TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 13, No. 1, Janua
ry 201
5, pp. 166 ~
173
DOI: 10.115
9
1
/telkomni
ka.
v
13i1.678
2
166
Re
cei
v
ed Au
gust 2, 201
4; Re
vised Sept
em
ber
18, 20
14; Accepted
Octob
e
r 16, 2
014
Yoruba Language and Numerals’ Offline Interpreter
Using Morphological and Template Matching
Olakanmi O. Olada
y
o
Electrical and Electron
ic
Engi
neer
ing,
T
e
chnology
Dr
ive, Office 6, N
e
w
Fac
u
lt
y
of Engi
neer
ing Building.
Univers
i
t
y
of Ibada
n, Ibada
n Niger
ia
email: o
l
aka
n
m
i
.ola
da
yo
@ui.
e
du.ng
A
b
st
r
a
ct
Yorub
a
as a l
ang
ua
ge has
passe
d throu
g
h
gen
erati
on r
e
formatio
n
s makin
g
so
me of
the old
docu
m
ents in t
he arch
ive to b
e
unre
a
d
abl
e by the pres
ent
readers. Ap
art from this, so
me Yorub
a
w
r
iters
usua
lly
mixe
d Engl
ish nu
mer
a
ls w
h
ile w
r
itin
g due to
br
evit
y and conc
ise
n
e
ss of Englis
h
nu
mer
a
ls co
mpare
to Yorub
a
n
u
m
er
als w
h
ic
h
are co
mbin
atio
n of sev
e
ra
l c
haracters. R
e
-typin
g suc
h
his
t
orical
docu
m
e
n
t
s
may
be ti
me c
onsu
m
ing, ther
efore a n
eed f
o
r an effici
e
n
t Optical Ch
arac
ter Read
er (OCR) w
h
ich w
ill
not
only effectivel
y recogn
i
z
e
Y
o
rub
a
texts bu
t also conv
erts all the En
glis
h nu
mer
a
ls in
the docu
m
ent
to
Yoruba num
er
als.Sever
a
l Optical
Char
acter
Reader
(O
CR) system
s
had been
dev
eloped to recogni
z
e
characters
or
texts of s
o
me
la
ngu
ag
es suc
h
as En
g
lis
h,
Ara
b
ic,
Ja
pa
nese,
Chin
ese, an
d Korea
n
,
h
o
w
e
ver
,
desp
i
te the
sig
n
ifica
n
t contri
b
u
tion
of Yor
u
b
a
la
ng
uag
e to
historic
al d
o
cu
me
ntatio
n a
nd
communic
a
tio
n
,
it
was observ
e
d that there is
no
partic
u
lar
OCR system
f
o
r the language. In this
paper corr
elation
and
temp
late
matc
hin
g
tech
niq
u
e
s
w
e
re use
d
to
deve
l
op
an
OC
R for the r
e
co
g
n
itio
n of Y
o
rub
a
bas
ed t
e
xts a
n
d
convert En
glis
h nu
merals
in
the doc
u
m
e
n
t to Yorub
a
n
u
m
erals. Exp
e
ri
mental r
e
sults s
how
the rel
a
tiv
e
l
y
hig
h
accuracy
of the deve
l
op
ed OCR w
hen i
t
w
a
s
test
ed on
all si
z
e
Yor
u
b
a
alp
h
a
bets an
d nu
mer
a
ls.
Ke
y
w
ords
: OCR, Yoruba, pattern
recog
n
iti
on, imag
e,
template
m
a
tching
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
Yorub
a
is the mo
st d
o
c
ume
n
ted
West Af
ri
can l
angu
age. Yo
ruba
is sp
o
k
en
by
18,850,0
00 p
eople i
n
Nige
ria. Th
e total
popul
ation
o
f
native sp
ea
kers i
n
all
co
untrie
s
i
s
ab
out
20,000,0
00.
The n
u
mbe
r
rise
s to 22,0
0
0
,000 if we
al
so in
clu
de
se
con
d
-la
ngu
ag
e sp
ea
kers. T
he
langu
age ha
s nume
r
o
u
s dialect
s
sp
oke
n
in different area
s o
f
Nigeria. Wi
thin Nige
ria
the
langu
age i
s
spoken in th
e
area
s of
Oyo, Ogun,
Ondo
Osu
n
, Kwa
r
a,
Lago
s a
nd t
he weste
r
n
p
a
r
t
of Kogi State
.
It is al
so
sp
oke
n
in
Beni
n,
Tog
o
, an
d
by immig
r
a
n
t
s in
the
Unit
ed King
dom
and
the USA. Yoruba i
s
o
ne of
the 12
Ede
k
iri lan
guag
es of the Yo
rub
o
id g
r
ou
p tha
t
also i
n
cl
ude
s
Igala. The Y
o
rub
o
id g
r
ou
p belon
gs to
t
he Defoid l
angu
age
s of
the Benue
-Congo g
r
o
up
and
ultimately to the Volta-Co
ngo, and Atl
antic-Co
ngo
grou
ps of the
Niger-Con
go
Family of 1419
langu
age
s m
o
stly spo
k
e
n
in Central and
South Africa
[6].
Image re
cog
n
ition is the process of ide
n
tify
ing and detecting an o
b
ject or a fea
t
ure in a
digital image
or video. This con
c
ept is used in
m
any applications li
ke syst
ems for fa
ctory
automation, toll
booth monitori
ng, and se
cu
rity
su
rveillan
c
e
.
Typical im
age
re
cog
n
ition
algorith
m
s in
clud
e:
1) Optical
cha
r
a
c
ter
recogniti
on
2)
Pattern and g
r
adie
n
t match
i
ng
3)
Face
re
cog
n
it
ion
4)
Lice
nse plate
matching
5) Scene
cha
n
g
e
detectio
n
It has be
co
me a tren
d to docume
n
t most
of the
document
s i
n
the archiv
es u
s
ing
scann
er, ho
wever, these d
o
cum
ents
ca
nnot be e
d
it
e
d
or read the
r
eafter
by co
mputer
syste
m
s.
Due to
the fa
ct that sca
n
n
e
r
scan
s do
cuments
a
s
a
n
imag
e not
as
encode
d
set of ch
aract
e
rs.
Optical
Ch
aracter Read
er (O
CR)
syste
m
doe
s el
ect
r
oni
c tra
n
sl
ati
on of h
and
written or p
r
inte
d
text into machine e
n
code
d text. OCR is
widely
used to
conve
r
t boo
ks and
document
s i
n
to
electroni
c file
s a
nd to
com
puteri
z
e
a
re
cord
ke
epin
g
system in
an
o
ffice. OCR m
a
ke
s it
po
ssi
b
l
e
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Yorub
a
Lan
g
uage a
nd Nu
m
e
rals
’
Offline Interpreter
Usi
ng… (Ola
kanm
i O. Ola
dayo)
167
to edit such
document,
se
arch fo
r a
wo
rd o
r
p
h
ra
se,
store it mo
re
com
p
a
c
tly, display o
r
p
r
int
a
copy an
d app
ly techniqu
es
su
ch a
s
machine tr
an
slati
on, text-to-sp
eech and text mining to it.
Optical Cha
r
acter Re
co
gn
ition
study
was star
ted by
Tyurin a Ru
ssi
an scie
ntist [1]. The
first mo
de
rn
chara
c
te
r reco
gnizers a
ppe
ared
in
th
e mi
ddle
of the
19
40s with
the
developm
ent
of
the digital
co
mputer. T
h
e
early
wo
rk
on the
aut
o
m
atic
re
cogn
ition of characters
ha
s b
e
e
n
con
c
e
n
trated
either
up
on well printe
d
t
e
xt
or
u
pon
small set of well disting
u
i
s
he
d
h
and
wri
tten
text or symb
o
l
s, althou
gh,
su
cc
essful
bu
t had b
een
im
plemente
d
m
o
stly for
Latin
ch
ara
c
te
rs
a
n
d
nume
r
al
s.
Beside
s some studies on Ja
pane
se, Ch
in
ese,
Heb
r
e
w
, Indian a
nd
Arabi
c chara
des
and num
eral
s in both pri
n
ted and ha
ndwritten ca
se
s we
re al
so con
s
ide
r
e
d
by some O
C
R
system
s. Th
e develo
p
me
nts in O
C
R
until 198
0s
suffered from
lack of adva
n
ce
d alg
o
rith
m,
powerful com
puting
h
a
rd
ware and optical
devices.
With the out
ward explo
s
ion
o
n
the
comp
uting
techn
o
logy d
e
velopme
n
t, the previou
s
ly prop
osed m
e
thodol
ogie
s
found a fe
rtile
environm
ent
for
rapid
g
r
o
w
th i
n
ma
ny ap
pli
c
ation
a
r
ea
s.
Pres
ently, re
newed vig
ours a
r
e
bei
ng
p
u
t in the
o
p
tical
cha
r
a
c
ter
re
cog
n
ition research. O
ne
of these
is
recognitio
n
of printe
d a
nd ha
nd
written
document
s.
More
sophi
st
icated
alg
o
rit
h
ms wh
ich u
t
ilize a
d
van
c
ed m
e
thod
ologie
s
a
r
e
be
ing
develop
ed.
In this work two method
o
l
ogie
s
are comb
ine
d
to achi
eve an efficient Yoru
ba OCR
system
whi
c
h
will be able
to recogni
ze
off-
line typed
and han
dwri
tten Yoruba
document
s a
n
d
conve
r
t Engli
s
h
nume
r
al
s t
o
Yoruba
nu
meral
s
. T
he
remainin
g p
a
rt
of this p
ape
r is
arran
ged
as
follows: se
cti
on 2 i
s
th
e review of
rela
ted wo
rks on
OCR
syste
m
s a
nd m
e
thodol
ogie
s
.
The
desi
gn meth
o
dology an
d worki
ng p
r
in
cip
l
e of the
syst
em are
expla
i
ned in sectio
n 3. Section
4
contai
ns the t
e
st re
sult
s an
d con
c
lu
sio
n
.
2. Related Works
Referen
c
e [8
] described
a complete
system
for t
he
recognitio
n
of u
n
con
s
trained
hand
written
Arabi
c words usin
g ove
r
-segmentatio
n
of cha
r
a
c
ters and va
riabl
e
duration hi
d
den
Markov mod
e
l
(VDHMM). I
n
this, a
seg
m
entation
al
g
o
rithm
wa
s u
s
ed to tran
sla
t
e the 2-D im
age
into
1-D se
q
uen
ce
of su
b
-
ch
aracte
r symbols.
Thi
s
seq
uen
ce of symbol
s wa
s
model
ed by
the
VDHMM.
Th
e shap
e inf
o
rmatio
n of
cha
r
a
c
te
r an
d sub-ch
ara
c
ter sym
bol
s
wa
s
comp
actly
rep
r
e
s
ente
d
by forty-five feature
s
i
n
th
e feat
ure spa
c
e. Th
e featu
r
e ve
ctor
wa
s model
ed a
s
an
indep
ende
ntly
distrib
u
ted multivariate discrete
di
stri
bution. And
the vari
able
d
u
ration
state
i
s
use
d
to resol
v
e the segme
n
tation ambi
g
u
ity among the con
s
e
c
utive
characte
rs.
Different
met
hodol
ogie
s
o
n
ho
w th
e q
uality of the
captu
r
ed
ca
mera
imag
e
coul
d b
e
improve
d
ha
d been th
oro
ughly co
nsi
d
ered
by vario
u
s resea
r
che
s
. For
examp
l
e, referen
c
e
[2
]
analyzed the
quality of such
ca
pture
d
image fo
r o
p
tical
cha
r
a
c
ter recognitio
n
. In their
work
different m
e
a
n
s
of imp
r
ovi
ng tran
scripti
on a
nd
re
co
g
n
ition
wa
s p
r
opo
sed. Al
so
, referen
c
e
[1
8
]
prop
osed a new
pe
rspe
ctive
re
ct
ificati
on
system
b
a
se
d o
n
van
i
shin
g p
o
int
detectio
n
. Th
eir
system a
c
hi
e
v
ed both the
desi
r
ed
effici
ency an
d a
c
cura
cy usi
ng
a multi-sta
g
e
strategy: at the
first stage, do
cume
nt
bo
un
darie
s and
st
raight
lin
es
are u
s
ed
to
co
mpute vani
sh
ing p
o
ints;
at
the
second stage, text baselines a
nd block aligns
are utilized; and at
the last stage, character t
i
l
t
orientatio
ns
are vote
d fo
r the ve
rtical
vanish
i
ng
p
o
int. A profit
functio
n
wa
s introdu
ce
d
to
evaluate the
reliability of
detecte
d vani
shing point
s
at each stag
e. If vanishing point
s at one
stage
are rel
i
able, then
rectificatio
n i
s
end
ed at
th
at stag
e. Ot
herwise, mult
i-stag
e
strate
gy
method conti
nue
s to obtai
n more
reliabl
e
vanishi
ng p
o
ints in the n
e
xt stage.
Re
sea
r
ch h
a
s
sho
w
n th
at Cha
r
a
c
ter d
egra
dation
af
fects m
a
chin
e pri
n
ted
ch
ara
c
ter
recognitio
n
. Two mai
n
re
aso
n
s fo
r de
grad
ati
on
we
re extrin
sic i
m
age d
egra
dation such
as
blurring an
d
low image
dimen
s
ion, a
nd intrin
si
c
degradatio
n cau
s
e
d
by font variation
s
.
A
recognitio
n
method that co
mbine
s
two complem
entary classifie
r
s is pro
p
o
s
ed in
referen
c
e [1
7].
The lo
cal fe
ature
ba
sed
cla
ssifie
r
extra
c
t
s
the lo
cal co
ntour dire
ctio
n
ch
ang
es, which
i
s
effe
ctive
f
o
r
cha
r
a
c
t
e
r
pat
t
e
r
n
s
wit
h
le
ss
st
ru
ct
ure
det
e
r
io
ra
tion. The
glo
bal featu
r
e
b
a
se
d
cla
ssifi
er
extract
s
the texture di
strib
u
tion of the chara
c
te
r ima
ge, whi
c
h i
s
effective whe
n
the ch
ara
c
t
e
r
stru
cture i
s
h
a
rd
to di
scrim
i
nate. Th
e two comp
le
men
t
ary cl
assifiers a
r
e
combin
ed by
ca
ndid
a
t
e
fusion
in
a
coa
r
se-to
-
fine
style. Expe
riment
s a
r
e
carrie
d o
n
d
egra
ded
Chi
nese
cha
r
a
c
ter
recognitio
n
.
Referen
c
e [1
3] wo
rked o
n
Ch
aracte
r
recognitio
n
system Telu
g
u
; one of th
e an
cient
langu
age
s of South India. It has a co
mpl
e
x orthog
ra
p
h
y with a larg
e numbe
r of distin
ct cha
r
a
c
ter
sha
p
e
s
com
p
ose
d
of simpl
e
and com
p
o
und ch
aracte
rs
. In this wo
rk, structu
r
al
feature
s
of the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 1, Janua
ry 2015 : 166 –
173
168
syllable a
nd
the com
pon
e
n
t model were com
b
ined t
o
extract mid
d
le zo
ne co
mpone
nts. T
h
e
sha
pe of the
middle
zon
e
comp
one
nts i
s
cl
osely rel
a
ted to a
circl
e
wh
ere
a
s ot
her
com
pon
e
n
ts
are foun
d wit
h
different top
o
logi
cal featu
r
es.
A simple a
nd effective
template matchin
g
m
e
thod for id
entification o
f
Musna
d
cha
r
a
c
ters
was introdu
ce
d
in refe
ren
c
e
[10[. T
he ch
ara
c
ters
were extracte
d from input im
a
g
e
and no
rmali
z
ed. Duri
ng re
cog
n
ition, the
extracted ch
ara
c
ter
wa
s compa
r
ed to e
a
ch templ
a
te in
the databa
se
to find the close
s
t re
pre
s
e
n
tation of
the input cha
r
a
c
t
e
r. The mat
c
hing metri
c
was
comp
uted u
s
i
ng 2
-
D
co
rrel
ation coeffici
ents a
p
p
r
oa
ch to identify
simila
r patterns b
e
twe
en t
h
e
test image an
d the databa
se image
s.
In refere
nce
[5], a novel approa
ch to
effici
ently re
cogni
ze ha
nd
written n
u
me
rals
wa
s
prop
osed. Th
is ap
pro
a
ch
exploits a two-sta
ge
fram
ewo
r
k
by usi
ng differe
nce
feature
s
. In the
first
stage,
a
reg
u
la
r SV
M is train
ed
on all
t
he
tra
i
ning data;
in
the se
con
d
stage, only
t
h
e
sampl
e
s mi
scl
assified i
n
the first
stage a
r
e
spe
c
ially con
s
id
ere
d
. The num
be
r of
miscl
assifications i
s
often
small because of
the good performance of
SVM. This will present
difficulties in
training
an a
c
curate SVM
engine
only for these miscla
ssifie
d
sa
mples.
We th
en
further propo
se
a multi-way to bina
ry
app
roa
c
h us
in
g d
i
ffe
r
enc
e
fea
t
u
r
es
. T
h
is a
p
p
r
oa
c
h
su
ccessfully transfo
rm
s m
u
lti-ca
te
go
ry cla
ssifi
cation
to binary classificatio
n
and expand
s
the
training sam
p
les
greatly.
2.1. O
v
er
v
i
e
w
o
f
Yo
ruba
Orhogr
aph
y
In its written f
o
rm, Yoruba
use
s
the
Ro
man alp
hab
et. It has 25 letters
as
sh
own
in fig. 2.
The letter 'p
' is always p
r
o
noun
ce
d as '
k
' an
d 'p'
co
mbined. Yo
ru
ba orth
ograp
hy does n
o
t use
the letters
c,
q, v, x, z. Yoruba
ha
s thre
e ba
si
c
ton
e
s, high, mid,
a
nd lo
w,
whi
c
h
are
indi
cate
d
in
the ortho
g
ra
p
h
y. The high
is marke
d
wit
h
an a
c
ut
e a
c
cent (e.g. á
)
, the low with
a grave a
ccent
(à), a
nd the
mid tone
usu
a
lly left unma
r
ke
d. Th
e
s
e
marks
are u
s
ually pla
c
ed
on the vo
wel
s
. In
some
circum
stan
ce
s the mid tone is i
ndicated
with
a 'macron'.
The lang
uag
e has b
een
written
sin
c
e th
e ea
rly 19
th
cent
ury, althou
gh
there
have
bee
n
ma
ny ch
ang
es i
n
aspe
cts
of
its
orthog
ra
phi
c rep
r
e
s
entatio
n. In the 196
0s, the t
hen
Ministry of E
ducati
on with
in the We
ste
r
n
Regi
on of Nig
e
ria, whi
c
h
was where mo
st of t
he Yoru
ba sp
ea
king
comm
unity is located, form
ed
two committe
es to
co
nsi
d
er a
stan
da
rd orth
ogr
aph
y for the la
n
guag
e. The
more i
n
fluent
ial of
these
two, t
he Yo
rub
a
Orthog
ra
phy
Committee
was
set
up in
196
6. The
report
whi
c
h
this
se
con
d
ortho
g
rap
h
y com
m
ittee submi
tted in 1966
becam
e the basis for th
e cre
a
tion a
nd
introdu
ction i
n
to school
s o
f
the standa
rd
Yoruba o
r
tho
g
rap
h
y [6].
Table 1. Engli
s
h num
erals
and their e
qui
valent Yorub
a
nume
r
al
s
English
1 2
3
4
5
Y
o
ruba
Eni Eji
Eta
Erin
Arun
English
6 7
8
9
10
0
Y
o
r
uba
Efa Eje
Ejo
Esan
Ew
a
O
d
o
.
Table 2. Yoru
ba upp
er an
d
lower al
pha
b
e
ts
2.2. Yoruba OCR Sy
stem Methodolo
g
y
OCR a
s
e
a
rl
ier
stated i
s
the scien
c
e t
hat ent
ail
s
th
e de
scription
or
cla
ssifi
ca
tion of
cha
r
a
c
ter m
e
asu
r
em
ents t
hat usu
a
lly base
d
on so
m
e
model
s. O
CR i
s
one of
the categ
o
rie
s
of
image
re
cog
n
i
tion. There i
s
variou
s
cha
r
acter re
co
gnit
i
on meth
od
s
use
d
in d
e
vel
oping
ch
ara
c
t
e
r
recogni
ze
r. These meth
ods a
r
e: ne
ural net
wo
rk,
moment ba
sed a
pproa
ch, contou
r b
a
se
d
approa
ch, template match
i
ng and mo
rp
hologi
cal a
p
p
r
oa
ch. In this work templat
e
matchin
g
a
nd
morp
holo
g
ica
l
techniqu
es
are u
s
ed to reco
gni
ze
Yoruba texts.Te
mplate match
i
ng refers to the
pro
c
e
ss of d
e
tecting a
n
object having
a certai
n si
ze, shap
e an
d orientatio
n
in an image
b
y
applying a
n
operator
con
t
aining po
siti
ve weight
s i
n
a regi
on resem
b
ling th
e obje
c
ts to
be
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Yorub
a
Lan
g
uage a
nd Nu
m
e
rals
’
Offline Interpreter
Usi
ng… (Ola
kanm
i O. Ola
dayo)
169
detecte
d and
containi
ng
negative wei
ghts in a
re
gi
on surrou
n
d
ing the p
o
sitive weight [15].
Morp
holo
g
y as de
rived from biology i
s
a bran
ch
of biology which d
eal
s with the form and
animal
s
a
nd
plants. It i
s
a
dopted
in thi
s
co
ntext as
a
tool for extra
c
ting im
age
compon
ents th
at
are u
s
eful in
the represe
n
tation and
descri
p
ti
on o
f
the region
shap
e. The
r
e are seve
ral
pro
c
ed
ural st
eps
eng
age
d in a
c
hievi
ng mo
rph
o
lo
gical te
ch
niq
ues. T
hese
inclu
de filteri
ng,
thinning, prun
ing, ero
s
ion a
nd dilation, o
penin
g
and
cl
osin
g.
3. Yoruba OCR Impleme
n
ta
tion usin
g Templa
te
Matchin
g
an
d Morpholog
ical Techniq
u
e
Template m
a
tchin
g
and
morph
o
logi
cal techniq
u
e
s a
s
state
d
earlie
r, a
r
e O
C
R
recognitio
n
techni
que
s. Th
ese al
gorith
m
s involve f
eature
s
extra
c
tio
n
and cl
assifi
er. In template
matchin
g
ima
ge pixel
s
a
r
e
use
d
a
s
the
feature
s
b
e
in
g extra
c
ted from both th
e i
nput cha
r
a
c
ter
and th
e
cla
s
sified characte
rs.
The
cl
assifier comp
ar
e
s
the
inp
u
t ch
ara
c
ter features
with
a
set
of
c
h
arac
ter template in the
c
h
arac
te
r
cla
ss.
I
n
t
h
i
s
co
nt
ex
t
t
he ch
a
r
acter
cla
s
s contain
s
num
e
r
als,
uppe
r and lo
wer
ca
se
s of Yorub
a
cha
r
a
c
ters as
sho
w
n in Figu
re 1 and Figu
re 2. The absol
ute
value of the classifier procedur
e whi
c
h i
s
the co
rrel
a
tion coeffici
ent
between the
input cha
r
a
c
t
e
r
and the
co
nsi
dere
d
cha
r
a
c
ter templ
a
te is use
d
to mo
rpholo
g
ically
d
e
termin
e the
template
with
a
clo
s
e
s
t correl
ation match.
Formally,
,
,
,
(
1
)
(
2
)
Whe
r
e:
The tran
sfo
r
mation functi
on
on charact
e
r
is
:
:
→
In the
cha
r
a
c
ter
cla
s
s som
e
of the
cha
r
acters we
re
written
in diff
erent
way
s
in
ord
e
r to
accomm
odat
e differe
nt wa
ys of
writing.
The p
r
o
p
o
s
e
d
Yoruba
O
C
R
system,
as sh
own in
fig
u
re
3, is g
r
oup
ed
into three
pro
c
e
ssi
ng level
s
whic
h are l
o
w level
pro
c
essing, inte
rmediate level
and
high level
proce
s
sing. T
h
ese
are impl
emented
u
s
i
ng 64
-bit M
a
tlab version
7.8.0.387 a
n
d
the
input texts are built with pa
int brush and
text.
3.1. Lo
w
L
e
v
e
l Processin
g
As
sho
w
n
in
the Fi
gure
3, low level
pro
c
e
ssi
ng i
n
volves ima
g
e
acqui
sition
and
pre
-
pro
c
e
ssi
ng of
the acqui
red
images. Ima
ge acq
u
isitio
n stage a
c
qui
res ima
ge of the docu
m
ent
or
cha
r
a
c
ters to
be
re
co
gni
zed. Mo
st tim
e
inp
u
t cha
r
acter ima
ge
is of
finite resol
u
tion
whi
c
h
ultimately affects the qu
ali
t
y of its transf
o
rmat
io
n, therefore, pre
-
proce
s
sing be
comes n
e
cessary.
The pre-pro
c
essing
stage
includ
es
col
our n
o
rm
aliz
ation, scaling
filtering and
thinning. Co
lou
r
norm
a
lization
is u
s
ed to
chang
e inp
u
t cha
r
a
c
ter fo
regro
und
col
o
ur to bl
ack a
nd ba
ckgrou
nd
colo
ur to
whit
e. To a
c
hiev
e this, hi
stog
ram techniqu
e
wa
s u
s
ed. T
he inp
u
t ch
aracter was
used
to form histog
ram of sin
g
le
cla
ss
whi
c
h
wa
s gro
upe
d into intervals.
Over ea
ch of
these interva
l
s
a vertical re
ctangle is dra
w
n with its a
r
ea propo
rt
io
nal to the number of point
falling into that
interval.
Th
e luminan
ce
of the
ima
ge wa
s
d
e
termi
ned
usi
ng equ
ation
3.
Figu
re
2a sho
w
s
in
p
u
t
image
befo
r
e
no
rmali
z
atio
n while fig
u
re 2b
an
d 2c depi
ct
the
in
put
ima
ge after normalizat
ion
and filtering
resp
ectively.
(
3
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 1, Janua
ry 2015 : 166 –
173
170
Normaliz
ation algorithm:
1)
Select the rel
e
vant part of the ch
ara
c
te
r.
2)
Determine th
e threshold fo
r the colo
ur n
o
rmali
z
atio
n
3)
Process the i
m
age from to
p corne
r
line by line
4)
Store the R,G
,
B value of each pixel
5) Determine
using e
quatio
n
1
6) If
< thre
sh
old
value then turn the
pixel black otherwi
se
white.
7)
Rep
eat for the whol
e input
image
The im
age
scalin
g
scale
s
the i
nput
ch
ara
c
ter ima
g
e
up
o
r
down de
pen
ding
on
the
origin
al si
ze. This was d
o
ne to redu
ce
the reco
gniti
on time and error rate as large character
image
s woul
d take lo
nge
r time to process whil
e
sm
all image ma
y be difficult to re
cog
n
ize. After
scaling the
chara
c
te
r be
comes bl
ocky and he
nce
the smooth
enin
g
filtering sta
ge rem
o
ves t
he
spi
k
e edg
es. This
sta
ge al
so co
ntain
s
smootheni
ng
fi
lter, lo
w p
a
ss filter. The
s
e
filters
are
u
s
e
d
to re
du
ce
blurring
and
n
o
ise. Also, i
m
ple
m
ented
in
the lo
w le
ve
l p
r
oc
es
s
i
n
g
is
the th
in
n
i
ng
w
h
ich
conve
r
ts a
n
y elong
ated pa
rts or
strip
s
in
the image
re
g
a
rdle
ss of the
i
r bits into n
a
rrow
stri
ps tha
t
are only ab
ou
t one pixel wi
de.
3.2. Interme
d
iate Lev
e
l Proces
sing
Intermedi
ate Level
Processing
(ILP
) in
the i
n
figu
re 3
involves imag
e
rotati
on a
n
d
segm
entation
.
Sometimes i
nput ch
aracte
r image
m
a
y not be prope
rly aligned in
angul
ar fa
shi
on
with re
spe
c
t to the cha
r
a
c
t
e
r template
set. An instan
ce of this will
be co
rre
cted
by realign th
e
image O
C
R. Segmentatio
n whi
c
h form
s the core
of IL pro
c
e
ssin
g
stage
parti
tions the in
p
u
t
image into its con
s
tituent chara
c
te
rs. Sh
own b
e
lo
w is
the algorith
m
use
d
for se
g
m
entation:
Segmenta
tio
n
algorithm:
1)
Scan the ima
ge from rig
h
t to left row wi
se
2)
Add and co
u
n
t all the x coordin
a
tes
3)
Determine th
e x-co
ordinat
e of the cent
roid u
s
in
g
∑
/
where
n is th
e
total
numbe
r of the centroid.
4)
Determine th
e y-coo
r
din
a
te of the centroid u
s
ing
∑
/
where n is the
total
numbe
r of the centroid.
3.3.
Repre
s
e
n
ta
tion and
Des
c
ription
Rep
r
e
s
entati
on m
aps the
scan
ned
ch
ara
c
ter imag
e to fo
rm
su
itable for
su
bse
que
nt
comp
uter p
r
o
c
e
ssi
ng while
descri
p
tion is a feature
sel
e
ction whi
c
h deal
s
with
ext
r
actin
g
features
in so
me q
u
a
n
titative man
ner
or
differe
ntiating on
e
cla
ss
of obj
e
c
ts from a
n
o
t
her. Thi
s
was
achi
eved u
s
in
g internal
cha
r
acte
ri
stics, t
hat is, the pixels comp
romi
sing the regio
n
.
3.4.
Kno
w
l
e
d
g
e Bas
e
The
kno
w
le
d
ge b
a
se con
t
ains th
e n
u
m
bers, p
u
n
c
tuation, u
ppe
r and
lower case
s
of
Yorub
a
alph
abets a
s
sh
own in Fi
gu
re 1 and 2.
It is basi
c
all
y
a databa
se of typed and
hand
written
English
alph
a
bets, nu
mbe
r
s, and
pu
nct
uation
s
. Individual
cha
r
a
c
t
e
r ima
g
e
s
in
the
kno
w
le
dge
b
a
se
are u
s
ed
to gen
erate
the correlat
io
n
values for th
e input
ch
ara
c
ter im
age
a
nd
output ch
ara
c
ter text.
Figure 1. Sch
e
matic of the
off-line Yoru
b
a
Optical
Cha
r
acte
r Read
er
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Yorub
a
Lan
g
uage a
nd Nu
m
e
rals
’
Offline Interpreter
Usi
ng… (Ola
kanm
i O. Ola
dayo)
171
Figure 2(a
)
. Input image
chara
c
te
r befo
r
e
norm
a
lization
Figure 2(b
)
. Input image te
xt after
norm
a
lization
Figure 2(c). Input image te
xt after filtered
Figure 3(a
)
. OCR han
dwri
tten Yoruba
chara
c
te
r kn
o
w
led
ge ba
se
Figure 3(b
)
. OCR typed Yorub
a
ch
aract
e
r kn
owl
edg
e
base
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 1, Janua
ry 2015 : 166 –
173
172
Figure 4(a
)
. OCR input of a scann
ed im
age
text documen
t
Figure 4(b
)
. OCR output o
f
the scan
ned
image
text documen
t in 4a
Figure 5(a
)
. OCR input of a scann
ed im
age
text documen
t
Figure 5(b
)
. OCR output o
f
the scan
ned
image
text documen
t in 5a
Figure 6(a
)
. OCR input of a scann
ed
hand
written i
m
age text docume
n
t
Figure 6(b
)
. OCR output o
f
a scan
ned
hand
written i
m
age text docume
n
t
4. Test and
Discus
s
ion
The OCR sy
stem wa
s subje
c
ted to different
set
of input text images in
orde
r to
determi
ne its re
cog
n
ition
efficien
cy. Th
e test
wa
s
carri
ed o
u
t on
both typed
and h
and
written
input texts. T
he inp
u
t ima
ges
as shown in Fig
u
re
4
(
a), 5
(
a
)
a
nd
6(a
)
a
r
e diffe
rent
set of in
put
texts created
usin
g the
pai
nt bru
s
h
a
s
p
en an
d p
a
int t
e
xt whi
c
h
rep
r
esent h
and
written an
d typ
e
d
Yorub
a
texts respe
c
tively.
The output
s of the O
CR system for the input text image are
sho
w
n
in
Figure 4
(
b), 5
(
b)
and
6(b).
T
he te
st re
sul
t
s we
re
quite
impre
s
sive. It wa
s o
b
se
rve
d
from th
e O
C
R
output in Fig
u
re 4
(
b) that
characte
rs
I and O
we
re
the only chara
c
ters not
recogni
ze
d. This
sho
w
s a
n
a
c
curacy
of 86
% for the
typed text with
executio
n tim
e
of 1
12
cha
r
/se
c
re
cog
n
ition
rate. Al
so, fo
r inp
u
t text in
Figu
re
5(b
)
it
wa
s
ob
serve
d
from
the
O
C
R outp
u
t in
Figure 5
(
b
)
th
at
all the E
nglish num
be
rs were
co
rrectly
recogni
ze
d a
nd
conve
r
ted
to the
Yoru
b
a
nu
merals.
Thi
s
s
h
ow
ed
acc
u
r
a
c
y
o
f
10
0% fo
r
the
nume
r
a
l
s
re
co
g
n
ition a
n
d
conversion.
T
he O
C
R
system
output in Fig
u
re 6
(
b)
whi
c
h rep
r
e
s
ent
s OCR output
for the han
dwritten input text in Figure 6(a),
also
re
co
rd
ed
an
accu
ra
cy
of 100%. It
was
ob
se
rve
d
t
hat the
devel
oped
Yoruba
OCR system’
s
perfo
rman
ce
unit is i
ndep
e
ndent a
nd
co
nstant fo
r ha
ndwritten a
n
d
typed text image
s of different
sizes. Also, t
he re
sult sh
o
w
ed that the
devel
oped
OCR syste
m
more effecti
v
ely recog
n
ized
nume
r
al
s tha
n
alpha
bets.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Yorub
a
Lan
g
uage a
nd Nu
m
e
rals
’
Offline Interpreter
Usi
ng… (Ola
kanm
i O. Ola
dayo)
173
Referen
ces
[1]
Isaac, Adejo
j
u
O.
T
he Search for a Yoruba
Orthogra
p
h
y
s
i
nce the 1
840s
Obstacles to the Ch
oice
of
the Arabic.
Sudanic Africa
. 2003; 77-
10
2.
[2]
A Jain, Karu K. Page Se
gmen
tation Usi
ng T
e
xt
ure A
nal
ys
is, Pattern Reco
g
n
itio
n. 200
6; 743-7
70.
[3]
Kund
u A, MITRE Cor
p
, McL
ean,
Hin
es T
,
Phill
ips
J, Hu
yck BD. Ara
b
ic
Han
d
w
r
iti
ng
R
e
cog
n
itio
n
Usi
n
g
Variable Duration
HMM.
IC
DAR. Ninth I
n
ternati
o
n
a
l C
onfere
n
ce
on
.
Docume
nt A
nal
ysis
an
d
Reco
gniti
on. 2
007.
[4]
Dueir
e
L
i
ns,
R, Pereir
a Sil
v
a, G, Gomes e Silv
a AR.
Assessin
g
an
d
Improvin
g th
e Qual
it
y
of
Docum
ent Images Acqu
ire
d
w
i
t
h
Portab
le D
i
gital
Camer
a
s.
ICDAR. Ninth
Internatio
na
l C
onfere
n
ce o
n
.
Docum
ent Ana
l
y
s
is an
d Reco
gniti
on. 20
07; 2: 569-5
73.
[5]
Yin,
Xu-
C
h
eng
, Sun, J
un,
Na
oi, S, F
u
j
i
moto
K. A
M
u
lti-Stage Strat
e
gy
t
o
Pe
rspectiv
e
R
e
ctificatio
n f
o
r
Mobil
e
Pho
ne
Camera-B
ase
d
Docume
nt Images. 200
7.
[6]
Sun, Jun, Hua
ng, Kaizh
u
, Ho
tta Y,
F
u
jimoto
K.
Degrade
d Char
acter Rec
ogn
ition b
y
C
o
mpleme
ntar
y
Classifi
ers Co
mbin
ation.
IC
DAR. Ninth I
n
ternati
o
n
a
l C
onfere
n
ce o
n
.
Docume
nt Anal
ysis
an
d
Reco
gniti
on. 2
007.
[7]
Pratap RL, Sat
y
a
p
ras
ad L, Sastr
y
A. Middl
e
Z
one Comp
on
ent Extracti
on
and R
e
cog
n
iti
o
n of
T
e
lugu.
ICDAR Ninth I
n
ternati
o
n
a
l C
onfere
n
ce o
n
Docu
ment Ima
ge Doc
u
ment
Analys
is and
Reco
gniti
on
.
200
7,
[8]
Mohamm
ed, Ali Q.
T
e
mplate Matching
Method
for R
e
cog
n
itio
n Mu
snad C
haract
e
rs base
d
o
n
Correlation Analy
s
is.
ACIT.
2011.
[9]
Hua
ng, Ka
izhu
, Sun, Jun, H
o
tta, Y, F
u
jimoto
K. An SVM-Based
Hig
h-acc
u
rate R
e
co
gniti
on Ap
pro
a
c
h
for Han
d
w
r
itte
n Num
e
rals
b
y
Usin
g D
i
ffere
nce F
e
atures.
ICDAR, Ni
nth
Internat
io
na
l C
onfere
n
ce
o
n
Docu
ment Ana
l
ysis an
d Reco
gniti
on
. 20
07;
589-
593.
[10]
RMK Sinh
a, e
t.al. H
y
br
idC
o
nte
x
tual T
e
xt
Reco
gniti
on
w
i
th String matc
hin
g
.
Pattern Analys
is
an
d
Machi
ne Intell
i
genc
e (PAMI).
199
7;
915-
925.
[11] F
u
kuna
ga
K.
Introductio
n
to Statistical Patter
n
Reco
gniti
on
. 199
0.
[12]
Hua
ng, Gar
y
,
Lear
ned-M
ill
er, Erik, McCa
llu
m, A
ndre
w
.
C
r
y
t
ogr
am D
e
co
din
g
for Optic
a
l C
har
act
e
r
Reco
gniti
on.
[13]
Kamalj
it, Kaur
, Balpre
et, Kaur. Ch
aracter
Re
co
gniti
on
of Hig
h Sec
u
rit
y
N
u
mb
er
Plates Us
in
g
Morph
o
lo
gica
l
Operator.
Int
e
rnati
ona
l J
o
u
r
nal
of C
o
mp
uter
Sci
ence
& Eng
i
ne
eri
n
g
T
e
chn
o
lo
gy.
IJCSET.
2013; 4(5).
[14]
Lin, Sh
an
g-Hu
ng. An Intro
d
u
c
tion to F
a
c
e
Reco
gniti
on T
e
chn
o
lo
g
y
.
Informing
Scie
nc
e spec
ial
issu
e
on Multi
m
edi
a Informin
g
T
e
ch
nol
ogi
es
. 200
0; 3(1).
[15]
Nad
eem, Dan
i
sh, Rizvi, Sale
ha.
Char
acter Reco
gniti
on Us
ing T
e
mp
late
Matchin
g
.
[16]
Na
w
a
z, T
abas
sam, Hassa
n, S
y
e
d
Ammar,
Naqv
i, S
h
a
h
, Rehm
an, Ha
bi
bur, F
a
iz, An
o
s
hia. Optica
l
Char
acter Rec
ogn
ition S
y
ste
m
for Urdu (Na
skh F
ont) Usin
g Pattern Matchin
g
T
e
chniqu
e.
[17]
Qing, Chen,
Emi, Petriul M.
Optical
Ch
ar
acter R
e
co
gnit
i
on
for M
ode
l-
base
d
Ob
ject
Reco
gniti
on
Appl
icatio
ns.
[18]
Saqi
b, Rash
ee
d, Asad, Na
e
e
m
, Omer,
Ishaq. Automated
Numb
er Plate
Reco
gniti
on U
s
ing H
o
u
g
h
Lin
e
s an
d T
e
mplate Match
i
ng.
Procee
di
n
g
s of W
o
rld C
ongr
ess E
ngi
n
eeri
ng an
d Co
mp
uter Scie
nc
e
WCECS.
2012.
[19] Ullma
nn
JR.
A
pplic
atio
n of Pattern Reco
gn
ition
. CRC Pr
es
s, Inc., 1987.
Evaluation Warning : The document was created with Spire.PDF for Python.