pp
.
17
66
-
1779,
2012
.
[38]
X.
Lu
,
e
t al
.
,
“
Multi
m
odal Fac
i
al
Gende
r and
E
thn
ic
ity
Ide
n
ti
fi
cati
on,
”
B
erlin,
Hei
del
ber
g
,
pp
.
554
-
561
,
2005
.
[39]
W
.
A.
P.
S.
J
.
Wu
and
E.
R.
Hanc
o
ck,
“
Gende
r
Cla
ss
ifi
cation
using
Shape
from
Shading,
”
BM
V
C
,
pp.
499
-
508
,
20
07.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Ha
w
raa
H.
Ab
bas
rec
ei
ved
th
e
PhD
in
computer
engi
n
ee
ring
fr
om
Cardi
ff
Univer
sit
y
,
UK
,
2017
.
She
recei
ved
her
B.
Sc.
degr
ee
in
computer
engi
n
ee
ring
from
Bag
hdad
Univer
si
t
y
,
Ira
q
and
M.Sc
.
degr
ee
in
compu
te
r
engi
n
ee
ring
f
rom
Baghda
d
U
nive
rsit
y
,
Ir
aq.
Her
rese
arc
h
intere
sts
inc
lud
e
3D
fac
e
m
odeling,
cl
assifi
ca
t
ion
of
fac
i
al
tr
aits,
i
m
age
proc
essin
g,
computer
n
etw
ork
design,
an
d
gene
t
ic
associati
ons.
Ammar
A.
Altame
em
i
was
bo
rn
in
Basra
h,
Ir
aq.
He
r
ec
e
ive
d
the
B.
Sc
.
and
M.Sc.
degr
ee
s
i
n
computer
eng
in
ee
ring
f
rom
Al
-
Nahra
in
Univ
ersity
,
B
aghda
d
in
1999
and
2002
respe
ctive
l
y
.
He
rec
e
ive
d
the
Ph
D i
n
computer
engi
nee
r
ing
from
The
Univer
sit
y
o
f
Quee
nsland,
Br
isbane
,
Aus
tra
l
ia
in
2013.
He
is
a
lectur
er
in
Univer
sit
y
of
Ker
bal
a
sin
ce
2002
.
His
rese
arc
h
i
nte
rests
in
cl
ud
e
computer
har
dw
are
,
m
ic
roproc
e
ss
ors
arc
hit
e
ct
ur
e,
embedde
d
s
ystems
design,
di
git
al
image
and
signal
proc
essin
g.
Curre
ntly
,
H
e
is
a
per
m
ena
nt
profe
ss
iona
l
staff
of
Biom
edi
cal
Engi
ne
eri
ng
Depa
rtment in
U
nive
rsit
y
of
Ker
bal
a
.
Hame
ed
R.
Fa
r
han
rec
ei
ved
hi
s
B.
Sc.
,
M.Sc
.
a
nd
Ph.D.
degr
ees
in
el
ectroni
c
e
ngine
er
ing
from
Univer
sit
y
of
T
e
chnol
og
y
,
Bagh
dad,
Ira
q
,
in
1986,
2011
and
2018
respe
ct
iv
ely
.
He
is
cur
ren
tly
a
le
c
ture
r
at
el
e
ct
r
ic
a
l
and
el
e
ct
ron
ic
engi
n
ee
ring
d
epa
rtment
,
colle
ge
of
engi
neerin
g,
Univer
sit
y
of
Kerba
la.
His
gr
e
at
in
te
r
ests
inclu
de
Digital
Elec
tr
onic
s,
Im
age
Proce
ss
ing,
Fa
ce
Rec
ognition,
an
d
Com
pute
r
vision
.
Evaluation Warning : The document was created with Spire.PDF for Python.