I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l
.
13
,
No
.
4
,
Dece
m
b
er
2
0
2
4
,
p
p
.
4
9
2
7
~
4
9
3
8
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai
.
v
13
.
i
4
.
pp
4
9
2
7
-
4
9
3
8
4927
J
o
ur
na
l
ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e
.
co
m
Enha
nced
scen
e t
ex
t
rec
o
g
nition
us
ing
deep
lea
rn
ing
ba
sed
hy
brid atte
ntion r
ecog
nition net
wo
rk
Ra
t
na
m
a
la
S
.
P
a
t
il,
G
ee
t
a
H
a
nji
,
Ra
k
esh
H
ud
e
d
1
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s
a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
,
P
D
A
C
o
l
l
e
g
e
o
f
En
g
i
n
e
e
r
i
n
g
,
K
a
l
a
b
u
r
a
g
i
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
No
v
1
1
,
2
0
2
3
R
ev
is
ed
Ma
r
2
2
,
2
0
2
4
Acc
ep
ted
Ap
r
1
7
,
2
0
2
4
Th
e
tec
h
n
i
q
u
e
o
f
a
u
to
m
a
ti
c
a
ll
y
re
c
o
g
n
izin
g
a
n
d
tra
n
sfo
rm
in
g
te
x
t
t
h
a
t
is
p
re
se
n
t
i
n
p
ict
u
re
s
o
r
sc
e
n
e
s
i
n
t
o
m
a
c
h
in
e
-
re
a
d
a
b
le
tex
t
is
k
n
o
w
n
a
s
sc
e
n
e
tex
t
re
c
o
g
n
it
i
o
n
.
It
fa
c
il
it
a
tes
a
p
p
l
ica
ti
o
n
s
l
ik
e
c
o
n
ten
t
e
x
trac
ti
o
n
,
tr
a
n
sla
ti
o
n
,
a
n
d
te
x
t
a
n
a
l
y
sis
i
n
re
a
l
-
wo
rl
d
v
isu
a
l
d
a
ta
b
y
e
n
a
b
li
n
g
c
o
m
p
u
ters
t
o
c
o
m
p
re
h
e
n
d
a
n
d
e
x
trac
t
te
x
tu
a
l
in
fo
rm
a
ti
o
n
fro
m
ima
g
e
s,
v
id
e
o
s,
o
r
d
o
c
u
m
e
n
ts
.
S
c
e
n
e
tex
t
re
c
o
g
n
it
i
o
n
is
e
ss
e
n
ti
a
l
fo
r
m
a
n
y
a
p
p
li
c
a
ti
o
n
s,
su
c
h
a
s
lan
g
u
a
g
e
tran
sla
ti
o
n
a
n
d
c
o
n
ten
t
e
x
trac
ti
o
n
fr
o
m
p
h
o
to
g
ra
p
h
s
.
T
h
e
h
y
b
ri
d
a
tt
e
n
ti
o
n
re
c
o
g
n
i
ti
o
n
n
e
two
r
k
(HA
RN),
u
n
iq
u
e
tec
h
n
o
lo
g
y
p
re
se
n
ted
in
th
is
re
se
a
rc
h
,
is
in
ten
d
e
d
t
o
g
re
a
tl
y
imp
r
o
v
e
e
fficie
n
c
y
a
n
d
a
c
c
u
ra
c
y
o
f
tex
t
re
c
o
g
n
it
i
o
n
in
c
o
m
p
li
c
a
ted
sc
e
n
e
situ
a
ti
o
n
s
.
HA
RN
m
a
k
e
s
u
se
o
f
c
u
tt
i
n
g
-
e
d
g
e
e
lem
e
n
ts
in
c
l
u
d
in
g
a
li
g
n
m
e
n
t
-
fre
e
se
q
u
e
n
c
e
-
to
-
se
q
u
e
n
c
e
(AFS
)
m
o
d
u
le,
c
re
a
ti
v
e
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
s,
a
n
d
h
y
b
ri
d
a
rc
h
it
e
c
tu
re
th
a
t
b
len
d
s
a
tt
e
n
ti
o
n
m
o
d
e
ls
wi
th
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
s
(CNN
s)
.
T
h
a
n
k
s
to
it
s
n
o
v
e
l
a
tt
e
n
ti
o
n
p
r
o
c
e
ss
e
s,
HA
RN
is
c
a
p
a
b
le
o
f
c
o
m
p
re
h
e
n
d
in
g
wid
e
ra
n
g
e
o
f
sc
e
n
e
tex
t
c
o
m
p
o
n
e
n
ts
b
y
c
a
p
tu
rin
g
b
o
t
h
lo
c
a
l
a
n
d
g
lo
b
a
l
c
o
n
tex
t
in
fo
rm
a
ti
o
n
.
T
h
ro
u
g
h
fa
ste
r
n
e
t
wo
rk
c
o
n
v
e
rg
e
n
c
e
,
sh
o
rter
train
in
g
ti
m
e
s,
a
n
d
b
e
tt
e
r
u
ti
li
z
a
ti
o
n
o
f
c
o
m
p
u
t
in
g
re
so
u
rc
e
s,
th
e
su
g
g
e
ste
d
tec
h
n
i
q
u
e
ra
ise
s
b
a
r
fo
r
sta
te
-
of
-
t
h
e
-
a
rt
.
HA
RN
’
s
v
e
rsa
ti
li
ty
m
a
k
e
s
it
a
g
o
o
d
c
h
o
ice
fo
r
ra
n
g
e
o
f
sc
e
n
e
tex
t
re
c
o
g
n
it
i
o
n
a
p
p
li
c
a
ti
o
n
s,
in
c
lu
d
i
n
g
m
u
l
ti
li
n
g
u
a
l
tex
t
a
n
a
ly
sis
a
n
d
d
a
ta
e
x
trac
ti
o
n
.
E
x
ten
si
v
e
tes
ts
a
re
c
o
n
d
u
c
ted
t
o
a
ss
e
ss
th
e
e
ff
e
c
ti
v
e
n
e
ss
o
f
HA
RN
a
p
p
ro
a
c
h
a
n
d
d
e
m
o
n
stra
te
it
i
s
a
b
il
it
y
t
o
g
re
a
tl
y
i
n
fl
u
e
n
c
e
re
a
l
-
wo
rld
a
p
p
li
c
a
ti
o
n
s wh
e
re
a
c
c
u
ra
t
e
a
n
d
e
fficie
n
t
te
x
t
re
c
o
g
n
it
i
o
n
is
e
ss
e
n
ti
a
l
.
K
ey
w
o
r
d
s
:
Alig
n
m
en
t
-
f
r
ee
s
eq
u
en
ce
-
to
-
s
eq
u
en
ce
Atten
tio
n
m
ec
h
an
is
m
s
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
Hy
b
r
id
atten
tio
n
r
ec
o
g
n
itio
n
n
etwo
r
k
Scen
e
tex
t 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
:
R
atn
am
ala
S Patil
Dep
ar
tm
en
t o
f
E
lectr
o
n
ics
an
d
C
o
m
m
u
n
icatio
n
,
PDA
C
o
lleg
e
o
f
E
n
g
in
ee
r
in
g
Kala
b
u
r
ag
i,
I
n
d
ia
E
m
ail: r
atn
am
ala_
1
2
@
r
ed
iifm
ail
.
co
m
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
lan
g
u
ag
e
s
ee
n
in
im
ag
es
o
f
th
e
n
atu
r
al
wo
r
l
d
co
n
tain
s
im
p
o
r
tan
t
a
n
d
p
r
ac
tical
in
f
o
r
m
atio
n
.
T
h
e
ap
p
licatio
n
s
in
d
icate
d
ab
o
v
e
ar
e
r
eg
u
lar
ly
u
s
ed
in
a
v
a
r
iety
o
f
r
ea
l
-
wo
r
ld
s
itu
atio
n
s
.
C
o
m
p
u
ter
v
is
io
n
ap
p
licatio
n
s
m
ay
b
e
f
o
u
n
d
i
n
m
an
y
d
if
f
er
en
t
f
ield
s
,
in
clu
d
in
g
g
eo
l
o
ca
tio
n
,
c
o
n
ten
t
-
b
ased
p
ictu
r
e
o
r
v
i
d
eo
r
etr
iev
al,
r
o
b
o
tic
n
av
ig
atio
n
,
au
to
m
atic
licen
s
e
p
late
id
e
n
tif
icatio
n
,
h
elp
in
g
p
eo
p
le
w
h
o
a
r
e
b
lin
d
o
r
v
is
u
ally
im
p
air
ed
,
an
d
im
ag
e
in
ter
p
r
et
atio
n
[
1
]
.
Desp
ite
s
ig
n
if
ican
t
p
r
o
g
r
ess
m
ad
e
in
th
e
f
ield
o
f
n
atu
r
al
s
ce
n
e
tex
t
r
ec
o
g
n
itio
n
,
a
n
u
m
b
er
o
f
is
s
u
es
s
till
n
ee
d
to
b
e
ad
d
r
ess
ed
an
d
r
eso
lv
e
d
.
T
h
er
e
a
r
e
m
an
y
d
if
f
icu
lties
an
d
co
m
p
lex
ity
in
th
e
r
ea
lm
o
f
im
ag
e
p
r
o
ce
s
s
in
g
.
Nu
m
e
r
o
u
s
ele
m
en
ts
,
in
clu
d
in
g
as
b
ac
k
g
r
o
u
n
d
n
o
is
e,
b
lu
r
r
in
ess
,
b
lo
ck
ag
e,
l
o
w
r
eso
lu
tio
n
,
in
tr
icate
b
ac
k
d
r
o
p
s
,
an
d
d
if
f
er
en
ce
s
in
tex
t
p
r
o
p
e
r
ties
lik
e
s
ize,
co
lo
r
,
an
d
o
r
ien
tatio
n
,
p
r
o
v
id
e
o
b
s
tacle
s
in
th
is
s
p
ec
if
ic
en
v
ir
o
n
m
en
t
[
2
]
.
Dee
p
lear
n
in
g
m
eth
o
d
s
co
m
e
with
a
h
ig
h
r
eso
u
r
ce
co
s
t,
in
clu
d
in
g
m
em
o
r
y
,
c
o
m
p
u
tin
g
p
o
wer
,
an
d
en
er
g
y
.
Dee
p
lear
n
i
n
g
tech
n
iq
u
es
o
r
th
e
f
o
u
n
d
atio
n
al
id
ea
s
o
f
co
m
p
u
ter
v
is
io
n
i
n
f
o
r
m
th
e
al
g
o
r
ith
m
s
u
s
ed
f
o
r
s
ce
n
e
tex
t
r
ec
o
g
n
itio
n
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l
.
13
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
4
9
2
7
-
4
9
3
8
4928
id
en
tific
atio
n
.
B
ec
au
s
e
o
f
th
e
ir
r
eso
u
r
ce
n
ee
d
s
,
th
ese
ap
p
r
o
ac
h
es
ca
n
b
e
ch
allen
g
in
g
to
ap
p
ly
in
r
ea
l
-
tim
e
em
b
ed
d
e
d
s
y
s
tem
s
,
p
a
r
ticu
lar
ly
th
o
s
e
r
u
n
n
in
g
o
n
h
ar
d
wa
r
e
th
at
o
n
ly
s
u
p
p
o
r
ts
in
teg
e
r
o
p
e
r
atio
n
s
.
Ma
n
y
m
eth
o
d
s
h
av
e
b
ee
n
d
ev
elo
p
ed
to
r
ec
o
g
n
ize
tex
t in
n
atu
r
alis
tic
s
ettin
g
s
[
3
]
,
[
4]
.
T
h
e
u
s
e
o
f
d
ee
p
lea
r
n
in
g
al
g
o
r
ith
m
s
h
as
led
to
s
ig
n
if
ican
t
ad
v
an
ce
m
e
n
ts
in
th
e
f
ield
o
f
s
ce
n
e
tex
t
r
ec
o
g
n
itio
n
in
r
ec
e
n
t
y
ea
r
s
.
Sig
n
if
ican
t
ad
v
an
ce
m
e
n
ts
h
av
e
b
ee
n
ac
h
ie
v
ed
in
th
e
f
ield
o
f
s
eq
u
en
ce
lear
n
i
n
g
-
b
ased
m
eth
o
d
s
.
I
n
c
o
n
tem
p
o
r
ar
y
ap
p
r
o
ac
h
es
[
5
]
,
[
6
]
,
th
e
e
m
p
lo
y
m
e
n
t
o
f
a
n
en
c
o
d
er
-
d
ec
o
d
er
ar
c
h
itectu
r
e
is
a
co
m
m
o
n
l
y
u
s
ed
s
tr
ateg
y
f
o
r
wo
r
d
d
etec
tio
n
in
p
ictu
r
es
o
f
n
atu
r
al
s
itu
atio
n
s
.
On
e
p
o
p
u
lar
m
eth
o
d
f
o
r
en
co
d
in
g
th
e
in
p
u
t
p
ictu
r
e
is
to
em
p
lo
y
co
n
v
o
lu
tio
n
al
n
eu
r
a
l
n
etwo
r
k
s
(
C
NNs
)
an
d
r
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
s
(
R
NNs
)
,
r
esp
ec
tiv
ely
.
R
NN
s
with
atten
tio
n
p
r
o
ce
s
s
es
[
7
]
o
r
co
n
n
ec
tio
n
is
t
tem
p
o
r
al
class
if
icatio
n
(
C
T
C
)
[
8
]
ar
e
th
e
m
eth
o
d
s
u
s
ed
t
o
d
ec
o
d
e
th
e
en
co
d
ed
i
n
p
u
t
in
to
th
e
tar
g
et
te
x
t
.
T
h
e
p
r
esen
t
p
ar
a
d
ig
m
f
o
r
s
ce
n
e
tex
t
r
ec
o
g
n
itio
n
m
o
s
tly
r
elies o
n
t
h
e
atten
tio
n
-
b
ased
e
n
co
d
er
-
d
ec
o
d
er
p
a
r
ad
ig
m
.
A
ty
p
ical
s
y
s
tem
d
esig
n
ed
to
d
ea
l
with
tex
t
s
ee
n
in
r
ea
l
-
wo
r
ld
s
ettin
g
s
m
u
s
t
h
av
e
tex
t
d
etec
tio
n
an
d
id
en
tific
atio
n
as
b
asic
f
ea
tu
r
es
.
T
ex
t
d
etec
tio
n
’
s
p
r
im
ar
y
o
b
j
ec
tiv
e
is
to
lo
ca
te
an
d
is
o
late
th
e
ex
ac
t
r
eg
io
n
o
f
a
p
ictu
r
e
co
n
tain
i
n
g
tex
t,
with
a
h
ig
h
lev
el
o
f
ac
cu
r
ac
y
b
ein
g
th
e
m
ajo
r
f
o
cu
s
.
T
ex
t
m
a
y
b
e
r
etr
iev
ed
f
r
o
m
th
e
o
r
ig
in
al
im
ag
e
af
ter
th
e
r
eg
io
n
h
as
b
ee
n
id
en
tifie
d
.
T
ex
t
r
e
co
g
n
itio
n
’
s
p
r
im
ar
y
o
b
jectiv
e
[
8
]
is
to
ex
tr
ac
t
an
d
ac
cu
r
ately
r
ec
o
g
n
ize
tex
t
f
r
o
m
a
well
-
d
ef
in
ed
p
ictu
r
e
.
Fig
u
r
e
1
d
ep
icts
th
e
tex
t
r
ec
o
g
n
itio
n
p
r
o
ce
d
u
r
e
.
Fig
u
r
e
1
.
T
e
x
t r
ec
o
g
n
itio
n
p
r
o
ce
s
s
Scen
e
tex
t
d
etec
tio
n
is
th
e
f
ir
s
t
an
d
m
o
s
t
im
p
o
r
tan
t
s
tep
in
th
e
s
ce
n
e
tex
t
r
ec
o
g
n
itio
n
p
r
o
ce
s
s
.
C
o
n
v
en
tio
n
al
a
p
p
r
o
ac
h
es
s
u
c
h
as
lin
k
e
d
co
m
p
o
n
en
t
-
b
ased
s
tr
ateg
ies
an
d
s
lid
in
g
win
d
o
w
-
b
ased
tech
n
iq
u
es
h
av
e
b
ee
n
u
s
ed
f
o
r
s
ce
n
e
tex
t
id
en
tific
atio
n
.
T
o
id
e
n
tify
p
o
s
s
ib
le
lo
ca
tio
n
s
f
o
r
s
ce
n
e
tex
t
r
ec
o
g
n
itio
n
,
s
lid
in
g
win
d
o
w
tech
n
iq
u
es
ar
e
u
tili
ze
d
.
T
h
e
m
eth
o
d
m
ak
es
u
s
e
o
f
tech
n
iq
u
es
th
at
en
tail
u
s
in
g
a
m
o
v
in
g
win
d
o
w
to
s
ca
n
th
e
in
p
u
t
im
a
g
e
o
f
a
n
a
tu
r
al
s
ce
n
e
.
C
o
n
n
ec
ted
c
o
m
p
o
n
en
t
-
b
ased
tec
h
n
iq
u
es
ar
e
a
n
ex
ce
llen
t
way
to
ex
tr
ac
t
ch
ar
ac
ter
s
f
r
o
m
s
ce
n
e
tex
t
tak
en
in
ac
tu
al
en
v
ir
o
n
m
en
ts
.
T
o
co
m
p
lete
th
is
g
o
al,
t
h
ese
alg
o
r
ith
m
s
u
s
e
ex
tr
em
e
ar
ea
ex
tr
ac
tio
n
a
n
d
c
o
lo
r
clu
s
ter
in
g
tech
n
iq
u
es
.
I
m
p
r
o
v
in
g
s
ce
n
e
tex
t
i
d
en
tific
atio
n
p
er
f
o
r
m
an
ce
is
th
e
aim
o
f
th
e
h
y
b
r
i
d
tech
n
i
q
u
e
.
T
h
e
ad
v
an
tag
es
o
f
lin
k
ed
co
m
p
o
n
en
t
-
b
ased
an
d
s
lid
in
g
win
d
o
w
-
b
ased
tech
n
iq
u
es
ar
e
co
m
b
i
n
ed
i
n
th
e
s
u
g
g
ested
m
eth
o
d
o
l
o
g
y
.
I
n
t
h
eir
in
v
esti
g
atio
n
,
th
ey
u
s
ed
th
e
co
n
tr
ast
-
en
h
an
ce
m
en
t
m
ax
im
ally
s
tab
le
ex
tr
em
al
r
eg
io
n
s
(
C
E
-
MSE
R
s
)
d
etec
to
r
.
T
h
e
d
e
tecto
r
u
s
es
a
wid
e
v
ar
iety
o
f
h
y
b
r
id
ap
p
r
o
ac
h
es
to
o
p
tim
ize
an
d
im
p
r
o
v
e
th
e
co
n
tr
ast
o
f
th
e
im
ag
e
.
T
h
e
m
ax
im
ally
s
tab
le
ex
tr
em
al
ar
ea
s
alg
o
r
ith
m
is
a
s
p
ec
ial
tech
n
iq
u
e
d
esig
n
ed
to
en
h
an
ce
th
e
d
is
tin
ctio
n
b
etw
ee
n
ar
ea
s
with
an
d
with
o
u
t te
x
t [
9
]
.
I
t ser
v
es a
s
th
e
b
asis
f
o
r
th
is
n
o
tio
n
.
Scen
e
tex
t r
ec
o
g
n
itio
n
is
a
cr
u
cial
s
tag
e
in
th
e
p
r
o
ce
s
s
th
at
h
as to
b
e
im
p
lem
en
ted
in
o
r
d
er
to
co
n
v
er
t
s
ce
n
e
tex
t
r
e
g
io
n
s
i
n
to
s
tr
in
g
tex
ts
.
Var
i
o
u
s
m
eth
o
d
s
h
a
v
e
b
ee
n
d
ev
elo
p
e
d
a
n
d
r
ef
in
ed
to
f
ac
ilit
ate
th
e
r
ec
o
g
n
itio
n
an
d
ca
te
g
o
r
izatio
n
o
f
tex
t
in
v
ar
io
u
s
co
n
tex
t
s
.
T
h
e
w
o
r
d
class
if
icatio
n
-
b
ased
m
eth
o
d
,
t
h
e
s
eq
u
en
ce
-
b
ased
m
eth
o
d
,
an
d
t
h
e
ch
ar
ac
ter
class
if
icatio
n
-
b
ased
ap
p
r
o
ac
h
ar
e
s
o
m
e
o
f
t
h
e
t
ec
h
n
iq
u
es
av
ailab
le
f
o
r
tex
t
r
ec
o
g
n
itio
n
in
s
ce
n
e
s
.
C
h
ar
ac
ter
ca
n
d
id
ates
ar
e
id
en
tifie
d
an
d
r
ec
o
g
n
ized
in
d
iv
id
u
ally
t
h
r
o
u
g
h
ch
ar
ac
ter
class
if
icatio
n
-
b
ased
p
r
o
ce
s
s
es,
ak
in
to
th
o
s
e
u
tili
ze
d
b
y
[
1
0
]
.
W
o
r
d
class
if
icatio
n
tech
n
iq
u
es
r
ely
o
n
wo
r
d
tr
ain
i
n
g
a
n
d
a
p
r
ee
x
is
tin
g
v
o
ca
b
u
lar
y
to
au
g
m
en
t
th
eir
e
f
f
icac
y
.
Fo
u
r
p
r
im
a
r
y
class
es
m
ay
b
e
d
is
tin
g
u
is
h
ed
am
o
n
g
th
e
ty
p
ic
al
ir
r
eg
u
lar
s
h
o
r
t
tan
d
em
r
ep
e
at
(
STR)
tech
n
iq
u
es:
th
e
atten
tio
n
-
b
ased
m
eth
o
d
,
th
e
m
u
lti
-
d
ir
ec
tio
n
en
co
d
in
g
-
b
ased
ap
p
r
o
ac
h
,
th
e
tex
t
s
h
ap
e
r
ec
tific
atio
n
-
b
ased
t
ec
h
n
iq
u
e,
an
d
th
e
C
T
C
-
b
ased
ap
p
r
o
ac
h
.
Sh
ap
e
r
ec
tific
atio
n
is
th
e
ac
t
o
f
elim
in
atin
g
d
is
to
r
tio
n
f
r
o
m
tex
t
,
wh
ich
p
r
o
d
u
ce
s
a
m
o
r
e
co
n
s
is
ten
t
v
is
u
al
r
ep
r
esen
tatio
n
o
f
wo
r
d
s
an
d
en
h
an
ce
s
ir
r
eg
u
lar
tex
t
r
ec
o
g
n
it
io
n
.
B
ef
o
r
e
u
s
in
g
tr
ad
itio
n
al
tex
t
r
ec
o
g
n
izer
s
,
ir
r
eg
u
lar
tex
t
is
f
ir
s
t
tr
an
s
f
o
r
m
ed
in
to
r
eg
u
lar
tex
t
u
s
in
g
s
h
ap
e
r
ec
tific
atio
n
tech
n
iq
u
es
[
1
1
]
.
T
h
e
f
ir
s
t
n
et
wo
r
k
t
o
b
e
p
r
esen
ted
f
o
r
tex
t
n
o
r
m
aliza
tio
n
was
th
e
s
p
atial
tr
an
s
f
o
r
m
er
n
etwo
r
k
(
STN)
.
T
h
e
s
y
s
tem
’
s
m
ain
g
o
al
was
to
co
r
r
ec
t
in
d
i
v
id
u
al
letter
s
as
well
as
wo
r
d
r
e
p
r
esen
tatio
n
s
in
th
eir
wh
o
le
.
T
o
im
p
r
o
v
e
it
i
s
p
er
f
o
r
m
a
n
ce
,
W
an
g
et
a
l
.
[
1
2
]
in
clu
d
ed
th
e
th
in
-
p
late
s
p
lin
e
(
T
PS
)
tr
an
s
f
o
r
m
atio
n
m
eth
o
d
i
n
to
its
tex
t
n
o
r
m
ali
za
tio
n
m
o
d
u
le
.
I
m
p
r
o
v
in
g
th
e
m
o
d
u
le
’
s
ca
p
ac
ity
to
m
an
ag
e
co
m
p
lex
tex
t
d
is
to
r
tio
n
s
was
th
e
aim
o
f
th
is
in
clu
s
io
n
.
T
h
e
g
r
o
win
g
u
b
iq
u
ity
o
f
s
o
p
h
is
ticated
r
ec
tific
atio
n
m
o
d
u
les
m
ay
b
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
n
h
a
n
ce
d
s
ce
n
e
text
r
ec
o
g
n
iti
o
n
u
s
in
g
d
ee
p
lea
r
n
in
g
b
a
s
ed
h
yb
r
id
a
tten
tio
n
…
(
R
a
tn
a
ma
l
a
S
.
P
a
til
)
4929
ascr
ib
ed
to
th
eir
p
iv
o
tal
f
u
n
ct
io
n
in
ef
f
ec
tiv
ely
h
an
d
lin
g
an
ex
ten
s
iv
e
ar
r
ay
o
f
d
is
to
r
tio
n
s
.
Nu
m
er
o
u
s
f
ac
to
r
s
af
f
ec
t
th
e
r
ec
o
g
n
itio
n
alg
o
r
ith
m
s
’
p
er
f
o
r
m
an
ce
an
d
m
em
o
r
y
u
s
e
.
T
h
e
u
s
e
o
f
C
T
C
h
as
s
u
b
s
tan
tially
ass
is
te
d
th
e
d
ev
elo
p
m
en
t
o
f
n
u
m
e
r
o
u
s
ap
p
licatio
n
s
,
s
u
ch
as
o
n
lin
e
h
an
d
wr
itten
ch
a
r
ac
ter
id
en
tif
icatio
n
an
d
s
p
ee
ch
r
ec
o
g
n
itio
n
.
I
n
t
h
e
STR
d
o
m
ain
,
th
e
C
T
C
tech
n
iq
u
e
is
a
well
-
k
n
o
wn
an
d
o
f
te
n
u
s
ed
p
r
ed
ictio
n
to
o
l
.
B
y
ap
p
ly
in
g
a
p
r
e
d
ictio
n
m
eth
o
d
b
ased
o
n
C
T
C
d
u
r
in
g
th
e
m
o
d
el
’
s
tr
ain
in
g
p
h
ase,
th
e
ea
r
ly
a
lg
o
r
ith
m
s
th
at
u
s
ed
C
T
C
in
STR
s
h
o
wed
b
etter
r
esu
lts
.
T
h
is
m
eth
o
d
m
ad
e
u
s
e
o
f
d
e
v
elo
p
m
e
n
ts
in
v
o
ice
r
ec
o
g
n
itio
n
.
Alth
o
u
g
h
STR
h
as
p
er
f
o
r
m
ed
s
atis
f
ac
to
r
ily
,
C
T
C
is
n
o
t
with
o
u
t
its
lim
its
.
W
o
r
k
in
g
with
n
ar
r
o
w
d
is
tr
ib
u
tio
n
s
m
ig
h
t
p
r
o
v
id
e
d
i
f
f
icu
lties
d
u
e
to
th
e
in
tr
icate
u
n
d
er
p
in
n
i
n
g
m
ec
h
a
n
is
m
o
f
th
e
C
T
C
m
eth
o
d
.
Pre
d
ictio
n
is
s
u
es
w
ith
ir
r
eg
u
lar
two
-
d
im
en
s
io
n
al
d
ata,
lik
e
STR,
m
ig
h
t n
o
t
b
e
a
g
o
o
d
f
it f
o
r
it
.
I
n
th
e
r
ea
lm
o
f
s
ce
n
e
te
x
t
r
ec
o
g
n
itio
n
,
th
e
n
ee
d
f
o
r
i
n
n
o
v
at
iv
e
s
o
lu
tio
n
s
to
a
d
d
r
ess
th
e
c
o
m
p
lex
ities
o
f
r
ea
l
-
wo
r
ld
s
ce
n
a
r
io
s
is
ev
er
-
in
cr
ea
s
in
g
.
T
h
e
m
o
tiv
atio
n
to
ad
v
an
ce
th
e
s
tate
-
of
-
t
h
e
-
ar
t in
th
is
d
o
m
ain
s
tem
s
f
r
o
m
th
e
r
ec
o
g
n
itio
n
o
f
th
e
c
h
allen
g
es
f
ac
ed
in
ac
c
u
r
ately
u
n
d
er
s
tan
d
in
g
tex
t
with
in
im
ag
es
.
T
h
ese
ar
e
d
r
iv
e
n
b
y
th
e
o
p
p
o
r
tu
n
ity
to
d
ev
elo
p
an
d
r
ef
in
e
tech
n
iq
u
es
th
at
ca
n
r
ev
o
lu
tio
n
ize
s
ce
n
e
tex
t
r
ec
o
g
n
itio
n
,
en
a
b
lin
g
ap
p
licatio
n
s
r
an
g
in
g
f
r
o
m
im
a
g
e
-
b
ased
tr
a
n
s
latio
n
to
c
o
n
ten
t
ex
tr
ac
tio
n
f
r
o
m
v
is
u
al
d
ata
.
W
ith
a
co
m
m
itm
en
t
to
im
p
r
o
v
in
g
ac
cu
r
ac
y
an
d
ef
f
icien
cy
,
th
e
aim
is
to
m
ak
e
a
m
ea
n
in
g
f
u
l
im
p
ac
t
o
n
th
e
f
ield
,
p
r
o
v
id
i
n
g
to
o
ls
an
d
in
s
ig
h
ts
th
at
ca
n
b
en
ef
it v
ar
io
u
s
in
d
u
s
tr
ies an
d
a
p
p
licatio
n
s
.
−
A
n
o
v
el
h
y
b
r
id
atten
tio
n
r
ec
o
g
n
itio
n
n
etwo
r
k
(
HARN)
is
d
ev
elo
p
e
d
wh
ich
im
p
r
o
v
es
t
ex
t
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
.
−
T
h
e
h
y
b
r
id
ap
p
r
o
ac
h
co
m
b
in
es
C
NN
s
wi
th
atten
tio
n
m
o
d
els,
p
r
o
v
id
in
g
a
ef
f
icien
t
s
o
lu
tio
n
f
o
r
lo
ca
l
co
n
tex
t m
o
d
elin
g
a
n
d
lo
n
g
-
ter
m
r
elatio
n
s
h
ip
m
o
d
elin
g
i
n
tex
t r
ec
o
g
n
itio
n
task
s
.
−
T
h
e
A
FS
m
o
d
u
l
e
is
d
e
v
el
o
p
e
d
t
h
a
t
i
n
t
r
o
d
u
c
e
s
a
d
v
a
n
c
e
d
a
l
i
g
n
m
e
n
t
t
ec
h
n
i
q
u
e
s
,
e
n
h
a
n
ci
n
g
r
e
c
o
g
n
i
t
i
o
n
a
c
c
u
r
a
c
y
,
e
s
p
e
ci
a
l
l
y
i
n
c
o
m
p
l
ex
a
n
d
l
e
n
g
t
h
y
s
e
q
u
e
n
c
e
s
,
a
d
d
r
es
s
i
n
g
a
c
r
i
t
ic
a
l
n
e
e
d
i
n
s
c
e
n
e
tex
t
r
e
c
o
g
n
i
t
i
o
n
.
−
T
h
e
in
teg
r
atio
n
o
f
atten
tio
n
m
ec
h
an
is
m
s
with
in
th
e
m
o
d
el
ar
ch
itectu
r
e
e
n
ab
les
th
e
ca
p
t
u
r
e
o
f
b
o
th
lo
ca
l
an
d
g
lo
b
al
co
n
te
x
t in
f
o
r
m
atio
n
,
im
p
r
o
v
in
g
th
e
ac
cu
r
ac
y
o
f
s
ce
n
e
tex
t r
ec
o
g
n
itio
n
.
T
h
e
r
esear
ch
is
o
r
g
a
n
is
ed
in
to
f
o
u
r
s
ec
tio
n
s
:
th
e
f
ir
s
t
s
ec
tio
n
in
tr
o
d
u
ce
s
;
f
ir
s
t
s
ec
tio
n
o
f
th
e
r
esear
ch
s
tar
ts
with
b
ac
k
g
r
o
u
n
d
o
f
s
c
en
e
tex
t
r
ec
o
g
n
itio
n
alo
n
g
w
ith
p
r
o
b
lem
ass
o
ciate
d
with
th
e
s
am
e
.
Seco
n
d
s
ec
tio
n
o
f
th
e
r
esear
ch
f
o
cu
s
es o
n
v
ar
io
u
s
s
tate
-
of
-
ar
t te
ch
n
i
q
u
es f
o
r
s
ce
n
e
tex
t r
ec
o
g
n
itio
n
.
I
n
th
ir
d
s
ec
tio
n
o
f
th
e
r
esear
ch
,
HARN
i
s
d
ev
elo
p
ed
an
d
its
ar
ch
itectu
r
e
is
d
esig
n
ed
.
At
last
,
HA
R
N
is
ev
alu
ated
in
f
o
u
r
th
s
ec
tio
n
co
n
s
id
er
in
g
d
if
f
er
e
n
t d
ataset
to
p
r
o
v
e
th
e
p
r
o
p
o
s
ed
m
o
d
el
ef
f
icien
c
y
.
2.
RE
L
AT
E
D
WO
RK
Scen
e
tex
t
r
ec
o
g
n
itio
n
is
wid
ely
r
ec
o
g
n
ized
as
a
ch
allen
g
in
g
an
d
l
o
n
g
-
s
tan
d
in
g
r
esea
r
ch
ar
ea
in
co
m
p
u
ter
v
is
io
n
[
1
3
]
.
Nu
m
er
o
u
s
ac
ad
em
ics
h
av
e
d
ed
icate
d
th
eir
ef
f
o
r
ts
to
en
h
a
n
cin
g
th
e
ca
p
ab
ilit
y
o
f
s
ce
n
e
tex
t
r
ec
o
g
n
itio
n
in
r
ea
l
-
wo
r
ld
s
ce
n
ar
io
s
.
Mo
r
eo
v
er
,
s
ce
n
e
te
x
t
r
ec
o
g
n
itio
n
s
er
v
es
a
m
u
ltit
u
d
e
o
f
a
p
p
licatio
n
s
,
in
clu
d
in
g
b
u
t
n
o
t
lim
ited
to
p
i
ctu
r
e
r
etr
iev
al,
v
is
u
al
n
av
ig
ati
o
n
s
y
s
tem
s
,
an
d
tex
t
r
ea
d
in
g
f
o
r
in
d
iv
id
u
als
wh
o
ar
e
b
lin
d
o
r
v
is
u
ally
im
p
air
e
d
.
T
h
e
task
o
f
s
ce
n
e
tex
t
id
e
n
tif
icatio
n
is
h
ig
h
ly
c
h
allen
g
in
g
d
u
e
to
th
e
p
r
esen
ce
o
f
n
u
m
er
o
u
s
s
ce
n
e
tex
t
o
r
ien
t
atio
n
s
in
o
u
r
s
u
r
r
o
u
n
d
in
g
s
.
T
h
e
d
if
f
er
en
t
s
ce
n
e
o
r
ien
tatio
n
s
in
clu
d
e
h
o
r
izo
n
tal,
r
an
d
o
m
,
cu
r
v
e
d
,
an
d
v
er
tical
o
r
ien
tatio
n
s
[
1
4
]
.
M
o
r
eo
v
er
,
p
h
o
to
s
th
at
in
clu
d
e
tex
t c
an
b
e
c
ateg
o
r
ized
in
to
two
d
is
tin
ct
g
r
o
u
p
s
:
im
ag
es
o
f
n
at
u
r
al
lan
d
s
ca
p
es
an
d
s
ca
n
n
ed
d
o
cu
m
en
ts
.
T
h
e
p
r
esen
ce
o
f
tex
t
in
p
h
o
to
g
r
ap
h
s
o
f
n
atu
r
al
s
ce
n
es o
f
ten
ex
h
i
b
its
v
ar
iatio
n
s
in
ap
p
ea
r
a
n
ce
a
n
d
is
o
f
ten
f
o
u
n
d
in
i
n
tr
icate
s
u
r
r
o
u
n
d
in
g
s
.
I
n
co
n
t
r
ast,
tex
t
o
b
s
er
v
ed
in
s
ca
n
n
e
d
p
ap
er
s
ty
p
ically
p
o
s
s
ess
e
s
co
n
s
is
ten
t
an
d
u
n
if
o
r
m
b
ac
k
g
r
o
u
n
d
s
.
T
h
e
p
r
o
ce
s
s
o
f
r
ec
o
g
n
izin
g
tex
t
ca
n
b
e
ac
co
m
p
lis
h
ed
u
s
in
g
a
co
n
v
en
tio
n
a
l
tech
n
iq
u
e
k
n
o
wn
as
o
p
tical
ch
ar
ac
ter
r
ec
o
g
n
itio
n
(
OC
R
)
.
OC
R
ca
n
ac
h
iev
e
a
h
ig
h
lev
el
o
f
ac
cu
r
a
cy
wh
e
n
p
r
o
ce
s
s
in
g
tex
t
th
at
is
d
is
p
lay
e
d
in
a
leg
i
b
le
f
o
n
t
with
a
clea
r
an
d
u
n
clu
tter
ed
b
ac
k
d
r
o
p
.
OC
R
is
n
o
t su
itab
le
f
o
r
tex
t r
ec
o
g
n
itio
n
in
n
atu
r
al
e
n
v
ir
o
n
m
en
ts
d
u
e
to
v
ar
io
u
s
is
s
u
es
.
T
h
e
ch
allen
g
es
en
co
m
p
ass
co
n
g
ested
an
d
in
tr
icate
en
v
ir
o
n
m
e
n
ts
,
lo
w
-
q
u
ality
an
d
b
lu
r
r
ed
im
ag
es,
d
iv
er
s
e
tex
t
o
r
ien
tatio
n
s
,
an
d
a
m
u
ltit
u
d
e
o
f
f
o
n
ts
.
T
h
e
p
er
f
o
r
m
a
n
ce
o
f
tex
t
r
ec
o
g
n
itio
n
alg
o
r
ith
m
s
ca
n
b
e
h
i
n
d
er
e
d
b
y
ch
allen
g
es
in
d
if
f
er
en
tiatin
g
b
etwe
en
tex
t
an
d
n
o
n
-
tex
t
elem
en
ts
,
p
ar
ticu
lar
ly
wh
e
n
ce
r
tain
item
s
clo
s
ely
r
esem
b
le
tex
t to
h
u
m
an
r
ea
d
er
s
.
T
h
e
i
n
t
e
g
r
ati
o
n
o
f
s
ce
n
e
te
x
t
d
et
ec
ti
o
n
a
n
d
s
ce
n
e
t
ex
t
r
e
co
g
n
it
io
n
h
as
g
a
r
n
e
r
e
d
c
o
n
s
i
d
e
r
a
b
le
i
n
t
er
est
in
r
e
ce
n
t
y
ea
r
s
,
wi
th
s
e
v
e
r
a
l
p
r
o
p
o
s
e
d
m
o
d
els
b
e
i
n
g
t
h
e
f
o
c
u
s
o
f
a
tte
n
t
io
n
.
V
ar
i
o
u
s
te
c
h
n
i
q
u
es
h
a
v
e
b
ee
n
d
e
v
e
lo
p
ed
f
o
r
t
h
e
d
et
ec
t
io
n
o
f
te
x
t
o
r
ie
n
t
ati
o
n
i
n
s
ce
n
es,
in
cl
u
d
i
n
g
c
u
r
v
ed
,
ar
b
i
tr
ar
y
,
a
n
d
h
o
r
iz
o
n
ta
ll
y
o
r
ie
n
te
d
tex
t
.
T
h
e
c
u
r
r
e
n
t
s
ta
te
-
of
-
t
h
e
-
a
r
t
s
o
l
u
ti
o
n
s
p
r
i
m
a
r
il
y
p
r
i
o
r
i
tiz
e
s
ce
n
e
te
x
ts
t
h
a
t
a
r
e
h
o
r
i
z
o
n
t
all
y
o
r
i
en
te
d
d
u
e
t
o
th
e
ir
s
u
p
er
io
r
r
ea
d
a
b
i
lit
y
.
T
h
e
a
p
p
r
o
a
ch
p
r
o
p
o
s
e
d
b
y
[
1
5
]
i
s
ce
n
t
er
e
d
ar
o
u
n
d
t
h
e
r
ec
o
g
n
i
tio
n
o
f
h
o
r
iz
o
n
ta
ll
y
o
r
ie
n
t
ed
s
ce
n
e
te
x
t
.
T
h
e
t
as
k
is
ac
h
i
ev
e
d
t
h
r
o
u
g
h
t
h
e
u
t
iliz
ati
o
n
o
f
a
p
a
tc
h
-
b
ase
d
c
l
ass
if
ica
ti
o
n
s
y
s
te
m
,
ef
f
e
cti
v
e
ly
p
r
es
er
v
i
n
g
t
h
e
d
i
s
tin
cti
v
e
a
r
ea
s
o
f
a
n
i
m
a
g
e
.
A
m
e
th
o
d
o
l
o
g
y
w
as
p
r
ese
n
ted
f
o
r
i
n
f
o
r
m
ati
o
n
r
et
r
i
ev
al
t
h
at
u
ti
liz
es t
h
e
h
i
d
d
e
n
m
a
r
k
o
v
m
o
d
e
l
(
HM
M)
t
o
l
o
ca
t
e
h
o
r
iz
o
n
tal
ly
o
r
ie
n
t
ed
d
at
e
in
f
o
r
m
at
io
n
.
Var
io
u
s
m
eth
o
d
o
lo
g
ies
[
1
6
]
h
av
e
b
ee
n
d
ev
is
ed
i
n
r
ec
e
n
t
s
ch
o
lar
ly
in
v
esti
g
atio
n
s
to
i
d
en
tify
tex
t
with
in
o
r
ien
ted
s
ce
n
es
.
T
h
e
t
ec
h
n
iq
u
es
ca
n
b
e
ca
teg
o
r
ized
in
to
two
m
ain
g
r
o
u
p
s
:
2
D
b
ased
tech
n
iq
u
es
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l
.
13
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
4
9
2
7
-
4
9
3
8
4930
r
ec
tific
atio
n
-
b
ased
tech
n
iq
u
es
.
R
ec
tific
atio
n
n
etwo
r
k
s
ar
e
u
tili
ze
d
b
y
r
ec
tific
atio
n
-
b
ase
d
m
eth
o
d
o
l
o
g
ies
to
tr
an
s
f
o
r
m
th
e
o
r
ien
te
d
s
ce
n
e
tex
t
in
to
a
s
tan
d
a
r
d
ized
f
o
r
m
at
.
T
h
e
co
r
r
ec
ted
tex
ts
ar
e
th
en
s
en
t
in
t
o
r
ec
o
g
n
izer
s
.
T
h
e
ch
ar
ac
ter
-
b
a
s
ed
ap
p
r
o
ac
h
es
em
p
lo
y
a
two
-
s
tep
p
r
o
ce
d
u
r
e
in
wh
ic
h
p
o
te
n
tial
ch
ar
ac
ter
s
ar
e
in
itially
lo
ca
lized
an
d
s
u
b
s
eq
u
en
tly
r
ec
o
g
n
ize
d
.
Sev
e
r
al
illu
s
tr
ativ
e
tech
n
iq
u
es
h
av
e
b
ee
n
p
r
o
p
o
s
ed
f
o
r
th
e
ca
n
d
id
ate
c
h
ar
ac
ter
lo
ca
lizatio
n
p
r
o
ce
s
s
.
T
h
e
u
tili
za
tio
n
o
f
m
ax
im
u
m
s
tab
le
ex
tr
em
al
r
eg
io
n
s
(
MSE
R
s
)
is
em
p
lo
y
ed
f
o
r
th
e
p
u
r
p
o
s
e
o
f
d
etec
tin
g
ch
ar
ac
ter
c
o
m
p
o
n
e
n
ts
.
Stro
k
es
ar
e
em
p
lo
y
e
d
b
y
[
1
5
]
f
o
r
ch
a
r
ac
ter
lo
ca
lizatio
n
.
I
n
a
d
d
itio
n
,
a
g
e
n
er
ativ
e
s
h
ap
e
m
o
d
el
h
as
b
ee
n
d
ev
elo
p
ed
to
e
f
f
ec
tiv
ely
e
x
tr
ac
t
ch
ar
ac
ter
tr
aits
f
r
o
m
a
lim
ited
s
et
o
f
d
is
tin
ct
p
h
o
to
g
r
a
p
h
s
.
T
h
e
p
r
o
d
u
ctio
n
o
f
wo
r
d
-
lev
el
r
ec
o
g
n
itio
n
r
esu
lts
ty
p
ically
in
v
o
lv
es
th
e
u
s
e
o
f
h
e
u
r
is
tic
alg
o
r
ith
m
s
o
r
lan
g
u
a
g
e
m
o
d
els
d
u
r
in
g
th
e
ca
n
d
id
ate
ch
ar
ac
ter
r
ec
o
g
n
itio
n
s
tag
e
.
Nev
er
th
eless
,
ch
ar
ac
ter
-
b
ased
tech
n
iq
u
es
ex
h
ib
it
ce
r
tain
li
m
itatio
n
s
in
ter
m
s
o
f
ac
cu
r
ately
d
eter
m
in
in
g
th
e
ex
ac
t
p
o
s
itio
n
s
o
f
ch
ar
ac
ter
s
.
T
h
e
p
r
im
ar
y
f
ac
to
r
s
co
n
tr
ib
u
t
in
g
to
th
is
is
s
u
e
in
clu
d
e
a
clu
tter
ed
b
ac
k
g
r
o
u
n
d
an
d
in
s
u
f
f
icien
t le
tter
s
p
ac
i
n
g
.
An
OC
R
task
r
ef
er
s
t
o
th
e
p
r
o
ce
s
s
o
f
r
ec
o
g
n
izin
g
an
d
ex
tr
ac
tin
g
te
x
t
f
r
o
m
s
ce
n
es
.
T
h
e
OC
R
tech
n
o
lo
g
y
p
e
r
f
o
r
m
s
ef
f
ec
tiv
e
ly
in
c
o
n
tr
o
lled
en
v
ir
o
n
m
e
n
ts
.
Ho
wev
er
,
it
en
co
u
n
ter
s
d
if
f
ic
u
lties
wh
en
d
ea
lin
g
with
co
m
p
lex
ty
p
ef
ac
es
an
d
b
ac
k
g
r
o
u
n
d
s
,
s
u
ch
as
ch
a
r
ac
ter
s
wr
itten
in
cu
r
s
iv
e
.
I
m
a
g
e
d
eter
io
r
atio
n
ca
n
o
cc
u
r
d
u
e
to
v
ar
io
u
s
en
v
ir
o
n
m
en
tal
f
ac
to
r
s
s
u
ch
as
b
lu
r
,
l
ig
h
t,
an
d
o
r
ien
tatio
n
.
T
h
ese
f
ac
to
r
s
ca
n
lead
to
s
ig
n
if
ican
t
is
s
u
es
.
Su
b
s
eq
u
en
t
iter
atio
n
s
in
th
e
d
ev
el
o
p
m
en
t
p
r
o
ce
s
s
en
co
m
p
ass
th
e
ex
tr
ac
tio
n
o
f
p
er
tin
en
t
elem
en
ts
f
r
o
m
te
x
tu
al
s
ec
tio
n
s
an
d
th
e
m
an
ag
em
e
n
t
o
f
v
ar
i
atio
n
s
in
s
ize,
co
lo
r
,
an
d
lay
o
u
t
.
T
h
e
ac
c
u
r
ac
y
o
f
tex
t
r
ec
o
g
n
itio
n
ca
n
b
e
in
f
l
u
en
ce
d
b
y
v
ar
io
u
s
f
ac
to
r
s
s
u
ch
as
f
o
n
t
s
izes,
s
ty
les,
co
lo
r
s
,
tex
tu
r
es,
an
d
o
r
ien
tatio
n
[
1
6
]
.
T
h
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
in
v
o
lv
es
u
tili
zi
n
g
ap
p
r
o
p
r
iate
tex
t
lo
ca
lizatio
n
tech
n
iq
u
es
to
ef
f
ec
tiv
ely
h
a
n
d
le
t
h
ese
co
m
p
lex
ities
.
Var
io
u
s
m
eth
o
d
o
lo
g
ies
h
av
e
b
ee
n
s
u
g
g
ested
to
ac
h
iev
e
th
e
task
s
o
f
ex
tr
ac
tio
n
,
r
ec
o
g
n
itio
n
,
an
d
lo
ca
tio
n
.
T
h
e
co
n
v
o
l
u
tio
n
al
r
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
(
C
R
NN)
s
y
s
tem
u
tili
ze
s
co
n
v
o
lu
tio
n
al
a
n
d
r
ec
u
r
r
en
t
n
eu
r
al
n
etwo
r
k
s
to
g
e
n
er
ate
c
h
ar
ac
ter
s
eq
u
en
ce
f
ea
tu
r
es
.
T
h
is
s
y
s
tem
is
d
esig
n
ed
as
an
en
d
-
to
-
e
n
d
s
o
lu
tio
n
.
T
h
e
v
an
ill
a
C
T
C
alg
o
r
ith
m
is
d
esig
n
ed
f
o
r
1
D
p
r
o
b
ab
ilit
y
d
i
s
tr
ib
u
tio
n
s
.
T
h
e
c
o
n
d
itio
n
al
p
r
o
b
ab
ilit
y
o
f
lab
els
f
r
o
m
2
D
d
is
tr
ib
u
tio
n
s
ca
n
b
e
co
m
p
u
ted
u
s
in
g
th
e
2
DC
T
C
ap
p
r
o
ac
h
,
as
d
escr
ib
ed
in
[
1
7
]
.
T
h
er
ef
o
r
e,
it
is
id
ea
lly
s
u
ited
f
o
r
a
n
aly
zin
g
u
n
p
r
ed
ictab
le
o
r
in
co
n
s
is
ten
t
tex
t
.
T
h
e
o
r
g
an
izatio
n
s
r
o
b
u
s
t
t
ex
t
r
ec
o
g
n
izer
with
au
to
m
atic
r
ec
tific
atio
n
(
R
AR
E
)
,
atten
tio
n
al
s
ce
n
e
tex
t r
ec
o
g
n
izer
(
ASTE
R
)
,
MO
R
AN,
an
d
r
esid
u
al
co
lo
r
ec
tal
n
eo
p
lasi
a
(
RCN
)
s
u
cc
ess
f
u
lly
tr
an
s
f
o
r
m
ed
th
e
in
co
n
s
is
ten
t
tex
t
im
ag
es
in
to
ac
cu
r
ate
v
er
s
io
n
s
.
Su
b
s
eq
u
en
tly
,
tex
t
r
ec
o
g
n
itio
n
was
ac
c
o
m
p
li
s
h
ed
b
y
em
p
l
o
y
in
g
an
atten
tio
n
-
b
ased
r
ec
o
g
n
itio
n
n
etwo
r
k
with
in
an
e
n
co
d
e
r
-
d
ec
o
d
er
f
r
am
ewo
r
k
.
T
h
e
T
P
S
tr
an
s
f
o
r
m
atio
n
m
eth
o
d
ca
n
b
e
em
p
l
o
y
ed
to
ac
cu
r
ately
r
ec
tify
in
co
n
s
is
ten
t
s
ce
n
e
tex
t
[
9
]
.
T
h
e
r
ec
tific
at
io
n
o
f
ir
r
e
g
u
lar
s
ce
n
e
tex
t
c
an
b
e
en
h
a
n
ce
d
b
y
i
n
co
r
p
o
r
atin
g
a
s
y
m
m
et
r
y
co
n
s
tr
ain
t
in
to
t
h
e
r
ec
tific
ati
o
n
n
etwo
r
k
,
as
s
u
g
g
ested
b
y
[
1
1
]
.
Gu
o
et
a
l
.
[
1
0
]
d
ev
el
o
p
ed
a
m
u
lti
-
o
b
ject
r
ec
tific
atio
n
n
etwo
r
k
t
h
at
o
p
e
r
ates
b
y
m
o
d
if
y
in
g
th
e
o
f
f
s
et
o
f
ea
ch
p
i
x
el
.
A
p
o
ten
tial
s
o
l
u
tio
n
f
o
r
co
r
r
ec
tin
g
in
d
iv
id
u
al
ch
a
r
ac
ter
s
in
an
in
p
u
t
p
ictu
r
e
is
to
em
p
lo
y
a
c
h
a
r
ac
ter
-
awa
r
e
n
eu
r
al
n
etwo
r
k
i
n
s
tead
o
f
co
r
r
ec
tin
g
th
e
en
tire
p
ictu
r
e
.
Fo
r
th
e
p
u
r
p
o
s
e
o
f
en
h
a
n
cin
g
s
p
atial
in
f
o
r
m
atio
n
r
eten
tio
n
,
2
D
-
b
ased
ap
p
r
o
ac
h
es
em
p
lo
y
en
co
d
in
g
tech
n
iq
u
es to
co
n
v
er
t th
e
in
p
u
t
p
ictu
r
e
in
to
2
D
r
ep
r
esen
tatio
n
s
.
A
d
u
al
r
e
lat
io
n
m
o
d
u
le
[
1
8
]
,
[
1
9
]
h
as
b
e
e
n
d
e
v
e
lo
p
e
d
s
p
ec
i
f
ical
ly
f
o
r
s
ce
n
e
te
x
t
r
ec
o
g
n
iti
o
n
i
n
o
r
d
e
r
to
ex
tr
ac
t
c
o
m
p
l
em
e
n
ta
r
y
c
h
a
r
a
cte
r
is
t
ics
s
i
m
u
lt
a
n
e
o
u
s
ly
.
T
h
e
m
o
d
u
le
co
n
s
is
ts
o
f
tw
o
b
r
an
ch
es:
a
l
o
n
g
-
r
a
n
g
e
co
n
t
ex
tu
al
b
r
a
n
ch
a
n
d
a
lo
ca
l v
is
u
al
b
r
a
n
c
h
.
T
h
e
l
o
ca
l
v
is
u
al
b
r
a
n
ch
u
til
iz
es
a
t
o
p
o
l
o
g
ica
ll
y
-
aw
ar
e
o
p
e
r
a
ti
o
n
t
o
ef
f
e
cti
v
e
ly
ca
p
t
u
r
e
t
h
e
u
n
i
q
u
e
ch
ar
ac
t
er
is
ti
cs
wi
th
in
i
n
d
i
v
i
d
u
al
c
h
ar
ac
t
er
s
a
n
d
i
d
e
n
t
if
y
d
is
ti
n
g
u
is
h
i
n
g
ele
m
e
n
ts
am
o
n
g
m
u
lti
p
l
e
c
h
a
r
a
cte
r
s
.
T
h
e
l
o
n
g
-
r
a
n
g
e
c
o
n
te
x
t
u
a
l
b
r
a
n
c
h
i
n
t
eg
r
a
tes
i
n
t
er
-
c
h
ar
ac
t
er
ass
o
ciat
io
n
s
i
n
t
o
f
e
at
u
r
e
m
a
p
s
t
h
r
o
u
g
h
a
s
t
r
ai
g
h
t
f
o
r
w
ar
d
y
et
e
f
f
ici
e
n
t
a
p
p
r
o
a
ch
.
T
h
e
d
u
al
r
e
lat
io
n
m
o
d
u
le
is
a
p
l
u
g
-
a
n
d
-
p
l
ay
ele
m
e
n
t th
at
ca
n
b
e
s
e
am
less
l
y
in
te
g
r
at
ed
i
n
t
o
c
o
n
t
em
p
o
r
ar
y
d
e
ep
a
r
c
h
it
ec
tu
r
es
.
T
h
e
u
s
e
r
h
as
p
r
o
v
i
d
e
d
t
h
e
t
ex
t
[
2
0
]
.
T
h
e
f
o
ll
o
w
in
g
t
ex
t
p
r
es
e
n
ts
a
m
o
d
e
l
s
p
e
ci
f
i
ca
ll
y
d
esi
g
n
e
d
f
o
r
att
e
n
ti
o
n
,
w
h
i
ch
a
im
s
to
e
n
h
a
n
c
e
att
en
ti
o
n
ali
g
n
m
e
n
t
b
y
a
c
q
u
ir
i
n
g
m
o
r
e
r
e
f
i
n
e
d
v
is
u
al
c
h
a
r
ac
te
r
is
t
ic
s
.
T
h
e
i
n
itia
l
s
ta
g
e
i
n
v
o
l
v
es
t
h
e
a
p
p
lic
ati
o
n
o
f
d
e
f
o
r
m
a
b
l
e
c
o
n
v
o
l
u
t
io
n
t
o
e
x
t
r
a
ct
h
i
g
h
-
q
u
ali
ty
v
is
u
al
f
ea
tu
r
es
.
T
h
e
s
u
b
s
e
q
u
e
n
t
s
ta
g
e
em
p
l
o
y
s
a
l
ay
e
r
e
d
s
elf
-
a
tte
n
t
io
n
a
r
c
h
it
ec
t
u
r
e
wi
th
in
te
r
m
e
d
ia
te
s
u
p
e
r
v
is
i
o
n
t
o
e
n
h
a
n
c
e
t
h
e
q
u
al
it
y
o
f
v
is
u
al
f
ea
t
u
r
e
r
e
p
r
es
en
tat
io
n
.
I
n
a
d
d
it
io
n
,
th
e
att
e
n
ti
o
n
ali
g
n
m
e
n
t
p
r
o
c
ess
is
f
a
cili
tat
ed
b
y
e
m
p
lo
y
i
n
g
s
p
ati
al
lo
ca
ti
o
n
e
n
c
o
d
i
n
g
.
3.
P
RO
P
O
SE
D
M
E
T
H
O
DO
L
O
G
Y
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
co
m
b
i
n
e
s
an
atten
tio
n
-
b
ased
en
co
d
er
,
an
AFS
m
o
d
u
le,
an
d
an
atten
t
io
n
-
b
ased
d
ec
o
d
er
,
n
am
e
d
HARN
co
n
s
is
ts
o
f
an
atten
tio
n
-
b
ased
en
co
d
er
,
a
n
AFS
an
d
an
atten
t
io
n
-
b
ased
d
ec
o
d
e
r
.
Atten
tio
n
-
b
ased
d
ec
o
d
e
r
co
n
s
is
ts
o
f
two
m
o
d
u
les
k
n
o
wn
at
ten
tio
n
d
ec
o
d
er
+tr
a
n
s
f
o
r
m
e
r
a
n
d
th
e
o
t
h
er
is
t
h
e
AFS
m
o
d
u
le
.
T
h
e
m
o
d
el
f
o
r
tr
ain
in
g
k
n
o
wn
as
th
e
atten
tio
n
d
ec
o
d
er
+tr
an
s
f
o
r
m
er
m
o
d
el
en
s
u
r
es
p
r
o
p
er
g
u
id
an
ce
f
o
r
b
etter
u
n
d
er
s
tan
d
in
g
o
f
an
d
r
ep
r
esen
tatio
n
o
f
f
e
atu
r
es
.
Up
o
n
ass
u
m
p
tio
n
it
is
co
n
s
id
er
ed
th
at
th
e
atten
tio
n
-
b
ased
e
n
co
d
er
an
d
th
e
AFS
m
o
d
u
le
th
at
p
r
e
d
i
cts
th
e
ch
ar
ac
ter
s
eq
u
e
n
ce
s
,
th
at
m
in
im
ize
th
e
p
ar
am
eter
s
in
co
m
p
ar
is
o
n
wit
h
th
e
r
ec
o
g
n
izer
.
Fig
u
r
e
2
s
h
o
ws th
e
p
r
o
p
o
s
ed
ar
ch
itectu
r
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
n
h
a
n
ce
d
s
ce
n
e
text
r
ec
o
g
n
iti
o
n
u
s
in
g
d
ee
p
lea
r
n
in
g
b
a
s
ed
h
yb
r
id
a
tten
tio
n
…
(
R
a
tn
a
ma
l
a
S
.
P
a
til
)
4931
Fig
u
r
e
2
.
Pro
p
o
s
ed
ar
c
h
itectu
r
e
3
.
1
.
At
t
ent
io
n
-
ba
s
ed
enco
der
C
o
n
ten
t
-
b
ased
atten
tio
n
h
as b
ee
n
em
er
g
ed
as a
m
ain
p
ar
t
o
f
th
e
s
eq
u
en
ce
m
o
d
el
an
d
r
elev
an
t m
o
d
els
f
o
r
d
if
f
er
en
t
task
s
wh
ich
ac
c
o
m
m
o
d
ate
lo
n
g
in
ter
ac
tio
n
s
with
o
u
t
co
n
s
id
er
in
g
th
e
d
is
tan
ce
.
Ho
wev
er
,
th
e
y
ar
e
n
o
t
ca
p
ab
le
o
f
ex
tr
ac
tin
g
f
in
e
-
g
r
ain
e
d
lo
ca
l
in
f
o
r
m
atio
n
.
T
h
e
C
NN
u
tili
ze
s
a
co
m
m
o
n
ly
em
p
lo
y
e
d
p
o
s
itio
n
-
b
ased
k
e
r
n
el
to
p
er
f
o
r
m
a
s
ca
n
o
n
t
h
e
two
-
d
im
e
n
s
io
n
al
in
p
u
t d
ata
.
T
h
e
m
ain
lim
i
tatio
n
o
f
th
e
s
y
s
tem
is
th
e
lack
o
f
g
lo
b
al
r
elatio
n
s
h
ip
m
o
d
elin
g
.
Ho
we
v
er
,
th
e
s
y
s
tem
is
ab
le
to
ca
p
tu
r
e
l
o
ca
l
f
ea
tu
r
e
p
atter
n
s
th
r
o
u
g
h
t
h
e
estab
lis
h
ed
lin
k
s
.
T
h
e
cr
ea
tio
n
o
f
a
r
ec
o
g
n
itio
n
m
o
d
el
i
n
v
o
lv
es
th
e
c
o
m
b
in
ati
o
n
o
f
co
n
v
o
lu
tio
n
a
l
an
d
atten
tio
n
m
o
d
els
.
T
h
e
co
n
v
o
lu
tio
n
al
m
o
d
u
le
is
p
r
im
ar
i
ly
d
ed
icate
d
to
lo
ca
l
co
n
te
x
t
m
o
d
elin
g
,
wh
er
ea
s
th
is
m
o
d
u
le
f
o
cu
s
es o
n
estab
lis
h
in
g
lo
n
g
-
ter
m
r
elatio
n
s
h
ip
s
.
3
.
2
.
E
x
t
ended
a
t
t
ent
io
n
T
h
e
two
-
lay
er
s
u
b
-
s
am
p
le
m
o
d
u
le
en
h
an
ce
s
m
em
o
r
y
u
tili
za
tio
n
b
y
elim
in
atin
g
d
is
cr
im
in
ativ
e
f
ea
tu
r
es,
as
o
p
p
o
s
ed
to
th
e
co
n
v
en
tio
n
al
f
ea
tu
r
e
s
elec
tio
n
s
tag
e
th
at
f
o
cu
s
es
o
n
ex
tr
ac
tin
g
v
is
u
al
f
ea
tu
r
es
u
s
in
g
d
ee
p
er
C
NNs
th
r
o
u
g
h
t
h
e
u
tili
za
tio
n
o
f
two
co
n
v
o
l
u
tio
n
al
lay
er
s
.
A
lin
ea
r
o
p
e
r
atio
n
is
ap
p
licab
le
f
o
r
r
esizin
g
th
e
en
co
d
e
r
-
len
g
t
h
v
e
cto
r
A
th
r
o
u
g
h
th
e
s
h
ap
e
o
f
(
V
,
f
s
)
wh
er
ein
V
d
en
o
tes
th
e
s
eq
u
en
ce
l
en
g
th
an
d
f
s
is
th
e
s
ize
.
On
ce
th
is
is
d
o
n
e
th
e
in
p
u
t
is
f
ed
to
th
e
e
n
co
d
er
wh
ich
c
o
n
s
is
ts
o
f
a
co
llectio
n
o
f
P
g
en
co
d
er
b
lo
ck
s
wh
ich
co
n
s
is
ts
o
f
a
f
o
r
war
d
p
ass
an
d
atten
tio
n
-
b
ased
co
n
v
o
lu
tio
n
al
m
o
d
u
le,
t
h
e
att
en
tio
n
m
ec
h
an
is
m
co
n
s
is
ts
o
f
a
PS
m
o
d
u
le
.
T
h
e
elem
en
twis
e
ad
d
itio
n
co
n
s
is
tin
g
o
f
wo
r
d
-
em
b
ed
d
i
n
g
with
tem
p
o
r
al
e
n
co
d
i
n
g
th
r
o
u
g
h
o
u
t
th
e
atten
tio
n
s
co
r
e
with
in
ea
c
h
lay
e
r
f
o
r
g
en
er
al
izatio
n
o
n
v
a
r
io
u
s
in
p
u
t
len
g
t
h
s
.
Fig
u
r
e
3
s
h
o
ws
th
e
atten
tio
n
-
b
ased
e
n
co
d
e
r
d
e
co
d
er
f
r
am
ewo
r
k
.
T
h
e
o
u
tp
u
t is co
m
p
u
ted
b
y
(
1
)
:
H
j
=
s
oft
ma
x
(
SM
V
+
ST
V
√
f
m
j
)
(
1
)
W
h
er
ein
S
d
ep
icts
th
e
q
u
er
ies
an
d
S
=
A
Y
s
,
Y
s
∈
T
f
s
,
f
m
j
,
M
d
e
p
icts
th
e
k
ey
s
as
M
=
A
Y
m
,
Y
m
∈
T
f
s
,
f
m
j
.
T
d
en
o
tes
p
o
s
itio
n
al
en
co
d
in
g
a
n
d
T
=
R
Y
t
,
Y
t
∈
T
f
s
,
f
m
j
,
R
∈
T
V
∗
f
s
.
T
h
e
o
u
tp
u
t
is
in
teg
r
ated
an
d
p
r
o
jecte
d
as
d
ep
icted
as (
2
)
:
PS
(
Z
)=
in
te
gr
a
te
(
H
1
,
H
2
,
…
…
…
.
.
,
H
j
)
Y
q
(
2
)
W
h
er
ein
Y
q
∈
T
f
x
∗
f
x
co
n
s
is
ts
o
f
a
lin
ea
r
tr
an
s
f
o
r
m
atio
n
as
f
x
=
j
∗
f
x
j
an
d
f
m
j
=
f
x
j
.
3
.
3
.
Sh
o
rt
-
r
a
ng
e
co
nte
x
t
m
o
dellin
g
Her
e,
th
e
m
o
d
el
in
teg
r
ates
a
co
n
v
o
lu
ti
o
n
al
m
o
d
u
le
with
th
e
s
elf
-
atten
tio
n
m
o
d
u
le
b
y
ca
p
t
u
r
in
g
b
o
th
g
lo
b
al
an
d
lo
ca
l
co
n
te
x
ts
.
Po
in
t
-
wis
e
co
n
v
o
lu
tio
n
in
s
id
e
th
e
s
elf
-
atten
tio
n
m
o
d
u
le
m
ak
es
u
p
th
e
C
o
n
v
m
o
d
u
le
.
T
h
e
g
ated
lin
ea
r
a
ctiv
atio
n
lay
er
,
s
o
m
etim
es
r
ef
er
r
ed
to
as
p
o
i
n
twis
e
co
n
v
o
lu
tio
n
v
ia
f
ac
to
r
ex
p
an
s
io
n
,
m
a
k
es
u
p
th
e
c
o
n
v
o
lu
tio
n
m
o
d
u
le
.
C
o
n
tex
t
m
o
d
ellin
g
is
ac
co
m
p
lis
h
ed
an
d
p
r
o
ce
s
s
in
g
is
r
ed
u
ce
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l
.
13
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
4
9
2
7
-
4
9
3
8
4932
with
a
1
-
D
(
d
im
en
s
io
n
)
d
e
p
th
co
n
v
o
l
u
tio
n
al
m
o
d
u
le
.
An
ac
tiv
atio
n
lay
er
th
at
tr
ain
s
th
e
r
eg
u
lar
izatio
n
o
f
d
ee
p
m
o
d
els
r
ec
eiv
es
th
e
b
atch
-
wis
e
o
u
tp
u
t
.
T
h
e
lo
ca
l
an
d
g
lo
b
al
p
er
s
p
ec
tiv
es
o
f
th
e
p
ictu
r
e
ar
e
s
u
cc
ess
f
u
lly
co
m
b
in
ed
b
y
th
e
s
elf
-
atten
tio
n
m
o
d
el
an
d
th
e
co
n
v
o
lu
tio
n
al
m
o
d
u
le
.
T
h
e
two
f
ee
d
-
f
o
r
war
d
lev
els
in
clu
d
e
s
eg
m
en
ts
th
at
m
a
k
e
u
p
th
e
f
o
r
war
d
p
ass
m
o
d
u
le
.
T
h
e
e
n
c
o
d
er
’
s
o
u
tp
u
t
is
r
ep
r
esen
ted
b
y
(
3
)
.
Her
e,
h
o
wev
er
FP
r
ef
er
s
to
th
e
f
o
r
war
d
p
as
s
,
C
o
n
v
d
en
o
tes
th
e
co
n
v
o
lu
tio
n
al
m
o
d
u
le
an
d
n
or
m
_
l
a
ye
r
r
ef
er
s
to
n
o
r
m
aliza
tio
n
.
Fo
r
in
p
u
t
A
k
to
th
e
en
co
d
er
b
lo
ck
k
,
Output
k
r
ef
er
s
to
th
e
o
u
tp
u
t
o
f
th
e
b
lo
ck
,
f
o
r
k
∈
[
1
,
P
g
]
.
A
k
=
A
k
+
0
.
5
FP
(
A
k
)
,
A
k
=
A
k
+
PS
(
A
k
)
A
k
=
A
k
+
C
on
v
(
A
k
)
,
Output
k
=
n
or
m
_
l
a
ye
r
(
A
k
+
0
.
5F
P
(
A
k
)
(
3
)
Fig
u
r
e
3
.
Atten
tio
n
-
b
ased
en
c
o
d
er
d
ec
o
d
er
f
r
am
ewo
r
k
3
.
4
.
At
t
ent
io
n dec
o
der+tr
a
ns
f
o
rm
er
mo
del
A
p
ictu
r
e
is
th
e
i
n
p
u
t
u
s
ed
b
y
th
e
two
-
la
y
er
s
u
b
s
am
p
lin
g
m
o
d
u
le
.
T
o
c
h
an
g
e
th
e
e
n
co
d
er
v
ec
to
r
’
s
len
g
th
,
th
e
m
o
d
u
le
p
er
f
o
r
m
s
a
lin
ea
r
o
p
er
atio
n
.
T
h
e
len
g
th
o
f
th
e
en
co
d
er
v
ec
to
r
is
m
o
d
if
ied
b
y
co
m
b
in
in
g
th
e
atten
tio
n
-
b
ased
e
n
co
d
e
r
with
a
f
ea
t
u
r
e
-
b
ased
r
ep
r
esen
tatio
n
.
T
h
e
AFS
m
o
d
u
le
a
n
d
th
e
d
ec
o
d
e
r
-
b
ased
tr
an
s
f
o
r
m
er
m
o
d
el
ar
e
th
e
two
m
o
d
u
les
th
at
im
p
lem
en
t
th
e
f
ea
tu
r
es
.
E
f
f
ec
tiv
e
ass
u
m
p
tio
n
is
th
e
m
ain
em
p
h
asis
o
f
th
e
d
ec
o
d
e
r
-
b
ase
d
tr
an
s
f
o
r
m
e
r
m
o
d
el,
a
n
d
th
e
AFS
m
o
d
u
le
aim
s
to
ac
h
iev
e
th
e
s
am
e
th
in
g
.
T
h
e
d
ec
o
d
er
b
lo
ck
,
wh
ich
m
ak
es
u
p
th
e
s
y
s
tem
,
is
s
ep
ar
ated
in
to
th
r
ee
p
a
r
ts
:
f
o
r
war
d
p
ass
,
en
s
em
b
le
atten
tio
n
,
an
d
f
o
r
war
d
s
elf
-
a
tten
tio
n
.
Usi
n
g
th
e
f
u
n
ctio
n
,
a
c
h
ar
ac
ter
-
l
ev
el
em
b
e
d
d
in
g
ap
p
r
o
ac
h
is
f
i
r
s
t
u
s
ed
to
tu
r
n
th
e
in
p
u
t
ch
ar
ac
ter
in
to
a
v
ec
to
r
.
E
v
er
y
em
b
ed
d
in
g
g
ets
an
ad
d
itio
n
al
co
p
y
o
f
th
e
lo
ca
tio
n
en
co
d
in
g
with
th
e
s
am
e
d
im
en
s
io
n
.
B
y
k
ee
p
in
g
p
o
s
itio
n
s
f
r
o
m
en
g
a
g
in
g
in
co
n
cu
r
r
en
t
p
o
s
itio
n
s
,
th
e
f
o
r
war
d
s
elf
-
atten
tio
n
m
eth
o
d
en
s
u
r
es
p
r
ec
is
e
p
r
e
d
ic
tio
n
an
d
co
n
ce
n
tr
ates
o
n
d
ata
th
at
co
m
es
b
ef
o
r
e
th
e
p
r
esen
t
p
o
s
itio
n
.
T
h
e
v
al
u
e
f
o
r
e
n
s
em
b
le
atten
tio
n
is
o
b
t
ain
ed
f
r
o
m
t
h
e
en
c
o
d
er
’
s
o
u
tp
u
t,
wh
ich
co
n
tain
s
t
h
e
r
e
q
u
ests
f
o
r
en
s
em
b
le
atten
tio
n
.
T
h
e
f
in
al
s
tag
e
o
f
c
o
m
p
u
tin
g
,
th
e
f
o
r
war
d
p
ass
,
is
s
ep
ar
ated
in
to
two
la
y
er
s
:
a
lin
ea
r
lay
er
with
t
h
e
lin
ea
r
ac
tiv
atio
n
f
u
n
ctio
n
an
d
a
lin
ea
r
lay
e
r
with
th
e
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
ac
tiv
atio
n
.
T
h
e
co
n
n
ec
tio
n
s
th
at
ar
e
s
till
p
r
esen
t in
th
e
d
ec
o
d
er
b
l
o
ck
'
s
s
u
b
-
lay
er
m
ak
e
u
p
th
e
n
o
r
m
aliza
tio
n
lay
e
r
.
3
.
5
.
A
lig
nm
ent
-
f
re
e
s
equence
-
to
-
s
equence
a
nd
deco
der
b
lo
ck
T
h
e
AFS
m
eth
o
d
is
wid
ely
r
e
co
g
n
ized
as
a
r
o
b
u
s
t
an
d
ef
f
e
ctiv
e
ap
p
r
o
ac
h
f
o
r
ad
d
r
ess
in
g
s
eq
u
en
tial
p
r
o
b
lem
s
.
T
h
is
is
ac
co
m
p
lis
h
ed
b
y
em
p
lo
y
in
g
d
y
n
am
ic
p
r
o
g
r
am
m
i
n
g
tech
n
iq
u
es
an
d
lev
er
a
g
in
g
.
I
t
is
im
p
o
r
tan
t to
r
ec
o
g
n
ize
th
at
th
e
ap
p
r
o
ac
h
r
elies o
n
s
ev
e
r
al
co
n
d
itio
n
al
in
d
e
p
en
d
e
n
ce
ass
u
m
p
tio
n
s
.
T
h
e
in
clu
s
io
n
o
f
th
ese
ad
d
iti
o
n
al
p
r
esu
m
p
tio
n
s
is
d
ee
m
e
d
ess
en
tial
f
o
r
atten
tio
n
d
ec
o
d
er
s
.
T
h
e
tr
ain
in
g
p
r
o
ce
s
s
o
f
th
e
atte
n
tio
n
d
ec
o
d
e
r
ca
n
b
e
ch
alle
n
g
in
g
wh
en
d
ea
lin
g
with
n
o
is
y
d
ata
a
n
d
lo
n
g
s
eq
u
en
ce
s
,
as
it
r
eq
u
ir
es
ac
h
iev
in
g
p
r
ec
is
e
alig
n
m
en
t
.
I
n
ad
d
itio
n
,
th
e
atten
tio
n
-
b
ased
m
o
d
el
g
en
er
ates
p
r
ed
ictio
n
s
at
ea
ch
tim
e
s
tep
b
y
lev
er
a
g
in
g
th
e
p
r
e
v
io
u
s
ch
ar
ac
ter
s
’
ac
tiv
ities
.
T
h
e
f
r
am
e
-
by
-
f
r
am
e
d
ec
o
d
in
g
p
r
o
ce
s
s
is
a
s
ig
n
if
ican
t
co
n
tr
i
b
u
tin
g
f
ac
t
o
r
to
t
h
e
d
if
f
ic
u
lties
en
co
u
n
ter
e
d
wh
en
u
tili
zin
g
tex
t
r
ec
o
g
n
itio
n
in
r
ea
l
-
wo
r
ld
s
ce
n
ar
io
s
.
T
h
e
r
ec
o
m
m
en
d
e
d
ap
p
r
o
ac
h
f
o
r
ad
d
r
ess
in
g
th
e
af
o
r
em
en
tio
n
e
d
is
s
u
es
is
to
u
tili
ze
an
atten
tio
n
d
ec
o
d
er
an
d
an
AFS
.
T
h
e
d
ec
o
d
er
c
o
n
s
is
ts
o
f
two
m
ain
c
o
m
p
o
n
en
ts
:
th
e
atten
tio
n
m
o
d
u
le
an
d
th
e
AFS
m
o
d
u
le
.
T
h
e
atten
tio
n
m
o
d
u
le
is
b
u
ilt
u
p
o
n
th
e
t
r
an
s
f
o
r
m
er
-
d
ec
o
d
er
ar
c
h
itectu
r
e
.
T
h
e
p
r
im
ar
y
f
u
n
ctio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
n
h
a
n
ce
d
s
ce
n
e
text
r
ec
o
g
n
iti
o
n
u
s
in
g
d
ee
p
lea
r
n
in
g
b
a
s
ed
h
yb
r
id
a
tten
tio
n
…
(
R
a
tn
a
ma
l
a
S
.
P
a
til
)
4933
o
f
th
e
AFS
m
o
d
u
le
is
to
tr
a
n
s
f
o
r
m
t
h
e
o
u
tp
u
t
o
f
th
e
en
c
o
d
er
in
to
a
n
AFS
f
o
r
m
.
T
h
is
tr
an
s
f
o
r
m
atio
n
is
ac
h
iev
ed
u
s
in
g
a
s
in
g
le
lin
ea
r
lay
er
.
T
h
e
atten
tio
n
m
o
d
u
le
b
ased
o
n
th
e
tr
an
s
f
o
r
m
er
-
d
ec
o
d
er
is
u
tili
ze
d
in
th
e
co
m
p
u
tatio
n
o
f
atten
tio
n
-
b
ased
p
r
o
b
ab
ilit
ies,
as
d
em
o
n
s
tr
ated
in
p
r
ev
io
u
s
r
esear
ch
.
T
h
e
a
tten
tio
n
m
ec
h
a
n
is
m
is
a
r
eg
u
lar
izatio
n
ap
p
r
o
ac
h
t
h
at
is
s
p
ec
if
ically
u
tili
ze
d
in
th
e
tr
ain
in
g
p
h
ase
.
I
t
is
im
p
lem
en
ted
in
to
th
e
AFS
m
o
d
u
le
.
I
n
o
r
d
er
f
o
r
th
e
AFS
f
o
r
war
d
-
b
ac
k
war
d
alg
o
r
ith
m
t
o
o
p
er
ate
c
o
r
r
ec
tly
,
it
is
c
r
u
cia
l
th
at
th
e
in
p
u
t
a
n
d
o
u
tp
u
t seq
u
e
n
ce
s
ar
e
alig
n
e
d
p
r
ec
is
ely
.
I
n
a
d
d
itio
n
,
th
e
in
clu
s
io
n
o
f
t
h
e
AFS
m
o
d
u
le
ca
n
ac
ce
le
r
ate
th
e
c
o
n
v
e
r
g
en
ce
r
ate
o
f
th
e
n
etwo
r
k
.
T
h
e
u
tili
za
tio
n
o
f
th
e
atten
tio
n
m
o
d
u
le
d
o
es
n
o
t
r
eq
u
ir
e
an
y
co
n
d
itio
n
al
in
d
ep
e
n
d
en
c
e
ass
u
m
p
tio
n
s
.
T
h
e
in
clu
s
io
n
o
f
th
is
f
ea
tu
r
e
en
h
a
n
ce
s
th
e
ea
s
e
o
f
co
n
f
ig
u
r
i
n
g
a
r
eliab
le
an
d
ef
f
icien
t
m
o
n
it
o
r
in
g
s
y
s
tem
d
u
r
in
g
th
e
tr
ain
in
g
p
r
o
ce
s
s
o
f
th
e
r
ec
o
g
n
izer
.
T
h
e
in
teg
r
atio
n
o
f
th
e
atten
tio
n
m
o
d
u
le
an
d
th
e
AFS
m
o
d
u
le
is
ac
h
iev
ed
b
y
u
tili
zin
g
a
d
ec
o
d
er
th
at
in
co
r
p
o
r
ates
b
o
th
A
FS
an
d
atten
tio
n
m
eth
o
d
s
.
T
h
e
u
tili
za
tio
n
o
f
th
e
s
h
ar
ed
en
co
d
er
n
etwo
r
k
b
y
t
h
e
d
ec
o
d
er
m
o
d
u
le
is
d
o
n
e
in
an
ef
f
icien
t
m
an
n
er
.
T
h
e
at
ten
tio
n
ar
ch
itectu
r
e
em
p
lo
y
s
en
co
d
i
n
g
an
d
d
ec
o
d
in
g
m
eth
o
d
s
to
ef
f
ec
tiv
el
y
u
t
ilize
f
ea
tu
r
e
in
f
o
r
m
atio
n
p
r
ec
ed
in
g
th
e
cu
r
r
en
t
p
o
s
itio
n
.
Mo
r
eo
v
er
,
th
r
o
u
g
h
th
e
in
teg
r
atio
n
o
f
th
e
g
lo
b
al
p
r
o
b
ab
ilit
y
ca
lc
u
latio
n
a
p
p
r
o
ac
h
in
t
o
th
e
AFS
d
esig
n
,
th
e
s
y
s
tem
ca
n
u
tili
ze
f
ea
tu
r
e
in
f
o
r
m
atio
n
th
at
g
o
es
b
ey
o
n
d
th
e
cu
r
r
en
t
p
o
s
itio
n
.
T
h
e
co
m
b
in
atio
n
o
f
atten
tio
n
d
ec
o
d
e
r
an
d
AFS
ac
ce
ler
ates
n
etwo
r
k
co
n
v
e
r
g
en
ce
an
d
en
h
a
n
ce
s
r
ec
o
g
n
itio
n
p
er
f
o
r
m
a
n
ce
.
T
h
e
o
b
jectiv
e
th
at
h
as b
ee
n
d
ec
lar
e
d
is
as
(
4
)
:
N
=
ϑ
l
oss
A
F
S
+(
1
-
ϑ
)
N
AL
(
4
)
ϑ
is
th
e
p
ar
am
eter
th
at
is
tu
n
ed
in
th
e
r
an
g
e
0
≤
ϑ
≤
1
.
N
AL
s
h
o
ws
th
e
atten
tio
n
lo
s
s
,
co
n
s
is
ts
o
f
lo
s
s
f
u
n
ctio
n
.
l
oss
A
F
S
is
th
e
AFS
lo
s
s
.
T
h
e
o
b
jectiv
e
is
to
m
in
im
ize
th
e
n
eg
ativ
e
lo
g
-
lik
elih
o
o
d
wh
e
n
c
o
m
p
ar
ed
t
o
th
e
g
r
o
u
n
d
t
r
u
th
.
l
oss
A
F
S
=
−
l
ogR
(
I
U
)
(
5
)
Her
e
R
d
ep
icts
th
e
g
r
o
u
n
d
-
t
r
u
th
lab
el
s
eq
u
en
ce
wh
er
ein
U
=
u
1
,
…
…
…
.
.
u
V
is
th
e
s
eq
u
en
ce
th
at
th
e
en
co
d
er
pr
o
d
u
ce
s
u
s
in
g
th
e
tr
ain
in
g
im
ag
e
.
K
.
E
ac
h
u
v
is
th
e
p
r
o
b
a
b
ilit
y
d
is
tr
ib
u
tio
n
f
o
r
t
h
e
ch
ar
ac
ter
s
et
E
,
co
m
p
r
is
in
g
t
h
e
task
’
s
wh
o
le
lab
el
s
et
as
well
as
th
e
b
lan
k
la
b
el
.
A
m
ap
p
in
g
f
u
n
ctio
n
O
is
d
e
f
in
ed
th
r
o
u
g
h
th
e
s
eq
u
en
ce
ρ
∈
E
V
.
O
m
ap
s
ρ
b
y
elim
in
atin
g
th
e
d
u
p
licate
lab
els
f
ir
s
t,
f
o
llo
wed
b
y
th
e
b
lan
k
lab
el
.
T
h
e
co
n
d
itio
n
al
p
r
o
b
a
b
ilit
y
is
s
u
m
m
ed
as th
e
p
r
o
b
ab
ilit
ies o
f
ρ
th
at
is
m
ap
p
ed
b
y
O
with
I
.
R
(
I
U
)
=
∑
ρ
:
O
(
ρ
)
=
I
R
(
ρ
U
)
(
6
)
T
h
e
p
r
o
b
ab
ilit
y
o
f
ρ
is
d
en
o
te
d
b
y
t
h
e
R
(
ρ
U
)
=
∑
V
v
=
1
u
v
ρ
v
,
wh
er
ein
u
v
ρ
v
is
th
e
p
r
o
b
ab
ilit
y
ass
o
ciate
d
with
lab
el
ρ
v
at
tim
e
-
s
tam
p
v
.
4.
P
E
RF
O
RM
A
NCE
E
VA
L
U
AT
I
O
N
I
n
th
e
p
e
r
f
o
r
m
an
ce
ev
al
u
atio
n
o
f
tex
t
r
ec
o
g
n
itio
n
m
eth
o
d
o
lo
g
ies
ac
r
o
s
s
th
r
ee
d
atasets
,
in
clu
d
in
g
s
tr
ee
t
v
iew
tex
t
-
p
er
s
p
ec
tiv
e
(
SVTP)
,
I
C
1
3
,
an
d
C
UT
E
8
0
,
v
ar
y
in
g
lev
els
o
f
ac
cu
r
ac
y
is
o
b
s
er
v
ed
.
T
h
e
ev
alu
atio
n
o
f
th
is
p
r
o
p
o
s
ed
f
r
a
m
ewo
r
k
is
co
n
d
u
cted
u
s
in
g
b
o
th
ir
r
eg
u
lar
an
d
r
eg
u
lar
b
en
ch
m
ar
k
d
atasets
.
T
h
e
p
r
o
p
o
s
ed
m
HARN
u
tili
ze
s
d
ee
p
lear
n
in
g
lib
r
ar
ies
an
d
is
im
p
lem
en
ted
o
n
a
s
y
s
tem
co
n
f
ig
u
r
atio
n
f
ea
t
u
r
in
g
a
4
GB
C
UDA
-
en
ab
led
g
r
ap
h
ics
ca
r
d
with
1
6
GB
o
f
R
AM
,
o
p
er
atin
g
o
n
th
e
W
in
d
o
ws
p
latf
o
r
m
,
th
e
r
esu
lts
ar
e
ev
alu
ated
in
th
e
f
o
r
m
o
f
g
r
ap
h
an
d
tab
les
.
4
.
1
.
Da
t
a
s
et
det
a
ils
4
.
1
.
1
.
s
t
re
et
v
iew
t
e
x
t
d
a
t
a
s
et
T
h
e
s
tr
ee
t
v
iew
tex
t
(
S
VT
)
[
2
1
]
d
ataset
is
a
co
llect
io
n
o
f
6
4
7
o
u
t
d
o
o
r
s
tr
ee
t
im
ag
es
s
o
u
r
ce
d
f
r
o
m
Go
o
g
le
Stre
et
View,
d
esig
n
ed
f
o
r
tex
t
r
ec
o
g
n
itio
n
r
esear
ch
.
E
ac
h
im
ag
e
is
m
eticu
lo
u
s
ly
an
n
o
tated
with
in
f
o
r
m
atio
n
ab
o
u
t
t
h
e
lo
ca
tio
n
,
s
ize,
o
r
ien
tatio
n
,
an
d
tr
an
s
c
r
ib
ed
tex
t
p
r
esen
t
with
in
it
.
T
h
e
d
ataset
’
s
n
o
ta
b
le
co
m
p
lex
ity
s
tem
s
f
r
o
m
th
e
wid
e
ar
r
ay
o
f
tex
t
in
s
tan
ce
s
it
co
n
tain
s
,
s
h
o
wca
s
in
g
v
ar
io
u
s
lan
g
u
ag
es,
f
o
n
ts
,
s
izes,
o
r
ien
tatio
n
s
,
an
d
ap
p
ea
r
an
ce
s
.
T
h
is
r
ich
d
iv
er
s
ity
m
ak
es
it
an
in
v
alu
ab
le
r
eso
u
r
ce
f
o
r
ev
alu
atin
g
tex
t
d
etec
tio
n
an
d
r
ec
o
g
n
itio
n
alg
o
r
ith
m
s
,
esp
ec
ially
in
th
e
c
o
n
tex
t
o
f
r
ea
l
-
wo
r
ld
ch
allen
g
e
s
s
u
ch
as
v
ar
iab
le
lig
h
tin
g
co
n
d
itio
n
s
,
o
cc
lu
s
io
n
s
,
an
d
im
ag
e
q
u
ality
.
T
h
e
s
tr
ee
t
v
iew
tex
t
(
SVT)
d
ataset
is
o
f
ten
s
p
lit
in
to
tr
ain
in
g
a
n
d
test
s
ets,
en
ab
lin
g
r
esear
ch
e
r
s
to
d
ev
el
o
p
a
n
d
ass
es
s
tex
t
r
ec
o
g
n
itio
n
m
o
d
els
f
o
r
h
an
d
lin
g
th
e
co
m
p
lex
ities
o
f
tex
t in
o
u
td
o
o
r
s
ce
n
es
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l
.
13
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
4
9
2
7
-
4
9
3
8
4934
4
.
1
.
2
.
I
CDAR
2
0
1
3
(
I
C1
3
)
T
h
e
I
C
DAR
2
0
1
3
(
I
C
1
3
)
[
2
2
]
d
ataset
is
a
co
llectio
n
o
f
1
,
0
1
5
s
ce
n
e
tex
t
im
ag
es,
m
an
y
o
f
wh
ich
ar
e
in
h
er
ited
f
r
o
m
th
e
I
C
DAR
2
0
0
3
(
I
C
0
3
)
d
ataset
.
T
h
ese
im
a
g
es
ar
e
cr
u
cial
f
o
r
r
esear
ch
a
n
d
ev
alu
atio
n
in
th
e
f
ield
o
f
s
ce
n
e
tex
t
r
ec
o
g
n
iti
o
n
.
E
ac
h
im
a
g
e
is
m
eticu
lo
u
s
ly
an
n
o
tated
,
p
r
o
v
id
i
n
g
i
n
f
o
r
m
atio
n
a
b
o
u
t
tex
t
r
eg
io
n
lo
ca
tio
n
,
s
ize,
o
r
ie
n
tatio
n
,
a
n
d
t
h
e
tr
a
n
s
cr
ib
ed
te
x
t
.
T
h
e
d
ataset
’
s
co
m
p
lex
ity
a
r
is
es
f
r
o
m
th
e
d
iv
er
s
e
r
an
g
e
o
f
tex
t
in
s
tan
ce
s
it
co
n
tain
s
,
in
clu
d
in
g
v
ar
iatio
n
s
in
lan
g
u
ag
es,
f
o
n
ts
,
s
izes,
o
r
ien
tatio
n
s
,
an
d
b
ac
k
g
r
o
u
n
d
s
.
R
esear
ch
er
s
u
s
e
I
C
1
3
to
d
ev
elo
p
an
d
ass
ess
t
ex
t
r
ec
o
g
n
itio
n
alg
o
r
ith
m
s
,
f
a
cin
g
ch
allen
g
es
lik
e
ch
an
g
in
g
lig
h
tin
g
c
o
n
d
itio
n
s
,
o
cc
lu
s
io
n
s
,
an
d
v
ar
iab
le
im
ag
e
q
u
ality
.
T
h
e
d
ataset
is
o
f
ten
d
iv
i
d
ed
in
to
tr
ain
in
g
an
d
test
s
ets
f
o
r
m
o
d
e
l
d
ev
elo
p
m
e
n
t
an
d
e
v
alu
atio
n
,
an
d
p
r
o
p
er
citatio
n
is
ess
en
ti
al
wh
en
u
tili
zin
g
it
f
o
r
r
esear
ch
to
ac
k
n
o
wled
g
e
it
s
cr
ea
to
r
s
an
d
s
o
u
r
ce
.
4
.
1
.
3
.
CUTE
-
80
T
h
e
C
UT
E
8
0
[
2
3
]
d
ataset
co
m
p
r
is
es
288
-
wo
r
d
im
ag
es
p
r
i
m
ar
ily
em
p
h
asizin
g
n
o
n
-
lin
ea
r
,
cu
r
v
ed
tex
t
lay
o
u
ts
.
T
h
is
d
ataset
is
p
ar
ticu
lar
ly
v
al
u
ab
le
f
o
r
r
ese
ar
ch
in
OC
R
an
d
co
m
p
u
ter
v
is
io
n
,
o
f
f
e
r
in
g
a
s
p
ec
ialized
f
o
cu
s
o
n
th
e
r
ec
o
g
n
itio
n
o
f
tex
t
f
o
llo
win
g
cu
r
v
ed
o
r
i
r
r
eg
u
la
r
p
ath
s
,
as
o
p
p
o
s
ed
to
tr
ad
itio
n
al
h
o
r
izo
n
tal
o
r
lin
ea
r
te
x
t
.
C
UT
E
8
0
en
ab
les
r
esear
ch
e
r
s
to
d
ev
elo
p
an
d
e
v
alu
ate
alg
o
r
ith
m
s
s
p
ec
if
ically
g
en
er
ated
to
ad
d
r
ess
th
e
p
a
r
ticu
lar
d
if
f
icu
lties
in
id
en
tif
y
in
g
an
d
in
te
r
p
r
etin
g
tex
t
in
u
n
c
o
n
v
en
tio
n
al
lay
o
u
ts
,
in
clu
d
in
g
te
x
t sh
o
wn
in
s
id
e
cr
ea
tiv
e
cr
ea
tio
n
s
o
r
o
n
cu
r
v
ed
s
u
r
f
ac
es
.
4
.
2
.
E
x
perim
ent
a
l
a
na
ly
s
is
T
h
e
ex
p
er
im
en
tal
a
n
aly
s
is
ev
alu
ates
th
e
p
e
r
f
o
r
m
an
ce
o
f
v
ar
io
u
s
tex
t
r
ec
o
g
n
itio
n
m
eth
o
d
o
lo
g
ies
ac
r
o
s
s
th
r
ee
d
atasets
:
SVT
P,
I
C
1
3
,
an
d
C
UT
E
8
0
.
T
h
e
p
r
o
p
o
s
ed
HARN
m
eth
o
d
o
lo
g
y
is
ev
alu
ated
ag
ain
s
t
th
e
ex
is
tin
g
s
tate
-
of
-
ar
t
tech
n
iq
u
e
s
.
T
ab
le
1
s
h
o
ws
th
e
r
esu
lts
o
f
SVTP
d
ataset
.
Fo
r
S
VT
P
d
a
taset
th
e
ac
cu
r
ac
y
co
m
p
ar
is
o
n
g
r
ap
h
is
p
lo
tted
,
as
s
h
o
wn
in
F
ig
u
r
e
4
th
e
R
AR
E
m
eth
o
d
d
ep
icts
an
ac
cu
r
ac
y
o
f
7
1
.
8
%,
d
em
o
n
s
tr
ates
m
o
d
e
r
ate
tex
t
r
ec
o
g
n
itio
n
ca
p
ab
ilit
ies
.
ADC
D
s
u
r
p
ass
es
R
A
R
E
with
an
ac
cu
r
ac
y
o
f
7
5
.
8
%,
in
d
icatin
g
im
p
r
o
v
ed
p
er
f
o
r
m
a
n
ce
.
Ar
b
itra
r
y
o
r
ien
tatio
n
n
etwo
r
k
(
AON)
r
ep
o
r
ts
an
ac
cu
r
ac
y
o
f
7
3
%,
wh
ile
AC
E
ac
h
iev
es
7
0
.
1
%,
b
o
t
h
f
a
llin
g
with
in
a
s
im
ilar
p
er
f
o
r
m
an
ce
r
an
g
e
.
D
ec
o
u
p
led
atten
ti
o
n
n
etwo
r
k
(
DAN)
s
tan
d
s
o
u
t
with
an
ac
cu
r
ac
y
o
f
8
0
%,
d
e
m
o
n
s
tr
atin
g
s
tr
o
n
g
t
ex
t
r
ec
o
g
n
itio
n
ca
p
a
b
ilit
ies
.
C
h
ar
-
n
et
r
e
p
o
r
ts
a
n
ac
cu
r
ac
y
o
f
7
8
.
9
%,
s
im
ilar
to
en
d
-
to
-
e
n
d
tr
ai
n
ab
le
s
ce
n
e
te
x
t
r
ec
o
g
n
itio
n
s
y
s
tem
(
E
SIR)
a
n
d
ASTE
R
,
wh
ich
ac
h
iev
e
ac
cu
r
ac
ies
o
f
7
9
.
6
a
n
d
7
8
.
5
,
r
esp
ec
tiv
ely
.
ScR
N
ex
h
ib
its
s
tr
o
n
g
p
er
f
o
r
m
an
ce
with
an
ac
cu
r
ac
y
o
f
80
.
8
.
C
ASR
-
DR
Net
lead
s
th
e
lis
t
wi
th
a
h
ig
h
ac
cu
r
ac
y
o
f
8
5
,
in
d
icatin
g
r
o
b
u
s
t
tex
t
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
,
wh
ile
PS
s
ig
n
if
ican
tly
o
u
tp
er
f
o
r
m
s
th
e
r
est with
an
ac
c
u
r
ac
y
s
co
r
e
o
f
9
5
.
0
4
%
.
T
ab
le
1
.
R
esu
lts
f
o
r
SVTP
d
ataset
M
e
t
h
o
d
V
a
l
u
e
R
A
R
E
[
2
4
]
71
.
8
A
D
C
D
[
2
5
]
75
.
8
AON
[
2
6
]
73
A
C
E
[
2
7
]
73
.
9
DAN
[
2
8
]
80
C
h
a
r
-
n
e
t
[2
7
]
78
.
9
ESI
R
[
29
]
79
.
6
A
S
TER
[
3
0
]
78
.
5
S
c
R
N
[
3
0
]
80
.
8
C
A
S
R
-
D
R
N
e
t
[
5
]
85
PS
95
.
04
Fig
u
r
e
4
.
Acc
u
r
ac
y
c
o
m
p
ar
is
o
n
f
o
r
SVTP
d
ataset
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
n
h
a
n
ce
d
s
ce
n
e
text
r
ec
o
g
n
iti
o
n
u
s
in
g
d
ee
p
lea
r
n
in
g
b
a
s
ed
h
yb
r
id
a
tten
tio
n
…
(
R
a
tn
a
ma
l
a
S
.
P
a
til
)
4935
4
.
2
.
1
.
I
C1
3
da
t
a
s
et
Fo
r
I
C
1
3
d
ataset
th
e
ac
cu
r
ac
y
co
m
p
ar
is
o
n
g
r
ap
h
is
p
lo
tte
d
,
as
s
h
o
wn
in
T
a
b
le
2
an
d
F
ig
u
r
e
5
in
clu
d
es
v
ar
io
u
s
tex
t
r
ec
o
g
n
it
io
n
m
eth
o
d
o
lo
g
ies,
ea
ch
ac
c
o
m
p
an
ied
b
y
its
r
esp
ec
tiv
e
ac
cu
r
ac
y
v
alu
e
.
I
B
SR
s
h
o
ws
a
s
tr
o
n
g
ac
cu
r
ac
y
o
f
8
9
.
6
%,
d
em
o
n
s
tr
atin
g
p
r
o
f
i
cien
t
tex
t
r
ec
o
g
n
itio
n
.
CA
-
FC
N
ex
ce
ls
with
an
ac
cu
r
ac
y
o
f
9
1
.
5
%,
in
d
icatin
g
r
o
b
u
s
t
tex
t
r
e
co
g
n
itio
n
ca
p
a
b
ilit
ies
.
C
R
F
ac
h
iev
es
an
ac
cu
r
a
cy
o
f
8
1
.
8
%,
wh
ile
R
AR
E
p
er
f
o
r
m
s
well
with
an
ac
cu
r
ac
y
o
f
8
8
.
6
.
R
2
AM
m
ain
tain
s
a
s
o
lid
p
er
f
o
r
m
an
ce
with
an
ac
cu
r
ac
y
o
f
9
0
.
E
DG
s
tan
d
s
o
u
t
wi
th
an
im
p
r
ess
iv
e
ac
cu
r
ac
y
o
f
9
2
.
9
.
E
SI
R
d
em
o
n
s
tr
ates
a
h
ig
h
ac
cu
r
ac
y
o
f
9
1
.
3
,
wh
ile
ASTE
R
ex
ce
ls
with
an
ac
cu
r
ac
y
o
f
9
1
.
8
.
ScR
N
p
er
f
o
r
m
s
ex
ce
p
tio
n
ally
well
with
a
h
ig
h
ac
cu
r
ac
y
o
f
9
3
.
9
.
C
AS
R
-
DR
Ne
t
lead
s
th
e
li
s
t
with
a
r
em
ar
k
ab
le
ac
cu
r
ac
y
o
f
9
5
.
3
,
s
h
o
wca
s
in
g
s
u
p
e
r
io
r
tex
t
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
.
PS
s
ig
n
if
ican
tly
o
u
tp
er
f
o
r
m
s
th
e
o
th
er
s
with
an
ac
cu
r
ac
y
s
co
r
e
o
f
9
8
.
42
.
T
ab
le
2
.
R
esu
lts
f
o
r
I
C
1
3
d
ata
s
et
M
e
t
h
o
d
A
c
c
u
r
a
c
y
I
B
S
R
[3
1
]
89
.
6
CA
-
F
C
N
[3
2
]
91
.
5
CRF
[3
3
]
81
.
8
R
A
R
E
[
2
4
]
88
.
6
R
2
A
M
[3
4
]
90
ED
G
[
27
]
92
.
9
ESI
R
[
29
]
91
.
3
A
S
TER
[
3
0
]
91
.
8
S
c
R
N
[
3
0
]
93
.
9
C
A
S
R
-
D
R
N
e
t
[
5
]
95
.
3
PS
98
.
42
Fig
u
r
e
5
.
Acc
u
r
ac
y
c
o
m
p
ar
is
o
n
f
o
r
I
C
1
3
d
ataset
4
.
2
.
2
.
CUTE
-
80
Fo
r
C
UT
E
8
0
d
ataset
th
e
ac
c
u
r
ac
y
co
m
p
ar
is
o
n
g
r
a
p
h
is
p
lo
tt
ed
,
as
s
h
o
w
n
in
T
ab
le
3
a
n
d
F
ig
u
r
e
6
th
e
lis
t
en
co
m
p
ass
es
v
ar
io
u
s
tex
t
r
ec
o
g
n
itio
n
m
eth
o
d
o
lo
g
ies,
ch
ar
ac
ter
atten
tio
n
f
u
lly
c
o
n
v
o
lu
tio
n
al
n
etwo
r
k
(
C
A
-
FC
N)
r
ep
o
r
ts
an
ac
cu
r
a
cy
o
f
7
9
.
9
%,
s
h
o
wca
s
in
g
c
o
m
m
en
d
ab
le
tex
t
r
ec
o
g
n
itio
n
ca
p
ab
ilit
ies
.
ADC
D
ac
h
iev
es
6
9
.
3
%,
in
d
icatin
g
m
o
d
er
ate
p
e
r
f
o
r
m
an
ce
in
c
o
m
p
ar
is
o
n
.
R
AR
E
ex
h
ib
its
an
ac
cu
r
ac
y
o
f
5
9
.
2
,
im
p
ly
in
g
m
o
r
e
m
o
d
est
tex
t
r
e
co
g
n
itio
n
p
er
f
o
r
m
an
ce
.
PR
N
ex
ce
ls
with
a
s
tr
o
n
g
ac
cu
r
ac
y
o
f
8
8
.
2
,
in
d
icatin
g
p
r
o
f
icien
t
te
x
t
r
ec
o
g
n
itio
n
.
SEE
D
d
em
o
n
s
tr
ates
s
o
lid
tex
t
r
ec
o
g
n
itio
n
with
an
ac
c
u
r
ac
y
o
f
8
3
.
6
,
wh
ile
E
SIR
r
ep
o
r
ts
a
co
m
p
etitiv
e
ac
cu
r
a
cy
o
f
8
3
.
3
.
ASTE
R
ex
h
ib
its
an
ac
c
u
r
ac
y
o
f
7
9
.
5
,
m
ar
k
in
g
a
n
o
tewo
r
th
y
p
er
f
o
r
m
an
ce
.
ScR
N
ac
h
iev
es
a
h
ig
h
ac
c
u
r
ac
y
o
f
8
7
.
5
,
d
e
m
o
n
s
tr
atin
g
s
tr
o
n
g
te
x
t
r
ec
o
g
n
itio
n
ca
p
a
b
ilit
ies
.
C
AS
R
-
DR
Ne
t
lead
s
th
e
lis
t
with
a
r
em
ar
k
ab
le
ac
c
u
r
ac
y
o
f
8
9
.
2
,
s
h
o
wca
s
in
g
s
u
p
er
i
o
r
t
ex
t
r
ec
o
g
n
itio
n
.
PS
s
ig
n
if
ican
tly
o
u
tp
e
r
f
o
r
m
s
th
e
o
th
er
s
with
an
ex
ce
p
tio
n
al
ac
c
u
r
ac
y
s
co
r
e
o
f
9
8
.
96
.
T
ab
le
3
.
R
esu
lts
f
o
r
C
UT
E
8
0
d
ataset
M
e
t
h
o
d
A
c
c
u
r
a
c
y
CA
-
F
C
N
[
3
2
]
79
.
9
A
D
C
D
[
2
5
]
69
.
3
R
A
R
E
[
2
4
]
59
.
2
P
R
N
[
3
5
]
88
.
2
S
EED
[
36
]
83
.
6
ESI
R
[
29
]
83
.
3
A
S
TER
[
3
0
]
79
.
5
S
c
R
N
[
3
0
]
87
.
5
C
A
S
R
-
D
R
N
e
t
[
ES]
[
5
]
89
.
2
PS
98
.
96
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l
.
13
,
No
.
4
,
Dec
em
b
er
2
0
2
4
:
4
9
2
7
-
4
9
3
8
4936
Fig
u
r
e
6
.
Acc
u
r
ac
y
c
o
m
p
ar
is
o
n
f
o
r
C
UT
E
-
8
0
d
ataset
4
.
3
.
Co
m
pa
ra
t
iv
e
a
na
ly
s
is
I
n
th
e
co
n
tex
t
o
f
te
x
t
r
ec
o
g
n
itio
n
ac
r
o
s
s
d
iv
e
r
s
e
d
atas
ets,
a
co
m
p
ar
ativ
e
an
aly
s
is
s
h
o
wca
s
e
s
s
ig
n
if
ican
t
v
ar
iatio
n
s
in
s
y
s
tem
p
er
f
o
r
m
an
ce
an
d
th
e
d
e
g
r
ee
o
f
im
p
r
o
v
em
en
t
ac
h
ie
v
ed
b
y
p
r
o
p
o
s
ed
s
o
lu
tio
n
s
.
Fo
r
th
e
SVTP
d
ataset,
th
e
ex
is
tin
g
s
y
s
tem
r
ep
o
r
ts
an
ac
cu
r
ac
y
o
f
8
5
%,
wh
ile
th
e
p
r
o
p
o
s
ed
s
y
s
tem
s
ig
n
if
ican
tly
en
h
a
n
ce
s
th
is
to
9
5
.
0
4
%,
r
esu
ltin
g
in
a
n
im
p
r
e
s
s
iv
e
1
1
.
8
1
%
im
p
r
o
v
em
e
n
t
.
Mo
v
in
g
to
th
e
I
C
1
3
d
ataset,
th
e
ex
is
tin
g
s
y
s
tem
s
tar
ts
at
9
5
.
3
%,
an
d
th
e
p
r
o
p
o
s
ed
s
y
s
tem
p
u
s
h
es
th
i
s
to
9
8
.
4
2
%,
s
ig
n
if
y
in
g
a
n
o
tewo
r
th
y
3
.
2
6
%
in
cr
ea
s
e
in
ac
cu
r
ac
y
.
I
n
th
e
ca
s
e
o
f
th
e
C
UT
E
-
8
0
d
ataset,
th
e
ex
is
tin
g
s
y
s
tem
s
tan
d
s
at
89
.
2
%,
an
d
th
e
p
r
o
p
o
s
ed
s
y
s
tem
ex
h
ib
its
a
m
o
r
e
m
o
d
est
y
et
v
alu
ab
le
im
p
r
o
v
em
e
n
t
o
f
0
.
8
5
%,
r
ea
ch
in
g
an
ac
cu
r
ac
y
o
f
8
9
.
9
6
%
.
T
h
ese
r
esu
lts
illu
s
tr
ate
th
e
ef
f
ica
cy
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
s
in
en
h
an
cin
g
tex
t
r
ec
o
g
n
itio
n
ac
r
o
s
s
v
ar
io
u
s
d
at
asets
,
with
th
e
SVTP
d
ataset
d
em
o
n
s
tr
atin
g
th
e
m
o
s
t
s
u
b
s
tan
tial
ad
v
an
ce
m
en
t,
f
o
llo
wed
b
y
I
C
1
3
a
n
d
C
UT
E
-
8
0
,
ea
c
h
s
h
o
wca
s
in
g
v
a
r
y
in
g
d
eg
r
ee
s
o
f
p
r
o
g
r
ess
in
tex
t
r
ec
o
g
n
itio
n
ca
p
ab
ilit
ies
.
T
ab
le
4
s
h
o
ws th
e
co
m
p
ar
is
o
n
an
aly
s
is
.
T
ab
le
4
.
C
o
m
p
a
r
is
o
n
an
aly
s
is
D
a
t
a
s
e
t
Ex
i
s
t
i
n
g
s
y
st
e
m
P
r
o
p
o
se
d
s
y
st
e
m
I
mp
r
o
v
i
s
a
t
i
o
n
i
n
%
S
V
TP
85
95
.
04
11
.
81
I
C
1
3
95
.
3
98
.
42
3
.
26
C
U
TE
8
0
89
.
2
89
.
96
0
.
85
5.
CO
NCLU
SI
O
N
I
n
c
o
n
clu
s
io
n
,
th
is
p
ap
er
in
tr
o
d
u
ce
s
th
e
HARN,
a
p
i
o
n
ee
r
in
g
m
eth
o
d
o
lo
g
y
in
s
ce
n
e
tex
t
r
ec
o
g
n
itio
n
.
HARN
’
s
in
n
o
v
ativ
e
d
esig
n
,
f
ea
tu
r
in
g
th
e
AFS
m
o
d
u
le,
ad
v
an
ce
d
atten
tio
n
m
ec
h
an
is
m
s
,
an
d
a
h
y
b
r
id
ar
ch
itectu
r
e,
s
ig
n
if
ican
tly
en
h
an
ce
s
tex
t
r
ec
o
g
n
itio
n
ac
cu
r
a
cy
an
d
ef
f
icien
cy
in
co
m
p
lex
s
ce
n
e
co
n
tex
ts
.
B
y
ad
d
r
ess
in
g
alig
n
m
e
n
t
ch
alle
n
g
es,
ca
p
tu
r
in
g
b
o
th
lo
ca
l
an
d
g
l
o
b
al
co
n
tex
t
in
f
o
r
m
at
io
n
,
an
d
ef
f
icien
tly
m
o
d
elin
g
lo
ca
l
c
o
n
tex
t
an
d
lo
n
g
-
ter
m
r
elatio
n
s
h
ip
s
,
HARN
p
r
esen
ts
a
v
er
s
atile
s
o
lu
tio
n
f
o
r
a
wid
e
r
an
g
e
o
f
tex
t
r
ec
o
g
n
itio
n
ch
allen
g
es
.
T
h
r
o
u
g
h
c
o
m
p
r
e
h
en
s
iv
e
e
x
p
er
im
en
ts
,
we
h
a
v
e
d
em
o
n
s
tr
ated
HARN
’
s
ef
f
ec
tiv
en
ess
in
im
p
r
o
v
in
g
ac
cu
r
ac
y
,
r
ed
u
cin
g
tr
ain
in
g
tim
e,
an
d
o
p
tim
izin
g
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
h
o
ld
s
p
o
ten
tial
f
o
r
n
u
m
e
r
o
u
s
r
ea
l
-
wo
r
ld
ap
p
licatio
n
s
,
wh
er
e
p
r
ec
i
s
e
an
d
ef
f
icien
t
tex
t
r
ec
o
g
n
itio
n
is
cr
u
cial,
f
r
o
m
d
a
ta
ex
tr
ac
tio
n
in
im
ag
es
to
m
u
l
tili
n
g
u
al
tr
an
s
latio
n
an
d
co
n
te
n
t
an
aly
s
is
.
As
w
e
lo
o
k
ah
ea
d
,
HARN
r
ep
r
esen
ts
a
n
o
tewo
r
th
y
s
tep
f
o
r
war
d
i
n
th
e
f
ield
o
f
s
ce
n
e
te
x
t
r
ec
o
g
n
itio
n
,
o
f
f
er
in
g
an
ef
f
icien
t a
n
d
a
d
ap
tab
le
to
o
l to
ad
d
r
ess
th
e
ev
o
lv
i
n
g
d
e
m
an
d
s
o
f
tex
t u
n
d
er
s
tan
d
in
g
in
d
iv
er
s
e
v
is
u
al
d
ata
.
RE
F
E
R
E
NC
E
S
[
1
]
P
.
D
a
i
,
H
.
Z
h
a
n
g
,
a
n
d
X
.
C
a
o
,
“
S
LO
A
N
:
sc
a
l
e
-
a
d
a
p
t
i
v
e
o
r
i
e
n
t
a
t
i
o
n
a
t
t
e
n
t
i
o
n
n
e
t
w
o
r
k
f
o
r
sce
n
e
t
e
x
t
r
e
c
o
g
n
i
t
i
o
n
,
”
I
EE
E
T
ra
n
s
a
c
t
i
o
n
s
o
n
I
m
a
g
e
Pro
c
e
ss
i
n
g
,
v
o
l
.
3
0
,
p
p
.
1
6
8
7
–
1
7
0
1
,
2
0
2
1
,
d
o
i
:
1
0
.
1
1
0
9
/
TI
P
.
2
0
2
0
.
3
0
4
5
6
0
2
.
[
2
]
A
.
A
.
C
h
a
n
d
i
o
,
M
.
A
si
k
u
z
z
a
ma
n
,
M
.
R
.
P
i
c
k
e
r
i
n
g
,
a
n
d
M
.
Le
g
h
a
r
i
,
“
C
u
r
si
v
e
t
e
x
t
r
e
c
o
g
n
i
t
i
o
n
i
n
n
a
t
u
r
a
l
s
c
e
n
e
i
mag
e
s
u
s
i
n
g
d
e
e
p
c
o
n
v
o
l
u
t
i
o
n
a
l
r
e
c
u
r
r
e
n
t
n
e
u
r
a
l
n
e
t
w
o
r
k
,
”
I
EEE
Ac
c
e
ss
,
v
o
l
.
1
0
,
p
p
.
1
0
0
6
2
–
1
0
0
7
8
,
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
2
.
3
1
4
4
8
4
4
.
[
3
]
O
.
Y
.
Li
n
g
,
L
.
B
.
Th
e
n
g
,
A
.
C
.
W
e
i
y
e
n
,
a
n
d
C
.
M
c
c
a
r
t
h
y
,
“
D
e
v
e
l
o
p
m
e
n
t
o
f
v
e
r
t
i
c
a
l
t
e
x
t
i
n
t
e
r
p
r
e
t
e
r
f
o
r
n
a
t
u
r
a
l
sce
n
e
i
m
a
g
e
s,”
I
EEE
A
c
c
e
ss
,
v
o
l
.
9
,
p
p
.
1
4
4
3
4
1
–
1
4
4
3
5
1
,
2
0
2
1
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
1
.
3
1
2
1
6
0
8
.
[
4
]
R
.
B
a
g
i
,
T
.
D
u
t
t
a
,
a
n
d
H
.
P
.
G
u
p
t
a
,
“
C
l
u
t
t
e
r
e
d
T
e
x
t
S
p
o
t
t
e
r
:
a
n
e
n
d
-
to
-
e
n
d
t
r
a
i
n
a
b
l
e
l
i
g
h
t
-
w
e
i
g
h
t
sce
n
e
t
e
x
t
s
p
o
t
t
e
r
f
o
r
c
l
u
t
t
e
r
e
d
e
n
v
i
r
o
n
m
e
n
t
,
”
I
EE
E
A
c
c
e
ss
,
v
o
l
.
8
,
p
p
.
1
1
1
4
3
3
–
1
1
1
4
4
7
,
2
0
2
0
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
0
.
3
0
0
2
8
0
8
.
[
5
]
M
.
Li
,
B
.
F
u
,
Z
.
Zh
a
n
g
,
a
n
d
Y
.
Q
i
a
o
,
“
C
h
a
r
a
c
t
e
r
-
a
w
a
r
e
sam
p
l
i
n
g
a
n
d
r
e
c
t
i
f
i
c
a
t
i
o
n
f
o
r
sc
e
n
e
t
e
x
t
r
e
c
o
g
n
i
t
i
o
n
,
”
I
E
EE
T
r
a
n
s
a
c
t
i
o
n
s
o
n
M
u
l
t
i
m
e
d
i
a
,
v
o
l
.
2
5
,
p
p
.
6
4
9
–
6
6
1
,
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
T
M
M
.
2
0
2
1
.
3
1
2
9
6
5
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.