I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
4
,
Augus
t
2025
,
pp.
3201
~
3213
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
4
.
pp
32
01
-
3213
3201
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
Op
t
imiz
in
g c
itr
u
s d
is
e
as
e
d
e
t
e
c
t
io
n
:
a
t
r
an
s
f
e
r
ab
le
c
on
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
o
r
k
m
od
e
l
e
n
h
an
c
e
d
w
it
h
t
h
e
f
r
u
i
t
f
ly
op
t
imiz
at
io
n
al
gor
it
h
m
Anoop
Gan
ad
al
u
L
in
gar
aj
u
,
As
h
a
M
an
ga
la
S
h
a
n
k
ar
e
gowda
,
B
ab
u
Ku
m
ar
S
at
h
iya
m
u
r
t
h
y
,
S
an
t
h
r
u
p
t
h
B
u
d
an
oor
Chan
n
e
gowda
,
S
h
r
u
t
i
Ja
lap
u
r
,
Chai
t
r
a
P
alah
all
i
Chenn
ak
e
s
h
ava
D
e
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ie
nc
e
a
nd E
ngi
ne
e
r
in
g,
S
c
hool
of
E
ngi
ne
e
r
in
g a
nd T
e
c
hnol
ogy,
C
H
R
I
S
T
U
ni
ve
r
s
it
y, B
e
nga
lu
r
u
, I
ndi
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
S
e
p
30
,
2024
R
e
vis
e
d
Apr
9
,
2025
Ac
c
e
pted
J
un
8
,
2025
Fu
n
g
a
l
,
b
ac
t
eri
a
l
,
an
d
v
i
ral
d
i
s
eas
e
s
s
i
g
n
i
f
i
can
t
l
y
t
h
rea
t
en
ci
t
ru
s
p
r
o
d
u
ct
i
o
n
an
d
q
u
al
i
t
y
w
o
rl
d
w
i
d
e,
p
r
o
mp
t
i
n
g
p
r
o
d
u
cers
t
o
e
x
p
l
o
re
t
ec
h
n
o
l
o
g
i
cal
s
o
l
u
t
i
o
n
s
t
o
mi
t
i
g
at
e
t
h
e
fi
n
an
c
i
al
i
m
p
act
o
f
t
h
es
e
d
i
s
e
as
es
.
Imag
e
an
a
l
y
s
i
s
t
ech
n
i
q
u
e
s
h
a
v
e
emerg
e
d
as
p
o
w
er
fu
l
t
o
o
l
s
f
o
r
d
e
t
ect
i
n
g
c
i
t
r
u
s
d
i
s
eas
e
s
b
y
d
i
ffere
n
t
i
at
i
n
g
b
et
w
ee
n
h
eal
t
h
y
an
d
d
i
s
eas
e
d
s
p
ec
i
men
s
t
h
ro
u
g
h
t
h
e
ex
t
rac
t
i
o
n
o
f
d
i
s
cri
m
i
n
a
t
i
v
e
feat
u
res
fro
m
i
n
p
u
t
i
mag
es
.
T
h
i
s
p
ap
er
i
n
t
ro
d
u
ce
s
a
v
al
u
ab
l
e
d
at
a
s
et
c
o
mp
r
i
s
i
n
g
9
5
3
co
l
o
r
i
ma
g
es
o
f
o
ran
g
e
l
eav
e
s
fro
m
t
h
e
s
p
ec
i
es
Ci
t
r
u
s
s
i
n
e
n
s
i
s
(L
.
)
O
s
b
eck
,
w
h
i
c
h
s
er
v
es
t
o
t
rai
n
,
ev
a
l
u
a
t
e,
an
d
c
o
mp
are
v
ari
o
u
s
al
g
o
r
i
t
h
ms
ai
me
d
at
i
d
e
n
t
i
fy
i
n
g
a
b
n
o
rmal
i
t
i
es
i
n
c
i
t
r
u
s
fru
i
t
s
.
T
h
e
d
e
v
el
o
p
me
n
t
o
f
au
t
o
ma
t
ed
d
et
ec
t
i
o
n
s
y
s
t
ems
i
s
cru
c
i
al
f
o
r
red
u
c
i
n
g
eco
n
o
mi
c
l
o
s
s
es
i
n
ci
t
ru
s
p
ro
d
u
c
t
i
o
n
,
w
i
t
h
t
h
i
s
res
earch
fo
c
u
s
i
n
g
o
n
t
w
e
l
v
e
s
p
eci
f
i
c
d
i
s
ea
s
es
a
n
d
n
u
t
ri
e
n
t
d
ef
i
ci
e
n
c
i
es
.
W
e
p
ro
p
o
s
e
a
n
o
v
e
l
ap
p
r
o
ach
t
o
ci
t
ru
s
p
l
a
n
t
d
i
s
ea
s
e
d
e
t
ect
i
o
n
u
t
i
l
i
zi
n
g
a
h
y
p
er
-
p
arame
t
er
t
u
n
ed
t
ran
s
ferrab
l
e
co
n
v
o
l
u
t
i
o
n
al
n
e
u
ra
l
n
et
w
o
r
k
(T
C
N
N
)
m
o
d
el
,
referre
d
t
o
a
s
t
h
e
en
h
a
n
ced
fru
i
t
fl
y
o
p
t
i
m
i
zat
i
o
n
al
g
o
ri
t
h
m
(
E
FO
A
)
-
T
CN
N
mo
d
el
.
T
h
i
s
mo
d
el
o
p
t
i
mi
zes
t
h
e
p
arame
t
ers
o
f
T
C
N
N
u
s
i
n
g
t
h
e
E
FO
A
an
d
en
h
an
ce
s
arch
i
t
ect
u
ral
d
e
s
i
g
n
b
y
i
n
c
o
rp
o
rat
i
n
g
t
h
ree
co
n
v
o
l
u
t
i
o
n
a
l
l
ay
er
s
al
o
n
g
s
i
d
e
an
en
erg
y
l
ay
er
i
n
s
t
ea
d
o
f
a
t
ra
d
i
t
i
o
n
a
l
p
o
o
l
i
n
g
l
a
y
er.
E
x
p
eri
me
n
t
a
l
res
u
l
t
s
d
emo
n
s
t
rat
e
t
h
a
t
t
h
e
p
ro
p
o
s
ed
E
F
O
A
-
T
CN
N
mo
d
el
o
u
t
p
erf
o
rms
e
x
i
s
t
i
n
g
s
t
a
t
e
-
of
-
the
-
ar
t
met
h
o
d
s
,
ach
i
ev
i
n
g
a
s
en
s
i
t
i
v
i
t
y
o
f
0
.
9
7
5
an
d
an
accu
rac
y
o
f
0
.
9
9
5
.
K
e
y
w
o
r
d
s
:
B
a
c
ter
ia
C
it
r
us
f
r
uit
dis
e
a
s
e
de
tec
ti
on
E
nha
nc
e
d
f
r
uit
f
ly
opti
mi
z
a
ti
on
a
lgor
it
hm
I
mage
a
na
lys
is
T
r
a
ns
f
e
r
r
a
ble
c
onvolut
ional
ne
ur
a
l
ne
twor
k
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
As
ha
M
a
nga
la
S
ha
nka
r
e
gowda
D
e
p
a
r
t
m
e
nt
o
f
C
o
m
pu
t
e
r
S
c
i
e
nc
e
a
n
d
E
n
g
i
ne
e
r
i
ng
,
S
c
h
o
o
l
o
f
E
n
g
in
e
e
r
i
n
g
a
n
d
T
e
c
h
n
o
l
og
y
,
C
HR
I
S
T
Un
i
v
e
r
s
i
t
y
B
e
nga
lur
u,
Ka
r
na
taka
,
I
ndia
E
mail:
a
s
ha
gowda
05
@gmail.
c
om
1.
I
NT
RODU
C
T
I
ON
S
pe
c
ies
that
a
r
e
indi
ge
nous
to
Aus
tr
a
li
a
,
M
e
lane
s
ia,
a
nd
c
e
r
tain
r
e
gions
of
As
ia
a
r
e
include
d
in
the
ge
nus
C
it
r
us
,
whic
h
is
c
ompr
is
e
d
of
the
f
r
uit
c
r
ops
that
a
r
e
the
mos
t
e
c
onomi
c
a
ll
y
va
luable
on
a
glob
a
l
s
c
a
le.
C
it
r
us
f
r
uit
s
,
whic
h
include
s
we
e
t
or
a
nge
s
a
nd
manda
r
ins
,
a
r
e
gr
own
in
mo
r
e
than
140
c
ountr
ie
s
a
nd
a
r
e
pr
im
a
r
il
y
c
ult
ivate
d
f
or
the
pur
pos
e
of
s
e
r
ving
the
mar
ke
ts
f
or
f
r
e
s
h
f
r
uit
a
nd
be
ve
r
a
ge
s
.
Ac
c
or
din
g
to
the
F
ood
a
nd
Agr
icultur
e
Or
ga
niza
ti
on
of
the
Unite
d
Na
ti
ons
(
F
AO
)
in
2020
[
1
]
,
s
we
e
t
or
a
nge
s
a
c
c
ount
f
or
65
pe
r
c
e
nt
o
f
the
tot
a
l
c
it
r
us
pr
oduc
ti
on
wor
ld
wide
[
2]
,
with
M
e
dit
e
r
r
a
ne
a
n
c
ountr
ies
be
ing
th
e
lea
ding
e
xpor
ter
s
of
f
r
e
s
h
f
r
ui
t.
How
e
ve
r
,
the
pe
r
s
is
ten
t
thr
e
a
t
pos
e
d
by
a
wide
va
r
iety
o
f
f
unga
l,
ba
c
te
r
ial,
a
nd
oomyce
te
dis
e
a
s
e
s
is
a
s
igni
f
ica
nt
ba
r
r
ier
to
th
e
e
f
f
icie
nt
tr
a
de
o
f
c
it
r
us
f
r
uit
s
,
bo
th
domes
ti
c
a
ll
y
a
nd
int
e
r
na
ti
ona
ll
y
[
3]
.
T
his
is
the
c
a
s
e
be
c
a
us
e
c
it
r
u
s
f
r
uit
s
a
r
e
s
us
c
e
pti
ble
to
a
wide
r
a
nge
o
f
dis
e
a
s
e
s
.
T
he
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
20
25
:
320
1
-
3213
3202
pr
e
s
e
nc
e
of
thes
e
pa
thogen
s
,
whic
h
include
P
lenodom
us
tr
ac
he
iphi
lus
a
nd
P
hy
topht
hor
a
s
p
e
c
ies
,
ha
s
the
potential
to
r
e
s
ult
in
yield
los
s
e
s
of
thi
r
ty
to
f
if
ty
pe
r
c
e
nt
dur
ing
c
r
uc
ial
pha
s
e
s
of
the
plant's
li
f
e
c
yc
le
[
2]
.
T
he
s
us
c
e
pti
bil
it
y
of
c
it
r
us
plants
to
dis
e
a
s
e
s
is
e
xa
c
e
r
ba
ted
by
their
a
c
id
pH
a
nd
high
wa
ter
c
onte
nt,
whic
h
lea
ve
s
them
vulner
a
ble
dur
ing
the
pr
e
-
ha
r
ve
s
t
a
nd
pos
t
-
ha
r
ve
s
t
s
tag
e
s
-
6
a
nd
7
r
e
s
pe
c
ti
ve
ly
[
4]
.
T
he
pur
pos
e
o
f
thi
s
s
tudy
is
to
inves
ti
ga
te
the
is
s
ue
of
c
it
r
us
dis
e
a
s
e
de
tec
ti
on,
whic
h
is
a
s
igni
f
ica
nt
obs
tac
le
that
a
f
f
e
c
ts
c
it
r
us
pr
oduc
ti
on
a
ll
ove
r
the
wor
ld.
B
e
c
a
us
e
of
the
de
mands
plac
e
d
on
c
omp
utational
r
e
s
our
c
e
s
a
nd
the
c
ons
tr
a
int
s
im
pos
e
d
by
ne
twor
k
de
s
ign,
the
mac
hine
lea
r
ning
a
nd
de
e
p
lea
r
ning
tec
hniques
that
a
r
e
c
ur
r
e
ntl
y
in
us
e
ha
ve
l
im
it
a
ti
ons
whe
n
i
t
c
omes
to
de
te
c
ti
ng
thes
e
dis
e
a
s
e
s
with
the
r
e
qui
r
e
d
leve
l
of
e
f
f
icie
nc
y.
T
he
r
e
a
s
on
f
or
thi
s
is
that
the
r
e
is
a
g
r
owing
de
mand
f
or
a
gr
icultur
a
l
pr
oduc
ts
that
a
r
e
of
high
qua
li
ty
a
nd
s
a
f
e
f
or
the
e
nvir
onment
[
5]
–
[
8
]
.
T
he
r
e
f
or
e
,
a
dva
nc
e
d
de
tec
ti
on
s
ys
tems
a
r
e
e
s
s
e
nti
a
l
in
or
de
r
to
mi
nim
ize
e
c
onomi
c
los
s
e
s
in
the
c
it
r
us
indus
tr
y.
T
he
a
ppr
oa
c
h
that
ha
s
be
e
n
pr
opos
e
d
pr
ovi
de
s
f
r
e
s
h
pe
r
s
pe
c
ti
ve
s
by
pr
e
s
e
nti
ng
a
tr
a
ns
f
e
r
r
a
ble
c
onv
olut
ional
ne
ur
a
l
ne
twor
k
(
T
C
N
N)
model
that
h
a
s
be
e
n
opti
mi
z
e
d
with
the
e
nha
nc
e
d
f
r
uit
f
ly
opti
mi
z
a
ti
on
a
lgor
it
hm
(
E
F
OA
)
.
Us
ing
thi
s
method,
dis
e
a
s
e
de
t
e
c
ti
on
is
im
pr
ove
d
while
maintaining
low
c
omput
a
t
ional
c
os
ts
,
whic
h
make
s
it
a
c
c
e
s
s
ibl
e
a
nd
pr
a
c
ti
c
a
l
f
or
a
ppli
c
a
ti
ons
in
the
r
e
a
l
wo
r
l
d
[
9]
.
As
s
ta
ted
in
t
he
i
nt
r
oduc
ti
o
n,
the
pa
pe
r
's
ma
in
goa
l
is
to
iden
ti
f
y
c
it
r
us
dis
e
a
s
e
s
thr
ough
im
a
ge
a
na
lys
is
a
nd
mac
hi
ne
lea
r
n
ing
.
H
owe
ve
r
,
i
t
s
hou
ld
be
no
ted
that
the
r
e
s
e
a
r
c
h
pr
e
s
e
n
ts
t
he
E
F
OA
-
T
C
N
N
mode
l
to
iden
ti
f
y
12
di
f
f
e
r
e
n
t
kinds
o
f
c
it
r
us
plan
t
il
lnes
s
e
s
a
nd
de
f
ici
ts
[
10
]
–
[
12
]
.
A
lt
h
ough
t
he
f
i
na
nc
ia
l
b
ur
de
n
o
f
c
it
r
us
il
l
ne
s
s
e
s
is
m
e
nti
one
d
in
the
a
bs
tr
a
c
t
,
the
i
ntr
o
duc
t
ion
w
il
l
be
e
xten
de
d
t
o
dis
c
us
s
the
im
po
r
tanc
e
o
f
a
utom
a
ted
de
t
e
c
t
ion
.
T
he
f
ina
nc
ial
r
a
mi
f
ica
t
ions
will
be
hi
ghl
ight
e
d
,
inc
lud
ing
a
v
oid
ing
lar
ge
c
r
o
p
los
s
e
s
,
mainta
ini
ng
qua
l
it
y
c
on
tr
ol
in
the
c
i
tr
us
s
e
c
to
r
,
a
n
d
the
s
ho
r
tc
omi
ngs
o
f
c
onve
nti
ona
l
man
ua
l
ins
pe
c
ti
ons
in
de
tec
ti
ng
t
he
s
e
il
lnes
s
e
s
.
T
he
pr
im
a
r
y
go
a
l
o
f
th
is
r
e
s
e
a
r
c
h
is
to
de
ve
l
op
t
he
E
F
OA
-
T
C
NN
mo
de
l
t
o
ident
if
y
twe
lve
dis
ti
nc
t
types
of
c
i
tr
us
dis
e
a
s
e
s
a
nd
nu
tr
it
io
na
l
de
f
icie
nc
ies
th
r
oug
h
im
a
ge
a
na
lys
is
a
nd
mac
hine
lea
r
n
ing
.
Al
tho
ugh
the
a
bs
t
r
a
c
t
r
e
f
e
r
e
nc
e
s
the
f
i
na
nc
ial
b
ur
de
n
o
f
c
i
tr
us
dis
e
a
s
e
s
[
1
3]
–
[
1
5]
,
t
he
in
tr
oduc
t
ion
will
be
e
xpa
nde
d
to
e
m
pha
s
ize
the
i
mpo
r
ta
nc
e
o
f
a
utom
a
ted
de
tec
t
ion
.
Ke
y
f
i
na
nc
i
a
l
i
mpl
ica
ti
ons
in
c
lude
t
he
pr
e
ve
n
ti
o
n
o
f
s
ubs
ta
nti
a
l
c
r
op
los
s
e
s
,
t
he
ma
int
e
na
nc
e
of
qua
l
it
y
c
ont
r
o
l
in
the
c
i
t
r
us
s
e
c
to
r
,
a
n
d
the
li
mi
tat
ions
of
t
r
a
di
ti
on
a
l
man
ua
l
ins
pe
c
ti
o
n
meth
ods
in
e
f
f
e
c
ti
v
e
ly
iden
ti
f
yi
ng
thes
e
dis
e
a
s
e
s
[
16]
.
M
or
e
ove
r
,
thi
s
int
r
oduc
ti
on
will
de
lve
de
e
pe
r
int
o
the
E
F
OA
a
nd
it
s
r
ole
in
f
ine
-
tuni
ng
the
T
C
NN
.
I
t
will
a
ls
o
e
lucida
te
the
innovative
incor
po
r
a
ti
on
of
a
n
e
ne
r
gy
laye
r
in
plac
e
of
the
c
onve
nti
ona
l
pooli
ng
laye
r
a
nd
dis
c
us
s
how
hype
r
pa
r
a
mete
r
opti
mi
z
a
ti
on
e
nha
nc
e
s
ove
r
a
ll
model
pe
r
f
or
manc
e
.
B
y
e
va
luating
the
E
F
OA
-
T
C
NN
model
us
ing
a
publi
c
ly
a
va
il
a
ble
da
tas
e
t
of
o
r
a
nge
lea
ve
s
,
thi
s
pa
pe
r
a
im
s
to
p
r
e
s
e
nt
a
c
ompr
e
he
ns
ive
s
olut
ion
f
or
a
c
c
ur
a
te
c
it
r
us
d
is
e
a
s
e
c
las
s
if
ica
ti
on.
S
pe
c
if
ica
ll
y,
the
pa
pe
r
is
or
ga
nize
d
a
s
f
oll
ows
.
T
he
li
ter
a
tur
e
that
is
pe
r
ti
ne
nt
to
the
top
ic
is
dis
c
us
s
e
d
in
s
e
c
ti
on
2
.
T
he
methodology
is
pr
e
s
e
nted
in
s
e
c
ti
on
3
.
T
he
r
e
s
ult
s
a
r
e
dis
c
us
s
e
d
in
s
e
c
ti
on
4
.
L
a
s
tl
y,
the
f
indi
ngs
a
r
e
s
um
mar
ize
d
in
s
e
c
ti
on
5
.
2.
RE
L
AT
E
D
WORKS
I
n
their
inves
ti
ga
ti
on
of
a
n
a
utom
a
ted
s
ys
tem
f
or
c
it
r
us
dis
e
a
s
e
c
las
s
if
ica
ti
on,
B
utt
e
t
al
.
[
17]
e
mpl
oye
d
de
e
p
lea
r
ning
c
ombi
ne
d
wi
th
opti
mal
f
e
a
tur
e
s
e
lec
ti
on
tec
hniques
.
T
he
ini
ti
a
l
pha
s
e
of
their
a
ppr
oa
c
h
invol
ve
d
da
ta
a
ugmenta
ti
on,
whic
h
e
ntails
ge
ne
r
a
ti
ng
ne
w
im
a
ge
s
f
or
the
tr
a
ini
n
g
da
tas
e
t
f
r
om
e
xis
ti
ng
e
xa
mpl
e
s
.
L
e
ve
r
a
ging
t
r
a
ns
f
e
r
le
a
r
ning,
the
a
uthor
s
r
e
tr
a
ined
two
p
r
e
-
e
xis
ti
ng
models
—
De
ns
e
Ne
t
-
201
a
nd
Ale
xNe
t
-
us
ing
the
e
nha
nc
e
d
da
tas
e
t
de
r
ived
f
r
om
lea
f
im
a
ge
s
.
T
he
ir
e
xp
e
r
im
e
nts
a
c
hieve
d
a
r
e
mar
ka
ble
p
r
e
c
is
ion
leve
l
o
f
99
.
6%
.
At
e
a
c
h
s
tage
,
the
pr
opos
e
d
f
r
a
mew
or
k
wa
s
c
om
pa
r
e
d
to
s
tate
-
of
-
the
-
a
r
t
methodologi
e
s
,
de
mons
tr
a
ti
ng
s
upe
r
ior
pe
r
f
or
manc
e
.
Ya
da
v
e
t
al
.
[
18]
de
ve
loped
a
c
omput
e
r
vis
ion
s
y
s
tem
c
a
pa
ble
of
a
utom
a
ti
c
a
ll
y
c
a
tegor
izing
f
r
uit
s
a
nd
lea
ve
s
,
ther
e
by
f
a
c
il
it
a
ti
ng
e
f
f
icie
nt
dis
e
a
s
e
mana
ge
me
nt
in
o
r
c
ha
r
ds
.
T
his
s
tudy
uti
li
z
e
d
f
e
a
tur
e
s
ge
ne
r
a
ted
by
C
NN
s
a
nd
mac
hine
lea
r
ning
c
las
s
if
ier
s
to
e
f
f
e
c
ti
ve
ly
de
tec
t
c
it
r
us
blac
k
s
pot
(
C
B
S
)
-
inf
e
c
ted
f
r
uit
s
a
nd
lea
ve
s
e
xhibi
ti
ng
c
a
nke
r
s
ympt
oms
.
T
he
c
us
tom
s
ha
ll
ow
C
NN
c
ombi
ne
d
with
r
a
dial
ba
s
is
f
unc
ti
on
(
R
B
F
)
s
uppor
t
ve
c
to
r
mac
hine
(
S
VM
)
a
c
hieve
d
a
n
ove
r
a
ll
a
c
c
ur
a
c
y
of
92.
1%
f
o
r
f
r
uit
s
a
f
f
e
c
ted
by
C
B
S
a
nd
f
our
other
c
ondit
ions
(
g
r
e
a
s
y
s
pot,
mela
nos
e
,
wind
s
c
a
r
,
a
nd
ma
r
ke
table
)
.
F
or
lea
ve
s
s
howing
c
a
nke
r
s
ympt
oms
a
longs
ide
f
our
other
c
ondit
ions
(
c
ont
r
ol,
g
r
e
a
s
y
s
pot,
mela
nos
e
s
,
a
nd
s
c
a
b)
,
the
VG
G1
6
model
with
R
B
F
S
VM
a
c
hieve
d
a
n
im
pr
e
s
s
ive
ove
r
a
ll
a
c
c
ur
a
c
y
of
93%
.
Ac
c
or
ding
to
Dhiman
e
t
a
l
.
[
3]
,
a
n
e
f
f
e
c
ti
ve
c
it
r
u
s
f
r
uit
dis
e
a
s
e
pr
e
diction
model
c
a
n
be
de
ve
loped
us
ing
hype
r
s
pe
c
tr
a
l
im
a
ging
(
HSI
)
s
ys
tems
a
nd
f
e
a
tur
e
s
e
xtr
a
c
ted
thr
ough
both
de
e
p
a
nd
s
ha
ll
ow
c
onvolut
ional
ne
ur
a
l
ne
twor
ks
,
c
ombi
ne
d
wi
th
ma
c
hine
lea
r
ning
c
las
s
if
ier
s
.
T
he
ir
p
r
opos
e
d
model
i
ntegr
a
tes
e
dge
c
omput
ing
with
de
e
p
lea
r
ning
a
r
c
hit
e
c
tur
e
s
,
s
pe
c
if
ica
ll
y
C
NN
a
nd
long
s
hor
t
-
ter
m
memor
y
(
L
S
T
M
)
ne
twor
ks
.
T
his
model
incor
po
r
a
tes
a
f
e
a
tur
e
-
f
us
ion
s
ubs
ys
tem,
a
down
-
s
a
mpl
ing
method,
a
nd
a
n
a
dva
nc
e
d
f
e
a
tur
e
-
e
xtr
a
c
ti
on
mec
ha
nis
m
to
e
ns
ur
e
a
c
c
ur
a
te
dis
e
a
s
e
de
tec
ti
on
in
c
it
r
us
f
r
uit
s
whi
le
e
na
bli
ng
s
ubs
tantial
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
Optimiz
ing
c
it
r
us
dis
e
as
e
de
tec
ti
on:
a
tr
ans
fer
able
c
onv
olut
ional
ne
ur
al
…
(
A
noop
Ganadalu
L
ingar
aju)
3203
identif
ica
ti
on
c
a
pa
bil
it
ies
.
T
he
s
tudy
uti
l
ize
d
2,
95
0
labe
led
im
a
ge
s
of
c
it
r
us
f
r
uit
s
identi
f
ied
a
s
a
f
f
e
c
ted
by
m
e
lanos
is
,
s
c
a
b
s
,
c
a
nke
r
s
,
blac
k
s
pots
,
or
gr
e
e
ning,
dr
a
wn
f
r
om
onli
ne
Ka
ggle
a
nd
vil
lage
da
tas
e
ts
.
P
e
r
f
or
manc
e
met
r
ics
s
uc
h
a
s
pr
e
c
is
ion,
r
e
c
a
ll
,
F
-
mea
s
ur
e
,
a
nd
s
uppor
t
we
r
e
e
mpl
oye
d
to
c
om
pa
r
e
the
pr
opos
e
d
model
with
e
xis
ti
ng
one
s
,
a
s
s
e
s
s
e
d
both
with
a
nd
without
f
e
a
tur
e
pr
uning
.
T
he
r
e
s
e
a
r
c
h
i
nc
luded
two
pha
s
e
s
:
the
f
ir
s
t
invol
ve
d
e
xpe
r
im
e
ntal
a
na
lys
i
s
us
ing
magnitude
-
b
a
s
e
d
pr
uning
(
M
B
P
)
,
while
the
s
e
c
ond
c
ombi
ne
d
M
B
P
with
pos
t
-
qua
nti
z
a
t
ion.
T
he
C
NN
-
L
S
T
M
model,
e
nha
nc
e
d
by
thes
e
te
c
hniques
,
outper
f
or
med
the
c
ur
r
e
nt
s
tate
-
of
-
the
-
a
r
t
C
NN
method,
a
c
hieving
a
c
c
ur
a
c
y
r
a
tes
of
97.
18
%
a
nd
98.
25%
,
r
e
s
pe
c
ti
ve
ly.
T
he
C
NN
-
L
S
T
M
M
ode
l
hybr
id
mode
l
us
e
d
f
or
maiz
e
dis
e
a
s
e
c
las
s
if
ica
ti
on
[
19]
.
Uğuz
e
t
al
.
[
8]
int
r
oduc
e
d
C
it
r
us
Ne
t,
a
nove
l
mo
de
l
ba
s
e
d
on
C
NN
s
f
or
c
las
s
if
ying
da
mage
d
a
nd
a
bnor
mal
c
it
r
us
f
r
uit
s
.
T
he
s
t
udy
ga
ther
e
d
5
,
14
9
f
r
uit
im
a
ge
s
f
r
om
c
it
r
us
gr
ove
s
in
Anta
lya,
T
ur
ke
y.
Among
the
f
ou
r
C
NN
models
tes
ted,
C
it
r
us
Ne
t
a
nd
R
e
s
Ne
t50
yielde
d
the
be
s
t
c
las
s
if
ica
ti
on
r
e
s
ult
s
.
T
he
s
e
c
ond
pha
s
e
of
their
r
e
s
e
a
r
c
h
e
va
luate
d
f
ive
dif
f
e
r
e
nt
C
NN
models
f
o
r
de
tec
ti
ng
two
c
omm
on
dis
e
a
s
e
s
in
T
ur
kis
h
c
it
r
us
:
a
lt
e
r
na
r
ia
a
lt
e
r
na
ta
a
nd
thr
ips
.
E
xpe
r
im
e
ntal
r
e
s
ult
s
indi
c
a
ted
that
YO
L
Ov5
a
nd
M
a
s
k
R
-
C
NN
we
r
e
the
mos
t
e
f
f
e
c
ti
ve
in
de
tec
ti
ng
c
it
r
u
s
dis
e
a
s
e
s
,
a
c
hieving
a
n
a
ve
r
a
ge
p
r
e
c
is
ion
(
AP
)
of
0
.
99.
W
a
ng
e
t
al
.
[
20]
buil
t
a
c
it
r
us
ye
ll
ow
s
hoot
dis
e
a
s
e
r
e
c
ognit
ion
model
ba
s
e
d
on
the
YO
L
Ov5s
a
nd
a
c
hives
a
n
a
c
c
ur
a
c
y
of
91.
3
%
S
a
ini
e
t
al
.
[
21]
pr
opos
e
d
a
de
e
p
C
NN
model
f
or
c
las
s
if
ying
c
it
r
us
plant
lea
f
a
nd
f
r
uit
dis
e
a
s
e
s
int
o
s
e
ve
n
c
a
tegor
ies
.
T
he
model’
s
pe
r
f
or
manc
e
wa
s
e
va
luate
d
us
ing
opti
mi
z
e
r
s
s
u
c
h
a
s
Ada
m,
s
t
oc
ha
s
ti
c
gr
a
dient
de
s
c
e
nt
(
S
GD
)
,
a
nd
R
M
S
pr
op,
with
Ada
m
a
c
hieving
the
h
ighes
t
pr
e
c
is
ion
of
98.
6
%
.
T
h
e
s
tudy
a
ls
o
s
howe
d
that
da
ta
a
ugmenta
ti
on,
a
long
with
va
r
iat
ions
in
e
poc
hs
,
ba
tch
s
ize
,
a
nd
dr
opout
,
i
mpr
ove
d
model
a
c
c
ur
a
c
y.
T
his
a
ppr
oa
c
h
de
mons
tr
a
tes
the
potential
of
AI
in
e
na
bli
ng
f
a
s
t
a
nd
a
c
c
ur
a
te
plant
dis
e
a
s
e
de
tec
ti
on.
Ar
thi
e
t
al
.
[
22]
pr
opos
e
d
a
nove
l
a
p
pr
oa
c
h
na
med
duc
k
opti
m
iza
ti
on
with
e
nha
nc
e
d
c
a
ps
ule
ne
tw
or
k
(
DO
E
C
N
)
ba
s
e
d
c
it
r
us
dis
e
a
s
e
de
tec
ti
on
f
or
s
us
taina
ble
c
r
op
mana
ge
ment
(
C
DD
C
M
)
wa
s
pr
opos
e
d
to
de
tec
t
a
nd
c
las
s
if
y
c
it
r
us
dis
e
a
s
e
s
e
f
f
e
c
ti
ve
ly.
T
he
method
int
e
gr
a
tes
pr
e
p
r
oc
e
s
s
ing
s
teps
,
us
e
s
DO
E
C
N
f
o
r
f
e
a
tur
e
e
xtr
a
c
ti
on
a
nd
hype
r
pa
r
a
mete
r
tuni
ng,
a
nd
e
mpl
oys
a
de
e
p
s
tac
ke
d
s
pa
r
s
e
a
utoenc
ode
r
(
DSS
AE
)
f
o
r
c
las
s
if
ica
ti
on.
A
C
NN
-
S
VM
hybr
id
model
wa
s
pr
opos
e
d
f
or
c
it
r
us
d
is
e
a
s
e
de
tec
ti
on,
us
ing
C
NN
f
o
r
f
e
a
tur
e
e
xtr
a
c
ti
on
a
nd
S
VM
f
or
c
las
s
if
ica
ti
on
[
23]
.
W
it
h
a
n
a
c
c
ur
a
c
y
of
92.
34%
,
the
model
e
f
f
e
c
ti
ve
ly
i
de
nti
f
ies
mul
ti
ple
c
it
r
us
dis
e
a
s
e
s
,
s
uppor
ti
ng
pr
e
c
is
ion
a
gr
icultur
e
a
nd
s
us
taina
ble
c
r
op
mana
ge
ment.
C
howdhur
y
e
t
al
.
[
24]
de
ve
loped
a
li
ghtwe
ight
C
NN
model
f
or
c
it
r
us
lea
f
dis
e
a
s
e
de
tec
ti
on
a
nd
c
ompar
e
d
with
pr
e
-
tr
a
ined
models
li
ke
R
e
s
Ne
t
-
50,
VG
G16,
a
nd
De
ns
e
Ne
t
va
r
iants
.
T
r
a
ined
on
a
n
a
u
gmente
d
da
tas
e
t
of
2800
im
a
ge
s
,
the
mo
de
l
a
c
hieve
d
a
97.
84%
va
li
da
ti
on
a
c
c
ur
a
c
y
a
nd
96%
F
1
-
s
c
or
e
.
S
ha
s
tr
i
e
t
al
.
[
7]
pr
opos
e
d
a
ne
w
a
ppr
oa
c
h
f
o
r
r
e
li
a
ble
a
nd
a
utom
a
ted
dis
e
a
s
e
identif
ica
ti
on
us
ing
C
NN
s
.
B
y
a
na
lyzing
a
s
ub
s
tantial
da
tas
e
t
of
im
a
ge
s
de
pi
c
ti
ng
dis
e
a
s
e
d
c
it
r
us
f
r
uit
s
a
nd
lea
ve
s
,
their
s
ugge
s
ted
E
-
C
NN
model
de
mons
tr
a
ted
e
xc
e
pti
ona
l
r
e
s
ult
s
in
both
r
e
c
ognit
ion
a
nd
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y.
Qiu
e
t
al
.
[
25]
e
xplor
e
d
s
e
man
ti
c
e
mbedding
methods
we
r
e
inves
ti
ga
ted
f
or
dis
e
a
s
e
im
a
ge
s
a
nd
s
tr
uc
tur
e
d
de
s
c
r
ipt
i
ve
texts
.
Vis
ua
l
f
e
a
tur
e
s
of
lea
ve
s
we
r
e
e
xtr
a
c
ted
us
ing
c
onvolut
ional
ne
twor
ks
of
va
r
ying
de
pths
,
including
VG
G16,
R
e
s
Ne
t50,
M
obil
e
Ne
tV2,
a
nd
S
huf
f
leN
e
tV2.
S
he
r
mi
la
e
t
al
.
[
26]
pr
opos
e
d
a
tailo
r
e
d
a
ppr
oa
c
h
that
i
ntegr
a
tes
a
C
NN
with
a
n
L
S
T
M
,
a
c
hieving
a
n
e
f
f
icie
nc
y
of
96%
.
2.
1
.
P
r
ob
lem
s
t
at
e
m
e
n
t
C
it
r
us
dis
e
a
s
e
s
,
c
a
us
e
d
by
va
r
ious
f
unga
l,
ba
c
ter
ial,
a
nd
vir
a
l
inf
e
c
ti
ons
,
lea
d
to
s
igni
f
ica
nt
f
inanc
ial
los
s
e
s
in
the
c
it
r
us
indus
tr
y
wor
ldwide.
M
a
n
ua
l
in
s
pe
c
ti
on
methods
a
r
e
ti
me
-
c
ons
umi
ng,
p
r
one
to
e
r
r
or
s
,
a
nd
r
e
quir
e
e
xpe
r
t
knowle
dge
,
making
them
ine
f
f
icie
nt
f
or
lar
ge
-
s
c
a
le
de
tec
ti
on.
T
he
r
e
f
or
e
,
ther
e
is
a
c
r
it
ica
l
ne
e
d
f
or
a
utom
a
ted
s
ys
tems
that
c
a
n
a
c
c
ur
a
tely
de
tec
t
a
nd
c
las
s
if
y
mul
ti
ple
dis
e
a
s
e
s
in
c
it
r
us
plants
to
mi
ti
ga
te
e
c
onomi
c
los
s
e
s
a
nd
e
n
s
ur
e
high
-
qua
li
ty
pr
oduc
ti
on.
T
he
pr
oblem
is
the
ine
f
f
icie
nc
y
of
t
r
a
dit
ional
manua
l
a
nd
non
-
opti
mi
z
e
d
mac
hine
lea
r
ning
methods
in
de
tec
ti
ng
mul
ti
ple
c
it
r
us
dis
e
a
s
e
s
,
whic
h
c
a
n
lea
d
to
e
c
onomi
c
los
s
e
s
a
nd
r
e
duc
e
d
pr
oduc
ti
on
qua
li
ty.
C
ha
ll
e
nge
s
in
c
it
r
us
dis
e
a
s
e
de
tec
ti
on:
c
it
r
us
dis
e
a
s
e
s
,
c
a
us
e
d
by
f
unga
l,
ba
c
ter
ial,
a
nd
vir
a
l
a
ge
nts
,
a
r
e
a
major
thr
e
a
t
to
global
c
it
r
us
pr
oduc
ti
on.
T
r
a
dit
ional
methods
of
dis
e
a
s
e
de
tec
ti
on
r
e
ly
on
manua
l
ob
s
e
r
va
ti
on,
whic
h
i
s
ti
me
-
c
ons
umi
ng
a
nd
e
r
r
o
r
-
pr
one
.
T
he
pr
opos
e
d
method
tac
kles
the
c
ha
ll
e
nge
by
a
uto
mating
the
de
tec
ti
on
pr
oc
e
s
s
us
ing
im
a
ge
-
ba
s
e
d
a
na
lys
is
,
a
ll
owing
e
a
r
ly
de
tec
ti
on
a
nd
mi
nim
izing
c
r
op
los
s
e
s
.
3.
P
ROP
OS
E
D
S
Y
S
T
E
M
T
he
pa
pe
r
pr
opos
e
s
a
n
E
F
OA
int
e
gr
a
ted
with
a
T
C
NN
.
T
he
s
tr
a
tegy
be
hind
us
ing
the
E
F
OA
is
to
opti
mi
z
e
the
hype
r
-
pa
r
a
mete
r
s
of
the
T
C
NN
,
a
ll
owing
f
o
r
be
tt
e
r
f
e
a
tur
e
e
xt
r
a
c
ti
on
a
nd
c
las
s
if
ica
ti
on
pe
r
f
or
manc
e
.
T
h
is
dir
e
c
tl
y
a
ddr
e
s
s
e
s
the
pr
oblem
by
im
pr
oving
the
a
c
c
ur
a
c
y
a
nd
s
e
ns
it
ivi
ty
of
dis
e
a
s
e
de
tec
t
ion.
T
his
r
e
s
e
a
r
c
h
pr
ovides
a
n
innovative
s
olut
ion
f
o
r
a
uto
mate
d
c
it
r
us
dis
e
a
s
e
de
tec
ti
on,
with
the
E
F
OA
-
T
C
NN
model
outper
f
o
r
mi
ng
c
onve
nti
ona
l
methods
.
T
he
de
tailed
a
na
lys
is
a
nd
c
ompar
is
on
o
f
models
va
li
da
te
the
r
obus
tnes
s
of
the
p
r
opos
e
d
a
ppr
oa
c
h,
of
f
e
r
in
g
s
igni
f
ica
nt
potential
f
or
r
e
a
l
-
wor
ld
a
ppli
c
a
ti
ons
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
20
25
:
320
1
-
3213
3204
T
he
r
e
s
e
a
r
c
h
pa
ve
s
the
wa
y
f
or
f
utur
e
wo
r
k
in
opti
mi
z
ing
s
im
il
a
r
models
f
or
other
plant
dis
e
a
s
e
s
a
nd
a
gr
icultur
a
l
c
ha
ll
e
nge
s
.
Ou
r
r
e
s
e
a
r
c
h
de
mons
tr
a
ted
that
the
E
F
OA
-
T
C
NN
model
s
igni
f
ica
ntl
y
e
nha
nc
e
s
the
de
tec
ti
on
of
c
it
r
us
dis
e
a
s
e
s
c
ompar
e
d
to
tr
a
di
ti
ona
l
methods
.
W
e
f
ound
that
the
s
e
ns
it
ivi
ty
of
the
E
F
OA
-
T
C
NN
model
r
e
a
c
he
d
0.
975
,
indi
c
a
ti
ng
a
h
igh
tr
ue
pos
it
ive
r
a
te
f
o
r
identi
f
ying
d
is
e
a
s
e
d
s
a
mp
les
.
3.
1.
Dat
a
d
e
s
c
r
ip
t
ion
Ga
ther
e
d
in
or
a
nge
or
c
ha
r
ds
in
the
nor
thea
s
ter
n
M
e
xica
n
s
tate
s
of
S
a
n
L
uis
P
otos
i,
the
c
oll
e
c
ti
on
c
ontains
953
c
olo
r
pho
tos
of
C
it
r
us
s
pe
c
ies
le
a
ve
s
.
Or
a
nge
lea
ve
s
in
the
da
tas
e
t
a
r
e
c
a
tegor
i
z
e
d
int
o
12
gr
oups
,
a
s
s
hown
in
T
a
ble
1.
T
he
s
e
gr
oups
include
he
a
lt
hy,
s
ick,
nutr
ient
de
f
icie
nt,
a
nd
pe
s
ts
.
I
n
a
ddit
ion,
F
igur
e
s
1(
a
)
to
1
(
l)
dis
plays
e
xa
mpl
e
s
of
e
a
c
h
a
n
omaly,
s
howing
how
the
lea
ve
s
'
textur
e
a
nd
c
olor
pa
tt
e
r
ns
c
ha
nge
a
s
pr
e
dicte
d
[
27]
.
T
a
ble
1.
C
las
s
-
wi
s
e
im
a
ge
de
li
ve
r
y
in
the
da
tas
e
t
C
la
s
s
n
a
me
A
bnor
ma
li
ty
t
ype
# I
ma
ge
s
G
r
e
a
s
y s
pot
D
is
e
a
s
e
100
Fe
I
r
on de
f
ic
ie
nc
y
100
Mg
M
a
gne
s
iu
m l
a
c
k
100
Zn
Z
in
c
de
f
ic
ie
nc
y
100
H
e
a
lt
hy
N
ot
a
bnor
ma
l
100
H
L
B
D
is
e
a
s
e
43
T
e
xa
s
mi
te
P
e
s
t
100
R
e
d s
c
a
l
e
P
e
s
t
30
R
e
d s
c
a
l
e
s
e
qu
e
la
e
P
e
s
t
100
C
it
r
us
l
e
a
f
mi
ne
r
P
e
s
t
100
F
igur
e
1.
He
r
e
a
r
e
the
lea
f
s
a
mpl
e
s
include
d
in
the
da
tas
e
t:
(
a
)
s
tr
ong,
(
b)
hua
nglongbi
ng
,
(
c
)
gr
e
a
s
y
s
pot,
(
d)
i
r
on
de
f
icie
nc
y,
(
e
)
magne
s
ium
lac
k,
(
f
)
manga
ne
s
e
de
f
icie
nc
y,
(
g)
nit
r
oge
n
de
f
icit,
(
h
)
z
inc
de
f
icie
nc
y
,
(
i)
t
he
c
it
r
us
mi
te
of
T
e
xa
s
,
(
j)
c
it
r
us
lea
f
mi
ne
r
,
(
k)
r
e
d
s
c
a
le
a
nd
it
s
a
f
ter
e
f
f
e
c
ts
,
a
nd
(
l
)
r
e
d
s
c
a
le
it
s
e
lf
3.
2.
Clas
s
if
icat
ion
u
s
in
g
d
e
e
p
lear
n
in
g
n
e
t
wor
k
c
on
s
t
r
u
c
t
ion
F
or
the
pur
pos
e
o
f
plant
dis
e
a
s
e
c
a
tegor
iza
ti
on,
t
he
E
F
OA
-
T
C
NN
a
r
c
hit
e
c
tur
e
is
s
hown.
T
he
r
e
a
r
e
thr
e
e
ke
y
a
s
pe
c
ts
of
the
pictur
e
that
the
r
e
c
om
mende
d
de
e
p
C
NN
take
s
int
o
a
c
c
ount:
I
f
the
s
ize
of
the
de
s
c
r
ipt
ion
pa
tt
e
r
n
is
e
qua
l
f
il
ter
will
be
a
ble
to
id
e
nti
f
y
it
.
S
e
c
ond,
va
r
ious
por
ti
ons
of
the
input
pic
tur
e
us
e
of
the
s
a
me
f
or
ms
or
pa
tt
e
r
ns
[
27
]
.
C
onvolvi
ng
t
he
whole
s
our
c
e
im
a
ge
is
a
ddit
ional
models
.
F
inally,
the
ge
ometr
y
of
the
s
our
c
e
pictur
e
is
una
f
f
e
c
ted
by
do
wn
s
a
mpl
e
d
pixels
,
whic
h
laye
r
.
T
wo
c
onvolut
ion
laye
r
s
make
up
the
s
ugge
s
ted
E
F
OA
-
T
C
NN
,
with
the
pooli
ng
laye
r
s
a
nd
the
E
L
be
ing
dir
e
c
ted
by
a
thi
r
d
c
on
volut
ion
laye
r
.
Af
t
e
r
that
,
the
f
ull
y
c
onne
c
ted
(
F
C
)
laye
r
is
pa
i
r
e
d
with
a
S
of
tM
a
x
laye
r
.
B
y
taking
a
n
a
ve
r
a
ge
o
f
the
output
,
E
L
s
umm
a
r
ize
s
the
f
e
a
tur
e
maps
of
t
he
f
inal
c
onvolut
ional
laye
r
.
E
a
c
h
f
e
a
tur
e
map
ge
ts
a
va
lue
ba
c
k
f
r
om
thi
s
,
whic
h
is
the
s
a
me
a
s
the
e
ne
r
gy
ba
nk.
T
his
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
Optimiz
ing
c
it
r
us
dis
e
as
e
de
tec
ti
on:
a
tr
ans
fer
able
c
onv
olut
ional
ne
ur
al
…
(
A
noop
Ganadalu
L
ingar
aju)
3205
de
s
ign
not
only
r
e
duc
e
s
the
number
of
laye
r
s
,
b
ut
it
a
ls
o
us
e
s
les
s
memor
y
a
nd
c
omput
a
ti
on
ti
me
whe
n
lea
r
ning
textur
e
f
unc
ti
ons
,
a
nd
it
doe
s
it
quit
e
we
ll
.
T
h
is
pe
r
f
or
manc
e
-
c
omput
ing
ti
me
tr
a
de
-
of
f
is
made
pos
s
ibl
e
by
E
L
.
T
o
laye
r
'
s
da
ta
f
low,
thi
s
laye
r
is
a
dde
d.
S
hor
tl
y
f
oll
owing
the
las
t
pooli
ng
laye
r
,
t
he
E
L
's
f
lattene
d
output
is
r
ou
ted
to
the
c
onc
a
tena
ti
on
lay
e
r
.
I
n
o
r
de
r
to
im
pa
r
t
inf
or
mation
on
the
im
a
ge
's
s
ha
pe
a
nd
textur
e
to
the
c
ompl
e
tely
li
nke
d
laye
r
,
thi
s
c
onne
c
ti
on
ge
ne
r
a
tes
a
f
r
e
s
h,
f
lattene
d
ve
c
tor
.
T
he
mathe
matica
l
c
omput
a
ti
on
of
the
c
onvolut
ion
laye
r
's
output
s
ize
is
given
by
(
1
)
a
s
:
=
−
+
2
+
1
(
1)
w
he
r
e
be
s
ides
c
ha
r
a
c
ter
ize
s
the
input
be
s
ides
f
il
ter
s
ize
c
or
r
e
s
pondingl
y,
S
s
igni
f
ies
t
he
pa
d
ding,
be
s
ides
ϱ
is
the
s
tr
ide
wor
th.
Ne
xt,
the
f
ir
s
t
two
laye
r
s
of
the
th
r
e
e
c
onvolut
io
n
laye
r
s
a
r
e
t
r
a
ined
us
ing
a
5×
5
ke
r
ne
l
s
ize
,
a
nd
their
output
s
a
r
e
16
a
nd
32
c
ha
nne
ls
,
c
or
r
e
s
pondi
ngly.
T
he
th
ir
d
c
onvolut
ional
laye
r
,
whic
h
ha
s
6
4
output
c
ha
nne
l
s
be
s
ides
a
ke
r
ne
l
s
ize
of
3×
3,
is
us
e
d
a
s
a
t
r
a
ns
it
ional
laye
r
to
e
xtr
a
c
t
textur
e
a
tt
r
ibut
e
s
[
28]
.
T
he
c
onvolut
ion
laye
r
may
only
lea
r
n
31,
744
pa
r
a
mete
r
s
,
whic
h
a
r
e
de
ter
mi
ne
d
u
s
ing
the
f
o
r
mul
a
in
(
2)
a
nd
(
3)
:
=
×
(
×
+
1
)
(
2)
=
+
×
×
(
3)
w
h
e
r
e
m
e
a
n
s
t
h
e
C
N
N
l
a
y
e
r
l
e
a
r
n
a
b
l
e
l
i
m
i
t
s
,
c
h
a
r
a
c
t
e
r
i
z
e
s
t
h
e
k
e
r
n
e
l
s
i
z
e
,
t
h
e
n
s
i
g
n
i
f
i
e
s
t
h
e
c
h
a
n
n
e
l
s
u
m
.
T
he
output
of
the
input
ne
ur
on
is
c
omput
e
d
by
e
a
c
h
c
onvolut
ion
laye
r
.
I
ts
we
ight
plus
the
lea
s
t
input
f
ield
a
s
s
oc
iate
d
with
it
a
r
e
mu
lt
ipl
ied
by
a
dot
pr
oduc
t
to
ge
t
the
c
omput
a
ti
on.
A
16
-
ke
r
ne
l
out
put
with
dim
e
ns
ions
of
32×
32×
16
is
s
ha
p
e
d
by
the
ini
ti
a
l
c
onvolut
ion
laye
r
.
Ac
c
or
ding
to
(
4
)
,
the
f
ir
s
t
c
on
volut
ion
laye
r
ne
ur
ons
'
output
is
:
=
∑
×
+
(
4)
W
he
r
e
r
e
p
r
e
s
e
nts
the
maps
,
c
ha
r
a
c
te
r
i
z
e
s
the
f
e
a
tur
e
maps
tha
t
w
e
r
e
s
upp
li
e
d
,
a
nd
T
s
tan
ds
f
or
the
we
ight
e
d
ma
p.
T
he
las
t
laye
r
e
ne
r
g
y
de
s
c
r
ipt
ion
a
s
it
s
ou
tpu
t.
Af
ter
the
thi
r
d
c
onv
olut
ion
laye
r
,
e
ne
r
g
y
laye
r
s
a
r
e
mi
xe
d
c
r
it
e
r
ia
.
L
ike
a
de
s
c
r
i
pto
r
,
it
de
s
c
r
ibes
th
e
tex
tu
r
e
in
a
s
im
il
a
r
wa
y.
I
n
(
5)
p
r
ov
ides
r
e
lati
ons
hip
a
s
:
(
,
)
=
[
∑
+
=
1
]
(
5)
W
he
r
e
E
L
(
ξ
,
ϑ
)
s
tands
f
or
the
E
L
laye
r
,
j
f
or
the
in
put
inf
luenc
e
s
,
a
nd
T
f
o
r
the
E
L
we
ight
e
d
ve
c
tor
.
T
he
r
e
a
r
e
f
e
we
r
pa
r
a
mete
r
s
that
c
a
n
be
lea
r
ne
d
be
c
a
us
e
the
c
onne
c
ti
on
a
mong
the
E
L
a
nd
F
C
laye
r
s
is
na
r
r
o
we
r
than
the
las
t
t
r
a
dit
ional
c
onvolut
ion
laye
r
c
onne
c
ti
vit
y
[
29]
.
F
o
r
wa
r
d
a
nd
ba
c
kwa
r
d
p
r
opa
ga
ti
on
a
ls
o
a
ll
o
w
E
L
to
lea
r
n,
a
nd
it
r
e
membe
r
s
the
e
ne
r
gy
da
ta
f
r
om
t
he
p
r
e
vious
laye
r
.
I
n
a
ddit
ion
to
e
nha
nc
ing
the
lea
r
ning
c
a
pa
bil
it
y
a
nd
s
im
pli
f
ying
the
pr
o
jec
ted
s
ys
tem,
E
L
he
lps
to
de
c
r
e
a
s
e
the
ve
c
tor
s
ize
of
laye
r
.
T
o
de
ter
mi
ne
the
E
L
pa
r
a
mete
r
s
that
c
a
n
be
lea
r
ne
d,
us
e
(
6)
a
s
:
=
×
−
1
(
6)
whe
r
e
is
th
e
E
L
lea
r
na
ble
li
m
it
s
,
is
the
c
ur
r
e
nt
ne
ur
on,
be
s
ides
−
1
is
the
ne
ur
on.
B
e
twe
e
n
the
c
onvolut
ion
the
ba
tch
a
c
ti
va
ti
on
pur
p
os
e
is
uti
li
z
e
d
to
pr
oc
e
s
s
.
T
o
e
li
mi
na
te
the
int
e
r
na
l
c
ova
r
iate
s
hif
t,
ba
tch
nor
maliza
ti
on
is
e
mpl
oye
d.
T
he
s
tanda
r
d
de
viation
a
nd
mea
n
c
a
n
be
no
r
ma
li
z
e
d
to
a
c
hieve
thi
s
.
I
n
(
7)
a
nd
(
8)
a
r
e
uti
li
z
e
d
in
the
b
ulk
nor
maliza
ti
on
pr
oc
e
dur
e
to
de
ter
m
ine
the
m
e
a
n
a
nd
va
r
ianc
e
.
=
1
∑
(
7)
=
1
×
∑
(
−
)
2
(
8)
W
he
r
e
a
nd
c
ha
r
a
c
ter
i
z
e
s
the
mea
n
be
s
ide
s
va
r
ianc
e
c
or
r
e
s
pondingl
y,
is
the
s
ize
of
c
omponent
of
a
tt
r
ibut
e
s
.
A
va
lue
of
64
is
us
e
d
f
or
n
in
ou
r
wor
k
.
W
he
n
us
ing
(
9
)
to
de
ter
mi
ne
the
ba
tch
nor
maliza
t
ion,
the
r
e
s
ult
is
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
20
25
:
320
1
-
3213
3206
=
−
√
2
+
+
(
9)
whe
r
e
a
nd
a
r
e
the
va
lues
of
the
ini
ti
a
li
z
a
ble
pa
r
a
mete
r
s
f
or
e
a
c
h
outpu
t
f
unc
ti
on.
B
y
pluggi
ng
r
e
c
ti
f
ied
li
ne
a
r
unit
(
R
e
L
U
)
int
o
the
a
c
ti
va
ti
on
f
unc
ti
on
in
(
10)
a
nd
then
c
a
lcula
ti
ng
the
output
of
(
11)
a
s
:
,
,
=
0
,
,
,
(
10)
=
(
(
(
,
)
)
)
(
11)
W
he
r
e
,
,
s
igni
f
ies
f
e
a
tur
e
s
a
nd
,
,
s
tands
f
o
r
the
inp
ut
e
leme
nt's
a
tt
r
ibut
e
.
As
a
c
ons
e
que
nc
e
of
the
c
ontr
ol
ne
twor
k's
ove
r
f
it
ti
ng
,
the
pool
ing
laye
r
s
ubs
e
que
ntl
y
s
hr
inks
the
f
e
a
tur
e
mappings
,
we
igh
ts
,
a
nd
c
omput
a
ti
ons
.
In
(
12)
is
us
e
d
to
c
a
lcula
te
the
laye
r
numer
ica
ll
y
:
=
(
0
,
∑
−
1
)
(
12)
W
he
r
e
s
tands
f
or
the
f
e
a
tur
e
maps
that
will
be
ou
tput
,
ϑ
f
or
the
f
e
a
tur
e
maps
that
will
be
input
,
Q
f
or
the
s
ize
of
the
pooli
ng,
a
nd
T
f
or
the
maximum
la
ye
r
f
or
the
ve
c
tor
.
T
wo
max
laye
r
s
,
e
a
c
h
with
a
2×
2
ke
r
ne
l
s
c
ope
,
a
r
e
uti
li
z
e
d
in
thi
s
s
tudy.
F
or
e
a
c
h
we
ight
e
d
upda
te,
the
dr
opout
laye
r
r
e
m
ove
s
a
f
r
a
c
ti
on
of
r
a
ndoml
y
c
hos
e
n
pa
r
a
mete
r
s
in
or
de
r
to
pr
e
ve
nt
ove
r
f
it
ti
ng
of
the
tr
a
ini
ng
da
ta
[
30
]
.
I
n
o
r
de
r
to
pr
e
ve
nt
tr
a
ini
ng
da
ta
f
r
om
be
ing
ove
r
f
it
,
dr
op
e
dit
ing
is
e
mpl
oye
d
in
to
c
onti
nua
ll
y
e
l
im
inate
a
p
a
r
a
mete
r
.
Ove
r
-
c
ompatibi
li
ty
o
f
t
r
a
ini
ng
da
ta
is
a
pr
oblem
f
or
F
C
laye
r
s
be
c
a
us
e
they
ha
ve
the
mos
t
pa
r
a
m
e
ter
s
in
the
ne
twor
k.
B
e
c
a
us
e
of
thi
s
,
the
dr
opout
laye
r
is
de
c
ided
upon
s
ubs
e
que
nt
to
the
F
C
laye
r
.
A
c
las
s
if
ier
that
make
s
us
e
of
the
los
s
f
unc
ti
on
is
the
S
of
tM
a
x
laye
r
.
F
or
S
of
tM
a
x
,
the
pos
s
ibl
e
outcome
s
mi
gh
t
take
on
va
lues
be
twe
e
n
z
e
r
o
a
nd
one
.
I
n
(
13)
the
los
s
f
unc
ti
on
is
e
xpr
e
s
s
e
d
mathe
m
a
ti
c
a
ll
y
a
s
:
=
δ
+
∑
(
)
(
13)
W
he
r
e
s
tands
f
or
the
ove
r
a
ll
los
s
a
nd
δ
with
the
i
-
t
h
ve
c
tor
e
leme
nt's
c
las
s
d.
As
s
hown
in
(
14)
us
ing
the
S
of
t
M
a
x
f
unc
ti
on,
the
c
las
s
if
ier
's
goa
l
is
to
mi
ni
mi
z
e
the
pr
oba
bil
it
y
dis
c
r
e
pa
nc
y
be
twe
e
n
the
a
c
tual
a
nd
e
s
ti
mate
d
labe
ls
.
=
∑
(
)
(
14)
W
he
n
thi
s
s
tep
is
f
ini
s
he
d,
E
F
OA
-
T
C
NN
moves
on
to
the
hype
r
-
pa
r
a
mete
r
tuni
ng
p
r
oc
e
s
s
,
whic
h
will
be
de
s
c
r
ibed
in
the
s
ubs
e
c
ti
on
that
f
oll
ows
.
As
you
c
a
n
s
e
e
f
r
om
T
a
ble
2
,
the
input
a
nd
output
dim
e
ns
io
ns
of
the
pr
ojec
ted
ne
twor
k
a
r
e
f
ull
y
labe
ll
e
d.
T
a
ble
2.
P
r
opos
e
d
E
F
OA
-
T
C
NN
a
r
c
hit
e
c
tur
e
laye
r
s
T
ype
s
P
a
ddi
ng
K
e
r
ne
l
s
iz
e
t
o f
or
m
e
a
c
h f
e
a
tu
r
e
ma
p
S
tr
id
e
O
ut
put
s
iz
e
I
nput
s
iz
e
C
onvolut
io
na
l
l
a
ye
r
1
[
1 1 1 1]
5×
5
[
1 1]
62×
62×
16
64×
64×
1
M
a
x
pool
in
g l
a
ye
r
1
[
1 1 1 1]
2×
2
[
2 2]
32×
32×
16
62×
62×
16
C
onvolut
io
na
l
l
a
ye
r
2
[
1 1 1 1]
5×
5
[
1 1]
30×
30×
32
32×
32×
16
R
e
L
U
M
a
x
pool
in
g l
a
ye
r
2
[
1 1 1 1]
2×
2
[
2 2]
16×
16×
32
30×
30×
32
C
onvolut
io
na
l
l
a
ye
r
3
[
1 1 1 1]
3×
3
[
1 1]
16×
16×
64
16×
16×
32
R
e
L
U
EL
-
-
-
128×
1
16×
16×
64
D
r
opout
-
-
-
128×
1
128×
1
F
C
1
-
-
-
1024×
1
128×
1
D
r
opout
-
-
-
1024×
1
1024×
1
F
C
2
-
-
-
2×
1
1024×
1
S
of
tM
a
x
l
a
ye
r
-
-
-
-
-
C
la
s
s
if
ic
a
ti
on
l
a
ye
r
-
-
-
-
-
3.
2.
1.
F
in
e
-
t
u
n
in
g
u
s
in
g
E
F
OA
T
his
s
e
c
ti
on
be
gins
by
looki
ng
int
o
the
or
ig
ins
of
loga
r
it
hmi
c
s
pir
a
l
pa
thwa
ys
.
Af
ter
that,
a
n
a
da
ptable
s
witch
(
r
a
ti
o)
is
buil
t
to
s
tr
ike
the
r
ig
ht
c
ombi
na
ti
on
of
e
xplo
r
a
ti
on
a
nd
e
xploi
tation
ba
s
e
d
on
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
Optimiz
ing
c
it
r
us
dis
e
as
e
de
tec
ti
on:
a
tr
ans
fer
able
c
onv
olut
ional
ne
ur
al
…
(
A
noop
Ganadalu
L
ingar
aju)
3207
va
r
iations
in
the
va
lues
o
f
e
xpr
e
s
s
ions
on
the
f
a
c
e
.
E
F
OA
is
us
e
d
f
or
opti
mi
z
ing
the
hype
r
-
pa
r
a
mete
r
s
of
the
T
C
NN
model,
s
uc
h
a
s
the
lea
r
ning
r
a
te
,
ba
tch
s
ize
,
number
o
f
f
i
lt
e
r
s
,
a
nd
dr
opout
r
a
tes
.
Optim
iza
ti
on
of
thes
e
hype
r
-
pa
r
a
mete
r
s
is
c
r
it
ica
l
to
a
c
hieving
the
be
s
t
pos
s
ibl
e
model
pe
r
f
or
manc
e
,
a
s
im
pr
ope
r
hype
r
-
pa
r
a
me
ter
tuni
ng
c
a
n
lea
d
to
poor
ge
ne
r
a
li
z
a
ti
on
or
ove
r
f
it
ti
ng.
I
n
E
F
OA
,
e
a
c
h
f
r
ui
t
f
ly
r
e
pr
e
s
e
nts
a
c
a
ndidate
s
olut
ion
(
a
s
e
t
of
hype
r
-
pa
r
a
mete
r
s
)
,
a
nd
the
a
lgor
i
thm
e
xplor
e
s
the
s
e
a
r
c
h
s
pa
c
e
to
f
ind
the
be
s
t
c
ombi
na
ti
on
of
pa
r
a
mete
r
s
that
mi
nim
ize
s
the
c
las
s
i
f
ica
ti
on
e
r
r
or
on
the
va
li
da
ti
on
da
tas
e
t.
T
he
s
e
a
r
c
h
pr
oc
e
s
s
in
E
F
OA
invol
ve
s
both
loca
l
a
nd
global
s
e
a
r
c
h
pha
s
e
s
to
ba
lanc
e
e
xplor
a
ti
on
a
nd
e
xploi
tation.
T
he
loca
l
s
e
a
r
c
h
r
e
f
ines
the
s
e
a
r
c
h
a
r
ound
p
r
omi
s
ing
s
olut
ions
,
while
the
glob
a
l
s
e
a
r
c
h
e
ns
ur
e
s
the
a
lgor
it
hm
doe
s
not
ge
t
s
tuck
in
lo
c
a
l
mi
nim
a
.
T
he
e
nha
nc
e
d
ve
r
s
ion
of
F
OA
int
r
oduc
e
s
im
pr
ove
ments
in
it
s
s
e
a
r
c
h
s
tr
a
tegie
s
,
includ
ing
a
da
pti
ve
pa
r
a
mete
r
tuni
ng
a
nd
mu
lt
i
-
dim
e
ns
ional
e
xplor
a
ti
on,
whic
h
he
lp
in
f
a
s
ter
c
onve
r
ge
nc
e
to
the
opti
mal
hype
r
-
pa
r
a
mete
r
s
.
T
he
int
e
gr
a
ti
on
o
f
a
ne
w
a
da
ptable
s
witch
(
r
a
ti
o)
a
nd
a
n
e
nha
nc
e
d
F
A
is
then
s
hown.
3.
2.
2.
De
s
ign
of
t
h
e
logarit
h
m
ic
s
p
iral
p
at
h
T
he
S
phe
r
e
f
unc
ti
on
(
)
=
∑
2
=
1
a
nd
S
c
hwe
f
e
l
f
unc
ti
on
(
)
=
∑
(
∑
=
1
)
2
=
1
(
xi
va
r
ies
be
twe
e
n
-
100
to
100,
a
nd
d
is
s
e
t
to
10)
a
r
e
c
h
os
e
n
a
s
the
be
nc
hmar
k
f
unc
ti
ons
.
Ou
r
r
e
s
e
a
r
c
h
i
nvolves
r
unning
tes
ts
50
ti
mes
to
c
a
lcula
te
the
mea
n
pe
r
f
o
r
manc
e
.
B
y
doing
thi
s
,
the
f
inal
r
e
s
ult
s
a
r
e
les
s
a
f
f
e
c
ted
by
unpr
e
dicta
bil
it
y
in
population
pos
it
ion
ini
ti
a
li
z
a
ti
on.
I
n
thes
e
50
r
uns
,
the
a
ppr
opr
iate
va
r
iable
s
f
or
e
ve
r
y
a
lgor
it
hm
a
r
e
a
ll
given
the
s
a
me
va
lue.
T
he
qua
nti
t
y
of
s
e
a
r
c
he
r
s
is
15
,
=
0
.
2
,
0
=
1
=
1
.
T
he
potential
f
o
r
loca
l
s
pa
c
e
uti
li
z
a
ti
on
ha
s
g
one
unnoti
c
e
d.
How
to
s
u
s
tain
e
xplor
a
ti
on
a
nd
e
xploi
tation
of
f
i
r
e
f
ly
is
a
n
int
r
igui
ng
topi
c
in
the
s
e
a
r
c
h
pr
oc
e
s
s
.
W
he
n
the
pr
oblem
is
s
olved,
the
opti
mi
z
e
r
's
e
f
f
e
c
ti
ve
ne
s
s
will
incr
e
a
s
e
a
nd
the
c
omput
a
ti
ona
l
load
will
de
c
r
e
a
s
e
.
W
e
s
olve
thi
s
c
onund
r
um
by
c
ons
ider
ing
lar
ge
r
a
pto
r
s
that
t
r
a
ve
l
in
a
logar
it
h
mi
c
s
pir
a
l,
li
ke
pe
r
e
gr
ine
f
a
lcons
in
s
e
a
r
c
h
of
f
o
od.
T
h
is
s
tr
a
tegy
is
ba
s
e
d
on
e
xpe
r
i
menta
l
bio
logy.
W
e
've
a
ls
o
not
ice
d
a
s
im
il
a
r
f
l
ight
pa
tt
e
r
n
f
o
r
f
ir
e
f
li
e
s
a
t
night
.
T
he
logar
i
thm
ic
s
pir
a
l
pa
th
is
one
tec
hnique
that
c
ould
be
us
e
d
to
i
mpr
ove
F
A
e
xploi
tation
.
T
he
de
s
ign
of
a
nove
l
pos
it
ion
upda
ti
ng
method
ba
s
e
d
on
thi
s
is
ho
w
the
logar
it
hm
ic
s
pir
a
l
looks
,
,
+
1
=
,
+
0
.
−
.
2
.
(
,
−
,
)
⨂
.
⨂
co
s
(
2
.
)
(
15)
I
n
(
15
)
,
l
is
a
n
e
ve
n
r
a
ndom
ve
c
tor
in
d
dim
e
ns
ion
s
in
[
−
1
,
1
]
;
the
logar
i
thm
ic
s
pir
a
l's
s
ha
pe
is
de
ter
mi
ne
d
by
the
c
ons
tant
b,
whic
h
ha
s
a
de
f
a
ult
va
lue
of
1
.
T
o
de
s
c
r
ibe
the
pos
it
ion
upda
te
methodology
in
(
1
6
)
,
two
c
a
us
e
s
de
ter
mi
ne
thi
s
pos
it
ion
va
r
iation
:
log
a
r
it
hm
ic
s
pir
a
l
pa
ths
a
nd
br
igh
tnes
s
int
e
ns
it
y.
T
he
c
oe
f
f
i
c
ients
in
mathe
matics
r
e
pr
e
s
e
nt
the
latter
.
⨂
co
s
(
2
.
)
.
3.
2.
3.
Adap
t
ive
s
wit
c
h
d
e
s
ign
I
n
or
de
r
to
e
qua
li
z
e
both
e
xplo
r
ing
a
nd
e
xploi
ti
ng
modes
,
thi
s
r
e
s
e
a
r
c
h
pr
ovides
the
a
da
pti
ve
s
witch
(
r
a
ti
o)
tec
hnique.
W
hich
a
pp
r
oa
c
h
will
be
e
mpl
oye
d
in
the
f
oll
owing
it
e
r
a
ti
on
:
{
ℎ
,
>
,
ℎ
,
≤
,
(
16)
W
he
r
e
u
is
a
ge
ne
r
a
ted
number
r
a
ndoml
y
wi
th
unif
or
m
dis
tr
ibu
ti
on
[
0
,
1]
a
nd
the
is
c
omput
e
d
in
the
pr
e
vious
it
e
r
a
ti
on.
T
he
e
xploi
tation
method
mus
t
be
s
e
lec
ted
with
a
gr
e
a
ter
pr
oba
bil
it
y
than
a
mode
of
e
xplor
a
ti
on
in
or
de
r
to
ha
s
ten
the
opti
mi
z
e
r
's
c
onve
r
g
e
nc
e
.
T
hus
,
we
de
s
c
r
ibe
the
c
ha
nge
ove
r
R
_(
t+1)
r
a
nging
f
r
o
m
[
0
.
5,
1]
.
S
e
tt
ing
the
be
ginni
ng
r
a
ti
o
t
o
0.
5
r
e
s
ult
s
in
the
va
lue
of
the
a
da
ptable
s
witch
:
+
1
=
{
1
1
+
e
x
p
(
−
∗
−
1
∗
)
,
⌊
lg
|
∗
|
⌋
≠
⌊
lg
|
−
1
∗
|
⌋
,
1
1
+
e
x
p
(
−
∗
−
.
⌊
∗
⌋
−
1
∗
−
.
⌊
−
1
∗
⌋
)
,
(
17)
I
n
(
17
)
,
∗
a
t
the
t
-
th
it
e
r
a
ti
on
,
is
the
be
s
t
f
unc
ti
o
n's
f
it
ne
s
s
va
lue
;
(
·
)
=
10
(
·
)
;
⌊
·
⌋
is
the
f
loor
's
pur
pos
e
.
T
he
c
a
lcula
ti
on
of
the
(
18
)
is
done
us
ing
t
he
f
or
mul
a
by
the
a
da
ptable
s
c
a
le
pa
r
a
mete
r
thr
e
s
hold
:
=
10
⌊
lg
|
∗
−
−
1
∗
|
⌋
+
1
(
18)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
20
25
:
320
1
-
3213
3208
W
e
s
ha
ll
e
xa
mi
ne
thr
e
e
s
it
ua
ti
ons
in
o
r
de
r
to
c
omp
r
e
he
nd
how
thi
s
f
o
r
mul
a
wo
r
ks
in
g
r
e
a
ter
de
tail:
−
T
he
f
ir
s
t
c
ondit
ion
,
∗
≫
−
1
∗
.
As
a
r
e
s
ult
,
ther
e
is
a
s
iza
ble
ga
p
be
twe
e
n
two
it
e
r
a
ti
ons
.
T
he
pr
oc
e
s
s
of
opti
mi
s
a
ti
on
now
ha
s
a
f
r
e
s
h,
be
tt
e
r
opti
on
than
ks
to
s
e
a
r
c
h
a
ge
nts
.
He
nc
e
,
a
n
a
da
pti
ve
r
a
ti
o
+
1
be
c
omes
1
a
nd
the
ne
xt
it
e
r
a
ti
on's
a
mode
o
f
e
xplo
r
a
ti
on
is
s
e
lec
ted
;
−
If
∗
≪
−
1
∗
,
then
s
hif
t
to
e
xplor
a
ti
on
with
r
e
ga
r
d
to
of
p
e
r
f
or
manc
e
de
gr
a
da
ti
on.
T
he
f
lexible
r
a
ti
o
+
1
will
be
0.
5
,
a
nd
we
'll
make
us
ing
a
mode
of
e
xplo
r
a
ti
on
mor
e
li
ke
ly
the
f
oll
owing
s
e
a
r
c
he
s
.
−
T
he
f
inal
p
r
e
r
e
quis
it
e
is
⌊
lg
|
∗
|
⌋
=
⌊
lg
|
−
1
∗
|
⌋
,
im
ply
th
e
d
is
c
ove
r
y
of
a
c
los
e
s
t
mi
nim
um.
B
y
the
i
tem,
we
a
djus
t
the
r
a
ti
o
by
(
19
)
to
incr
e
a
s
e
the
s
e
ns
it
ivi
ty
of
the
a
da
pti
ve
s
witch.
I
n
thi
s
s
c
e
na
r
io,
the
r
e
will
be
a
high
li
ke
li
hood
of
s
e
a
r
c
h
a
ge
nts
e
s
c
a
ping
potential
tr
a
ps
.
∗
−
.
⌊
∗
⌋
−
1
∗
−
.
⌊
−
1
∗
⌋
(
19)
T
he
s
c
a
li
ng
f
a
c
tor
⌊
lg
|
∗
−
−
1
∗
|
⌋
c
a
n
a
utom
a
t
ica
ll
y
de
tec
t
the
f
indi
ng
s
tatus
.
T
hus
,
our
a
da
pti
ve
s
witch
im
pr
ove
s
c
onve
r
ge
nc
e
e
ve
n
mor
e
.
F
o
ll
owi
ng
that,
the
logi
s
ti
c
f
unc
ti
on
is
us
e
d
to
c
onve
r
t
the
va
r
iation
to
a
pr
oba
bil
it
y
.
T
he
a
da
ptable
s
witc
h
r
a
ti
o
is
then
de
ter
mi
ne
d.
B
a
s
e
d
on
the
idea
o
f
a
f
lexible
s
witc
h
layout,
we
dis
c
ove
r
that
thi
s
pa
r
ti
c
ular
s
witch
ha
s
a
gr
e
a
ter
c
a
pa
c
it
y:
a
ge
nts
that
s
e
a
r
c
h
a
c
ti
vit
y
ha
s
incr
e
a
s
e
d
in
c
hoos
ing
a
mode
of
e
xplor
a
ti
on
whe
n
the
r
e
is
a
n
int
e
r
r
upti
on
du
r
ing
the
p
r
oc
e
s
s
of
s
e
a
r
c
hing
to
e
ns
ur
e
the
opti
mi
z
a
ti
on
a
lgor
it
hm
to
f
ind
a
gr
e
a
ter
idea
l.
3.
2.
4.
F
ire
f
ly'
s
u
p
d
a
t
e
d
algori
t
h
m
T
he
upda
ted
loca
ti
on
f
or
mul
a
is
now
dis
playe
d
a
s
(
20)
:
,
+
1
=
{
,
+
0
.
−
.
2
.
(
,
−
,
)
+
.
[
−
0
.
5
]
⨂
,
>
,
,
+
0
.
−
.
2
.
(
,
−
,
)
⨂
.
⨂
co
s
(
2
.
)
,
≤
.
(
20)
Our
r
e
vis
e
d
f
ir
e
f
ly
method,
ter
med
the
a
da
ptable
logar
it
hmi
c
s
pir
a
l
f
i
r
e
f
ly
a
lgor
it
hm,
whic
h
in
tegr
a
tes
the
a
dva
ntage
s
of
F
A
with
a
moder
nize
d
the
logar
i
th
mi
c
s
pir
a
l
pa
thwa
y
c
ontr
oll
e
d
by
a
s
mar
t
a
da
pti
v
e
s
witch.
T
he
ps
e
udo
-
c
ode
de
mons
tr
a
tes
the
modi
f
ica
ti
ons
that
we
r
e
made
f
or
the
c
onve
nti
ona
l
F
A
f
r
a
me
wor
k
to
a
c
c
omm
oda
te
our
f
lexible
s
witch.
T
he
mos
t
r
e
c
e
n
t
va
lue
of
the
f
it
ne
s
s
f
unc
ti
on
f
_t
^
*
is
logged
to
c
onf
igur
e
the
f
lexible
s
witch.
W
e
will
e
mpl
oy
a
tr
a
dit
ional
t
e
c
hnique
in
whic
h
the
va
lue
o
f
the
s
witch
is
f
ixed
a
t
0.
5
to
a
s
s
e
s
s
the
e
f
f
e
c
ti
ve
ne
s
s
of
the
c
a
pa
bil
it
y
of
the
a
da
pti
ve
s
witch
a
nd
the
logar
it
hm
s
pir
a
l
pa
th.
I
n
(
21)
is
the
method
f
or
upda
ti
ng
p
os
it
ion
:
,
+
1
=
{
,
+
0
.
−
.
2
.
(
,
−
,
)
+
.
[
−
0
.
5
]
⨂
,
>
0
.
5
,
,
+
0
.
−
.
2
.
(
,
−
,
)
⨂
.
⨂
co
s
(
2
.
)
,
≤
0
.
5
.
(
21)
4.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
T
o
c
onduc
t
the
r
e
s
e
a
r
c
h,
an
I
ntel
C
or
e
i5
-
7200
C
P
U
be
s
ides
8
GB
o
f
int
e
r
na
l
memor
y
is
uti
li
z
e
d.
T
he
pr
oc
e
s
s
or
is
a
c
c
ompl
is
he
d
of
r
unning
a
t
2.
7
G
Hz
.
De
voted
Us
e
r
I
nter
f
a
c
e
(
U
I
)
be
s
ides
J
upyter
N
otebook
(
P
ython
3
.
7)
pe
r
f
or
m
the
ope
r
a
ti
ons
on
W
indows
1
0,
a
6
4
-
bit
ope
r
a
ti
ng
s
ys
tem
na
tur
a
l
s
e
tt
ing
.
4.
1.
Valid
a
t
ion
a
n
alys
is
of
p
r
op
os
e
d
m
od
e
l
wi
t
h
e
xis
t
in
g
p
r
oc
e
d
u
r
e
s
T
a
bl
e
3
p
r
o
vi
de
s
t
he
e
x
pe
r
im
e
n
tal
in
ve
s
ti
ga
ti
on
of
p
r
e
dic
ta
bl
e
f
a
u
lt
les
s
w
it
h
e
x
is
t
in
g
p
r
o
c
e
d
u
r
e
s
i
n
t
e
r
ms
o
f
di
f
f
e
r
e
n
t
met
r
ics
.
I
n
T
a
bl
e
3
m
e
a
ns
t
ha
t
t
he
va
l
ida
t
io
n
s
t
ud
y
o
f
p
r
oj
e
c
ted
f
a
u
lt
les
s
w
it
h
e
x
is
t
in
g
t
e
c
hn
iq
ue
s
.
I
n
t
his
i
nv
e
s
t
ig
a
t
io
n
,
t
he
M
L
P
t
e
c
h
ni
que
a
tt
a
i
ne
d
s
e
ns
it
iv
it
y
a
s
0
.
9
15
b
e
s
i
de
s
s
p
e
c
i
f
ic
it
y
a
s
0
.
9
7
a
nd
a
c
c
u
r
a
c
y
o
f
0
.
95
7
a
nd
F
-
mea
s
u
r
e
o
f
0
.
9
37
a
nd
p
r
e
c
is
i
on
a
s
0
.
8
40
c
or
r
e
s
p
on
di
ng
ly
.
T
h
e
n
t
he
a
u
to
e
n
c
o
de
r
t
e
c
hn
iq
ue
a
t
ta
ine
d
s
e
ns
i
t
iv
it
y
a
s
0
.
9
35
,
s
pe
c
i
f
ic
it
y
a
s
0
.
9
8
,
a
c
c
u
r
a
c
y
of
0
.
9
53
,
F
-
mea
s
u
r
e
o
f
0
.
9
46
,
a
n
d
p
r
e
c
is
i
on
a
s
0
.
8
87
c
o
r
r
e
s
po
nd
in
g
ly
.
T
he
n
t
he
de
e
p
b
e
l
ie
f
n
e
two
r
k
(
DB
N)
te
c
h
ni
qu
e
a
tt
a
i
ne
d
s
e
n
s
i
ti
vi
t
y
a
s
0
.
9
21
,
s
pe
c
i
f
i
c
i
ty
a
s
0
.
99
,
a
c
c
u
r
a
c
y
o
f
0
.
96
4
,
F
-
mea
s
u
r
e
o
f
0
.
94
2
,
a
n
d
p
r
e
c
is
io
n
a
s
0
.
8
92
c
o
r
r
e
s
p
on
di
ng
ly
.
T
h
e
n
t
he
C
NN
tec
hn
iq
ue
a
t
ta
ine
d
s
e
ns
it
iv
it
y
a
s
0
.
9
36
,
s
pe
c
if
i
c
i
ty
a
s
0
.
99
,
a
c
c
u
r
a
c
y
o
f
0
.
9
88
,
F
-
mea
s
u
r
e
o
f
0
.
9
6
5
,
a
n
d
p
r
e
c
is
i
on
a
s
0
.
91
3
c
o
r
r
e
s
po
nd
i
ng
ly
.
T
he
n
t
he
T
C
NN
te
c
h
ni
qu
e
a
t
ta
in
e
d
s
e
ns
it
i
vi
ty
a
s
0
.
95
6
,
s
pe
c
i
f
ic
it
y
a
s
0
.
96
,
a
c
c
u
r
a
c
y
o
f
0
.
9
81
,
F
-
m
e
a
s
ur
e
o
f
0
.
9
71
,
a
n
d
pr
e
c
is
io
n
a
s
0
.
9
37
c
o
r
r
e
s
po
nd
in
gl
y
.
T
he
n
t
he
E
F
OA
-
T
C
NN
t
e
c
hn
iq
ue
a
t
ta
ine
d
s
e
ns
i
t
iv
it
y
a
s
0
.
9
75
,
s
pe
c
i
f
ic
it
y
a
s
1
.
0
0
,
a
c
c
u
r
a
c
y
of
0
.
9
95
,
F
-
mea
s
u
r
e
o
f
0
.
9
86
,
a
n
d
p
r
e
c
is
i
on
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
Optimiz
ing
c
it
r
us
dis
e
as
e
de
tec
ti
on:
a
tr
ans
fer
able
c
onv
olut
ional
ne
ur
al
…
(
A
noop
Ganadalu
L
ingar
aju)
3209
a
s
0
.
9
59
c
o
ng
r
u
e
n
tl
y
.
T
h
e
F
i
gu
r
e
s
2
(
a
)
a
n
d
2
(
b
)
s
hows
t
he
v
is
u
a
l
r
e
p
r
e
s
e
n
ta
ti
on
o
f
p
r
o
je
c
te
d
c
las
s
i
c
a
l
a
nd
c
om
pa
r
is
o
n
a
na
l
ys
is
o
f
pr
op
os
e
d
w
it
h
e
xis
ti
ng
m
ode
ls
f
o
r
pl
a
n
t
d
is
e
a
s
e
d
e
t
e
c
t
i
on
r
e
s
pe
c
t
ive
l
y
.
T
a
ble
3
.
Va
li
da
ti
on
a
na
lys
is
of
p
r
opos
e
d
textbook
with
e
xis
ti
ng
tec
hniques
C
la
s
s
if
ie
r
s
S
e
ns
it
iv
it
y
S
pe
c
if
ic
it
y
A
c
c
ur
a
c
y
F
-
me
a
s
ur
e
P
r
e
c
is
io
n
M
L
P
0.915
0.97
0.957
0.937
0.840
AE
0.935
0.98
0.953
0.946
0.887
D
B
N
0.921
0.99
0.964
0.942
0.892
C
N
N
0.936
0.99
0.988
0.965
0.913
T
C
N
N
0.956
0.96
0.981
0.971
0.937
E
F
O
A
-
T
C
N
N
0.975
1.00
0.995
0.986
0.959
(
a
)
(
b)
F
igur
e
2
.
P
e
r
f
or
manc
e
a
nd
c
ompar
a
ti
ve
a
na
lys
is
,
(
a
)
vis
ua
l
r
e
pr
e
s
e
ntation
of
pr
ojec
ted
c
las
s
ica
l
a
nd
(
b)
c
ompar
is
on
a
na
lys
is
of
pr
opos
e
d
with
e
xis
ti
ng
models
f
or
plant
dis
e
a
s
e
de
tec
ti
on
T
he
im
pr
ove
ment
in
pr
e
c
is
ion
a
nd
F
-
mea
s
ur
e
indi
c
a
tes
be
tt
e
r
ba
lanc
e
in
ha
ndli
ng
both
tr
ue
pos
it
ives
a
nd
f
a
ls
e
pos
it
ives
.
T
his
table
pr
ovides
a
c
onc
is
e
ye
t
c
ompr
e
he
ns
ive
c
ompar
is
on,
c
lea
r
ly
de
mons
tr
a
ti
ng
the
e
f
f
e
c
ti
ve
ne
s
s
of
the
p
r
opos
e
d
model.
T
h
is
a
ppr
oa
c
h
ke
e
ps
the
manus
c
r
ipt
mana
ge
a
ble
while
s
howc
a
s
ing
the
s
igni
f
ica
nt
pe
r
f
or
manc
e
ga
ins
o
f
the
E
F
OA
-
T
C
NN
.
Va
li
da
ti
on
a
na
lys
is
c
ompar
ing
the
pe
r
f
or
manc
e
of
the
pr
opos
e
d
E
F
OA
-
T
C
NN
model
with
e
xis
ti
ng
tec
hniques
,
s
pe
c
if
ica
ll
y
the
r
e
f
e
r
e
nc
e
model
[
4]
,
is
pr
e
s
e
nted
in
T
a
ble
4.
T
he
a
na
lys
is
highl
ight
s
ke
y
metr
ics
s
uc
h
a
s
s
e
n
s
it
ivi
ty,
s
pe
c
if
icity
,
a
c
c
ur
a
c
y,
F
-
mea
s
ur
e
,
a
nd
pr
e
c
is
ion.
T
he
r
e
f
e
r
e
nc
e
model
a
c
h
ieve
d
a
s
e
ns
it
ivi
ty
of
95%
,
s
pe
c
if
icity
o
f
96%
,
a
c
c
ur
a
c
y
of
96
%
,
F
-
mea
s
ur
e
of
95
%
,
a
nd
pr
e
c
is
ion
of
96%
.
I
n
c
ontr
a
s
t,
the
E
F
OA
-
T
C
NN
model
outper
f
or
med
the
r
e
f
e
r
e
nc
e
model
with
a
s
e
ns
it
ivi
ty
of
97
.
5%
,
pe
r
f
e
c
t
s
pe
c
if
icity
of
100%
,
a
n
a
c
c
ur
a
c
y
o
f
99.
5%
,
a
n
F
-
mea
s
ur
e
of
98
.
6%
,
a
nd
a
pr
e
c
is
ion
o
f
95.
9
%
.
T
his
indi
c
a
tes
that
the
E
F
OA
-
T
C
NN
model
de
mons
tr
a
tes
s
upe
r
i
or
pe
r
f
o
r
manc
e
a
c
r
os
s
a
ll
metr
ics
,
s
igni
f
ica
ntl
y
e
nha
nc
ing
de
tec
ti
on
c
a
pa
bil
it
ies
in
th
e
c
ontext
of
the
s
tudy.
S
uc
h
f
indi
ngs
unde
r
li
ne
the
e
f
f
e
c
ti
ve
ne
s
s
of
int
e
gr
a
ti
ng
opti
mi
z
a
ti
on
tec
hniques
li
ke
E
F
OA
in
de
e
p
lea
r
ni
ng
models
f
or
im
pr
ove
d
c
las
s
if
ica
ti
on
tas
ks
,
a
s
s
uppor
ted
by
va
r
ious
s
tudi
e
s
in
the
f
ield
o
f
mac
hine
lea
r
ning
a
nd
im
a
ge
pr
oc
e
s
s
ing.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
20
25
:
320
1
-
3213
3210
T
a
ble
4
.
Va
li
da
ti
on
a
na
lys
is
of
p
r
opos
e
d
textbook
with
e
xis
ti
ng
tec
hniques
C
la
s
s
if
ie
r
s
S
e
ns
it
iv
it
y (
%
)
S
pe
c
if
ic
it
y (
%
)
A
c
c
ur
a
c
y (
%
)
F
-
me
a
s
ur
e
(
%
)
P
r
e
c
is
io
n (
%
)
R
e
f
e
r
e
nc
e
[
4]
95
96
96
95
96
E
F
O
A
-
T
C
N
N
0.975
1.00
0.995
0.986
0.959
4.
2
.
Dis
c
u
s
s
ion
T
his
r
e
s
e
a
r
c
h
p
r
e
s
e
nts
the
de
ve
lopm
e
nt
of
a
n
opti
mi
z
e
d
model
f
or
c
it
r
us
dis
e
a
s
e
de
tec
ti
on,
f
oc
us
ing
on
im
pr
oving
a
c
c
ur
a
c
y
a
nd
e
f
f
icie
nc
y
thr
ough
the
int
e
gr
a
ti
on
of
the
E
F
OA
with
a
T
C
NN
.
T
he
e
xpe
r
im
e
ntal
r
e
s
ult
s
de
mons
tr
a
te
that
the
pr
opos
e
d
E
F
OA
-
T
C
NN
model
s
igni
f
ica
ntl
y
outpe
r
f
or
ms
t
r
a
dit
ional
mac
hine
lea
r
ning
a
nd
de
e
p
lea
r
n
ing
tec
hniques
in
de
tec
ti
ng
dis
e
a
s
e
s
f
r
om
im
a
ge
s
of
or
a
nge
lea
ve
s
.
T
he
E
F
O
A
-
T
C
NN
model
a
c
hieve
d
r
e
mar
ka
ble
pe
r
f
or
manc
e
a
c
r
os
s
ke
y
metr
ics
,
with
a
n
a
c
c
ur
a
c
y
of
0.
995
,
s
e
ns
it
ivi
ty
of
0.
975
,
a
nd
pr
e
c
is
ion
of
0.
959
.
T
he
s
e
r
e
s
ult
s
indi
c
a
te
the
model's
high
c
a
pa
c
it
y
to
c
or
r
e
c
tl
y
identif
y
dis
e
a
s
e
d
lea
ve
s
while
mi
nim
izing
f
a
ls
e
pos
it
ives
a
nd
ne
ga
ti
ve
s
.
T
he
model's
us
e
o
f
a
n
e
ne
r
gy
laye
r
in
plac
e
of
a
tr
a
dit
ional
pooli
ng
laye
r
a
ll
owe
d
f
or
be
tt
e
r
f
e
a
tur
e
e
xtr
a
c
ti
o
n,
pr
e
s
e
r
ving
c
r
uc
ial
im
a
ge
de
tails
that
c
ont
r
ibut
e
d
to
it
s
s
upe
r
ior
c
las
s
if
ica
ti
on
pe
r
f
or
manc
e
.
T
h
is
modi
f
ic
a
ti
on,
c
ombi
ne
d
wi
th
the
hype
r
-
pa
r
a
mete
r
tuni
ng
pr
ovided
by
E
F
OA
,
r
e
s
ult
e
d
in
im
pr
ove
d
s
e
ns
it
ivi
ty
a
nd
a
c
c
ur
a
c
y
c
ompar
e
d
to
the
ba
s
e
li
ne
C
NN
model
a
nd
other
c
las
s
if
ier
s
s
uc
h
a
s
M
L
P
a
nd
AE
.
T
he
opti
m
iza
ti
on
o
f
T
C
NN
pa
r
a
mete
r
s
us
ing
E
F
O
A
playe
d
a
c
r
it
ica
l
r
ole
in
the
model
’
s
s
uc
c
e
s
s
.
B
y
f
ine
-
tuni
ng
hype
r
-
pa
r
a
mete
r
s
s
uc
h
a
s
lea
r
ning
r
a
te,
f
il
ter
s
ize
,
a
nd
dr
opout
r
a
tes
,
the
E
F
OA
s
ign
if
ica
ntl
y
e
nha
nc
e
d
the
T
C
NN
’
s
a
bil
it
y
to
c
a
ptur
e
int
r
ica
te
dis
e
a
s
e
p
a
tt
e
r
ns
in
the
input
im
a
ge
s
.
T
his
de
mons
tr
a
tes
the
im
por
tanc
e
of
a
pplyi
ng
opt
im
iza
ti
on
a
lgo
r
it
hm
s
to
im
pr
ove
de
e
p
lea
r
ning
a
r
c
hit
e
c
tur
e
s
f
or
s
pe
c
if
ic
a
ppli
c
a
ti
ons
,
s
uc
h
a
s
a
gr
icultu
r
a
l
dis
e
a
s
e
de
tec
ti
on.
I
n
c
ompar
is
on
to
other
c
las
s
if
ier
s
,
the
p
r
opos
e
d
E
F
OA
-
T
C
NN
model
c
ons
is
tently
outper
f
or
med
M
L
P
,
AE
,
DB
N,
a
nd
C
NN
in
a
ll
ke
y
metr
ics
.
T
h
e
c
los
e
s
t
c
ompeting
model,
C
NN
,
a
c
hieve
d
a
n
a
c
c
ur
a
c
y
of
0.
988,
but
the
E
F
OA
-
T
C
NN
model
s
ti
ll
s
ur
p
a
s
s
e
d
it
with
a
notable
0.
995
a
c
c
ur
a
c
y.
T
he
im
p
r
ove
ment
in
s
e
ns
it
ivi
ty
(
0.
975)
a
nd
s
pe
c
if
icity
(
1.
00
)
s
hows
the
model's
e
xc
e
pti
ona
l
ba
lanc
e
be
twe
e
n
c
or
r
e
c
tl
y
ide
nti
f
ying
dis
e
a
s
e
d
lea
ve
s
a
nd
a
voidi
ng
f
a
ls
e
pos
it
ives
,
whic
h
is
c
r
it
ica
l
in
r
e
a
l
-
wor
ld
a
gr
icult
ur
a
l
s
e
tt
ings
whe
r
e
ove
r
-
diagnos
ing
dis
e
a
s
e
s
c
a
n
lea
d
to
unne
c
e
s
s
a
r
y
tr
e
a
tm
e
nts
a
nd
c
os
ts
.
T
he
r
e
s
e
a
r
c
h
f
indi
ngs
pr
ovide
s
tr
ong
e
videnc
e
th
a
t
the
pr
opos
e
d
E
F
OA
-
T
C
NN
model
is
a
highl
y
e
f
f
e
c
ti
ve
tool
f
or
a
utom
a
ted
c
it
r
us
dis
e
a
s
e
de
te
c
ti
on.
B
y
int
e
gr
a
ti
ng
th
e
E
F
OA
,
the
model
s
igni
f
ica
ntl
y
outper
f
or
ms
c
onve
nti
ona
l
tec
hniques
a
nd
de
m
ons
tr
a
tes
potential
f
o
r
r
e
a
l
-
wor
ld
im
pleme
ntati
on.
T
he
inclus
ion
of
opti
mi
z
a
ti
on
a
lgo
r
it
hms
s
uc
h
a
s
E
F
OA
he
lps
br
idge
the
ga
p
be
twe
e
n
theor
e
ti
c
a
l
model
de
ve
lopm
e
nt
a
nd
pr
a
c
ti
c
a
l
a
pp
li
c
a
ti
ons
,
making
thi
s
a
ppr
oa
c
h
s
c
a
lable
f
or
br
oa
de
r
.
5.
CONC
L
USI
ON
De
e
p
lea
r
ning
ha
s
a
c
hieve
d
s
igni
f
ica
nt
a
dva
nc
e
m
e
nts
in
a
gr
icultur
e
,
pa
r
ti
c
ular
ly
in
a
utom
a
ti
ng
the
identif
ica
ti
on
o
f
p
lant
dis
e
a
s
e
s
while
mi
nim
izing
t
he
r
e
li
a
nc
e
on
e
xtens
ive
human
invol
ve
m
e
nt.
T
his
r
e
s
e
a
r
c
h
f
oc
us
e
d
on
de
ve
lopi
ng
a
n
a
utom
a
ted
s
ys
tem
f
o
r
d
is
e
a
s
e
identif
ica
ti
on
in
c
it
r
us
lea
ve
s
thr
ough
de
e
p
lea
r
ning
a
nd
opti
mal
f
e
a
tur
e
s
e
lec
ti
on.
I
nit
ially
,
da
ta
a
ugme
ntation
wa
s
e
mpl
oye
d
to
e
xpa
nd
the
da
tas
e
t,
e
nha
n
c
ing
the
r
obus
tnes
s
of
de
e
p
lea
r
ning
r
e
pr
e
s
e
ntations
.
T
o
f
u
r
ther
im
pr
ove
the
pr
e
c
is
ion
a
nd
e
f
f
icie
nc
y
o
f
plan
t
dis
e
a
s
e
de
tec
ti
on,
thi
s
s
tudy
int
r
oduc
e
d
a
nove
l
a
ppr
oa
c
h
that
int
e
gr
a
tes
mul
ti
ple
methodologi
e
s
.
S
pe
c
if
ica
ll
y,
a
hype
r
pa
r
a
mete
r
-
tuned
T
C
NN
wa
s
uti
li
z
e
d
to
e
nha
nc
e
c
las
s
if
ic
a
ti
on
pe
r
f
or
manc
e
.
T
his
a
ppr
oa
c
h
s
tr
e
a
ml
ined
the
a
r
c
hit
e
c
tur
e
by
r
e
plac
ing
c
onve
nti
ona
l
pooli
n
g
laye
r
s
with
jus
t
thr
e
e
e
ne
r
gy
laye
r
s
,
ther
e
by
m
a
king
the
model
mor
e
a
c
c
e
s
s
ibl
e
a
nd
e
f
f
e
c
ti
ve
.
T
he
us
e
o
f
th
e
E
F
OA
f
a
c
il
it
a
ted
e
f
f
icie
nt
hype
r
pa
r
a
mete
r
tuni
n
g,
w
hich
c
ontr
ibut
e
d
to
im
p
r
ove
d
model
pe
r
f
or
manc
e
.
T
h
e
e
xtr
a
c
ted
f
e
a
tur
e
s
we
r
e
s
ubs
e
que
ntl
y
c
a
tegor
ize
d
us
ing
va
r
ious
s
upe
r
vis
e
d
lea
r
ning
a
lgor
it
hms
.
Although
the
f
us
ion
of
s
e
lec
ted
de
e
p
f
e
a
tur
e
s
s
igni
f
ica
ntl
y
e
nha
nc
e
d
de
tec
ti
on
a
c
c
ur
a
c
y,
it
wa
s
a
c
c
ompanie
d
b
y
a
tr
a
de
-
of
f
in
incr
e
a
s
e
d
c
omput
a
ti
ona
l
ti
me.
L
ooking
a
he
a
d,
thi
s
r
e
s
e
a
r
c
h
lays
the
gr
oundwor
k
f
or
im
pleme
nti
n
g
r
e
a
l
-
ti
me
moni
tor
ing
s
ys
tems
f
or
c
it
r
us
plant
dis
e
a
s
e
de
tec
ti
on.
F
utu
r
e
wo
r
k
wi
ll
e
xplor
e
the
int
e
g
r
a
ti
o
n
of
thi
s
model
wi
th
int
e
r
ne
t
of
thi
ngs
(
I
o
T
)
de
v
ice
s
a
nd
e
dge
c
omput
ing
s
ys
tems
,
e
na
bli
ng
r
e
a
l
-
ti
me
dis
e
a
s
e
de
tec
ti
on
in
a
gr
icultur
a
l
f
ields
.
S
uc
h
int
e
gr
a
ti
o
n
would
f
a
c
il
it
a
te
r
a
pid
on
-
s
it
e
a
na
lys
is
of
plant
he
a
lt
h
w
it
hout
r
e
lyi
ng
on
c
e
ntr
a
li
z
e
d
s
e
r
ve
r
s
,
a
ll
owing
f
a
r
mer
s
to
r
e
c
e
ive
im
media
te
f
e
e
dba
c
k
a
nd
make
ti
mely
int
e
r
ve
nti
ons
.
M
or
e
ove
r
,
f
utu
r
e
s
tudi
e
s
c
ould
f
oc
us
on
e
nha
nc
ing
the
model's
e
f
f
icie
nc
y
thr
ough
opti
mi
z
a
ti
on
tec
hniques
that
r
e
duc
e
c
omput
a
t
ional
ove
r
he
a
d
while
maintaining
a
c
c
ur
a
c
y.
T
he
e
xplor
a
ti
on
of
a
da
pti
ve
lea
r
ning
s
ys
tems
that
c
onti
nuous
ly
im
pr
ove
with
ne
w
da
ta
c
a
n
a
ls
o
be
a
ke
y
a
r
e
a
f
o
r
de
ve
lopm
e
nt
.
T
his
r
e
s
e
a
r
c
h
unde
r
s
c
or
e
s
the
potential
of
AI
-
powe
r
e
d
dis
e
a
s
e
de
tec
ti
on
s
ys
tems
in
a
gr
icultu
r
e
a
nd
s
e
ts
the
s
tage
f
o
r
f
ur
ther
a
dva
nc
e
ments
in
pr
e
c
is
ion
f
a
r
m
ing
tec
hnolog
ies
.
B
y
c
ombi
ning
a
dva
nc
e
d
i
mage
a
n
a
lys
is
a
nd
opti
mi
z
a
ti
on
tec
hniques
,
we
c
a
n
unl
oc
k
ne
w
a
ve
nue
s
f
or
s
us
taina
ble
c
r
op
pr
otec
ti
on
methods
,
u
lt
im
a
tely
be
ne
f
it
ing
both
p
r
oduc
e
r
s
a
nd
c
ons
umer
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.