I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
p
ute
r
E
ng
in
ee
ring
(
I
J
E
CE
)
Vo
l.
11
,
No
.
2
,
A
p
r
il
2
0
2
1
,
p
p
.
1
6
3
4
~
1
6
4
0
I
SS
N:
2088
-
8708
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
ec
e
.
v
11
i
2
.
pp
1
6
3
4
-
1
6
4
0
1634
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Rea
l
-
ti
m
e
Arabic
scene t
ex
t
de
tecti
o
n using
f
ully
con
v
o
lutiona
l
neura
l net
w
o
rk
s
Ra
j
a
e
M
o
u
m
e
n,
Ra
dd
o
ua
ne
Chih
eb,
R
do
ua
n F
a
izi
Na
ti
o
n
a
l
S
c
h
o
o
l
o
f
C
o
m
p
u
ti
n
g
a
n
d
S
y
ste
m
A
n
a
l
y
sis
(
ENS
I
A
S
)
,
M
o
h
a
m
m
e
d
V
Un
iv
e
rsity
,
M
o
ro
c
c
o
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
No
v
28
,
2
0
19
R
ev
i
s
ed
A
u
g
18
,
20
20
A
cc
ep
ted
Oct
1
3
,
2
0
2
0
T
h
e
a
i
m
o
f
th
is
re
se
a
rc
h
is
to
p
ro
p
o
se
a
f
u
ll
y
c
o
n
v
o
lu
ti
o
n
a
l
a
p
p
r
o
a
c
h
t
o
a
d
d
re
ss
th
e
p
r
o
b
lem
o
f
r
eal
-
ti
m
e
sc
e
n
e
tex
t
d
e
tec
ti
o
n
f
o
r
A
ra
b
ic
lan
g
u
a
g
e
.
T
e
x
t
d
e
tec
ti
o
n
is
p
e
rf
o
rm
e
d
u
sin
g
a
tw
o
-
ste
p
s
m
u
lt
i
-
sc
a
le
a
p
p
ro
a
c
h
.
T
h
e
f
irst
ste
p
u
se
s
li
g
h
t
-
w
e
ig
h
ted
f
u
ll
y
c
o
n
v
o
lu
ti
o
n
a
l
n
e
tw
o
rk
:
T
e
x
tBl
o
c
k
De
te
c
to
r
F
CN
,
a
n
a
d
a
p
tati
o
n
o
f
V
G
G
-
1
6
to
e
li
m
in
a
te
n
o
n
-
tex
tu
a
l
e
lem
e
n
t
s,
lo
c
a
li
z
e
w
id
e
sc
a
l
e
tex
t
a
n
d
g
iv
e
te
x
t
sc
a
le
e
sti
m
a
ti
o
n
.
T
h
e
se
c
o
n
d
ste
p
d
e
term
in
e
s
n
a
rro
w
sc
a
le
ra
n
g
e
o
f
te
x
t
u
sin
g
f
u
ll
y
c
o
n
v
o
lu
ti
o
n
a
l
n
e
tw
o
rk
f
o
r
m
a
x
i
m
u
m
p
e
rf
o
r
m
a
n
c
e
.
T
o
e
v
a
lu
a
te
th
e
s
y
ste
m
,
w
e
c
o
n
f
ro
n
t
th
e
re
su
l
ts
o
f
th
e
f
ra
m
e
w
o
rk
to
th
e
re
su
lt
s o
b
tai
n
e
d
w
it
h
sin
g
le V
GG
-
1
6
f
u
ll
y
d
e
p
lo
y
e
d
f
o
r
tex
t
d
e
tec
ti
o
n
i
n
o
n
e
-
sh
o
t;
i
n
a
d
d
it
i
o
n
to
p
re
v
io
u
s
re
su
l
ts
in
th
e
sta
te
-
of
-
th
e
-
a
rt
.
F
o
r
train
i
n
g
a
n
d
tes
ti
n
g
,
w
e
in
it
iate
a
d
a
tas
e
t
o
f
575
ima
g
e
s
m
a
n
u
a
ll
y
p
ro
c
e
ss
e
d
a
lo
n
g
w
it
h
d
a
ta
a
u
g
m
e
n
tatio
n
to
e
n
ric
h
train
in
g
p
r
o
c
e
ss
.
T
h
e
s
y
ste
m
sc
o
re
s
a
p
re
c
isio
n
o
f
0
.
6
5
1
v
s
0
.
6
4
in
t
h
e
sta
te
-
of
-
th
e
-
a
rt
a
n
d
an
F
P
S
o
f
2
4
.
3
v
s 3
1
.
7
f
o
r
a
V
G
G
-
1
6
f
u
ll
y
d
e
p
lo
y
e
d
.
K
ey
w
o
r
d
s
:
A
r
ab
ic
t
ex
t d
etec
tio
n
C
o
n
v
o
lu
tio
n
a
l n
e
u
r
al
n
e
t
w
o
r
k
s
Natu
r
al
la
n
g
u
a
g
e
p
r
o
ce
s
s
in
g
Scen
e
te
x
t d
etec
tio
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
:
R
aj
ae
Mo
u
m
en
Natio
n
al
Sc
h
o
o
l o
f
C
o
m
p
u
ti
n
g
an
d
S
y
s
te
m
An
al
y
s
i
s
Mo
h
a
m
m
ed
V
Un
iv
er
s
it
y
Av
e
n
u
e
Mo
h
a
m
ed
B
en
A
b
d
ell
ah
R
e
g
r
ag
u
i
,
R
ab
at,
Mo
r
o
cc
o
E
m
ail: r
aj
ae
.
m
o
u
m
en
@
u
m
5
s
.
n
et.
m
a
1.
I
NT
RO
D
UCT
I
O
N
Scen
e
tex
t
d
etec
tio
n
s
ta
n
d
s
f
o
r
lo
ca
lizin
g
i
n
a
n
ac
cu
r
ate
w
a
y
e
x
is
tin
g
tex
t
ar
ea
i
n
a
s
ce
n
e
in
n
a
tu
r
al
en
v
ir
o
n
m
e
n
t
w
it
h
n
a
tu
r
al
co
n
d
itio
n
s
.
T
ec
h
n
ical
l
y
,
i
t
co
n
s
is
t
s
o
n
lo
ca
lizi
n
g
t
h
e
b
o
u
n
d
in
g
b
o
x
o
f
ea
ch
tex
tu
al
ele
m
e
n
t.
T
h
er
e
ar
e
m
an
y
ch
al
l
en
g
e
s
in
t
h
is
ta
s
k
;
f
ir
s
t
o
f
all,
th
e
h
eter
o
g
e
n
eit
y
o
f
t
h
e
s
ce
n
e
,
th
at
m
a
y
co
n
tai
n
,
s
ev
er
al
te
x
ts
in
d
i
f
f
er
en
t
s
ca
le
r
an
g
e,
m
u
lti
-
o
r
ien
ted
a
n
d
cu
r
v
ed
;
I
m
a
g
e
q
u
al
it
y
th
at
m
a
y
p
r
esen
t
b
l
u
r
,
lack
o
f
lig
h
t,
etc.
Scen
e
te
x
t
d
etec
tio
n
h
a
s
r
ec
eiv
ed
an
in
cr
ea
s
in
g
a
m
o
u
n
t
o
f
in
ter
est
t
h
r
o
u
g
h
last
y
ea
r
s
,
as
it
i
s
a
m
an
d
ato
r
y
s
tep
f
o
r
s
ce
n
e
s
co
m
p
r
e
h
e
n
s
io
n
an
d
a
n
al
y
s
i
s
u
s
e
d
in
m
an
y
ap
p
licatio
n
s
s
u
c
h
as
v
id
eo
an
d
i
m
a
g
e
u
n
d
er
s
ta
n
d
in
g
,
v
is
u
all
y
i
m
p
ai
r
ed
ass
is
ta
n
ce
,
au
to
m
atic
d
r
i
v
in
g
,
r
o
b
o
t
s
en
s
i
n
g
,
etc.
T
h
e
f
o
cu
s
o
f
r
e
s
ea
r
ch
i
n
th
is
ar
ea
is
to
p
r
o
d
u
ce
a
tr
a
n
s
c
r
ip
tio
n
o
f
s
ce
n
e
te
x
ts
w
it
h
a
h
i
g
h
ac
cu
r
ac
y
a
n
d
a
r
ea
s
o
n
ab
le
v
elo
cit
y
i
n
o
r
d
er
to
ev
o
lv
e
to
w
ar
d
s
a
r
ea
l
-
ti
m
e
ac
c
u
r
ate
tex
t d
etec
tio
n
an
d
r
ec
o
g
n
itio
n
.
Ma
in
s
t
u
d
ies
co
n
ce
r
n
E
n
g
li
s
h
lan
g
u
a
g
e,
an
d
f
e
w
r
esear
c
h
es
f
o
cu
s
o
n
A
r
ab
ic
s
ce
n
e
te
x
t
p
r
o
ce
s
s
in
g
.
A
ct
u
all
y
,
f
o
r
m
a
n
y
y
ea
r
s
,
t
h
e
f
o
cu
s
o
f
r
esear
ch
ar
o
u
n
d
A
r
ab
ic
w
a
s
o
n
tex
t
d
etec
tio
n
in
clo
s
e
m
ater
ial,
s
u
c
h
as
d
o
cu
m
en
ts
,
p
r
i
n
ted
o
r
h
a
n
d
wr
itten
at
a
h
an
d
-
h
eld
d
is
ta
n
ce
r
an
g
e.
T
h
e
p
r
o
m
in
e
n
t
p
ar
ad
ig
m
i
n
t
h
is
ar
ea
i
s
s
eg
m
e
n
tatio
n
o
f
te
x
t
s
,
th
e
n
wo
r
d
s
an
d
id
en
ti
f
icatio
n
o
f
s
i
n
g
le
ch
ar
ac
ter
s
,
m
a
k
in
g
t
h
e
p
r
o
ce
s
s
a
cla
s
s
i
f
ica
t
io
n
p
r
o
b
lem
.
T
h
is
m
o
d
el
s
h
o
w
e
d
its
li
m
its
f
ir
s
t
b
ec
au
s
e
o
f
th
e
le
n
g
t
h
y
p
r
o
ce
s
s
,
an
d
t
h
e
m
u
lti
-
a
m
b
i
g
u
o
u
s
ch
ar
ac
ter
is
tic
o
f
A
r
ab
ic
r
ep
r
esen
tat
io
n
.
Ot
h
er
s
e
g
m
e
n
tat
io
n
f
r
ee
m
o
d
els
s
h
o
w
ed
u
p
lat
er
,
s
u
ch
a
s
h
id
d
en
m
ar
k
o
v
m
o
d
els
(
HM
M)
an
d
n
eu
r
al
n
et
w
o
r
k
s
r
es
u
lti
n
g
i
n
s
p
ee
d
in
g
u
p
p
r
o
ce
s
s
in
g
ti
m
e
a
n
d
ac
h
ie
v
i
n
g
b
etter
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8708
R
ea
l
-
time
A
r
a
b
ic
s
ce
n
e
text
d
e
tectio
n
u
s
in
g
fu
lly
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
ks (
R
a
ja
e
Mo
u
men
)
1635
ac
cu
r
ac
y
.
Mo
r
e
r
ec
en
tl
y
,
th
e
s
p
ec
tr
u
m
o
f
tex
t
a
n
al
y
s
is
e
x
p
an
d
ed
to
co
m
p
u
ter
v
is
io
n
ap
p
licatio
n
s
s
u
ch
as
te
x
t
d
etec
tio
n
an
d
r
ec
o
g
n
itio
n
f
r
o
m
i
m
a
g
es,
s
ce
n
es
a
n
d
v
id
eo
s
.
H
o
w
e
v
er
,
r
ea
l
-
ti
m
e
s
ce
n
e
te
x
t
p
r
o
ce
s
s
in
g
r
eq
u
ir
es
a
h
ig
h
ca
p
ac
it
y
ap
p
r
o
ac
h
in
te
r
m
s
o
f
p
r
o
ce
s
s
i
n
g
f
u
n
ct
io
n
alit
y
,
ef
f
icie
n
c
y
a
n
d
laten
c
y
.
I
n
a
d
d
itio
n
to
m
e
m
o
r
y
ef
f
icien
t a
n
d
f
a
s
t p
r
o
ce
s
s
i
n
g
h
ar
d
w
ar
e,
m
ak
in
g
it d
i
f
f
ic
u
lt
f
o
r
m
o
b
ile
d
ev
ice
s
to
h
o
s
t s
u
c
h
ap
p
lic
atio
n
s
.
Fo
r
m
an
y
y
ea
r
s
,
t
h
e
tr
ad
itio
n
al
s
lid
in
g
w
i
n
d
o
w
-
b
ased
m
et
h
o
d
s
w
er
e
w
id
el
y
u
s
ed
[
1
,
2
]
.
I
t
co
n
s
is
t
s
o
n
m
o
v
in
g
a
m
u
lti
-
s
ca
le
w
in
d
o
w
o
v
er
an
i
m
a
g
e
a
n
d
d
eter
m
i
n
e
i
f
th
e
co
n
te
n
t
is
tex
t
o
r
n
o
n
-
te
x
t.
A
n
o
t
h
er
p
o
p
u
lar
ap
p
r
o
ac
h
is
th
e
co
n
n
ec
ted
co
m
p
o
n
e
n
t
(C
C
)
a
s
u
s
ed
in
[
3
,
4
]
,
an
al
g
o
r
ith
m
d
esig
n
ed
to
ex
tr
ac
t
m
ax
i
m
a
ll
y
s
tab
le
ex
tr
e
m
al
r
eg
io
n
s
(
MSE
R
s
)
as
ch
ar
ac
te
r
ca
n
d
id
ates.
T
h
e
h
ig
h
er
p
r
o
b
ab
ilit
ies
o
f
tex
t
ca
n
d
id
ates
co
r
r
esp
o
n
d
in
g
to
n
o
n
-
te
x
t
ar
e
e
s
ti
m
ated
w
it
h
a
c
h
ar
ac
ter
class
if
ier
.
T
h
ese
m
et
h
o
d
s
o
b
tain
m
o
d
es
t
r
esu
lt
s
b
ec
au
s
e
t
h
e
tas
k
i
s
d
i
v
id
ed
to
s
ev
er
al
s
tep
s
,
te
x
t
li
n
e
d
etec
tio
n
,
an
d
c
h
ar
ac
ter
clas
s
if
icatio
n
.
E
ac
h
s
tep
w
it
h
a
n
er
r
o
r
r
ate,
th
e
in
f
er
r
ed
o
v
er
-
all
er
r
o
r
r
ate
is
s
ig
n
i
f
ica
n
t.
R
ec
e
n
tl
y
,
tex
t
d
etec
tio
n
is
m
ai
n
l
y
u
n
d
er
tak
e
n
w
it
h
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
es
.
T
h
ese
m
et
h
o
d
s
ca
n
b
e
d
iv
id
e
d
in
to
th
r
ee
s
ec
tio
n
s
:
R
e
g
r
ess
i
o
n
-
b
ased
m
e
th
o
d
s
,
s
eg
m
e
n
tatio
n
-
b
ased
m
eth
o
d
s
a
n
d
h
y
b
r
id
m
e
th
o
d
s
.
R
eg
r
es
s
io
n
-
b
ased
m
e
th
o
d
s
u
s
e
th
e
b
o
u
n
d
in
g
b
o
x
co
n
ce
p
t
co
n
s
id
er
in
g
tex
t
ele
m
e
n
ts
a
s
o
b
j
ec
t
an
d
tr
ea
tin
g
te
x
t
d
etec
tio
n
as
an
o
b
j
ec
t
d
etec
tio
n
p
r
o
b
le
m
.
Fo
r
ex
a
m
p
le,
w
o
r
k
i
n
[
5
]
a
n
d
T
ex
tB
o
x
es
[
6
]
th
at
p
r
ed
icts
th
e
tex
t
b
o
x
b
y
ap
p
ly
in
g
a
f
u
ll
y
co
n
v
o
lu
tio
n
al
n
et
w
o
r
k
.
T
ex
tB
o
x
es++
[
7
]
u
s
es
q
u
ad
r
ila
ter
a
l
r
eg
r
ess
io
n
f
o
r
tex
t
d
etec
tio
n
.
E
A
ST
[
8
]
u
s
e
p
ix
el
-
le
v
el
r
eg
r
ess
io
n
f
o
r
d
etec
tio
n
m
u
l
ti
-
o
r
ien
ted
tex
t
s
.
T
h
ese
m
et
h
o
d
s
p
r
esen
t
t
h
e
i
n
co
n
v
en
ien
t
to
r
en
d
er
m
o
d
est
r
es
u
lt
s
w
h
e
n
it
co
m
e
s
to
cu
r
v
ed
tex
t.
R
eg
r
es
s
io
n
-
b
ased
m
et
h
o
d
s
w
o
r
k
i
n
t
w
o
m
eth
o
d
s
,
o
n
e
-
s
ta
g
e
m
et
h
o
d
an
d
t
w
o
-
s
tag
e
m
e
th
o
d
s
;
th
e
t
w
o
-
s
ta
g
e
m
et
h
o
d
co
n
s
id
er
s
a
s
ec
o
n
d
s
tep
f
o
r
r
ef
in
i
n
g
th
e
r
esu
lt
s
o
f
f
ir
s
t
s
tep
.
I
n
liter
atu
r
e
,
th
e
t
w
o
-
s
tep
s
s
tag
e
ac
h
iev
e
s
b
etter
r
esu
lts
th
a
n
th
e
o
n
e
-
s
tep
m
eth
o
d
s
.
Seg
m
en
tatio
n
-
b
ased
m
eth
o
ds
lik
e
[9
-
11]
u
s
u
all
y
p
r
o
ce
ed
b
y
s
e
g
m
e
n
ti
n
g
t
h
e
b
ac
k
g
r
o
u
n
d
an
d
u
s
i
n
g
p
ix
el
-
lev
el
p
r
ed
ictio
n
.
E
x
a
m
p
le
in
P
SENe
t
[
1
2
]
,
au
th
o
r
s
p
r
o
p
o
s
e
a
n
o
v
el
ap
p
r
o
ac
h
to
d
etec
t
tex
t
w
it
h
ar
b
itra
r
y
s
h
ap
es,
P
SENe
t
g
en
er
ates
d
if
f
er
e
n
t
s
ca
le
s
o
f
k
e
r
n
els
f
o
r
ea
ch
tex
t
i
n
s
ta
n
ce
an
d
g
r
ad
u
all
y
e
x
p
an
d
s
th
e
k
er
n
el
to
th
e
co
m
p
lete
s
h
ap
e
o
f
th
e
tex
t
in
s
ta
n
ce
,
s
tu
d
y
in
[
1
3
]
p
r
o
p
o
s
es
a
n
o
v
el
s
eg
m
en
tatio
n
b
ased
o
n
ca
s
ca
d
ed
co
n
v
o
lu
tio
n
n
e
u
r
al
n
et
w
o
r
k
s
.
Au
t
h
o
r
s
in
[
1
4
]
p
r
o
p
o
s
e
to
p
er
f
o
r
m
te
x
t
d
etec
tio
n
w
it
h
a
d
ee
p
ap
p
r
o
ac
h
u
s
in
g
co
n
n
ec
ted
co
m
p
o
n
en
ts
.
A
r
ab
ic
tex
t
d
etec
tio
n
r
esear
c
h
w
o
r
k
s
ar
e
li
m
ited
;
Au
th
o
r
s
in
[
1
5
]
p
r
o
p
o
s
e
a
h
y
b
r
id
ap
p
r
o
ac
h
f
o
r
Far
s
i/
A
r
ab
ic
te
x
t
d
etec
tio
n
an
d
lo
c
aliza
tio
n
i
n
v
id
eo
f
r
a
m
e
s
,
w
h
er
e
t
h
e
i
m
a
g
e
i
s
d
iv
id
ed
in
to
m
ac
r
o
b
lo
ck
s
an
d
f
ed
i
n
to
s
u
p
p
o
r
t v
ec
to
r
m
a
ch
in
e
(
SVM)
clas
s
i
f
ier
to
ca
teg
o
r
ize
th
e
m
i
n
to
tex
t a
n
d
n
o
n
-
tex
t
g
r
o
u
p
.
I
n
[
1
6
]
,
au
th
o
r
s
u
s
e
a
C
NN
-
R
NN
h
y
b
r
id
ar
ch
itect
u
r
e
b
y
tr
a
n
s
cr
ib
i
n
g
co
n
v
o
l
u
tio
n
a
l
f
ea
tu
r
e
s
f
r
o
m
t
h
e
i
n
p
u
t
i
m
a
g
e
to
a
s
eq
u
en
ce
o
f
tar
g
et
lab
el
s
.
I
n
[
1
7
]
,
a
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
e
t
w
o
r
k
is
u
s
ed
as
a
d
ee
p
class
i
f
ier
to
d
etec
t
s
ce
n
e
ch
ar
ac
ter
s
;
t
h
e
n
et
w
o
r
k
is
tr
ai
n
ed
w
ith
d
i
s
ti
n
ct
lear
n
in
g
r
at
es.
I
n
[
1
8
]
,
a
d
ee
p
f
u
ll
y
co
n
v
o
lu
tio
n
al
n
et
w
o
r
k
s
(
FC
N)
m
u
lt
i
-
o
r
ie
n
ted
s
y
s
te
m
f
o
r
r
ea
l
-
ti
m
e
te
x
t
d
etec
tio
n
.
I
n
[
1
9
]
,
au
th
o
r
s
p
r
o
p
o
s
e
a
d
ee
p
s
ce
n
e
te
x
t
d
etec
to
r
f
o
r
A
r
ab
ic
tex
t d
etec
tio
n
.
I
n
th
i
s
p
ap
er
,
w
e
ap
p
l
y
a
m
u
ltis
ca
le
ap
p
r
o
ac
h
to
p
er
f
o
r
m
A
r
ab
ic
s
ce
n
e
tex
t
d
etec
tio
n
.
W
e
u
s
e
th
e
id
ea
b
eh
in
d
th
e
ap
p
r
o
ac
h
p
r
o
p
o
s
ed
in
[
2
0
]
f
o
r
A
r
ab
ic
lan
g
u
ag
e,
a
t
w
o
-
s
tep
s
ap
p
r
o
ac
h
;
f
i
r
s
t
a
n
o
v
el
n
et
w
o
r
k
,
s
ca
le
-
b
ased
r
e
g
io
n
n
et
w
o
r
k
(
S
P
R
N)
a
m
u
lti
s
ca
le
f
r
a
m
e
w
o
r
k
,
ai
m
i
n
g
to
eli
m
in
ate
n
o
n
-
te
x
t
u
al
ele
m
en
ts
o
f
th
e
s
ce
n
e
i
m
ag
e,
p
r
o
v
id
in
g
tex
t
u
al
ar
ea
s
an
d
esti
m
atio
n
o
f
s
c
ale
r
an
g
e
o
f
ea
ch
ele
m
en
t.
A
f
u
ll
y
co
n
v
o
lu
tio
n
al
n
et
w
o
r
k
is
t
h
en
u
s
ed
to
d
eter
m
i
n
e
n
ar
r
o
w
r
an
g
e
tex
t
ar
ea
s
.
T
ex
t
d
e
tectio
n
is
p
r
o
ce
s
s
ed
in
t
w
o
-
s
tep
s
f
o
r
m
u
lti
tex
t
u
al
ar
ea
s
s
ce
n
es.
T
h
e
ad
v
an
ta
g
e
o
f
th
i
s
m
et
h
o
d
is
it
s
s
p
ee
d
i
n
co
m
p
ar
i
s
o
n
w
it
h
o
th
er
m
u
lti
-
s
ca
le
ap
p
r
o
ac
h
es
m
a
k
i
n
g
it
m
o
r
e
s
u
itab
le
w
it
h
r
ea
l
-
ti
m
e
p
r
o
ce
s
s
in
g
.
T
h
e
s
ec
o
n
d
s
tep
u
s
e
s
a
f
u
ll
y
co
n
v
o
l
u
tio
n
a
l
n
eu
r
al
n
et
w
o
r
k
to
lo
ca
lize
te
x
t
i
n
a
n
ac
cu
r
ate
w
a
y
s
i
n
ce
t
h
e
o
u
tp
u
t
o
f
th
e
f
ir
s
t
s
tep
co
n
tai
n
s
n
o
n
o
i
s
e.
W
e
s
h
o
w
t
h
at
th
e
t
w
o
-
s
tep
s
m
et
h
o
d
s
co
r
es
b
etter
r
u
n
ti
m
e
r
es
u
lt
s
th
a
n
o
n
e
-
s
tep
ap
p
r
o
ac
h
es
in
liter
atu
r
e
r
elate
d
to
A
r
ab
ic
p
r
o
ce
s
s
in
g
,
co
m
p
ar
ab
le
r
esu
lts
to
E
n
g
lis
h
s
ce
n
e
tex
t
d
etec
tio
n
an
d
b
etter
r
esu
lts
th
a
n
a
o
n
e
-
s
tep
ap
p
r
o
ac
h
b
ased
o
n
VGG
-
16.
2.
RE
S
E
ARCH
M
E
T
H
O
D
T
o
o
u
r
k
n
o
w
led
g
e,
A
r
ab
ic
te
x
t
d
etec
tio
n
h
as
n
e
v
er
b
ee
n
u
n
d
er
ta
k
en
w
i
th
t
w
o
-
s
tep
s
a
p
p
r
o
ac
h
es.
T
h
er
ef
o
r
e,
w
e
p
r
o
p
o
s
e
to
s
eg
m
en
t
tex
t
d
etec
tio
n
i
n
to
t
w
o
s
tep
s
.
First,
d
etec
tin
g
w
id
e
s
ca
le
r
an
g
e
tex
t,
w
h
ic
h
m
ea
n
s
e
li
m
in
at
in
g
n
o
n
-
te
x
t
u
a
l
ar
ea
s
,
e
s
ti
m
ate
t
h
e
s
ca
le
o
f
ea
ch
te
x
t
in
s
ta
n
ce
,
w
e
w
ill
ca
ll
t
h
is
f
ir
s
t
s
tep
te
x
t
b
lo
ck
d
etec
tio
n
a
n
d
t
h
e
f
ir
s
t
n
et
w
o
r
k
T
ex
tB
lo
ck
L
o
ca
lizer
;
w
e
f
ee
d
t
h
e
r
e
s
u
l
tin
g
i
m
a
g
e
s
t
o
a
tex
t
d
etec
to
r
w
e
n
a
m
e
T
ex
tDetec
to
r
,
w
h
ich
r
o
le
w
i
ll
b
e
d
etec
tin
g
n
ar
r
o
w
s
ca
le
r
an
g
e
te
x
t.
Fi
g
u
r
e
1
d
es
cr
ib
es
th
e
o
v
er
all
p
r
o
ce
s
s
o
f
tex
t d
etec
tio
n
.
I
n
t
h
is
w
o
r
k
,
w
e
ass
u
m
e
th
a
t:
-
E
le
m
e
n
ts
o
f
a
te
x
t b
lo
ck
h
av
e
s
a
m
e
o
r
ien
ta
tio
n
,
s
ize
a
n
d
f
o
n
t
-
B
o
u
n
d
in
g
b
o
x
es a
r
e
q
u
ad
r
ilate
r
al,
b
u
t n
o
t n
ec
es
s
ar
il
y
r
ec
ta
n
g
le
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
SN
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
2
,
A
p
r
il
2
0
2
1
:
1
6
3
4
-
1640
1636
Fig
u
r
e
1
.
P
r
o
ce
s
s
o
f
tex
t d
etec
tio
n
f
r
a
m
e
w
o
r
k
2
.
1
.
I
m
ple
m
ent
a
t
io
n det
a
ils
First
s
tep
is
p
er
f
o
r
m
ed
w
i
th
a
f
u
ll
y
co
n
v
o
lu
tio
n
al
n
et
w
o
r
k
(
FC
N)
n
a
m
ed
T
ex
tB
lo
ck
L
o
ca
lizer
,
a
n
ad
ap
tatio
n
o
f
VGG
-
1
6
w
it
h
a
s
er
ies
o
f
m
o
d
i
f
icatio
n
s
to
k
e
ep
th
e
n
et
w
o
r
k
li
g
h
t
-
w
eig
h
te
d
w
i
th
a
r
eg
ar
d
o
n
ef
f
icien
t
r
u
n
ti
m
e.
VG
G
is
u
s
e
d
f
o
r
m
a
n
y
r
ea
s
o
n
s
,
f
ir
s
t
it
co
n
s
id
er
s
lo
ca
l
an
d
g
lo
b
al
in
f
o
r
m
atio
n
,
it
i
s
tr
ai
n
ed
en
d
-
to
-
en
d
a
n
d
i
t
h
as
s
h
o
w
n
its
e
f
f
icie
n
c
y
i
n
p
i
x
el
lab
eli
n
g
.
Fir
s
t,
T
ex
tB
lo
ck
L
o
ca
lizer
f
ea
tu
r
e
e
x
tr
ac
tio
n
is
d
er
iv
ed
f
r
o
m
V
GG
-
1
6
w
it
h
3
*
3
co
n
v
o
l
u
tio
n
la
y
er
s
s
tac
k
ed
s
tr
aig
h
t
f
o
r
w
ar
d
.
A
s
er
ie
o
f
m
o
d
i
f
icatio
n
s
ar
e
p
r
o
ce
s
s
ed
to
th
e
n
et
w
o
r
k
i
n
o
r
d
er
to
o
p
ti
m
ize
r
u
n
ti
m
e,
s
u
ch
as
eq
u
a
l
c
h
an
n
el
w
i
d
t
h
t
o
in
cr
ea
s
e
f
o
r
w
ar
d
s
p
ee
d
.
Slo
w
i
n
g
o
p
er
atio
n
s
s
u
ch
as
e
x
ce
s
s
i
v
e
g
r
o
u
p
co
n
v
o
lu
tio
n
,
ele
m
e
n
t
-
w
is
e
o
p
er
atio
n
s
a
n
d
n
et
w
o
r
k
f
r
ag
m
e
n
tatio
n
ar
e
o
p
ti
m
ized
t
o
in
cr
ea
s
e
t
h
e
n
et
w
o
r
k
s
p
ee
d
.
Nu
m
b
er
o
f
k
er
n
el
s
i
s
d
r
asti
c
all
y
r
ed
u
ce
d
to
1
6
k
er
n
el
s
,
th
e
q
u
ar
ter
in
VGG
-
1
6
.
W
e
r
ely
o
n
ex
p
er
i
m
e
n
t
al
s
tu
d
ies
i
n
[
2
1
]
,
th
at
u
s
e
less
p
ar
a
m
eter
s
an
d
p
er
f
o
r
m
s
r
an
d
o
m
i
n
itializa
tio
n
f
o
r
s
ce
n
e
f
ea
t
u
r
e
e
x
tr
ac
tio
n
.
Fo
r
f
ea
t
u
r
e
f
u
s
io
n
,
a
1
*
1
co
n
v
o
lu
tio
n
i
s
u
s
ed
to
n
o
r
m
alize
c
h
an
n
el
w
id
t
h
,
an
d
a
d
ec
o
n
v
o
lu
tio
n
la
y
er
to
u
p
-
s
a
m
p
le
s
p
ac
ial
r
eso
lu
t
io
n
.
T
ex
tB
lo
ck
L
o
ca
lizer
FC
N
p
er
f
o
r
m
s
t
w
o
tas
k
s
:
First
ta
s
k
i
s
lo
ca
lizi
n
g
te
x
t
b
l
o
ck
s
,
t
h
e
n
et
w
o
r
k
d
ea
ls
it
a
s
a
class
if
icatio
n
p
r
o
b
lem
b
y
p
er
f
o
r
m
i
n
g
ca
teg
o
r
izatio
n
o
f
tex
t
u
al
a
n
d
n
o
n
-
te
x
tu
a
l
r
eg
io
n
s
;
t
h
is
s
tep
ai
m
s
to
f
i
lter
o
u
t
n
o
n
-
te
x
t
u
al
ar
ea
s
,
th
at
co
n
s
is
t
s
n
o
is
y
an
d
s
lo
w
i
n
g
e
le
m
e
n
t
s
f
o
r
th
e
d
etec
tio
n
to
o
l.
T
ex
t
lo
c
aliza
tio
n
i
s
p
er
f
o
r
m
ed
a
s
a
cla
s
s
i
f
icatio
n
p
r
o
b
lem
at
a
p
ix
el
-
lev
e
l.
No
n
-
te
x
t
u
al
ar
ea
s
ar
e
f
ilter
ed
o
u
t
an
d
th
is
s
te
p
o
u
tp
u
ts
w
id
e
s
ca
le
r
an
g
e
tex
tu
al
in
s
t
a
n
ce
s
.
W
e
m
ak
e
t
h
e
p
r
es
u
m
p
t
io
n
t
h
at
p
ix
els
w
it
h
i
n
t
h
e
b
o
u
n
d
i
n
g
b
o
x
ar
e
p
o
s
iti
v
e
te
x
t
i
n
s
ta
n
ce
;
f
ir
s
t,
b
ec
au
s
e
t
h
e
r
en
d
er
ed
o
u
tp
u
t
is
to
b
e
r
ef
i
n
ed
in
t
h
e
s
ec
o
n
d
s
tep
a
n
d
in
eit
h
er
w
a
y
s
,
r
e
g
io
n
s
b
et
w
ee
n
ch
ar
ac
ter
s
ar
e
d
if
f
er
e
n
t i
n
co
n
tr
ast to
n
o
n
-
tex
tu
al
i
n
s
ta
n
ce
s
.
Seco
n
d
tas
k
is
to
r
e
n
d
er
s
ca
l
e
esti
m
atio
n
o
f
ea
ch
tex
t
in
s
t
an
ce
;
t
h
e
f
r
a
m
e
w
o
r
k
a
f
f
ec
ts
t
o
ea
ch
tex
t
in
s
ta
n
ce
t
h
r
ee
p
o
s
s
ib
le
v
al
u
e
s
(
B
ig
–
No
r
m
al
–
B
ig
)
.
I
n
a
r
e
g
r
ess
i
v
e
w
a
y
,
ca
te
g
o
r
y
o
f
a
b
o
u
n
d
in
g
te
x
t b
o
x
is
d
eter
m
i
n
ed
w
i
th
t
h
e
f
o
llo
w
i
n
g
eq
u
atio
n
ta
k
i
n
g
i
n
to
co
n
s
id
er
atio
n
b
o
u
n
d
in
g
b
o
x
d
i
m
e
n
s
i
o
n
s
:
(
)
=
{
(
)
<
(
)
∈
[
,
]
(
)
>
Stan
d
in
g
f
o
r
th
e
s
u
p
er
io
r
th
r
esh
h
o
ld
to
tex
t o
f
s
m
all
s
ca
le
Stan
d
i
n
g
f
o
r
th
e
s
u
p
er
io
r
th
r
es
h
o
ld
to
tex
t o
f
n
o
r
m
al
s
ca
le
Seco
n
d
s
tep
i
s
t
h
e
T
ex
tDe
tect
o
r
;
it
is
also
an
FC
N
w
it
h
t
h
e
t
ask
o
f
d
etec
ti
n
g
te
x
t
in
an
ac
c
u
r
ate
w
a
y
.
I
t is b
ased
o
n
VGG
-
1
6
f
u
ll
y
d
ep
lo
y
ed
f
o
llo
w
i
n
g
th
e
d
esi
g
n
d
escr
ib
ed
in
Fig
u
r
e
2
.
Fig
u
r
e
2
.
VGG
-
16
a
r
ch
itect
u
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8708
R
ea
l
-
time
A
r
a
b
ic
s
ce
n
e
text
d
e
tectio
n
u
s
in
g
fu
lly
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
ks (
R
a
ja
e
Mo
u
men
)
1637
2.
2
.
Da
t
a
s
et
W
e
co
u
n
t
s
ev
er
al
d
atasets
ai
m
ed
f
o
r
h
an
d
w
r
itte
n
A
r
ab
ic
o
r
o
f
f
-
li
n
e
p
r
in
ted
d
o
cu
m
e
n
ts
s
u
ch
as
A
P
T
I
[
2
2
]
an
d
f
e
w
d
ata
s
ets
f
o
r
A
r
ab
ic
tex
t
d
etec
tio
n
an
d
r
ec
o
g
n
i
tio
n
i
n
v
id
eo
s
i.e
.
A
cT
iv
[
2
3
]
,
A
R
AST
I
[
2
4
]
.
T
o
o
u
r
k
n
o
w
led
g
e,
th
er
e
is
a
s
i
n
g
le
d
ataset
E
A
ST
R
[
2
5
]
f
o
r
A
r
ab
ic
s
ce
n
e
tex
t
d
etec
tio
n
,
u
n
f
o
r
t
u
n
al
t
y
i
t
is
n
o
t
av
ailab
le
p
u
b
licall
y
.
Fo
r
t
h
is
r
ea
s
o
n
,
w
e
i
n
itiated
a
d
ataset
f
o
r
th
e
n
ee
d
s
o
f
th
is
r
esear
c
h
w
o
r
k
.
I
n
T
ab
le
1
,
s
tatis
t
ics o
f
t
h
e
d
ataset:
T
ab
le
1
.
Data
s
et
s
tatis
tic
s
A
t
t
r
i
b
u
t
e
N
u
mb
e
r
I
mag
e
s
5
7
5
T
e
x
t
u
a
l
i
n
st
a
n
c
e
s
7
6
2
W
o
r
d
s
1
1
2
0
P
e
r
c
e
n
t
a
g
e
o
f
c
u
r
v
e
d
t
e
x
t
2
0
.
8
6
%
P
e
r
c
e
n
t
a
g
e
o
f
i
mag
e
s p
r
e
se
n
t
i
n
g
b
l
u
r
1
0
.
0
8
%
P
e
r
c
e
n
t
a
g
e
o
f
i
mag
e
s p
r
e
se
n
t
i
n
g
mi
s
si
n
g
/
h
i
d
d
e
n
c
h
a
r
a
c
t
e
r
s
1
1
.
8
2
%
N
u
mb
e
r
o
f
f
o
n
t
s
41
P
e
r
c
e
n
t
a
g
e
o
f
i
mag
e
s w
i
t
h
n
o
t
e
x
t
1
4
.
8
%
2.
3
.
T
ec
hn
ica
l e
nv
iro
n
m
ent
W
e
b
u
ilt
m
an
u
all
y
t
h
e
a
s
s
o
ciate
d
g
r
o
u
n
d
tr
u
t
h
co
r
r
esp
o
n
d
in
g
to
i
m
a
g
es,
a
n
d
d
eter
m
i
n
ed
th
e
b
o
u
n
d
in
g
b
o
x
o
f
ea
ch
w
o
r
d
i
n
s
e
m
i
-
a
u
t
m
atic
m
o
d
e.
T
h
e
s
ca
le
ca
teg
o
r
y
i
s
d
eter
m
in
ed
a
s
d
escr
ib
ed
in
2
.
1
(
s
m
all,
n
o
r
m
al,
b
ig
)
.
B
ec
au
s
e
o
f
t
h
e
li
m
ited
n
u
m
b
er
o
f
i
m
a
g
es
i
n
th
e
d
ata
s
et,
w
e
e
x
p
lo
it
d
ata
au
g
m
e
n
tatio
n
b
y
u
s
in
g
o
n
l
in
e
a
u
g
m
e
n
tatio
n
f
o
r
t
w
o
r
ea
s
o
n
s
:
f
ir
s
t
cr
ea
ted
i
m
ag
e
s
ar
e
n
o
t
s
to
r
ed
,
h
en
ce
d
o
n
o
t
h
a
v
e
ad
d
itio
n
al
m
e
m
o
r
y
r
eq
u
ir
e
m
en
ts
;
a
n
d
s
ec
o
n
d
s
y
s
te
m
d
o
es
n
o
t
g
o
th
r
o
u
g
h
s
a
m
e
i
m
ag
e
t
w
ice.
Fo
r
th
is
p
u
r
p
o
s
e,
w
e
u
s
e
Ker
as
lib
r
ar
y
i
m
a
g
e
d
ata
g
en
er
ato
r
[
2
6
]
w
it
h
s
p
ec
i
f
ic
co
n
f
i
g
u
r
atio
n
to
av
o
id
u
n
n
ec
e
s
s
ar
y
tr
an
s
f
o
r
m
atio
n
s
in
o
r
d
er
to
m
ai
n
tai
n
r
ea
lis
tic
n
at
u
r
al
i
m
ag
es,
f
o
r
ex
a
m
p
le
n
o
v
er
tica
l
o
r
h
o
r
izo
n
tal
f
lip
n
ee
d
ed
,
zo
o
m
r
an
g
e
li
m
ited
to
t
w
ice
t
h
e
i
n
itial s
ize
.
T
h
e
t
w
o
n
et
w
o
r
k
s
ar
e
tr
ain
ed
o
n
th
e
d
ataset,
T
ex
tB
lo
ck
lo
ca
lizer
to
ca
teg
o
r
ize
tex
tu
al
an
d
n
o
n
-
tex
t
u
al
i
m
a
g
es
an
d
g
i
v
e
s
ca
le
esti
m
atio
n
,
an
d
T
ex
tDetec
to
r
to
d
etec
t
te
x
t
in
an
ac
cu
r
ate
w
a
y
.
T
h
e
d
atab
ases
ar
e
d
iv
id
ed
in
to
tr
ain
in
g
a
n
d
test
s
u
b
-
s
ets.
W
e
ex
p
lo
it
A
DA
M
o
p
ti
m
izer
[
2
7
]
to
u
p
d
ate
n
et
w
o
r
k
w
ei
g
h
ts
in
s
tead
o
f
s
to
c
h
asti
c
g
r
ad
ien
t
f
o
r
m
an
y
r
ea
s
o
n
s
;
f
ir
s
t,
A
d
a
m
o
p
ti
m
ize
s
co
m
p
u
tatio
n
,
r
eq
u
ir
es
li
g
h
t
m
e
m
o
r
y
,
an
d
o
p
ti
m
ize
s
tr
ain
in
g
i
n
t
h
e
ca
s
e
o
f
n
o
is
y
d
ata
p
r
o
b
le
m
s
,
a
s
it
i
s
t
h
e
ca
s
e
i
n
te
x
t
d
etec
t
io
n
.
A
ll
d
e
v
elo
p
m
e
n
t
is
m
ad
e
w
it
h
P
y
t
h
o
n
o
n
Ker
as
.
3.
RE
SU
L
T
S
A
ND
D
I
SCU
SS
I
O
NS
Fig
u
r
e
3
s
h
o
w
s
s
o
m
e
r
esu
lts
o
f
th
e
p
r
o
p
o
s
ed
f
r
am
e
w
o
r
k
i
n
v
ar
io
u
s
s
it
u
atio
n
s
an
d
s
ce
n
e
co
n
f
i
g
u
r
atio
n
.
W
e
n
o
tice
th
at
s
y
s
te
m
p
er
f
o
r
m
s
w
ell
an
d
g
i
v
es
s
atis
f
y
in
g
r
esu
lts
.
T
o
d
em
o
n
s
tr
ate
th
e
s
tr
en
g
th
o
n
t
h
e
p
r
o
p
o
s
ed
f
r
a
m
e
w
o
r
k
i
n
p
er
f
o
r
m
a
n
ce
a
n
d
t
i
m
e
co
s
t
i
m
p
r
o
v
e
m
en
t,
w
e
p
er
f
o
r
m
te
x
t
d
etec
tio
n
w
i
th
a
VGG
-
1
6
f
u
l
l
y
d
ep
lo
y
ed
.
T
h
e
n
et
w
o
r
k
i
s
tr
ai
n
ed
an
d
te
s
ted
o
n
s
a
m
e
d
ataset
a
n
d
w
e
co
n
f
r
o
n
t
r
esu
lts
o
f
b
o
t
h
ap
p
r
o
ac
h
es.
T
a
b
le
2
s
h
o
w
s
p
r
ec
is
io
n
/R
ec
all
/
an
d
F
-
m
ea
s
u
r
e
an
d
a
v
er
ag
e
ti
m
e
co
s
t
f
o
r
th
e
p
r
o
p
o
s
ed
f
r
a
m
e
w
o
r
k
in
co
m
p
ar
i
s
o
n
w
it
h
b
est
-
r
ec
o
r
d
ed
s
tu
d
ies (
to
o
u
r
k
n
o
w
led
g
e)
an
d
th
e
VGG
-
1
6
r
esu
lt
s
:
-
B
est r
esu
lt
s
in
E
n
g
l
is
h
s
ce
n
e
t
ex
t d
etec
tio
n
-
B
est r
esu
lt
s
in
A
r
ab
ic
s
ce
n
e
d
etec
tio
n
-
VGG
-
1
6
p
er
f
o
r
m
a
n
ce
T
h
e
p
r
o
p
o
s
ed
f
r
am
e
w
o
r
k
ac
h
i
ev
es
s
at
is
f
y
in
g
r
esu
lts
i
n
p
er
f
o
r
m
a
n
ce
,
o
u
tp
er
f
o
r
m
in
g
s
tate
-
of
-
t
h
e
-
ar
t
p
er
f
o
r
m
a
n
ce
i
n
A
r
ab
ic
tex
t
d
etec
tio
n
.
Fo
r
tim
e
co
s
t,
co
m
p
ar
i
s
o
n
b
et
w
ee
n
a
VGG
-
1
6
tim
e
co
s
t
a
n
d
th
e
p
r
o
p
o
s
ed
f
r
a
m
e
w
o
r
k
s
h
o
w
s
a
s
ig
n
i
f
ica
n
t
i
m
p
r
o
v
e
m
e
n
t
eq
u
i
v
alen
t
to
2
3
.
3
4
%.
Ho
w
ev
er
,
p
er
f
o
r
m
an
ce
is
to
b
e
i
m
p
r
o
v
ed
in
s
o
m
e
ca
s
e
s
.
Fig
u
r
e
4
(a
-
f
)
s
h
o
w
s
ex
a
m
p
les
t
h
at
th
e
f
r
a
m
e
w
o
r
k
f
ailed
to
p
r
o
ce
s
s
,
th
e
y
p
r
esen
t
r
esp
ec
tiv
el
y
p
r
o
b
lem
s
o
f
ca
ll
ig
r
ap
h
y
,
s
m
all
f
o
n
t
s
ize,
c
u
r
v
ed
tex
t,
v
er
tical
tex
t,
m
i
s
s
i
n
g
/u
n
c
lear
ch
ar
ac
ter
,
h
ig
h
b
lu
r
.
T
ab
le
2
.
P
er
f
o
r
m
a
n
ce
co
m
p
ar
i
s
o
n
o
f
d
etec
tio
n
ap
p
r
o
ac
h
es
A
l
g
o
r
i
t
h
m
P
r
e
c
i
si
o
n
R
e
c
a
l
l
F
-
m
e
a
su
r
e
T
i
me
c
o
st
(
f
p
s)
P
r
o
p
o
se
d
f
r
a
me
w
o
r
k
0
.
6
5
1
0
.
7
1
4
0
.
6
8
2
4
.
3
VGG
-
16
0
.
6
6
1
0
.
7
5
2
0
.
6
9
3
1
.
7
A
r
a
b
i
c
b
e
st
p
e
r
f
o
r
man
c
e
[
1
9
]
0
.
6
4
0
.
7
2
0
.
7
0
N
.
A
En
g
l
i
sh
b
e
st
p
e
r
f
o
r
man
c
e
[
2
0
]
0
.
7
6
9
8
0
.
8
4
8
5
0
.
8
0
7
2
1
6
.
5
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
SN
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
2
,
A
p
r
il
2
0
2
1
:
1
6
3
4
-
1640
1638
Fig
u
r
e
3
.
E
x
a
m
p
le
i
m
ag
e
s
f
o
r
m
d
ataset
(
a)
(
b
)
(
c)
(
d
)
(
e)
(
f
)
Fig
u
r
e
4
.
I
m
a
g
es
w
h
er
e
tex
t d
etec
tio
n
f
a
iled
,
(
a)
p
r
o
b
lem
s
o
f
ca
lli
g
r
ap
h
y
,
(
b
)
s
m
a
ll f
o
n
t si
ze
,
(
c)
cu
r
v
ed
tex
t,
(
d
)
v
er
tical
tex
t,
(
e)
m
is
s
in
g
/
u
n
clea
r
ch
ar
ac
ter
,
an
d
(
f
)
h
i
g
h
b
lu
r
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8708
R
ea
l
-
time
A
r
a
b
ic
s
ce
n
e
text
d
e
tectio
n
u
s
in
g
fu
lly
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
ks (
R
a
ja
e
Mo
u
men
)
1639
4.
CO
NCLU
SI
O
N
I
n
th
i
s
p
ap
er
,
w
e
p
r
esen
ted
a
n
o
v
el
ap
p
r
o
ac
h
to
d
ea
l
w
ith
r
ea
l
-
ti
m
e
A
r
ab
ic
s
ce
n
e
te
x
t
d
etec
tio
n
u
s
i
n
g
a
t
w
o
-
s
tep
s
ap
p
r
o
ac
h
in
s
tead
o
f
a
s
in
g
le
o
n
e
,
w
e
co
n
f
r
o
n
te
d
th
e
r
esu
lts
to
a
o
n
e
-
ti
m
e
f
r
a
m
e
w
o
r
k
b
ased
o
n
VGG
-
1
6
an
d
o
u
r
ap
p
r
o
ac
h
g
i
v
es
s
atis
f
y
i
n
g
r
e
s
u
l
ts
w
i
th
o
p
t
i
m
ized
r
u
n
ti
m
e.
Ho
w
e
v
er
,
th
e
f
r
a
m
e
w
o
r
k
is
to
b
e
i
m
p
r
o
v
ed
in
f
o
llo
w
i
n
g
asp
ec
t
s
:
F
ir
s
t,
th
e
f
r
a
m
e
w
o
r
k
d
o
es
n
o
t
r
eso
lv
e
cu
r
v
ed
tex
t
d
etec
tio
n
,
s
ec
o
n
d
,
a
s
in
g
le
lan
g
u
a
g
e
co
u
ld
b
e
p
r
o
ce
s
s
ed
,
in
o
u
r
ca
s
e
A
r
ab
ic,
in
f
o
r
m
a
t
io
n
in
m
u
l
tili
n
g
u
al
s
ce
n
e
s
is
n
o
t
f
u
ll
y
r
etr
ie
v
ed
.
Fin
all
y
,
y
et
i
m
p
o
r
ta
n
tl
y
,
th
e
l
ac
k
o
f
A
r
ab
ic
d
atas
et
s
f
o
r
s
ce
n
e
tex
t
d
etec
tio
n
a
n
d
r
ec
o
g
n
iti
o
n
is
a
b
ig
b
r
ak
e
t
o
ac
h
iev
in
g
b
etter
r
esu
lt
s
t
h
u
s
b
u
ild
in
g
a
lar
g
e
d
atase
t f
o
r
th
i
s
p
u
r
p
o
s
e
is
an
u
r
g
en
t
m
atter
.
RE
F
E
R
E
NC
E
S
[1
]
T
.
S
h
a
n
g
x
u
a
n
,
e
t
a
l.
,
“
T
e
x
t
F
lo
w
:
A
Un
if
ied
T
e
x
t
D
e
tec
ti
o
n
S
y
ste
m
in
Na
tu
ra
l
S
c
e
n
e
Im
a
g
e
s,
”
in
2
0
1
5
IEE
E
In
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
C
o
mp
u
ter
Vi
si
o
n
(
ICCV)
,
S
a
n
ti
a
g
o
,
C
h
il
e
,
2
0
1
5
,
p
p
.
1
-
9
.
[2
]
Z.
Zh
a
n
g
,
e
t
a
l.
,
“
S
y
m
m
e
tr
y
-
b
a
s
e
d
tex
t
li
n
e
d
e
tec
ti
o
n
i
n
n
a
t
u
ra
l
sc
e
n
e
s,
”
in
2
0
1
5
IEE
E
Co
n
fer
e
n
c
e
o
n
Co
mp
u
te
r
Vi
sio
n
a
n
d
P
a
tt
e
rn
Rec
o
g
n
it
i
o
n
(
CVP
R)
,
B
o
sto
n
,
M
A
,
USA
,
2
0
1
5
,
p
p
.
2
5
5
8
-
2
5
6
7
.
[3
]
S
.
L
e
i,
e
t
a
l.
,
“
A
ro
b
u
st
a
p
p
r
o
a
c
h
f
o
r
tex
t
d
e
te
c
ti
o
n
f
ro
m
n
a
tu
ra
l
sc
e
n
e
i
m
a
g
e
s,
”
Pa
tt
e
rn
Rec
o
g
n
it
i
o
n
,
v
o
l.
4
8
,
n
o
.
9
,
p
p
.
2
9
0
6
-
2
9
2
0
,
2
0
1
5
.
[4
]
Y.
X
u
-
C
h
e
n
g
,
e
t
a
l.
,
“
Ro
b
u
st T
e
x
t
De
tec
ti
o
n
in
Na
tu
ra
l
S
c
e
n
e
Im
a
g
e
s,
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
P
a
tt
e
rn
An
a
lys
is a
n
d
M
a
c
h
in
e
In
telli
g
e
n
c
e
,
v
o
l
.
3
6
,
n
o
.
5
,
p
p
.
9
7
0
-
9
8
3
,
2
0
1
4
.
[5
]
M
.
L
iao
,
e
t
a
l.
,
“
Re
a
l
-
ti
m
e
S
c
e
n
e
T
e
x
t
De
tec
ti
o
n
w
it
h
Diff
e
re
n
ti
a
b
le Bi
n
a
riza
ti
o
n
,
”
a
rXiv
:
1
9
1
1
.
0
8
9
4
7
,
2
0
1
9
.
[6
]
M.
L
iao
,
e
t
a
l.
,
“
T
e
x
tBo
x
e
s: A
F
a
st T
e
x
t
De
t
e
c
to
r
w
it
h
a
S
in
g
le De
e
p
Ne
u
ra
l
Ne
tw
o
rk
,
”
a
rXiv: 1
6
1
1
.
0
6
7
7
9
,
2
0
1
6
.
[7
]
M.
L
iao
,
e
t
a
l.
,
“
T
e
x
tBo
x
e
s+
+
:
A
S
in
g
le
-
S
h
o
t
Orie
n
ted
S
c
e
n
e
T
e
x
t
De
tec
to
r,
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
Ima
g
e
Pro
c
e
ss
in
g
,
v
o
l
.
2
7
,
p
p
.
3
6
7
6
-
3
6
9
0
,
2
0
1
8
.
[8
]
X
.
Z
h
o
u
,
e
t
a
l.
,
“
EA
S
T
:
A
n
Eff
i
c
ien
t
a
n
d
A
c
c
u
ra
t
e
S
c
e
n
e
Tex
t
De
tec
to
r,
”
in
C
o
n
fer
e
n
c
e
o
n
C
o
mp
u
ter
Vi
sio
n
a
n
d
Pa
tt
e
rn
Rec
o
g
n
it
io
n
,
2
0
1
7
,
p
p
.
5
5
5
1
-
5
5
6
0
.
[9
]
D.
D
e
n
g
,
e
t
a
l.
,
“
P
ix
e
lL
in
k
:
De
te
c
ti
o
n
sc
e
n
e
tex
t
v
ia
in
sta
n
c
e
s
e
g
m
e
n
tatio
n
,
”
in
t
h
e
th
irty
-
se
c
o
n
d
AA
AI
Co
n
fer
e
n
c
e
o
f
Arti
f
icia
l
I
n
telli
g
e
n
c
e
,
2
0
1
8
.
[1
0
]
L
.
S
h
a
n
g
b
a
n
g
,
e
t
a
l.
,
“
T
e
x
tS
n
a
k
e
:
A
F
le
x
ib
le
Re
p
re
se
n
tatio
n
f
o
r
De
tec
ti
n
g
T
e
x
t
o
f
A
rb
it
ra
r
y
S
h
a
p
e
s,
”
Eu
ro
p
e
a
n
Co
n
fer
e
n
c
e
o
n
Co
m
p
u
ter
Vi
si
o
n
ECCV
,
2
0
1
8
,
p
p
.
1
9
-
3
5
.
[1
1
]
Z.
Zh
a
n
g
,
e
t
a
l.
,
“
M
u
lt
i
-
Orie
n
ted
T
e
x
t
De
tec
ti
o
n
w
it
h
F
u
ll
y
Co
n
v
o
lu
ti
o
n
a
l
Ne
tw
o
rk
s,
”
in
Co
n
f
e
re
n
c
e
o
n
C
o
mp
u
te
r
Vi
sio
n
a
n
d
P
a
tt
e
rn
Rec
o
g
n
it
i
o
n
,
2
0
1
6
,
p
p
.
4
1
5
9
-
4
1
6
7
.
[1
2
]
W
.
W
a
n
g
,
e
t
a
l
.,
“
S
h
a
p
e
Ro
b
u
s
t
T
e
x
t
D
e
tec
ti
o
n
w
it
h
P
ro
g
r
e
ss
iv
e
S
c
a
l
e
Ex
p
a
n
sio
n
Ne
t
w
o
rk
,
”
Co
n
fer
e
n
c
e
o
n
Co
mp
u
ter
V
isio
n
a
n
d
Pa
t
ter
n
Rec
o
g
n
it
io
n
,
2
0
1
9
,
p
p
.
9
3
3
6
-
9
3
4
5
.
[1
3
]
T
a
n
g
,
Yu
a
n
e
t
a
l.
,
“
S
c
e
n
e
T
e
x
t
D
e
tec
ti
o
n
a
n
d
S
e
g
m
e
n
tatio
n
Ba
se
d
o
n
Ca
sc
a
d
e
d
Co
n
v
o
lu
ti
o
n
Ne
u
ra
l
Ne
tw
o
rk
s
,
”
in
IEE
E
t
ra
n
s
a
c
ti
o
n
s
o
n
ima
g
e
p
r
o
c
e
ss
in
g
:
a
p
u
b
li
c
a
ti
o
n
o
f
t
h
e
IEE
E
S
ig
n
a
l
Pro
c
e
ss
in
g
S
o
c
iety
,
v
o
l.
2
6
,
n
o
.
3
,
p
p
.
1
5
0
9
-
1
5
2
0
,
2
0
1
7
.
[1
4
]
J.
F
a
n
,
e
t
a
l.
,
“
De
e
p
S
c
e
n
e
T
e
x
t
De
tec
ti
o
n
w
it
h
Co
n
n
e
c
ted
Co
m
p
o
n
e
n
t
P
ro
p
o
sa
ls,
”
Co
n
fer
e
n
c
e
o
n
Co
mp
u
ter
Vi
si
o
n
a
n
d
Pa
tt
e
rn
Rec
o
g
n
it
io
n
,
2
0
1
7
,
p
p
.
1
-
10
.
[1
5
]
M
.
M
o
h
ied
d
i
n
a
n
d
M
.
S
a
e
e
d
,
“
Hy
b
rid
a
p
p
ro
a
c
h
f
o
r
F
a
rsi/A
ra
b
ic
tex
t
d
e
tec
ti
o
n
a
n
d
lo
c
a
li
sa
ti
o
n
i
n
v
id
e
o
f
ra
m
e
s,
”
IET
Ima
g
e
Pro
c
e
ss
in
g
,
v
o
l.
7
,
n
o
.
2
,
p
p
.
1
5
4
-
1
6
4
,
2
0
1
2
.
[1
6
]
M.
Ja
in
,
e
t
a
l
.,
“
Un
c
o
n
stra
in
e
d
sc
e
n
e
t
e
x
t
a
n
d
v
id
e
o
tex
t
re
c
o
g
n
it
io
n
f
o
r
A
ra
b
ic
sc
rip
t,
”
in
1
st
In
ter
n
a
ti
o
n
a
l
W
o
rk
sh
o
p
o
n
Ara
b
ic
S
c
rip
t
An
a
ly
sis a
n
d
Rec
o
g
n
it
i
o
n
,
Na
n
c
y
,
p
p
.
1
-
5
,
2
0
1
7
.
[1
7
]
S
.
B.
A
h
m
e
d
,
e
t
a
l
.
,
“
De
e
p
lea
rn
i
n
g
b
a
se
d
iso
late
d
A
ra
b
ic
sc
e
n
e
c
h
a
ra
c
ter
re
c
o
g
n
it
io
n
,
”
i
n
1
s
t
In
ter
n
a
ti
o
n
a
l
W
o
rk
sh
o
p
o
n
Ara
b
ic
S
c
rip
t
An
a
ly
sis a
n
d
Rec
o
g
n
it
i
o
n
,
Na
n
c
y
,
p
p
.
1
-
6
,
2
0
1
7
.
[1
8
]
M
.
S
.
H
.
S
a
ss
i
,
e
t
a
l
.,
“
M
u
lt
i
-
Or
ien
ted
Re
a
l
-
T
i
m
e
A
ra
b
ic
S
c
e
n
e
T
e
x
t
De
tec
ti
o
n
w
it
h
De
e
p
F
u
ll
y
Co
n
v
o
l
u
ti
o
n
a
l
Ne
tw
o
rk
s,
”
in
1
6
th
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
C
o
mp
u
ter
S
y
ste
ms
a
n
d
Ap
p
li
c
a
ti
o
n
s
,
A
b
u
D
h
a
b
i,
Un
it
e
d
A
ra
b
Em
irate
s
,
2
0
1
9
,
p
p
.
1
-
6
.
[1
9
]
I.
Be
lt
a
ief
a
n
d
M
.
B.
Ha
li
m
a
,
“
D
e
e
p
F
CN
f
o
r
A
r
a
b
ic
S
c
e
n
e
T
e
x
t
De
tec
ti
o
n
,
”
in
IEE
E
2
n
d
In
ter
n
a
t
io
n
a
l
W
o
rk
sh
o
p
o
n
Ar
a
b
ic
a
n
d
De
riv
e
d
S
c
rip
t
A
n
a
lys
is a
n
d
Rec
o
g
n
i
ti
o
n
,
L
o
n
d
o
n
,
p
p
.
1
2
9
-
1
3
4
,
2
0
1
8
.
[2
0
]
W
.
He
,
e
t
a
l.
,
“
Re
a
lt
im
e
m
u
lt
i
-
sc
a
le
sc
e
n
e
te
x
t
d
e
tec
ti
o
n
w
it
h
sc
a
le
-
b
a
se
d
re
g
io
n
p
ro
p
o
sa
l
n
e
t
w
o
rk
,
”
Pa
tt
e
rn
Rec
o
g
n
it
io
n
,
v
o
l.
9
8
,
2
0
2
0
.
[2
1
]
W.
He
,
e
t
a
l
.,
“
M
u
lt
i
-
Orie
n
te
d
a
n
d
M
u
lt
i
-
L
in
g
u
a
l
S
c
e
n
e
T
e
x
t
De
tec
ti
o
n
w
it
h
Dire
c
t
Re
g
re
ss
io
n
,
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
Ima
g
e
Pro
c
e
ss
in
g
,
v
o
l.
2
7
,
n
o
.
1
1
,
p
p
.
5
4
0
9
-
5
4
1
9
,
2
0
1
8
.
[2
2
]
F
.
S
li
m
a
n
e
,
e
t
a
l
.,
“
A
Ne
w
A
ra
b
ic
P
rin
ted
T
e
x
t
I
m
a
g
e
Da
t
a
b
a
s
e
a
n
d
Ev
a
lu
a
ti
o
n
P
r
o
to
c
o
ls,
”
in
1
0
th
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Do
c
u
me
n
t
An
a
lys
i
s a
n
d
Rec
o
g
n
it
i
o
n
,
Ba
rc
e
lo
n
a
,
2
0
0
9
,
p
p
.
9
4
6
-
9
5
0
.
[2
3
]
O.
Zay
e
n
e
,
e
t
a
l.
,
“
A
d
a
tas
e
t
f
o
r
A
ra
b
ic
tex
t
d
e
tec
ti
o
n
,
trac
k
in
g
a
n
d
re
c
o
g
n
i
ti
o
n
i
n
n
e
w
s
v
id
e
o
s
-
Ac
T
iV
,
”
in
1
3
t
h
In
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
D
o
c
u
me
n
t
A
n
a
lys
is
a
n
d
Rec
o
g
n
it
i
o
n
,
T
u
n
is
,
2
0
1
5
,
p
p
.
9
9
6
-
1
0
0
0
.
[2
4
]
M
.
T
o
u
n
si,
e
t
a
l.
,
“
A
R
A
S
T
I:
A
d
a
tab
a
se
f
o
r
A
r
a
b
ic
sc
e
n
e
tex
t
re
c
o
g
n
it
io
n
,
”
in
1
st
In
ter
n
a
t
io
n
a
l
W
o
rk
sh
o
p
o
n
Ara
b
ic
S
c
rip
t
A
n
a
lys
is a
n
d
Rec
o
g
n
it
io
n
,
Na
n
c
y
,
p
p
.
1
4
0
-
1
4
4
,
2
0
1
7
.
[2
5
]
S
.
B.
A
h
m
e
d
,
e
t
a
l
.,
“
A
No
v
e
l
Da
tas
e
t
f
o
r
En
g
li
sh
-
A
ra
b
ic
S
c
e
n
e
T
e
x
t
Re
c
o
g
n
it
io
n
(EA
S
TR)
-
4
2
K
a
n
d
It
s
Ev
a
lu
a
ti
o
n
Us
in
g
In
v
a
rian
t
F
e
a
tu
re
Ex
trac
ti
o
n
o
n
De
tec
ted
Ex
tre
m
a
l
Re
g
io
n
s,
”
IEE
E
Acc
e
ss
,
v
o
l.
7
,
p
p
.
1
9
8
0
1
-
1
9
8
2
0
,
2
0
1
9
.
[2
6
]
“
I
m
a
g
e
d
a
ta p
re
p
ro
c
e
ss
in
g
,
”
Ke
ra
s
,
[
O
n
li
n
e
]
.
A
v
a
il
a
b
le:
h
tt
p
s://
k
e
ra
s.io
/ap
i/
p
re
p
ro
c
e
ss
in
g
/i
m
a
g
e
/.
[2
7
]
P
.
K.
Die
d
e
rik
a
n
d
B.
Jim
m
y
,
“
Ad
a
m
:
A
m
e
th
o
d
f
o
r
sto
c
h
a
stic o
p
ti
m
iz
a
ti
o
n
,
”
Co
R
R
,
2
0
1
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
SN
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
2
,
A
p
r
il
2
0
2
1
:
1
6
3
4
-
1640
1640
B
I
O
G
RAP
H
I
E
S
O
F
AUTH
O
RS
Ra
ja
e
M
o
u
m
e
n
is
a
P
h
D
stu
d
e
n
t
a
t
th
e
Na
ti
o
n
a
l
S
c
h
o
o
l
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
S
y
ste
m
s
A
n
a
l
y
si
s
a
t
M
o
h
a
m
m
e
d
V
Un
i
v
e
rsit
y
,
Ra
b
a
t,
M
o
ro
c
c
o
.
S
h
e
h
a
s
a
d
e
g
re
e
o
f
e
n
g
in
e
e
r
in
Co
m
p
u
ter
sc
ien
c
e
s
f
ro
m
th
e
sa
m
e
sc
h
o
o
l
a
n
d
h
a
s
1
0
y
e
a
rs
e
x
p
e
rien
c
e
a
s
I
T
m
a
n
a
g
e
r.
He
r
re
se
a
rc
h
in
tere
sts
a
re
in
th
e
a
re
a
o
f
n
e
u
ra
l
n
e
tw
o
rk
s,
n
a
tu
ra
l
la
n
g
u
a
g
e
p
ro
c
e
ss
in
g
,
e
sp
e
c
ially
A
ra
b
ic
.
Ra
d
d
o
u
a
n
e
Ch
i
h
e
b
is
a
p
ro
f
e
ss
o
r
o
f
a
p
p
l
ied
m
a
th
e
m
a
ti
c
s
a
t
th
e
Na
ti
o
n
a
l
S
c
h
o
o
l
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
S
y
ste
m
s
A
n
a
l
y
sis
a
t
M
o
h
a
m
m
e
d
V
U
n
iv
e
rsity
,
Ra
b
a
t,
M
o
ro
c
c
o
.
He
o
b
tain
e
d
h
is
M
a
ste
r
f
ro
m
th
e
Na
ti
o
n
a
l
In
sti
tu
t
e
o
f
A
p
p
li
e
d
S
c
ien
c
e
s
o
f
L
y
o
n
a
n
d
P
h
D
f
ro
m
th
e
Je
a
n
M
o
n
n
e
t
Un
i
v
e
rsit
y
o
f
S
a
in
t
-
Et
ien
n
e
.
His
re
se
a
rc
h
in
tere
sts
a
re
in
th
e
a
re
a
o
f
S
e
m
a
n
ti
c
A
n
a
l
y
sis,
S
tru
c
tu
ra
l
Op
ti
m
iza
ti
o
n
,
Ed
u
c
a
ti
o
n
,
Op
ti
m
iza
ti
o
n
o
f
th
e
lo
g
isti
c
s
c
h
a
in
,
M
a
c
h
in
e
lea
rn
i
n
g
,
a
n
d
V
a
lu
e
A
n
a
l
y
sis.
He
su
p
e
rv
ise
d
o
v
e
r
1
0
stu
d
e
n
ts.
P
r
o
f
.
Ra
d
d
o
u
a
n
e
Ch
ih
e
b
is
P
re
sid
e
n
t
o
f
th
e
M
o
ro
c
c
a
n
A
ss
o
c
iatio
n
f
o
r
t
h
e
V
a
lu
e
A
n
a
l
y
sis
Rd
o
u
a
n
F
a
izi
h
a
s
a
P
h
.
D
.
in
l
in
g
u
isti
c
s
f
ro
m
M
o
h
a
m
m
e
d
V
Ag
d
a
l
Un
iv
e
rsit
y
,
Ra
b
a
t,
M
o
ro
c
c
o
.
Cu
rre
n
tl
y
,
h
e
tea
c
h
e
s
En
g
li
sh
a
t
Eco
le
Na
ti
o
n
a
le
S
u
p
é
rieu
re
d
’I
n
f
o
rm
a
ti
q
u
e
e
t
d
’A
n
a
l
y
s
e
d
e
s
S
y
st
è
m
e
s (E
NSI
A
S
),
M
o
h
a
m
m
e
d
V
S
o
u
issi Un
iv
e
rsity
.
Evaluation Warning : The document was created with Spire.PDF for Python.