TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 15, No. 2, August 201
5, pp. 346 ~
351
DOI: 10.115
9
1
/telkomni
ka.
v
15i2.837
6
346
Re
cei
v
ed Ma
y 7, 2015; Re
vised June
3
0
, 2015; Acce
pted Jul
y
14,
2015
Isolated
Handwritten Eastern Arabic Numerals
Recognition Using Support Vectors Machines
B. El Kessab
*, C. Daoui, B. Bouikh
ale
n
e, R. Salouan
Lab
orator
y of Informatio
n
Pro
c
essin
g
an
d Decisio
n
Sup
por
t, F
a
culty
of Scienc
e an
d T
e
chno
log
y
,
BP 523, Ben
i
M
e
lla
l, Morocco
*Corres
p
o
ndi
n
g
author, em
ail
:
bade1
0@
hot
mail.fr
Ab
stra
ct
In this pa
per,
w
e
present
a c
o
mparis
on
bet
w
een the
di
ffer
ent vari
atio
ns
of virtual r
e
tin
a
(grid s
i
z
e
)
in features extr
action w
i
th the
support vector
s mach
in
es cla
ssifier for isolat
ed ha
ndw
ritten
Eastern Arabi
c
nu
mer
a
ls reco
gniti
on. F
o
r this purpos
e w
e
have us
ed
for pre-pr
ocessi
ng
each nu
meral
imag
e the med
i
a
n
filter, the thres
hol
din
g
, nor
ma
li
z
a
tio
n
a
nd th
e
centeri
ng tech
niq
ues. F
u
rthe
rmor
e, the ex
p
e
re
me
nts resul
t
s
that w
e
hav
e
obtai
ne
d d
e
m
o
n
strate re
a
l
l
y
th
a
t
th
e
m
o
st
po
we
rfu
l
m
e
thod
i
s
th
a
t
vi
rtua
l re
ti
na
si
z
e
equ
al
20x2
0
. T
h
is w
o
rk has ach
i
ev
e
d
ap
proxi
m
atel
y 85% of
succ
ess rate for Easter
n Arabic
nu
mer
a
ls d
a
tab
a
s
e
identification.
Ke
y
w
ords
:
isolat
ed
ha
nd
w
r
itten easter
n
ar
abic
n
u
m
er
als,
medi
an fi
lter, thre
shol
din
g
, ce
nterin
g
,
nor
mal
i
z
a
ti
on, retina
l codi
ng
meth
od, an
d th
e supp
ort vectors mac
h
i
nes
Copy
right
©
2015 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
Optical
Cha
r
acter
Re
cog
n
i
tion (OCR) i
s
con
s
id
ered
re
ce
ntly as
a ve
ry dyn
a
m
ic field
given that its appli
c
a
b
ility in many different
doma
i
ns such as
postal
so
rtin
g, bank
che
que
pro
c
e
ssi
ng a
nd autom
atic data
entry, etc. Moreove
r
, the OC
R can be a
pplie
d on both
ca
se
s
printed o
r
ha
ndwritten. In fact reco
gni
tion for
han
dwritten
ca
se
is more co
mplex than that
printed
due t
o
varying
writ
ing style
s
fro
m
perso
n
to anothe
r even
so ju
st for o
ne given pfe
r
son
whi
c
h
will ma
ke thi
s
ki
nd o
f
recognition
very difficu
lt
whi
c
h requi
re
s for
re
solvin
g this p
r
o
b
le
m to
use
seve
ral
e
ffecient techn
i
que
s in e
a
ch
of the
thre
e
prin
cipal
pha
se
s formi
ng a
ce
rtain
syste
m
of reco
gnitio
n
whi
c
h are firstly the pre
-
proces
sin
g
then secondly
the features extraction th
en
finally learnin
g
and cl
assifi
cation o
r
quit
e
simply
re
cognition. In this fram
wo
rk, several stud
ies
has be
en
do
ne fo
r
recogn
ition of i
s
olat
ed h
and
wri
tte
n Ara
b
ic o
r
L
a
tin cha
r
a
c
ter or num
erals
by
usin
g in the f
eature
s
extra
c
tion p
h
a
s
e t
he retin
a
l cod
i
ng metho
d
in
one h
and
or
in the lea
r
nin
g
-
cla
ssifi
cation
pha
se the su
pport vecto
r
s machin
es [6
-8] on the ot
her ha
nd. He
nce, con
c
erni
ng
this app
roa
c
h
,
we are inte
rested to isola
t
ed hand
writt
en Easte
r
n Arabic n
u
me
ral
s
re
cog
n
ition.
Therefore, i
n
this sen
s
e a
n
d
in o
r
de
r to
achi
eve this t
a
sk
we h
a
ve
pre
-
p
r
o
c
e
s
sed ea
ch
nume
r
al im
a
ge by the
m
edian
filter, the thre
sholdi
ng, the
ce
ntering
an
d th
e no
rmai
zati
on
techni
que
s
while we extra
c
ted th
e feat
ure
s
of
ea
ch
nume
r
al
by
the retin
a
l
co
ding, ab
out t
h
e
recognitio
n
of each un
kn
o
w
n num
eral
we have u
s
e
d
the supp
ort
vectors ma
chine
s
. In fact, our
targeted p
u
rp
ose is b
e
ing
able to comp
are bet
wee
n
the pre
c
isi
o
n
of the differ
ent variation
s
of
grid si
ze
s in feature
s
extra
c
tion with the
sup
por
t vecto
r
s ma
chi
n
e
s
cla
ssifie
r
on the other
side
for
isolate
d
hand
written Ea
ste
r
n Arabi
c nu
meral
s
re
c
o
g
n
ition. Anyway, this pape
r is organi
ze
d in
the followi
ng
mann
er. First, in section
1 the p
r
op
ose
d
reco
gni
tion syste
m
i
s
sch
e
mati
ze
d,
Section
2 d
e
s
cribe
s
te
ch
nique
s fo
r i
m
age
pre-p
r
oce
s
sing. Se
ction
3 intro
duces the
re
tinal
codi
ng. In Se
ction 4, the
suppo
rt vecto
r
s ma
chin
es
classifier i
s
p
r
ese
n
ted. Se
ction 5 sho
w
s
the
experim
ental
results. Finall
y
, t
he study is ended by a concl
u
si
on.
2. Recog
n
ition Sy
stem
The re
cog
n
ition system th
at we have o
p
ted in
this study ispre
s
e
n
ted in the followin
g
figure:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Isolated
Han
d
written Ea
st
ern Ara
b
ic
Nu
m
e
rals
Reco
gnition Using
Suppo
rt… (B. El Kessab
)
347
Figure 1. The
propo
se
d re
cognition
syste
m
3. Pre-Proce
ssing
The first pha
se in e
a
ch O
CR
system t
he pre-pro
c
e
ssi
ng
who
s
e
the goal i
s
to
remove
each n
eedle
s
s pixel in
clu
d
ing
noise a
nd redu
nda
nt
inform
ation i
n
orde
r to
re
nder in
a b
e
st
quality the
nu
meral
imag
e
so th
at it
can
be
used
in
a
n
efficie
n
t ma
nner in
the fo
llowing
p
h
a
s
e
whic
h is
the features
extrac
tion. Of this
fac
t, to
achie
v
e this task, we have
pre
-
pro
c
e
s
sed in
this
resea
r
ch the image
s by the following te
chniqu
es:
1)
The medi
an filter applie
d for perfo
rmin
g a filtration of image.
2)
The th
re
shol
ding
used to
ren
d
e
r
ea
ch
image
cont
ains only the
bla
c
k and
white
colo
rs a
c
co
rd
ing a pre
-
sele
cted thresh
ol
d.
3)
The ce
nteri
n
g
exploited for locali
zing the
num
e
r
al ju
stly in center of
its image.
The no
rmali
z
ation with sta
ndar
d si
ze of each nume
r
al
image.
4. Featur
es
Extrac
tion
Features ext
r
actio
n
play
enormou
s
ly a ve
ry impo
rtant role i
n
each
OCR system,
esp
e
ci
ally for ha
nd
writte
n opti
c
al
ch
a
r
acte
r
re
co
g
n
ition, in fa
ct
the preci
s
io
n
of an
certain
system
re
co
g
n
ition d
epen
d
s
h
eavily to f
eature
s
ex
tra
c
tion
op
erati
on i
n
rea
s
o
n
of if a
n
g
r
e
a
t
discrmi
nation
betwee
n
cha
r
acte
rs is
trul
y realize
d
its recognitio
n
wi
ll be at that time very corre
c
t.
More pre
c
i
s
ely,
feature extraction m
e
thod
s
c
an
be divide
d into two p
r
in
cipal
cate
gories:
stru
ctural [10
-
17] an
d stati
s
tical [1
-5] feature
s
The fi
rst catego
ry is ba
sed o
n
l
o
cal
stru
cture
o
f
nume
r
al im
a
ge
while
the
se
con
d
i
s
int
e
re
sted to
st
atistical
information’s lo
ca
lized
in
cha
r
a
c
ter
image by wa
y of example within this co
ntext t
here are the momen
t
s of images
esp
e
ci
ally those
invariant
s.
In this frame
w
ork, we h
a
ve cho
s
e
n
a st
ructu
r
al meth
od whi
c
h i
s
Retinal co
ding
method.
4.1. Retin
a
l
Codi
ng
The process
of retinal co
di
ng that we
ha
ve used i
s
explaine
d as foll
ow:
Each ima
ge i
s
a bla
ck
co
n
t
aining a nu
m
e
ral
writting i
n
white colo
r and ha
s firstly an size
equal to
30x
30 pixel
s
. First of all, give
n a virtual
gri
d
or
retin
a
h
a
ving a
si
ze
equal to
2N/
3
x
2N/3 pixel
s
while this la
st
of each num
eral ima
ge i
s
equal to
NxN pixels, the
r
efore in o
r
d
e
r
to
applie
d this method a
s
it should the i
m
age mu
st
be re
sized to
2Nx2
N pixels, afterwards
the
retina i
s
pla
c
ed o
n
the fi
rst zone
of im
age th
e on
seco
nd
zon
e
and
so
on
u
n
til the la
st zone
while
at starti
ng from th
e top lo
cated to
the left
of the image in
ea
ch putting in
zone of the
ret
i
na
the numb
e
r
of white pixe
ls is
cal
c
ulat
ed whi
c
h
will
allow the
r
ea
fter ultimately to conve
r
t the
image to a vector.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 346 –
351
348
Moreve
r, in orde
r to well fix the ideas the schema b
e
l
o
w illust
rate
s this mech
ani
sm
(we
tak
e
N=
12):
Figure 2. The
process of re
tinal codi
ng
method
5. Recog
n
ition
An SVM [6-8] is con
s
ide
r
ed a
s
an st
atistica
l an
d sup
e
rvised
method it is basi
cally
defined fo
r t
w
o-cla
s
s p
r
o
b
lem
sep
a
rat
i
on, and
it
finds
an opti
m
al
hype
rpla
ne which ca
n
maximize the
margin b
e
tween the ne
arest exampl
e
s
of both classe
s, named
sup
port vect
ors
(SVs
).
First of all, given a training
databa
se of M data: X
i
, i=
1,2…..M.
Figure 3. The
determin
a
tio
n
of optimal hyperpla
ne, vectors suppo
r
t
s
, maximum
Marg
e and va
lid
hyperpl
ane
s
The linea
r SVM classifie
r
is then defined
as:
f(X,
w
,
b) : x
y
(1)
f
(
X
)
=
w
X
+
b
(2)
Whe
r
e w a
n
d
b are the pa
rameters of the cla
ssifie
r
y is the label.
The line
a
r SV
M can
be extende
d to a no
n-line
a
r
cla
ssi
fier by repl
aci
ng the inn
e
r p
r
odu
c
t
betwe
en the i
nput vectors
x
and the SVMs, throu
gh a
kern
el functi
on K defined
as:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Isolated
Han
d
written Ea
st
ern Ara
b
ic
Nu
m
e
rals
Reco
gnition Using
Suppo
rt… (B. El Kessab
)
349
Table 1. Examples of diffe
rent ke
rn
el function
s u
s
ed i
n
SVM
Kernel linear
xy
Kernel pol
y
nomi
a
l of degree n
Gaussian radial basis function
(GRBF) of a
standa
rd deviation
σ
:
‖
‖
2
The metho
d
descri
bed
ab
ove is de
sig
ned fo
r a
problem of two
classe
s o
n
l
y
, many
studie
s
treat
a gene
rali
zati
on of theSVM to a multi-cla
ssifi
cation [8] among the
s
e
studie
s
we cit
e
the two
strat
egie
s
fre
quen
tly used: the
first ap
pr
o
a
ch
isba
se
d to u
s
e
N de
ci
sio
n
functio
n
s (one
again
s
t all
)
al
lowin
g
to
ma
ke
a di
scrimi
nation
of
a
cl
ass
contai
ns
a on
e ve
ctor label
ed
by the
value 1 agai
nstall othe
r vectors existe
d in a ot
her
cla
ss o
ppo
sit
e
having a la
bel equal to
-1.
Therefore the deci
s
ion rule used in this
case is
usuall
y
the maximum such that wewill assign an
unkno
wn vect
or X into a cla
ss a
s
so
ciated
with an outp
u
t of SVM is the larg
est.
Class
e (X) =
arg
,,…,
(3)
6. Experiments and
Res
u
lts
First of
all, we
p
r
e
s
ent an example of some
Ea
ste
r
n
Arabi
c h
and
written n
u
mera
ls that
we have u
s
e
d
in our stu
d
y:
Figure 4. Example of som
e
isolated h
a
n
d
written Ea
st
ern Ara
b
ic n
u
m
eral
s
We have
cho
s
en the follo
wing data:
1)
Each o
r
igin
al nume
r
al imag
e has a
size equal to 30x3
0
pixels.
2)
The si
ze of th
e virtual retin
a
equal to 5x
5, 10x10, 15x15 and 2
0
x20
pixels.
3)
Each o
r
igin
al nume
r
al imag
e is re
sized to 60x60 pixel
s
.
4)
Each nu
mera
l is tran
sform
ed to a vector.
The sta
nda
rd
deviation of the GRBF
ke
rnel functio
n
is equal to 0.1.
No
w, we
gro
up the valu
es
of the re
co
g
n
ition rate
τ
g
(
giv
e
n in
%
)
for
ea
ch num
eral and
also tho
s
e of the global rat
e
re
co
g
n
ition i.e. of all numeral
s (
giv
e
n in
%
) which we have obtain
ed
in the followin
g
table:
Table 2. The
obtaine
d re
co
gnition rate
s
τ
n and
τ
g by e
a
ch meth
od o
f
extraction
Numera
ls
τ
n
(RC)
τ
n
(RC)
τ
n
(RC)
τ
n
(RC)
5 X 5
10 X 10
15 X 15
20 X 20
٠
83 87 89 90
١
70 79 92 97
٢
67 80 81 89
٣
66 67 68 70
٤
74 75 77 80
٥
77 78 81 93
٦
60 70 75 80
٧
80 81 84 88
٨
79 83 84 90
٩
60 62 65 69
τ
g
71,6 76,2
79,6 84,6
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 346 –
351
350
Table 3. The
obtaine
d re
co
gnition rate
s for all num
eral
s
Numera
ls
٠
١
٢
٣
٤
٥
٦
٧
٨
٩
٠
90.00
3.00 0.00
0.01 0.40 0.27 0.22 0.17
0.36 0.21
١
7.19
97.00
0.23
0.64 0.02 0.48 0.44 0.18
0.33 2.59
٢
0.19 0.00
89.00
23.30
9.00 0.65 5.00 3.41
1.20 0.93
٣
0.16 0.00 9.96
70.00
8.00 0.45 3.00 6.23
2.95 0.42
٤
0.22 0.00 0.00
1.70
80.00
0.85 3.30 0.28
0.59 0.83
٥
0.13 0.00 0.00
0.00 0.60
93.00
0.03 0.22
0.58 0.92
٦
0.29 0.00 0.81
3.00 0.40 0.43
80.00
0.00 1.78
23.45
٧
0.15 0.00 0.00
0.64 0.00 0.74 0.00
88.00
0.84 0.79
٨
0.82 0.00 0.00
0.26 0.08 0.34 0.70 1.12
90.00
0.86
٩
0.85 0.00 0.00
0.45 1.50 2.79 7.31 0.39
1.37
69.00
The graphi
cal
representatio
n to re
cogniti
on rate of ea
ch nume
r
al
τ
n
is
:
Figure 5. The
graphi
cal
rep
r
esentation of
reco
gnition
rate
τ
n
of each
method of extraction
The
gra
phi
ca
l re
pre
s
e
n
tation to
re
co
gni
tion rate of
a
ll nume
r
al
s
τ
g
is presente
d
in th
e
following figure:
Figure 6. The
graphi
cal
rep
r
esentat
ion of
global rate
re
cog
n
ition
τ
g
of each meth
od
of extraction
6.1.
Analy
s
is and Comment
Takin
g
into a
c
count all the
results that
we
obtai
ned,
we really ca
n
to con
c
lurethat:The
most
perfo
rm
ant meth
od i
s
the
retinal
co
ding
with
virtual retina
size e
qual
2
0
X 20
followed
byretinal
codi
ngwith virtu
a
l retina
size eq
ual 15
X 15 then
the retin
a
l codi
ng with
virtual retina
size
equal 1
0
X 10
then finaly the retinal
cod
i
ng withvirtual
retinasi
z
e e
q
ual 5
X 5
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Isolated
Han
d
written Ea
st
ern Ara
b
ic
Nu
m
e
rals
Reco
gnition Using
Suppo
rt… (B. El Kessab
)
351
7. Conclusio
n
In this p
ape
r,
we
have p
r
e
s
ente
d
a
co
mpari
s
o
n
bet
wee
n
the
performan
ce
s of
seve
ral
sizes of virtu
a
l retinawhich a
r
e th
e
su
pport
vecto
r
s ma
chine
s
u
s
ed
for re
co
g
n
ition of i
s
ol
a
t
ed
Eastern Ara
b
ic ha
nd
written nume
r
al
s.
In this
sen
s
e we have
verified that the recognit
i
on
system
s use
d
in
thi
s
app
roach wh
i
c
h
contain
s
in
the
preprocessin
g
ph
ase the
median
filter,
the
thresholdi
ng
and th
e
cent
ering
an
d th
e
su
ppo
rt ve
ct
ors ma
chi
ne i
n
th
e
recogn
ition ph
asere
a
lly
sho
w
s t
hat
t
h
e mo
st
p
o
w
e
r
f
ul re
cog
n
it
io
n sy
st
em i
s
t
hat contain
s
t
he
retinal
co
d
i
ng
with vi
rtu
a
l
retina
size eq
ual
20 X 20
.
Ackn
o
w
l
e
dg
ements
We
are
ve
ry grateful
to ou
r p
r
ofesso
rs M
i
ster
Ch
erki
Da
oui a
n
d
Miste
r
BelaidBoui
kh
alene f
o
r th
ei
r continu
o
u
s
encourage
me
nts, their g
r
e
a
t co
ope
ratio
n
, their
pertin
ent
advice
s
, their appro
p
ri
ate guida
nce in the reali
z
at
io
n
of this work. Many thanks
again to them
.
Referen
ces
[1]
Shahr
ul Niz
am
Yaako
b1, Put
eh Sa
ad.
Kraw
tchouk Mo
me
n
t
Invariant An
d
Gaussia
n
AR
T
M
AP Neural
Netw
ork: A Combi
natio
n T
e
chni
ques F
o
r
Imag
e Cl
ass
i
ficatio
n
. KUK
U
M Engi
ne
eri
ng R
e
searc
h
Semin
a
r. 200
6.
[2]
Hamid R
e
za
Boveiri. On
P
a
ttern Class
ific
ation Us
ing St
atistical Mome
nts.
Internatio
nal Jo
urna
l of
Sign
al Process
i
ng, Image Pro
c
essin
g
an
d Pattern Reco
gn
ition
. 20
10; 3(4):
2010.
[3]
Anass El affar
Khalid
Ferdous, Abdeljabbar Cher
kaoui Hakim El fadil, Hassan Qjidaa. Kraw
tchouk
Moment F
e
atu
r
e E
x
tractio
n
f
o
r Ne
ura
l
Ara
b
i
c Ha
nd
w
r
itten
W
o
rds R
e
co
g
n
itio
n.
IJCSNS International
Journ
a
l of Co
mputer Scie
nce
and N
e
tw
ork Security.
200
9; 9(1).
[4]
P Nag
a
b
hush
a
n
, SA Ang
a
d
i
, BS Anami. A f
u
zz
y statistical
ap
proac
h of
Kann
ad
a vo
w
e
l reco
gniti
o
n
base
d
on i
n
vari
ant moments. NCDA
R. Man
d
y
, Indi
a. 200
3: 275-
285.
[5]
Ibrahim A E
l
r
ube'
, Moh
a
me
d T
El Sonni,
Soha S S
a
l
eh.
Printed Ar
abic
sub-
w
o
rd rec
o
gniti
on us
in
g
moments.
W
o
rld Acade
my of
Scienc
e, Engi
n
eeri
ng an
d T
e
c
hno
logy
. 2
010;
42.
[6]
Srivastava DK,
Bhambh
u L. Data classificati
o
n
usin
g
Su
pp
or
t
Vector
M
a
chi
ne.
Inter
natio
n
a
l J
ourn
a
l
of
T
heoretic
al an
d Appl
ied Infor
m
ati
on T
e
ch
no
logy.
20
09; 12(
1): 1-7.
[7]
J Jind
al, A
Dhir
R. Ra
ni R.
Dia
gon
al fe
atures
and
SVM class
i
fier for
han
d
w
r
i
tten Gurum
u
k
h
i ch
aract
e
r
recog
n
itio
n.
Internatio
nal J
o
urna
l of Adva
nced R
e
se
arc
h
in C
o
mput
er Scienc
e a
nd Softw
ar
e
Engi
neer
in
g.
2012; 2(5): 5
05-
508.
[8]
DC Shu
b
h
ang
i
,
PS Hiremath.
Hand
w
r
itten E
nglis
h
char
acter and
dig
i
t recogn
ition
usin
g
multiclas
s
SVM cl
assifi
er a
nd
usi
ng
structural m
i
cro featur
es.
Int
e
rnati
ona
l J
o
u
r
nal
of R
e
ce
n
t
T
r
ends
i
n
Engi
neer
in
g.
2009; 2(2).
[9]
G Sinh
a, J k
u
mar. Ara
b
ic
numer
al R
e
co
gniti
on
Usin
g
SVM Class
ifie
r.
Internatio
nal
Journ
a
l
o
f
Emer
gi
ng Res
earch i
n
Mana
ge
me
nt & T
e
chnol
ogy
. 20
13; 2(5).
[10]
A Hen
n
ig, N S
herkat. Expl
oiti
ng zo
nin
g
bas
ed o
n
ap
pro
x
i
m
ating s
p
li
nes
in cursiv
e scri
p
t recog
n
itio
n
.
Pattern Reco
g
n
itio
n. 200
2; 35(2): 445-
45
4.
[11]
B El kessab,
C Dao
u
i, B Bouik
hal
en
e, M F
a
ki
r, K Moro. Extractio
n
Method of H
a
nd
w
r
itte
n Dig
it
Reco
gniti
on T
e
sted on the
MNIST
Database.
Internati
ona
l Journ
a
l
of Advance
d
Scienc
e a
n
d
T
e
chno
logy
. 2
013; 50.
[12] RS
H
ega
di.
Re
cogn
ition
of Pr
i
n
ted K
a
n
n
a
da
Nu
mera
ls
bas
e
d
o
n
Z
o
nin
g
M
e
thod
. I
n
ternati
ona
l Jo
urn
a
l
of Com
puter
Appl
icatio
ns (
097
5–
887
8) o
n
N
a
tion
al C
onfere
n
ce
of Advanc
ed Co
mputin
g
a
n
d
Commun
i
cati
o
n
s – NCAC
C
. 201
2.
[13]
A Hen
n
ig, N S
herkat. Expl
oiti
ng zo
nin
g
bas
ed o
n
ap
pro
x
i
m
ating s
p
li
nes
in cursiv
e scri
p
t recog
n
itio
n
.
Pattern Reco
g
n
itio
n.
200
2; 35(2): 445- 4
54.
[14]
D Imped
ovo,
G Pirlo. Z
oni
n
g
metho
d
s fo
r
han
d
w
ritten
character r
e
co
gniti
on: A sur
v
e
y
.
Pattern
Reco
gniti
on.
2
014; 47(
3): 969
-981.
[15]
M Kamel, A
Camp
ilh
o. A
n
Iris R
e
co
g
n
itio
n Meth
od
Base
d On
Z
i
gzag
Co
llar
e
tte Area
an
d
As
y
mmetric
al
Supp
ort Vector
Machin
es.
ICIAR
. 2007: 8
54-
865.
[16]
M Padma
a
, Y
Venkatar
ama
n
i
.
Z
IG-Z
A
G PVD – A
No
ntrad
i
tion
al A
ppro
a
c
h
.
Internati
o
n
a
l
Journ
a
l
of
Co
mp
uter Appl
icatio
ns.
201
0; 5(7).
[17]
K Ro
y, P Bh
attachar
ya. A
n
Iris Rec
o
g
n
i
tion M
e
tho
d
Based O
n
Z
i
gzag
Col
l
ar
ette Area
an
d
As
y
mmetric
al
Supp
ort Vector
Machin
es.
IEEE.
2006: 86
1-8
65.
Evaluation Warning : The document was created with Spire.PDF for Python.