Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
Vol
.
4
,
No
. 5, Oct
o
ber
2
0
1
4
,
pp
. 71
9~
72
9
I
S
SN
: 208
8-8
7
0
8
7
19
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Te
mpla
te
Ma
tc
hing
Me
tho
d
fo
r Re
cognition of St
on
e Ins
crip
t
ed
Kannada Characters of Differe
nt Time
Fr
a
m
e
s
Ba
se
d o
n
Correl
at
ion Anal
ysis
Ra
jith
kum
ar B
K
*
, H
.
S.
Mohan
a
**
*Department of ECE,
**Depar
tment
of Instrum
e
ntation
Technolog
y
Malnad Co
lleg
e
of Engin
eerin
g,
Visvesvaray
a
Technol
ogical University
(VTU)
,
Belgaum, K
a
rnataka, Ind
i
a
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
May 27, 2014
Rev
i
sed
Au
g
13
, 20
14
Accepted Aug 25, 2014
Stone in-script
e
d liter
a
tur
e
speaks about the his
t
or
y
,
l
a
nguage
of differen
t
regions of
the
world. Preserv
a
tion of
such
document th
rough digitalization
process is beco
me ver
y
important.
To
stop d
e
gr
adation and
missing further
,
the an
al
y
s
is of t
h
e sam
e
will
thr
ough light on
hi
storical ev
ents o
f
that
region
.
In this connecti
on present work propos
es a si
m
p
le m
e
thod of
digitizatio
n
using ordinar
y
digital camera furt
her, the p
r
e-processing algorithm is
implemented to
enhance th
e image and
im
pr
ove the r
ead
abi
lit
y.
Here
it
recognizes
th
e Kannada char
acters
ba
sed
on template matching. In
this
method is normally
implemen
ted b
y
first p
i
ckin
g template
and then it
call
the s
e
a
r
ch
im
age
,
then
b
y
s
i
m
p
l
y
com
p
aring th
e
tem
p
lat
e
over
ea
ch point
in
the search image and it calcu
late the su
m of
products between th
e coefficien
t.
Based on this
calcu
l
ated product valu
e
it recog
n
izes th
e ch
aracter.
Cross
correl
a
tion t
ech
nique is
im
plem
ented in m
a
tch
i
ng the chara
c
t
e
r
s
coeffici
ent
.
Experimental r
e
sults shows, it dem
onstrates r
e
la
tive
l
y high
a
ccura
c
y
in
recognizing Sto
n
e inscrip
tions
charac
ters of both Ho
y
s
ala, Ganga time
frames and with
better time eff
i
ciency
when
comp
ared
to pr
evious
methods.
Keyword:
Tem
p
late Matc
h
i
ng
Cro
s
s Correlatio
n
Copyright ©
201
4 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Raj
ith
ku
m
a
r B K,
Depa
rt
m
e
nt
of
EC
E,
Malnad C
o
llege of E
ngi
neeri
n
g,
Visves
va
ra
y
a
Tech
nol
ogi
cal
Uni
v
ersi
t
y
(VT
U
)
,
Belgaum
,
Karnataka,
India
1.
INTRODUCTION
Ind
i
a is
p
r
aised
for its
rich
past and
th
e cu
ltu
re.
T
h
e ri
c
h
heri
t
a
ge
o
f
t
h
e
co
unt
ry
ha
s
b
een ca
rri
e
d
ove
r ge
nerat
i
o
n t
h
r
o
ug
h t
h
e m
a
nuscri
p
t
s
an
d hi
st
o
r
i
c
wri
t
i
ngs
. R
a
pi
d
gr
o
w
t
h
o
f
t
ech
nol
ogy
an
d p
r
e
v
al
ent
us
e
of
com
put
er i
n
t
h
e
bu
si
ness
a
n
d
ot
her
areas
,
m
o
re and
m
o
re o
r
ga
ni
zat
i
o
n
a
r
e c
o
n
v
ert
i
ng
t
h
ei
r
pa
per
d
o
c
u
m
e
nt
in
to
electro
n
i
c do
cu
m
e
n
t
s that can
be processed
b
y
com
p
u
t
er [1
]. R
eco
gn
itio
n of
an
y ston
e in
scrip
tion
s
ch
ar
acter
w
ith r
e
sp
ect to
an
y
lan
g
u
a
g
e
is
d
i
f
f
i
cu
lt.
K
a
n
n
a
d
a
langu
ag
e has go
t a
h
i
sto
r
y o
f
m
o
r
e
t
h
an
2000
year
s and K
a
nn
ad
a in
scr
i
p
tion
s
f
oun
d on h
i
sto
r
ical
h
e
ro
Sto
n
e
, co
i
n
and
te
m
p
le wall, p
i
llar, tab
l
et an
d ro
ck
edi
c
t
[
20]
.
A
n
al
y
s
i
s
of
any
l
a
ng
ua
ge
wi
t
h
r
i
ch
heri
t
a
ge
an
d
hi
st
o
r
y
i
s
ve
ry
i
m
port
a
nt
t
o
un
de
rst
a
n
d
t
h
e l
i
f
e
and
cul
t
u
re
of
t
h
at
pe
ri
o
d
.
It
i
s
necessa
ry
t
o
d
i
gi
t
i
ze St
one
i
n
scri
pt
i
o
ns
by
m
odern t
e
c
hni
que
.
Here in the present work, the im
age is
processed
suc
h
that its
character
is recognized. The
m
a
jor
p
r
ob
lem
wh
ich arises
wh
ile i
d
en
tifying
t
h
e
characte
r
s in a
stone
insc
ripti
o
n is the
difference in the sty
l
e in
literatu
re. Temp
late
m
a
tch
i
n
g
, o
r
m
a
trix
m
a
tch
i
n
g
, is on
e o
f
th
e m
o
st commo
n
classificatio
n
m
e
th
o
d
s. In
t
e
m
p
l
a
t
e
m
a
t
c
hi
ng, i
n
di
vi
dual
im
age pi
xel
s
are used as feat
u
r
es [3]
.
C
l
assi
f
i
cat
i
on i
s
perf
o
r
m
e
d by
co
m
p
ari
n
g
an input cha
r
a
c
ter im
age w
ith a set
of templates from
each c
h
aracter
class. Eac
h
c
o
mparis
on res
u
lts in a
sim
i
l
a
ri
ty
m
e
asure
b
e
t
w
ee
n
t
h
e i
n
p
u
t
cha
r
act
er an
d
t
h
e
t
e
m
p
l
a
t
e
. One
m
easure i
n
c
r
eases t
h
e
am
ount
o
f
si
m
ilarit
y
wh
en
a p
i
x
e
l in
the o
b
serv
ed
ch
aracter is id
en
tical to
th
e same p
i
x
e
l in
t
h
e te
m
p
late i
m
ag
e
.
If t
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 5
,
O
c
tob
e
r
20
14
:
719
–
7
29
72
0
pi
xel
s
di
f
f
e
r
t
h
e
m
easure of s
i
m
i
l
a
ri
t
y
m
a
y
be dec
r
eas
ed
.
After all te
m
p
lates h
a
v
e
b
e
en co
m
p
ared
with
th
e
o
b
s
erv
e
d ch
aracter im
ag
e, th
e ch
aracter's id
en
tity is a
ssig
n
e
d
as th
e i
d
en
tity o
f
th
e m
o
st si
m
i
lar te
m
p
late.
Struct
ural clas
sification m
e
thods utilize
structural
features and d
ecision rules to classify
characters.
St
ruct
ural
feat
ures
m
a
y
be de
fi
ne
d i
n
t
e
rm
s of c
h
a
r
act
er
st
roke
s, c
h
aracte
r
holes,
or ot
her cha
r
acter attri
butes
suc
h
as c
onca
v
ities. For a
cha
r
acter im
age input, the structural features a
r
e
extracted and
a rule
-base
d
sy
stem
is applied to
classify the character.Tem
plate
m
a
tc
hi
ng fo
r cha
r
act
er
reco
g
n
i
t
i
on i
s
st
rai
ght
f
o
rwa
r
d a
n
d
rel
i
a
bl
e. T
h
i
s
m
e
t
hod i
s
m
o
re t
o
l
e
ra
nt
t
o
n
o
i
se t
h
an
st
r
u
ct
u
r
al
anal
y
s
i
s
m
e
t
h
o
d
.
2.
RELATED WORKS
Th
e
rev
i
ew of
th
e literatu
re
pertain
i
ng
to
t
h
e p
r
esen
t top
i
c i
s
p
r
esen
ted
to
t
h
e read
ers.
In [1
] au
thors
conce
n
trate
on Tem
p
late
Matching m
e
thod for Rec
o
gnitio
n Musna
d
c
h
ara
c
ters base
d
on
correlation a
n
a
l
ysis.
In th
is
p
a
p
e
r,
we ex
tend
ed
t
h
at
work an
d
ap
p
lied th
at
alg
o
rith
m
fo
r reco
gn
ize
Ston
e
in
scrip
tion
s
Kan
n
a
da
characte
r
s.
In [3] authors
c
o
n
cent
r
at
e
on t
h
e
Era
Ide
n
t
i
f
i
cat
i
on a
n
d R
eco
g
n
i
t
i
on
of
St
o
n
e
In
-sc
r
i
p
t
e
d
Ka
nna
d
a
Ch
aracters
Usi
n
g
Artificial Neural Network
s
. In
th
is
paper we use same Gau
ssian
filter for filtering, th
e
Gau
s
sian
filter sm
o
o
t
h
i
ng
t
h
e im
ag
e an
d it h
e
lp
s
find edg
e
s
o
f
characters accu
r
ately. In
[5
]
au
tho
r
s
co
n
c
en
trate Pri
n
ted
Nu
m
b
er Reco
gn
itio
n
u
s
in
g
M
A
TLAB.
In
th
is
p
a
p
e
r
we
u
s
e sam
e
th
inn
i
ng
an
d cro
p
p
i
ng
proce
d
ure for
to extract
desi
red c
h
a
r
acters
sha
p
e.
In
[2] authors conc
entrate on E
x
t
r
action
of Ka
nna
da
characte
r
s
using SIFT
. In this
pape
r
we e
x
tended that
work
and applied t
h
a
t
al
go
rithm
for
im
age M
o
saic.
3.
PROP
OSE
D
ALGO
RITH
M
The
pr
o
p
o
s
ed
m
e
t
hod c
o
nsi
s
t
s
o
f
f
o
l
l
o
wi
n
g
st
eps see
i
n
Fi
g
u
re
1
1)
Create
a Te
mplate
of Kannada c
h
ar
ac
ters and e
a
ch imag
e in a template is i
n
size o
f
24x4
2
dimensions
.
2)
Test imag
es
C
a
pt
ure
Ka
n
n
a
d
a st
one
i
n
sc
ri
pt
i
o
n
s
c
h
aract
e
r
s
usi
n
g
or
di
na
ry
di
gi
t
a
l
cam
e
ra
of
1
6
M
e
ga
p
i
xel
res
o
l
u
t
i
o
n
3)
Image
Mosaic
b
a
sed
on
SIF
T
algorith
m
Step
s i
n
vo
lv
ed in
th
is is
In th
is it read
t
e
st i
m
ag
es
It
pe
rf
orm
m
o
sai
c
base
d
on
ke
y
poi
nt
s f
o
un
d
and
E
u
cl
i
d
i
a
n
di
st
ance
bet
w
e
e
n t
e
st
i
m
ages
If k
e
y po
in
ts are fou
n
d
,
th
en
it d
ecid
e
s th
at is th
e co
n
tinu
a
tio
n
o
f
th
at im
ag
e an
d
it fu
se tho
s
e im
ag
e, else
it discard and a
g
ain
select another test im
age.
4)
Pre-pr
ocessing
If M
o
saic
done
succes
sful the
n
it pe
rform
pre-proce
ssing. In
this
it involves
a.
R
e
m
ovi
n
g
noi
s
e
an
d R
e
si
zi
n
g
of
al
l
pre
-
pr
oc
essed i
m
ages i
n
t
o
fi
xe
d
pi
xel
si
ze an
d
di
m
e
n
s
i
o
n
b.
Fi
ndi
ng
E
dge
i
n
a
n
i
m
age usi
n
g
S
obel
e
d
ge
det
ect
i
o
n
c.
Perform
d
ilatio
n
an
d Use top
-
h
a
t filtering
t
o
co
rrect un
ev
en illu
m
i
n
a
tio
n
d.
Re
m
o
v
e
all objects in
th
e im
a
g
e co
n
t
ain
i
ng
fewer th
an
8
0
pix
e
ls
e.
Reco
n
s
t
r
u
c
tion of
im
ag
e b
y
r
e
co
nstr
u
c
ting
its bo
und
ar
y and
f.
Fillin
g
its ho
les
g.
Th
inn
i
ng
o
f
characters and
tak
e
co
m
p
le
m
e
n
t
of im
ag
e for clear v
i
sib
ility
5)
Ch
arac
ters C
r
oppi
n
g
a.
Th
is is a u
s
er
blo
c
k
;
h
e
re u
s
er
can
crop
an
y ch
aracters in an
i
m
ag
e for Reco
gn
itio
n
6)
Cros
s c
o
rrelation
a.
In
t
h
i
s
m
e
t
hod
we
per
f
o
r
m
C
r
oss c
o
r
r
el
at
i
o
n
bet
w
ee
n Tem
p
l
a
t
e
and
ext
r
act
ed c
h
aract
e
r
7)
Reco
gni
t
i
o
n
o
f
K
a
nn
ad
a
S
t
on
e ins
criptions characters
a.
B
a
sed
on
C
r
o
ss co
rrel
a
t
i
o
n
A
n
al
y
s
i
s
val
u
e i
t
rec
o
gni
z
e
C
a
pt
u
r
ed
K
a
nna
da
st
o
n
e
i
n
scri
pt
i
o
ns
characte
r
b.
If
val
u
e o
f
a
n
y
t
w
o i
m
ages shoul
d be
hi
g
h
t
h
en i
t
di
spl
a
y
s
t
h
e rec
o
g
n
i
z
e
d
charact
e
r
an
d i
f
co
rrel
a
t
i
o
n
v
a
lu
e app
ears l
o
w th
en
it d
i
spla
y character not rec
o
gnize
d.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Temp
la
te Ma
tch
i
ng
Met
h
o
d
f
o
r Recogn
itio
n
o
f
S
t
o
n
e
In
scri
p
t
ed
K
a
n
nad
a
Ch
a
r
a
c
ters
o
f
… (Ra
jith
kumar B K)
72
1
Fi
gu
re
1.
A
Fl
ow
cha
r
t
di
agr
a
m
of t
h
e P
r
o
p
o
se
d Al
g
o
ri
t
h
m
4.
METHO
D
OL
OGY AND
IMPLEME
N
T
A
TION
4.
1.
I
m
age Mosaic
Based on
Simplified
SI
FT
Th
is is th
e v
e
ry i
m
p
o
r
tan
t
step
in
ou
r
p
r
o
posal a
nd we ca
n’t capt
u
re
d all Characters in
an stone i
n
-
scrip
t
ed b
y
si
ng
le im
ag
e with
h
i
gh
reso
l
u
tio
n, for to
m
a
i
n
tain
g
ood
reso
lu
tion
an
d to
av
o
i
d
o
v
e
rlappin
g
o
f
characte
r
s
between two im
a
g
es
in M
o
saic, he
re
we
use
n
e
w M
o
sai
c
t
e
c
hni
que
cal
l
e
d
‘Im
age M
o
sai
c
B
a
sed
o
n
SI
FT
A
l
gor
ith
m
’
, th
e
m
a
in
adv
a
n
t
ag
es
o
f
th
is algor
ith
m
is
it w
i
l
l
m
o
saic
th
e tw
o
im
ag
es w
ith
g
ood
resol
u
tion a
n
d it elim
inate overla
ppi
ng of
characte
r
s
b
e
tween
t
w
o im
a
g
es and
it prod
u
c
e
o
u
t
pu
t
m
o
saic
im
age like naturally capture
d im
age. This
m
o
saic imag
e was helpe
d
fut
u
re steps t
o
extract all Kanna
da
characte
r
s easil
y.
4.
1.
1. SIFT
SIFT
key
p
o
i
n
t
s
of
ob
ject
s are
fi
rs
t extracte
d
from
a set of refere
nce im
ages and st
ore
d
in a
database
.
An object is
re
cognized in a
new im
age by
indivi
dually
c
o
m
p
aring eac
h
feature
from
the ne
w im
age to this
dat
a
base a
nd fi
ndi
ng ca
ndi
dat
e
m
a
t
c
hi
ng fea
t
ures base
d o
n
Eucl
i
d
ea
n di
st
ance o
f
t
h
ei
r f
eat
ure vect
ors
.
From
t
h
e f
u
l
l
set
o
f
m
a
t
c
hes, su
bse
t
s of
key
poi
nt
s t
h
at
ag
ree
on
t
h
e o
b
j
ect
an
d
i
t
s
l
o
cat
i
on,
sc
al
e, an
d
ori
e
nt
at
i
on i
n
th
e n
e
w im
ag
e are id
en
tified
t
o
filter
ou
t go
od
m
a
tch
e
s.
Th
e d
e
term
in
atio
n
o
f
con
s
isten
t
cl
u
s
ters is
p
e
rfo
r
med
rapi
dl
y
by
usi
n
g a
n
e
ffi
ci
ent
h
a
sh t
a
bl
e i
m
pl
em
ent
a
t
i
on
of t
h
e ge
neral
i
zed
Ho
u
g
h
t
r
a
n
sf
or
m
.
Each cl
ust
e
r
of
3
or m
o
re features that agree on an
o
b
j
ect
an
d i
t
s
pose i
s
t
h
en su
bje
c
t
t
o
furt
her det
a
i
l
e
d
m
odel
veri
fi
cat
i
on an
d
su
bsequ
e
n
tly o
u
tliers are
d
i
scard
e
d
.
Fin
a
lly th
e p
r
o
b
a
b
ility th
at a p
a
rticu
l
ar set o
f
featu
r
es ind
i
cates th
e
prese
n
ce
of a
n
object is compute
d
,
give
n the accuracy of fit and num
b
er
of
probable false
m
a
tches. Object
matches that pass all these tes
t
s can
be i
d
en
ti
fied as
correct
w
ith
h
i
gh
co
nf
i
d
en
ce.
4.
2. Pre-pr
oce
ssi
ng
In
th
is we first rem
o
v
i
n
g
th
e no
ise u
s
ing
Gaussian
filter
4.
2.
1 Medi
a
n
f
i
l
t
er
Med
i
an
filtering
h
e
l
p
s
u
s
b
y
erasing
th
e
b
l
ack
do
ts
, called th
e Pep
p
e
r, and it also
fills in
wh
ite ho
les
in
th
e im
ag
e, called
Salt “i
m
p
u
l
se no
ise”. It's lik
e th
e m
ean
filter bu
t is
b
e
tter in
1-
Preserv
i
n
g
sh
arp
edg
e
s 2
-
Th
e m
e
d
i
an
valu
e is m
u
ch
lik
e
n
e
igh
bou
rho
o
d
p
i
x
e
ls a
nd will no
t affect th
e
o
t
h
e
r p
i
x
e
ls sig
n
i
fican
tly -th
i
s
means that the
mean does t
h
at.
Med
i
an
filtering
is p
o
p
u
l
ar in re
m
o
v
i
ng
salt n
p
a
p
e
r no
ise an
d
wo
rk
s b
y
rep
l
acing
th
e p
i
x
e
l v
a
lue
with
th
e m
e
d
i
an
v
a
lu
e i
n
t
h
e
n
e
igh
bou
rho
od of th
at
p
i
x
e
l.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
JECE Vo
l. 4
,
N
o
. 5
,
O
c
tob
e
r
20
14
:
719
–
7
29
72
2
4.
2.
2 E
d
ge
det
ecti
o
n
The E
dge
Det
ect
i
on bl
oc
k
com
put
es t
h
e
aut
o
m
a
t
i
c
t
h
reshol
d
usi
n
g t
h
e m
ean of t
h
e g
r
a
d
i
e
nt
m
a
gni
t
ude
s
q
u
a
red i
m
age. H
o
we
ve
r, y
o
u
c
a
n a
d
j
u
st
t
h
i
s
t
h
res
h
ol
d
usi
n
g
t
h
e T
h
res
h
ol
d
scal
e fact
o
r
(u
sed
t
o
au
to
m
a
tical
ly
calcu
late th
resh
o
l
d
v
a
lu
e) param
e
ter. Th
e b
l
o
c
k
m
u
ltip
lies th
e v
a
lu
e yo
u
en
ter wi
th
th
e
au
to
m
a
tic th
resh
o
l
d
v
a
lu
e t
o
determin
e a n
e
w th
resho
l
d
val
u
e. He
re
we a
r
e
use
d
s
obel
e
d
g
e
det
ect
i
o
n.
4.
2.
3 Di
l
a
ti
on
The
dilation operator ta
kes two pieces
of
dat
a
as inputs
. T
h
e first is the image which is t
o
be dilated.
The second is a set of coordi
nate poi
nts known as a stru
c
t
uring elem
ent (also known as a kernel
). It is this
stru
cturing
elemen
t th
at d
e
termin
es th
e
p
r
eci
se effect of th
e d
ilatio
n
on
th
e in
pu
t i
m
ag
e. No
te th
at in
th
i
s
an
d
sub
s
eq
ue
nt
di
a
g
ram
s
, f
o
re
gr
o
u
n
d
pi
xel
s
are
rep
r
ese
n
t
e
d
by
1'
s an
d
bac
k
g
r
ou
n
d
pi
xel
s
by
0'
s. Si
m
p
l
e
di
l
a
t
i
on
ope
rat
i
o
n i
s
as
sho
w
n i
n
fi
gu
r
e
Fi
gu
re
2.
A
3×
3 s
q
u
a
re
st
ruct
uri
n
g
el
em
ent
I
n
abov
e f
i
gu
re is
3
×
3
str
u
ctu
r
ing
elem
en
t, th
e ef
f
ect o
f
th
is o
p
e
r
a
tion is to
set
to
th
e f
o
r
e
g
r
ound
col
o
r any
back
gr
o
u
n
d
pi
xel
s
t
h
at
have a ne
i
g
h
b
o
u
r
i
n
g fo
r
e
gr
o
u
n
d
pi
x
e
l
(assum
i
ng 8
-
c
o
n
n
ect
ed
ness
).
Such
pi
xel
s
m
u
st
l
i
e
at
t
h
e e
dges
o
f
whi
t
e
regi
ons
,
and
s
o
t
h
e
p
r
ac
t
i
cal
ups
hot
i
s
t
h
at
f
o
re
g
r
o
u
n
d
re
gi
o
n
s
gr
o
w
(
a
n
d
h
o
l
es in
side a
reg
i
o
n
shrink
). Dilatio
n is the du
al
o
f
ero
s
i
o
n i.e. d
ilatin
g fo
reg
r
ou
nd
p
i
x
e
ls is equ
i
v
a
l
e
n
t
to
ero
d
i
n
g
t
h
e ba
ckg
r
ou
n
d
pi
xel
s
4.
2.
4 Rem
ovi
n
g
of
Sm
al
l
ob
j
ect
Th
e d
ilated
imag
e con
t
ain
s
some s
m
a
ll o
b
j
ect, so
we rem
o
v
e
all s
m
a
ll o
b
j
ect u
s
in
g
B
W
AREAOPEN
o
p
e
ration
,
th
is o
p
e
ration
will
rem
o
v
e
all s
m
all p
i
x
e
l o
b
j
ect an
d
it re
m
o
v
e
s
m
a
ll p
i
x
e
l
o
b
j
ect b
a
sed
on
u
s
er
need
an
d
he
re i
n
ou
r
wo
r
k
we
rem
ovi
ng
al
l
s
m
al
l
object
w
h
ose si
ze
l
e
ss t
h
an
8
0
pi
xel
4.
2.
5 Reco
nstr
ucti
o
n
of
i
m
a
g
e
The dilated i
m
age contain som
e
br
eaki
ng bo
r
d
er
s
o
In
t
h
i
s
we
re
constructs the
character
by
el
im
i
n
at
i
ng i
t
s
brea
ki
n
g
b
o
r
der
,
f
o
r r
eco
n
s
t
r
uct
i
o
n i
t
us
e t
h
i
s
bl
oc
k f
o
r
reco
nst
r
uct
bo
rde
r
a
nd
I
M
FILL
ope
rat
i
o
n f
o
r fi
l
l
hol
es i
n
i
m
ages
4.
2.
6 T
h
i
nni
n
g
Thi
n
ni
n
g
p
r
oc
ess rem
oves sel
ect
ed part
s of f
o
re
g
r
o
u
n
d
pi
xel
s
of a b
i
nary
im
age. The t
h
i
n
ni
n
g
o
p
e
ration
is related
to
th
e h
i
t-and
-
m
i
ss tran
sfo
r
m
an
d
can
b
e
ex
pressed
qu
ite si
m
p
ly
in
term
s o
f
it
. Th
e
thinni
ng of a
n
i
m
age I
by a st
ructuring elem
e
n
t J is
thin(
I
,J
)
= I – h
it
and
m
i
ss(I,J
)
(1
)
4.
3
T
e
mpl
a
te
Ma
tchi
n
g
Met
h
od
Tem
p
late
matc
h
i
ng
is o
n
e
of th
e
Ch
aracter Reco
gn
itio
n tech
n
i
q
u
e
s.
It is th
e pro
c
ess
o
f
find
ing
the
lo
catio
n
o
f
a su
b
im
ag
e calle
d
a te
m
p
late i
n
sid
e
an
im
ag
e. Once a num
b
e
r o
f
corresp
ond
ing
tem
p
l
a
tes is
fo
u
n
d
,
t
h
ei
r
c
e
nt
res a
r
e
use
d
as c
o
r
r
es
p
o
ndi
ng
p
o
i
n
t
s
t
o
d
e
term
in
e th
e reg
i
stration p
a
ram
e
ters. Te
m
p
lat
e
match
i
n
g
inv
o
lv
es d
e
term
in
in
g
sim
ilarities
b
e
tween
a g
i
v
e
n
tem
p
late an
d
windo
ws
of th
e sam
e
siz
e
in
an
im
age and i
d
e
n
t
i
f
y
i
ng t
h
e w
i
nd
ow t
h
at
p
r
od
uces t
h
e hi
ghe
st
sim
i
l
a
ri
ty
m
easure.
It wo
rk
s by
com
p
ari
n
g
deri
ved im
age feature
s
of t
h
e
im
age and the
te
m
p
late
for ea
ch
possible
dis
p
lacem
ent of t
h
e tem
p
late.
Thi
s
pr
ocess
i
n
vol
ves
t
h
e use
of
a dat
a
base
o
f
c
h
aract
ers
o
r
te
m
p
lates. Th
ere ex
ists a temp
late fo
r all
pos
sible input
characte
r
s. For recogniti
on to occur, t
h
e current input cha
r
act
er is com
p
ared t
o
each te
m
p
la
te
to
find
eith
er an
ex
act m
a
tch
,
o
r
t
h
e tem
p
late
with
th
e
closest represen
tation
of th
e i
n
put c
h
aracter. If I(x, y) is
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Temp
la
te Ma
tch
i
ng
Met
h
o
d
f
o
r Recogn
itio
n
o
f
S
t
o
n
e
In
scri
p
t
ed
K
a
n
nad
a
Ch
a
r
a
c
ters
o
f
… (Ra
jith
kumar B K)
72
3
th
e in
pu
t ch
aracter, TN(x, y) is th
e te
m
p
l
a
te n
,
th
en
th
e
m
a
tch
i
n
g
functio
n
s (I, TN) will retu
rn
a v
a
l
u
e
in
d
i
cating
how well te
m
p
late n
m
a
tch
e
s th
e in
pu
t ch
aracter. So
m
e
o
f
t
h
e
m
o
re co
mm
o
n
m
a
tch
i
n
g
fun
c
tio
n
s
are
based on t
h
e followi
ng
Form
ulas
S
I.
T
n
|
i,
j
T
n
i,
j
|
(2
)
S
I.
T
n
|
i,
j
T
n
i,
j
|
2
(3
)
S
I.
T
n
|
i,
j
Tn
i,
j
|
(4
)
S
I.
T
n
∑∑
|
i,
j
|
I
|
T
n
i,
j
T
n
|
|
∑∑
|
i,
j
|
I
|
2
Tn
i,
j
T
n
2|
(5
)
M
a
t
c
hi
ng a
p
p
r
oac
h
es:
(
2
)
C
i
t
y
bl
ock, (
3
) E
u
cl
i
d
ea
n di
st
ance,
(4
)
C
r
oss C
o
r
r
el
at
i
on,
(5
) 2
-
D
Norm
alized
Co
rrelatio
n.
4.3.1 Cros
s-c
o
rrelation
Cross
-
correlation is a m
easure of sim
ilarity
of two
wav
e
form
s as a fun
c
ti
o
n
of a ti
m
e
-lag
ap
p
lied to
one
o
f
t
h
em
. Thi
s
i
s
al
so
kn
ow
n as a sl
i
d
i
ng
d
o
t
pr
o
d
u
c
t
or sl
i
d
i
ng i
n
n
e
r-
pr
o
duct
.
It
i
s
com
m
onl
y
used f
o
r
searchi
n
g a long signal for a s
h
orter,
kno
wn
featu
r
e. It
h
a
s
ap
p
lication
s
in p
a
ttern recog
n
itio
n
,
sing
le
p
a
rticle
an
alysis, electro
n to
m
o
g
r
aph
i
c, av
er
ag
ing
,
cr
yp
tan
a
lysis, an
d n
e
ur
oph
ysio
log
y
.
For
c
ont
i
n
u
ous
f
unct
i
o
ns
f a
n
d
g, t
h
e c
r
os
s-c
o
r
r
el
at
i
on i
s
de
fi
ne
d as:
f
∗g
τ
f
t
g
t
τ
dt
(6
)
Whe
r
e
f*
de
n
o
t
es t
h
e c
o
m
p
l
e
x c
o
n
j
ugat
e
o
f
f a
nd i
s
t
h
e
ti
me lag
.
Sim
ila
rly, fo
r
d
i
screte fu
n
c
tion
s
, the cro
ss-
correlation is define
d as:
f
∗g
n
f
m
g
m
n
(7
)
4.
3.
2 Im
pl
eme
nta
t
i
o
n
of
K
a
nna
da
C
h
ar
ac
ter
Reco
gni
t
i
o
n
Th
e im
p
l
e
m
en
tatio
n
o
f
Kann
ad
a ch
aracter recog
n
ition
is d
o
n
e
b
y
firstly refin
i
n
g
the ex
tracted
ch
aracters to
fi
t th
e
m
in
to
a
wind
ow withou
t wh
ite sp
aces o
n
all th
e fou
r
sid
e
s and
creatin
g
th
e temp
late for
each extracte
d
character. T
h
e te
m
p
la
tes
are norm
alize
d
to 42x24 pixe
ls and stored in the dat
a
base.
No
rm
al
i
z
at
i
on
i
s
do
ne
usi
n
g
wi
n
d
o
w
t
o
vi
e
w
po
rt
t
r
a
n
sf
or
m
a
t
i
on. T
h
i
s
m
a
ppi
ng
i
s
us
ed
t
o
m
a
p every
p
i
xel
o
f
th
e orig
inal i
m
ag
e to
t
h
e co
rrespo
n
d
i
ng
p
i
xel in
th
e
n
o
rmalized im
age.
The e
x
tract
ed
characte
r
of the input
test i
m
ag
e, after norm
a
l
i
zati
o
n, is m
a
tch
e
d
with
all
the
cha
r
acters i
n
the data
ba
se usin
g
2-
D
no
rmalized
co
rrelatio
n
co
efficien
ts app
r
oach
to id
en
tify si
m
ilar p
a
tterns
betwee
n a t
e
st im
age and
the standard
database
i
m
ag
es. Th
is ap
pro
ach is sh
ow
n in
Equ
a
tion 4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 5
,
O
c
tob
e
r
20
14
:
719
–
7
29
72
4
5.
R
E
SU
LTS AN
D ANA
LY
SIS
5.
1 Resul
t
s
Step
1
:
Cr
eate
a Tem
p
late o
f
K
a
nn
ad
a ch
ar
acter
s
and
each
i
m
ag
e in
a temp
late is in
size o
f
24x42
di
m
e
nsi
ons as
sho
w
n i
n
Fi
gu
r
e
3.
Figure
3. Ka
nnada c
h
aracte
r
s
Tem
p
late images
St
ep 2:
C
a
pt
ur
e St
one i
n
s
c
ri
pt
i
o
n
s
Ka
nna
d
a
charact
ers i
m
ages usi
n
g o
r
di
nary
di
gi
t
a
l
cam
e
ra of 1
6
M
e
ga
Pix
e
ls r
e
so
l
u
tio
n, as show
n in
Fi
g
u
r
e
4
Fi
gu
re 4.
C
a
pt
ure
d
Ka
nna
da St
one
I
n
sc
ri
pt
i
ons
C
h
a
r
act
ers
Step
3: Im
age
Mosaic ba
sed
on SIFT al
gori
thm
Thi
s
i
s
t
h
e
ver
y
im
port
a
nt
st
ep i
n
o
u
r
pr
o
p
o
s
al
and
we ca
n’t capt
u
re
d all
Characters
in a
n
st
one
in-
scrip
t
ed b
y
si
ng
le im
ag
e with
h
i
gh
reso
l
u
tio
n, for to
m
a
i
n
tain
g
ood
reso
lu
tion
an
d to
av
o
i
d
o
v
e
rlappin
g
o
f
charact
e
r
s
bet
w
een
t
w
o i
m
ages i
n
M
o
sai
c
,
here
we
use
n
e
w M
o
sai
c
t
echni
que
cal
l
e
d
‘
I
m
a
ge M
o
sai
c
B
a
se
d
o
n
SIFT al
g
o
ri
th
m
’
, th
e m
a
in
adv
a
n
t
ag
es
o
f
th
is algo
rith
m
is it will
m
o
saic two im
ag
es
with
g
ood
resolu
tio
n
and it elim
inate overla
ppi
ng
of c
h
a
r
acters
between t
w
o images a
n
d it pro
duce
o
u
t
p
ut
m
o
sai
c
ki
ng
i
m
ag
e l
i
k
e
naturally captured im
age. Thi
s
Mosaic im
age was hel
p
ed
fu
ture step
s to
ex
tract
all Kannada cha
r
acters
easily.
The st
e
p
s
d
u
ri
ng
M
o
sai
c
bas
e
d
on
S
I
FT
Al
go
ri
t
h
m
resul
t
s
are s
h
ow
n i
n
f
i
gu
re
5.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Temp
la
te Ma
tch
i
ng
Met
h
o
d
f
o
r Recogn
itio
n
o
f
S
t
o
n
e
In
scri
p
t
ed
K
a
n
nad
a
Ch
a
r
a
c
ters
o
f
… (Ra
jith
kumar B K)
72
5
Fi
gu
re 5.
Im
age
M
o
sai
c
base
d on
SI
FT.
Step
4: Pre-preprocessi
ng
In this ste
p
we
extract the
eac
h
Kanna
d
a C
h
aract
ers a
n
d st
eps i
n
volve
d
a
r
e shown i
n
Figure
6
Figure
6. Im
age Processing
and E
x
traction
of Each Ka
nna
d
a C
h
aracter
In
th
is
we rem
o
v
i
n
g
no
ise con
t
en
t in
an
captu
red
im
ag
es u
s
in
g
Gaussian
filter an
d
we R
e
sizin
g
o
f
all
pre
-
p
r
ocesse
d i
m
ages i
n
t
o
fi
x
e
d pi
xel
si
ze and
di
m
e
nsi
on a
nd
we Fi
n
d
i
n
g
Ed
ge i
n
an i
m
age usi
ng
So
bel
edge
d
e
tectio
n
and
th
en
we Perfo
r
m d
ilatio
n
an
d Use top
-
h
a
t fi
lterin
g
to
co
rrect u
n
e
v
e
n
illumin
a
tio
n
.
We re
m
o
ve
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
JECE Vo
l. 4
,
N
o
. 5
,
O
c
tob
e
r
20
14
:
719
–
7
29
72
6
al
l
ob
ject
s i
n
t
h
e i
m
age cont
a
i
ni
ng
fe
wer t
h
a
n
3
0
pi
xel
s
a
n
d we
rec
o
n
s
t
r
u
c
t
i
on
of i
m
age by
rec
onst
r
uct
i
ng i
t
s
b
oun
d
a
ry an
d
fillin
g
its h
o
l
es, fin
a
lly b
y
calc
u
latin
g
its co
nn
ected
co
m
p
onen
t
s we ex
tract
th
e ch
aracter in
an
im
age.
St
ep 5.C
r
oss
c
o
r
r
el
at
i
on
In this
we perform
cross correl
atio
n
b
e
tween
pre-p
r
o
cessed
im
ag
e with
te
m
p
late. Ba
sed
on
th
ei
r
result value we
rec
o
gnize
the
charact
er
s as
sh
own
in Figur
e 7
,
8,
9
.
Fi
gu
re
7.
R
ecg
oni
zat
i
o
n
of
K
a
nna
da
st
o
n
e i
n
sci
p
t
i
o
ns c
h
a
r
act
er G
h
a.
Fi
gu
re
8.
R
ecg
oni
zat
i
o
n
of
K
a
nna
da
st
o
n
e i
n
sci
p
t
i
o
ns c
h
a
r
act
er R
a
.
Fi
gu
re
9.
R
ecg
oni
zat
i
o
n
of
K
a
nna
da
st
o
n
e i
n
sci
p
t
i
o
ns c
h
a
r
act
er Ka
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
Temp
la
te Ma
tch
i
ng
Met
h
od
f
o
r Recogn
itio
n
o
f
S
t
o
n
e
In
scri
p
t
ed
K
a
n
nad
a
Ch
a
r
a
c
ters
o
f
… (Ra
jith
kumar B K)
72
7
5.
2. E
x
peri
me
ntal
resul
ts
Exp
e
r
i
m
e
n
t
s hav
e
b
e
en
p
e
rf
or
m
e
d
to
test the pr
opo
sed m
e
th
od
. M
A
TLAB (
R
200
9
a
)
is
th
e so
f
t
w
a
re
tool that was used for Recognition
of Ka
nnada Characte
r
s
.
The Experi
m
e
nt
s were
per
f
o
rm
ed capt
u
re
d m
a
ny
st
one i
n
scri
pt
i
ons
cha
r
act
ers
of
Ho
ys
al
a
and
Ga
ng
a
t
i
m
e
s fram
e
s and T
a
bl
e 1 a
n
d Ta
b
l
e 2 gi
ves t
h
e
r
e
sul
t
s
of Recognition rate
betwee
n t
h
e c
h
aracte
r
i
m
ages and t
h
eir Tem
p
lates images.
Th
e
Reco
gn
itio
n
rate
can
b
e
o
b
t
ain
e
d
b
y
form
u
l
a
(8
)
Whe
r
e
= Recog
n
ition
rate,
=
s
u
m o
f
co
rr
ec
t
ma
tc
h
,
= sum
of incorrect
m
a
tch,
= Num
b
er of test
sam
p
les
Tab
l
e
1
.
The
Reco
gn
itio
n
rate an
alysis of
Kann
ad
a st
o
n
e
in
scri
p
tio
n
c
h
aracters for Hoysala
tim
e
frames.
Test i
m
ages
The Recognition r
a
te analysis of Kannada H
o
ysala Period stone inscription characters
Recognized Character
%
JA 18
2
20
90%
H 19
1
20
95%
M
A
18
2
20
90%
KA 19
1
20
95%
GA 18
2
20
90%
YA 19
1
20
95%
NA 18
2
20
90%
SHA 19
1
20
95%
KHA 19
1
20
95%
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 5
,
O
c
tob
e
r
20
14
:
719
–
7
29
72
8
Tab
l
e
2
.
Th
e Reco
gn
itio
n rate an
alysis of
Kan
n
a
d
a
ston
e inscrip
tion
c
h
ara
c
ters f
o
r
Ga
n
g
a
tim
e
fram
e
s.
Test i
m
ages
The Recognition r
a
te analysis of Kannada
Ganga Period stone inscription characters
Recognized
Character
%
JHA 18
2
20
90%
YA 18
2
20
90%
M
A
18
2
20
90%
DA 19
1
20
95%
BU 19
1
20
95%
SE 19
1
20
95%
E
18
2
20
90%
VA 18
2
20
90%
From
Tabl
e 1 and Tabl
e 2
,
i
t
i
s
show
n t
h
at
our p
r
op
osa
l
recog
n
i
t
i
on r
a
t
e
shoul
d be
92.7%
accuracy
in
r
ecogn
itio
n
of
st
one i
n
sc
ri
pt
i
ons c
h
aract
er
s
of Hoysala time fram
e
s
and
9
1
.
87%
accuracy for
Ganga time
fram
e
s.
6.
CO
NCL
USI
O
N
A sim
p
le an
d effectiv
e Reco
gn
itio
n m
e
th
o
d
fo
r id
en
tifi
catio
n
o
f
d
i
fferen
t
tim
e frames Kann
ad
a
st
one i
n
sc
ri
pt
i
ons c
h
aract
e
r
s
were i
n
t
r
od
uc
ed i
n
t
h
i
s
pap
e
r. Thi
s
p
r
ovi
des a go
o
d
t
o
ol
fo
r t
h
e pe
o
p
l
e
fo
r
id
en
tificatio
n
o
f
t
h
e ston
e i
n
scrip
tion
s
.
It also
h
e
lp
s a
co
m
m
o
n
m
a
n
k
nowing
th
e
presen
t Kann
ad
a literatu
re
to
read
th
e an
cien
t literatu
re
Here
we are
u
s
ed
o
r
d
i
n
a
ry m
o
b
ile cam
era o
f
1
6
Meg
a
p
i
x
e
l
reso
l
u
tio
n camera to
cap
ture ch
aracters so
it is v
e
ry easy to
i
m
p
l
e
m
en
t an
d
it i
s
co
stless wh
en
co
m
p
are to
o
t
h
e
r m
e
th
o
d
s. Fo
r
recognition process, the e
x
tract
ed characte
r
was com
p
ared
to each te
m
p
la
te in the database to find the c
l
osest
represe
n
tation
of t
h
e input c
h
aracter. T
h
e m
a
tching m
e
tric
was co
m
p
u
t
ed
using
2-D correlatio
n co
effi
cien
ts
ap
pro
ach
t
o
iden
tify si
m
ilar
p
a
ttern
s
b
e
tween
th
e test im
a
g
e and
th
e d
a
t
a
b
a
se im
ag
es. Exp
e
rim
e
n
t
al resu
lts
sho
w
t
h
at
t
h
e
p
r
o
p
o
sed
m
e
t
hod i
s
e
ffi
ci
ent
f
o
r i
d
e
n
t
i
f
i
cat
i
o
n
Ka
nna
da
st
o
n
e
i
n
scri
pt
i
o
ns c
h
aract
er
s.
REFERE
NC
ES
[1]
Moham
m
e
d Ali
Qatran. T
e
m
p
late m
a
tch
i
ng m
e
thod for
recog
n
ition Musnad charact
e
rs based on correlati
o
n
analy
s
is. Depar
tment of Computer
Sc
ie
nce
,
Amra
n Uni
v
e
r
sity
, Ye
me
n.
[2]
Dr HS Mohana,
et al
. Extractio
n of Stone In-scripted Kannada
Charact
ers
Us
i
ng S
i
ft Algorithm
Bas
e
d Im
age
Mosa
ic
.
In
ternational Journal
of Electronics
&
C
o
mmunication T
echnolog
y, IJEC
T
. 2014; 5(2
)
.
[3]
Dr HS
Mohana,
et al.
Era Ident
i
f
i
cation and
Recognition of Ston
e
In-scripted Ka
nnada Char
acters Using Artifici
a
l
Neural Networks
. 2nd
Nation
a
l Conference on
Inn
ovation
in Comp
uting
and Communication
Techn
o
log
y
. 2014
.
[4]
M
aalin
ee R
a
m
u
.
Printed Number
Recognition using Matlab
.
UNIVERSITY TEK
NOLOGI,
MALAYSIA.
[5]
Ahm
a
d Abdulkader Matthew R
Casey
.
Low Co
st Correcti
on of
OCR Errors Using Learning
i
n
a Multi-Engin
e
Environm
ent”
. I
EEE10
.
1109/IC
DAR.2009.242.
Evaluation Warning : The document was created with Spire.PDF for Python.