Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
1
3
,
No.
2
,
Feb
r
uar
y
201
9
, pp.
5
98
~
605
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
2
.pp
598
-
605
598
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Handw
ritten tifin
ag
h
char
acter
recogniti
on u
sing si
mp
le
geom
etr
i
c s
h
apes and gra
ph
s
Y.
Ou
ad
id,
B.
El
ba
la
ou
i,
M.
Bouta
ou
n
te, M. Fa
kir, B.
Minaoui
Sulta
n
Moul
a
y
S
li
m
ane
Univ
ersity
,
Facu
lty
of
Sci
enc
e
and
T
ec
hni
ques,
Ben
i
-
Mel
l
al
,
Morocc
o
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
M
ay
8
, 2
018
Re
vised
Oct
11
, 2
018
Accepte
d
Nov
2
1
, 201
8
In
thi
s
pape
r
,
a
g
rap
h
base
d
h
and
writt
en
Ti
fin
agh
cha
ra
cter
re
cogn
it
ion
s
y
s
te
m
is
pre
sen
te
d.
In
pre
proc
essing
Z
hang
Suen
a
lgor
it
hm
is
enha
n
ced.
In
fe
at
ur
es
ext
ra
ct
ion
,
a
nov
el
ke
y
poin
t
ext
r
a
ct
ion
a
lgori
thm
i
s
pre
sent
ed.
Im
a
ges
ar
e
the
n
rep
rese
nt
ed
b
y
adj
a
ce
nc
y
m
at
ri
ce
s
def
ini
ng
g
ra
phs
where
nod
e
s
rep
res
ent
fea
tur
e
poin
ts
ex
tra
c
te
d
b
y
a
nov
e
l
al
gor
it
hm
.
Th
es
e
gra
phs
ar
e
cl
as
sifie
d
using
a
gra
ph
m
at
ch
i
ng
m
et
hod.
Ex
per
imental
r
esul
ts
are
obt
ai
ned
using
two
dat
ab
ase
s
to
te
st
the
eff
ec
t
ive
n
ess.
The
s
y
s
te
m
show
s
good
r
esult
s
i
n
t
erms
of
rec
ogni
ti
on
r
ate.
Ke
yw
or
d
s
:
Am
azi
gh
Gr
a
ph m
at
ching
OCR
O
ptica
l C
ha
rac
te
r
Re
co
gn
it
io
n
Tifinag
h
C
ha
ra
ct
ers
Copyright
©
201
9
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
:
Y.
O
uad
i
d
,
Faculty
of S
ci
e
nce a
nd Tec
hn
i
qu
e
s
,
Su
lt
an
M
oula
y Sl
i
m
ane Univ
ersit
y
,
Be
ni
-
Me
ll
al
, Mo
r
occo
.
Em
a
il
:
yo.o
ua
di
d@gm
ail.co
m
1.
INTROD
U
CTION
Op
ti
cal
C
har
ac
te
r
Re
c
ogniti
on
(
OCR)
is
a
branc
h
of
patte
r
n
rec
ogniti
on
a
nd
a
rtific
ia
l
int
el
li
gen
ce.
It
con
t
rib
utes
en
or
m
ou
sly
to
hum
an
-
m
achine
com
m
un
ic
at
i
on.
It
is
the
c
om
pu
ta
ti
on
al
process
of
c
on
ver
ti
ng
i
m
ages
of
pr
i
nt
ed
or
ha
ndw
ritt
en
te
xt
into
m
achine
-
r
eada
ble
te
xt
that
is
easi
ly
rep
r
oducible
by
a
c
om
pu
te
r
syst
e
m
(su
ch
a
s
Mi
cro
s
of
t
W
ord).
OCR
is
m
or
e
sp
eci
fica
ll
y
con
cer
ne
d
with
f
orm
s
of
inf
or
m
at
ion
acqu
i
red
from
a
pr
i
nted
pap
e
r
docum
ent,
su
c
h
as
i
nvoi
ces,
c
om
pu
te
rized
receipts,
pa
ssport
docum
ents,
ba
nk
sta
te
m
ents,
le
tt
ers,
bo
oks
or
a
ny
a
ppr
opria
te
docum
entat
ion.
T
he
reali
za
ti
on
of
s
uc
h
a
proces
s
is
ve
ry
c
om
plica
te
d
bec
ause
there
is
a
n
i
nf
i
nity
of
represe
ntati
on
s
o
f
w
riti
ng
.
I
ndeed
,
e
ach
pe
rson
has
a
uniq
ue
wr
it
ing
sty
le
an
d
th
ere
a
r
e
m
any
sty
le
s
and
f
onts
in
pri
nt
ed
cha
racters.
Th
us
,
c
har
act
er
rec
ogniti
on
syst
e
m
s
are
ad
apted
to
the
ty
pe
of
wr
it
in
g
e
nv
isa
ged (
pr
i
nted, m
anuscript
or c
ursive
).
Tha
nk
s
to
r
ece
nt
a
dv
a
nce
s
in
com
pu
te
r
sci
e
nc
e,
m
any
te
chni
qu
es
of
ha
ndw
riti
ng
rec
ogniti
on
hav
e
al
s
o
been
im
pr
oved
,
es
pecial
ly
f
or
Lat
in
an
d
A
ra
bic
wr
it
in
g
[1,
2].
T
hus,
in
r
ecent
ye
ars
,
wi
th
the
gro
wth
of
the
m
eans
of
c
omm
un
ic
at
ion
,
ot
her
al
pha
bets
su
c
h
a
s
t
he
Tifina
gh
al
pha
be
t
o
f
the
Am
azi
gh
la
ng
uag
e
have
integrate
d
in
th
e
inform
at
ion
syst
e
m
s.
This
ha
s
le
d
to
the
a
ppeara
nce
of
othe
r
ty
pes
of
doc
um
ents
wh
ere
wr
it
in
g
is
not
ye
t
ha
ndle
d
a
nd
the
refore
m
or
e
dif
ficul
t
to
rec
ognize
.
Text
r
eco
gnit
ion
of
su
c
h
doc
um
ents
re
qu
ire
s
m
or
e
sp
eci
fic
proces
sing t
ech
nique
s.
In
this
pa
pe
r,
the
ai
m
is
to
app
ly
a
n
OCR
syst
em
on
t
he
Am
azi
gh
ha
ndwr
it
in
g
(Tif
inag
h).
T
he
Am
azi
gh
la
ng
uag
e
is
s
poke
n
tod
ay
by
a
la
r
ge
nu
m
ber
of
popula
ti
on
s
al
l
over
the
w
or
l
d,
m
ai
nly
in
th
e
N
or
t
h
Africa
.
It
is
no
r
m
al
l
y
wr
it
te
n
f
ro
m
le
ft
to
right
and
ver
ti
cal
ly
from
top
to
bo
tt
o
m
.
The
Am
azi
gh
wr
it
in
g
is
non
-
cur
si
ve
w
hich
si
m
plifie
s
the
segm
entat
i
F
on
of
c
har
act
er
s
in
a
te
xt
i
m
age.
Fig
ur
e
1
il
lustrate
s
the
Ti
fina
gh
char
act
e
rs
a
dopted by
IRCA
M.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Han
dwrit
te
n
ti
fi
nagh char
acte
r recog
niti
on
usi
ng sim
ple g
e
om
et
ric
s
hape
s
and gra
phs
(
Y
. Ou
ad
i
d
)
599
Com
par
ing
with
Rom
an,
Ar
a
bic
or
Chi
nese
,
researc
h
on
th
e
recog
niti
on
of
Am
azi
gh
wr
i
ti
ng
has
not
reache
d
perfect
ion
.
As
far
as
a
uthors
know,
ve
ry
few
at
te
m
pts
hav
e
bee
n m
ade
on
t
he rec
ogniti
on
of
Am
a
zi
gh
hand
wr
it
in
g.
Am
ro
uc
h
et
al
.
[
3]
desc
ribe
d
a
global
ap
pr
oach
f
or
t
he
r
ecognit
ion
of
hand
wr
it
t
en
Ti
fina
gh
char
act
e
r
i
n
w
hich
they
use
d
H
ough
tra
ns
f
or
m
as
featu
re
ext
racti
on
m
et
hod
an
d
Hi
dden
Ma
rko
v
Mod
el
s
(H
MM
)
for
cl
a
ssific
at
ion
.
T
he
syst
e
m
pr
oduc
ed
pr
om
is
ing
r
esult, h
owe
ve
r
the
disc
rim
in
at
ion
of
this
m
od
el
is
no
t
good
en
ou
gh
due
to
le
ar
ning
ste
p
wh
e
n
a
sin
gle
i
m
a
ge
is
us
e
d
f
or
ever
y
ch
aracte
r.
El
Ayac
hi
et
al
.
[4
]
pr
ese
nted
a
s
ol
ution
of
ro
ta
ti
on
a
nd
noise
prob
le
m
s
by
us
i
ng
dy
nam
ic
pr
ogram
m
ing
,
i
nvaria
nt
m
o
m
ents
an
d
Walsh
tra
nsfo
r
m
.
The
syst
em
pro
du
ce
d
go
od
res
ults
ho
wever
c
om
bin
ing
m
ul
ti
ple
desc
r
ipto
rs
m
ake
the
syst
e
m
slow
e
r.
El
Kes
sab
et
al
.
[
5]
co
m
bin
ed
Mult
i
-
Lay
er
Pe
rcep
t
r
on
an
d
HMM
f
or
m
or
e
discri
m
inati
ve
syst
em
wh
ic
h
pro
ven
to
be
r
e
li
able
in
te
rm
of
rec
ogniti
on
rate
but
slo
wer
in
te
rm
of
C
P
U
ti
m
e.
Gou
na
ne
et
al
.
[6
]
pr
opos
e
d
a
hybri
d
a
ppr
oach
by
c
om
bin
ing
ne
ural
ne
tworks
an
d
b
-
gram
.
The
sy
stem
pr
od
uced
sat
isfact
or
y
r
esults,
howe
ver
rec
ogniti
on
rate
ca
n
be
im
pr
oved
usi
ng
bigge
r
c
orpu
s
.
Es
-
saa
dy
et
al
.
[
7]
pr
e
se
nted
a
syst
em
wh
e
re
the
horizo
ntal
and
ver
ti
cal
ba
sel
ine
po
sit
io
ns
are
th
e
featu
r
es
o
f
the
c
har
a
ct
er.
T
he
syst
e
m
sh
ow
e
d
bett
er
res
ults
wh
e
n baseli
nes
are replace
d b
y ce
nterli
ne.
ya
ⴰ
a
y
aḥ
ⵃ
ḥ
y
aṛ
ⵕ
ṛ
y
ab
ⴱ
b
y
aƐ
ⵄ
Ɛ
y
aɣ
ⵖ
ɣ
y
ag
ⴳ
g
y
ax
ⵅ
x
y
as
ⵙ
s
y
agᵂ
ⴳⵯ
g
ᵂ
y
aq
ⵇ
q
y
aṣ
ⵚ
ṣ
y
ad
ⴷ
d
yi
ⵉ
i
y
a
c
ⵛ
c
y
aḍ
ⴹ
ḍ
y
a
j
ⵊ
j
y
a
t
ⵜ
t
y
e
y
ⴻ
e
y
a
l
ⵍ
l
y
a
ṭ
ⵟ
ṭ
y
af
ⴼ
f
y
am
ⵎ
m
y
aw
ⵡ
w
y
ak
ⴽ
k
y
an
ⵏ
n
y
a
y
ⵢ
y
y
akᵂ
ⴽⵯ
kᵂ
yu
ⵓ
u
y
a
z
ⵣ
z
y
ah
ⵀ
h
y
ar
ⵔ
r
y
a
ẓ
ⵥ
ẓ
Figure
1.
Tifinag
h
c
ha
ra
ct
ers
Com
par
ing
with
E
ng
li
s
h,
Ar
a
bic
or
Chi
nese,
researc
h
on
th
e
recog
niti
on
of
Am
azi
gh
wr
it
ing
has
no
t
reache
d
perfect
ion
.
As
far
as
a
uthors
know,
ve
ry
few
at
te
m
pts
hav
e
bee
n m
ade
on
t
he rec
ogniti
on
of
Am
a
zi
gh
hand
wr
it
in
g.
Am
ro
uc
h
et
al
.
[
3]
desc
ribe
d
a
global
ap
pr
oach
f
or
t
he
r
ecognit
ion
of
hand
wr
it
te
n
Ti
fina
gh
char
act
e
r
i
n
w
hich
they
use
d
H
ough
tra
ns
f
or
m
as
featu
re
ext
racti
on
m
et
hod
an
d
Hi
dden
Ma
rko
v
Mod
el
s
(H
MM
)
for
cl
a
ssific
at
ion
.
T
he
syst
e
m
pr
oduc
ed
pr
om
is
ing
r
esult, h
owe
ve
r
the
disc
rim
in
at
ion
of
this
m
od
el
is
no
t
g
oo
d
en
ou
gh
due
to
le
ar
ning
ste
p
wh
e
n
a
sin
gle
i
m
a
ge
is
us
e
d
f
or
ever
y
ch
aracte
r.
El
Ayac
hi
et
al
.
[4
]
pr
ese
nted
a
s
ol
ution
of
ro
ta
ti
on
a
nd
noise
prob
le
m
s
by
us
i
ng
dy
nam
ic
pr
ogram
m
ing
,
i
nvaria
nt
m
o
m
ents
an
d
Walsh
tra
nsfo
r
m
.
The
syst
em
pro
du
ce
d
go
od
res
ults
ho
wever
c
om
bin
ing
m
ul
ti
ple
desc
riptors
m
ake
the
syst
e
m
slow
e
r.
El
Kes
sab
et
al
.
[
5]
co
m
bin
ed
Mult
i
-
Lay
er
Pe
rcep
t
r
on
an
d
HMM
f
or
m
or
e
discri
m
inati
ve
syst
em
wh
ic
h
pro
ven
to
be
r
e
li
able
in
te
rm
of
rec
ogniti
on
rate
but
slo
wer
in
te
rm
of
C
P
U
ti
m
e.
Gou
na
ne
et
al
.
[6
]
pr
opos
e
d
a
hybri
d
a
ppr
oach
by
c
om
bin
ing
ne
ural
ne
tworks
an
d
b
-
gram
.
The
sy
stem
pr
od
uced
sat
isfact
or
y
r
esults,
howe
ver
rec
ogniti
on
rate
ca
n
be
im
pr
oved
usi
ng
bigge
r
c
orpu
s
.
Es
-
saa
dy
et
al
.
[
7]
pr
e
se
nted
a
syst
em
wh
e
re
the
horizo
ntal
and
ver
ti
cal
ba
sel
ine
po
sit
io
ns
are
th
e
featu
r
es
of
the
c
har
a
ct
er.
T
he
syst
e
m
sh
ow
e
d
bett
er
res
ults
wh
e
n baseli
nes
are replace
d b
y ce
nterli
ne.
To
su
m
m
arize
,
al
l
resea
rch
pro
po
se
d
f
or
t
he
rec
ogniti
on
of
Am
azi
gh
hand
wr
it
in
g
use
d
sta
ti
sti
cal
appr
oach,
w
here
feat
ur
es
a
re
r
epr
es
e
nted
as
a
dim
ension
al
point
in
the
c
or
re
sp
on
ding
vect
or
s
pa
ce
,
w
hich
al
lo
w
the use
of
vect
or sp
ace
pr
op
e
rtie
s to resol
ve c
la
ssific
at
ion
pro
blem
s.
In
t
his
w
ork,
a
hand
wr
it
te
n
c
ha
racter
re
co
gnit
ion
syst
em
based
on
a
str
uctur
al
a
ppr
oach
i
s
pro
posed
.
In
this
ap
proac
h
t
he
patte
r
ns
are
de
vised
int
o
seve
ral
par
ts
the
n
re
pr
ese
nt
ed
by
a
gr
a
ph.
This
grap
h
des
cribes
the pr
op
e
rtie
s
and s
patia
l po
s
it
ion
of t
ho
se
parts i
nclu
di
ng their
relat
ion
s
hi
p betwee
n
eac
h othe
r.
Th
is
pa
per
is
organize
d
as
f
ollows:
in
s
ect
io
n
2,
t
he
pr
e
pro
cessi
ng
phase
i
s
descr
i
be
d.
Se
ct
ion
3
deal
s
with
the
str
uctur
al
featu
res
e
xtracti
on
ph
as
e.
Sect
io
n
4
de
al
s
with
the
cl
assifi
cat
ion
ph
ase
an
d
finall
y,
in
t
he
la
st sect
ion
e
xperim
ental
r
esults are
disc
us
s
ed.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
3
, N
o.
2
,
Fe
bru
ary
201
9
:
5
98
–
605
600
2.
PREP
ROCES
SIN
G
The
purpose
of
prep
r
ocessi
ng
is
to
produce
a
cl
ean
ve
rsion
of
the
i
nput
i
m
age
by
keepi
ng
releva
nt
inf
or
m
at
ion
which
dif
fer
e
ntiat
e the p
at
te
r
n
i
n
the
f
eat
ur
e
spac
e.
In
t
his
ste
p,
T
he
RGB
im
age
is
co
nv
e
rted
int
o
a
bin
a
ry
im
a
ge.
In
fact,
the
well
-
kn
own
Ot
su
Me
th
od
[13]
is
ap
plied
to
t
he
Re
d,
G
reen
a
nd
bl
ue
channels
of
th
e
input
i
m
age.
Gathe
rin
g
th
ose
three
m
on
oc
hrom
e
i
m
ages
giv
e
s
us
a
bin
a
ry
im
a
ge
with
t
he
m
axim
u
m
relevant
inf
or
m
at
ion
of
the
patte
rn
.
T
his
ste
p
al
lo
we
d
t
o
rem
ov
e lo
w
to
m
edium
level of n
oise
from
t
he
im
age.
Ba
sed
on
t
he
m
on
ochro
m
e
ver
si
on
of
the
input
im
age,
the
unwa
nted
a
reas
are
rem
oved
f
r
om
the
i
m
age.
It
co
ns
i
sts
of
sca
nnin
g
the
ve
rtic
al
histogram
ho
riz
on
t
al
ly
and
the
hor
iz
on
ta
l
histo
gr
a
m
ver
ti
cal
ly
t
o
fi
nd
char
act
e
r borde
rs.
The
sp
ace
betwee
n
th
os
e
bor
der
s
and t
he
i
m
a
ge
bor
de
r
s is r
em
ov
e
d.
To
sim
plify
the
ext
racti
on
st
ru
ct
ur
al
featu
r
e;
the
norm
al
i
zed
m
onoc
hro
m
e
i
m
age
is
t
hinned
us
in
g
Zha
ng
-
S
uen algori
thm
[
8]. Fo
r
a
bette
r
t
hinning res
ults, t
his
al
gorithm
is u
sed
as
foll
ow
:
1.
Apply
thi
nn
i
ng
on
co
nnect
ed
com
po
ne
nts
i
nst
ead
of
a
pp
ly
ing
it
to
the
e
ntire
im
age.
T
his
al
lows
us
to
tre
at
the case
of r
em
ov
e
d
c
om
po
ne
nts.
2.
Sm
oo
th the sk
el
et
on
by eli
m
i
nating pi
xels t
hat do
not af
fe
ct
the con
necti
vity
an
d t
opol
ogy o
f
t
he
c
har
a
ct
er
Figure
2
s
umm
arize t
he
ste
ps
of prep
r
ocessin
g of a
hand
wr
it
te
n
Tifin
ag
h
c
ha
racter.
(a)
(b)
Figure
2.
Thinnin
g res
ult exam
ple: (a) S
ta
nd
a
rd thin
ning
Algorithm
, (
b)
En
ha
nced th
inn
in
g
al
gorith
m
Figure
3
,
il
lustrate
s
the
exec
ut
ion
tim
e
of
th
e
process
of
pre
-
proce
ssin
g
of
the
Tifina
gh
hand
wr
it
te
n
char
act
e
rs
an
d
tho
se
of
t
he
pr
i
nted
c
har
a
ct
ers
us
e
d
in
pr
e
vious
wor
k
[11].
T
he
pr
eprocessi
ng
ti
m
e
of
hand
wr
it
in
g
is
ob
viously
gre
at
er
tha
n
that
of
the
pri
nte
d
w
riti
ng
be
cau
se
the
la
tt
er
r
equ
i
res
m
uch
m
or
e
pr
e
processi
ng.
Figure
3.
Pr
e
-
proces
sin
g
tim
e o
f han
dw
ritt
en
an
d pr
i
nt
ed
c
har
act
e
rs
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Ya
Yag
Yad
Yey
Yak
Yah
Ya
ᶓ
Yaq
Yaj
Yam
Yu
Yaṛ
Yas
Yac
Yaṭ
Yay
Yaẓ
CPU TIME
(S)
han
d
w
ri
t
ti
ng
Printed
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Han
dwrit
te
n
ti
fi
nagh char
acte
r recog
niti
on
usi
ng sim
ple g
e
om
et
ric
s
hape
s
and gra
phs
(
Y
. Ou
ad
i
d
)
601
3.
STRUCT
URAL FEAT
U
R
E E
X
T
R
AC
T
ION
The
str
uctu
ral
descr
i
ptions
c
on
sist
of
e
xtra
ct
ing
the
to
polog
ic
al
an
d
geom
et
rical
pr
ope
rtie
s
of
th
e
char
act
e
r
with
a
high
tolera
nc
e
to
disto
rtion
s
and
va
riat
ions
of
sty
le
.
This
ty
pe
o
f
re
pr
e
sent
at
ion
ca
n
al
s
o
e
ncode
so
m
e
kn
owle
dge
ab
out
the
s
tructu
re
of
the
obj
ect
or
can
prov
i
de
so
m
e
knowle
dge
ab
ou
t
the
c
om
ponen
ts
const
it
uting
t
his ob
j
ect
.
Gr
a
ph
represe
nt
at
ion
is
one
of
the
m
os
t
ade
quat
e
ways
t
o
e
nc
od
e
the
str
uctu
re
in
f
or
m
at
ion
of
a
pa
tt
ern
(in
our
case
T
ifinag
h
c
har
ac
te
r)
.
It
is
a
form
al
m
at
he
m
a
ti
cal
rep
rese
nt
at
ion
of
a
set
of
obj
e
ct
s
an
d
their
relat
ion
s
hip
s
.
Each
obj
ect
is
cal
le
d
ve
rtic
es.
The
relat
io
nsh
ips
betwee
n
ob
j
ect
s
a
re
cal
le
d
edg
e
s.
M
or
e
f
orm
ally,
a
gra
ph
is
defi
ned
as
an
o
r
de
red
pair
=
(
,
)
whe
re
is
a
set
of
ver
ti
ces
an
d
is
a
set
of
e
dg
e
s
that
def
i
ne
the
con
necti
vity
b
et
we
en
a
pair
of
vert
ic
es.
The
a
dj
ac
ency
m
a
trix
is
a
w
ay
to
re
pr
ese
nt
the
gra
ph
s
.
It
is
a
bin
a
ry
squ
are
m
at
rix
where
the
nu
m
ber
of
ver
t
ic
es
|
|
is
it
s
siz
e.
The
ent
ry
in
r
ow
i
a
nd
c
olum
n
j
is
no
nze
r
o
if
an
d
on
ly
if
the
ed
ge
(
,
)
is
in
the gra
ph whic
h
m
eans:
{
(
,
)
=
1
,
(
,
)
=
0
,
(
1
)
In
this
case,
a
uthors
wor
ked
on
un
directed
gra
ph
s
whose
ed
ges
directi
on
is
not
im
po
rta
nt.
Th
i
s
represe
ntati
on
is
fast
an
d
c
om
pact.
Howe
ver,
there
is
a
redu
nd
a
ncy
of
inf
orm
ation
since
the
m
at
rix
is
sy
m
m
e
tric
.
To
re
prese
nt
Tifina
gh
c
ha
racters
by
gr
a
phs,
ke
y
po
i
nts
are
e
xtracted
first.
They
will
ser
ve
as
delim
i
te
r
for
ch
aracte
r
se
gm
entat
ion
into
li
ne
se
gm
ents
and
no
des
in
grap
h
represe
ntati
on
.
In
a
pe
rf
e
ct
char
act
er
sk
e
le
ton,
a
f
or
e
gro
und
pix
el
is
c
on
si
der
e
d
as
a
ke
y
point
by
l
ooki
ng
to
th
e
nu
m
ber
of
f
oregro
und
pi
xels
in
it
s
neig
hborh
ood
as
show
n
in
F
i
gure
4.
I
f
the
num
ber
of
neig
hbors
is
one
th
en
that
pi
xel
is
con
si
der
e
d
as
an
en
d
po
i
nt
(as
s
how
n
i
n
F
ig
ure
4.
a
).
I
f
t
he
num
ber
of
neig
hbors
is
m
or
e
tha
n
t
wo
th
en
that
pix
el
is
a
n
i
nters
ect
ion
po
i
nt (
see
F
ig
ure 4.b).
H
owev
er,
a sk
el
et
on
is
ra
rely
p
er
fect
no
m
at
te
r
wh
a
t
thin
ning
al
go
rithm
is
us
e
d,
t
wo o
r
m
or
e
-
pix
el
wi
dth
pa
rt
a
re
of
t
en
fou
nd
in
th
e
sk
el
et
on
.
T
o
rem
edy
to
this
issue,
the
num
ber
of
t
he
t
ransi
ti
ons
from
a
backg
r
ound
pix
el
t
o
a
f
or
e
gro
und
pi
xel
is
c
onside
r
ed
i
ns
te
ad
of
t
he
nu
m
ber
of
f
or
e
gro
und
pi
xe
ls
(see
F
igure
4.c)
.
Wh
ic
h m
eans
tha
t
a
pix
el
is
cl
assifi
ed a
s
a
pri
m
ary
key
point
if
the
num
ber
of
transiti
ons
is
one
or
m
or
e than
t
wo.
0
1
0
0
1
0
0
0
0
0
1
0
0
1
1
0
1
0
0
1
0
1
0
0
1
1
(a)
(b)
(c)
Figure
4.
Exam
ple o
f pi
xel n
ei
ghbor
ho
od, (a): e
nd poi
nt, (b
): inters
ec
ti
on
po
i
nt, (c):
none key
po
i
nt p
ixel
since the
num
ber
of
t
ran
sact
i
ons
from
w
hite
pix
el
to
b
la
c
k pixel is t
wo
En
d
a
nd
inters
ect
ion
points
a
r
e
no
t
en
ough
to
div
ide
the
c
ha
r
act
er
s
kelet
on
into
se
ver
al
li
ne
segm
ents.
Seco
nd
a
ry
key
points
are
nee
ded
betwee
n
s
egm
ents
that
a
re
c
urvier
the
n
strai
ght.
To
do
s
o,
a
n
al
gorithm
is
pro
po
se
d w
hic
h
is
sim
il
ar to
t
he
alg
or
it
hm
au
th
or
s
pro
po
s
e
d
i
n [
14]
that
a
ll
ow
s
to
div
ide
a
cur
vy se
gm
e
nt i
nto
sever
al
li
ne
se
gm
ents.
It
is
ba
sed
on
the
fac
t
that
the
sm
all
est
distance
be
tween
tw
o
po
i
nts
(i
n
our
cas
e
th
ose
two
points
are
the
e
nd
of
se
gme
nts
e
xtracted
base
d
on
e
nd
a
nd
inter
sect
ion
po
i
nts)
is
the
st
raig
ht
li
ne
c
on
necti
ng
them
.
The
al
gorithm
sta
rts
by
e
xtr
act
ing
segm
ents
bet
ween
pr
i
m
ary
key
points
(end
a
nd
i
ntersecti
on
po
i
nts).
Ev
ery
segm
ent
is
cl
assifi
ed
int
o
a
st
r
ai
gh
t
li
ne
or
c
urvy
li
ne.
This
is
done
us
i
ng
a
thres
hold
w
hi
ch
is
the
le
ngth
of
s
egm
ent
div
i
de
d
by
the
E
uclidia
n
distance
it
s
end.
T
he
va
lue
of
t
he
t
hr
e
sh
ol
d,
cal
culat
ed
e
m
pirical
ly
,
is
0.2.
T
he
e
xam
ple
for
“y
a
h”
c
har
act
er
key
points
extracti
on
us
in
g
aut
hor
s
pr
opos
e
d
al
gorithm
sh
ow
n
in
Fi
gur
e 5
.
{
d
D
≥
0
.
2
t
he
seg
men
t
is
an
ar
c
d
D
<
0
.
2
t
he
seg
men
t
is
a
li
g
ne
(
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
3
, N
o.
2
,
Fe
bru
ary
201
9
:
5
98
–
605
602
Figure
5.
“y
ah” ch
a
racte
r key
points e
xt
racti
on
us
i
ng
auth
or
s
pr
opose
d
al
go
rithm
Wh
e
n
a
n
a
rc
is
detect
e
d,
the
ort
hogo
nal
distance
bet
ween
th
e
se
gm
ent
el
em
ents
an
d
the
strai
gh
t
li
ne
is
cal
culat
ed.
The
el
em
ent
that
has
the
m
axi
m
u
m
per
pe
nd
i
cular
dista
nce
is
desig
nated
a
s
an
inflect
io
n
po
i
nt.
Ever
y
ti
m
e
a
po
int
is
a
dd
e
d
t
o
a
se
gm
ent,
two
n
e
w
se
gm
e
nts
re
place
the
form
er
segm
e
nt
in
t
he
se
gme
nt
li
st.
The
al
gorithm
u
se
d
f
or the
ext
racti
on
of the
key points
is as
foll
ow
s:
1.
Extract
pr
im
ary key points
.
2.
Extract se
gm
e
nts
based o
n
t
he
se points.
3.
Fo
r
e
ver
y se
gm
ents, ch
ec
ks
i
f
it
is an
arc
.
4.
Fo
r
e
ver
y a
r
c,
m
ake th
e
po
i
nt
that ha
ve
t
he b
igg
est
or
t
hogo
nal d
ist
a
nce as
a k
ey
po
i
nt.
5.
Divid
e
arc
segm
ent into
t
wo
segm
ents w
he
r
e the
new k
ey
po
i
nts
beco
m
e an
e
nd of t
he n
ew
se
gm
ents.
6.
Update the
li
st of segm
ents.
7.
Re
peat ste
p 3
to 6
un
ti
l n
o
a
rc
seg
m
ent is d
et
ect
ed.
T
o
e
valu
at
e
th
e
pe
rfor
m
ance
of
the
pro
pose
d
al
gorithm
,
for
eac
h
Tifina
gh
ha
ndwr
it
te
n
char
act
e
r,
t
he
nu
m
ber
of
key
points
e
xtract
ed
by
this
al
gorithm
is
com
par
e
d
t
o
that
e
xtracted
by
t
he
well
-
known
Harris
al
gorithm
[1
5]
an
d
t
he
op
ti
m
al
nu
m
ber
of
po
i
nts
(calc
ulate
d
em
pirical
ly
)
necessa
r
y
for
t
he
gr
a
ph
ic
al
represe
ntati
on
;
an
d
on
the
ot
he
r
hand,
the
ex
ecuti
on
ti
m
es
of
these
al
gorit
hm
s.
T
his
com
par
iso
n,
il
lust
rated
in
F
igure 6
an
d
7,
shows
that
ou
r
al
gorithm
is
m
or
e
e
ff
ic
ie
nt
t
ha
n
Ha
rr
is
i
n
te
r
m
s
of
the
sp
ee
d
an
d
t
he
ext
ra
ct
ion
of the
key
po
i
nt
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Han
dwrit
te
n
ti
fi
nagh char
acte
r recog
niti
on
usi
ng sim
ple g
e
om
et
ric
s
hape
s
and gra
phs
(
Y
. Ou
ad
i
d
)
603
Figure
6.
Com
par
ison o
f
the
nu
m
ber
of
po
i
nts
ob
ta
ine
d by the
H
a
rr
is
m
et
ho
d, the
pr
opos
e
d
al
gorithm
an
d
the opti
m
al
n
um
ber
o
f p
oin
ts
Figure
7.
Exec
ution t
i
m
e
of the
H
a
rr
is a
lgorit
hm
an
d
t
he pr
opos
e
d
al
gorithm
, case of
hand
wr
it
te
n
Tifinag
h
char
act
e
rs
Now
t
hat
th
e
char
act
e
r
s
kele
ton
is
div
ide
d
into
se
ve
ral
li
ne
se
gm
ents
a
nd
key
points
,
a
gr
a
phic
al
represe
ntati
on
can
be
a
pp
li
ed
wh
e
re
the
nodes
re
pr
ese
nt
the
key
po
ints
an
d
e
dg
e
s
re
pr
ese
nt
se
gm
ents
connecti
ng t
he
se points i
n pairs. Th
e
grap
h
c
on
st
ru
ct
io
n
al
gorithm
is p
r
oce
eded as
fo
ll
ow
s:
1.
Let
be
a
n
×
ad
ja
cency m
at
rix
wh
e
re
n
is t
he nu
m
ber
of k
ey
po
i
nts.
2.
Fo
r
e
ve
ry
se
gm
ent,
f
our
in
f
orm
ation
a
re
ext
racted:
first
ke
y
point,
sec
ond
key
po
i
nt,
seg
m
ent
ori
entat
io
n
,
segm
ent leng
th
.
3.
Ba
sed on t
his i
nfor
m
at
ion
, t
he
adjace
ncy m
at
rix
is c
onstruc
te
d
as
fo
ll
ow:
{
AM
(
i
,
j
)
=
ω
,
key
po
in
t
i
is
con
ne
cte
d
to
key
p
oint
j
AM
(
i
,
j
)
=
0
,
el
se
(3)
Wh
e
re,
ω
=
2
×
L
+
O
(4)
0
10
20
30
40
50
60
Ya
Yab
Yag
Yag
ᵂ
Yad
Yaḍ
Yey
Yef
Yak
Yakᵂ
Yah
Yaḥ
Ya
ᶓ
Yax
Yaq
Yi
Yaj
Yal
Yam
Yan
Yu
Yar
Yaṛ
Yaᵹ
Yas
Yaṣ
Yac
Yat
Yaṭ
Yaw
Yay
Yaz
Yaẓ
Optima
l
nu
m
be
r of key p
oin
ts
H
ar
ri
s
algorithm
Propos
e
d a
l
gorith
m
0
0.005
0.01
0.015
0.02
0.025
0.03
Ya
Yag
Yad
Yey
Yak
Yah
Ya
ᶓ
Yaq
Yaj
Yam
Yu
Yaṛ
Yas
Yac
Yaṭ
Yay
Yaẓ
CPU TIME
(S)
CHARAC
TE
R
NAME
Propos
e
d a
l
gorith
m
H
ar
ri
s
algorithm
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
1
3
, N
o.
2
,
Fe
bru
ary
201
9
:
5
98
–
605
604
This
represe
ntati
on
is
faster
than
i
n
t
he
cas
e
of
the
use
of
key
points
e
xtracted
by
Ha
rr
is.
This
is
il
lustrate
d
in
F
i
gure
8,
w
hic
h
giv
es
t
he
exec
ut
ion
tim
e
of
the
al
gorithm
of
the
gra
ph
ic
al
re
pr
ese
ntati
on
f
or
each
char
act
e
r.
Figure
8.
Exec
ution t
i
m
e
of the
grap
h re
pr
ese
ntati
on al
gorithm
ch
arac
te
rs
pe
r key
points extracte
d b
y t
he
al
gorithm
an
d Har
ris alg
or
it
hm
4.
CLASSIFI
C
A
TION
The
use
d
s
pect
ral
m
et
ho
d [9
]
aim
s at find
ing a coh
e
ren
t a
gree
m
ent b
et
wee
n
tw
o
grap
hs
. T
his is don
e
by
cal
culat
in
g
Eigen
values
a
nd
Ei
gen
vecto
rs
of
the
gr
a
ph
pro
duct
(or
Sim
il
arity
m
at
rix)
of
the
tw
o
gr
aphs.
the
biggest
Eig
en val
ue o
f
t
he
gra
ph
pro
du
ct
desig
ns t
he
de
gr
ee
of sim
il
ari
ty
b
et
wee
n t
he
two g
ra
ph
s
.
Readers
are in
vited t
o
c
heck the
ori
gina
l pap
e
r
t
hat is
ci
te
d
f
or
m
or
e
detai
ls.
5.
E
X
PERI
MEN
TAL RES
UL
TS
To
e
valuate
t
he
overall
perf
or
m
ance
of
t
he
pro
posed
sy
stem
,
auth
or
s
us
e
d
tw
o
data
base
set
s
of
hand
wr
it
te
n
Ti
fina
gh
cha
racters.
T
he
first
database
[
10
]
i
s
com
po
se
d
of
1376
im
ages;
43
im
ages
for
each
char
act
e
r.
It
wa
s
create
d
wit
h
t
he
help
of
30
pe
op
le
a
nd
oth
e
r
13
to
ta
ke
in
c
on
si
der
at
io
n
i
nc
li
nation
iss
ues
.
T
he
seco
nd
da
ta
bas
e
[
12
]
,
is
c
ompo
s
ed
of
25
740
im
ages;
78
0
i
m
ages
f
or
eac
h
c
harac
te
r.
It
was
c
reated
with
th
e
help
of
60
p
ers
on w
it
h
de
fer
e
nt
a
ges. Durin
g
te
sts,
80% (1
100
im
ages
f
or
the
f
i
rst d
at
ab
ase
a
nd 2
05
92 i
m
ages
for
the
sec
ond
databa
se)
of
t
hese
databas
es
are
us
e
d
as
re
fer
e
nce
im
ages
an
d
20%
(
276
i
m
ages
f
or
th
e
first
database
a
nd 5148
im
ages
for
the
sec
ond
dat
abase)
a
re u
se
d
as
te
st
im
ages.
T
hese
e
xperi
m
ents
are
do
ne
us
i
ng
a lapto
p w
it
h 2
.6
G
Hz
s
pee
d
i
7 du
al
c
ore
pro
cesso
r,
8 GB R
AM and M
AT
LAB s
of
t
war
e
.
The
rec
ogniti
on
rate
res
ults
obta
ined
a
re
il
lu
strat
ed
i
n
T
abl
e
1.
A
rec
ognit
ion
rate
of
94
%
is
obta
ine
d
on
the
fi
rst
database
a
nd
85
%
f
or
the
sec
ond
database.
T
he
decr
e
ased
r
ecognit
ion
rate
f
ro
m
the
first
to
t
he
seco
nd
data
bas
e
is
due
to
t
he i
ncr
easi
ng
of
ba
dly
wr
it
te
n c
ha
racters,
t
he i
nc
reasin
g
of
nois
e
le
vel
or
the
loss
of
i
m
po
rtant i
nform
at
ion
in
t
he
a
cqu
isi
ti
on
ph
as
e.
Table
1.
c
om
par
iso
n of rec
ogniti
on
a
nd e
rro
r
rate
of the
pr
opos
e
d
syst
em
w
it
h othe
r
syst
e
m
s
First datab
ase [
1
0
]
Seco
n
d
datab
ase [
1
2
]
Local Dat
ab
ase
Au
th
o
rs pro
p
o
sed
syste
m
Reco
g
n
itio
n
r
ate (%
)
94
85
-
Erro
r rate
(
%
)
6
15
-
A
m
rou
ch
et
al.
[
3
]
Reco
g
n
itio
n
r
ate (%
)
-
-
9
0
.4
Erro
r rate
(
%
)
-
-
9
.6
El
Kess
ab
et
al.
[
5
]
Reco
g
n
itio
n
r
ate (%
)
-
-
9
2
.3
Erro
r rate
(
%
)
-
-
7
.7
Es
-
saad
y
et
al
.
[
7
]
Reco
g
n
itio
n
r
ate (%
)
-
9
4
.96
-
Erro
r rate
(
%
)
-
5
.04
-
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
Ya
Yag
Yad
Yey
Yak
Yah
Ya
ᶓ
Yaq
Yaj
Yam
Yu
Yaṛ
Yas
Yac
Yaṭ
Yay
Yaẓ
CPU TIME
(S)
CHARAC
TE
R
NAME
Propos
e
d a
l
gorith
m
H
ar
ri
s
algorithm
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Han
dwrit
te
n
ti
fi
nagh char
acte
r recog
niti
on
usi
ng sim
ple g
e
om
et
ric
s
hape
s
and gra
phs
(
Y
. Ou
ad
i
d
)
605
6.
CONCL
US
I
O
N
In
t
his
pa
per,
recently
,
pro
po
s
ed
O
ptica
l
Characte
r
Re
cogniti
on
(
O
CR
)
syst
e
m
i
s
ap
plied
to
hand
wr
it
te
n
Tifina
gh
cha
ract
ers.
The
m
ai
n
con
t
rib
ution
of
this
pap
e
r
is
t
he
a
pp
li
cat
io
n
of
a
novel
key
point
extracti
on
te
ch
nique
a
nd
gra
ph
m
at
ching
on
hand
wr
it
te
n
T
ifinag
h
c
ha
ra
ct
ers.
T
his
te
c
hn
i
qu
e
sho
we
d
gr
e
at
resu
lt
s
i
n
te
rm
of
patte
r
n
de
sc
riptio
n
a
nd
s
pe
ed
c
om
par
ed
t
o
t
he
well
-
known
Harris
m
eth
od.
E
xperim
e
ntati
on
on
tw
o
diff
e
re
nt
data
bases
pro
ve
d
the
go
od
perform
ance
of
the
syst
em
.
H
oweve
r,
m
a
ny
en
ha
ncem
e
nts
are
require
d,
e
sp
ec
ia
ll
y
in
the
cl
assifi
cat
ion
proc
ess.
A
bette
r
cl
assifi
cat
ion
te
c
hn
i
qu
e
will
be
pro
po
se
d
i
n
the
fu
t
ure
work.
REFERE
NCE
S
[1]
Mansi
S,
Jeth
ava
G
B
.
A
Li
t
eratur
e
R
evi
ew
On
Ha
nd
W
rit
t
en
chara
ct
er
Re
cognition
.
Ind
ian
Streams
Re
search
Journa
l
.
2013;
3(2).
[2]
Lori
go
M
L
,
Gov
inda
ra
ju
V.
Offl
i
ne
Arabi
c
handw
rit
ing
r
ec
ogn
it
io
n:
a
surve
y
.
I
EEE
Tr
ansacti
ons
o
n
Pat
t
ern
Analysis
and
Mac
hin
e
In
t
el
li
g
ence
;
2006.
[3]
Am
rouc
h
M,
Rac
hidi
A
,
El
Yass
a
M,
Mam
m
ass
D.
Handwrit
t
en
Am
az
igh
Charac
te
r
Re
cognition
Based
On
Hidde
n
Markov
Models.
Dec
ember
2010
;
ICGST
-
GVIP
J
ournal
,
10(5)
.
[4]
El
A
y
ac
h
i
R,
Mo
ro
K,
Fakir
M,
B
ouikha
l
ene
B
.
On
The
R
ec
ogni
tion
of
Ti
fin
aghe
Script
s.
Journal
of
Theoretical
a
nd
Appl
ie
d
Informa
ti
on
Te
chnol
ogy
.
2010.
[5]
El
Kess
ab
B,
D
a
oui
C,
Moro
K,
Bouikha
le
n
e
B
,
Fak
ir
M.
Re
cog
nit
ion
of
Han
dwritt
en
Ti
f
ina
gh
C
har
acte
rs
Us
ing
a
Multi
lay
e
r
Neur
al
Networks
an
d
Hidden
Mark
ov
Model.
Glob
al
Journal
of
C
omputer
Sci
en
ce
and
Technol
og
y
.
Septe
m
ber
2011;
11(15).
[6]
Gounane
S
,
Fa
kir
M,
Bou
ikhalen
B.
Handwri
tt
en
T
ifi
nagh
T
ext
Rec
ogn
it
ion
Us
ing
Fuz
z
y
K
-
NN
and
B
i
-
g
ram
La
nguag
e
Mode
l.
IJ
ACSA
Sp
ec
i
al
Iss
ue
on
Selec
t
ed
Pape
rs
fr
om
Thir
d
int
ernati
onal
sympos
i
um
on
Aut
omati
c
Amazigh
proce
ss
ing
(
SITACAM’ 13)
.
Beni
Me
ll
a
l
,
2013
.
[7]
Es
-
Saad
y
Y,
Am
rouc
h
M,
R
ac
hid
i
A,
E
l
Yass
a
M,
Mam
m
ass
D.
H
andwri
tt
en
Ti
fin
agh
R
ec
ogni
ti
on
Us
ing
B
ase
li
n
es
Dete
c
ti
on
Fe
at
ur
es.
Int
ernati
onal
Journal
o
f
S
ci
en
ti
fic
&
Engi
nee
ri
ng
Re
search
.
20
14;
5(4).
[8]
Zha
ng
T
Y,
Sue
n
C
Y.
A
Fast
-
P
ara
l
le
l
Algorit
h
m
for
Thi
nning
Digit
al
Patterns.
Comm
unic
ati
ons
of
th
e
ACM
.
19
84
;
27(3).
[9]
Le
orde
anu
M,
Hebe
rt
M.
A sp
ec
t
ral
te
chn
ique for
cor
responde
n
ce
proble
m
s using pairwise constrai
nts.
Int
ernati
ona
l
Confe
renc
e
of
C
omputer
Vi
sion
.
2005.
[10]
Benc
har
ef
O,
C
hiha
b
Y,
Mous
aid
N,
Oujaour
a
N.
Dat
a
set
for
Ti
fin
agh
h
andwr
it
ing
ch
aract
er
r
ec
ogni
ti
on.
Data
in
Brie
f
.
2015;
4:11
-
13.
[11]
Ouadid
Y,
Fa
kir
M,
Minao
ui
B.
T
ifi
n
ag
h
Printe
d
Cha
rac
t
er
R
ec
ogni
t
ion
through
S
truc
tur
al
Feat
ur
e
Ext
ra
ct
ion
.
In
te
r
nati
onal Journal
of
Comput
er
V
ision
and
Imag
e
P
roce
ss
ing
(
IJCV
IP)
.
2016;
6(2)
:
4
2
-
53.
[12]
Es
Saad
y
Y
,
Ra
chi
di
A,
El
Yas
sa
M,
Mam
m
ass
D.
AM
HCD
:
A
Dat
aba
se
for
Am
az
igh
Hand
writt
en
Ch
ara
c
ter
Rec
ognition R
ese
arc
h
.
In
te
rnatio
nal
Journal
of
C
omputer
Applica
ti
ons
.
Augus
t
20
11;
27(4):
44
-
48.
[13]
Otsu
N.
A
Thres
hold
Sele
c
ti
on
Method
from
Gre
y
-
le
v
el
Histog
rams
.
IEE
E
Tr
a
nsacti
ons
on
sy
stems,
M
an
and
Cybe
rnet
ic
s
.
197
9;
SM
C
-
9:
pp.
6
2
-
66.
[14]
Bouta
ounte
M,
Ouadid
Y.
Ti
f
in
agh
Char
acte
rs
Rec
ognition
Us
i
ng
Sim
ple
Geo
m
et
ric
Shap
es.
I
ndonesia
n
Journ
al
of
El
e
ct
ri
ca
l
Eng
in
ee
ring
and
Com
pute
r
Sc
ie
nc
e. 2
016
Jul 1;
3(1)
:2
35
-
9.
[15]
Harri
s
C,
Stephens
M.
A
combined
co
rne
r
and
e
dge
detec
tor
.
Pr
oce
ed
ings
of
the
Fourth
Alve
y
V
ision
Confe
ren
c
e,
Unive
rs
it
y
of
Sh
ef
field
Printing
Unit,
Man
che
st
e
r
.
1988;
pp
.
147
-
151.
Evaluation Warning : The document was created with Spire.PDF for Python.