TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3281 ~ 32
8
8
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.4947
3281
Re
cei
v
ed O
c
t
ober 1
0
, 201
3; Revi
se
d Novem
b
e
r
27, 2013; Accept
ed De
cem
b
e
r
14, 2013
A New License Plate Fault-tolerant Characters
Recognition Algorithm
Guo
w
e
i
Yan
g
*
1
, Xiaochun Wang
1
, Ya
ng Yang
2
1
Colle
ge of Aut
o
matio
n
Eng
i
n
eeri
ng, Qing
da
o Univ
ersit
y
,
No.30
8
, Nin
gxi
a
Roa
d
, Qingd
ao, Shan
do
ng,
266
07
1, Chin
a, Ph./F
ax: 053
2
-
859
560
69
2
Qingzho
u Cig
arette F
a
ctor
y
,
Shan
do
ng T
obacco Industr
y
Co., Ltd,
No.18
18, Li
ngl
ongs
ha
nbe
i Ro
ad, Qingzh
ou,
Shan
do
ng, 26
2
500, Ch
in
a, Ph./F
ax: 0531-
82
599
00
1
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:
y
g
w
_
ustb
@1
63.com*,
w
a
ng
xi
aoc
hun
_
y
t@
126.com,
yy_b
onn
ie@
1
6
3
.co
m
A
b
st
r
a
ct
Accordi
ng to th
e lic
ense
pl
ate
recog
n
itio
n pr
o
b
le
m,
this
pa
pe
r did th
e res
ear
ch ab
out l
i
cens
e pl
at
e
locati
on
an
d c
haracters
reco
gniti
on. It pr
op
osed
tw
o new
alg
o
rith
ms, the
y
sep
a
rately
ar
e lic
ens
e l
o
cati
o
n
a
l
go
ri
thm
b
a
s
ed
on
co
l
o
r segm
en
ta
ti
on
a
n
d
fa
u
l
t-to
le
ra
nt
characters
rec
ogn
ition
a
l
gor
ithm b
a
se
d o
n
BP
neur
al
netw
o
rk
. In the
pre-
pro
c
essin
g
stag
e,
it pro
pose
d
image
en
ha
nc
e
m
ent a
l
gor
ith
m
w
h
ich co
uld
make
the imag
e mor
e
easi
l
y an
aly
z
ed by co
mp
ute
r
. In the lo
catio
n
stage, it ma
d
e
utili
z
a
t
i
on
of color a
nd sh
ap
e
infor
m
ati
on, an
d then pr
op
ose
d
locati
on
alg
o
r
ithm. In t
he r
e
cogn
ition sta
g
e
,
it fu
lly ma
de t
he co
nsid
eratio
n
of characters
’ fault-toler
ant, and th
en
made th
e us
e
of improve
d
BP neur
al n
e
tw
ork to recog
n
i
z
e
characters. It did so
me exp
e
ri
ment by MA
T
L
AB. Exper
i
m
e
n
ts show
that the s
peci
a
l
licens
e pl
ate fault-
tolera
nt chara
c
ters recogn
iti
on al
gorith
m
i
s
more
acc
u
r
a
te than the
origi
n
a
l
lice
n
s
e
plate rec
o
g
n
itio
n
meth
ods, an
d i
t
s recogniti
on r
a
te has b
een i
m
pr
ove
d
greatl
y
.
Ke
y
w
ords
:
ch
aracters reco
g
n
itio
n, color se
gmentati
on, fa
ult-toler
ant, BP neura
l
netw
o
rk, MAT
L
AB
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
Intelligent tra
n
sp
ortation
system has b
e
com
e
an im
portant di
re
ction of curre
n
t
traffic
manag
eme
n
t and
develo
p
m
ent, and
license pl
ate reco
gnition
as a
kind
of traffic informati
o
n
acq
u
isitio
n techn
o
logy ha
s incre
a
si
ngly attracte
d more and more a
ttention. It can automatical
ly
authenti
c
ate t
he identity of the v
ehicle,
and ma
ke ve
hicle m
ana
ge
ment, traffic flow control a
nd
intersectio
n
p
a
yment highl
y automated, so it
has a
wi
de ran
ge of p
r
acti
cal ap
plication.
For a li
cen
s
e
plate re
cog
n
i
tion system,
the re
cog
n
ition pro
c
e
s
s ge
nerally in
clud
es the
fo
llo
w
i
ng
s
t
ep
s
:
lic
e
ns
e pla
t
e
imag
e
pr
e
p
r
o
c
e
s
s
i
ng,
location,
ch
ara
c
ter segm
entation, feat
ure
extraction, chara
c
te
r
re
cogniti
on a
n
d
post-processing. Fi
rstly,
usin
g the color
seg
m
ent
ation
techn
o
logy fi
nd the p
o
ssi
b
le re
gion t
h
roug
h
colo
r
histog
ram, th
en test the
l
ength-width
ratio
,
length, heigh
t and plate texture of the region to
o
b
tain the best location. Seco
ndly, did the
cha
r
a
c
ter
seg
m
entation an
d feature vectors extra
c
ti
o
n
. Thirdly, put the vectors
to the improved
BP neural n
e
twork fo
r training. L
a
stly
, recogni
z
e
d
the cha
r
a
c
ters.
Figu
re
1
is li
cen
s
e
p
l
ate
recognitio
n
system.
Figure 1. Lice
nse Plate Re
cog
n
ition System Block
Di
agra
m
2.
License Plate Image Pre-proces
sing
2.1. Image Enhanc
ement
Curre
n
tly, th
ere a
r
e seve
ral lice
n
se pl
ate
locatio
n
algorith
m
s [1
, 2]. When u
s
ing the
histog
ram eq
ualization me
thod to adjust the bright
ne
ss of the lice
n
se plate ima
ge, we find that
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3281 – 32
88
3282
light and dark cont
ra
st betwee
n
the chara
c
te
rs
an
d backg
rou
n
d
in the image is we
ake
ned,
whi
c
h will red
u
ce the a
c
cu
racy of location and re
cog
n
ition. So we
adapt the followin
g
step
s to
adju
s
t the bri
ghtne
ss of th
e plate image
gradu
ally.
Statistic the n
u
mbe
r
of blue
pixels of the
i
m
age, then fi
x the value of a and b to m
eet the
requi
rem
ents
that the number of
pixels which b
r
ightn
e
ss value x
[min,a] acco
unts for 5% of
the
total numbe
r of blue pixels and the nu
mber of pixel
s
whi
c
h b
r
igh
t
ness value x
[b,max] als
o
accou
n
ts for
5% of the total numbe
r of blue pixel
s
.
Set the b
r
ig
htness val
u
e
which
is le
ss t
han
min to
a,
and
the
b
r
ig
htness val
u
e
whi
c
h i
s
greate
r
than
max to b.
Similarly, do the same
steps to the red and g
r
e
e
n
pixels. Fro
m
doing a
b
o
v
e , we
accompli
sh th
e brightn
e
ss
adju
s
tment.
After the bri
ghtne
ss a
d
ju
stment, we
sho
u
ld do t
he filtering t
o
make the
further
enha
ncement
.
2.2. License
Plate Loca
t
i
on
In the pro
c
e
ss of ima
g
e
analysi
s
an
d pro
c
e
s
sing
, it is consi
dera
b
le to select a
n
approp
riate thre
shol
d to separate
the t
a
rget from b
a
ckgroun
d. In
this pa
pe
r, a new th
re
shold
segm
entation
algorithm i
s
as follo
ws:
Assum
e
co
n
s
ide
r
R co
mp
onent firstly.
L mean
s the gray scale of the image,
i
n
mean
s
the numbe
r o
f
pixels who
s
e gray value
are i, and N
mean
s the total numbe
r of pixels.
L
i
i
N
1
n
(1)
Cal
c
ulate pi:
1
,
0
,
1
L
i
i
i
i
i
p
p
N
n
p
(2)
The im
age
i
s
divid
e
d
into two
p
a
rts.
They
are
C0 (ta
r
get
)an
d
C1
(b
ackg
ro
und).
C0
mean
s pixel
s
who
s
e
gray scale a
r
e [
1
,…,k],
and
C1 m
ean
s p
i
xels who
s
e
gray
scale a
r
e
[k
+
1
,…,L].
)
(
1
0
k
w
p
k
i
i
w
(3)
)
(
1
1
1
k
w
p
L
k
i
i
w
(4)
Average g
r
ay
scal
e
of C0 a
nd C1 a
r
e:
0
1
0
u
w
ip
k
i
i
i
(5)
1
1
1
u
w
ip
L
i
k
i
i
(6)
Average g
r
ay
scal
e
of the imagine i
s
:
1
1
0
0
1
u
u
w
u
w
ip
L
i
i
i
T
(7)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A New Li
cen
s
e Plate Fault-tolerant Cha
r
ac
ters
Recog
n
ition Algorith
m
(Guowei Y
ang)
3283
Within-cla
ss varian
ce of C0 and C1 are:
0
1
2
0
2
0
)
1
(
w
p
u
k
i
i
i
(8)
1
1
2
1
2
1
)
1
(
w
p
u
L
i
k
i
i
(9)
Between
-cl
a
ss va
rian
ce o
f
C0 and C1 and po
pulatio
n varian
ce a
r
e:
2
0
1
1
0
2
1
1
2
0
0
2
)
(
)
(
)
(
u
u
w
w
u
u
w
u
u
w
T
T
B
(10)
L
i
i
i
T
T
p
u
i
1
2
2
)
(
(11)
Set paramete
r
:
2
2
T
B
(12)
Whe
n
takes to th
e
maximum,
the corre
s
pondi
ng
k
value is th
e be
s
t
threshold.Sim
ilarly, do
the
same
cal
c
ulat
ion to
G
and
B com
pon
ent
, then
get the
be
st threshol
d.
In Chin
a, the
r
e a
r
e th
ree
kind
s of
ratio
:
polic
e ca
rs and
milita
r
y
vehicl
e
(white backg
rou
n
d
)
is
3.8, large ve
hicle
(ba
c
k p
l
ate) is
2.0, and the
re
m
a
ining vehi
cl
es i
s
3.6. We sel
e
ct the
ratio
arou
nd 3.6 b
e
ca
use we
mainly
study
the blue-whi
t
e plate. Figure 2 a
r
e th
e locatio
n
image
s
based on
col
o
r se
gme
n
tation.
Figure 2. The
Licen
s
e Plat
e Locat
ion Ba
sed o
n
Col
o
r
Segmentatio
n
3.
Char
acters
Reco
gnition
Bas
e
d on Fe
ature Ex
tra
c
tion and BP Neur
al Net
w
ork
3.1. Chara
c
ter Segmenta
tion
Firstly, tra
n
sf
orm th
e lo
cat
i
on ima
ge i
n
to bin
a
ry im
a
ge
whi
c
h
backgroun
d i
s
b
a
ck a
n
d
target i
s
whit
e. Plate
regi
on bi
nari
z
atio
n is go
od
or
bad
whi
c
h
di
rectly im
pa
ct the
accu
ra
cy of
cha
r
a
c
ter
se
gmentation a
nd re
cog
n
itio
n. There
a
r
e
several traditi
onal charact
e
r se
gme
n
tation
method
s
su
ch as ho
rizont
al proje
c
tion
[3], template matchin
g
[4], clu
s
ter
analy
s
is [5], et
c.
We
adapt the ho
ri
zontal p
r
oje
c
t
i
on method h
o
rizontal p
r
oj
ection.
3.2. Chara
c
ter Image Nor
m
alization
The
cha
r
a
c
te
rs segme
n
te
d from
the
pl
ate imag
e a
r
e not i
n
the
same
size.
In
order to
recogni
ze
the
cha
r
a
c
ters
convenie
n
tly, we
sho
u
ld tu
rn them into
the same
si
ze
. It adapt bilin
ear
interpol
ation
algorith
m
to transfo
rm th
ese cha
r
a
c
te
r image
s into 20*35. Assume
H as
th
e
origin
al imag
e’s h
o
ri
zontal
proje
c
tion
a
nd V mean t
he ori
g
inal i
m
age’
s verti
c
al proj
ectio
n
. M
mean
s the h
e
ighth after
n
o
rmali
z
atio
n
and
N mea
n
s
the
width a
fter norm
a
lization. Point(m,n)
after normalization is expre
s
sed a
s
[6]:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3281 – 32
88
3284
i
k
j
k
k
H
M
k
H
1
1
)
(
)
(
m
(13)
i
k
j
k
k
V
N
k
V
1
1
)
(
)
(
n
(14)
3.3. Chara
c
ter Feature E
x
tra
c
tion
It adapt coa
r
se g
r
id featu
r
e extra
c
tion
in this pa
pe
r. The ba
si
c idea of coa
r
se
grid
feature
extra
c
tion i
s
: eq
u
a
lly divide t
he
cha
r
a
c
ter image
which is
after
si
ze
and l
o
cation
norm
a
lization
into M*N gri
d
s, then stati
s
tics the nu
mber of whit
e pixels in e
v
ery grid
s. If the
numbe
r of wh
ite pixels ove
r
20% of total
pixels, set
fe
ature valu
e of this gri
d
to1, otherwise to
0.
In this
pap
er,
we
no
rmali
z
e the
s
e
ch
aracters into
70
*50, then
divi
de them
into
7*5 g
r
id
s. So
we
can
get th
e f
eature
vecto
r
of ea
ch
cha
r
acte
r,
they are all
3
5
-di
m
ensi
onal
ve
ctors whi
c
h are
comp
osed by
0 and 1.
3.4. Chara
c
ter Rec
ogniti
on
3.4.1. Chara
c
ter Fa
ult-tol
e
rant
We ad
apt the method whi
c
h combi
ned f
eatur
e extraction and BP neural n
e
two
r
k. When
recogni
zin
g
the characte
rs, take fully
accou
n
t
of chara
c
te
r fault
-
tolerant. An
d calcul
ate n
o
isy
sampl
e
s whi
c
h
are
theo
retically allo
wed of e
a
ch
chara
c
te
r. Ta
ke the
Chin
ese characte
rs for
example t
o
e
x
plain. Ta
ble
1 i
s
fe
ature
vectors
of th
e several
Chi
nese
cha
r
a
c
t
e
rs fro
m
the
31
Chin
ese ch
aracters.
Table 1. Feat
ure Ve
ctors o
f
the Several Chin
ese Cha
r
acters
藏
1 1 1
1 1
0 1
1
1 0
1 1 1
1
0
1
1
1
1
0
1
1
1
1
0
1
1
1
1
0 0 0
0 0
0
川
0 0 0
1 0
1 1
1
1 0
1 1 1
1
0
1
1
1
1
0
1
1
1
1
0
1
1
1
1
0 1 0
0 1
0
鄂
1 1 1
1 1
1 1
1
1 1
1 1 1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
1
1
1 0 1
1 1
0
赣
1 1 1
1 1
1 1
1
1 1
1 1 1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1 1 1
1 1
1
贵
0 0 0
0 0
1 1
1
1 0
1 1 1
1
0
1
1
1
1
0
1
1
1
1
0
1
1
1
1
0 1 1
0 1
1
桂
1 1 1
1 0
1 1
1
1 0
1 1 1
1
1
1
1
1
1
0
1
1
1
1
1
1
1
1
1
0 0 1
0 1
1
黑
0 1 1
1 0
0 1
1
1 1
0 1 1
1
1
0
1
1
1
0
0
1
1
1
0
1
1
1
1
1 0 0
0 0
0
沪
1 1 1
1 0
1 1
1
1 1
1 1 1
0
1
1
1
1
1
1
1
1
1
0
0
1
1
1
0
0 1 1
1 0
0
吉
1 1 1
1 1
1 1
1
1 1
1 1 1
1
0
1
1
1
1
1
1
1
1
1
1
1
0
0
1
1 1 1
1 1
1
冀
1 1 1
1 0
1 1
1
1 0
1 1 1
1
0
1
1
1
1
0
1
1
1
1
0
1
1
1
1
1 0 0
0 0
1
津
1 0 1
0 0
0 1
1
1 1
1 1 1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1 1 0
1 0
0
晋
1 1 1
1 1
1 1
0
1 1
1 1 1
1
1
1
1
1
0
0
0
1
0
1
0
1
1
1
1
0 0 1
1 1
0
京
1 1 1
1 0
1 1
1
1 0
1 1 1
1
0
1
1
1
1
0
1
1
1
1
0
1
1
1
1
1 1 1
1 0
1
辽
1 1 1
1 1
0 1
0
1 0
1 1 1
1
0
1
1
1
1
0
1
1
1
1
0
1
1
1
1
0 1 1
1 1
1
鲁
0 1 0
0 0
1 1
1
1 0
1 1 1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1 0 1
1 1
1
蒙
0 1 1
1 0
0 1
1
1 0
0 0 1
0
0
0
1
1
0
0
0
1
1
1
0
0
1
1
1
1 0 0
1 0
0
闽
0 0 1
1 0
0 0
1
0 1
1 1 1
1
1
1
1
1
1
1
1
0
1
1
1
1
1
0
1
1 0 0
0 0
0
琼
0 0 1
0 0
1 1
1
1 1
1 1 1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1 1 0
1 1
1
陕
0 0 0
0 0
1 1
1
1 0
1 1 1
1
1
1
1
0
1
0
1
1
1
1
1
1
1
1
1
0 1 0
1 0
1
苏
1 1 1
1 1
0 1
1
1 0
1 1 1
1
0
0
1
1
1
1
1
1
1
1
1
0
1
1
1
0 1 1
1 1
0
湘
0 0 1
1 0
0 0
1
1 1
0 0 1
1
1
0
1
1
1
1
1
1
1
1
1
1
0
1
1
1 0 0
0 0
0
新
1 1 1
1 1
1 1
1
1 0
1 1 1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1 0 1
0 1
1
渝
0 0 0
0 0
1 0
1
1 0
0 1 1
1
0
0
1
1
0
0
0
1
1
1
0
1
1
1
1
1 1 1
1 0
0
豫
1 1 1
1 1
1 1
1
1 1
1 1 1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1 0 0
1 1
0
粤
1 1 1
1 1
1 1
1
1 1
1 1 1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
1
1
0 0 0
1 1
0
云
0 1 1
1 0
0 1
1
1 0
0 0 0
0
0
1
1
1
1
1
0
1
1
0
0
1
1
0
1
0 1 1
1 1
1
浙
1 0 1
1 1
0 1
1
1 0
0 0 1
1
1
0
1
1
1
1
1
0
1
1
1
1
0
1
1
1 1 1
1 0
1
青
0 1 1
1 1
0 1
1
1 1
1 1 1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1 1 1
0 1
1
皖
1 1 1
1 1
1 1
1
0 1
1 1 1
1
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1 1 1
1 1
1
宁
1 1 1
1 0
1 1
0
1 1
1 1 1
1
1
0
0
1
1
0
0
0
1
1
0
0
0
1
1
0 0 1
1 0
0
甘
0 1 0
1 1
0 1
1
1 1
0 1 0
1
1
0
1
1
1
1
0
1
0
1
1
0
1
0
1
1 0 1
1 1
0
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
”
藏
”and oth
e
r
s a
r
e:
7,10,11,10,7,
8,11,12,12,4,
10,
7,6,13,11,
12,14,12,8,1
3
,8,
15,8,8,15,1
4
,9,11,13,17;
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A New Li
cen
s
e Plate Fault-tolerant Cha
r
ac
ters
Recog
n
ition Algorith
m
(Guowei Y
ang)
3285
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
川
”a
nd oth
e
rs a
r
e :
7,13,12,3,8,1
1
,12,13,7,11,
13,7,9,10,12,
13,9,7,11,14,
22,
20,11,1
1
,1
4,17,11,12,1
6
,
18
；
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
鄂
”a
nd oth
e
rs a
r
e :
10,13,3,14,7,
12,9,6,10,8,8,
9,10,7,15,12,
8,14,6,13,4,1
7
,2,2,17,12,6,
5,11,9;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
赣
”a
nd oth
e
rs a
r
e :
11,12,3,11,6,
13,8,3,9,7,9,
6
,
7,6,18,13,5,1
1
,7,14,3,
16,3,
5,14,9,3,2,14,
12;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
贵
”a
nd oth
e
rs a
r
e :
10,3,14,11,6,
13,13,12,8,1
2
,
14,7,
8,7,17,1
6
,8,6,12,17,1
0
,9
,14,14,13,
16,10,11,1
7
,1
9;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
桂
”a
nd oth
e
rs a
r
e :
7,8,7,6,7,11,1
0
,9,5,11,
10,6,
7,6,16,12,9,9,
9,14,3,
16,7,7,
14,13,7,8,12,
16;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
黑
”a
nd oth
e
rs a
r
e :
8,11,12,13,1
4
,
11,13,16,8,1
0
,12,
11,14,1
3
,
7,10,12,14,1
4
,8,12,
11,10,
12,15,14,1
0
,1
5,11,10;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
沪
”a
nd oth
e
rs a
r
e :
11,12,9,8,13,
10,13,11,1
1
,9
,11,
8,11,12,1
6
,15,11,13,1
1
,
16,11,
16,9,9,
12,15,11,1
0
,1
1,18;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
吉
”a
nd oth
e
rs a
r
e :
12,13,6,3,12,
9,16,11,10,1
0
,
12,
7,8,9,19,1
4
,8,13,8,15,6,
17,6,8,13,10,
6,3,15,13;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
冀
”a
nd oth
e
rs a
r
e :
12,7,10,9,8,5,
8,11,10,10,1
2
,
3,8,9,11,12,1
0
,10,12,13,6,
13,8,10,15,1
4
,
10,9,13,19;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
津
”a
nd oth
e
rs a
r
e :
4,11,8,7,12,1
1
,10,9,10,10,
14,9,12,9,15,
8,4,8,10,9,10,
13,6,8,17,10,
8,9,15,15;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
晋
”a
nd oth
e
rs a
r
e :
10,13,8,9,14,
10,12,11,1
2
,1
2,14,
11,8,14,
15,18,14,1
6
,1
2,19,10,
15,8,
8,17,18,12,1
1
,9,13;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
京
”a
nd oth
e
rs a
r
e :
7,7,9,6,7,6,11
,8,7,3,9,11,5,8,12,
15,9,9,9,
16,7,10,9,
11,
12,11,9,6,12,
18;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
辽
”a
nd oth
e
rs a
r
e :
6,9,10,7,8,7,1
4
,11,8,8,
12,8,
5,11,15,18,1
2
,
13,6,19,8,
14,
10,10,11,1
2
,8
,7,13,17;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
鲁
”a
nd oth
e
rs a
r
e :
13,10,7,6,7,6,
13,12,9,9,9,
1
4
,8,11,16,13,
5,7,11,14,5,
1
2
,7,9,14,13,7,
8,17,12;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
蒙
”a
nd oth
e
rs a
r
e :
11,12,15,1
8
,1
7,16,7,16,19,
11,15,
15,1
2
,1
5,16,17,18,1
7
,13,12,15,
10,
15,15,12,1
5
,1
7,18,14,14;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
闽
”a
nd oth
e
rs a
r
e :
12,13,12,1
3
,1
6,12,10,15,1
4
,
12,8,
18,15,1
8
,13,17,10,1
4
,16,7,12,
17,1
0
,12,19,14,1
0
,
13,17,15;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
琼
”a
nd oth
e
rs a
r
e :
14,9,8,5,8,9,1
2
,11,8,10,4,
1
4
,9,12,5,18,1
0
,6,12,11,8,
1
3
,6,8,15,12,6,
7,17,15;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
陕
”a
nd oth
e
rs a
r
e :
12,7,14,11,6,
9,14,13,13,1
0
,
8,
16,9,13,7,1
7
,14,6,14,15,
12,
11,12,1
2
,1
7,14,12,13,1
7
,
19;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
苏
”a
nd oth
e
rs a
r
e :
8,11,6,7,12,9,
14,11,8,12,1
0
,
12,9,6,11,13,
16,12,14,1
5
,8
,
15,8,5,13,9,8
,
7,13,11;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
湘
”a
nd oth
e
rs a
r
e :
13,14,13,1
4
,1
7,14,8,16,15,
13,9,
19,16,1
9
,
14,12,7,11,1
5
,15,13,
14,1
1
,
13,18,9,11,1
6
,16,14;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
新
”a
nd oth
e
rs a
r
e :
8,11,4,3,10,3,
12,11,6,6,10,
10,7,8,5,15,1
2
,8,12,8,13,
1
7
,4,6,15,10,4,
5,15,13;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
渝
”a
nd oth
e
rs a
r
e :
15,10,17,1
6
,9
,16,11,16,17,
13,13,
15,1
0
,1
4,12,10,17,1
3
,11,15,14,
17,
17,19,16,1
5
,1
7,16,16,16;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
豫
”a
nd oth
e
rs a
r
e :
8,11,2,3,14,7,
10,9,6,8,6,
8,9
,
10,7,15,10,6,
12,8,11,4,
17,
2,17,12,6,5,1
3
,11;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
粤
”a
nd oth
e
rs a
r
e :
8,11,2,5,14,7,
12,9,8,10,8,
8,
11,10,9,15,1
2
,
8,12,5,13,6,
1
9
,2,17,14,8,7,
11,11,;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
云
”a
nd oth
e
rs a
r
e :
15,14,17,1
4
,1
3,14,15,12,1
3
,
15,
17,17,12,
11,14,12,1
9
,1
5,17,13,
18,1
5
,
16,17,17,15,
13,14,20,1
5
;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
浙
”a
nd oth
e
rs a
r
e :
14,17,12,9,1
6
,
13,14,15,10,
14,10,18,
1
1
,1
2,13,15,14,1
2
,14,9,9,
10,15,
12,14,15,1
0
,1
1,16,15;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
青
”a
nd oth
e
rs a
r
e :
9,11,6,3,10,7,
10,11,6,10,8,
12,9,8,7,17,1
0
,6,12,8,11,
4,
17,6,8,13,10,
5,17,11;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
皖
”a
nd oth
e
rs a
r
e :
11,12,5,2,11,
8,15,10,3,9,9,
11,6,7,8,18,1
3
,7,13,7,16,
5,
16,5,7,14,11,
5,16,14;
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
宁
”a
nd oth
e
rs a
r
e :
13,16,11,1
4
,1
7,12,11,11,1
5
,
13,15,
9,12,1
3
,17,14,17,1
7
,
17,13,16,
15,
16,13,11,2
0
,1
6,17,16,16;
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3281 – 32
88
3286
The num
ber
of different bits of feature v
e
ctor b
e
twe
e
n
“
甘
”a
nd oth
e
rs a
r
e :
17,18,9,12,1
9
,
16,10,18,13,
19,15,13,
1
8
,1
7,12,14,15,1
5
,19,11,4,
13,1
6
,11,11,15,1
5
,
11,14,16.
The
numb
e
r
of differe
nt bi
ts of fe
ature
vect
or is al
so
the
Hammi
n
g
di
stan
ce
b
e
twee
n
different cha
r
acters. We
co
uld cal
c
ulate
noisy
fe
ature
vectors
if we kno
w
th
e
mini
mum Hammi
ng
distan
ce
η
. Since some ch
ara
c
ters noi
sy sample
s are t
heoreti
c
all
y
allowed are
too much. we are
here to
set
a
theoreti
c
al p
r
emi
s
e, 1) Refuse to
re
co
gnize charact
e
r which is
n
o
t clo
s
e to a
n
y
c
h
ar
ac
te
r
.
2) fa
u
l
t-
to
le
ra
nt s
e
lec
t
e
d
as
2. If the
minimum
Ha
mming di
sta
n
ce
between
a
cha
r
a
c
ter
and
others i
s
ove
r
than
4, we
should
cal
c
ul
ate 2-dimen
s
io
nal noi
sy sam
p
les
whi
c
h
are
theoreti
c
ally allowed and
then put the
m
to the cl
a
ssifie
r
whi
c
h
is de
sign
ed
according to
the
followin
g
met
hod fo
r trainin
g
. If the mini
mum
Hammi
ng di
stan
ce
b
e
twee
n a
cha
r
acte
r
and
ot
hers
is not more than 4, we sho
u
ld do secon
dary re
co
gniti
on.
3.4.2. Classifier Design
Whe
n
trainin
g
the stand
a
r
d sa
mple
s (65 cla
s
s ch
a
r
acte
rs), we must have m
o
re than
one time
sam
p
les. In this
p
aper,
we ta
ke
10 time
s sa
mples. We
d
e
sig
n
68 cla
s
sifiers.
They are
Chin
ese character cl
assifie
r
, digital
cla
s
sifier
and
alp
habet
cla
ssifi
er, othe
r 65
classifiers for t
he
other 65
cl
a
s
s cha
r
a
c
ters. Th
e featu
r
e ve
ctor
through
the
m
e
thod
of rou
gh g
r
id
feat
ure
extraction
is
35-di
men
s
ion
a
l, so
we
ada
pt a 3-
l
a
yer B
P
neural n
e
twork
whi
c
h
contain
s
a
hid
d
e
n
layer, its inp
u
t
node
s i
s
35
, output
node
s i
s
1 a
nd th
e nu
mbe
r
of
neuron
s
whi
c
h hid
den
lay
e
r
contai
ned is
different from
each othe
r. Therefore,
in
the actual de
sign,
the nu
mber of neu
rons
whi
c
h hid
den
layer co
ntain
ed ca
n cal
c
ul
ated by the empiri
cal form
ula (15
)
.
51
.
0
35
.
0
77
.
0
54
.
2
12
.
0
43
.
0
2
n
m
n
mn
s
(15)
Thro
ugh
cal
c
ulation a
nd repeate
d
expe
riment,
we ult
i
mately determined the
nu
mber
of
neuron
s whi
c
h contai
ned
by Chine
s
e
cha
r
a
c
ter n
e
twork’
s hidd
e
n
layer is 2
0
,
the numbe
r of
neuron
s whi
c
h
containe
d by
alpha
bet
netwo
rk’
s
hid
den laye
r i
s
16,and th
e n
u
mbe
r
of n
e
u
r
on
s
whi
c
h contain
ed by digital netwo
rk’
s
hid
den layer i
s
8
[7].
Firstly de
sign
3 classifie
r
s:
Chine
s
e
cha
r
acte
r cl
assifi
er, alpha
bet
cla
ssifie
r
and
digital
cla
ssifie
r
. Ta
ke Chi
n
e
s
e
ch
ara
c
ter
cla
s
si
fier a
s
an
exa
m
ple. First training th
e BP neu
ral n
e
two
r
k
and
dema
n
d
that
when
i
nput the
31
Chin
ese
cha
r
acters’
sta
n
d
a
rd
vecto
r
, t
he o
u
tput i
s
1,
otherwise the
output i
s
0. The
traini
ng
is
su
cce
ssfu
l
until the
ou
tput erro
r i
s
l
e
ss tha
n
0.0
1
.
De
sign al
pha
bet cla
ssifie
r
and digital
cla
ssifie
r
as the
same p
r
in
cipl
e.
Secon
d
ly, design 6
5
ch
ara
c
ter
classifie
r
s whi
c
h me
a
n
s that de
sig
n
a cla
ssifie
r
for each
c
h
ar
ac
te
r
.
Tak
e
“
鲁
”
as an
example. Fi
rst training th
e
BP neural net
work
and
de
mand th
at wh
en
input the sta
ndard vecto
r
of “
鲁
”, the output is 1,
otherwise th
e output is 0
.
The trainin
g
is
su
ccessful u
n
til the outpu
t erro
r is le
ss than 0.0
1
. Input 2-dim
e
nsio
nal noi
sy
sampl
e
vect
ors
whi
c
h a
r
e the
o
retically allo
wed a
nd oth
e
r
ch
ara
c
te
rs,
then ret
r
ain t
he network a
nd dem
and t
hat
whe
n
input 2
-
dime
nsi
onal
noisy sample
vectors,
the output is 1, o
t
herwi
se
the
output is 0. T
he
training i
s
su
ccessful u
n
til the out
put e
r
ror i
s
le
ss th
a
n
0.01. Fig
u
re 3 is th
e trai
ning result of
“
鲁
”
cla
ssifie
r
with
out noise. Figure 4 is the training result of “
鲁
” c
l
ass
i
fier with nois
e
.
Figure 3. Trai
ning Result of “
鲁
” Class
i
fier
without Noise
Figure 4. Trai
ning Result of “
鲁
” Class
i
fier with
Noi
s
e
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A New Li
cen
s
e Plate Fault-tolerant Cha
r
ac
ters
Recog
n
ition Algorith
m
(Guowei Y
ang)
3287
3.4.3. Chara
c
ter
Reco
gni
tion
Save the su
ccessful n
e
u
r
al n
e
two
r
k’
s weig
ht and
bias
matrix,
and the
n
g
e
t re
sult
throug
h
ope
ration the
s
e
matrix with
u
n
kn
own
cha
r
acter.
First in
put the
un
kn
own
cha
r
act
e
r to
Chin
ese, alp
habet an
d di
gital re
cog
n
itions
sep
a
ratel
y
. If
it is Chin
ese, inp
u
t the
cha
r
a
c
ter to
31
Chin
ese cla
s
sifiers for recogni
zing, the
n
we
shoul
d get 31 floatin
g-poi
nts which betwe
en 0
and
1. Take “
鲁
” a
s
an exampl
e
,
the 31 floating-p
o
ints a
r
e
as follo
ws:
(0.867
3, 0.2
231, 0.176
5,
0.3943, 0.3
754, 0.248
7, 0.4612,
0.4
752, 0.321
1, 0.2513,
0.2435, 0.18
23, 0.1286, 0.3341,
0.27
14, 0.1123, 0.3522, 0.34
78, 0.4013, 0.2231, 0.24
63
,
0.2847, 0.26
4
5
, 0.2336, 0.3
156, 0.374
6, 0.
2568, 0.29
3
1
, 0.7156, 0.2
460, 0.302
5).
Comp
are the
s
e floatin
g-po
ints an
d
we
could
find th
at
0.8673 i
s
m
u
ch bi
gge
r tha
n
othe
rs
and it
clo
s
e
s
to 1. And
0.
8673
is the
output
corre
s
pondi
ng
“
鲁
”
cla
ssif
i
e
r
.
S
o
we
c
an m
a
k
e
a
con
c
lu
sio
n
that the unkn
o
w
n characte
r
is “
鲁
”.
Cha
r
a
c
ter
wh
ose
Hammin
g
distan
ce be
tween it and
stand
ard
“
鲁
”
cha
r
a
c
ter i
s
not more
than 2
can
b
e
re
co
gnized
as
“
鲁
”. Cha
r
a
c
ter wh
ose
Hamming dist
a
n
ce between it
and
othe
rs is
more tha
n
2 can b
e
refu
sed to recogni
ze. If the con
s
eq
uent is no
t good enou
g
h
, we sh
ould
do
se
con
dary
r
e
cog
n
it
ion.
3.4.4. Secon
d
ar
y
Recogn
ition
For the
s
e ch
ara
c
ters wh
o
s
e Ha
mming
distan
ce is n
o
t more than
4, the result is not so
good
by ado
pting the a
b
o
v
e method,
so we
do th
e
seco
nda
ry re
cognition. T
a
ke alph
abet a
s
an
example,
cha
r
acte
rs
who
s
e minimu
m
Hamming
dista
n
ce
is
not mo
re tha
n
4
are
D E F K
Q
R
V
W. We a
dopt
13-p
o
int feature extra
c
tio
n
to extr
act the 8 charact
e
rs’ fe
at
ure. De
sign cla
ssi
fiers
for the
8
ch
aracters.
Then
input the
feat
ure
vect
o
r
s t
o
the
BP ne
u
r
al
network fo
r trai
ning.
Ta
ke
“D” as an ex
ample, dema
nd that
when
input “D”
sta
ndard vecto
r
,
the output is 1, otherwi
se
the
output
i
s
0. The
t
r
ainin
g
i
s
su
ccessful
until
the
outp
u
t erro
r i
s
l
e
ss than
0.0
1
. Next,
save t
he
su
ccessful ne
ural
net
work’
s
weig
ht
an
d bias
ma
trix, a
nd the
n
g
e
t result th
ro
ugh
operation th
e
s
e
matrix with
u
n
kn
own
cha
r
acter.
At the
same
ti
me,
al
so get som
e
floating-p
o
int
s
whi
c
h between
0 and 1. Find
the bigge
st point whi
c
h m
ean
s we
c
oul
d kno
w
the re
sult. As an ex
ample, the "D",
finally get a
set of that
n
u
mbe
r
: (0.9
8
58,
0.231
2, 0.1432,
0.2
8
97,
0.381
3, 0.1783,
0.4
3
3
2
,
0.2526
). 0.9
858 is m
u
ch
bigge
r than
others and
clo
s
e to 1, so we
can
co
nclu
de that the
unkno
wn cha
r
acte
r is
“D”.
4. Conclu
sion
No matter n
o
rmal trai
nin
g
or noisy traini
ng, we
could ch
oo
se
the same pa
ramete
rs.
Acco
rdi
ng to
seve
ral exp
e
rime
nts a
n
d
the analy
s
i
s
of the
net
work
stru
ctu
r
e, we
set th
e
para
m
eters a
s
Table 2.
Table 2. The
Initial Value of the Netwo
r
k Para
meters
Net
w
ork t
y
pes
Learning r
a
te
Momentum const
ant
Ma
ximum c
y
cles
Display
inte
rval
Output e
rro
r
Chinese 0.08
0.9
5000
30
0.01
alphabet
0.08
0.9
5000
30
0.01
digital 0.08
0.9
5000
30
0.01
In the testing
stage,
we
choo
se 1
00 te
sting
sam
p
le
s for
ea
ch
cl
assifier a
nd t
he pa
rtial
results a
r
e a
s
Table 3.
Table 3. Part
of the Test Result
s
Classifiers
Testing samples number
Correct num
ber
Accuracy
“
鲁
”classifier
100
90
90%
“A” classifier
100
93
93%
“3” classifier
100
96
95%
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3281 – 32
88
3288
Since
use th
e col
o
r
and
sha
pe info
rm
ation in th
e l
o
catio
n
sta
g
e
,
the location
plate i
s
more
a
c
curat
e
which offe
rs many
adv
antage
s
fo
r t
he
cha
r
a
c
ter se
gmentatio
n an
d featu
r
e
extraction.
In
the
re
cog
n
ition
stage,
the
accu
ra
cy h
a
s
im
prove
d
g
r
eatly b
e
cau
s
e of
usi
ng th
e
improve
d
BP neural network to re
cogni
ze the fault-tolerant cha
r
a
c
ters.
Although the
BP algorithm
has
a solid t
heoretic
al ba
sis and high versatility,
it
also ha
s
s
o
me
weak
points
[8],
The improve
m
ent is a
s
follows:
Optimize
the
initial
weig
hts. Sele
ct
ran
dom
numb
e
r between
0
and
1 to
be
the initial
weig
hts in ge
neral.
Adjust lea
r
ni
ng rate
adap
tively. To ensure
syst
em’
s
stability, sele
ct a sm
aller l
earni
ng
rate between
0.08 and 0.1.
Incre
a
se the
momentum t
e
rm. Mome
ntum te
rm ha
s
the effect of smooth
and
cu
shio
n,
whi
c
h
coul
d
help im
prove
the stability of the netwo
rk’
s
convergence proces
s and also
solve t
he
probl
em
s of local mini
mu
m to some ex
tent.
Ackn
o
w
l
e
dg
ements
This
work wa
s finan
cially sup
porte
d by
National
Nat
u
ral Sci
e
n
c
e
Found
ation o
f
China
unde
r Grant No. 612
720
7
7
, 60973
048.
Referen
ces
[1]
F
r
ank Y, Shih
, Souxia
n Che
ng. Automatic
seede
d re
gio
n
gro
w
i
n
g for color im
age s
egme
n
tatio
n
.
Imag
e an
d Visi
on Co
mputeri
n
g
. 2005; 2
3
: 87
7-88
6.
[2]
Li Yuch
eng,
Yang Gua
ngm
ing, W
ang M
u
shu. St
ud
y a
nd Desi
gn for
S
y
stem of Licens
e Plate
Extractio
n
an
d
Reco
gniti
on.
C
o
mputer Me
as
ure
m
e
n
t and C
ontrol
. 20
11; 1
9
(1): 158-
16
0.
[3]
Hu
Xi
aofe
ng, Z
hao
Hui. V
i
su
al
C+
+
/
MAT
L
AB Image
Process
i
ng
and
Rec
o
g
n
itio
n. Bei
jin
g:
Peop
le P
o
st
Press. 2004: 9
4
-10
1
.
[4]
F
an W
e
iqi, M
u
Ch
ang
jia
ng.
A method
of licens
e pl
ate
character rec
ogn
ition
bas
ed
on Ch
in
es
e
character struc
t
ure.
Instrume
n
t
and Meter Jo
urna
l
. 200
3; 24
(4): 472-4
74.
[5]
Li W
e
nju,
Li
an
g D
equ
n. T
he
alg
o
rithm
of Q
ualit
y
de
gra
dat
ion
lic
ense
p
l
a
t
e char
acter s
egme
n
tatio
n
.
Co
mp
uter Aide
d Desi
gn a
nd
Graphics Jo
urn
a
l
. 200
4; 6(5): 697-
700.
[6]
W
ang N
i
an,
Xi
ong Y
u
a
n
, Z
hao H
a
ifen
g, R
en Bi
n.Veh
i
cle
Lice
nse P
l
ate
Reco
gniti
on
Automatica
ll
y
Based o
n
Ne
ur
al Net
w
o
r
k.
Anhui U
n
ivers
i
ty Journ
a
l
, 20
00; (3): 45-50. ch
ap
ter, 10.
[7]
Amir Sedig
h
i, Mansur Vafa
du
st. A ne
w
and r
obust
metho
d
for character se
gmentati
on an
d recog
n
itio
n
in lice
n
se p
l
ate
images.
Exper
t Systems w
i
th Appl
icatio
ns.
2
011; 28(
11): 13
497
–1
350
4.
[8]
Yin Z
h
a
o
q
i
ng,
Yin H
ao. Artific
i
al Inte
lli
ge
nce
and E
x
pert S
ystem. Beiji
ng:
Chin
a W
a
ter R
e
sourc
e
s a
n
d
H
y
dro
p
o
w
e
r
Press. 2002.
Evaluation Warning : The document was created with Spire.PDF for Python.