I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
5
,
O
c
to
be
r
2025
, pp.
3757
~
3770
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
5
.pp
3757
-
3770
3757
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
M
ob
i
l
e
C
h
i
l
i
N
e
t
:
c
on
vol
u
t
i
on
al
n
e
u
r
a
l
n
e
t
w
or
k
f
or
c
h
i
l
i
l
e
ave
s
c
l
ass
i
f
i
c
at
i
on
S
ayu
t
i
R
ah
m
an
1
, M
ar
is
c
h
a E
lv
e
n
y
2
, M
ar
w
an
R
a
m
li
3
, D
io
n
ik
xon
M
an
u
r
u
n
g
1
1
D
e
pa
r
t
m
e
nt
of
I
nf
or
m
a
t
i
c
s
, F
a
c
ul
t
y
of
E
ngi
ne
e
r
i
ng
, U
ni
ve
r
s
i
t
a
s
M
e
da
n A
r
e
a
,
M
e
da
n,
I
ndone
s
i
a
2
D
e
pa
r
t
m
e
nt
of
D
a
t
a
S
c
i
e
nc
e
a
nd A
r
t
i
f
f
i
c
i
a
l
I
nt
e
l
l
i
ge
nc
e
,
F
a
c
ul
t
y of
C
om
put
e
r
S
c
i
e
nc
e
a
nd I
nf
or
m
a
t
i
on T
e
c
hnol
ogy
,
U
ni
ve
r
s
i
t
a
s
S
um
a
t
e
r
a
U
t
a
r
a
, M
e
d
a
n. I
ndone
s
i
a
3
D
e
pa
r
t
m
e
nt
of
M
a
t
he
m
a
t
i
c
s
,
F
a
c
ul
t
y of
M
a
t
he
m
a
t
i
c
s
a
nd N
a
t
ur
a
l
S
c
i
e
nc
e
s
,
U
n
i
ve
r
s
i
t
a
s
S
yi
a
h K
ua
l
a
, B
a
nd
a
A
c
e
h, I
ndone
s
i
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
N
ov 7, 2024
R
e
vi
s
e
d
14
J
ul
, 2025
A
c
c
e
pt
e
d
A
ug 6, 2025
Chili
pepper
(
Capsicum
annuum
)
is
an
important
crop
in
many
cou
ntries,
including
Indonesia,
which
plays
an
essential
role
in
the
local
econo
my
and
food
production.
To
meet
the
high
demand,
effective
agricultural
manageme
nt,
especially
the
diagnosis
and
treatment
of
plant
diseases,
is
essential
.
This
study
aims
to
improve
the
accuracy
of
chili
leaf
disease
classifi
cation
while
reducing
the
computat
ional
cost
so
that
it
can
be
a
pplied
to
low
-
cost
smart
farming
systems.
Through
the
development
of
the
MobileChiliNet
architectur
e,
which
is
the
result
of
pruning
and
fine
-
tuning
of
MobileNetV2,
this
model
achieves
the
best
accuracy,
better
than
other
convolut
ional
neural
networks
(
CNNs
)
such
as
residual
network
(
Res
Net50
)
and
visual
geometry
group
(
VGG
)
16.
Testing
with
various
optimize
rs
and
learning r
ate sche
dulers shows
that AdamW
with PolynomialDe
cay
pr
ovides
the
best
performance
by
increasing
the
validation
accuracy
to
96.48
%.
The
reduced
model
complexit
y
directly
translates
into
faster
inference
tim
es
and
lower
hardware
requirements,
allowing
the
model
to
run
on
edge
d
evices
such
as
Raspberry
Pi
o
r
smartphones.
This
makes
MobileChiliNet
highly
practical
for
smallholder
farmers
and
rural
agricultural
settings,
where
computat
ional
resources
are
limit
ed
.
By
balancing
high
classif
ication
performance
with
minimal
computational
demands,
MobileChiliNet
supports
scalable,
affordable
,
and
real
-
time
disease
monitoring
for
precision
agricult
ure.
K
e
y
w
o
r
d
s
:
C
hi
li
l
e
a
ve
s
C
hi
li
l
e
a
ve
s
c
la
s
s
if
ic
a
ti
on
C
la
s
s
if
ic
a
ti
on
C
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
M
obi
le
C
hi
li
N
e
t
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
M
a
r
is
c
ha
E
lv
e
ny
D
e
pa
r
tm
e
nt
of
D
a
ta
S
c
ie
nc
e
a
nd A
r
ti
f
f
ic
ia
l
I
nt
e
ll
ig
e
nc
e
F
a
c
ul
ty
of
C
om
put
e
r
S
c
ie
nc
e
a
nd I
nf
or
m
a
ti
on T
e
c
hnol
ogy
,
U
ni
ve
r
s
it
a
s
S
um
a
te
r
a
U
ta
r
a
M
e
da
n
,
I
ndone
s
ia
E
m
a
il
:
m
a
r
is
c
ha
e
lv
e
ny@
us
u.a
c
.i
d
1.
I
N
T
R
O
D
U
C
T
I
O
N
A
ll
s
ta
n
da
r
d
pa
pe
r
c
om
po
ne
n
ts
ha
v
e
be
e
n s
pe
c
if
ie
d
f
o
r
t
he
s
u
s
ta
i
na
b
le
de
ve
l
op
m
e
n
t
goa
ls
(
S
D
G
s
)
e
m
ph
a
s
iz
e
t
he
im
po
r
ta
nc
e
of
c
o
ns
u
m
pt
io
n
a
nd
pr
od
uc
t
io
n
in
im
pr
ov
in
g
t
he
qua
li
ty
of
l
if
e
i
n
s
oc
ie
ty
,
pa
r
ti
c
ul
a
r
ly
i
n
t
he
a
g
r
i
c
ul
tu
r
a
l
s
e
c
t
or
,
w
h
ic
h
pl
a
ys
a
ke
y
r
ol
e
i
n
e
ns
ur
in
g
f
o
od
s
e
c
u
r
i
ty
a
n
d
r
e
duc
in
g
pove
r
ty
.
A
gr
ic
ul
tu
r
e
a
ls
o
pl
a
ys
a
n
i
m
po
r
ta
n
t
r
ol
e
i
n
na
ti
on
-
bu
il
di
ng
[
1
]
.
O
n
e
a
r
e
a
o
f
f
oc
us
is
th
e
c
ul
ti
v
a
t
io
n
a
n
d
m
a
in
te
n
a
nc
e
o
f
c
h
il
i
pl
a
nt
s
,
s
pe
c
i
f
ic
a
ll
y
r
e
d
c
h
il
i
(
C
a
ps
ic
um
ann
uum
)
,
w
h
ic
h
is
a
n
im
p
o
r
ta
nt
c
r
o
p
in
m
a
n
y
c
ou
nt
r
ie
s
,
i
nc
l
ud
in
g
I
n
do
ne
s
ia
[
2
]
,
[
3
]
.
R
e
d
c
hi
li
no
t
on
ly
c
o
nt
r
ib
u
te
s
t
o
th
e
lo
c
a
l
e
c
on
om
y
bu
t
a
ls
o
t
o
f
ood
pr
od
uc
t
io
n,
s
u
c
h
a
s
s
a
u
c
e
p
r
o
duc
t
io
n
[
4
]
,
m
e
di
c
in
e
[
5
]
,
a
nd
c
hi
li
pow
de
r
[
6]
.
T
he
d
e
m
a
nd
f
or
c
h
i
li
is
ve
r
y
h
ig
h,
w
h
ic
h
r
e
q
ui
r
e
s
e
f
f
e
c
ti
ve
a
g
r
i
c
ul
tu
r
a
l
m
a
n
a
ge
m
e
n
t,
e
s
pe
c
ia
ll
y
in
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 5, O
c
to
be
r
2025
:
3757
-
3770
3758
di
a
g
nos
is
a
n
d
t
r
e
a
tm
e
nt
o
f
p
la
n
t
d
is
e
a
s
e
s
.
T
hi
s
is
c
r
uc
ia
l
to
s
upp
or
t
s
us
ta
i
na
b
le
a
gr
ic
u
lt
u
r
a
l
pr
a
c
t
ic
e
s
a
s
ta
r
ge
t
e
d
b
y
t
he
S
D
G
s
[
7
]
.
E
a
r
ly
d
e
te
c
ti
on
of
di
s
e
a
s
e
s
i
n
c
h
il
i
le
a
ve
s
c
a
n
h
e
l
p
r
e
duc
e
th
e
r
is
k
of
c
r
op
y
ie
ld
lo
s
s
a
nd
i
m
p
r
o
ve
pr
odu
c
t
io
n
q
ua
l
it
y.
I
f
c
hi
li
le
a
f
di
s
e
a
s
e
s
a
r
e
not
de
te
c
te
d
e
a
r
ly
,
th
e
im
pa
c
t
c
a
n
be
hi
ghl
y
s
ig
ni
f
ic
a
nt
on
pr
oduc
ti
on
a
nd
c
r
op
qua
li
ty
.
D
is
e
a
s
e
d
pl
a
nt
s
w
il
l
e
xpe
r
ie
nc
e
s
tu
nt
e
d
gr
ow
th
,
w
hi
c
h
w
il
l
ul
ti
m
a
te
ly
a
f
f
e
c
t
th
e
qua
nt
it
y
a
nd
qua
li
ty
of
th
e
f
r
ui
t
pr
oduc
e
d.
S
om
e
of
th
e
im
pa
c
ts
c
a
us
e
d
b
y
c
hi
li
le
a
f
di
s
e
a
s
e
s
in
c
lu
de
s
tu
nt
e
d
gr
ow
th
,
f
lo
w
e
r
dr
op,
unde
r
s
iz
e
d
f
r
ui
ts
,
a
nd
e
ve
n
r
ot
[
8]
.
A
ddi
ti
on
a
ll
y,
unde
te
c
te
d
di
s
e
a
s
e
s
pr
e
a
d
c
a
n
le
a
d
to
w
id
e
s
pr
e
a
d
in
f
e
c
ti
on
th
r
oughout
th
e
e
nt
ir
e
f
a
r
m
in
g
a
r
e
a
,
dr
a
s
ti
c
a
ll
y
in
c
r
e
a
s
in
g
th
e
c
os
t
of
tr
e
a
tm
e
nt
a
nd
di
s
e
a
s
e
c
ont
r
ol
,
s
uc
h
a
s
th
e
m
or
e
f
r
e
que
nt
a
nd
in
te
ns
iv
e
u
s
e
of
pe
s
ti
c
id
e
s
.
T
hi
s
not
onl
y
c
a
us
e
s
f
in
a
nc
ia
l
lo
s
s
e
s
f
or
f
a
r
m
e
r
s
but
a
ls
o ha
s
t
he
pot
e
nt
ia
l
to
ha
r
m
t
he
e
nvi
r
on
m
e
nt
.
F
a
il
ur
e
to
de
te
c
t
c
hi
li
le
a
f
di
s
e
a
s
e
s
e
a
r
ly
c
a
n
a
l
s
o
e
xt
e
nd
th
e
r
e
c
ove
r
y
ti
m
e
of
th
e
pl
a
nt
s
,
th
us
a
f
f
e
c
ti
ng
th
e
ne
xt
pl
a
nt
in
g
c
yc
le
.
A
s
a
r
e
s
ul
t,
lo
w
e
r
c
r
op
yi
e
ld
s
m
a
y
a
f
f
e
c
t
th
e
s
uppl
y
of
c
hi
li
in
th
e
m
a
r
ke
t,
dr
iv
in
g
up
pr
ic
e
s
,
a
nd
unde
r
m
in
in
g
th
e
e
c
onomi
c
s
t
a
bi
li
ty
of
f
a
r
m
e
r
s
[
9]
.
T
he
r
e
f
or
e
,
a
n e
f
f
ic
ie
nt
a
nd
a
c
c
ur
a
te
a
ppr
oa
c
h
to
c
hi
li
le
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on
is
c
r
uc
ia
l
to
m
a
i
nt
a
in
a
gr
ic
ul
tu
r
a
l
pr
oduc
ti
vi
ty
a
nd
e
ns
ur
e
th
e
s
us
ta
in
a
bi
li
ty
of
c
hi
li
f
a
r
m
in
g.
A
c
c
ur
a
te
c
la
s
s
if
ic
a
ti
on
of
c
hi
li
le
a
f
di
s
e
a
s
e
s
is
e
s
s
e
nt
ia
l
f
or
e
a
r
ly
di
a
gnos
is
,
e
na
bl
in
g
f
a
r
m
e
r
s
to
ta
ke
ti
m
e
ly
a
c
ti
on
in
di
s
e
a
s
e
m
a
na
g
e
m
e
nt
.
T
r
a
di
ti
ona
l
m
e
th
ods
of
di
s
e
a
s
e
de
te
c
ti
on
a
r
e
of
te
n
ti
m
e
-
c
ons
um
in
g
a
nd
pr
one
to
e
r
r
or
s
.
T
he
r
e
f
or
e
,
a
ut
om
a
ti
ng
th
e
pr
oc
e
s
s
of
c
hi
li
le
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on
us
in
g
im
a
ge
-
ba
s
e
d
m
a
c
hi
n
e
le
a
r
ni
ng
m
e
th
ods
is
hi
ghl
y
be
ne
f
ic
ia
l.
T
hr
ough
c
la
s
s
if
ic
a
ti
on
te
c
hni
que
s
,
f
a
r
m
e
r
s
c
a
n
obt
a
in
a
c
c
ur
a
te
a
nd
r
e
li
a
bl
e
in
f
or
m
a
ti
on
a
bout
th
e
c
ondi
ti
on
of
th
e
ir
pl
a
nt
s
,
w
hi
c
h
w
il
l
he
lp
th
e
m
m
a
ke
be
tt
e
r
de
c
is
io
ns
.
S
im
il
a
r
a
ppr
oa
c
he
s
ha
ve
be
e
n
a
ppl
ie
d
to
ot
he
r
c
r
ops
,
s
uc
h
a
s
th
e
di
a
gnos
is
of
A
lt
e
r
na
r
ia
di
s
e
a
s
e
a
nd
L
e
a
f
m
in
e
r
pe
s
t
on
to
m
a
t
o
le
a
ve
s
u
s
in
g
im
a
ge
pr
oc
e
s
s
in
g
te
c
hni
que
s
,
de
m
ons
tr
a
ti
ng t
he
e
f
f
e
c
ti
ve
ne
s
s
of
a
ut
om
a
te
d vi
s
ua
l
a
n
a
ly
s
is
i
n
pr
e
c
is
io
n a
gr
ic
ul
tu
r
e
[
10]
.
D
e
e
p
le
a
r
ni
ng,
pa
r
ti
c
ul
a
r
ly
c
onvolut
io
na
l
n
e
ur
a
l
ne
twor
ks
(
C
N
N
)
,
ha
s
ga
in
e
d
a
lo
t
of
a
tt
e
nt
io
n
due
to
it
s
be
tt
e
r
pe
r
f
or
m
a
nc
e
in
im
a
ge
c
la
s
s
if
ic
a
ti
on
ta
s
k
s
[
11]
,
in
c
lu
di
ng
pl
a
nt
di
s
e
a
s
e
de
te
c
ti
on
[
12]
.
C
N
N
ha
s
be
e
n
w
id
e
ly
us
e
d
in
va
r
io
us
a
gr
ic
ul
tu
r
a
l
a
ppl
ic
a
ti
ons
be
c
a
us
e
o
f
it
s
a
bi
li
ty
to
a
ut
om
a
ti
c
a
ll
y
e
xt
r
a
c
t
im
por
ta
n
t
f
e
a
tu
r
e
s
f
r
om
im
a
ge
s
w
it
hout
th
e
ne
e
d
f
or
m
a
nua
l
f
e
a
tu
r
e
e
ngi
ne
e
r
in
g.
T
hi
s
m
a
ke
s
C
N
N
a
n
id
e
a
l
c
hoi
c
e
f
or
c
hi
li
le
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on,
a
s
it
c
a
n
h
a
ndl
e
c
om
pl
e
x
v
is
ua
l
pa
tt
e
r
ns
a
nd
va
r
ia
ti
ons
in
le
a
f
im
a
g
e
s
,
r
e
s
ul
ti
ng i
n m
or
e
a
c
c
ur
a
te
pr
e
di
c
ti
ons
[
13]
.
R
e
c
e
nt
s
tu
di
e
s
s
how
th
a
t
C
N
N
m
ode
l
pe
r
f
or
m
a
nc
e
c
a
n
be
e
nha
nc
e
d
th
r
ough
te
c
hni
que
s
s
uc
h
a
s
tr
a
ns
f
e
r
le
a
r
ni
ng
[
14]
,
f
in
e
-
tu
ni
ng
[
15]
,
a
nd
pr
uni
ng
[
16]
.
T
r
a
n
s
f
e
r
le
a
r
ni
ng
a
ll
ow
s
f
or
le
ve
r
a
gi
ng
pr
e
-
tr
a
in
e
d
m
ode
ls
to
im
pr
ove
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
on
ne
w
da
ta
s
e
ts
w
it
h
m
in
im
a
l
tr
a
in
in
g
ti
m
e
,
w
hi
le
f
in
e
-
tu
ni
ng
f
ur
th
e
r
opt
i
m
iz
e
s
th
e
m
ode
l
by
a
dj
us
ti
ng
s
pe
c
if
ic
la
ye
r
s
.
O
n
th
e
ot
he
r
ha
nd,
pr
uni
ng
he
lp
s
r
e
duc
e
m
ode
l
s
iz
e
by e
li
m
in
a
ti
ng unne
c
e
s
s
a
r
y pa
r
a
m
e
te
r
s
, r
e
s
ul
ti
ng i
n m
or
e
e
f
f
ic
ie
nt
c
om
put
a
ti
on w
it
hout
s
a
c
r
if
ic
in
g a
c
c
ur
a
c
y.
I
n
a
ddi
ti
on,
r
e
c
e
nt
a
dva
nc
e
m
e
nt
s
in
li
ght
w
e
ig
ht
C
N
N
a
r
c
hi
te
c
tu
r
e
s
ha
ve
de
m
ons
tr
a
te
d
s
tr
ong
pot
e
nt
ia
l
f
or
a
g
r
ic
ul
tu
r
a
l
a
ppl
ic
a
ti
ons
on
e
dge
de
vi
c
e
s
.
A
r
e
c
e
nt
s
tu
dy
in
tr
oduc
e
d
a
n
ul
tr
a
-
li
ght
w
e
ig
h
t
ne
twor
k w
it
h a
l
ow
numbe
r
of
pa
r
a
m
e
te
r
s
, ye
t
c
a
pa
bl
e
of
a
c
hi
e
vi
ng c
om
pe
ti
ti
ve
a
c
c
ur
a
c
y i
n pl
a
nt
di
s
e
a
s
e
a
nd
pe
s
t
de
te
c
ti
on
w
hi
le
m
a
in
ta
in
in
g
m
in
im
a
l
c
om
put
a
ti
ona
l
c
o
m
pl
e
xi
ty
[
17]
.
A
not
he
r
s
tu
dy
im
pl
e
m
e
nt
e
d
a
M
obi
le
N
e
tV3L
a
r
ge
-
ba
s
e
d
m
od
e
l
f
or
r
e
a
l
-
ti
m
e
gr
a
pe
le
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on
on
a
n
e
dge
d
e
vi
c
e
(
J
e
ts
on
N
a
no)
,
a
c
hi
e
vi
ng
ove
r
99%
a
c
c
ur
a
c
y
a
lo
ng
w
it
h
e
xpl
a
in
a
b
il
it
y
f
e
a
tu
r
e
s
us
in
g
G
r
a
d
-
C
A
M
[
18]
.
T
he
s
e
de
ve
lo
pm
e
nt
s
r
e
in
f
or
c
e
th
e
pr
a
c
ti
c
a
li
ty
of
li
ght
w
e
ig
ht
C
N
N
s
in
r
e
a
l
-
w
or
ld
a
gr
ic
ul
tu
r
a
l
e
nvi
r
onm
e
nt
s
,
pa
r
ti
c
ul
a
r
ly
i
n r
ur
a
l
or
l
ow
-
r
e
s
our
c
e
s
e
tt
in
gs
w
it
h l
im
it
e
d c
om
put
in
g powe
r
.
I
n
th
is
s
tu
dy,
w
e
p
r
opos
e
M
obi
le
C
hi
li
N
e
t,
a
li
ght
w
e
ig
ht
a
nd
a
c
c
ur
a
te
m
ode
l
de
ve
lo
pe
d
by
c
om
bi
ni
ng
pr
uni
ng
a
nd
f
in
e
-
tu
ni
ng
te
c
hni
que
s
on
M
obi
le
N
e
tV
2.
T
he
obj
e
c
ti
ve
i
s
to
c
r
e
a
te
a
c
om
p
a
c
t
m
ode
l
w
it
h f
e
w
e
r
pa
r
a
m
e
te
r
s
a
nd high c
la
s
s
if
ic
a
ti
on a
c
c
ur
a
c
y, s
ui
ta
bl
e
f
or
de
pl
oym
e
nt
on l
ow
-
pow
e
r
ha
r
dw
a
r
e
s
uc
h
a
s
s
m
a
r
tp
hone
s
or
R
a
s
pbe
r
r
y
P
i.
T
a
bl
e
1
s
um
m
a
r
iz
e
s
r
e
l
a
te
d
r
e
s
e
a
r
c
h
on
le
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on
us
in
g
va
r
io
us
m
e
th
ods
a
nd t
he
ir
a
c
c
ur
a
c
ie
s
.
T
he
m
e
th
ods
li
s
te
d
in
T
a
bl
e
1
de
m
on
s
tr
a
te
va
r
io
us
te
c
hni
qu
e
s
f
or
c
hi
li
le
a
f
c
la
s
s
if
ic
a
ti
on.
F
in
e
-
tu
ni
ng
m
ode
ls
li
ke
S
huf
f
le
N
e
t
r
e
s
ul
ts
in
hi
gh
a
c
c
ur
a
c
y,
but
onl
y
f
or
two
c
la
s
s
e
s
.
F
or
a
la
r
ge
r
num
be
r
of
c
la
s
s
e
s
,
s
uc
h
a
s
f
iv
e
c
la
s
s
e
s
in
m
ode
l
s
li
ke
s
uppor
t
v
e
c
to
r
m
a
c
hi
ne
+
r
e
c
ur
r
e
nt
ne
ur
a
l
ne
twor
k
(
S
V
M
+
R
N
N
)
,
e
xt
r
e
m
e
i
nc
e
pt
io
n
(
X
c
e
pt
io
n
)
, a
nd E
f
f
ic
ie
nt
L
e
a
f
N
e
tB
4, t
he
a
c
c
ur
a
c
y r
e
m
a
in
s
r
e
la
ti
ve
ly
l
ow
, w
it
h a
m
a
xi
m
um
a
c
c
ur
a
c
y
of
92.10%
.
T
hi
s
in
di
c
a
te
s
th
a
t
th
e
r
e
is
s
ti
ll
a
ne
e
d
f
or
im
pr
ove
d
a
c
c
ur
a
c
y
in
c
hi
li
le
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on t
o e
n
s
ur
e
i
ts
a
ppl
ic
a
bi
li
ty
i
n s
m
a
r
t
a
gr
ic
ul
tu
r
a
l
s
ys
te
m
s
.
T
hi
s
r
e
s
e
a
r
c
h
a
im
s
to
de
ve
lo
p
a
m
ode
l
th
a
t
not
onl
y
i
m
pr
ove
s
a
c
c
ur
a
c
y
but
a
ls
o
r
e
du
c
e
s
c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
.
B
y
ut
il
iz
in
g
pr
uni
ng
te
c
hni
que
s
c
o
m
bi
ne
d
w
it
h
f
in
e
-
tu
ni
ng,
w
e
ta
r
ge
t
a
c
om
pa
c
t
a
nd
e
f
f
ic
ie
nt
a
r
c
hi
te
c
tu
r
e
th
a
t
c
a
n
a
c
hi
e
ve
hi
ghe
r
a
c
c
ur
a
c
y
w
it
h
lo
w
e
r
c
om
put
a
ti
ona
l
c
os
ts
. T
he
ul
ti
m
a
te
go
a
l
of
th
is
r
e
s
e
a
r
c
h
is
to
pr
oduc
e
a
m
ode
l
th
a
t
c
a
n
be
us
e
d
in
r
e
a
l
a
gr
ic
ul
tu
r
a
l
a
ppl
ic
a
ti
ons
,
im
pr
ovi
ng
di
s
e
a
s
e
de
te
c
ti
on
in
c
hi
li
pl
a
nt
s
w
hi
le
m
in
im
iz
in
g
c
om
put
a
ti
ona
l
c
os
ts
.
T
he
m
a
jo
r
c
ont
r
ib
ut
io
ns
of
th
is
s
tu
dy
a
r
e
s
um
m
a
r
iz
e
d a
s
:
i)
p
r
opos
e
s
M
obi
le
C
hi
li
N
e
t,
a
l
ig
ht
w
e
ig
ht
de
e
p
l
e
a
r
ni
ng mode
l
opt
im
iz
e
d f
r
om
M
obi
le
N
e
tV2
th
r
ough
pr
uni
ng
a
nd
f
in
e
-
tu
ni
ng
f
or
c
hi
li
le
a
f
di
s
e
a
s
e
c
l
a
s
s
if
i
c
a
ti
on
;
ii
)
a
c
hi
e
ve
s
a
va
li
da
ti
on
a
c
c
ur
a
c
y
of
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
obi
le
C
hi
li
N
e
t:
c
onv
ol
ut
io
nal
ne
ur
al
ne
tw
or
k
f
or
c
hi
li
l
e
av
e
s
c
la
s
s
if
ic
at
io
n
(
M
ar
is
c
ha E
lv
e
ny
)
3759
96.48%
,
out
pe
r
f
o
r
m
in
g
s
e
ve
r
a
l
s
ta
te
-
of
-
th
e
-
a
r
t
C
N
N
s
a
c
r
os
s
f
iv
e
di
s
e
a
s
e
c
la
s
s
e
s
;
ii
i)
e
va
lu
a
te
s
m
ul
ti
pl
e
opt
im
iz
a
ti
on
a
lg
or
it
hm
s
a
nd
le
a
r
ni
ng
r
a
te
s
c
h
e
dul
e
r
s
,
id
e
nt
if
yi
n
g
A
da
m
W
w
it
h
P
ol
ynomi
a
lDe
c
a
y
a
s
th
e
m
os
t
s
ta
bl
e
a
nd
a
c
c
ur
a
te
c
om
bi
na
ti
on
;
iv
)
e
ns
ur
e
s
r
e
a
l
-
ti
m
e
de
pl
oym
e
nt
f
e
a
s
ib
il
it
y
on
lo
w
-
pow
e
r
e
dge
de
vi
c
e
s
s
uc
h a
s
R
a
s
pbe
r
r
y P
i
a
nd s
m
a
r
tp
hone
s
, pr
om
ot
in
g a
c
c
e
s
s
ib
il
it
y i
n r
e
s
our
c
e
-
li
m
it
e
d a
gr
ic
ul
tu
r
a
l
e
nvi
r
on
m
e
nt
s
;
a
nd v)
c
ont
r
ib
ut
e
s
t
o pr
e
c
is
io
n f
a
r
m
in
g a
nd S
D
G
s
by e
na
bl
in
g e
a
r
ly
di
s
e
a
s
e
d
e
te
c
ti
on, r
e
duc
in
g c
r
op l
os
s
,
a
nd
s
uppor
ti
ng s
m
a
ll
hol
de
r
f
a
r
m
e
r
s
.
T
he
r
e
m
a
in
de
r
of
t
hi
s
pa
pe
r
i
s
or
ga
ni
z
e
d a
s
:
s
e
c
ti
on 2 de
s
c
r
ib
e
s
t
he
r
e
s
e
a
r
c
h m
e
th
odol
ogy, inc
lu
di
ng
da
ta
s
e
t
pr
e
pa
r
a
ti
on,
m
ode
l
pr
uni
ng,
a
nd
f
in
e
-
tu
n
in
g
pr
oc
e
s
s
e
s
. S
e
c
ti
on
3
pr
e
s
e
nt
s
th
e
r
e
s
ul
ts
a
nd
di
s
c
us
s
io
ns
,
in
c
lu
di
ng
c
om
pa
r
is
ons
w
it
h
e
xi
s
ti
ng
C
N
N
s
,
e
va
lu
a
ti
on
of
va
r
io
us
opt
im
iz
e
r
s
,
a
nd
th
e
e
f
f
e
c
t
of
di
f
f
e
r
e
nt
le
a
r
ni
ng r
a
te
s
c
he
dul
e
r
s
. S
e
c
ti
on 4 pr
ovi
de
s
t
he
c
onc
lu
s
io
ns
a
nd i
m
pl
ic
a
ti
ons
of
t
he
p
r
opos
e
d M
obi
le
C
hi
li
N
e
t
m
ode
l
f
or
s
m
a
r
t
a
gr
ic
ul
tu
r
a
l
s
ys
te
m
s
.
T
a
bl
e
1. R
e
s
e
a
r
c
h r
e
la
t
e
d t
o l
e
a
ve
s
c
la
s
s
if
ic
a
ti
on
M
e
t
hods
N
um
be
r
of
c
l
a
s
s
e
s
A
c
c
ur
a
c
y
(%)
V
G
G
N
e
t
[
19]
3
97.00
S
V
M
+R
N
N
[
20]
5
92.10
G
L
C
M
+K
N
N
[
21]
2
94.00
F
i
ne
t
uni
ng S
huf
f
l
e
N
e
t
[
22]
2
99.30
I
nc
e
pt
i
on V
3
[
23]
4
93.00
X
c
e
pt
i
on
[
24]
5
79.56
E
f
f
i
c
i
e
nt
L
e
a
f
N
e
t
B
4
[
25]
5
92.00
E
f
f
i
c
e
nt
N
e
t
[
26]
4
91.00
2.
M
E
T
H
O
D
T
hi
s
r
e
s
e
a
r
c
h
a
im
s
to
de
ve
lo
p
a
n
opt
im
a
l
C
N
N
m
ode
l
w
it
h
be
t
te
r
a
c
c
ur
a
c
y
a
nd
f
a
s
te
r
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
.
T
he
r
e
s
e
a
r
c
h
is
di
vi
de
d
in
to
f
our
m
a
in
s
te
ps
:
da
ta
s
e
t
pr
e
pa
r
a
ti
on,
te
s
ti
ng
e
xi
s
ti
ng
C
N
N
m
ode
ls
,
pr
uni
ng
th
e
be
s
t
-
pe
r
f
or
m
in
g
C
N
N
,
a
nd
f
in
e
-
tu
ni
ng
th
e
hype
r
pa
r
a
m
e
te
r
s
of
th
e
pr
une
d
C
N
N
m
ode
l.
T
h
e
m
e
th
odol
ogy us
e
d i
n t
hi
s
s
tu
dy i
s
i
ll
us
tr
a
te
d i
n F
ig
ur
e
1
.
F
ig
ur
e
1. R
e
s
e
a
r
c
h
m
e
th
odol
ogy
A
s
s
how
n
in
F
ig
ur
e
1,
th
e
da
ta
s
e
t
is
s
pl
it
in
to
80%
f
or
tr
a
in
in
g
a
nd
20%
f
or
va
li
da
ti
on.
T
hi
s
da
t
a
is
us
e
d
to
te
s
t
e
xi
s
ti
ng
C
N
N
m
ode
ls
s
uc
h
a
s
M
obi
le
N
e
t
[
27]
,
R
e
s
N
e
t
[
28]
,
V
G
G
N
e
t
[
29]
,
A
le
xne
t
[
30]
,
a
nd
S
huf
f
le
N
e
t
[
31]
,
w
hi
c
h
ha
ve
pr
ove
n
e
f
f
e
c
ti
ve
in
c
la
s
s
if
yi
ng
le
a
f
di
s
e
a
s
e
s
.
T
he
e
xi
s
ti
ng
C
N
N
m
ode
l
w
it
h
th
e
be
s
t
a
c
c
ur
a
c
y
is
s
e
le
c
te
d
f
or
pr
uni
ng
to
c
r
e
a
te
a
m
or
e
c
om
pa
c
t
a
nd
f
a
s
te
r
c
la
s
s
if
ic
a
ti
on
m
ode
l.
T
he
pr
une
d
m
ode
l
is
th
e
n
f
in
e
-
tu
ne
d
to
f
u
r
th
e
r
im
pr
ove
a
c
c
ur
a
c
y,
r
e
s
ul
ti
ng
in
a
f
a
s
te
r
a
nd
m
or
e
a
c
c
ur
a
te
m
ode
l
f
or
c
la
s
s
if
yi
ng c
hi
li
l
e
a
f
di
s
e
a
s
e
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 5, O
c
to
be
r
2025
:
3757
-
3770
3760
2.1. Dat
as
e
t
s
T
he
da
t
a
s
e
t
ut
il
iz
e
d
in
th
i
s
r
e
s
e
a
r
c
h
i
s
c
om
pr
is
e
d
of
im
a
ge
s
de
p
ic
ti
ng
va
r
io
us
di
s
e
a
s
e
s
th
a
t
a
f
f
e
c
t
r
e
d
c
hi
li
le
a
ve
s
,
obt
a
in
e
d
f
r
om
M
e
nde
le
y
d
a
ta
.
T
he
s
e
im
a
ge
s
a
r
e
c
a
te
gor
iz
e
d
in
to
f
iv
e
di
s
ti
nc
t
di
s
e
a
s
e
c
la
s
s
e
s
[
24]
.
T
he
da
ta
s
e
t
ha
s
be
e
n
pr
oc
e
s
s
e
d
in
to
two
va
r
ia
nt
s
:
a
ugm
e
nt
e
d
a
nd
non
-
a
ugm
e
nt
e
d.
I
ni
ti
a
ll
y,
th
e
da
ta
s
e
t
c
ont
a
in
e
d
531
im
a
ge
s
;
how
e
ve
r
,
a
f
te
r
a
ppl
yi
ng
da
ta
a
ugm
e
nt
a
ti
on
te
c
hni
que
s
,
th
e
da
ta
s
e
t
e
xpa
nd
e
d
s
ig
ni
f
ic
a
nt
ly
t
o 2,128 im
a
ge
s
. T
a
bl
e
2 pr
ovi
de
s
a
de
ta
il
e
d br
e
a
k
dow
n of
t
he
a
ugm
e
nt
e
d da
ta
s
e
t.
T
a
bl
e
2
.
A
ugm
e
nt
e
d
d
a
ta
s
e
t
C
l
a
s
s
n
a
m
e
I
m
a
ge
T
r
a
i
ni
ng
V
a
l
i
da
t
i
on
T
ot
a
l
P
ow
de
r
y
m
i
l
de
w
486
122
608
H
e
a
l
t
hy
l
eaf
221
55
276
M
ur
da
c
om
pl
e
x (
m
i
t
e
s
, t
hr
i
ps
)
342
86
428
L
e
a
f
s
pot
(
C
e
r
c
os
por
a
)
326
82
408
N
ut
r
i
e
nt
d
e
f
i
c
i
e
nc
y
327
81
408
T
ot
a
l
1
,
702
426
2
,
128
T
a
bl
e
2
out
li
ne
s
th
e
di
s
tr
ib
ut
io
n
of
th
e
a
ugm
e
nt
e
d
da
ta
s
e
t
us
e
d
in
th
is
s
tu
dy
to
c
la
s
s
if
y
di
s
e
a
s
e
s
a
f
f
e
c
ti
ng
r
e
d
c
hi
li
le
a
ve
s
.
T
he
da
ta
s
e
t
e
nc
om
pa
s
s
e
s
f
iv
e
di
s
ti
nc
t
di
s
e
a
s
e
c
a
te
gor
ie
s
:
pow
de
r
y
m
il
de
w
,
he
a
lt
hy
le
a
f
,
m
ur
da
c
om
pl
e
x
(
m
it
e
s
a
nd
th
r
ip
s
)
,
le
a
f
s
pot
(
C
e
r
c
o
s
por
a
)
,
a
nd
nut
r
ie
nt
de
f
ic
ie
nc
y
.
E
a
c
h
di
s
e
a
s
e
c
la
s
s
is
r
e
pr
e
s
e
nt
e
d
by
a
s
e
t
of
im
a
ge
s
,
w
hi
c
h
ha
ve
b
e
e
n
f
ur
th
e
r
di
vi
de
d
in
to
tr
a
in
in
g
a
nd
va
li
da
ti
on
s
ubs
e
t
s
.
A
f
te
r
a
ugm
e
nt
a
ti
on,
th
e
da
ta
s
e
t
to
ta
ls
2,128
im
a
ge
s
,
w
it
h
1,702
de
s
ig
na
te
d
f
or
tr
a
in
in
g
a
nd
426
f
o
r
va
li
da
ti
on.
T
he
or
ig
in
a
l
im
a
ge
s
, w
hi
c
h ha
ve
unde
r
gone
a
ugm
e
nt
a
ti
on, a
r
e
de
pi
c
te
d i
n F
ig
ur
e
2.
F
ig
ur
e
2.
D
a
ta
a
ugm
e
nt
a
ti
on
2.2
.
M
ob
il
e
N
e
t
V
2
a
r
c
h
it
e
c
t
u
r
e
T
he
M
obi
le
N
e
tV2
a
r
c
hi
te
c
tu
r
e
ut
il
iz
e
s
de
pt
hw
i
s
e
s
e
p
a
r
a
bl
e
c
onvolut
io
ns
a
nd
in
ve
r
te
d
r
e
s
id
ua
l
bl
oc
ks
w
it
h
a
li
ne
a
r
bot
tl
e
ne
c
k,
de
s
ig
ne
d
to
im
pr
ove
c
om
put
a
ti
ona
l
e
f
f
ic
ie
nc
y
a
nd
pe
r
f
or
m
a
nc
e
on
de
vi
c
e
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
obi
le
C
hi
li
N
e
t:
c
onv
ol
ut
io
nal
ne
ur
al
ne
tw
or
k
f
or
c
hi
li
l
e
av
e
s
c
la
s
s
if
ic
at
io
n
(
M
ar
is
c
ha E
lv
e
ny
)
3761
w
it
h
li
m
it
e
d
r
e
s
our
c
e
s
.
I
n
th
e
in
it
ia
l
s
ta
ge
,
a
s
t
a
nda
r
d
c
onvolut
io
na
l
la
ye
r
w
it
h
32
f
il
te
r
s
a
nd
a
s
tr
id
e
of
2
i
s
us
e
d
to
e
xt
r
a
c
t
in
it
ia
l
f
e
a
tu
r
e
s
f
r
om
th
e
in
put
im
a
ge
.
T
he
ne
twor
k
th
e
n
pr
ogr
e
s
s
e
s
th
r
ough
a
s
e
r
ie
s
of
bot
tl
e
ne
c
k
bl
oc
ks
,
s
ta
r
ti
ng
w
it
h
a
n
e
xp
a
ns
io
n
f
a
c
to
r
of
1,
r
e
s
ul
ti
ng
in
16
f
il
te
r
s
w
it
hout
c
ha
ngi
ng
th
e
out
put
s
iz
e
.
A
f
te
r
th
a
t,
a
bot
tl
e
ne
c
k
bl
oc
k
w
it
h
a
n
e
xpa
n
s
io
n
f
a
c
to
r
of
6
is
r
e
pe
a
te
dl
y
a
ppl
ie
d,
w
he
r
e
th
e
f
e
a
tu
r
e
s
iz
e
is
t
e
m
por
a
r
il
y e
xpa
nde
d be
f
or
e
be
in
g c
om
pr
e
s
s
e
d a
ga
in
. T
hi
s
p
r
oc
e
s
s
r
e
s
ul
ts
i
n c
ha
nge
s
i
n output
di
m
e
ns
io
ns
f
r
om
112
×
112
to
56
×
56,
28
×
28,
14
×
14, a
nd
f
in
a
ll
y
7
×
7,
w
it
h
t
he
num
be
r
of
f
il
te
r
s
gr
a
dua
ll
y
in
c
r
e
a
s
in
g
f
r
om
24,
32,
64,
96,
160,
to
320.
A
f
te
r
p
a
s
s
in
g
th
r
ough
th
e
f
in
a
l
c
onvolut
io
na
l
la
ye
r
w
it
h
1,280
f
il
te
r
s
,
a
ve
r
a
g
e
pool
in
g
is
a
ppl
ie
d
to
r
e
duc
e
th
e
f
e
a
tu
r
e
di
m
e
ns
io
ns
to
1
×
1.
T
he
f
in
a
l
la
ye
r
is
a
f
ul
ly
c
onne
c
te
d
la
ye
r
w
it
h
5
ne
ur
ons
,
w
hi
c
h
i
s
us
e
d
to
c
la
s
s
if
y
th
e
5
c
l
a
s
s
e
s
of
c
hi
li
le
a
f
di
s
e
a
s
e
s
.
T
he
de
ta
il
e
d
a
r
c
hi
te
c
tu
r
e
of
M
obi
le
N
e
tV2 f
or
c
hi
li
l
e
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on i
s
s
how
n i
n
T
a
bl
e
3.
T
a
bl
e
3
. M
obi
le
N
e
t
V
2
a
r
c
hi
t
e
c
tu
r
e
L
a
ye
r
t
ype
t
C
n
s
I
nput
s
i
z
e
O
ut
put
s
i
z
e
C
onv2D
-
32
1
2
224
×
224
×
3
112
×
112
×
32
B
ot
t
l
e
ne
c
k
1
16
1
1
112
×
112
×
32
112
×
112
×
16
B
ot
t
l
e
ne
c
k
6
24
2
2
112
×
112
×
16
56
×
56
×
24
B
ot
t
l
e
ne
c
k
6
32
3
2
56
×
56
×
24
28
×
28
×
32
B
ot
t
l
e
ne
c
k
6
64
4
2
28
×
28
×
32
14
×
14
×
64
B
ot
t
l
e
ne
c
k
6
96
3
1
14
×
14
×
64
14
×
14
×
96
B
ot
t
l
e
ne
c
k
6
160
3
2
14
×
14
×
96
7
×
7
×
160
B
ot
t
l
e
ne
c
k
6
320
1
1
7
×
7
×
160
7
×
7
×
320
C
onv2D
-
1280
1
1
7
×
7
×
320
7
×
7
×
1280
A
vg
p
ool
i
ng
-
-
1
-
7
×
7
×
1280
1
×
1
×
1280
F
ul
l
y c
onne
c
t
e
d
(
FC
)
-
5
1
-
1
×
1
×
1280
1
×
1
×
5
I
n T
a
bl
e
3, s
e
ve
r
a
l
im
por
ta
nt
pa
r
a
m
e
te
r
s
e
xpl
a
in
t
he
c
onf
ig
ur
a
ti
on of
e
a
c
h l
a
ye
r
:
t
(
e
xpa
ns
io
n f
a
c
to
r
)
in
di
c
a
te
s
how
m
u
c
h
th
e
num
be
r
of
c
h
a
nne
ls
w
il
l
b
e
e
xpa
nd
e
d
be
f
or
e
pe
r
f
or
m
in
g
th
e
de
pt
hw
is
e
c
onvolut
io
n;
c
(
out
put
c
ha
nne
ls
)
r
e
pr
e
s
e
nt
s
th
e
num
be
r
of
out
put
c
ha
nne
l
s
f
r
om
e
a
c
h
la
y
e
r
or
bl
oc
k;
n
(
num
be
r
of
r
e
pe
a
ts
)
s
how
s
how
m
a
ny
ti
m
e
s
th
e
bot
tl
e
ne
c
k
bl
oc
k
i
s
r
e
pe
a
te
d
to
in
c
r
e
a
s
e
c
om
pl
e
xi
ty
a
nd
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
c
a
pa
bi
li
ty
;
a
nd s
(
s
tr
id
e
)
in
di
c
a
t
e
s
th
e
s
hi
f
t
s
te
p
of
th
e
ke
r
ne
l
du
r
in
g
th
e
c
onvolut
io
n,
w
hi
c
h
a
f
f
e
c
ts
th
e
s
pa
ti
a
l
s
iz
e
of
th
e
out
put
.
A
s
tr
id
e
va
lu
e
g
r
e
a
te
r
th
a
n
1
(
e
.g.,
s
=
2)
w
il
l
r
e
s
ul
t
in
a
r
e
duc
ti
on
of
th
e
r
e
s
ol
ut
io
n
(
dow
ns
a
m
pl
in
g)
in
th
e
out
put
,
w
he
r
e
a
s
a
s
tr
id
e
of
1
ke
e
ps
t
he
out
put
s
iz
e
th
e
s
a
m
e
a
s
th
e
in
put
.
T
he
s
e
pa
r
a
m
e
te
r
c
om
bi
na
ti
ons
he
lp
to
unde
r
s
ta
nd
th
e
s
tr
uc
tu
r
e
a
nd
f
unc
ti
ona
li
ty
of
e
a
c
h
la
ye
r
in
th
e
M
obi
le
N
e
tV2
a
r
c
hi
te
c
tu
r
e
. T
he
bot
tl
e
ne
c
k i
n
M
obi
le
N
e
tV2 is
i
ll
u
s
tr
a
te
d a
s
s
h
ow
n i
n F
ig
ur
e
3.
F
ig
ur
e
3. I
ll
us
tr
a
ti
on of
t
he
b
ot
tl
e
ne
c
k i
n M
obi
le
N
e
tV2
T
he
bot
tl
e
ne
c
k
r
e
s
id
ua
l
bl
oc
k
in
M
obi
le
N
e
tV2
h
a
s
di
f
f
e
r
e
nt
c
ha
r
a
c
te
r
is
ti
c
s
d
e
pe
ndi
ng
on
th
e
s
tr
id
e
va
lu
e
. A
t
s
tr
id
e
=
1, a
s
s
how
n i
n F
ig
ur
e
3, t
he
r
e
i
s
a
br
a
nc
hi
ng pr
oc
e
s
s
w
h
e
r
e
t
he
f
ir
s
t
br
a
nc
h s
im
pl
y pa
s
s
e
s
t
he
or
ig
in
a
l
in
put
w
it
hout
a
ny
c
ha
nge
s
,
w
hi
le
th
e
s
e
c
ond
br
a
nc
h
pe
r
f
or
m
s
s
e
ve
r
a
l
ope
r
a
ti
ons
.
T
he
s
e
c
ond
br
a
nc
h
c
ons
is
ts
of
a
1
×
1
c
onvolut
io
n
w
it
h
a
R
e
L
U
6
a
c
ti
va
ti
on
f
unc
ti
on,
f
ol
lo
w
e
d
by
a
3
×
3
de
pt
hw
is
e
c
onvolut
io
n
w
it
h
a
R
e
L
U
6
a
c
ti
va
ti
on
f
unc
ti
on,
a
nd
th
e
n
a
not
he
r
1
×
1
c
onv
ol
ut
io
n
w
it
hout
a
ny
a
c
ti
va
ti
on
f
unc
ti
on.
T
he
s
e
two
br
a
nc
he
s
a
r
e
th
e
n
s
um
m
e
d
to
pr
oduc
e
th
e
f
in
a
l
out
put
of
th
e
bl
oc
k.
O
n
th
e
ot
he
r
ha
nd,
a
t
s
tr
id
e
=
2,
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 5, O
c
to
be
r
2025
:
3757
-
3770
3762
pr
oc
e
s
s
is
s
li
ght
ly
di
f
f
e
r
e
nt
a
s
it
a
im
s
to
dow
ns
a
m
pl
e
th
e
f
e
a
tu
r
e
s
.
A
t
th
is
s
tr
id
e
va
lu
e
,
th
e
in
it
ia
l
c
onvolut
io
n
la
ye
r
in
th
e
s
e
c
ond
br
a
nc
h
r
e
m
a
in
s
th
e
s
a
m
e
,
but
th
e
3
×
3
de
pt
hw
is
e
c
onvolut
io
n
w
it
h
s
tr
id
e
2
is
a
ppl
ie
d
to
r
e
duc
e
th
e
s
p
a
ti
a
l
s
iz
e
of
th
e
out
put
f
e
a
tu
r
e
s
.
A
ddi
ti
ona
ll
y,
th
e
f
ir
s
t
br
a
nc
h
doe
s
not
m
e
r
e
ly
p
a
s
s
th
e
in
put
but
a
ls
o
unde
r
goe
s
a
dj
us
tm
e
nt
s
to
m
a
tc
h
th
e
di
m
e
ns
io
n
s
of
th
e
s
e
c
ond
br
a
nc
h.
T
he
out
put
s
of
bot
h
br
a
nc
he
s
a
r
e
not
di
r
e
c
tl
y s
um
m
e
d but a
r
e
c
om
bi
ne
d a
t
th
e
f
in
a
l
s
ta
ge
t
o pr
od
uc
e
m
or
e
c
om
pa
c
t
a
nd de
n
s
e
f
e
a
tu
r
e
s
.
2.
3
.
P
r
op
os
e
d
c
on
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
or
k
a
r
c
h
it
e
c
t
u
r
e
T
he
pr
opos
e
d
C
N
N
a
r
c
hi
te
c
tu
r
e
is
th
e
r
e
s
ul
t
of
p
r
uni
ng
th
e
M
obi
le
N
e
tV2
m
ode
l,
w
hi
c
h
pr
e
vi
ous
ly
s
how
e
d
th
e
be
s
t
a
c
c
ur
a
c
y
in
c
la
s
s
if
yi
ng
c
hi
li
le
a
f
di
s
e
a
s
e
s
.
T
h
e
pr
uni
ng
pr
oc
e
s
s
w
a
s
c
a
r
r
ie
d
out
by
r
e
duc
in
g
th
e
num
be
r
of
bot
tl
e
ne
c
k
la
ye
r
s
in
M
obi
le
N
e
tV2,
f
r
om
s
e
ve
n
la
ye
r
s
to
onl
y
th
r
e
e
bot
tl
e
ne
c
k
l
a
ye
r
s
w
it
h
di
f
f
e
r
e
nt
out
put
c
ha
nne
ls
.
T
hi
s
s
t
e
p
a
im
s
to
de
c
r
e
a
s
e
c
o
m
put
a
ti
ona
l
c
os
ts
w
it
hout
s
a
c
r
if
ic
in
g
m
ode
l
pe
r
f
or
m
a
nc
e
.
A
f
te
r
th
e
la
ye
r
r
e
duc
ti
on,
th
e
pa
r
a
m
e
te
r
s
t,
c
,
n,
a
nd
s
w
e
r
e
r
e
c
onf
ig
ur
e
d
to
de
te
r
m
in
e
th
e
opt
im
a
l
a
r
c
hi
te
c
tu
r
e
s
e
tu
p. A
n i
ll
us
tr
a
ti
on of
t
he
M
obi
le
N
e
tV2
pr
uni
ng pr
oc
e
s
s
i
s
s
how
n i
n F
ig
ur
e
4.
F
ig
ur
e
4.
I
ll
us
tr
a
ti
on
of
t
h
e
M
obi
l
e
N
e
t
V
2
pr
u
ni
ng pr
oc
e
s
s
A
s
s
how
n
in
F
ig
ur
e
4,
e
a
c
h
pa
r
a
m
e
te
r
c
om
bi
na
ti
on
w
a
s
te
s
te
d
a
nd
tr
a
in
e
d
ove
r
10
e
poc
hs
.
T
he
c
om
bi
na
ti
on
w
it
h
th
e
be
s
t
va
li
da
ti
on
a
c
c
ur
a
c
y
w
a
s
th
e
n
r
e
tr
a
in
e
d
f
or
up
to
50
e
poc
hs
to
e
va
lu
a
te
th
e
f
in
a
l
pe
r
f
or
m
a
nc
e
of
th
e
r
e
s
ul
ti
ng
C
N
N
m
ode
l.
B
a
s
e
d
on
th
e
s
e
e
x
pe
r
im
e
nt
s
,
w
e
s
uc
c
e
s
s
f
ul
ly
id
e
nt
if
ie
d
th
e
be
s
t
c
om
bi
na
ti
on,
w
hi
c
h
pr
oduc
e
d
a
n
opt
im
a
l
C
N
N
a
r
c
hi
te
c
tu
r
e
f
or
c
hi
li
le
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on,
w
hi
c
h
w
e
na
m
e
d M
obi
le
C
hi
li
N
e
t.
T
he
c
om
pl
e
te
s
tr
uc
tu
r
e
of
t
he
M
obi
le
C
hi
li
N
e
t
a
r
c
hi
te
c
tu
r
e
i
s
pr
e
s
e
nt
e
d i
n
T
a
bl
e
4.
T
a
bl
e
4. M
obi
le
C
hi
li
N
e
t
a
r
c
hi
te
c
tu
r
e
L
a
ye
r
t
ype
t
C
n
s
I
nput
s
i
z
e
O
ut
put
s
i
z
e
C
onv2
D
-
32
1
2
3
×
128
×
128
32
×
64
×
64
B
ot
t
l
e
ne
c
k
1
24
1
1
32
×
64
×
64
24
×
64
×
64
B
ot
t
l
e
ne
c
k
6
32
3
2
24
×
64
×
64
32
×
32
×
32
B
ot
t
l
e
ne
c
k
6
96
4
2
32
×
32
×
32
96
×
16
×
16
C
onv2D
-
1
,
280
1
1
96
×
16
×
16
1280
×
16
×
16
A
da
pt
i
ve
A
vgP
ool
2D
-
-
1
-
1280
×
16
×
16
1280
×
1
×
1
F
ul
l
y
c
onne
c
t
e
d (
l
i
ne
a
r
)
-
5
1
-
1280
5
T
he
M
obi
le
C
hi
li
N
e
t
a
r
c
hi
te
c
tu
r
e
,
s
how
n
in
T
a
bl
e
4
,
c
ons
is
t
s
of
s
e
v
e
r
a
l
ke
y
la
y
e
r
s
de
s
ig
ne
d
to
e
xt
r
a
c
t
im
por
ta
nt
f
e
a
tu
r
e
s
f
r
om
c
hi
li
le
a
f
im
a
g
e
s
.
T
he
f
ir
s
t
la
y
e
r
is
a
C
onv2D
l
a
ye
r
w
it
h
32
out
put
c
ha
nne
l
s
a
nd
a
s
tr
id
e
of
2,
s
e
r
vi
ng
a
s
th
e
in
it
ia
l
la
ye
r
to
r
e
duc
e
th
e
i
nput
im
a
ge
s
iz
e
f
r
om
128
×
128
to
64
×
64
a
nd
e
xt
r
a
c
t
ba
s
ic
f
e
a
tu
r
e
s
.
N
e
xt
,
bot
tl
e
ne
c
k
la
y
e
r
s
a
r
e
us
e
d
to
opt
im
iz
e
th
e
num
be
r
of
p
a
r
a
m
e
te
r
s
by
ut
il
iz
in
g
th
e
c
onf
ig
ur
a
ti
on
of
th
e
e
xpa
ns
io
n
f
a
c
to
r
(
t)
,
th
e
num
be
r
of
out
put
f
il
te
r
s
(
C
)
,
th
e
num
be
r
of
r
e
pe
a
ts
(
n)
,
a
nd
th
e
s
tr
id
e
(
s
)
.
I
ni
ti
a
ll
y,
a
bot
tl
e
ne
c
k
w
it
h
a
n
e
xpa
n
s
io
n
f
a
c
to
r
of
1
a
nd
24
f
i
lt
e
r
s
is
a
ppl
ie
d
w
it
hout
c
h
a
ngi
ng
th
e
out
put
s
iz
e
. T
he
n, a
bot
tl
e
ne
c
k w
it
h a
n e
xpa
ns
io
n f
a
c
to
r
of
6
a
n
d 32 f
i
lt
e
r
s
i
s
us
e
d t
hr
e
e
t
i
m
e
s
w
it
h a
s
tr
id
e
of
2,
r
e
duc
in
g
th
e
f
e
a
tu
r
e
s
iz
e
to
32
×
32.
A
s
im
il
a
r
pr
oc
e
s
s
is
a
ppl
ie
d
in
th
e
ne
xt
la
ye
r
w
i
th
96
f
il
te
r
s
a
nd
4
r
e
pe
a
ts
, f
ur
th
e
r
r
e
duc
in
g t
he
s
pa
ti
a
l
s
iz
e
t
o 16
×
16.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
obi
le
C
hi
li
N
e
t:
c
onv
ol
ut
io
nal
ne
ur
al
ne
tw
or
k
f
or
c
hi
li
l
e
av
e
s
c
la
s
s
if
ic
at
io
n
(
M
ar
is
c
ha E
lv
e
ny
)
3763
A
f
te
r
w
a
r
d,
a
C
onv2D
la
ye
r
w
it
h
1,280
out
pu
t
c
ha
nne
ls
is
a
ppl
ie
d
to
e
nr
ic
h
th
e
f
e
a
tu
r
e
r
e
pr
e
s
e
nt
a
ti
on,
f
ol
lo
w
e
d
by
a
ve
r
a
ge
pool
in
g
us
in
g
A
da
pt
iv
e
A
vgP
ool
2D
,
r
e
duc
in
g
th
e
f
e
a
tu
r
e
s
iz
e
to
1
×
1.
F
in
a
ll
y,
a
f
ul
ly
c
onne
c
te
d
la
ye
r
w
it
h
5
ne
ur
ons
is
us
e
d
a
s
th
e
c
la
s
s
if
ic
a
ti
on
la
ye
r
to
de
te
r
m
in
e
th
e
c
hi
li
le
a
f
di
s
e
a
s
e
c
la
s
s
e
s
.
T
he
num
be
r
of
ne
ur
ons
in
th
is
la
y
e
r
c
or
r
e
s
ponds
to
th
e
num
be
r
of
c
la
s
s
e
s
to
be
c
la
s
s
if
ie
d,
w
hi
c
h i
s
5 t
ype
s
of
c
hi
li
l
e
a
f
di
s
e
a
s
e
s
.
W
it
h
th
is
c
om
bi
na
ti
on
of
la
ye
r
s
,
M
obi
le
C
hi
li
N
e
t
a
c
hi
e
ve
s
opt
i
m
a
l
pe
r
f
or
m
a
nc
e
in
c
hi
li
le
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on
w
hi
le
m
a
in
ta
in
in
g
a
lo
w
e
r
num
be
r
of
pa
r
a
m
e
t
e
r
s
c
om
pa
r
e
d
to
th
e
s
ta
nda
r
d
M
obi
le
N
e
tV2
a
r
c
hi
te
c
tu
r
e
,
m
a
ki
ng
it
m
or
e
e
f
f
ic
ie
nt
in
c
om
put
a
ti
on
a
nd
m
e
m
or
y
us
a
ge
.
S
e
ve
r
a
l
r
e
a
s
on
s
w
hy
a
m
ode
l
w
it
h
f
e
w
e
r
la
ye
r
s
c
a
n
be
m
or
e
a
c
c
ur
a
te
in
c
lu
de
:
i)
r
e
duc
e
d
o
ve
r
f
it
ti
ng:
C
N
N
s
w
it
h
de
e
pe
r
la
ye
r
s
a
r
e
pr
one
to
ove
r
f
it
ti
ng
a
s
th
e
y
te
nd
to
le
a
r
n
to
o
m
a
ny
p
a
tt
e
r
ns
f
r
om
th
e
tr
a
in
in
g
da
ta
,
in
c
lu
di
ng
noi
s
e
.
A
s
im
pl
e
r
m
ode
l
c
a
n
c
a
pt
ur
e
m
or
e
ge
ne
r
a
l
a
nd
im
por
ta
nt
pa
tt
e
r
ns
,
of
te
n
le
a
di
ng
to
be
tt
e
r
a
c
c
ur
a
c
y
[
32]
;
a
nd
ii
)
m
or
e
opt
im
a
l
pa
r
a
m
e
te
r
tu
ni
ng:
th
i
s
a
ll
ow
s
f
or
m
or
e
e
f
f
ic
ie
nt
pa
r
a
m
e
te
r
op
ti
m
iz
a
ti
on,
a
s
th
e
s
m
a
ll
e
r
pa
r
a
m
e
te
r
s
pa
c
e
is
e
a
s
ie
r
f
or
opt
im
iz
a
ti
on a
lg
or
it
hm
s
t
o e
xpl
or
e
[
33]
.
3.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
T
hi
s
s
e
c
ti
on
pr
e
s
e
nt
s
th
e
r
e
s
ul
ts
f
r
om
te
s
ti
ng
th
e
M
obi
le
C
h
il
iNe
t
m
ode
l,
w
hi
c
h
ha
s
unde
r
gone
s
e
ve
r
a
l
s
t
a
ge
s
of
opt
im
iz
a
ti
on.
T
he
r
e
s
e
a
r
c
h
be
ga
n
by
te
s
ti
ng
va
r
io
us
e
xi
s
ti
ng
C
N
N
a
r
c
hi
te
c
tu
r
e
s
,
f
ol
lo
w
e
d
by
th
e
pr
uni
ng
pr
oc
e
s
s
on
M
obi
le
N
e
tV2,
a
nd
c
on
c
lu
de
d
w
i
th
f
in
e
-
tu
ni
ng
th
e
pa
r
a
m
e
te
r
s
.
T
he
a
c
c
ur
a
c
y
c
om
pa
r
is
on
r
e
s
ul
ts
of
e
a
c
h
opt
im
iz
e
r
u
s
e
d
dur
in
g
th
e
tr
a
in
in
g
of
M
obi
le
C
hi
li
N
e
t,
a
s
w
e
ll
a
s
th
e
im
pl
e
m
e
nt
a
ti
on of
va
r
io
us
l
e
a
r
ni
ng
r
a
te
s
c
he
dul
e
r
s
, a
r
e
out
li
ne
d i
n t
he
r
e
le
va
nt
t
a
bl
e
s
. T
he
di
s
c
us
s
io
n f
oc
us
e
s
on
th
e
im
pa
c
t
of
e
a
c
h
m
e
th
od
on
im
pr
ovi
ng
a
c
c
ur
a
c
y
a
nd
th
e
s
ta
bi
li
ty
of
th
e
m
ode
l
in
c
hi
li
le
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on.
3.1
.
C
om
p
ar
is
on
o
f
e
xi
s
t
in
g
c
on
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
or
k
s
T
he
f
ir
s
t
s
te
p
in
th
is
r
e
s
e
a
r
c
h
w
a
s
ut
il
iz
in
g
e
xi
s
ti
ng C
N
N
a
r
c
hi
t
e
c
tu
r
e
s
to
c
la
s
s
if
y
c
hi
li
le
a
f
di
s
e
a
s
e
s
.
T
he
c
hi
li
le
a
f
di
s
e
a
s
e
da
ta
s
e
t,
w
hi
c
h
c
on
s
is
t
s
of
f
iv
e
c
la
s
s
e
s
,
w
a
s
us
e
d
a
s
tr
a
in
in
g
a
nd
va
li
da
ti
on
da
ta
f
or
th
e
e
xi
s
ti
ng
C
N
N
m
ode
ls
.
A
ll
e
xi
s
ti
ng
C
N
N
s
us
e
d
in
th
i
s
s
tu
dy
w
e
r
e
r
e
tr
a
in
e
d
on
th
e
d
a
ta
s
e
t
w
it
hout
us
in
g
w
e
ig
ht
s
f
r
om
tr
a
ns
f
e
r
le
a
r
ni
ng.
T
hi
s
w
a
s
done
to
id
e
nt
if
y
th
e
be
s
t
a
r
c
hi
te
c
tu
r
e
th
a
t
c
oul
d
be
f
ur
th
e
r
de
ve
lo
pe
d. T
r
a
in
in
g w
a
s
c
onduc
te
d us
in
g t
he
S
G
D
opt
im
iz
a
ti
o
n f
unc
ti
on w
it
h a
m
om
e
nt
um
of
0.9, a
l
e
a
r
ni
ng
r
a
te
of
0.01, a
nd a
ba
tc
h s
iz
e
of
64.
T
he
t
r
a
in
in
g r
e
s
ul
ts
of
t
he
e
xi
s
ti
ng C
N
N
m
ode
ls
a
r
e
s
how
n i
n T
a
bl
e
5
.
T
a
bl
e
5. A
c
c
ur
a
c
y
c
om
pa
r
is
on of
e
xi
s
ti
ng
C
N
N
s
M
e
t
hods
T
r
a
i
n
a
c
c
ur
a
c
y
(%)
V
a
l
i
da
t
i
on
a
c
c
ur
a
c
y
(%)
M
obi
l
e
N
e
t
V
2
85.01
89.43
R
e
s
N
e
t
50
84.13
86.61
V
G
G
16
74.32
76.29
A
l
e
x
N
et
67.21
68.30
S
huf
f
l
e
N
e
t
77.02
77.46
T
a
bl
e
5
s
how
s
th
a
t
th
e
M
obi
le
N
e
tV2
a
r
c
hi
te
c
tu
r
e
a
c
hi
e
ve
d
th
e
hi
ghe
s
t
a
c
c
ur
a
c
y
a
m
ong
th
e
e
va
lu
a
te
d
m
ode
l
s
,
w
it
h
85.01%
tr
a
in
in
g
a
c
c
ur
a
c
y
a
nd
89.
43%
va
li
da
ti
on
a
c
c
ur
a
c
y.
B
e
yond
it
s
b
e
tt
e
r
pe
r
f
or
m
a
nc
e
,
M
obi
le
N
e
tV2
w
a
s
s
e
le
c
te
d
a
s
th
e
ba
s
e
a
r
c
h
it
e
c
tu
r
e
due
to
it
s
li
ght
w
e
ig
ht
d
e
s
ig
n
a
nd
c
om
put
a
ti
ona
l
e
f
f
ic
ie
nc
y,
w
hi
c
h
a
r
e
c
r
it
ic
a
l
f
or
de
pl
oym
e
nt
i
n
r
e
s
our
c
e
-
li
m
it
e
d
e
nvi
r
onm
e
nt
s
.
T
he
m
ode
l
ut
il
iz
e
s
de
pt
hw
is
e
s
e
pa
r
a
bl
e
c
onvolut
io
ns
a
nd
in
ve
r
te
d
r
e
s
id
ua
l
bl
oc
ks
w
it
h
li
ne
a
r
bot
tl
e
ne
c
ks
,
s
ig
ni
f
ic
a
nt
ly
r
e
duc
in
g
th
e
num
be
r
of
pa
r
a
m
e
te
r
s
a
nd
c
om
put
a
ti
ona
l
c
os
t
w
it
hout
c
om
pr
om
is
in
g
r
e
pr
e
s
e
nt
a
ti
ona
l
c
a
pa
c
it
y.
T
he
s
e
c
ha
r
a
c
te
r
is
ti
c
s
m
a
ke
it
w
e
ll
-
s
ui
te
d
f
or
r
e
a
l
-
ti
m
e
a
ppl
ic
a
t
io
ns
on
lo
w
-
pow
e
r
de
vi
c
e
s
s
uc
h
a
s
R
a
s
pb
e
r
r
y
P
i
or
s
m
a
r
tp
hone
s
,
a
li
gni
ng
w
it
h
th
e
obj
e
c
ti
ve
of
th
is
r
e
s
e
a
r
c
h.
T
h
e
pr
uni
ng
pr
oc
e
s
s
,
c
om
bi
ne
d
w
it
h
hype
r
pa
r
a
m
e
te
r
tu
ni
ng,
le
d
to
th
e
de
ve
lo
pm
e
nt
of
a
m
or
e
c
om
pa
c
t
a
nd
e
f
f
ic
ie
nt
a
r
c
hi
te
c
tu
r
e
na
m
e
d
M
obi
le
C
hi
li
N
e
t.
T
he
c
om
pa
r
is
on be
twe
e
n M
obi
le
C
hi
li
N
e
t
a
nd
e
xi
s
ti
ng C
N
N
s
i
s
pr
e
s
e
nt
e
d i
n
T
a
bl
e
6.
T
a
bl
e
6. A
c
c
ur
a
c
y
c
om
pa
r
is
on of
M
obi
le
C
hi
li
N
e
t
a
nd
e
xi
s
ti
ng
C
N
N
s
M
e
t
hods
T
r
a
i
n
a
c
c
ur
a
c
y
(%)
V
a
l
i
da
t
i
on
a
c
c
ur
a
c
y
(%)
M
obi
l
e
N
e
t
V
2
85.01
89.43
R
e
s
N
e
t
50
84.13
86.61
V
G
G
16
74.32
76.29
A
l
e
x
N
et
67.21
68.30
S
huf
f
l
e
N
e
t
77.02
77.46
M
obi
l
e
C
hi
l
i
N
e
t
95.35
94.13
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 5, O
c
to
be
r
2025
:
3757
-
3770
3764
T
a
bl
e
6
s
how
s
th
a
t
a
f
te
r
th
e
pr
uni
ng
a
nd
f
in
e
-
tu
ni
ng
pr
oc
e
s
s
on
M
obi
le
N
e
tV2,
th
e
r
e
s
ul
ti
ng
a
r
c
hi
te
c
tu
r
e
,
M
obi
le
C
hi
li
N
e
t,
a
c
hi
e
ve
d
a
tr
a
in
in
g
a
c
c
ur
a
c
y
of
95.35%
a
nd
a
va
li
da
ti
on
a
c
c
ur
a
c
y
of
94.13%
.
T
hi
s
is
s
ig
ni
f
ic
a
nt
ly
be
tt
e
r
th
a
n
th
e
ot
h
e
r
C
N
N
a
r
c
hi
te
c
tu
r
e
s
te
s
te
d,
s
uc
h
a
s
R
e
s
N
e
t5
0, V
G
G
16,
A
le
xN
e
t,
a
nd
S
huf
f
le
N
e
t.
M
obi
le
C
hi
li
N
e
t
de
m
ons
tr
a
te
d
a
s
ubs
ta
nt
ia
l
pe
r
f
or
m
a
nc
e
im
pr
ove
m
e
nt
,
pa
r
ti
c
ul
a
r
ly
in
te
r
m
s
of
va
li
da
ti
on a
c
c
ur
a
c
y, w
hi
c
h i
ndi
c
a
t
e
s
be
tt
e
r
m
ode
l
ge
ne
r
a
li
z
a
ti
on f
or
c
la
s
s
if
yi
ng c
hi
li
l
e
a
f
di
s
e
a
s
e
s
.
3.2
.
R
e
s
u
lt
s
of
op
t
im
iz
e
r
t
u
n
in
g on
M
ob
il
e
C
h
il
iNe
t
O
ne
a
ppr
oa
c
h
to
im
pr
ovi
ng
C
N
N
a
c
c
ur
a
c
y
is
tu
ni
ng
th
e
o
pt
im
iz
e
r
f
unc
ti
on
[
33]
,
[
34]
.
I
n
th
is
r
e
s
e
a
r
c
h,
w
e
te
s
t
e
d
va
r
io
us
opt
im
iz
a
ti
on
f
unc
ti
ons
to
e
nha
nc
e
th
e
pe
r
f
or
m
a
nc
e
of
M
obi
le
C
hi
li
N
e
t.
S
e
ve
r
a
l
opt
im
iz
e
r
s
te
s
te
d
in
c
lu
de
A
da
G
r
a
d,
A
da
m
,
A
da
m
W
,
S
D
G
M
,
a
nd
R
M
S
pr
op,
a
ll
of
w
hi
c
h
a
r
e
c
om
m
onl
y
us
e
d
in
C
N
N
opt
im
iz
a
ti
on.
E
a
c
h
opt
im
iz
e
r
ha
s
a
uni
que
m
e
c
ha
ni
s
m
f
or
upda
ti
ng
w
e
ig
ht
s
,
a
im
in
g
to
f
in
d
th
e
opt
im
a
l
s
ol
ut
io
n
dur
in
g
tr
a
in
in
g.
T
he
gr
a
ph
in
F
ig
ur
e
5
s
how
s
th
e
pe
r
f
or
m
a
nc
e
of
A
d
a
G
r
a
d
on
M
obi
le
C
hi
li
N
e
t,
pr
ovi
di
ng a
vi
s
ua
l
ove
r
vi
e
w
of
i
ts
i
m
pa
c
t
on m
ode
l
a
c
c
ur
a
c
y.
F
ig
ur
e
5.
T
r
a
in
in
g
a
nd
va
l
id
a
ti
on a
c
c
ur
a
c
y
of
M
obi
l
e
C
hi
l
iNe
t
w
it
h A
d
a
G
r
a
d
o
pt
im
i
z
e
r
A
s
s
how
n
in
F
ig
ur
e
5,
th
e
us
e
of
th
e
A
da
G
r
a
d
opt
im
iz
e
r
s
u
c
c
e
s
s
f
ul
ly
im
pr
ove
d
th
e
a
c
c
ur
a
c
y
of
M
obi
le
C
hi
li
N
e
t,
bot
h
in
tr
a
in
in
g
a
nd
va
li
da
ti
on
da
ta
.
W
it
h
a
le
a
r
ni
ng
r
a
te
of
0.01,
th
e
tr
a
in
in
g
a
c
c
ur
a
c
y
r
e
a
c
he
d
99.18%
,
w
hi
le
th
e
va
li
da
ti
on
a
c
c
ur
a
c
y
r
e
a
c
he
d
95.31
%
.
H
ow
e
ve
r
,
it
c
a
n
be
s
e
e
n
th
a
t
th
e
va
li
da
ti
on
a
c
c
ur
a
c
y
e
xpe
r
ie
n
c
e
d
f
lu
c
tu
a
ti
ons
,
in
di
c
a
ti
ng
pe
r
f
or
m
a
nc
e
i
ns
ta
bi
li
ty
.
T
hi
s
s
ugge
s
t
s
th
a
t
hi
gh
tr
a
in
in
g
a
c
c
ur
a
c
y
doe
s
not
a
lwa
y
s
gua
r
a
nt
e
e
opt
im
a
l
va
li
da
ti
on
a
c
c
ur
a
c
y.
N
e
xt
,
a
n
e
xpe
r
im
e
nt
w
a
s
c
onduc
te
d
us
in
g
th
e
A
da
m
opt
im
iz
e
r
t
o c
om
pa
r
e
pe
r
f
or
m
a
nc
e
, a
s
s
how
n i
n F
ig
u
r
e
6.
F
ig
ur
e
6.
T
r
a
i
ni
ng
a
nd
va
l
id
a
ti
on a
c
c
ur
a
c
y
of
M
obi
l
e
C
hi
l
iNe
t
w
it
h A
d
a
m
o
pt
im
iz
e
r
T
he
A
da
m
opt
im
iz
e
r
a
ls
o
s
uc
c
e
s
s
f
ul
ly
im
pr
ove
d
th
e
a
c
c
ur
a
c
y
of
M
obi
le
C
hi
li
N
e
t.
A
s
s
how
n
in
F
ig
ur
e
6,
w
it
h
a
le
a
r
ni
ng
r
a
te
of
0.0001,
A
da
m
a
c
hi
e
ve
d
a
hi
ghe
s
t
tr
a
in
in
g
a
c
c
ur
a
c
y
of
96.30%
a
nd
a
hi
ghe
s
t
va
li
da
ti
on
a
c
c
ur
a
c
y
of
94.84%
.
A
lt
hough
th
e
gr
a
ph
s
how
s
s
li
ght
f
lu
c
tu
a
ti
ons
,
th
e
a
c
c
ur
a
c
y
c
ons
is
te
nt
ly
in
c
r
e
a
s
e
d
w
it
h
e
a
c
h
e
po
c
h.
O
ve
r
a
ll
,
A
da
m
de
m
ons
tr
a
te
d
s
t
a
bl
e
pe
r
f
or
m
a
nc
e
dur
in
g
tr
a
in
in
g.
N
e
xt
,
a
n
e
xpe
r
im
e
nt
w
a
s
c
onduc
te
d
u
s
in
g
th
e
S
G
D
opt
im
iz
e
r
w
it
h
M
om
e
nt
um
0.9
to
c
om
pa
r
e
pe
r
f
or
m
a
nc
e
,
a
s
s
how
n
in
F
ig
ur
e
7.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
M
obi
le
C
hi
li
N
e
t:
c
onv
ol
ut
io
nal
ne
ur
al
ne
tw
or
k
f
or
c
hi
li
l
e
av
e
s
c
la
s
s
if
ic
at
io
n
(
M
ar
is
c
ha E
lv
e
ny
)
3765
F
ig
ur
e
7.
T
r
a
in
in
g
a
nd
va
l
id
a
ti
on a
c
c
ur
a
c
y
of
M
obi
l
e
C
hi
l
iNe
t
w
it
h S
G
D
m
om
e
nt
um
0.
9
W
he
n M
obi
le
C
hi
li
N
e
t
us
e
d t
he
S
G
D
opt
im
iz
e
r
w
it
h
m
om
e
nt
um
0.9, the
a
c
c
ur
a
c
y r
e
s
ul
ts
w
e
r
e
l
ow
e
r
c
om
pa
r
e
d
to
us
in
g
A
d
a
G
r
a
d
a
nd
A
da
m
.
A
s
s
ho
w
n
in
F
ig
ur
e
7,
w
it
h
a
le
a
r
ni
ng
r
a
te
of
0.01,
S
G
D
a
c
hi
e
ve
d
a
hi
ghe
s
t
tr
a
in
in
g
a
c
c
ur
a
c
y
of
95.36%
a
nd
a
hi
ghe
s
t
va
li
da
ti
on
a
c
c
ur
a
c
y
of
94.13%
.
A
lt
hough
it
s
pe
r
f
or
m
a
nc
e
w
a
s
good,
th
e
a
c
c
ur
a
c
y
r
e
m
a
in
e
d
lo
w
e
r
th
a
n
th
e
pr
e
vi
ous
opt
im
iz
e
r
s
.
A
f
te
r
c
ha
ngi
ng
th
e
m
om
e
nt
um
to
0.99,
a
s
s
ho
w
n
in
F
ig
ur
e
8,
th
e
a
c
c
ur
a
c
y
f
lu
c
tu
a
te
d
w
it
h
e
a
c
h
e
po
c
h, but
th
e
r
e
s
ul
ts
di
d
not
s
how
im
pr
ove
m
e
nt
a
nd
e
ve
n e
xpe
r
ie
nc
e
d
a
de
c
li
ne
c
om
pa
r
e
d t
o t
he
i
ni
ti
a
l
m
om
e
nt
um
.
F
ig
ur
e
8.
T
r
a
in
in
g
a
nd
va
l
id
a
ti
on a
c
c
ur
a
c
y
of
M
obi
l
e
C
hi
l
iNe
t
w
it
h S
G
D
m
om
e
nt
um
0.
99
F
ig
ur
e
8
s
how
s
th
a
t
c
ha
ngi
ng
th
e
m
om
e
nt
um
in
S
G
D
f
r
om
0.9
to
0.99
di
d
not
s
ig
ni
f
ic
a
nt
ly
im
pa
c
t
th
e
a
c
c
ur
a
c
y
im
pr
ove
m
e
nt
.
W
it
h
a
le
a
r
ni
ng
r
a
te
o
f
0.001,
th
e
m
ode
l
a
c
hi
e
ve
d
a
hi
ghe
s
t
tr
a
in
in
g
a
c
c
ur
a
c
y
of
95.83%
a
nd
a
hi
ghe
s
t
va
li
da
ti
on
a
c
c
ur
a
c
y
of
93.43%
.
T
he
m
o
m
e
nt
um
c
ha
nge
di
d
not
s
ig
ni
f
ic
a
nt
ly
e
nha
nc
e
th
e
m
ode
l'
s
pe
r
f
or
m
a
nc
e
a
nd
e
ve
n
r
e
s
ul
te
d
in
a
s
li
ght
de
c
r
e
a
s
e
in
va
li
da
ti
on
a
c
c
ur
a
c
y.
A
f
te
r
w
a
r
d,
th
e
e
xpe
r
im
e
nt
c
ont
in
ue
d
by
s
w
it
c
hi
ng
th
e
opt
im
iz
e
r
to
A
da
m
W
,
a
nd
th
e
tr
a
in
in
g
r
e
s
ul
ts
f
or
th
is
opt
im
iz
e
r
a
r
e
s
how
n i
n F
ig
ur
e
9.
F
ig
ur
e
9.
T
r
a
in
in
g
a
nd
va
l
id
a
ti
on a
c
c
ur
a
c
y
of
M
obi
l
e
C
hi
l
iNe
t
w
it
h A
d
a
m
W
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
. 14, No. 5, O
c
to
be
r
2025
:
3757
-
3770
3766
I
n F
ig
ur
e
9,
t
he
gr
a
ph
s
how
s
t
ha
t
us
in
g a
l
e
a
r
ni
ng r
a
te
of
0.001
w
it
h t
he
A
da
m
W
opt
im
iz
e
r
p
r
ovi
de
d
m
or
e
s
ta
bl
e
r
e
s
ul
ts
a
nd highe
r
a
c
c
ur
a
c
y c
om
pa
r
e
d t
o
ot
he
r
l
e
a
r
ni
ng r
a
te
va
lu
e
s
. T
he
hi
ghe
s
t
tr
a
in
in
g a
c
c
ur
a
c
y
r
e
a
c
he
d
97.30%
,
w
hi
le
th
e
hi
ghe
s
t
v
a
li
da
ti
on
a
c
c
ur
a
c
y
w
a
s
95.31%
.
A
lt
hough
th
e
va
li
da
ti
on
a
c
c
ur
a
c
y
is
c
om
pa
r
a
bl
e
to
th
e
r
e
s
ul
ts
obt
a
in
e
d
us
in
g
A
d
a
G
r
a
d,
th
e
pe
r
f
or
m
a
nc
e
w
it
h
A
da
m
W
w
a
s
m
or
e
s
ta
bl
e
th
r
oughout t
he
t
r
a
in
in
g
pr
oc
e
s
s
. N
e
xt
, t
he
e
xpe
r
im
e
nt
w
a
s
c
ont
i
nue
d by s
w
it
c
hi
ng t
he
opt
im
iz
e
r
t
o R
M
S
P
r
op
,
a
nd t
he
a
c
c
ur
a
c
y r
e
s
ul
ts
f
or
e
a
c
h e
poc
h u
s
in
g R
M
S
P
r
op a
r
e
s
how
n i
n F
ig
ur
e
10.
F
ig
ur
e
10. T
r
a
in
in
g a
nd
va
li
da
ti
on a
c
c
ur
a
c
y
of
M
obi
le
C
hi
li
N
e
t
w
it
h R
M
S
P
r
op
A
s
s
how
n
in
F
ig
ur
e
10,
th
e
tr
a
in
in
g
a
nd
va
li
da
ti
on
a
c
c
ur
a
c
y
of
M
obi
le
C
hi
li
N
e
t
w
he
n
opt
im
iz
e
d
us
in
g
R
M
S
P
r
op
w
a
s
ve
r
y
lo
w
.
T
he
hi
ghe
s
t
a
c
c
ur
a
c
y
a
c
hi
e
ve
d w
it
h
a
le
a
r
ni
ng
r
a
te
of
0.0001
w
a
s
88.08%
f
or
tr
a
in
in
g
a
c
c
ur
a
c
y
a
nd
87.32%
f
or
va
li
da
ti
on
a
c
c
ur
a
c
y.
T
hi
s
in
di
c
a
te
s
th
a
t
R
M
S
P
r
op
is
not
s
ui
ta
bl
e
f
or
c
hi
li
le
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on.
T
he
r
e
s
ul
ts
of
a
ll
opt
im
iz
e
r
e
xp
e
r
im
e
nt
s
,
in
c
lu
di
ng
A
da
G
r
a
d,
A
da
m
,
S
G
D
,
A
da
m
W
, a
nd R
M
S
P
r
op, a
r
e
pr
e
s
e
nt
e
d i
n T
a
bl
e
7
.
T
a
bl
e
7
. A
c
c
ur
a
c
y
c
om
pa
r
is
on of
opt
im
iz
e
r
us
a
ge
O
pt
i
m
i
z
e
r
L
e
a
r
ni
ng
r
a
t
e
T
r
a
i
ni
ng
a
c
c
ur
a
c
y (
%
)
V
a
l
i
da
t
i
on
a
c
c
ur
a
c
y (
%
)
A
da
gr
a
d
0.0001
64.28
63.85
0.001
91.77
87.09
0.01
99.18
95.31
A
da
m
0.0001
96.30
94.84
0.001
97.12
94.37
0.01
89.60
85.45
S
G
D
M
om
e
nt
um
=0.9
0.0001
92.89
89.91
0.001
94.42
91.08
0.01
95.36
94.13
S
G
D
M
om
e
nt
um
=0.99
0.0001
94.42
90.85
0.001
95.83
93.43
0.01
84.96
83.10
A
da
m
W
0.0001
97.36
95.07
0.001
97.30
95.31
0.01
90.64
89.80
R
M
S
P
r
op
0.0001
88.08
87.32
0.001
76.33
72.54
0.01
38.37
43.43
F
r
om
T
a
bl
e
7
,
it
c
a
n
be
s
e
e
n
th
a
t
th
e
A
da
G
r
a
d
opt
im
iz
e
r
pr
ovi
de
d
th
e
be
s
t
r
e
s
ul
ts
f
or
M
obi
le
C
hi
li
N
e
t
w
it
h
a
le
a
r
ni
ng
r
a
te
of
0.01,
a
c
hi
e
vi
ng
a
tr
a
in
in
g
a
c
c
ur
a
c
y
of
99.18%
a
nd
a
va
li
da
ti
on
a
c
c
ur
a
c
y
of
95.31%
.
H
ow
e
ve
r
,
f
lu
c
tu
a
ti
on
s
c
a
u
s
e
d
in
s
ta
bi
li
ty
dur
in
g
tr
a
in
in
g.
I
n
c
ont
r
a
s
t,
A
d
a
m
W
of
f
e
r
e
d
be
tt
e
r
s
ta
bi
li
ty
w
it
h
th
e
s
a
m
e
va
li
da
ti
on
r
e
s
ul
t
of
95.31%
.
O
n
t
he
ot
he
r
ha
nd,
R
M
S
P
r
op
yi
e
ld
e
d
m
uc
h
lo
w
e
r
r
e
s
ul
ts
,
w
it
h
uns
a
ti
s
f
a
c
to
r
y
tr
a
in
in
g
a
nd
va
li
da
ti
on
a
c
c
ur
a
c
y,
in
di
c
a
ti
ng
th
a
t
th
is
opt
im
iz
e
r
is
le
s
s
s
ui
ta
bl
e
f
or
c
hi
li
l
e
a
f
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on i
n t
hi
s
m
ode
l.
T
he
ob
s
e
r
ve
d
f
lu
c
tu
a
ti
ons
in
opt
im
iz
e
r
pe
r
f
or
m
a
nc
e
c
a
n
be
a
tt
r
ib
ut
e
d
to
di
f
f
e
r
e
nc
e
s
in
how
e
a
c
h
a
lg
or
it
hm
ha
ndl
e
s
gr
a
di
e
nt
upda
te
s
.
S
G
D
r
e
li
e
s
on
a
f
ix
e
d
le
a
r
ni
ng
r
a
te
a
nd
s
im
pl
e
m
om
e
nt
um
,
w
hi
c
h
c
a
n
le
a
d
to
uns
t
a
bl
e
c
onv
e
r
ge
nc
e
,
e
s
pe
c
ia
ll
y
w
he
n
na
vi
ga
ti
ng
n
oi
s
y
or
c
om
pl
e
x
lo
s
s
s
ur
f
a
c
e
s
.
I
n
c
ont
r
a
s
t,
A
da
m
W
c
om
bi
ne
s
a
da
pt
iv
e
le
a
r
ni
ng
r
a
te
s
w
it
h
w
e
ig
ht
de
c
a
y
r
e
gul
a
r
iz
a
ti
on,
a
ll
ow
in
g
it
to
a
dj
us
t
le
a
r
ni
ng
Evaluation Warning : The document was created with Spire.PDF for Python.