TELKOM
NIKA
, Vol.11, No
.11, Novemb
er 201
3, pp. 6420
~6
426
e-ISSN: 2087
-278X
6420
Re
cei
v
ed Ma
y 6, 2013; Re
vised June
1
8
, 2013; Acce
pted Jul
y
10,
2013
Application of Data Fusion in Computer Facial
Recognition
Wang
Aiqian
g*, Han Min
Dep
a
rtment of Information En
gin
eeri
ng, He
n
an Pol
y
tec
h
n
i
c, Z
hengzh
ou H
ena
n 45
004
6, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: 5783
51
601
@
qq.com
*
A
b
st
r
a
ct
T
he recog
n
iti
o
n rate of sing
le
recogn
itio
n method is
i
neffici
ency in co
mput
er facial rec
o
g
n
itio
n. W
e
prop
osed a n
e
w
confluent facial reco
gn
ition
meth
od us
ing
data fusio
n
techno
logy, a vari
ety of recogniti
o
n
alg
o
rith
m ar
e combi
ned to for
m
the fus
i
on-
b
a
sed fac
e
reco
gniti
on syste
m
to improve th
e
recog
n
itio
n rat
e
in
ma
ny w
a
ys. Data fusion c
ons
iders thre
e lev
e
ls of data fu
si
on, feature l
e
v
e
l fusio
n
an
d d
e
cisio
n
lev
e
l fu
sion.
And the
data l
a
yer us
es a si
mp
le w
e
i
ghted
avera
ge a
l
g
o
rithm, w
h
ich
is e
a
sy to i
m
pl
e
m
e
n
t. Artificial n
e
u
r
a
l
netw
o
rk al
gorit
hm w
a
s sel
e
cted i
n
fe
ature
la
yer an
d fu
zz
y
r
easo
n
in
g
alg
o
ri
thm w
a
s
use
d
i
n
d
e
cisi
on
laye
r
.
F
i
nally, w
e
co
mp
are
d
w
i
th the BP ne
ural
netw
o
rk
alg
o
rit
h
m i
n
the MA
T
L
AB experi
m
ental p
l
atfor
m
.
T
he
result show
s th
at the recog
n
iti
on rate h
a
s be
en gre
a
tly i
m
pr
oved after a
d
o
p
ting d
a
ta fusi
on tech
nol
ogy
i
n
computer faci
al
recogn
ition.
Ke
y
w
ords
:
dat
a fusion tec
hno
logy, facia
l
rec
ogn
ition, fu
zz
y
reaso
n
in
g, neu
ral netw
o
rk
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The a
c
cu
rate
re
sults of f
a
cial
re
co
gnit
i
on pl
ay a i
m
porta
nt rol
e
in p
o
sitio
n
in
g faci
al
feature
and
contra
st of i
c
o
n
in vid
eo
su
rveillanc
e. So
far, both
dom
estic an
d fo
re
ign resea
r
che
r
s
have pro
p
o
s
e
d
a variety of algorith
m
s for face re
co
g
n
ition. Literature [1] propo
se
s the algo
rith
m
for fa
ce
dete
c
tion i
n
th
e m
o
vement
of the a
c
to
r,
whi
c
h ma
ke
s
it e
a
sy
fo
r
vide
o retrieval, but the
recognitio
n
ra
te is not high
for the algo
rit
h
m, while
the
actor
remai
n
s still. Literatu
re [2] pro
p
o
s
es
to use BP ne
ural n
e
two
r
k
thinkin
g
. Although thi
s
me
thod improve
s
the a
c
cura
cy and efficien
cy
of recognition, the algorithm still has
shortc
omi
ngs that it
cannot
meet the real-ti
m
e
requi
rem
ents of video re
trieval. The
article,
b
a
se
d on data f
u
sio
n
tech
ni
que, processes
sep
a
rately th
e colle
cted
origin
al information in
accordan
ce
with the type of sensor. Th
en,
grad
ually increase the re
co
gnition effect
accord
ing to the three-tier
f
r
amework
think
i
ng.
The m
a
in p
u
rpose of thi
s
article
is to
solv
e the
sho
r
t
c
omin
gs that
the faci
al recognition
rate is lo
w t
o
use
sin
g
le
identification
method an
d
to use data
fusion te
chn
i
que to comb
ine
variou
s id
enti
f
ication m
e
th
ods t
o
omni
b
earin
gly fo
rm
fuse
d faci
al reco
gnition
sy
stem to im
prove
the overall pe
rforma
nce of face re
co
gniti
on.
2.
Data Fusion
Techniqu
e Principle
The data fu
si
on take full a
d
vantage of t
he data
re
so
urces
coll
ect
ed by many
sensors in
different time
and
sp
ace
and
use
s
th
e compute
r
techn
o
logy
t
o
ma
ke anal
ysis, synthe
sis,
dominatio
n a
nd cal
c
ulatio
n unde
r ce
rtain crite
r
ia
of
the data gained by m
any sensors in time
orde
r, as wel
l
as achi
eve the corre
s
po
nding de
ci
sio
n
s and e
s
tim
a
tes for the
obje
c
ts ob
se
rved
[3]. The n
a
ture of the
data
fusion
is the
pro
c
e
s
s
of m
u
lti-so
urce
inf
o
rmatio
n fu
si
on a
nd
ab
stract
treating from l
o
w-l
e
vel to high-level, which is divi
ded in
to the three le
vels of the da
ta level fusion
,
feature level fusio
n
and d
e
c
isi
on level fu
sion [4].
The data fu
si
on pu
rpo
s
e i
s
to try to kee
p
the initial in
formation. T
h
erefo
r
e, in thi
s
level,
fuse directly the origin
al data dete
c
te
d by ev
ery sensor, witho
u
t giving more su
btle out
put.
However, the initial data obtai
ned by
the sensor
has the
char
acteri
stics
of instability and
uncertainty, whi
c
h re
sults
in
the blind
ness defe
c
t
of the data
l
e
vel in di
re
ct
integration.
The
feature
level
in the
ab
ove
level is to
extract
the fe
atu
r
e i
n
form
atio
n an
d m
a
ke
comp
re
hen
si
ve
analysi
s
an
d
treatment of these da
ta and fi
nal
ly, make cl
assificatio
n
, gatheri
ng an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Applicatio
n of Data Fu
sion
in Com
put
er Facial Re
cog
n
ition
(Wang Aiqiang
)
6421
synthe
sizatio
n
, in orde
r to make it ea
sy to
treat and redu
ce
real
-time the com
m
unication traffic
of the sy
ste
m
, comp
re
ss the data
duri
ng the
extr
action. The o
u
tput and
the p
r
etre
ated extract
informatio
n in the second l
e
vel is to be identified
by the de
cisio
n
-makin
g level, of which the
task
is to make fu
ll use of the input inform
ation and in
a
c
cordan
ce
with the algorith
m
of this level,
pro
c
e
ss i
n
formation, mainl
y
on the cha
r
acteri
stic
of face. In this
way, the out
put re
sults
could
kee
p
ac
cu
rat
e
.
3.
Data Fusion
in Facial Rec
ognition
The fa
cial
re
cog
n
ition
designed
in the
article,
ma
ke
s u
s
e
of the
thinki
ng of
dat
a thre
e-
le
ve
l fu
s
i
on
an
d
h
i
er
ar
ch
ica
lly p
r
oc
es
s th
e
or
ig
i
nal i
n
formation
coll
ected. T
he d
a
ta acqui
sitio
n
informatio
n is from vario
u
s sen
s
o
r
s,
sen
s
or
1, 2,…,n,
maybe ho
m
ogen
eou
s o
r
hetero
gen
eo
us.
Und
e
r
norma
l circum
stan
ces, the
data
obtai
ne
d by
the hom
og
eneo
us
se
nsor, after
bei
ng
pretreated by
the
data l
e
vel, will
be directly taken to the featur
e l
e
vel for fusi
on. However, t
h
e
data by the h
e
terog
ene
ou
s sen
s
o
r
mu
st go throu
gh t
he feature extraction
pro
c
ess, in ord
e
r
to
be take
n to the he de
cisi
on
-ma
k
ing level
to output.
Therefore, th
e article a
dop
ts two way
s
to pr
o
c
e
ss the
pretre
ated o
u
tput informat
ion from
the data
level
.
The
data fro
m
the
homo
g
eneo
us sen
s
or
will b
e
di
re
ctly take
n to t
he featu
r
e
level,
the other
dat
a will
directly
go th
rough feature extraction
and will not
participat
e in the fusion in
the feature l
e
vel. In theory, this method
will
have b
e
tter re
co
gnitio
n
effects th
a
n
to use
sin
g
l
e
factor to pro
c
ess the data obtaine
d by sen
s
o
r
.
The facial recognit
i
on pro
c
e
s
s desi
gne
d in the
article i
s
sh
o
w
n in the figu
re belo
w
:
Figure 1. Faci
al Re
cog
n
ition Process
The ap
propri
a
te algo
rithm
not only ca
n
redu
ce th
e computation
a
l
compl
e
xity, but also
improve the a
c
cura
cy. Therefore, we
m
u
st
sele
ct
sev
e
ral algo
rit
h
m
s
out of many
that could giv
e
full play to th
e effect of
da
ta fusio
n
in t
he a
r
ticle. After the
re
se
arch
and co
mp
arison of
vari
ous
fusion
algo
rithm pri
n
ci
ple
s
an
d the a
d
vantage
s a
nd di
sadva
n
tage
s,
it wa
s found that
the
weig
hted
ave
r
age
al
gorith
m
is si
mple
a
nd e
a
sy
to
condu
ct a
nd
meet the
req
u
irem
ents of
data
layer p
r
etreat
ment. The B
P
neural net
work i
s
t
he
a
l
gorithm
ba
se
d on trainin
g
, whi
c
h
can
gi
ve
play to its strong poi
nts in
the f
eature l
e
vel. The de
cisi
on-ma
king
level of the last level, ado
pt
s
the fuzzy rea
s
oni
ng to ach
i
eve the
final output re
sult of human face.
3.1. Data Level Fusion
In the comp
uter faci
al re
cog
n
ition, we
must
first g
e
t the data. If the original
data is
obtaine
d by t
he came
ra, b
a
se
d on
the
con
s
id
erati
o
n
about th
eir
chara
c
te
risti
c
of unsta
bility, i
t
need
s to p
r
ep
rocess a
nd e
x
tract the feat
ure valu
es. T
he main
goal
of prep
ro
ce
ss is to re
du
ce
or
eliminate ima
ge noi
se. In o
r
de
r to ma
ke
it easy
to extract the hu
ma
n face from the ba
ckgroun
d,
it need
s th
e
edge
dete
c
tio
n
. Wh
en
the
comp
uter is
p
r
ocessin
g
the
imag
e, the
g
r
ayscale
ima
g
e
is co
nsi
d
e
r
ed
to redu
ce th
e cal
c
ulatio
n compl
e
xi
ty and re
so
urce
con
s
um
pti
on.
Therefo
r
e, it
still
need
s the gra
y
scal
e
pro
c
e
s
sing.
Pre
t
re
at
Vide
o
Surv
eill
a
nce
Sensor1
Sensor2
Sensor n
Pre
t
re
at
Pre
t
re
at
F
eature
Level
Fusion
Recogniti
on
F
eature
Extra
c
tio
Decis
i
on-
making
Leve
l
Fusion
Recogniti
on
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 642
0 – 6426
6422
The feature point shoul
d b
e
first determ
i
ned fo
r the feature extraction. The differen
c
e
s
coul
d be final
ly found on the face of th
e different
a
c
tors. Sele
ct the lips, no
se,
eyes, eyebrow
human fa
cial
feature poi
nts, order O
as
pixel,
，
o
the poi
nt within the neigh
borhoo
d
()
o
qq
，
q
in the followin
g formul
a is pre
s
e
n
te
d as the gra
y
scal
e
of q
，
the numbe
r
of pixels of
neigh
bou
rho
o
d
o
is
pr
es
en
ted
as
o
n
。
The fe
ature
wei
ght
cal
c
ulatio
n d
e
s
ign
ed i
n
thi
s
pap
er
is divided into
three step
s. The wei
ght value bri
n
g
s
us gre
a
t conve
n
ien
c
e to extract feature. T
h
e
weig
ht cal
c
ul
ation is a
s
follows:
Step 1: Requ
est mean
squ
a
re e
rro
r of grayscale
2
()
1
o
qq
q
oq
o
n
(1)
Step 2: Project into nucl
e
a
r
spa
c
e
ex
p
o
oq
r
oq
oq
o
n
(2)
Step 3: Weig
ht calculation
formula
oq
oq
oq
qo
(3)
After the g
r
ey
Pro
c
e
ssi
ng
o
f
face i
m
ag
e i
n
the
above
steps, the
influ
ence of
noi
se
point
on the ce
nter point of the gray
-valu
e
coul
d be
eliminated. In addition, u
s
e anti-sh
arp
en
membrane
to
enha
nce ed
ges, fold
ima
ge an
d hig
h
-
pass filter
wit
h
the im
p
a
ct
respon
se
P, the
margi
nal information ca
n b
e
obtaine
d. The final imag
e pro
c
e
ssi
ng
results a
r
e a
s
follows
()
QL
P
Q
P
Q
(4)
Orde
r L as id
entity operato
r
. Add a factor
(1
)
，
whi
c
h is u
s
ed to enh
an
ce the high
-
freque
ncy pa
rt
：
()
sh
QL
P
Q
P
Q
(5)
Define a
n
ti-sharp
en ma
sk filter:
(1
)
I
LP
(6)
After filtering
of anti-sha
r
pen ma
sk, th
e ultimate id
eal imag
e ca
n be o
b
taine
d
. Then
unde
rtake fusion pro
c
e
s
sin
g
of the image of differ
ent sen
s
o
r
throu
gh every level. Presently, the
algorith
m
s a
pplicable
to t
he level
such a
s
wei
ghte
d
ave
r
ag
e m
e
thod
and
wavelet tra
n
sf
orm
method
co
uld
all b
e
u
s
e
d
f
o
r th
e fu
sion
cal
c
ulatio
n. F
o
r th
e
simple
and
co
nvenie
n
t cal
c
ul
ation
is
con
s
id
ere
d
,
adopt
the weighted average
fu
sion al
gor
ithm. F
o
r
the se
nsors
that the ori
g
i
nal
image
s a
r
e f
r
om
are
diffe
rent, the
differen
c
e
s
co
ul
d be
ca
used
in the m
a
tching. Thi
s
article
aban
don
s th
ese
con
s
id
erations a
nd in
tegrate
s
t
he redun
dant inf
o
rmatio
n of the ori
g
inal i
m
age
to obtain the better sig
nal
-to-noi
se
ratio.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Applicatio
n of Data Fu
sion
in Com
put
er Facial Re
cog
n
ition
(Wang Aiqiang
)
6423
Oder
(,
)
x
y
to be pre
s
ented as result fu
nction after fusion
，
o
r
der
12
3
(,
)
(
,
)
(,
)
A
xy
A
x
y
A
xy
、、
to be
presen
ted a
s
the
im
age from th
e
sen
s
o
r
，
of which
(,
)
x
y
is
the coo
r
din
a
te positio
n of certai
n point i
n
the
image.
The imag
e fusion fun
c
tion
to be use
d
is
as
follows
：
12
3
(,
)
(
,
)
(,
)
(
,
)
1
x
ya
A
x
y
b
A
x
y
c
A
x
y
a
c
、b
、
(7)
Of which
a
、
b、
c
is
weig
hting co
efficient
，
The
size of the
weighting
coeff
i
cient can be
determi
ned
accordi
ng to the quality of the image
，
but it sh
ould meet
++
=
1
ab
c
。，
W
h
en
1
==
=
3
ab
c
，
it is pre
s
ent
ed as the ave
r
age fu
sion.
In the syste
m
desig
n, three image
s of
the fa
ce of the sam
e
act
o
r are to be colle
cted,
after the loca
tion, characte
r seg
m
entati
on and no
rm
alizatio
n, thre
e sampl
e
s of
the same fa
ce
coul
d b
e
o
b
tained. It i
s
different fro
m
the
pa
st t
hat only
con
s
ide
r
s
single
-
frame
al
gori
t
hm,
becau
se thi
s
frame ta
ke
n
may have
qu
ality probl
em.
The
arti
cle
u
nderta
ke
s averag
ed fu
sio
n
of
pixels of the three
sam
p
le
s, the obtained
resu
lt
s saved and spe
c
ial
level input are from he
re.
3.2. Featur
e Lev
e
l Fusion
The
se
con
d
l
e
vel of data
fusio
n
, the fea
t
ure leve
l, the work is
rather diffic
u
lt. It forms the
fusion
vecto
r
and co
ndu
cts
the nonlin
ea
r
o
perati
on
o
n
the
data
fro
m
differe
nt
sensors
and
well
solve
s
the
d
e
fects of the
linea
r op
era
t
ion. The
r
e a
r
e al
so
many
fusio
n
alg
o
ri
thms in
featu
r
e
level. The
alg
o
rithm
s
often
use
d
a
r
e th
e
para
m
et
er te
mplate al
go
rithm, BP ne
ural net
work,
RBF
neural net
wo
rk al
gorith
m
and so on. T
he neu
ral n
e
t
work can b
e
use
d
in no
n
-
linea
r ma
ppi
n
g
system that demands
se
l
f
-learning and self
-organi
zation. Its ability to
learn
is strong and it
c
o
nd
uc
ts
linea
r
pr
oc
es
s
i
ng
o
n
the
in
fo
r
m
a
t
ion
out
put by different se
ns
ors,
whi
c
h m
eet
s the
nonlin
ear n
e
ed. The co
mmonly use
d
neural net
work alg
o
rith
ms are the data fusion
and
cla
ssifi
cation
of BP neural
network a
n
d
RBF neu
ral
network. Th
e article
cho
o
se
s BP neu
ral
netwo
rk that i
s
ba
sed o
n
the training,
be
cau
s
e the BP
training is
co
nvenient.
The input level of BP neural netwo
rk re
ceive
s
the ou
tput after pretreatme
nt by data level
or
sen
s
o
r
, then
sen
d
s t
he info
rmatio
n to the int
e
rme
d
iate le
vel and final
ly transmit
s
the
informatio
n to
every
neu
ro
n of the
o
u
tp
ut level. Use
sigmoi
d fun
c
t
i
on a
s
th
e ex
citation fu
ncti
on,
its expre
ssi
on
is as follo
ws:
1
()
1e
x
p
(
)
t
t
(8)
In addition, th
is fun
c
tion
ca
n also b
e
u
s
e
d
as the ex
citation fun
c
tion
of the out
put
node.
The
slop
e of
the fun
c
tion i
s
p
r
e
s
ente
d
as
paramete
r
，
The
slop
e
of the fun
c
tio
n
will
chang
e
will the value of
。
Wh
en
de
cre
a
ses
，
th
e
slop
e of th
e
function
will
also
de
crea
se
，
wh
en
，
the Sigmoid
function be
comes the ju
m
p
function.
The inp
u
t vector of e
a
ch neuron i
s
set
to be
1
1
12
,(
,
,
.
.
.
)
n
T
n
uR
u
u
u
u
，
and
is
with
2
n
个
outpu
ts
2
2
12
,(
,
,
.
.
.
)
n
T
n
vR
v
v
v
v
.
1
1
n
j
ii
j
i
u
(9)
1
()
1
e
xp(
)
jj
j
(10
)
Do de
rivation
：
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 642
0 – 6426
6424
e
xp(
)
1
()
1
(
)
(
)
1
e
xp(
)
1
e
xp(
)
j
jj
jj
jj
(11
)
Set the n
u
m
ber of n
euro
n
s i
n
the
hi
d
den l
e
vel of
the first level
to be
n, the
output is
12
,(
,
,
.
.
.
)
nT
n
x
Rx
x
x
x
，
The
se
co
nd level
has m
neu
ron
s
,
the out
put is
12
,(
,
,
.
.
.
)
mT
m
x
Rx
x
x
x
. Set
ij
as the
weight from
the input lev
e
l to the hidd
en level of
the first level
,
the thre
shol
d value i
s
j
，
orde
r
j
k
as
the
weight from t
he input level to the
hidde
n level of the se
con
d
level
j
is set to be the thre
shold value
；
t
he wei
ght fro
m
the hidde
n
level of the seco
nd level t
o
the outp
u
t level is
ij
w
，
the thre
shol
d valu
e is
j
. So the
neuron
s
outputs of the
two levels are respe
c
tively:
0
0
12
0
()
,
0
,
1
,
2
,
.
.
.
,
(
)
,
0
,
1
,
2
,
...,
(
)
,
0
,
1
,
2
,
...
1
n
ji
j
i
j
i
m
kj
k
j
k
j
m
kl
l
k
x
xj
n
x
xk
m
yx
l
n
(12
)
After setting
t
he n
e
two
r
k p
a
ram
e
ters
according
to th
e ne
ed
s of
th
e recognition
system,
con
d
u
c
t traini
ng for the
ne
ural n
e
two
r
k, and
con
s
tantl
y
adjust p
a
ra
meters du
rin
g
the traini
ng
to
achi
eve goo
d
convergen
ce
effect, finally,
cond
uct ide
n
t
ification on face ima
ge.
3.3. Decision
-making Lev
e
l Fusion
The high
er l
e
vel of syste
m
fusion i
s
the fusio
n
of deci
s
ion
-
ma
king level. T
he final
deci
s
io
n-m
a
ki
ng ba
sis is
the output result of
the deci
s
io
n-m
a
ki
ng level fusi
on. Therefore,
rega
rdl
e
ss o
f
the output
from the
feature
le
vel
or th
e dire
ctly req
u
e
s
te
d output
of the
characteri
stics of the data
of the
sensor as we
ll as the auxiliary inf
o
rmat
ion data, the deci
sion-
makin
g
level
must ta
ke full
advanta
ge o
f
them an
d
starts from the
sp
ecifi
c
issu
es, a
nd finall
y
output the accurate
re
sul
t
s. The deci
s
ion-m
a
ki
ng a
nd judge
men
t
basis is the
interrelatio
n
of
each facto
r
.
First, it ne
ed
s the tran
smi
ssi
on b
and
wi
dth with lo
we
r sen
s
or i
n
formation, seco
nd,
the st
rong anti-interference ability,
then it is
effective t
o
reflec
t on different
information.
But
t
h
e
factor h
a
s fu
zzi
ne
ss, so in the articl
e, we u
s
e
the f
u
zzy rea
s
o
n
ing tech
niqu
e. The un
stru
ct
ured
data processi
ng is ta
ken a
s
the
outp
u
t of the neural
netwo
rk. Th
e
fuzzy re
ason
ing tech
niqu
e
is
suitabl
e for the comp
utatio
n of
the results structu
r
ed
knowl
edge.
Duri
ng the fu
zzy rea
s
onin
g
, the sch
em
atic diag
ram
we u
s
e is a
s
follows:
Figure 2. Fuzzy Rea
s
o
n
ing
System Sch
e
matic Di
ag
ram
The fu
zzy
re
aso
n
ing
met
hod
s in
clud
e
:
Take
maxi
mum an
d mi
nimum
com
p
osition
a
l
operation
as the Ma
mdani
i
n
feren
c
e
met
hod
of co
mpo
s
itional
rul
e
.
Whe
n
requ
est, do re
asonin
g
Fuz
z
y
Decis
i
on
Non-
f
u
zzif
i
cation
Treat
m
e
nt
Quit
Control
Variable
featur
e
Leve
l
Output
Fuz
z
y
Quantiz
ation
Trea
tm
ent
Fuz
z
y
Control
Rule
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Applicatio
n of Data Fu
sion
in Com
put
er Facial Re
cog
n
ition
(Wang Aiqiang
)
6425
or fusio
n
of incentive in
tensity, use prod
uct op
eration Larse
n
reasonin
g
method. Zad
eh
reasoni
ng m
e
thod defined by using
different
fuzzy relations.
The arti
cle adopts Mamdani
rea
s
oni
ng me
thod, of which the empha
sis is on t
he
se
lection of sub
o
rdin
ating de
gree fun
c
tion.
The a
r
ticl
e
cho
o
ses fuzzy stati
s
tical
method
in
the dete
r
min
a
tion of m
e
mbershi
p
function, a
s
t
he metho
d
is based o
n
th
e numb
e
r
of
tests
and
ha
s ce
rtain p
r
a
c
tical b
a
si
s. Fi
rst
determi
ne a
domain of di
scourse
U, variabl
e cle
a
r
set A, total numbe
r of tests n, confirmi
ng
factor a, so the requ
est formula of
mem
bership frequ
ency is a
s
foll
ows:
f
aA
n
(13
)
When n is
gradually increasing, f wi
ll
become st
abilized. The stable value is the
membe
r
ship
value of a to A.
In addition, d
u
ring
obfu
s
cation, the theory
to dete
r
mine the me
mbershi
p
fun
c
tion is:
arrang
e
m
fuzzy sub
s
ets in lunyu
12
,,
.
.
.
m
B
BB
，
more
over, for every
i
B
it has
sub
o
rdi
nating
degre
e
functi
on
()
i
B
x
，
for any
0
x
U
，：
if
01
0
2
0
0
()
m
a
x
(
)
,
()
,
.
.
.
()
im
Bx
B
x
B
x
B
x
(14
)
It can judge t
hat
0
x
b
e
l
on
gs
to
i
A
。
It needs to
a
nalyze i
n
form
ation on th
e
detecti
o
n
re
gi
on, the environment a
nd t
he targ
et
obje
c
t and
make
rea
s
o
nable a
nd correct de
ci
si
on-m
a
ki
ng o
u
tput, thus get the accurate
identificatio
n
of human fa
ces. First, dete
r
mine th
e
de
cision
-ma
k
ing
factor.
Use th
e feature val
u
e
from the feat
ure level
out
put or
extract
ed from
se
nsor a
s
a
ba
sis for faci
al re
cognition. Fin
a
lly,
the deci
s
io
n fusio
n
stru
ctu
r
e to desig
n facial re
co
gnitio
n
is as follo
ws:
Figure 3. Faci
al Re
cog
n
ition
De
cisi
on-m
a
kin
g
Level
Model
Afte
r
th
e
ne
ur
a
l
ne
tw
ork
is
ba
se
d
on
tr
a
i
n
i
n
g
, extra
c
t the
differe
nt feature
s
of huma
n
face. As the para
m
eters set may chang
e durin
g trai
n
i
ng in each n
e
twork, it lead to the different
output an
d t
he recognitio
n
rate
of the
sam
e
face
by different
neural net
wo
rk,
so it n
e
e
d
to
con
s
id
er this
to determine
the membe
r
ship func
tion.
After every neural n
e
two
r
k has identifie
d
the same im
a
ge, it forms a matrix M. After figur
e out the re
cog
n
itio
n rate of each
neural n
e
two
r
k,
norm
a
lize it into matrix N. Memb
e
r
ship functio
n
is det
ermin
ed a
s
the prod
uct of two matri
c
e
s
:
*
QN
M
(15
)
Finally, figure
out the maximum value corre
s
p
ondin
g
to the human face a
s
the final
res
u
lt
s of
t
he
f
u
zzy
de
ci
sio
n
:
()
m
a
x
m
a
x
*
i
A
xQ
N
M
(16
)
F
eature
Level Output
Fuzzif
i
cation
Fuzzy Logical
Reasoning
Deblur
r
i
ng
Fuzzif
i
cation
Control Rule
Output
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 642
0 – 6426
6426
4.
Simulation Results a
nd Analy
s
is
The a
r
ticle
u
s
e
s
MATLAB
softwa
r
e. M
A
TLAB c
an
b
e
ea
sily used
in expe
rime
nt and its
starting
point
of learni
ng i
s
not that hi
gh a
s
ot
he
r
softwa
r
e. Th
e articl
e con
duct
s
sim
u
lat
i
on
totally three times. The p
r
e
v
ious two tim
e
s are di
re
cte
d
at BP neura
l
network. Th
e last time is for
simulatio
n
d
a
t
a fusio
n
alg
o
rithm flo
w
.
Colle
ct 14
4
sets of im
age
s fro
m
the vi
deo
su
rveilla
nce.
Takin
g
into a
c
count that the neural network i
s
t
he net
work ba
se
d o
n
training, we
sele
ct 90 sets
of them a
s
th
e sa
mple
s to
be u
s
ed
in training
net
work. The t
r
aini
n
g
is
co
ndu
cte
d
for b
e
tter te
st,
then the
rem
a
ining i
m
age
s a
r
e u
s
e
d
f
o
r te
st. Thro
ugh fa
cial
re
cog
n
ition a
n
d
simul
a
tion in
the
article, the fol
l
owin
g re
sults statistics co
u
l
d be obtain
e
d
.
Table 1. Fa
ci
al Re
cog
n
ition Statisticss
Classifi
cation
Model
Sample
（）
90
Test
（）
54
Correc
t
Identification
Wrong
Identification
Correc
t
Identification
rate/%
Correc
t
Identification
Wrong
Identification
Correc
t
Identification
rate /%
BP net
w
o
rk1
88
2
97.8
45
9
83.3
BP net
w
o
rk2
90
0
100
48
6
88.9
Fusion
Sy
s
t
e
m
90 0
100
52
2
96.3
From th
e tabl
e, it can
be fi
gure
d
out th
a
t
afte
r the dat
a fusio
n
, the f
a
cial
re
cog
n
ition rate
has g
r
eatly improve
d
.
5.
Summar
y
and Outlook
For
data fu
si
on, it takes
a
d
vantage
of
a vari
ety of
reco
gnition
al
gorithm
s, lev
e
l by level
to improve the ability of recognition. It i
s
an em
er
gi
ng research di
rection.
In the article, after
readi
ng
a lot
of literatu
r
e, I
am
dee
ply in
spired to
a
ppl
y the ide
a
of
data fu
sion
to
the
com
puter
facial recogni
tion system. I
t
need
s to co
nsid
er t
hei
r-o
w
n
cha
r
a
c
teri
stics an
d pu
rposes. Fin
a
ll
y,
determi
ne th
e ap
propri
a
te
fusio
n
al
go
rithm that
use t
he
weig
hted
averag
e al
go
rithm in
the
d
a
ta
level, sele
ct
BP neural
n
e
twork in fe
ature le
vel
a
nd u
s
e fu
zzy reasonin
g
algorith
m
in
the
deci
s
io
n-m
a
ki
ng level. Fi
na
lly, the experi
m
ents
sh
ow
that the m
e
th
od p
r
op
osed
in the a
r
ticl
e i
s
effective. To
test on th
e
MATLAB sim
u
lation pl
atform, an
alyze t
he expe
rime
ntal re
sult
s a
n
d
throug
h com
pari
s
on, finall
y
find that accura
cy of fa
cial recogniti
on gain
ed b
y
applying d
a
ta
fusion te
chni
que can be
over 90%.Co
m
pared with
only BP neural net
wo
rk pro
c
e
ss, in
the
article, we ca
n get better reco
gnition re
sults.
Fore
ca
st the appli
c
ation of
data fusion t
e
ch
ni
qu
e in face recogniti
on. The future work
inclu
d
e
s
two
asp
e
ct
s: first,
develop
and
improve
dat
a fusio
n
theo
ry and th
e th
eory i
s
p
r
acti
cal
guida
nce, the data fusion requires
m
o
re authoritative
theor
etical
su
pport, in orde
r to go furthe
r in
sci
en
ce. Se
cond, the
excellent
system
often
ha
s strong
le
arning and adapt
ive
ca
pabilitie
s,
the
three l
e
vel p
r
ocessin
g
al
g
o
rithm
s
of th
e cu
rrent
d
a
ta fusio
n
h
a
ve ce
rtain
ad
aptivity. But the
appli
c
able
ra
nge is n
o
t wi
de eno
ugh a
nd the algo
rit
h
m
of the de
cisi
on-ma
king
level often has to
rely on peo
pl
e's expe
rie
n
ce to set.
Referen
ces
[1]
Kakuma
nu P, Makrogi
an
nis
S, Bourbakis N
.
A surve
y
of skin-col
o
r mod
e
ling
and d
e
tect
ion meth
ods.
Pattern Reco
g
n
itio
n. 200
7; 40(3): 110
6-1
1
2
2
.
[2]
Z
hou Ji
an
hua.
Vide
o faci
al
recog
n
itio
n of
PCA an
d SV
M multi-bi
ome
t
ric feature fu
sion
. Jiam
usi
Univers
i
ty Jour
nal: Natur
a
l Sci
ence Ed
itio
n
. 2010; 28 (4): 4
8
5
-48
8
.
[3]
John Sa
lem
o
. Informatio
n
F
u
sion: A Hig
h-lev
e
l
Architectur
e
Overvie
w
. Ann
apll
i
s: ISIF
2002.
[4]
Z
hang
Cu
i. Us
ed for
inform
ation
fusio
n
tec
h
niq
ue r
e
se
arch
of u
n
d
e
r
w
ater
target
ide
n
tific
a
tion.
Xi'
a
n
:
North
w
e
s
tern
Pol
y
t
e
ch
nica
l Univers
i
t
y
. 2
0
0
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.