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
.
3
,
J
u
n
e
2
0
2
5
,
p
p
.
2547
~
2
5
5
8
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
14
.i
3
.
p
p
2
5
4
7
-
2
5
5
8
2547
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
A
no
v
el
lig
ht
-
wei
g
ht
co
nv
o
lutiona
l
neura
l
net
wo
rk
for
rice
lea
f
disea
se
cla
ss
ificat
io
n
P
a
rt
ha
s
a
ra
t
hi
J
a
y
a
ra
m
a
n
1,
2
,
M
uthula
k
s
hm
i
P
a
la
niy
a
nd
i
1
1
D
e
p
a
r
t
me
n
t
of
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
F
a
c
u
l
t
y
of
S
c
i
e
n
c
e
a
n
d
H
u
ma
n
i
t
i
e
s,
S
R
M
I
n
st
i
t
u
t
e
of
S
c
i
e
n
c
e
a
n
d
Te
c
h
n
o
l
o
g
y
,
K
a
t
t
a
n
k
u
l
a
t
h
u
r
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
of
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s
,
F
a
c
u
l
t
y
of
S
c
i
e
n
c
e
a
n
d
H
u
ma
n
i
t
i
e
s,
S
R
M
I
n
st
i
t
u
t
e
of
S
c
i
e
n
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
V
a
d
a
p
a
l
a
n
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
Ma
r
2
7
,
2
0
2
4
R
ev
is
ed
Feb
7
,
2
0
2
5
Acc
ep
ted
Ma
r
1
5
,
2
0
2
5
Rice
is
one
of
th
e
p
rima
ry
so
u
rc
e
s
of
sta
p
le
m
e
a
ls.
It
may
tu
rn
o
u
t
to
be
a
d
isa
ste
r
as
th
e
p
ro
d
u
c
ti
o
n
of
a
g
r
icu
lt
u
ra
l
p
r
o
d
u
c
ts
is
d
e
c
li
n
e
d
due
to
d
ise
a
se
s
a
n
d
th
e
re
fo
re
it
is
re
q
u
ired
to
stra
i
g
h
ten
up
th
e
situ
a
ti
o
n
by
tak
i
n
g
p
re
c
a
u
ti
o
n
a
r
y
m
e
a
su
re
s.
G
e
n
e
r
a
ll
y
,
d
e
e
p
lea
rn
in
g
(DL)
a
rc
h
it
e
c
tu
re
s
a
re
e
m
p
lo
y
e
d
f
o
r
t
h
e
id
e
n
ti
f
ica
ti
o
n
of
p
lan
t
lea
f
d
ise
a
se
s
a
n
d
it
is
o
b
s
e
rv
e
d
th
a
t
th
e
re
is
a
trad
e
-
o
ff
b
e
twe
e
n
th
e
a
c
c
u
ra
c
y
a
n
d
p
a
ra
m
e
ters
.
Th
is
stu
d
y
in
tro
d
u
c
e
s
a
li
g
h
t
-
we
i
g
h
t
a
rc
h
it
e
c
tu
re
c
a
ll
e
d
rice
lea
f
d
ise
a
se
c
la
ss
ifi
c
a
ti
o
n
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
(R
LDC
-
CNN
).
Th
e
o
b
jec
ti
v
e
of
t
h
e
p
ro
p
o
se
d
a
rc
h
it
e
c
tu
re
is
to
imp
ro
v
e
t
h
e
a
c
c
u
ra
c
y
a
n
d
re
d
u
c
e
th
e
lo
ss
by
u
sin
g
a
c
o
m
b
in
a
ti
o
n
of
c
o
n
v
o
lu
ti
o
n
a
l
lay
e
rs,
m
a
x
p
o
o
li
n
g
lay
e
rs,
a
n
d
f
u
ll
y
c
o
n
n
e
c
ted
lay
e
rs.
T
h
e
se
lay
e
rs
u
se
a
c
ti
v
a
ti
o
n
fu
n
c
ti
o
n
f
o
r
n
o
n
-
li
n
e
a
rit
y
,
d
r
o
p
o
u
t
fo
r
re
g
u
lariz
a
ti
o
n
a
n
d
imp
lem
e
n
ts
h
y
p
e
rp
a
ra
m
e
ter
tu
n
i
n
g
wit
h
v
a
rio
u
s
o
p
ti
m
ize
rs
th
a
t
i
n
c
lu
d
e
Ad
a
m
,
R
M
S
p
r
o
p
,
st
o
c
h
a
stic
g
ra
d
ien
t
d
e
sc
e
n
t
(S
G
D)
a
n
d
a
d
a
p
ti
v
e
g
ra
d
ien
t
(
Ad
a
G
ra
d
)
.
Ex
p
e
rime
n
ts
a
re
c
o
n
d
u
c
ted
on
t
h
e
d
a
tas
e
t
of
7
,
0
9
6
ima
g
e
s
with
b
a
tch
siz
e
of
32
u
n
d
e
r
v
a
rio
u
s
lea
rn
in
g
ra
tes
.
Th
e
b
e
h
a
v
i
o
r
is
a
n
a
l
y
z
e
d
by
c
o
m
p
a
r
in
g
th
e
e
x
isti
n
g
m
o
d
e
ls
a
n
d
th
e
c
o
u
n
t
of
p
a
ra
m
e
ters
(in
m
il
li
o
n
s)
e
q
u
ip
p
e
d
by
RLDC
-
CNN
,
De
n
se
Ne
t1
2
1
,
VGG
-
16,
a
n
d
Re
sN
e
t5
0
is
0
.
6
5
,
8
.
4
9
,
1
5
.
4
4
,
2
6
.
4
9
with
a
c
c
u
ra
c
y
of
9
9
.
1
5
%
,
9
8
.
9
4
%
,
9
7
.
8
2
%
,
9
6
.
4
8
%
re
sp
e
c
ti
v
e
ly
.
K
ey
w
o
r
d
s
:
Ad
am
o
p
tim
izer
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
Deep
lear
n
in
g
R
ice
leaf
d
is
ea
s
e
T
r
an
s
f
er
lear
n
i
n
g
T
h
is
is
an
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r
th
e
CC
BY
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Mu
th
u
lak
s
h
m
i
Palan
iy
an
d
i
Dep
ar
tm
en
t
of
C
o
m
p
u
ter
Scie
n
ce
,
Facu
lty
of
Scien
ce
an
d
H
u
m
an
ities
SR
M
I
n
s
titu
te
of
Scien
ce
an
d
T
ec
h
n
o
lo
g
y
Kattan
k
u
lath
u
r
,
C
h
e
n
n
ai,
T
am
iln
ad
u
,
6
0
3
2
0
3
,
I
n
d
ia
E
m
ail:
m
u
th
u
lap
@
s
r
m
is
t.e
d
u
.
i
n
1.
I
NT
RO
D
UCT
I
O
N
In
th
e
m
o
d
er
n
s
p
h
er
e
of
ag
r
i
cu
ltu
r
e,
th
e
p
r
im
ar
y
f
o
c
u
s
h
a
s
s
h
if
ted
to
war
d
s
t
h
e
cr
itical
is
s
u
es
of
en
s
u
r
in
g
f
o
o
d
s
ec
u
r
ity
an
d
o
p
t
im
izin
g
th
e
cr
o
p
y
ield
s
.
R
ice
cr
o
p
s
,
a
d
ietar
y
s
tap
le
f
o
r
a
s
u
b
s
tan
tial
p
o
r
tio
n
of
th
e
g
lo
b
al
p
o
p
u
latio
n
,
ar
e
c
o
n
s
tan
tly
u
n
d
er
th
e
l
o
o
m
in
g
th
r
e
at
of
v
ar
io
u
s
d
is
ea
s
es
th
at
ar
e
p
o
ten
tial
to
im
p
air
b
o
th
th
e
q
u
an
tity
an
d
q
u
ality
of
y
ield
s
.
T
h
e
s
y
m
p
t
o
m
s
an
d
in
d
icatio
n
s
in
d
u
ce
d
by
th
e
p
at
h
o
g
en
s
ar
e
u
tili
ze
d
f
o
r
th
e
p
u
r
p
o
s
e
of
i
d
en
tify
in
g
an
d
ca
teg
o
r
izin
g
th
e
d
is
ea
s
es
t
h
at
af
f
ec
t
th
e
r
ice
cr
o
p
s
.
Gen
er
ally
,
th
ese
p
ath
o
g
e
n
s
ar
e
f
u
n
g
i
an
d
v
ir
u
s
es,
th
at
in
d
u
ce
th
e
illn
ess
es
s
u
ch
as
b
ac
ter
ial
b
lig
h
t,
b
last
,
b
r
o
wn
s
p
o
t,
an
d
tu
n
g
r
o
.
Mo
s
tly
,
th
ese
d
is
ea
s
es
im
p
ac
t
th
e
f
o
liag
e
of
th
e
r
ice
p
lan
t.
T
h
e
im
ag
es
r
ep
r
esen
tin
g
th
e
p
ath
o
g
en
s
a
r
e
s
h
o
wn
in
Fig
u
r
e
1.
Fig
u
r
e
1
(
a)
illu
s
tr
ates
b
ac
ter
ial
b
lig
h
t,
a
p
r
e
v
alen
t
d
is
ea
s
e
in
r
ice.
T
h
e
d
is
ea
s
e
m
an
if
ests
as
ch
lo
r
o
s
is
of
f
o
liag
e
or
d
esiccatio
n
of
s
ee
d
s
.
Fig
u
r
e
1
(
b
)
illu
s
tr
ates
th
e
o
cc
u
r
r
e
n
ce
of
a
b
last
,
wh
ich
is
a
r
esu
lt
of
f
u
n
g
al
i
n
f
ec
tio
n
.
T
h
e
en
tire
lea
f
ex
h
ib
its
a
ch
ar
r
e
d
lo
o
k
,
r
esu
ltin
g
in
a
d
ec
lin
e
in
g
r
ai
n
q
u
ality
.
T
h
e
f
u
n
g
al
s
p
o
t,
also
k
n
o
wn
as
b
r
o
wn
s
p
o
t,
is
d
e
p
icted
in
Fig
u
r
e
1
(
c)
.
It
in
itially
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
.
3
,
J
u
n
e
2
0
2
5
:
2
5
4
7
-
2
5
5
8
2548
ap
p
ea
r
s
as
a
b
r
o
wn
dot
a
n
d
th
en
tak
es
on
a
c
y
lin
d
r
ical
or
cir
cu
lar
s
h
ap
e.
In
s
ev
er
e
s
itu
ati
o
n
s
,
it
can
af
f
ec
t
up
to
50%
of
th
e
y
ield
.
Fig
u
r
e
1
(
d
)
d
e
p
icts
tu
n
g
r
o
,
a
d
is
ea
s
e
ca
u
s
ed
by
th
e
p
r
esen
ce
of
t
wo
d
is
tin
ct
v
ir
u
s
es.
T
u
n
g
r
o
illn
ess
m
an
if
ests
as
s
t
u
n
ted
g
r
o
wth
an
d
d
ec
r
ea
s
ed
g
r
ain
p
r
o
d
u
ctio
n
.
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
1
.
I
n
f
ec
ted
r
ice
leav
es (
a)
b
ac
ter
ial
b
lig
h
t,
(
b
)
b
last
,
(
c
)
b
r
o
wn
s
p
o
t
,
an
d
(
d
)
tu
n
g
r
o
T
h
e
m
an
u
al
i
d
en
tific
atio
n
o
f
th
ese
d
is
ea
s
es
p
o
s
es
s
ig
n
if
ican
t
ch
allen
g
es
an
d
r
eq
u
ir
es
a
s
u
b
s
tan
tial
am
o
u
n
t o
f
tim
e.
T
h
e
im
p
o
r
tan
ce
o
f
s
wif
tly
an
d
ac
cu
r
ately
id
en
tify
in
g
th
ese
d
is
ea
s
es c
an
n
o
t b
e
o
v
er
s
tated
as it
is
v
ital
f
o
r
th
e
tim
ely
d
ep
lo
y
m
en
t
o
f
ef
f
ec
tiv
e
s
tr
ateg
ies
to
s
af
eg
u
ar
d
ag
r
icu
ltu
r
al
p
r
o
d
u
c
tiv
ity
an
d
ec
o
n
o
m
ic
s
tab
ilit
y
[
1
]
.
C
u
r
r
en
tly
,
s
ev
er
al
ar
tific
ial
in
tellig
en
ce
(
AI
)
ap
p
r
o
ac
h
es
ar
e
u
s
ed
to
d
etec
t
an
d
class
if
y
cr
o
p
d
is
ea
s
e
in
clu
d
in
g
k
-
n
ea
r
est
n
e
ig
h
b
o
r
s
(
k
NN)
,
d
ec
is
io
n
tr
ee
(
DT
)
,
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
i
n
e
(
SVM)
[
2
]
,
lo
g
is
tic
r
eg
r
ess
io
n
,
an
d
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN)
[
3
]
,
[
4
]
.
Am
o
n
g
s
ev
e
r
al
AI
a
p
p
r
o
ac
h
es,
C
NN
is
th
e
m
o
s
t
k
n
o
wn
an
d
wid
el
y
ap
p
lied
m
eth
o
d
f
o
r
im
ag
e
an
aly
s
is
in
ag
r
icu
ltu
r
al
r
esear
ch
[
5
]
,
[
6
]
.
C
NNs
ar
e
a
ty
p
e
of
d
ee
p
,
f
ee
d
-
f
o
r
war
d
ar
tific
i
al
n
eu
r
al
n
etwo
r
k
s
(
ANN)
th
at
h
as
g
ain
ed
s
ig
n
if
ican
t
p
r
o
m
in
e
n
ce
in
th
e
f
iel
d
of
co
m
p
u
ter
v
is
io
n
(
C
V)
an
d
its
ap
p
licatio
n
s
[
7
]
,
[
8
]
.
It
is
f
o
u
n
d
th
at
th
ese
t
ec
h
n
o
lo
g
ies
can
q
u
ick
ly
lear
n
co
m
p
lex
p
r
o
b
lem
s
b
ec
au
s
e
of
th
e
ab
ilit
y
to
s
h
a
r
e
weig
h
ts
an
d
k
n
o
wn
f
o
r
e
n
co
u
r
ag
in
g
s
ca
lab
ilit
y
an
d
m
ass
iv
e
p
ar
alleliza
tio
n
[
9
]
,
[
1
0
]
.
T
h
e
m
ain
ad
v
an
tag
e
of
th
e
C
NN
m
o
d
el
is
its
o
r
d
er
e
d
s
tr
u
ctu
r
e
an
d
h
u
g
e
lear
n
in
g
ca
p
ac
ity
th
at
w
h
ich
m
a
y
s
o
lv
e
c
o
m
p
lex
ch
allen
g
es
with
f
lex
ib
ilit
y
an
d
ad
ap
tab
ilit
y
[
1
1
]
,
[
1
2
]
.
T
h
o
u
g
h
th
e
m
o
d
els
of
tr
an
s
f
er
lear
n
in
g
(TL)
h
av
e
p
r
o
v
e
d
th
eir
lear
n
in
g
ab
ilit
ies
on
lar
g
er
d
atasets
,
it
is
f
o
u
n
d
f
r
o
m
th
o
s
e
m
o
d
els
th
at
ar
e
p
r
o
p
o
s
ed
in
r
ec
e
n
t
p
ast
ar
e
ca
p
ab
le
of
lear
n
in
g
f
r
o
m
s
m
aller
d
at
asets
[
1
3
]
,
[
1
4
]
.
In
co
n
tr
ast,
tr
ain
in
g
r
eq
u
ir
e
m
en
ts
f
o
r
C
NN
co
u
ld
be
more
p
r
ac
tical
wh
en
co
n
s
id
er
in
g
s
c
alab
ilit
y
is
s
u
es
th
at
f
o
cu
s
on
tim
e
[
1
5
]
,
[
1
6
]
.
C
NN
can
in
cr
ea
s
e
th
e
lik
elih
o
o
d
of
ac
cu
r
ate
class
if
icatio
n
;
wh
en
e
x
ten
s
iv
e
d
atasets
ar
e
p
r
o
v
i
d
ed
to
d
escr
ib
e
th
e
p
r
o
b
lem
[
1
7
]
,
[
1
8
]
.
Du
r
in
g
th
e
s
u
r
v
ey
,
it
is
s
ee
n
th
at
m
an
y
r
esear
ch
er
s
h
av
e
em
p
lo
y
ed
d
iv
er
s
e
d
ee
p
lear
n
in
g
(
DL
)
alg
o
r
ith
m
s
to
d
etec
t
d
is
ea
s
es
in
r
ice
p
la
n
ts
.
R
ah
m
an
et
al
.
[
1
9
]
p
r
o
p
o
s
ed
an
a
r
ch
itectu
r
e
th
at
ad
d
r
ess
es
th
e
in
ter
class
v
ar
iatio
n
p
er
tain
in
g
to
v
ar
io
u
s
s
tag
es
of
s
p
ec
if
ic
d
i
s
ea
s
e.
T
h
e
ex
p
er
im
en
t
u
s
es
1
,
426
im
ag
es
of
r
ice
p
lan
t
an
d
r
esu
lts
in
an
ac
cu
r
a
cy
of
9
4
.
3
3
%
with
p
ar
a
m
eter
co
u
n
t
of
0
.
8
m
illi
o
n
.
Kr
is
h
n
a
m
o
o
r
th
y
et
a
l.
[
2
0
]
p
r
o
p
o
s
es
a
m
o
d
el
ex
clu
s
iv
ely
f
o
r
th
e
class
if
icatio
n
of
r
ice
l
ea
f
d
is
ea
s
es,
th
at
ad
d
r
ess
es
ef
f
ec
tiv
ely
v
an
is
h
in
g
g
r
ad
ien
t
a
n
d
d
eter
io
r
atio
n
.
Al
s
o
,
th
e
m
eth
o
d
in
co
r
p
o
r
ates
g
lo
b
al
av
er
a
g
e
p
o
o
lin
g
to
d
ec
r
ea
s
e
th
e
p
ar
am
ete
r
co
u
n
t.
T
h
e
m
o
d
el
u
s
es
5
,
2
0
0
i
m
ag
es
an
d
th
e
ac
c
u
r
ac
y
r
ate
is
f
o
u
n
d
to
be
9
5
.
6
7
%.
Seth
y
et
a
l.
[
2
1
]
co
n
d
u
cted
an
ex
p
er
im
en
t
u
s
in
g
11
d
is
tin
ct
C
NN
m
o
d
els
u
s
in
g
TL
a
p
p
r
o
ac
h
.
T
h
e
d
ataset
co
n
s
is
ts
of
5
,
932
im
a
g
es
th
at
r
ep
r
esen
t
f
o
u
r
d
if
f
e
r
en
t
ty
p
es
of
illn
ess
es
af
f
ec
tin
g
r
ice
leav
es.
T
h
e
p
er
f
o
r
m
an
ce
of
one
of
th
e
11
m
o
d
els
th
at
im
p
lem
en
ts
R
esN
et5
0
an
d
SVM
is
f
o
u
n
d
to
be
good
with
F1
-
s
co
r
e
of
9
8
.
9
8
%.
Ho
wev
er
,
it
is
o
b
s
er
v
ed
th
at
th
e
u
tili
za
tio
n
of
TL
am
p
lifie
s
th
e
in
tr
ic
ac
y
of
th
e
m
o
d
el.
Deb
et
a
l.
[
2
2
]
d
ep
lo
y
ed
f
i
v
e
d
is
t
in
ct
TL
m
o
d
els
to
a
n
a
ly
ze
a
c
o
l
lec
ti
o
n
of
7
,
096
im
a
g
es
of
r
ic
e
le
a
v
es
,
ea
ch
r
e
p
r
ese
n
t
in
g
o
n
e
of
f
i
v
e
d
if
f
e
r
en
t
ill
n
ess
g
r
o
u
p
s
.
T
h
e
m
o
d
els
e
m
p
l
o
y
e
d
in
t
h
e
s
t
u
d
y
i
n
cl
u
d
e
A
le
x
Ne
t,
V
GG
-
16,
R
esNet
-
18,
Mo
b
ileNet
-
V2
,
a
n
d
I
n
ce
p
tio
n
V3
.
T
h
e
p
er
f
o
r
m
an
ce
of
I
n
ce
p
tio
n
V3
d
em
o
n
s
t
r
ates
its
ef
f
icien
cy
with
an
ac
cu
r
ac
y
r
ate
of
9
6
.
2
3
%.
Saleem
et
a
l.
[
2
3
]
p
r
esen
t
an
ap
p
r
o
ac
h
th
at
u
s
es
m
u
ta
n
t
p
a
r
ticle
s
war
m
o
p
tim
izatio
n
(
MU
T
-
PS
O)
to
d
etec
t
r
ice
leaf
d
is
ea
s
es
to
id
en
tify
th
e
m
o
s
t
o
p
tim
al
C
NN
ar
ch
itectu
r
e.
In
t
h
is
s
tu
d
y
,
two
d
atasets
th
at
co
n
s
is
ts
of
f
o
u
r
class
es
each
ass
es
s
an
d
o
b
s
er
v
ed
an
a
v
er
ag
e
ac
cu
r
ac
y
of
9
6
.
6
2
%.
C
h
en
et
a
l.
[
2
4
]
p
r
o
p
o
s
ed
a
lig
h
tweig
h
t
in
ce
p
tio
n
n
etwo
r
k
f
o
r
th
e
class
if
icatio
n
of
r
ice
leaf
d
is
ea
s
es,
wh
ich
i
n
co
r
p
o
r
ates
th
e
Mo
b
ileNet
ar
ch
itectu
r
e.
T
h
e
s
tu
d
y
u
s
es
a
d
ataset
co
n
s
is
t
s
of
12
d
is
tin
ct
class
es
an
d
ex
p
er
im
en
t
o
u
tco
m
es
r
esu
lts
with
an
ac
cu
r
ac
y
of
9
7
.
8
9
%.
J
o
s
h
i
et
a
l.
[
2
5
]
in
tr
o
d
u
ce
d
a
C
NN
f
r
a
m
ewo
r
k
d
esig
n
ed
s
p
ec
if
ically
f
o
r
r
ice
leaf
im
ag
es.
T
h
e
ar
ch
itectu
r
e
in
v
o
lv
es
a
s
er
ies
of
co
n
v
o
lu
tio
n
al
lay
er
s
th
at
ar
e
u
tili
ze
d
to
id
en
tify
an
d
class
if
y
leaf
co
n
d
itio
n
s
an
d
f
o
u
n
d
t
h
at
th
e
m
o
d
el
ca
m
e
up
with
an
ac
cu
r
a
cy
of
9
3
.
7
5
%.
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
A
n
o
ve
l lig
h
t
-
w
eig
h
t c
o
n
vo
l
u
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
fo
r
r
ice
lea
f d
is
ea
s
e
…
(
P
a
r
th
a
s
a
r
a
th
i
Ja
ya
r
a
ma
n
)
2549
B
ased
on
th
e
liter
atu
r
e
s
u
r
v
ey
,
it
is
f
o
u
n
d
th
at
th
e
ex
is
tin
g
wo
r
k
s
ar
e
to
d
etec
t
th
e
d
is
ea
s
es
in
r
ice
leaf
h
av
e
u
s
ed
TL
an
d
tailo
r
ed
C
NN
ar
ch
itectu
r
es.
T
y
p
ically
,
TL
ar
ch
itectu
r
es
em
p
lo
y
a
lar
g
e
n
u
m
b
er
of
p
ar
am
eter
s
an
d
y
ield
h
ig
h
lev
els
of
ac
cu
r
ac
y
[
2
6
]
.
Ho
wev
er
,
th
e
u
tili
za
tio
n
of
b
esp
o
k
e
ar
c
h
itectu
r
es
r
esu
lts
in
a
r
ed
u
ctio
n
of
p
ar
a
m
eter
s
,
alb
eit
at
th
e
ex
p
e
n
s
e
of
ac
cu
r
ac
y
[
2
7
]
.
T
h
e
r
e
is
a
s
co
p
e
to
cr
ea
t
e
a
C
NN
m
o
d
el
th
at
en
h
an
ce
s
th
e
ac
c
u
r
ac
y
p
r
o
v
id
e
d
by
TL
m
o
d
els
wh
ile
m
in
im
izin
g
th
e
n
u
m
b
er
of
p
ar
am
eter
s
.
T
h
is
r
ed
u
ctio
n
in
p
ar
am
eter
s
lead
s
to
less
tim
e
co
m
p
lex
ity
of
th
e
m
o
d
el.
T
h
e
p
r
o
p
o
s
ed
wo
r
k
aim
s
to
cr
ea
te
a
C
NN
ar
ch
itectu
r
e
by
co
m
b
in
in
g
d
o
m
ain
e
x
p
er
tis
e,
m
o
d
e
r
n
n
eu
r
al
n
etwo
r
k
a
r
ch
itectu
r
es,
an
d
th
e
wea
lth
of
k
n
o
wled
g
e
en
co
d
e
d
ex
is
tin
g
m
o
d
els.
T
h
e
ar
ch
itectu
r
e
m
ay
d
is
tin
g
u
is
h
b
etwe
en
th
e
v
ar
io
u
s
d
is
ea
s
e
m
an
if
estatio
n
s
th
at
af
f
ec
t
r
ice
c
r
o
p
s
,
t
h
at
in
clu
d
es
b
ac
ter
ial
b
lig
h
t
,
b
last
,
b
r
o
wn
s
p
o
t,
a
n
d
tu
n
g
r
o
.
T
h
e
s
ig
n
if
ican
t
c
o
n
tr
ib
u
tio
n
s
of
th
e
p
r
o
p
o
s
ed
s
tu
d
y
in
clu
d
e
,
i)
t
he
d
ev
elo
p
m
en
t
of
a
s
p
ec
ialized
C
NN
ar
ch
itectu
r
e
k
n
o
wn
as
r
ice
leaf
d
is
ea
s
e
clas
s
if
icatio
n
(
R
L
DC
)
-
C
NN
with
f
ewe
r
p
ar
am
eter
s
.
ii)
t
he
f
ea
tu
r
e
ex
tr
ac
tio
n
p
r
o
ce
s
s
f
r
o
m
th
e
i
n
p
u
t
im
ag
e
th
at
u
tili
ze
s
a
s
er
ie
s
of
co
n
v
o
lu
tio
n
al
lay
er
s
,
m
ax
pool
lay
er
s
,
a
n
d
ac
tiv
atio
n
f
u
n
ctio
n
s
at
v
ar
io
u
s
s
tag
es
to
in
tr
o
d
u
ce
n
o
n
-
lin
e
ar
b
eh
a
v
io
r
.
iii)
t
he
m
o
d
el
u
s
es
Ad
am
,
R
MSp
r
o
p
,
SGD
an
d
ad
ap
tiv
e
g
r
ad
ie
n
t
(
Ad
ag
r
ad
)
o
p
tim
izer
s
an
d
co
m
p
ar
ed
th
e
lev
el
of
o
p
tim
izatio
n
.
iv
)
t
he
e
v
alu
atio
n
of
th
e
m
o
d
el
u
s
es
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all
an
d
F1
-
Sco
r
e.
v
)
t
he
r
esu
lts
ar
e
co
m
p
ar
ed
with
th
e
s
tate
-
of
-
th
e
-
ar
t
(
SOTA
)
m
o
d
els.
T
h
e
s
u
b
s
eq
u
en
t
s
ec
tio
n
s
of
th
is
wo
r
k
a
r
e
o
r
g
an
ize
d
as
f
o
llo
ws,
s
ec
tio
n
2
p
r
o
v
id
es
a
co
m
p
r
eh
en
s
iv
e
an
aly
s
is
of
th
e
d
ataset
u
s
ed
f
o
r
tr
ain
in
g
an
d
ev
alu
atio
n
.
T
h
e
s
im
u
latio
n
f
in
d
in
g
s
ar
e
ex
p
l
ain
ed
in
s
ec
tio
n
3.
Sectio
n
4
f
o
cu
s
es
on
th
e
c
o
n
c
lu
s
io
n
,
wh
er
e
th
e
s
u
m
m
ar
y
of
th
e
w
o
r
k
an
d
p
r
o
s
p
ec
tiv
e
d
ir
ec
tio
n
s
f
o
r
f
u
r
th
er
r
esear
ch
ar
e
s
u
g
g
ested
.
2.
M
E
T
H
O
D
T
h
is
s
ec
tio
n
o
u
tlin
es
th
e
s
e
q
u
en
tial
p
r
o
ce
d
u
r
es
in
v
o
lv
e
d
in
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
.
T
h
e
p
r
o
ce
d
u
r
e
co
m
m
en
ce
s
with
g
ath
er
in
g
th
e
d
ataset,
wh
ich
is
s
u
b
s
eq
u
en
tly
s
u
b
jecte
d
to
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
.
T
h
is
en
tails
d
o
in
g
o
p
e
r
atio
n
s
s
u
ch
as
ad
ju
s
tin
g
th
e
s
ize,
en
h
an
cin
g
,
an
d
s
tan
d
ar
d
izin
g
th
e
im
ag
es.
Su
b
s
eq
u
en
tly
,
an
R
L
DC
-
C
N
N
m
o
d
el
is
em
p
lo
y
e
d
to
e
x
tr
ac
t
f
ea
tu
r
es
f
r
o
m
th
e
d
at
a.
Hy
p
er
-
p
ar
am
eter
o
p
tim
izatio
n
in
v
o
lv
es
a
d
ju
s
tin
g
th
e
v
alu
es
of
th
e
ep
o
ch
c
o
u
n
t,
lear
n
in
g
r
ate,
an
d
o
p
tim
ize
r
s
.
Ultim
ately
,
th
e
d
is
ea
s
es
ar
e
ca
teg
o
r
ized
ac
co
r
d
in
g
to
th
e
r
etr
iev
e
d
ch
a
r
ac
ter
is
tics
.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
co
n
s
is
ts
of
th
r
ee
s
ec
tio
n
s
n
am
ely
,
i)
d
ata
p
r
e
p
ar
atio
n
an
d
p
r
e
-
p
r
o
ce
s
s
in
g
;
ii)
f
ea
tu
r
e
ex
t
r
ac
tio
n
;
a
n
d
iii)
h
y
p
er
p
ar
am
eter
o
p
tim
izatio
n
.
T
h
e
wo
r
k
f
lo
w
of
th
e
p
r
o
p
o
s
ed
w
o
r
k
is
s
h
o
wn
in
Fig
u
r
e
2.
Fig
u
r
e
2.
W
o
r
k
f
lo
w
of
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
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
.
3
,
J
u
n
e
2
0
2
5
:
2
5
4
7
-
2
5
5
8
2550
2
.
1
.
Da
t
a
prepa
ra
t
i
o
n
a
nd
prepro
ce
s
s
ing
T
h
e
s
tu
d
y
u
s
es
d
ataset
cu
r
ated
f
r
o
m
b
e
n
ch
m
a
r
k
ed
Me
n
d
eley
an
d
Kag
g
le
r
ep
o
s
ito
r
y
[
2
8
]
,
[
2
9
]
.
It
en
co
m
p
ass
es
a
co
llectio
n
of
7
,
0
9
6
r
ice
leaf
im
ag
es
th
at
s
h
o
ws
th
e
p
r
esen
ce
of
d
is
ea
s
es
th
at
m
ay
be
class
if
ied
ac
r
o
s
s
f
o
u
r
class
es
n
am
ely
,
i)
b
ac
ter
ial
b
lig
h
t;
ii)
b
last
;
iii)
b
r
o
wn
s
p
o
t;
a
n
d
iv
)
tu
n
g
r
o
.
Als
o
,
th
e
d
ataset
is
in
clu
d
ed
with
im
ag
es
of
h
ea
lth
y
leav
es.
T
h
e
d
ataset
co
m
p
r
is
es
of
im
ag
es
of
p
l
an
ts
th
at
ar
e
g
r
o
wn
u
n
d
er
d
if
f
er
e
n
t
en
v
ir
o
n
m
e
n
tal
co
n
d
itio
n
s
,
g
r
o
wth
s
tag
es,
an
d
d
is
ea
s
e
s
ev
er
ities
.
T
h
e
im
ag
es
ar
e
c
o
n
v
e
r
ted
to
th
e
s
ize
to
f
it
th
e
r
eq
u
ir
e
m
e
n
t
of
th
e
DL
p
ip
elin
e.
A
s
er
i
es
of
p
r
e
-
p
r
o
ce
s
s
in
g
s
tep
s
ar
e
tak
e
n
to
s
tan
d
ar
d
ize
an
d
s
tab
ilize
th
e
s
u
itab
ilit
y
of
t
h
e
d
atase
t
f
o
r
tr
ai
n
in
g
an
d
ev
alu
atio
n
.
T
h
e
s
ize
of
th
e
i
m
ag
es
is
r
esized
to
2
2
4
×
2
2
4
p
ix
els
f
r
o
m
th
eir
o
r
ig
i
n
al
s
ize
of
300
×
300
p
ix
els.
T
h
e
u
s
e
of
d
ata
n
o
r
m
aliza
tio
n
en
s
u
r
es
th
at
each
p
ix
el
v
alu
e
s
h
o
u
ld
be
in
th
e
r
an
g
e
b
etwe
en
0
an
d
1,
wh
ic
h
lead
s
to
q
u
ick
c
o
n
v
e
r
g
en
ce
.
T
h
e
p
r
e
-
p
r
o
ce
s
s
ed
d
ataset
is
d
iv
id
ed
in
to
two
ca
teg
o
r
ies
n
am
ely
,
i)
th
e
d
ataset
f
o
r
tr
ain
in
g
an
d
ii)
th
e
d
ataset
f
o
r
test
in
g
.
Am
o
n
g
th
e
7
,
096
i
m
ag
es
co
n
s
id
er
ed
f
o
r
co
n
d
u
cti
n
g
th
e
ex
p
e
r
im
en
t,
80%
of
th
e
im
a
g
es
ar
e
u
s
ed
f
o
r
tr
ain
in
g
a
n
d
t
h
e
r
em
ai
n
in
g
20%
im
a
g
es
ar
e
u
s
ed
f
o
r
test
in
g
.
C
o
n
s
eq
u
e
n
tly
,
th
e
d
ataset
is
b
alan
ce
d
,
as
th
e
7
,
096
im
a
g
es
ar
e
d
is
tr
ib
u
te
d
n
ea
r
ly
eq
u
itab
ly
am
o
n
g
th
e
f
iv
e
class
es
.
T
h
e
d
is
tr
ib
u
tio
n
of
tr
ain
an
d
test
im
ag
es
is
illu
s
tr
ated
in
Fig
u
r
e
3.
Fig
u
r
e
3.
R
ice
leaf
d
is
ea
s
e
d
at
aset
d
escr
ip
tio
n
2
.
2
.
F
e
a
t
ure
ex
t
r
a
ct
io
n
T
h
e
f
ea
tu
r
e
e
x
tr
ac
tio
n
p
r
o
ce
d
u
r
e
ac
ce
p
ts
p
r
e
-
p
r
o
ce
s
s
ed
d
at
a
as
in
p
u
t
an
d
co
n
v
er
ts
th
e
r
a
w
d
ata
in
to
n
u
m
er
ical
f
ea
tu
r
es,
wh
ile
m
ai
n
tain
in
g
t
h
e
in
f
o
r
m
atio
n
f
r
o
m
th
e
o
r
ig
in
al
d
ata
s
et.
T
h
e
R
L
DC
-
C
NN
m
o
d
el
is
co
m
p
o
s
ed
of
m
u
ltip
le
lay
e
r
s
th
at
ar
e
u
s
ed
to
e
x
tr
ac
t
f
ea
tu
r
es.
T
h
ey
ar
e
i)
co
n
v
o
lu
tio
n
a
l
lay
er
;
ii)
p
o
o
lin
g
lay
er
;
iii)
ac
tiv
atio
n
f
u
n
ctio
n
;
i
v
)
d
r
o
p
o
u
ts
;
v)
f
latten
lay
er
;
a
n
d
v
i)
f
u
lly
co
n
n
ec
ted
(
FC
)
lay
er
an
d
is
s
h
o
wn
in
Fig
u
r
e
4.
Fig
u
r
e
4
.
T
h
e
p
r
o
p
o
s
ed
R
L
DC
-
C
NN
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
A
n
o
ve
l lig
h
t
-
w
eig
h
t c
o
n
vo
l
u
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
fo
r
r
ice
lea
f d
is
ea
s
e
…
(
P
a
r
th
a
s
a
r
a
th
i
Ja
ya
r
a
ma
n
)
2551
T
h
e
p
r
o
p
o
s
ed
R
L
DC
-
C
NN
ar
ch
itectu
r
e
co
n
s
is
ts
of
an
in
p
u
t
lay
er
,
13
f
ea
tu
r
e
ex
tr
ac
tio
n
l
ay
er
s
th
at
in
clu
d
es
5
co
n
v
o
lu
tio
n
al
lay
e
r
s
,
5
m
ax
p
o
o
lin
g
lay
er
,
1
f
latten
lay
er
,
an
d
2
FC
lay
er
s
.
T
h
e
wo
r
k
f
lo
w
s
tar
ts
with
th
e
in
p
u
t
lay
er
f
o
llo
wed
by
th
e
alter
n
ativ
e
ar
r
an
g
e
m
en
t
of
c
o
n
v
o
lu
tio
n
al
lay
e
r
an
d
m
ax
p
o
o
lin
g
lay
er
an
d
th
en
co
n
tin
u
es
with
th
e
f
latten
lay
er
s
to
r
ea
ch
out
th
e
last
two
d
en
s
e
FC
lay
er
s
th
at
p
r
o
ce
ed
to
th
e
class
if
icatio
n
of
im
ag
es,
th
e
o
u
tp
u
t.
T
h
e
tr
ain
in
g
p
r
o
ce
s
s
b
eg
in
s
with
tak
in
g
in
th
e
in
p
u
t
im
ag
e
an
d
th
en
p
r
o
ce
s
s
ed
th
r
o
u
g
h
a
co
n
v
o
lu
t
io
n
al
lay
er
th
at
c
o
n
s
is
ts
of
16
k
er
n
els,
each
of
s
ize
(3
×
3
)
.
T
h
e
co
n
v
o
lu
tio
n
p
r
o
ce
d
u
r
e
g
en
er
ates
16
f
ea
t
u
r
e
m
ap
s
,
each
of
d
im
e
n
s
io
n
s
(
2
2
2
×
2
2
2
)
.
T
h
ese
f
e
atu
r
e
m
ap
s
ar
e
s
u
b
s
eq
u
en
tl
y
p
ass
es
on
to
th
e
m
ax
p
o
o
lin
g
l
ay
er
th
at
ex
tr
ac
t
th
e
v
ital
f
ea
tu
r
es
an
d
r
e
d
u
ce
s
th
e
im
ag
e
d
im
en
s
io
n
s
to
th
e
s
ize
(
1
1
1
×
1
1
1
)
.
T
h
e
p
r
o
ce
s
s
is
r
ep
ea
ted
by
p
ass
in
g
th
e
g
e
n
er
ated
f
ea
tu
r
e
m
a
p
s
to
th
e
n
ex
t
s
u
b
s
eq
u
en
t
co
n
v
o
l
u
tio
n
lay
er
s
co
n
tain
in
g
32,
64,
1
2
8
k
er
n
els
th
at
ar
e
ac
co
m
p
an
ie
d
by
m
ax
p
o
o
lin
g
lay
er
.
In
o
r
d
e
r
to
en
co
u
r
ag
e
non
-
lin
ea
r
ity
,
th
e
r
ec
tifie
d
lin
ea
r
u
n
i
t
(
R
eL
U)
ac
tiv
atio
n
f
u
n
ctio
n
is
u
s
ed
in
co
m
b
in
atio
n
with
all
c
o
n
v
o
lu
t
io
n
o
p
er
atio
n
s
.
To
m
itig
ate
o
v
er
f
itti
n
g
,
th
e
p
r
o
p
o
s
ed
wo
r
k
u
s
ed
a
d
r
o
p
o
u
t
lay
e
r
with
th
e
r
ate
of
0
.
2
an
d
th
e
s
a
m
e
is
im
p
lem
en
ted
a
f
ter
th
e
f
o
u
r
th
c
o
n
v
o
lu
tio
n
o
p
er
atio
n
.
T
h
is
lay
er
r
an
d
o
m
ly
d
ea
ctiv
ates
a
f
r
ac
tio
n
of
n
eu
r
o
n
s
d
u
r
in
g
each
tr
ai
n
in
g
iter
at
io
n
an
d
t
h
er
ef
o
r
e
an
im
p
r
o
v
e
m
en
t
in
th
e
m
o
d
el'
s
r
esil
ien
ce
s
h
all
be
o
b
s
er
v
ed
.
T
h
e
p
r
o
ce
s
s
co
n
tin
u
es
with
th
e
f
latten
lay
er
th
at
c
o
n
v
e
r
ts
th
e
co
m
b
in
e
d
f
ea
tu
r
e
m
ap
s
th
at
ar
e
o
b
tain
ed
ac
r
o
s
s
th
e
lay
er
s
to
f
o
r
m
a
s
in
g
le
-
d
i
m
en
s
io
n
al
v
ec
to
r
of
s
ize
3
,
2
0
0
.
T
h
e
v
ec
to
r
is
f
ed
as
in
p
u
t
to
th
e
FC
lay
er
,
wh
ic
h
is
r
esp
o
n
s
ib
le
f
o
r
c
o
m
b
i
n
in
g
th
e
lo
w
-
lev
el
f
ea
tu
r
es
th
at
a
r
e
lear
n
e
d
f
r
o
m
th
e
co
n
v
o
l
u
tio
n
al
lay
er
s
in
o
r
d
e
r
to
f
o
r
m
th
e
h
ig
h
-
lev
el
f
ea
tu
r
e
m
ap
s
.
In
th
is
s
tu
d
y
,
th
e
f
ir
s
t
FC
lay
er
is
co
n
f
ig
u
r
ed
with
128
n
eu
r
o
n
s
wh
ich
is
f
o
llo
wed
by
an
o
u
tp
u
t
FC
lay
er
with
5
n
eu
r
o
n
s
t
h
at
is
eq
u
iv
alen
t
to
ca
teg
o
r
ies
of
d
is
ea
s
es
to
be
class
if
ied
.
T
h
e
So
f
tMa
x
ac
tiv
atio
n
f
u
n
cti
o
n
is
f
o
llo
wed
by
th
e
o
u
tp
u
t
lay
er
th
at
g
iv
es
th
e
p
r
o
b
ab
ilis
tic
o
u
tco
m
e
th
at
m
ay
f
it
th
e
r
ice
leaf
d
is
e
ase
in
one
of
th
e
5
class
es.
T
a
b
le
1
r
ep
r
esen
ts
th
e
co
n
f
ig
u
r
atio
n
s
u
m
m
ar
y
of
R
L
DC
-
C
NN
m
o
d
el
th
at
p
r
o
v
i
d
e
s
an
o
v
er
v
iew
of
th
e
ar
c
h
itectu
r
e.
T
h
e
s
u
m
m
ar
y
in
clu
d
es
th
e
d
etails
of
th
e
lay
er
s
th
at
d
eter
m
in
es
th
e
o
u
tp
u
t
s
h
ap
e
an
d
th
e
n
u
m
b
er
of
p
ar
am
eter
s
u
s
ed
ac
r
o
s
s
lay
er
s
.
T
ab
le
1.
R
L
DC
-
C
NN
co
n
f
ig
u
r
atio
n
s
u
m
m
ar
y
La
y
e
r
O
u
t
p
u
t
s
h
a
p
e
P
a
r
a
me
t
e
r
s
C
o
n
v
2
D
(
N
o
n
e
,
2
2
2
,
2
2
2
,
1
6
)
4
4
8
M
a
x
p
o
o
l
2
D
(
N
o
n
e
,
1
1
1
,
1
1
1
,
1
6
)
0
C
o
n
v
2
D
(
N
o
n
e
,
1
0
9
,
1
0
9
,
3
2
)
4
,
6
4
0
M
a
x
p
o
o
l
2
D
(
N
o
n
e
,
5
4
,
5
4
,
3
2
)
0
C
o
n
v
2
D
(
N
o
n
e
,
5
2
,
5
2
,
6
4
)
18
,
496
M
a
x
p
o
o
l
2
D
(
N
o
n
e
,
2
6
,
2
6
,
6
4
)
0
C
o
n
v
2
D
(
N
o
n
e
,
2
4
,
2
4
,
1
2
8
)
73
,
856
M
a
x
p
o
o
l
2
D
(
N
o
n
e
,
1
2
,
1
2
,
1
2
8
)
0
D
r
o
p
o
u
t
(
N
o
n
e
,
1
2
,
1
2
,
1
2
8
)
0
C
o
n
v
2
D
(
N
o
n
e
,
1
0
,
1
0
,
1
2
8
)
1
4
7
,
5
8
4
M
a
x
p
o
o
l
2
D
(
N
o
n
e
,
5
,
5
,
1
2
8
)
0
F
l
a
t
t
e
n
(
N
o
n
e
,
3
2
0
0
)
0
D
e
n
se
(
N
o
n
e
,
1
2
8
)
4
0
9
,
7
2
8
D
e
n
se
(
N
o
n
e
,
5
)
6
4
5
A
c
t
i
v
a
t
i
o
n
(
N
o
n
e
,
5
)
0
To
t
a
l
P
a
r
a
m
e
t
e
r
s
6
5
5
,
3
9
7
2
.
3
.
H
y
perpa
ra
m
e
t
er
o
ptim
i
za
t
io
n
T
h
is
s
ec
tio
n
p
r
esen
ts
th
e
o
p
ti
m
izatio
n
tech
n
iq
u
e
th
at
is
im
p
lem
en
ted
in
th
e
p
r
o
p
o
s
ed
m
o
d
el.
T
h
e
o
b
jectiv
e
f
u
n
ctio
n
can
be
r
ep
r
esen
ted
as
θ
th
at
m
a
x
im
izes
th
e
ac
cu
r
ac
y
an
d
m
in
im
izes
th
e
l
o
s
s
.
T
h
e
o
b
jectiv
e
f
u
n
ctio
n
can
be
d
ef
in
e
d
u
s
in
g
th
e
p
ar
am
eter
s
th
at
in
clu
d
e
‘D
-
T
r
ain
’
,
‘
n
’
,
‘
M/k
’
,
wh
er
e
D
-
T
r
ain
r
ep
r
esen
ts
th
e
n
u
m
b
er
of
tr
ain
in
g
im
ag
es,
‘
n
’
s
tan
d
s
f
o
r
n
u
m
b
er
of
ep
o
c
h
s
an
d
‘
M/k
’
is
th
e
r
atio
f
o
r
ev
e
r
y
b
atch
‘
k
’
of
s
ize
32
with
th
e
t
r
ain
in
g
im
ag
es
‘
M’
.
H
e
r
e
,
t
h
e
e
p
o
c
h
s
r
e
p
r
e
s
e
n
t
t
h
e
n
u
m
b
e
r
of
t
i
m
e
s
t
h
a
t
t
h
e
m
o
d
e
l
e
x
e
c
u
t
e
s
t
h
e
t
r
a
i
n
i
n
g
d
a
t
a
s
e
t
D
-
T
r
a
i
n
.
F
o
r
ev
e
r
y
e
p
o
c
h
,
t
h
e
D
-
T
r
a
i
n
is
s
p
l
i
t
i
n
t
o
b
a
t
c
h
e
s
.
E
a
c
h
b
a
t
c
h
c
o
n
t
ai
n
s
t
h
e
s
a
m
p
le
s
t
h
at
a
r
e
c
o
n
s
i
d
e
r
e
d
f
o
r
t
h
e
u
p
d
a
t
i
n
g
of
t
h
e
m
o
d
e
l
p
a
r
a
m
e
t
e
r
s
.
T
h
e
l
o
s
s
f
u
n
c
ti
o
n
(
L
)
is
a
p
p
l
i
e
d
to
th
e
b
a
t
c
h
e
s
o
b
t
ai
n
e
d
t
h
a
t
r
e
p
r
e
s
e
n
t
s
t
h
e
d
i
s
c
r
e
p
a
n
c
y
b
e
t
w
e
e
n
t
h
e
a
ct
u
a
l
la
b
e
l
y
i
a
n
d
t
h
e
p
r
e
d
i
c
t
e
d
l
a
b
e
l
y
^
i
.
‘
L
’
is
s
h
o
wn
as
(
1
)
:
=
−
∑
c
i
=
1
y
i
(
l
og
(
y
^
i
)
)
(
1
)
T
he
m
ea
n
of
lo
s
s
es
of
all
b
atc
h
es
is
r
ep
r
esen
ted
as
co
s
t
f
u
n
c
tio
n
(
C
)
an
d
is
r
e
p
r
esen
ted
as
(
2
)
:
=
1
n
∑
n
i
=
1
∑
[
y
ij
∗
l
og
(
y
ij
^
)
]
c
j
=
1
(
2
)
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
.
3
,
J
u
n
e
2
0
2
5
:
2
5
4
7
-
2
5
5
8
2552
In
g
e
n
er
al,
o
p
tim
izer
s
ad
j
u
s
t
th
e
v
alu
e
of
p
ar
am
ete
r
s
to
m
i
n
im
ize
th
e
lo
s
s
f
u
n
ctio
n
.
T
h
e
p
r
o
p
o
s
e
d
wo
r
k
em
p
lo
y
s
s
to
ch
asti
c
g
r
ad
ien
t
d
escen
t
(
SGD)
,
Ad
ag
r
ad
,
R
MSp
r
o
p
,
an
d
Ad
am
o
p
tim
iz
er
s
to
m
in
im
ize
th
e
lo
s
s
f
u
n
ctio
n
.
W
h
en
R
L
DC
-
C
NN
is
co
m
p
letely
tr
ain
ed
,
th
e
weig
h
t
an
d
b
ias
v
alu
es
ar
e
ad
ju
s
ted
with
r
esp
ec
t
to
L
R
.
T
h
e
weig
h
t
an
d
b
ias
v
a
lu
es
ar
e
ca
lcu
lated
u
s
in
g
t
h
e
(
3
)
an
d
(
4
)
:
W
t
←
W
t
−
1
−
η
∇
W
(
l
oss
(
W
)
)
(
3
)
b
t
←
b
t
−
1
−
η
∇
W
(
l
oss
(
b
)
)
(
4
)
w
h
er
e,
W
t
−
1
an
d
b
t
−
1
ar
e
th
e
v
alu
es
of
th
e
weig
h
t
a
n
d
b
ias
o
b
tain
e
d
d
u
r
in
g
th
e
p
r
e
v
io
u
s
ex
ec
u
tio
n
,
η
is
th
e
lear
n
in
g
r
ate
(
L
R
)
,
∇
r
ep
r
esen
ts
th
e
r
ate
of
c
h
an
g
e
in
th
e
d
er
iv
ativ
e
with
r
esp
ec
t
to
weig
h
t.
T
h
e
p
r
o
p
o
s
ed
R
L
DC
-
C
NN
m
o
d
el
u
p
d
ates
th
e
o
b
jectiv
e
f
u
n
ctio
n
θ
af
ter
th
e
c
o
m
p
l
etio
n
of
30
tr
ain
i
n
g
e
p
o
ch
s
with
LR
of
0
.
0
0
1
.
W
h
er
e
co
n
tain
s
t
h
e
m
in
im
ized
ar
g
u
m
en
ts
f
o
r
weig
h
t,
b
ias,
lo
s
s
an
d
co
s
t
can
be
ap
p
lied
on
u
n
s
e
en
d
ataset
D
-
T
est.
T
h
e
lear
n
in
g
p
a
r
am
eter
s
u
s
ed
f
o
r
th
e
im
p
lem
en
tatio
n
of
R
L
DC
-
C
NN
is
g
iv
en
in
T
ab
le
2.
T
ab
le
2.
R
L
DC
-
C
NN
lear
n
in
g
p
ar
am
eter
s
S
.
N
o
.
P
a
r
a
me
t
e
r
s
V
a
l
u
e
1
O
p
t
i
mi
z
e
r
S
G
D
,
A
d
a
g
r
a
d
,
R
M
S
p
r
o
p
,
A
d
a
m
2
Lo
ss
f
u
n
c
t
i
o
n
S
p
a
r
se
c
a
t
e
g
o
r
i
c
a
l
c
r
o
ss
e
n
t
r
o
p
y
3
Le
a
r
n
i
n
g
r
a
t
e
0
.
0
0
1
4
B
a
t
c
h
si
z
e
32
5
Ep
o
c
h
s
30
3.
RE
SU
L
T
S
AND
D
I
SCU
SS
I
O
N
T
h
is
s
tu
d
y
in
v
esti
g
ates
th
e
e
f
f
ec
ts
of
cu
s
to
m
ized
C
NNs
f
o
r
R
L
DC
.
It
is
f
o
u
n
d
f
r
o
m
t
h
e
ex
is
tin
g
wo
r
k
s
g
iv
en
in
s
ec
tio
n
1
t
h
at
th
e
m
o
d
els
u
s
e
TL
m
o
d
els
ar
e
not
ad
d
r
ess
in
g
th
e
in
f
lu
en
ce
of
o
p
tim
izer
s
th
at
m
ay
in
cr
ea
s
e
th
e
ef
f
icien
cy
of
th
e
m
o
d
el.
T
h
e
p
r
o
p
o
s
ed
R
L
DC
-
C
NN
m
o
d
el
o
u
tp
er
f
o
r
m
s
th
e
TL
m
o
d
els
in
ter
m
s
of
ac
cu
r
ac
y
a
n
d
p
a
r
am
e
ter
r
ed
u
ctio
n
.
T
h
e
in
f
r
astru
ctu
r
e
s
etu
p
f
o
r
co
n
d
u
ctin
g
th
e
ex
p
er
im
en
t
in
clu
d
es,
an
I
n
tel
-
co
r
e
i5
-
8
2
6
5
0
C
PU
o
p
er
atin
g
at
a
f
r
e
q
u
en
c
y
of
1
.
6
0
GHz
with
8
GB
of
R
AM
.
G
o
o
g
le
C
o
lab
o
r
at
o
r
y
GPU
is
u
tili
ze
d
as
h
ar
d
war
e
ac
ce
ler
ato
r
in
th
e
Py
t
h
o
n
f
r
am
e
wo
r
k
th
at
u
s
es
T
en
s
o
r
Flo
w
an
d
Ker
as.
T
h
e
R
L
DC
-
C
NN
m
o
d
el
is
tr
ain
ed
u
s
in
g
th
e
d
ataset
of
s
ize
5
,
6
7
6
,
wh
ich
is
s
aid
to
h
a
v
e
b
alan
ce
d
p
r
o
p
o
s
itio
n
of
im
a
g
es
with
in
f
ec
tio
n
s
th
at
can
be
class
if
ied
ac
r
o
s
s
f
iv
e
class
es,
v
iz.
,
b
ac
ter
ial
b
lig
h
t;
b
last
;
b
r
o
wn
s
p
o
t;
tu
n
g
r
o
;
an
d
h
ea
lth
y
.
T
h
e
ef
f
icac
y
of
th
e
p
r
o
p
o
s
ed
m
o
d
el
is
ass
e
s
s
ed
to
g
au
g
e
th
e
o
v
er
all
p
er
f
o
r
m
an
ce
an
d
q
u
ality
,
th
r
o
u
g
h
th
e
co
n
s
id
er
atio
n
of
v
a
r
io
u
s
lev
els
of
s
ev
e
r
ity
an
d
cir
cu
m
s
tan
ce
s
.
T
h
e
o
p
tim
izer
s
ar
e
in
d
iv
id
u
all
y
co
m
b
in
ed
with
th
e
m
o
d
el
to
a
n
aly
ze
th
e
p
er
f
o
r
m
an
ce
t
h
at
en
tails
a
d
ataset
of
1
,
420
im
ag
es
tak
e
n
f
r
o
m
D
-
T
est.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
is
ev
alu
ated
u
s
in
g
th
e
m
etr
ics,
ac
cu
r
ac
y
;
p
r
ec
is
io
n
;
r
ec
all
;
an
d
F1
-
s
co
r
e
is
r
e
p
r
esen
ted
u
s
in
g
th
e
(
5
)
to
(
8
)
as
:
A
c
c
ura
c
y
=
TP
+
TN
TP
+
TN
+
FP
+
FN
(
5
)
Pr
e
c
isio
n
=
TP
TP
+
FP
(
6
)
R
e
c
a
l
l
=
TP
TP
+
FN
(
7
)
F
1
−
Score
=
2
X
Pr
ecis
i
o
n
X
Recal
l
Pr
ecis
i
o
n
+
Recal
l
(
8
)
wh
er
e
,
(
TP
)
r
ep
r
esen
ts
tr
u
e
p
o
s
itiv
e
,
(
TN
)
is
tr
u
e
n
eg
ativ
e
,
(
FP
)
in
d
i
ca
tes
f
alse
p
o
s
itiv
e
,
(
FN
)
s
p
ec
if
ies
f
alse
n
eg
ativ
e
.
T
h
e
an
aly
s
is
of
th
e
R
L
DC
-
C
NN
is
d
escr
ib
ed
in
th
e
f
o
llo
wi
n
g
s
u
b
s
ec
tio
n
s
.
S
ec
tio
n
3
.
1
ad
d
r
ess
es
th
e
p
er
f
o
r
m
an
ce
of
th
e
p
r
o
p
o
s
e
d
m
o
d
el
im
p
lem
en
tin
g
o
n
e
o
p
tim
izer
at
a
tim
e.
Sectio
n
3
.
2
d
is
cu
s
s
es
th
e
ef
f
icien
cy
of
th
e
p
r
o
p
o
s
ed
m
o
d
el
with
o
th
er
p
r
e
-
tr
ain
ed
m
o
d
els.
Sectio
n
3
.
3
p
r
esen
ts
th
e
c
o
m
p
ar
is
o
n
of
R
L
DC
-
C
NN
m
o
d
el
an
d
ex
is
tin
g
liter
atu
r
es.
3
.
1
.
P
er
f
o
r
m
a
nce
ev
a
lua
t
io
n
of
RL
DC
-
CNN
wit
h
diff
er
e
nt
o
ptim
izer
s
T
h
e
p
r
o
p
o
s
ed
R
L
DC
-
C
NN
m
o
d
el
im
p
lem
en
ts
th
e
o
p
tim
ize
r
s
n
am
ely
,
Ad
am
;
R
MSp
r
o
p
;
SGD;
an
d
Ad
ag
r
ad
.
T
h
e
im
p
lem
en
tatio
n
co
m
b
in
es
th
e
p
r
o
p
o
s
ed
m
o
d
el
an
d
one
of
th
e
o
p
tim
izer
s
p
er
tim
e
to
g
et
th
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
A
n
o
ve
l lig
h
t
-
w
eig
h
t c
o
n
vo
l
u
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
fo
r
r
ice
lea
f d
is
ea
s
e
…
(
P
a
r
th
a
s
a
r
a
th
i
Ja
ya
r
a
ma
n
)
2553
o
p
tim
ized
r
esu
lts
.
It
is
o
b
s
er
v
ed
f
r
o
m
th
e
r
esu
lts
th
at
th
e
A
d
am
o
p
tim
izer
ac
h
ie
v
es
th
e
h
ig
h
est
ac
cu
r
ac
y
of
9
9
.
1
5
%.
It
in
co
r
p
o
r
ates
th
e
m
o
s
t
ad
v
an
tag
eo
u
s
ch
ar
ac
ter
is
tics
of
o
th
er
o
p
tim
izer
s
,
s
p
ec
if
i
ca
lly
Ad
ag
r
ad
a
n
d
R
MSp
r
o
p
,
wh
ich
is
o
b
s
er
v
e
d
to
y
ield
im
p
r
o
v
ed
c
o
n
v
e
r
g
en
ce
of
th
e
m
o
d
el.
T
h
e
p
er
f
o
r
m
a
n
ce
of
th
e
R
L
DC
-
C
NN
m
o
d
el
u
s
in
g
d
if
f
er
en
t
o
p
tim
izer
s
is
p
r
esen
ted
in
T
ab
le
3.
T
ab
le
3.
R
L
DC
-
C
NN
m
o
d
el
with
d
if
f
er
en
t
o
p
tim
izer
s
O
p
t
i
mi
z
e
r
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
R
e
c
a
l
l
(
%)
F1
-
S
c
o
r
e
(
%)
A
d
a
m
9
9
.
1
5
9
9
.
1
5
9
9
.
1
3
9
9
.
1
3
R
M
S
p
r
o
p
9
8
.
3
1
9
8
.
3
9
8
.
3
9
8
.
2
9
S
G
D
9
7
.
9
6
9
7
.
9
6
9
7
.
9
2
9
7
.
9
3
A
d
a
g
r
a
d
7
9
.
7
2
8
0
.
8
7
9
.
5
7
8
0
.
1
8
Fig
u
r
es
5
to
8
s
h
o
ws
th
e
ac
cu
r
ac
y
a
n
d
lo
s
s
p
lo
ts
,
wh
ich
r
ep
r
esen
ts
th
e
in
f
o
r
m
ati
o
n
of
th
e
co
n
v
er
g
en
ce
of
th
e
R
L
DC
-
C
NN
m
o
d
el
th
at
u
s
es
d
if
f
er
en
t
o
p
tim
izer
s
.
It
is
ev
id
en
t
f
r
o
m
Fig
u
r
e
5
(
a)
t
h
at
th
e
Ad
am
o
p
tim
izer
e
x
ce
ls
in
te
r
m
s
of
ac
cu
r
ac
y
w
h
en
co
m
p
ar
ed
to
o
t
h
er
o
p
tim
izer
s
.
T
h
e
m
o
d
e
l
e
x
h
i
b
i
t
s
a
c
o
n
v
e
r
g
e
n
c
e
r
a
t
e
t
h
a
t
is
e
x
c
e
e
d
i
n
g
l
y
r
a
p
i
d
,
w
i
t
h
an
a
c
c
u
r
a
c
y
of
9
9
%
.
F
i
g
u
r
e
5
(
b
)
d
e
m
o
n
s
t
r
a
t
e
s
t
h
a
t
t
h
e
A
d
am
o
p
t
i
m
i
z
e
r
r
e
d
u
c
es
t
h
e
m
o
d
e
l'
s
lo
s
s
in
t
h
e
1
0
t
h
e
p
o
c
h
a
n
d
t
h
a
t
th
e
r
e
s
u
l
ts
do
not
e
x
h
i
b
i
t
a
n
y
f
o
r
m
of
o
v
e
r
f
i
t
t
i
n
g
.
(
a)
(
b
)
Fig
u
r
e
5.
T
r
ain
in
g
a
n
d
v
alid
atio
n
p
lo
ts
of
R
L
DC
-
C
NN
w
ith
Ad
am
o
p
tim
izer
(
a)
ac
cu
r
ac
y
an
d
(
b
)
lo
s
s
T
h
e
im
p
ac
t
of
t
h
e
R
MSp
r
o
p
o
p
tim
izer
is
illu
s
tr
ated
in
Fig
u
r
e
6.
Fig
u
r
e
6
(
a)
ex
h
ib
its
th
e
m
o
d
el'
s
ab
ilit
y
to
attain
an
ac
cu
r
ac
y
of
9
8
%.
Ad
d
itio
n
ally
,
th
e
m
o
d
el'
s
lo
s
s
is
illu
s
tr
ated
in
Fig
u
r
e
6
(
b
)
,
wh
ich
co
n
tain
s
n
u
m
e
r
o
u
s
p
ea
k
s
.
N
ev
er
th
eless
,
th
e
p
ea
k
s
o
b
s
er
v
ed
in
th
e
p
lo
ts
s
u
g
g
est
th
at
a
r
eg
u
lar
izatio
n
is
war
r
an
ted
,
wh
ic
h
co
u
l
d
p
o
te
n
tially
in
ten
s
if
y
th
e
m
o
d
el'
s
co
m
p
lex
ity
.
(
a)
(
b
)
Fig
u
r
e
6.
T
r
ain
in
g
a
n
d
v
alid
atio
n
p
lo
ts
of
R
L
DC
-
C
NN
w
ith
R
MSp
r
o
p
o
p
tim
izer
(
a
)
ac
cu
r
a
cy
an
d
(
b
)
l
o
s
s
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
.
3
,
J
u
n
e
2
0
2
5
:
2
5
4
7
-
2
5
5
8
2554
T
h
e
b
e
h
av
io
u
r
of
SGD
an
d
A
d
ag
r
ad
is
illu
s
tr
ated
in
Fig
u
r
es
7
an
d
8,
an
d
it
is
ev
id
e
n
t
th
at
th
e
m
o
d
el
co
n
v
er
g
es
af
ter
a
co
n
s
id
er
ab
le
p
er
io
d
.
T
h
e
ac
c
u
r
ac
y
of
th
e
S
GD
o
p
tim
izer
,
wh
ich
r
eq
u
ir
es
ad
d
itio
n
al
tim
e
to
co
n
v
er
g
e,
is
d
ep
icted
in
Fig
u
r
e
7
(
a
)
.
Fig
u
r
e
7
(
b
)
illu
s
tr
a
tes
th
e
SGD
o
p
tim
izer
'
s
lo
s
s
,
wh
ich
is
n
ea
r
ly
co
n
v
er
g
in
g
at
th
e
3
0
th
ep
o
c
h
,
but
s
till
ex
h
i
b
its
p
ea
k
s
.
F
i
g
u
r
e
8
(
a
)
i
l
l
u
s
t
r
a
t
es
t
h
e
A
d
a
g
r
a
d
o
p
t
i
m
i
z
e
r
'
s
a
c
c
u
r
a
c
y
,
w
h
i
c
h
i
s
s
i
g
n
i
f
i
c
a
n
tl
y
l
o
w
e
r
t
h
a
n
t
h
a
t
o
f
o
t
h
e
r
o
p
t
i
m
i
z
e
r
s
d
u
e
t
o
i
t
s
s
l
u
g
g
i
s
h
l
e
a
r
n
i
n
g
r
a
t
e
.
A
d
d
i
t
i
o
n
a
l
l
y
,
F
i
g
u
r
e
8
(
b
)
d
e
m
o
n
s
t
r
a
t
es
t
h
a
t
a
s
u
b
s
t
a
n
t
i
a
l
l
o
s
s
is
o
b
s
e
r
v
e
d
w
h
e
n
c
o
m
p
a
r
e
d
w
i
t
h
t
h
e
A
d
a
m
o
p
t
i
m
i
z
e
r
.
(
a)
(
b
)
Fig
u
r
e
7.
T
r
ain
in
g
a
n
d
v
alid
atio
n
p
lo
ts
of
R
L
DC
-
C
NN
w
ith
SGD
o
p
tim
izer
(
a)
ac
cu
r
ac
y
a
n
d
(
b
)
lo
s
s
(
a)
(
b
)
Fig
u
r
e
8.
T
r
ain
in
g
a
n
d
v
alid
atio
n
p
lo
ts
of
R
L
DC
-
C
NN
w
ith
Ad
ag
r
ad
o
p
tim
izer
(
a
)
ac
cu
r
ac
y
an
d
(
b
)
lo
s
s
Fig
u
r
es
9
an
d
10
s
h
o
ws
th
e
co
n
f
u
s
io
n
m
atr
ix
th
at
e
x
h
ib
it
th
e
p
er
f
o
r
m
a
n
ce
s
of
th
e
R
L
DC
-
C
N
N
m
o
d
el
with
th
e
o
p
tim
izer
s
:
Ad
am
,
R
MSp
r
o
p
,
SGD
,
an
d
Ad
ag
r
ad
.
Fro
m
th
e
co
n
f
u
s
io
n
m
a
tr
ix
in
Fig
u
r
e
9
(
a)
,
it
can
be
s
h
o
wn
th
at
Ad
a
m
o
p
tim
izer
p
e
r
f
o
r
m
s
b
etter
th
an
th
e
o
th
er
o
p
tim
izer
s
,
with
th
e
f
ewe
s
t
m
is
class
if
icatio
n
s
.
T
h
e
c
o
n
f
u
s
io
n
m
atr
ix
in
Fig
u
r
e
9
(
b
)
il
lu
s
tr
ates
th
e
p
er
f
o
r
m
a
n
ce
of
R
L
DC
-
C
NN
with
R
MSp
r
o
p
o
p
tim
izer
in
d
is
tin
g
u
is
h
in
g
b
last
d
is
ea
s
e
f
r
o
m
o
th
er
d
is
ea
s
es,
in
d
icatin
g
d
if
f
icu
lties
in
ac
cu
r
ate
class
if
icatio
n
.
T
h
e
co
n
f
u
s
io
n
m
atr
ix
d
is
p
lay
ed
in
Fig
u
r
e
1
0
(
a
)
clea
r
ly
in
d
icate
s
th
at
th
er
e
ar
e
a
p
p
r
o
x
im
ately
28
in
s
tan
ce
s
of
m
is
class
if
icati
o
n
s
.
W
h
ich
is
h
ig
h
e
r
c
o
m
p
ar
e
d
to
th
e
r
esu
lts
o
b
tain
e
d
with
th
e
ad
am
o
p
tim
izer
s
h
o
wn
in
Fig
u
r
e
9
(
a)
.
Fig
u
r
e
1
0
(
b
)
d
e
m
o
n
s
tr
ates
th
e
Ad
ag
r
ad
o
p
tim
izer
.
W
h
ich
h
as
2
8
8
m
is
class
if
icatio
n
s
,
in
d
icatin
g
th
at
th
is
o
p
tim
izer
p
er
f
o
r
m
s
p
o
o
r
ly
.
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
A
n
o
ve
l lig
h
t
-
w
eig
h
t c
o
n
vo
l
u
tio
n
a
l n
e
u
r
a
l n
etw
o
r
k
fo
r
r
ice
lea
f d
is
ea
s
e
…
(
P
a
r
th
a
s
a
r
a
th
i
Ja
ya
r
a
ma
n
)
2555
(
a)
(
b
)
Fig
u
r
e
9.
C
o
n
f
u
s
io
n
m
atr
i
x
(
a)
R
L
DC
-
C
NN
-
Ad
am
an
d
(
b
)
R
L
DC
-
C
NN
-
R
MSp
r
o
p
(
a)
(
b
)
Fig
u
r
e
10.
C
o
n
f
u
s
io
n
m
atr
ix
(
a)
R
L
DC
-
C
N
N
-
SGD
an
d
(
b
)
R
L
DC
-
C
NN
-
Ad
ag
r
ad
3
.
2
.
P
er
f
o
r
m
a
nce
ev
a
lua
t
io
n
of
RL
DC
-
CNN
wit
h
diff
er
e
nt
pre
-
t
ra
ined
m
o
dels
T
h
e
p
r
o
p
o
s
ed
R
L
DC
-
C
NN
m
o
d
el
co
m
b
in
es
each
o
n
e
of
th
e
o
p
tim
izer
s
h
as
s
h
o
wn
t
h
at
t
h
e
m
o
d
el
p
er
f
o
r
m
s
b
etter
with
Ad
am
o
p
tim
izer
.
In
th
is
s
ec
tio
n
,
a
s
tu
d
y
is
m
ad
e
to
ass
e
s
s
th
e
p
er
f
o
r
m
an
ce
of
th
e
p
r
o
p
o
s
ed
R
L
DC
-
C
NN
m
o
d
el
with
Ad
am
o
p
tim
izer
w
h
en
co
m
p
ar
ed
with
p
r
e
-
tr
ai
n
ed
m
o
d
els
Viz.
,
Den
s
eNe
t1
2
1
,
VGG1
6
,
an
d
R
esNet5
0
.
T
h
e
s
tu
d
y
in
v
o
lv
es
7
,
096
im
a
g
es
of
r
ice
leaf
d
is
ea
s
es
to
co
n
d
u
ct
th
e
ex
p
er
im
en
t
of
th
e
f
ac
to
r
s
n
am
ely
,
i)
th
e
n
u
m
b
er
of
p
a
r
am
eter
s
an
d
ii)
ac
c
u
r
ac
y
.
T
ab
le
4
s
h
o
ws
th
e
ef
f
ec
tiv
en
ess
of
th
e
p
r
o
p
o
s
ed
m
o
d
el
th
at
r
esu
lts
in
an
ac
cu
r
ac
y
of
9
9
.
1
5
%
co
n
s
id
er
in
g
a
m
in
im
u
m
of
0
.
6
5
m
illi
o
n
p
ar
am
eter
s
wh
en
co
m
p
ar
in
g
th
e
p
ar
am
eter
s
with
o
th
er
o
p
tim
izer
s
.
T
ab
le
4.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
R
L
DC
-
C
N
N
with
v
ar
io
u
s
p
r
e
-
tr
ain
ed
m
o
d
els
M
o
d
e
l
P
a
r
a
me
t
e
r
s
(
i
n
M
i
l
l
i
o
n
)
A
c
c
u
r
a
c
y
(
%)
R
LD
C
-
C
N
N
+
A
d
a
m
0
.
6
5
9
9
.
1
5
D
e
n
seN
e
t
1
2
1
8
.
4
9
9
8
.
9
4
V
G
G
1
6
1
5
.
4
4
9
7
.
8
2
R
e
sN
e
t
5
0
2
6
.
4
9
9
6
.
4
8
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
.
3
,
J
u
n
e
2
0
2
5
:
2
5
4
7
-
2
5
5
8
2556
3.
3
.
Co
m
pa
ra
t
iv
e
s
t
ud
y
of
R
L
DC
-
CNN
wit
h
ex
is
t
ing
CN
N
m
o
dels
T
ab
le
5
p
r
esen
ts
a
co
m
p
ar
ativ
e
an
aly
s
is
of
th
e
R
L
D
C
-
C
N
N
m
o
d
el
with
o
th
er
m
o
d
els
ad
d
r
ess
ed
in
th
e
liter
atu
r
e,
as
d
is
cu
s
s
ed
in
s
ec
tio
n
1.
T
h
e
s
u
p
er
io
r
ity
of
th
e
R
L
DC
-
C
NN
is
ap
p
ar
en
t,
as
it
s
u
r
p
ass
es
th
e
o
th
er
m
o
d
els
in
b
o
th
p
ar
am
e
ter
u
s
ag
e
an
d
ac
cu
r
ac
y
.
T
h
e
ex
is
tin
g
m
o
d
els
u
tili
ze
p
r
e
-
tr
ain
ed
ar
ch
itectu
r
es
s
u
ch
as
I
n
ce
p
tio
n
R
esNetV2
,
I
n
ce
p
tio
n
V3
,
an
d
R
esNet5
0
as
p
ar
t
of
th
ei
r
DL
ar
ch
itectu
r
e
s
.
It
is
im
p
o
r
tan
t
to
m
en
tio
n
t
h
at
p
r
e
-
tr
ain
ed
m
o
d
els
n
ee
d
a
s
u
b
s
tan
tial
n
u
m
b
er
of
p
ar
a
m
eter
s
,
th
at
m
ay
lead
to
h
ig
h
co
m
p
u
tatio
n
al
an
d
m
em
o
r
y
r
eq
u
ir
em
en
ts
.
T
h
is
m
ay
b
ec
o
m
e
a
s
er
io
u
s
is
s
u
e
wh
en
atte
m
p
t
to
im
p
lem
en
t
in
em
b
ed
d
e
d
s
y
s
tem
s
.
T
h
e
p
r
o
p
o
s
ed
R
L
DC
-
C
NN
ar
ch
itectu
r
e
is
ca
r
ef
u
lly
d
esig
n
ed
to
im
p
r
o
v
e
th
e
ac
cu
r
ac
y
with
m
in
im
u
m
n
u
m
b
er
of
la
y
e
r
s
th
at
r
esu
lts
in
good
p
er
f
o
r
m
an
ce
with
o
p
tim
ized
p
ar
am
ete
r
s
.
T
ab
le
5.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
R
L
DC
-
C
N
N
with
ex
is
tin
g
m
o
d
els
in
liter
atu
r
e
A
u
t
h
o
r
s
DL
a
r
c
h
i
t
e
c
t
u
r
e
P
a
r
a
me
t
e
r
s
(
M
i
l
l
i
o
n
)
A
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
R
e
c
a
l
l
(
%)
F1
-
s
c
o
r
e
(
%)
K
r
i
s
h
n
a
m
o
o
r
t
h
y
et
a
l
.
[
20
]
I
n
c
e
p
t
i
o
n
R
e
sn
e
t
V
2
5
5
.
9
9
5
.
6
7
9
6
.
5
9
6
.
5
9
6
.
5
S
e
t
h
y
et
a
l
.
[
21
]
R
e
sN
e
t
5
0
+
S
V
M
2
5
.
6
9
8
.
2
5
9
8
.
2
5
9
8
.
2
6
9
8
.
2
6
D
e
b
et
a
l
.
[
22
]
I
n
c
e
p
t
i
o
n
V3
2
3
.
9
9
6
.
2
3
9
6
.
4
9
6
.
2
9
6
.
2
P
r
o
p
o
se
d
m
e
t
h
o
d
o
l
o
g
y
R
LD
C
-
C
N
N
0
.
6
5
9
9
.
1
5
9
9
.
1
5
9
9
.
1
3
9
9
.
1
3
4.
CO
NCLU
SI
O
N
R
ice
is
th
e
p
r
im
ar
y
f
o
o
d
s
o
u
r
c
e
f
o
r
a
m
ajo
r
ity
of
th
e
wo
r
ld
'
s
p
o
p
u
latio
n
.
Var
io
u
s
d
is
ea
s
es
im
p
ac
t
th
e
r
ice
p
lan
t
at
d
i
f
f
er
en
t
p
h
ases
of
its
g
r
o
wth
.
T
h
e
u
tili
za
tio
n
of
CV
in
c
o
n
ju
n
ctio
n
with
DL
ar
ch
itectu
r
e
can
ef
f
ec
tiv
ely
id
e
n
tify
ce
r
tain
d
is
o
r
d
er
s
.
T
h
is
r
esear
c
h
p
r
esen
ts
R
L
DC
-
C
NN,
a
n
o
v
el
lig
h
t
weig
h
t
cu
s
to
m
ize
d
DL
m
o
d
el
to
class
if
y
th
e
r
ice
leaf
d
is
ea
s
es.
T
h
e
m
o
d
el
is
ev
alu
ated
u
s
in
g
7
0
9
6
i
m
ag
es
tak
en
f
r
o
m
Me
n
d
eley
an
d
Kag
g
le
r
ep
o
s
ito
r
ies.
T
h
e
ar
ch
itectu
r
e
of
th
e
m
o
d
el
is
m
ad
e
up
of
s
er
ies
of
co
n
v
o
l
u
t
io
n
al
an
d
m
ax
pool
lay
er
s
th
at
ar
e
a
r
r
an
g
e
d
alter
n
ativ
ely
.
T
h
ese
lay
er
s
in
th
e
m
o
d
el
ar
e
u
s
ed
f
o
r
f
ea
t
u
r
e
e
x
tr
ac
tio
n
.
In
o
r
d
er
to
in
tr
o
d
u
ce
n
o
n
-
lin
ea
r
ity
,
th
e
m
o
d
el
in
c
o
r
p
o
r
ates
th
e
R
eL
U
a
ctiv
atio
n
f
u
n
ctio
n
.
T
h
e
d
r
o
p
o
u
t
lay
er
is
u
s
ed
to
d
ea
ctiv
ate
ce
r
tain
n
eu
r
o
n
s
th
at
m
ay
r
esu
lts
in
a
s
im
p
le
m
o
d
el
with
g
r
ea
t
ef
f
icien
cy
.
To
en
s
u
r
e
o
p
tim
izatio
n
,
th
e
m
o
d
el
u
s
es
Ad
a
m
Op
tim
izer
wh
ich
h
as
a
lear
n
i
n
g
r
at
e
of
0
.
0
0
1
an
d
s
p
ar
s
e
ca
teg
o
r
ical
en
tr
o
p
y
lo
s
s
f
u
n
ctio
n
.
T
h
e
o
p
tim
izatio
n
p
r
o
ce
s
s
is
p
er
f
o
r
m
ed
o
v
e
r
30
e
p
o
ch
s
,
with
each
ep
o
c
h
c
o
n
s
i
s
tin
g
of
a
b
atch
s
ize
of
3
2
.
T
h
is
m
et
h
o
d
o
lo
g
y
f
ac
ili
tates
th
e
ef
f
ec
tiv
e
co
n
v
er
g
en
c
e
an
d
t
h
e
u
s
e
of
m
in
im
u
m
n
u
m
b
er
of
p
a
r
am
eter
s
.
T
h
e
R
L
DC
-
C
N
N
m
o
d
el
ac
h
iev
ed
an
ac
cu
r
ac
y
r
ate
of
9
9
.
1
5
%
wh
ile
u
tili
zin
g
a
r
elativ
ely
s
m
all
p
ar
am
eter
co
u
n
t
of
0
.
6
5
m
illi
o
n
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
is
co
m
p
a
r
e
d
with
v
ar
io
u
s
p
r
e
-
tr
ain
ed
m
o
d
els
th
at
in
clu
d
e
Den
s
eNe
t1
2
1
,
VGG1
6
,
an
d
R
esNet5
0
,
an
d
th
e
p
r
o
p
o
s
ed
R
L
DC
-
C
NN
m
o
d
el
o
u
tp
er
f
o
r
m
s
o
th
er
co
m
p
ar
e
d
m
o
d
els.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
wo
u
ld
d
ef
i
n
itely
h
elp
t
h
e
f
ar
m
er
s
to
id
en
tify
th
e
r
ice
leaf
d
is
ea
s
es,
wh
ich
m
ay
lead
to
c
h
o
o
s
e
th
e
ap
p
r
o
p
r
iate
p
r
o
ce
d
u
r
e
to
c
u
r
e
t
h
e
s
am
e.
T
h
e
ac
cu
r
ate
an
d
tim
ely
d
iag
n
o
s
is
s
h
all
in
cr
ea
s
e
th
e
p
r
o
d
u
ctiv
ity
at
v
er
y
h
ig
h
e
r
r
ate.
T
h
e
s
tu
d
y
'
s
s
ig
n
if
ican
t
co
n
s
tr
ain
t
is
th
at
v
ar
iatio
n
s
in
illu
m
in
atio
n
or
th
e
p
r
esen
ce
of
s
h
a
d
o
ws
f
r
o
m
o
n
e
im
ag
e
to
an
o
th
er
do
n
o
t
in
f
l
u
en
ce
th
e
d
ataset.
It
h
as
th
e
p
o
ten
tial
to
d
im
in
is
h
ac
cu
r
ac
y
.
In
f
u
tu
r
e
s
tu
d
ies,
it
wo
u
ld
be
b
en
ef
icial
to
in
co
r
p
o
r
ate
d
atasets
in
clu
d
in
g
im
ag
es
of
th
is
n
atu
r
e.
Mo
r
eo
v
er
,
it
can
be
e
x
ten
d
ed
to
in
clu
d
e
ex
p
lain
ab
le
AI
,
wh
ich
wo
u
ld
p
r
o
v
id
e
f
ar
m
er
s
with
p
r
ec
is
e
ad
v
ic
e
b
ased
on
d
is
co
v
er
ed
ailm
e
n
ts
an
d
ef
f
ec
tiv
e
s
o
lu
tio
n
s
to
r
eso
l
v
e
th
e
p
r
o
b
lem
.
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
Par
th
asar
ath
i Jay
ar
am
an
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Mu
th
u
lak
s
h
m
i Pala
n
iy
an
d
i
✓
✓
✓
✓
✓
✓
✓
✓
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
Evaluation Warning : The document was created with Spire.PDF for Python.