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.
14
,
No
.
6
,
Dec
em
b
er
2
0
2
5
,
p
p
.
5
0
3
8
~
5
0
4
8
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
14
.i
6
.
p
p
5
0
3
8
-
5
0
4
8
5038
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
K
a
nna
da
hand
wr
itt
en
numera
l
rec
o
g
nition throug
h
deep
lea
rning
and
o
pti
mized
hy
per
pa
ra
meter
tuning
Uj
wa
la
B
.
S
.
1
,
P
ra
mo
d K
um
a
r
S
.
1
, H
.
R
.
M
a
ha
dev
a
s
wa
my
2
,
Su
m
a
t
hi K
.
1
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
En
g
i
n
e
e
r
i
n
g
,
F
a
c
u
l
t
y
o
f
E
l
e
c
t
r
o
n
i
c
s
a
n
d
C
o
mm
u
n
i
c
a
t
i
o
n
,
J
a
w
a
h
a
r
l
a
l
N
e
h
r
u
N
e
w
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
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
s
i
t
y
,
S
h
i
v
a
mo
g
g
a
,
I
n
d
i
a
2
Te
c
h
n
i
c
a
l
E
d
u
c
a
t
i
o
n
D
i
v
i
s
i
o
n
,
J
a
g
a
d
g
u
r
u
S
r
i
S
h
i
v
a
r
a
t
h
r
e
e
sh
w
a
r
a
U
n
i
v
e
r
s
i
t
y
,
N
o
i
d
a
,
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
Mar
10
,
2
0
2
5
R
ev
is
ed
Sep
12
,
2
0
2
5
Acc
ep
ted
Oct
18
,
2
0
2
5
Th
e
c
las
sifica
ti
o
n
o
f
h
a
n
d
writ
ten
n
u
m
e
ra
ls
is
a
v
it
a
l
a
n
d
c
h
a
ll
e
n
g
i
n
g
tas
k
i
n
d
e
v
e
lo
p
in
g
a
u
to
m
a
ted
sy
ste
m
s,
i
n
c
lu
d
in
g
p
o
sta
l
a
d
d
re
ss
so
rti
n
g
a
n
d
li
c
e
n
se
p
late
re
c
o
g
n
it
i
o
n
.
T
h
e
p
re
se
n
t
stu
d
y
e
lu
c
i
d
a
tes
a
n
e
w
m
e
th
o
d
o
lo
g
y
fo
r
re
c
o
g
n
izin
g
Ka
n
n
a
d
a
h
a
n
d
writt
e
n
n
u
m
e
ra
ls
u
sin
g
d
e
e
p
lea
rn
i
n
g
Re
sN
e
t
a
n
d
VG
G
a
rc
h
it
e
c
tu
re
with
tran
sfe
r
lea
rn
in
g
.
Th
e
c
h
a
ll
e
n
g
e
i
n
Ka
n
n
a
d
a
h
a
n
d
writ
ten
re
c
o
g
n
it
io
n
is
c
o
m
p
li
c
a
ted
stru
c
t
u
ra
l
h
iera
rc
h
y
a
n
d
lar
g
e
v
o
c
a
b
u
lary
.
Th
e
m
a
jo
r
p
ro
b
lem
in
d
e
e
p
n
e
u
ra
l
n
e
two
r
k
s
is
v
a
n
is
h
in
g
g
ra
d
ien
t
,
wh
ich
c
a
n
lea
d
t
o
d
e
g
r
a
d
a
ti
o
n
in
c
h
a
ra
c
ter
re
c
o
g
n
it
i
o
n
,
a
n
d
wa
s
a
d
d
re
ss
e
d
u
si
n
g
o
u
r
n
e
w
m
e
th
o
d
o
lo
g
y
u
si
n
g
Re
sN
e
t
a
rc
h
it
e
c
tu
re
.
We
a
p
p
l
y
th
e
p
ro
p
o
se
d
Re
sN
e
t
m
e
th
o
d
in
v
a
rio
u
s
re
a
l
-
wo
r
ld
a
p
p
li
c
a
ti
o
n
s
a
n
d
c
o
m
p
a
re
it
wit
h
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s
(
CNN
)
a
rc
h
it
e
c
tu
re
,
V
GG
.
Th
e
e
x
p
e
rime
n
t
wa
s
imp
lem
e
n
ted
wit
h
th
e
G
o
o
g
le
Co
la
b
so
ftwa
re
v
e
r
sio
n
o
n
a
se
lf
-
c
re
a
ted
d
a
tas
e
t
,
with
h
a
n
d
wr
it
ten
Ka
n
n
a
d
a
n
u
m
e
ra
ls
fe
d
a
s
t
h
e
in
p
u
t
t
o
th
e
re
c
o
g
n
i
ti
o
n
p
r
o
c
e
ss
.
Ou
r
p
ro
p
o
se
d
m
e
th
o
d
a
c
h
iev
e
d
a
h
ig
h
a
c
c
u
ra
c
y
o
f
9
9
.
2
0
%
o
n
train
i
n
g
sa
m
p
les
a
n
d
a
g
e
n
e
ra
li
z
a
ti
o
n
a
c
c
u
ra
c
y
o
f
9
7
.
5
%
o
n
tes
t
sa
m
p
les
,
in
d
ica
ti
n
g
o
u
r
m
e
th
o
d
'
s
e
ffe
c
ti
v
e
n
e
ss
in
re
c
o
g
n
izin
g
h
a
n
d
writt
e
n
Ka
n
n
a
d
a
n
u
m
e
ra
ls.
K
ey
w
o
r
d
s
:
Dee
p
lear
n
in
g
Han
d
wr
itten
r
ec
o
g
n
itio
n
Kan
n
ad
a
n
u
m
er
als
Op
tim
ized
h
y
p
e
r
p
ar
am
ete
r
s
T
r
an
s
f
er
lear
n
i
n
g
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
:
Ujwala
B
.
S.
Dep
ar
tm
en
t o
f
E
lectr
o
n
ics an
d
C
o
m
m
u
n
icatio
n
E
n
g
in
ee
r
i
n
g
,
Facu
lty
o
f
E
lectr
o
n
ics an
d
C
o
m
m
u
n
icatio
n
J
awa
h
ar
lal
Neh
r
u
New
C
o
lleg
e
o
f
E
n
g
in
ee
r
in
g
Sh
iv
am
o
g
g
a
,
Kar
n
atak
a
State,
I
n
d
ia
E
m
ail:
u
jwalar
av
i2
0
0
4
@
jn
n
ce
.
ac
.
in
1.
I
NT
RO
D
UCT
I
O
N
I
n
m
a
n
y
b
u
s
in
ess
an
d
wo
r
k
f
lo
w
p
latf
o
r
m
s
p
ap
er
f
o
r
m
s
f
r
o
m
d
o
c
u
m
en
t
d
ig
itizatio
n
a
n
d
m
a
n
ag
em
en
t
to
law
en
f
o
r
ce
m
en
t
an
d
s
ec
u
r
ity
s
y
s
tem
s
.
Ma
in
tain
in
g
lar
g
e
am
o
u
n
ts
o
f
p
h
y
s
ical
d
o
cu
m
en
ts
is
ted
io
u
s
to
m
an
ag
e
an
d
s
to
r
e.
Pap
e
r
less
m
an
ag
em
en
t
o
f
d
ata
is
an
id
ea
l
s
o
lu
tio
n
th
at
in
v
o
lv
es
s
ca
n
n
in
g
th
e
d
o
c
u
m
en
t
in
to
an
im
ag
e,
b
u
t
it
ca
lls
f
o
r
a
lo
t
o
f
m
an
u
al
in
ter
v
e
n
tio
n
,
an
d
th
e
p
r
o
ce
s
s
tak
es
a
lo
t
o
f
tim
e.
T
h
e
b
est
s
o
lu
tio
n
is
to
co
n
v
er
t
tex
t
im
a
g
es
in
to
tex
t
d
ata
u
s
in
g
o
p
tical
ch
ar
ac
ter
r
ec
o
g
n
itio
n
(
OC
R
)
tech
n
o
lo
g
y
.
OC
R
tr
an
s
f
o
r
m
s
s
ca
n
n
ed
d
o
cu
m
e
n
ts
in
to
m
ac
h
in
e
-
en
co
d
ed
tex
t.
I
t
is
r
eq
u
ir
ed
to
r
ed
u
ce
th
e
la
b
o
r
co
s
t,
to
en
h
a
n
ce
th
e
ef
f
icien
cy
,
an
d
to
s
av
e
th
e
tim
e.
T
h
is
p
r
o
ce
s
s
is
a
n
ee
d
o
f
th
e
d
ay
f
o
r
all
lan
g
u
a
g
es,
esp
ec
ially
f
o
r
I
n
d
ic
s
cr
ip
ts
,
as
th
er
e
ar
e
a
lo
t
o
f
co
m
p
o
u
n
d
ch
ar
ac
ter
s
th
at
n
e
ed
to
b
e
c
o
n
s
id
er
ed
alo
n
g
wi
th
b
asic
ch
a
r
ac
ter
s
.
I
n
s
p
ite
o
f
tr
e
m
en
d
o
u
s
ad
v
an
c
es
th
at
h
av
e
b
ee
n
in
c
o
r
p
o
r
ate
d
in
r
ec
en
t
y
ea
r
s
in
r
ec
o
g
n
izi
n
g
n
u
m
er
als
wr
itten
b
y
h
an
d
,
th
ey
s
till
n
ee
d
to
b
e
s
o
lv
ed
in
ac
cu
r
ate
r
ec
o
g
n
itio
n
.
Ma
n
y
tr
ad
itio
n
al
tech
n
i
q
u
es
ar
e
u
s
ed
f
o
r
OC
R
;
f
ea
tu
r
e
d
etec
tio
n
tech
n
iq
u
es
lik
e
g
r
ad
ie
n
t
f
ea
tu
r
es
(
So
b
el
o
p
er
at
o
r
)
an
d
d
en
s
ity
f
ea
tu
r
es
wer
e
u
s
ed
[
1
]
.
Ma
ch
in
e
lear
n
i
n
g
tec
h
n
iq
u
es
lik
e
h
is
to
g
r
am
o
f
o
r
ien
te
d
g
r
ad
ien
ts
(
Ho
G
)
f
o
r
ex
tr
ac
tin
g
f
ea
tu
r
es,
K
-
n
ea
r
est
n
eig
h
b
o
r
s
(
KNN
)
,
an
d
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
alg
o
r
ith
m
s
ar
e
u
s
ed
t
o
class
if
y
Dev
an
ag
a
r
i
an
d
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
K
a
n
n
a
d
a
h
a
n
d
w
r
itten
n
u
mera
l reco
g
n
itio
n
th
r
o
u
g
h
d
ee
p
le
a
r
n
in
g
…
(
Ujw
a
la
B
.
S
.
)
5039
B
en
g
ali
ch
ar
ac
ter
s
[
2
]
.
A
wo
r
k
o
n
Kan
n
ad
a
n
u
m
er
al
id
en
tifi
ca
tio
n
is
im
p
lem
en
ted
u
s
in
g
H
OG
d
escr
ip
to
r
s
f
o
r
ex
tr
ac
tin
g
f
ea
t
u
r
es
an
d
SVM
f
o
r
c
ateg
o
r
izatio
n
[
3
]
.
A
n
o
v
el
ap
p
r
o
ac
h
f
o
r
r
ec
o
g
n
izin
g
San
s
k
r
it
ch
ar
ac
ter
s
u
s
in
g
th
e
co
n
v
en
ts
'
d
ee
p
lea
r
n
in
g
a
p
p
r
o
ac
h
as
class
if
ier
s
f
o
r
I
n
d
ic
OC
R
s
is
im
p
lem
e
n
ted
[
4
]
.
Dif
f
er
e
n
t
m
ac
h
in
e
lear
n
in
g
a
p
p
r
o
ac
h
es
h
av
e
b
ee
n
i
n
co
r
p
o
r
ated
f
o
r
h
an
d
wr
itten
d
ig
it
r
ec
o
g
n
itio
n
,
s
u
ch
as
SVM
an
d
KNN.
I
n
th
e
last
f
ew
y
ea
r
s
,
r
esear
ch
er
s
h
av
e
witn
ess
ed
a
d
ee
p
en
in
g
o
f
d
ee
p
lear
n
i
n
g
r
e
s
ea
r
ch
,
th
e
tr
ain
in
g
tim
e
o
f
d
ee
p
lear
n
in
g
n
etwo
r
k
s
h
as
g
r
ea
tly
r
ed
u
ce
d
with
g
r
a
p
h
ics
p
r
o
c
ess
in
g
u
n
its
(
GPUs
)
,
an
d
th
e
ef
f
icien
cy
[
5
]
an
d
ac
cu
r
ac
y
h
a
v
e
b
ee
n
g
r
ea
tly
im
p
r
o
v
e
d
[
6
]
.
T
h
e
o
f
f
i
c
ia
l
l
a
n
g
u
ag
e
u
s
ed
f
o
r
a
d
m
in
i
s
t
r
a
t
iv
e
p
u
r
p
o
s
e
s
i
n
K
a
r
n
a
t
ak
a
s
t
a
t
e
i
s
K
a
n
n
a
d
a
,
w
i
t
h
n
e
ar
l
y
5
6
.
9
m
i
l
l
i
o
n
s
p
e
a
k
e
r
s
wo
r
ld
w
i
d
e
.
T
h
e
K
an
n
ad
a
l
a
n
g
u
a
g
e
i
s
d
e
r
iv
e
d
f
r
o
m
th
e
o
f
f
i
c
i
a
l
K
a
n
n
ad
a
s
c
r
i
p
t.
T
h
e
K
a
n
n
ad
a
s
c
r
i
p
t
co
n
s
i
s
t
s
o
f
a
t
o
t
a
l
o
f
f
o
r
ty
-
n
i
n
e
l
e
t
t
er
s
.
K
a
n
n
ad
a
s
c
r
i
p
t
co
n
s
i
s
t
s
o
f
v
o
w
e
l
s
(
s
w
a
r
a
s
)
a
n
d
c
o
n
s
o
n
an
t
s
.
L
e
t
t
e
r
s
r
ep
r
e
s
en
t
in
g
v
y
a
n
j
an
a
s
an
d
y
o
g
a
v
a
h
a
k
s
a
r
e
c
o
m
b
in
e
d
w
i
t
h
n
o
n
-
i
n
t
er
v
en
i
n
g
v
o
w
e
l
s
t
o
f
o
r
m
d
i
g
r
a
p
h
s
(
k
ag
u
n
i
t
a
s
)
.
Se
v
e
r
a
l
r
e
s
e
a
r
ch
e
r
s
h
av
e
m
ad
e
s
i
g
n
i
f
i
c
an
t
c
o
n
t
r
ib
u
t
io
n
s
to
a
u
to
m
a
t
in
g
O
C
R
.
H
o
w
e
v
er
,
h
a
n
d
w
r
i
t
t
en
ch
a
r
ac
t
e
r
r
ec
o
g
n
i
t
io
n
s
t
i
l
l
n
e
e
d
s
t
o
b
e
c
o
m
p
l
e
t
ed
.
I
t
c
o
n
t
i
n
u
e
s
t
o
b
e
a
p
iv
o
t
a
l
c
h
a
l
l
en
g
e
in
th
e
f
i
e
l
d
o
f
p
a
tt
e
r
n
r
e
co
g
n
i
t
i
o
n
,
a
t
tr
i
b
u
t
e
d
to
t
h
e
e
x
te
n
s
i
v
e
v
o
c
ab
u
l
ar
y
,
c
o
m
p
l
e
x
s
tr
u
c
tu
r
a
l
h
i
e
r
ar
c
h
y
,
an
d
t
h
e
d
i
v
e
r
s
e
ar
r
a
y
o
f
h
a
n
d
w
r
i
t
in
g
s
ty
l
e
s
e
x
h
i
b
i
t
ed
b
y
i
n
d
i
v
id
u
a
l
s
.
H
a
n
d
w
r
i
t
i
n
g
r
e
co
g
n
i
t
i
o
n
,
a
l
s
o
k
n
o
wn
a
s
h
an
d
wr
i
t
i
n
g
O
C
R
o
r
cu
r
s
i
v
e
O
C
R
,
co
n
s
ti
t
u
t
e
s
a
s
u
b
s
e
t
o
f
O
C
R
t
e
ch
n
o
l
o
g
y
f
o
cu
s
e
d
o
n
d
e
c
i
p
h
e
r
in
g
h
a
n
d
w
r
i
t
t
en
t
e
x
t,
en
c
o
m
p
a
s
s
i
n
g
b
o
t
h
m
a
n
u
s
c
r
i
p
t
an
d
cu
r
s
i
v
e
f
o
r
m
s
.
M
an
u
s
c
r
ip
t
-
s
t
y
le
te
x
t
,
c
h
a
r
ac
t
e
r
i
ze
d
b
y
s
ep
a
r
a
te
b
lo
c
k
l
e
t
t
er
s
,
i
s
ty
p
ic
a
l
l
y
e
a
s
ie
r
to
r
e
c
o
g
n
iz
e
c
o
m
p
a
r
e
d
to
cu
r
s
i
v
e
w
r
it
i
n
g
.
H
o
w
e
v
er
,
c
u
r
s
iv
e
h
a
n
d
w
r
i
t
in
g
i
s
d
if
f
i
cu
l
t
a
s
i
t
c
o
n
s
i
s
t
s
o
f
j
o
i
n
ed
ch
ar
a
c
t
e
r
s
.
S
i
n
ce
h
an
d
w
r
i
t
t
en
ch
a
r
a
ct
e
r
r
e
c
o
g
n
i
t
io
n
i
s
v
e
r
y
e
s
s
e
n
t
i
a
l
f
o
r
m
u
l
t
ip
l
e
ap
p
l
ic
a
t
i
o
n
s
,
o
u
r
wo
r
k
w
i
l
l
h
e
lp
to
id
e
n
t
if
y
t
h
e
h
a
n
d
w
r
i
t
t
en
c
h
a
r
ac
t
e
r
.
L
a
n
g
u
a
g
e
s
o
th
e
r
t
h
a
n
w
id
e
l
y
u
s
e
d
l
an
g
u
a
g
e
s
,
l
i
k
e
a
r
e
a
s
p
e
c
i
f
ic
a
n
d
v
er
n
a
c
u
l
a
r
la
n
g
u
ag
e
s
,
p
r
e
s
e
r
v
e
o
u
r
cu
l
t
u
r
e
a
n
d
h
er
it
a
g
e
o
f
v
a
r
io
u
s
co
m
m
u
n
i
t
i
e
s
a
n
d
m
a
in
t
a
in
th
e
l
eg
a
c
y
o
f
h
is
t
o
r
y
,
an
d
cr
e
a
t
e
a
p
o
s
i
t
i
v
e
i
m
p
a
c
t
th
a
t
s
t
r
e
n
g
t
h
en
s
g
lo
b
a
l
i
n
t
er
c
o
n
n
e
c
t
e
d
n
e
s
s
.
Kan
n
ad
a
lan
g
u
ag
e
is
m
o
s
t
co
m
m
o
n
ly
u
s
ed
.
Dr
a
v
id
ian
lan
g
u
ag
es
p
r
o
p
ag
ated
m
ain
ly
in
Kar
n
atak
a,
alo
n
g
with
T
am
il
Nad
u
,
An
d
h
r
a
Pra
d
esh
,
an
d
Ma
h
a
r
ash
tr
a
S
tates
[
7
]
.
Mo
r
e
th
an
5
0
m
illi
o
n
in
d
iv
id
u
als
co
n
tr
ib
u
te
to
th
e
lan
g
u
ag
e
v
itality
,
an
d
it is
wr
itten
u
s
in
g
th
e
Kan
n
ad
a
s
cr
ip
t.
T
h
e
Kan
n
a
d
a
s
cr
ip
t c
o
n
s
is
ts
o
f
a
to
tal
o
f
f
o
r
t
y
-
n
in
e
letter
s
.
Kan
n
ad
a
s
cr
ip
t
co
n
s
is
ts
o
f
v
o
wels
an
d
co
n
s
o
n
an
ts
.
Kan
n
ad
a
ch
ar
ac
ter
r
ec
o
g
n
itio
n
is
ch
allen
g
in
g
b
ec
au
s
e
Kan
n
ad
a
ch
ar
ac
ter
s
h
av
e
a
lar
g
er
ch
ar
ac
ter
s
et
an
d
s
im
ilar
ities
o
f
c
h
ar
ac
ter
s
.
T
h
e
wid
th
o
f
ch
ar
ac
ter
s
v
ar
ies
—
th
e
p
r
esen
ce
o
f
k
a
g
u
n
itas
an
d
o
ttak
s
h
ar
as
lead
s
to
th
e
u
n
ev
en
s
p
ac
in
g
b
etwe
en
ch
ar
ac
ter
s
an
d
s
en
ten
ce
s
.
T
h
e
co
n
v
e
r
s
io
n
o
f
h
a
n
d
wr
itten
Kan
n
ad
a
n
u
m
er
als
o
n
d
o
cu
m
en
ts
in
to
m
ac
h
i
n
e
-
r
ea
d
ab
le
f
o
r
m
is
v
er
y
s
ig
n
if
ic
an
t
in
th
e
elec
tr
o
n
ic
f
ield
.
Ma
n
y
alg
o
r
ith
m
s
an
d
tech
n
iq
u
es
h
av
e
alr
ea
d
y
b
ee
n
d
ev
elo
p
e
d
f
o
r
in
d
iv
i
d
u
al
d
ig
it
id
en
tific
atio
n
.
Fig
u
r
e
1
s
h
o
ws
s
am
p
le
im
ag
es o
f
h
a
n
d
wr
itten
o
n
es.
Fig
u
r
e
1
.
Sam
p
le
im
a
g
es o
f
K
an
n
ad
a
h
an
d
wr
itten
d
ig
its
OC
R
is
an
ess
en
tial
tech
n
o
lo
g
y
f
o
r
r
ec
o
g
n
izin
g
an
d
e
x
tr
ac
tin
g
tex
t
f
r
o
m
im
a
g
e
d
atasets
,
wh
ich
h
as
b
ee
n
u
s
ed
in
v
a
r
io
u
s
d
o
m
ain
s
s
u
ch
as
d
o
cu
m
en
t
d
i
g
itizatio
n
,
id
en
tific
atio
n
,
a
n
d
d
ata
en
tr
y
.
Han
d
wr
itten
OC
R
is
m
o
r
e
ch
allen
g
in
g
th
a
n
m
ac
h
in
e
-
p
r
i
n
ted
OC
R
d
u
e
to
th
e
h
ig
h
d
eg
r
ee
o
f
d
if
f
er
en
ce
an
d
v
ar
iab
ilit
y
in
wr
itin
g
s
ty
les
an
d
in
d
iv
id
u
al
v
a
r
iatio
n
s
.
I
n
th
is
liter
atu
r
e
s
u
r
v
ey
,
we
r
ev
iewe
d
class
ical,
d
ee
p
lear
n
in
g
,
tr
a
n
s
f
er
lear
n
in
g
,
s
p
a
r
s
e
lear
n
in
g
an
d
co
m
b
in
e
d
ap
p
r
o
ac
h
es
f
o
r
r
e
co
g
n
izin
g
h
an
d
wr
itten
Kan
n
a
d
a
n
u
m
er
als
u
s
in
g
OC
R
.
T
h
e
ea
r
lies
t
ap
p
r
o
ac
h
e
s
to
h
an
d
wr
itin
g
r
ec
o
g
n
itio
n
wer
e
b
ased
o
n
r
u
le
-
b
ased
m
eth
o
d
s
an
d
p
atter
n
r
ec
o
g
n
itio
n
tech
n
iq
u
es.
Ver
y
co
m
m
o
n
ly
u
s
ed
p
atter
n
r
ec
o
g
n
itio
n
tech
n
i
q
u
es
f
o
r
h
a
n
d
wr
itin
g
r
ec
o
g
n
itio
n
in
clu
d
e
n
ea
r
est
g
r
ad
ie
n
t
f
ea
tu
r
es,
d
en
s
ity
f
ea
tu
r
es,
n
eig
h
b
o
r
class
if
icatio
n
,
d
ec
is
io
n
tr
ee
s
,
an
d
SVMs
[
8
]
.
T
h
ese
tech
n
iq
u
es
ar
e
b
ased
o
n
m
an
u
ally
c
r
af
ted
f
ea
t
u
r
es
an
d
ar
e
co
n
s
tr
ain
ed
in
th
eir
ca
p
ab
ilit
y
to
ca
p
tu
r
e
co
m
p
lex
f
ea
t
u
r
es in
th
e
d
ata.
Ho
G
f
ea
tu
r
es c
an
r
ec
o
g
n
ize
h
an
d
wr
itten
v
o
wels a
n
d
co
n
s
o
n
an
ts
.
Dee
p
lear
n
in
g
h
as
tr
an
s
f
o
r
m
e
d
th
e
f
ield
o
f
h
a
n
d
wr
itin
g
r
ec
o
g
n
itio
n
in
r
ec
en
t
y
ea
r
s
.
Dee
p
lear
n
i
n
g
-
b
ased
ap
p
r
o
ac
h
es
h
av
e
au
to
m
atic
f
ea
tu
r
e
ex
tr
ac
tio
n
,
wh
i
ch
is
o
n
e
o
f
th
e
s
alien
t
f
ea
tu
r
es,
an
d
ex
tr
ac
t
d
ata
f
r
o
m
it
a
n
d
h
av
e
s
h
o
w
n
f
o
r
m
id
ab
le
p
er
f
o
r
m
an
ce
wh
e
n
co
m
p
a
r
ed
t
o
tr
ad
itio
n
al
p
atter
n
r
ec
o
g
n
itio
n
tech
n
iq
u
es.
C
o
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs)
[
9
]
ar
e
m
o
s
t
ex
ten
s
iv
ely
u
s
ed
in
h
an
d
wr
itin
g
r
ec
o
g
n
itio
n
task
s
b
ec
au
s
e
th
ey
ca
p
tu
r
e
s
p
atial
p
atter
n
s
o
r
f
ea
t
u
r
es
in
th
e
im
ag
es.
T
r
an
s
f
er
lear
n
in
g
f
a
cilitates
th
e
lear
n
in
g
f
r
o
m
o
n
e
e
n
v
ir
o
n
m
en
t
a
n
d
g
en
er
alizin
g
to
o
th
er
b
u
t
r
elate
d
p
r
o
b
lem
,
s
er
v
in
g
as
a
v
al
u
ab
le
tech
n
i
q
u
e
in
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.
14
,
No
.
6
,
Dec
em
b
er
20
25
:
5
0
3
8
-
5
0
4
8
5040
m
ac
h
in
e
lear
n
i
n
g
.
T
r
an
s
f
er
l
ea
r
n
in
g
h
as
s
h
o
wn
p
r
o
m
in
e
n
t
r
esu
lts
in
h
an
d
wr
itin
g
r
ec
o
g
n
itio
n
task
s
,
as
it
allo
ws
u
s
in
g
p
r
e
-
in
itialized
m
o
d
els
with
lar
g
e
c
o
r
p
o
r
a
to
o
p
tim
ize
o
u
tco
m
es
o
n
s
m
aller
d
atasets
.
I
n
tr
an
s
f
er
lear
n
in
g
,
th
e
p
r
e
-
d
ev
elo
p
e
d
m
o
d
el
is
ca
lib
r
ated
o
n
th
e
ta
r
g
et
d
ataset
to
ad
ap
t
to
th
e
s
p
ec
if
ic
task
.
So
m
e
p
o
p
u
l
ar
p
r
e
-
le
ar
n
e
d
m
o
d
els
u
s
ed
f
o
r
tr
an
s
f
er
le
a
r
n
in
g
i
n
h
a
n
d
wr
it
in
g
r
ec
o
g
n
iti
o
n
tas
k
s
in
cl
u
d
e
R
esN
et
[
1
0
]
a
n
d
VGG
[
1
1
]
.
S
p
ar
s
e
le
ar
n
i
n
g
ap
p
r
o
ac
h
es
h
av
e
b
ee
n
m
a
jo
r
ly
u
s
e
d
i
n
i
m
a
g
e
r
ec
o
g
n
iti
o
n
t
ask
s
,
i
n
c
lu
d
i
n
g
h
a
n
d
w
r
itt
e
n
d
i
g
it
r
ec
o
g
n
iti
o
n
.
T
h
ese
a
p
p
r
o
a
c
h
es
ai
m
to
a
cq
u
ir
e
s
p
a
r
s
e
f
ea
t
u
r
e
r
e
p
r
ese
n
t
ati
o
n
o
f
t
h
e
i
n
p
u
t
d
ata
.
On
e
o
f
t
h
e
p
o
p
u
l
ar
s
p
a
r
s
e
l
ea
r
n
in
g
a
lg
o
r
it
h
m
s
is
t
h
e
s
p
ar
s
e
a
u
t
o
e
n
c
o
d
e
r
(
S
AE
)
[
1
2
]
.
A
n
SA
E
is
a
n
u
n
s
u
p
e
r
v
is
e
d
l
ea
r
n
i
n
g
al
g
o
r
i
th
m
th
at
m
i
n
im
iz
es
th
e
d
if
f
e
r
en
ce
b
etw
ee
n
i
n
p
u
t
d
ata
an
d
r
e
co
n
s
t
r
u
ct
ed
o
u
t
p
u
t.
B
y
r
e
c
o
n
s
tr
u
c
ti
n
g
t
h
e
in
p
u
t
,
t
h
e
m
o
d
e
l
le
ar
n
s
a
c
o
m
p
ac
t
a
n
d
s
p
a
r
s
e
f
e
at
u
r
e
r
ep
r
ese
n
ta
ti
o
n
.
A
n
o
t
h
e
r
p
o
p
u
la
r
ap
p
r
o
ac
h
f
o
r
h
a
n
d
wr
itt
en
r
e
c
o
g
n
iti
o
n
o
f
t
h
e
ch
ar
ac
te
r
is
t
h
e
co
m
b
i
n
e
d
a
p
p
r
o
ac
h
,
w
h
i
c
h
c
o
m
b
i
n
es
m
u
lt
ip
le
f
e
at
u
r
e
e
x
t
r
a
cti
o
n
m
et
h
o
d
s
o
r
m
o
d
els
t
o
b
o
o
s
t
th
e
r
ec
o
g
n
iti
o
n
p
e
r
f
o
r
m
an
ce
.
F
o
r
e
x
a
m
p
le
,
Mo
n
ta
ze
r
i
et
a
l
.
[
1
3
]
o
u
tlin
ed
a
tech
n
iq
u
e
m
er
g
i
n
g
d
ee
p
lear
n
in
g
an
d
d
ictio
n
ar
y
lea
r
n
in
g
tec
h
n
iq
u
es
to
an
aly
ze
E
n
g
lis
h
h
an
d
wr
itten
d
ig
its
.
T
h
r
o
u
g
h
a
m
u
lti
-
lay
er
ed
f
r
am
ewo
r
k
,
th
e
y
attain
ed
an
ac
cu
r
ac
y
r
ate
ex
ce
ed
in
g
9
9
%.
T
h
eir
in
n
o
v
ativ
e
ap
p
r
o
ac
h
in
teg
r
ate
s
s
in
g
u
lar
v
alu
e
d
ec
o
m
p
o
s
itio
n
with
in
a
m
u
lti
-
lay
er
C
NN
m
o
d
el.
I
n
th
is
p
ap
er
,
o
n
e
o
f
th
e
m
o
s
t
p
o
p
u
lar
an
d
wid
ely
u
s
ed
m
o
d
els,
R
esNet,
is
p
lay
in
g
a
k
e
y
f
ac
to
r
in
s
o
lv
in
g
th
e
p
r
o
b
lem
o
f
v
an
is
h
in
g
g
r
ad
ien
t/ex
p
l
o
d
in
g
g
r
ad
ien
t,
wh
ic
h
is
m
o
s
t
co
m
m
o
n
in
C
NN
-
b
ased
ar
c
h
itectu
r
es.
T
h
e
m
o
s
t
ch
allen
g
in
g
p
ar
t is to
h
a
n
d
le
Kan
n
ad
a
n
u
m
er
als d
u
e
to
s
lan
ts
,
s
tr
o
k
es,
cu
r
v
es,
an
d
lo
o
p
s
.
2.
M
E
T
H
O
D
As
s
h
o
wn
in
F
ig
u
r
e
2
,
o
u
r
n
o
v
el
ar
ch
itectu
r
e
co
n
tain
s
two
p
h
ases
:
tr
ain
in
g
an
d
test
in
g
.
T
h
e
tr
ain
in
g
p
ar
t
h
as
its
f
ir
s
t
s
tep
as p
r
e
-
p
r
o
ce
s
s
in
g
th
e
d
ata,
wh
ich
later
b
u
ild
s
th
e
n
etwo
r
k
ar
ch
itectu
r
e.
T
h
e
d
is
tin
ct
s
tep
s
in
er
ec
tin
g
th
e
n
etwo
r
k
ar
ch
ite
ctu
r
e
ar
e
:
‒
Sp
lit d
ata
s
et
(
Kan
n
ad
a
h
a
n
d
w
r
itten
im
ag
es)
in
to
tr
ain
i
n
g
,
v
a
lid
atio
n
,
an
d
test
in
g
d
atasets
.
‒
Select
lo
s
s
f
u
n
ctio
n
an
d
o
p
tim
izatio
n
alg
o
r
ith
m
(
e.
g
.
,
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
,
a
d
ap
tiv
e
m
o
m
en
t
esti
m
atio
n
(
Ad
am
),
an
d
o
p
tim
izer
)
.
‒
T
r
ain
th
e
m
o
d
el
b
y
ad
ju
s
tin
g
h
y
p
er
p
ar
am
eter
s
as n
ee
d
e
d
.
Fig
u
r
e
2
.
T
h
e
p
r
o
p
o
s
ed
a
r
ch
it
ec
tu
r
e
2
.
1
.
P
re
-
pro
ce
s
s
ing
Her
e
,
th
e
d
ataset
was
clea
n
ed
an
d
n
o
r
m
alize
d
(
e.
g
.
,
r
esizin
g
im
ag
es,
co
n
v
e
r
tin
g
to
g
r
ay
s
ca
le)
,
an
d
au
g
m
en
tatio
n
tech
n
i
q
u
es
(
e
.
g
.
,
r
o
tatio
n
,
f
lip
p
in
g
)
wer
e
ap
p
lied
to
en
h
a
n
ce
m
o
d
el
r
o
b
u
s
tn
ess
.
R
esca
lin
g
is
u
s
ed
to
r
e
d
u
ce
th
e
tar
g
et
v
alu
es.
T
h
e
o
r
ig
in
al
im
ag
es
a
r
e
r
ep
r
esen
ted
u
s
in
g
R
GB
co
ef
f
i
cien
ts
r
an
g
in
g
f
r
o
m
0
to
2
5
5
;
s
u
ch
v
alu
es
wo
u
ld
b
e
to
o
h
ig
h
f
o
r
p
r
o
ce
s
s
in
g
,
s
o
we
u
s
e
s
ca
lin
g
to
co
n
v
er
t
p
ix
e
l
in
ten
s
ities
to
(
0
,
1)
b
y
m
u
ltip
ly
in
g
with
1
/
2
5
5
[
1
4
]
.
T
h
is
tech
n
iq
u
e
is
ca
lled
n
o
r
m
aliza
tio
n
.
So
m
e
o
f
th
e
p
r
e
-
p
r
o
ce
s
s
in
g
tech
n
iq
u
es u
s
ed
h
er
e
ar
e
s
h
ea
r
r
an
g
e,
z
o
o
m
r
an
g
e
,
an
d
h
o
r
iz
o
n
tal
f
lip
.
2
.
2
.
Resid
ua
l
net
wo
r
k
Dee
p
n
eu
r
al
n
etwo
r
k
s
(
DNN)
ar
e
p
r
ed
o
m
in
an
tly
u
s
ed
to
e
x
tr
ac
t
f
ea
tu
r
es
a
n
d
class
if
y
im
a
g
es
b
ased
o
n
th
e
lear
n
in
g
o
f
m
u
ltip
le
le
v
els
o
f
r
ep
r
esen
tatio
n
s
o
f
th
e
d
ata
o
r
f
ea
tu
r
es.
Ho
wev
e
r
,
D
NN
h
as
s
o
m
e
is
s
u
es
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
K
a
n
n
a
d
a
h
a
n
d
w
r
itten
n
u
mera
l reco
g
n
itio
n
th
r
o
u
g
h
d
ee
p
le
a
r
n
in
g
…
(
Ujw
a
la
B
.
S
.
)
5041
th
at
m
ak
e
it
d
i
f
f
icu
lt
to
wo
r
k
s
atis
f
ac
to
r
ily
in
all
s
ce
n
ar
i
o
s
.
On
e
o
f
th
e
d
r
awb
ac
k
s
in
tr
ain
in
g
DNN
is
v
an
is
h
in
g
g
r
ad
ien
ts
.
R
esNet
i
s
o
n
e
o
f
th
e
m
eth
o
d
s
u
s
ed
to
s
o
lv
e
th
e
p
r
o
b
lem
.
I
t
u
s
es
th
e
s
k
ip
co
n
n
ec
tio
n
co
n
ce
p
t
to
o
v
er
co
m
e
t
h
e
p
r
o
b
l
em
.
T
h
e
v
an
is
h
in
g
g
r
ad
ie
n
t
p
r
o
b
lem
ar
is
es
wh
en
g
r
ad
ien
ts
a
r
e
b
ac
k
p
r
o
p
ag
ate
d
to
ea
r
lier
lay
er
s
an
d
b
ec
o
m
e
ex
tr
em
ely
s
m
all,
im
p
ed
i
n
g
ef
f
ec
tiv
e
lear
n
in
g
.
W
h
en
co
m
p
lex
ity
in
cr
ea
s
es
in
DNN
,
tr
ain
in
g
ac
cu
r
ac
y
in
cr
e
ases
,
b
u
t
v
alid
atio
n
ac
cu
r
ac
y
d
ec
r
ea
s
es
b
ec
au
s
e
th
er
e
is
n
o
lear
n
in
g
o
v
er
t
h
e
s
y
s
tem
d
u
e
to
th
e
v
an
is
h
in
g
g
r
ad
ien
t
p
r
o
b
lem
.
T
o
ad
d
r
ess
th
e
ch
allen
g
e
o
f
v
an
is
h
in
g
o
r
e
x
p
lo
d
in
g
g
r
ad
ie
n
ts
,
th
is
ar
ch
itectu
r
e
i
n
tr
o
d
u
ce
d
r
esid
u
al
b
lo
ck
s
,
wh
ich
e
m
p
lo
y
s
k
ip
c
o
n
n
ec
tio
n
s
.
T
h
ese
c
o
n
n
ec
tio
n
s
lin
k
t
h
e
lay
er
’
s
ac
tiv
atio
n
s
p
r
o
p
a
g
ated
to
s
u
b
s
eq
u
e
n
t
lay
er
s
b
y
b
y
p
a
s
s
in
g
ce
r
tain
in
ter
m
ed
iate
lay
e
r
s
,
th
er
eb
y
f
o
r
m
in
g
a
r
esid
u
al
b
lo
ck
.
R
esNets
ar
e
co
n
s
tr
u
cted
b
y
s
tack
in
g
th
ese
r
esid
u
al
b
lo
ck
s
.
T
h
e
co
n
ce
p
t o
f
s
k
ip
co
n
n
ec
tio
n
is
to
ad
d
th
e
in
p
u
t
o
f
b
lo
ck
to
t
h
e
o
u
tp
u
t
o
f
t
h
e
co
n
v
o
l
u
tio
n
b
l
o
ck
.
T
h
e
R
esNet
is
co
m
p
r
is
ed
o
f
r
esid
u
al
b
lo
ck
s
.
Fig
u
r
e
3
illu
s
tr
ates
th
e
s
tr
u
ctu
r
e
o
f
th
ese
r
esid
u
al
b
lo
ck
s
,
s
h
o
wca
s
in
g
ce
r
tain
lay
er
s
co
n
n
ec
ted
v
ia
s
k
ip
co
n
n
ec
tio
n
s
.
C
o
n
s
id
er
a
n
e
u
r
a
l n
etwo
r
k
th
at
h
as x
as th
e
in
p
u
t a
n
d
Y(
x
)
as th
e
o
u
t
p
u
t o
f
th
e
b
lo
ck
.
T
h
e
o
u
tp
u
t
Y(
x
)
is
d
ef
in
ed
in
(
1
)
.
(
)
=
(
)
+
(
1
)
F(x
)
is
th
e
r
esid
u
al
f
u
n
ctio
n
th
at
lear
n
s
th
e
r
esid
u
al
m
ap
p
in
g
b
etwe
en
th
e
in
p
u
t
x
an
d
th
e
o
u
tp
u
t
Y(
x
)
.
F(x
)
is
r
ep
r
esen
ted
b
y
c
o
n
v
o
lu
ti
o
n
al
lay
er
s
co
u
p
le
d
with
n
o
r
m
aliza
tio
n
an
d
ac
tiv
a
tio
n
f
u
n
ctio
n
s
,
s
u
ch
as
b
atch
n
o
r
m
aliza
tio
n
an
d
r
e
ctif
ied
lin
ea
r
u
n
it
(
R
eL
U
)
.
T
h
e
m
ain
l
o
g
ic
b
eh
in
d
r
esid
u
al
n
etwo
r
k
s
is
to
m
ak
e
F(x
)
=0
,
b
ec
au
s
e
th
e
tar
g
et
is
t
o
m
ak
e
Y(
x
)
=
x
.
T
h
e
id
ea
h
er
e
is
to
ad
d
th
e
ac
tu
al
in
p
u
t
x
to
y
o
u
r
lo
s
s
f
u
n
ctio
n
F(x
)
an
d
tr
y
t
o
m
a
k
e
th
e
lo
s
s
f
u
n
ctio
n
ze
r
o
.
I
n
m
a
n
y
C
NN
ar
ch
itectu
r
es,
x
is
th
e
in
p
u
t,
Y(
x
)
is
th
e
o
u
tp
u
t,
a
n
d
Y=
F(x
)
.
Her
e,
th
e
al
g
o
r
ith
m
l
ea
r
n
s
o
r
is
tr
ain
ed
f
r
o
m
th
e
v
a
lu
e
o
f
Y.
B
u
t
in
R
esNet
,
th
e
al
g
o
r
ith
m
lear
n
s
o
r
is
tr
ain
ed
f
r
o
m
th
e
v
alu
e
o
f
F(x
)
.
T
h
is
o
b
s
er
v
atio
n
lead
s
u
s
to
r
ec
o
g
n
ize
th
at
t
h
e
lay
er
s
with
in
th
e
r
esid
u
al
b
l
o
ck
aim
to
lear
n
th
is
r
esid
u
al
f
u
n
ct
io
n
,
wh
ich
i
n
tu
r
n
y
ield
s
an
(
2
)
.
(
)
=
(
)
−
x
(
2
)
W
ith
th
e
in
co
r
p
o
r
atio
n
o
f
s
k
ip
co
n
n
ec
tio
n
s
,
t
h
e
o
r
ig
in
a
l
f
u
n
ctio
n
tr
an
s
f
o
r
m
s
in
to
F(x
)
+x
as
illu
s
tr
ated
in
Fig
u
r
e
3
.
T
h
ese
s
k
ip
co
n
n
ec
tio
n
s
f
ac
ilit
ate
th
e
tr
an
s
m
is
s
io
n
o
f
lar
g
e
r
g
r
ad
ie
n
ts
to
ea
r
lier
lay
er
s
,
en
ab
lin
g
th
em
to
lear
n
at
a
co
m
p
ar
ab
le
p
ac
e
to
th
e
f
in
al
lay
er
s
.
C
o
n
s
eq
u
en
tly
,
th
is
em
p
o
wer
s
th
e
tr
ain
in
g
o
f
DNN
,
ef
f
ec
tiv
ely
a
d
d
r
ess
in
g
th
e
is
s
u
e
o
f
v
an
is
h
in
g
g
r
ad
ie
n
ts
[
1
5
]
.
R
esNet
co
m
p
r
is
es
t
wo
b
lo
c
k
ty
p
es:
th
e
"
id
en
tity
b
lo
ck
"
(
s
tan
d
ar
d
R
esNet)
an
d
th
e
"
co
n
v
o
l
u
tio
n
b
lo
ck
"
(
m
o
d
if
ie
d
R
esNet)
.
An
"id
en
tity
b
lo
c
k
"
f
ea
tu
r
es
a
s
k
ip
p
ed
co
n
n
ec
tio
n
r
u
n
n
in
g
p
ar
allel
to
th
e
m
ain
s
eq
u
en
tial
co
n
n
ec
tio
n
p
at
h
,
m
ain
tain
in
g
id
en
tical
d
im
en
s
io
n
s
at
th
e
s
p
litt
in
g
an
d
m
er
g
in
g
n
o
d
es.
C
o
n
v
e
r
s
ely
,
a
co
n
v
o
lu
tio
n
b
lo
ck
with
in
th
e
s
k
ip
p
ath
s
clo
s
ely
r
esem
b
les th
e
id
en
tity
b
lo
c
k
b
u
t in
teg
r
ates a
co
n
v
o
lu
tio
n
lay
er
.
R
esNet5
0
co
n
s
titu
tes
a
m
o
d
if
ied
v
ar
ian
t
o
f
th
e
R
esNet
a
r
ch
itectu
r
e
th
at
h
as
5
0
lay
er
s
in
to
tal
.
I
t
co
n
s
is
ts
o
f
5
s
tag
es,
ea
ch
co
n
tain
in
g
s
ev
er
al
r
esid
u
al
b
lo
ck
s
.
T
h
e
o
v
e
r
all
s
tr
u
ctu
r
e
o
f
R
esNet5
0
,
s
h
o
wn
in
F
ig
u
r
e
4
,
ca
n
b
e
s
u
m
m
ar
ized
i
n
T
ab
le
1
.
Fig
u
r
e
3
.
Sk
ip
co
n
n
ec
tio
n
o
f
R
esNet
Fig
u
r
e
4
.
R
esNet5
0
ar
ch
itectu
r
e
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.
14
,
No
.
6
,
Dec
em
b
er
20
25
:
5
0
3
8
-
5
0
4
8
5042
T
ab
le
1
.
Ar
c
h
itectu
r
e
d
etails o
f
R
esNet5
0
La
y
e
r
t
y
p
e
D
e
t
a
i
l
s
I
n
p
u
t
l
a
y
e
r
C
o
n
v
o
l
u
t
i
o
n
l
a
y
e
r
s
M
a
x
p
o
o
l
i
n
g
l
a
y
e
r
s
R
e
si
d
u
a
l
b
l
o
c
k
s
F
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
s
G
l
o
b
a
l
a
v
e
r
a
g
e
p
o
o
l
i
n
g
F
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
O
u
t
p
u
t
l
a
y
e
r
I
n
p
u
t
i
m
a
g
e
r
e
s
i
z
e
d
a
n
d
n
o
r
ma
l
i
z
e
d
t
o
a
f
i
x
e
d
s
i
z
e
.
7
×
7
c
o
n
v
o
l
u
t
i
o
n
,
s
t
r
i
d
e
=
2
→
r
e
d
u
c
e
s
sp
a
t
i
a
l
r
e
s
o
l
u
t
i
o
n
.
P
o
o
l
s
i
z
e
2
×
2
,
s
t
r
i
d
e
=
2
→
r
e
d
u
c
e
s
s
p
a
t
i
a
l
r
e
s
o
l
u
t
i
o
n
F
o
u
r
s
t
a
g
e
s
o
f
r
e
si
d
u
a
l
b
l
o
c
k
s,
e
a
c
h
w
i
t
h
d
i
f
f
e
r
e
n
t
n
u
m
b
e
r
s
o
f
b
l
o
c
k
s
.
E
a
c
h
b
l
o
c
k
=
mu
l
t
i
p
l
e
c
o
n
v
o
l
u
t
i
o
n
a
l
l
a
y
e
r
s +
b
a
t
c
h
n
o
r
m
a
l
i
z
a
t
i
o
n
+
R
e
LU
a
c
t
i
v
a
t
i
o
n
+
s
k
i
p
c
o
n
n
e
c
t
i
o
n
R
e
d
u
c
e
s
e
a
c
h
f
e
a
t
u
r
e
m
a
p
t
o
a
si
n
g
l
e
v
a
l
u
e
→
o
u
t
p
u
t
s
a
c
o
mp
a
c
t
v
e
c
t
o
r
.
D
e
n
se
l
a
y
e
r
f
o
r
c
l
a
ssi
f
i
c
a
t
i
o
n
i
n
t
o
c
a
t
e
g
o
r
i
e
s
S
o
f
t
M
a
x
a
c
t
i
v
a
t
i
o
n
,
a
ss
i
g
n
s
p
r
o
b
a
b
i
l
i
t
i
e
s
a
c
r
o
ss c
a
t
e
g
o
r
i
e
s.
2
.
3
.
VG
G
1
6
a
rc
hite
ct
ure
VGG
r
ep
r
esen
ts
a
d
ee
p
co
n
v
o
lu
tio
n
al
f
ea
tu
r
e
lear
n
in
g
n
etwo
r
k
ar
c
h
itectu
r
e
u
s
ed
f
o
r
im
ag
e
class
if
icatio
n
,
p
r
o
p
o
s
ed
b
y
th
e
v
is
u
al
g
e
o
m
etr
y
g
r
o
u
p
at
t
h
e
Un
iv
er
s
ity
o
f
Ox
f
o
r
d
in
2
0
1
4
.
T
h
e
VGG
n
etwo
r
k
is
n
o
tab
le
f
o
r
its
s
im
p
licity
a
n
d
th
e
u
s
e
o
f
s
m
all
c
o
n
v
o
l
u
tio
n
al
f
ilter
s
(
3
×
3
)
th
r
o
u
g
h
o
u
t
th
e
ar
ch
itectu
r
e
,
as
s
h
o
wn
in
F
ig
u
r
e
5
.
T
h
e
o
r
ig
i
n
al
VGG
ar
ch
itectu
r
e
in
clu
d
es
1
6
o
r
1
9
lay
er
s
a
n
d
is
n
a
m
ed
VGG1
6
o
r
VGG1
9
,
r
esp
ec
tiv
ely
.
VGG1
6
co
m
p
r
is
es
1
6
lay
er
s
,
in
clu
d
in
g
1
3
co
n
v
o
lu
tio
n
al
lay
e
r
s
an
d
th
r
ee
f
u
l
ly
co
n
n
ec
te
d
lay
er
s
[
1
6
]
.
T
h
e
g
e
n
er
al
ar
c
h
itectu
r
e
o
f
th
e
VGG1
6
ca
n
b
e
s
u
m
m
ar
ized
in
T
ab
le
2
.
Fig
u
r
e
5
.
VGG1
6
m
o
d
el
T
ab
le
2
.
Ar
c
h
itectu
r
e
d
etails o
f
VGG1
6
La
y
e
r
t
y
p
e
D
e
t
a
i
l
s
I
n
p
u
t
l
a
y
e
r
C
o
n
v
o
l
u
t
i
o
n
l
a
y
e
r
s
M
a
x
p
o
o
l
i
n
g
l
a
y
e
r
s
F
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
s
O
u
t
p
u
t
l
a
y
e
r
I
n
p
u
t
i
m
a
g
e
r
e
s
i
z
e
d
t
o
f
i
x
e
d
s
i
z
e
,
n
o
r
mal
i
z
e
d
1
3
c
o
n
v
l
a
y
e
r
s
,
f
i
l
t
e
r
si
z
e
3
×
3
,
st
r
i
d
e
=
1
,
e
a
c
h
f
o
l
l
o
w
e
d
b
y
R
e
LU
a
c
t
i
v
a
t
i
o
n
a
n
d
B
a
t
c
h
n
o
r
ma
l
i
z
a
t
i
o
n
.
D
e
p
t
h
o
f
f
e
a
t
u
r
e
ma
p
s i
n
c
r
e
a
s
e
s wi
t
h
l
a
y
e
r
s
P
o
o
l
s
i
z
e
2
×
2
,
s
t
r
i
d
e
=
2
,
r
e
d
u
c
e
s s
p
a
t
i
a
l
r
e
s
o
l
u
t
i
o
n
o
f
f
e
a
t
u
r
e
ma
p
s.
F
l
a
t
t
e
n
e
d
f
e
a
t
u
r
e
ma
p
s →
3
d
e
n
se
l
a
y
e
r
s,
e
a
c
h
w
i
t
h
4
0
9
6
n
e
u
r
o
n
s,
R
e
LU
a
c
t
i
v
a
t
i
o
n
,
a
n
d
B
a
t
c
h
N
o
r
m
a
l
i
z
a
t
i
o
n
S
o
f
t
M
a
x
a
c
t
i
v
a
t
i
o
n
,
o
u
t
p
u
t
s
c
l
a
ss
p
r
o
b
a
b
i
l
i
t
i
e
s
2
.
4
.
O
pti
m
izer
s
W
ith
th
e
o
b
jectiv
e
o
f
o
b
tain
i
n
g
o
p
tim
al
s
o
lu
tio
n
a
n
d
t
o
r
e
d
u
ce
th
e
lo
s
s
es
in
f
r
o
n
t
an
d
b
ac
k
p
ass
es,
o
p
tim
ize
r
s
[
1
7
]
g
u
id
e
th
e
m
o
d
el
to
u
p
d
ate
m
o
d
el
-
p
ar
am
ete
r
s
o
f
n
eu
r
al
n
etwo
r
k
s
b
y
m
o
v
in
g
m
o
d
el
weig
h
ts
u
s
in
g
th
e
g
r
ad
ien
ts
th
at
m
in
i
m
ize
th
e
lo
s
s
,
m
o
d
u
late
b
ias
-
v
ar
ian
ce
tr
a
d
eo
f
f
,
an
d
co
n
t
r
o
l
s
tep
s
ize
in
ea
ch
iter
atio
n
to
o
b
tain
o
p
tim
al
lea
r
n
in
g
r
ates.
Op
tim
izer
s
ar
e
th
e
b
ac
k
b
o
n
e
f
o
r
an
y
ty
p
e
o
f
r
e
co
g
n
itio
n
.
W
ith
o
u
t
p
r
o
p
er
o
p
tim
izatio
n
,
th
e
m
o
d
e
l m
ay
tak
e
to
o
lo
n
g
to
lear
n
p
a
tter
n
s
o
r
f
ail
to
c
o
n
v
er
g
e.
2
.
4
.
1
.
G
ra
dient
des
ce
nt
T
o
m
in
im
ize
p
r
e
d
ictio
n
er
r
o
r
o
r
co
s
t
f
u
n
ctio
n
s
in
th
e
m
o
d
e
l,
a
f
u
n
d
am
e
n
tal
o
p
tim
izatio
n
tech
n
iq
u
e
ca
lled
g
r
ad
ien
t
d
escen
t
[
1
8
]
is
u
s
ed
.
First
,
it
s
tar
t
s
with
an
in
itial
g
u
ess
o
f
th
e
p
ar
am
eter
s
an
d
s
tag
e
b
y
s
tag
e
it
o
p
tim
izes
th
e
p
a
r
am
eter
s
in
t
o
war
d
s
th
e
s
teep
est
d
ec
r
e
a
s
e
o
f
th
e
co
s
t
f
u
n
ctio
n
.
T
h
is
h
ap
p
e
n
s
with
th
e
h
el
p
o
f
lear
n
in
g
r
ate
u
n
til
th
e
m
in
im
um
is
r
ea
ch
ed
.
I
t is d
escr
ib
e
d
in
(
3
)
.
=
−
(
3
)
2
.
4
.
2
.
Ada
ptiv
e
g
ra
dient
des
ce
nt
T
h
is
o
p
tim
izer
d
y
n
am
ically
c
h
an
g
es
th
e
lear
n
in
g
r
ate
‘
η
’
a
t
ev
er
y
tim
e
s
tep
‘
t’
b
ased
o
n
g
r
ad
ien
t
in
f
o
r
m
atio
n
.
T
h
is
h
elp
s
in
f
ast co
n
v
er
g
e
n
ce
an
d
im
p
r
o
v
es st
ab
ilit
y
.
T
h
e
p
r
in
cip
l
e
o
f
a
d
a
p
ti
v
e
g
r
ad
ien
t d
escen
t
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
K
a
n
n
a
d
a
h
a
n
d
w
r
itten
n
u
mera
l reco
g
n
itio
n
th
r
o
u
g
h
d
ee
p
le
a
r
n
in
g
…
(
Ujw
a
la
B
.
S
.
)
5043
(
Ad
ag
r
ad
)
is
th
at
it
wo
r
k
s
o
n
cu
m
u
lativ
e
s
u
m
o
f
s
q
u
ar
e
d
g
r
ad
ien
ts
.
T
h
is
o
p
tim
izatio
n
is
v
er
y
m
u
ch
s
u
itab
le
f
o
r
s
p
ar
s
e
d
ata
[
1
9
]
.
an
d
th
is
is
s
h
o
wn
f
r
o
m
(
4
)
t
o
(
6
)
.
=
−
1
−
|
−
1
(
4
)
|
=
√
+
(
5
)
=
∑
(
)
2
=
1
(
6
)
2
.
4
.
3
.
Ro
o
t
m
ea
n sq
ua
re
pro
pa
g
a
t
io
n
R
o
o
t
m
ea
n
s
q
u
a
r
e
p
r
o
p
a
g
atio
n
(
R
MSp
r
o
p
)
[
2
0
]
is
a
p
r
ac
tic
al
an
d
r
o
b
u
s
t
o
p
tim
izatio
n
alg
o
r
ith
m
th
at
tr
ies
to
im
p
r
o
v
e
A
d
a
g
r
a
d
.
I
t
t
ak
es
th
e
‘
ex
p
o
n
en
tial
d
ec
a
y
in
g
m
o
v
i
n
g
av
e
r
ag
e’
f
o
r
ea
c
h
p
ar
am
eter
.
Mo
v
in
g
av
er
ag
e
co
n
ce
p
t
p
r
ev
en
ts
er
r
atic
m
o
v
em
e
n
ts
,
u
s
es
s
im
p
le
h
y
p
er
p
ar
am
eter
s
a
n
d
p
r
o
m
o
tes
s
tab
ilit
y
.
T
h
e
f
u
n
ctio
n
ality
is
d
escr
ib
e
d
f
r
o
m
(
7
)
t
o
(
1
0
)
.
=
+
(
1
−
)
2
(
7
)
=
+
(
1
−
)
2
(
8
)
=
0
−
[
√
+
]
(
9
)
=
0
−
[
√
+
]
(
1
0
)
2
.
4
.
4
.
Ada
ptiv
e
mo
m
ent
esti
m
a
t
io
n
I
t
wo
r
k
s
with
ex
p
o
n
en
tial
d
ec
ay
in
g
av
er
a
g
es
an
d
p
ast
s
q
u
ar
ed
g
r
ad
ien
ts
.
T
h
e
a
d
ap
tiv
e
le
a
r
n
in
g
r
ate,
in
teg
r
atio
n
o
f
th
e
m
o
m
en
tu
m
,
an
d
r
eliab
ilit
y
ac
r
o
s
s
ar
ch
itectu
r
es
h
as
m
ad
e
Ad
am
a
d
ef
a
u
lt
o
p
tim
izer
[
2
1
]
.
Ad
am
is
th
e
co
m
b
in
atio
n
o
f
R
MSPr
o
p
an
d
m
o
m
e
n
tu
m
s
h
o
wn
f
r
o
m
(
1
1
)
to
(
1
8
)
.
T
h
e
d
if
f
e
r
en
t
h
y
p
er
p
ar
am
eter
s
ar
e
K,
ξ,
β1
,
an
d
β2
.
=
0
,
=
0
=
0
=
0
(
1
1
)
=
1
+
(
1
−
1
)
2
(
1
2
)
=
1
+
(
1
−
1
)
2
(
1
3
)
=
2
+
(
1
−
2
)
(
1
4
)
=
2
+
(
1
−
2
)
(
1
5
)
|
=
1
−
1
|
=
1
−
1
(
1
6
)
|
=
1
−
2
|
=
1
−
2
(
1
7
)
=
0
−
[
|
√
|
+
]
=
0
−
[
|
√
|
+
]
(
1
8
)
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
Acc
ura
cy
a
nd
lo
s
s
co
mp
a
riso
n
T
h
e
ex
p
er
im
e
n
t
is
im
p
lem
en
ted
u
s
in
g
Go
o
g
le
C
o
lab
s
o
f
twa
r
e
v
er
s
io
n
;
th
e
p
r
o
p
o
s
ed
R
esNet
m
eth
o
d
r
ec
eiv
es
in
p
u
t
as
th
e
Kan
n
a
d
a
h
an
d
wr
itten
n
u
m
e
r
als
(
0
t
o
9
)
d
o
c
u
m
en
t
im
a
g
e
b
y
tak
in
g
i
m
ag
es
o
f
d
if
f
er
e
n
t
u
s
er
s
in
v
a
r
io
u
s
s
ty
les.
I
n
th
e
s
elf
-
d
ataset,
a
to
tal
o
f
2
5
0
0
im
ag
es
ar
e
tak
en
:
2
0
0
im
a
g
es
o
f
ea
c
h
n
u
m
er
al
f
o
r
tr
ain
in
g
,
2
5
f
o
r
v
alid
atio
n
,
an
d
2
5
f
o
r
test
in
g
in
th
e
r
atio
8
0
:1
0
:1
0
.
Han
d
wr
itten
ch
ar
ac
t
er
im
ag
es
f
r
o
m
th
e
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.
14
,
No
.
6
,
Dec
em
b
er
20
25
:
5
0
3
8
-
5
0
4
8
5044
d
ataset
wer
e
co
n
v
er
ted
t
o
2
8
×
2
8
im
ag
es.
T
h
e
e
x
p
er
im
e
n
tal
r
esu
lt
o
b
tain
ed
af
ter
all
th
e
ep
o
ch
s
f
o
r
th
e
g
iv
en
in
p
u
t
is
p
ictu
r
ed
b
elo
w.
At
f
ir
s
t
,
th
e
p
r
e
-
p
r
o
ce
s
s
in
g
p
r
o
ce
d
u
r
es
a
r
e
co
m
p
leted
;
af
t
er
th
at,
t
h
e
f
ea
tu
r
e
ex
tr
ac
tio
n
tec
h
n
iq
u
e
is
ap
p
lie
d
to
th
e
p
r
e
-
p
r
o
ce
s
s
ed
im
ag
e.
Af
ter
o
b
s
er
v
in
g
th
e
tr
en
d
o
f
tr
ain
in
g
ac
cu
r
ac
y
,
test
in
g
,
an
d
lo
s
s
v
alu
es,
th
e
n
u
m
b
er
o
f
e
p
o
ch
s
is
co
n
f
ig
u
r
ed
t
o
v
alu
es f
o
r
b
o
th
v
alid
atio
n
an
d
tr
ain
in
g
s
am
p
les
to
o
p
tim
ize
th
e
tr
ad
eo
f
f
b
etwe
en
ac
cu
r
ac
y
an
d
ep
o
ch
.
I
n
th
is
ex
p
er
im
en
t,
th
e
R
esNet5
0
m
o
d
el
is
s
u
cc
es
s
f
u
l in
o
b
tain
in
g
a
n
ac
cu
r
ac
y
o
f
9
8
.
2
0
%
d
u
r
in
g
th
e
t
r
ain
in
g
p
h
ase,
an
ac
cu
r
ac
y
o
n
th
e
v
alid
atio
n
s
et
is
9
7
.
5
%,
an
d
u
n
s
ee
n
test
d
ata
is
o
f
9
7
.
2
%,
as
s
h
o
wn
in
F
ig
u
r
es
6
a
n
d
7
,
an
d
th
e
VGG1
6
m
o
d
el
is
s
u
c
ce
s
s
f
u
l
in
o
b
tain
in
g
ev
alu
atio
n
ac
c
u
r
ac
y
o
v
er
t
h
e
tr
ain
in
g
s
am
p
les
is
o
f
9
8
.
1
0
%,
a
r
ec
o
g
n
itio
n
r
ate
o
n
v
ali
d
atio
n
s
am
p
les
is
o
f
9
7
.
3
3
%,
an
d
a
g
en
er
aliza
tio
n
ac
cu
r
ac
y
o
n
t
h
e
test
d
ata
s
et
is
o
f
9
7
%,
as
s
h
o
wn
in
F
ig
u
r
es
8
an
d
9
.
B
y
em
p
lo
y
i
n
g
th
is
m
eth
o
d
o
lo
g
y
,
we
s
u
cc
ess
f
u
lly
ac
h
iev
ed
th
e
b
est
ac
c
u
r
ac
y
an
d
a
lo
w
l
o
s
s
r
ate.
B
u
t
d
u
e
to
th
e
s
im
ilar
ity
b
etwe
en
s
o
m
e
ch
ar
ac
ter
s
,
ac
cu
r
ac
y
is
a
li
ttle
less
.
T
ab
le
3
s
h
o
w
s
th
e
ar
ch
itectu
r
e
a
n
d
p
er
f
o
r
m
an
ce
co
m
p
ar
is
o
n
b
etwe
en
R
esNet5
0
an
d
VGG1
6
.
Fig
u
r
e
6
.
T
r
ain
in
g
a
n
d
v
alid
atio
n
ep
o
ch
wis
e
ac
cu
r
ac
y
g
r
ap
h
f
o
r
R
esNet5
0
o
n
th
e
Ka
n
n
ad
a
n
u
m
b
e
r
s
d
ataset
Fig
u
r
e
7
.
T
r
ain
in
g
a
n
d
v
alid
atio
n
lo
s
s
v
is
u
aliza
tio
n
g
r
ap
h
f
o
r
R
esNet5
0
o
n
th
e
Ka
n
n
ad
a
n
u
m
b
e
r
s
d
ataset
Fig
u
r
e
8
.
T
r
ain
in
g
a
n
d
v
alid
atio
n
e
p
o
ch
wis
e
a
cc
u
r
ac
y
g
r
ap
h
f
o
r
VGG1
6
o
n
t
h
e
Kan
n
ad
a
n
u
m
b
er
s
d
ataset
Fig
u
r
e
9
.
T
r
ain
in
g
a
n
d
v
alid
atio
n
lo
s
s
v
is
u
aliza
tio
n
g
r
ap
h
f
o
r
VGG1
6
o
n
t
h
e
Kan
n
ad
a
n
u
m
b
er
s
d
ataset
T
ab
le
3
.
Ar
c
h
itectu
r
e
an
d
p
er
f
o
r
m
an
ce
c
o
m
p
ar
is
o
n
b
etwe
e
n
R
esNet5
0
an
d
VGG1
6
O
p
t
i
mi
z
e
r
s
R
e
sN
e
t
5
0
V
G
G
1
6
To
t
a
l
p
a
r
a
me
t
e
r
s
2
,
5
6
,
9
6
,
1
3
8
1
,
5
2
,
5
0
,
2
5
0
Tr
a
i
n
a
b
l
e
p
a
r
a
m
e
t
e
r
s
N
o
n
-
t
r
a
i
n
a
b
l
e
p
a
r
a
met
e
r
2
1
,
0
8
,
4
2
6
2
,
3
5
,
8
7
,
7
1
2
5
,
3
5
,
5
6
2
1
,
4
7
,
1
4
,
6
8
8
Ep
o
c
h
s
Tr
a
i
n
i
n
g
sam
p
l
e
s
Tr
a
i
n
i
n
g
a
c
c
u
r
a
c
y
V
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
Te
st
i
n
g
a
c
c
u
r
a
c
y
O
p
t
i
mi
z
e
r
15
2
0
0
0
9
8
.
2
0
%
9
7
.
5
%
9
7
.
2
%
A
d
a
m
15
2
0
0
0
9
8
.
1
0
%
9
7
.
3
3
%
9
7
%
A
d
a
m
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
K
a
n
n
a
d
a
h
a
n
d
w
r
itten
n
u
mera
l reco
g
n
itio
n
th
r
o
u
g
h
d
ee
p
le
a
r
n
in
g
…
(
Ujw
a
la
B
.
S
.
)
5045
3
.
2
.
O
pti
m
izer
a
na
ly
s
is
Ou
t
o
f
d
if
f
er
e
n
t
d
ee
p
lear
n
i
n
g
p
ar
am
eter
s
,
o
p
tim
izer
s
p
lay
a
cr
u
cial
r
o
le
wh
en
co
n
f
ig
u
r
in
g
a
n
eu
r
al
n
etwo
r
k
[
2
2
]
.
T
h
e
tab
le
p
r
o
v
i
d
es r
esu
lts
o
f
ac
cu
r
ac
y
co
m
p
a
r
is
o
n
f
o
r
1
5
ep
o
ch
s
.
T
h
e
b
est o
p
tim
izer
h
er
e
is
th
e
Ad
am
o
p
tim
izer
[
2
3
]
.
T
ab
le
2
s
h
o
ws
an
ac
cu
r
ac
y
co
m
p
a
r
is
o
n
o
f
R
esNet5
0
an
d
VGG1
6
.
F
ig
u
r
e
1
0
s
h
o
ws
th
e
co
m
p
ar
is
o
n
o
f
ac
cu
r
ac
y
with
d
if
f
er
en
t
o
p
tim
izer
s
.
T
ab
le
4
s
h
o
ws b
atch
s
ize
an
aly
s
is
f
o
r
d
i
f
f
er
en
t o
p
tim
izer
s
.
T
ab
le
4
.
Acc
u
r
ac
y
co
m
p
ar
is
o
n
o
f
R
esNet5
0
an
d
VGG1
6
O
p
t
i
mi
z
e
r
s
R
e
sN
e
t
5
0
V
G
G
1
6
S
G
D
9
4
.
5
2
8
2
.
7
0
A
d
a
g
r
a
d
9
0
.
7
5
6
6
.
4
5
R
M
S
p
r
o
p
A
d
a
m
9
4
.
8
0
9
8
.
2
0
8
2
.
7
5
9
8
.
1
0
Fig
u
r
e
1
0
.
C
o
m
p
ar
is
o
n
o
f
o
p
tim
izer
s
f
o
r
R
esNet5
0
an
d
VG
G1
6
f
o
r
Kan
n
ad
a
n
u
m
e
r
als
3
.
3
.
B
a
t
ch
s
ize
a
na
ly
s
is
B
atc
h
s
iz
e
is
a
n
i
m
p
o
r
t
an
t
h
y
p
er
p
a
r
a
m
e
te
r
t
h
at
af
f
ec
ts
t
h
e
p
e
r
f
o
r
m
a
n
ce
o
f
t
h
e
n
et
wo
r
k
.
I
t
is
d
e
f
i
n
e
d
as
th
e
t
o
tal
n
u
m
b
e
r
o
f
s
am
p
l
es
p
ass
e
d
t
h
r
o
u
g
h
th
e
n
et
wo
r
k
i
n
b
o
t
h
f
o
r
wa
r
d
a
n
d
b
ac
k
w
a
r
d
p
ass
es
[
2
4
]
.
T
a
b
l
e
5
s
h
o
ws
th
e
co
m
p
ar
is
o
n
o
f
ac
cu
r
ac
y
an
d
b
atch
s
ize
f
o
r
b
o
th
R
esNet5
0
an
d
VGG1
6
tr
an
s
f
er
lear
n
in
g
m
eth
o
d
s
.
Fig
u
r
e
1
1
s
h
o
ws th
e
co
m
p
a
r
is
o
n
o
f
b
atch
s
ize
an
aly
s
is
f
o
r
R
esNet5
0
an
d
VGG1
6
f
o
r
Kan
n
ad
a
n
u
m
e
r
als.
T
ab
le
5
.
B
atch
s
ize
co
m
p
a
r
is
o
n
o
f
R
esNet5
0
an
d
VGG1
6
O
p
t
i
mi
z
e
r
A
c
c
u
r
a
c
y
(
%)
R
e
sN
e
t
5
0
B
a
t
c
h
si
z
e
=
3
2
A
c
c
u
r
a
c
y
(
%)
R
e
sN
e
t
5
0
B
a
t
c
h
si
z
e
=
6
4
A
c
c
u
r
a
c
y
(
%)
V
G
G
1
6
B
a
t
c
h
si
z
e
=
3
2
A
c
c
u
r
a
c
y
(
%)
V
G
G
1
6
B
a
t
c
h
si
z
e
=
6
4
S
G
D
9
4
.
5
2
9
0
.
5
8
8
2
.
7
0
7
7
.
1
4
A
d
a
g
r
a
d
9
0
.
7
5
8
5
.
4
1
6
6
.
4
5
5
9
.
0
4
R
M
S
p
r
o
p
A
d
a
m
9
4
.
8
0
9
8
.
2
0
8
9
.
6
9
9
7
.
6
7
8
2
.
7
5
9
8
.
1
0
7
9
.
3
6
8
9
.
7
8
Fig
u
r
e
1
1
.
C
o
m
p
ar
is
o
n
o
f
b
atc
h
s
ize
an
aly
s
is
f
o
r
R
esNet5
0
an
d
VGG1
6
f
o
r
Kan
n
ad
a
n
u
m
e
r
als
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.
14
,
No
.
6
,
Dec
em
b
er
20
25
:
5
0
3
8
-
5
0
4
8
5046
3
.
4
.
B
enchm
a
r
k
ing
t
he
o
bta
ined re
s
u
lt
s
wit
h e
a
rlier
re
s
e
a
r
ch
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
h
as
s
h
o
wn
th
e
b
est
r
esu
lts
co
m
p
ar
ed
with
ea
r
lier
r
esear
ch
wo
r
k
s
.
T
h
e
b
est
tr
ain
i
n
g
ac
c
u
r
ac
y
is
9
8
.
2
0
%,
as
s
h
o
wn
in
T
a
b
le
6
.
T
h
e
u
s
e
o
f
th
e
Go
o
g
le
C
o
lab
lab
o
r
ato
r
y
an
d
th
e
o
p
tim
izatio
n
tech
n
iq
u
es
o
f
t
h
e
h
y
p
e
r
p
ar
am
eter
s
p
r
o
v
id
e
d
ee
p
er
in
s
ig
h
ts
in
lear
n
in
g
a
r
eliab
le
ch
ar
ac
ter
r
ec
o
g
n
itio
n
f
r
am
ewo
r
k
with
r
o
b
u
s
t
,
s
u
p
er
io
r
ac
cu
r
ac
y
an
d
e
f
f
icien
t p
r
o
ce
s
s
in
g
s
p
ee
d
.
T
ab
le
6
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
f
r
o
m
l
iter
atu
r
e
r
e
v
iew
A
u
t
h
o
r
s
F
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
a
r
c
h
i
t
e
c
t
u
r
e
C
l
a
s
si
f
i
e
r
A
c
c
u
r
a
c
y
(
%)
H
u
[
2
5
]
R
e
sN
e
t
5
0
R
e
sN
e
t
5
0
9
7
.
3
8
C
h
a
n
d
r
a
k
a
l
a
a
n
d
Th
i
p
p
e
sw
a
m
y
[
2
6
]
R
e
sN
e
t
5
0
R
e
sN
e
t
5
0
15
D
u
t
t
a
e
t
a
l
.
[
2
7
]
R
e
sN
e
t
5
0
S
V
M
90
P
r
o
p
o
se
d
R
e
sN
e
t
5
0
R
e
sN
e
t
5
0
9
8
.
2
0
4.
CO
NCLU
SI
O
N
T
h
is
m
an
u
s
cr
ip
t
p
r
o
p
o
s
es
a
n
ew
m
eth
o
d
o
lo
g
y
f
o
r
r
ec
o
g
n
izin
g
Kan
n
ad
a
h
an
d
wr
itten
n
u
m
er
als
u
s
in
g
d
ee
p
lear
n
in
g
R
esNet
ar
ch
itectu
r
e
with
tr
an
s
f
er
lear
n
in
g
.
T
h
e
p
r
o
p
o
s
ed
R
esNet
an
d
V
GG
m
eth
o
d
o
lo
g
ies
wer
e
ev
alu
ated
u
s
in
g
p
er
f
o
r
m
an
ce
p
ar
am
eter
s
o
n
a
s
elf
-
c
r
e
ated
d
ataset,
an
d
t
h
e
o
b
tain
e
d
r
esu
lts
u
n
d
er
wen
t
co
m
p
ar
is
o
n
with
s
tate
-
of
-
t
h
e
-
ar
t
liter
atu
r
e.
T
h
e
p
r
o
p
o
s
ed
R
esNet
m
eth
o
d
ad
d
r
ess
ed
th
e
is
s
u
e
o
f
th
e
v
an
is
h
in
g
g
r
ad
ien
t
p
r
o
b
lem
,
wh
ich
ca
n
l
ea
d
to
d
e
g
r
ad
atio
n
in
c
h
ar
ac
te
r
r
ec
o
g
n
itio
n
.
T
h
e
ex
p
er
im
en
t
was
im
p
lem
en
ted
u
s
in
g
th
e
Go
o
g
le
C
o
lab
s
o
f
twar
e
v
er
s
io
n
,
with
h
a
n
d
wr
itten
Kan
n
ad
a
n
u
m
er
al
im
ag
es
f
e
d
as
th
e
in
p
u
t
to
th
e
r
ec
o
g
n
itio
n
p
r
o
ce
s
s
.
T
h
e
p
r
o
p
o
s
ed
R
esNet
m
eth
o
d
s
h
o
w
s
th
e
r
esu
lts
o
f
h
ig
h
ev
alu
atio
n
ac
cu
r
ac
y
o
v
er
th
e
tr
ain
in
g
s
am
p
les
o
f
9
8
.
2
%
an
d
g
en
er
aliza
tio
n
ac
c
u
r
ac
y
o
n
th
e
test
d
ataset
o
f
9
7
.
2
%
wh
ich
in
d
icate
s
th
e
ac
co
m
p
lis
h
m
en
t
a
n
d
e
f
f
icien
cy
o
f
r
ec
o
g
n
i
z
in
g
h
an
d
wr
itten
Kan
n
a
d
a
n
u
m
er
als,
wh
ich
ca
n
b
e
p
o
ten
tially
ap
p
lied
in
v
ar
i
o
u
s
r
ea
l
-
life
ap
p
licatio
n
s
,
s
u
ch
as
d
ig
it
r
ec
o
g
n
itio
n
,
OC
R
,
an
d
d
o
c
u
m
en
t
d
i
g
itizatio
n
.
Ov
er
all,
th
e
p
r
o
p
o
s
ed
R
esNet
m
eth
o
d
p
r
o
v
id
es
a
p
r
o
m
is
in
g
s
o
lu
tio
n
f
o
r
au
to
m
ated
s
y
s
tem
s
r
eq
u
ir
in
g
ac
cu
r
ate
h
an
d
wr
itten
Kan
n
ad
a
n
u
m
er
al
r
ec
o
g
n
itio
n
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
Au
th
o
r
s
s
tate
n
o
f
u
n
d
in
g
in
v
o
lv
ed
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
Ujwala
B
.
S
.
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Pra
m
o
d
Ku
m
ar
S
.
✓
✓
✓
✓
✓
✓
H
.
R
.
Ma
h
ad
ev
aswam
y
✓
✓
✓
✓
✓
✓
✓
✓
✓
Su
m
ath
i K
.
✓
✓
✓
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
Au
th
o
r
s
s
tate
n
o
co
n
f
lict o
f
in
t
er
est.
DATA AV
AI
L
AB
I
L
I
T
Y
T
h
e
d
ata
th
at
s
u
p
p
o
r
t
th
e
f
in
d
i
n
g
s
o
f
th
is
s
tu
d
y
a
r
e
av
ailab
le
f
r
o
m
th
e
co
r
r
esp
o
n
d
i
n
g
au
t
h
o
r
,
[
UB
S],
u
p
o
n
r
ea
s
o
n
ab
le
r
eq
u
est.
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
K
a
n
n
a
d
a
h
a
n
d
w
r
itten
n
u
mera
l reco
g
n
itio
n
th
r
o
u
g
h
d
ee
p
le
a
r
n
in
g
…
(
Ujw
a
la
B
.
S
.
)
5047
RE
F
E
R
E
NC
E
S
[
1
]
A
.
M
.
M
.
O
.
C
h
a
c
k
o
a
n
d
P
.
M
.
D
h
a
n
y
a
,
“
M
u
l
t
i
p
l
e
c
l
a
ss
i
f
i
e
r
sy
s
t
e
m
f
o
r
o
f
f
l
i
n
e
M
a
l
a
y
a
l
a
m
c
h
a
r
a
c
t
e
r
r
e
c
o
g
n
i
t
i
o
n
,
”
Pr
o
c
e
d
i
a
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
v
o
l
.
4
6
,
p
p
.
8
6
–
9
2
,
2
0
1
5
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
p
r
o
c
s
.
2
0
1
5
.
0
1
.
0
6
1
.
[
2
]
D
.
P
.
Y
a
d
a
v
a
n
d
M
.
K
u
mar,
“
K
a
n
n
a
d
a
c
h
a
r
a
c
t
e
r
r
e
c
o
g
n
i
t
i
o
n
i
n
i
m
a
g
e
s
u
si
n
g
h
i
st
o
g
r
a
m
o
f
o
r
i
e
n
t
e
d
g
r
a
d
i
e
n
t
s
a
n
d
ma
c
h
i
n
e
l
e
a
r
n
i
n
g
,
”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
2
n
d
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
e
r
V
i
si
o
n
&
I
m
a
g
e
Pro
c
e
ss
i
n
g
,
S
i
n
g
a
p
o
r
e
:
S
p
r
i
n
g
e
r
S
i
n
g
a
p
o
r
e
,
2
0
1
8
,
v
o
l
.
7
0
4
,
p
p
.
2
6
5
–
277
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
981
-
10
-
7
8
9
8
-
9
_
2
2
.
[
3
]
S
.
K
a
r
t
h
i
k
a
n
d
K
.
S
.
M
u
r
t
h
y
,
“
H
a
n
d
w
r
i
t
t
e
n
K
a
n
n
a
d
a
n
u
m
e
r
a
l
s
r
e
c
o
g
n
i
t
i
o
n
u
s
i
n
g
h
i
st
o
g
r
a
m
o
f
o
r
i
e
n
t
e
d
g
r
a
d
i
e
n
t
d
e
scr
i
p
t
o
r
s
a
n
d
su
p
p
o
r
t
v
e
c
t
o
r
ma
c
h
i
n
e
s
,
”
i
n
Em
e
r
g
i
n
g
I
C
T
f
o
r
Br
i
d
g
i
n
g
t
h
e
F
u
t
u
re
–
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
4
9
t
h
A
n
n
u
a
l
C
o
n
v
e
n
t
i
o
n
o
f
t
h
e
C
o
m
p
u
t
e
r
S
o
c
i
e
t
y
o
f
I
n
d
i
a
C
S
I
,
2
0
1
5
,
p
p
.
5
1
–
57
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
319
-
1
3
7
3
1
-
5
_
7
.
[
4
]
M
.
A
v
a
d
e
sh
a
n
d
N
.
G
o
y
a
l
,
“
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
f
o
r
S
a
n
sk
r
i
t
u
s
i
n
g
c
o
n
v
o
l
u
t
i
o
n
n
e
u
r
a
l
n
e
t
w
o
r
k
s
,
”
i
n
2
0
1
8
1
3
t
h
I
AP
R
I
n
t
e
r
n
a
t
i
o
n
a
l
W
o
rks
h
o
p
o
n
D
o
c
u
m
e
n
t
An
a
l
y
si
s
S
y
st
e
m
s (D
A
S
)
,
A
p
r
.
2
0
1
8
,
p
p
.
4
4
7
–
452
,
d
o
i
:
1
0
.
1
1
0
9
/
D
A
S
.
2
0
1
8
.
5
0
.
[
5
]
C
.
B
o
u
f
e
n
a
r
,
A
.
K
e
r
b
o
u
a
,
a
n
d
M
.
B
a
t
o
u
c
h
e
,
“
I
n
v
e
st
i
g
a
t
i
o
n
o
n
d
e
e
p
l
e
a
r
n
i
n
g
f
o
r
o
f
f
-
l
i
n
e
h
a
n
d
w
r
i
t
t
e
n
A
r
a
b
i
c
c
h
a
r
a
c
t
e
r
r
e
c
o
g
n
i
t
i
o
n
,
”
C
o
g
n
i
t
i
v
e
S
y
st
e
m
s R
e
se
a
rc
h
,
v
o
l
.
5
0
,
p
p
.
1
8
0
–
1
9
5
,
A
u
g
.
2
0
1
8
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
c
o
g
sy
s
.
2
0
1
7
.
1
1
.
0
0
2
.
[
6
]
B
.
R
.
K
a
v
i
t
h
a
a
n
d
C
.
S
r
i
ma
t
h
i
,
“
B
e
n
c
h
mar
k
i
n
g
o
n
o
f
f
l
i
n
e
h
a
n
d
w
r
i
t
t
e
n
T
a
mi
l
c
h
a
r
a
c
t
e
r
r
e
c
o
g
n
i
t
i
o
n
u
si
n
g
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
s,”
J
o
u
r
n
a
l
o
f
K
i
n
g
S
a
u
d
U
n
i
v
e
rsi
t
y
,
v
o
l
.
3
4
,
n
o
.
4
,
p
p
.
1
1
8
3
–
1
1
9
0
,
A
p
r
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
j
k
s
u
c
i
.
2
0
1
9
.
0
6
.
0
0
4
.
[
7
]
V
.
C
.
H
a
l
l
u
r
a
n
d
R
.
S
.
H
e
g
a
d
i
,
“
H
a
n
d
w
r
i
t
t
e
n
K
a
n
n
a
d
a
n
u
mera
l
s
r
e
c
o
g
n
i
t
i
o
n
u
s
i
n
g
d
e
e
p
l
e
a
r
n
i
n
g
c
o
n
v
o
l
u
t
i
o
n
n
e
u
r
a
l
n
e
t
w
o
r
k
(
D
C
N
N
)
c
l
a
ss
i
f
i
e
r
,
”
C
S
I
T
ra
n
s
a
c
t
i
o
n
s
o
n
I
C
T
,
v
o
l
.
8
,
n
o
.
3
,
p
p
.
2
9
5
–
3
0
9
,
2
0
2
0
,
d
o
i
:
1
0
.
1
0
0
7
/
s
4
0
0
1
2
-
020
-
0
0
2
7
3
-
9.
[
8
]
T.
K
.
B
h
o
w
m
i
k
,
P
.
G
h
a
n
t
y
,
A
.
R
o
y
,
a
n
d
S
.
K
.
P
a
r
u
i
,
“
S
V
M
-
b
a
se
d
h
i
e
r
a
r
c
h
i
c
a
l
a
r
c
h
i
t
e
c
t
u
r
e
s
f
o
r
h
a
n
d
w
r
i
t
t
e
n
B
a
n
g
l
a
c
h
a
r
a
c
t
e
r
r
e
c
o
g
n
i
t
i
o
n
,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
n
D
o
c
u
m
e
n
t
A
n
a
l
y
si
s
a
n
d
R
e
c
o
g
n
i
t
i
o
n
(
I
J
D
A
R)
,
v
o
l
.
1
2
,
n
o
.
2
,
p
p
.
9
7
–
1
0
8
,
Ju
l
.
2
0
0
9
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
0
0
3
2
-
009
-
0
0
8
4
-
x.
[
9
]
S
.
A
h
l
a
w
a
t
,
A
.
C
h
o
u
d
h
a
r
y
,
A
.
N
a
y
y
a
r
,
S
.
S
i
n
g
h
,
a
n
d
B
.
Y
o
o
n
,
“
I
mp
r
o
v
e
d
h
a
n
d
w
r
i
t
t
e
n
d
i
g
i
t
r
e
c
o
g
n
i
t
i
o
n
u
s
i
n
g
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
s
(
C
N
N
)
,
”
S
e
n
s
o
rs
,
v
o
l
.
2
0
,
n
o
.
1
2
,
Ju
n
.
2
0
2
0
,
d
o
i
:
1
0
.
3
3
9
0
/
s
2
0
1
2
3
3
4
4
.
[
1
0
]
M
.
M
h
a
p
se
k
a
r
,
P
.
M
h
a
p
s
e
k
a
r
,
A
.
M
h
a
t
r
e
,
a
n
d
V
.
S
a
w
a
n
t
,
“
I
mp
l
e
me
n
t
a
t
i
o
n
o
f
r
e
si
d
u
a
l
n
e
t
w
o
r
k
(
R
e
sN
e
t
)
f
o
r
D
e
v
a
n
a
g
a
r
i
h
a
n
d
w
r
i
t
t
e
n
c
h
a
r
a
c
t
e
r
r
e
c
o
g
n
i
t
i
o
n
,
”
i
n
A
d
v
a
n
c
e
d
C
o
m
p
u
t
i
n
g
T
e
c
h
n
o
l
o
g
i
e
s
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s
,
2
0
2
0
,
p
p
.
1
3
7
–
148
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
9
8
1
-
15
-
3
2
4
2
-
9
_
1
4
.
[
1
1
]
C
.
S
h
a
r
ma,
S
.
S
h
a
r
m
a
,
S
a
k
s
h
i
,
a
n
d
H
.
-
Y
.
C
h
e
n
,
“
A
d
v
a
n
c
e
me
n
t
s
i
n
h
a
n
d
w
r
i
t
t
e
n
D
e
v
a
n
a
g
a
r
i
c
h
a
r
a
c
t
e
r
r
e
c
o
g
n
i
t
i
o
n
:
a
st
u
d
y
o
n
t
r
a
n
sf
e
r
l
e
a
r
n
i
n
g
a
n
d
V
G
G
1
6
a
l
g
o
r
i
t
h
m,
”
D
i
s
c
o
v
e
r
Ap
p
l
i
e
d
S
c
i
e
n
c
e
s
,
v
o
l
.
6
,
n
o
.
1
2
,
N
o
v
.
2
0
2
4
,
d
o
i
:
1
0
.
1
0
0
7
/
s
4
2
4
5
2
-
024
-
0
6
2
1
7
-
1.
[
1
2
]
M
.
A
a
mi
r
,
N
.
M
.
N
a
w
i
,
H
.
B
.
M
a
h
d
i
n
,
R
.
N
a
se
e
m,
a
n
d
M
.
Zu
l
q
a
r
n
a
i
n
,
“
A
u
t
o
-
e
n
c
o
d
e
r
v
a
r
i
a
n
t
s
f
o
r
s
o
l
v
i
n
g
h
a
n
d
w
r
i
t
t
e
n
d
i
g
i
t
s
c
l
a
ss
i
f
i
c
a
t
i
o
n
p
r
o
b
l
e
m
,
”
I
n
t
e
rn
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
Fu
zz
y
L
o
g
i
c
a
n
d
I
n
t
e
l
l
i
g
e
n
t
S
y
st
e
m
s
,
v
o
l
.
2
0
,
n
o
.
1
,
p
p
.
8
–
1
6
,
M
a
r
.
2
0
2
0
,
d
o
i
:
1
0
.
5
3
9
1
/
I
JF
I
S
.
2
0
2
0
.
2
0
.
1
.
8
.
[
1
3
]
A
.
M
o
n
t
a
z
e
r
i
,
M
.
S
h
a
m
si
,
a
n
d
R
.
D
i
a
n
a
t
,
“
U
s
i
n
g
a
n
e
w
a
p
p
r
o
a
c
h
i
n
d
e
e
p
d
i
c
t
i
o
n
a
r
y
l
e
a
r
n
i
n
g
t
o
h
a
n
d
w
r
i
t
i
n
g
n
u
m
b
e
r
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
i
n
2
0
2
0
2
5
t
h
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
m
p
u
t
e
r
C
o
n
f
e
re
n
c
e
,
C
o
m
p
u
t
e
r
S
o
c
i
e
t
y
o
f
I
ra
n
(
C
S
I
C
C
)
,
T
e
h
r
a
n
,
Ja
n
.
2
0
2
0
,
p
p
.
1
–
8
,
d
o
i
:
1
0
.
1
1
0
9
/
C
S
I
C
C
4
9
4
0
3
.
2
0
2
0
.
9
0
5
0
0
6
8
.
[
1
4
]
P
.
S
h
a
r
ma
,
P
.
H
a
n
s
,
a
n
d
S
.
C
.
G
u
p
t
a
,
“
C
l
a
ss
i
f
i
c
a
t
i
o
n
o
f
p
l
a
n
t
l
e
a
f
d
i
se
a
ses
u
si
n
g
mac
h
i
n
e
l
e
a
r
n
i
n
g
a
n
d
i
ma
g
e
p
r
e
p
r
o
c
e
ssi
n
g
t
e
c
h
n
i
q
u
e
s
,
”
i
n
2
0
2
0
1
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
C
l
o
u
d
C
o
m
p
u
t
i
n
g
,
D
a
t
a
S
c
i
e
n
c
e
&
En
g
i
n
e
e
ri
n
g
(
C
o
n
f
l
u
e
n
c
e
)
,
N
o
i
d
a
,
I
n
d
i
a
,
Ja
n
.
2
0
2
0
,
p
p
.
4
8
0
–
4
8
4
,
d
o
i
:
1
0
.
1
1
0
9
/
C
o
n
f
l
u
e
n
c
e
4
7
6
1
7
.
2
0
2
0
.
9
0
5
7
8
8
9
.
[
1
5
]
L.
-
K
.
H
u
a
n
g
,
H
.
-
T.
T
se
n
g
,
C
.
-
C
.
H
si
e
h
,
a
n
d
C
.
-
S
.
Y
a
n
g
,
“
D
e
e
p
l
e
a
r
n
i
n
g
b
a
sed
t
e
x
t
d
e
t
e
c
t
i
o
n
u
si
n
g
R
es
N
e
t
f
o
r
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
,
”
Mu
l
t
i
m
e
d
i
a
T
o
o
l
s
a
n
d
Ap
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
8
2
,
n
o
.
3
0
,
p
p
.
4
6
8
7
1
–
4
6
9
0
3
,
D
e
c
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
1
0
4
2
-
0
2
3
-
1
5
4
4
9
-
z.
[
1
6
]
S
.
S
i
n
g
h
,
N
.
K
.
G
a
r
g
,
a
n
d
M
.
K
u
m
a
r
,
“
V
G
G
1
6
:
O
f
f
l
i
n
e
h
a
n
d
w
r
i
t
t
e
n
D
e
v
a
n
a
g
a
r
i
w
o
r
d
r
e
c
o
g
n
i
t
i
o
n
u
s
i
n
g
t
r
a
n
sf
e
r
l
e
a
r
n
i
n
g
,
”
Mu
l
t
i
m
e
d
i
a
T
o
o
l
s
a
n
d
Ap
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
8
3
,
n
o
.
2
9
,
p
p
.
7
2
5
6
1
–
7
2
5
9
4
,
F
e
b
.
2
0
2
4
,
d
o
i
:
1
0
.
1
0
0
7
/
s1
1
0
4
2
-
0
2
4
-
1
8
3
9
4
-
7.
[
1
7
]
D
.
W
u
,
B
.
S
u
n
,
a
n
d
M
.
S
h
a
n
g
,
“
H
y
p
e
r
p
a
r
a
me
t
e
r
l
e
a
r
n
i
n
g
f
o
r
d
e
e
p
l
e
a
r
n
i
n
g
-
b
a
se
d
r
e
c
o
mm
e
n
d
e
r
sy
s
t
e
ms
,
”
I
EE
E
T
r
a
n
s
a
c
t
i
o
n
s
o
n
S
e
r
v
i
c
e
s
C
o
m
p
u
t
i
n
g
,
v
o
l
.
1
6
,
n
o
.
4
,
p
p
.
2
6
9
9
–
2
7
1
2
,
J
u
l
.
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
TSC.
2
0
2
3
.
3
2
3
4
6
2
3
.
[
1
8
]
X
.
W
a
n
g
,
L.
Y
a
n
,
a
n
d
Q
.
Zh
a
n
g
,
“
R
e
se
a
r
c
h
o
n
t
h
e
a
p
p
l
i
c
a
t
i
o
n
o
f
g
r
a
d
i
e
n
t
d
e
sc
e
n
t
a
l
g
o
r
i
t
h
m
i
n
ma
c
h
i
n
e
l
e
a
r
n
i
n
g
,
”
i
n
2
0
2
1
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
C
o
m
p
u
t
e
r
N
e
t
w
o
rk,
E
l
e
c
t
r
o
n
i
c
a
n
d
A
u
t
o
m
a
t
i
o
n
(
I
C
C
N
EA)
,
X
i
’
a
n
,
C
h
i
n
a
,
2
0
2
1
,
p
p
.
1
1
–
1
5
.
d
o
i
:
1
0
.
1
1
0
9
/
I
C
C
N
EA
5
3
0
1
9
.
2
0
2
1
.
0
0
0
1
4
.
[
1
9
]
L.
S
h
e
n
,
C
.
C
h
e
n
,
F
.
Z
o
u
,
Z
.
Ji
e
,
J
.
S
u
n
,
a
n
d
W
.
L
i
u
,
“
A
u
n
i
f
i
e
d
a
n
a
l
y
s
i
s
o
f
a
d
a
g
r
a
d
w
i
t
h
w
e
i
g
h
t
e
d
a
g
g
r
e
g
a
t
i
o
n
a
n
d
mo
me
n
t
u
m
a
c
c
e
l
e
r
a
t
i
o
n
,
”
I
EEE
T
r
a
n
s
a
c
t
i
o
n
s
o
n
N
e
u
ra
l
N
e
t
w
o
r
k
s
a
n
d
L
e
a
r
n
i
n
g
S
y
s
t
e
m
s
,
v
o
l
.
3
5
,
n
o
.
1
0
,
p
p
.
1
4
4
8
2
–
1
4
4
9
0
,
O
c
t
.
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
TN
N
LS.
2
0
2
3
.
3
2
7
9
3
8
1
.
[
2
0
]
R
.
El
s
h
a
m
y
,
O
.
A
.
-
El
n
a
sr
,
M
.
E
l
h
o
s
e
n
y
,
a
n
d
S
.
E
l
mo
u
g
y
,
“
I
mp
r
o
v
i
n
g
t
h
e
e
f
f
i
c
i
e
n
c
y
o
f
R
M
S
P
r
o
p
o
p
t
i
mi
z
e
r
b
y
u
t
i
l
i
z
i
n
g
N
e
s
t
r
o
v
e
i
n
d
e
e
p
l
e
a
r
n
i
n
g
,
”
S
c
i
e
n
t
i
f
i
c
R
e
p
o
rt
s
,
v
o
l
.
1
3
,
n
o
.
1
,
M
a
y
2
0
2
3
,
d
o
i
:
1
0
.
1
0
3
8
/
s
4
1
5
9
8
-
0
2
3
-
3
5
6
6
3
-
x.
[
2
1
]
S
.
B
o
c
k
a
n
d
M
.
W
e
i
s,
“
A
p
r
o
o
f
o
f
l
o
c
a
l
c
o
n
v
e
r
g
e
n
c
e
f
o
r
t
h
e
A
d
a
m
o
p
t
i
m
i
z
e
r
,
”
i
n
2
0
1
9
I
n
t
e
rn
a
t
i
o
n
a
l
J
o
i
n
t
C
o
n
f
e
re
n
c
e
o
n
N
e
u
ra
l
N
e
t
w
o
rks
(
I
J
C
N
N
)
,
B
u
d
a
p
e
st
,
H
u
n
g
a
r
y
,
J
u
l
.
2
0
1
9
,
p
p
.
1
–
8
,
d
o
i
:
1
0
.
1
1
0
9
/
I
J
C
N
N
.
2
0
1
9
.
8
8
5
2
2
3
9
.
[
2
2
]
S
.
M
.
Za
ma
n
,
M
d
.
M
.
H
a
sa
n
,
R
.
I
.
S
a
k
l
i
n
e
,
D
.
D
a
s
,
a
n
d
M
d
.
A
.
A
l
a
m
,
“
A
c
o
m
p
a
r
a
t
i
v
e
a
n
a
l
y
si
s
o
f
o
p
t
i
m
i
z
e
r
s
i
n
r
e
c
u
r
r
e
n
t
n
e
u
r
a
l
n
e
t
w
o
r
k
s
f
o
r
t
e
x
t
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
i
n
2
0
2
1
I
EE
E
As
i
a
-
P
a
c
i
f
i
c
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
D
a
t
a
En
g
i
n
e
e
ri
n
g
(
C
S
D
E)
,
B
r
i
s
b
a
n
e
,
A
u
s
t
r
a
l
i
a
,
D
e
c
.
2
0
2
1
,
p
p
.
1
–
6
,
d
o
i
:
1
0
.
1
1
0
9
/
C
S
D
E
5
3
8
4
3
.
2
0
2
1
.
9
7
1
8
3
9
4
.
[
2
3
]
K
.
M
.
H
o
s
n
y
,
D
.
E
l
sh
o
u
r
a
,
E
.
R
.
M
o
h
a
me
d
,
E.
V
r
o
c
h
i
d
o
u
,
a
n
d
G
.
A
.
P
a
p
a
k
o
st
a
s,
“
D
e
e
p
l
e
a
r
n
i
n
g
a
n
d
o
p
t
i
mi
z
a
t
i
o
n
-
b
a
se
d
me
t
h
o
d
s
f
o
r
sk
i
n
l
e
si
o
n
s
s
e
g
me
n
t
a
t
i
o
n
:
a
r
e
v
i
e
w
,
”
I
EE
E
A
c
c
e
ss
,
v
o
l
.
1
1
,
p
p
.
8
5
4
6
7
–
8
5
4
8
8
,
A
u
g
.
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
3
.
3
3
0
3
9
6
1
.
[
2
4
]
I
.
A
.
U
sma
n
i
,
M
.
T
.
Q
a
d
r
i
,
R
.
Zi
a
,
F
.
S
.
A
l
r
a
y
e
s,
O
.
S
a
i
d
a
n
i
,
a
n
d
K
.
D
a
s
h
t
i
p
o
u
r
,
“
i
n
t
e
r
a
c
t
i
v
e
e
f
f
e
c
t
o
f
l
e
a
r
n
i
n
g
r
a
t
e
a
n
d
b
a
t
c
h
s
i
z
e
t
o
i
m
p
l
e
me
n
t
t
r
a
n
sf
e
r
l
e
a
r
n
i
n
g
f
o
r
b
r
a
i
n
t
u
m
o
r
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
El
e
c
t
ro
n
i
c
s
,
v
o
l
.
1
2
,
n
o
.
4
,
F
e
b
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
e
l
e
c
t
r
o
n
i
c
s1
2
0
4
0
9
6
4
.
[
2
5
]
Q
.
H
u
,
“
E
v
a
l
u
a
t
i
o
n
o
f
d
e
e
p
l
e
a
r
n
i
n
g
mo
d
e
l
s
f
o
r
K
a
n
n
a
d
a
h
a
n
d
w
r
i
t
t
e
n
d
i
g
i
t
r
e
c
o
g
n
i
t
i
o
n
,
”
i
n
2
0
2
0
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
C
o
m
p
u
t
i
n
g
a
n
d
D
a
t
a
S
c
i
e
n
c
e
(
C
D
S
)
,
S
t
a
n
f
o
r
d
,
C
A
,
U
S
A
,
A
u
g
.
2
0
2
0
,
p
p
.
1
2
5
–
1
3
0
,
d
o
i
:
1
0
.
1
1
0
9
/
C
D
S
4
9
7
0
3
.
2
0
2
0
.
0
0
0
3
1
.
[
2
6
]
H
.
T.
C
h
a
n
d
r
a
k
a
l
a
a
n
d
G
.
T
h
i
p
p
e
sw
a
my
,
“
D
e
e
p
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
s
f
o
r
r
e
c
o
g
n
i
t
i
o
n
o
f
h
i
s
t
o
r
i
c
a
l
h
a
n
d
w
r
i
t
t
e
n
K
a
n
n
a
d
a
c
h
a
r
a
c
t
e
r
s
,
”
i
n
Fro
n
t
i
e
rs
i
n
I
n
t
e
l
l
i
g
e
n
t
C
o
m
p
u
t
i
n
g
:
T
h
e
o
ry
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
1
0
1
4
.
2
0
2
0
,
p
p
.
6
9
–
77
,
2
0
2
0
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
9
8
1
-
13
-
9
9
2
0
-
6
_
7
.
[
2
7
]
K
.
K
.
D
u
t
t
a
,
A
.
H
e
r
l
e
,
L.
A
p
p
a
n
n
a
,
A
.
Tu
sh
a
r
,
a
n
d
K
.
Te
j
a
sw
i
n
i
,
“
C
l
a
ss
i
f
i
c
a
t
i
o
n
o
f
K
a
n
n
a
d
a
h
a
n
d
w
r
i
t
t
e
n
a
l
p
h
a
b
e
t
s
u
s
i
n
g
m
u
l
t
i
-
c
l
a
ss
su
p
p
o
r
t
v
e
c
t
o
r
m
a
c
h
i
n
e
w
i
t
h
c
o
n
v
o
l
u
t
i
o
n
n
e
u
r
a
l
n
e
t
w
o
r
k
s,
”
i
n
I
n
t
e
l
l
i
g
e
n
t
C
o
m
p
u
t
i
n
g
P
a
r
a
d
i
g
m
a
n
d
C
u
t
t
i
n
g
-
e
d
g
e
T
e
c
h
n
o
l
o
g
i
e
s
(
I
C
I
C
C
T
2
0
2
0
)
,
C
h
a
m
,
S
w
i
t
z
e
r
l
a
n
d
:
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
s
h
i
n
g
,
A
p
r
.
2
0
2
1
,
p
p
.
4
5
5
–
4
6
3
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
0
3
0
-
6
5
4
0
7
-
8
_
4
0
.
Evaluation Warning : The document was created with Spire.PDF for Python.