TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.4, April 201
4, pp. 2724 ~ 2
7
3
4
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i4.4307
2724
Re
cei
v
ed Au
gust 14, 20
13
; Revi
sed O
c
t
ober 2
2
, 201
3; Acce
pted
No
vem
ber 1
2
,
2013
License Plate Recognition Model Research Based on
the Multi-Feature Techn
o
logy
Li Ju-xia
Coll
eg
e of Information Sci
enc
e and En
gi
neer
ing, Sha
n
x
i Agr
i
cultur
al Un
iver
sit
y
T
a
igu, Shan
xi
030
80
1, Chin
a
email: li
j
x
sn@
1
26.com
A
b
st
r
a
ct
Due to the i
m
pact of poll
u
ti
on, envir
on
me
nt and so
on
in actua
l
scen
e
, it is difficult for the
traditio
nal si
ng
l
e
feature rec
o
g
n
itio
n mod
e
l to
obtai
n a
hi
gh
er
accuracy of th
e lice
n
se p
l
ate
recog
n
itio
n. T
he
pap
er pr
op
ose
d
a
new
lic
ens
e pl
ate
imag
e
recog
n
itio
n
mo
del. F
i
rst, the s
t
ructural fe
atur
es a
nd
gray
le
ve
l
features
of the
lice
n
se
pl
ate,
such
as
the
contour
a
nd st
roke
order
ar
e
extracted. T
h
en, the
pr
inci
p
a
l
compo
nent a
n
a
lysis is use
d
to carry out the fusi
on, di
me
nsio
nal
ity redu
ction an
d redu
nda
ncy re
mov
a
l
process
i
ng for
the tw
o kinds
of features, a
n
d
a fu
zz
y
fu
si
o
n
mod
e
l for d
i
fferenti
a
ted fe
atures is i
n
trod
u
c
e
d
to ens
ure
min
i
mal
loss f
o
r t
he fe
ature
in
fusion. F
i
nal
ly, the fi
nal
res
u
lt of th
e l
i
ce
nse
plate
i
m
a
g
e
recog
n
itio
n is
a
c
hiev
ed
in
acc
o
rda
n
ce w
i
th
hi
gh d
egr
ee
of c
onfid
enc
e crite
r
ion w
h
en th
e
slop
e i
n
terfere
n
ce
of the licens
e
plate is co
nsi
d
ered ful
l
y. Simulati
on re
su
lts show
that the licens
e pl
ate i
m
a
ge rec
o
g
n
iti
o
n
mo
de
l bas
ed
o
n
multi-fe
ature
combi
nat
io
n ca
n solv
e the pr
o
b
le
m
of the si
n
g
le fe
ature rec
ogn
ition
mod
e
l
,
improve
the
ac
curacy
of lic
en
se p
l
ate r
e
co
g
n
itio
n w
h
ic
h
is
up to
9
9
%. Mo
reover, th
e
mo
del
has
a
faste
r
recog
n
itio
n spe
ed an
d can b
e
app
lie
d to
the actual l
i
cens
e
plate rec
o
g
n
itio
n.
Ke
y
w
ords
: lic
ense p
l
ate rec
ogn
ition, struct
ural featur
es, g
r
ey scale featu
r
es, princi
pal c
o
mpo
nent a
nal
ysis
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
With the in
crea
se in th
e
numbe
r of
vehi
cle
s
, in
telligent traffic control
pl
ays an
increa
singly i
m
porta
nt rol
e
in the d
e
vel
opment
pro
c
ess of the
transportatio
n
indu
stry [1]. Car
licen
se
plate
re
co
gnition
is a
n
im
po
rtant ele
m
ent
in
th
e s
t
ud
y o
f
in
te
lligent traffic
c
o
ntrol,
esp
e
ci
ally in compl
e
x environm
ent [2].
Automatic license plate re
cog
n
ition is
a cl
as
si
c t
w
o
-
cla
s
s pro
b
le
ms,
inclu
d
ing
licens
e
plate image
acqui
sition,
the image
automatic l
o
catio
n
, lice
n
se plate
s
feature extra
c
tion
sep
a
ratio
n
a
nd licen
se pl
ate re
cog
n
itio
n. The a
c
curacy an
d spe
ed of lice
n
se
plate re
co
gn
ition
rep
r
e
s
ent th
e entire pe
rf
orma
nce me
rits of t
he a
u
tomatic lice
n
se pl
ate re
cog
n
ition sy
ste
m
.
Lice
nse plate
recognitio
n
model i
s
ba
si
cally ba
sed
o
n
the structu
r
al ch
ara
c
te
ristics of the
bin
a
ry
image
with a
d
vantage
s of
high
re
cogniti
on spee
d an
d
feature extra
c
tion,
etc.
wh
ose preci
s
ion
is
high u
nde
r n
o
rmal
circu
m
stan
ce
s [3]. Actually, the
licen
se
plat
e photo
sh
o
o
t in a com
p
le
x
environ
ment
is differe
nt from the a
c
tua
l
licen
se
pl
ate image
s, su
ch a
s
the li
cense plate
was
slud
ge shelte
red,
licen
se plate
ru
st ca
use
d
by
lon
g
useful life,
shooting i
n
no
n-ide
a
l weath
e
r
con
d
ition
s
(to
rre
ntial rain,
clou
dy, high tem
perature,
and the stro
ng su
nshine
), and there a
r
e
even big
differen
c
e
s
bet
wee
n
licen
se
plate ima
g
e
s
shot at ni
g
h
t and
duri
n
g the d
a
y [4, 5].
Therefore,
ori
g
inal
li
ce
nse plate conve
r
sion with
tradit
i
onal bina
ry model will
l
e
ad
to a se
rio
u
s
loss of ima
g
e
informatio
n
and lo
w
re
co
gnition
a
c
cu
racy [6]. Later sc
hola
r
s hav
e propo
se
d the
gray featu
r
e
of licen
se pl
a
t
e image
can
be u
s
ed to
co
mpletely re
se
rve the initial
informatio
n a
nd
achi
eve the a
u
tomatic li
ce
nse pl
ate re
cognition [7
]. But the feature dime
nsi
o
n
of the model
get
is too high
and the co
m
putational co
mplexity
is increa
sed ex
pone
ntially which ma
ke
s
th
e
recognitio
n
time too long
can
not meet
the real
-time
intelligent traf
fic mana
gem
ent req
u
ire
m
ent.
In addition, th
e identify effect is n
o
t sup
e
r
ior to th
e structural cha
r
a
c
teri
stic
s of the bina
ry ima
ge.
Therefore, h
o
w
to imp
r
ove
t
he accu
ra
cy of licen
se pl
a
t
e recognitio
n
remai
n
s
an
open
pro
b
le
m-
s
o
lving [8].
To thi
s
e
n
d
,
a licen
se
plate i
m
ag
e recognitio
n
mo
del
ba
sed
on
mul
t
i-feature
combi
nation,
whi
c
h integ
r
a
t
es the adva
n
tage
s of
both stru
ctural feature a
nd grey level feature,
is pro
p
o
s
ed i
n
this pap
er.
First, the structural
features an
d gray
level
feature
s
of the license
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Lice
nse Plate Recognitio
n
Model Resea
r
ch Ba
se
d on
the Multi-Fea
t
ure Te
chn
o
lo
gy (Li
Ju
-xi
a
)
2725
plate, such
as th
e conto
u
r a
nd
stro
ke order a
r
e
extracted.
Th
en, the p
r
in
cipal comp
on
ent
analysi
s
is u
s
ed to carry out the fusi
on, dimen
s
io
nality redu
ction and
red
u
ndan
ce remo
val
pr
oc
es
s
i
ngs
f
o
r
the tw
o k
i
nd of featur
es
, and a fu
zzy fus
i
on model for
differ
entiated features
is
introdu
ce
d to
ensure
mini
mal loss fo
r the featur
e in fus
i
on. Finally, the s
t
ruc
t
ural featur
e
recognitio
n
m
odel
and
gre
y
level feature re
co
gniti
on
model
a
r
e e
s
tabli
s
he
d re
spe
c
tively u
s
i
ng
the sup
port
vector ma
chi
ne, and the
final resu
lt of the license
plate image
recognition
is
achi
eved in
accordan
ce
with the
high
deg
ree
of
confiden
ce
criterion. Sim
u
la
tion re
sult
s
show
that the licen
se plate reco
gnition a
c
curacy of t
he mo
del is up to 9
9
% and re
co
gnition time is less
than 1.55m
s.
2.
License Plate Reco
gnitio
n
Principles
The li
cen
s
e
plate recognit
i
on p
r
inci
ple
can
be
d
e
scribed a
s
: Fi
rst
,
the lice
n
se
plate is
automatically extracted from an
imag
e. Then ima
ge feature is
extracted. F
i
nally, the license
plate is
re
cog
n
ize
d
which realizes i
n
telligent
vehi
cle
monitori
ng a
n
d
mana
gem
e
n
t. Known fro
m
the licen
se
p
l
ate re
cog
n
ition pri
n
ci
ple, licen
se
pl
ate
cha
r
a
c
ter pl
ays vital role
in recognitio
n
accuracy. Structural ch
ara
c
teri
st
ics and
the image gray feature of
the traditional
sepa
rate ima
g
e
can o
n
ly describ
e frag
me
nt informatio
n in a lice
n
se
plate whi
c
h
can
not fully reflect the lice
n
se
plate catego
ry [9]. Multi-feature
com
b
in
ation lic
e
n
se
plate re
co
gnit
i
on sy
stem is a good
sol
u
tion
to the defect of traditional identificatio
n, mainly
inclu
d
e
four part
s
: Structu
r
al fea
t
ure extra
c
tio
n
,
the gray
-scal
e
feature
extraction, P
C
A dimen
s
ion
a
lity redu
ctio
n treatme
nt, sup
port ve
ctor
machi
ne mult
i-cla
s
sificatio
n
. The pro
c
e
s
s is sho
w
n in
Figure 1
Figure 1. Lice
nse Plate Re
cog
n
ition Flo
w
Ch
art with
Multi-feature Combi
nation
Multi-feature
combin
ation
licen
se plat
e re
cognitio
n
model first
extracts th
e binary
stru
ctural ch
ara
c
teri
stics
and g
r
ay
scale ch
aracte
ri
stics of th
e
licen
se
plat
e, red
u
ces t
he
dimen
s
ion
of
the extract f
eature
to
eliminate
d
upli
c
ate informati
on b
e
twe
en f
eature
s
t
h
ro
u
g
h
Princi
pal Co
mpone
nt
An
alysis (PCA
),
and sele
ct
s the feature
most h
e
lpfu
l to improve
the
licen
se pl
ate
recognition
accuracy. Th
en lice
n
se pl
ate stru
ctural
feature
s
cla
ssifi
cation m
odel
and lice
n
se p
l
ate gray feature cl
assification model
a
r
e
establi
s
hed
sep
a
rately by
Support Vect
or
Machi
ne (SV
M
) with nonl
inear a
nd int
e
lligent ca
pa
bilities. Finall
y
, the discri
m
ination re
sult
corre
s
p
ondin
g
to highe
r de
gree
of co
nfid
ence
from
th
e
two
mod
e
ls i
s
sele
cted
a
s
a licen
se
plat
e
final re
co
gniti
on result. Th
e mod
e
l ta
ke
s a
d
vantag
e
of high
pe
rfo
r
man
c
e
an
d
simpli
city of the
stru
ctural feature
s
, the adv
antage
of, at the sam
e
time makes u
p
the losi
ng info
rmation d
e
fects
cau
s
e
d
by t
he bi
nary
conversion
fo
r the
st
ru
ct
u
r
al cha
r
a
c
teristics
by usi
ng
of grayscale
cha
r
a
c
teri
stics, effectively
redu
ce
s th
e f
eatur
e
dim
e
n
s
ion
after PCA pro
c
e
s
sing
and
ha
s
hig
h
pre
c
isi
on, less time-con
su
ming advanta
ges.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2724 – 2
734
2726
3.
License Plate Reco
gnitio
n
Model
w
i
th Multi-feature Combination
3.1. License
Plate Struc
t
ural Feature
s
Extrac
tion
(1) Lic
e
nse p
l
ate image binariza
tion
The licen
se
plate st
ru
ctural f
eature
s
in
clud
e outlin
e
feature
s
a
n
d
stro
ke fe
atures
with
the adva
n
tag
e
s
of si
mple
extraction
an
d excell
ent re
cog
n
ition perf
o
rma
n
ce whi
c
h are
no
rma
lly
extracted
fro
m
the bin
a
ri
zation imag
e
of the plat
e chara
c
te
rs.
F
o
r
u
s
ual plate image
s,
they are
binari
z
e
d
first
in which the
y
are tran
sfe
rre
d to
the grey model wit
h
black/white
two colo
rs. By
histog
ram t
r
a
n
sform meth
o
d
, the imag
es can
be tr
an
sferred to bi
na
rization im
age
s fast
and
wit
h
high qu
ality.
The tran
sfe
r
red bina
rizatio
n
image is
sh
own in Fig
u
re
2.
Figure 2. Binarized Li
cen
s
e Plate Image
(2) O
u
tline fe
ature
s
ex
tra
c
tion
The outline f
eature of the
binari
z
ation
im
age
s ca
n well de
scrib
e
the cha
r
acte
r frame
informatio
n whi
c
h ha
s two types-int
ernal a
nd ex
ternal outlin
e
features. Th
e internal ou
tline
feature
s
(I
NL
) defin
e the
amount
s of i
n
ternal
bl
a
c
k pixels, n
a
me
ly amount
s o
f
pixels fro
m
the
first white
-
bl
a
ck j
o
int point
to black-whit
e join
t point;
the external
outline featu
r
es (OUL)
defi
n
e
amount
s of p
i
xels from th
e image o
u
tside to t
he first white pixel
.
When
com
pute the outli
ne
feature
s
of the binari
z
e
d
image, first th
e cha
r
a
c
te
r i
m
age is divid
ed into nH su
b-ima
g
e
s
fro
m
row
dire
ction. n
H
intern
al outli
ne featu
r
e
s
a
nd n
H
exte
rn
al outline
feat
ure
s
can
be
o
b
tained f
r
om l
e
ft
and
right
sid
e
s
whi
c
h a
r
e
totally 4*n
H
o
u
tline f
eatu
r
e
s
. The
n
, the
whol
e bin
a
ri
zed ima
ge
ca
n
be
segm
ented
to
nK su
b-im
ag
es from row
dire
ction.
Sim
ilarly, nK inte
rnal
outline f
eature
s
and
nK
external o
u
tline features
can be o
b
tain
ed from
u
p
a
nd do
wn di
re
ction
s
whi
c
h
are 4*
nH
outline
feature
s
. Totally 4*(n
H+ n
K
) outline fea
t
ur
es of the
whole imag
e can be obtai
ne
d.
3.2. License
Plate Principle Image Fea
t
ures Ex
tra
c
tion
PCA is
a
ki
nd of hi
gh
efficient
statistical
analy
s
i
s
meth
od fo
r featu
r
e di
mensi
o
n
s
redu
ction in
whi
c
h an opti
m
al
feature subset includi
ng small
am
ount of unrel
ated synthe
si
ze
d
factors i
s
used to
re
pla
c
e
the initial
m
u
ltiple f
eatu
r
e
facto
r
s. It
ca
n p
r
e
s
erve
th
e initial fe
atu
r
e
informatio
n to the maxim
u
m extend in
orde
r to
sim
p
lify the initial feature set and remove t
h
e
redu
nda
nt inf
o
rmatio
n am
ong initial
fe
ature
s
[10,
1
1
]. For th
e d
a
ta set with
N
sampl
e
s,
and
feature
s
of
x
(i
=1, 2, …., N), the princi
pal
comp
one
nt selectio
n pro
c
edure is a
s
follows.
First, the mea
n
value m of each feature
i
n
the data set
is obtaine
d [12-1
4
].
1
1
N
i
i
mx
N
(1)
After the mea
n
of each feat
ure sampl
e
is obt
ained, the
covaria
n
ce matrix of the data set
is gen
erate
d
.
1
1
()
N
T
i
i
R
xm
N
(2)
Then, the
Ja
cobi metho
d
is use
d
to solv
e P feature v
a
lue
s
1
>
2
>…
>
p
(after so
rting
pro
c
e
s
s)
whi
c
h a
r
e l
a
rger than 0 fo
r t
he featu
r
e fu
nction
||
0
RI
. The
co
rre
sp
ondi
ng
feature vecto
r
s for ea
ch fea
t
ure value
j
are
as follows.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Lice
nse Plate Recognitio
n
Model Resea
r
ch Ba
se
d on
the Multi-Fea
t
ure Te
chn
o
lo
gy (Li
Ju
-xi
a
)
2727
()
()
()
()
12
(
,
,
...,
,
)
,
(
1
,
2
,
.
..,
)
jj
j
j
p
CC
C
C
j
p
(3)
Each featu
r
e
vectors satisf
y the followin
g
con
d
ition
s
.
()
(
)
()
(
)
1
1(
)
0(
)
qq
p
jk
jk
q
j
k
CC
CC
j
k
(
4
)
The initial
fe
ature
s
a
r
e
m
appe
d to p
p
r
inci
ple
com
p
onent
s Z1,
Z
2
, …, Zp. If the ratio
11
()
/
(
)
p
m
j
j
jj
a
of the cova
ri
ance su
m of the previo
us
m prin
ciple
compon
ents i
s
large
r
,
namely the
p
r
eviou
s
m
pri
n
cipl
e comp
o
nents are
p
r
e
s
erve
d p
r
in
ci
ple comp
one
nts, are
sele
cted
for further analys
is
. If
a
l
a
rge
r
th
an
0
.
85, ba
sically the p
r
eviou
s
m
pri
n
ci
pl
e compo
nent
s
pre
s
e
r
ve initi
a
l feature
info
rmation. T
h
u
s
, 0.85
can
b
e
used a
s
th
resh
old to d
e
termin
e the va
lue
of m.
3.3. Featur
e Dimensional
i
t
y
Reductio
n
Due to that
must be o
b
tain
ed first in the
identification
pro
c
e
ss, an
d that
is a three-
d
i
me
ns
io
na
l ve
c
t
o
r
wh
ich is comprised of
,,
x
yz
. There
f
ore, in orde
r to simplify the
cal
c
ulatio
n, the dim
e
n
s
ion
a
lity redu
ctio
n ope
ratio
n
i
s
in
nee
d for
multidimen
sio
nal vecto
r
s. The
dimen
s
ion
a
lity redu
ction
method b
a
se
d on p
r
in
cip
a
l com
pon
en
t analysis i
s
adopted i
n
this
pape
r, so a
s
to redu
ce the
dimen
s
ion of
the feature
sp
ace.
The pri
n
ci
pal
comp
onent
analysi
s
met
hod is a lin
e
a
r dime
nsi
o
n
a
lity redu
ction method,
whi
c
h
ca
n o
b
tain the
mi
nimum m
ean
sq
uare e
r
ro
r. The
metho
d
proje
c
t the
origi
nal fe
ature
vector into
smaller
sub
-
sp
ace, so as to
redu
ce
the di
mensi
on of the origin
al feature vecto
r
.
Assu
ming tha
t
the n-dimen
s
ion
a
l ran
d
o
m
vector can
be expre
s
sed
as
1
(
,
......
)
T
n
x
xx
,
x
Ex
, the correl
ation matrix
be
[]
T
x
R
Ex
x
, and the cova
rian
ce matrix can be
r
e
pr
es
e
n
t
ed
b
y
[(
)(
)
]
T
x
CE
x
x
x
x
. The d
i
mensi
onality
red
u
ctio
n
pro
c
e
s
s can
be
rep
r
e
s
ente
d
the pro
c
e
ss that tran
sformin
g
x
into
1
(
,
......
)
T
n
yy
y
by the
orthog
onal
trans
form. It is
c
a
lc
ulated as
follows
:
1
2
12
(
,
,
....
)
.
.
T
T
TT
n
T
n
u
u
yU
x
u
u
u
x
x
u
(5)
Her
e
,
T
ii
yu
x
,
1
,
2....
in
.
x
can be exp
r
e
s
sed u
s
ing th
e followin
g
formula:
1
2
11
12
1
(
)
(
)
(
,
,
.
.
...
)
.
.
n
TT
T
yy
n
i
i
i
n
y
y
x
Uy
U
y
U
x
U
y
U
u
u
u
y
u
y
(6)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2724 – 2
734
2728
I
f
just
a
sub
s
et
1
{
,
....
}
n
yy
of vector
y
is
res
e
rved for the es
timate of
x
, an
d the
remai
n
ing
co
mpone
nts a
r
e
repla
c
ed u
s
i
ng
i
b
, the formula for the est
i
mate is as fol
l
ows:
11
mn
ii
ii
ii
m
x
yu
b
u
(7)
The erro
r is:
2
(
)
[(
)
(
)]
T
mE
x
x
x
x
2
1
[(
)
]
n
ii
im
E
yb
(8)
By a kn
owl
e
dge
of differe
ntial calculus,
it
is
obtain
e
d
that the
error i
s
mi
nimu
m wh
en
[]
[
]
TT
ii
i
i
bE
y
u
E
x
u
x
. Then:
22
11
()
[
(
)
]
nn
T
ii
i
x
i
im
i
m
mE
y
b
u
C
u
(9)
In orde
r to make the valu
e of
2
()
m
be mini
mum, the differential
cal
c
u
l
us meth
od is
use
d
and by:
0
i
J
u
Her
e
,
1
[(
1
)
]
n
TT
ix
i
i
i
i
im
Ju
C
u
u
u
.
It can be de
rived that:
,
1
,
....
xi
i
i
Cu
u
i
m
n
(10)
Whe
r
ein, rep
r
esents the
eigenvalu
e
s
of the covari
ance matrix of
x
,
i
u
denote
s
the
corre
s
p
ondin
g
eigenve
c
tor.
Then,
2
1
2
1
2....
()
[(
)
]
[(
)(
)
]
n
i
im
ii
i
TT
yi
i
X
n
m
Ey
y
C
E
yy
yy
U
C
U
(11)
Thus,
by th
e p
r
in
cipal
comp
one
nt
analysi
s
dim
ensi
onality
redu
ction met
hod,
the
eigenvalu
e
compon
ent ca
n
be
p
r
e
s
erv
ed com
p
lete
l
y
, not only
si
mplifying co
mputation,
bu
t also
save a lot of origin
al information as m
u
ch a
s
po
ssi
bl
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Lice
nse Plate Recognitio
n
Model Resea
r
ch Ba
se
d on
the Multi-Fea
t
ure Te
chn
o
lo
gy (Li
Ju
-xi
a
)
2729
3.4. Crea
te L
i
cense Plate
Reco
gnition
Multi-re
co
gnit
i
on data fu
sio
n
and d
e
ci
sio
n
al
go
rithm is used i
n
so
n
a
r si
gnal p
r
o
c
e
ssi
ng
to detect the
subm
ari
ne ta
rget a
s
e
a
rly
as t
he
197
0s. After years
of developm
e
n
t, a variety of
intelligent dat
a fusion
stru
cture
s
, such
as fu
zzy
logi
c theo
ry, ne
ural n
e
two
r
ks, DS infere
nce
algorith
m
and
so forth, hav
e been
cre
a
te
d.
The dat
a fusi
on an
d de
ci
sion alg
o
rithm
prop
osed in
this pa
pe
r is based o
n
th
e fuzzy
logic the
o
ry, and the sy
ste
m
block diag
ram is a
s
follows:
Figure 3. Multi-se
nsor Imag
e Fuzzy Fusi
on and
De
cisi
on Sche
matic
The fu
zzy l
o
gic
and
de
cision meth
od i
s
a
dopte
d
to
integrate
the
passive im
ag
ing a
n
d
the ca
pture
d
l
i
cen
s
e
plate
data, so
as to
determi
ne
the s
a
fety of the s
u
rfac
e of the licen
se pl
ate.
Based o
n
fuzzy logic, the data fusion te
chn
o
logy
ca
n
effectively reflect the expert opinion in the
desi
gn
of me
mbershi
p
fun
c
tion. T
he
al
gorithm
is
ea
sy to im
plem
ent an
d
simil
a
r to
the
hu
man
way of
thin
ki
ng. Th
e u
n
certaintie
s i
n
t
he d
a
ta i
s
p
r
ocesse
d
by
pre
-
setting th
e mem
b
e
r
ship
func
tion of fuz
z
y
sets
.
The fu
zzy fu
sion
and
de
cision
pro
c
e
ss of the mu
lti-feature d
a
ta
is divide
d int
o
three
step
s: First, the fuzzy pro
c
essing of the
data is
carrie
d out in acco
rdan
ce with p
r
e-defin
ed fuzzy
set and its
membe
r
ship
function. The
n
, the obt
ained topog
rap
h
i
cal feature data is used
to
achi
eve the reasonin
g
re
sults of
the terrain
se
curity
throug
h the u
s
e of fuzzy p
r
opo
sition
s
se
t
descri
bed by
the langu
age.
The fuzzy inferen
c
e rule
s used in thi
s
pape
r are sh
own in Ta
ble
1.
Finally, the
n
u
meri
cal
a
s
sessed
value
of the li
cen
s
e
plate i
dentifi
ability is
obta
i
ned th
rou
g
h
the
defuzz
i
fication s
t
eps
(The s
e
t of s
a
fety ass
e
s
s
m
ent values
is
[0
,
1
]
).
Table 1. The
Fuzzy Inferen
c
e Rule
s
T
e
x
t
ur
e
Slope of the license plate
Ver
y
steep
Steep
Flat
Ver
y
flat
Ver
y
rough
Poor
Poor
Lo
w
Lo
w
Rough
Poor
Lo
w
Lo
w
Lo
w
Smooth Lo
w
Lo
w
Medium
High
Ver
y
smooth
Lo
w
Lo
w
High
High
As
sho
w
n
in
Table
1, the
fuzzy set to
re
pre
s
ent
the
p
a
ssive im
age
ro
ugh
ness is define
d
to contain fo
ur elem
ents,
that is, {very roug
h,
rou
gh,
smooth, very smooth}. A
nd fuzzy set
to
rep
r
e
s
ent th
e
slo
pe
data
of the li
cen
s
e plate
obtai
ned from th
e
imagin
g
e
q
u
ipment i
s
al
so
defined to
co
ntain fou
r
ele
m
ents, n
a
me
ly {very st
ee
p, steep, flat,
very flat}. The fuzzy
set
to
rep
r
e
s
ent the
security data con
s
ist
s
of four
element
s, namely {p
oor, low, me
dium, high}.
The
trape
zoid
al membe
r
ship
function (Fig
ure 4
)
determined em
piri
cally ca
n be
set in acco
rd
ance
with the expe
rt opinion
s. In addi
tion, the membershi
p
function is
set reasona
bly, so so to meet
the re
co
gniti
on requi
rem
e
nts of th
e a
r
ea, where th
e security is
"low". Thu
s
,
the a
s
sessm
e
n
t
results
of the lice
n
se pl
ate re
co
gniti
on can b
e
use
d
to p
r
o
v
ide su
ppo
rt
for the
rele
vant
depa
rtment
s.
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2724 – 2
734
2730
Figure 4. Rou
ghne
ss Mem
bership F
u
n
c
tion Base
d on
the Image Da
ta
The gen
erate
d
fuzzy logi
c
discrimi
nant surface
in accorda
n
ce with
the inferen
c
e
rule
s in
Table 1 is
sh
own in Fig
u
re
5:
Figure 5. The
Fuzzy Logic
Discri
mi
natio
n Surface of the Secu
rity
3.5. License
Plate Re
cog
n
ition
Acco
rdi
ng to
the PCA
di
mensi
onality
redu
ct
ion
me
thod a
bove,
the colle
cted
feature
vector sp
ace of
the
license
plate can be pro
c
e
s
s
ed
wit
h
dime
nsio
na
lity redu
ction
operation. Th
e
Erro
r v
e
cto
r
is a multi-dim
ensi
onal ve
ctor, and i
s
bro
k
en d
o
wn int
o
two sepa
ra
te sub
-
spa
c
e
by PCA dime
nsio
nality red
u
ction m
e
tho
d
: the feature
sub
s
p
a
ce
an
d the non
-fea
ture sub
s
pa
ce.
They
a
r
e expre
s
sed by
R
and
R
respe
c
tively. The
simila
rity of the M l
o
w-lati
tude ei
gen
vectors in fea
t
ure su
bspa
ce
is:
2
2
1
()
/2
1
/
2
2
1
1
()
ex
p
ex
p
(
)
2
2
(
)
()
()
(2
)
(2
)
M
i
i
i
FF
MN
M
M
i
i
y
PP
P
(12)
Whe
r
ein,
()
F
P
is t
he real
edg
e
den
sity of su
bsp
a
ce
R;
R
is the e
dge
e
s
timated
dens
i
ty;
i
y
is the main comp
onent;
2
()
is the resi
dual
ene
rgy; The weig
ht param
eter
ca
n
be expre
s
sed
by the mean value of the c
hara
c
te
risti
c
values of the
sub
-
spa
c
e
R
as
follows
:
1
1
N
i
iM
NM
(13)
R
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Lice
nse Plate Recognitio
n
Model Resea
r
ch Ba
se
d on
the Multi-Fea
t
ure Te
chn
o
lo
gy (Li
Ju
-xi
a
)
2731
Here, both
()
i
P
and
()
E
P
su
bject to the t
w
o-dime
nsio
nal Ga
ussia
n
distrib
u
tion, a
nd meet:
1/
2
1
1/
2
2
1/
2
1
1/
2
2
()
(2
)
()
(2
)
r
D
E
r
D
i
eE
P
E
eE
I
P
I
(14)
Whe
r
ein,
is
the c
o
varianc
e
matrix.
The
step
s to
cal
c
ulate th
e
matchin
g
si
m
ilarity of som
e
tested t
w
o
-
dimen
s
ion
a
l l
i
cen
s
e
plate featu
r
e
sampl
e
k
I
to a
particula
r two
-
dime
nsi
onal
licen
se
plate
s
libra
ry
sampl
e
j
I
are
a
s
follows
: Firs
t, us
e
to s
ubtrac
t
j
I
, take the result a
s
vecto
r
, and then map it to the formul
a
(14
)
. Th
en,
calcul
ate the
()
i
P
and
()
E
P
b
a
se
d
o
n
the
eige
n v
e
ctors of the
prima
r
y
comp
one
nt o
f
the intra-cla
ss
and inte
r-cla
ss
Gau
ssi
an den
sity function. Fin
a
ll
y, calculate t
he
degree of matchin
g
acco
rding to formu
l
a (12). In
order to simplif
y the calculat
ion, two albin
o
vectors are a
dded to ea
ch
image of the libra
ry:
1/
2
j
II
j
iI
,
1/
2
j
II
j
iI
(15)
Her
e
,
and
are the maxi
mum eigenva
l
ue diago
nal
matrix and ei
genve
c
tor m
a
trix
corre
s
p
ondin
g
with the
s
e
eigenvalu
e
s
of
I
and
E
resp
ectively, and
the dimen
s
io
nality
of the corre
s
p
ondin
g
su
b-spaces a
r
e
I
M
and
E
M
res
p
ec
tively.
Simplify the probability calcul
ations menti
oned above to the simple
cal
c
ulation of
eucli
dean di
st
ance:
1/
2
1
1/
2
/2
1/
2
1
1/
2
/2
()
(2
)
()
(2
)
T
I
D
T
E
D
eI
P
I
eE
P
E
(16)
Based
at the approximate
match in thi
s
archit
ectu
re, t
here i
s
a mo
re simpl
e
form
, since
only an albi
n
o
vector
j
i
for each image
is sto
r
ed. After t
he compl
e
tion of the
cal
c
ulatio
n for
albino vecto
r
k
i
of the test samples, the si
milarity of
the degre
e
of asso
ciation
can
be cal
c
ulate
d
usin
g the followin
g
formul
a:
2
1/
2
'
1/
2
/2
()
(2
)
jk
I
D
I
ei
i
SP
(17)
Based
o
n
the
above
an
alysis, it
can
be
see
n
that
the
sim
p
le
cal
c
ul
ation b
a
sed
o
n
two
-
dimen
s
ion
a
l licen
se plate
simila
rity is transfo
rme
d
into the calcul
ation of euro
pean ge
omet
ric
k
I
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2724 – 2
734
2732
distan
ce, an
d the distan
ce can be
seen a
s
t
he degree of asso
ciation
of the
multi-feat
ure
cha
nge
s cau
s
ed by external ch
ang
es
and the reco
gnition. If the sco
re i
s
abo
ve 1/2, then the
degree of a
s
so
ciation
of this
cha
nge i
s
hig
her
and
the re
cog
n
ition re
sult di
splays the
rig
h
t
licen
se pl
ate, otherwise, it
is not the licen
se
pl
ate. Thus, the p
r
opo
se
d sim
p
lified simila
rity
cal
c
ulatio
n method ad
ded
with the asso
ciati
on d
egre
e
is more sim
p
le and effe
ctive.
4. Simulation
Experiment
4.1. Experiment Data
In orde
r to simulate the licen
se pl
ate image
s in co
mplicate
d
en
vironme
n
t, all of the
licen
se
plate
image
s i
n
the expe
rime
nt are
taken
in
different climate, su
ch
as rai
n
y,
su
nny,
clou
dy, wind
y and at
nig
h
t whi
c
h i
n
cl
ude diffe
rent
types of pl
ates,
su
ch a
s
l
a
rge
freig
h
t cars,
buses, co
mp
act ca
rs a
nd motors. After
the licen
se
pl
ates are obta
i
ned, first the
plates are p
r
e-
pro
c
e
s
sed. T
he cha
r
a
c
ters in th
e plate
s
a
r
e i
s
olate
d
to re
move
the Chi
n
e
s
e
cha
r
a
c
ters. T
hen
each cha
r
a
c
ter i
s
store
d
i
n
thre
e different form
ats-i
n
itial ch
aract
e
r ima
ge, bi
n
a
rized
ch
ara
c
ter
image a
nd
g
r
ey characte
r image to
b
u
ild the
cha
r
acter data
b
a
s
e in th
e ex
perim
ent. The
cha
r
a
c
ter lib
rary contai
ns 10945 cha
r
acter im
age
s in which 5
0
characte
rs of each type
o
f
c
h
ar
ac
te
r
imag
e
a
r
e
r
a
nd
omly s
e
le
c
t
ed
a
s
te
s
t
sa
mp
le
.
4.2. Compari
s
on Models
In orde
r to verify the perf
o
rma
n
ce of t
he lice
n
se pl
ate re
cogniti
on model
with multi-
feature co
mb
ination,
the e
x
perime
n
t
est
ablishe
s
5
co
mpari
s
o
n
mo
dels-si
ngle
st
ructu
r
al
featu
r
e
model (mod
e
l
1), singl
e g
r
ey feature
model (mo
d
e
l
2); initial feature
without
PCA dimen
s
ion
redu
ction
(m
odel 3
)
,
sing
le mod
e
l ba
sed
on
stru
ctural featu
r
e
s
and
grey feature
s
a
nd t
h
e
dimen
s
ion
s
o
f
the initial feature
s
a
r
e re
duced
by PCA (model
4),
and the p
a
ral
l
el model
ba
sed
on stru
ctu
r
al feature
s
an
d g
r
ey feature
s
wit
hout initial
feature
s
PCA
process (mo
del 5).
4.3. Results and An
aly
s
is
The digit
s
a
nd alph
abet
s recognitio
n
results of th
e licen
se
pla
t
e in each m
odel a
r
e
sho
w
n
in ta
bl
e 2.
Du
ring
the
re
cognitio
n
trai
ni
ng
pro
c
ed
ure,
the
d
i
git “0
” a
nd
al
phab
et “O”
a
r
e
rega
rd
ed a
s
“0” and the di
g
i
t “1” and al
ph
abet “I” a
r
e re
gard
ed a
s
“1
”.
Table 2. Lice
nse Plate Ima
ge Re
co
gnitio
n
Preci
s
io
ns
of each Mo
de
l
Digits
Alphabets
Precision(%)
Time(ms/piece)
Precision(%)
Time(ms/piece)
Model 1
94.71
1.58
95.64
1.48
Model 2
95.43
0.94
96.35
0.91
Model 3
91.25
1.95
88.68
1.92
Model 4
92.53
1.26
90.19
1.21
Model 5
96.36
2.59
97.71
2.47
Proposed Model
99.37
1.47
99.35
1.32
From the
abo
ve results, re
cog
n
ition p
r
e
c
isi
o
n
s
of the
prop
osed license plate m
odel with
multi-feature combi
nation are highe
st
a
m
ong all
of the mo
dels an
d they are ab
ove 99%. Th
ere
are
follo
wing con
c
lu
sio
n
s.
(1)
Com
pare
d
the grey feat
ure
model
and st
ru
ctural feat
ure m
odel, the recognition
pre
c
isi
o
n
s
of
the structu
r
al
feature m
ode
l are
highe
r t
han tho
s
e
of grey featu
r
e
model. Be
ca
use
the feature
di
mensi
on i
s
m
u
ch
simpl
e
r, t
he con
s
um
e
d
time of the st
ructu
r
al fe
atu
r
e mo
del i
s
le
ss
than that
of g
r
ey mo
del. It’
s o
b
viou
s tha
t
the
stru
ct
ural featu
r
e
s
a
r
e bette
r tha
n
grey fe
ature
s
for
simple
cha
r
a
c
t
e
r
s
re
co
gnit
i
on.
(2) Ea
ch
mod
e
l is
comp
are
d
with
the
sa
me mo
del
after
dimen
s
io
n
redu
ction
with PCA.
The re
co
gniti
on preci
s
ion
s
can b
e
slig
htly incre
a
se
d after PCA
pro
c
e
ss
and
the con
s
um
ed
recognitio
n
time ca
n be
g
r
eatly re
du
ce
d. Espe
cia
lly
for the
single
cha
r
a
c
ter i
n
the re
cog
n
ition
model with multi-feature combi
nation, the
time
co
st ca
n be
redu
ced
by 1m
s.
Obviou
sly, the
feature dime
nsio
ns can
be gre
a
tly redu
ced
with
reco
gnition
time reduced by ensu
r
ing
recognitio
n
preci
s
ion
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Lice
nse Plate Recognitio
n
Model Resea
r
ch Ba
se
d on
the Multi-Fea
t
ure Te
chn
o
lo
gy (Li
Ju
-xi
a
)
2733
(3) T
he sim
p
le parall
e
l model utilize
s
t
he
advanta
g
e
s of grey fe
at
ure
s
and
structu
r
al
feature
s
. The
reco
gnition p
r
eci
s
io
ns a
r
e
obviou
s
ly
higher than tho
s
e of single fe
ature mod
e
l, but
the feature d
i
mensi
o
n
s
are highe
st. In all of
the compa
r
ison m
odel
s, the co
nsum
ed time
is
longe
st.
(4) The
multi
-
feature
com
b
ination
re
co
gniti
on m
ode
l ca
n effe
ctively increa
se
licen
se
plate re
co
gnit
i
on preci
s
ion
s
by parallel
modelin
g
of stru
ctural fea
t
ures
and g
r
ey feature
s
. The
recognitio
n
p
r
eci
s
io
ns for
digits
and
alp
habet
s a
r
e
m
o
re
than
99%
. PCA i
s
u
s
e
d
to
red
u
ce t
h
e
dimen
s
ion
s
a
nd the time
con
s
um
ption
is effectively redu
ce
d in
whi
c
h ea
ch
digit only ne
eds
1.47ms a
nd
each alph
abe
t only needs
1.32ms. Th
e rec
ognitio
n
preci
s
ion
s
of all
cha
r
act
e
rs a
r
e
sho
w
n in figu
re 6. Except “D/0”, “H”, “U” and “M
”, the
reco
gnition p
r
eci
s
io
ns of o
t
her ch
aracte
rs
can rea
c
h up
to 100%.
Figure 6. Lice
nse Plate Ch
ara
c
ters Recognition
Preci
s
ion
s
with Mu
lti-feature
Co
mbination
5. Conclu
sion
In ord
e
r to
solve the
pro
b
l
ems i
n
the
li
cen
s
e
plate
reco
gnition, th
is p
ape
r p
r
o
p
o
se
s
a
licen
se pl
ate
recognitio
n
model
with
multi-f
eature com
b
inati
on. This
m
odel utilizes the
advantag
es
of the stru
ctu
r
al feat
ure of
simpli
city and high effici
e
n
cy to not o
n
ly increa
se
the
recognitio
n
p
r
eci
s
io
ns
of the syste
m
, but also
save
recognitio
n
time. Meanwhile, the pap
er
establi
s
h
e
s
parall
e
l grey feature mo
del whi
c
h
solves the d
e
fects of ea
si
ly losing im
age
informatio
n fo
r the
st
ruct
ural featu
r
e
s
.
More
over,
PC
A is
us
ed
to
re
d
u
c
e
th
e
d
i
me
ns
io
ns
of th
e
grey feature
s
and remove
redun
dan
cy. Therefo
r
e,
the multi-feat
ure combin
ation re
cog
n
ition
model can no
t only incre
a
se the re
cog
n
ition pre
c
i
s
ion
to large exten
d
, but also en
sure less tim
e
con
s
um
ption.
Referen
ces
[1]
John W
r
ig
ht, Allen Y
ang, Arvi
nd Ganes
h, Shankar Sastr
y
, Yi Ma.
Robust F
a
ce Reco
gniti
onvi
a
Spa
r
s
e
Repr
esentati
o
n
.
PAMI. Napa. 200
8: 256-
263.
[2]
D Don
oho. Co
mpresse
d sens
ing.
IEEE Trans. on Informati
on Theory
. 2
0
0
6
; 52(4): 12
89–
130
6.
[3]
RG Barann
iuk.
Compressiv
e
sensi
ng.
IEEE Signal Process
i
ng Maga
z
i
ne
.
200
7; 24(4): 11
8–1
21.
[4]
Dala
l N, T
r
igg
s
B.
Histogr
a
m
s
of ori
ented
grad
ie
nts for
hu
ma
n d
e
tecti
on Proc
ee
din
g
s
of CVPR
’
0
5
.
IEEE Computer Societ
y
.
SanDiego. 2010:886
-
99
3.
[5]
Shan Ji
ali
ng. L
i
cens
e Plate R
e
cog
n
itio
n S
y
s
t
em
Based o
n
W
i
reless Net
w
ork Reco
gniti
o
n
Algor
ithm.
Co
mp
uter Mea
s
ure
m
e
n
t & Control
. 201
1; 19(
1): 124-1
26.
[6]
Mikola
jcz
y
k K, Schmid C. A p
e
rformanc
e ev
alu
a
tion
of loca
l descri
p
tors.
IEEE Transactions on P
a
ttern
Analys
is & Machin
e Intelli
ge
n
c
e
. 2005; 2
7
(1
0): 1615
-
1
630
.
[7]
E Cand`
es, J Romb
erg, T
Tao Ro
bust. un
certaint
y pr
i
n
ci
ples: E
x
act sig
nal rec
onstruct
i
onfrom h
i
gh
l
y
incom
p
lete fre
que
nc
y
inform
ation.
IEEE Trans. on Information Theory
. 20
06; 52(2): 4
89–
509.
[8]
W
e
i Gao, Hai
z
hou Ai, Sh
ih
ong
Lao.
Ad
a
p
tive Co
ntour
F
eatures in O
r
iente
d
Granu
l
a
r Space fo
r
Hu
ma
n Detecti
on an
d Seg
m
e
n
tation
. CVPR.
San F
r
ancisco
. 2011:2
55-3
6
3
.
[9]
Don
g
Yan, Z
hu Yongs
he
ng,
Liu Co
ng. Ap
plicat
i
on of Image F
u
si
on T
e
chn
o
lo
g
y
in
Lice
nse Plat
e
Reco
gniti
on.
C
o
mputer Me
as
ure
m
e
n
t & Con
t
rol
. 2013; 2
1
(3
): 791-79
3.
[10]
Bar
y
on J, F
l
e
e
t D. Bea
u
ch
e
m
in S.
Perfor
mance
of o
p
ti
cal flo
w
tech
ni
ques.
Itern
atio
nal J
our
nal
of
Co
mp
uter Visi
on
. 199
4; 12(1)
: 163-17
8.
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
012
34
5678
9
A
BCDE
FG
HIJK
LMN
OP
QRST
U
V
WXYZ
Evaluation Warning : The document was created with Spire.PDF for Python.