I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
15
,
No
.
6
,
Decem
b
er
20
25
,
p
p
.
5
9
1
4
~
5
9
2
3
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
15
i
6
.
pp
5
9
1
4
-
5
9
2
3
5914
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Dev
a
na
g
a
ri
o
ptical cha
ra
cter
r
ecog
nition
o
f
prin
ted
t
e
x
t
M
a
la
t
hi P
.
1
,
Cha
nd
ra
k
a
nth
G
.
P
uja
ri
2
1
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
D
r
.
A
mb
e
d
k
a
r
I
n
st
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
B
e
n
g
a
l
u
r
u
,
a
f
f
i
l
i
a
t
e
t
o
V
i
s
v
e
s
v
a
r
a
y
a
Te
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
si
t
y
,
B
e
l
a
g
a
v
i
2
D
e
p
a
r
t
me
n
t
o
f
M
a
st
e
r
o
f
C
o
m
p
u
t
e
r
A
p
p
l
i
c
a
t
i
o
n
s
,
D
r
.
A
m
b
e
d
k
a
r
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
B
e
n
g
a
l
u
r
u
,
a
f
f
i
l
i
a
t
e
t
o
V
i
sv
e
sv
a
r
a
y
a
Te
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
si
t
y
,
B
e
l
a
g
a
v
i
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Dec
1
8
,
2
0
2
4
R
ev
is
ed
J
u
l 1
6
,
2
0
2
5
Acc
ep
ted
Sep
1
4
,
2
0
2
5
Hu
n
d
re
d
s
o
f
n
a
ti
v
e
lan
g
u
a
g
e
s
a
n
d
sc
rip
ts
a
re
m
a
k
in
g
t
h
e
ir
wa
y
o
n
d
ig
i
tal
p
latfo
rm
to
su
sta
in
i
n
m
u
lt
i
p
le
d
a
ta
fo
rm
a
ts.
Op
ti
c
a
l
c
h
a
ra
c
ter
re
c
o
g
n
i
ti
o
n
(OCR)
is
o
n
e
s
u
c
h
d
ime
n
sio
n
w
h
e
re
th
e
l
o
w
re
so
u
rc
e
lan
g
u
a
g
e
s
a
re
y
e
t
to
fin
d
t
h
e
ir
sta
b
i
li
ty
.
De
v
a
n
a
g
a
ri
O
CR
is
o
n
e
su
c
h
lo
w
re
so
u
rc
e
sc
rip
t
p
r
o
b
lem
to
b
e
d
e
a
lt
with
,
th
o
u
g
h
it
is
th
e
fo
u
rth
wi
d
e
ly
u
se
d
g
l
o
b
a
l
sc
ri
p
t.
Re
c
e
n
t
wo
rk
s
c
a
rried
o
n
OCR
h
a
v
e
f
o
c
u
se
d
o
n
w
o
rd
le
v
e
l
a
p
p
r
o
a
c
h
a
n
d
fa
c
e
c
h
a
ll
e
n
g
e
s
o
f
sp
irali
n
g
c
o
m
p
lex
it
y
a
s
lan
g
u
a
g
e
a
lp
h
a
b
e
t
se
t
siz
e
c
ro
ss
e
s
h
u
n
d
re
d
s.
M
o
st
o
f
t
h
e
se
OCR
w
o
rk
s
a
re
d
o
n
e
i
n
c
o
n
stra
in
e
d
e
n
v
iro
n
m
e
n
t
,
with
h
u
g
e
d
a
tas
e
ts
a
n
d
larg
e
c
o
m
p
u
tatio
n
a
l
re
so
u
rc
e
s.
As
a
re
su
l
t,
e
ffe
c
ti
v
e
b
e
n
c
h
m
a
rk
e
v
a
l
u
a
ti
o
n
o
f
t
h
e
wo
r
k
s
a
g
a
in
st o
n
e
a
n
o
t
h
e
r
o
n
d
e
fin
e
d
m
e
tri
c
s
is
sc
a
rc
e
.
Aim
h
e
re
is
to
e
x
p
l
o
re
c
h
a
ra
c
ter
lev
e
l
De
v
a
n
a
g
a
ri
OCR
wi
th
p
rin
te
d
tex
t
ima
g
e
s
a
s
in
p
u
t.
P
a
tt
e
r
n
re
c
o
g
n
it
i
o
n
(P
R)
p
r
in
c
ip
les
fo
r
d
iac
rit
ic
c
las
sifica
ti
o
n
a
n
d
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
(CNN
)
f
o
r
b
a
se
c
h
a
ra
c
ter
c
las
sifica
ti
o
n
a
re
u
se
d
.
w
o
rd
e
rro
r
ra
te
(
WE
R)
o
f
2
4
.
4
7
%
is
a
tt
a
in
e
d
.
Ho
we
v
e
r,
t
h
e
train
in
g
d
a
tas
e
t
c
o
m
p
lex
it
y
is
re
d
u
c
e
d
b
y
4
.
3
5
ti
m
e
s.
Th
e
ten
m
u
lt
i
c
las
s m
o
d
e
ls,
train
i
n
g
ti
m
e
ra
n
g
e
fro
m
4
5
m
in
u
tes
to
2
.
5
h
o
u
rs.
F
u
rt
h
e
r
th
e
m
o
d
e
ls
c
a
n
b
e
train
e
d
i
n
p
a
ra
ll
e
l
to
c
o
m
p
lete
th
e
train
i
n
g
p
ro
c
e
ss
in
3
-
4
h
o
u
rs.
Th
u
s,
t
h
e
a
p
p
r
o
a
c
h
u
se
d
fo
r
te
x
t
c
las
sifica
ti
o
n
fa
c
il
it
a
tes
th
e
De
v
a
n
a
g
a
ri
OCR so
l
u
ti
o
n
t
o
b
e
o
ffe
re
d
in
o
ff
-
th
e
-
sh
e
lf
c
o
m
p
u
ti
n
g
d
e
v
ice
s.
K
ey
w
o
r
d
s
:
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
Data
s
et
co
m
p
lex
ity
Dev
an
ag
ar
i scr
ip
t
Diac
r
itic
L
ex
ico
n
L
o
w
r
eso
u
r
ce
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
:
Ma
lath
i P.
Dep
ar
tm
en
t o
f
I
n
f
o
r
m
atio
n
Scien
ce
an
d
E
n
g
in
ee
r
i
n
g
,
Dr
.
Am
b
ed
k
ar
I
n
s
titu
te
o
f
T
ec
h
n
o
l
o
g
y
,
af
f
iliate
to
Vis
v
esv
ar
ay
a
T
ec
h
n
o
l
o
g
ical
Un
iv
er
s
ity
B
en
g
alu
r
u
-
5
6
0
0
5
6
,
Kar
n
atak
a,
I
n
d
ia
E
m
ail: m
alath
ip
.
is
@
d
r
ait.
ed
u
.
in
1.
I
NT
RO
D
UCT
I
O
N
Hu
n
d
r
ed
s
o
f
n
ativ
e
lan
g
u
ag
e
s
an
d
s
cr
ip
ts
ar
e
m
ak
in
g
th
ei
r
way
o
n
d
ig
ital
p
lat
f
o
r
m
t
o
s
u
s
tain
in
m
u
ltip
le
d
ata
f
o
r
m
ats
lik
e
te
x
t,
im
ag
es,
a
u
d
io
an
d
v
i
d
eo
.
I
n
th
e
p
r
esen
t
s
ce
n
a
r
io
,
u
n
d
er
d
ig
ital
tex
t
h
a
n
d
lin
g
o
p
tical
ch
ar
ac
ter
r
ec
o
g
n
itio
n
(
OC
R
)
f
o
r
m
s
th
e
co
r
n
er
s
to
n
e
as
v
ast
am
o
u
n
t
o
f
tex
t
to
b
e
ex
tr
ac
ted
s
till
lay
s
in
im
ag
e
f
o
r
m
at
wh
ic
h
ca
n
o
n
ly
b
e
a
d
d
r
ess
ed
with
o
p
e
n
-
s
o
u
r
ce
s
o
lu
tio
n
s
.
Fu
r
t
h
er
,
th
e
s
o
lu
tio
n
s
d
ev
elo
p
ed
a
r
e
m
o
r
e
f
o
c
u
s
ed
to
war
d
s
n
u
r
tu
r
in
g
th
e
h
ig
h
r
eso
u
r
ce
la
n
g
u
a
g
es
wh
ich
h
av
e
n
o
tab
le
b
en
c
h
m
ar
k
d
atasets
f
o
r
ad
d
r
ess
in
g
th
e
OC
R
p
r
o
b
lem
.
E
m
er
g
en
ce
o
f
v
a
r
iatio
n
s
in
O
C
R
lik
e
s
ce
n
e
tex
t
OC
R
,
o
n
lin
e
OC
R
h
av
e
f
u
r
th
er
clu
tter
ed
th
e
r
esear
ch
wo
r
k
s
o
n
t
h
e
m
o
r
e
n
o
tab
le,
v
ital
an
d
d
e
f
au
lt
f
o
r
m
,
th
e
d
o
cu
m
en
t
OC
R
.
Natu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
d
ea
ls
wi
th
s
em
an
tic
h
an
d
lin
g
o
f
d
ig
iti
ze
d
tex
t
an
d
ca
n
b
e
p
ar
t
o
f
p
o
s
t
p
r
o
ce
s
s
in
g
th
e
r
esu
lts
o
f
OC
R
.
R
esear
ch
in
n
atu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
(
NL
P)
th
o
u
g
h
s
tar
ted
late
t
h
a
n
OC
R
h
as
g
ain
ed
b
etter
m
o
m
en
tu
m
in
h
an
d
lin
g
d
ata
b
o
th
in
tex
t
an
d
au
d
io
an
d
v
id
eo
f
o
r
m
at
an
d
s
h
ad
o
wed
its
in
ter
ests
.
State
-
of
-
th
e
-
ar
t
(
SOTA
)
d
ee
p
lear
n
in
g
tech
n
iq
u
es
ad
o
p
ted
f
o
r
OC
R
s
o
lu
tio
n
s
th
r
iv
e
o
n
lar
g
e
d
atasets
an
d
o
f
ten
tig
h
tly
co
u
p
led
with
p
r
e
-
p
r
o
ce
s
s
in
g
a
n
d
p
o
s
t
-
p
r
o
ce
s
s
in
g
s
tag
es
in
th
e
p
ip
elin
e
th
at
ef
f
ec
tiv
e
m
etr
ics
f
o
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
-
8
7
0
8
Dev
a
n
a
g
a
r
i o
p
tica
l c
h
a
r
a
cter reco
g
n
itio
n
o
f
p
r
in
ted
text
(
M
a
la
th
i P
.
)
5915
co
m
p
ar
is
o
n
am
o
n
g
th
e
co
n
te
m
p
o
r
ize
w
o
r
k
s
is
f
u
zz
y
.
E
m
e
r
g
in
g
tec
h
n
o
l
o
g
ies
lik
e
lar
g
e
l
ea
r
n
in
g
m
o
d
els
th
at
f
o
cu
s
o
n
en
d
-
to
-
en
d
s
o
lu
tio
n
s
ar
e
f
u
r
th
er
tak
in
g
to
n
e
x
t
lev
el
o
f
s
o
p
h
is
ticatio
n
.
T
h
u
s
,
lacu
n
ae
o
f
b
en
c
h
m
ar
k
d
atasets
f
o
r
d
o
cu
m
en
t
OC
R
o
f
m
an
y
r
eg
io
n
al
lan
g
u
ag
es
l
an
d
s
th
em
as
lo
w
r
eso
u
r
ce
lan
g
u
ag
es
an
d
I
n
d
ic
s
cr
ip
ts
ar
e
p
ar
t
o
f
th
is
ca
teg
o
r
y
.
Dev
an
a
g
ar
i
OC
R
is
o
n
e
a
m
o
n
g
t
h
e
lo
w
r
eso
u
r
ce
I
n
d
ic
s
cr
ip
ts
to
b
e
d
ea
lt
with
,
ev
en
th
o
u
g
h
it
is
th
e
f
o
u
r
th
wid
ely
u
s
ed
an
d
p
o
p
u
lar
g
lo
b
al
s
cr
ip
t.
W
o
r
k
s
ca
r
r
ied
o
u
t
o
n
Dev
a
n
ag
ar
i
OC
R
ar
e
f
o
u
n
d
to
b
e
h
ig
h
ly
i
n
f
lu
en
ce
d
with
h
ig
h
r
eso
u
r
ce
OC
R
s
an
d
h
en
ce
th
eir
d
is
tin
g
u
is
h
in
g
f
ea
tu
r
es
ar
e
less
ad
d
r
ess
ed
.
Pre
s
en
tly
th
e
p
r
o
p
r
ietar
y
s
o
lu
tio
n
s
d
ev
elo
p
e
d
f
o
r
Dev
an
a
g
ar
i
OC
R
ar
e
ch
ar
ac
ter
ized
with
th
e
u
s
e
o
f
lar
g
e
d
atasets
f
o
r
tr
ain
in
g
m
ac
h
in
e
lear
n
in
g
m
o
d
el
s
,
co
n
s
u
m
in
g
ab
u
n
d
an
t
co
m
p
u
tin
g
r
eso
u
r
ce
s
in
d
is
tr
ib
u
ted
an
d
clo
u
d
co
m
p
u
ti
n
g
en
v
ir
o
n
m
e
n
t.
T
h
is
im
p
ac
ts
th
e
af
f
o
r
d
ab
ilit
y
,
c
u
s
to
m
izatio
n
an
d
f
lex
ib
ilit
y
to
ad
ap
t
to
ap
p
licatio
n
s
in
d
i
v
er
s
e
d
o
m
ain
s
an
d
h
en
ce
t
h
e
s
u
s
ten
an
ce
o
f
th
e
Dev
an
a
g
ar
i
O
C
R
.
Aim
h
er
e
is
to
ex
p
lo
r
e
th
e
u
s
e
o
f
co
n
v
en
tio
n
al
p
atter
n
r
ec
o
g
n
itio
n
tech
n
iq
u
es
an
d
m
ac
h
in
e
lear
n
in
g
tech
n
iq
u
es
in
a
b
alan
ce
d
s
y
s
tem
ic
m
an
n
er
to
a
d
d
r
ess
th
e
c
h
allen
g
es
m
e
n
tio
n
ed
ab
o
v
e
an
d
co
n
t
r
ib
u
te
to
b
u
ild
in
g
o
p
en
-
s
o
u
r
ce
Dev
an
ag
ar
i
OC
R
f
o
r
p
r
in
ted
t
ex
t.
T
h
e
w
o
r
k
h
an
d
les
in
p
u
t
i
m
ag
es
at
ch
ar
ac
ter
lev
el
d
is
ti
n
ct
f
r
o
m
th
e
STOA
wo
r
k
s
wh
ich
ad
d
r
ess
at
t
h
e
wo
r
d
,
p
h
r
ase
o
r
lin
e
lev
el
i
n
p
u
t
im
a
g
es.
Acc
u
r
ac
y
o
f
7
5
.
7
3
%
in
ch
ar
ac
te
r
p
r
ed
ictio
n
is
ac
h
iev
ed
with
th
e
d
ataset
co
m
p
lex
ity
r
ed
u
ce
d
b
y
4
.
3
5
tim
es.
C
o
n
s
eq
u
en
tially
th
e
s
o
lu
tio
n
ca
n
b
e
u
s
ed
in
th
e
o
f
f
th
e
s
h
elf
co
m
p
u
tin
g
d
e
v
ices o
r
in
co
r
p
o
r
ated
i
n
em
b
ed
d
ed
s
y
s
tem
s
.
2.
P
RE
VIOU
S WO
RK
2
.
1
.
E
x
is
t
ing
s
y
s
t
em
Go
o
g
le
p
r
o
v
id
es
t
h
e
co
n
v
en
ti
o
n
al
tess
er
ac
t
4
as
an
o
p
en
-
s
o
u
r
ce
en
g
in
e
b
ased
o
n
L
STM
a
v
ailab
le
in
s
tan
d
alo
n
e
f
o
r
m
.
W
h
ile
Go
o
g
le
clo
u
d
v
is
io
n
is
th
eir
p
r
o
p
r
i
etar
y
im
ag
e
an
aly
s
is
web
p
la
tf
o
r
m
with
C
UDA
s
u
p
p
o
r
t
f
o
r
im
a
g
e
lab
ellin
g
,
f
ac
e
an
d
lan
d
m
ar
k
d
etec
tio
n
,
OC
R
,
an
d
s
af
e
s
ea
r
ch
b
ase
d
o
n
d
ee
p
lear
n
in
g
p
r
in
cip
les.
T
h
am
m
ar
ak
et
a
l.
[
1
]
i
n
th
eir
wo
r
k
o
n
“Co
m
p
ar
ativ
e
an
aly
s
is
o
f
T
ess
er
ac
t
an
d
Go
o
g
le
C
lo
u
d
Vis
io
n
f
o
r
T
h
ai
v
eh
icle
r
eg
is
tr
atio
n
ce
r
tific
ate”
b
r
in
g
s
o
u
t
th
e
d
if
f
er
en
ce
s
b
etwe
en
th
e
two
OC
R
p
latf
o
r
m
s
.
Su
m
m
ar
y
o
f
t
h
eir
ac
cu
r
ac
y
b
ased
o
n
in
p
u
t
im
ag
es
s
ize
is
s
h
o
wn
in
th
e
T
ab
le
1
.
C
lear
ly
licen
s
ed
C
lo
u
d
Vis
io
n
is
f
ar
ef
f
ec
tiv
e
th
a
n
o
p
en
-
s
o
u
r
ce
T
ess
er
ac
t
4
a
n
d
also
p
r
o
v
id
es
ad
d
itio
n
al
f
e
atu
r
es
to
f
ac
ilit
ate
au
to
m
atio
n
p
ip
elin
e
.
T
ab
le
1
.
T
ess
er
ac
t v
s
Go
o
g
le
C
lo
u
d
v
is
io
n
f
o
r
T
h
ai
p
late
r
e
co
g
n
itio
n
I
mag
e
S
i
z
e
A
c
c
u
r
a
c
y
(
%)
Te
ssera
c
t
C
l
o
u
d
V
i
si
o
n
La
r
g
e
6
0
.
2
1
9
4
.
7
2
S
t
a
n
d
a
r
d
/
M
e
d
i
u
m
4
1
.
1
7
7
9
.
8
5
S
mal
l
4
1
.
2
1
7
1
.
1
7
A
v
e
r
a
g
e
4
7
.
0
2
8
4
.
4
3
Su
n
et
a
l.
[
2
]
h
av
e
m
ad
e
elab
o
r
ate
o
b
s
er
v
atio
n
s
ab
o
u
t
th
e
d
ev
elo
p
m
en
t
o
f
th
e
Go
o
g
le
C
lo
u
d
Vis
io
n
m
o
d
els
u
s
in
g
p
r
etr
ain
ed
R
esNet
-
1
0
1
an
d
th
e
G
o
o
g
les
in
t
er
n
al
d
ataset
J
FT
-
3
0
0
M
b
ac
k
ed
b
y
m
eticu
lo
u
s
p
r
ep
r
o
ce
s
s
in
g
p
h
ase.
C
lass
if
icatio
n
s
ize
o
f
1
8
,
2
9
1
lab
els
cle
ar
ly
r
ef
lects
th
e
co
m
p
lex
ity
o
f
th
e
m
o
d
els
tr
ain
ed
an
d
th
e
i
n
teg
r
atio
n
o
f
d
is
tr
ib
u
ted
co
m
p
u
tin
g
r
eso
u
r
ce
s
f
o
r
th
e
tr
ai
n
in
g
p
r
o
ce
s
s
.
Hig
h
lig
h
t
is
th
e
u
s
e
o
f
asy
n
ch
r
o
n
o
u
s
g
r
ad
ien
t
d
escen
t
tr
ain
in
g
o
n
5
0
NVI
DI
A
K8
0
GPUs
.
1
7
p
ar
am
eter
s
er
v
er
s
wer
e
u
s
ed
to
s
to
r
e
an
d
u
p
d
ate
th
e
m
o
d
el
weig
h
ts
.
T
o
ac
co
m
m
o
d
ate
tr
ain
in
g
o
f
s
u
ch
co
m
p
le
x
ar
ch
itectu
r
e,
m
o
d
els
ar
e
s
p
lit
v
er
tically
in
to
5
0
eq
u
al
s
ized
s
u
b
-
f
u
lly
co
n
n
ec
ted
lay
er
an
d
d
is
tr
ib
u
ted
ar
o
u
n
d
d
if
f
er
en
t
p
ar
am
eter
s
er
v
er
s
.
Sear
ch
in
g
th
e
r
ig
h
t
s
et
o
f
h
y
p
er
-
p
ar
a
m
eter
s
r
eq
u
ir
es
s
ig
n
if
ican
t
ef
f
o
r
ts
.
E
x
ten
t
o
f
tim
e
t
ak
en
f
o
r
tr
ain
in
g
a
J
FT
m
o
d
el
is
h
in
ted
wh
er
e
4
e
p
o
ch
s
n
ee
d
e
d
2
m
o
n
th
s
o
n
5
0
K
-
8
0
GPUs
.
Su
ch
g
ig
an
tic
m
o
d
els
ar
e
less
id
ea
l
an
d
f
ea
s
ib
le
f
o
r
ev
o
lv
in
g
d
y
n
a
m
ic
n
ee
d
s
o
f
t
h
e
OC
R
u
s
e
ca
s
es.
E
asy
OC
R
[
3
]
is
an
o
p
e
n
ac
ce
s
s
OC
R
s
o
lu
tio
n
o
f
f
er
ed
b
y
J
AI
DE
D
AI
co
m
m
u
n
ity
g
r
o
u
p
.
T
h
e
ar
ch
itectu
r
e,
o
p
tio
n
s
f
o
r
p
r
e
-
p
r
o
ce
s
s
in
g
,
p
o
s
t
-
p
r
o
ce
s
s
in
g
p
h
ases
an
d
f
lex
ib
ilit
y
t
o
ex
ten
d
to
o
th
er
lan
g
u
ag
es
ar
e
m
en
tio
n
ed
in
th
ei
r
d
o
cu
m
en
tatio
n
.
I
n
t
h
eir
o
f
f
icial
p
a
p
e
r
,
it
is
s
tated
th
at
F1
-
s
co
r
e
r
ea
ch
es
to
0
.
8
5
6
.
B
r
ief
in
s
ig
h
t
to
th
e
ef
f
o
r
t
in
v
o
lv
ed
i
s
ex
p
r
ess
ed
as
1
4
h
to
tr
ain
with
m
in
im
u
m
m
an
u
al
s
u
p
er
v
is
io
n
f
o
r
2
5
k
iter
atio
n
with
8
R
T
X
3
0
9
0
T
i
s
y
s
tem
r
eso
u
r
ce
s
.
Half
o
f
th
e
GPU
wa
s
ass
ig
n
ed
f
o
r
tr
ain
i
n
g
,
an
d
h
a
lf
o
f
GPU
ass
ig
n
ed
f
o
r
s
u
p
er
v
is
io
n
s
ettin
g
.
T
h
is
clea
r
ly
ass
er
ts
th
e
STOA
s
o
lu
tio
n
s
o
f
f
er
ed
ar
e
o
u
t
o
f
th
e
r
ea
ch
o
f
th
e
o
f
f
s
h
elf
o
r
m
ed
iu
m
s
ized
co
m
p
u
tin
g
r
eso
u
r
ce
s
.
I
g
n
at
et
a
l.
[
4
]
in
th
eir
wo
r
k
h
av
e
f
ac
ilit
ated
p
u
b
licly
a
v
ailab
le
n
o
v
el
b
e
n
ch
m
ar
k
,
OC
R
4
MT
,
co
n
s
is
tin
g
o
f
r
ea
l
an
d
s
y
n
th
e
tic
d
ata,
en
r
ich
ed
with
n
o
is
e,
f
o
r
6
0
lo
w
-
r
eso
u
r
ce
lan
g
u
ag
es
in
lo
w
r
eso
u
r
ce
s
cr
ip
ts
.
Ma
in
o
b
s
er
v
atio
n
o
f
th
eir
r
esu
lts
was
th
at
th
e
b
est
av
ail
-
ab
le
OC
R
s
y
s
tem
s
wo
r
k
well
o
n
L
atin
s
cr
ip
ts
an
d
p
er
f
o
r
m
s
ig
n
if
ican
tly
w
o
r
s
e
o
n
n
o
n
-
L
atin
an
d
n
o
n
-
E
u
r
o
p
e
an
s
cr
ip
ts
(
e.
g
.
,
Per
s
o
-
Ar
ab
ic,
Kh
m
er
)
.
Mo
n
o
lin
g
u
al
d
atasets
th
u
s
f
r
a
m
ed
s
er
v
es
as
a
v
alu
ab
le
s
o
u
r
ce
o
f
d
ata
a
u
g
m
e
n
tatio
n
f
o
r
f
u
tu
r
e
r
esear
c
h
in
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
6
,
Decem
b
e
r
20
25
:
5
9
1
4
-
5
9
2
3
5916
im
p
r
o
v
in
g
m
ac
h
i
n
e
tr
an
s
latio
n
f
o
r
lo
w
r
eso
u
r
ce
lan
g
u
ag
es
.
I
t
clea
r
ly
h
ig
h
lig
h
ts
th
e
lac
u
n
ae
o
f
b
en
ch
m
ar
k
d
atasets
till
r
ec
en
t y
ea
r
s
.
Fan
et
a
l.
[
5
]
,
W
en
ze
k
et
a
l.
[
6
]
,
Go
y
al
et
a
l.
[
7
]
in
th
eir
f
in
d
in
g
s
h
av
e
ex
p
r
ess
ed
th
at
v
er
y
lar
g
e
f
r
ac
tio
n
o
f
th
e
lan
g
u
ag
es
s
p
o
k
en
b
y
th
e
wo
r
l
d
’
s
p
o
p
u
latio
n
h
av
e
lo
w
r
eso
u
r
ce
s
u
p
p
o
r
t
f
o
r
d
ev
el
o
p
in
g
th
e
OC
R
s
o
lu
tio
n
s
.
Sm
ith
et
a
l.
[
8
]
,
W
ick
et
a
l.
[
9
]
in
th
ei
r
o
b
s
er
v
atio
n
h
av
e
in
f
e
r
r
ed
t
h
a
t
m
o
s
t
o
f
th
e
OC
R
m
o
d
els
h
av
e
o
n
ly
b
ee
n
ev
alu
ated
o
n
a
h
an
d
f
u
l
o
f
lan
g
u
ag
e
s
.
No
n
av
ailab
ilit
y
o
f
p
u
b
lic
b
en
ch
m
ar
k
s
f
o
r
lo
w
r
eso
u
r
ce
s
cr
ip
ts
an
d
lan
g
u
a
g
e
s
is
th
e
b
o
ttlen
ec
k
.
As
a
r
esu
lt
,
m
etr
ics
an
d
to
o
ls
f
o
r
co
m
p
r
e
h
en
s
iv
e
ev
alu
atio
n
o
f
OC
R
s
o
lu
tio
n
s
,
p
ar
ticu
lar
ly
f
o
r
lo
w
-
r
eso
u
r
ce
lan
g
u
ag
es a
n
d
s
cr
ip
ts
,
is
s
till
an
o
p
en
p
r
o
b
lem
.
R
ijh
wan
i
et
a
l.
[
1
0
]
f
o
u
n
d
th
at
en
d
an
g
er
e
d
lan
g
u
a
g
e
lin
g
u
is
tic
ar
ch
iv
es
co
n
tain
th
o
u
s
an
d
s
o
f
s
ca
n
n
ed
d
o
cu
m
en
ts
an
d
s
u
ch
lan
g
u
ag
e
d
o
cu
m
en
ts
o
f
ten
c
o
n
tain
a
tr
an
s
latio
n
in
to
an
o
t
h
er
(
u
s
u
ally
h
ig
h
r
eso
u
r
ce
)
lan
g
u
ag
e.
T
h
ey
als
o
o
b
s
er
v
e
d
th
at
th
e
av
ailab
le
g
en
er
al
p
u
r
p
o
s
e
OC
R
s
o
lu
tio
n
s
ar
e
n
o
t
r
o
b
u
s
t
to
th
e
d
ata
-
s
ca
r
ce
s
ettin
g
o
f
en
d
an
g
er
ed
lan
g
u
a
g
es.
T
o
ad
d
r
es
s
th
is
p
r
o
b
lem
,
OC
R
p
o
s
t
-
co
r
r
ec
tio
n
m
o
d
u
les
ar
e
d
ev
elo
p
e
d
wh
ich
ar
e
tailo
r
ed
to
ea
s
e
th
e
tr
ain
in
g
p
r
o
ce
s
s
.
T
h
is
en
lig
h
ten
s
th
e
n
ee
d
f
o
r
lex
ico
n
f
r
ee
OC
R
ap
p
r
o
ac
h
es to
f
ac
ilit
ate
h
is
to
r
ical
d
o
cu
m
e
n
ts
tr
an
s
latio
n
.
Din
g
et
a
l.
[
1
1
]
wo
r
k
is
b
ased
o
n
co
m
p
r
ess
in
g
C
NN
-
DB
L
STM
m
o
d
els
f
o
r
OC
R
u
s
in
g
teac
h
er
-
s
tu
d
en
t
lear
n
in
g
a
n
d
T
u
ck
e
r
d
ec
o
m
p
o
s
itio
n
ap
p
r
o
ac
h
.
I
n
te
g
r
ated
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
a
n
d
d
ee
p
b
id
ir
ec
ti
o
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
DB
L
STM
)
b
ased
ch
ar
ac
ter
m
o
d
els
h
av
e
ac
h
iev
ed
ex
ce
llen
t
r
ec
o
g
n
itio
n
a
cc
u
r
ac
ies
o
n
OC
R
task
s
.
Mo
d
els
th
u
s
d
ev
is
ed
co
m
p
r
is
e
o
f
h
u
g
e
am
o
u
n
t
o
f
m
o
d
el
p
ar
a
m
eter
s
an
d
m
ass
iv
e
co
m
p
u
tatio
n
co
s
t.
So
,
to
d
ep
lo
y
C
NN
-
DB
L
S
T
M
m
o
d
els
in
p
r
o
d
u
cts
with
C
PU
s
er
v
er
,
u
r
g
en
t
n
ee
d
to
co
m
p
r
ess
an
d
ac
ce
ler
ate
th
em
as
m
u
ch
as
p
o
s
s
ib
le
is
s
u
g
g
ested
.
E
s
p
ec
ially
th
e
f
o
cu
s
o
n
C
NN
p
ar
t,
wh
ich
d
o
m
i
n
ates
b
o
th
p
ar
am
e
ter
s
an
d
co
m
p
u
tatio
n
is
p
o
in
t
ed
.
T
h
e
s
tu
d
y
clea
r
ly
d
r
aws
a
tten
tio
n
to
th
e
n
ee
d
o
f
in
d
e
p
th
f
o
cu
s
o
n
p
er
f
o
r
m
a
n
ce
an
d
e
f
f
icien
cy
f
ac
to
r
s
o
f
m
ac
h
in
e
lear
n
in
g
b
ased
m
o
d
el
s
.
Go
n
g
id
i
a
n
d
J
awa
h
ar
[
1
2
]
wo
r
k
is
o
n
h
a
n
d
wr
itten
te
x
t
r
ec
o
g
n
itio
n
f
o
r
I
n
d
ic
s
cr
ip
t
s
.
Sp
atial
tr
an
s
f
o
r
m
er
n
etwo
r
k
(
STN)
w
ith
af
f
in
e
tr
an
s
f
o
r
m
atio
n
(
AT
N)
,
an
d
th
in
-
p
late
s
p
lin
e
tr
an
s
f
o
r
m
atio
n
(
T
PS
)
f
o
r
r
ec
tify
in
g
v
ar
iatio
n
s
in
i
n
p
u
t
im
ag
es
o
f
p
h
r
ases
.
Dee
p
n
eu
r
al
n
etwo
r
k
s
,
VGG
an
d
R
E
SNET
f
o
r
f
ea
t
u
r
e
ex
tr
ac
tio
n
.
B
id
ir
ec
tio
n
al
L
S
T
M
(
B
L
STM
)
is
u
s
ed
f
o
r
s
eq
u
en
cin
g
th
e
ex
tr
ac
ted
f
ea
t
u
r
es.
C
o
n
n
ec
tio
n
is
t
tem
p
o
r
al
class
if
icatio
n
(
C
T
C
)
f
o
r
s
eq
u
en
ce
p
r
e
d
ictio
n
.
T
h
e
wo
r
k
is
in
f
lu
en
ce
d
b
y
E
asy
OC
R
.
T
h
e
to
tal
p
ar
am
eter
s
v
ar
y
f
r
o
m
8
.
3
5
M
to
1
7
.
1
5
M
f
o
r
m
o
d
el
v
er
s
io
n
s
.
On
B
en
g
ali
lan
g
u
ag
e
o
f
Dev
an
ag
ar
i
s
cr
ip
t,
C
E
R
=4
.
8
5
%,
W
E
R
=1
4
.
7
7
%
f
o
r
t
h
e
v
o
ca
b
u
lar
y
s
et
tr
a
in
ed
,
C
E
R
=3
.
7
1
%,
W
E
R
=1
6
.
6
5
%
o
u
ts
id
e
th
e
v
o
ca
b
u
lar
y
co
n
s
id
er
ed
.
I
n
-
s
p
i
te
o
f
in
-
b
u
ilt
wo
r
d
er
r
o
r
co
r
r
ec
tio
n
p
r
in
cip
le
u
s
ed
an
d
s
tan
d
ar
d
d
ee
p
lear
n
in
g
ar
ch
itectu
r
es
tr
ied
;
er
r
o
r
r
ate
i
s
n
o
tewo
r
th
y
.
T
h
e
r
esu
lts
in
d
icate
ch
allen
g
es,
u
n
iq
u
e
o
f
th
e
I
n
d
ic
s
cr
ip
ts
to
b
e
h
an
d
led
.
2
.
2
.
Su
m
m
a
ry
Do
m
ain
s
p
ec
if
ic
cu
s
to
m
izatio
n
is
th
e
m
ain
n
ee
d
f
o
r
r
ea
l
ti
m
e
ap
p
licatio
n
s
o
f
OC
R
.
Fo
r
ex
am
p
le,
n
ee
d
o
f
E
n
g
lis
h
n
u
m
er
als,
s
elec
ted
s
et
o
f
s
p
ec
ial
s
y
m
b
o
ls
to
b
e
p
ar
t
o
f
Dev
a
n
ag
ar
i
OC
R
alp
h
ab
et
s
et
h
as
to
b
e
co
n
s
id
er
e
d
.
I
n
s
u
ch
cir
cu
m
s
tan
ce
s
,
m
ac
h
in
e
lear
n
in
g
b
as
ed
m
o
d
els
f
all
ap
a
r
t
as
th
ey
a
r
e
r
ig
id
a
n
d
ar
e
n
o
t
ea
s
ily
ex
ten
s
ib
le.
C
u
s
to
m
izatio
n
is
as
g
o
o
d
as
b
u
ild
in
g
an
d
tr
ain
in
g
th
e
m
o
d
el
f
r
o
m
th
e
s
cr
atch
,
f
o
r
co
m
p
lex
m
o
d
els
th
is
r
esu
lts
to
b
e
d
au
n
tin
g
task
.
T
h
is
asp
ec
t
f
o
r
lar
g
e
alp
h
a
b
et
s
ets
tr
ig
g
er
s
to
e
x
p
lo
r
e
th
e
ef
f
ec
tiv
e
way
s
o
f
r
ed
u
cin
g
th
e
co
m
p
l
ex
ity
o
f
OC
R
s
o
lu
tio
n
s
.
Fu
r
t
h
er
co
n
v
en
tio
n
al
ap
p
r
o
ac
h
es
ar
e
lex
ico
n
b
ased
,
wh
ich
ag
ain
m
ak
es
th
e
s
o
lu
ti
o
n
s
in
ef
f
ec
tiv
e
wh
en
th
e
y
h
a
v
e
to
ac
co
m
m
o
d
ate
n
ew
wo
r
d
s
.
T
h
u
s
,
ch
ar
ac
te
r
lev
el
OC
R
s
u
f
f
ice
lex
ico
n
f
r
ee
ap
p
r
o
ac
h
r
e
q
u
ir
em
en
t a
ls
o
.
3.
P
RO
P
O
SE
D
WO
RK
3
.
1
.
Desig
n
T
h
e
d
e
s
i
g
n
o
f
t
h
e
p
r
o
p
o
s
e
d
w
o
r
k
f
o
c
u
s
e
s
o
n
t
w
o
m
ai
n
c
o
m
p
o
n
e
n
t
s
:
t
h
e
a
l
p
h
a
b
e
t
s
e
t
a
n
d
p
r
e
p
r
o
c
e
s
s
i
n
g
.
3
.
1
.
1
.
Alph
a
bet
s
et
Dev
an
ag
ar
i,
s
cr
ip
t
u
s
ed
as
b
a
s
e
f
o
r
San
s
k
r
it,
Pra
k
r
it,
Hin
d
i
,
Gu
jar
ati,
Ma
r
ath
i,
B
en
g
ali,
Nep
ali
an
d
o
th
er
r
eg
io
n
al
lan
g
u
ag
es,
f
in
d
s
an
ce
s
tr
al
r
o
o
ts
in
B
r
ah
m
i
s
cr
ip
t,
f
r
o
m
wh
ich
all
th
e
m
o
d
er
n
I
n
d
ian
wr
itin
g
s
y
s
tem
s
ar
e
d
er
iv
ed
.
W
ith
7
5
0
+
m
illi
o
n
g
lo
b
al
u
s
er
s
,
is
th
e
f
o
u
r
th
m
o
s
t
wid
ely
ad
o
p
ted
wr
itin
g
s
y
s
tem
in
th
e
wo
r
ld
.
Sti
ll,
Dev
an
ag
ar
i
OC
R
is
o
n
e
s
u
ch
lo
w
r
eso
u
r
ce
d
s
cr
i
p
t
s
h
o
r
t
o
f
b
en
ch
m
ar
k
e
d
d
atasets
an
d
p
ar
am
etr
ic
o
b
s
er
v
atio
n
s
.
Mo
d
er
n
Dev
an
a
g
ar
i,
s
cr
ip
t
is
co
m
p
o
s
ed
o
f
4
6
p
r
im
ar
y
ch
ar
ac
te
r
s
,
in
clu
d
in
g
1
3
v
o
wels
an
d
3
3
co
n
s
o
n
an
ts
.
I
t
is
an
ab
u
g
id
a,
wh
ich
m
ea
n
s
th
e
wr
itin
g
s
y
s
t
em
h
as
co
n
s
o
n
an
ts
with
an
in
h
er
en
t
v
o
wel
s
o
u
n
d
ca
lled
as
d
iacr
itic.
On
e
o
f
th
e
m
o
s
t
r
ec
o
g
n
izab
le
f
ea
tu
r
es
o
f
Dev
an
ag
ar
i
is
th
e
h
o
r
iz
o
n
tal
li
n
e
at
t
h
e
to
p
o
f
th
e
wo
r
d
ca
lled
S
h
ir
o
r
ek
h
a
th
at
g
r
o
u
p
s
th
e
ch
ar
ac
te
r
s
.
C
o
n
ju
n
ct
co
n
s
o
n
a
n
ts
,
wh
er
e
two
o
r
m
o
r
e
co
n
s
o
n
a
n
ts
co
m
b
in
e
to
f
o
r
m
a
s
in
g
le
ch
a
r
ac
ter
,
ar
e
co
m
m
o
n
in
Dev
an
a
g
ar
i
an
d
ad
d
to
th
e
r
ich
n
ess
an
d
co
m
p
lex
ity
o
f
th
e
s
cr
ip
t.
T
h
e
b
asic
co
n
s
o
n
an
t
ch
ar
ac
ter
s
ar
e
o
r
g
an
ized
in
g
r
o
u
p
s
d
ep
en
d
in
g
o
n
th
eir
m
a
n
n
er
o
f
ar
ticu
latio
n
,
i.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
-
8
7
0
8
Dev
a
n
a
g
a
r
i o
p
tica
l c
h
a
r
a
cter reco
g
n
itio
n
o
f
p
r
in
ted
text
(
M
a
la
th
i P
.
)
5917
v
o
iced
,
v
o
icele
s
s
,
an
d
n
asal.
T
h
e
p
r
o
n
u
n
ciatio
n
o
f
th
e
c
h
ar
ac
ter
s
m
ig
h
t
v
ar
y
s
lig
h
tly
d
ep
en
d
in
g
o
n
th
e
lan
g
u
ag
e
u
s
in
g
th
e
Dev
a
n
ag
ar
i
s
cr
ip
t.
I
t
h
as
its
o
wn
n
u
m
b
er
r
e
p
r
esen
tatio
n
th
at
f
o
llo
ws
Hin
d
u
-
Ar
ab
i
c
n
u
m
er
al
s
y
s
tem
.
T
h
e
alp
h
ab
et
s
et
S
co
n
s
id
er
e
d
f
o
r
th
e
cu
r
r
en
t
wo
r
k
is
th
e
m
a
jo
r
s
u
b
s
et
o
f
m
o
d
e
r
n
Dev
an
ag
ar
i
s
cr
ip
ts
f
u
ll
-
f
led
g
e
d
alp
h
ab
et
s
et.
I
t
co
n
s
is
ts
o
f
:
i)
p
r
im
a
r
y
ch
ar
ac
ter
s
et
o
f
v
o
wels
{
13
}
,
co
n
s
o
n
an
t
s
{
33
}
,
c
o
m
m
o
n
ly
u
s
ed
co
n
s
o
n
a
n
t
co
n
ju
n
cts
{
3
}
s
h
o
wn
in
T
ab
le
2
with
th
eir
Un
ico
d
e
r
ep
r
esen
tatio
n
s
er
ies
an
d
ii)
B
ar
ak
h
ad
i
ch
ar
ac
ter
s
et
is
co
m
p
o
s
ed
o
f
c
o
n
ju
n
cts f
o
r
m
ed
f
r
o
m
co
m
b
in
atio
n
o
f
co
n
s
o
n
an
ts
with
v
o
we
l d
iacr
itics
.
Sam
p
le
s
et
o
f
B
ar
ak
h
ad
i is sh
o
wn
in
T
ab
le
3
.
T
ab
le
2
.
Un
ico
d
e
o
f
Dev
an
a
g
a
r
i sy
m
b
o
ls
X
0
1
2
3
4
5
6
7
8
9
A
B
C
D
E
F
U
+
0
9
0
X
ं
ं
अ
आ
इ
ई
उ
ऊ
ऋ
ए
U
+
0
9
1
X
ऐ
ओ
औ
क
ख
ग
घ
ङ
च
छ
ज
झ
ञ
ट
U
+
0
9
2
X
ठ
ड
ढ
ण
त
थ
द
ध
न
प
फ
ब
भ
म
य
U
+
0
9
3
X
र
ल
ळ
व
श
ष
स
ह
ं
िं
U
+
0
9
4
X
ं
ं
ं
ं
ं
ं
ं
ं
T
ab
le
3
.
B
ar
ak
h
a
d
i sam
p
le
V
o
w
e
l
s
अ
आ
इ
ई
उ
ऊ
ऋ
ए
ऐ
ओ
औ
अ
अ
D
i
a
c
r
i
t
i
c
ं
िं
ं
ं
ं
ं
ं
ं
ं
ं
ं
ं
B
a
r
a
k
h
a
d
i
S
a
mp
l
e
क
क
िक
क
क
क
क
क
क
क
क
क
क
T
o
t
a
l
s
i
z
e
o
f
t
h
e
al
p
h
a
b
e
t
s
e
t
S
c
o
n
s
i
d
e
r
e
d
f
o
r
cl
a
s
s
i
f
i
c
a
t
i
o
n
a
g
g
r
e
g
a
t
e
s
t
o
4
8
1
(
4
3
2
+
4
9
)
i
n
t
h
e
p
r
e
s
e
n
t
w
o
r
k
.
L
e
t
=
,
=
,
=
,
i
n
p
u
t
s
i
z
e
o
f
e
a
c
h
i
m
a
g
e
f
o
r
=
,
th
e
n
B
ar
a
k
h
a
d
i
co
n
t
ai
n
s
(
−
1
)
∗
(
+
)
d
is
ti
n
ct
c
h
a
r
a
cte
r
s
.
T
h
e
o
r
et
ica
ll
y
wit
h
c
o
n
v
en
ti
o
n
al
a
p
p
r
o
ac
h
w
h
ic
h
u
s
e
wo
r
d
,
p
h
r
ase
o
r
s
e
n
te
n
ce
le
v
el
i
n
p
u
t
f
o
r
tr
ai
n
i
n
g
t
h
e
d
atas
et,
c
o
m
p
le
x
it
y
f
o
r
r
ec
o
g
n
it
io
n
ca
n
th
e
n
b
e
r
e
p
r
ese
n
t
e
d
as
:
=
[
+
+
+
(
−
1
)
∗
(
+
)
]
∗
(
1
)
W
ith
ch
ar
ac
ter
lev
el
ap
p
r
o
ac
h
th
e
d
ataset
co
m
p
lex
ity
f
o
r
r
ec
o
g
n
itio
n
ca
n
b
e
r
ep
r
esen
ted
as
:
=
[
+
+
+
(
−
1
)
+
(
+
)
]
∗
(
2
)
T
h
en
r
e
d
u
ctio
n
o
f
d
ataset
co
m
p
lex
ity
ca
n
th
e
n
b
e
f
o
r
m
u
lated
as
:
(
,
,
,
)
=
[
x
+
y
+
z
+
(
x
−
1
)
∗
(
y
+
z
)
]
∗
s
[
+
+
+
(
−
1
)
+
(
+
)
]
∗
(
3
)
Fo
r
Dev
an
ag
ar
i
alp
h
ab
et
s
et
c
o
n
s
id
er
ed
in
th
e
p
r
esen
t w
o
r
k
-
=
13
,
=
33
,
=
3
,
=
2000
.
Su
b
s
titu
tin
g
in
(
3
)
,
r
e
d
u
ctio
n
o
f
d
ataset
co
m
p
lex
ity
(
,
,
,
)
=
7
.
2
tim
es.
Als
o
,
th
e
n
u
m
b
er
o
f
in
p
u
t
im
ag
e
s
am
p
les
tak
en
f
o
r
ea
ch
o
f
th
e
class
f
o
r
tr
ain
in
g
th
e
m
o
d
el
ca
n
b
e
e
v
en
ly
d
is
tr
ib
u
ted
.
T
h
is
in
t
u
r
n
h
elp
s
to
f
ig
u
r
e
o
u
t th
e
class
es
f
o
r
wh
ich
th
e
class
if
icatio
n
m
o
d
el
ef
f
icien
cy
is
less
an
d
co
r
r
ec
tio
n
s
to
b
e
tak
en
.
W
ith
co
n
v
en
tio
n
al
ap
p
r
o
ac
h
,
d
is
tr
ib
u
tio
n
o
f
in
p
u
t
s
am
p
les
ac
r
o
s
s
th
e
alp
h
ab
et
s
et
is
n
eith
er
m
ain
tain
ed
n
o
r
ass
es
s
ed
.
I
n
co
n
s
eq
u
e
n
ce
th
e
y
ad
o
p
t
to
r
ev
er
s
e
en
g
in
ee
r
t
h
e
d
ataset
s
ize
f
o
r
ex
p
er
im
e
n
ts
b
ased
o
n
r
esu
lts
.
Als
o
,
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
u
s
ed
,
an
d
tim
e
ta
k
en
will
b
e
t
u
n
ed
to
m
ee
t
th
e
ex
p
ec
ted
e
f
f
i
cien
cy
b
y
t
r
ial
an
d
er
r
o
r
.
3
.
1
.
2
.
P
re
pro
ce
s
s
ing
Sam
p
le
im
ag
es
o
f
d
ataset
co
n
s
id
er
ed
f
o
r
test
in
g
is
s
h
o
wn
in
Fig
u
r
e
1
(
a
)
.
E
asy
OC
R
API
[
3
]
ap
p
lied
o
n
th
e
im
ag
es
g
av
e
to
tally
wr
o
n
g
p
r
ed
ictio
n
s
.
Go
o
g
le
C
lo
u
d
Vis
io
n
API
[
1
3
]
r
esp
o
n
d
e
d
as
b
ad
d
ata.
T
h
e
im
ag
es
wer
e
th
en
b
in
a
r
ized
,
Ots
u
th
r
esh
o
ld
an
d
th
i
n
n
in
g
o
p
er
atio
n
s
wer
e
ap
p
lied
ex
p
licitly
as
s
h
o
wn
in
Fig
u
r
e
1
(
b
)
.
b
ef
o
r
e
ca
llin
g
th
e
E
asy
OC
R
API
an
d
Go
o
g
le
C
lo
u
d
Vis
io
n
API
to
r
esp
o
n
d
a
p
p
r
o
p
r
iately
.
T
h
u
s
,
clar
ity
o
f
in
p
u
t
im
ag
es
also
im
p
ac
ts
th
e
r
esu
lt
o
f
class
if
icatio
n
.
T
h
is
h
ig
h
lig
h
ts
th
e
r
o
le
o
f
v
ar
i
o
u
s
p
r
e
-
p
r
o
ce
s
s
in
g
o
p
e
r
atio
n
s
to
b
e
u
s
ed
o
n
t
h
e
cu
s
to
m
er
e
n
d
to
u
s
e
th
e
ex
is
tin
g
s
o
lu
tio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
6
,
Decem
b
e
r
20
25
:
5
9
1
4
-
5
9
2
3
5918
(
a)
(
b
)
Fig
u
r
e
1
.
Sam
p
le
i
n
p
u
t im
a
g
es (
a)
b
ef
o
r
e
p
r
e
-
p
r
o
ce
s
s
in
g
an
d
(
b
)
af
ter
p
r
e
-
p
r
o
ce
s
s
in
g
3
.
2
.
Alg
o
rit
hm
M
o
d
u
l
e
s
d
e
v
e
l
o
p
e
d
i
n
c
la
s
s
i
f
ica
t
i
o
n
a
n
d
r
e
c
o
g
n
i
t
i
o
n
p
h
a
s
es
o
f
t
h
e
O
C
R
p
i
p
el
i
n
e
is
s
h
o
w
n
i
n
F
i
g
u
r
e
2
.
a.
Pre
-
p
r
o
ce
s
s
in
g
:
T
h
e
in
p
u
t
g
r
e
y
im
ag
e
is
f
ir
s
t
co
n
v
er
ted
in
to
b
in
ar
y
im
ag
e
u
s
in
g
h
is
to
g
r
am
-
b
ased
Ots
u
b
in
ar
izatio
n
m
eth
o
d
.
T
h
is
h
e
lp
s
to
elim
in
ate
th
e
o
u
tlier
p
ix
els
in
tr
o
d
u
ce
d
d
u
r
in
g
d
o
c
u
m
en
t
s
ca
n
n
in
g
p
r
o
ce
s
s
.
T
h
e
w
o
r
d
im
ag
es
ar
e
n
ex
t
p
r
o
ce
s
s
ed
f
o
r
s
k
ew
a
n
d
s
lan
t
co
r
r
ec
tio
n
s
.
Scen
e
t
ex
t
OC
R
h
as
to
h
an
d
le
co
l
o
r
ed
in
p
u
t im
ag
es.
C
o
n
v
er
s
io
n
to
g
r
ey
im
a
g
es f
o
r
m
s
an
ad
d
itio
n
al
s
tep
in
th
at
c
o
n
tex
t.
b.
T
h
in
n
in
g
:
I
t
h
elp
s
to
f
o
cu
s
o
n
v
ital
p
ix
els
th
at
tr
an
s
f
o
r
m
to
f
ea
tu
r
es
o
f
th
e
c
h
ar
ac
ter
s
an
d
also
r
ed
u
ce
th
e
o
v
er
all
co
m
p
u
tatio
n
s
in
v
o
lv
e
d
.
Ma
jo
r
ity
o
f
th
e
p
r
in
ts
d
o
es
n
o
t
h
a
v
e
ch
ar
ac
te
r
s
o
f
u
n
if
o
r
m
th
ick
n
ess
to
ap
p
ly
g
lo
b
al
th
in
n
in
g
o
p
er
a
tio
n
wh
ich
ca
n
r
esu
lt
in
m
is
lead
in
g
f
ea
tu
r
es.
R
ef
in
ed
p
r
o
ce
s
s
ca
lled
s
k
eleto
n
izatio
n
av
ailab
le
as
l
ib
r
ar
y
is
m
ad
e
u
s
e
with
f
u
r
t
h
er
s
co
p
e
to
f
in
e
t
u
n
e
it.
I
t
f
o
r
m
s
a
way
o
f
n
o
r
m
aliza
tio
n
w.
r
.
t
th
ick
n
ess
o
f
ch
ar
ac
ter
in
im
ag
es.
I
t
also
b
o
o
s
ts
to
g
en
er
alize
th
e
s
o
lu
ti
o
n
,
v
o
id
o
f
f
o
n
t
d
ep
en
d
e
n
ce
.
c.
C
h
ar
ac
ter
s
eg
m
en
tatio
n
:
C
h
al
len
g
es
f
ac
ed
f
o
r
ch
ar
ac
ter
im
ag
es
s
eg
m
en
tatio
n
is
d
e
p
icted
in
Fig
u
r
e
3
(
a
)
.
On
m
ask
in
g
S
h
ir
o
r
ek
h
a
in
wo
r
d
im
a
g
e,
th
e
s
y
m
b
o
l
im
a
g
es
s
eg
r
eg
atio
n
s
h
o
wn
in
Fig
u
r
e
3
(
b
)
o
n
ly
lead
s
to
wr
o
n
g
p
r
ed
ictio
n
s
.
B
u
t
s
em
an
tically
th
e
ch
ar
ac
ter
im
a
g
es
to
b
e
s
eg
r
eg
ated
is
s
h
o
wn
in
Fi
g
u
r
e
3
(
c
)
.
e
v
en
b
ef
o
r
e
t
h
e
p
r
o
ce
s
s
o
f
tex
t r
ec
o
g
n
itio
n
.
d.
C
h
ar
ac
ter
lis
tin
g
:
T
h
e
ch
ar
ac
ter
s
s
eg
m
en
ted
ef
f
ec
tiv
ely
h
as
to
b
e
ass
em
b
led
in
th
e
wo
r
d
in
th
e
s
am
e
o
r
d
er
at
th
e
en
d
o
f
r
ec
o
g
n
itio
n
p
r
o
c
ess
as
we
h
av
e
u
s
ed
wo
r
d
im
ag
es
f
o
r
test
in
g
th
e
b
u
ilt
class
if
icatio
n
m
o
d
el
d
u
r
in
g
ex
p
e
r
im
en
ts
.
L
is
tin
g
i
s
im
p
o
r
tan
t,
p
ar
ticu
lar
ly
wh
e
n
m
u
ltip
le
wo
r
d
im
a
g
es
ar
e
p
ass
ed
as
b
atch
in
p
u
t to
th
e
OC
R
p
ip
elin
e
d
u
r
i
n
g
ex
p
er
im
en
ts
.
e.
Diac
r
itic
s
eg
m
en
tatio
n
:
T
h
is
p
h
ase
f
o
r
m
s
th
e
f
o
ca
l
p
o
in
t
f
o
r
ef
f
ec
tiv
en
ess
o
f
th
e
OC
R
.
T
h
e
p
o
s
s
ib
le
v
o
wel
d
iacr
itic
p
o
s
itio
n
,
r
elativ
e
to
th
e
b
ase
co
n
s
o
n
a
n
t
in
D
ev
an
ag
ar
i
s
cr
ip
t
is
d
e
p
icted
in
Fig
u
r
e
5
.
T
h
e
ap
p
r
o
ac
h
u
s
ed
to
s
eg
m
en
t
th
e
d
iacr
itic
f
r
o
m
th
e
b
ase
co
n
s
o
n
an
t
v
ar
ies
f
o
r
ea
ch
r
eg
i
o
n
.
Seg
m
en
tin
g
th
e
d
iacr
itic
th
at
lies
in
th
e
b
o
tto
m
r
eg
io
n
is
th
e
m
o
s
t
ch
allen
g
i
n
g
task
.
Po
s
itio
n
o
f
th
e
d
iacr
itic
r
elativ
e
to
th
e
b
ase
co
n
s
o
n
an
t
u
s
ed
v
ar
ies
ac
r
o
s
s
th
e
ch
ar
ac
ter
s
in
th
e
al
p
h
ab
et
s
et
as
s
h
o
wn
in
Fig
u
r
e
4
.
Hen
ce
,
th
is
p
h
ase
p
lay
s
a
p
iv
o
tal
r
o
le
in
t
h
e
ef
f
icien
cy
o
f
th
e
d
esig
n
ed
O
C
R
s
o
lu
tio
n
s
.
f.
Diac
r
itic
p
r
ed
ictio
n
:
Patter
n
m
atch
in
g
b
ased
tec
h
n
iq
u
e
is
u
s
ed
to
id
en
tif
y
th
e
d
iac
r
itic.
Par
ticu
lar
ly
h
is
to
g
r
am
s
wer
e
u
s
ed
f
o
r
h
o
r
izo
n
tal
an
d
v
er
tical
p
r
o
f
ilin
g
f
o
llo
wed
b
y
cr
o
s
s
s
ec
t
a
n
aly
s
is
d
o
n
e
o
n
th
e
d
iacr
itic
r
eg
io
n
o
f
th
e
in
p
u
t
i
m
ag
es.
Size
o
f
t
h
e
s
eg
r
eg
ate
d
d
iacr
itic
is
co
m
p
a
r
ab
ly
s
m
all
er
th
an
t
h
e
b
ase
co
n
s
o
n
an
t
in
co
n
ju
n
ct
ch
ar
ac
t
er
s
.
Als
o
d
if
f
er
en
ce
s
b
etwe
en
d
iacr
itics
is
r
ep
r
esen
ted
b
y
v
er
y
f
ew
p
ix
els
in
th
e
in
p
u
t
im
ag
es
esp
ec
ially
af
ter
th
in
n
in
g
p
r
o
ce
s
s
u
s
ed
in
s
t
ep
2
.
So
,
it
h
as
to
b
e
p
ad
d
ed
o
n
all
f
o
u
r
s
id
es
an
d
en
lar
g
ed
u
s
in
g
in
ter
p
o
latio
n
tech
n
i
q
u
e
to
f
ac
ilit
ate
in
d
e
p
en
d
en
t
d
iacr
itic r
ec
o
g
n
itio
n
o
f
th
e
co
n
j
u
n
ct.
g.
B
ase
ch
ar
ac
ter
lo
ca
lizatio
n
:
Me
an
s
m
o
o
th
in
g
o
f
th
e
s
h
o
r
t
h
o
r
izo
n
tal
an
d
v
e
r
tical
lin
es
th
at
ar
e
p
ar
t
o
f
v
o
wel
o
r
co
n
s
o
n
an
t
a
p
ar
t
f
r
o
m
S
h
ir
o
r
e
k
h
a
a
n
d
M
atr
a
f
ac
i
litates
f
in
e
tu
n
in
g
o
f
ch
ar
ac
te
r
s
h
ap
es
in
t
h
e
im
ag
es.
R
esu
r
r
ec
tio
n
h
elp
s
to
r
esto
r
e
th
e
f
ea
tu
r
e
p
ix
els lo
s
t d
u
r
in
g
th
in
n
in
g
in
s
tep
2
.
h.
Key
f
ea
tu
r
e
ex
tr
ac
tio
n
:
T
o
p
a
n
d
b
o
tto
m
p
en
d
an
t
v
er
tices
a
r
e
th
e
m
ajo
r
k
ey
f
ea
tu
r
es
in
v
o
wels
o
r
b
ase
co
n
s
o
n
an
ts
o
f
t
h
e
ch
ar
ac
ter
.
T
h
eir
id
en
tific
atio
n
h
elp
s
to
d
if
f
er
en
tiate
th
e
m
u
lti
cla
s
s
lab
els
d
u
r
in
g
class
if
icatio
n
.
i.
B
ase
ch
ar
ac
ter
(
v
o
wel
o
r
co
n
s
o
n
an
t
in
c
o
n
ju
n
ct)
p
r
e
d
ictio
n
:
Statis
tical
m
atch
in
g
,
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
s
(
C
NN)
an
d
tr
an
s
f
o
r
m
er
v
is
io
n
p
r
in
cip
les
wer
e
weig
h
ed
.
I
t
is
th
is
m
o
d
u
le
wh
ich
d
eter
m
in
es
th
e
co
m
p
lex
ity
o
f
th
e
e
n
tire
s
y
s
tem
.
T
h
e
m
ain
o
b
jectiv
es
to
ap
p
r
o
ac
h
th
is
task
wer
e
:
i)
Ov
er
all
OC
R
s
o
lu
tio
n
to
b
e
m
ed
iu
m
s
ized
s
o
lu
tio
n
,
ii)
Ma
n
u
al
ef
f
o
r
t
n
ee
d
e
d
to
p
r
ep
ar
e
d
ata
s
et
h
as
to
b
e
k
ep
t
in
b
o
u
n
d
s
,
iii)
C
o
m
p
u
tin
g
r
eso
u
r
ce
s
r
eq
u
ir
ed
f
o
r
th
e
co
m
p
lete
p
r
o
ce
s
s
s
h
o
u
ld
b
e
im
p
lem
en
ted
wi
th
o
f
f
th
e
s
h
elf
d
ev
ices.
Statis
tical
m
atch
in
g
,
C
NN
an
d
tr
an
s
f
o
r
m
er
v
is
io
n
p
r
in
cip
les
wh
ich
ar
e
b
ased
o
n
d
ee
p
lea
r
n
in
g
m
ac
h
in
es
wer
e
weig
h
ed
.
C
o
n
s
id
er
in
g
th
e
p
r
io
r
ities
o
f
all
t
h
e
m
en
ti
o
n
ed
o
b
jectiv
es
C
NN
p
r
in
cip
le
was
f
o
u
n
d
to
b
e
id
ea
l.
B
asic
s
h
ap
e
f
ea
tu
r
es
lik
e
cu
r
v
atu
r
es,
lin
es
an
d
th
eir
alig
n
m
en
t
i
n
b
ase
ch
ar
ac
ter
wer
e
co
n
s
id
er
ed
to
ar
r
i
v
e
at
a
m
u
lti
-
lev
el
class
if
icatio
n
ap
p
r
o
ac
h
.
Acc
o
r
d
in
g
ly
th
e
class
es
ca
teg
o
r
ized
as
=
[
क
,
फ
,
ऋ
]
,
[
त
,
ल
]
,
[
न
,
ग
]
,
[
च
,
ज
,
ञ
,
न
,
ब
,
व
,
,
]
,
[
थ
,
ध
,
भ
]
,
[
घ
,
ण
,
प
,
म
,
य
,
ष
]
,
[
ख
,
झ
,
स
,
श
,
,
अ
,
आ
,
ओ
,
औ
]
,
[
द
,
ह
,
इ
,
ई
]
,
[
र
,
ए
,
ऐ
]
,
[
ङ
,
छ
,
ट
,
ठ
,
ड
,
ढ
,
उ
,
ऊ
]
.
T
h
is
r
esu
lts
in
ten
ef
f
ec
tiv
e
s
m
all
s
ized
tr
ain
ed
m
o
d
els th
an
a
m
e
d
iu
m
s
ized
s
in
g
le
class
if
ier
m
o
d
el
f
o
r
u
s
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
-
8
7
0
8
Dev
a
n
a
g
a
r
i o
p
tica
l c
h
a
r
a
cter reco
g
n
itio
n
o
f
p
r
in
ted
text
(
M
a
la
th
i P
.
)
5919
j.
C
h
ar
ac
ter
p
r
ed
ictio
n
:
T
h
e
r
es
u
lts
o
f
d
iacr
itic
clas
s
if
icatio
n
an
d
b
ase
v
o
wel/
co
n
s
o
n
an
t
cl
ass
if
icatio
n
ar
e
co
m
b
in
ed
th
r
u
r
u
le
-
b
ased
d
e
cisi
o
n
s
.
C
a
s
es
wh
er
e
cr
o
s
s
c
h
ec
k
an
d
au
t
o
co
r
r
ec
tio
n
r
eq
u
ir
ed
f
o
r
v
alid
ch
ar
ac
ter
p
r
e
d
ictio
n
is
p
er
f
o
r
m
ed
else u
n
s
u
cc
ess
f
u
l p
r
e
d
ictio
n
is
in
d
icate
d
.
k.
W
o
r
d
p
r
ed
ictio
n
:
T
h
e
ch
ar
ac
t
er
s
p
r
ed
icted
ar
e
ass
em
b
led
i
n
to
wo
r
d
as
p
a
r
t
o
f
wo
r
d
lis
t.
T
h
e
in
p
u
t
wo
r
d
im
ag
es
wer
e
s
u
p
p
lied
in
b
atch
es
d
u
r
in
g
test
in
g
p
u
r
p
o
s
e.
T
h
is
h
elp
ed
u
s
to
co
m
p
ar
e
th
e
C
r
o
s
s
Sect
O
C
R
d
esig
n
with
Go
o
g
le
C
lo
u
d
Vi
s
io
n
an
d
E
asy
OC
R
m
o
d
els
o
n
b
o
th
C
E
R
an
d
W
E
R
as
in
T
a
b
le
4
i
n
s
tead
o
f
ju
s
t CER
co
m
p
ar
is
o
n
as c
h
ar
a
cter
lev
el
OC
R
is
o
u
r
p
r
im
ar
y
ap
p
r
o
ac
h
.
Fig
u
r
e
2
.
C
r
o
s
s
s
ec
t O
C
R
alg
o
r
ith
m
(
a)
(
b
)
(
c)
Fig
u
r
e
3
.
C
h
ar
ac
ter
s
eg
m
en
tat
io
n
(
a)
wo
r
d
im
a
g
e,
(
b
)
s
y
m
b
o
l seg
r
eg
atio
n
,
a
n
d
(
c
)
ch
ar
ac
te
r
s
eg
r
eg
atio
n
Fig
u
r
e
4
.
Diac
r
itic p
o
s
itio
n
in
g
Fig
u
r
e
5
.
Dev
a
n
ag
ar
i c
h
ar
ac
te
r
lay
o
u
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
6
,
Decem
b
e
r
20
25
:
5
9
1
4
-
5
9
2
3
5920
4.
E
XP
E
R
I
M
E
N
T
S
E
T
UP
C
o
m
m
o
n
ly
u
s
ed
,
Un
ico
d
e
b
a
s
ed
Ak
s
h
ar
a
f
o
n
t
was
u
s
ed
t
o
g
en
e
r
ate
2
0
0
0
im
a
g
es
f
o
r
e
ac
h
o
f
t
h
e
b
ase
ch
ar
ac
ter
(
v
o
wel/co
n
s
o
n
an
t)
s
et
o
f
s
ize
3
8
as
p
ar
t
o
f
th
e
in
itial
tr
ain
d
ataset.
Go
o
g
le
C
o
lab
Py
th
o
n
p
latf
o
r
m
was u
s
ed
f
o
r
d
ataset
p
r
ep
ar
atio
n
,
m
o
d
el
tr
ain
in
g
an
d
OC
R
p
ip
elin
e
ex
ec
u
tio
n
.
I
m
ag
e
d
ata
g
en
er
ato
r
s
wer
e
u
s
ed
to
g
en
er
ate
im
ag
es
to
co
m
p
lem
en
t
th
e
tr
ain
d
ataset
d
y
n
am
ically
.
T
h
is
f
ac
ilit
ates
th
e
tr
ain
ed
class
if
icatio
n
m
o
d
els
to
b
e
in
d
ep
en
d
e
n
t
o
f
f
o
n
t
u
s
ed
an
d
f
o
n
t
s
ize.
C
NN
[
1
4
]
–
[
1
6
]
was
f
o
u
n
d
as
an
o
p
tim
u
m
ap
p
r
o
ac
h
k
ee
p
in
g
MN
I
ST
[
1
7
]
as
r
e
f
er
en
ce
.
Acc
o
r
d
in
g
ly
,
th
r
ee
m
o
d
els
wer
e
d
esig
n
ed
th
e
as
s
h
o
wn
in
Fig
u
r
e
6
an
d
e
x
p
er
im
e
n
ted
.
E
f
f
icien
cy
o
f
m
o
d
el
I
I
I
o
n
tr
a
in
in
g
was
f
o
u
n
d
to
b
e
3
%
-
4
%
b
etter
th
an
m
o
d
el
I
I
a
n
d
5
%
-
6
%
b
etter
th
an
m
o
d
el
I
.
Als
o
,
th
e
to
tal
tr
ain
in
g
p
ar
am
eter
s
o
f
C
NN
in
m
o
d
el
I
I
I
is
g
r
ea
tly
r
ed
u
ce
d
c
o
m
p
ar
ed
to
m
o
d
el
I
an
d
I
I
.
Mo
d
el
I
I
I
ar
c
h
itectu
r
e
i
s
s
h
o
wn
in
Fig
u
r
e
7
.
Fig
u
r
e
6
.
C
NN
m
o
d
els co
n
f
ig
u
r
atio
n
f
o
r
b
ase
c
h
ar
ac
ter
ex
p
er
im
en
ted
Fig
u
r
e
7
.
Mo
d
el
I
I
I
ar
ch
itectu
r
e
I
n
a
ll th
e
th
r
ee
m
o
d
els
R
e
L
U
[
1
8
]
was
t
h
e
a
cti
v
a
ti
o
n
f
u
n
cti
o
n
u
s
ed
f
o
r
t
h
e
h
id
d
en
la
y
er
s
a
n
d
So
f
t
Ma
x
ac
ti
v
ati
o
n
f
u
n
c
ti
o
n
i
n
t
h
e
o
u
t
p
u
t
la
y
e
r
.
St
o
c
h
as
tic
g
r
a
d
i
en
t
d
esce
n
t
(
SG
D)
wit
h
m
o
m
en
tu
m
[
1
9
]
–
[
2
1
]
was
th
e
lo
s
s
ev
al
u
at
io
n
f
u
n
ct
io
n
u
s
e
d
f
o
r
tr
ai
n
i
n
g
d
at
a
i
n
b
a
tc
h
es
.
Dr
o
p
o
u
t
[
2
2
]
m
ec
h
a
n
is
m
w
as
u
s
e
d
t
o
p
r
e
v
e
n
t
o
v
e
r
f
itti
n
g
o
f
t
h
e
d
atas
et
.
T
h
e
b
es
t
c
h
o
s
e
n
d
esi
g
n
was
th
e
C
NN
m
o
d
el
I
I
I
[
2
3
]
,
[
2
4
]
.
E
a
ch
m
u
lt
i
c
lass
t
a
k
es
10
-
1
5
it
er
ati
o
n
s
o
f
b
a
tc
h
d
at
aset
t
o
r
e
ac
h
a
cc
u
r
ac
y
o
f
9
0
+
a
n
d
to
tal
tim
e
o
f
2
0
-
1
5
0
m
i
n
u
tes
f
o
r
t
r
a
in
in
g
d
e
p
e
n
d
i
n
g
o
n
i
ts
s
i
ze
.
T
h
e
tr
a
in
i
n
g
d
e
tai
ls
o
f
a
s
a
m
p
le
m
u
l
t
i
cl
ass
[
घ
,
ण
,
प
,
म
,
य
,
ष
]
is
s
h
o
wn
i
n
F
i
g
u
r
e
8
.
L
o
w
est
v
al
id
ati
o
n
l
o
s
s
o
n
t
r
a
i
n
i
n
g
th
e
te
n
d
i
f
f
er
e
n
t
m
u
l
ti
cl
ass
m
o
d
els
w
er
e
r
ea
ch
ed
at
d
i
f
f
er
e
n
t
e
p
o
c
h
s
.
S
till
,
th
e
c
o
n
s
is
t
e
n
c
y
o
f
th
e
te
n
m
u
lt
i c
lass
m
o
d
els
a
cr
o
s
s
t
h
e
b
asi
c
m
et
r
i
cs w
as
f
o
u
n
d
a
n
d
is
d
e
p
i
c
ted
i
n
Fi
g
u
r
e
9
.
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
-
8
7
0
8
Dev
a
n
a
g
a
r
i o
p
tica
l c
h
a
r
a
cter reco
g
n
itio
n
o
f
p
r
in
ted
text
(
M
a
la
th
i P
.
)
5921
Fig
u
r
e
8
.
Mu
lti
-
class
tr
ain
in
g
o
f
[
घ
,
ण
,
प
,
म
,
य
,
ष
]
Fig
u
r
e
9
.
Mu
lti
-
class
tr
ain
in
g
m
etr
ics co
n
v
er
g
e
n
ce
5.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
Fo
r
test
in
g
,
r
ef
in
ed
s
u
b
s
et
o
f
Mile
Dev
an
ag
ar
i
d
ataset
[
2
5
]
is
u
s
ed
.
T
ex
t
d
o
cu
m
en
ts
ar
e
s
ca
n
n
ed
u
s
in
g
UM
AX
Ast
r
a
5
4
0
0
,
HP
Scan
-
jet
2
2
0
0
c
s
ca
n
n
e
r
s
at
3
0
0
d
p
i
r
eso
lu
tio
n
an
d
s
to
r
e
d
in
8
-
b
it
g
r
e
y
f
o
r
m
at.
T
h
e
im
ag
es
ar
e
e
x
tr
ac
ted
f
r
o
m
n
ewsp
ap
er
s
,
m
a
g
az
in
es,
b
o
o
k
s
an
d
laser
p
r
i
n
ted
d
o
cu
m
en
ts
.
Var
iatio
n
s
in
p
r
in
tin
g
s
ty
le
an
d
s
izes
was
e
n
s
u
r
ed
wh
ile
s
elec
tin
g
.
Ab
o
u
t
1
0
0
p
ag
es
ar
e
s
ca
n
n
e
d
an
d
a
r
e
s
eg
m
en
ted
in
to
wo
r
d
s
o
r
p
h
r
ase
im
ag
es
b
y
an
au
to
m
atic
p
r
o
ce
s
s
.
R
ef
in
ed
s
u
b
s
et
o
f
1
0
k
was
u
s
ed
f
o
r
test
in
g
p
u
r
p
o
s
e
wh
e
r
e
th
e
wo
r
d
s
ar
e
f
r
o
m
2
to
8
ch
a
r
ac
ter
s
in
len
g
th
.
T
h
e
two
s
tan
d
ar
d
m
etr
ics
u
s
ed
f
o
r
ass
ess
i
n
g
th
e
p
e
r
f
o
r
m
an
ce
o
f
OC
R
s
y
s
tem
s
ar
e
ch
ar
ac
ter
er
r
o
r
r
ate
(
C
E
R
)
.
(
ℎ
)
=
.
ℎ
.
ℎ
∗
100%
(
)
=
.
.
∗
100%
.
Acc
o
r
d
in
g
ly
,
th
e
r
esu
lts
o
b
tai
n
ed
f
r
o
m
th
e
e
x
p
er
im
e
n
ts
ar
e
s
h
o
wn
in
T
ab
le
4.
T
ab
le
4
.
R
esu
lts
Er
r
o
r
R
a
t
e
A
c
c
u
r
a
c
y
(
%)
G
o
o
g
l
e
C
l
o
u
d
V
i
si
o
n
O
C
R
Ea
sy
O
C
R
C
r
o
ss S
e
c
t
O
C
R
C
ER
1
4
.
9
3
2
5
.
3
2
9
.
2
3
W
ER
9
.
7
3
2
0
.
1
7
2
4
.
4
7
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
6
,
Decem
b
e
r
20
25
:
5
9
1
4
-
5
9
2
3
5922
E
r
r
o
r
r
ate
o
b
s
er
v
e
d
in
o
u
r
w
o
r
k
C
r
o
s
s
Sect
OC
R
ar
e
d
u
e
to
m
is
s
clas
s
if
icatio
n
o
f
g
r
o
u
p
s
o
f
clo
s
e
r
esem
b
lin
g
ch
a
r
ac
ter
s
.
I
n
g
r
o
u
p
s
lik
e
[
ङ
,
ड
]
[
अ
अ
अ
]
th
e
d
o
t
p
i
x
els
ca
n
b
e
l
o
s
t
d
u
r
in
g
th
in
n
in
g
s
tag
e
o
f
OC
R
p
r
o
ce
s
s
a
s
o
n
o
u
tlier
o
r
n
o
is
e.
I
n
(
3
)
,
n
ew
v
al
u
es
o
f
v
ar
iab
les
will
th
en
b
e
x
=
1
1
,
y
=
3
2
,
z
=
3
an
d
s
=
2
0
0
0
.
(
,
,
,
)
=
4
.
3
5
.
Hen
c
e
th
e
d
ataset
co
m
p
lex
ity
r
e
d
u
ctio
n
f
ac
to
r
is
4
.
3
5
in
p
lace
o
f
7
.
2
.
I
t
s
u
g
g
ests
to
ex
p
lo
r
e
alter
n
ativ
es
t
o
co
u
n
ter
th
e
s
id
e
e
f
f
ec
ts
o
f
th
in
n
in
g
s
tag
e.
E
r
r
o
r
r
ate
is
also
d
u
e
t
o
th
e
c
o
n
ju
n
cts
lik
e
ऌ
wh
er
e
th
e
d
iacr
itic
◌
ं
ca
n
n
o
t
b
e
d
is
tin
g
u
is
h
ed
f
r
o
m
th
e
b
ase
co
n
s
o
n
an
t
ल
d
u
r
in
g
class
i
f
icatio
n
as
s
ep
ar
ate
en
titi
es.
T
h
is
ca
n
b
e
h
an
d
led
in
p
o
s
t
p
r
o
ce
s
s
in
g
s
tag
e
o
f
OC
R
u
s
in
g
n
-
g
r
am
s
in
th
e
v
o
ca
b
u
lar
y
t
o
r
ed
u
ce
C
E
R
.
E
asy
O
C
R
clas
s
if
icatio
n
m
o
d
el
[
2
6
]
h
as
8
.
3
M
p
ar
a
m
eter
s
.
Go
o
g
le
C
lo
u
d
Vis
io
n
OC
R
m
o
d
els
[
2
7
]
,
[
2
8
]
w
h
ich
is
b
ased
o
n
v
is
io
n
tr
an
s
f
o
r
m
er
s
p
r
in
cip
le
h
as
1
5
M
-
1
4
7
M
p
a
r
am
eter
s
.
T
h
e
p
r
esen
t
wo
r
k
m
o
d
el
h
as o
n
ly
2
1
0
.
5
3
k
B
p
a
r
am
eter
s
.
6.
CO
NCLU
SI
O
N
W
e
o
b
s
er
v
e
th
at
C
r
o
s
s
Sect
OC
R
d
ev
elo
p
ed
in
th
is
wo
r
k
la
g
s
in
ter
m
s
o
f
ac
c
u
r
ac
y
b
y
4
.
3
%
-
1
4
.
7
%
co
m
p
ar
ed
with
STOA
s
o
lu
tio
n
s
.
At
th
e
s
am
e
tim
e
d
ataset
s
ize
u
s
ed
f
o
r
tr
ain
in
g
is
r
e
d
u
ce
d
b
y
4
.
3
5
tim
es
an
d
C
NN
m
o
d
el
u
s
ed
f
o
r
class
if
icatio
n
is
m
u
ch
s
im
p
ler
.
T
h
e
id
ea
l
d
ataset
co
m
p
lex
ity
r
ed
u
ctio
n
f
o
r
th
e
d
ataset
co
n
s
id
er
ed
s
h
o
u
ld
b
e
7
.
2
tim
es.
T
h
e
in
v
esti
g
atio
n
h
as
th
u
s
p
r
o
v
id
e
d
,
in
s
ig
h
t
in
to
alter
n
ate
ap
p
r
o
ac
h
es
to
tack
le
m
ac
h
in
e
tr
an
s
latio
n
o
f
lo
w
r
eso
u
r
ce
lan
g
u
ag
es.
Fe
asib
ilit
y
o
f
in
co
r
p
o
r
atin
g
th
e
OC
R
s
o
lu
tio
n
s
o
n
s
tan
d
alo
n
e
an
d
o
f
f
th
e
s
h
elf
elec
tr
o
n
ic
d
ev
ices
f
o
r
lan
g
u
ag
es
wh
er
e
th
e
lar
g
e
alp
h
ab
et
s
et
ca
n
b
e
tr
an
s
f
o
r
m
e
d
to
s
im
p
le
r
ep
r
esen
tatio
n
is
p
r
o
v
ed
.
T
h
e
ef
f
icien
cy
o
f
th
e
p
r
o
p
o
s
ed
OC
R
s
o
lu
tio
n
ca
n
b
e
en
h
an
ce
d
b
y
f
u
r
th
e
r
wo
r
k
in
g
o
n
th
in
n
in
g
p
r
o
ce
s
s
u
s
ed
,
b
y
f
in
e
tu
n
in
g
th
e
C
NN
b
ase
m
o
d
els,
b
y
u
s
in
g
ef
f
ec
tiv
e
way
s
f
o
r
d
iacr
itics
s
ep
ar
atio
n
an
d
with
a
d
o
p
tio
n
o
f
p
o
s
t
p
r
o
ce
s
s
in
g
co
r
r
ec
tio
n
tech
n
iq
u
es.
T
h
e
lex
ico
n
f
r
ee
ap
p
r
o
ac
h
o
f
t
h
e
p
r
o
p
o
s
ed
OC
R
i
s
n
o
t
co
n
f
in
e
d
to
an
y
v
o
ca
b
u
lar
y
is
an
ad
d
ed
ad
v
an
tag
e
.
T
h
e
ap
p
r
o
ac
h
c
an
b
e
u
s
ed
f
o
r
o
th
er
I
n
d
ic
s
cr
ip
ts
with
d
iacr
itic a
n
d
co
n
ju
n
ct
lig
at
u
r
es sy
s
tem
s
.
RE
F
E
R
E
NC
E
S
[
1
]
K
.
Th
a
mm
a
r
a
k
,
P
.
K
o
n
g
k
l
a
,
Y
.
S
i
r
i
sa
t
h
i
t
k
u
l
,
a
n
d
S
.
I
n
t
a
k
o
su
m,
“
C
o
m
p
a
r
a
t
i
v
e
a
n
a
l
y
si
s
o
f
T
e
ssera
c
t
a
n
d
G
o
o
g
l
e
C
l
o
u
d
V
i
si
o
n
f
o
r
Th
a
i
v
e
h
i
c
l
e
r
e
g
i
s
t
r
a
t
i
o
n
c
e
r
t
i
f
i
c
a
t
e
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
rn
a
l
o
f
E
l
e
c
t
ri
c
a
l
a
n
d
C
o
m
p
u
t
e
r
En
g
i
n
e
e
r
i
n
g
,
v
o
l
.
1
2
,
n
o
.
2
,
p
p
.
1
8
4
9
–
1
8
5
8
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
e
c
e
.
v
1
2
i
2
.
p
p
1
8
4
9
-
1
8
5
8
.
[
2
]
C
.
S
u
n
,
A
.
S
h
r
i
v
a
st
a
v
a
,
S
.
S
i
n
g
h
,
a
n
d
A
.
G
u
p
t
a
,
“
R
e
v
i
si
t
i
n
g
u
n
r
e
a
s
o
n
a
b
l
e
e
f
f
e
c
t
i
v
e
n
e
ss o
f
d
a
t
a
i
n
d
e
e
p
l
e
a
r
n
i
n
g
e
r
a
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
I
E
EE
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
C
o
m
p
u
t
e
r
Vi
s
i
o
n
,
v
o
l
.
2
0
1
7
-
O
c
t
o
b
,
p
p
.
8
4
3
–
8
5
2
,
2
0
1
7
,
d
o
i
:
1
0
.
1
1
0
9
/
I
C
C
V
.
2
0
1
7
.
9
7
.
[
3
]
Jai
d
e
d
A
I
,
“
Ea
s
y
O
C
R
,
”
J
AI
D
ED
AI
.
h
t
t
p
s
:
/
/
w
w
w
.
j
a
i
d
e
d
.
a
i
/
e
a
s
y
o
c
r
(
a
c
c
e
ss
e
d
D
e
c
1
8
,
2
0
2
4
)
.
[
4
]
O
.
I
g
n
a
t
,
J
.
M
a
i
l
l
a
r
d
,
V
.
C
h
a
u
d
h
a
r
y
,
a
n
d
F
.
G
u
z
má
n
,
“
O
C
R
i
m
p
r
o
v
e
s
mac
h
i
n
e
t
r
a
n
s
l
a
t
i
o
n
f
o
r
l
o
w
-
r
e
so
u
r
c
e
l
a
n
g
u
a
g
e
s,
”
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
A
n
n
u
a
l
M
e
e
t
i
n
g
o
f
t
h
e
Asso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s
,
p
p
.
1
1
6
4
–
1
1
7
4
,
2
0
2
2
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
2
.
f
i
n
d
i
n
g
s
-
a
c
l
.
9
2
.
[
5
]
A
.
F
a
n
e
t
a
l
.
,
“
B
e
y
o
n
d
En
g
l
i
s
h
-
C
e
n
t
r
i
c
M
u
l
t
i
l
i
n
g
u
a
l
M
a
c
h
i
n
e
Tr
a
n
sl
a
t
i
o
n
,
”
Pre
p
r
i
n
t
,
a
r
Xi
v
:
2
0
1
0
.
1
1
1
2
5
,
2
0
2
0
.
[
6
]
G
.
W
e
n
z
e
k
e
t
a
l
.
,
“
C
C
N
e
t
:
e
x
t
r
a
c
t
i
n
g
h
i
g
h
q
u
a
l
i
t
y
m
o
n
o
l
i
n
g
u
a
l
d
a
t
a
set
s
f
r
o
m
w
e
b
c
r
a
w
l
d
a
t
a
,
”
L
RE
C
2
0
2
0
-
1
2
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
L
a
n
g
u
a
g
e
Re
s
o
u
r
c
e
s
a
n
d
Ev
a
l
u
a
t
i
o
n
,
C
o
n
f
e
re
n
c
e
Pr
o
c
e
e
d
i
n
g
s
,
2
0
2
0
,
p
p
.
4
0
0
3
–
4
0
1
2
.
[
7
]
N
.
G
o
y
a
l
e
t
a
l
.
,
“
T
h
e
F
LO
R
ES
-
1
0
1
e
v
a
l
u
a
t
i
o
n
b
e
n
c
h
mar
k
f
o
r
l
o
w
-
r
e
so
u
r
c
e
a
n
d
m
u
l
t
i
l
i
n
g
u
a
l
ma
c
h
i
n
e
t
r
a
n
s
l
a
t
i
o
n
,
”
T
r
a
n
sa
c
t
i
o
n
s
o
f
t
h
e
Asso
c
i
a
t
i
o
n
f
o
r
C
o
m
p
u
t
a
t
i
o
n
a
l
L
i
n
g
u
i
st
i
c
s
,
v
o
l
.
1
0
,
p
p
.
5
2
2
–
5
3
8
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
6
2
/
t
a
c
l
_
a
_
0
0
4
7
4
.
[
8
]
R
.
S
m
i
t
h
,
“
A
n
o
v
e
r
v
i
e
w
o
f
t
h
e
t
e
ss
e
r
a
c
t
O
C
R
e
n
g
i
n
e
,
”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
D
o
c
u
m
e
n
t
A
n
a
l
y
si
s a
n
d
Re
c
o
g
n
i
t
i
o
n
,
I
C
D
AR
,
2
0
0
7
,
v
o
l
.
2
,
p
p
.
6
2
9
–
6
3
3
,
d
o
i
:
1
0
.
1
1
0
9
/
I
C
D
A
R
.
2
0
0
7
.
4
3
7
6
9
9
1
.
[
9
]
C
.
W
i
c
k
,
C
.
R
e
u
l
,
a
n
d
F
.
P
u
p
p
e
,
“
C
a
l
a
mari
−
a
h
i
g
h
-
p
e
r
f
o
r
ma
n
c
e
Te
n
s
o
r
F
l
o
w
-
b
a
s
e
d
d
e
e
p
l
e
a
r
n
i
n
g
p
a
c
k
a
g
e
f
o
r
o
p
t
i
c
a
l
c
h
a
r
a
c
t
e
r
r
e
c
o
g
n
i
t
i
o
n
,
”
D
i
g
i
t
a
l
H
u
m
a
n
i
t
i
e
s
Q
u
a
rt
e
rl
y
,
v
o
l
.
1
4
,
n
o
.
2
,
2
0
2
0
.
[
1
0
]
S
.
R
i
j
h
w
a
n
i
,
A
.
A
n
a
st
a
so
p
o
u
l
o
s,
a
n
d
G
.
N
e
u
b
i
g
,
“
O
C
R
p
o
st
c
o
r
r
e
c
t
i
o
n
f
o
r
e
n
d
a
n
g
e
r
e
d
l
a
n
g
u
a
g
e
t
e
x
t
s,”
i
n
EM
N
L
P
2
0
2
0
-
2
0
2
0
C
o
n
f
e
re
n
c
e
o
n
Em
p
i
r
i
c
a
l
Me
t
h
o
d
s
i
n
N
a
t
u
r
a
l
L
a
n
g
u
a
g
e
Pro
c
e
ss
i
n
g
,
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
C
o
n
f
e
re
n
c
e
,
2
0
2
0
,
p
p
.
5
9
3
1
–
5
9
4
2
,
d
o
i
:
1
0
.
1
8
6
5
3
/
v
1
/
2
0
2
0
.
e
m
n
l
p
-
m
a
i
n
.
4
7
8
.
[
1
1
]
H
.
D
i
n
g
,
K
.
C
h
e
n
,
a
n
d
Q
.
H
u
o
,
“
C
o
m
p
r
e
ss
i
n
g
C
N
N
-
D
B
LS
TM
m
o
d
e
l
s
f
o
r
O
C
R
w
i
t
h
t
e
a
c
h
e
r
-
st
u
d
e
n
t
l
e
a
r
n
i
n
g
a
n
d
T
u
c
k
e
r
d
e
c
o
m
p
o
si
t
i
o
n
,
”
Pa
t
t
e
r
n
R
e
c
o
g
n
i
t
i
o
n
,
v
o
l
.
9
6
,
p
.
1
0
6
9
5
7
,
D
e
c
.
2
0
1
9
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
p
a
t
c
o
g
.
2
0
1
9
.
0
7
.
0
0
2
.
[
1
2
]
S
.
G
o
n
g
i
d
i
a
n
d
C
.
V
Jawa
h
a
r
,
“
I
I
I
T
-
I
N
D
I
C
-
HW
-
w
o
r
d
s:
a
d
a
t
a
s
e
t
f
o
r
I
n
d
i
c
h
a
n
d
w
r
i
t
t
e
n
t
e
x
t
r
e
c
o
g
n
i
t
i
o
n
,
”
i
n
L
e
c
t
u
r
e
N
o
t
e
s
i
n
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
(
i
n
c
l
u
d
i
n
g
s
u
b
seri
e
s
L
e
c
t
u
r
e
N
o
t
e
s
i
n
Art
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
L
e
c
t
u
r
e
N
o
t
e
s
i
n
Bi
o
i
n
f
o
rm
a
t
i
c
s)
,
v
o
l
.
1
2
8
2
4
LN
C
S
,
2
0
2
1
,
p
p
.
4
4
4
–
4
5
9
.
[
1
3
]
G
o
o
g
l
e
,
“
G
o
o
g
l
e
C
l
o
u
d
V
i
si
o
n
d
o
c
u
me
n
t
a
t
i
o
n
,
”
G
o
o
g
l
e
C
l
o
u
d
.
2
0
2
4
,
A
c
c
e
ss
e
d
:
D
e
c
1
8
,
2
0
2
4
.
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s
:
/
/
c
l
o
u
d
.
g
o
o
g
l
e
.
c
o
m
/
v
i
si
o
n
/
d
o
c
s/
[
1
4
]
Y
.
Le
C
u
n
,
L.
B
o
t
t
o
u
,
Y
.
B
e
n
g
i
o
,
a
n
d
P
.
H
a
f
f
n
e
r
,
“
G
r
a
d
i
e
n
t
-
b
a
s
e
d
l
e
a
r
n
i
n
g
a
p
p
l
i
e
d
t
o
d
o
c
u
m
e
n
t
r
e
c
o
g
n
i
t
i
o
n
,
”
P
ro
c
e
e
d
i
n
g
s
o
f
t
h
e
I
EEE
,
v
o
l
.
8
6
,
n
o
.
1
1
,
p
p
.
2
2
7
8
–
2
3
2
3
,
1
9
9
8
,
d
o
i
:
1
0
.
1
1
0
9
/
5
.
7
2
6
7
9
1
.
[
1
5
]
T.
G
u
o
,
J.
D
o
n
g
,
H
.
L
i
,
a
n
d
Y
.
G
a
o
,
“
S
i
m
p
l
e
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
o
n
i
m
a
g
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
i
n
2
0
1
7
I
E
EE
2
n
d
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
B
i
g
D
a
t
a
An
a
l
y
si
s
,
I
C
BD
A
2
0
1
7
,
2
0
1
7
,
p
p
.
7
2
1
–
7
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
I
C
B
D
A
.
2
0
1
7
.
8
0
7
8
7
3
0
.
[
1
6
]
M
.
D
.
Ze
i
l
e
r
a
n
d
R
.
F
e
r
g
u
s
,
“
V
i
su
a
l
i
z
i
n
g
a
n
d
u
n
d
e
r
st
a
n
d
i
n
g
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
t
w
o
r
k
s,”
i
n
L
e
c
t
u
r
e
N
o
t
e
s
i
n
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
(
i
n
c
l
u
d
i
n
g
su
b
ser
i
e
s
L
e
c
t
u
re
N
o
t
e
s
i
n
Art
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
L
e
c
t
u
re
N
o
t
e
s
i
n
B
i
o
i
n
f
o
rm
a
t
i
c
s)
,
v
o
l
.
8
6
8
9
LN
C
S
,
n
o
.
P
A
R
T
1
,
S
p
r
i
n
g
e
r
I
n
t
e
r
n
a
t
i
o
n
a
l
P
u
b
l
i
sh
i
n
g
,
2
0
1
4
,
p
p
.
8
1
8
–
8
3
3
.
[
1
7
]
“
P
a
p
e
r
s
w
i
t
h
c
o
d
e
-
i
ma
g
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
o
n
I
mag
e
N
e
t
.
”
2
0
2
2
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s
:
/
/
p
a
p
e
r
sw
i
t
h
c
o
d
e
.
c
o
m/
s
o
t
a
/
i
ma
g
e
-
c
l
a
ss
i
f
i
c
a
t
i
o
n
-
on
-
i
ma
g
e
n
e
t
.
(
a
c
c
e
sse
d
D
e
c
1
8
,
2
0
2
4
)
.
[
1
8
]
V
.
N
a
i
r
a
n
d
G
.
E
.
H
i
n
t
o
n
,
“
R
e
c
t
i
f
i
e
d
l
i
n
e
a
r
u
n
i
t
s
i
m
p
r
o
v
e
R
e
st
r
i
c
t
e
d
B
o
l
t
z
ma
n
n
ma
c
h
i
n
e
s,
”
i
n
I
C
ML
2
0
1
0
-
P
ro
c
e
e
d
i
n
g
s
,
2
7
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
Ma
c
h
i
n
e
L
e
a
r
n
i
n
g
,
2
0
1
0
,
p
p
.
8
0
7
–
8
1
4
.
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
-
8
7
0
8
Dev
a
n
a
g
a
r
i o
p
tica
l c
h
a
r
a
cter reco
g
n
itio
n
o
f
p
r
in
ted
text
(
M
a
la
th
i P
.
)
5923
[
1
9
]
N
.
K
.
S
i
n
h
a
a
n
d
M
.
P
.
G
r
i
sc
i
k
,
“
A
st
o
c
h
a
st
i
c
a
p
p
r
o
x
i
m
a
t
i
o
n
m
e
t
h
o
d
,
”
I
EEE
T
ra
n
s
a
c
t
i
o
n
s
o
n
S
y
st
e
m
s,
M
a
n
,
a
n
d
C
y
b
e
rn
e
t
i
c
s
,
v
o
l
.
S
M
C
-
1
,
n
o
.
4
,
p
p
.
3
3
8
–
3
4
4
,
O
c
t
.
1
9
7
1
,
d
o
i
:
1
0
.
1
1
0
9
/
TS
M
C
.
1
9
7
1
.
4
3
0
8
3
1
6
.
[
2
0
]
I
.
S
u
t
sk
e
v
e
r
,
J.
M
a
r
t
e
n
s,
G
.
D
a
h
l
,
a
n
d
G
.
H
i
n
t
o
n
,
“
O
n
t
h
e
i
m
p
o
r
t
a
n
c
e
o
f
i
n
i
t
i
a
l
i
z
a
t
i
o
n
a
n
d
mo
me
n
t
u
m
i
n
d
e
e
p
l
e
a
r
n
i
n
g
,
”
i
n
3
0
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
Ma
c
h
i
n
e
L
e
a
r
n
i
n
g
,
I
C
ML
2
0
1
3
,
2
0
1
3
,
n
o
.
P
A
R
T
3
,
p
p
.
2
1
7
6
–
2
1
8
4
.
[
2
1
]
M
.
A
b
a
d
i
e
t
a
l
.
,
“
T
e
n
s
o
r
F
l
o
w
:
a
sy
s
t
e
m
f
o
r
l
a
r
g
e
-
s
c
a
l
e
ma
c
h
i
n
e
l
e
a
r
n
i
n
g
,
”
i
n
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
1
2
t
h
U
S
EN
I
X
S
y
m
p
o
si
u
m
o
n
O
p
e
r
a
t
i
n
g
S
y
s
t
e
m
s
D
e
si
g
n
a
n
d
I
m
p
l
e
m
e
n
t
a
t
i
o
n
,
O
S
D
I
2
0
1
6
,
2
0
1
6
,
p
p
.
2
6
5
–
2
8
3
.
[
2
2
]
N
.
S
r
i
v
a
st
a
v
a
,
G
.
H
i
n
t
o
n
,
A
.
K
r
i
z
h
e
v
sk
y
,
I
.
S
u
t
s
k
e
v
e
r
,
a
n
d
R
.
S
a
l
a
k
h
u
t
d
i
n
o
v
,
“
D
r
o
p
o
u
t
:
a
s
i
mp
l
e
w
a
y
t
o
p
r
e
v
e
n
t
n
e
u
r
a
l
n
e
t
w
o
r
k
s
f
r
o
m o
v
e
r
f
i
t
t
i
n
g
,
”
J
o
u
r
n
a
l
o
f
M
a
c
h
i
n
e
L
e
a
r
n
i
n
g
R
e
se
a
rc
h
,
v
o
l
.
1
5
,
n
o
.
1
,
p
p
.
1
9
2
9
–
1
9
5
8
,
2
0
1
4
.
[
2
3
]
P
.
P
a
t
r
o
,
K
.
K
u
m
a
r
,
a
n
d
G
.
S
u
r
e
s
h
K
u
m
a
r
,
“
A
p
p
l
i
c
a
t
i
o
n
s
o
f
t
h
r
e
e
l
a
y
e
r
C
N
N
i
n
i
ma
g
e
p
r
o
c
e
ss
i
n
g
,
”
J
o
u
r
n
a
l
o
f
A
d
v
a
n
c
e
d
Re
se
a
rc
h
i
n
D
y
n
a
m
i
c
a
l
a
n
d
C
o
n
t
r
o
l
S
y
st
e
m
s
,
v
o
l
.
2
0
1
8
,
n
o
.
1
S
p
e
c
i
a
l
I
ssu
e
,
p
p
.
5
1
0
–
5
1
2
,
2
0
1
8
.
[
2
4
]
F
.
Le
i
,
X
.
L
i
u
,
Q
.
D
a
i
,
a
n
d
B
.
W
.
K
.
Li
n
g
,
“
S
h
a
l
l
o
w
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
f
o
r
i
m
a
g
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
S
N
Ap
p
l
i
e
d
S
c
i
e
n
c
e
s
,
v
o
l
.
2
,
n
o
.
1
,
2
0
2
0
,
d
o
i
:
1
0
.
1
0
0
7
/
s
4
2
4
5
2
-
0
1
9
-
1
9
0
3
-
4.
[
2
5
]
M
I
LE
-
G
r
o
u
p
,
“
M
I
LE
d
a
t
a
se
t
,
”
M
I
LE
-
G
r
o
u
p
2
0
1
2
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s:
/
/
s
i
t
e
s.
g
o
o
g
l
e
.
c
o
m
/
a
/
mi
l
e
g
r
o
u
p
.
n
e
t
/
w
w
w
/
h
o
m
e
(
a
c
c
e
s
se
d
D
e
c
1
8
,
2
0
2
4
).
[
2
6
]
B
.
S
h
i
,
X
.
B
a
i
,
a
n
d
C
.
Y
a
o
,
“
A
n
e
n
d
-
to
-
e
n
d
t
r
a
i
n
a
b
l
e
n
e
u
r
a
l
n
e
t
w
o
r
k
f
o
r
i
mag
e
-
b
a
se
d
s
e
q
u
e
n
c
e
r
e
c
o
g
n
i
t
i
o
n
a
n
d
i
t
s
a
p
p
l
i
c
a
t
i
o
n
t
o
sce
n
e
t
e
x
t
r
e
c
o
g
n
i
t
i
o
n
,
”
I
EE
E
T
r
a
n
s
a
c
t
i
o
n
s
o
n
P
a
t
t
e
r
n
A
n
a
l
y
si
s
a
n
d
M
a
c
h
i
n
e
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
3
9
,
n
o
.
1
1
,
p
p
.
2
2
9
8
–
2
3
0
4
,
2
0
1
7
,
d
o
i
:
1
0
.
1
1
0
9
/
TPA
M
I
.
2
0
1
6
.
2
6
4
6
3
7
1
.
[
2
7
]
Z.
Z
o
n
g
e
t
a
l
.
,
“
S
e
l
f
-
sl
i
mm
e
d
v
i
si
o
n
t
r
a
n
s
f
o
r
m
e
r
,
”
Pre
p
r
i
n
t
,
a
rX
i
v
:
2
1
1
1
.
1
2
6
2
4
,
2
0
2
1
.
[
2
8
]
Z
.
Zo
n
g
,
“
S
e
l
f
-
sl
i
mm
e
d
V
i
si
o
n
Tr
a
n
sf
o
r
mer
(
EC
C
V
2
0
2
2
)
,
”
G
i
t
H
u
b
,
2
0
2
2
.
h
t
t
p
s
:
/
/
g
i
t
h
u
b
.
c
o
m/
S
e
n
s
e
-
X
/
S
i
T
(
a
c
c
e
ss
e
d
D
e
c
1
8
,
2
0
2
4
).
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
Ma
la
t
h
i
P
.
is an
a
ss
istan
t
p
ro
f
e
ss
o
r
in
th
e
De
p
a
rtme
n
t
o
f
In
fo
r
m
a
ti
o
n
S
c
ien
c
e
a
n
d
En
g
i
n
e
e
rin
g
a
t
Dr.
Am
b
e
d
k
a
r
In
s
ti
tu
te
o
f
Tec
h
n
o
lo
g
y
.
S
h
e
h
a
s
m
o
r
e
t
h
a
n
2
8
y
e
a
r
s
o
f
t
e
a
c
h
i
n
g
e
x
p
e
r
i
e
n
c
e
a
n
d
i
s
c
u
r
r
e
n
t
l
y
p
u
r
s
u
i
n
g
h
e
r
d
o
c
t
o
r
a
l
r
e
s
e
a
rc
h
u
n
d
e
r
V
i
s
v
e
s
v
a
ra
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
s
i
t
y
(
V
T
U
)
,
Be
l
a
g
a
v
i
.
He
r
re
s
e
a
r
c
h
i
n
t
e
re
s
t
s
i
n
c
l
u
d
e
c
o
m
p
u
t
e
r
v
i
s
i
o
n
a
n
d
o
p
t
i
c
a
l
c
h
a
r
a
c
t
e
r
r
e
c
o
g
n
i
t
i
o
n
(
O
C
R
)
f
o
r
I
n
d
i
c
s
c
r
i
p
ts
.
S
h
e
c
a
n
b
e
c
o
n
t
a
c
te
d
a
t
e
m
a
i
l
:
m
a
l
a
t
h
i
p
.
i
s
@
d
ra
i
t
.
e
d
u
.
i
n
.
Cha
n
d
r
a
k
a
n
t
h
G
.
Pu
j
a
r
i
is
a
p
r
o
fe
ss
o
r
i
n
t
h
e
De
p
a
rtm
e
n
t
o
f
M
a
ste
r
o
f
Co
m
p
u
ter
Ap
p
li
c
a
ti
o
n
s
a
t
Dr.
Am
b
e
d
k
a
r
In
stit
u
te
o
f
Tec
h
n
o
l
o
g
y
.
W
it
h
o
v
e
r
3
0
y
e
a
rs
o
f
tea
c
h
in
g
e
x
p
e
rien
c
e
,
h
is
a
c
a
d
e
m
ic
a
n
d
re
se
a
rc
h
in
tere
sts
c
o
v
e
r
s
o
ftwa
re
e
n
g
in
e
e
ri
n
g
,
f
u
z
z
y
lo
g
ic,
a
n
d
a
rt
ifi
c
ial
in
telli
g
e
n
c
e
.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
la
v
b
a
n
g
a
l
o
re
6
@g
m
a
il
.
c
o
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.