Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
9
, No
.
5
,
Octo
ber
201
9
, pp.
4311
~4
320
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v9
i
5
.
pp4311
-
43
20
4311
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
The
imp
act of th
e imag
e proce
ssi
ng in the
ind
exation s
yste
m
Youssef
El
fak
ir
, G
hiz
lane K
ha
issi
di,
M
os
t
afa Mr
ab
ti
,
Dri
ss C
he
nouni
La
bora
tor
y
of
C
om
puti
ng
and
In
te
rdisc
ipl
in
ar
y
P
h
y
sics
,
ENS,
Sid
i
Moham
ed
B
en Abdell
ah
Univer
sit
y
,
Morroco
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Des 24
, 201
8
Re
vised
A
pr
15
, 2
01
9
Accepte
d Apr
2
8
, 201
9
Thi
s
pape
r
pre
se
nts
an
eff
ic
i
ent
word
spotti
ng
s
y
stem
appl
ie
d
to
handwri
tte
n
Arabi
c
docume
nts,
where
imag
es
are
rep
r
ese
nt
ed
with
bag
-
of
-
visual
-
SIF
T
desc
ript
ors
and
a
slidi
ng
window
appr
oac
h
is
used
to
loc
at
e
th
e
reg
ions
tha
t
are
m
ost
sim
il
ar
to
the
quer
y
b
y
foll
owing
th
e
quer
y
-
by
-
exa
m
p
le
par
agon
.
First,
a
pre
-
pro
c
essing
step
is
used
to
pr
oduc
e
a
bet
t
er
rep
rese
nt
a
ti
on
of
the
m
ost
informati
ve
feature
s.
Seco
ndl
y
,
a
reg
ion
-
b
ase
d
fra
m
ework
is
depl
o
y
e
d
to
rep
rese
n
t
e
ac
h
local
r
egion
b
y
a
bag
-
of
-
visual
-
SIF
T
desc
ript
ors
.
After
ward,
som
e
expe
riments
are
in
orde
r
to
dem
onstrat
e
th
e
cod
ebook
size
inf
lue
n
ce
on
th
e
eff
ic
i
ency
of
the
s
y
stem,
b
y
anal
y
zi
ng
th
e
cur
se
of
dimensional
ity
c
urve
.
In
the
end
,
to
m
ea
sure
the
sim
il
ari
t
y
scor
e
,
a
floa
ti
n
g
dista
nc
e
base
d
on
the
desc
rip
to
r’s
num
ber
for
ea
ch
qu
er
y
is
a
dopte
d.
The
expe
riment
al
r
esult
s
prove
th
e
ef
fic
i
en
c
y
of
the
p
roposed
proc
essi
ng
steps
in
the
word spot
ti
n
g
s
y
st
em.
Ke
yw
or
d
s
:
Ba
g
-
of
-
vis
ual
word
Floati
ng sim
i
lar
it
y dist
ance
Hand
wr
it
te
n
a
r
abic doc
um
ents
Scal
e
-
in
var
ia
nt
-
featu
re tr
ans
form
Wor
d
s
po
tt
in
g
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
:
E
lfakir
Y
ousse
f,
Lab
or
at
ory
of
Com
pu
ti
ng
a
nd
In
te
r
discipli
na
ry P
hysic
s,
Nati
on
al
Supe
r
ior Sch
ool,
Un
i
ver
sit
y Si
di
Moham
m
ed
Be
n Abdell
ah, F
es, Mo
rocco
.
Em
a
il
:
yous
sef
.elfakir
1@usm
ba.
ac.m
a
1.
INTROD
U
CTION
Old
ha
ndw
ritt
en
A
rab
ic
doc
um
ents
are
par
t
of
the
richest
c
ultur
al
he
rita
ge
and
c
on
ta
ins
a
wealt
h
of
inf
or
m
at
ion
.
The
re
petit
ive
m
anipu
la
ti
on
of
these
m
anu
scripts
sho
uld
be
avo
i
ded
as
it
cou
ld
destr
oy
them
.
To
ex
plo
it
this w
eal
th of in
f
orm
at
ion
co
ntain
ed
in these m
a
nu
s
cripts
, d
igit
al
iz
at
ion
is a c
onve
nient so
l
ut
ion
to
pr
ese
r
ve
them
.
The
recent
ad
van
ce
s
in
patte
rn
rec
ogniti
on,
storag
e
,
an
d
ne
twork
te
ch
nolog
y
ha
ve
pa
ve
d
th
e
way
for
m
an
y
dig
it
iz
at
ion
proj
ect
s,
wh
i
c
h
treat
Lat
in
scripts,
su
c
h
as
m
anu
scri
pts
Be
tt
er
Access
t
o
Ma
nu
sc
ripts a
nd Br
owsin
g of
Im
ages [
1
]
, E
le
ct
ronic Acces
s
to Medie
val Manusc
ripts
(
E
A
MM
S)
[
2
]
, etc.
This
pap
e
r
dea
ls
with
t
he
pro
blem
of
w
ord
-
by
-
e
xam
ple
spott
ing
i
n
ha
ndwr
it
te
n
Ar
a
bic
docum
ents.
F
ro
m
the
surv
ey
of
word
spott
ing
syst
em
,
we
f
ound
tha
t
a
few
r
esear
cher
s
t
reat
the
hand
wr
it
te
n
Ar
a
bic
do
c
um
ents,
w
her
e
a
m
illi
on
do
c
um
ents
had
bee
n
w
riti
ng
in
var
io
us
dis
ci
plines.
I
n
th
e
Ar
a
bic
hand
wr
it
te
n
case, the
r
ec
og
niti
on
syst
e
m
is face
d wit
h va
rio
us
pro
blem
s
, which
can
b
e
su
m
m
arized as foll
ow
s:
-
Curs
i
vity
o
f
t
he
Arabic sc
ript
-
Ar
a
bic lan
guag
e co
ntains m
any d
ia
crit
ic
m
ar
ks
-
Fo
rm
o
f
the
sa
m
e let
te
r
at
the b
e
ginnin
g
a
nd end
of a
word
can
be
c
hange
-
Peo
ple writ
e wi
th their
own
s
cript
The
w
ord
s
po
t
ti
ng
proces
s
ne
eds
en
ough
ti
m
e
and
effor
t
to
be
pe
rfor
m
ed
by
m
anu
al
insp
ect
io
n.
To
facil
it
at
e the search
in
nu
m
erical
d
ocu
m
ent i
m
ages,
num
ero
us
word
s
po
tt
in
g
resea
rc
her
s
had
based on
text
li
ne
or
word
segm
entat
ion
ste
ps
[
3
-
6
]
.
F
irst,
an
init
ia
l
ste
p
is
per
f
orm
ed
to
s
egm
ent
te
xt
into
word
cand
i
dates
[7]
.
Then,
can
did
a
te
s
are
rep
re
se
nted
by
their
s
equ
e
nces
of
fe
at
ur
es
[
8,
9]
.
I
n
the
en
d,
to
c
om
par
e
the
qu
e
ry
w
ord
an
d
these
ca
nd
i
dates,
a
si
m
il
arit
y
m
easur
e
based
on
Dy
nam
ic
Ti
m
e
W
arp
i
ng
[
8]
or
Hidde
n
Ma
rkov
M
ode
l
[10]
is
us
e
d.
The
m
ai
n
pro
blem
with
these
ap
proach
e
s
is
that
they
are
ver
y
se
ns
it
ive
s
an
d
need
to
perf
orm
a
costly
segm
entat
ion
ste
p
to
sel
ect
can
di
date
re
gions.
Wh
e
n,
the
se
gm
entat
ion
ste
p
is
not
us
ua
ll
y
easy
a
nd
a
ny
er
ror
af
fects
the
re
pr
e
s
entat
ion’s
w
or
d,
the
r
ef
ore
th
e
m
at
ching
ste
ps
.
T
his
e
xp
la
i
ns
w
hy
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19
:
4
3
1
1
-
4
3
2
0
4312
researc
h
on
word
s
po
tt
in
g
an
d
retrieval
is
ori
ente
d
t
ow
a
r
ds
se
gm
e
ntati
on
-
f
ree
m
et
hods
over
the
la
st
few y
ears.
Lat
e
researc
h
on
w
ord
s
pott
ing
a
rch
it
ec
ture
has
pro
pose
d
ap
proac
he
s
that
do
not
need
a
ny
segm
entat
ion
ste
p.
I
n
[
1
1]
,
L
ey
dier
et
al
.
avo
id
the
se
gm
e
ntati
on
ste
p
i
n
the
wor
d
retrie
val
syst
e
m
,
by
us
in
g
featur
e
s
fitt
ed
to
any
ty
pe
of
al
ph
a
bet
by
c
om
pu
ti
ng
loca
l
key
po
i
nts
usi
ng
a
sim
ple
descr
i
ptor
bas
ed
on
inf
or
m
at
ion
’s
gr
a
dient.
I
n
t
he
sam
e
ap
proach,
Z
ha
ng
and
Ta
n
us
e
featur
e
s
base
d
on
t
he
Heat
Ke
rn
el
Sign
at
ur
e
[
1
2]
.
The
dr
a
wb
a
ck
of
t
hese
pr
opos
e
d
m
et
ho
ds
is
that
no
t
scal
able
to
la
r
ge
da
ta
set
s,
beca
use
the
us
e
a
co
stl
y
distance
c
om
pu
ta
ti
on
i
n
the
m
at
ching
ph
ase
.
I
n
this
way,
Rothac
ker
et
a
l.
pro
pose
i
n
[
1
3]
t
o
exp
l
oit t
he use
of b
a
g
-
of
-
visu
a
l
-
w
ord rep
rese
ntati
on
with
H
i
dd
e
n
Ma
rko
v M
od
el
s t
o
a
vo
i
d
se
gm
entat
ion
step.
In
[
1
4]
,
Alm
az
án
et
al
.
a
vo
i
d
the
segm
entat
ion
ste
p
by
re
presenti
ng
doc
um
ents
with
a
gri
d
of
H
OG
s
descr
i
ptors,
where
a
sli
ding
-
w
indow
par
a
gon
is
us
ed
to
l
oca
te
the
locat
ion
s
that
are
m
os
t
si
m
il
ar
to
the
qu
er
y
in the dat
aset
. T
he
n,
t
hey u
se
SV
Ms st
ru
ct
ure to
get a
bette
r
re
pr
e
sentat
io
n of
t
he qu
e
ry.
To
s
olv
e
the
prob
le
m
of
m
e
m
or
y,
la
te
ly
,
the
auth
or
s
m
ov
e
d
to
us
e
t
he
co
nc
ept
of
bag
-
of
-
featur
e
re
pr
e
se
ntati
on
.
T
he
m
et
hod
pro
po
se
d
by
M
arçal
[
1
5]
,
us
e
qu
e
ry
-
by
-
e
xam
ple
[
16
]
par
a
di
gm
wh
ere
t
he
l
ocal
patc
hes
a
r
e
desc
ribing
w
it
h
a
bag
-
of
-
vis
ual
-
words
m
od
el
powe
red
by
Scal
e
-
in
var
ia
nt
-
featu
re
-
t
ran
s
f
or
m
descr
i
ptors.
The
n,
the
sp
at
ia
l
pyram
id
-
m
at
ching
fr
am
ework i
s u
se
d.
In
[
1
7]
,
Rodri
gu
ez
us
es
a
Mod
el
-
base
d
a
ppr
oach
to
m
e
asur
e
the
sim
i
la
rity
between
sequ
e
nce’s
vecto
r,
w
her
e
s
ever
al
featu
res
are
e
xtracte
d
f
or
al
l
the
im
ages
by
us
i
ng
sli
ding
window
s
uch
as
l
ocal
gradient
histo
gr
am
[1
8]
,
the
zo
ning
fe
at
ur
es
[
1
9]
and
the
colum
n
featur
es
[
20
]
.
Th
is
sequ
e
nce
is
m
app
ed
t
o
a
H
MM
s
and
a
sim
il
arity
m
easur
e
is
c
om
pu
te
d
betw
een
them
.
I
n
[
2
1
]
,
Peti
tjean
exp
la
in
s
how
t
he
te
m
plate
m
at
ching
influ
e
nce
in
t
he
co
ntext
of
patte
rn
s
pott
ing
in
hist
or
ic
al
do
c
um
ent
i
m
ages
by
inte
grat
ing
an
d
e
valuati
ng
diff
e
re
nt tem
pl
at
e
m
at
ching
m
et
ho
ds.
The
de
gr
a
dation
on
t
he
ha
ndw
ritt
en
do
cum
ent
and
can
ta
ke
di
fferent
form
s,
su
c
h
as
t
he
discol
or
at
io
n
of
in
k,
inter
fe
ring
patte
r
ns
li
ke
ink
bleed
-
th
rou
gh,
sho
w
-
t
hro
ugh
[
22
]
,
e
tc
.
Ther
e
f
or
e
,
befor
e
any
pr
ocess,
a
s
featu
re
e
xtra
ct
ion
or
te
xt
s
egm
entat
ion
,
a
ppr
opriat
e
pre
-
processi
ng
is
essenti
al
in
order
t
o
correct
the d
e
gradati
on [
23
]
. In
the
prese
nt
w
ork,
t
h
e docum
ent
im
ages
are p
re
-
pr
ocesse
d,
in
or
der
t
o
e
nhance
them
and
to
el
i
m
inate
the
stron
gly
interfe
rin
g
bac
kgrou
nd,
this
ste
p
i
m
pr
ove
the
extracti
on
ph
ase
.
To
e
nab
l
e
an
ef
fici
ent
fea
ture
e
xtracti
on,
fin
ding
e
ff
ect
i
ve
a
nd
r
obus
t
f
eat
ur
es
is
a
n
i
m
po
rtant
ta
s
k,
wh
ic
h
af
fects
t
o
the
word
retrieval
perform
ance
[2
4
]
.
I
n
this
cas
e,
the
scal
e
inva
riant
featu
re
trans
f
or
m
al
go
r
it
h
m
(S
IF
T
)
ha
s
bee
n
app
li
ed
t
o
extr
act
and
to
cha
racteri
ze
intere
sti
ng
points
in
the
do
c
um
ent.
This
al
gorith
m
has
sh
own
their
eff
ic
ie
ncy
in
p
rev
i
ou
s
resea
rc
h
[
2
5,
26]
.
T
o
so
lve
t
he
pr
oble
m
of
com
pu
te
r
m
e
m
or
y
cause
d
by
descri
pto
r'
s
dim
ension
,
we
propose
t
o
use
bag
-
of
-
featu
r
es
appr
oach
[
25
]
,
the
S
IF
T
de
scripto
rs
hav
e
been
us
e
d
to
create
the h
ist
og
ram
s,
an
d
K
-
Me
an
s
cl
us
te
rin
g has
been ap
plied
f
or
cl
us
te
ri
ng
t
o creat
e the
ba
g
-
of
-
vis
ual
-
descri
pto
rs
[27
,
28]
.
T
he
n,
we
re
pr
e
sent
the
i
m
age'
s
r
egio
ns
as
hist
og
ram
s
by
us
ing
t
he
ba
g
of
visu
al
w
ords
[
29,
30
]
m
et
ho
d.
At
this
sta
ge
,
a
nd
us
i
ng
data
i
n
hi
gh
-
dim
ensi
on
al
s
paces
,
th
e
co
debo
ok
siz
e
or
t
he
cl
ust
er
'
s
nu
m
ber
becam
e
a
ver
y
i
m
po
rtant
ta
s
k,
w
hich
af
fects
to
the
re
gion'
s
represe
ntati
on
,
subse
qu
e
ntly
,
the
m
at
ching
phase
.
Fo
r
this,
we
ch
os
e
t
he
best
siz
e
of
c
od
e
book
by
analy
zi
ng
t
he
c
ur
se
of
di
m
ension
al
it
y
curve
[3
1,
32]
.
The
la
st
op
ti
on
is
the
histogram
s
di
stan
ce
com
pu
ta
ti
on
[33
]
,
a
pr
es
entat
ion
of
a
pro
posed
floati
ng
dista
nce
to
m
easure
the sim
i
la
rity
s
cor
e
w
il
l f
ollo
w.
The
rem
ai
nd
er
of
this
pa
per
is
orga
nized
as
f
ollows.
In
Sect
ion
2,
w
e
fi
rst
pr
ese
nt
t
he
pr
e
-
proce
ssin
g
sta
ge
to
e
nha
nce
the
de
gr
a
dation
on
t
he
hand
wr
it
te
n
docum
ent.
Sect
ion
3
descr
i
be
s
the
propose
d
word
sp
otti
ng
syst
em
.
Af
te
rw
a
r
d,
Sect
ion
4
stu
dy
the
influ
e
nc
e
of
the
proc
essing
ste
p
in
the
pr
op
os
e
d
syst
e
m
wh
e
re
we
re
port
e
xp
e
rim
ent
al
resu
lt
s
an
d
analy
sis.
Final
ly
,
con
cl
us
io
n
and
f
ur
t
her
r
esea
rch
are
drawn
i
n
sect
ion
5.
2.
POST P
RE
-
P
ROCESSI
NG
The
dig
it
iz
at
io
n
of
Ar
a
bic
ha
ndw
ritt
en
docum
ents
app
ea
rs
tod
ay
as
a
necessit
y
to
pr
eser
ve
the
integrity
and
r
arit
y
of
sp
ace,
Howev
e
r,
the
dig
it
iz
at
ion
is
the
first
ste
p
in
a
process
of
cl
assifi
cat
ion
an
d
ind
e
xing
to
ex
plo
it
al
l
wealt
h
of
i
nfor
m
at
i
on.
F
or
this
r
easo
n,
we
have
ad
op
te
d
a
n
ind
e
xing
m
et
ho
d
f
or
scan
ned
A
ra
bic
handwrit
te
n
do
c
um
ents.
Im
ages
docum
ents,
and
es
pecial
ly
scann
ed
ha
ndw
ritt
en
docu
m
ents,
are
c
om
plex
a
nd
co
ntain
a
la
rg
e
am
ou
nt
of
releva
nt
in
for
m
at
ion
.
Mo
st
of
this
data
is
connecte
d
by
r
el
at
ion
s
of
c
ol
or
s
or
in
te
ns
it
ie
s.
A
naly
sis
and
pre
-
proces
sin
g
of
docum
ent
i
m
ages
are
a
voide
d
in
s
om
e
scenario
of
word
spott
in
g
[
14,
15]
,
w
he
re
the
var
ia
ti
on
betw
een
c
olors
or
inte
nsi
ti
es
in
the
docum
ent
is
n
ot
la
rg
e.
Fo
r
this
,
and
i
n
orde
r
to
ov
e
rco
m
e
this
exi
sti
ng
va
riat
ion
in
oth
er
doc
um
ents
li
ke
as
I
bn
Si
na
databa
se
we
pro
po
se
to st
ar
t t
he
in
dex
i
ng s
yst
e
m
b
y pr
e
-
proces
sin
g
ste
p.
The
te
xt
se
par
a
ti
on
f
ro
m
i
m
ag
e
backgro
und
i
s
a
ver
y
va
st
dom
ai
n,
wh
er
e
m
any
r
esearc
h
address
thi
s
pro
blem
by
a
r
ough
est
im
a
ti
on
of
the
te
xt
a
nd
bac
kgrou
nd
reg
io
ns
[
3
4
-
3
7]
.
I
n
[
3
6]
,
to
identify
the
te
xt
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Th
e i
mpact
of the ima
ge pr
oc
essing i
n
t
he
in
dexa
ti
on
syste
m
(
Elf
akir Y
ou
ssef
)
4313
backg
rou
nd
cl
asses,
a
global
bin
arizat
io
n
thres
holdin
g
is
us
e
d.
The
n,
to
adjust
the
thr
esh
old
val
ue,
a
no
ise
m
od
el
is
bu
il
di
ng
a
nd
us
ed
.
I
n
auth
or
wor
k
[2
3]
,
an
d
in
order
to
ide
ntify
the
te
xt,
backg
rou
nd,
and
un
c
ertai
n
pix
el
s,
a
bin
a
r
iz
at
ion
par
a
go
n
is
presente
d,
the
n,
t
he
unc
ertai
n
pix
el
s
it
bi
nar
iz
ed
us
in
g
a
cl
assifi
er
trai
ne
d
base
d on the te
xt and
bac
kgr
ound classe
s.
In
this
pa
per,
we
ad
dr
ess
th
e
enh
a
ncin
g
and
re
stori
ng
pro
blem
s
of
Arabic
hand
wr
it
te
n
docum
ent
i
m
ages
by
us
i
ng
a
se
ries
of
m
ulti
-
le
vel
cl
ass
ifie
rs
[
2
3]
.
W
e
have
m
od
el
ed
the
pre
-
proces
sing
im
ages
ste
p
as
sh
ow
n
in
Fig
ure
1
base
d
on
these
cl
assifi
er
s,
w
hich
can
be
us
e
d
to
an
e
nh
a
ncem
ent
or
resto
rati
on
m
et
hod.
These m
ulti
-
lev
el
classi
fiers
are m
aps
that e
xtract releva
nt
inf
or
m
at
ion
f
or
d
iffe
ren
t l
evel
s: l
ocal, r
egi
onal
an
d
global.
A
nd
pr
ov
i
de
a
val
ue
f
or
eac
h
pi
xel
in
the
im
age.
Ther
e
a
re
seve
ra
l
cl
assifi
ers.
I
n
this
wor
k
,
we
us
e
the
est
i
m
at
ed
backgro
und, the st
r
ok
e
grey
level,
and e
dg
e
pr
of
i
le
.
-
Estim
at
ed
backgro
un
Is
a
hi
gh
-
le
vel
cl
assifi
er
[
3
8]
,
us
e
m
any
ot
he
r
cl
assifi
er
s
to
arr
ive
at
a
n
est
i
m
at
e
back
gr
ound
a
s
near
as possible
to
t
h
e tr
ue bac
kgr
ound
of the im
age as
sho
wn in
Fi
gure
1
-
b
.
-
Stroke
gr
ay
le
ve
This
cl
assifi
er
prov
i
des
a
gray
value
f
or
e
ach
pi
xel
[
2
3]
,
the
est
i
m
at
ed
intensit
ie
s
for
a
stroke
is
cal
culat
ed
by
aver
a
ging
t
he
i
ntensiti
es
of
th
e
pi
xels,
a
n
i
nterpolat
ed
val
ue
will
b
e
assi
gned
f
or
the
non
-
te
xt
pix
el
s
(b
ac
kgr
ound
, fi
gure
s i
nterf
e
re
nce,
et
c.).
A
s s
how
n
i
n
Fi
gure
1
-
c
.
-
Ed
ge pr
of
il
e
The
ed
ge
pr
ofi
le
is
a
cl
assifi
er
us
ed
t
o
ov
erco
m
e
the
interfer
e
nce
pr
oble
m
of
the
inf
or
m
at
ion
.
The
cal
culat
io
n
of
the
ed
ge
prof
il
e
is
base
d
on
the
gradi
ent
of
histogra
m
in
each
regi
on
in
the
im
a
ge
as
Figure
1
-
d
.
(a)
(b)
(c)
(d)
Figure
1. Pr
e
-
proces
sin
g
im
age steps, a)
Orig
inal im
age, b)
Estim
at
ed
backgro
und,
c) S
tr
oke
gr
ay
scal
e, d
)
E
dge
prof
il
e
Figure
2
s
ho
ws
the
pr
oces
s
of
pre
-
proc
essing
ste
p.
We
a
pp
ly
the
resto
rati
on
m
et
ho
d
t
o
the
hand
wr
it
te
n Ara
bic doc
um
ents. F
ig
ure
3
s
ho
w
s th
e
pr
e
-
pro
cessi
ng r
es
ults:
Figure
2
.
Pr
e
-
proces
sin
g proc
ess
Figure
3
.
Pr
e
-
proces
sed
im
age
s
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19
:
4
3
1
1
-
4
3
2
0
4314
3.
PROP
OSE
D SYSTE
M
We
a
ddress
t
he
w
ords
s
pott
ing
pro
blem
by
us
in
g
a
Ba
g
of
Vis
ual
desc
riptors
m
od
el
powe
red
by
scal
e
-
inv
a
riant
featur
e
tra
nsf
or
m
,
wh
ic
h
c
onsist
s
to
desc
r
ibe
each
detec
te
d
interest
po
int.
As
we
ca
n
see
in
[2
7]
,
the
pe
r
form
ance
of
th
is
m
od
el
dep
e
nds
on
the
nu
m
ber
of
visu
al
de
scripto
rs
e
xtr
act
ed
from
the
i
m
age.
In
order
t
o
re
pr
ese
nt
each
r
egio
n
in
the
i
m
ages
by
a
hi
stogram
and
ta
king
acco
unt
diff
e
re
nt
wor
ds
siz
e,
we
den
sel
y
div
ide
d
the
im
age
into
a
set
of
local
re
gions.
Fo
r
this,
we
de
fine
th
ree
wi
dth
s
H
*H,
2H*H
an
d
3H*
H
to
be
sy
nchr
on
iz
e
d
wit
h
diff
e
ren
t
l
oc
al
s
reg
i
on
'
s
siz
es.
T
he
ai
m
of
this
m
ulti
-
scale
represe
ntati
on
is
to
captu
re all
di
fferent
w
ords
siz
e as s
how
n
in
Figure
4.
Figure
4
.
Ha
nd
wr
it
te
n Ar
a
bic
words in
d
i
ff
e
r
ent size
s
H*H,
H*2
H
a
nd H*
3H
The
pro
po
se
d
m
et
ho
d
it
ha
s
been
ap
plied
t
o
diff
e
ren
t
ha
ndw
ritt
en
doc
um
ent
i
m
ages.
Figure
5
s
how
s
the pr
ocess of
the
pro
po
se
d w
ord
s
pott
ing
syst
e
m
:
Figure
5
.
The
process
of the
pro
po
se
d w
ord
spott
ing sy
ste
m
3.1.
Ima
ges
re
pres
ent
at
io
n
In
[
3
9]
,
Lla
do
s
et
al
.
sh
ow
in
g
the
influ
e
nce
of
wor
d
represe
ntati
on
s
f
or
ha
ndwr
it
te
n
wor
ds
pott
ing
i
n
histor
ic
al
doc
um
ents,
and
how
a
bag
of
vis
ual
wor
ds
re
presentat
ion
us
in
g
SIFT
desc
ri
ptors
can
ef
fec
ti
vely
perform
the
cl
assic
al
appro
ac
hes,
s
uc
h
as
D
T
W
base
d
on
s
equ
e
nce
featu
r
es.
He
re,
we
use
the
Scal
e
-
Inv
ariant
Transf
or
m
Feat
ur
e
al
gorithm
,
due
to
ca
noni
zat
ion
;
desc
ri
ptors
a
re
in
va
riant
to
tra
ns
la
ti
on
s,
r
otati
ons
a
nd
scal
ing
,
an
d
w
e
sho
w
the
im
pact
of
the
pre
-
proces
sin
g
sta
ge
in
t
he
in
de
xation
syst
em
,
then,
we
sho
w
how
we
can
perform
t
he
re
su
lt
s
by
us
in
g
floati
ng
distance
sim
il
a
rity
.
The
Si
ft
detect
or
extrac
ts
the
interest
po
i
nts
from
the
i
m
ages
an
d
t
hen
we
desc
ribe
them
,
the
ta
ken
al
go
rithm
in
our
im
plem
entat
ion
is
ins
pi
re
d
by
t
he
one
ta
ken
by L
owe
et.
[
40]
.
The
m
ai
n
draw
back
of
t
his
a
ppr
oac
h
at
this
sta
ge
is
that
th
ey
us
e
a
costly
distance
com
pu
ta
ti
on
an
d
the
need
a
gr
e
at
co
m
pu
te
r
m
e
m
or
y,
w
hich
is
no
t
scal
able
to
la
rg
e
datase
ts.
The
col
os
sa
l
nu
m
ber
of
de
te
ct
ed
key
points
in
t
he
doc
um
ent
c
ause
this
prob
l
e
m
.
i.e.
the
ave
rag
e
num
ber
of
the
key
points
at
each
re
gion
is
94,
wh
ic
h
is
re
pre
sented
by
a
de
scripto
r
of
94*128,
each
doc
um
ent
i
m
age
hav
in
g
in
a
ve
ra
ge
m
or
e
tha
n
10
00
0
reg
i
on
s
,
in
this
case,
we
requi
re
a
pproxim
at
e
ly
114
MB
of
RAM
to
st
or
e
each
im
age.
T
o
s
olv
e
this
prob
le
m
cause
d
by
dim
ensio
n
of
thes
e
descr
i
pto
r
s,
instea
d
to
repr
esent
the
im
ag
e'
s
reg
ion
s
by
their
desc
riptors
w
e
encodin
g
each
re
gion
by
a
histo
gr
am
us
in
g
a
ba
g
-
of
-
vis
ual
-
descr
i
pto
r
s
fram
ewo
rk,
wh
ic
h
is
ins
pir
ed
by
m
od
el
s
us
ed
in
natu
ral
la
ng
uag
e
pr
ocessin
g,
t
his
te
ch
nique
is
base
d
on
a
spa
rse
histo
gr
am
of
occ
urrenc
e
counts
of v
is
ua
l wor
ds
.
The
m
ai
n
ste
ps o
f
this te
c
hn
i
que a
re:
-
Feat
ur
es
ex
t
rac
ti
on
-
Cl
assifi
cat
ion
"cod
e
book"
-
Qu
a
ntif
ic
at
ion
-
Con
st
ru
ct
t
he
r
egio
n'
s
histo
gra
m
s u
sin
g
c
odeboo
k
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Th
e i
mpact
of the ima
ge pr
oc
essing i
n
t
he
in
dexa
ti
on
syste
m
(
Elf
akir Y
ou
ssef
)
4315
In
the
sec
ond
s
te
p,
local
reg
i
ons
are
re
pr
ese
nt
ed
by
us
in
g
hi
stogram
s
of
B
ag
-
of
-
Visu
al
-
D
escripto
rs.
To
ac
hieve
thi
s
re
pr
ese
ntati
on,
we
us
e
10%
of
al
l
des
cript
or
s
to
quantiz
e
them
into
K
di
ff
ere
nt
cl
ust
er
s
us
i
ng
the
k
-
m
eans
a
lgori
thm
.
In
the
en
d,
al
l
re
gions
are
des
cribe
d
with
th
ei
r
histo
gr
am
s
by
assigni
ng
each
descr
i
ptor
in
t
he
re
gion
to
t
he
near
est
visu
a
l
descr
ipt
or
i
n
the
co
debo
ok.
So
,
eac
h
r
egi
on
is
re
pr
ese
nte
d
by
a
histo
gr
am
of
accum
ulate
s
f
reque
ncies
1
,
=
1
,
,
2
,
,
3
,
,
4
,
,
5
,
,
…
…
,
,
wh
e
re
k
repr
esent
t
he
nu
m
ber
of
the
visu
al
desc
rip
tor
in
the
c
od
eboo
k
a
nd
,
rep
rese
nt
the
cu
m
ula
ti
ve
fr
e
quency
of
k
visua
l
descr
i
ptor in
th
e j
em
e
reg
i
on.
3.2.
Ima
ges
re
pres
ent
at
io
n
The
in
form
at
i
on
e
xtracti
on
from
han
dwrit
te
n
docum
ent
i
m
ages
is
on
e
of
the
m
ajor
chall
eng
i
ng
top
ic
s
i
n
the
fi
el
d
of
docum
ent
im
age
analy
sis.
I
n
this
pa
r
t,
we
s
how
ho
w
the
pre
-
proc
essing
im
ages
ste
p
influ
e
nce
i
n
th
e
ind
e
xatio
n
s
yst
e
m
scenario.
The
w
ord
"
‘
مو
ي
"
wr
it
te
n
i
n
three
dif
fer
e
nt
wa
ys
in
t
he
sam
e
d
oc
um
ent
as
s
how
n
in
Fig
ure
6,
w
hich
are
wr
it
te
n
with
diff
e
ren
t
col
or
s
a
nd
diff
e
re
nt
diacrit
ic
al
m
ark
s
,
wh
il
e
so
m
e
reg
ion
s
in
these
wor
ds
are
degrad
e
d
due
to
the
antiquity
of
these
m
anu
scri
pts
and
the
m
anu
al
m
anipu
la
ti
on.
Fo
r
th
e
reas
on
abov
e,
an
d
in
or
de
r
to
com
par
e
the
pr
e
-
pro
cessi
ng
im
pact
,
we
ha
ve
te
ste
d
the
pro
po
se
d
syst
e
m
in
the
gr
ay
s
cal
e
and
pr
e
-
tr
eat
ed
i
m
ages.
We
extra
ct
the
interest
po
i
nts
from
each
word,
an
d
for
each
one
in
the
first
wor
d,
we
sea
rch
t
he
si
m
il
ar
on
e
in
t
he
sec
ond
wor
ds
as
s
how
n
in
Figure
7.
As
we
see
in
T
a
ble
1,
the
nu
m
ber
of
the
m
at
ched
points
in
the
gray
scal
e
i
m
age
is
high
than
the
pr
e
-
tr
eat
ed
im
age
due
t
o
the
fa
ults
dete
ct
ed.
T
he
m
os
t
of
this
fau
lt
s
keys
points
a
re
com
ing
f
rom
no
n
-
te
xt
re
gions
a
nd
doe
s
not
descr
i
be
the
word trait
in
the
i
m
age.
Figure
6
.
Wo
r
d ‘
مو
ي
’
w
ritt
en
in
thr
ee
d
if
fere
nt
w
ay
s i
n
the
sa
m
e d
oc
um
ent
Table
1.
N
um
ber
of
t
he
m
at
ched
po
i
nts
W
o
rds
m
a
tch
in
g
W
ith
prep
rocess
in
g
step
W
ith
out
p
reproces
sin
g
step
1
em
e
m
a
tch
in
g
67
143
2
em
e
m
a
tch
in
g
119
145
3
em
e
m
a
tch
in
g
98
131
Figure
7
.
The
m
at
ched
poi
nts
To
exam
ine
the
influ
e
nce
of
these
fa
ults
det
ect
ion
s
an
d
c
om
par
e
the
resul
ts
between
pr
e
-
treat
ed
a
nd
no
pr
e
-
treat
ed
i
m
ages
in
th
e
wo
r
d
-
spott
in
g
fr
am
ewo
r
k,
we
ap
ply
the
pro
po
se
d
syst
em
Figu
re
5.
I
n
th
e
si
m
il
arity
ste
p,
we
c
hoos
e
t
o
fix
th
res
ho
l
d
(
Tf),
an
d
we
re
tur
n
each
dista
nce
sim
i
la
rity
le
ss
of
Tf.
The
n
w
e
cal
culat
e
the
te
st'
s
accuracy
of
the
syst
em
Fscore,
wh
ic
h
is
a
har
m
on
ic
m
ean
of
prec
isi
on
an
d
rec
al
l
that
m
eans th
e a
bili
t
y of
t
he
syst
e
m
to provide
all
r
el
eva
nt so
l
ut
ion
s
and
rej
ect
oth
e
rs.
The F
-
sc
or
e
is
cal
culat
e as fol
lows
:
F
S
c
ore
=
2
.
1
1
r
ecal
l
+
r
ecision
=
2
.
rec
a
ll
+
pr
e
c
isio
n
rec
a
ll
+
pr
e
c
isio
n
(1)
To
e
valuate
th
e
pe
rfor
m
ance
of
the
ap
proac
h
i
n
hand
wr
it
te
n
Ar
a
bic
doc
um
ents,
we
ch
ang
e
the
siz
e
of
c
odeb
ook.
As
s
how
n
in
F
igure
8,
the
F
-
scor
e
res
ults
de
pendin
g
on
the
co
de
book
s
iz
e
and
t
he
ap
proac
h
base
d
on
pr
e
-
processi
ng
pr
ovide
a
good
pe
rfor
m
ance
in
te
r
m
of
F
-
sc
or
e
Ta
ble
2,
the
best
m
ean
F
-
sc
or
e
(0,77
8)
is
obtai
ned f
or 30
0
c
ode
wor
ds
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19
:
4
3
1
1
-
4
3
2
0
4316
(a)
(b)
Figure
8. F
-
sco
re
at
diff
e
re
nt c
od
e
book
siz
es,
(a)
pre
-
pr
ocess
ed do
c
um
ent, (b)
no
pre
-
proc
essed d
ocu
m
ent
Table
2
. N
um
ber
F
-
sc
or
e at
di
ff
e
ren
t c
odeb
ook
siz
es
Variable
100
200
300
400
F
-
Sco
re
f
o
r
p
re
-
p
r
o
cess
ed
d
o
cu
m
en
t
0
,62
7
0
,76
0
,78
0
,62
9
F
-
Sco
re
f
o
r
p
re
-
p
r
o
cess
ed
d
o
cu
m
en
t
0
,53
0
,61
0
,63
0
,51
3.3.
T
he curse
of d
im
ension
ality
im
pa
ct
in wor
d spot
ting s
ys
t
em
In
this
sect
ion,
we
exp
la
in
th
e
cur
se
of
dim
ensio
nalit
y
i
mp
act
in
the
bag
-
of
-
visu
al
w
ord
syst
e
m
,
this
te
rm
was
inv
e
nted
by
Ri
cha
r
d
Be
ll
m
an
in
[
31
,
3
2]
,
t
he
go
al
is
to
see
di
ve
rse
ph
e
nom
ena
that
a
pp
ea
r
wh
e
n
us
in
g
data
in
hi
gh
-
dim
ensional
sp
aces
and
a
naly
sing
them
.
Fo
r
this, w
e
use
two
dif
fer
e
nt
A
rab
ic
ha
ndwr
it
te
n
do
c
um
ents
fo
r
m
Gall
ic
a,
wh
ic
h
is
the
dig
it
a
l
li
br
ary
of
the
Fr
e
nch
Nati
onal
Librar
y,
in
open
acce
s
s.
It
br
i
ngs
dig
it
iz
ed
hand
wr
it
te
n
do
c
um
ents,
m
agazine
s,
im
ages..
.
Fir
st,
we
e
xtract
t
he
inter
est
poi
nts
f
ro
m
each
reg
i
on
in
the
pre
-
tr
ea
te
d
im
ages
us
i
ng
S
IF
T
al
gor
it
h
m
.
The
n,
i
n
the
le
ar
ni
ng
ste
p,
t
he
descri
pto
rs
of
t
he
4
first
i
m
ages
of
t
he
do
c
um
ent
are
gro
up
e
d
to
pro
vid
e
th
e
cl
us
te
r
centr
es
(c
ode
book)
.
I
n
this
s
ta
ge,
we
us
e
k
-
m
eans
al
gorithm
with
diff
e
ren
t
nu
m
ber
of
k
(ce
ntr
es),
100
to
900
centres.
T
he
n,
to
cal
culat
e
the
si
m
i
la
rity
bet
ween
the h
ist
ogram
s o
f
each
r
e
gions
in
the
im
age'
s
do
c
um
ent and
the que
ry'
s h
ist
ogram
, w
e
us
e
the cos
distanc
e
def
i
ned b
y:
S
=
1
−
∑
H
i
,
j
R
i
N
i
=
1
√
∑
H
i
,
j
2
N
i
=
1
√
∑
R
i
2
N
i
=
1
(2)
Wh
e
re
H
i,j
represent
the
occ
urren
ce
of
t
he
i
e
m
e
centre
of
the
co
debo
ok
i
n
the
j
em
e
reg
i
on,
an
d
R(
i
)
represe
nt
the
oc
currence
of
t
he
i
em
e
centre
of
the
c
od
e
book
in
the
query.
T
o
j
ud
ge
that
a
reg
i
on
j
is
sim
i
la
r
to
the
qu
e
ry,
the
cos
distance
S
sh
oul
d
be
le
ss
than
a
certai
n
thres
hold.
F
or
this,
f
or
eac
h
cod
e
book
siz
e,
we
us
e
var
i
ou
s
th
res
hold
bet
ween
0:
0.
05:
0.7
an
d
we
cal
culat
e
t
he
recall
and
pr
eci
sio
n
m
eas
ur
es
.
The
rec
a
ll
and
pr
eci
sio
n
c
urve
s sho
w
that t
he
r
es
ult de
pend on t
he
t
hr
es
hold a
nd code
book size
as
s
how
n
in
Fig
ure
9.
Figure
9. Re
cal
l and p
recisi
on
at
d
iffe
re
nt cod
eboo
k
siz
es a
nd se
ver
al
dif
fere
nt th
reshold
To
evalu
at
e
the
i
m
pact
of
the
cod
e
book
siz
e,
we
cal
culat
e
the
F_
sorce
m
easur
e
by
usi
ng
ch
oosin
g
the
best
th
res
hold
of
eac
h
si
ze
as
show
n
i
n
Fig
ure
10
.
We
rem
ark
th
at
the
best
res
ult
is
giv
e
n
f
or
k=
300,
the Fs
or
ce
d
ec
r
ease bey
ond
t
hi
s size d
ue
to
t
he
im
pac
t of
t
he
d
im
ension
al
sp
aces.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Th
e i
mpact
of the ima
ge pr
oc
essing i
n
t
he
in
dexa
ti
on
syste
m
(
Elf
akir Y
ou
ssef
)
4317
Figure
10. T
he
F
-
s
or
c
e c
urve
Now,
we
sea
r
ch
the
influ
e
nc
e
of
the
co
de
book
siz
e
on
the
thres
hold,
f
ro
m
the
(
3
)
,
w
hen
the
siz
e
increases
(
N),
the
pr
oba
bili
ty
to
no
t
find
al
l
visu
al
s
descr
ipto
rs
in
the
co
debo
ok
for
a
giv
e
n
r
egio
n’
s
descr
i
ptors
increase
,
tha
t’s
m
ean
the
pro
bab
il
it
y
P
(
)
=
0
increase,
a
nd
s
ubs
equ
e
ntly
P
(
,
∗
)
=
0
increase.
Wher
e
R
is
the
hist
ogram
o
f
the
qu
e
ry
an
d
H
j
is
the
histo
gra
m
of
j
em
e
reg
ion
in
the doc
um
ent.
So
D
=
∑
H
i
,
j
R
i
N
i
=
1
√
∑
H
i
,
j
2
N
i
=
1
√
∑
R
i
2
N
i
=
1
(
3)
Decr
ease
d
,
Ther
ea
fter
, wh
en
the
size
N
i
ncr
ease
s,
t
he
si
m
il
arity d
ist
an
ce S
will
incr
e
ase (S
=
1
-
D).
↗
⇒
↘
⇒
↗
Figure
1
1
s
hows
the
c
urve
of
the
best
thr
esh
old
f
or
eac
h
co
de
book
siz
e,
as
we
see,
the
thres
hol
d
dep
e
nd
on
the
siz
e
of
the
cod
eb
ook.
For
th
is
reaso
n,
an
d
to
ov
e
rco
m
e
t
he
pro
blem
li
n
ked
to
the
c
urve
of
dim
ension
al
it
y,
w
e
u
s
e a c
ode
book
with
300 visual
desc
ript
or
s
th
at
giv
e t
he
b
est
res
ult.
Figure
11. Best
thr
es
hold
f
or
e
ach c
od
e
book
siz
e
At
this
sta
ge,
we
have
de
m
on
strat
e
that
the
Fsc
ore
m
easur
e
dep
e
nd
on
t
he
co
debo
ok
siz
e
.
To
pe
rfor
m
the
exp
e
rim
ents,
we
shou
l
d
se
arch
the
im
pact
of
the
desc
riptor'
s
nu
m
ber
(n)
on
the
thre
sh
ol
d.
Fo
rm
(
3
)
,
wh
e
n
the
num
ber
of
interest
point
s
n
inc
reases
,
the
pr
ob
a
bili
ty
to
fin
d
al
l
vis
ua
ls
descr
i
ptors
for
in
histo
gr
am
fo
r
a
giv
e
n
re
gi
on’s
descr
i
ptors
incr
ease,
th
at
m
ean
probabil
it
y
of
P
(
)
≠
0
)
in
crease
to
o.
Ther
ea
fter
,
the
pro
ba
bili
ty
P
(
,
∗
)
≠
0
inc
rease.
S
o
D
i
ncr
ease
s,
with
0
≤
≤
1
.
The
reafte
r,
wh
e
n
n
increases
,
the
si
m
il
arity
dist
ance
S
will
decr
ease
(
S=1
-
D):
↗
⇒
↗
⇒
↘
,
w
hich
e
xp
la
in
t
hat
th
e
nu
m
ber
of
inte
rest
points
in
fluen
ce
on
the
t
hr
es
hold,
or
ea
ch
w
ord
has
a
certai
n
num
ber
of
descr
i
ptors.
For
this,
the
distan
ce
si
m
i
la
rity
sh
ould
be
ta
ken
account
the
num
ber
of
desc
r
iptor
s
by
us
in
g
a
floati
ng
thre
sh
ol
d.
F
igure
1
2
r
epr
ese
nt
s
the
threshol
d
curves
acc
ording
to
t
he
nu
m
ber
of
po
i
nts
of
i
nt
erest.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19
:
4
3
1
1
-
4
3
2
0
4318
By
analy
sing
t
hese
c
urves
,
w
e
can
use
the
relat
ion
of
th
e
pro
gressi
on
li
ne
as
a
floati
ng
t
hr
es
hold
f
or
each
qu
e
ry.
Or,
if
t
wo
dif
fer
e
nt
r
egio
ns
hav
i
ng
even
num
ber
of
points
of
in
te
rest
will
ha
ve
the
sam
e
threshold,
bu
t
di
ff
e
ren
t
hi
stogram
s,
becau
se
they
do
not
hav
e
the
sam
e
interest
po
i
nts,
and
t
her
ea
fter
,
diff
e
re
nt
distances
of sim
il
arit
y between
t
hese t
w
o regi
on
s
and t
he qu
e
ry.
Figure
12. T
hr
esh
old
acc
ordi
ng to
t
he nu
m
ber
of
desc
riptor
s
4.
E
X
PERI
MEN
TAL RES
UL
TS
We
te
ste
d
our
m
e
tho
dolo
gy
on
Ar
a
bic
ha
ndw
ritt
en
do
cum
ent
fr
om
the
di
gital
li
br
ary
Gall
ic
a,
Figure
13
pr
es
ents
qu
al
it
at
ive
res
ults
f
or
tw
o
dif
fer
e
nt
do
c
um
ents.
The
us
e
d
syst
em
is
based
on
S
IF
T
-
B
oV
W
descr
i
ptors
with
300
vis
ual
words,
a
nd
flo
at
ing
th
res
ho
l
d.
W
e
re
port
he
re
s
om
e
qu
eri
es
w
her
e
the
s
yst
e
m
yi
el
ds
autom
atical
ly
,
wh
ic
h
a
re
sim
il
ar
to
the
qu
e
ry
an
d
without
ch
os
e
th
e
K
be
st
si
m
ilar
res
ults,
w
hi
ch
is
a
pro
blem
in
oth
er
syst
em
[1
4,
41]
.
T
hen,
we
us
e
a
filt
er
ing
ste
p
t
o
sel
ect
on
e
be
st
r
esult
w
he
n
c
onf
u
sio
n
reg
i
on
s
are
r
et
urnin
g, base
d o
n
thei
r
sim
il
ari
ty
sco
res
and
posit
ion
s
.
In
c
om
par
iso
n
with
sta
te
-
of
-
the
-
a
rt,
we
s
how
the
retrie
v
al
perform
ance
in
te
rm
s
of
m
AP
as
s
how
n
i
n
Table
3,
we
c
an
see
ho
w
pre
-
processi
ng
by
ke
epi
ng
th
e
inf
orm
ative
interest
po
i
nts
in
eac
h
w
ord
a
nd
discrim
inati
ng
the
oth
ers
.
In
add
it
ion,
to
overc
om
e
the
p
roblem
li
nk
ed
to
the
var
io
us
nu
m
ber
of
i
nterest
po
i
nts in
d
if
fere
nt r
e
gions
by
us
in
g floati
ng t
hr
es
hold.
Figure
13. T
he
r
et
rieve
d
im
ages for so
m
e q
ue
ries in t
he
t
wo
evaluate
d d
oc
um
ents
Table
3.
Per
for
m
ance of th
e
prop
os
ed
m
et
ho
d
a
nd o
t
her w
orks
Metho
d
Precisio
n
Al
m
az
án
et al
.
[
14
]
6
8
,4%
Prop
o
sed
m
e
th
o
d
83%
Ho
we
e
t a
l
. [
42]
79%
Li
a
n
g
e
t
a
l
.[
4
3
]
67%
F
i
sc
he
r
e
t
a
l.[
4
4
]
62%
Elf
ak
ir
et al
.
[
4
5
]
81%
Ter
asa
wa
e
t
a
l
.
[
4
6
]
79%
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Th
e i
mpact
of the ima
ge pr
oc
essing i
n
t
he
in
dexa
ti
on
syste
m
(
Elf
akir Y
ou
ssef
)
4319
5.
CONCL
US
I
O
N
In
t
his
pa
per,
we
ha
ve
pr
e
se
nted
a
n
ef
fici
ent
fr
e
e
-
se
gm
entat
ion
w
ord
s
pott
ing
a
ppr
oac
h
f
or
A
rab
ic
m
anu
scri
pts.
T
he
pro
pose
d
m
et
hod
prese
nt
an
excell
ent
re
su
lt
wh
e
n
c
ompari
ng
oth
e
r
m
et
hods
in
li
te
ra
ture.
We
ha
ve
s
ho
wn
ho
w
t
he
proces
sin
g
ste
p
can
im
pr
ov
e
the
re
su
lt
,
i
n
Ba
g
-
of
-
de
scrip
tors
of
S
IF
T
s
yst
e
m
,
by
inf
or
m
at
ive
and
disc
rim
inati
ve
featur
es
.
The
n,
we
hav
e
sh
own
how
w
e
can
i
m
pr
ove
d
the
res
ult
by,
first
choosi
ng
t
he
be
st
siz
e
of
co
de
book
by
anal
ysi
ng
th
e
c
ur
se
of
dim
ension
a
li
ty
cur
ve,
an
d
seco
nd
ly
,
t
he
us
e
of
floati
ng
cos
di
sta
nce.
Fin
al
ly
,
we
ha
ve
pr
es
ented
a
c
om
par
iso
n
with
othe
r
m
et
ho
ds
,
w
e
te
ste
d
our
m
et
hod
us
in
g
e
xperim
e
ntal set
up
base
d on MAT
LA
B co
de
a
pp
li
ed
to Gal
li
ca data
base
REFERE
NCE
S
[1]
Cal
abr
et
to
S.
,
Bozz
i
A.
,
Pinon
J.,
"
Digit
i
za
t
ion
of
m
edi
eva
l
m
anusc
ript
s (i
n
Frenc
h
),
"
the
European
projec
t
B
AMBI
,
in:
Proce
ed
ings
of
the
conference
Tow
ards
a
ne
w
erudit
ion:
dig
it
izati
on
and
res
earc
h
in
book
hi
story
(in
Frenc
h)
,
Renc
ontr
es
Jac
q
ues
Cartier
,
L
y
o
n
,
De
c
1999
.
[2]
htt
p://ww
w.hm
m
l.
org/e
amm
s/inde
x.
htm
l
[3]
Rat
h
T.
M.
,
Ma
nm
at
ha
R.
,
"
W
ord
image
m
at
c
hing
using
d
y
n
amic
ti
m
e
warpi
ng,
"
Proceedi
n
gs
Inte
rnational
Confe
renc
e
on
C
omputer
Vi
sion
and
Pattern
R
ecogniti
on
,
pp
.
52
1
-
527
,
Feb
2013
.
[4]
El
fak
ir
Y.,
Kha
issidi
G.,
Mrab
ti
M.,
Chenoun
i
D.
,
"
Handwrit
te
n
Arabi
c
Doc
um
e
nts
Inde
xat
i
on
using
HO
G
Feat
ure
,
"
Int
ernati
onal Journal of
Computer
Ap
pli
cations
,
vo
l.
1
26,
no
.
9
,
pp
14
-
18
, S
ep
2015
.
[5]
Saee
d
K
.
,
Alb
a
koor
M.,
"
Region
growing
bas
ed
segm
ent
a
ti
on
al
gor
it
hm
for
t
y
pewri
tten
a
nd
handwri
tten
t
ext
rec
ogni
ti
on,
"
Ap
pli
ed
Soft Comp
uti
ng
,
vol.
9
,
pp
.
608
-
617
,
2009
.
[6]
Loul
oudis
G.
,
Gatos
B.
,
Prati
k
ak
is
I.
,
et
a
l.
,
"
Te
x
t
li
n
e
and
word
segm
ent
at
ion
of
handwri
tten
doc
um
ent
s
,"
Pat
t
ern
Re
cogn
it
ion
,
vo
l
.
42
,
pp
.
3169
-
3
183
,
2009
.
[7]
Fitri
ani
ngsih
F.
,
Made
nda
S.,
Erna
stuti
E.,
W
idodo
S.,
Rodiah
R.
,
"
Cursive
handwri
ti
ng
seg
m
ent
at
ion
using
ide
a
l
dista
n
ce
a
ppro
ac
h,
"
In
te
rn
ati
onal
Journal
of
Elec
tri
cal
an
d
Computer
En
gine
ering
(
IJE
C
E)
,
vol.
7
,
no
.
5
,
pp.
2863
,
2017
.
[8]
Rodrigue
z
-
Serr
a
no
J.,
Perronni
n
F.,
"
A
m
odel
-
base
d
seque
n
ce
sim
ilarity
w
it
h
appl
i
ca
t
i
on
to
h
andwri
tte
n
word
-
spotti
ng,
"
IEE
E
Tr
ans.
Pattern
Ana
l. Mac
h
.
Intell
,
vol
.
34
,
p
p.
2108
-
2120
,
2
012
.
[9]
Marti
U.
-
V.,
Bu
nke
H.
,
"
Us
ing
a
stat
isti
c
al
la
ngu
age
m
odel
to
improve
the
per
for
m
anc
e
of
an
HM
M
base
d
cur
si
ve
handwri
ti
ng
re
co
gnit
ion
s
y
st
ems
,
"
Int. J.
Pa
ttern
Re
cogn
it
.
Art
if
.
I
nte
ll
,
pp
.
65
-
90
,
2001
.
[10]
Griffi
ths
Rat
h
T.
,
Manm
at
h
a
R.
,
"
W
ord
spotti
n
g
for
historical
do
cuments
,
"
Int.
J
.
Doc.
Anal
.
Re
cogn
it
,
pp
.
139
-
152
,
20
07
.
[11]
Ley
di
er
Y.,
Ouji
A.,
Le
Bour
geoi
s
F.,
et
al.
,
"
Towa
rds
an
om
nil
ingua
l
word
ret
rie
v
a
l
s
y
s
te
m
for
anc
ie
n
t
m
anusc
ript
s,"
P
att
ern
Re
cogn
it
i
on
,
vol
.
42
,
no
.
9
,
pp
2089
-
2105
,
2009
.
[12]
Zha
ng
X.
,
Ta
n
C.
L.
,
"
Segm
ent
at
ion
-
fre
e
ke
y
word
spotti
ng
f
or
handwri
t
te
n
documents
base
d
on
he
at
k
ern
e
l
signat
ure
,
" i
n
Int
ernati
onal
Confer
enc
e
on
Docum
ent
Anal
ysis
and
Recogni
t
ion
,
pp
.
827
-
831
,
2013
.
[13]
Rotha
ck
er
L.,
Rusiñol
M.,
Fink
G.
,
"
Bag
-
of
-
feat
ure
s
HM
M
s
for
segm
ent
at
ion
-
fr
ee
word
sp
ott
ing
in
handwri
tt
en
documents,
"
1
2th
Inte
rnat
io
nal
Confe
ren
c
e
on
Docum
ent
Ana
ly
sis
and
Re
cogn
it
i
on,
Proc
ee
din
gs
,
pp.
1305
-
1309
,
2013
.
[14]
Alm
az
án
J.
,
Gordo
A.,
For
nés
A.,
et
al
.
,
"
Segm
ent
a
ti
o
n
-
fre
e
w
ord
spotti
ng
wi
th
e
xemplar
SV
Ms,
"
Pat
te
rn
Recogni
t
ion
,
vo
l.
47,
no.
12,
pp
.
3967
-
39
78
,
2014
.
[15]
Marc
al
M.,
Ald
ave
rt
D.,
Toledo
R.
,
e
t
al
.
,
"
Eff
i
ci
en
t
segm
ent
a
tion
fre
e
ke
y
word
spotti
ng
in
h
istori
cal
documen
t
col
l
ec
t
ions,"
Pa
t
te
rn R
ec
ogni
ti
on
,
v
ol
.
48
,
no
.
2
,
p
p.
545
-
555
,
Feb
2015
.
[16]
Doungpaisan
P.
and
Mingkhwan
A.
,
"
Quer
y
b
y
Exa
m
ple
of
Speake
r
Audio
Sig
nal
s
using
Pow
er
Spect
rum
and
MF
CCs
,
"
Inte
r
nati
onal Journal of
E
le
c
tric
al
&
Computer
Engi
n
ee
ring
(
IJE
C
E)
,
vol.
7
,
no
.
6
,
pp
.
2088
-
8708,
201
7
.
[17]
Rodrigue
z
-
Serr
a
no
J.,
Perronni
n
F.
"
A
m
ode
l
-
base
d
seque
n
c
e
sim
il
arit
y
wi
th
appl
i
cation
to
handwri
t
te
n
word
-
sp
ott
ing,
"
IEE
E
Tr
ans.
Pattern
Ana
l. Mac
h
.
Intell
,
pp
.
2108
-
2120
,
2012
.
[18]
Rodrıgue
z
JA
.
,
Perronnin
F.
,
"
L
oca
l
gr
adi
en
t
hi
stogram
fea
ture
s
for
word
spotti
ng
in
unco
nstrained
handwri
t
ten
document,
" In
In
te
rnational
con
f
ere
nce on fronti
ers i
n
handwrit
i
ng
rec
ogni
ti
on
,
2008.
[19]
Bunke
H.,
Beng
io
S.,
Vinci
a
rell
i
A.
,
"
Offline
re
cogni
ti
on
of
un
constra
in
ed
han
dwritt
en
t
ext
s
using
HM
M
s
and
stat
isti
ca
l
sta
ti
sti
ca
l
la
ngu
age models,
"
I
EE
E
Tr
ans P
attern A
nal
Mac
h
Int
el
l
,
vo
l.
26,
no.
6
,
pp
.
70
9
-
720
,
2004
.
[20]
Marti
U
-
V.,
Bunke
H.
,
"
Us
ing
a
stat
isti
ca
l
l
anguage
m
odel
to
imp
rove
t
he
per
for
m
anc
e
of
an
HM
M
-
base
d
cur
sive
handwri
ti
ng
re
co
gnit
ion
s
y
st
em,
"
Int
J
Pa
tt
ern
Rec
ognit
Arti
f
Intell
,
vol.
15,
no.
1,
p
p
.
65
-
90
,
2001
.
[21]
En
S.,
Petitj
ea
n
C.
,
Ni
colas
S.,
e
t
al
.
,
"
Patt
ern
lo
ca
l
iz
a
ti
on
in
h
istori
cal
docume
n
t
images
vi
a
te
m
pla
t
e
m
at
ch
ing
,"
23rd Int
ernati
on
al
Conf
ere
nce o
n
Pattern
R
ec
og
nit
ion
(
ICPR
)
,
C
anc
un,
pp.
2054
-
2059
,
2016
.
[22]
Moghadda
m
R.
F.,
Cheriet
M.
,
"
Low
qual
i
t
y
doc
um
ent
image
m
odel
ing
and
enh
anc
ement
,
"
Int
ernati
onal
Journal
on
Document
An
aly
sis and Re
cog
nit
ion
,
vol
.
11
,
n
o.
4
,
pp
.
183
-
20
1
,
Mar
2009
.
[23]
Moghadda
m
R.
F.,
RS
LDI
M.
C.
,
"
Restora
ti
on
of
single
-
sided
lo
w
-
qual
ity
docum
ent
images,
"
Pattern
Re
cognition
,
vol.
42
,
no
.
12
,
p
p
.
3355
-
3364
,
D
ec
2009
.
[24]
Marina
i
S.
,
Miott
i
B.
,
Soda
G.
,
"
Digit
a
l
Librari
es
and
Do
cument
Im
age
Ret
ri
eva
l
T
ec
h
n
ique
s:
A
Survey
,
"
Learnin
g Struc
ture
and Sc
he
ma
s
fro
m D
oc
um
e
nts,
S
pringe
rlink
,
no.
375,
p
p.
181
-
204
,
201
1
.
[25]
Shekhar
R.
,
Ja
waha
r
C.
,
"
W
ord
image
ret
r
i
ev
al
using
bag
of
visual
words
,
"
In
Proce
edi
ngs
of
the
10th
IAPR
Inte
rnational
W
orkshop on,
Doc
ument
Ana
ly
sis
Syste
ms
(
DAS)
,
2012.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19
:
4
3
1
1
-
4
3
2
0
4320
[26]
Rotha
ck
er
L.,
Rusinol
M.,
Fink
G.
,
"
Bag
-
of
-
feat
ure
s
HM
M
s
for
segm
ent
at
ion
-
fr
ee
word
sp
ott
ing
in
handwri
tt
en
documents,
"
In
Proce
ed
ings
of
t
he
21st
Inte
rnati
onal
Confe
renc
e
on
Document
A
naly
sis
and
Re
co
gnit
ion
(
ICDAR)
,
2013.
[27]
No
wak
E.
,
Juri
e
F.,
Tr
iggs
B.
,
"
Sam
pli
ng
strat
eg
ie
s
for
bag
-
o
f
-
fea
tur
es
image
class
ifi
cation
,
"
in
European
Confe
renc
e
on
C
omputer
Vi
sion
,
Lect
ure
No
te
s in
Computer
Sc
ie
n
ce
(
LN
CS
)
,
vol. 3
954,
pp
.
490
-
50
3
,
2006
.
[28]
Sankar
K.P.
,
Ma
nm
at
ha
R.
,
Jawa
har
C.
V
.
,
"
La
rge
-
sc
al
e
document
image
r
et
ri
ev
a
l
b
y
aut
om
atic
wo
rd
annotati
on
,
"
Inte
rnational
Jo
urnal
on
Docum
ent
Anal
ysis
and
Recogni
t
ion
(
IJ
DAR)
,
vol.
17,
n
o.
1
,
pp
.
1
-
17
,
2
013.
[29]
La
z
ebni
k
S.
,
Sc
hm
id
C.
,
Ponc
e
J.
,
"
Be
y
ond
ba
gs
of
feature
s:
s
pat
i
al
p
y
r
amid
m
at
chi
ng
for
re
cog
nizing
na
tur
al
sce
ne
ca
t
egor
i
es,
"
in
Proceedi
n
gs
of
the
IE
EE
Computer
Socie
ty
Conf
ere
nc
e
o
n
Computer
V
ision
and
Pat
t
ern
Re
cogn
it
ion
,
pp
.
2169
-
2178
,
200
6
.
[30]
El
fak
ir
y
.
,
Kha
i
ss
idi
G.,
Mrab
ti
M.,
et
a
l.
,
"
Ba
g
-
of
-
desc
ript
ors
of
SIF
T
for
Seg
m
ent
at
ion
-
fr
ee
word
sp
ott
ing
i
n
Handwrit
te
n
Ar
abi
c
documents
,
"
Sec
ond
Inte
r
nati
onal
Confe
r
enc
e
on
Natur
al
Sci
en
ce
s
and
Technol
ogy
i
n
Manuscript
Ana
l
ysis,
Proc
ee
d
ing
s
,
Ham
burg,
Ger
m
an
y
(ICNTMA
)
2016.
[31]
Erne
st B
el
lman
R.
Dynamic
pro
gram
ming
,
Princ
et
on
Univ
ersity
Press
,
Rand
Cor
pora
ti
on
,
1957
.
[32]
Erne
st B
el
lman
R.
,
Adapt
i
ve c
on
trol
proce
ss
es:
a
guide
d
tour
,
Pri
nce
ton
Univer
sit
y
Pr
ess
,
1961
.
[33]
Jégou
H.,
Douz
e
M.,
Schm
id
C.
,
"
Produc
t
quan
ti
z
at
ion
for
ne
ar
est
ne
ighbor
sea
rch
,
"
I
EE
ETran
s.
Pattern
Ana
l.
Mac
h.
Int
el
l
,
vo
l.
33,
no.
1
,
pp
.
11
7
-
128
,
2011
.
[34]
Hedja
m
R.
,
Far
rah
i
Mog
hada
m
R.
,
Cher
ie
t
M.
,
"
A
spati
al
l
y
ad
apt
iv
e
statistical
m
et
hod
for
the
bina
ri
za
t
ion
of
histori
c
al
m
anus
cri
p
ts
and
d
egr
a
ded
document
i
m
age
s,"
Pattern
Re
cogn
,
vo
l.
44,
pp
.
2184
-
2196
,
2011
.
[35]
Chen
Y.
,
L
ee
d
ham
G.
,
"
De
c
om
pose
al
gorit
hm
for
thre
sh
oldi
ng
d
egr
aded
historic
al
do
cument
imag
es,
"
IEE
Proc.
–
Vi
s
.
Image
Signal P
r
oce
ss
,
vol
.
152
,
pp
.
702
-
714
,
20
05
.
[36]
Don
H.
S.
,
"
A
noise
at
tri
bu
te
t
hre
sholding
m
et
hod
for
docume
nt
image
bina
ri
za
t
ion,
"
Int.
J
.
Document
Ana
l
.
Re
cogn
,
vo
l.
4,
p
p
.
131
-
138
,
200
1
.
[37]
Shokri
M., T
i
zh
oosh H.,
"
Q(k)
-
b
ase
d
imag
e
thr
esholdi
ng
,
"
in
C
V
R'04
,
pp
.
504
-
50
8
,
2004
.
[38]
Cheri
e
t
M.,
Farr
ahi
Moghadd
am
R.
,
Hedja
m
R.
,
"
A
le
arn
ing
fra
m
ework
for
the
opti
m
iz
ation
an
d
aut
om
at
ion
o
f
document
bin
ariza
t
ion
m
et
hods,"
CVIU
,
vol. 117, pp.
269
-
280
,
20
13
.
[39]
Ll
ados
J.,
Rusiñ
ol
M.,
Fornée
A.
,
et
al
.
,
"
On
the
i
nflue
nc
e
of
word
rep
rese
ntations
for
handwri
tt
en
words
pott
ing
in
histori
c
al
do
cuments,
"
In
t. J
.
Patt
ern
Recogni
t
.
Arti
f. Intell
,
26
,
20
12
.
[40]
Lowe
D.
G.,
"
Distinc
ti
v
e
Im
age
Feat
ure
s
fro
m
Scal
e
-
Inv
ari
a
nt
Ke
y
poin
ts,"
Inte
rnational
Jo
urnal
of
Compu
te
r
Vi
sion
,
vol
.
60
,
pp.
91
-
110
,
200
4
.
[41]
Rusiñol
M.,
Al
dave
rt
D
.
,
Llad
os
R.
J.
,
"
Eff
ici
ent
segm
entati
o
n
-
fre
e
k
e
y
word
spotti
ng
in
hist
oric
a
l
document
col
l
ec
t
ions,"
Pa
t
te
rn R
ec
ogni
ti
on
,
vol
.
48
,
pp
.
545
-
555
,
2015
.
[42]
How
e
H.,
Rat
h
T.
,
R
anmatha
R.
,
"
Boosted
dec
is
ion
tre
es
for
word
rec
ognition
in
handwri
tten
doc
um
ent
ret
rie
v
al
,
”
in
Proce
edi
ngs
of
the
Annual
Inte
rnational
AC
M
SIGIR
Confe
renc
e
on
Re
searc
h
and
Dev
el
opment
in
Informati
on
Re
tri
ev
al
,
pp
.
37
7
—
383
,
2005
.
[43]
Li
ang
Y
.
,
Fa
irh
urst
M.,
Guest
R.
,
"
A
s
y
nthe
si
ze
d
word
app
ro
ac
h
to
word
ret
rie
va
l
in
h
andwri
tt
en
document
s
,
"
Pat
te
rn
Recogni
t
ion
,
vo
l.
45,
no.
12,
pp
.
4225
-
42
36
,
2012
.
[44]
Fis
che
r
A.,
Keller
A.,
Frinken
V.,
et
al
.,
"
L
exi
co
n
-
fre
e
handwri
t
t
en
wor
d
spotti
ng
usin
g
cha
rac
t
er
HM
Ms
,
"
Pat
te
rn
Re
cogn
it
.
Let
t
,
v
ol.
33
,
no
.
7
,
pp
.
934
-
942,
2012
.
[45]
El
fak
ir
Y.,
K
ha
i
ss
idi
G.,
Mrabti
M.,
Chenouni
D
.
,
El
Ya
coubi
M.
,
"
W
ord
spotti
ng
in
handwri
tten
A
rab
ic
do
c
um
ent
s
using ba
g
-
of
-
des
cri
ptors,
"
Contem
porar
y
Engi
ne
ering
Sc
ie
nc
es
,
v
ol.
9
,
no
.
8
,
pp
.
1
349
-
1357
,
2016
.
[46]
Te
rasa
wa
K.
,
T
ana
ka
Y
.
,
"
Slit
st
y
l
e
HO
G
fe
ature
for
docume
nt
image
wo
rd
spotti
ng
,
"
in
Proce
ed
ing
of
th
e
Inte
rnational
Co
nfe
renc
e
on
Doc
ument
Ana
ly
sis
and
Recogni
t
ion
,
pp
.
116
-
120
,
2
009
.
Evaluation Warning : The document was created with Spire.PDF for Python.