Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
1
3
,
No.
2
,
Febr
uar
y
201
9
, pp.
7
87
~793
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
2
.pp
7
87
-
7
93
787
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
AI
-
B
ase
d Targ
eted Adv
ertising
System
Te
w
J
ia
Yu, Chi
n P
oo Lee,
Kian Mi
n
g
Li
m, S
iti
Fa
tima
h Abdul
R
az
ak
Facul
t
y
of
Infor
m
at
ion
Sci
ence and
T
ec
hnolo
g
y
,
Multi
m
edia
Uni
ver
sit
y
,
Mal
a
y
s
i
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Oct
1
5
, 201
8
Re
vised
Dec
16
, 2
018
Accepte
d
Dec
30
, 201
8
The
m
ost
comm
on
te
chnol
og
y
used
in
t
ar
get
ed
adve
rt
isi
ng
is
fac
ial
rec
ogni
ti
on
and
vehi
c
le
re
cogni
t
i
on.
Eve
n
though
the
re
ar
e
exi
sti
ng
sy
st
ems
serving
for
th
e
t
arg
eting
purpose
s,
m
ost
propose
li
m
it
ed
fun
ctio
nal
ities
and
the
s
y
s
te
m
per
form
anc
e
is
no
rm
al
l
y
unknown.
Thi
s
p
ape
r
pre
sents
a
n
int
ellige
n
t
t
arg
eted
adve
r
ti
sing
s
y
stem
with
m
ultiple
fun
ct
ion
al
i
ties,
name
l
y
fac
i
al
re
cogni
t
io
n
for
gende
r
a
nd
age
,
veh
ic
l
e
rec
ognition,
a
nd
m
ult
ipl
e
obje
c
t
detec
ti
on.
The
m
ai
n
purpo
se
is
to
improve
the
eff
ective
n
ess
of
outdoo
r
adve
rt
ising
thro
ugh
biometri
cs
a
pproa
che
s
and
m
ac
hine
l
ea
rnin
g
te
chnol
o
g
y
.
Mac
hine
le
arn
i
ng
al
gorit
hm
s
are
implement
e
d
for
highe
r
rec
ogni
ti
on
ac
cur
acy
and
he
nce
ac
h
ie
ved
be
t
te
r targ
et
ed
adv
e
rti
sing
eff
e
ct
.
Ke
yw
or
ds:
Ag
e
esti
m
a
ti
on
Gende
r
rec
ogni
ti
on
Object
recog
niti
on
Targ
et
e
d
a
dver
ti
sing
Veh
ic
le
recog
ni
ti
on
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
:
Chin
Poo Lee
,
Faculty
of In
form
ation
Scie
nc
e an
d
Tec
hnol
og
y,
Mult
im
ed
ia
U
ni
ver
sit
y,
Jal
an Aye
r Ker
oh Lam
a, 7
54
50 B
uk
it
Ber
ua
ng, Mel
a
ka,
Mal
ay
sia
.
Em
a
il
: cplee
@m
m
u.
edu
.m
y
1.
INTROD
U
CTION
Bi
ll
bo
ard
a
dverti
sing
is
a
t
ype
of
O
ut
-
Of
-
Ho
m
e
adv
ert
isi
ng
that
gr
a
bs
the
cha
nce
s
of
outd
oo
r
prom
otion
w
hi
ch
ty
pical
ly
achieves
desir
able
res
ults.
I
n
co
ntrast
to
tradit
ion
al
bi
ll
bo
ar
ds
wit
h
sta
tic
m
essages,
di
gital
bill
bo
ards
with
m
or
e
flex
ibil
it
y
and
up
-
t
o
-
date
m
essages
are
ta
kin
g
t
heir
way
of
re
placi
ng
them
.
This
m
or
e
ad
van
ce
d
f
orm
of
adv
e
rtisi
ng
is
known
as
the
Digital
Out
of
H
om
e
(D
O
OH).
As
te
ch
nolo
gy
evo
l
ved,
di
gital
adv
e
rtisi
ng
be
com
es
increasing
ly
popula
r
no
t
on
ly
bec
au
se
of
it
s
te
nde
nc
y
of
lo
wer
c
os
t,
but
al
so
it
s tar
getin
g
a
nd interact
i
ve feat
ur
e
s
with the
use
of
ca
m
eras,
se
ns
ors
and o
t
her ad
d
-
on d
e
vices.
Digital
bill
bo
a
rd
s
a
re
cal
le
d
“sm
art”
or
“i
ntell
igent”
wi
th
their
ca
pabi
li
t
ie
s
of
rec
ognizin
g
a
par
ti
cula
r
ob
j
e
ct
and
disp
la
y
releva
nt
co
ntent
to
it
.
The
se
bill
bo
a
rd
s
a
re
connecte
d
t
o
de
vices
f
or
c
ollec
ti
ng
inputs, a
nd a s
yst
e
m
is w
orki
ng b
e
hind a
s th
e co
ntr
ol. Alo
ng
with all
the a
lgorit
hm
s an
d processi
ng func
ti
on
s
,
the
bill
boar
d
will
be
ta
r
geting
a
certai
n
gro
up
an
d
dis
play
releva
nt
ad
ver
ti
sem
ent
for
bette
r
at
te
ntion
a
nd
influ
e
nces.
Se
ver
al
iss
ues
ar
e
identifie
d
in
the
existi
ng
a
dverti
sin
g
syst
e
m
s
as
the
fo
ll
ow
i
ng
:
(1)
un
s
uitable
adv
e
rtise
m
ents
are
disp
la
ye
d
to
the
au
dience
,
(2)
una
ble
to
ta
rg
et
outd
oor
aud
ie
nce
with
out
any
act
ivit
y,
and
(3)
lim
i
te
d
functi
on
al
it
ie
s.
These
pro
ble
m
s
resu
lt
in
wastage
of
res
ources,
high
costs
in
adv
e
rtisi
ng,
an
d
ineff
ect
i
ve
a
dverti
sing
.
In
vie
w
of
this,
an
intel
li
gen
t
ta
rg
et
ed
ad
vert
isi
ng
syst
e
m
i
s
propose
d
f
or
the
fo
ll
ow
i
ng
obj
ect
ives
:
(1)
disp
la
y
be
tt
er
-
ta
rg
et
e
d
adv
e
rtise
m
ents
to
the
aud
ie
nce,
(
2)
im
pr
ov
e
the
e
ff
ec
ti
ven
ess
of
outd
oo
r
adv
e
rtisi
ng,
a
nd
(
3)
offe
r
wi
de
range
of
f
un
c
ti
on
al
it
ie
s
in
a
sing
le
syst
em
.
The
propose
d
syst
e
m
is
capab
le
of
recog
nising
ge
nd
e
r
an
d
age
of
detect
ed
face
s
as
well
as
var
io
us
obj
ect
ca
te
gories
su
c
h
as
veh
ic
le
s,
el
ec
tro
nic
dev
ic
es
,
a
nd
f
ood.
Be
st
-
s
uited
a
dv
e
rtise
m
ent
rele
van
t
t
o
the
real
-
ti
m
e
dem
og
ra
ph
ic
s
will
be
retriev
ed
a
nd
disp
la
ye
d
on
the
bill
bo
a
rd,
s
olv
in
g
the
pro
blem
s
enco
unt
ered
in
existi
ng
a
dverti
sin
g
s
yst
e
m
s
as
m
en
ti
one
d
befor
e
.
The
rest
of
the
pa
per
is
orga
nized
as
fo
ll
ows.
I
n
Se
ct
ion
I,
an
intr
oduct
ion
of
bill
bo
a
r
d
a
dv
e
rtisi
ng
and
ta
r
geted
a
dv
e
rtisi
ng
is
pro
vid
e
d.
S
ub
se
qu
e
ntly
,
s
om
e
existi
ng
w
orks
in
sm
art
bill
boar
d
a
dverti
sin
g
a
re
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
3
, N
o.
2
,
Fe
bru
ary
201
9
:
7
8
7
–
7
9
3
788
rev
ie
wed
i
n
Se
ct
ion
II.
T
he
sy
stem
f
low
an
d
functi
onal
it
ie
s ar
e d
et
ai
le
d
in
Sect
ion
III.
Se
ct
ion
IV
pr
ese
nts the
exp
e
rim
ental
r
esults o
f beta
t
est
ing
a
nd Sect
ion
V
c
oncl
ude
s the
pa
per
.
2.
LIT
ERATUR
E REVIE
W
Existi
ng
ad
ve
rtisi
ng
syst
em
s
are
stu
died
an
d
a
c
om
par
iso
n
is
m
ade
in
te
rm
s
of
their
f
unct
ion
al
i
ti
es.
Yaho
o
sm
art
bi
ll
bo
ar
d
[1
]
reli
es
on
a
co
nce
pt
cal
le
d
gro
up
li
zat
ion
,
wh
ic
h
pri
ors
on
the
m
ajo
rity
to
gai
n
m
or
e
at
te
ntion
f
r
om
people
ar
ound
it
.
Yahoo
sm
art
bill
bo
ar
d
is
usi
ng
im
age
rec
ogniti
on
te
c
hn
ology
w
orkin
g
with
ca
m
eras
to
c
ollec
t
data
f
or
i
de
nti
ficat
ion
of
dem
og
ra
ph
ic
c
har
act
erist
ic
s.
Ra
ther
tha
n
ju
st
ob
ta
i
ning
i
m
ages,
the
syst
e
m
of
Yaho
o
is
capab
le
of
sou
nd
c
aptu
rin
g
thr
ough
the
use
of
m
ic
ro
phones,
t
o
colle
ct
keyw
ords
sp
oke
n
by
a
gro
up.
A
n
ad
diti
on
al
m
et
ho
d
us
e
d
to
ens
ur
e
at
te
ntion
is
the
ey
e
tr
acking
te
chn
iq
ue,
detect
i
ng
visio
n of
passe
rs
-
by usi
ng se
nsors
equi
pp
e
d on the
bill
boar
d.
NEC
Digital
Bi
ll
bo
ar
d
[2
]
is
desig
ne
d
sp
eci
fical
ly
to
disp
la
y
adv
erti
sem
e
nts
that
ref
le
ct
passer
s
-
by
per
s
onal
intere
sts.
NEC
Di
gital
Bi
l
lbo
ar
d
use
s
wireless
te
chnolo
gy
ta
gs
,
wh
ic
h
are
al
s
o
kn
own
as
R
adio
Fr
e
qu
e
ncy
I
de
ntific
at
ion
(RF
ID)
c
hip
s
.
As
nowa
days
RF
I
D
c
hip
s
are
inc
reasin
gly
bei
ng
inc
orp
orat
ed
su
c
h
as
cred
it
car
ds
an
d
m
ob
il
e
ph
on
es,
these
chi
ps
are
act
ing
li
ke
inv
isi
ble
la
bels
carried
by
pe
ople
al
l
the
way
they
go.
The
m
et
hod
is
that
these
chips
ar
e
enc
oded
with
inf
or
m
at
ion
about
ind
i
viduals,
s
o
the
dig
it
al
adverti
sing
bo
a
r
d
co
uld
id
entify
a
per
s
on
wh
e
n
they
pa
ss
by,
by
read
i
ng
the
ta
r
get'
s
RFID
data.
N
EC
Digital
Bi
l
lbo
a
rd
al
so
im
ple
m
ent
ed faci
al
r
ec
ogniti
on
t
o
i
den
ti
fy sho
pp
e
r'
s g
e
nd
e
r, et
hnic
it
y, an
d
a
ppr
ox
im
at
e age.
The
face
-
rec
ogniti
on
bill
boar
d
in
Lo
ndon
[3]
is
us
ed
by
a
gl
ob
al
child
ren’
s
char
it
y,
Plan
UK
in
thei
r
“B
ecause
I’
m
a
Girl”
ca
m
paign,
to
raise
awar
e
ness
f
or
e
qu
al
op
portu
ni
ty
and
acce
ss
to
ed
ucati
o
n
f
or
bo
t
h
sexes,
as
well
as
raisi
ng
f
und
to
sp
onsor
ed
ucati
on
for
girl
s
in
dev
el
opin
g
co
un
try
.
T
he
m
ai
n
pu
rpose
of
thi
s
face
-
recog
niti
on
bill
bo
ar
d
is
t
o
detect
ge
nd
e
r
an
d
s
how
it
s
entire
co
nte
nt
on
ly
to
w
om
en.
To
achie
ve
th
is,
th
e
bill
bo
a
rd
is
e
quip
ped
wi
t
h
a
“hig
h
def
i
niti
on
”
cam
era
to
s
can
people
fac
es,
detect
ing
t
he
ir
ge
nder
us
i
ng
face
recog
niti
on
te
c
hn
i
qu
e
,
with
a
high
su
cc
ess
rate.
T
he
ey
e
tracki
ng
te
ch
ni
qu
e
is
al
s
o
use
d
t
o
e
ns
ure
t
hat
th
e
ta
rg
et
ed
p
e
rs
on is lo
ok
i
ng at the
bill
bo
a
rd.
Astra
Girl
Det
ect
ion
Bi
ll
bo
ar
d
[
4]
is
locat
ed
ou
tsi
de
of
a
pub
i
n
Ham
bu
r
g,
Ger
m
any
as
pa
rt
of
a n
ew
adv
e
rtisi
ng
ca
m
paign
fo
r
As
tra.
Ra
ther
tha
n
just
fo
c
us
es
on
prom
oting
beer
to
w
om
e
n,
the
bill
boar
d
eve
n
sm
art
ly
avo
ids
the
youngs
te
r
s
under
t
he
le
gal
dr
i
nk
i
ng
a
ge
of
sixtee
n
.
W
it
h
a
bu
il
t
-
in
cam
era
and
the
i
m
ple
m
ented
gen
de
r
-
detect
io
n
softwa
re,
the
bill
bo
ar
d
of
Astra
is
capa
ble
of
detect
in
g
the
gende
r
of
peopl
e
lookin
g
at
it
, n
o
m
at
te
r
it
is an
in
div
i
du
al
or
a gro
up.
Lex
us
is
m
ov
i
ng
it
s
way
t
o
a
bette
r
a
ppr
oac
h
of
ad
ver
ti
sin
g
by
intr
oduci
ng
sm
art
dig
it
al
bill
bo
a
rds
to
pr
om
ote
the
cars
of
Le
xus
by
trig
ge
ring
a
per
s
onal
iz
ed
m
essage
to
dr
i
ver
s
c
orr
esp
onds
t
o
the
br
a
nd
,
m
od
el
,
and
col
our
of
the
ve
hicle
s
[5
]
.
In
order
to
ca
ptu
re
al
l
the
passing
traff
ic
,
Lex
us
bill
bo
ar
ds
rel
y
on
a
s
eries
of
high
ro
ta
ti
on
cam
er
as.
The
ca
ptur
ed
im
ages
are
sent
to
the
A
P
N
O
utdoor
Cl
assifi
er,
w
hic
h
is
in
charge
of
m
at
chin
g
them
to
it
s
databa
se
of
ve
hicle
m
akes,
m
od
el
s,
and
co
lours,
as
well
a
s
the
oth
er
va
ri
ables.
Pers
on
al
iz
ed
m
essage is
d
is
play
ed
f
or the
ta
r
geted ve
hicle
bei
ng
rec
ognise
d.
Ci
sco
is
placi
ng
this
co
nnect
ed
bill
boar
d
f
or
the
intenti
on
to
highli
gh
t
the
c
on
ce
pt
of
“t
he
In
te
r
net
of
ever
yt
hi
ng
”
in
adv
e
rtisi
ng
a
nd
to
showcase
it
s
latest
te
chn
ology.
The
Ci
sco
bill
bo
a
r
d
syst
e
m
[6
]
us
es
a
series
of
A
P
Is
c
onne
ct
ed
to
real
-
ti
m
e
traff
ic
se
nsors
t
o
get
the
t
raffic
co
nd
it
io
ns
,
i
n
c
onjun
ct
ion
with
th
e
usa
ge
of
m
aps
an
d
bac
k
-
e
nd
net
work.
A
m
essage
with
di
ff
e
ren
t
le
ng
th
is
dis
play
ed
base
d
on
the
ve
hicle
s
sp
ee
d,
sel
ect
ed
acco
r
di
ng
t
o
the
s
pee
d ran
ge
it
f
al
ls
i
nto
.
To
gether
with
the
sm
art
data
stora
ge
com
pan
y
Cl
oudian,
a
Japan
ese
a
dve
rtisi
ng
com
pany
Den
tsu
is
a
pro
gram
cor
r
esp
onding
t
o
i
ntell
igent
bill
bo
ar
ds
has
bee
n
la
unche
d
[
7].
The
syst
em
includes
the
abil
it
y
to
analy
se
traf
fic
volum
es
to
enab
le
highly
e
ff
ect
ive
ta
rg
et
ed
r
oa
ds
ide
ad
ver
ti
sin
g.
The
bill
bo
a
rd
syst
em
s
are
i
m
ple
m
ented
us
in
g
big
data
and
dee
p
le
a
rn
i
ng.
Dee
p
le
anin
g
an
al
ysi
s
is
carried
out
at
it
s
first
sta
ge
to
at
tribu
te
the
re
cogniti
on
with
autom
at
ic
feat
ur
e
e
xtracti
on
of
tra
ff
ic
patte
rn
s
a
nd
volum
e
,
an
d
al
so
a
uto
m
atic
veh
ic
le
recog
niti
on
.
3.
INTEL
LIGE
NT TA
RGET
ED AD
VERT
ISING S
YS
T
EM
This
sect
ion
de
scribes
the
sy
stem
flow
and
trai
nin
g
arc
hitec
ture
of
intel
li
gen
t
ta
rg
et
e
d
adv
e
rtisi
ng
syst
e
m
.
3.1.
S
ystem
Architec
tu
re
The
syst
e
m
flo
w
ca
n
be
basic
al
ly
div
ide
d
i
nto
f
our
sta
ge
s,
nam
el
y
i
m
age
a
cqu
isi
ti
on,
m
ulti
ple
obj
ect
detect
ion, m
ajo
rity
grou
p bid
ding a
nd tar
geted
a
dv
e
rtisi
ng. T
he
syst
em
f
low
is i
ll
us
trat
e
d i
n
Fig
ure
1.
At
the
init
ia
l
st
age,
im
ages
are
captur
e
d
by
vid
e
o
cam
era
and
se
rv
e
d
as
input
f
or
the
sy
stem
.
These
i
m
ages
underg
o
processi
ng
in
the
syst
e
m
and
are
pass
ed
to
the
al
go
rithm
s
in
la
te
r
sta
ge.
At
the
seco
nd
sta
ge,
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
AI
-
B
as
e
d Tar
ge
te
d
A
dverti
sin
g System
(
Tew
Ji
a
Y
u
)
789
m
ul
ti
ple
obj
ect
detect
io
n
is
de
plo
ye
d
to
dete
ct
diff
e
re
nt
ty
pe
s
of
obj
ect
s
i
n
a
sin
gle
im
age
or
f
ram
e.
Detect
ed
faces
a
nd
ve
hi
cl
es
are
f
urt
her
recog
nised
th
e
ge
nd
e
r
a
nd
a
ge,
or
the
cat
e
gories
re
sp
ect
i
vely
.
Th
e
rec
ogniti
on
processes
a
re
carried
ou
t
ba
sed
on
the
pre
-
trai
ne
d
m
od
el
s.
An
a
naly
sis
is
do
ne
to
de
te
rm
ine
the
m
ajorit
y
gro
up at that
ti
m
e p
oin
t
a
nd t
he results a
re i
nput int
o
th
e
bid
di
ng alg
or
it
hm
f
or
adve
rtise
m
ent sele
ct
ion
.
The
biddin
g
outp
ut
is
tra
nsm
itted
to
th
e
database
co
nta
ining
the
a
dve
rtisem
ents
in
t
he
nex
t
sta
ge.
The
ch
os
e
n
co
ntent
is
retriev
ed
from
database
and
the
n
tr
ansf
e
rr
e
d
to
th
e
dig
it
al
bill
board
to
be
dis
play
ed
.
The
fi
nal
sta
ge
is
the
ad
ver
ti
s
e
m
ent
releva
nt
to
the
real
-
ti
m
e
dem
og
ra
ph
ic
is
disp
la
ye
d.
The
process
it
erates
wh
e
n
t
he pr
of
il
e of
detect
ed
c
r
owd or t
ra
ff
ic
changes
.
Figure
1
.
Th
e
s
yst
e
m
f
low
of intel
li
gen
t t
ar
ge
te
d
ad
ve
rtisi
ng syst
e
m
3.2
.
Tr
aining
A
rchi
tectu
re
s
The
m
od
el
s
use
d
f
or
rec
ogni
ti
on
are
pr
e
-
tr
ai
ned
before
be
ing
im
ple
m
ented
into
t
he
syst
e
m
.
The
m
od
el
fo
r
re
c
ognisin
g
ge
nd
er
an
d
age
a
nd
the
m
od
el
f
or
recog
nisin
g
var
io
us
ty
pes
of
obj
e
ct
s
are
trai
ned
separ
at
el
y
by
diff
e
re
nt
arc
hitec
tures.
F
or
ge
nd
e
r
a
nd
age
r
ecognit
ion,
fac
e
casca
des
intr
oduce
d
by
Mi
cro
s
oft
is used
as the
f
ace detec
ti
on
fram
ewo
rk. I
ts
functi
on is to det
ect
f
aces fr
om
ca
m
era
i
m
a
ges
f
or
gende
r
and
a
ge
recog
niti
on
.
T
he
ide
a
[
8]
be
hind
this
arc
hitec
ture
is
to
com
bin
e
face
al
ign
m
ent
with
detect
ion.
Pr
el
i
m
inary
stud
ie
s
sho
wed
that
al
ig
ned
fa
ces
are
able
t
o
pro
vid
e
bette
r
featur
e
s
to
e
nhance
face
cl
ass
ific
at
ion
proce
ss.
I
n
the
casca
de
f
ra
m
ewo
r
k,
bo
os
t
ed
casca
de
str
uc
ture
an
d
si
m
p
le
featur
es
pri
nc
iples
are
i
m
pl
e
m
ented
to
en
han
ce
the
detect
io
n
eff
ic
ie
ncy.
As
in
[
9],
bo
os
ti
ng
is
per
f
orm
ed
on
t
ho
se
sim
pl
e
cl
assifi
ers,
or
in
ot
her
w
ords,
th
e
weak
cl
assifi
er
s
extracte
d
are
com
bin
ed
for
bette
r
perf
or
m
ance
com
par
ed
to
the
si
m
ple
cl
assif
ie
rs
al
one.
I
n
[8
]
,
t
he
cas
cad
e
detect
or
not
only
ta
kes
shorter
ti
m
e
fo
r
face
detect
ion,
it
al
so
outpe
r
form
s
oth
er
sim
il
ar
so
luti
ons i
n detec
ti
on
unde
r
c
halle
ng
i
ng c
onditi
on
s
su
c
h
a
s
poor li
ghti
ng
s
, l
arg
e
view
poin
ts, an
d occlusi
on.
Fo
r
t
he
m
ulti
p
le
obj
ect
recogn
it
io
n,
the
M
ob
il
eNet
arc
hit
ect
ur
e
is
us
e
d
to
trai
n
the
im
ple
m
ented
m
od
el
s.
As
de
scribe
d
in
[
10]
,
Mob
il
eNet
is
a
li
gh
t
weig
ht
deep
neural
ne
tworks
a
rch
it
ect
ur
e,
wh
ic
h
is
bu
il
t
us
in
g
de
pt
h
-
w
ise
separ
a
ble
conv
olu
ti
ons,
or
known
a
s
f
act
or
ise
d
c
onvoluti
o
ns
.
M
ob
i
le
Net
consi
sts
of
28
la
ye
rs,
com
pr
i
sing
de
pth
-
wis
e
an
d
po
i
ntwi
se
co
nvol
ution
la
ye
rs.
On
ly
the
fir
st
la
ye
r
of
the
Mo
bileNet
structu
re
is
bu
il
t
on
fu
ll
c
on
vo
l
ution
j
us
t
li
ke
oth
e
r
ty
pic
al
ly
seen
ne
ur
al
netw
orks.
I
n
this
arc
hitec
ture,
a
sta
nd
a
rd co
nvo
luti
on
is
facto
ri
sed
int
o
tw
o di
ff
e
ren
t c
onvolu
ti
on
s, nam
el
y th
e
dep
t
h
-
wise
conv
olu
ti
on a
nd the
1
×
1
point
w
ise
conv
olu
ti
on.
In
de
pth
-
w
ise
conv
olu
ti
on,
in
put
cha
nnel
s
are
filt
ered
but
no
t
i
nst
antly
com
bin
ed
t
o
cr
eat
e
ne
w
feat
ures.
It
r
eq
uires
an
ad
diti
on
al
la
ye
r,
w
hich
is
the
pointwise
conv
olu
ti
on
la
ye
r
to
com
pu
te
a
li
ne
ar
com
bin
at
ion
of
the
outp
ut
of
de
pth
-
wise
conv
olu
ti
on
vi
a
the
1
×
1
con
voluti
on.
By
hav
i
ng
two
s
epa
rate l
a
ye
rs,
the
co
m
pu
ta
ti
on
s
, m
od
e
l si
ze, and c
ompu
ta
ti
onal
c
os
t
are m
uch
r
e
du
ced.
Mob
il
e
Net
is
use
d
to
trai
n
on
Com
m
on
Obje
ct
s
in
Con
te
xt
(COCO
)
dataset
fo
r
obj
ect
de
te
ct
ion
an
d
recog
niti
on
in
the
propose
d
syst
e
m
.
The
COCO
datas
et
is
pr
esente
d
by
Mi
cro
s
oft
m
ai
nly
fo
r
obj
ect
recog
niti
on
.
T
he
dataset
co
nt
ai
ns
sam
ple
p
ho
t
os
of
91
obj
ect
cat
eg
or
i
e
s,
inclu
ding
al
l
the
cat
ego
rie
s
from
PA
SC
AL
V
O
C
ad
su
pe
r
cat
egories
as
in
[
11
]
.
I
n
this
da
ta
set
,
sh
ape
m
ask
is
us
e
d
to
detect
obj
ect
s
us
in
g
boundi
ng bo
x appr
oach,
provi
din
g a m
or
e a
ccur
at
e m
easure
of the
artic
ul
at
ed
ob
j
ect
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
3
, N
o.
2
,
Fe
bru
ary
201
9
:
7
8
7
–
7
9
3
790
Othe
r
than
de
te
ct
ing
an
d
re
cog
nisin
g
ob
je
ct
s,
si
m
il
arity
scor
es
of
the
rec
ognise
d
obj
ect
s
are
com
pu
te
d
in
t
he
syst
em
.
The
scor
i
ng
funct
ion
is
im
ple
m
e
nted
by
us
in
g
the
Sin
gle
Sho
t
Mult
iB
ox
De
te
ct
or
(S
S
D)
arc
hitec
ture.
I
n
a
rece
nt
stud
y
i
n
[12],
f
or
each
detec
te
d
obj
ect
in
a b
ou
nd
i
ng b
ox
,
p
re
dicti
ng
sc
ores
ar
e
com
pu
te
d
for
e
ach
obj
ect
cat
e
gory.
Adjustm
ents
are
the
n
pe
rfor
m
ed
on
t
he
boun
ding
box
to
bette
r
m
atch
the
obj
ect
sh
a
pe.
S
SD
is
al
so
able
to
encap
s
ulate
al
l
co
m
pu
ta
ti
on
into
a
sin
gle
netw
ork.
T
his
m
akes
SSD
ea
sy
to
trai
n
an
d
le
ss
com
plex
to
be
integrate
d
int
o
the
syst
em
fo
r
detect
ion
pur
poses.
D
ur
i
ng
t
he
SS
D
trai
ning,
for
each
obj
ect
in
vo
l
ved,
it
on
ly
req
uires
a
n
input
i
m
age
a
nd
groun
d
trut
h
boxes
for
th
e
detect
ion
.
A
t
each
locat
ion
,
t
her
e
is
an
evaluati
on
of
t
he
de
fa
ult
boxes
of
di
ff
ere
nt
aspect
rati
os
with
diff
e
ren
t
scal
es.
The
evaluati
on
is
bein
g
processe
d
in
s
eve
ral
f
eat
ur
e
m
aps
[
13
]
.
These
de
f
ault
boxes
are
then
m
at
ched
to
the
gro
und
tr
uth [
14]
bo
xes
in
the
trainin
g p
hase.
3.3
.
Ge
nder
an
d
Age
Rec
ogniti
on
In
ge
nder
a
nd
age
recog
n
it
ion
process
,
face
i
m
ages
are
a
c
qu
i
red
by
t
he
syst
e
m
as
inpu
t
data.
T
he
i
m
ages
are
pre
-
proce
ssed
a
nd
passe
d
to
the
face
detect
ion
functi
ons.
The
exact
face
po
sit
io
n
w
il
l
be
com
pu
te
d
a
nd
crop
ped
out
f
r
om
the
unnece
ssary
bac
kgr
ou
nd
t
o
op
ti
m
ise
the
rec
ogniti
on
pr
ocess.
C
r
oppe
d
face
im
age
then
un
dergo
e
s
the
featu
re
e
xtracti
on
proc
ess.
Sig
nifica
nt
featur
e
po
i
nt
s
are
extract
ed
by
Mi
cro
s
of
t
face
casca
des
al
go
rithm
to
form
a
com
plete
fac
e
m
ap.
The
ob
ta
ined
face
m
ap
is
a
naly
sed
a
nd
the
ou
t
pu
t
resu
lt
s
con
t
rib
uted
to
the
cl
assifi
cat
ion
process
.
T
he
syst
e
m
finally
gen
er
at
es
the
pr
e
dicte
d
gender
a
nd
age
of that
par
t
ic
ular
fa
ce.
In
the
syst
em
,
gender
a
nd
a
ge
rec
ogniti
on
wer
e
im
ple
mented
usi
ng
Mi
cro
s
of
t
Fac
e
AP
I
.
T
he
Mi
cro
s
of
t
Fac
e
AP
I
off
ers
a
wide
ra
ng
e
of
functi
on
al
it
ie
s
include
d
face
identific
at
ion
,
si
m
i
la
r
face
search
,
and
f
ace
gro
uping
. O
nly
two
f
ace
at
tribu
te
s
are
co
nf
ig
ured
t
o
retu
rn
thei
r
va
lues
as
requir
ed,
an
d
the
ret
urn
of
face
ID
a
nd
fa
ce
la
nd
m
ark
s
va
lues
are
disa
bl
ed.
The
ca
ptur
ed
i
m
ages
are
tran
sm
itted
over
the
In
te
rn
et
to
the
Mi
cro
s
of
t
Co
gnit
ive
ser
ver
f
or
recog
niti
on
,
and
t
he
res
ults
are
retrie
ved
in
a
parsed
li
s
t,
wh
ic
h
inclu
de
s
the
par
am
et
ers
of
gende
r
a
nd ag
e
. S
am
ple r
esult
s of
gende
r
a
nd ag
e
rec
ogniti
on a
re
disp
la
ye
d i
n
Fi
gure
2.
Figur
e
2. The
s
a
m
ple o
ut
put o
f gen
der an
d
a
ge reco
gnit
ion
3.4
.
Multiple
Ob
ject Rec
ognition
Mult
iple
obj
ec
t
reco
gnit
ion
a
i
m
s
to
detect
a
nd
recog
nise
di
ff
ere
nt
kinds
of
obj
ect
s
in
r
eal
tim
e
by
us
in
g
Te
nsor
flow
rea
dily
trai
ned
obje
ct
m
o
dels.
Vide
o
im
ages
ar
e
ca
pt
ured
an
d
passe
d
t
o
the
syst
em
as
input
data.
T
he
ca
pt
ur
e
d
im
ages
w
il
l
un
de
r
go
pre
-
proce
ssin
g
s
uc
h
as
resizi
ng.
Using
the
pre
-
t
raine
d
m
od
el
s
loade
d
into
the
syst
e
m
,
sign
ific
ant
featur
e
s
are
e
xt
racted
f
ro
m
t
he
te
st
vid
e
o
f
ram
es
and
m
atch
ed
with
the
trai
ne
d
obj
ect
m
od
el
s.
The
e
xtracted
featur
e
s
a
re
th
en
cl
assifi
e
d
as
the
m
os
t
sim
i
l
ar
obj
ect
cl
ass
.
Cl
asses
in
volv
ed
i
n
the
syst
e
m
are
car,
bicy
cl
e,
m
oto
rcycl
e,
bus,
truc
k,
anim
al
,
bag
,
um
br
el
la
,
ti
e,
su
it
cas
e,
bo
tt
le
,
fruit
,
fo
od
,
la
pto
p, cel
l
phon
e
, a
nd bo
ok.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
AI
-
B
as
e
d Tar
ge
te
d
A
dverti
sin
g System
(
Tew
Ji
a
Y
u
)
791
The
m
od
el
i
m
plem
ented
in
t
he
syst
e
m
is
t
he
SS
D
Mob
il
eNet
COCO
m
od
el
.
T
his
m
o
del
is
trai
ned
us
in
g
the
co
nvol
ution
al
ne
ural
networ
ks
on
the
Mi
cro
s
oft
COCO
data
set
.
The
num
ber
of
ob
j
ect
cl
asses
def
i
ned
in
this
trai
ned
m
od
el
is
90
.
For
det
ect
ing
obj
ect
s
,
SSD
i
s
i
m
ple
m
ented
as
the
al
gorithm
that
detect
s
obj
ect
s
i
n
im
a
ges
us
in
g
a
sing
le
dee
p
neural
net
work,
by
pu
tt
in
g
boundi
ng
bo
xes
over
detect
ed
obj
ect
featur
e
s
acco
rdi
ng
to
t
he
featu
re
m
ap.
The
ne
twork
will
then
ge
nerat
e
scores
for
each
obje
ct
cat
ego
ry
in
ea
c
h
b
ox
an
d
f
ur
the
r
produce
ad
ju
st
m
ents
to
the
box
t
o
bette
r
m
at
ch
the
obj
e
ct
sh
a
pe,
as
de
scribe
d
i
n
[4
]
.
Labels
for
eac
h object
class i
s loa
de
d i
nto
t
he
syst
e
m
as a f
il
e ty
pe
acce
ssible
by t
he
Te
nsor
flo
w t
echnolo
gy.
In
orde
r
t
o
perform
reco
gnit
ion
pr
ocess,
th
e
captu
re
d
a
nd
save
d
vid
e
o
fra
m
e
is
conve
rted
int
o
data
arr
ay
.
A
sessi
on
is
c
reated
f
or
a
ne
w
gra
ph
of
e
xecu
ti
on
and
res
ources
al
locat
ion
.
Ne
cessary
va
riabl
es
are
init
ia
li
sed
in the s
ession. For Ten
s
orflo
w
co
m
pu
ta
ti
on
p
ur
po
s
es, the
ar
ra
y of
i
m
age is e
xp
a
nd
ed by ad
ding the
m
issi
ng
dim
ension
s
re
qu
ire
d
for
ser
ving
the
te
ns
or
as
in
put
to
the
f
unct
ions.
D
ur
i
ng
t
he
s
ession
r
unning
tim
e,
the
co
nf
i
dence
value
of
the
de
te
ct
ed
obj
ect
cl
ass
is
retu
rn
e
d
in
a
vect
or
,
corres
pondin
g
to
the
in
de
x
of
cl
ass
la
bels in
t
he
m
od
el
set
up
proc
ess. T
he
sam
ple outp
ut of m
ulti
ple o
bject
rec
ogniti
on is il
lu
strat
ed
in
Fig
ur
e 3
.
Figure
3. The
s
a
m
ple o
ut
of m
ulti
ple ob
j
ect
re
cogniti
on
3.5
.
Advertis
ement
Sele
ctio
n
The
sel
ect
io
n
of
a
dverti
sin
g
vid
e
o
is
base
d
on
t
he
m
axi
m
u
m
nu
m
ber
of
t
he
rec
og
nised
obj
ect
cat
egories.
As
the
ge
nder
a
nd
age
recog
niti
on
a
nd
m
ulti
ple
obj
ect
recog
niti
on
a
re
im
ple
m
ented
with
dif
f
eren
t
AP
I
s
a
nd
m
odel
s,
the
rec
ogni
ti
on
re
su
lt
s
a
r
e
retrie
ved
se
par
at
el
y.
F
or
i
ns
ta
nce
,
if
m
os
t
ad
ult
w
om
e
n
a
re
recog
nised
at
a
tim
e
po
int,
a
dv
e
rtise
m
ent
relevan
t
to
t
his
gro
up
will
be
disp
la
ye
d
on
the
bill
boar
d.
On
t
he
oth
e
r
ha
nd,
if
t
he
num
ber
of
recog
nised
obje
ct
s
is
gr
eat
er
than
the
num
ber
of
rec
ognise
d
pe
rs
on,
the
obj
ect
cat
egory with
the
gr
eat
est
num
ber
w
il
l be
re
ferred
to
sel
ec
t
an
a
dverti
sem
ent r
el
at
ed
to
it
.
4.
RESU
LT
S
A
ND AN
ALYSIS
The
m
od
el
s
f
or
rec
ogniti
on
wh
ic
h
are
im
p
lem
ented
in
th
e
syst
e
m
is
trai
ned
us
in
g
t
he
co
ncep
t
of
m
achine
le
arni
ng
.
T
he
first
ste
p
for
m
ac
hin
e
le
ar
ning
is
data
acqu
is
it
ion
.
Ra
w
da
ta
are
colle
ct
ed
an
d
cl
assifi
ed
int
o
three
s
et
s,
nam
el
y
trai
nin
g
set
,
vali
dation
set
,
an
d
te
sti
ng
se
t,
ty
pical
ly
with
the
per
ce
nta
ge
of
70,
20,
a
nd
10.
All
three
data
set
s
are
ge
ner
a
te
d
ra
ndom
ly
a
nd
c
onsist
of
s
a
m
ples
from
all
the
ou
t
pu
t
cl
a
sses
to
ens
ur
e
ef
fici
ent
trai
ning.
T
he
trai
ni
ng
set
is
us
ed
t
o
trai
n
the
m
od
el
s
for
rec
ogniti
on,
validat
io
n
set
to
tu
ne
the
m
od
el
par
a
m
et
ers
to
m
ini
m
ise
the
ou
tp
ut
err
or
rates,
a
nd
te
sti
ng
set
to
assess
the
perf
or
m
ance
of
t
he
final
m
od
el
.
The
refi
ned
a
nd
c
om
plete
d
m
od
el
s
are
finall
y
placed
i
nto
a
pp
li
ca
ti
on
.
Re
co
gnit
ion
is
no
w
bas
ed
on
the n
e
w data
from
the r
eal
wo
rld.
Test
ing
ens
ur
es
the
le
vel
of
pe
rfor
m
anc
e,
sta
bili
ty
,
a
nd
acce
ptance
,
th
us
br
i
ngs
sig
nificant
i
m
pr
ovem
ents
and
ref
i
nem
ents
to
the
syst
em
.
Beta
te
sti
n
g
wa
s
co
nduct
ed
f
or
t
he
syst
e
m
by
real
so
f
tware
us
ers
t
o
ens
ure
that
the
syst
e
m
can
handle
the
require
d
and
si
gn
i
ficant
ta
sk
s
in
real
-
world
sce
nar
i
os.
The
syst
e
m
h
as p
as
sed
al
l t
he
test
s and the
test
f
ie
lds ar
e
li
ste
d
in
Tab
le
1.
The
te
st
co
nducted
on
t
he
s
yst
e
m
is
div
ided
into
te
n
f
ie
lds.
A
s
the
c
ore
aspects
of
t
he
syst
em
,
detect
ion
an
d
recog
niti
on
f
unct
ions
ar
e
ra
nk
e
d
a
s
to
p
pri
or
it
ie
s
to
be
te
ste
d.
T
her
e
sh
oul
d
be
s
uc
cessf
ul
detect
ion
in
a
n
acce
pta
ble
di
sta
nce
and
t
he
recogn
it
io
n
accuracy
f
or
t
he
rec
ogniti
on
processes
s
hould
be
abo
ve
80
pe
rc
ent.
F
or
cam
era
i
m
age
capturi
ng
,
vid
e
o
str
e
a
m
ing
,
s
uitable
adv
e
rtise
m
ent
disp
la
ys,
an
d
pro
per
resu
lt
text
displ
ay
s,
cl
ear r
es
ol
ution
s a
re
require
d
as
well
as er
r
or
-
fr
ee
pro
cesses. T
he pr
oc
ess of
detect
ion a
nd
recog
niti
on
is
al
so
r
e
quire
d
t
o be c
o
m
plete
d
in a
n
acce
ptab
le
tim
e p
eriod.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
3
, N
o.
2
,
Fe
bru
ary
201
9
:
7
8
7
–
7
9
3
792
Table
1.
Res
ults o
f
Be
ta
Testi
ng
ID
Test Field
Exp
ected R
esu
lts
Pass
/
Fail?
1
Face detectio
n
Ab
le to d
etect hu
m
an
f
aces in an
acc
e
p
tab
le dis
tan
ce
Pass
2
Gen
d
er
an
d
age
recog
n
itio
n
Ab
le to p
erfo
r
m
ge
n
d
er
an
d
age
recog
n
ito
in
bas
ed
on
det
ected f
aces witho
u
t
an
y
er
ror
Reco
g
n
itio
n
acc
u
racy
is
abo
v
e 80
%
Pass
3
Veh
icle t
y
p
e
recog
n
itio
n
Ab
le to p
erfo
r
m
ve
h
icle ty
p
e
recog
n
itio
n
bas
ed
on
detected
veh
icles with
o
u
t
an
y
er
ror
Reco
g
n
itio
n
acc
u
racy
is
abo
v
e 80
%
Pass
4
Variou
s o
b
ject
catego
ry
recog
n
itio
n
Ab
le to p
erfo
r
m
re
co
g
n
itio
n
bas
ed
o
n
detected
ob
jects w
ith
o
u
t any
err
o
r
Reco
g
n
itio
n
acc
u
racy
is
abo
v
e 80
%
Pass
5
Multip
le ob
ject
d
etectio
n
Ab
le to d
etect a
n
d
recog
n
ise
m
u
ltip
le
k
in
d
s o
f
ob
jects o
n
the sa
m
e
i
m
ag
e
S
m
o
o
th
pro
cess
wi
th
o
u
t er
ror
Pass
6
Ca
m
er
a i
m
ag
e
capt
u
ring
Cap
tu
red i
m
ag
es a
re
clea
r
S
m
o
o
th
pro
cess
wi
th
o
u
t any
d
elay
or
err
o
r
Pass
7
Vid
eo
strea
m
in
g
Vid
eo
s d
isp
lay
ed
in acceptab
le r
eso
lu
tio
n
S
m
o
o
th
pro
cess
wi
th
o
u
t any
d
elay
or
err
o
r
Pass
8
Disp
lay
o
f
ap
p
rop
riate
ad
v
ertise
m
en
t
Relev
an
t adv
ertisem
e
n
t of
the largest
detected
ob
ject ca
teg
o
ry
is sele
cted
a
n
d
d
isp
lay
ed
Selection
based
on
r
eal
-
ti
m
e
de
m
o
g
ra
p
h
ics
Pass
9
Disp
lay
of
pro
p
er
resu
lt text
Text d
isp
lay
ed
on
syste
m
in
terface is
bas
ed
on
r
ecog
n
itio
n
r
esu
lts
Sh
o
ws an
acc
u
rate
nu
m
b
e
r
o
f
detecte
d
perso
n
or ob
jects
Up
to d
ate with th
e
r
eal
-
ti
m
e
r
e
co
g
n
itio
n
r
esu
lts
Pass
10
Reco
g
n
itio
n
sp
eed
The p
rocess
of
detectio
n
and
r
ecog
n
itio
n
is in an
acc
e
p
ta
b
le ti
m
e
perio
d
Pass
5.
CONCL
US
I
O
N
This
pa
per
pr
esents
a
n
i
ntell
igent
ta
r
geted
a
dverti
sin
g
syst
e
m
that
ai
m
s
at
pr
ovidi
ng
a
bette
r
adv
e
rtisi
ng
e
xperie
nce
f
or
both
the
a
dv
e
rtise
r
an
d
the
a
udie
nce.
T
he
in
te
ll
igent
ta
rg
et
ed
ad
ver
ti
sin
g
syst
e
m
consi
sts
of
sev
eral
integ
rated
functi
onal
it
ie
s,
inclu
ding
ge
nder
a
nd
ag
e
re
cogniti
on,
veh
i
cl
e
ty
pe
recog
niti
on,
and
m
ulti
ple
ob
j
ect
detect
io
n.
The
m
ulti
ple
obj
ect
detect
io
n
te
ch
nolo
gy
pro
vid
es
the
ca
pa
bili
ti
es
of
detect
ing
diff
e
re
nt
kinds
of
obj
ect
s
on
a
sing
le
i
m
ag
e,
and
f
ur
t
her
enab
li
ng
the
de
te
ct
ion
and
r
ecognit
ion
of
hu
m
an
faces,
veh
ic
le
s,
and
va
rio
us
ki
nd
s
of
ob
j
ect
s.
Faci
al
recogn
i
ti
on
is
im
ple
mented
to
rec
ognise
ge
nder
a
nd
ag
e
base
d
on
faci
al
featur
es.
Mult
iple
obj
ect
recog
niti
on
te
chnolo
gy
is
us
ed
f
or
ve
hicle
t
ypes
and
di
ff
ere
nt
cat
egories o
f o
bj
ect
rec
ogniti
on b
ase
d o
n
th
ei
r
uniq
ue
c
harac
te
risti
cs. A
ll
of
t
he
m
od
el
s
us
e
d
in t
he
syst
e
m
f
or
recog
niti
on
a
r
e
pr
e
-
trai
ne
d
with
the
c
onc
ept
of
m
achine
le
arn
in
g
for
highly
accu
ra
te
resu
lt
s
a
nd
bette
r
perform
ance.
W
it
h
th
e
syst
e
m
abili
t
y
to
disp
la
y
ta
rg
et
ed
adv
e
rtise
m
ent
con
te
nt,
it
benefit
s
the
ad
ver
t
ise
rs
a
s
they
co
uld
sig
ni
ficantl
y
reduc
e
their
prom
otion
al
c
os
ts
due
to
the
ef
fecti
ve
ness
of
ta
r
gete
d
a
dv
e
rtisi
ng.
As
for
the
aud
ie
nce,
t
hey
will
be
ex
po
s
ed
to
c
on
te
nt
that
a
re
m
o
re
releva
nt
to
them
,
and
be
offe
red
product
s
and
serv
ic
es
that th
ey
m
igh
t re
qu
i
re.
ACKN
OWLE
DGE
MENTS
Re
search
re
ported
i
n
this
pa
per
was
sup
ported
by
Mult
im
edia
Un
i
ver
s
it
y
Mi
ni
Fu
nd
,
G
ran
t
N
o.
MM
UI
/1
70032
an
d
MM
U
I
\
180182.
The
resea
rch
w
as
al
so
spo
nsore
d
by
M
MU
CA
PEX
F
un
d
MM
UI
/C
AP
E
X 180
011. T
he
Tit
an Xp use
d f
or
t
his r
e
sear
ch was
donate
d by the
N
VID
IA
C
orp
or
at
io
n.
REFERE
NCE
S
[
1
]
Kim
,
E.
(2016).
“
Yahoo
has
an
ide
a
for
“
sm
art
”
bil
lbo
ard
th
at
m
onit
ors
peopl
e
with
a
ca
m
er
a
and
m
ic
rophone
|
Business
Insider
Malay
si
a”
.
Re
tri
ev
ed
Jul
y
19
,
2017,
from
ht
t
p:/
/www
.
busines
sinsider.
m
y
/
y
ah
oo
-
pat
en
ts
-
sm
art
-
bil
lboa
rd
-
with
-
c
amera
-
and
-
m
ic
r
ophone
-
2016
-
10
/
?
r=US&IR=
T
[
2
]
Digit
al
Signage
.
(n.
d.
)
.
Ret
ri
ev
ed
Mar
ch
1,
2018,
from
htt
p://m
y
.
n
ec.c
o
m
/e
n_MY
/solut
ions/di
git
a
l_sign
age
/
inde
x.
h
tml
[
3
]
“
Faci
al
Re
cognition
Bil
lbo
ard
Onl
y
L
et
s
W
om
en
See
The
Full
Ad”.
(n.
d
.
)
.
Ret
ri
eve
d
Ma
rch
1,
2018,
fr
om
htt
ps://
ww
w.psfk.c
om
/2012/
02/f
ac
i
al
-
re
cognition
-
bil
lboard.ht
m
l
[
4
]
Hutchi
ngs, E.
(2
015).
“
Gende
r
-
D
et
e
ct
ion
Bil
lboa
r
d
Advert
ises
B
e
er
Onl
y
to
W
o
m
en”
.
Ret
r
ie
v
ed
J
ul
y
22,
2017
,
fro
m
htt
ps://
ww
w.psfk.c
om
/2015/
06/
a
stra
-
bee
r
-
g
ende
r
-
det
e
ct
ion
-
b
il
lbo
ard
-
adve
r
ti
ses
-
to
-
wom
en.
html
[
5
]
Foote,
B.
(20
17).
“
5
E
y
e
-
Popping
Le
xus
Bil
lboa
rds
–
Clubl
exus”
.
Ret
ri
eve
d
Jul
y
22,
2017,
fr
om
htt
ps://
ww
w.clu
ble
xus.c
om
/how
-
tos/
slide
show
s/5
-
e
y
e
-
popping
-
l
exus
-
bil
lbo
ard
s
-
468623#1
-
thi
s
-
is
-
the
-
new
-
l
exus
[
6
]
Finebe
rg,
S.
(20
14).
“
Cisco
La
u
nche
s
Connecte
d
Bil
lboa
rd
in
San
Franc
isco
|
BtoB
–
AdA
ge”
.
Ret
ri
eve
d
Jul
y
2
2,
2017,
from
ht
tp:
/
/a
dag
e.com/art
i
c
le
/bt
ob
/c
is
co
-
la
u
nche
s
-
connect
ed
-
bil
lboa
rd
-
san
-
fr
anc
isco
/293168/
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
AI
-
B
as
e
d Tar
ge
te
d
A
dverti
sin
g System
(
Tew
Ji
a
Y
u
)
793
[
7
]
Mell
or,
C.
(201
6).
“
Cloudi
an
clobbers
ca
r
drivers
with
ta
rge
t
ed
ads
•
The
Regi
ster”
.
Re
trieve
d
Augus
t
24,
2017,
from
htt
ps://
ww
w.t
her
eg
iste
r
.
co
.
uk/2016/
06/21/c
loudi
an_
coul
d_c
lobbe
r_c
ar_
driv
es_with_t
arg
et
e
d_ads/
[
8
]
Chen,
D.
,
Ren
,
S.,
W
ei
,
Y.,
C
ao
,
X.,
&
Sun,
J.
(
2014).
“
Joint
cas
ca
de
f
ac
e
detec
ti
on
and
al
ignme
nt”
.
Le
ct
ure
No
t
es
in
Computer
Sci
en
ce
(
Inc
luding
Subseries
L
ec
ture
Not
es
in
Arti
fi
c
ial
Int
el
li
g
ence
and
Lect
ure
Note
s
in
Bi
oinf
orm
atics)
,
8694
LN
CS
(PART
6), 109
–
122.
htt
ps:/
/doi.o
rg/1
0.
1007/978
-
3
-
31
9
-
10599
-
4_8.
[
9
]
Ferre
ira,
a
J.,
&
Figuei
red
o
,
M.
a
T.
(2012).
“
Boosting
A
lgori
thms
:
A
R
evi
ew
of
Methods,
The
or
y
,
a
nd
Applic
a
ti
ons”.
Ensemble
Mac
hine
Le
arn
ing:
Methods
and
Applic
a
ti
ons,
35.
htt
ps://
do
i.
org/1
0.
1007/9781441
993267_2
[
1
0
]
How
ard
,
A.
G.,
Zhu,
M.,
Ch
en,
B.
,
Kal
eni
ch
enk
o,
D.,
W
ang,
W
.
,
W
e
y
and,
T
.
,
…
Adam
,
H.
(2017).
“
Mobile
Nets:
Eff
icient Convol
uti
onal Neural
N
et
works
for
Mob
il
e
Vision
Appli
ca
t
ions”. ht
tps:
//doi.org/arXiv:
17
04.
04861.
[
1
1
]
Li
n,
T
.
Y.,
Mair
e,
M.,
Bel
ong
ie,
S.,
Ha
y
s,
J.
,
Perona
,
P.,
Ramana
n,
D.,
…
Zi
tnick
,
C.
L.
(2014).
“
Microsoft
COCO
:
Com
m
on
obje
ct
s
in
cont
ex
t”.
Lect
ure
No
te
s
in
Computer
Sci
en
ce
(
Inc
ludi
ng
Su
bseries
Lect
ure
Note
s
in
Artif
i
c
i
a
l
Inte
lligen
ce
and
Le
ct
u
re
Notes
i
n
Bioi
nform
at
ics
),
8693
LNCS(P
ART
5),
740
–
755.
htt
ps
://doi
.
org/10.
1007/978
-
3
-
319
-
10602
-
1_48.
[
1
2
]
Li
u,
W
.
,
Angue
lov,
D.,
Erh
an,
D.,
Szege
d
y
,
C
.
,
Ree
d,
S.
,
Fu,
C.
Y.,
&
Berg,
A.
C.
(2016).
“
SS
D:
Single
shot
m
ult
ibox
detec
to
r”.
Le
ct
ure
Not
e
s
in
Computer
S
ci
en
ce
(
Inc
ludi
n
g
Subseries
Lectur
e
Note
s
in
Artific
ia
l
Int
e
lligen
c
e
and
Lect
ure
Not
es
in Bi
oin
formatic
s)
,
9905
LN
C
S
,
21
–
37
.
ht
tps:/
/
doi.
org/10
.
1007/
978
-
3
-
319
-
46448
-
0_2.
[
1
3
]
Ren,
S.
,
He,
K
.
,
Girshick,
R
.
,
Z
hang,
X.
,
&
Sun
,
J.
(2015)
.
“
Objec
t
De
tecti
on
N
et
works
on
Convolut
ional
Fea
tu
re
Maps”,
1
–
8
.
ht
tp
s://
doi.
org
/10.
11
0
9/T
PA
MI.2016.
2601099.
[
1
4
]
Ric
hter,
S
.
R.
,
Vinee
t
,
V.
,
Ro
th
,
S.
,
&
Koltun
,
V.
(2016).
“
Pla
y
ing
for
d
at
a
:
Gr
ound
trut
h
from
computer
g
ame
s”.
Le
c
ture
Not
es
in
Com
pute
r
Scie
n
ce
(In
cl
uding
Su
bserie
s
Lectu
r
e
Notes
in
Artifici
al
Int
el
l
ige
n
ce
a
nd
Le
c
ture
Not
e
s
in
Bioi
nfo
rm
at
ics
),
9906
LNCS, 102
–
118.
ht
tps:/
/
doi.
org/10
.
1007/
978
-
3
-
319
-
46475
-
6_7.
BIOGR
AP
HI
ES OF
A
UTH
ORS
Te
w
Jia
Yu
w
as
gra
duat
ed
fr
om
Multi
m
edi
a
Univer
sit
y
,
Mela
k
a
holdi
ng
a
Bac
hel
or
of
Inform
at
ion
Tec
hnolog
y
(Honou
rs)
Secur
ity
T
echnolog
y
.
She
cu
rre
ntly
li
ves
in
Muar
Johor
and
works
as
an
As
sistant
Mana
ger
in
NS
J
Tra
ding.
Her
rese
a
rc
h
int
ere
sts
inclu
de
computer
vision
and
sec
ur
i
t
y
t
ec
hnolog
ie
s.
Chin
Poo
Le
e
is
a
Senior
Le
c
tur
er
in
the
Facu
lty
of
Inform
at
ion
Scie
nce
and
Technol
og
y
at
Multi
m
edi
a
Uni
ve
rsit
y
,
Ma
lay
si
a.
She
comple
t
ed
her
Master
s
of
Scie
nc
e
an
d
Ph.D.
in
Inform
at
ion
T
echnolog
y
in
the
a
rea
of
abnor
m
al
beha
viour
detec
t
ion
and
gait
r
ec
o
gnit
ion.
She
is
a
ce
rt
ified
P
rofe
ss
iona
l
Tec
hnologi
st
sinc
e
2018
and
cur
ren
tly
th
e
Depu
t
y
Dire
ct
or
of
AD
EPT
(
Offic
e
of
Aca
demic
Deve
lopment
fo
r
Exc
e
ll
en
ce
in
Program
m
es
a
nd
Te
a
chi
ng)
,
Chai
rpe
rson
of
t
he
Adm
ission
and
Credi
t
Tr
ansfe
r
Com
m
it
te
e
,
a
senior
rese
a
rc
her
of
a
few
Mini
Fund
project
s,
and
Project
Le
ad
er
of
an
ex
te
rna
l
gra
nt
proj
ec
t
fund
ed
b
y
MO
HE.
Her
rese
arc
h
in
te
r
est
s
inc
lude
action
rec
ognition,
co
m
pute
r
vision,
gai
t
re
cogni
t
ion,
and
dee
p
le
arn
ing.
Kian
Ming
Li
m
rec
ei
v
ed
B.
I
T
(Hons
)
in
I
nform
at
ion
S
y
st
ems
Engi
nee
rin
g,
Master
of
Engi
ne
eri
ng
Sci
enc
e
(MEngSc)
and
Ph.D.
(I
.
T
.
)
degr
e
es
from
Multi
m
edi
a
Uni
v
ersity
.
H
e
is
cur
ren
t
l
y
a
Le
c
t
ure
r
with
th
e
Facult
y
of
Infor
m
at
ion
Scie
n
ce
and
Technol
og
y,
Multi
m
edia
Univer
sit
y
.
His
rese
arc
h
in
te
rest
s
inc
lud
e
m
ac
h
i
ne
l
ea
rning
,
d
eep
learni
ng
,
com
pute
r
v
ision
and
pa
tt
ern
re
co
gnit
ion.
Siti
Fati
m
ah
Abdul
Raz
ak
r
ecei
ve
d
he
r
B.
Sc
(Hons
)
with
edu
ca
t
ion
where
she
m
aj
ors
in
Mathe
m
at
i
cs
an
d
Inform
at
ion
T
ec
hnolog
y
and
Master
of
Infor
m
at
ion
t
ec
hnolo
g
y
m
aj
oring
in
Scie
nc
e
and
S
y
stem
Mana
gement
from
the
Nat
iona
l
Univer
si
t
y
of
Malay
si
a
in
2004.
Sh
e
complet
ed
h
er
p
ostgradua
t
e
stud
ie
s
in
Inform
at
i
on
Te
chno
log
y
from
Multi
m
edi
a
Univer
sit
y
.
She
is
cur
ren
tly
a
le
c
ture
r
in
Facult
y
of
Infor
m
at
ion
Scie
nc
e
and
T
ec
hnolog
y
,
Mul
ti
m
edia
Univer
sit
y
.
Her
rese
arc
h
intere
st
includes
rule
m
ini
ng,
information
s
y
stems
developm
ent
and
educ
a
ti
ona
l te
ch
n
olog
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.