Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
Vol.
3, No. 6, Decem
ber
2013, pp. 863~
874
I
S
SN
: 208
8-8
7
0
8
8
63
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Faci
al I
m
age Verificati
on a
nd Quality Assessm
ent System -
FaceIVQA
Omidio
ra
E.
O.
1
,
Ola
b
i
y
isi S. O.
1
, Oj
o J. A.
2
, Abay
omi-Alli
A.
3
, Abayo
mi-Alli O.
3
and
Erameh K.
B.
4
1Department of Computer
Scien
ce and
Engineering, Ladoke Akin
tola Universi
ty
o
f
Technolog
y
,
O
gbomoso, Nigeria
2
Department
of
Electronic and Electrical
Engin
e
ering,
Ladok
e Ak
intola Univers
ity of Techno
log
y
,
Ogbomoso, Nigeria
3
Department of
Computer Scien
ce, Feder
a
l Univ
ersity
of
Ag
ricu
lture, Abeokuta,
Nigeria
4
Departm
e
nt of
Ele
c
tri
cal
and
E
l
ectron
i
c
Engin
e
ering, Univ
ersity
of Benin
,
B
e
nin, Nigeria
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Oct 3, 2013
Rev
i
sed
No
v 2, 201
3
Accepted Nov 12, 2013
Although sever
a
l techniqu
es have been
propos
ed for pr
edictin
g biometric
s
y
s
t
em
perform
a
n
ce us
ing qu
ali
t
y
v
a
lues
, m
a
n
y
of the r
e
s
ear
ch
works
were
based on no-reference asse
ssment techniq
u
e using a single quality
attr
ibute
m
eas
ured direc
t
l
y
from
the d
a
ta.
Thes
e t
echn
i
ques
have pro
v
ed to be
inappropri
a
te fo
r faci
al ver
i
fic
a
tion s
cena
r
ios
a
nd ineffi
cien
t b
ecaus
e
no
s
i
ngle quali
t
y
att
r
ibute c
a
n s
u
ffici
ent m
eas
ure the
qualit
y of a fa
cia
l
im
age. In
this
res
e
a
r
ch w
o
rk, a f
a
c
i
al
i
m
a
ge
verif
i
cation and
quality a
sse
ssme
n
t
framework (FaceIVQA) was developed
.
Different algorithms and methods
were implemented in FaceIVQA to extr
act
the f
acen
ess, pose, illumination
,
contrast and
similarity
qu
ality attr
ibutes usin
g an objective
full-ref
e
ren
c
e
im
age quali
t
y
as
s
e
s
s
m
e
nt approach
. S
t
ru
ctured im
age
verific
a
tion
experiments wer
e
condu
cted
on
the survei
llan
c
e camera (SCfa
ce) datab
a
se
to
collect
individu
al quality
scor
es
and algor
ith
m matching scores from
FaceIVQA using three recognition algorit
hms
namely
pr
incip
a
l component
anal
ys
is
(P
CA), linear d
i
s
c
rim
i
nant an
al
ys
is
(LDA) and a com
m
e
rcia
l
recognition SDK. FaceIVQA p
r
oduced accura
te and consistent facial image
a
sse
ssme
n
t da
ta
. The
Re
sult shows tha
t
it ac
c
u
rate
ly
a
ssigns qua
lity
sc
ore
s
to
probe image samples. The resulting qua
lity
s
c
or
e can be assigned to images
captur
e
d for
enr
o
lm
ent or re
cog
n
ition
and
c
a
n b
e
used as
an inp
u
t to qua
lit
y-
driven b
i
ometric fusion s
y
s
t
ems.
Keyword:
Algo
rith
m
s
Aut
h
entication
Facial
Im
age
Qu
ality
Reco
gn
itio
n
Verificatio
n
Copyright ©
201
3 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Ab
ay
o
m
i-Alli A.
Depa
rt
m
e
nt
of
C
o
m
put
er Sci
e
nce,
Fed
e
ral Un
iv
ersity o
f
Agricu
lt
u
r
e,
Abe
o
kuta, Ogun State,
Nige
ri
a
Em
a
il: ab
ayo
m
iallia@u
n
aab.ed
u
.ng
1.
INTRODUCTION
Im
ag
e q
u
a
lity is a ch
aracteristic o
f
an
imag
e th
at
m
e
asu
r
es th
e perceiv
e
d
im
ag
e d
e
grad
atio
n
;
t
y
pi
cal
l
y
, co
m
p
are
d
t
o
an i
d
eal
or per
f
ect
im
age [1]
.
Im
agi
n
g sy
st
em
s
m
a
y i
n
t
r
od
uc
e som
e
am
ount
s of
d
i
sto
r
tion
o
r
artifacts in
th
e sig
n
a
l, so
th
e
quality asse
ss
m
e
n
t
is an
im
p
o
r
tan
t
p
r
ob
lem
.
T
h
e prim
ary g
o
a
l o
f
i
m
ag
e qu
ality
assessm
en
t is to
sup
p
l
y t
h
e
q
u
a
lity metrics th
at can
p
r
ed
ict p
e
rceiv
e
d im
ag
e
qu
ality
au
to
m
a
tical
ly.
By d
e
fin
i
ng
i
m
ag
e q
u
a
lity in
term
s o
f
a dev
i
atio
n
fro
m
t
h
e id
eal situ
atio
n, qu
ality
m
e
asu
r
es
becom
e
techni
cal in the sens
e that
th
ey can b
e
obj
ectiv
ely d
e
term
in
ed
in term
s o
f
d
e
v
i
atio
n
s
fro
m
th
e id
eal
m
o
d
e
ls. Im
ag
e
q
u
a
lity can
, ho
wev
e
r, also
be related
to
th
e su
bj
ectiv
e
p
e
rcep
tio
n
of an
i
m
ag
e, e.g., a hu
m
a
n
looking at a
photogra
p
h. E
x
a
m
ples are how colors a
r
e
re
presente
d in a
black-a
nd-wh
ite
im
age, as
well as in
co
lor im
ag
es,
o
r
th
at th
e redu
ctio
n of im
ag
e qu
ality fro
m
n
o
i
se d
e
p
e
n
d
s
o
n
how th
e no
i
s
e correlates
with
th
e
in
fo
rm
atio
n
the v
i
ewer seek
s
in
th
e im
ag
e rath
er t
h
an its overall streng
th.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Fa
cia
l
Imag
e Verifica
tio
n
and
Qua
lity Assessmen
t S
y
stem -Fa
c
eIVQA
(Ab
a
y
omi-Alli O)
86
4
Im
ag
e q
u
a
lity v
a
l
u
es can
be u
s
ed
i
n
d
i
fferen
t
stag
es of
b
i
o
m
etric ap
p
licatio
n
s
, some o
f
t
h
ese
in
clu
d
e
: enro
ll
men
t
-ph
a
se
q
u
ality assess
m
e
n
t
,
v
e
rifi
cation/id
en
tificatio
n
q
u
a
lity assessmen
t
, pred
ictio
n
of
alg
o
rith
m
failu
re,
qu
ality-b
ased
ad
ap
tation
of th
e pro
c
e
ssing
p
h
a
se an
d mu
lti
m
o
d
a
l b
i
o
m
etric fusion
[2
]
-
[6
].
While steady
progress is
registered each year in
face recognition
research, r
eal worl
d
deploym
e
nt of
bi
om
et
ri
c veri
f
i
cat
i
on sy
st
em
s per
f
o
r
m
far l
e
ss t
h
an t
h
e
resu
lts ob
tain
ed
in
th
e laborato
r
y. Th
e reaso
n
is
si
m
p
le b
i
o
m
e
t
ric syste
m
p
e
rfo
r
m
a
n
ce is d
i
rectly affected
b
y
th
e q
u
ality
o
f
th
e im
ag
es cap
tured
in
real wo
rl
d
an
d tho
s
e
p
r
esen
t in
t
h
e
d
a
tab
a
se. Th
at is,
if th
e
qu
ality o
f
t
h
e
b
i
o
m
etric i
m
ag
es is po
or, th
e
recogn
itio
n
syste
m
’s p
e
rform
a
n
ce is certain
to
b
e
red
u
c
ed
[7
].
V
a
riation
s
d
u
e
to
low
q
u
a
lity i
m
ag
es p
l
aqu
e
all b
i
o
m
etric syste
m
s, su
ch
variab
ility is d
u
e to
a long
list o
f
facto
r
s
wh
ich
in
clud
es facial ex
p
r
essio
n
s
, illu
min
a
tio
n
con
d
ition
s
, p
o
se, presen
ce o
r
absen
ce
of eye
g
l
asses and
facial h
a
irs, o
c
clu
s
ion
,
ag
ing
,
e.t.c. [8
]. Th
ese v
a
riatio
n
s
i
n
i
m
ag
e q
u
a
lity v
a
ry sig
n
i
fi
can
tl
y
d
e
p
e
nd
ing
o
n
wh
ere and
wh
en
t
h
e syst
e
m
o
p
e
rates.
[3
] Po
sit th
at
th
e
q
u
a
lity o
f
b
i
o
m
etric d
a
ta is
ope
rat
i
o
nal
l
y
im
port
a
nt
beca
use i
t
di
rect
l
y
i
n
fl
ue
nces rec
o
g
n
i
t
i
on
per
f
o
rm
ance whi
l
e
[9]
co
ncl
u
de
d t
h
at
a
maj
o
r research area is th
e stu
d
y
o
f
face recog
n
ition
ov
er
a
wid
e
rang
e of q
u
a
lity facto
r
s. Alth
ou
gh
th
ere h
a
s
been a signifi
cant im
prove
ment in
face recognition
pe
rform
a
nce
duri
ng
the past de
cade,
it
is
still
below
acceptable le
vels for
use i
n
many app
lications [10] [11].
This is
beca
us
e
differe
n
t face rec
ognition al
gorithm
s
are d
e
sign
ed
t
o
b
e
robu
st to
particu
l
ar su
bsets o
f
th
es
e
factors.
Hence, a
h
i
g
h
qu
ality i
m
a
g
e for on
e al
g
o
rith
m
is not necess
a
rily of the sam
e
quality for a
not
her. Th
e
r
efore
,
quality shoul
d be learne
d for a s
p
ecifi
c face
match
i
n
g
al
g
o
rith
m
[12
]
. Th
e
p
e
rform
a
n
ce of a
facial reco
gn
itio
n
al
go
rithm
is d
i
rectly affected
b
y
th
e qu
ality
of the facial images capture
d by th
e sens
o
r
(p
ro
be
) and t
h
e o
n
e prese
n
t
i
n
t
h
e dat
a
bas
e
(gal
l
e
ry
). Al
t
h
o
u
gh
p
r
i
n
cip
l
ed
q
u
a
lity
m
easu
r
es h
a
v
e
b
een
develo
p
e
d
for fing
erp
r
i
n
t sa
m
p
les lik
e th
e NIST Fin
g
e
rp
rin
t
i
m
ag
e
q
u
a
lity (NFIQ), th
e facial imag
e qu
ality p
r
o
b
l
em
stil
l remain
s o
p
e
n
[12
]
. Th
e
kn
owl
e
d
g
e
of su
ch
bio
m
e
t
ric
i
m
ag
e q
u
a
lity
p
r
i
o
r t
o
recognitio
n
can be
u
s
ed
to im
p
r
o
v
e
th
e
o
p
e
ration
an
d p
e
rform
a
n
ce of th
e system.
Sev
e
ral research
ers
h
a
v
e
m
a
d
e
atte
m
p
ts to
m
easure bi
ometric syste
m
perform
a
nce using im
age
q
u
a
lity assessmen
t
and
p
r
edictio
n
b
u
t
m
a
n
y
o
f
th
es
e research work
s
were b
a
sed on no
-referen
ce
q
u
a
lity
assessm
en
t tec
h
n
i
q
u
e
s an
d
the assessm
en
t e
v
alu
a
tion
is u
s
u
a
lly fo
cu
sed
on
th
e b
i
o
m
etric sa
m
p
les th
e
m
s
e
lv
es,
th
ereb
y u
s
i
n
g
q
u
a
lity m
easu
r
es d
i
rectly calcu
lated
fro
m
t
h
e
d
a
ta, su
ch
as d
e
n
o
i
si
n
g
t
ech
n
i
q
u
e
s [13
]
, the
si
gnal
-
t
o
-
n
oi
se-rat
i
o
[
1
4]
, si
m
i
l
a
ri
ty
surfac
e
anal
y
s
i
s
[15]
,
m
odel
l
i
ng re
cog
n
i
t
i
on si
m
i
lari
t
y
scores [6
]
,
hi
gh
fre
que
ncy
com
p
o
n
e
n
t
s
o
f
di
s
c
ret
e
cosi
ne t
r
ansf
o
r
m
a
ti
on [
16]
,
di
ffe
re
nce
i
n
im
age i
n
t
e
nsi
t
y
[1
7]
and
im
age
act
i
v
i
t
y
estim
at
i
on i
n
b
o
t
h
h
o
ri
z
ont
al
an
d vert
i
cal
di
rect
i
on [
1
8]
. C
ont
r
a
ry
t
o
[1
9]
wh
i
c
h concl
u
d
e
t
h
at
n
o
sin
g
l
e
q
u
a
lity metric can
reliab
l
y
m
easu
r
e
p
e
rform
a
n
ce,
all th
ese tech
n
i
q
u
e
s u
s
ed
o
n
l
y o
n
e
p
r
o
p
e
rty o
f
th
e
b
i
o
m
etric i
m
a
g
e to
assess qu
ality an
d
m
e
a
s
u
r
e
p
e
rform
a
n
ce. Seco
nd
ly, th
ese tech
n
i
qu
es h
a
v
e
proved
to
be
inappropriate for ve
rification scenari
o
s where the perform
a
nce of a rec
ogn
itio
n
algo
rithm is
a fu
n
c
tion o
f
th
e
p
r
ob
e im
ag
e’s
q
u
a
lity wh
en co
m
p
ared
with
t
h
e
g
a
llery im
a
g
e
[20
]
.
Th
is p
a
p
e
r focu
ses on
d
e
v
e
lo
p
i
ng
an
im
a
g
e qu
a
lity fea
t
u
r
e ex
traction syste
m
fo
r fu
ll-referen
c
e
o
b
j
ectiv
e im
ag
e qu
ality assessm
en
t u
s
in
g statistical an
d
geo
m
etric featu
r
es
o
f
th
e faci
al i
m
ag
e. Hence,
a
facial i
m
age ve
rification a
n
d quality
assessment system
(FaceIVQA)
was
de
vel
ope
d t
o
e
x
tract selected
i
m
age
q
u
a
lity featu
r
es th
at will co
rrelate with
th
e v
a
riatio
ns in
th
e p
r
ob
e im
a
g
e and
th
e al
g
o
rith
m
reco
gn
itio
n
score
s
.
2.
R
E
SEARC
H M
ETHOD
The approac
h
for de
veloping FaceIVQA was base
d on
[20], which c
onc
ludes
that for a verification
task, when
a probe
im
age
is com
p
ared a
g
ainst t
h
e
gallery im
age
of th
e claim
e
d
id
en
tity
u
s
i
ng
recogn
itio
n
algo
rith
m
, if th
e p
r
ob
e sam
p
les
are o
f
un
ifo
r
m
l
y h
i
g
h
qu
ality t
h
en
th
e pro
b
e
sa
m
p
le’s q
u
a
lity i
s
su
fficien
t
to
pred
ict alg
o
r
ith
m
’s per
f
o
r
m
a
nce. The m
a
t
c
hi
ng al
go
ri
t
h
m
will p
r
o
d
u
ce a reco
gn
itio
n
score
fo
r a
gi
ve
n
pai
r
o
f
i
m
ages:
Ǎ
,
(1
)
If th
e
reco
gn
itio
n score
is
ab
ov
e a
p
r
ed
efin
ed
thresh
o
l
d
,
th
e
v
e
rificatio
n
task
is co
n
s
i
d
ered to be
successful.
Fac
e
IVQA was
de
veloped to c
o
m
b
ine feature
extraction techniques for fi
ve
quality m
easures of
the face im
ag
es through an
integration of their ge
om
etr
i
c and statis
tical inform
ation.
This approac
h
was
ai
m
e
d
at ex
tractin
g
im
ag
e qu
ality v
a
lu
es t
h
at is ef
fective and
will h
i
gh
ly co
rrelate
with
th
e reco
gn
itio
n
matching sc
ore
s
. The c
o
ncept
of sim
ilarit
y
as m
easure of
facial quality was introduc
e
d
because this re
search
stu
d
y
b
e
liev
e
s
th
at with
ou
t a
su
itab
l
e con
c
ep
tu
alizatio
n
an
d m
easure of t
r
ue si
m
i
l
a
r
ity b
e
tween facial images,
a tru
e
m
easu
r
e
o
f
qu
ality d
i
sparity b
e
tween
a prob
e and
g
a
llery im
ag
e can
no
t b
e
don
e in verificatio
n scenario
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJECE Vol. 3, No. 6, D
ecem
ber 2013
:
863 – 874
8
65
2.
1.
FaceI
VQ
A Im
a
g
e Q
u
al
i
t
y Fe
a
t
ures
The c
o
m
p
lete
FaceIVQA arc
h
itecture
is shown
on fi
gure
1. The
syst
em
w
ill assess the quality of the
facial im
ages using
five
feat
ures
nam
e
ly: faceness, pose
,
cont
rast, illum
i
nation, a
n
d si
m
ilarity. The
m
e
thods
and
al
g
o
ri
t
h
m
fo
r eac
h
qual
i
t
y
feat
ure
i
s
di
scusse
d
bel
o
w:
Facenes
s me
asure
The
facene
ss
measure is a
c
o
m
b
ination
of occlus
ion a
n
d distance
bet
w
een the
eyes (DBE). T
h
e
a
m
ount
of the
face
re
gion a
v
ailable for
re
cognition is
determ
ined by
t
h
e occlusi
o
n
from
non-c
ooperative
subjects due to objects or acc
essories
(e.g sunglasses
,
scarf, m
a
sks, etc)
and the size of the face due t
o
face-
to-cam
era distance (m
easure
d
as distance betwee
n eyes).
Th
u
s
, th
is research
set o
u
t
to
co
m
b
in
e th
e two
qualities as the
facenes
s feature since
bot
h is
depe
ndent
on
the am
ount of
th
e face area that is detected by the
alg
o
rith
m
.
Figure
1. Complete FaceIVQA a
r
chitecture
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
Facial Image Verification and Qu
ality Asse
ssme
n
t Syste
m
-Face
I
VQA
(Ab
a
y
omi-Alli O)
86
6
The face
-t
o
-
ca
m
e
ra di
st
ance
i
s
recom
m
ended by
[
2
1]
t
o
be bet
w
een
1.
2–
2.
5m
i
n
a ty
pi
cal
ph
ot
o
studi
o a
nd
dist
ance-betwee
n-eyes (DBE
)
to be 120 pixels
. DBE
is
inve
rs
ely
related to t
h
e size
of the
face in
an
im
ag
e, th
u
s
it can
b
e
u
s
ed
as a qu
ality esti
mate fo
r th
e su
bj
ect’s d
i
stan
ce fro
m
th
e cam
era [2
2
]
. To
measure the fa
ceness
quality, the face in
the
probe im
age is tracke
d
. If a face
is not
detected then t
h
ere is no
facenes
s m
eas
ure a
n
d an error m
e
ssage
is given. Howeve
r, if a face is
de
tected the
n
the
distance
betwe
e
n t
h
e
ey
es
(DB
E
)
i
s
obt
ai
ne
d wi
t
h
equat
i
o
n 2.
To fi
nd t
h
e di
st
ance
bet
w
een t
h
e t
w
o
ey
es (p
oi
nt
s)
wh
ose
pi
xel
c
o
o
r
di
nat
e
s are
gi
ve
n. Let
,
and
,
b
e
the po
in
ts
represen
tin
g
t
h
e left and
ri
ght eyes
re
spectively. From
the right
angl
e
d
t
r
i
a
ngl
e
,
t
h
e
di
st
ance
b
e
t
w
een t
h
e
poi
nt
s
and
i
s
gi
v
e
n
by
:
2
(2
)
Th
e facen
e
ss qu
ality
in
th
is
research is m
e
a
s
u
r
ed
as
p
e
rcen
tage
of the
di
stance
betwee
n the eyes
(DBE
)
of the
probe image
with
resp
ect to
th
e
stand
a
rd g
a
llery im
ag
e
of t
h
at subj
ect in
th
e
d
a
tab
a
se.
100
(3
)
Pose meas
ure
Pose is a m
a
j
o
r c
ova
riate tha
t
determ
ines the
usa
b
ility of the
face im
age
in
recognition [23]. The
am
ount
of
face
regi
on a
v
ailable for
re
c
o
gnition is
directly affected
by th
e
subject’s pose. A good
quality face
i
m
ag
e
m
a
y
n
o
t
b
e
u
s
efu
l
du
ri
n
g
recog
n
ition
d
u
e
to
sev
e
re po
se v
a
riation
s
.
In
t
h
i
s
researc
h
, t
h
e
opt
i
cal
fl
ow
t
ech
ni
q
u
e
pr
o
pose
d
by
L
u
cas a
n
d
Kana
de i
n
[
2
4]
was
adapt
e
d
wi
t
h
slight m
odifications from
[25] an
d [26]. T
h
e L
u
cas
–Ka
n
ade m
e
thod as
su
m
e
s that the displacem
ent of t
h
e
im
age cont
e
n
t
s
bet
w
ee
n t
w
o
near
by
i
n
st
ant
s
(
fram
e
s) i
s
sm
al
l
and
ap
pr
oxi
m
a
t
e
ly
const
a
nt
wi
t
h
i
n
a
n
e
igh
bor
hoo
d o
f
th
e po
in
t
p
u
nde
r c
o
nsi
d
e
r
at
i
on.
The l
o
cal im
ag
e flow
(vel
ocity) vect
or
m
u
st satisfy
(4
)
Wh
ere
are the p
i
x
e
ls in
si
de th
e
window, and
are
the
pa
rtial
d
e
ri
v
a
tiv
es of th
e i
m
ag
e
wi
th
resp
ect to
p
o
s
ition
x
,
y
and
ti
me
t
, evaluated at the point
and at the
cu
rren
t tim
e. Th
ese equ
a
tio
n
s
can
b
e
written
in
m
a
trix
form
, wh
ere
(5
)
Th
us t
h
e
o
p
t
i
cal
fl
ow e
q
uat
i
o
n
ca
n be
ass
u
m
e
d t
o
ho
ld
for
all
p
i
x
e
ls with
in
a windo
w cen
tered
at
p
.
In
o
r
de
r t
o
t
r
a
c
k t
h
e
face
,
w
e
l
l
-
t
e
xt
ure
d
fac
i
al
feat
ures
wit
h
in
t
h
e targ
et
reg
i
on
wh
ich
is th
e stan
d
a
rd
gallery
i
m
ag
e is first id
en
tified
and
t
h
en th
e co
rresp
ond
ing
o
p
tica
l
flow i
n
eac
h subject
probe
im
age is calculated
using a two-fram
e
gradie
nt-base
d
m
e
thod devel
ope
d by
[24]. T
h
e task of m
a
tching a
face in t
h
e st
anda
rd
g
a
llery im
ag
e
to a
tar
g
et
(p
ro
be
) im
age
in
th
e past
fram
e
1
is g
e
ne
rally
refe
rre
d
to as
a
regi
st
rat
i
o
n
pr
obl
em
. Opt
i
cal
fl
ow i
s
a
regi
st
rat
i
on m
e
t
h
o
d
t
h
at
p
r
ovi
de
s a
m
easure
of
t
h
e ap
pare
nt
m
o
ti
on
within seque
n
c
e
of im
ages. B
a
sed
on t
h
e
position
of th
e
fe
ature
points i
n
each im
age and t
h
e
position
of the
feat
ure
poi
nt
s
(aft
er
t
h
e
t
r
ack
i
ng
p
r
oce
ss)
i
n
t
h
e st
a
n
da
rd
i
m
age, optical flow vect
ors
were calcu
lated
.
Th
is
measu
r
e is
referred
to in
t
h
is res
earch approa
ch as
the
pose
measure
.
M
odi
fi
cat
i
o
n i
n
t
h
i
s
resea
r
ch
st
udy
t
o
L
u
cas
and
Ka
na
de
19
81
al
g
o
ri
t
h
m
w
e
re i
n
t
e
rm
s of:
(a)
Area
of a
pplic
ation to im
age pos
e m
easure
m
ent
(b
)
Use
of
Ga
ussi
a
n
e
r
r
o
r
di
st
ri
bu
t
i
on
rat
h
er
t
h
a
n
t
h
e l
east
s
q
u
a
r
e
d a
p
p
r
oach
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJECE Vol. 3, No. 6, D
ecem
ber 2013
:
863 – 874
8
67
(c)
Inc
o
rporation
of textured faci
al f
eature
e
x
tra
c
tion for face detection.
Table 1 s
u
mmarizes the proce
d
ure to
obtain op
tical flow vectors
for each s
u
bject and their
corres
ponding varying probe face
im
ages
as adapte
d from
[26].
Tabl
e 1. Al
g
o
ri
t
h
m
for obt
ai
ni
ng
o
p
t
i
cal
t
h
e f
l
ow vect
o
r
s.
I
npu
t:
fa
ce i
m
ag
es.
Let
∗
,
,
…
,
,
,
,…,
denote f
a
ce images
.
M repr
esent
s
t
h
e
nu
m
b
er
of
i
m
a
g
es f
o
r e
a
c
h
pers
o
n
, N
repr
esent
s
t
h
e
n
u
m
ber
of
pers
ons
.
Ou
t
p
u
t
: f
a
ce
op
tica
l
flo
w
,
,
…,
,
,
,
…
,
.
1:
f
a
ce
i
m
ages
are
aver
a
g
ed
by
∑∑
(6
)
2:
f
a
ce
i
m
ages
are
n
o
r
m
al
i
z
e
d
by s
u
bt
ract
i
n
g
aver
age
f
r
a
m
e
.
3: for
E
a
ch face image
and
, op
tica
l
flo
w
do
4
:
ca
lcu
l
a
t
e the op
tica
l
flo
w
,
,
…,
,
,
,
…
,
.
5:
en
d
6:
en
d.
So
urce:
[2
6]
Whe
r
e,
denote
optical flow
between face
image
1
,
2
,…,
,
1
,
2
,…,
and
.
,
∑
,
100%
(7
)
Contra
st
and illumina
tio
n
mea
sure
Stru
ct
u
r
al Sim
i
larity in
d
e
x
(SSIM) is an
im
a
g
e qu
ality
m
e
tric. SSIM is com
p
u
t
ed
fo
r t
h
e i
m
ag
e with
resp
ect to
t
h
e referen
ce im
ag
e
.
Th
e referen
c
e i
m
ag
e u
s
u
a
lly
n
eed
s t
o
b
e
o
f
p
e
rfect q
u
ality. Th
is is co
nsisten
t
with
th
e ap
pro
ach of t
h
is stu
d
y
h
e
n
c
e the SSIM ind
e
x was u
s
ed
t
o
o
b
t
ain a
qu
an
t
itativ
e v
a
lu
e
fo
r two
p
a
ram
e
ters n
a
mely u
n
e
v
e
n
illu
m
i
n
a
tio
n
(l
u
m
in
an
ce) an
d
con
t
rast
qu
ality
measu
r
e between
the stan
d
a
rd
gallery im
age
and the ta
rget (probe
) image
. SSIM
can be
use
d
as a benc
hm
ark t
o
c
h
ec
k t
h
e
p
e
rform
a
n
ce of o
t
h
e
r im
ag
e p
r
o
cessi
n
g
al
go
rith
m
s
[18
]
an
d
it is an
im
p
r
ov
em
en
t to
Un
iv
ersal Im
ag
e Qu
ality
In
de
x (
U
IQ
I)
pr
o
pose
d
by
[
27]
.
Th
e S
S
I
M
al
go
ri
t
h
m
se
p
a
rates ou
t the similarity
measu
r
em
en
ts into
three
di
ffe
re
nt
com
pone
nt
s bet
w
ee
n t
h
e t
w
o n
o
n
-
n
egat
i
v
e i
m
age si
gnal
s
:
Lum
i
nance
L(g,p)
, C
ont
ra
st
C(g
,
p)
and
Struct
ural
S(
g,
p)
b
u
t
th
e stru
ctu
r
al
v
a
lu
e is ou
tsid
e t
h
e in
terest o
f
th
is
research.
A use
f
ul m
eas
ure
of face im
a
g
e quality is th
e contra
st of the skin area
of t
h
e face. SSIM
determ
ines
cont
rast
by
t
h
e
st
anda
rd
de
vi
at
i
on
o
f
t
h
e
si
g
n
a
l
fr
om
t
h
e t
w
o
im
ages.
,
(8
)
Hen
c
e, th
e contrast qu
ality
measu
r
em
en
t b
e
t
w
een th
e
standard g
a
llery im
a
g
e
i
g
a
n
d the
probe
im
age
i
p
of
t
h
e
su
bj
ect i
n
th
e datab
a
se is
d
e
noted
b
y
M
C
wh
ich
is equ
i
v
a
lent to
C
(
i
g
, i
p
)
i
n
equat
i
o
n 8.
Variation in ill
um
ination conditions
pose
s
a signi
ficant
problem
for the
face recognition task [28]
[29
]
. Facto
r
s su
ch as illu
m
i
n
a
tio
n
d
i
rection
an
d in
ten
s
ity
of th
e ligh
t
so
urce can sev
e
rel
y
alter th
e appearan
ce
of an individua
l
'
s
face and s
u
bseque
ntly wea
k
en ge
nuine m
a
tch sc
ores
.
Th
e SSIM index
d
e
term
in
es th
e lu
m
i
n
a
n
ce b
e
tween
two
im
ag
es b
y
th
e m
ean
in
tensity o
f
th
eir
si
gnal
s
.
,
µ
µ
µ
µ
(9
)
Sub
s
equ
e
n
tly,
u
n
e
v
e
n illu
min
a
tio
n qu
ality measu
r
e
b
e
tween
t
h
e stand
a
rd
g
a
llery im
ag
e
and t
h
e
probe im
age
of an
y
o
f
t
h
e s
u
bj
ect
i
n
t
h
e
dat
a
base
i
s
de
not
e
d
by
M
L
wh
ich is
eq
u
i
v
a
len
t
to
,
in
equat
i
o
n 9.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Fa
cia
l
Imag
e Verifica
tio
n
and
Qua
lity Assessmen
t S
y
stem -Fa
c
eIVQA
(Ab
a
y
omi-Alli O)
86
8
Similarity me
asure
Si
m
ilarit
y
in
facial reco
gn
itio
n
is
d
e
fin
e
d
as th
e Euclidea
n distance
bet
w
een t
w
o
face
im
ages whe
n
represe
n
ted i
n
a pri
n
cipal c
o
m
ponent
(PCA) feat
ure
vect
or s
p
ace [30]. T
h
is approac
h
is
capa
b
le of
providing
a set of
ge
nerat
i
ng
dim
e
nsions
that
can acc
urately represe
n
t faces. T
h
e
r
e e
x
ist seve
ral m
e
thods for m
eas
uri
ng
the dista
n
ce
be
tween im
ages
and/
or faces; t
h
ey includ
e ta
nge
nt
distance, ge
nerali
zed Haus
dorff
distance and
Eucl
i
d
ea
n di
st
ance.
Am
ong
al
l
t
h
e im
age m
e
t
r
i
c
s, Eucl
i
d
ean
di
st
ance i
s
t
h
e m
o
st
com
m
onl
y
used d
u
e
t
o
i
t
s
si
m
p
licit
y. Howev
e
r, th
e trad
itio
n
a
l Eu
cli
d
ean
d
i
stan
ce
m
e
tric su
ffers fro
m
a h
i
g
h
sen
s
itiv
ity to sm
al
l
def
o
rm
at
i
on b
e
t
w
een i
m
ages an
d
doe
s
not
t
a
ke i
n
t
o
acc
o
unt
t
h
e s
p
at
i
a
l
rel
a
t
i
ons
hi
p
be
t
w
een
pi
xel
s
.
Henc
e
,
[31] presente
d the Im
age Euclidean
Distan
ce (IMED)
metric wh
ich
wa
s
ad
op
te
d
an
d
in
co
rp
or
a
t
ed
in
to
FaceIVQA for facial
simila
rit
y
m
easure.
The
choice of
IMED fo
r the
sim
i
l
a
rity
m
easure i
s
base
d it:
(a)
Ro
bu
stn
e
ss to
sm
a
ll ch
an
g
e
s
b
e
tween
im
ag
es;
(b
)
Si
m
p
licity o
f
co
m
p
u
t
atio
n
;
(c)
Ease o
f
in
corpo
r
ation
in
to
mo
st o
f
th
e im
ag
e reco
gn
ition
tech
n
i
q
u
e
s su
ch as Rad
i
al Bas
i
s Fu
n
c
ti
on
Su
pp
ort
Vect
or M
achi
n
es (R
B
F
-S
VM
s)
,
Pri
n
ci
pal
C
o
m
ponent
A
n
al
y
s
i
s
(PC
A
) a
nd B
a
y
e
si
an
si
m
ilarit
y
.
For
an
M
by
N
im
age in an
MN
dim
e
nsional Euclidean
space
(im
a
ge space),
,
,…
,
will
form
a coordi
nate system
of the im
age spa
ce, whe
r
e
corresponds
to an ideal
poi
nt source
with unit
in
ten
s
ity at l
o
catio
n
,
.
If I
m
ag
e
,
,⋯,
,
w
h
er
e
is
the co
ord
i
n
a
te with
resp
ect
t
o
and the m
e
tric coefficients
;
,
1
,
2
,⋯,
,
ar
e de
fi
ne
d as:
,
,
,
∙
cos
(1
0)
Whe
r
e <,>
is the scalar
product and
is the
angle
betwee
n
and
. The E
u
c
lidean
distance
of t
w
o im
ages
x
,
y is written
b
y
:
,
∑
,
(1
1)
Th
e symmetric
m
a
trix
will
be referred
t
o
as m
e
tric
m
a
trix
. Fo
r im
ag
es of
fix
e
d
size
M
by
N
,
every
MN
th
o
r
d
e
r symm
etric
an
d po
sitiv
e
d
e
fi
n
ite
m
a
trix
G indu
ces a
Eu
clid
ean
d
i
st
an
ce.
If t
h
e metric
co
efficien
ts d
e
p
e
nd
p
r
op
erly
o
n
t
h
e
p
i
x
e
l d
i
stan
ces, the obtain
e
d
Eu
clid
ean
d
i
stan
ce is
in
sen
s
itiv
e t
o
sm
a
l
l
d
e
fo
rm
atio
n
.
Th
e ap
p
eali
n
g prop
erties are
based
on
its
satisfyin
g three con
d
ition
s
[31
]
,
wh
ich
states that:
(a)
The m
e
tric coefficient
depe
nds
o
n
t
h
e
di
s
t
ance bet
w
een
pi
xel
s
and
. Let
f
rep
r
esen
t
th
is
depe
n
d
ency
;
(b
)
f
is con
tin
uou
s, and
dec
r
eases
m
onot
o
n
i
cal
l
y
as
increases;
(c)
The f
u
nct
i
onal
depe
n
d
ency
f
i
s
a uni
ve
rsal
f
unct
i
o
n.
That
i
s
, i
t
i
s
not
f
o
r i
m
ages of a
par
t
i
c
ul
ar si
ze
o
r
reso
lu
tion
.
Fin
a
lly, th
e si
milarit
y
measu
r
e (M
S
) is d
e
fi
n
e
d
as a facial i
m
ag
e q
u
a
lity
m
easu
r
e in
term
s o
f
th
e
sim
ilarity between t
h
e standard
gallery image
and the
probe im
age
of a
pa
rticular subject in the
dat
a
base
.
Ms
,
⁄
, and
(1
2)
2.
2.
Over
al
l
Q
u
al
i
t
y Sc
ore F
u
si
on
An ov
erall-no
rmalized
sco
r
e
is ob
tain
ed b
y
th
e
fu
si
o
n
of
th
e norm
a
l
i
zed
qu
ality sco
r
es
′
using
th
e Su
m
ru
le wh
ich
is si
m
p
ly th
e su
m
o
f
all n
o
r
m
a
lized
q
u
a
lity
m
easu
r
e scores. Th
us a co
m
p
o
s
ite
sco
r
e
k
nown as t
h
e
ov
erall
q
u
a
lity sco
r
e
) i
s
deri
ve
d as:
∑
′
(1
3)
Th
is ov
erall qu
ality sco
r
e (OQS) is ex
p
ected
to
b
e
p
r
ed
ict
i
v
e
o
f
t
h
e con
t
ribu
tio
n
o
f
th
e p
r
o
b
e
im
ag
e to
th
e
perform
a
nce of the
rec
ogn
itio
n
algor
ith
m
s
u
s
ed
.
2.3. FaceI
VQA Recogni
t
ion
Algorithms
FaceIVQA c
o
m
b
ines thre
e rec
o
gnition algor
ithm
s
and ret
u
rns
thei
r recognition score
s
si
m
u
ltaneously
. The
face re
cognition
algorithm
s
used
in Face
IVQA
are PCA
[30]
, LDA [32] and a
co
mmercial reco
gn
itio
n en
g
i
n
e
[33
]
lux
a
nd
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJECE Vol. 3, No. 6, D
ecem
ber 2013
:
863 – 874
8
69
2.4.
Faci
al Ve
rification E
x
periement
A fa
cial im
age ve
rification
expe
rim
e
nt was conducte
d
on Face
IV
QA a
n
d the
face
a
u
thentication
pr
ot
oc
ol
p
r
op
o
s
ed
by
[3
4]
w
a
s ad
opt
e
d
. F
o
l
l
o
wi
n
g
t
h
e
da
y
-
t
i
m
e
and ni
g
h
t
-
t
i
m
e
t
e
st
scenari
o
s 2
,
9
9
0
i
m
ages
from
all 130 s
u
bjects in t
h
e
Scface s
u
rveillance cam
era database [29]
was
utilized. Frontal
m
ug s
hots
of eac
h
su
bj
ect
(130
)
will represen
t
th
e g
a
llery
o
f
kn
own
h
i
gh
qu
ality i
m
a
g
es
wh
ile th
e p
r
ob
e d
a
tab
a
se for
v
e
rification
tri
a
ls will in
clu
d
e th
e 1
3
0
h
i
gh q
u
a
lity i
m
ag
e
s
o
f
each
subject an
d
th
eir
oth
e
r (22
x13
0) i
m
ag
es
with
co
n
s
i
d
erab
le session
and
q
u
a
lity v
a
riatio
n
s
.
Each
sub
j
ect
was enrolled
with
a sing
le h
i
g
h
qu
ality
m
ug
sh
o
t
im
ag
e fo
r th
e g
a
llery d
a
tab
a
se, pro
b
e
i
m
ag
es were tak
e
n
fro
m
th
e 8
su
rv
eillan
ce ca
m
e
ras at 3
d
i
fferen
t
distances: clos
e, m
e
diu
m
and far. Eac
h
s
u
bject’s gallery image was com
p
are
d
(v
eri
f
icatio
n) with
th
e
23
probe
i
m
ag
es o
f
v
a
ry
in
g
qu
ality.
3.
R
E
SU
LTS AN
D ANA
LY
SIS
FaceIVQA wa
s success
f
ully im
ple
m
ented.
Whe
n
tested
a
nd
use
d
for t
h
e
expe
rim
e
nt it
was obse
rve
d
to pe
rform
the following tas
k
s
accurately:
(a)
Accept a
probe
or
gallery im
a
g
e
from
a file or folde
r
direct
ory;
(b
)
Take live
im
ag
es from
the com
puter we
bca
m
;
(c)
Detect face i
n
i
m
ages and ca
rry out
rec
o
gnition;
(d
)
Ex
tract selected
q
u
a
lity features fro
m
th
e prob
e
im
ag
e and
sav
e
th
e d
a
ta on a tab
l
e i
n
d
a
tab
a
se;
(e)
Out
put
rec
o
g
n
i
t
i
on resul
t
s
or
err
o
r
m
e
ssages.
Table 2 sum
m
aries the re
sult of the
verifica
tion e
x
perim
e
nt with
FaceIVQA t
h
rough the
p
e
rform
a
n
ce of th
e th
ree reco
gn
itio
n algo
ri
th
m
s
. Th
e
result was gen
e
rally p
o
o
r
acro
s
s th
e three
recogn
ition
al
go
ri
t
h
m
s
. Th
i
s
i
s
co
nsi
s
t
e
nt
wi
t
h
t
h
e
resul
t
s rep
o
r
t
e
d
by
[2
9]
an
d
[
35]
wh
ose
eval
uat
i
ons
we
re
base
d
on
PCA and
Mace co
rrelatio
n
fi
lter alg
o
r
ith
m
resp
ectiv
ely.
It p
r
o
v
e
d
th
at the lo
w qu
ality o
f
prob
e im
ag
es fro
m
the Sc
face
database
provide
d
a ve
ry
diffic
u
lt test to
the
recognition al
gorithm
s
i
m
plemented in Fac
e
IVQA
al
so. L
u
xan
d
S
DK
ha
d
2,
71
8
fal
s
e re
ject
(
F
R
)
w
h
i
l
e
PC
A
and
LD
A
ha
d
2,
85
0
res
p
ect
i
v
el
y
.
Al
t
h
o
u
g
h
PC
A
an
d LDA seems to
h
a
v
e
th
e sa
m
e
p
e
rform
a
n
ce, PC
A
h
a
d slig
h
tly h
i
gh
er
mean
reco
gn
itio
n sco
r
e
(MRS) th
an
LDA. Fifty-four (54) im
ages
failed-
to-acquire (FTA) beca
use the face de
tection algorithm
could not detect
th
e face in
t
h
e
i
m
ag
es du
e to
ex
trem
ely lo
w q
u
a
lity.
Tab
l
e 2
.
Su
mmary
of v
e
rificat
io
n
exp
e
rim
e
n
t
with
reco
gn
itio
n algo
rith
m
’
s
p
e
rform
a
n
ce
Algor
ith
m
SR
FTA
TA
FR
FA
TR
MRS
L
uxand SDK
2,
936
54
217
2,
718
0
0
0.
083
PCA
2,
936
54
130
2,
805
0
0
0.
072
LDA
2,
936
54
130
2,
805
0
0
0.
067
** Decision threshold = 0.4
SR = Successful Recognition
TA =
True
Accept
FTA =
Failure
to
Acqui
re (f
ailure to
detect f
ace in i
m
a
g
e)
FR = False
Reject
FA = F
a
lse A
ccept
TR =
True Rej
ect
MRS =
Mean Re
c
ognition Score
Fig
u
res
2
-
4 sh
ows
o
t
h
e
r exp
e
rim
e
n
t
al resu
lts su
ch
as th
e effect
o
f
varyin
g
cam
era
qu
ality on
algorithm
perform
a
nce, the
effect
of face
-to-cam
era distance
on algori
thm
perform
ance and the
effect of
face-to-cam
era distance on av
erage
recognition tim
e. In order to re
du
ce t
h
e num
ber of fa
lse rej
ects (FR
)
, the
recogn
itio
n
t
h
resho
l
d
was
set
at 0
.
4
d
u
e
to
t
h
e v
e
ry low
q
u
a
lity o
f
th
e
probe i
m
ag
es.
Fi
gu
re
2 sh
o
w
s t
h
at
cam
era
7 ha
d t
h
e hi
g
h
est
n
u
m
b
er o
f
fai
l
u
re-t
o-ac
qui
re (F
TA
) f
o
l
l
o
we
d
by
ca
m
e
ra 6 whil
e ca
m
e
ras 3, 5 and
9 (frontal
day) ha
d
none
.
It was obse
rved on
figure 3
that Face-to-ca
m
era
di
st
ance ha
d a si
gni
fi
ca
nt
effe
ct
on pe
rform
ance especially at distance 1 (4
.
2
m
)
but
at
dist
ance 2 (2.6m) the
perform
a
nce im
prove
d. This
is consistent with th
e rec
o
mmendations for face im
age data on c
o
ndit
i
ons for
t
a
ki
ng
pi
ct
u
r
es
i
n
[
21]
.
In a
d
di
t
i
on t
o
t
h
i
s
,
cam
e
ra 7 an
d
cam
e
ra 1 f
r
o
n
t
a
l
-
day
ret
u
r
n
s
t
h
e hi
g
h
est
a
n
d l
o
west
avera
g
e rec
o
g
n
i
t
i
on t
i
m
e
of 5.
05 a
nd
1.
82
seco
nds re
spec
t
i
v
el
y
as show
n o
n
fi
g
u
re 4
.
Al
l
t
h
ese resul
t
are
co
nsisten
t
with th
ose repo
rted b
y
[29
]
[35
]
.
Tab
l
e 3
shows
th
at p
o
se im
ag
e q
u
a
lity (QP)
h
a
d
t
h
e
h
i
gh
est co
rrelatio
n
coefficien
t
o
f
R= 0
.
9
3
6
wit
h
Ov
erall Qu
ality Sco
r
es
(OQS) wh
ile on
tab
l
e 4
sim
i
lar
ity q
u
a
lity (QS)
h
a
d
th
e
h
i
gh
est co
rrelatio
n
co
efficien
t
o
f
R=0
.
85
5
with
Al
go
rith
m
Match
i
n
g
Sco
r
es (AMS).
Th
e lu
m
i
n
a
n
ce quality (QL) and
co
n
t
rast
q
u
a
lity (QC
)
had the least c
o
rrelation
coe
f
ficient for OQS and
AMS.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Fa
cia
l
Imag
e Verifica
tio
n
and
Qua
lity Assessmen
t S
y
stem -Fa
c
eIVQA
(Ab
a
y
omi-Alli O)
87
0
Figure
2. Graph s
h
owing the
effect
of
va
rying cam
era qual
ity on algorithm
perform
ance
Figure
3. Graph s
h
owing the
effect
of f
ace
-to-cam
era dista
n
ce
on algorithm
perform
ance
100
150
200
250
300
350
400
Cam
1C
a
m
2C
a
m
3C
a
m
4C
a
m
5C
a
m
6C
a
m
7F
r
o
n
t
a
l
Night
Cam
8
Frontal
Day
Cam
9
No.
Of
Failure
‐
to
‐
Acquire
(FTA)
No.
Successful
Recognition
Varying
I
mage
Camera
No.
of
Successful
Recogniton/No.
of
Failure
‐
to
‐
Acquire
(FTA)
877
901
898
129
130
33
9
12
1
0
Distance
1
(4.2m)
Distance
2
(2.6m)
Distance
3
(1m)
Frontal
Night
Cam
8F
r
o
n
t
a
l
Day
Cam
9
No.
Successful
Recognition
No.
Of
Failure
‐
to
‐
Acquire
(FTA)
No.
of
Successful
Recogniton/No.
of
Failure
‐
to
‐
Acquire
(FTA)
Face
‐
to
‐
Camera
Distance
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJECE Vol. 3, No. 6, D
ecem
ber 2013
:
863 – 874
8
71
Fig
u
re
4
.
Graph
sh
owing th
e
effect
o
f
face
-to
-
cam
era d
i
stan
ce
o
n
av
erag
e reco
gn
itio
n time
Tab
l
e
3
.
C
o
rrel
atio
n
o
f
ov
erall
qu
ality sco
r
es
(OQS)
with
ind
i
v
i
du
al im
ag
e qu
ality sco
r
es
QP
QF
QL
QC
QS
OQS
Pearson Correlatio
n
0.
936*
*
0.
840*
*
0.
266*
*
0.
262*
*
0.
670*
*
Sig
.
(2
-tailed
)
0.
000
0.
000
0.
000
0.
000
0.
000
N
2936
2936
2936
2936
2936
** Cor
r
e
lation is significant at the 0.
01 level (
2
-
t
ailed)
Tab
l
e
4
.
C
o
rrel
atio
n
o
f
algo
rit
h
m
m
a
tch
i
n
g
sco
r
es (AMS)
with
ind
i
v
i
d
u
al i
m
ag
e qu
ality sco
r
es
QP
QF
QL
QC
QS
AMS
Pearson Correlatio
n
0.
599*
*
0.
379*
*
0.
168*
*
0.
048*
*
0.
855*
*
Sig
.
(2
-tailed
)
0.
000
0.
000
0.
000
0.
009
0.
000
N
2936
2936
2936
2936
2936
** Cor
r
e
lation is significant at the 0.
01 level (
2
-
t
ailed)
Th
e
o
v
e
rall i
m
ag
e qu
ality scores (OQS)
was categ
ori
zed
into
fiv
e
qu
ality
classes as shown
on
tab
l
e
5 a
n
d each image
verification a
n
d quality assessm
ent
(IVQA) num
b
er is a pre
d
iction
of t
h
e
recognition
al
go
ri
t
h
m
’
s perf
orm
a
nce an
d
t
h
e co
nt
ri
b
u
t
i
o
n o
f
t
h
e
pr
o
b
e
im
age to the overall pe
rf
orm
a
nce of the
biometric
facial reco
gn
itio
n system
. Th
e i
m
p
licatio
n
o
f
th
is categ
orizatio
n
is th
at
1
,
7
1
8
and
1
,
02
0 im
ag
es with
in
t
h
e
“unacce
pta
b
le” and “poor”
ca
tegory wa
s
dis
carde
d
from
the experim
e
ntal da
tabas
e
. T
h
at
is 93.3% (2,738)
of
t
h
e i
m
ages was
rem
oved
an
d
onl
y
6.
7%
(1
9
8
)
was
l
e
ft
t
o
f
o
rm
a ne
w
dat
a
base.
Tab
l
e 5
.
Catego
rization
o
f
d
a
t
a
b
a
se
p
r
ob
e imag
es across
q
u
ality scales
Overall quality Sc
ore range
IVQA
nu
m
b
e
r
Description
0.
9 -
1.
0
5
E
x
cellent
0.
80 – 0.
89
4
Good
0.60 – 0.79
3
Acceptable
0.
40-
0.
59
2
Poor
0 – 0.39
1
Unacceptable
3.38
3.09
3.46
3.78
3.5
3.58
3.79
3.4
4.31
4.41
3.54
3.44
4.04
3.4
3.49
3.67
3.42
3.65
5.05
4.47
4.52
2.98
1.82
Cam
1_dist
1
Cam
1_dist
2
Cam
1_dist
3
Cam
2_dist
1
Cam
2_dist
2
Cam
2_dist
3
Cam
3_dist
1
Cam
3_dist
2
Cam
3_dist
3
Cam
4_dist
1
Cam
4_dist
2
Cam
4_dist
3
Cam
5_dist
1
Cam
5_dist
2
Cam
5_dist
3
Cam
6_dist
1
Cam
6_dist
2
Cam
6_dist
3
Cam
7_dist
1
Cam
7_dist
2
Cam
7_dist
3
Cam8_Frontal
…
Cam9_Frontal
…
Average
Recognition
Time
(Secs)
Average
R
ecognition
Time
(Secs)
Camera
Types
with
Face
‐
to
‐
Camera
Distance
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Fa
cia
l
Imag
e Verifica
tio
n
and
Qua
lity Assessmen
t S
y
stem -Fa
c
eIVQA
(Ab
a
y
omi-Alli O)
87
2
The
ne
w data
base
now c
o
ntains im
ages of accepta
ble
(55), good (13)
or e
x
cellent
quality (130).
Hen
c
e, t
h
e
p
e
rform
a
n
ce o
f
the b
i
o
m
etric reco
gn
itio
n syst
em
was g
r
eatly
i
m
p
r
ov
ed on
t
h
e
n
e
w
d
a
tab
a
se with
100%
accurac
y
of 198 true
a
ccept (TA), ze
ro false
re
j
ect
(FR) and a
m
e
an
rec
ognition score
(MRS) of 0.76
acro
s
s th
e th
ree reco
gn
itio
n alg
o
rith
m
as sho
w
n
on
tab
l
e
6.
Tabl
e
6.
Sum
m
a
ry
o
f
reco
g
n
i
t
i
on al
go
ri
t
h
m
’
s pe
rf
orm
a
nce o
n
t
h
e
ne
w
dat
a
base.
Algor
ith
m
SR
FTA
TA
FR
FA
TR
MRS
L
uxand SDK
198
0
198
0
0
0
0.
88
PCA
198
0
198
0
0
0
0.
72
LDA
198
0
198
0
0
0
0.
67
** Decision thr
e
shold = 0.
6
4.
CO
NCL
USI
O
N
Thi
s
pape
r
de
scri
bes
t
h
e
de
vel
o
pm
ent
and
im
pl
em
ent
a
t
i
on
o
f
di
f
f
ere
n
t
m
e
t
hods
t
o
m
easure t
h
e
quality of
facial i
m
ages using the
ge
om
etric and statistical features
of
the face through a
proposed facial
i
m
age verification and quality assessm
ent s
y
ste
m
(FaceIVQ
A). The quality of the f
acial i
m
age is expresse
d
b
y
i
m
p
l
e
m
en
t
i
n
g
m
easu
r
es an
d
algorithm
s
fo
r fi
v
e
imag
e q
u
a
lity attrib
u
t
es such
as facen
e
ss, po
se,
illu
m
i
n
a
tio
n
,
co
n
t
rast, an
d
si
milarit
y
. Th
e fu
ll-referen
ce
ob
j
ective qu
ality
m
easu
r
em
en
t tech
n
i
qu
e
for was
em
ployed in F
aceIVQA. T
h
e
distance
bet
w
een the
eyes
(DBE) and
t
h
e am
ount of
face
area detected by
the
algorithm
was use
d
to m
easure the
faceness
quality, a m
odi
fied a
n
d a
d
apt
e
d
optical
fl
ow techni
que
wa
s use
d
for th
e
p
o
s
e quality, stru
ctu
r
al
si
m
ilari
ty in
d
e
x
(SSIM
)
was
u
s
ed
fo
r
u
n
e
v
e
n
illu
min
a
tio
n
an
d
con
t
rast qu
ality
measu
r
e wh
ile th
e
im
ag
e
Eu
clid
ean
d
i
stan
ce (IMED)
m
e
tr
ic was
u
s
ed
fo
r t
h
e sim
ilarit
y
q
u
a
lity
m
easu
r
e.
The Results of evaluating Fa
ceIVQA shows that it
accurately assigns quality scores to probe i
m
age
sam
p
les. These indivi
dual
quality scores have shown
bo
th to
be hi
ghly
correlated
wit
h
each
othe
r a
nd also
p
r
ed
ictiv
e
o
f
th
e algorith
m
’
s m
a
tch
i
n
g
scores (AMS). Th
ey d
i
sclo
sed
a co
rrelatio
n b
e
t
w
een
d
i
fferen
t qu
ality
metr
ics an
d f
a
ce r
e
co
gn
ition
p
e
rf
or
m
a
n
ce lead
ing
to
th
e
p
o
ssib
l
e in
co
rpo
r
atio
n
o
f
q
u
a
lity m
easu
r
es in
a f
ace
perform
a
nce prediction sche
me to re
duce t
h
e ne
gative effect of
poor qu
ality sa
m
p
les
in face databases.
A
m
eans o
f
q
u
a
n
t
i
f
y
i
ng m
a
t
c
h per
f
o
r
m
a
nce w
a
s de
vel
o
ped
,
t
h
e res
u
lt shows that norm
aliz
ed dis
p
a
r
ate quality
attrib
u
t
e sco
r
es pred
icts m
a
t
c
h
p
e
rf
orm
a
n
ce, and
co
m
b
in
es m
u
ltip
le q
u
a
lity
measu
r
es in
to
a sing
le
sco
r
e
(OQS). Th
e resu
ltin
g
q
u
a
lity sco
r
e can
b
e
assig
n
e
d
to
im
ag
es cap
tured
for en
ro
ll
m
e
n
t
o
r
recogn
itio
n
and
can
b
e
u
s
ed
as an
i
n
pu
t to qu
ality-driv
e
n
b
i
o
m
et
ric fu
sion
system
s.
ACKNOWLE
DGE
M
ENTS
The aut
h
ors wi
sh to tha
nk Profes
s
o
rs Mislav Grgic,
Kresi
m
ir Del
ac and
Sonja
Grgic for the release
of the
SCFace
-surveillance ca
m
e
ras face da
tabase. Special
thanks to the
Luxand in
c
o
rporation for rele
asing
t
h
e
eval
uat
i
o
n versi
o
n
o
f
t
h
e Lu
xan
d
SD
K 4
.
0
l
i
b
rary
.
REFERE
NC
ES
[1]
DP
D’Am
ato. “Bes
t pract
ices
fo
r taking fa
ce pho
tographs and face im
age qual
i
t
y
M
e
trics
”
.
NIST Biometric Quality
Work
shop
. 2006.
[2]
WJ Sc
he
i
r
e
r
,
et a
l
.,
“
Predic
ting
biometric fa
cia
l
recognit
i
on fai
l
ure
with simila
rity surfaces an
d support vecto
r
machines
”. in
Proc. of
IEEE Co
mputer
Society
Workshop on Biometrics,
in
association with CV
PR. 2008.
[3]
P
Grother and E
Tabas
s
i
.
“
P
erform
ance of Biom
etri
c Quali
t
y
M
e
as
ures
”.
I
EEE Transactions on Pattern
Analysis
and Machin
e In
t
e
llig
ence
. 2007;
29(4): 531-543
.
[4]
M
T
Chan,
et
a
l
. “
Incorporating
quality metrics
in
multimodal
biometric fusion
”.
in Proc. of I
EEE Computer
Society
Worksh
op on Biometr
i
cs, in
a
ssociation
with CVPR. 200
6: 67-79.
[5]
K Nandakum
ar,
et a
l
. “
Q
ual
i
t
y
-b
ased score
lev
e
l
fusion in m
u
ltibi
o
m
e
tric s
y
s
t
em
s”.
In In
ternation
a
l Conferen
ce o
n
Pattern
Recognition
. 2006
.
[6]
P
W
a
ng, et al
.
“
Modeling and predicting
face recognition syst
em
performance based on analysis of similarit
y
sc
ore
s
”. in
Proc. of IEEE Tr
ansacti
ons on
Pattern
Analy
s
is and
M
ach
in
e Intelligen
ce. 2007; 29(4):
665–670.
[7]
E Tabassi. “NIST Fingerpr
i
nt
imag
e qu
ality
and r
e
lation
to PIV”.
NIST Image Group Technical rep
o
rt
. 2005.
[8]
F Perronnin. “A Probabilist
i
c Model of
Fa
ce M
a
pping Appl
ied
to Person Reco
gnition”.
An
Un
published Ph.D
Thesis submitte
d to th
e Depart
m
ent of
Mult
i-
Media Commun
i
cations
, École
Poly
techniqu
e
F
e
derale Lausann
e
(EPFL), Fran
ce.
2004.
[9]
F Hua, et al
. “
I
m
p
act of Out-of-Focus Blur on
Face Recog
n
ition Perform
a
n
ce Based on
Modular Transf
er
Function”.
Inter
national Con
f
erence on
Biometrics’12
. 2012
256-
272.
[10]
PJ Phillips,
et
al
. “
F
ace R
ecogni
ti
on Vendor
Test
2002”.
Eva
l
uation Report, Tech. Report NISTIR
6965
. 2003
.
Evaluation Warning : The document was created with Spire.PDF for Python.