TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.1, Jan
uary 20
14
, pp. 584 ~ 5
9
1
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i1.3097
584
Re
cei
v
ed Ma
y 4, 2013; Re
vised July 2
3
, 2013; Accept
ed Augu
st 7, 2013
Rectification of License Plate Images Based on Hough
Transformation and Projection
Hongy
ao Deng*
1, 2
, Qingxin Zhu
1
, Jingsong Tao
3
, Hao Feng
4
1
School of Infor
m
ation & Soft
w
a
re Eng
i
n
eeri
n
g, Univer
s
i
t
y
of
Electronic Sci
ence a
nd T
e
chnol
og
y of Ch
in
a,
No. 200
6, Xi
yu
an Ave, W
e
st Hi-T
e
ch Z
one, Che
ngd
u, 611
731, Sich
ua
n Chin
a, Ph/F
ax:
+
86 28 83
201
864
2
School of Mat
hematics & Co
mputer
Scie
nc
e, Yangtze N
o
r
m
al Univ
ersit
y
,
No. 98 Julo
ng
Rd., Lidu F
u
l
i
n
g
District, Chon
g
q
in
g, 408
000,
Chin
a. Ph: +
86 23 072
90
08
8
3
School of Elec
trical Eng
i
ne
eri
ng, W
uhan U
n
i
v
ersit
y
, L
uoj
ias
han, W
u
cha
ng,
W
uhan, 43
007
2, Chin
a, Ph:
+
86 27 68
77
07
76
4
School of Co
mputer Scie
nc
e and En
gi
neer
ing, Guil
in
Un
iv
ersit
y
Of Electronic T
e
chnol
og
y (GUET
)
, No.1
Jinji R
o
a
d
, Guil
in, 541
00
4, Guang
xi Chi
na, P
h
/F
ax: +
86-7
7
3
5896
52
8
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: h
y
den
g_
200
4@1
63.com*
1,2
, qxzh
u@u
e
stc.edu.cn
1
,
jamso
n_tao
@1
63.com
3
, fengh
@gu
e
t.edu.cn
4
A
b
st
r
a
ct
It is crucial to
seg
m
e
n
t char
acters correctl
y
and
i
m
prov
e
rate of correc
t
character rec
ogn
iti
o
n
w
hen proc
essi
ng a
u
to
mo
bil
e
lice
n
se
plates
correctio
ns. In this
pap
er, tw
o algor
ith
m
s
are pr
op
ose
d
to
obtai
n the h
o
ri
z
o
n
t
al tilt a
nd
vertic
al sh
ear
ang
les. T
he transfor
m
ati
on
ma
trix for i
m
a
ges rectificati
o
n is
give
n an
d the
subp
ixel
issue
is solve
d
. So
me exp
e
ri
me
nts w
e
re don
e to test t
he al
gorith
m
s. Exper
i
m
en
t
a
l
results show
that the alg
o
rith
m is
rob
u
st, flexible
and
effective.
Ke
y
w
ords
: lic
ense p
l
ate; hor
i
z
o
n
ta
l tilt; vertical
she
a
r; subp
ixel issu
e; plat
e correctio
n
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
LPR (Li
c
e
n
se Plates Re
cognition
) is a
cru
c
ial elem
ent for imple
m
entation of the ITS
(Intelligent T
r
an
spo
r
tation
System). Generally
, a typical alg
o
rit
h
m for the LPR tech
niq
u
e
con
s
i
s
ts of th
ree m
a
jo
r pa
rts: licen
se pl
ate lo
cation,
l
i
cen
s
e plate cha
r
a
c
ter se
gmentation
a
n
d
licen
se
plate
cha
r
a
c
ter id
entification. T
he convent
io
nal ap
pro
a
ch
es to l
o
catin
g
lice
n
se pla
t
es
attempt to find the lice
n
se
plate by me
ans of li
ce
n
s
e plates i
m
a
ge features
such a
s
the
color,
sha
pe, symm
etry, geomet
ry or the
gra
d
i
ent [1, 2].
For
example,
in
our earli
er wo
rk,
li
cen
s
e pl
ate
regio
n
s were
extracte
d by
both the
dynamic RGB (whi
ch a
r
e th
e thre
e colo
r cha
nnel
s:
Red,
Gree
n
and
Blue) thre
sh
old
formul
a a
nd
the g
r
adie
n
t
of pixel inte
n
s
ities over th
e ho
rizontal
rows
[3]. The next task i
s
li
cen
s
e ch
ara
c
te
r segmentatio
n.
Ho
wever, the
licen
se
plate
s
extra
c
ted fro
m
image
s are not usu
a
lly rectan
gula
r
d
ue to per
spe
c
tive or othe
r deformation
s and extre
m
ely
disa
dvantag
e
ous for
seg
m
entation of
licen
se plat
e cha
r
a
c
ters.
Therefo
r
e, it is nece
s
sa
ry
to
rectify the lice
n
se pl
ate ima
ge
before ch
a
r
acte
rs are
se
gmented.
To corre
c
t license plate
s
, t
here
are two
para
m
et
ers t
hat play imp
o
r
tant roles.
O
ne is the
hori
z
ontal tilt; another is t
he vertical tilt, also
called
shea
r angle.
At present many corre
c
t
i
on
algorith
m
s h
a
ve been propo
sed. Literature [4] pro
posed differe
ntial proje
c
ti
on algo
rithm
to
corre
c
t the h
o
rizontal tilt (the
auth
o
r'
s representati
on). In fact, it is a hori
z
o
n
tal differenti
a
l
histog
ram. T
he p
r
in
cipal
method
sum
s
up
the di
ff
erent i
n
ten
s
ities of
all im
mediate
adja
c
ent
pixels when
an angl
e is o
b
tained by ro
tating t
he lice
n
se pl
ate ima
ge. The an
gl
e co
rre
sp
ondi
ng
clo
s
ely to the maximum is
the hori
z
ontal
tilt angle.
It is evident that
the algorithm
is con
s
id
era
b
ly
sen
s
itive to n
o
ise. [5,
6] chec
ke
d the
u
pper an
d lo
wer lin
es on
th
e licen
se pl
ate imag
e u
s
in
g
only Hou
gh t
r
an
sform
a
tion
. They con
s
i
dere
d
a thin
g justified
when the u
p
p
e
r line i
s
al
ways
parall
e
l
with t
he lo
we
r li
ne
witho
u
t resp
ect to
the
def
ormatio
n
of
the li
cen
s
e
pl
ate. The
r
e
is
an
error ma
de in
[5]. That is, that she
a
r correctio
n
wa
s u
s
ed to corre
c
t licen
se plate
region
s by the
angle
only ob
tained th
roug
h Ho
ugh
Tra
n
sformation.
Literatu
re [6]
use
d
the
Hou
gh tra
n
sfo
r
m
to
determi
ne th
e four corn
er coo
r
dinate
s
of a pl
ate region a
nd correcte
d the distorte
d ima
ges
throug
h a bili
near i
n
terp
ol
ation tran
sformation. For t
h
is alg
o
rithm,
it is very difficult to find the
actual vertex
coo
r
din
a
tes if
image qualit
y is not
very
high. The
r
e is anothe
r me
thod that can
be
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 584 – 5
9
1
585
see
n
in Litera
ture [7, 8]. Th
e straig
ht line
fitting method wa
s use
d
to corre
c
t the hori
z
ontal tilt of
licen
se
plate, and the h
o
ri
zontal proje
c
tion metho
d
was u
s
e
d
to co
rre
ct the verti
c
al tilt by findin
g
the minimum
variance. The dra
w
ba
ck of this appro
a
ch to lice
n
se plate co
rre
ction is that it is
sen
s
itive to noise. To overcome the i
s
sue, t
he author cho
s
e a suitabl
e de-n
o
isin
g arithm
etic
before
corre
c
tion. In literature
[9], a
rob
u
st al
gor
ith
m
wa
s
propo
se
d to
co
rre
ct li
cen
s
e
plate
s
.
It
obtaine
d
the corre
c
ted parameter
s tho
u
gh the ge
om
etric
config
uration bet
wee
n
two nei
ghb
oring
nume
r
ic
cha
r
acter bl
ob
s. Of course, there a
r
e
ma
ny other met
hod
s to cop
e
with the plate
corre
c
tion i
s
sue. Here no
more tha
n
menti
on of key
exempla
r
s of
these a
pproa
che
s
.
The focus
of this pap
er i
s
ho
w to obt
ai
n the two t
y
pe paramet
ers to
co
rrect license
plates. It i
s
o
r
gani
ze
d a
s
follows. T
he
next se
ction
pre
s
ent
s th
e
Hou
gh
Tra
n
sformation,
an
d
prop
oses a
n
algorith
m
by whi
c
h the horizontal co
rre
c
tion param
eter wa
s obtain
ed. In section
3,
a proje
c
tion
algorith
m
is
prop
osed to
get the
sh
e
a
r
pa
ram
e
ter. The corre
c
tio
n
tran
sformat
i
on
formula
is
de
scribe
d in
Section
4. Ad
ditionally, th
e
su
bfixels i
s
sue i
s
solved
in this secti
on.
Experimental
results an
d concl
u
si
on are
presented in
Section 5.
2. Obtaining
the Tilt Correction Parameter
Lice
nse plate
s
re
quire ho
ri
zont
al tilt co
rrection, which can
b
e
execu
t
ed by mean
s of the
hori
z
ontal
tilt angle. Majo
r approa
ch
e
s
check th
e e
d
g
e
s
of li
cen
s
e
plates. S
o
fa
r, there
a
r
e m
any
typical algo
rit
h
ms a
r
e
pro
p
o
se
d such a
s
Sobel Edge
Dete
ctor, Pre
w
itt Edge Detector,
Rob
e
rt
s
Edge
Dete
ctor, La
pla
c
ian
of Gau
s
sian
(Lo
G
)
Dete
ctor, Zero-Cro
ssi
ng
s Detector, and
Can
n
y
Edge
Dete
cto
r
[10-12]. The
s
e al
gorith
m
s som
e
ti
mes
cannot
che
c
k the ed
ge
co
rrectly due to
the
influen
ce of u
neven illumi
n
a
tion, noi
se a
nd def
o
r
mati
on.
The stan
dard Ho
ugh Tran
sfo
r
mati
on
is
a method to
che
c
k straigh
t
lines. This li
ne dete
c
tion
algorith
m
doe
s not limit performan
ce, ev
en
with noi
se an
d fragme
n
tation. In this pa
per, we ch
ecked the
uppe
r and lo
we
r li
nes o
n
lice
n
se
plates
by me
ans
of the Hough T
r
an
sfo
r
mati
on, a
n
d
found the
chara
c
te
r blob
s bet
ween t
w
o
lines.
The
h
o
rizontal tilt
angle
was o
b
tained
by t
he two lin
es and
ge
omet
ric configu
r
at
ion
feature
s
of ch
ara
c
ter bl
ob
s.
2.1. Hough T
r
ansformatio
n
The p
r
in
ciple
of Ho
ugh
Tra
n
sformation
can b
e
seen
from the lite
r
at
ure [1
3, 14].
Give a
set of p
o
ints i
n
an im
age
(t
ypically a
bin
a
ry im
ag
e),
with the Houg
h
Tran
sfo
r
mati
on, we
con
s
i
der
a point
)
,
(
i
i
y
x
and a
ll the lines th
a
t
pass thro
ug
h it. Infinitely
many line
s
p
a
ss thro
ugh
)
,
(
i
i
y
x
, all
of whi
c
h
satisfy the slo
pe-i
n
tercept e
qua
tion
b
ax
y
i
i
for some
value
s
of
a a
nd b.
Writin
g
this
equatio
n a
s
i
i
y
a
x
b
and co
nsi
deri
ng
the
ab
-pl
a
ne (al
s
o call
e
d
pa
ram
e
ter spa
c
e
)
yiel
ds
the
equatio
n of
a
sin
g
le
li
ne fo
r a
fixed p
a
ir
)
,
(
i
i
y
x
. Furthe
rmo
r
e
,
a second
p
o
int
)
,
(
j
j
y
x
also
ha
s a
line in pa
ra
meter
spa
c
e
asso
ciated
with it, and th
is line i
n
terse
c
ts the lin
e asso
ciated
wit
h
)
,
(
i
i
y
x
at
)
,
(
b
a
, where
a
is the slope a
n
d
b
the intercept of the line contai
ning
both
)
,
(
i
i
y
x
and
)
,
(
j
j
y
x
in the
xy
-pl
ane. In
fa
ct, all poi
nts
contai
ned
on
this lin
e h
a
ve line
s
in
pa
ramete
r
spa
c
e that int
e
rsect at
)
,
(
b
a
. Fig
u
re 1(a) illustra
tes these concepts.
In prin
ciple, t
he pa
ramete
r-sp
ace line
s
corre
s
p
ondin
g
to all imag
e point
s
)
,
(
i
i
y
x
can b
e
plotted, and
then im
age li
nes
co
uld
be
identifi
ed
by whe
r
e l
a
rge
numbe
rs of p
a
ram
e
ter-spa
c
e
lines inte
rsect. A practical
difficulty wi
th this app
ro
ach
,
howeve
r
, is
that
a
(the sl
ope of the lin
e)
approa
che
s
i
n
finity as the l
i
ne ap
pro
a
ch
es the ve
rtic
a
l
dire
ction. O
ne way aro
u
n
d
this difficulty is
to use the no
rmal rep
r
e
s
ent
ation of a line
:
sin
cos
y
x
(1)
Figure 1(b
)
ill
ustrate
s
the g
eometri
c interpretation of t
he pa
ramete
rs
and
, and it represents
the family of lines th
at p
a
ss through
a pa
rticula
r
point
)
,
(
i
i
y
x
. The intersec
tion point
)
,
(
corre
s
p
ond
s to the line that passe
s throu
gh both
)
,
(
i
i
y
x
and
)
,
(
j
j
y
x
.
T
o
de
te
c
t
the lin
es
in image
via
H
o
ug
h
T
r
a
n
s
f
or
ma
tio
n
,
th
e
-p
ara
m
eter
sp
ace i
s
divid
ed i
n
to
so-call
ed
accum
u
lator cell
s
, as illust
rat
ed in Figure 1(c), where
)
,
(
max
min
and
)
,
(
max
min
are
expecte
d ra
nge
s of the param
eter
values. Us
u
a
lly, the maximum rang
e of values is
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Re
ctification
of Licen
s
e Pl
ate Im
ages Base
d on Hou
gh Tra
n
sfo
r
m
a
tion and … (Hon
gyao Den
g
)
586
90
90
and
D
D
, wh
ere
D i
s
the
dista
n
ce
between
corne
r
s i
n
the
image. T
he
cell
at coo
r
di
nate
s
)
,
(
j
i
, with acc
u
mulator value
)
,
(
j
i
A
, co
rre
sp
on
ds to th
e sq
uare
a
s
soci
ated with
para
m
eter sp
ace
co
ordinat
es
)
,
(
j
i
. Initially, these cells are set to
zero.
Then, fo
r ev
ery no
n-
backg
rou
nd point
)
,
(
k
k
y
x
in the
image
plan
e, we
let
equ
al ea
ch
of th
e allo
we
d
su
bdivisio
n
values on
the
axi
s
a
n
d
solve for th
e
co
rrespon
ding
u
s
ing
th
e eq
ua
tio
n
sin
cos
k
k
y
x
.
The resulting
-val
ue
s
a
r
e
th
en roun
ded
o
ff to the ne
arest all
o
wed
cell value
alon
g the
-a
xis
.
The corre
s
p
o
nding accum
u
lator cell
i
s
t
hen
in
creme
n
t
ed. At the e
nd of thi
s
pro
c
ed
ure,
a val
u
e
of
Q
in
)
,
(
j
i
A
, means that
Q
points in the
xy
-plane
lie on the line
j
j
j
y
x
sin
cos
. The
para
m
eter
corre
s
p
ondin
g
to the
)
,
(
j
i
A
, in whi
c
h
the
value i
s
more
than
othe
rs,
is t
h
e
inclin
ation an
gle by whi
c
h
we can
corre
c
t the tilt of plate image
s.
Figure 1. (a)
slop
e-inte
rce
p
t equation; (b) no
rmal
rep
r
esentation; (c)
Dete
cting lines via the HT
2.2. Computi
ng Horizonta
l
Tilt Angle
The two
straight line
s
o
n
the licen
se plate re
gi
on we
re d
e
tected via th
e Hou
gh
Tran
sfo
r
mati
on Alg
o
rithm
.
Let the
av
erag
e va
l
ue
of their ho
ri
zontal tilt a
n
g
l
es
be
n
. W
e
sea
r
ched
the
cha
r
a
c
ter bl
obs
between
the line
s
. In
f
a
ct, they a
r
e
not ch
aracte
r blob
s b
u
t so
me
con
n
e
c
ted d
o
m
ain a
r
ea
s b
e
ca
use of the
fragme
n
t
ed,
overlap
p
ing
a
nd conn
ecte
d
cha
r
a
c
ters. I
n
our ea
rlie
r work [15], we
propo
se
d a
n
algorit
h
m
to solve the
issue, and o
b
tained n
o
rm
al
c
h
ar
ac
te
r b
l
ob
s
.
Le
t th
e
s
u
c
c
e
s
s
i
ve
c
h
ara
c
te
r b
l
ob
s be
}
,
,
{
1
n
s
s
S
, then the
minimum
co
n
t
ain
recta
ngle
of each cha
r
a
c
t
e
r blob
(M
CR
of cha
r
a
c
ter blob
)
ca
n b
e
lo
cated.
Suppo
se
that
the
hori
z
ontal di
stan
ce bet
ween the cen
t
er po
si
tions of two nei
ghbo
ring
cha
r
acte
r blo
b
s are
}
1
,
1
|
{
n
i
OD
OD
i
, and the vert
ical di
stan
ce
are
}
1
,
1
|
{
n
i
OH
OH
i
(
F
ig
ur
e
2)
. A s
e
t
of angle
s
is then given by
}
1
,
,
1
),
/
arctan(
|
{
n
i
OD
OH
i
i
i
i
(2)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 584 – 5
9
1
587
Figure 2. The
geometri
cs configurat
ion b
e
twee
n neig
h
borin
g ch
aracters.
The final ho
ri
zontal tilt angl
e is
n
i
i
n
1
*
1
(3)
Above all, the algorithm
ca
n be se
en in
Table 1.
3. Shear Ang
l
e Parameter
As we saw e
a
rlie
r, the deformatio
n
of Lice
n
s
e plate
not only has
the tilt on the
horizon
but also vert
ical shea
r. The focu
s of the se
ction i
s
on the sh
ear issu
e. Commonly, sh
ear
con
s
i
s
ts of X axis and Y axis dire
ction d
e
format
io
n. The deformati
on of the she
a
r mainly co
mes
from the
verti
c
al
dire
ction
after
the
tilt correctio
n
of li
cen
s
e
plate
s
.
The
othe
r d
e
f
ormation
ha
rdly
affects licen
se plates. The
r
efore, we
sha
ll not con
s
ide
r
them.
Assu
me that
the sh
ear
an
gle
ψ
is
different from –
φ
to
+
φ
,
ψ
∈
[–
φ
, +
φ
]. The she
a
r an
gle
is just the vertical
shear.
To
o
b
tain th
e
sh
ea
r a
ngle,
we
p
r
oje
c
t t
he M
C
R of
chara
c
te
r bl
ob
s in
some
directio
n. Let the M
C
R of
cha
r
a
c
te
r blo
b
s
be
}
,
1
|
{
n
i
s
s
i
N
the length
of projectio
n
be
i
S
L
(Figure 3), th
en we d
e
fine
a deci
s
io
n function a
s
}
{
)
(
)
(
N
S
Si
i
S
L
L
(4)
whe
r
e
ψ
∈
[-
φ
, +
φ
].
The functio
n
L
(
ψ
) can
con
v
ey the quality of shear an
gles, an
d the
minimum of the
L
(
ψ
) is th
e
optimal. To o
p
timize th
e a
l
gorithm, a fl
ag was
set. I
f
the proj
ecti
on is
overl
a
p
p
ing, the flag
is
false, otherwi
s
e it is t
r
ue. I
n
t
he iterative process of ca
lcul
ation, the cal
c
ul
ati
on i
n
its di
rection will
be aba
ndo
ne
d if the flag is false, beca
u
se it
is
not the be
st opti
m
al, and
will cho
o
se anot
her
angle to
p
r
o
c
ess. Let the
be the ite
r
ati
v
e step
si
ze,
whi
c
h
d
e
term
ines the a
c
cu
racy
of the
she
a
r an
gle. The algo
rithm
is described i
n
Table 2.
4. License Plate Rec
t
ifica
t
ion
4.1. Recti
f
ica
t
ion Trans
f
o
r
mation
The ho
rizont
al tilt
corre
ction paramete
r
θ
*
and she
a
r param
eter
ψ
*
were obtai
n
ed in the
third and fou
r
th sectio
ns. In this se
ction
,
t
he two parameters will
be us
ed to re
ctify the license
plate ima
ge.
The tran
sformation m
a
tri
x
of ho
rizo
nt
al tilt co
rrecti
on a
nd
sh
ea
r
corre
c
tion
were
defined respe
c
tively as
1
0
0
0
cos
sin
0
sin
cos
)
(
*
*
*
*
*
R
1
0
0
0
1
tan
0
0
1
)
(
*
*
SH
whe
r
e the
θ
*
and
ψ
*
are th
e hori
z
ontal ti
lt correctio
n
a
nd sh
ear p
a
rameters.
i
OD
i
OH
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Re
ctification
of Licen
s
e Pl
ate Im
ages Base
d on Hou
gh Tra
n
sfo
r
m
a
tion and … (Hon
gyao Den
g
)
588
Table 1. Obta
ining the ho
ri
zontal tilt angl
e
/*
A
l
gorit
hm
1: C
o
mp
uti
ng
the
til
t
correc
t
ion
par
a
meter*/
1.
Eliminate the points on the backgr
ound image, ima
ge binar
y
z
ation,
create the image
space
];
,
[
Y
X
I
2.
//check the line via HT algorithm.
3.
Define a pa
ra
meter space
]
,
[
Hu
, initi
a
l it w
i
th 0;
4.
for each point
I
y
x
I
i
)
,
(
{ //searching all points on the image
5.
for(
;
360
;
0
){
6.
Computin
g
sin
cos
y
x
;
7.
1
)
,
(
)
,
(
i
i
i
i
Hu
Hu
; // if hit a cell, then its accumulator increment b
y
1.
8.
}
9.
}
10.
}
max{
)
'
,
'
(
Hu
Hu
;
11. Computing
the
n
, and define ne
w
im
age space
]
,
[
'
Y
X
I
; //Th
e
region bet
ween
the t
w
o lines.
12.
//Searching the c
onnected domain
s
.
13.
]
,
[
'
Y
X
I
round its outline extended 1 p
i
xel w
i
dth
w
i
th 0;
//The purpose is to
unify
the processing method
w
i
th 8
-
unit-neibo
r
i
ng scheme.
14.
Label all the destination points
w
i
th no
tag and
d
e
fine a set of MC
R blobs,
}
,
,
,
{
]
[
2
2
1
1
y
x
y
x
n
S
;
0
k
;
15.
repeat: sea
r
ching the points on
'
I
16.
Find a desti
nation point,
)
,
(
y
x
p
i
;
17.
if(
)
,
(
y
x
p
i
has no tag){ /
/take the
point
p
as seed.
18.
)
,
,
,
(
]
[
y
y
x
x
s
k
S
;
)
,
(
)
,
(
y
x
p
y
x
p
i
; //Initiate the coordinates of MCR
19.
rep
eat
//Obtain a M
C
R
20.
Pus
h
the destination
points of 8-unit-n
e
iboring of
)
,
(
y
x
p
to stac
k;
21.
Pop
a point from the t
op of the stack{
22.
L
abel the point
)
,
(
i
i
y
x
p
w
i
t
h
tag;
23.
A
d
just the coordinates of the MCR
{
24.
if
)
(
1
x
x
i
then
i
x
x
1
; if
)
(
1
y
y
i
then
i
x
x
1
;
25.
if
)
(
2
x
x
i
then
i
x
x
2
; if
)
(
2
y
y
i
then
i
y
y
2
;
26.
}
27.
)
,
(
)
,
(
i
i
y
x
p
y
x
p
; goto ro
w
20;
28.
}
29.
until(th
e stack is nothin
g
);
30.
}
31.
until(all the points on
'
I
are searc
hed)
32.
Computing the
}
1
,
,
1
|
{
n
i
i
b
y
the M
CR blo
b
s;
33.
return
n
i
i
n
1
1
; //retu
r
n the ho
rizontal tilt angle,
Note: In a MCR,
)
,
,
,
(
2
2
1
1
y
x
y
x
S
,
)
,
,
,
(
2
2
1
1
y
x
y
x
notates th
e upp
er left and t
he lo
w
e
r rig
h
t coord
i
nates.
Figure 3. The
different values of the proj
ec
tion func
tion in different direc
t
ions
.
)
0
(
))
0
(
(
L
)
20
(
))
20
(
(
L
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 584 – 5
9
1
589
Table 2. Obta
ining the verti
c
al shea
r ang
le.
/*
A
l
gorit
hm
2: C
o
mp
uti
ng
the
v
e
rtical til
t
an
gle
correcti
on p
a
ra
meter*/
1.
Define the step size
∆ψ
; Let a pro
j
ection angle be
ψ
=-
φ
+
∆ψ
, candi
date optimal shear angle be
ψ
*
=0
, a projection
function of MCR
be L(
ψ
), can
d
idat
e shortest project
i
on function be L
*
(
ψ
*
);
2. repeat
3.
adjust projection direction,
ψ←
(
ψ
+
∆ψ
);
4.
Project each character MC
R {S
0
,…, S
n
} in direction
ψ
{
5.
if(projection
overlap)
6.
goto (3);
7.
else
8.
compute
r
L(
ψ
)
←Σ
si
∈
Sn
L(
S
i
);
9.
}
10.
if(L(
ψ
)< L
*
(
ψ
*
))
11.
L
*
(
ψ
*
)
←
L(
ψ
);
12. until(
ψ
>
φ
)
13. return
(
ψ
*
)
The final co
rrection tra
n
sfo
r
mation i
s
T
T
y
x
A
y
x
)
1
,
,
(
)
1
,
'
,
'
(
(5)
whe
r
e
A
=
R
(
θ
*
)*
SH
(
ψ
*
), (
x
,
y
) i
s
th
e p
o
i
n
t on th
e o
r
i
g
inal im
age
and
(
x
',
y
'
)
i
s
the p
o
int af
te
r
transfo
rmatio
n.
4.2. Decimal Fractio
n Iss
u
e
Obviou
sly, the pixel
coo
r
di
nates
(
x
',
y
'
)
obtaine
d by
equatio
n (5
)
may be d
e
ci
mal. We
call the
m
su
bpixel. Ho
we
ver, the
coo
r
dinate
s
of di
gital imag
es
must b
e
inte
ger val
u
e
s
. If the
pixel co
ordi
n
a
te (
x
',
y
'
)
is decim
al, it should
be retained
be
cau
s
e it is al
so t
he poi
nt on t
h
e
image. Thi
s
issue is a
ddre
s
sed a
s
follo
ws.
Assu
me th
at the coordinat
e's val
u
e
s
of
the poi
nt
)
'
,
'
(
y
x
p
are de
cimal. A
nd its
4-unit-
neigh
bori
ng
points of wh
ich
are
)
3
,
2
,
1
,
0
(
i
p
i
. The squa
re
wa
s divide
d int
o
4
re
ctangl
es
)
3
,
2
,
1
,
0
(
i
R
i
by the su
b
p
ixel point
p
(Figure 4). L
e
t the are
a
of the
i
R
is
)
3
,
2
,
1
,
0
(
i
A
i
respe
c
tively, then, the su
m of them is
1,
3
0
1
i
i
A
. Assum
e
t
hat the probability that the
p
hits
the sq
uare (n
ot containi
ng
the four
corn
er poi
nts) i
s
e
qual p
r
ob
abili
ty. We can e
a
sily obtain t
hat
the
i
i
A
R
p
)
(
and
3
0
)
(
i
i
R
p
=1. Let
the
)
(
f
be th
e inten
s
ity of
the pixel. Note that the im
age
is not the bl
ack-a
n
d
-
white image, but
a gray
ima
g
e
. Then the
decim
al fra
c
tion issue
ca
n
be
cop
ed with e
quation (6).
)
(
)
(
)
(
)
'
(
i
i
i
p
f
p
f
A
p
p
f
(6)
whe
r
e
)
(
i
p
f
notate the intensity before p
r
o
c
e
s
sing,
)
(
'
i
p
f
notate the inten
s
ity after pro
c
e
ssi
n
g
,
)
(
i
A
p
is the probabilit
y that
the point
p
hits the rectan
gle
i
R
.
Figure 4. The
geometri
cs c
onfiguration o
f
subpixel
0
A
1
A
2
A
3
A
2
p
0
p
1
p
3
p
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
Re
ctification
of Licen
s
e Pl
ate Im
ages Base
d on Hou
gh Tra
n
sfo
r
m
a
tion and … (Hon
gyao Den
g
)
590
5. Experiment and Concl
u
sion
To evaluate
the flexibility and effectivene
ss, mo
re
than 200 i
m
age
s, shot
rando
mly
unde
r vario
u
s
condition
s,
were teste
d
. The si
ze
of the imag
es
was 64
0 ×
48
0 pixels, an
d
th
e
format
was 2
4
-bit BMP o
r
JPG. Th
e license plate
s
e
x
traction we
re
finish
ed
i
n
our earli
er work
[3]. The next tasks a
r
e
chara
c
te
r seg
m
entation [1
5] and recog
n
ition. The t
e
ch
niqu
es in
this
pape
r p
r
ep
ro
ce
ss th
e ima
ge befo
r
e
ch
ara
c
ter
se
gm
entation, the
purpo
se i
s
to improve t
he
con
d
ition of
chara
c
te
r seg
m
entation. A
s
was
disc
u
s
sed
previo
usl
y
, it consi
s
ts
of the hori
z
o
n
tal
tilt and vertical she
a
r tra
n
sformation. Th
e typica
l expe
rimental resul
t
s we
re shown in Table 3.
As far
as th
e
compl
e
xity co
nce
r
ne
d, it is
O(
n
2
) for the
entire
algo
rith
ms, coming
p
r
imarily
from the
HT
and M
C
R a
l
gorithm
s. Ho
wever, th
e q
uantity of proce
s
sing
dat
a is
small
fo
r a
licen
se plate.
The algorith
m
was optimi
z
ed by cu
ttin
g
the hard da
ta set (See Algorithm 2
)
. The
runni
ng time
is not m
o
re
than o
ne milli
se
con
d
on th
e VC6.0
platform in th
e e
n
v
ironme
n
t of the
XP OS and
2M mem
o
ry.
Above all, th
e con
c
lu
si
on
can
be
drawn that
we
de
veloped
a
si
mple
and novel al
gorithm
to re
ctify
licen
se
plates, and
the al
gorith
m
is
ch
aracte
ristically
rob
u
s
t,
flexible, and effective.
Table 3. The
typical experi
m
ental co
rrection results
The license plate
extracted from
th
e
images.
The Hou
gh trans
formation
and the geom
etri
cs
configuration of
characters
The horizontal tilt
cor
r
e
ction
The vertical shear
cor
r
e
ction
1
2
3
4
5
6
Akno
w
l
e
d
ag
ement
This
wo
rk is partially
su
rpporte
d by V
i
si
ting S
c
hola
r
shi
p
of Stat
e Key Lab
oratory of
Powe
r Tran
smissio
n
Eq
uipment &
System
Security and
New T
e
chnol
ogy (Chon
g
q
ing
University) un
der g
r
ant No. 2007
DA10
51
2711
412.
Referen
ces
[1]
Christos
Nik
ol
aos E, A
n
a
g
nostop
o
u
l
os, I
oan
nis E.
A
Lice
nse P
l
ate-
Reco
gniti
on
A
l
gorit
hm for
Intelli
gent T
r
ansportati
on
S
y
stem A
p
p
lica
t
ions.
IEEE T
r
ansactions
on Inte
lligent Transportation
System
.
20
06; 7(3): 371-
39
1.
[2]
Guangr
on
g Ji, Hon
g
ji
e Yi.
T
he Reco
gniti
on
of the Vehicl
e
Lice
nse Plat
e Based o
n
Mod
u
lar N
e
tw
orks
.
Procee
din
g
s, Part I. Internationa
l S
y
m
pos
iu
m on Ne
ural N
e
t
w
o
r
ks. Dal
i
a
n
, Chin
a. 20
04
; 3173: 1
015
-
102
0.
[3]
Den
g
H
ong
ya
o
;
Song
Xiu
li.
A
nove
l
a
ppro
a
c
h
for l
i
cens
e
pl
ate l
o
catio
n
i
n
natura
l
i
m
ages
.
Applic
atio
n
s
of Digita
l
Information a
nd W
e
b T
e
chnolo
g
ies
.
Londo
n. 200
9
:
563-56
8.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 2302-4
046
TELKOM
NIKA
TELKOM
NIKA
Vol. 12, No
. 1, Janua
ry 2014: 584 – 5
9
1
591
[4]
W
u
Che
n
g
d
o
ng, F
an Y
u
q
uan.
Lice
nse
Plate
Char
a
c
ter Segme
n
t
a
tion
Base
d
on D
i
fferenc
in
g
Projecti
on a
nd
Preferab
l
y
S
e
g
m
ented C
har
a
c
ter.
Journal of
Northeaster
n
Univers
i
ty
(Nat
ural Sci
enc
e).
200
8; 29(7): 92
0-92
3.
[5]
Z
hou Ka
iju
n, Chen S
anb
ao. R
e
searc
h
of Veh
i
cle
L
i
cens
e Pl
ate Loc
ation
a
nd Ch
aracter S
egme
n
tatio
n
Und
e
r Comp
le
x Sce
nes.
Co
mputer Eng
i
n
eeri
n
g
. 200
7; 33(4)
: 198-20
0.
[6]
Che
n
Z
h
e
n
-
X
u
e
, W
ang Gu
o-
You. A N
o
vel
Algorit
hm of C
haracter S
egm
entat
io
n i
n
L
i
c
ense P
l
ate
s
Reco
gniti
on.
Microel
ectron
ics & Computer.
24(2): 42-44.
[7]
Kaush
i
k
D
eb, Andre
y
Vavi
lin.
Lice
nse Plate
T
ilt
Correcti
on Based
on
the
Straight
Lin
e
F
i
tting Met
h
o
d
and Pro
j
ecti
on
.
ICCCI 2010, P
a
rt III,
LNAI 6423. 201
0: 184-
193.
[8]
Kaush
i
k D
eb,
M
y
Ha
Le.
Autom
a
tic Vehicle Identi
fica
tion by
Plate
Reco
gniti
on
for Intelli
ge
nt
T
r
ansportati
on System
Ap
plic
ations
. IEA/AIE 2011, Part II,
LN
AI 6704. 20
11: 163-
172.
[9]
DJ Kan
g
. D
y
na
mic Programm
i
ng-b
a
sed M
e
th
od for E
x
tractio
n
of Lic
ens
e Pl
ate Num
bers o
f
Speed
in
g
Vehic
l
es on th
e Hig
h
w
a
y
.
Internati
ona
l Jour
nal of Auto
moti
ve T
e
chno
lo
gy.
2009; 1
0
(2): 2
05-2
10.
[10]
W
e
i Jia
ng, K
i
n-Man
Lam.
Efficient Ed
ge
Detecti
on
U
s
ing
Simp
lifie
d Gab
o
r Wa
velets.
IEEE
Transactio
n
s o
n
Systems, Ma
n and Cy
ber
ne
tics—Part B: C
y
bern
e
tics.
200
9; 39(4): 10
36-
104
7.
[11]
Bai Ho
ngl
ia
ng,
Liu Ch
ang
pi
n
g
.
A hybrid Lic
ense Pl
ate Extraction Meth
od
Based On Ed
ge Statistics
and Mor
p
h
o
lo
g
y
. Pattern Rec
ogn
ition, ICPR
2004. Proc
ee
din
g
s of the 17
th Internatio
nal
Confere
n
ce
.
200
4: 23-26.
[12]
Vahi
d Ab
olg
h
a
semi, Al
ireza
Ahmad
y
far
d
. An ed
ge-
bas
ed co
lor-a
id
ed
method f
o
r l
i
cens
e pl
at
e
detectio
n
Imag
e and Vis
i
o
n
C
o
mputi
ng. 20
0
9
; 27(8): 11
34-
114
2.
[13]
Chutata
pe O,
L
i
nfen
g Guo.
A
modifi
ed
Hou
g
h
transform f
o
r
line
d
e
tectio
n a
nd its
perform
a
n
ce.
Pattern
recog
n
itio
n
. 19
99; 32(2): 1
81-
192.
[14]
Kamat, Varsha.
An effici
ent i
m
p
l
e
m
e
n
tatio
n
of the
Ho
ug
h t
r
ansfor
m
for
de
tecting v
ehic
l
e
licens
e
plat
es
usin
g DSP'
S
. Real-T
ime T
e
chno
log
y
and A
pplic
atio
ns S
y
mposi
u
m.199
5
:
58-59.
[15]
Den
g
Hon
g
y
a
o
,
Song Xiul
i.
Li
cense Pl
ate C
haracters Se
g
m
e
n
tatio
n
Usin
g Projecti
on a
nd T
e
mpl
a
te
Matching
. Inter
natio
nal
C
onfe
r
ence
on
Infor
m
ation
T
e
chno
log
y
an
d
C
o
mp
uter
Sci
ence. w
u
h
an Chi
n
a
.
200
9; 1: 534-5
37.
Evaluation Warning : The document was created with Spire.PDF for Python.