Int
ern
at
i
onal
Journ
al of Inf
orm
at
ic
s
and
Co
m
munic
at
i
on
Tec
hn
olog
y (IJ
-
I
CT)
Vo
l.
7
,
No.
2
,
A
ugus
t
201
8
,
pp
.
96
~
104
IS
S
N:
22
52
-
8776
, DO
I: 10
.11
591/ijict
.
v7
i
2
.
pp96
-
104
96
Journ
al h
om
e
page
:
http:
//
ia
escore.c
om/j
ourn
als/i
ndex.
ph
p/IJI
C
T
Angula
r Symm
etric A
xis Const
ell
atio
n
M
odel fo
r
O
ff
-
lin
e
Odia
Han
dw
ritten
Charact
ers Re
cognit
ion
Py
ari Moh
an
Jena
1
, S
oumy
a Ranj
an
N
ayak
2
*
1
Depa
rt
m
ent of
Com
pute
r
Scie
n
ce
and
Engi
ne
ering,
C
ol
le
ge
of
E
ngine
er
ing
an
d
Te
chno
log
y
,
Bh
ubane
sw
ar,
Indi
a
2
Depa
rt
m
ent of
Com
pute
r
Scie
n
ce
and
Engi
ne
ering
,
K
L
Univ
ersi
t
y
,
Andhra
Prade
sh,
India
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
May
03, 201
8
Re
vised Ju
n 2
6, 20
18
Accepte
d
J
u
l
10, 2
018
Optic
a
l
cha
r
ac
t
e
r
rec
ognition
is
one
of
the
emer
ging
rese
ar
ch
to
pic
s
in
the
fie
ld
of
image
p
roc
essing,
and
it
has
ext
ensive
a
rea
of
applic
at
io
n
in
pat
t
ern
rec
ogni
ti
on.
Od
ia
handwri
t
te
n
script
is
the
m
ost
rese
arc
h
co
nce
rn
ar
ea
bec
ause
it
h
as
eldest
and
m
ost
likable
la
nguag
e
i
n
the
sta
te
of
od
isha,
Indi
a
.
Odia
cha
r
acte
r
i
s
a
usually
han
dwritt
en
,
which
was
gene
ra
lly
occ
upi
ed
b
y
sca
nner
int
o
m
ac
hin
e
rea
d
abl
e
form
.
In
thi
s
reg
ard
seve
ra
l
rec
ognition
te
chn
ique
hav
e
bee
n
evo
lve
d
f
or
var
ia
n
ce
kin
d
of
la
nguag
es
but
writi
n
g
pat
t
ern
of
od
ia
cha
ra
cter
is
just
li
ke
as
cur
ve
a
ppea
ran
ce;
Henc
e
it
is
m
or
e
diffi
cu
lt
for
re
co
gnit
ion.
In
thi
s
a
rti
cle
we
hav
e
pr
ese
nte
d
the
nove
l
appr
o
ac
h
for
Odia
cha
ra
cter
rec
ogni
ti
on
b
ase
d
on
the
diff
ere
nt
ang
le
base
d
sy
m
m
et
ric
axi
s
feature
ex
tr
ac
t
ion
technique
which
give
s
hi
gh
ac
cur
acy
of
rec
ogni
ti
on
pat
t
ern
.
Thi
s
e
m
piri
ca
l
m
odel
gene
r
at
es
a
u
nique
ang
le
b
ase
d
bound
a
r
y
point
s
on
ever
y
skel
et
on
ised
cha
ra
ct
e
r
images.
Th
ese
point
s
are
int
er
conne
c
te
d
with
each
othe
r
i
n
orde
r
to
ex
tr
act
row
and
col
um
n
s
y
m
m
et
r
y
axi
s.
W
e
ex
tract
ed
feature
m
atri
x
havi
ng
m
ea
n
dista
nc
e
of
row,
m
ea
n
angle
of
row,
m
ea
n
dista
nce
of
co
lumn
and
m
ea
n
angl
e
of
col
um
n
from
ce
ntre
o
f
the
imag
e
to
m
idpoi
nt
of
the
s
y
m
m
et
ric
axi
s
r
espe
ctively
.
The
s
y
stem
uses
a
10
fol
d
val
idatio
n
to
the
ran
dom
fore
st
(RF)
cl
assifie
r
and
SV
M
for
fea
tur
e
m
at
rix.
W
e
h
av
e
consid
ere
d
th
e
stand
ard
da
tabase
on
200
images
hav
ing
ea
ch
of
47
Odia
cha
ra
ct
er
and
10
Odia
num
eri
c
for
sim
ula
ti
on.
As
we
have
note
d
out
come
of
sim
ula
ti
on
o
f
SV
M
and
RF
y
ie
lds
96
.
3%
and
98.
2%
ac
cur
acy
r
ate
o
n
NIT
Rourkela
Odia
cha
r
ac
t
e
r
dat
ab
ase
and
88.
9%
and
93.
6%
from
ISI
Kolkat
a
Odia
nu
m
eri
ca
l
da
ta
base
.
Ke
yw
or
d
s
:
Op
ti
cal
char
a
ct
er
recogn
it
io
n
featur
e
v
e
ct
or
Patt
ern
recog
ni
ti
on
Ra
ndom
f
or
est
SV
M
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Soum
ya
Ranj
an Nay
ak
,
Dep
a
rt
m
ent o
f
Com
pu
te
r
Scie
nce a
nd
E
ng
i
ne
erin
g
,
K
L
Unive
rsity
,
Gr
ee
nf
ie
ld
s, V
add
e
swa
ram
,
Guntur,
An
dhr
a Pr
a
des
h,
I
nd
i
a 5225
02
.
Em
a
il
:
nayak.s
ou
m
ya
@
kluniversity
.in
1.
INTROD
U
CTION
In
t
he
era
of
di
gital
i
m
age
processin
g,
t
he
c
har
act
er
rec
ogniti
on
is
on
e
of
the
sig
nifica
nt
and
us
ef
ul
e
m
erg
in
g
re
se
arch
t
opic
s
is
the
area
of
patte
rn
recog
niti
on
.
T
he
m
ai
n
inten
d
of
c
ha
ra
ct
er
rec
ogniti
on
is
to
translat
e
hum
a
n
rea
da
ble
cha
racter
to
m
achine
rea
dab
le
c
od
e
s
o
that
m
a
chine
ca
n
ef
fici
ently
reco
gniz
e
the
char
act
e
r.
T
here
are
m
ai
nly
t
wo
broa
d
cat
e
gory
of
cha
ract
er
rec
ogniti
on
syst
e
m
are
fo
und
s
uc
h
as
offli
ne
an
d
on
li
ne
recog
niti
on
pro
cess.
I
n
case
of
on
li
ne
char
act
er
rec
ogniti
on
proces
s
,
it
rep
rese
nts
the
two
dim
ens
ion
al
co
-
ordinates
of
su
cce
ssive
points
of
the
ha
ndw
riti
ng
as
a
functi
on
of
tim
e
are
stored
in
par
ti
cula
r
order
descr
i
bed
by
[
1].where
a
s
in
case
of
the
offli
ne
ha
ndw
riti
ng,
only
the
c
om
plete
d
wr
it
ing
is
a
vaila
ble
as
a
n
i
m
age
descr
i
be
by
[
2].
In
t
his
pa
pe
r,
our
researc
h
i
nte
nd
co
nf
i
ne
d
with
offli
ne
ha
ndw
ritt
en
ch
aracte
r
recog
niti
on
.
O
ur
recog
niti
on
sta
ge
com
pr
ise
s
of
t
hr
ee
broa
d
sta
ges
i
nclu
di
ng
ac
qu
isi
ti
on
,
featu
re
ext
ra
ct
ion
and
cl
as
sific
at
io
n
ste
p.
Be
si
de
that
a
rec
ogniti
on
syst
em
m
os
tl
y
dep
ends
up
on
a
well
-
de
fine
d
f
eat
ur
e
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
ICT
IS
S
N:
22
52
-
8776
An
gu
l
ar
Symm
et
ric
Axis Co
nst
el
lati
on
M
od
el
for Off
-
li
ne.
.
.
(
Soumya
R
an
j
an N
aya
k
)
97
extracti
on
pro
cedure
al
ong
with
a
good
c
la
ssifie
r,
i
n
order
to
ac
hieve
high
su
cces
s
rate
[
3].
In
or
der
to
achieve
a
good
rec
ogniti
on
syst
e
m
fo
r
handwrit
te
n
form
a
t
is
qu
it
e
still
chall
eng
i
ng
be
cause
of
va
riat
ion
i
n
wr
it
in
g
sk
il
ls,
sh
a
pes
an
d
or
i
entat
ion
.
Va
rio
us
ap
proac
hes
are
f
ollow
e
d
up
by
diff
e
re
nt
researc
her
t
o
va
rio
us
scripts
li
ke
Arabic,
Chinese
,
and
En
glish
et
c
are
rep
ort
e
d
[4
]
.
Ba
sic
al
ly
od
ia
script
la
ngua
ge
is
one
th
e
la
nguag
e
w
hic
h
is
de
rive
d
f
r
om
de
vangiri
s
cripts.
It
is
on
e
of
t
he
re
giona
l
la
ng
ua
ges
of
India,
m
os
tl
y
s
poke
n
at
east
ern
pa
rt
(Odisha
)
an
d
s
om
e
so
uth
,
nor
th
par
t
of
Indi
a.
To
ac
hieve
a
good
accu
ra
cy
of
rec
ogniti
on
for
hand
wr
it
te
n
ch
aracte
rs
of
o
dia
char
a
ct
er
is
qu
it
e
i
m
pr
essi
ve.
T
hough
a
good
num
ber
of
w
orks
has
done
for
Indian
reg
i
on
a
l
la
ng
ua
ges
but
a
le
ss
in
num
ber
relat
ed
to
O
d
ia
script.
In
the
se
past
r
ecent
ye
ars
dif
fer
e
nt
auth
or
s
m
ake
an
at
te
m
pt
fo
r
analy
sis
with
r
espect
to
O
dia
scripts
a
re
repor
te
d
in
[
5].
T
he
feat
ur
e
ext
r
act
ion
te
chn
iq
ue
f
or
recog
niti
on
of
ha
ndw
ritt
en
c
har
act
er
is
a
chall
eng
i
ng
ta
sk
i
n
the
rese
arch
fiel
d
of
patte
r
n
recog
niti
on
.
I
n
this
reg
ar
d
a
la
rg
e
nu
m
ber
of
featu
re
extra
ct
ion
te
chn
i
qu
e
and
cl
assifi
c
at
ion
al
gorith
m
hav
e
been
pr
e
se
nte
d
in
recent
ye
ar
descr
i
bed
by
[
6].
Seve
ral
cha
racter
rec
ogniti
on
te
c
hn
i
qu
e
of
dif
fer
e
nt
la
ng
uag
e
is
found
in
m
any
li
te
ratur
es
[7
-
9].
In
li
ne
to
cha
racter
re
cogniti
on
t
he
extensi
ve
sur
ve
y
has
bee
n
re
ported
base
d
on
diff
e
ren
t
kinds
of
f
eat
ur
e
extracti
on
t
ech
nique
[
10
]
.
I
n
this
sur
vey
pap
e
r,
aut
hor
re
ported
di
ff
e
ren
t
featur
e
e
xtract
ion
te
c
hniq
ue
app
li
ed
on
te
m
pla
te
m
at
chi
ng,
pro
j
ect
ion
histo
gram
s,
def
orm
able
templa
te
s
,
con
t
our
pro
file
s,
un
it
ary
im
age
trans
form
s,
zon
i
ng,
gr
a
ph
descr
i
ption,
zern
ike
m
ome
nts,
s
pline
cur
ve
appr
ox
im
at
ion
and
f
ourier
de
scripto
rs
ha
s
been
a
pp
li
e
d
on
gray
le
vel
char
act
er
,
bi
nary
char
act
er,
c
ha
racter
con
t
our
,
cha
ra
ct
er s
kelet
on
a
nd
c
har
act
er
graph
im
age r
eprese
ntati
on
fo
r
m
in
the p
re p
r
ocessin
g
ste
ps.
A
s the
Indian
la
ng
uage
is
con
cer
ne
d,
the
op
ti
cal
ch
aracte
r
rec
ogni
ti
on
play
s
a
vital
ro
le
now
da
ys.
In
this
pa
pe
r
we
hav
e
m
ade
an
at
tem
pt
to
desig
n
a
nove
l
app
r
oac
h
that
eff
ic
ie
ntly
recogn
iz
e
th
e
od
ia
cha
rac
te
r
by
i
m
ple
m
enting
angular
m
easur
em
ent
and
e
uc
li
dian
distanc
e
by
ta
kin
g
the
m
idp
oin
t
fro
m
the
axis,
whic
h
was
gen
e
rated
by
ta
king
the
m
idpoint
of
tw
o
bounda
ry
ed
ge
of
row
sym
m
et
ri
c
axis
as
well
as
colum
n
sy
m
m
et
ric
axis
to
the
ce
nt
re
of
t
he
im
ages.
O
dish
a
sta
te
,
so
far
has
be
en
able
t
o
up
ho
l
d
the
pr
i
de
of
ha
ving
the
l
arg
est
nu
m
ber
of
p
al
m
le
af
m
anu
scripts
(
ov
e
r
20,
000
m
anu
scri
pt
s)
in
the
w
or
l
d.
[
11]
.
Mi
ll
ion
books
would
hav
e
been
pr
i
nted
f
r
om
start
ing
wh
ere “N
e
w
Test
a
m
ent” that got
p
rinted
in
1809
was
first pu
blishe
d.
[
12]
. Odia g
ot
cl
assic
al
sta
tus
excep
t
5
ot
her
Indian
la
ng
ua
ges
on
the
basi
s
of
it
s
li
te
rar
y
her
it
age
fo
ll
owin
g
ap
pro
val
of
th
e
Un
i
on cabi
net.
2.
RELATE
D
BACKG
ROUN
D
W
ORK
Od
ia
sc
ript
ha
s
bee
n
e
xtracte
d
us
in
g
B
hr
am
i
scripts
a
nd
one
of
t
he
m
os
t
ancient
la
ngua
ges
am
ong
Indian
re
giona
l
la
ng
ua
ge
m
os
t
sp
oken
east
ern
par
t
of
Ind
ia
ba
sic
al
ly
in
sta
te
Od
isha,
West
-
Be
ng
al
,
Guja
rat
et
c.
The
m
os
t
i
m
po
rtant
sce
na
rio
of
this
la
ngua
ge
that
it
has
no
l
ow
e
r
and
uppe
r
cas
e
form
at
.
Her
e
in
the
script
is
has
no
uppe
r
case
lowe
r
struct
ure.
A
certai
n
well
-
de
fine
d
appr
oach
es
a
re
adopted
by
di
ff
ere
nt
researc
hers
to
achieve
hi
gh
r
ecognit
ion
rate.
Re
cogniti
on
is
the
pr
oce
ss
of
acce
ptin
g
the
unknown
sam
ples
of
hand
wr
it
te
n
c
har
act
er
im
age
or
w
ords
a
nd
then
pr
oceeds
into
a
patte
r
n
rec
ogniti
on
pro
blem
fo
r
te
sti
ng
.
Re
cogniti
on
pr
ocess
ca
n
be
a
chieve
d
ei
the
r
i
n
three
im
po
rtant
way,
w
hich
is
desc
ribe
d
a
s
tem
plate
m
atch
in
g,
sta
ti
sti
cal
te
ch
nique
a
nd
ne
ur
al
netw
ork
te
c
hn
i
qu
e
s.
T
hese
cha
racter
recogn
it
io
n
a
ppr
oa
ches
us
es
ei
the
r
to
p
dow
n
ap
proac
hes
or
analy
ti
cal
strat
egies
for
recog
niti
on
.
Tem
plate
m
a
tch
in
g
is
the
s
i
m
plest
form
of
t
rainin
g
and
rec
ogniti
on.
He
re
is
the
i
dea
is
to
m
at
c
h
the
store
d
pr
edef
i
ned
proto
ty
pe
with
the
unknow
n
ha
ndwr
it
te
n
char
act
e
rs.
I
n
this
m
a
tc
hin
g
t
echn
i
qu
e
only
sel
ect
ed
pix
el
are
com
par
ed
with
data
sam
ples
an
d
r
uled
base
d
decisi
on
tree
a
nal
ysi
s.
Rule
ba
sed
decisi
on
t
echn
i
qu
e
w
ere
us
e
d
by
c
haudhu
ri
et
.
at
i
n
2002
[13].
Stat
ist
ic
al
te
chn
iq
ue
co
nsi
der
e
d
as
m
or
e
eff
ect
ive
w
hile
recogn
it
io
n
of
O
dia
cha
rac
te
rs.
In
this
re
gard
obai
dull
ah
et
al
[14]
in
2014
us
es
t
he
li
nea
r
log
ist
ic
re
gre
ssion
m
od
el
by
us
in
g
higher
orde
r
sta
ti
sti
cal
decisi
on
m
od
el
t
o
pro
vid
e
bette
r
perform
ance
rather
tha
n
the
li
near
m
od
el
in
perform
ance.
I
n
20
07
pal
et
al
[1
5]
use
d
qu
a
dr
at
ic
functi
on
f
or
cl
assifi
cat
ion
is
base
d
on
Ba
ye
sia
n
est
im
a
ti
on
.
I
n
20
09
a
nd
2005
a
sim
i
la
r
te
chn
i
ques
of
ps
e
udo
Ba
ye
sia
n
est
i
m
at
ion
te
ch
niqu
e
was
ad
opte
d
by
wa
xab
ya
s
hi
et
al
[1
5]
,
an
d
r
oy
at
al
[1
6]
for
odia
ha
nd
wr
it
te
n
nu
m
erical
reco
gnit
ion.
They
us
ed
c
onve
ntion
al
quad
rati
c
discrim
inant
functi
on.
In
2006
Hidde
n
Ma
rko
v
Mod
el
(
HMM)
was
purpose
d
by
Bho
wm
ik
et
al
[1
7].
This
is
us
ed
non
ho
m
og
e
neous
qu
a
drat
ic
m
et
h
od
f
or
trai
ning
an
d
recog
niti
on
of
ha
ndwr
it
te
n
nu
m
erical
.
In
2014
Dash
et
al
[18
]
-
[
19]
hav
e
a
dopt
ed
a
Discrim
inati
ve
Lear
ning
Ba
s
ed
Q
ua
dr
at
ic
Discrim
inant
Cl
assifi
er
(
DL
QDF)
a
nd
N
on
-
redu
nd
a
nt
St
oc
kwel
l
trans
form
b
ased f
eat
ur
e
ex
tra
ct
ion
fo
r
ha
nd
wr
it
te
n digit
r
e
cogniti
on. Neu
ral n
et
w
ork
is
the
par
al
le
l p
rocessi
ng
m
et
ho
d
ha
ving
interco
nnect
io
n
of
neurons
i
ns
ide
t
his
te
ch
nique.
It
pe
rfo
rm
co
m
pu
ta
ti
on
at
hi
gh
e
r
s
pe
ed
in
com
par
ison
with
sta
ti
sti
cal
and
tem
plate
m
a
t
chin
g.
Ne
ural
netw
ork
can
be
per
f
orm
ed
either
in
tw
o
wa
ys
li
ke
feed
f
orward
ne
twork
(FFN
A
)
a
nd
bac
k
pro
pag
at
io
n
net
w
ork
(BPNN
).
I
n
2013
m
ishra
et
al
[20]
pe
rfo
rm
the
cl
assifi
cat
ion
with
BPN
N
an
d
got
a
hig
h
ac
cur
acy
of
90.
44
pe
rcen
ta
ge.
I
n
2011
Ma
jhi
et
al
[2
1]
auth
or
s
have
pro
po
se
d
a
no
nlinear
ne
ural
netw
ork
cl
assi
fier
it
is
a
n
a
na
log
y
of
f
un
ct
ion
al
li
nk
a
rtif
ic
ia
l
neu
ral
ne
twork
(F
L
ANN)
cl
as
sifie
r.
In
2012
C
handa
et
al
[22]
propos
e
a
m
et
ho
d
f
or
w
rite
r
ide
nt
ific
at
ion
f
ro
m
O
dia
hand
wr
it
in
gs
wh
ic
h
us
es
the
SV
M
for
cl
as
sific
at
ion
.
I
n
2015
kaly
an
et
al
[23]
pur
pos
ed
BES
AC
sy
m
m
e
tric
axis
const
el
la
tio
n
m
od
el
us
in
g
cl
assifi
er
SVM
,
n
earest
nei
ghbour
an
d
r
a
ndom
fo
rest
ha
ving
accu
racy
98.90,
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
IJ
-
ICT
Vo
l.
7
, No
.
2
,
A
ugus
t
201
8
:
96
–
104
98
99.48,
96.
76
per
ce
ntage
res
pectf
ully
.
The
detai
ls
com
par
iso
n
of
r
eco
gn
it
io
n
acc
ur
a
ci
es
are
de
scri
bed
in
belo
w
T
a
ble.
1.
Table
1.
List
of
All
Feat
ur
e
s i
n
Re
c
ogniti
on
of Odia
Cha
rac
te
rs
with
A
cc
uracy
Metho
d
Featu
res
Clas
sif
ier
Databas
e
Accurac
y
(
%)
Pal et
al
.
(20
0
7
a)
Gradien
t
+
cu
rvatu
re
MQDF
Od
ia
b
asic
ch
ara
cters
o
n
II
TBBS
9
4
.60
W
ak
ab
ay
ash
i et
al
.
(
2
0
0
9
)
W
eig
h
ted
grad
ien
t
MQDF
Od
ia
b
asic
ch
ara
cters
o
n
II
TBBS
9
5
.14
Ro
y
et
al
.
(20
0
5
)
Direction
al
Qu
ad
ratic
Od
ia nu
m
erals
on
I
IT
BB
S
9
4
.12
Bh
o
w
m
ik
e
t al.
(
2
0
0
6
)
Scalar
HMM
Od
ia
n
u
m
e
rals
o
n
ISI
Ko
lk
ata
9
0
.50
Bh
o
w
m
ik
e
t al.
(
2
0
0
6
)
Scalar
HMM
Od
ia nu
m
erals
on
I
IT
BB
S
9
1
.26
Mish
ra
et
al.
(20
1
3
)
DCT
BP NN
Od
ia
n
u
m
e
rals
o
n
ISI
Ko
lk
ata
9
2
.00
Mish
ra
et
al.
(20
1
3
)
DCT
BP NN
Od
ia nu
m
erals
on
I
IT
BB
S
9
0
.44
Dash
et
al.
(
2
0
1
4
b
)
Hy
b
rid to
p
o
lo
g
y
DLQDF
Od
ia
n
u
m
e
rals
o
n
ISI
Ko
lk
ata
9
8
.50
Dash
et
al.
(
2
0
1
4
a)
Sto
ck
well
trans
f
o
r
m
k
-
Near
est
n
eig
h
b
o
r
Od
ia
n
u
m
e
rals
o
n
ISI
Ko
lk
ata
9
8
.80
Dash
et
al.
(
2
0
1
4
b
)
Hy
b
rid to
p
o
lo
g
y
DLQDF
Od
ia nu
m
erals
on
I
IT
BB
S
9
8
.28
Dash
et
al.
(
2
0
1
4
a)
Sto
ck
well
trans
f
o
r
m
k
-
NN
Od
ia nu
m
erals
on
II
TBBS
9
9
.00
Kal
y
an
S
Dash
et
al.
(20
1
6
)
BESAC
Ran
d
o
m
f
o
rest
SVM
Near
est n
eig
h
b
o
r
Od
ia
n
u
m
e
rals
o
n
ISI
Ko
lk
ata
9
8
.44
9
9
.02
9
9
.35
Kaly
an
S
Dash
et
a
l.
(20
1
6
)
BESAC
Ran
d
o
m
f
o
rest
SVM
Near
est n
eig
h
b
o
r
Od
ia nu
m
erals
on
I
IT
BB
S
9
7
.30
9
8
.56
9
8
.90
3.
PROP
OSE
D HA
NDWRIT
TE
N
C
HAR
A
CTER
REC
O
GNI
TI
ON SY
STE
M
In
this
sect
io
n,
we
ha
ve
m
ade
a
novel
te
ch
nique
t
hat
ef
fici
ently
recogn
i
zes
the
o
dia
c
har
act
er
.
T
he
com
plete
pr
op
os
e
d
m
et
ho
d
is
desc
ribe
d
gra
ph
ic
al
ly
in
Fig
ure
.
1
.
Thes
e
pro
posed
syst
em
s
are
carried
ou
t
by
includi
ng
t
he
certai
n
ste
ps
li
ke
Im
age
l
ike
Im
age
acqu
isi
ti
on,
pr
e
-
processi
ng,
fe
at
ur
e
e
xtracti
on,
a
nd
cl
assifi
cat
ion
.
The
detai
ls
dis
cussion ca
n be
m
ade in
se
vera
l su
b
-
c
hap
te
rs
i
n
s
ubseq
ue
nt s
ect
ion
.
Figure
1.
Sc
he
m
itic
M
od
el
of Rec
ogniti
on
Mod
el
3.1.
Im
ag
e
A
c
quisi
tion
As
per
our
pro
po
s
ed
m
et
ho
dolog
y
desc
ribe
d
ab
ove
we
ha
ve
co
ns
ide
r
t
he
sta
nd
a
rd
datab
ase
of
odiy
a
char
act
e
r
nam
e
d
as
Nit
R
ourkel
a
Od
ia
data
ba
se, w
hic
h
was
de
velo
pe
d
at
N
IT
,
R
ourk
el
a
b
y
Mi
shra
et
al
.
[
20]
.
In
this
data
bas
e
they
had
co
m
po
sed
of
va
ri
ou
s
15040
num
ber
s
of
i
m
ages
of
both
cha
r
act
er
and
num
erals.
In
this researc
h
analy
sis, w
e h
a
ve
co
ns
ide
red
47
c
har
act
er
s havin
g
200 nu
m
ber
s
of
sam
ples f
or
our
ex
perim
ental
stud
y.
The
m
od
er
n
O
dia
scri
pt
co
ns
ist
s
of
12
vowels,
3
vo
wel
m
od
ifie
rs,
37
sim
ple
con
s
on
a
nts,
10
nu
m
erical
dig
it
s
an
d
a
bout
159
c
om
po
s
it
e
char
act
ers
(
j
ukta
s)
.
O
dia
s
cript
is
a
cu
rv
e
d
ap
pea
ran
ce
of
w
riti
ng
patte
r
ns
on
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
ICT
IS
S
N:
22
52
-
8776
An
gu
l
ar
Symm
et
ric
Axis Co
nst
el
lati
on
M
od
el
for Off
-
li
ne.
.
.
(
Soumya
R
an
j
an N
aya
k
)
99
p
al
m
le
aves
wh
ic
h
ha
ve
bee
n
secur
e
f
ro
m
tear
if
wr
it
er
use
s
too
m
any
str
ai
gh
t
li
nes.
Ta
ble.
2
de
scribe
s
about
o
dia c
ha
racters
w
it
h
t
heir
p
honetic
s.
Table
1.
List
of A
ll
Feat
ur
e
s i
n
Re
c
ogniti
on
of Odia
Cha
rac
te
rs
with
A
cc
uracy
Phonetics
Letter
Phonetics
lett
er
Phonetics
Letter
Phonetics
lett
er
Phonetics
lett
er
a
Ka
Da
Ma
Ek
Aa
Kh
a
Dh
a
Ja
Du
e
E
Ga
Na
Ra
Tini
Ee
Gh
a
Ta
La
Ch
ari
U
Un
a
Tha
Laa
Paan
ch
Uu
Ch
a
Da
Ya
Ch
h
a
Ru
Ch
h
a
Dh
a
Sh
a
Saath
Ru
u
Ja
Na
Sh
a
Aath
A
Jh
a
Pa
Sa
Na
Ae
Ny
a
Ph
a
Ha
Sh
u
n
O
Ta
Ba
Kh
y
a
Ou
Tha
Bh
a
3.2
.
Im
ag
e
Pr
e Proces
sing
Pr
e
-
proces
sin
g
is
an
im
po
rtant
ste
p
du
rin
g
the
im
age
acqu
isi
ti
on
proc
ess
in
orde
r
t
o
get
higher
accuracy
re
su
lt
by
m
eans
pro
du
ci
ng
noise
f
ree
i
m
ages
as
well
as
fr
ee
of
sk
ew
nes
.
I
n
this
analy
sis
ste
p,
our
pre
-
pr
ocessin
g
ste
ps
are
do
ne
by
us
in
g
dif
fer
e
nt
pha
ses
li
ke
no
ise
redu
ct
ion
,
n
orm
ali
zat
ion
,
s
kew
or
sla
nt
adjustm
ent
and
segm
entat
ion
.
T
he
detai
ls
descr
i
ption
of
these
pre
-
proc
essing
ste
ps
ar
e
su
m
m
arised
in
the
fo
ll
owin
g
s
ub
sect
ion
s.
3.
3
.
N
oise
Re
duct
i
on
No
ise
is
the
un
wan
te
d
outp
ut
com
es
with
the
pix
el
intensit
y
value
in
the
sc
ann
e
d
doc
um
e
nt
wh
e
rea
s
reducti
on
of
no
ise
is the
proce
ss of eli
m
inatin
g sp
uri
ous
po
ints d
ue
t
o
the
poor sam
pling
rate o
f
the
sca
nner
.
3.
4
.
N
orm
alizati
on
Norm
al
iz
a
ti
on
is
the
pr
ocess
of
se
par
at
in
g
wh
at
data
we
get
and
w
hat
da
ta
we
req
ui
re
d.
W
e
ad
opt
bin
a
rizat
ion
a
s
the
inte
ns
it
y
norm
al
iz
ation
in
the
pr
e
-
pr
oces
sing
ste
p.
T
he
n
we
a
dju
st
t
he
siz
e
of
eac
h
s
a
m
ple
as 81
*81 dim
e
ns
io
ns
for
siz
e
norm
al
iz
a
ti
on
.
3.
5
.
Skew
or Sl
an
t
Adj
ustm
ent
Sk
e
w
ness
in
the
i
m
age
unde
rgoes
s
om
e
r
otati
on
of
sca
nn
e
d
im
age.
This
is
ver
y
i
m
po
rtant
to
el
i
m
inate
ro
ta
ti
on
i
n
the
pr
e
-
proces
sin
g
ste
p.
Rotat
ion
ca
n
be
el
i
m
inate
d
by
i
m
ple
m
entin
g
t
he
el
im
inati
on
of
degree
of ti
lt
an
gle a
nd
ro
ta
ti
on of
opposit
e
directi
on.
3.
6
.
Se
gmen
t
at
i
on
Segm
entat
ion
is
the
process
of
se
par
at
i
on
of
te
xt
a
nd
non
te
xt
area
i
n
the
sca
nned
hand
wr
it
te
n
do
c
um
ent.
It
is
the
c
halle
ngin
g
par
t
f
or
pre
-
proces
sin
g
the
re
are
2
ty
pes
of
segm
entat
ion
s
can
ha
ve
in
the
pre
-
processi
ng
ste
ps
,
e
xte
rn
al
se
gm
entat
ion
per
f
or
m
se
par
at
ion
of
pa
ra
gr
a
ph,
w
ords
or
se
ntence
f
ro
m
s
cann
e
d
do
c
um
ents wh
ereas inte
rn
al
s
egm
entat
ion
is
the pr
ocess of
separ
at
io
n
c
ha
r
act
er fro
m
each wo
rd.
3.
7
.
Fe
at
ure
Extr
act
i
on
Feat
ur
e
e
xtract
ion
te
c
hn
i
qu
es
are
us
e
d
t
o
eva
luate
the
uniq
ue
ness
of
eac
h
c
har
ac
te
r
im
age
by
w
hic
h
they
diff
e
rs
f
r
om
the
rest
ch
aracte
r
im
ages
.
In
this
sect
io
n
we
ha
ve
im
plem
ented
a
un
iq
ue
al
gorith
m
fo
r
evaluati
on
of
f
eat
ur
e
vect
or
by
con
side
rin
g
the
m
ean
distance
of
row,
m
e
an
an
gle
of
r
ow,
m
ean
dista
nce
of
colum
n
an
d
m
ean
a
ng
le
of
c
o
lum
n
from
ce
ntre
of
t
he
im
a
ge
to
m
idp
oi
nt
of
t
he
sym
m
etr
ic
axis
re
sp
ect
ively
.
All
the
operati
on
s
wer
e
pe
rfo
rm
ed
ov
e
r
s
kel
et
on
iz
ed
im
age
of
hand
wr
it
te
n
c
har
act
e
rs.
O
ur
feat
ur
e
extr
act
ion
i
m
ple
m
entat
io
n
is
m
ai
nly
fo
c
us
es
on
five
un
iqu
e
ste
ps,
a
nd
this
is
c
on
side
red
as
t
he
key
featur
e
values
of
ou
r
pro
po
se
d
syst
e
m
. Th
e d
et
ai
ls
of the
five st
ep
s ar
e
desc
ribe
d as f
ollo
ws:
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
IJ
-
ICT
Vo
l.
7
, No
.
2
,
A
ugus
t
201
8
:
96
–
104
100
1.
This
un
i
qu
e
te
chn
i
qu
e
ext
rac
te
d
the
featu
re
of
im
age
accord
i
ng
t
o
bin
a
ry
i
m
age
point
pa
ssing
thr
ough
the
an
gle
of
45°,
135°,
18
0°
,
225°,
270°,
31
5°
,
360°
resp
ect
ively
from
the
centre
of
the
im
a
ge
represe
nted
i
n F
ig
ure
3
.
2.
Extract
the
point
posit
io
ns
of
t
he
sam
pl
e
i
m
age
hav
i
ng
passe
s
thr
ough
these
a
ngle
s
re
pr
ese
nt
ed
in F
ig
ure
4
.
3.
Plott
ing
all
t
he po
i
nts cr
eat
e
a
un
i
qu
e
poly
go
n
s
ha
ped im
ag
e f
or
eac
h sam
ple
re
pr
e
sente
d i
n
F
i
gure
5
.
4.
Find r
ow sym
m
et
ry axis and
colum
n
sym
m
e
try
ax
is
base
d
on Fig
ure
6
&
9 resp
ect
ively
.
5.
Estim
at
e
m
ean
angle an
d dist
ance
of both
th
e sym
m
et
ry axis b
ase
d on Fig
ure
8 &
11 r
e
s
pecti
vely
.
Fo
r
the
a
bove
al
l
e
m
pirical
c
al
culat
ion
of
our
im
ple
m
entation
m
et
ho
do
l
ogy,
we
ha
ve
de
velo
ped
tw
o
al
gorithm
s w
hich were
d
e
picte
d
in
A
l
gorith
m
. I
an
d Al
gor
it
h
m
. I
I res
pect
ively
.
The
pro
posed
char
act
e
r
recogn
it
io
n
m
e
thod
are
div
i
ded
t
he
i
m
ages
into
two
par
ts
of
operati
on
an
d
the
first
par
t
operati
on
inclu
de
d
a
chor
d
that
is
dr
aw
n
from
each
bounda
r
y
pix
el
to
strai
gh
t
of
boun
dary
pix
el
in
row
wise
a
nd
t
he
seco
nd
par
t
co
ns
ist
in
g
a
ch
o
r
d
that
is
dr
aw
n
f
r
om
each
bounda
ry
to
strai
ght
of
it
s
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
ICT
IS
S
N:
22
52
-
8776
An
gu
l
ar
Symm
et
ric
Axis Co
nst
el
lati
on
M
od
el
for Off
-
li
ne.
.
.
(
Soumya
R
an
j
an N
aya
k
)
101
bounda
ry
pix
el
s
in
colum
n
wise
and
t
he
com
plete
descr
i
ption
of
t
hese
tw
o
ste
ps
a
re
disc
us
se
d
in
Al
gori
thm
1
and
Algorithm
2
respec
ti
vely
.
For
N
no
of
bounda
ry
pix
el
and
K
no
of
boun
dar
ie
s;
the
nu
m
ber
o
f
ava
il
able
cord
is
(
N/2)*
k
in
row
wise
a
nd
col
um
n
wis
e.
H
oweve
r,
w
e
discar
d
th
os
e
boun
dar
y
c
hor
ds
w
hich
ha
ving
le
ss
than
3
pi
xels
in
that
co
rds.
T
he
rem
ai
nin
g
c
ord
is
cal
le
d
r
ow
ch
ords
a
nd
c
olu
m
n
chords
because
these
chor
ds
are
pr
e
sent
in
the
sam
e
ro
w
and
sam
e
colum
n.
These
chord
s
are
par
al
le
l
pr
esent
in
r
ow
ch
ords,
wh
i
ch
is
pr
ese
nted
in
Fi
gure
6
a
nd
t
he
cord
a
re
ver
ti
c
al
ly
pr
ese
nt
in
colum
n
ch
ords
presente
d
i
n
F
igure
9
resp
ect
ively
.
In
our
s
ubseq
ue
nt
ste
p
we
ha
ve
gro
up
the
r
ow
ch
ords
a
nd
col
um
n
chor
ds,
in
or
der
to
f
ind
sym
m
et
ry
axis
from
par
al
le
l
r
ow
ch
ords
a
nd
ve
rtic
al
ly
colu
m
n
cho
r
ds
.
T
he
m
idp
oin
t
of
the
par
al
le
l
r
ow
ch
ords
an
d
ve
rtic
al
colum
n
ch
ords
co
uld
ge
ner
at
e
a
num
ber
of
row
sym
m
e
try
axes
a
s
w
el
l
as
colum
n
sym
m
et
ry
axes
w
hi
ch
are
pr
ese
nted
i
n
Fi
gur
e
7
a
nd
Fig
ur
e
10
res
pecti
vely
.
In
orde
r
to
fin
d
the
acc
urat
e
sy
m
m
e
try
axes
to
represe
nt
the
per
ce
ptu
al
pa
rts, w
e
pr
opos
e
m
idp
oin
t c
rite
r
ia
o
f
the
res
pec
ti
ve
ch
ords
t
o be
ver
ifie
d
i
n
t
he follo
wing m
et
hod
.
Mi
dpoin
t
of
r
ow c
hor
ds
=
Y
r
−
X
r
2
Wh
e
re
r=(1,
2,3,...
.
n) No
of
boun
dar
ie
s
Y
r
=Y
poi
nt of
r
’th r
ow bo
undar
y
point
(1)
X
r
=X
poi
nt of
r
’th r
ow bo
undar
y
point
Mi
dpoin
t
of
c
ol
um
n
chord
s=
Y
c
−
X
c
2
Wh
e
re c=
(1,2,
3,
.
...n
)
N
o o
f b
oundaries
Y
c
=Y
poi
nt of
c ’
t
h
c
olu
m
n
boun
dar
y
point
(2)
X
c
=X
point
of
c ’th col
um
n
boun
dar
y
point
Af
te
r
s
uccess
f
ully
analy
sed
of
a
bove
im
pl
e
m
entat
ion
m
o
del,
we
ha
ve
obta
ined
t
he
fin
al
set
of
r
ow
and
col
um
n
sy
m
m
et
ry
axis.
The
n
w
e
ha
ve
de
velo
ped
t
he
c
on
ste
ll
at
io
n
m
od
el
acco
r
ding
to
t
heir
r
el
a
ti
ve
sy
m
m
e
tric
axis
pix
el
posit
io
n
an
d
m
idp
oint
pix
el
an
gle
from
centre
of
the
im
age.
This
con
ste
ll
at
ion
m
od
el
gen
e
rates
tw
o
set
of
pa
ram
et
e
r
for
each r
ow sy
m
m
e
try
axis
and
c
olu
m
n
sym
m
e
try
axis.
Wh
e
re
one
pa
r
a
m
et
er
sh
ow
m
ean
val
ue
of
relat
ive
distance
of
e
ve
ry
sy
m
m
e
try
axis
pi
xel
posit
ion
t
o
ce
ntre
of
the
im
age
and
oth
e
r
par
am
et
er
sh
ows
the
a
ngle
be
tween
t
he
m
id
po
i
nts
of
the
s
ymm
et
ry
axis
to
the
ce
ntre
of
the
im
age.
The
reafte
r
we
f
ound
fou
r
par
am
et
er
of
each
im
age
hav
in
g
two
par
a
m
et
er
each
fo
r
row
sy
m
m
e
try
axis
and
col
um
n
sy
m
m
e
try
ax
is
pr
ese
nted
in
Figure
8
a
nd
Fi
gure
11
res
pecti
vely
.
Fig
ure
2. Sam
ple I
m
age
Figure
3. S
kelet
on
iz
ed
I
m
ages
Figure
4. A
ngle
Pixel
P
oin
t
Extra
ct
io
n
Figure
5. Plott
ing
Pixel
Po
i
nt
Figure
6. Ro
w
Sy
m
m
e
try
A
xi
s
Fig
ure
7. Mi
dpoin
t
of
Ro
w
Sy
m
m
e
try
A
xi
s
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
IJ
-
ICT
Vo
l.
7
, No
.
2
,
A
ugus
t
201
8
:
96
–
104
102
Figure
8. A
ngle
and
Dist
ance
from
C
entre
Figure
9. Colu
m
n
Sy
m
m
et
ry
Ax
is
Figure
10. Mid
po
i
nt of
Co
lum
n
to
M
id
po
i
nt in
Row
Sym
m
et
r
y
Figure
11.
A
ngle
and D
ist
a
nc
e from
Centre t
o
Mi
dp
oin
t i
n
Colum
n
Sym
metry
Axis
3.
8
.
Cl
as
sific
ati
on
Cl
assifi
cat
ion
is
on
e
of
the
i
m
po
rtant
phases
of
any
recogn
it
io
n
m
od
el
.
Acco
r
ding
to
ou
r
i
m
ple
m
entat
io
n
m
od
el
we
ha
ve
a
dopted
a
two
way
strat
e
gy
f
or
rec
ognit
ion
.
I
n
t
his
re
ga
rd
we
ha
ve
c
ho
s
e
n
two
well
li
ked
cl
assifi
er
nam
el
y
su
pport
ve
ct
or
m
achine
(
SV
M)
[
24
]
an
d
ra
ndom
fo
re
st
tree
(RFT
)
[
25
]
f
or
recog
niti
on
o
f
h
an
dwrit
te
n
c
ha
racters. After ev
al
uatin
g
the
d
esi
re
d
key
fea
ture
values
we
process
t
hese vect
or
to
cl
assifi
er
se
par
at
el
y
and
note
d
down
t
he
ov
e
rall
recog
niti
on
acc
ur
ac
y.
We
ha
ve
fir
st
evaluated
th
e
SV
M
[16]
cl
assifi
ers
wh
ic
h
are
m
ulti
-
cl
ass
cl
assifi
er
an
d
s
up
e
r
vised
one. Secon
dl
y
ran
dom
fo
re
st
tree
[25]
w
hich
is
work
base
d
on
the
idea
of
baggin
g
an
d
rando
m
sel
ec
ti
on
of
featu
r
es.
All
the
pe
rfor
m
ance
wa
s
li
ste
d
dep
e
ndin
g upo
n
the
v
al
ue of
the m
ean square err
or.
A
nd tel
ls ab
ou
t
w
hich cl
assifi
er is the
b
est
one.
4.
RESU
LT
A
N
D DIS
CUSSI
ON
All
the
im
plem
entat
ion
of
our
pro
posed
m
e
tho
d
were
carrie
d
out
with
the
sy
stem
hav
i
ng
sp
eci
ficat
io
n
w
it
h
windows
8
,
64
bit
operati
ng
syst
em
,
and
In
te
l
(R)
i
7
–
4770
CPU
@
3.40
G
Hz,
a
nd
al
l
the
si
m
ulati
on
is
do
ne
th
r
ough
m
at
la
b1
4
(a)
ov
e
r
a
sta
nd
a
rd
da
ta
base.
As
per
sta
nd
a
rd
Datab
ase
con
ta
ini
ng
20
0
sam
ples
fr
om
each
of
the
47
cat
ego
ries
na
m
ed
as
NI
T
R
our
kela
Od
ia
datab
ase
a
nd
consi
der
i
ng
nu
m
eri
c
database
f
r
om
IS
I
K
ol
kata
ha
ving
16
sam
pl
es
fr
om
each
of
the
10
cat
egorised
.
Af
te
r
getti
ng
the
f
our
key
featur
e
vect
or
values
from
e
ach
data
base
a
s
m
ean
distan
ce
of
row,
m
ean
an
gle
of
r
ow,
m
ean
distance
of
colum
n
and
m
ean
an
gle
of
colum
n
from
ce
ntre
of
the
im
age
to
m
idp
oin
t
of
the
sym
m
et
ric
axis
fr
om
each
i
m
age.
Hence
total
siz
e
of
i
nput
f
or
Od
ia
c
ha
racter
be
com
es
4*9400
a
nd
num
eric
char
act
er
bec
om
es
4*9400
and
m
akes
the
se
as
input
to
well
def
i
ned
cl
assifi
er
su
c
h
a
s
SV
M
an
d
ra
ndo
m
fo
rest
a
nd
al
s
o
pe
rfo
r
m
ed
the
validat
io
n
by
im
ple
m
enting
10
f
old
-
cr
os
s
validat
io
ns
to
the
syst
e
m
.
C
on
s
eq
ue
ntly
all
the
ob
ser
vation
w
as
counted
to
ce
rtai
n
as
75
,
25
r
at
io
as
trai
nin
g
and
te
sti
ng.
A
t
first
SV
M
cl
assifi
er
is
i
m
ple
m
ented
fo
ll
ow
ed
up
by r
an
do
m
f
or
est
classi
fier.
We h
a
ve
al
so
m
ade a
com
par
ison
a
naly
ses am
on
g
these tw
o
cl
assifi
ers,
a
nd
li
ste
d
93.6%
as
the
r
ecognit
ion
rate
for
SV
M
a
nd
98.2%
f
or
the
r
andom
fo
rest
f
or
NI
T
O
dia
char
act
er
,
sim
i
l
arly
for
IS
I
nu
m
eric
char
act
er
the
r
ecognit
ion
rate
fo
r
bo
t
h
S
VM
and
rand
om
fo
rest
as
88.91%
a
nd
96.3%
resp
ect
ively
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
ICT
IS
S
N:
22
52
-
8776
An
gu
l
ar
Symm
et
ric
Axis Co
nst
el
lati
on
M
od
el
for Off
-
li
ne.
.
.
(
Soumya
R
an
j
an N
aya
k
)
103
5.
CONCL
US
I
O
N
In
t
his
pa
per,
we
ha
ve
pr
ese
nted
a
n
a
ngul
ar
sym
m
et
ric
const
el
la
ti
on
te
chn
i
qu
e
f
or
offli
ne
O
dia
char
act
e
rs
rec
ogniti
on.
T
his
syst
e
m
us
es
r
ow
a
nd
c
olu
m
n
sym
m
et
ric
a
xis
f
or
ge
ner
at
ing
four
key
f
eat
ur
e
vecto
r
val
ues
f
ro
m
each
data
base
as
m
ean
di
sta
nce
of
r
ow,
m
ean
an
gle
of
row,
m
ean
dis
ta
nce
of
col
umn
an
d
m
ean
an
gle
of
colum
n
f
r
om
centre
of
the
im
age
to
m
idpoint
of
the
sym
m
e
tric
axis
f
ro
m
each
im
a
ge.
F
or
cl
assifi
cat
ion
pur
po
se
,
SV
M
and
R
F
m
od
el
is
us
ed.
A
n
ex
per
im
ental
resu
lt
fr
om
this
researc
h
giv
es
sat
isfact
or
y
re
cogniti
on
res
ult
ov
er
t
he
sta
ndar
d
dataset
,
but
sti
ll
the
devel
op
m
ent
is
in
it
s
infan
cy
.
F
ur
t
her
,
oth
e
r
te
ch
niqu
es are t
o be e
xplo
red f
or b
et
t
er r
ec
ogniti
on
accuracy.
REFERE
NCE
S
[1]
H.
Sw
et
hal
aksh
m
i
,
et
al.
,
“
Onli
ne
handwri
tten
cha
ra
cter
rec
ogn
it
ion
of
deva
n
a
gar
i
and
te
lu
gu
cha
ra
cters
using
support
vector
m
ac
hin
es,
”
2006
I
nte
rnational
Wo
rkshop on
Frontiers
in
Handwr
it
i
ng
Recogni
t
ion.
2006.
[2]
U.
Pal
,
et
al
.
,
“
A
sy
stem
for
off
-
li
ne
Ori
y
a
h
andwri
tten
ch
ar
ac
t
er
re
cognitio
n
using
cur
v
at
u
re
fe
at
ur
e,
”
200
7
int
ernati
ona
l co
nfe
renc
e
on
inf
o
rm
ati
on
technol
ogy,
Ind
ia,
pp.
2
27
-
229
,
2007
.
[3]
V.
K.
Govindan
.
,
A.
P
.
Shivapr
asa
d,
“
Char
ac
t
e
r
rec
ogn
it
ion
a
rev
ie
w
,
”
Pat
t
ern
Recogni
t
io
n,
vol
.
23,
pp
.
671
-
683,
1990
.
[4]
J.
Manta
s
,
“
An o
ver
vie
w of
ch
aract
er
re
cogni
t
ion m
et
hodologi
es,
”
Pattern
R
ec
ogn
i
ti
on,
vol. 19
,
pp
.
425
-
430,
1986.
[5]
U.
Pal.,
B.
B.
C
haudhur
i,
“
Indi
a
n
script
ch
aract
e
r
rec
ogni
ti
on:
a
surve
y
,
”
Pa
ttern
Re
cogn
it
ion
,
vo
l.
37,
pp
.
1887
-
1899,
2004
.
[6]
R.
Plamondon,
S.
N.
Srihari,
“
On
-
Li
ne
and
o
f
f
-
li
ne
h
andwri
tten
rec
ogn
it
ion
:
A
comprehe
nsive
surve
y
,
”
IEEE
Tr
ans on
PA
MI,
vol.
22,
pp.
62
-
84
,
2000
.
[7]
N.
Aric
a
and
F.
Vural
,
''
An
Overvi
ew
of
C
har
acte
r
R
ec
ogn
it
ion
Focused
o
n
Off
-
Li
ne
Han
dwriti
ng,
''
I
E
E
E
Tr
ansacti
ons on System
Man
Cyb
erne
tics
-
Part
C:
Appl
ic
a
ti
ons and
Revie
ws
,
vol
.
31,
pp.
216
-
233,
20
01.
[8]
L.
M.
Lor
igo
an
d
V.
Govindar
aju,
''
Off
li
ne
Arab
ic
Handwrit
ing
Rec
ognition:
A
Surve
y
,
''
IEEE
Tr
ansacti
ons
on
Pat
te
rn
Analysis and
Mac
h
ine In
te
lligen
ce
,
vol
.
2
2,
pp
.
712
-
724
,
2006.
[9]
G.
Nag
y
,
''
Chin
e
se
Chara
c
te
r
R
e
cogni
ti
on
,
A
twe
nt
y
f
ive
y
e
ars
re
trospec
t
ive
,
''
19
9
8
IEE
E
int
ernat
i
onal
conf
er
ence
on
pattern rec
og
nit
ion,
It
aly
,
pp
.
109
-
114
,
1988
.
[10]
A.
K.
Jain
,
T
.
T
axt
,
''
Fe
at
ur
e
Ex
tra
c
ti
on
Me
thod
s
for
Chara
c
te
r
Rec
ognition
-
A
Surve
y
,
”
Pat
t
ern
Re
cogn
it
ion
,
v
ol.
29,
pp
.
641
-
662
,
1996.
[11]
htt
p://ww
w.odi
sham
useum
.
nic
.
i
n/
?
q=d
ept
/manu
script
s
[12]
P.
Patt
n
ai
k
,
“
Pre
sentat
ion
on
dig
i
ti
zation
o
f
Odia
books”
in
Utka
l U
nive
rsit
y
(2
1
Februa
r
y
2014
).
[13]
B.
B.
Chaudhur
i
,
et al
.
,
“
Autom
at
ic
r
ec
ogni
ti
on
of
printed
Ori
y
a
sc
ript
,
”
Sadhana
,
vol.
27
,
pp
.
23
-
3
4,
2002
.
[14]
M.
Obaidul
l
ah
,
et
al
.
,
“
Struct
ur
al
fe
at
ur
e
base
d
appr
oac
h
o
r
scr
ipt
ide
n
ti
f
ic
a
ti
on
from
print
ed
in
dia
n
document
,
”
2014
IEEE inte
r
nati
onal conf
ere
nce
on
signal
proce
ss
ing
and
integrated
n
et
work
s,
India
,
pp
.
120
-
124
,
2014
.
[15]
T.
W
aka
b
a
y
ashi
,
et
a
l.
,
“F
-
rat
io
base
d
weighted
fea
tur
e
extrac
tion
for
sim
il
ar
shape
ch
aract
er
r
ec
ogni
ti
on,
”
200
9
IEE
E
internatio
nal
con
fe
renc
e
o
n
docume
nt
anal
ysis
and
rec
ogni
ti
on,
Spain
,
pp
.
196
-
200
,
2009
.
[16]
K.
Ro
y
,
et
al.
,
“
Oriy
a
handwri
t
te
n
num
era
l
r
ecogniti
on
s
y
s
te
m
,
”
2005
IE
EE
8
th
int
ernat
ional
conf
ere
n
ce
o
n
docume
nt
ana
ly
s
is and
rec
ogn
it
io
n,
South
Korea,
pp.
770
-
774
,
20
05
.
[17]
T.
K.
Bhowm
ik
,
et
al.
,
“
An
HM
M
base
d
rec
o
gnit
ion
sch
eme
for
handwri
tte
n
Ori
y
a
num
erals,”
2006
I
EE
E
int
ernati
ona
l co
nfe
renc
e
on
inf
o
rm
ati
on
technol
ogy,
Ind
ia,
pp.
1
05
-
110
,
2006
.
[18]
K.
S.
Dash
,
et
al.
,
“
A
h
y
brid
f
ea
tur
e
and
discr
iminant
class
ifi
e
r
for
high
ac
cur
acy
Odia
hand
writt
en
n
um
er
a
l
rec
ogni
ti
on,
”
20
14
IEEE tec
hn
ical
sympos
ium,
M
alay
sia,
pp.
531
-
535
,
2014
.
[19]
K.
S.
Dash
,
et
al
.
,
,
“
Non
-
red
undant
Stockwel
l
tr
a
nsform
base
d
fea
ture
ex
tracti
on
for
handwri
tt
en
digi
t
re
cogni
t
ion
,
”
2014
IEEE
sign
al
proce
ss
ing
an
d
communic
at
io
n,
India
,
pp
.
1
-
4
,
2014
.
[20]
T.
K.
Mishra
,
et
al.
,
“
A
compara
ti
ve
ana
l
y
sis
of
image
tra
nsform
at
ions
for
handwri
tt
en
Odia
num
e
ral
rec
ogn
it
ion
,
”
2013
IEE
E
int
er
nati
onal
co
nfe
re
nce
on
advan
ce
s
in
computi
ng
,
c
omm
unic
ati
ons
and
inf
orm
ati
cs,
India,
pp
.
790
–
793
,
2013
.
[21]
B.
Majhi
,
et
al.
,
“
Eff
ic
ie
nt
r
ec
o
gnit
ion
of
Odi
y
a
num
era
ls
using
low
complexi
t
y
neu
ral
c
la
ss
if
ie
r,”
2011
IEEE
int
ernati
ona
l co
nfe
renc
e
on
en
ergy
,
au
tomati
on
,
and
signal, Indi
a
,
pp
.
1
-
4
,
2011
[22]
S.
Chanda
,
e
t
al
.
,
,
“
Te
x
t
ind
epe
n
dent
writ
er
id
entificat
ion
fo
r
Ori
y
a
scrip
t
,
”
2012
IEE
E
int
ernat
io
nal
workshop
on
docume
nt
ana
ly
s
is sy
stems, Australia
,
pp.
369
-
37
3
,
2012
.
[23]
K.
S.
Dash
,
et
al
.
,
“
BESAC
:
Bin
ar
y
Externa
l
S
y
m
m
et
ry
Axis
Constel
lation
for
u
nconstra
in
ed
han
dwritt
en
cha
r
ac
t
e
r
rec
ogni
ti
on,
”
Pa
tt
ern
Re
cogn
it
io
n
Letters
,
vol
.
38
,
pp
.
413
-
422
,
2
016.
[24]
B.
Xu
,
et
a
l.
,
“
An
improved
ran
dom
fore
st
class
ifi
er
for
image
c
la
ss
ifica
ti
on,
”
2012
IE
EE
In
te
rnation
a
l
Confe
renc
e
on
I
nformation
and
Aut
omation
,
Chi
na,
pp
.
795
-
800
,
2012.
[25]
C.
Mitra
and
A.
K.
Pujar
i,
“
Mining
Intelli
g
e
nce
and
Know
l
edge
Expl
ora
ti
o
n.
Editor.
Mining
Inte
l
li
gen
ce
an
d
K
nowledge
Exp
l
ora
ti
on
,
”
2009
S
pringer
Lect
ure
Note
s in
Comput
er
Scienc
e, I
ndia
,
pp
.
82
-
84
,
201
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
IJ
-
ICT
Vo
l.
7
, No
.
2
,
A
ugus
t
201
8
:
96
–
104
104
BIOGR
AP
HI
ES
OF
A
UTH
ORS
P
y
ari
Mohan
J
ena
Com
pleted
his
BS
C
and
MS
C
in
computer
sc
ie
nc
e
f
rom
Rave
nshaw
Univer
sit
y
,
cut
t
a
ck
which
is
a
rep
ute
d
and
Oldest
insti
tute
of
India
aft
er
He
recei
ved
hi
s
exc
e
ll
en
c
y
M.T
ec
h
degr
ee
in
c
om
pute
r
scie
nc
e
and
engi
ne
eri
n
g
from
CET
,
BBS
R
under
Bij
u
P
at
nai
k
Unive
rsi
t
y
of
Technol
og
y
,
Odish
a,
Indi
a
in
2017.
His
re
sea
rch
in
te
rest
is
in
the
fi
el
d
of
Im
age
s proc
essi
ng,
ch
aracter
r
ecogniti
on
and
p
att
ern
re
cognition.
Soum
y
a
R
anj
an
Na
y
ak
r
ec
e
ive
d
his
exc
ellen
c
y
in
bat
chelor
degr
e
e
in
2009
and
m
aste
r
degr
ee
in
2012
of
comput
er
scie
nc
e
and
e
ngine
er
ing
from
Bij
u
Patna
ik
Uni
ver
sit
y
of
T
ec
hn
olog
y
,
Odisha,
India
.
BP
UT,
R
ourke
la
is
one
of
la
rge
st
and
old
e
st
Univer
sit
y
in
India
.
Curr
entl
y
,
he
is
working
as
As
sistia
nt
Profess
or
in
K
L
U
nive
rsit
y
,
Andhr
a
Prade
s
h
.
His
rese
arc
h
ar
ea
of
i
nte
rest
includes
image
an
aly
sis b
y
m
e
ans
of
fr
act
al
g
eometr
y
,
co
l
or
and te
x
ture feature
s.
Evaluation Warning : The document was created with Spire.PDF for Python.