I
nte
rna
t
io
na
l J
o
urna
l o
f
I
nfo
rm
a
t
ics a
nd
Co
m
m
un
ica
t
io
n
T
ec
hn
o
lo
g
y
(
I
J
-
I
CT
)
Vo
l.
1
5
,
No
.
4
,
D
e
c
em
b
er
20
2
6
,
p
p
.
1
5
1
0
~1
5
2
2
I
SS
N:
2252
-
8
7
7
6
,
DOI
:
1
0
.
1
1
5
9
1
/iji
ct
.
v
1
5
i
4
.
pp
1
5
1
0
-
1
5
2
2
1510
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ict.
ia
esco
r
e.
co
m
Ara
bic callig
ra
ph
y
style recog
nitio
n wit
h
edge e
nha
ncement
and
a
tt
ention
B
ila
l A
bd
ulra
hm
a
n T
ua
m
a
1,
2
,
F
a
rha
n M
o
ha
m
ed
1,
3
1
D
e
p
a
r
t
me
n
t
o
f
Emer
g
e
n
t
C
o
m
p
u
t
i
n
g
,
F
a
c
u
l
t
y
o
f
C
o
mp
u
t
i
n
g
,
U
n
i
v
e
r
si
t
y
o
f
Te
c
h
n
o
l
o
g
y
M
a
l
a
y
si
a
,
J
o
h
o
r
B
a
h
r
u
,
M
a
l
a
y
si
a
2
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
C
o
l
l
e
g
e
o
f
A
p
p
l
i
e
d
S
c
i
e
n
c
e
,
U
n
i
v
e
r
s
i
t
y
o
f
S
a
m
a
r
r
a
,
S
a
marr
a
,
I
r
a
q
3
M
e
d
i
a
a
n
d
G
a
me
I
n
n
o
v
a
t
i
o
n
C
e
n
t
e
r
o
f
E
x
c
e
l
l
e
n
t
,
I
n
st
i
t
u
t
e
o
f
H
u
ma
n
C
e
n
t
e
r
e
d
E
n
g
i
n
e
e
r
i
n
g
,
U
n
i
v
e
r
si
t
i
Te
k
n
o
l
o
g
i
M
a
l
a
y
si
a
,
K
u
a
l
a
Lu
m
p
u
r
,
M
a
l
a
y
si
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
an
3
1
,
2
0
2
6
R
ev
is
ed
J
u
n
1
8
,
2
0
2
6
Acc
ep
ted
J
u
l 5
,
2
0
2
6
Ara
b
ic
c
a
l
li
g
r
a
p
h
y
s
t
y
le
re
c
o
g
n
i
t
io
n
is
a
c
h
a
l
le
n
g
i
n
g
ta
s
k
b
e
c
a
u
s
e
v
i
su
a
l
l
y
sim
i
la
r
st
y
les
o
f
te
n
s
h
a
re
c
u
r
v
e
d
s
tr
o
k
e
s
,
c
o
n
n
e
c
te
d
st
r
u
c
t
u
re
s
,
a
n
d
s
u
b
tl
e
o
r
n
a
m
e
n
ta
l
v
a
r
iat
i
o
n
s
.
T
h
i
s
s
t
u
d
y
p
r
o
p
o
se
s
a
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
r
a
l
n
e
tw
o
r
k
(CNN
)
a
rc
h
it
e
c
t
u
r
e
t
h
a
t
c
o
m
b
i
n
e
s
le
a
r
n
a
b
le
e
d
g
e
e
n
h
a
n
c
e
m
e
n
t
,
s
q
u
e
e
z
e
-
a
n
d
-
e
x
c
it
a
t
i
o
n
(
S
E)
a
t
te
n
t
i
o
n
,
a
n
d
o
p
ti
o
n
a
l
B
i
LS
T
M
-
b
a
se
d
se
q
u
e
n
ti
a
l
m
o
d
e
l
in
g
fo
r
Ar
a
b
ic
c
a
ll
i
g
r
a
p
h
y
s
t
y
le
c
la
s
sif
ica
ti
o
n
.
T
h
e
e
d
g
e
m
o
d
u
le
is
d
e
si
g
n
e
d
t
o
p
re
se
r
v
e
fi
n
e
-
g
ra
i
n
e
d
s
tr
o
k
e
b
o
u
n
d
a
r
ies
,
w
h
il
e
t
h
e
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
e
m
p
h
a
s
ize
s
i
n
fo
rm
a
ti
v
e
c
h
a
n
n
e
l
re
s
p
o
n
se
s
a
n
d
t
h
e
Bi
L
S
T
M
m
o
d
e
ls
w
i
d
t
h
-
wise
c
o
n
te
x
t
u
a
l
d
e
p
e
n
d
e
n
c
i
e
s
.
E
x
p
e
r
ime
n
ts
we
re
c
o
n
d
u
c
te
d
o
n
a
p
u
b
l
ic
Ara
b
ic
c
a
l
l
i
g
ra
p
h
y
s
t
y
l
e
d
a
ta
se
t
u
s
i
n
g
a
s
tra
t
if
ie
d
7
0
/
1
5
/
1
5
t
ra
in
-
v
a
li
d
a
ti
o
n
-
tes
t
s
p
li
t
.
T
h
e
b
e
s
t
c
o
n
fi
g
u
r
a
t
i
o
n
a
c
h
ie
v
e
d
9
8
.
2
2
%
a
c
c
u
ra
c
y
a
n
d
9
8
.
2
0
%
F1
-
sc
o
re
o
n
t
h
e
tes
t
s
u
b
se
t.
T
h
e
se
f
i
n
d
i
n
g
s
i
n
d
ic
a
te
t
h
a
t
a
d
a
p
t
iv
e
e
d
g
e
-
a
wa
r
e
fe
a
t
u
re
lea
r
n
i
n
g
c
a
n
im
p
r
o
v
e
t
h
e
d
i
sc
r
imi
n
a
t
io
n
o
f
v
is
u
a
ll
y
s
imi
l
a
r
Ara
b
ic
c
a
ll
i
g
ra
p
h
y
s
ty
les
.
Ne
v
e
r
t
h
e
les
s
,
b
ro
a
d
e
r
v
a
l
i
d
a
t
i
o
n
o
n
a
d
d
i
ti
o
n
a
l
d
a
tas
e
ts
a
n
d
re
p
e
a
te
d
r
u
n
s
is
st
i
ll
n
e
e
d
e
d
t
o
c
o
n
fir
m
g
e
n
e
ra
l
iza
b
il
it
y
.
K
ey
w
o
r
d
s
:
Ar
ab
ic
ca
llig
r
ap
h
y
Atten
tio
n
m
ec
h
an
is
m
C
NN
E
d
g
e
ex
tr
ac
tio
n
Sty
le
r
ec
o
g
n
itio
n
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Far
h
an
Mo
h
am
e
d
Dep
ar
tm
en
t o
f
E
m
er
g
e
n
t Co
m
p
u
tin
g
,
Facu
lty
o
f
C
o
m
p
u
tin
g
,
Un
iv
er
s
ity
o
f
T
ec
h
n
o
l
o
g
y
Ma
lay
s
ia
J
o
h
o
r
B
ah
r
u
,
Ma
lay
s
ia
E
m
ail:
f
ar
h
an
@
u
tm
.
m
y
1.
I
NT
RO
D
UCT
I
O
N
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
le
r
ec
o
g
n
itio
n
is
an
im
p
o
r
tan
t
p
r
o
b
lem
in
d
o
cu
m
en
t
im
ag
e
an
aly
s
is
,
esp
ec
ially
f
o
r
d
ig
ital
ar
ch
iv
i
n
g
,
h
is
to
r
ic
al
m
an
u
s
cr
ip
t
p
r
o
ce
s
s
in
g
,
an
d
cu
ltu
r
al
h
er
itag
e
r
etr
ie
v
al.
I
t
is
im
p
o
r
tan
t
to
d
is
tin
g
u
is
h
th
is
task
f
r
o
m
o
p
tical
ch
ar
ac
ter
r
ec
o
g
n
itio
n
(
OC
R
)
.
W
h
ile
OC
R
aim
s
to
d
ec
o
d
e
tex
tu
al
co
n
ten
t
at
th
e
ch
ar
ac
ter
o
r
wo
r
d
lev
el,
c
allig
r
ap
h
y
s
ty
le
r
ec
o
g
n
itio
n
f
o
cu
s
es
o
n
id
en
tify
in
g
th
e
v
is
u
al
s
ty
le
ca
teg
o
r
y
o
f
th
e
s
cr
ip
t
im
ag
e
r
eg
ar
d
less
o
f
th
e
wr
itten
tex
t
[
1
]
.
Ar
a
b
ic
s
cr
ip
t
in
tr
o
d
u
ce
s
ad
d
itio
n
al
d
if
f
icu
lty
b
ec
au
s
e
it
is
wr
itten
f
r
o
m
r
i
g
h
t
t
o
lef
t,
ch
a
r
ac
ter
s
ch
an
g
e
s
h
ap
e
d
e
p
en
d
i
n
g
o
n
th
eir
p
o
s
itio
n
i
n
th
e
wo
r
d
,
a
n
d
ca
llig
r
ap
h
i
c
s
ty
les m
ay
s
h
ar
e
h
ig
h
ly
s
im
ilar
v
is
u
al
s
tr
u
ctu
r
es [
2
]
,
[
3
]
.
Ar
ab
ic
ca
llig
r
ap
h
y
is
n
o
t
o
n
ly
a
wr
itin
g
s
y
s
tem
b
u
t
als
o
an
ar
tis
tic
p
r
ac
tice
ch
ar
ac
ter
ized
b
y
h
ar
m
o
n
y
,
r
h
y
th
m
,
an
d
s
ty
lis
tic
v
ar
iatio
n
.
Sty
les
s
u
ch
as
T
h
u
lu
th
,
R
u
q
aa
,
an
d
Nask
h
e
x
h
i
b
it
d
is
tin
ctiv
e
v
is
u
al
p
atter
n
s
,
y
et
m
a
n
y
o
f
th
ese
d
if
f
er
en
ce
s
ar
e
en
co
d
ed
in
s
u
b
tl
e
s
tr
o
k
e
c
u
r
v
atu
r
e,
th
ick
n
ess
,
s
p
ac
in
g
,
an
d
o
r
n
am
en
tatio
n
.
C
o
n
v
en
tio
n
al
h
an
d
c
r
af
ted
d
escr
ip
to
r
s
c
an
ca
p
tu
r
e
p
ar
t
o
f
th
is
s
tr
u
ctu
r
e,
b
u
t
t
h
eir
g
en
er
aliza
tio
n
is
o
f
ten
lim
ited
wh
en
s
ty
les
ar
e
v
is
u
ally
c
lo
s
e
o
r
wh
en
s
am
p
les
ar
e
n
o
is
y
,
d
eg
r
ad
e
d
,
o
r
lo
w
in
r
eso
lu
tio
n
[
4
]
,
[
5
]
.
T
h
is
m
o
ti
v
ates
th
e
u
s
e
o
f
d
ee
p
m
o
d
els
th
at
ca
n
p
r
eser
v
e
b
o
th
s
tr
u
ct
u
r
al
an
d
c
o
n
tex
tu
al
cu
es.
Fig
u
r
e
1
illu
s
tr
ates
th
e
s
am
e
Ar
ab
ic
tex
t
p
r
esen
ted
in
d
if
f
er
e
n
t
ca
llig
r
ap
h
ic
s
ty
les,
h
ig
h
lig
h
tin
g
th
e
v
is
u
al
v
ar
iatio
n
s
th
at
m
ak
e
Ar
ab
ic
ca
llig
r
a
p
h
y
s
ty
le
r
ec
o
g
n
itio
n
ch
allen
g
i
n
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
A
r
a
b
ic
ca
llig
r
a
p
h
y
s
tyle
r
ec
o
g
n
itio
n
w
ith
ed
g
e
en
h
a
n
ce
me
n
t
a
n
d
a
tte
n
tio
n
(
B
ila
l A
b
d
u
lr
a
h
ma
n
Tu
a
ma
)
1511
Fig
u
r
e
1
.
T
h
e
s
am
e
tex
t b
u
t in
d
if
f
er
en
t c
allig
r
a
p
h
y
s
ty
les [
4
]
E
d
g
e
d
etec
tio
n
f
r
o
m
im
ag
es
is
an
ess
en
tial
is
s
u
e
in
im
a
g
e
-
b
a
s
ed
class
if
icatio
n
p
r
o
b
lem
s
,
p
ar
ticu
lar
ly
f
o
r
c
o
m
p
lex
s
ce
n
ar
io
s
lik
e
Ar
ab
ic
ca
llig
r
ap
h
y
a
n
d
v
is
u
al
s
im
ilar
ities
[
5
]
.
Un
lik
e
th
e
p
r
in
t
ed
L
atin
t
y
p
ef
ac
es,
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
les
ex
h
i
b
it
o
v
er
lap
p
i
n
g
s
h
ap
es,
cu
r
v
e
d
lin
es
an
d
s
m
o
o
th
n
ess
in
s
tr
o
k
e
tr
an
s
itio
n
s
th
at
n
o
r
m
ally
c
o
m
p
licate
s
ty
le
cla
s
s
if
icatio
n
.
Stan
d
ar
d
co
n
v
o
lu
t
io
n
al
n
etwo
r
k
s
m
ay
h
av
e
d
i
f
f
icu
lty
in
m
o
d
elin
g
th
ese
f
in
e
-
s
ca
le
d
etails
,
r
el
y
in
g
s
o
lely
o
n
t
h
e
am
p
litu
d
e
o
f
r
aw
p
ix
els
[
6
]
.
T
o
o
v
er
co
m
e
th
is
,
ed
g
e
ex
tr
ac
tio
n
tech
n
i
q
u
es
ar
e
o
f
ten
em
p
lo
y
e
d
to
illu
m
in
a
te
th
e
s
tr
u
ctu
r
al
b
o
u
n
d
ar
ies
a
n
d
o
u
tlin
es
o
f
th
e
g
ly
p
h
s
.
C
o
n
v
en
tio
n
al
m
eth
o
d
s
s
u
ch
as
So
b
el
[
7
]
,
[
8
]
f
ilter
s
ap
p
ly
p
r
ed
e
f
in
ed
k
er
n
els
to
e
n
h
an
ce
h
o
r
izo
n
tal,
v
er
tical,
o
r
d
iag
o
n
al
g
r
ad
ien
ts
,
th
u
s
em
p
h
asizin
g
s
h
ap
e
b
o
u
n
d
ar
ies.
Ho
wev
er
,
th
ese
s
tatic
o
p
er
ato
r
s
ar
e
n
o
t
ad
ap
tiv
e
an
d
m
ay
f
ail
to
ca
p
tu
r
e
s
ty
le
-
s
p
ec
if
ic
s
tr
o
k
e
v
a
r
i
atio
n
s
in
d
iv
er
s
e
ca
llig
r
a
p
h
y
ty
p
es.
I
n
c
o
n
tr
ast,
lear
n
ab
le
ed
g
e
e
x
tr
ac
to
r
s
in
te
g
r
ated
in
to
d
ee
p
n
e
u
r
al
n
etwo
r
k
s
p
r
o
v
i
d
e
a
d
ata
-
d
r
iv
en
alte
r
n
ativ
e
[
9
]
.
T
h
an
k
s
to
lear
n
in
g
p
r
o
p
er
ed
g
e
-
en
h
an
cin
g
f
ilter
s
in
tr
ain
in
g
,
th
ey
ar
e
ab
le
to
s
elec
tiv
ely
em
p
h
asize
th
e
m
o
s
t
d
is
cr
im
in
ativ
e
s
tr
o
k
e
p
atter
n
s
f
o
r
d
if
f
er
e
n
t
s
ty
les.
T
h
is
n
o
t o
n
ly
m
a
k
es
f
ea
tu
r
es
m
o
r
e
r
o
b
u
s
t,
b
u
t
it
also
m
ak
es
m
o
d
els m
o
r
e
in
ter
p
r
eta
b
le
as s
tr
u
ctu
r
al
cu
es th
at
ar
e
n
ee
d
ed
t
o
d
is
cr
im
in
ate
s
im
ilar
ty
p
ef
ac
es a
r
e
r
etain
ed
.
An
ex
am
p
le
o
f
ed
g
e
d
etec
tio
n
o
p
e
r
atio
n
o
n
A
r
ab
ic
letter
s
is
p
r
esen
ted
in
Fig
u
r
e
2
.
Fig
u
r
e
2
.
E
x
am
p
le
o
f
ed
g
e
ex
t
r
ac
tio
n
I
n
co
m
p
lex
s
cr
ip
ts
s
u
ch
as
Ar
ab
ic,
atten
tio
n
m
ec
h
an
is
m
s
h
elp
th
e
n
etwo
r
k
f
o
c
u
s
o
n
th
e
m
o
s
t
in
f
o
r
m
ativ
e
v
is
u
al
cu
es,
wh
ile
s
eq
u
en
ce
m
o
d
elin
g
ca
n
ca
p
tu
r
e
co
n
tex
tu
al
d
ep
en
d
en
cies
b
e
twee
n
n
eig
h
b
o
r
in
g
s
tr
u
ctu
r
es.
SE
atten
tio
n
im
p
r
o
v
es
ch
an
n
el
-
wis
e
f
ea
tu
r
e
r
ec
alib
r
atio
n
b
y
em
p
h
asizin
g
i
n
f
o
r
m
ativ
e
r
esp
o
n
s
es
an
d
s
u
p
p
r
ess
in
g
ir
r
elev
an
t
ac
tiv
atio
n
s
[
1
0
]
,
[
1
1
]
.
T
h
is
is
p
ar
ticu
lar
ly
u
s
ef
u
l
in
Ar
ab
ic
ca
llig
r
ap
h
y
,
w
h
er
e
s
m
all
s
tr
u
ctu
r
al
d
if
f
er
en
ce
s
ca
n
h
av
e
s
tr
o
n
g
s
ty
lis
tic
s
ig
n
if
ican
ce
[
1
2
]
.
T
h
is
s
tu
d
y
ev
alu
ates
wh
eth
er
ad
ap
tiv
e
ed
g
e
en
h
an
ce
m
en
t,
ch
an
n
el
atten
tio
n
,
an
d
wid
th
-
wis
e
s
eq
u
en
tial
m
o
d
elin
g
ca
n
im
p
r
o
v
e
Ar
a
b
ic
ca
llig
r
ap
h
y
s
ty
le
class
if
icatio
n
.
T
h
e
test
ed
v
ar
ian
ts
in
clu
d
e
a
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN
)
with
atten
tio
n
,
a
C
NN
with
So
b
el
ed
g
e
m
a
p
s
,
a
C
NN
with
a
lear
n
ab
le
ed
g
e
ex
tr
ac
to
r
,
a
n
d
a
b
aselin
e
C
NN
w
ith
o
u
t
s
eq
u
en
tial
m
o
d
elin
g
.
C
o
m
p
ar
e
d
with
ea
r
lier
s
tu
d
ies
th
at
em
p
h
asized
h
a
n
d
cr
a
f
ted
f
ea
t
u
r
es,
co
n
v
e
n
tio
n
al
class
if
ier
s
,
o
r
tr
an
s
f
er
lear
n
in
g
,
th
e
p
r
esen
t
wo
r
k
in
v
esti
g
ates
th
e
jo
in
t
r
o
le
o
f
lear
n
ab
le
e
d
g
e
-
awa
r
e
en
h
a
n
ce
m
en
t,
atte
n
tio
n
-
g
u
id
ed
f
ea
tu
r
e
r
ef
in
em
en
t,
an
d
co
n
tex
t
u
al
m
o
d
elin
g
i
n
a
u
n
if
ied
ar
c
h
itectu
r
e.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
s
o
f
th
is
wo
r
k
ar
e
s
u
m
m
ar
ized
as f
o
ll
o
ws
:
−
W
e
d
esig
n
a
n
o
v
el
C
NN
ar
ch
itectu
r
e
f
o
r
A
r
ab
ic
f
o
n
t
class
if
icatio
n
th
at
in
teg
r
ates
atten
tio
n
an
d
ed
g
e
en
h
an
ce
m
e
n
t m
o
d
u
les
.
−
W
e
co
m
p
ar
e
th
e
ef
f
ec
tiv
e
n
ess
o
f
f
ix
ed
(
So
b
el)
an
d
lear
n
a
b
le
ed
g
e
ex
tr
ac
tio
n
in
h
i
g
h
lig
h
tin
g
s
tr
u
ctu
r
al
f
o
n
t f
ea
tu
r
es
.
−
W
e
d
em
o
n
s
tr
ate
th
e
b
en
e
f
it o
f
B
iLST
M
lay
er
s
in
m
o
d
elin
g
s
p
atial
d
ep
en
d
e
n
cies f
o
r
class
if
icatio
n
.
−
W
e
p
r
o
v
id
e
a
co
m
p
r
e
h
en
s
iv
e
ev
alu
atio
n
o
n
a
p
u
b
lic
Ar
ab
ic
f
o
n
t d
ataset
with
class
b
alan
cin
g
an
d
s
tr
atif
ied
v
alid
atio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
4
,
D
e
c
em
b
er
20
2
6
:
1
5
1
0
-
1
5
2
2
1512
T
h
e
r
est
o
f
t
h
is
p
ap
er
is
o
r
g
an
ized
as
f
o
llo
ws:
Sect
io
n
2
r
e
v
iews
th
e
r
elate
d
wo
r
k
a
n
d
h
ig
h
lig
h
ts
th
e
tech
n
iq
u
es
o
f
e
x
is
tin
g
ap
p
r
o
ac
h
es.
Sectio
n
3
d
escr
ib
es
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
,
in
clu
d
in
g
d
ataset
p
r
ep
ar
atio
n
an
d
m
o
d
el
ar
ch
it
ec
tu
r
e.
Sectio
n
4
p
r
esen
ts
th
e
ex
p
er
im
en
tal
s
etu
p
an
d
r
e
s
u
lts
co
m
p
ar
is
o
n
s
.
Fin
a
lly
,
Sectio
n
5
co
n
clu
d
es t
h
e
p
ap
er
an
d
o
u
tlin
es d
ir
ec
tio
n
s
f
o
r
f
u
tu
r
e
wo
r
k
.
2.
RE
L
AT
E
D
WO
RK
S
Ver
y
lim
ited
wo
r
k
h
as
b
ee
n
co
n
d
u
cte
d
in
th
e
d
ev
elo
p
m
e
n
t
o
f
Ar
ab
ic
ca
llig
r
ap
h
y
r
ec
o
g
n
i
tio
n
f
r
o
m
co
m
p
u
tatio
n
al
an
d
d
ataset
co
n
s
tr
u
ctio
n
p
o
in
ts
of
v
iew.
Ka
o
u
d
ja
et
a
l.
[
1
3
]
in
t
r
o
d
u
ce
d
a
n
ew
s
y
s
tem
f
o
r
th
e
class
if
icatio
n
o
f
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
les
b
ased
o
n
h
an
d
-
m
a
d
e
tex
tu
r
e
-
b
ased
f
ea
tu
r
es
an
d
en
s
em
b
le
lear
n
er
s
.
T
h
e
au
t
h
o
r
s
id
e
n
tifie
d
a
n
d
b
u
ilt
a
n
ew
d
ataset
o
f
1
6
8
5
im
a
g
es
in
n
in
e
c
o
m
m
o
n
s
ty
les
(
e.
g
.
)
Nask
h
,
Diwan
i,
T
h
u
lu
th
,
Ku
f
ic
f
r
o
m
v
a
r
io
u
s
s
o
u
r
ce
s
s
u
ch
as
h
is
to
r
ical
m
a
n
u
s
cr
ip
ts
an
d
p
r
in
te
d
d
o
cu
m
e
n
ts
.
Acc
u
r
ate
an
d
d
is
cr
im
in
ativ
e
v
is
u
al
f
ea
t
u
r
es
wer
e
ex
tr
ac
ted
u
s
in
g
lo
ca
l
p
h
ase
q
u
an
tizatio
n
(
L
PQ)
th
at
is
r
o
b
u
s
t
u
n
d
er
n
o
is
e
an
d
b
lu
r
.
C
lass
if
icatio
n
was
c
o
n
d
u
cte
d
b
y
v
ar
i
o
u
s
class
if
ier
s
:
m
u
lti
-
lay
er
p
er
ce
p
tr
o
n
,
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
i
n
e
,
an
d
k
-
n
ea
r
est
n
eig
h
b
o
r
(
M
L
P,
SVM
an
d
KNN)
,
an
d
d
if
f
er
en
t
d
ec
is
io
n
-
lev
el
f
u
s
io
n
m
eth
o
d
s
wer
e
in
v
esti
g
ated
.
T
h
e
Su
m
r
u
le
co
m
b
in
atio
n
s
h
o
wed
th
e
b
es
t
p
er
f
o
r
m
an
ce
with
a
n
ac
cu
r
ac
y
o
f
9
6
.
9
1
\
%
o
n
av
er
ag
e
an
d
s
u
r
p
ass
ed
in
d
iv
id
u
al
class
if
ier
s
.
An
o
t
h
e
r
s
i
g
n
i
f
ic
an
t
s
t
u
d
y
,
Sa
b
er
et
a
l.
[
1
4
]
in
tr
o
d
u
ce
d
an
o
f
f
lin
e
r
ec
o
g
n
iti
o
n
s
y
s
te
m
o
f
ar
ti
s
tic
A
r
a
b
i
c
ca
ll
ig
r
a
p
h
y
s
cr
ip
ts
h
a
v
i
n
g
c
o
m
p
le
x
s
t
r
u
ct
u
r
al
t
ex
tu
r
es
,
p
att
er
n
s
,
a
n
d
d
i
f
f
e
r
e
n
t
f
o
r
m
s
o
f
l
ett
e
r
s
h
a
p
es.
T
h
e
is
s
u
es
p
o
s
e
d
b
y
d
is
c
r
im
in
ati
n
g
ca
lli
g
r
a
p
h
ic
s
t
y
l
es
s
u
c
h
as
i
n
t
er
-
l
ett
er
o
v
er
la
p
a
n
d
v
is
u
al
s
i
m
il
a
r
it
y
,
w
er
e
ta
ck
le
d
wit
h
th
e
g
en
e
r
ati
o
n
o
f
d
is
ti
n
c
ti
v
e
f
ea
t
u
r
es
f
o
r
ea
c
h
s
cr
ip
t
t
y
p
e
i
n
a
f
ea
t
u
r
e
v
ec
t
o
r
.
W
e
u
s
ed
t
h
e
g
en
eti
c
a
lg
o
r
it
h
m
tec
h
n
i
q
u
e
f
o
r
e
x
t
r
a
cti
n
g
an
o
p
t
im
al
s
et
o
f
f
e
at
u
r
es
a
n
d
an
i
m
a
g
e
r
es
o
l
u
ti
o
n
t
h
at
c
o
u
ld
i
m
p
r
o
v
e
t
h
e
class
i
f
i
ca
t
io
n
r
es
u
l
t.
T
h
e
c
lass
if
i
er
was
a
n
eu
r
a
l
n
etw
o
r
k
.
T
h
e
s
y
s
t
em
w
as
t
este
d
u
s
i
n
g
t
wo
d
at
ase
ts
:
a
lo
ca
l
d
at
ase
t,
a
n
d
t
h
e
f
i
r
s
t
p
u
b
li
c
d
a
taset
t
h
a
t
co
n
s
is
t
o
f
t
e
n
d
i
f
f
er
e
n
t
s
y
n
t
h
e
tic
A
r
a
b
i
c
f
o
n
ts
.
T
h
e
r
e
p
o
r
te
d
er
r
o
r
r
at
es
wer
e
8
.
0
2
a
n
d
7
.
5
5
%
f
o
r
t
h
e
tw
o
ex
p
e
r
i
m
e
n
ts
a
b
o
v
e
r
es
p
e
cti
v
e
ly
,
i
n
d
i
ca
t
in
g
th
at
t
h
e
p
r
o
p
o
s
e
d
m
e
th
o
d
ac
h
ie
v
e
d
g
o
o
d
p
er
f
o
r
m
a
n
c
e
b
o
th
i
n
h
a
n
d
-
wr
itt
en
c
all
ig
r
a
p
h
y
an
d
c
o
m
p
u
te
r
-
g
e
n
e
r
at
e
d
ca
l
li
g
r
a
p
h
y
r
e
c
o
g
n
iti
o
n
.
Ab
o
o
d
et
a
l.
[
1
5
]
in
t
r
o
d
u
ce
d
an
in
itiativ
e
f
o
r
a
s
p
ec
ialized
d
ataset
an
d
tex
t
co
r
p
u
s
to
lev
er
ag
em
ac
h
in
e
lear
n
in
g
ap
p
r
o
ac
h
es.
T
h
e
au
th
o
r
s
co
llecte
d
3
,
4
6
7
im
ag
es
f
o
r
Ar
ab
ic
ca
llig
r
ap
h
ic
letter
s
in
3
2
letter
ty
p
es
an
d
ad
d
itio
n
ally
ex
tr
ac
ted
5
4
4
d
if
f
er
en
t
p
h
r
ases
f
r
o
m
ca
llig
r
ap
h
ic
ar
t
p
iece
s
.
T
h
ey
p
r
o
g
r
am
m
ed
an
a
u
to
m
at
ic
h
an
d
wr
itin
g
an
al
y
s
t
an
d
u
s
ed
I
m
ag
eJ
to
m
a
n
u
ally
a
n
n
o
tate
an
d
o
b
tain
p
o
s
itio
n
d
eltas
f
o
r
ea
ch
letter
.
Fo
r
th
e
class
if
icatio
n
,
two
f
ea
tu
r
e
ex
tr
ac
tio
n
tec
h
n
iq
u
es
wer
e
u
s
ed
,
B
ag
o
f
SUR
F
Featu
r
es
an
d
HOG
f
ea
tu
r
es
co
m
b
in
ed
with
SVM.
T
h
e
HOG
-
b
ased
m
o
d
el
h
a
d
a
b
etter
class
if
icatio
n
p
er
f
o
r
m
an
ce
(
6
0
%)
th
an
th
e
SUR
F
-
b
ased
m
o
d
el
(
5
7
%).
I
m
an
e
et
a
l.
[
1
6
]
p
r
o
p
o
s
ed
an
o
p
tical
f
o
n
t
r
ec
o
g
n
itio
n
(
OFR
)
f
r
am
ewo
r
k
f
o
r
Ar
ab
ic
ca
llig
r
ap
h
y
,
wh
ic
h
o
v
er
c
o
m
es
th
e
lim
itatio
n
s
th
at
ar
e
en
co
u
n
ter
ed
b
y
p
r
ev
io
u
s
m
eth
o
d
s
d
u
e
to
d
eg
r
ad
e
d
an
d
h
is
to
r
ical
d
o
c
u
m
en
ts
.
T
h
e
f
ea
tu
r
es
a
r
e
ex
t
r
a
cted
at
p
i
x
el
lev
el
b
ased
o
n
ed
g
e
d
ir
ec
tio
n
m
atr
i
ce
s
(
E
DM
S)
th
at
en
co
d
e
d
is
tin
c
tiv
e
tex
tu
r
e
p
atter
n
s
in
th
e
ca
llig
r
ap
h
y
f
o
n
ts
.
A
b
ac
k
p
r
o
p
ag
atio
n
n
eu
r
al
n
etwo
r
k
(
B
PNN)
clas
s
if
icatio
n
was
im
p
lem
en
ted
,
wh
ich
ac
h
iev
es
b
etter
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
o
f
f
o
n
t
ty
p
es.
T
h
e
m
o
d
el
o
b
tain
e
d
a
te
x
tu
r
e
g
en
er
atio
n
(
9
2
.
8
%)
a
n
d
r
e
co
g
n
itio
n
ac
cu
r
ac
y
(4
3
.
7
%)
o
n
Ar
a
b
ic
o
u
tp
u
t
d
eg
r
ad
ed
im
ag
es
,
r
esp
ec
tiv
ely
.
Hu
s
s
ein
[
1
7
]
also
p
r
o
p
o
s
ed
th
e
p
r
in
ted
Ar
ab
ic
tex
t
s
et
(
PAT
S
-
A0
1
)
,
wh
ic
h
co
n
tain
s
2
,
7
6
6
tex
t
lin
es
cr
o
p
p
e
d
f
r
o
m
class
ical
Ar
ab
ic
b
o
o
k
s
o
f
eig
h
t
f
o
n
ts
:
Ar
ial,
T
ah
o
m
a,
Ak
h
b
ar
,
T
h
u
lu
th
,
Nask
h
,
Simp
lifie
d
A
r
ab
ic
An
d
alu
s
an
d
tr
ad
itio
n
al
Ar
ab
ic.
Mu
ltip
le
f
ea
tu
r
e
ty
p
es
wer
e
an
aly
ze
d
with
t
wo
n
eu
r
al
m
o
d
els
:
ex
tr
em
e
lear
n
i
n
g
m
ac
h
in
e
(
E
L
M)
an
d
f
ast
lear
n
in
g
n
e
two
r
k
(
FLN)
.
E
x
p
er
im
e
n
ts
r
ev
ea
led
th
at
t
h
e
p
er
f
o
r
m
an
ce
o
f
FLN
an
d
E
L
M
was
s
en
s
itiv
e
to
f
ea
tu
r
e
ty
p
e
an
d
ne
two
r
k
s
ize.
T
h
is
d
em
o
n
s
tr
ates
th
e
p
r
o
m
is
e
o
f
lig
h
tweig
h
t
a
n
d
f
ast
-
lear
n
in
g
m
o
d
els
in
Ar
ab
ic
f
o
n
t
r
ec
o
g
n
itio
n
,
th
o
u
g
h
th
e
r
o
le
o
f
d
iv
e
r
s
e
an
d
co
n
ten
t
-
r
ich
d
atasets
is
cr
itica
l
f
o
r
lear
n
in
g
r
o
b
u
s
t
class
if
ier
s
.
As
f
o
r
tr
an
s
f
er
lear
n
in
g
tech
n
iq
u
es,
Gü
r
er
a
n
d
Gö
k
b
ay
[
1
8
]
a
n
aly
ze
d
two
d
if
f
er
en
t c
allig
r
ap
h
y
d
atasets
to
in
v
esti
g
ate
ca
llig
r
ap
h
ic
wo
r
k
s
.
A
f
ir
s
t d
ataset
with
9
u
n
iq
u
e
ca
llig
r
a
p
h
y
s
ty
les
an
d
th
e
s
ec
o
n
d
o
n
e
f
o
r
th
e
class
if
icatio
n
task
s
,
co
n
s
is
tin
g
o
f
class
es
r
ep
r
esen
tin
g
ch
ar
ac
ter
s
in
d
iv
er
s
e
ca
llig
r
ap
h
i
c
s
ty
les
(
i.e
.
,
n
o
o
v
er
lap
p
in
g
p
atter
n
s
)
.
T
h
e
ex
p
e
r
im
en
tal
r
esu
lts
s
h
o
w
good
ac
cu
r
ac
y
o
f
ab
o
u
t 9
0
% a
n
d
7
9
% f
o
r
F1
s
co
r
es o
f
s
ty
les
an
d
l
etter
class
if
icatio
n
.
Fin
ally
,
th
e
wo
r
k
in
[
1
9
]
attem
p
ted
to
s
tu
d
y
th
e
im
p
ac
t
o
f
ap
p
l
y
in
g
tex
tu
r
e
d
escr
ip
to
r
s
f
o
r
th
e
p
u
r
p
o
s
e
o
f
r
ec
o
g
n
is
in
g
Ar
ab
ic
ar
tis
tic
wr
it
in
g
.
T
h
e
s
ix
p
o
p
u
lar
an
d
lan
d
m
ar
k
tex
tu
r
e
d
escr
ip
to
r
s
(
GL
C
M,
L
B
P,
W
L
D,
L
PQ,
HOG
an
d
B
SIF)
wer
e
u
s
ed
i
n
o
u
r
s
tu
d
y
.
T
h
ese
d
escr
ip
to
r
s
wer
e
f
e
d
to
f
iv
e
d
if
f
er
en
t
m
ac
h
in
e
lear
n
in
g
cl
ass
if
ier
s
(
SVM
)
,
r
an
d
o
m
f
o
r
e
s
t
(
R
F
)
,
d
ec
is
io
n
tr
ee
(
DT
)
,
K
-
n
ea
r
est
n
eig
h
b
o
r
s
(
KNN
)
an
d
lin
ea
r
d
is
cr
im
in
at
e
an
aly
s
is
(
L
DA
)
.
T
h
e
test
is
d
o
n
e
o
n
a
p
r
o
p
r
ietar
y
d
ataset
th
at
co
n
s
is
ts
o
f
n
in
e
s
ty
les
o
f
Ar
ab
ic
ca
llig
r
ap
h
y
.
Fro
m
th
e
ex
p
e
r
im
en
tal
r
esu
lts
,
we
f
in
d
th
at
th
e
f
ilter
-
b
ase
d
f
ea
tu
r
es
ar
e
m
o
r
e
u
s
ef
u
l
th
an
s
tatis
tical
f
ea
tu
r
es
f
o
r
t
h
is
task
.
I
n
p
ar
ticu
lar
,
th
e
h
ig
h
est
ac
cu
r
ac
y
o
f
9
4
.
7
%
wa
s
o
b
tain
ed
with
th
e
co
m
b
in
atio
n
o
f
B
SIF
d
escr
ip
to
r
an
d
SVM
class
if
ier
.
Fin
ally
,
it
is
al
s
o
o
b
s
er
v
ed
th
at
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
was
th
e
wo
r
s
t f
o
r
T
h
u
lu
th
an
d
Mo
h
ak
ik
s
ty
les as a
r
esu
lt
o
f
t
h
eir
s
im
ilar
s
tr
u
ctu
r
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
A
r
a
b
ic
ca
llig
r
a
p
h
y
s
tyle
r
ec
o
g
n
itio
n
w
ith
ed
g
e
en
h
a
n
ce
me
n
t
a
n
d
a
tte
n
tio
n
(
B
ila
l A
b
d
u
lr
a
h
ma
n
Tu
a
ma
)
1513
Ov
er
all,
p
r
e
v
io
u
s
s
tu
d
ies
d
em
o
n
s
tr
ate
th
e
v
al
u
e
o
f
h
a
n
d
cr
af
t
ed
tex
tu
r
e
d
escr
ip
to
r
s
,
class
ica
l
m
ac
h
in
e
lear
n
in
g
,
an
d
tr
an
s
f
er
lear
n
i
n
g
f
o
r
Ar
ab
ic
s
cr
ip
t
an
aly
s
is
.
Ho
wev
er
,
th
r
ee
g
ap
s
r
em
ai
n
.
First,
lim
ited
wo
r
k
h
as
ex
p
licitly
ex
am
in
ed
ad
a
p
tiv
e
ed
g
e
en
h
an
ce
m
en
t
f
o
r
p
r
eser
v
in
g
s
ty
le
-
s
p
ec
if
ic
s
tr
o
k
e
s
tr
u
ctu
r
e
in
Ar
ab
ic
ca
llig
r
ap
h
y
.
Seco
n
d
,
t
h
e
jo
i
n
t
co
n
tr
ib
u
tio
n
o
f
ch
a
n
n
el
a
tten
tio
n
an
d
s
eq
u
en
tial
m
o
d
elin
g
h
as
n
o
t
b
ee
n
s
u
f
f
icien
tly
an
aly
z
ed
f
o
r
f
in
e
-
g
r
ain
ed
s
ty
le
class
if
icatio
n
.
T
h
ir
d
,
m
a
n
y
s
tu
d
ies
r
ep
o
r
t
s
tr
o
n
g
r
esu
lts
b
u
t
p
r
o
v
id
e
lim
ited
d
is
cu
s
s
io
n
o
f
r
ep
r
o
d
u
cib
ilit
y
,
s
tab
ilit
y
,
an
d
th
e
in
ter
p
r
etatio
n
o
f
lear
n
ed
s
tr
u
ctu
r
al
c
u
es.
T
h
is
s
tu
d
y
ad
d
r
ess
es
th
ese
g
ap
s
b
y
c
o
m
b
in
in
g
lear
n
ab
le
ed
g
e
ex
tr
ac
t
io
n
,
SE
atte
n
tio
n
,
a
n
d
B
iLST
M
-
b
ased
co
n
tex
tu
al
m
o
d
elin
g
with
in
a
u
n
if
ied
class
if
icatio
n
f
r
am
ewo
r
k
[
2
0
]
,
[
2
1
]
.
3.
RE
S
E
ARCH
M
E
T
H
O
D
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
was
d
ev
elo
p
ed
to
im
p
r
o
v
e
f
in
e
-
g
r
ain
ed
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
le
class
if
icatio
n
b
y
c
o
m
b
in
in
g
e
x
p
licit
s
tr
u
ctu
r
al
en
h
a
n
ce
m
en
t
with
d
ee
p
v
i
s
u
al
r
ep
r
esen
tatio
n
lear
n
in
g
.
T
h
e
o
v
er
all
p
i
p
elin
e
th
at
s
h
o
wed
in
Fig
u
r
e
3
co
n
s
is
ts
o
f
in
p
u
t
p
r
ep
r
o
ce
s
s
in
g
,
e
d
g
e
ex
tr
ac
tio
n
,
h
ier
a
r
ch
ical
c
o
n
v
o
lu
ti
o
n
al
f
ea
tu
r
e
lear
n
in
g
,
o
p
tio
n
al
wid
th
-
wis
e
s
eq
u
en
ce
m
o
d
elin
g
,
a
n
d
f
i
n
al
class
if
ica
tio
n
.
T
h
e
m
o
d
el
is
s
tr
u
ctu
r
ed
in
a
s
eq
u
en
ce
o
f
m
o
d
u
lar
s
tag
es,
b
eg
in
n
in
g
with
in
p
u
t
p
r
ep
r
o
c
ess
in
g
an
d
ed
g
e
ex
t
r
ac
tio
n
,
f
o
llo
wed
b
y
h
ier
ar
c
h
ical
co
n
v
o
lu
tio
n
al
l
ay
er
s
f
o
r
f
ea
tu
r
e
lear
n
in
g
,
a
n
d
en
d
in
g
with
a
class
if
icatio
n
b
lo
ck
.
B
y
in
co
r
p
o
r
atin
g
b
o
th
s
p
atia
l
a
n
d
co
n
t
ex
tu
al
in
f
o
r
m
atio
n
,
th
e
ar
c
h
itectu
r
e
aim
s
to
o
f
f
er
r
o
b
u
s
t p
er
f
o
r
m
a
n
ce
ac
r
o
s
s
v
ar
ied
ca
llig
r
ap
h
ic
s
ty
les.
Fig
u
r
e
3
.
T
h
e
p
r
o
p
o
s
ed
s
y
s
te
m
ar
ch
itectu
r
e
3
.
1
.
Da
t
a
s
et
I
n
th
is
s
tu
d
y
,
we
u
s
ed
th
e
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
le
cl
ass
if
icatio
n
d
ataset
in
tr
o
d
u
ce
d
in
[
1
3
]
.
T
h
e
d
ataset
was
s
elec
ted
b
ec
au
s
e
it
p
r
o
v
id
es
m
u
ltip
le
c
lass
ical
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
les
with
v
is
ib
le
in
tr
a
-
class
v
ar
iatio
n
an
d
s
u
f
f
i
cien
t
s
am
p
les
f
o
r
s
u
p
e
r
v
is
ed
ev
alu
atio
n
.
T
o
p
r
eser
v
e
class
d
is
tr
ib
u
tio
n
ac
r
o
s
s
s
u
b
s
ets,
th
e
im
ag
es
wer
e
d
iv
id
ed
u
s
in
g
a
s
tr
atif
ied
s
p
lit
o
f
7
0
%
f
o
r
tr
ain
in
g
,
1
5
%
f
o
r
v
ali
d
atio
n
,
an
d
1
5
%
f
o
r
test
in
g
.
T
h
e
v
alid
atio
n
s
u
b
s
et
was
u
s
ed
f
o
r
m
o
d
el
s
elec
tio
n
an
d
ea
r
ly
s
to
p
p
in
g
,
w
h
ile
th
e
test
s
u
b
s
et
wa
s
r
eser
v
ed
o
n
ly
f
o
r
f
in
al
e
v
alu
at
io
n
.
All
im
ag
es
wer
e
r
esized
t
o
2
2
4
×
2
2
4
p
i
x
els
b
ef
o
r
e
tr
ain
in
g
.
No
ad
d
itio
n
al
d
ata
au
g
m
en
tatio
n
was
ap
p
li
ed
in
th
e
cu
r
r
en
t
ex
p
er
im
en
t
s
.
T
h
is
allo
wed
a
co
n
tr
o
lled
ev
alu
atio
n
o
f
th
e
p
r
o
p
o
s
ed
ar
ch
itectu
r
e
o
n
th
e
av
ailab
le
d
ataset,
alth
o
u
g
h
b
r
o
ad
e
r
au
g
m
en
tatio
n
a
n
d
r
e
p
ea
ted
-
r
u
n
an
aly
s
is
r
em
ain
im
p
o
r
tan
t d
ir
ec
tio
n
s
f
o
r
f
u
tu
r
e
wo
r
k
.
T
h
e
d
ataset
o
f
f
e
r
s
s
ev
er
al
im
p
o
r
tan
t c
h
ar
ac
ter
i
s
tics
:
−
Div
er
s
ity
o
f
s
ty
les
: I
t c
o
v
er
s
a
wid
e
s
p
ec
tr
u
m
o
f
class
ical
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
les.
−
Ad
eq
u
ate
s
ize
:
E
ac
h
s
ty
le
i
n
clu
d
es
a
s
u
f
f
icien
t
n
u
m
b
e
r
o
f
lab
eled
im
ag
es
t
o
s
u
p
p
o
r
t
lear
n
in
g
-
b
ased
ap
p
r
o
ac
h
es.
−
I
n
tr
a
-
class
v
ar
iatio
n
: T
h
e
d
ata
s
et
in
clu
d
es
v
ar
iatio
n
s
i
n
s
ca
le
an
d
p
h
r
ase
len
g
th
with
in
ea
c
h
s
ty
le
ca
teg
o
r
y
,
wh
ich
en
r
ich
es th
e
m
o
d
el’
s
ab
ilit
y
to
g
en
er
alize
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
4
,
D
e
c
em
b
er
20
2
6
:
1
5
1
0
-
1
5
2
2
1514
Sp
ec
if
ically
,
th
e
d
ataset
co
n
s
is
ts
o
f
n
in
e
ca
llig
r
a
p
h
ic
s
ty
les:
Nask
h
,
Diwan
i,
Par
s
i,
R
ek
aa
,
T
h
u
lu
th
,
Ma
g
h
r
ib
i,
Ku
f
ic,
Mo
h
a
k
ek
,
a
n
d
Sq
u
ar
e
-
Ku
f
ic,
with
ea
ch
s
t
y
le
r
ep
r
esen
tin
g
a
u
n
iq
u
e
class
lab
el,
as
s
h
o
wn
in
Fig
u
r
e
4
.
A
t
o
tal
o
f
1
6
8
5
im
a
g
es
ar
e
in
clu
d
e
d
,
with
1
8
0
t
o
1
9
5
im
a
g
es
p
er
class
.
T
h
e
im
ag
es
wer
e
c
ap
tu
r
e
d
u
s
in
g
an
1
8
-
m
eg
a
p
ix
el
ca
m
er
a,
en
s
u
r
in
g
h
ig
h
r
eso
lu
tio
n
an
d
clar
ity
.
T
o
en
s
u
r
e
v
i
s
u
al
an
d
co
n
tex
t
u
al
d
iv
er
s
ity
,
s
am
p
les
wer
e
co
llected
f
r
o
m
v
ar
i
o
u
s
s
o
u
r
ce
s
s
u
ch
as
b
o
o
k
s
,
m
an
u
s
cr
ip
ts
,
an
d
o
n
lin
e
r
eso
u
r
ce
s
.
E
ac
h
im
ag
e
was
m
an
u
ally
s
eg
m
en
ted
to
p
r
o
d
u
c
e
p
h
r
ases
o
f
v
ar
y
in
g
len
g
th
s
,
an
d
all
s
am
p
les
wer
e
m
eticu
lo
u
s
ly
ca
teg
o
r
ize
d
ac
co
r
d
in
g
to
th
eir
co
r
r
esp
o
n
d
in
g
c
allig
r
ap
h
y
s
ty
le
.
Fig
u
r
e
4
.
E
x
am
p
les o
f
ca
llig
r
a
p
h
ic
s
ty
les in
th
e
d
ataset
3
.
2
.
E
dg
e
ex
t
r
a
ct
io
n
I
n
th
is
wo
r
k
,
we
co
m
p
ar
e
two
ed
g
e
en
h
an
ce
m
e
n
t
s
tr
ateg
ies:
a
f
ix
ed
So
b
el
-
b
ased
b
r
an
ch
an
d
a
lear
n
ab
le
ed
g
e
e
x
tr
ac
tio
n
b
r
a
n
ch
.
T
h
e
o
b
jectiv
e
o
f
b
o
th
b
r
an
ch
es
is
to
h
ig
h
lig
h
t
s
tr
o
k
e
c
o
n
to
u
r
s
an
d
b
o
u
n
d
ar
y
tr
an
s
itio
n
s
th
at
ar
e
in
f
o
r
m
ativ
e
f
o
r
s
ty
le
d
is
cr
im
in
atio
n
.
We
in
v
esti
g
ate
f
ix
ed
an
d
lear
n
ab
le
ed
g
e
en
h
an
ce
m
en
t
m
o
d
u
les
an
d
co
m
p
ar
e
th
eir
ef
f
icien
cy
in
th
e
ap
p
licatio
n
o
f
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
le
class
if
icatio
n
.
I
n
o
u
r
p
r
esen
t
im
p
lem
en
tatio
n
,
we
u
s
e
s
tan
d
ar
d
So
b
el
f
ilter
s
[
2
2
]
,
[
2
3
]
f
o
r
ed
g
e
d
etec
tio
n
as
d
e
p
icted
in
Fig
u
r
e
5
,
wh
ich
co
n
s
is
t
o
f
t
h
e
h
o
r
izo
n
t
al
an
d
v
er
tical
o
p
e
r
ato
r
s
.
T
h
e
s
e
f
ilter
s
ar
e
ap
p
lied
to
g
r
ay
s
ca
le
v
er
s
io
n
s
o
f
th
e
in
p
u
t im
ag
es to
h
ig
h
lig
h
t
d
ir
ec
tio
n
al
g
r
ad
ien
ts
alo
n
g
b
o
th
ax
e
s.
Fo
r
th
e
So
b
el
-
b
ased
b
r
a
n
ch
,
t
h
e
R
GB
im
ag
e
is
f
ir
s
t
co
n
v
er
t
ed
to
g
r
ay
s
ca
le
an
d
t
h
en
p
r
o
ce
s
s
ed
u
s
in
g
h
o
r
izo
n
tal
a
n
d
v
e
r
tical
So
b
el
k
er
n
els.
T
h
e
r
esu
ltin
g
g
r
a
d
ien
t
r
esp
o
n
s
es
ar
e
co
m
b
in
ed
to
p
r
o
d
u
ce
a
s
in
g
le
ed
g
e
m
ag
n
itu
d
e
m
ap
,
wh
ich
is
co
n
c
aten
ated
with
th
e
o
r
i
g
in
al
R
GB
ch
an
n
els.
Fo
r
th
e
lear
n
a
b
le
b
r
an
ch
,
th
e
n
etwo
r
k
p
r
ed
icts
a
s
o
f
t
ed
g
e
m
ap
d
i
r
e
ctly
f
r
o
m
R
GB
in
p
u
t
th
r
o
u
g
h
tr
ain
ab
le
co
n
v
o
lu
tio
n
al
f
ilter
s
,
en
ab
lin
g
th
e
ed
g
e
r
ep
r
esen
tatio
n
to
ad
a
p
t
to
ca
ll
ig
r
ap
h
y
-
s
p
ec
if
ic
s
tr
o
k
e
p
atter
n
s
.
Fo
r
m
ally
,
let
I
d
e
n
o
te
th
e
in
p
u
t
R
GB
im
ag
e.
T
h
e
lear
n
ab
le
e
d
g
e
e
x
tr
ac
to
r
i
s
d
ef
in
ed
as
(
1
)
:
=
(
2
(
(
(
1
(
)
)
)
)
)
(
1
)
wh
er
e
C
o
n
v
1
an
d
C
o
n
v
2
ar
e
c
o
n
v
o
lu
ti
o
n
al
lay
er
s
,
B
N
d
en
o
tes
b
atch
n
o
r
m
aliza
tio
n
,
σ
(
.
)
is
R
e
L
U
ac
tiv
atio
n
f
u
n
ctio
n
a
n
d
δ
(
.
)
is
Sig
m
o
id
ac
tiv
atio
n
f
u
n
cti
o
n
.
T
h
e
o
u
tp
u
t
E
i
s
a
s
o
f
t
ed
g
e
m
a
p
th
at
is
lear
n
ed
jo
in
tly
with
th
e
class
if
icatio
n
n
etwo
r
k
.
I
n
th
e
So
b
el
-
b
ased
e
d
g
e
en
h
a
n
ce
m
en
t
m
o
d
u
le,
th
e
in
p
u
t
im
ag
es
ar
e
f
ir
s
t
co
n
v
er
ted
to
g
r
a
y
s
ca
le
b
y
av
er
ag
in
g
th
e
R
GB
ch
an
n
els.
Fix
ed
h
o
r
iz
o
n
tal
an
d
v
e
r
tical
So
b
el
k
er
n
els
ar
e
th
en
ap
p
lied
to
ex
tr
ac
t
th
e
ed
g
e
m
ap
in
b
o
t
h
d
ir
ec
tio
n
s
,
wh
ic
h
ar
e
co
m
b
i
n
ed
to
p
r
o
d
u
ce
a
co
m
p
o
s
ite
ed
g
e
m
a
g
n
itu
d
e
m
ap
.
I
n
co
n
tr
ast,
th
e
lear
n
in
g
ed
g
e
ex
tr
ac
ts
p
r
o
ce
s
s
R
GB
in
p
u
t
d
ir
ec
tly
an
d
u
s
es
tr
ain
ab
le
co
n
v
en
es
lay
er
s
to
lear
n
o
p
tim
al
ed
g
e
r
ep
r
esen
tatio
n
s
d
u
r
in
g
tr
ain
in
g
.
T
h
is
allo
ws
th
e
m
o
d
el
to
b
e
f
av
o
r
a
b
ly
d
etec
tin
g
th
e
wo
r
k
-
s
p
ec
if
ic
s
h
ap
e
a
n
d
s
tr
o
k
e
p
atter
n
,
w
h
ich
p
r
o
v
id
es
m
o
r
e
f
le
x
ib
ilit
y
th
an
th
e
s
tatic
ed
g
e
f
ilter
[
2
4
]
.
Fig
u
r
e
5
.
Stan
d
a
r
d
s
o
b
el
f
ilter
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
A
r
a
b
ic
ca
llig
r
a
p
h
y
s
tyle
r
ec
o
g
n
itio
n
w
ith
ed
g
e
en
h
a
n
ce
me
n
t
a
n
d
a
tte
n
tio
n
(
B
ila
l A
b
d
u
lr
a
h
ma
n
Tu
a
ma
)
1515
I
n
co
n
tr
ast to
f
ix
ed
ed
g
e
d
etec
tio
n
m
eth
o
d
,
we
also
ex
p
er
im
en
t w
ith
a
lear
n
ab
le
ed
g
e
ex
tr
ac
to
r
b
ased
o
n
tr
ain
ab
le
co
n
v
o
lu
tio
n
al
lay
er
s
.
T
h
is
m
o
d
u
le
r
ec
eiv
es
r
aw
R
G
B
in
p
u
t
an
d
ap
p
lies
t
wo
co
n
s
ec
u
tiv
e
3
×3
co
n
v
o
l
u
tio
n
al
lay
er
s
with
b
at
ch
n
o
r
m
aliza
tio
n
a
n
d
R
elu
ac
tiv
atio
n
as
s
h
o
wn
in
Fig
u
r
e
6
.
T
h
e
f
in
al
lay
er
p
r
o
d
u
ce
s
a
s
in
g
le
ch
a
n
n
el
e
d
g
e
m
a
p
u
s
in
g
a
s
ig
m
o
i
d
ac
tiv
atio
n
,
allo
win
g
th
e
m
o
d
el
to
lea
r
n
s
o
f
t
ed
g
e
r
ep
r
esen
tatio
n
s
th
at
ar
e
o
p
tim
ized
f
o
r
t
h
e
d
o
w
n
s
tr
ea
m
class
if
icatio
n
ta
s
k
.
T
h
is
lear
n
ab
le
ed
g
e
en
h
an
ce
m
en
t
en
ab
les
th
e
n
etwo
r
k
to
ad
ap
ti
v
ely
f
o
cu
s
o
n
task
r
elev
an
t
co
n
to
u
r
s
an
d
s
tr
o
k
e
p
atter
n
s
,
wh
ich
ar
e
o
f
ten
s
u
b
tle
an
d
co
m
p
lex
in
Ar
a
b
ic
ca
llig
r
ap
h
y
s
ty
les.
Fig
u
r
e
6
.
T
h
e
p
r
o
p
o
s
ed
lea
r
n
a
b
le
ed
g
e
e
x
tr
ac
to
r
Alth
o
u
g
h
b
o
th
b
r
a
n
ch
es
aim
to
em
p
h
asize
s
tr
u
ctu
r
al
in
f
o
r
m
atio
n
,
th
e
lear
n
ab
le
ex
tr
ac
t
o
r
i
s
ex
p
ec
ted
to
b
e
m
o
r
e
f
lex
ib
le
b
ec
au
s
e
it c
an
ad
ap
t to
c
u
r
v
at
u
r
e,
s
tr
o
k
e
th
ick
n
ess
v
ar
iatio
n
,
an
d
s
u
b
tle
co
n
to
u
r
tr
an
s
itio
n
s
th
at
f
ix
ed
g
r
a
d
ien
t
k
er
n
els
m
ay
n
o
t
ca
p
tu
r
e
ef
f
ec
tiv
ely
[
2
5
]
,
[
2
6
]
.
I
n
b
o
t
h
ca
s
es,
th
e
g
o
al
is
to
h
ig
h
lig
h
t
s
tr
o
k
e
co
n
to
u
r
s
an
d
b
o
u
n
d
a
r
y
tr
an
s
itio
n
s
th
at
ar
e
ch
ar
ac
ter
is
tic
o
f
c
allig
r
ap
h
ic
p
atter
n
s
.
T
h
e
o
u
tp
u
t o
f
th
is
s
tag
e
is
an
ed
g
e
m
ap
o
f
s
h
ap
e
(
1
×
2
2
4
×
2
2
4
)
.
T
ab
le1
s
h
o
ws
a
co
m
p
ar
is
o
n
b
etwe
en
So
b
el
an
d
lea
r
n
ab
le
ed
g
e
ex
tr
ac
to
r
s
in
n
u
m
b
er
o
f
asp
ec
ts
.
T
ab
le1
.
C
o
m
p
ar
is
o
n
b
etwe
en
s
o
b
el
an
d
lear
n
ab
le
ed
g
e
ex
tr
a
cto
r
s
A
sp
e
c
t
S
o
b
e
l
e
d
g
e
e
x
t
r
a
c
t
o
r
Le
a
r
n
a
b
l
e
e
d
g
e
e
x
t
r
a
c
t
o
r
I
n
p
u
t
t
y
p
e
C
o
n
v
e
r
t
s
R
G
B
i
n
p
u
t
t
o
g
r
a
y
sca
l
e
b
e
f
o
r
e
p
r
o
c
e
ss
i
n
g
U
ses
r
a
w
R
G
B
i
n
p
u
t
w
i
t
h
t
h
r
e
e
c
h
a
n
n
e
l
s
F
i
l
t
e
r
TY
P
E
F
i
x
e
d
h
o
r
i
z
o
n
t
a
l
a
n
d
v
e
r
t
i
c
a
l
S
o
b
e
l
k
e
r
n
e
l
s
Tr
a
i
n
a
b
l
e
c
o
n
v
o
l
u
t
i
o
n
a
l
f
i
l
t
e
r
s
l
e
a
r
n
e
d
f
r
o
m
d
a
t
a
Ed
g
e
o
r
i
e
n
t
a
t
i
o
n
C
a
p
t
u
r
e
s
o
n
l
y
h
o
r
i
z
o
n
t
a
l
a
n
d
v
e
r
t
i
c
a
l
g
r
a
d
i
e
n
t
s
C
a
n
a
d
a
p
t
i
v
e
l
y
l
e
a
r
n
c
o
m
p
l
e
x
,
m
u
l
t
i
-
d
i
r
e
c
t
i
o
n
a
l
e
d
g
e
f
e
a
t
u
r
e
s
Tr
a
i
n
a
b
i
l
i
t
y
N
o
n
-
t
r
a
i
n
a
b
l
e
(
w
e
i
g
h
t
s
a
r
e
f
i
x
e
d
a
n
d
man
u
a
l
l
y
d
e
f
i
n
e
d
)
F
u
l
l
y
t
r
a
i
n
a
b
l
e
e
n
d
-
to
-
e
n
d
w
i
t
h
i
n
t
h
e
n
e
t
w
o
r
k
A
d
a
p
t
a
b
i
l
i
t
y
t
o
d
a
t
a
C
a
n
n
o
t
a
d
a
p
t
t
o
s
p
e
c
i
f
i
c
s
t
r
o
k
e
v
a
r
i
a
t
i
o
n
s
o
r
f
o
n
t
st
y
l
e
s
Le
a
r
n
s e
d
g
e
p
a
t
t
e
r
n
s
t
a
i
l
o
r
e
d
t
o
t
h
e
d
a
t
a
se
t
a
n
d
t
a
s
k
P
e
r
f
o
r
ma
n
c
e
c
o
m
p
l
e
x
F
o
n
t
s
Li
mi
t
e
d
e
f
f
e
c
t
i
v
e
n
e
ss
o
n
v
i
s
u
a
l
l
y
si
m
i
l
a
r
o
r
h
i
g
h
l
y
st
y
l
i
z
e
d
c
a
l
l
i
g
r
a
p
h
y
H
i
g
h
e
r
a
c
c
u
r
a
c
y
i
n
c
a
p
t
u
r
i
n
g
su
b
t
l
e
d
i
f
f
e
r
e
n
c
e
s
a
n
d
s
t
r
u
c
t
u
r
a
l
d
e
t
a
i
l
s
T
h
e
class
if
icatio
n
b
ac
k
b
o
n
e
c
o
n
s
is
ts
o
f
f
o
u
r
co
n
v
o
l
u
tio
n
al
b
lo
ck
s
f
o
llo
we
d
b
y
b
atc
h
n
o
r
m
aliza
tio
n
,
R
eL
U
ac
tiv
atio
n
,
an
d
o
p
tio
n
al
SE
atten
tio
n
.
SE
atten
tio
n
r
ec
alib
r
ates
ch
an
n
el
-
wis
e
f
ea
tu
r
e
r
esp
o
n
s
es
an
d
h
elp
s
th
e
n
etwo
r
k
f
o
c
u
s
o
n
d
i
s
cr
im
in
ativ
e
v
is
u
al
cu
es.
Giv
en
an
in
ter
m
e
d
iate
f
ea
tu
r
e
te
n
s
o
r
X
in
R
^(
C
×
H
×
W
)
,
th
e
s
q
u
ee
ze
s
tep
co
m
p
u
te
s
as
(
2
)
:
=
1
∑
∑
=
1
=
1
(
,
)
(
2
)
an
d
th
e
e
x
citatio
n
s
tep
lear
n
s
ch
an
n
el
weig
h
ts
,
co
m
p
u
ted
as
(
3
)
:
=
(
₂
(
₁
)
)
(
3
)
T
h
e
r
ec
alib
r
ated
o
u
tp
u
t is th
en
o
b
tain
ed
b
y
(
4
)
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
4
,
D
e
c
em
b
er
20
2
6
:
1
5
1
0
-
1
5
2
2
1516
̃
=
.
(
4
)
wh
e
r
e
X
is
th
e
i
n
p
u
t
f
ea
t
u
r
e
ten
s
o
r
wi
th
C
c
h
a
n
n
els
a
n
d
s
p
at
ial
d
im
en
s
io
n
s
H
×W.
T
h
e
s
q
u
ee
ze
o
p
e
r
a
ti
o
n
co
m
p
u
tes
a
c
h
an
n
el
d
esc
r
i
p
t
o
r
u
s
i
n
g
g
l
o
b
a
l
a
v
e
r
a
g
e
p
o
o
li
n
g
.
T
h
e
e
x
c
ita
ti
o
n
o
p
e
r
at
i
o
n
th
e
n
g
en
er
at
es
ch
an
n
el
-
wis
e
w
ei
g
h
ts
t
h
r
o
u
g
h
two
f
u
l
ly
c
o
n
n
ec
te
d
tr
a
n
s
f
o
r
m
ati
o
n
s
f
o
ll
o
we
d
b
y
R
eL
U
an
d
s
ig
m
o
id
a
cti
v
a
ti
o
n
s
.
Fin
all
y
,
e
ac
h
c
h
a
n
n
el
is
r
es
ca
l
ed
b
y
its
c
o
r
r
es
p
o
n
d
i
n
g
w
ei
g
h
t
to
p
r
o
d
u
ce
t
h
e
r
ec
ali
b
r
at
e
d
f
e
atu
r
e
m
a
.
Af
ter
h
ier
ar
ch
ical
f
ea
tu
r
e
ex
tr
ac
tio
n
,
s
p
atial
d
o
wn
-
s
am
p
lin
g
is
ap
p
lied
to
r
etain
th
e
m
o
s
t
in
f
o
r
m
ativ
e
s
em
an
tic
co
n
ten
t
wh
ile
r
e
d
u
c
in
g
r
eso
lu
tio
n
.
W
h
en
s
eq
u
e
n
t
ial
m
o
d
elin
g
is
en
a
b
led
,
th
e
o
u
tp
u
t
o
f
th
e
last
co
n
v
o
l
u
tio
n
al
b
l
o
ck
is
r
esh
a
p
ed
alo
n
g
th
e
h
o
r
izo
n
tal
a
x
is
an
d
p
r
o
ce
s
s
ed
b
y
a
B
iLST
M.
T
h
is
wid
th
-
wis
e
d
esig
n
is
m
o
tiv
ated
b
y
th
e
co
n
n
ec
t
ed
n
atu
r
e
o
f
Ar
a
b
ic
wr
it
in
g
,
in
wh
ich
n
eig
h
b
o
r
in
g
v
is
u
al
elem
en
ts
alo
n
g
th
e
h
o
r
izo
n
tal
d
ir
ec
tio
n
o
f
ten
ca
r
r
y
co
n
tex
tu
al
s
ty
lis
tic
in
f
o
r
m
atio
n
.
3
.
3
.
Cla
s
s
if
ica
t
io
n
m
o
del
I
n
th
is
wo
r
k
,
we
p
r
o
p
o
s
e
a
n
o
v
el
d
ee
p
lear
n
in
g
ar
c
h
itectu
r
e
tailo
r
ed
f
o
r
th
e
class
if
i
ca
tio
n
o
f
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
les.
On
ce
th
e
in
p
u
t
ten
s
o
r
is
p
r
ep
ar
ed
f
r
o
m
t
h
e
p
r
ev
io
u
s
s
tag
e,
it
i
s
p
ass
ed
th
r
o
u
g
h
a
s
er
ies
o
f
f
o
u
r
co
n
v
o
lu
tio
n
al
b
l
o
ck
s
d
es
ig
n
ed
t
o
e
x
tr
ac
t
in
c
r
ea
s
in
g
ly
ab
s
tr
ac
t
an
d
d
is
cr
im
in
ativ
e
s
p
atial
f
ea
tu
r
es
f
r
o
m
th
e
im
ag
e.
E
ac
h
b
l
o
c
k
c
o
n
s
is
ts
o
f
a
c
o
n
v
o
lu
tio
n
al
la
y
er
f
o
llo
wed
b
y
b
atch
n
o
r
m
ali
za
tio
n
an
d
a
R
elu
ac
tiv
atio
n
f
u
n
ctio
n
.
T
o
e
n
h
an
ce
th
e
m
o
d
el’
s
ab
ilit
y
to
f
o
c
u
s
o
n
in
f
o
r
m
ati
v
e
f
ea
tu
r
e
m
a
p
s
,
ea
ch
b
lo
ck
o
p
tio
n
all
y
in
c
o
r
p
o
r
ates
a
s
q
u
ee
ze
-
a
n
d
-
e
x
citatio
n
(
SE)
atten
tio
n
m
o
d
u
le
[
2
7
]
,
[
2
8
]
th
at
s
h
o
wn
i
n
Fig
u
r
e
7
,
wh
ich
ad
ap
tiv
ely
r
ec
alib
r
ates
ch
an
n
el
-
wis
e
f
ea
tu
r
e
r
esp
o
n
s
es
b
y
m
o
d
elin
g
in
ter
d
ep
en
d
e
n
cies
b
etwe
en
ch
an
n
els.
Af
ter
f
ea
tu
r
e
tr
an
s
f
o
r
m
atio
n
,
s
p
atial
d
o
wn
s
am
p
lin
g
is
p
e
r
f
o
r
m
ed
u
s
in
g
m
ax
p
o
o
lin
g
o
r
ad
ap
tiv
e
av
er
ag
e
p
o
o
lin
g
,
wh
ich
p
r
o
g
r
ess
iv
ely
r
ed
u
ce
s
th
e
s
p
atial
r
eso
lu
tio
n
wh
ile
r
etain
in
g
s
em
an
tic
in
f
o
r
m
atio
n
.
Acr
o
s
s
th
e
f
o
u
r
b
lo
c
k
s
,
th
e
f
e
atu
r
e
r
ep
r
esen
tatio
n
is
tr
an
s
f
o
r
m
ed
f
r
o
m
lo
w
-
lev
el
p
i
x
el
p
a
tter
n
s
to
h
ig
h
-
lev
el
s
em
an
tic
ab
s
tr
ac
tio
n
s
s
u
ch
a
s
ch
ar
ac
ter
ed
g
es,
c
u
r
v
es,
a
n
d
s
ty
lis
tic
elem
en
ts
u
n
iq
u
e
to
s
p
ec
if
ic
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
les.
I
n
co
n
f
i
g
u
r
atio
n
s
wh
er
e
s
eq
u
en
tial
m
o
d
elin
g
is
en
ab
led
,
th
e
o
u
tp
u
t
o
f
th
e
last
co
n
v
o
l
u
tio
n
al
b
lo
ck
(
o
f
s
h
ap
e
B
×5
1
2
×7
×7
)
is
r
esh
ap
ed
to
m
o
d
el
th
e
h
o
r
izo
n
tal
s
tr
u
ctu
r
e
o
f
th
e
im
ag
e
w
h
er
e
(
B
)
r
ep
r
esen
t
th
e
b
atch
s
ize.
S
p
ec
if
ically
,
th
e
ten
s
o
r
is
p
er
m
u
ted
an
d
f
latten
ed
to
f
o
r
m
a
s
eq
u
en
ce
o
f
len
g
th
7
(
wid
th
-
wis
e)
,
wh
er
e
ea
ch
s
te
p
in
th
e
s
eq
u
en
ce
h
as
a
f
ea
t
u
r
e
d
im
e
n
s
io
n
o
f
5
1
2
×7
.
T
h
i
s
s
eq
u
en
ce
is
th
en
p
ass
ed
th
r
o
u
g
h
a
b
id
ir
ec
tio
n
al
L
o
n
g
S
h
o
r
t
-
T
er
m
Me
m
o
r
y
(
B
iLST
M)
lay
er
,
wh
ich
ca
p
tu
r
es
d
ep
en
d
e
n
ce
an
d
r
elatio
n
s
h
ip
o
n
lo
ca
ls
in
b
o
th
f
o
r
war
d
an
d
b
ac
k
war
d
d
ir
ec
tio
n
s
.
T
h
is
m
ec
h
an
is
m
is
p
ar
ticu
lar
ly
u
s
ef
u
l
f
o
r
Ar
ab
ic
ca
llig
r
ap
h
y
,
wh
er
e
th
e
s
h
ap
e
an
d
co
n
n
ec
tiv
ity
o
f
th
e
letter
o
f
ten
d
ep
e
n
d
s
o
n
th
e
r
e
f
er
en
ce
with
in
th
e
wo
r
d
o
r
lin
e
[
2
9
]
.
B
iLST
M
en
ab
les
m
o
d
el
to
in
teg
r
ate
s
p
atia
l
r
ef
er
en
ce
at
th
e
wid
th
o
f
th
e
im
ag
e,
ef
f
ec
tiv
ely
ca
p
tu
r
es st
r
u
ctu
r
al
d
e
p
en
d
e
n
c
e
b
etwe
en
n
eig
h
b
o
r
i
n
g
v
is
u
al
elem
en
ts
.
Fig
u
r
e
7
.
SE
atten
tio
n
m
ec
h
a
n
is
m
T
h
e
f
in
al
class
if
ier
p
r
ed
icts
th
e
ca
llig
r
ap
h
y
s
ty
le
lab
el
u
s
in
g
eith
er
th
e
B
iLST
M
r
ep
r
esen
t
atio
n
o
r
a
f
u
lly
co
n
n
ec
ted
h
e
ad
with
o
u
t
s
eq
u
en
tial
m
o
d
elin
g
.
T
r
ain
i
n
g
is
p
er
f
o
r
m
e
d
with
weig
h
ted
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
,
d
ef
in
ed
as
(
5
)
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
A
r
a
b
ic
ca
llig
r
a
p
h
y
s
tyle
r
ec
o
g
n
itio
n
w
ith
ed
g
e
en
h
a
n
ce
me
n
t
a
n
d
a
tte
n
tio
n
(
B
ila
l A
b
d
u
lr
a
h
ma
n
Tu
a
ma
)
1517
=
−
∑
=
1
(
̂
)
(
5
)
wh
er
e
K
is
th
e
n
u
m
b
er
o
f
class
es,
d
en
o
tes
th
e
class
we
ig
h
t
f
o
r
class
k
an
d
is
th
e
g
r
o
u
n
d
-
tr
u
th
lab
el
.
T
h
is
weig
h
tin
g
r
ed
u
ce
s
th
e
ef
f
ec
t
o
f
class
im
b
alan
ce
an
d
s
u
p
p
o
r
ts
a
f
air
er
co
n
tr
i
b
u
tio
n
o
f
ea
ch
class
d
u
r
in
g
o
p
tim
izatio
n
.
Fin
ally
,
th
e
f
ea
tu
r
e
r
ep
r
esen
tatio
n
is
p
ass
ed
th
r
o
u
g
h
a
class
if
icatio
n
h
ea
d
to
p
r
ed
ic
t
th
e
f
o
n
t
ca
teg
o
r
y
.
I
f
th
e
B
iLST
M
is
u
s
ed
,
th
e
last
h
id
d
en
s
tate
o
f
th
e
B
iLST
M
is
f
ed
in
to
a
f
u
lly
co
n
n
ec
ted
la
y
er
to
g
en
er
ate
lo
g
its
co
r
r
esp
o
n
d
in
g
to
ea
ch
class
.
Oth
er
wis
e,
th
e
f
ea
tu
r
e
m
ap
f
r
o
m
th
e
last
co
n
v
o
lu
tio
n
al
b
lo
c
k
is
f
latten
ed
an
d
p
ass
ed
th
r
o
u
g
h
a
s
er
ies
o
f
f
u
lly
c
o
n
n
ec
te
d
lay
e
r
s
with
d
r
o
p
o
u
t
an
d
R
elu
ac
tiv
atio
n
,
wh
ich
s
er
v
e
to
r
ed
u
ce
o
v
er
-
f
itti
n
g
an
d
in
t
r
o
d
u
ce
n
o
n
-
lin
ea
r
ity
.
T
h
e
f
in
al
o
u
tp
u
t
lay
er
p
r
o
d
u
ce
s
a
v
e
cto
r
o
f
class
lo
g
its
,
wh
ich
ar
e
in
ter
p
r
eted
as u
n
-
n
o
r
m
alize
d
p
r
o
b
ab
ilit
ies o
v
er
th
e
s
et
o
f
Ar
ab
ic
ca
llig
r
a
p
h
y
s
ty
les.
T
h
ese
lo
g
its
ar
e
th
en
u
s
ed
to
co
m
p
u
te
th
e
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
d
u
r
in
g
tr
ai
n
in
g
an
d
to
ass
ig
n
th
e
m
o
s
t
p
r
o
b
ab
le
lab
el
d
u
r
in
g
in
f
er
en
ce
.
T
a
b
le
2
p
r
esen
ts
a
d
etailed
ex
p
lan
atio
n
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el.
T
ab
le
2
.
Descr
ip
tio
n
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
S
t
a
g
e
La
y
e
r
/
O
p
e
r
a
t
i
o
n
S
h
a
p
e
I
n
p
u
t
R
G
B
i
ma
g
e
B
×
3
×
2
2
4
×
2
2
4
Ed
g
e
e
x
t
r
a
c
t
S
o
b
e
l
o
r
l
e
a
r
n
a
b
l
e
e
d
g
e
B
×
1
×
2
2
4
×
2
2
4
C
o
n
c
a
t
e
n
a
t
i
o
n
R
G
B
+
E
d
g
e
c
h
a
n
n
e
l
B
×
4
×
2
2
4
×
2
2
4
B
l
o
c
k
1
C
o
n
v
2
D
(
4
→
6
4
,
3
×
3
)
B
×
6
4
×
2
2
4
×
2
2
4
BN
R
e
l
u
S
EB
l
o
c
k
B
×
6
4
×
1
1
2
×
1
1
2
M
a
x
P
o
o
l
2
d
(
2
×
2
)
B
l
o
c
k
2
C
o
n
v
2
D
(
6
4
→1
2
8
,
3
×
3
)
B
×
1
2
8
×
1
1
2
×
1
1
2
BN
R
e
l
u
S
EB
l
o
c
k
B
×
1
2
8
×
5
6
×
5
6
M
a
x
P
o
o
l
2
d
(
2
×
2
)
B
l
o
c
k
3
C
o
n
v
2
D
(
1
2
8
→
2
5
6
,
3
×
3
)
B
×
2
5
6
×
5
6
×
5
6
BN
R
e
l
u
S
EB
l
o
c
k
(
o
p
t
i
o
n
a
l
)
,
B
×
2
5
6
×
2
8
×
2
8
M
a
x
P
o
o
l
2
d
(
2
×
2
)
B
l
o
c
k
4
C
o
n
v
2
D
(
2
5
6
→
5
1
2
,
3
×
3
)
B
×
5
1
2
×
2
8
×
2
8
BN
R
e
l
u
S
EB
l
o
c
k
(
o
p
t
i
o
n
a
l
)
,
B
×
5
1
2
×
7
×
7
A
d
a
p
t
i
v
e
A
v
g
P
o
o
l
(
7
×
7
)
LSTM
b
r
a
n
c
h
R
e
s
h
a
p
e
:
P
e
r
m
u
t
e
+
F
l
a
t
t
e
n
(
t
o
s
e
q
u
e
n
c
e
)
B
×
7
×
3
5
8
4
B
i
LST
M
(
i
n
p
u
t
=
3
5
8
4
,
h
i
d
d
e
n
=
2
5
6
×
2
)
B
×
7
×
5
1
2
S
e
l
e
c
t
l
a
s
t
s
t
e
p
B
×
5
1
2
C
l
a
s
si
f
i
e
r
(
w
i
t
h
LST
M
)
Li
n
e
a
r
(
5
1
2
→
n
u
m_
c
l
a
sses)
B
×
n
u
m
_
c
l
a
ss
e
s
C
l
a
s
si
f
i
e
r
(
n
o
LSTM
)
F
l
a
t
t
e
n
(
B
×
5
1
2
×
7
×
7
→
B
×
2
5
0
8
8
)
B
×
2
5
0
8
8
Li
n
e
a
r
(
2
5
0
8
8
→
5
1
2
)
B
×
5
1
2
D
r
o
p
o
u
t
(
0
.
4
)
,
R
e
l
u
B
×
5
1
2
Li
n
e
a
r
(
5
1
2
→
n
u
m_
c
l
a
sses)
B
×
n
u
m
_
c
l
a
ss
e
s
O
u
t
p
u
t
C
l
a
s
s l
o
g
i
t
s (r
a
w
s
c
o
r
e
s)
B
×
n
u
m
_
c
l
a
ss
e
s
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
A
co
m
p
r
eh
en
s
iv
e
r
a
n
g
e
o
f
e
x
p
er
im
en
ts
was
co
n
d
u
cted
to
ev
alu
ate
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
C
NN
s
ar
ch
itectu
r
e
p
r
o
p
o
s
ed
f
o
r
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
le
cl
ass
if
icatio
n
.
T
h
ese
ex
p
e
r
im
en
ts
ex
am
in
ed
th
e
co
n
tr
ib
u
tio
n
o
f
v
ar
io
u
s
ar
c
h
itectu
r
al
co
m
p
o
n
en
ts
,
in
cl
u
d
in
g
atten
tio
n
m
ec
h
a
n
is
m
s
,
ed
g
e
ex
t
r
ac
to
r
s
(
b
o
th
h
a
n
d
-
c
r
af
ted
an
d
lear
n
e
d
)
,
an
d
s
eq
u
en
ce
m
o
d
elin
g
,
w
h
ich
u
s
es
B
iLST
M
lay
er
s
f
o
r
o
v
er
all
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
T
h
is
s
ec
tio
n
u
n
d
er
lin
es
ex
p
er
im
en
tal
s
ettin
g
s
,
ab
latio
n
co
n
f
ig
u
r
atio
n
,
an
d
p
er
f
o
r
m
an
ce
r
esu
lts
.
T
h
e
i
n
p
u
t
im
a
g
es
r
ep
r
esen
tin
g
Ar
ab
ic
ca
llig
r
a
p
h
y
ar
e
f
ir
s
t
r
e
s
ized
to
a
u
n
if
o
r
m
s
p
atial
r
eso
lu
tio
n
o
f
2
2
4
×
2
2
4
p
ix
els
to
en
s
u
r
e
co
n
s
is
ten
cy
ac
r
o
s
s
th
e
d
ataset.
T
h
e
im
ag
es
ar
e
th
en
n
o
r
m
alize
d
to
a
r
an
g
e
b
etwe
en
-
1
a
n
d
1
u
s
in
g
m
ea
n
an
d
s
tan
d
ar
d
d
ev
i
atio
n
v
alu
es
o
f
0
.
5
p
er
ch
a
n
n
e
l.
T
h
is
n
o
r
m
aliza
tio
n
s
tep
h
el
p
s
ac
ce
ler
ate
m
o
d
el
co
n
v
er
g
en
ce
d
u
r
in
g
tr
ai
n
in
g
an
d
s
tab
ilizes
th
e
lear
n
in
g
p
r
o
ce
s
s
.
T
h
e
d
ataset
was
s
p
lit
in
to
tr
ain
in
g
,
v
alid
atio
n
,
a
n
d
test
s
u
b
s
ets
u
s
in
g
a
s
tr
atif
ied
s
tr
ateg
y
(
7
0
/1
5
/1
5
)
.
T
h
e
v
alid
atio
n
s
et
w
as
u
s
ed
f
o
r
ea
r
ly
s
to
p
p
in
g
a
n
d
m
o
d
el
s
elec
tio
n
,
wh
ile
th
e
test
s
et
was
r
eser
v
ed
f
o
r
f
in
al
e
v
alu
atio
n
o
n
l
y
.
T
ab
le
3
s
u
m
m
ar
izes
th
e
k
ey
tr
ain
i
n
g
p
a
r
am
eter
s
an
d
h
y
p
e
r
p
ar
a
m
eter
s
u
s
ed
in
o
u
r
ex
p
er
im
en
ts
.
T
o
ev
alu
ate
th
e
in
d
iv
id
u
al
c
o
n
tr
ib
u
tio
n
o
f
th
e
m
ain
ar
ch
i
tectu
r
al
co
m
p
o
n
en
ts
,
we
p
er
f
o
r
m
ed
an
ab
latio
n
s
tu
d
y
u
s
in
g
f
o
u
r
m
o
d
el
v
ar
ian
ts
.
T
h
e
co
m
p
a
r
is
o
n
f
o
cu
s
es
o
n
th
r
ee
f
ac
to
r
s
:
ch
an
n
el
atten
tio
n
th
r
o
u
g
h
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
4
,
D
e
c
em
b
er
20
2
6
:
1
5
1
0
-
1
5
2
2
1518
SE
b
lo
ck
s
,
ed
g
e
e
x
tr
ac
tio
n
t
y
p
e,
an
d
wid
th
-
wis
e
B
iLST
M
m
o
d
elin
g
.
All
r
esu
lts
wer
e
co
m
p
u
ted
o
n
t
h
e
h
eld
-
o
u
t te
s
tin
g
s
u
b
s
et
u
s
in
g
t
h
e
s
am
e
tr
ain
in
g
p
r
o
to
c
o
l.
T
ab
le
3
.
T
r
ai
n
in
g
c
o
n
f
ig
u
r
atio
n
an
d
h
y
p
e
r
p
ar
am
eter
s
P
a
r
a
me
t
e
r
V
a
l
u
e
O
p
t
i
mi
z
e
r
A
d
a
m
I
n
i
t
i
a
l
L
e
a
r
n
i
n
g
R
a
t
e
0
.
0
0
0
1
B
a
t
c
h
S
i
z
e
32
Ep
o
c
h
s (m
a
x
)
20
Ea
r
l
y
S
t
o
p
p
i
n
g
P
a
t
i
e
n
c
e
3
e
p
o
c
h
s
Lo
ss F
u
n
c
t
i
o
n
C
r
o
ss
-
E
n
t
r
o
p
y
-
L
o
ss (w
e
i
g
h
t
e
d
)
D
e
v
i
c
e
G
P
U
(
N
V
I
D
I
A
R
TX
4
0
6
0
)
4
.
1
.
Abla
t
io
n
s
t
ud
y
T
o
ev
alu
ate
th
e
in
d
iv
id
u
al
co
n
tr
ib
u
tio
n
s
o
f
th
e
co
r
e
co
m
p
o
n
en
ts
o
f
o
u
r
p
r
o
p
o
s
ed
ar
c
h
itectu
r
e,
we
co
n
d
u
cte
d
an
ab
latio
n
s
tu
d
y
co
m
p
r
is
in
g
two
d
is
tin
ct
ex
p
er
im
en
tal
s
etu
p
s
.
E
ac
h
co
n
f
ig
u
r
atio
n
is
o
lates
o
r
co
m
b
in
es
s
p
ec
if
ic
m
o
d
u
les:
a
tten
tio
n
(
SEBl
o
ck
)
with
s
eq
u
en
ce
m
o
d
elin
g
v
ia
B
iLST
M
an
d
e
d
g
e
ex
tr
ac
tio
n
(
f
ix
ed
o
r
lear
n
ab
le)
.
T
h
is
s
ec
tio
n
p
r
o
v
i
d
es
a
d
etailed
an
aly
s
is
o
f
ea
ch
ex
p
er
im
en
t
an
d
its
im
p
ac
t
o
n
p
er
f
o
r
m
an
ce
in
ter
m
s
o
f
ac
cu
r
ac
y
an
d
F1
-
s
co
r
e
[
3
0
]
.
T
ab
l
e
4
s
h
o
ws
a
co
m
p
ar
is
o
n
o
f
m
o
d
el
v
ar
ian
ts
with
d
if
f
er
e
n
t a
tten
tio
n
,
ed
g
e
ex
tr
ac
tio
n
,
an
d
B
iLST
M
co
m
p
o
n
e
n
ts
o
n
th
e
test
in
g
s
u
b
s
et.
T
ab
le
4
.
Mo
d
el
v
ar
ian
ts
an
d
p
er
f
o
r
m
a
n
ce
co
m
p
ar
is
o
n
N
o
.
Ex
p
e
r
i
m
e
n
t
V
a
r
i
a
n
t
A
t
t
e
n
t
i
o
n
(
S
E
B
l
o
c
k
)
Ed
g
e
Ex
t
r
a
c
t
i
o
n
B
i
LST
M
Te
st
A
c
c
u
r
a
c
y
F1
-
sc
o
r
e
1
C
N
N
+
A
t
t
e
n
t
i
o
n
✓
X
✓
0
.
9
7
9
2
0
.
9
4
1
8
2
C
N
N
+
A
t
t
e
n
t
i
o
n
+
S
o
b
e
l
✓
S
o
b
e
l
(
F
i
x
e
d
)
✓
0
.
9
7
6
3
0
.
9
6
7
5
3
C
N
N
+
A
t
t
e
n
t
i
o
n
+
L
e
a
r
n
e
d
E
d
g
e
✓
Le
a
r
n
a
b
l
e
C
N
N
✓
0
.
9
8
2
2
0
.
9
8
2
0
4
C
N
N
–
N
o
LST
M
X
X
X
0
.
9
7
0
3
0
.
9
2
9
0
T
o
ass
ess
th
e
r
o
le
o
f
SE
atte
n
tio
n
an
d
B
iLST
M
m
o
d
elin
g
,
we
co
m
p
ar
ed
th
e
atten
tio
n
-
b
ased
C
NN
with
th
e
b
aselin
e
C
NN
-
No
L
S
T
M
co
n
f
i
g
u
r
atio
n
.
T
h
e
atten
ti
o
n
-
b
ased
m
o
d
el
ac
h
iev
e
d
9
7
.
9
2
%
ac
cu
r
ac
y
an
d
9
4
.
1
8
%
F1
-
s
co
r
e,
wh
er
ea
s
t
h
e
C
NN
-
No
L
STM
co
n
f
ig
u
r
atio
n
ac
h
iev
e
d
9
7
.
0
3
%
ac
c
u
r
ac
y
a
n
d
9
2
.
9
0
%
F1
-
s
co
r
e.
T
h
ese
r
esu
lts
s
u
g
g
est
th
at
ch
an
n
el
r
ec
alib
r
atio
n
im
p
r
o
v
es
f
ea
tu
r
e
d
is
cr
im
in
atio
n
,
wh
ile
wid
th
-
wis
e
co
n
tex
tu
al
m
o
d
elin
g
h
elp
s
p
r
eser
v
e
s
ty
lis
tic
d
ep
en
d
e
n
cies
th
at
a
r
e
d
i
f
f
icu
lt
to
ca
p
t
u
r
e
u
s
in
g
p
u
r
ely
l
o
ca
l
co
n
v
o
l
u
tio
n
al
r
esp
o
n
s
es.
4
.
1
.
1
.
E
x
perim
ent
1
:
E
f
f
ec
t
o
f
a
t
t
ent
io
n
m
ec
ha
nis
m
a
nd
L
ST
M
T
o
ev
alu
ate
th
e
r
o
le
o
f
ch
an
n
e
l
-
wis
e
atten
tio
n
,
we
c
o
m
p
ar
e
d
th
e
b
aselin
e
C
NN
ar
ch
itectu
r
e
with
an
d
with
o
u
t
SE
b
lo
c
k
s
.
As
s
h
o
wn
in
T
ab
le
4
,
th
e
m
o
d
el
with
a
tten
tio
n
m
o
d
u
les
(
C
NN+
Atten
ti
o
n
)
ac
h
iev
ed
an
ac
cu
r
ac
y
o
f
9
7
.
9
2
%
an
d
an
F1
-
s
co
r
e
o
f
9
4
.
1
8
%,
co
m
p
a
r
ed
to
9
7
.
0
3
%
an
d
9
2
.
9
0
%
i
n
th
e
n
o
n
-
atten
tiv
e
co
n
f
ig
u
r
atio
n
(
C
NN
–
No
L
STM
)
.
T
h
is
in
d
icate
s
th
at
SE
b
lo
ck
s
en
h
an
ce
f
ea
tu
r
e
d
is
cr
im
in
atio
n
b
y
r
ewe
ig
h
tin
g
in
f
o
r
m
ativ
e
c
h
an
n
els.
I
n
th
e
s
am
e
co
n
te
x
t,
an
d
also
f
r
o
m
th
e
s
am
e
ex
p
er
im
en
t
we
ca
n
d
et
ec
t
th
e
im
p
o
r
tan
ce
o
f
L
STM
o
n
th
e
r
esu
lts
.
W
ith
o
u
t
L
STM
,
th
e
m
o
d
el
lo
s
t
th
e
ab
ilit
y
to
m
o
d
el
tem
p
o
r
al/s
p
atial
d
ep
en
d
en
cies,
r
ed
u
cin
g
its
o
v
er
all
d
is
cr
im
in
a
tiv
e
p
o
wer
.
Fig
u
r
e
8
s
h
o
ws co
m
p
ar
is
o
n
b
etwe
en
th
e
two
m
o
d
els
4
.
1
.
2
.
E
x
perim
ent
2
:
E
f
f
ec
t
o
f
edg
e
ex
t
r
a
ct
io
n t
y
pe
T
o
d
eter
m
in
e
t
h
e
co
n
t
r
ib
u
tio
n
o
f
ed
g
e
in
f
o
r
m
atio
n
to
o
u
r
class
if
icatio
n
p
er
f
o
r
m
a
n
ce
,
we
also
test
ed
two
ed
g
e
-
e
n
h
an
ce
d
b
aselin
e
C
NN
v
ar
ian
ts
:
o
n
e
with
f
ix
e
d
S
o
b
el
f
ilter
s
an
d
an
o
th
er
in
co
r
p
o
r
atin
g
a
lear
n
a
b
le
ed
g
e
ex
tr
ac
t
o
r
(
b
o
th
in
teg
r
ate
d
in
a
n
SE
atten
tio
n
s
ettin
g
)
.
T
h
e
So
b
el
-
b
ased
c
o
n
f
ig
u
r
ati
o
n
cr
ea
ted
a
s
tatic
ed
g
e
m
a
p
,
wh
ic
h
was
co
n
ca
ten
ated
with
th
e
R
GB
in
p
u
t
to
g
et
4
-
c
h
an
n
el
ten
s
o
r
.
C
o
m
p
ar
ed
with
f
ix
e
d
So
b
el
f
ilter
in
g
(
ac
c
u
r
ac
y
9
7
.
6
3
%,
F1
0
.
9
6
7
5
)
,
th
e
le
ar
n
ab
le
ed
g
e
ex
tr
ac
t
o
r
im
p
r
o
v
ed
p
er
f
o
r
m
an
ce
(
ac
cu
r
ac
y
9
8
.
2
2
%,
F1
0
.
9
8
2
0
)
,
in
d
icatin
g
th
at
ad
ap
tiv
e
ed
g
e
lear
n
in
g
y
ield
s
b
etter
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
f
o
r
Ar
ab
ic
ca
llig
r
ap
h
y
.
Su
ch
im
p
r
o
v
em
en
t
d
em
o
n
s
tr
ates
th
e
b
en
ef
it
o
f
tr
ain
in
g
task
-
s
p
ec
if
ic
e
d
g
e
f
ea
tu
r
es
wh
ich
ar
e
b
etter
ad
ap
ted
to
th
e
d
elic
ate
s
tr
o
k
e
v
ar
iatio
n
s
in
Ar
ab
ic
ca
llig
r
ap
h
y
.
I
t
is
wo
r
th
m
en
tio
n
in
g
th
at
lear
n
in
g
ed
g
e
ex
tr
ac
ti
o
n
a
n
d
in
te
g
r
atin
g
atten
tio
n
m
ec
h
an
is
m
a
n
d
L
STM
m
ad
e
th
e
m
o
d
el
m
o
r
e
r
o
b
u
s
t,
g
e
n
er
aliza
b
le.
Alt
o
g
eth
er
,
th
ese
r
esu
lts
d
em
o
n
s
tr
ate
th
at
ad
ap
tiv
e
ed
g
e
lear
n
in
g
is
s
u
p
er
io
r
to
f
i
x
ed
f
ilter
i
n
g
esp
ec
ially
with
th
e
u
s
e
o
f
Atten
tio
n
m
o
d
elin
g
.
C
o
m
p
ar
is
o
n
b
etwe
en
t
h
e
two
m
o
d
els
is
g
iv
en
in
Fig
u
r
e
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
A
r
a
b
ic
ca
llig
r
a
p
h
y
s
tyle
r
ec
o
g
n
itio
n
w
ith
ed
g
e
en
h
a
n
ce
me
n
t
a
n
d
a
tte
n
tio
n
(
B
ila
l A
b
d
u
lr
a
h
ma
n
Tu
a
ma
)
1519
Fig
u
r
e
8
.
R
esu
lt c
o
m
p
a
r
is
o
n
o
f
ex
p
er
im
e
n
t 1
Fig
u
r
e
9
.
R
esu
lt c
o
m
p
a
r
is
o
n
o
f
ex
p
er
im
e
n
t 2
Ov
er
all,
th
e
q
u
a
n
titativ
e
r
esu
lts
s
h
o
w
a
co
n
s
is
ten
t
p
a
tter
n
:
th
e
f
u
ll
m
o
d
el
th
at
co
m
b
in
es
lear
n
ab
le
ed
g
e
ex
t
r
ac
tio
n
,
S
E
atten
tio
n
,
an
d
B
iLST
M
ac
h
iev
es
th
e
s
tr
o
n
g
est
p
er
f
o
r
m
an
ce
o
n
th
e
cu
r
r
en
t
d
ataset.
T
h
e
b
ase
lin
e
C
NN
r
ea
ch
ed
9
7
.
0
3
%
ac
cu
r
ac
y
a
n
d
9
2
.
9
0
%
F1
-
s
co
r
e
.
W
ith
ad
d
itio
n
al
atten
tio
n
an
d
c
o
n
tex
tu
al
m
o
d
elin
g
,
th
e
s
co
r
es
im
p
r
o
v
e
d
,
a
n
d
th
e
b
e
s
t
co
n
f
ig
u
r
atio
n
ac
h
iev
e
d
9
8
.
2
2
%
ac
cu
r
ac
y
a
n
d
9
8
.
2
0
% F1
-
s
co
r
e.
A
p
lau
s
ib
le
ex
p
lan
atio
n
f
o
r
th
is
tr
en
d
is
th
at
th
e
th
r
ee
co
m
p
o
n
en
ts
p
r
o
v
i
d
e
co
m
p
lem
e
n
tar
y
b
en
ef
its
.
T
h
e
lear
n
ab
le
e
d
g
e
ex
tr
ac
t
o
r
p
r
eser
v
es
f
in
e
-
g
r
ain
ed
s
tr
u
ct
u
r
al
cu
es,
SE
atten
tio
n
em
p
h
asizes
in
f
o
r
m
ativ
e
ch
an
n
el
r
esp
o
n
s
es,
an
d
th
e
B
iLST
M
ca
p
tu
r
es
wid
th
-
wis
e
co
n
tex
tu
al
d
ep
en
d
en
cies
ac
r
o
s
s
co
n
n
ec
ted
s
cr
ip
t
p
atte
r
n
s
.
T
o
g
eth
er
,
th
ese
elem
en
ts
im
p
r
o
v
e
th
e
d
is
cr
im
in
a
tio
n
o
f
v
is
u
ally
s
im
ilar
Ar
ab
ic
ca
llig
r
ap
h
y
s
ty
les.
T
h
is
s
tu
d
y
co
n
d
u
cted
a
s
er
ies
o
f
ex
p
er
im
en
ts
to
e
v
alu
ate
th
e
ef
f
ec
tiv
en
ess
o
f
atte
n
tio
n
m
e
ch
an
is
m
s
,
tem
p
o
r
a
l
m
o
d
elin
g
,
a
n
d
ed
g
e
ex
tr
ac
tio
n
tech
n
iq
u
es
in
Ar
ab
ic
ca
llig
r
a
p
h
y
class
if
icatio
n
.
Star
tin
g
f
r
o
m
a
b
aselin
e
C
NN,
th
e
m
o
d
el
was
p
r
o
g
r
ess
iv
ely
en
h
an
ce
d
with
SE
atten
tio
n
b
lo
ck
s
,
L
STM
u
n
its
,
an
d
t
wo
ty
p
es
o
f
ed
g
e
ex
tr
ac
to
r
s
:
f
ix
ed
So
b
el
f
ilter
s
an
d
a
lear
n
a
b
le
C
NN
-
b
ased
ed
g
e
m
o
d
u
le.
Fig
u
r
e
1
0
s
h
o
ws
a
co
m
p
ar
is
o
n
o
f
all
m
o
d
els in
ac
cu
r
ac
y
an
d
F1
s
co
r
e.
T
h
e
b
aselin
e
C
NN
ac
h
iev
ed
9
7
.
0
3
%
ac
cu
r
ac
y
an
d
0
.
9
2
8
9
F1
-
s
co
r
e.
Ad
d
in
g
SE
atten
tio
n
im
p
r
o
v
e
d
ac
cu
r
ac
y
a
n
d
b
alan
ce
,
r
ea
c
h
in
g
9
7
.
9
2
%/0
.
9
4
1
8
.
I
n
c
o
r
p
o
r
atin
g
L
STM
f
u
r
t
h
er
b
o
o
s
t
ed
p
er
f
o
r
m
a
n
ce
to
9
8
.
0
9
%/0
.
9
5
6
3
,
b
y
m
o
d
elin
g
s
eq
u
en
tial
p
atter
n
s
.
Usi
n
g
So
b
el
f
ilter
s
s
lig
h
tly
r
ed
u
ce
d
ac
c
u
r
ac
y
(
9
7
.
6
3
%)
b
u
t
in
cr
ea
s
ed
F1
-
s
co
r
e
(
0
.
9
6
7
5
)
,
in
d
icatin
g
b
etter
p
r
ec
is
io
n
-
r
ec
all
tr
a
d
e
-
o
f
f
d
u
e
t
o
e
d
g
e
e
n
h
an
ce
m
e
n
t.
T
h
e
b
est
r
esu
lts
ca
m
e
f
r
o
m
th
e
lear
n
ab
le
ed
g
e
ex
tr
ac
to
r
co
m
b
in
ed
with
atten
tio
n
an
d
L
STM
,
ac
h
iev
in
g
9
8
.
2
2
% a
cc
u
r
ac
y
a
n
d
0
.
9
8
2
0
F1
-
s
co
r
e,
o
u
t
p
er
f
o
r
m
in
g
all
o
t
h
er
co
n
f
ig
u
r
atio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.