Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
15
,
No.
1
,
Febr
uary
20
25
, pp.
940
~
948
IS
S
N:
20
88
-
8708
, DO
I:
10
.11
591/ij
ece.v
15
i
1
.
pp
940
-
948
940
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om
T
o
m
a
t
o
p
l
a
n
t
d
i
s
e
a
s
e
p
r
e
d
i
c
t
i
o
n
s
y
s
t
e
m
w
i
t
h
a
n
e
w
f
r
a
m
e
w
o
r
k
S
S
M
A
N
u
s
i
n
g
a
d
v
a
n
c
e
d
d
e
e
p
l
e
a
r
n
i
n
g
t
ec
hn
i
ques
Sa
r
avanan
M
ad
deri
Siv
alingam
1
, L
ak
sh
mi
D
e
vi Bada
bagni
2
1
Dep
artm
en
t of
Co
m
p
u
ter
Sci
en
ce a
n
d
E
n
g
in
eering
,
Sav
eeth
a Scho
o
l of
En
g
in
eering
,
Sav
eeth
a I
n
stitu
te of
Medi
cal and
T
echn
ical
Scien
ces, Ch
en
n
ai,
I
n
d
ia
2
Sav
eeth
a Scho
o
l of E
n
g
in
eering
,
Sa
v
eeth
a I
n
stitu
te of
Medical and
T
echn
ical Sciences,
Sri
Pad
m
av
ati
Mahila
Visv
av
id
y
alay
am
,
Tir
u
p
ath
i,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
un
10, 2
024
Re
vised
A
ug 13, 2
024
Accepte
d
Aug 20,
2024
Agric
ult
ur
e
p
la
y
s
a
p
ivot
a
l
ro
le
i
n
India
's
e
cono
my,
and
the
t
imely
d
etec
t
ion
of
pl
ant
inf
ec
t
io
ns
is
essenti
a
l
to
safe
gu
ard
cro
ps
and
pr
eve
nt
fur
t
her
spre
ad
of
disea
ses.
Th
e
conve
n
ti
ona
l
ap
proa
ch
invol
ves
ma
nua
l
inspe
ct
i
on
of
pl
ant
le
av
es
to
ide
n
ti
f
y
the
s
pec
i
fic
ty
pe
of
d
isea
se
,
a
ta
sk
typ
ic
a
ll
y
carrie
d
ou
t
by
far
me
rs
or
p
la
n
t
pat
ho
logi
sts.
In
pre
vious
stu
die
s,
you
only
look
once
(YO
LO)
and
fa
ster
reg
ion
-
base
d
convol
ut
iona
l
neur
al
net
work
(R
-
CNN
),
ma
ch
ine
learni
n
g
al
gor
it
hms
we
re
appl
i
ed
to
d
a
ta
sets
for
d
e
tecting
objects
on
tom
a
to
l
ea
v
e
s
which
include
s
a
total
of
imag
es
2403
and
go
t
accuracie
s
of
86
and
82
p
erc
en
t.
In
th
is
p
ape
r,
a
d
ee
p
co
nvolut
ional
neur
al
net
work
(DCN
N)
mode
l
proposed
with
a
new
fra
me
work
sepa
ra
te,
shift
,
and
m
erg
e
base
d
AlexNe
t5
0
al
gori
t
hm
(SS
MA
N)
is
used
to
pre
dict
th
e
di
sea
se
a
t
an
ea
rl
ie
r
stag
e
wi
t
h
higher
ac
cu
racy.
Among
var
io
us
pre
-
tr
ai
ned
d
ee
p
models,
AlexNe
t
em
erg
e
s
as
the
top
per
forme
r,
a
chi
ev
in
g
the
h
ighe
st
accura
cy
in
disea
se
c
la
ss
ifica
t
ion.
SS
MA
N
ca
n
addr
ess
anom
a
li
es
in
im
ag
es
by
em
ployi
ng
a
c
las
s
dec
omposi
ti
o
n
appr
o
ac
h
to
scruti
ni
ze
cl
ass
b
oundar
ie
s.
AlexNe
t
exh
ibi
t
s
a
nota
bl
e
a
cc
u
rac
y
of
98
.
30%
in
succ
essfully
i
dent
ifyi
ng
tom
at
o
l
ea
f
d
ise
ase
s
from
im
ag
e
s,
with
pr
e
-
tr
ai
n
ed
new
fr
amework,
superior
to
th
e
or
igi
na
l
AlexNe
t
arc
h
it
e
ct
ure
as
we
ll
a
s
tra
di
ti
ona
l
cl
a
ss
ifi
ca
ti
on
me
thods wi
th
ot
her
a
lgori
th
ms.
Ke
yw
or
d
s
:
Accurac
y
A
gri
cultu
re
Cl
assifi
cat
ion
Conv
olu
ti
onal
neural
netw
ork
Deep l
ear
ning
Tom
at
o
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
BY
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Sara
van
a
n Ma
dd
e
ri Si
valinga
m
Dep
a
rtme
nt
of
Com
pu
te
r
Scie
nce a
nd E
ng
i
ne
erin
g,
Saveet
ha
Sc
hool
of E
ng
i
neer
i
ng, S
a
veetha
I
ns
ti
tute o
f
M
e
dical
and T
echn
ic
al
S
ci
en
ces
Tha
nd
al
am
–
602 1
05, C
he
nnai
, Ind
ia
Emai
l:
saran
e
na
du@
gm
ai
l.co
m
1.
INTROD
U
CTION
Id
e
ntify
i
ng
pl
ant
diseases
i
s
par
a
mou
nt
to
guara
ntee
the
culti
vatio
n
of
he
al
thy
agr
ic
ultur
al
pro
du
ct
io
n.
M
it
igati
ng
t
he
de
trime
ntal
ef
fec
ts
of
s
uc
h
dise
ases
is
ess
enti
al
,
especial
ly
c
on
si
der
i
ng
the
wi
de
arr
a
y
of
f
ood
from
popula
r
plants
s
uc
h
a
s
tomato
an
d
onio
n.
Far
mers
encou
nter
c
hal
le
ng
es
in
acc
urat
el
y
recog
nizing
s
ympto
ms
of
pla
nt
diseases,
pa
rtic
ularly
w
he
n
deali
ng
with
s
ub
tl
e
mani
festat
ion
s.
T
om
at
oe
s
not
on
l
y
f
ulfill
dai
ly
nutrit
io
nal
r
equ
i
reme
nts
but
al
so
ha
ve
a
high
ec
onom
ic
value
a
nd
pl
ay
a
si
gn
ific
a
nt
ro
le
in
man
y
local
an
d
natio
nal
eco
nomies.
D
eep
l
earn
i
ng
has
a
ppli
cat
ion
s
in
va
rio
us
fiel
ds
of
arti
fici
al
intel
l
igence,
par
ti
cula
rly
c
omp
uter
visio
n,
w
her
e
it
has
achieve
d
s
ucce
ss
in
plant
le
a
ves
reg
i
on
of
i
nterest
detect
ion
[
1]
.
The
dee
p
le
ar
ni
ng
models
use
d
the
c
onvolu
ti
on
al
ne
ural
ne
twork
(C
NN)
model
t
o
s
hape
the
data
incl
ud
i
ng
you
on
l
y
lo
ok
on
ce
(
YOLO
)
and
faster
re
gio
n
-
based
co
nv
olu
ti
onal
ne
ura
l
netw
ork
(R
-
CNN
)
,
for
detect
ing
tomatoes
a
nd
i
mpro
ving
dete
ct
ion
acc
uracy
[
2]
.
Dee
p
le
ar
ning
models
i
n
To
mato
de
te
ct
ion
an
d
co
mp
a
re
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
Toma
t
o pla
nt di
sease pre
dicti
on system
wi
th
a new fra
mew
ork
SSMAN
…
(
Sa
r
av
anan M
adde
ri Siv
alinga
m
)
941
stren
gth
s
a
nd
li
mit
at
ion
s
of
diff
e
re
nt
arc
hitec
tures.
H
owe
ver,
the
la
bor
-
intensive
an
d
ineff
ic
ie
nt
nat
ur
e
of
man
ual
ha
rv
e
s
ti
ng
prese
nts
a
sign
ific
a
nt
c
halle
ng
e
to
th
e
tomato
farm
ing
in
dustr
y
[
3]
.
The
ha
rv
e
sti
ng
proce
dure
is
c
os
tl
y,
raisi
ng
the
final
c
os
t
t
o
t
he
c
onsume
r,
a
nd
sl
ow,
de
la
yin
g
the
product's
delive
r
y
t
o
the
mar
ket
an
d
l
oweri
ng
it
s
nutr
it
ion
al
an
d
aest
hetic
value
.
T
oma
toe
s
a
re
now
f
re
qu
e
ntly
c
ulti
vated
in
s
uc
cessi
ve
crops.
S
om
e
i
mporta
nt
relat
ed
st
ud
ie
s
in
di
cat
e
that
to
ma
toes
a
re
a
ff
li
ct
ed
by
a
bout
30
dif
fer
e
nt
t
ypes
of
fun
gal
infecti
ons.
I
n
Chi
na,
t
om
at
o
in
fecti
ons
hav
e
bec
om
e
a
majo
r
pro
bl
em,
res
ulti
ng
in
a
10%
decr
e
ase
in
ou
t
pu
t.
Co
mp
l
et
e
crop
fail
ure
has
bee
n
re
ported
in
reg
i
on
s
w
he
re
the
se
diseases
ha
ve
had
a
sig
ni
ficant
influ
e
nce
[4]
.
Du
e
to
t
he
va
g
ue
sig
ns
of
tomato
in
fecti
ons
in
their
earl
y
sta
ge
s,
farm
ers
s
om
et
imes
fail
to
diag
nose
an
d
t
reat
these
il
lne
sses.
O
ften
,
thi
s
error
mea
ns
that
the
best
ti
me
to
pr
e
ven
t
and
c
ontr
ol
dis
ease
is
missed.
The
s
pr
a
ying
of
e
nor
mous
am
ou
nts
of
f
ungicides
pr
ov
e
d
to
be
inef
fectua
l
as
tomato
il
lnesse
s
worse
ned.
A
no
ther
set
of
farmers
fin
ds
it
di
ff
ic
ult
to
deter
mine
w
het
her
t
heir
to
matoes
are
co
ntaminat
ed
an
d
is
unable
to
de
te
rmin
e
the
e
xt
ent
of
the
il
lnesses
[5]
.
As
a
r
esult,
the
y
hea
vily
rely
on
th
e
us
e
of
f
ungic
ides
to
pr
e
ve
nt
an
d
co
ntr
ol
disease
.
Re
gr
et
ta
bl
y,
c
on
ti
nue
d
us
e
of
the
se
met
hods
res
ults
in
an
ov
e
r
us
e
of
f
un
gicides,
end
a
ngeri
ng
both
human
hea
lt
h
and
t
he
en
vi
ronme
nt
[6]
.
Ther
e
has
bee
n
a
rise
in
inte
rest
in
tomat
o
disease
researc
h
i
n
r
e
cent
yea
rs
due
to
the
gr
ow
i
ng
need
f
or
pr
ompt
an
d
e
ff
i
ci
ent
tomat
o
di
sease
identifi
cat
ion
,
detect
ion,
an
d
thera
py
a
ppli
cat
ion
s.
Acc
ur
at
el
y
dif
fer
e
ntiat
ing
the
uniq
ue
trai
ts
of
var
i
ous
diseases
is
diff
ic
ult
wh
e
n
usi
ng
e
xi
sti
ng
le
a
rn
i
ng
m
od
el
s
direct
ly
on
t
om
at
o
di
sease
de
te
ct
ion
ta
s
ks
.
T
his
r
est
rict
ion
f
requen
tl
y
resu
lt
s i
n
a siza
ble num
be
r of
error
s
or
om
iss
ion
s
.
As
a
res
ult,
bo
t
h
the
aca
de
mic
and
a
gr
i
cultural
c
om
m
un
it
ie
s
now
c
onfro
nt
simi
la
r
issues
in
integrati
ng
with
some
existi
ng
le
arn
i
ng
methods
with
avai
l
able
old
know
le
dg
e
base
an
d
pr
e
dict
at
early
sta
ge
with
higher
pe
rformance
li
ke
accu
racy
th
rough
j
oin
t
“
data
m
od
el
knowle
dg
e
”
a
ppro
ac
he
s
[
7]
.
T
his
res
earch
intends
to
c
ombine
dee
p
le
arn
i
ng
models
with
disease
relat
ed
tomat
o
le
aves
in
form
at
ion
to
overc
om
e
t
he
unbalance
d
di
stribu
ti
on
of
disease
sa
mp
l
es
an
d
the
la
ck
of
cl
ea
r
ar
ti
c
ulati
on
.
Ta
ki
ng
i
nto
acc
ount
t
he
com
plexity
of
tomato
diseas
e
data
a
nd
ma
king
us
e
of
e
xisti
ng
inf
orm
at
ion
,
t
he
st
udy
f
oc
us
es
on
t
om
at
o
diseases
t
hat
a
rise
in
c
omple
x
backd
rops
[
8]
.
T
her
e
are
504
a
rtic
le
s
in
G
oogle
Sc
hola
r
a
nd
10
0
a
rtic
le
s
on
IEEE
from
20
18
t
o
2024.
T
he
researc
h
unve
il
s
a
novel
de
ep
-
le
ar
ning
fr
a
mew
ork
e
ng
i
ne
ered
t
o
cat
eg
or
iz
e
a
wide
sp
ect
rum
of
plant
dise
ases.
De
par
ti
ng
from
co
nve
ntion
al
mac
hin
e
le
ar
ning
pa
rad
i
gm
s
that
man
date
man
ual
featu
r
e
extracti
on
an
d
data
feed
i
ng,
the
pro
pose
d
m
od
el
ca
pital
iz
es
on
c
onvoluti
on
al
neural
net
works
to
sig
nificantl
y
au
gme
nt
it
s
performa
nce.
Within
t
he
domain
of
deep
le
arn
in
g
mod
el
s,
two
f
unda
mental
op
e
rati
ons
are
cond
ucted
.
I
ni
ti
al
ly,
an
arti
f
ic
ia
l
neu
ral
ne
twork
is
de
ployed
t
o
ide
ntif
y
patte
r
ns
within
a
n
input
ima
ge.
S
ub
s
eq
ue
ntly,
the
model
aut
onomo
us
l
y
rend
ers
decisi
ons.
As
a
res
ult,
de
ep
le
a
rn
i
ng
models
dem
onstrat
e
e
nh
a
nce
d
e
ff
ic
a
cy
in
ima
ge
c
la
ssific
at
ion
ta
sk
s
c
ompa
red
to
tra
diti
on
al
l
earn
i
ng
al
gorit
hm
s
.
Util
iz
ing
a
de
ep
-
le
ar
ni
ng
neural
netw
or
k
f
or
ima
ge
cl
assifi
cat
ion
yields
nota
ble
eff
ect
ive
ness.
This
innov
at
ive
a
ppro
ac
h
outpe
rfo
rms
tradit
io
nal
meth
ods,
s
uc
h
as
ma
nual
in
sp
ect
io
n
of
in
di
vid
ual
pla
nt
l
eaves,
by d
el
ive
rin
g f
ast
er and mo
re
accurate
res
ults.
Farme
rs
hav
e
t
he
op
portu
nity
to
util
iz
ing
th
e
smartp
hone
cameras
t
o
det
ect
diff
e
ren
t
pl
ant
diseases
and
ta
ke
proa
ct
ive
meas
ures
to
c
urb
their
s
pr
ea
d
by
int
egr
at
in
g
with
oth
e
r
i
nterfac
e
a
pp
li
cat
ions.
M
a
ny
exp
e
rts strive t
o
bolst
er
to
mato
le
aves
, by
un
iqu
e
featur
e
sel
ect
ion
of
plant
diseases th
rou
gh the c
ombine
d
us
e
of
c
onve
ntio
na
l
methods
a
nd
neural
net
work
-
base
d
te
ch
niques.
Fi
gure
1
s
hows
th
e
proc
ess
of
dee
p
le
a
rn
i
ng
models
f
or
i
m
age
cl
assifi
cat
ion.
T
his
f
ram
ewor
k
ser
ves
as
the
f
oundat
ion
for
the
pr
opos
e
d
m
odel
,
wh
ic
h
le
ver
a
ges
th
e
s
epar
at
e,
sh
i
ft, a
nd me
rg
e
pre
-
t
raine
d
le
ar
ning
model.
Figure
1. Ge
ne
ral proce
ss
of
“
deep le
ar
ning
model f
or ima
ge
cl
assifi
cat
io
n
”
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
940
-
948
942
Figure
1
ta
ke
n
as
base
f
or
pro
posed
f
ra
mew
ork
se
pa
r
at
e,
sh
ift,
an
d
mer
ge
base
d
Alex
Net5
0
(S
S
MAN
)
to
com
bin
e
with
Alex
Net5
0,
a
deep
le
a
rn
i
ng
m
od
el
or
al
gorithm
.
Sect
ion
2
delineat
es
the
method
ology
of
th
e
tomat
o
dataset
,
offe
r
ing
in
sig
hts
in
to
it
s
structu
re
and
ac
quisi
ti
on.
I
n
s
ect
io
n
3,
the
pro
po
se
d
a
l
gor
it
hm
is
e
xpounded
up
on,
pro
vi
din
g
a
detai
le
d
acc
ount
of
it
s
impleme
ntati
on,
the
intric
ac
ie
s
of
the
S
S
M
mode
l,
an
d
an
e
xha
us
ti
ve
anal
ys
is
of
the
d
at
aset
.
Mo
ving
to
s
e
ct
ion
4,
a
th
oroug
h
e
xami
nat
ion
of
the
res
ults
wi
th
pe
rformanc
e
par
a
mete
rs
wer
e
discu
ss
ed
f
or
plant
di
sease
detect
ion
s
a
re
pr
ese
nted.
Con
cl
ud
i
ng the
stu
dy,
s
ect
io
n 5 e
ncapsulat
es
our fin
dings a
nd
offer
s
conclu
sive r
e
mar
ks
.
2.
RESEA
R
CH MET
HO
D
In
the
existi
ng
sy
ste
m,
man
ua
l
insp
ect
io
n
methods
an
d
a
fe
w
cl
assifi
ca
ti
on
meth
od
s
i
nv
e
sti
gating
ind
iv
i
du
al
plant
le
a
ves
or
c
lusters,
no
t
de
li
ver
in
g
faster
an
d
m
ore
ac
cur
at
e
res
ults.
T
her
e
fore,
thi
s
pap
e
r
fo
ll
ows
the
de
ep
le
a
rn
i
ng
te
c
hn
i
qu
e
as
a
ba
se
to
s
olv
e
the
earl
y
pr
e
dicti
on
of
to
mato
pl
ant
disease
.
F
or
thi
s
CNN
model
us
e
d.
I
n
Fi
gur
e
1
it
shows
the
ba
sic
str
uc
ture
of
C
NN
model.
I
n
th
is
CNNs
sta
rt
with
conv
olu
ti
onal
layers
t
hat appl
y
filt
ers or ker
nels to
the i
nput image.
Filt
ers
tra
ver
se t
he i
nput,
ei
ther b
y sl
iding
or
c
onvolvi
ng,
to
ide
ntify
fe
at
ur
es
li
ke
e
dges,
te
xtures,
or
patte
r
ns
.
Ea
ch
c
onvoluti
on
process
pro
duces
a
featur
e
ma
p
th
at
acce
ntu
at
es
sp
eci
fic
detai
ls
of
t
he
in
pu
t.
Pooli
ng
la
yer
s
fo
ll
ow
“
c
onvoluti
onal
la
yers
”
fo
r
dimensi
onal
it
y
reducti
on
of
t
he
in
put.
T
he
tradit
ion
al
methods
us
e
d
f
or
CNN
i
nclu
de
max
poolin
g,
wh
ic
h
sel
ect
s
the
hi
ghest
valu
e
in
a
s
pecified
re
gion,
an
d
ave
r
age
poolin
g
c
al
culat
es
the
a
ver
a
ge
val
ue
i
n
t
hat
reg
i
on.
P
ooli
ng
decr
ea
ses
c
omp
utati
on
al
l
oa
d
a
nd
e
nhanc
es
the
net
wor
k'
s
abili
ty
t
o
ha
nd
le
in
pu
t
va
riat
ion
s.
The
outp
uts
f
r
om
t
he
c
onvol
ution
al
an
d
po
oling
la
ye
rs
a
r
e
flat
te
ne
d
int
o
a
one
-
dime
ns
i
on
al
vect
or
,
w
hich
is
then
i
nput
into
the
f
ully
c
onne
ct
ed
la
yers.
I
n
f
ully
c
onnect
ed
la
ye
rs,
eac
h
neur
on
is
li
nk
ed
to
e
ve
ry
ne
uro
n
in
the
nex
t
la
ye
r.
Th
ese
la
yer
s
l
earn
high
-
le
vel
feat
ur
es
a
nd
ge
ner
at
e
predict
ion
s
by
co
m
bin
in
g
t
hese
fea
t
ur
es
.
CNNs
un
derg
o
trai
ning
us
in
g
la
bele
d
dataset
s.
Th
r
oughout
this
process
,
the
netw
ork
it
erati
vely
adj
us
ts
it
s
internal
pa
ram
et
ers,
w
hich
i
nclu
de
weig
ht
s
an
d
biases
,
to
mi
nimize
t
he
discre
pa
nc
y
betwee
n
pre
dicte
d
ou
t
pu
ts
and t
he
provide
d
la
bels.
Durin
g
t
rai
ning,
a
loss
f
un
ct
i
on
is
ap
plied
t
o
mea
sure
the
disp
a
rity
betw
een
pr
e
dicte
d
ou
t
pu
ts
an
d
true
la
bels,
a
nd
a
n
op
ti
miza
ti
on
al
gorith
m,
s
uch
as
st
ochast
ic
gradie
nt
desce
nt,
is
em
ployed
t
o
it
era
ti
vel
y
update
the
net
work
'
s
pa
rame
te
rs.
T
his
it
erati
ve
adjustme
n
t
of
par
a
mete
r
s
ena
bles
CN
Ns
to
a
uton
omo
us
ly
extract
hierar
c
hical
featu
res
f
rom
in
put
ima
ges.
This
i
nher
ent
capa
bili
ty
po
sit
io
ns
t
hem
as
po
werfu
l
t
ools
f
or
var
i
ou
s
ima
ge
-
relat
ed
ta
sks,
i
nclu
ding
obje
ct
dete
ct
ion
,
ima
ge
cl
assifi
cat
io
n,
an
d
se
gme
ntati
on
.
Sim
ultane
ously,
a
CN
N
m
odel
was
trai
ne
d
f
or
a
to
mato
pla
nt
image
s
dataset
colle
ct
ed
from
t
he
And
hr
a
Pr
a
desh
r
egi
on,
total
in
g
2
,
403
image
s,
in
this
80%
we
re
al
lo
cat
ed
for
t
raini
ng
data,
w
hile
the
remaini
ng
20%
wer
e
sel
ect
ed
for
te
sti
ng
dat
a.
This
ex
per
i
mental
co
nfi
gurati
on
was
i
mp
le
me
nted
at
the
M
achi
ne
Learn
i
ng
Lab
or
at
or
y
within
the
A
rtific
ia
l
I
ntell
igence
Dep
a
rtme
nt
at
Saveetha
Sc
ho
ol
of
E
ng
i
neer
i
ng,
pa
rt
of
Sa
ve
et
ha
In
sti
tute
of
Me
dical
an
d
Te
chn
ic
al
Scie
nc
es,
locat
e
d
in
Chen
nai,
Ta
mil
Nadu.
Th
e
subse
qu
e
nt
s
ect
ion
ou
tl
ines
the
pr
opos
e
d
S
S
M
m
od
el
.
3.
PROP
OSE
D SSM
A
N MO
DEL WITH
DE
EP L
EAR
NI
NG
The
propose
d
method
ology
e
mp
lo
ys
t
he
sep
arate
,
SS
MAN
model
of
dee
p
CNN
for
cl
ass
ifying
plant
diseases
with
higher
accu
rac
y.
Fig
ur
e
2
pr
esents
a
detai
le
d
dep
ic
ti
on
of
the
Ale
xN
et
deep
le
ar
ning
model
,
sh
owin
g
how
t
he
va
rio
us
la
ye
rs
are
i
ntegrat
ed
be
f
or
e
a
nd
after
the
know
le
dg
e
tra
nsfo
r
mati
on
process
.
T
hi
s
fig
ur
e
highli
ghts
the
strat
e
gi
c
placement
and
util
iz
at
ion
of
the
se
co
mpon
e
nts
to
i
ncr
ease
t
he
or
der
of
performa
nce
i
n
ov
e
rall
ef
fici
ency
a
nd
fun
ct
ion
al
it
y
of
t
he
m
odel
.
In
t
his
m
odel
,
w
e
util
iz
ed
an
A
le
xN
et
model
pre
-
t
rained
on
Ima
ge
Net
dataset
.
T
he
cl
ass
dec
omp
os
it
ion
la
ye
r
di
vid
es
a
set
of
im
ages
in
di
ff
e
ren
t
n
su
bc
la
sses,
w
it
h
each
s
ubcl
ass
bein
g
processe
d
in
de
pende
ntly.
T
he
cl
ass
c
ompo
sit
io
n
c
ompone
nt
su
bse
que
ntly
r
ecom
bin
es
t
he
se
subcl
asses
t
o
f
orm
the
fin
al
set
of
res
ult
s
with
cl
assifi
e
d
syst
em
f
o
r
ori
gin
al
image
dataset
[
8]
–
[
10]
.
In
the
Ale
xN
et
arc
hitec
ture,
i
t
i
s
note
d
that
ei
gh
t
la
yer
s
c
onta
in
tr
ai
na
ble
par
a
mete
rs.
It
i
s
obser
ve
d
that
amo
ng
the
se
la
yer
s
,
fi
ve
util
iz
e
a
com
bin
at
ion
of
c
onvoluti
onal
ope
ra
ti
on
s
a
nd
ma
x
poolin
g,
w
hile
the
remaini
ng
thr
e
e
co
ns
ist
of
f
ully
c
onnecte
d
la
yer
s
.
It
i
s
al
so
me
ntio
ne
d
that
recti
fie
d
li
nea
r
un
it
(
Re
LU
)
act
ivati
on
is
a
pp
li
ed
to
al
l
la
yer
s
e
xcep
t
t
he
outp
ut
la
ye
r.
N
otably,
du
e
to
t
h
e
de
pth
of
t
he
Alex
N
et
50
arch
it
ect
ure,
t
he
auth
ors
intr
oduce
d
pa
dd
i
ng
to
mit
igate
th
e
sign
i
ficant
r
edu
ct
io
n
i
n
th
e
siz
e
of
the
f
eat
ure
maps
.
It
I
s
sta
te
d
that
the
in
pu
t
to
this
m
odel
c
omprises
images
siz
ed
a
t
227
×
227
×
3.
The
arc
hitec
tu
re
of
Alex
Net5
0
re
port
e
dly
inclu
de
s
co
nvol
utio
na
l
la
yer
s
al
on
g
with
ma
x
-
poolin
g
la
ye
rs.
If
yo
u
a
re
unf
amil
ia
r
with
cal
c
ulati
ng
the
out
pu
t
si
ze
of
a
c
onvol
ution
al
la
yer
of
the
Alex
Net5
0
arc
hitec
ture,
the
first
c
onvoluti
on
a
l
la
yer
ap
plies
96
filt
ers
siz
ed
at
11
×
11
with
a
stri
de
o
f
4.
The
act
ivati
on
f
unct
ion
e
mp
l
oy
e
d
in
this
la
yer
is
Re
LU.
Subse
quently
, th
e
r
e
sul
ti
ng
f
eat
ur
e
m
ap
is si
ze
d
at
55
×
55
×
96.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
Toma
t
o pla
nt di
sease pre
dicti
on system
wi
th
a new fra
mew
ork
SSMAN
…
(
Sa
r
av
anan M
adde
ri Siv
alinga
m
)
943
=
(
(
−
)
/
)
+
1
(1)
Figure
2. Ale
xNet
,
deep le
ar
ni
ng
proces
s m
odel
It
i
s
me
ntio
ne
d
that
in
t
he
s
ub
s
eq
ue
nt
ste
p,
the
num
ber
of
filt
ers
deter
mines
t
he
c
ha
nn
el
s
in
t
he
ou
t
pu
t
featur
e
map.
Fo
ll
ow
ing
this
,
the
f
irst
max
poolin
g
la
ye
r
is
introd
uced,
siz
ed
at
3
×
3
siz
e
of
2.
Con
se
quently
,
the
res
ulti
ng
feature
ma
p
is
ob
ta
ine
d
with
dim
e
ns
io
ns
of
27
×
27
×
96.
Fo
ll
owin
g
t
ha
t,
the
seco
nd
co
nvol
ution
op
e
rati
on
is
a
ppli
ed,
w
it
h
a
reducti
on
in
filt
er
siz
e
t
o
5
×
5,
c
onsist
ing
of
256
filt
e
rs.
The
stride
is
set
to
1,
wit
h
a
paddi
ng
of
2.
F
ollo
wing
t
he
ap
plica
ti
on
of
the
s
econd
c
onvolu
ti
on
operati
on,
Re
L
U
act
ivati
on
fun
ct
ion
is
on
ce
again
util
iz
ed.
C
onseq
ue
nt
ly,
the
outp
ut
siz
e
obta
ined
is
27
×
27
×
256.
Subseque
ntly,
a
max
-
po
oling
la
yer
of
siz
e
3
×
3
with
a
stri
de
of
2
is
a
pp
l
ie
d,
res
ulti
ng
i
n
a
featu
re
ma
p
with
dimensi
ons
of
13
×
13
×
25
6.
N
ext,
t
he
t
hird
c
on
voluti
on
ope
rati
on
is
c
ondu
ct
ed,
e
mp
l
oy
i
ng
384
filt
ers
siz
ed
at
3
×
3
with
a
str
ide
of
1
an
d
pa
dd
i
ng
of
1.
T
he
Re
L
U
act
iv
at
ion
f
unct
ion
is
app
li
ed
onc
e
more,
yieldi
ng
an
ou
t
pu
t
featur
e
map
with
dim
ensio
ns
of
13
×
13
×
384.
F
ollow
i
ng
t
he
thir
d
co
nvol
ution
op
e
rati
on,
the
fourt
h
conv
olu
ti
on
op
erati
on
is
e
xec
uted,
util
iz
ing
384
filt
ers
siz
e
d
at
3
×
3,
with
a
stri
de
of
1
a
nd
pa
ddin
g
al
s
o
set
to
1.
A
dd
it
io
na
ll
y,
the
act
ivati
on
functi
on
is
app
li
ed
su
c
h
a
s
Re
LU,
is
ap
pl
ie
d.
Subse
que
ntly,
the
outp
ut
size
remains
unal
te
red
a
t
13
×
13
×
384.
Fo
ll
owin
g
this,
the
fina
l
convo
l
ution
l
ayer,
co
mprisi
ng
256
filt
ers
siz
ed
a
t
3
×
3,
is
intr
oduc
ed.
B
oth
the
stride
an
d
paddin
g
are
set
t
o
on
e
,
an
d
th
e
Re
LU
act
ivati
on
f
un
ct
io
n
is
app
li
ed
.
Con
se
quently
, t
he
res
ulti
ng
fe
at
ur
e ma
p mai
ntains a s
ha
pe
of
13
×
13
×
256. Obser
ving the
arch
it
ect
ure th
us
f
a
r,
it
i
s
eviden
t
th
at
the
numb
e
r
of
filt
ers
incre
ases
as
we
del
ve
dee
per.
T
his
pro
gr
essi
ve
increase
facil
it
at
es
the
extracti
on
of m
or
e
intricat
e fe
at
ur
es a
s
we
a
dvance
th
rou
gh
arch
it
ect
ure
.
Fu
rt
hermo
re,
it
i
s
w
or
t
h
no
ti
ng
that
t
he
filt
er
siz
e
is
decr
ea
sing
pro
gressi
vely.
T
his
im
pl
ie
s
that
the
init
ia
l
filt
ers
w
ere
la
r
ge
r,
an
d
as
we
pr
ogres
s,
the
filt
er
siz
e
re
duces,
co
nse
quently
le
adi
ng
to
a
re
du
ct
i
on
in
the
s
ha
pe
of
th
e
feat
ur
e
ma
ps.
F
ollo
wing
t
his
obse
rvat
i
on,
the
thi
rd
ma
x
-
poolin
g
la
yer,
siz
ed
at
3
×
3
w
it
h
a
stride
of
2,
is
app
li
ed
.
T
his
op
e
rati
on
yield
s
a
featur
e
ma
p
with
dime
ns
ion
s
of
6
×
6
×
25
6.
S
ubseq
ue
ntly,
the
first
dr
opout
la
yer
is
intr
oduc
ed
with
a
rate
of
0.5,
co
ntin
ua
ti
on
of
t
his
f
ul
ly
co
nn
ect
e
d
l
ayer
is
implem
ented,
util
iz
ing
the
R
eLU
act
ivati
on
functi
on.
T
he
ou
t
pu
t
siz
e
is
s
et
at
40
96
ima
ges.
S
ubse
qu
e
ntly,
a
no
t
her
dro
pout
la
yer
is
a
p
plie
d
wit
h
a
rate
of
0.5
.
T
his
is
s
ucceede
d
by
a
seco
nd
f
ully
“
connecte
d
la
ye
r
”
co
mprisi
ng
4
,
096
neur
on
s
,
wit
h
Re
LU
act
ivat
ion
fu
c
ntio
n.
Con
cl
ud
i
ng
t
he
arch
it
ect
ure,
the
fi
nal
la
ye
r
co
ns
ist
s
of
1
,
000
neur
on
s
, corr
e
s
pondin
g
to the 1
,
000
cl
asses i
n
the
datas
et
. T
his last
f
ully c
onnected la
yer
util
iz
es the
So
f
tMax
act
ivati
on
fun
ct
ion
.
T
he
a
rc
hitec
ture
of
th
e
Alex
Net
m
odel
co
mprise
s
a
total
of
62
.3
mil
li
on
le
ar
nab
le
par
a
mete
rs
dis
tribu
te
d
acr
oss
8
la
yer
s
.
T
he
in
put
to
t
he
m
odel
co
nsi
sts
of
RGB
images
.
It
f
eat
ur
es
5
c
onvoluti
on
la
ye
rs
acc
om
pan
ie
d
by
a
c
ombinati
on
of
ma
x
-
poolin
g
la
yer
s.
T
his
a
r
chite
ct
ur
e
co
m
pr
ise
s
“
3
-
fu
ll
y
co
nne
ct
ed
la
ye
rs
”
,
each
util
iz
ing
the
Re
LU
ac
ti
vation
f
unct
ion.
A
ddit
ion
al
ly,
it
integ
rates
two
Dropo
ut
la
yer
s
.
T
he
outp
ut
la
yer
util
iz
es
the
S
of
t
M
ax
act
iv
at
ion
f
un
ct
io
n.
The
total
par
a
mete
r
c
ounts
f
or
thi
s
config
ur
at
io
n a
moun
ts
to 6
2.3
mil
li
on
.
Alex
Net
sta
nds
as
a
gro
undbrea
king
CN
N
re
now
ne
d
f
or
it
s
pro
wess
in
im
age
re
co
gnit
ion
a
nd
cl
assifi
cat
ion
.
Its
tri
umph
in
the
2012
Ima
ge
Net
la
r
ge
sca
le
vi
s
ual
recog
niti
on
c
halle
nge
ma
rk
e
d
a
pi
vo
ta
l
mo
me
nt
in
th
e
ad
van
ce
men
t
of
dee
p
le
ar
ning
[
11]
.
T
he
arch
it
ect
ure
of
Ale
xN
et
,
char
act
e
rized
by
it
s
pioneeri
ng
im
plementat
io
n
of
c
onvoluti
ona
l
la
yer
s
an
d
Re
LU,
est
ablish
ed
a
co
r
ner
st
one
for
c
onte
m
porary
deep
le
a
rn
i
ng
models.
Its
im
pact
has
bee
n
instrumenta
l
in
pro
pelli
ng
a
dvance
men
ts
i
n
com
pu
te
r
visi
on
a
nd
patte
rn
rec
ogni
ti
on
a
pp
li
cat
ion
s
.
The
a
dopt
ion
of
Re
L
U
as
a
n
act
ivati
on
f
unct
io
n
nota
bly
e
xp
e
dit
ed
t
he
trai
ning
proce
ss,
en
ha
ncin
g
it
s
sp
eed
by
ne
arly
six
f
old
[
12]
.
Additi
ona
ll
y,
integ
rati
on
of
dro
pout
la
yer
s
eff
ect
ivel
y
mit
igate
d
over
fitt
i
ng
c
on
ce
rns,
bolst
erin
g
t
he
model'
s
gen
e
ra
li
zat
ion
capa
bi
li
ti
es.
M
oreo
ve
r,
t
he
model
underg
oe
s
trai
ni
ng
us
i
ng
the
Ima
ge
N
et
dataset
,
w
hich
c
omprises
ne
arly
14
mil
li
on
ima
ges
distri
bu
te
d
acro
s
s
on
e
th
ousa
nd
cl
asses.
The
le
af
ima
ge
dataset
wa
s
util
iz
ed
t
o
fine
-
t
un
e
a
pre
-
trai
ned
CN
N
model
or
i
gin
al
ly
trai
ned
on
Ima
ge
Net,
em
ployin
g
s
hallow
tu
ni
ng
[13
]
.
A
ddit
ion
al
ly,
we
can
denote
th
e
cl
ass
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
940
-
948
944
cat
egory
L
as
a
vect
or
of
le
ng
t
h
n,
w
he
re
eac
h
el
eme
nt
c
orr
esp
onds
to
t
he
cl
ass
la
bel
of
t
he
res
pecti
ve
i
mage.
Her
e
, k r
e
prese
nts the
num
ber o
f
cl
asses
[
14]
.
To
br
ea
k
dow
n
the
dataset
,
this
stu
dy
em
pl
oy
e
d
k
-
mea
ns
cl
us
te
rin
g.
Eac
h
patte
rn
withi
n
the
ne
w
cl
ass
cat
egor
y
L
receive
d
a
c
la
ss
la
bel
ba
se
d
on
the
near
e
st
centr
oid
,
det
ermine
d
by
th
e
square
d
E
ucl
idean
distance
(S
E
D)
[1
5]
.
Ce
ntro
i
ds
are
la
be
le
d
as
cj
.
Be
cause
of
the
rest
rict
ed
avail
abili
ty
of
t
rainin
g
dataset
s,
the
stochastic
gr
a
dient
desce
nt
(S
G
D
)
met
hod
may
disp
la
y
su
bst
antia
l
flu
ct
uations
in
t
he
le
arn
i
ng
f
un
ct
ion
.
Conv
olu
ti
onal
neural
netw
ork
s
are
e
mp
l
oy
e
d
i
n
t
his
stu
dy
f
or
the
cat
e
gorizat
ion
of
dif
f
eren
t
pla
nt
dis
eases
[16]
.
At
the
onset
,
t
he
c
onvoluti
onal
la
ye
r
of
t
he
m
odel
extracts
in
for
mati
on
f
rom
t
he
i
nput
ima
ge
s.
T
he
conv
olu
ti
on
la
yer
ap
plies
m
ul
ti
ple
filt
ers
to
an
i
mage
,
ge
ner
at
in
g
a
feat
ur
e
ma
p
to
ext
ract
uniq
ue
fe
at
ur
es.
Af
te
r
the
c
onvoluti
on
la
yer
s
,
po
oling
la
ye
r
s
are
us
ed
.
Th
ese
la
ye
rs
re
duce
t
he
num
be
r
of
par
amet
e
r
s
an
d
dow
ns
cal
e
the
image
.
C
om
m
on
l
y
us
e
d
pool
ing
te
ch
niques
inclu
de
max
poolin
g,
ave
ra
ge
poolin
g,
an
d
s
um
poolin
g
[17]
–
[
19]
.
(
)
=
[
0
,
]
(2)
These m
odel
s ex
hi
bit
high p
r
edict
ive
pe
rformance due
to
t
heir
e
xtensi
ve
trai
ning o
n
vis
ua
l
data.
I
n
this study,
the YOL
O
a
nd
R
-
CN
N mo
del
s ad
a
pted
w
it
h
“
trans
fer
lea
rn
i
ng
”
trai
ni
ng techn
i
qu
e
[
20]
.
2.1
.
A
bout
d
atase
t
This
researc
h
study
us
e
d
2,403
t
om
at
o
plant
ph
oto
s
an
d
the
se
ima
ge
s
we
re
cl
assifi
ed
int
o
fou
r
cat
egories:
‘
le
af
blig
ht’,
‘
black
r
ot’,
‘esca
(
black
mea
sle
s)
’,
a
nd
‘h
eal
th
y
(isario
ps
is
le
a
f
spot)
’
.
T
he
da
ta
set
consi
sts
of
42
3
images
s
howi
ng
healt
hy
le
a
ves,
1
,
180
ima
ges
of
le
a
ves
i
nf
ect
e
d
with
bl
ack
r
ot,
1,383
i
mages
of
le
aves
af
fec
te
d
by
esca,
an
d
1,076
i
mage
s
of
le
a
ves
a
ffec
te
d
by
le
af
bl
igh
t
[
21]
–
[
23]
.
T
he
data
em
pl
oy
e
d
in
this
s
tudy
is
so
urce
d
f
rom
r
ur
al
far
m
fi
el
ds
an
d
is
acce
ssible
at
“
https:/
/ww
w.k
agg
le
.c
om
/
data
set
s/e
mmare
x/p
la
ntd
ise
ase
”
,
known
as
the
Plant
Vill
age
c
ollec
ti
on
,
this
dataset
com
pr
ise
s
over
55,000 ima
ges
cate
gorized
in
to 38
cl
asse
s,
c
ov
e
rin
g 1
4 dist
inct plant
sp
eci
es.
Figure
s
3(
a
)
a
nd
3
(b)
il
lustr
at
es
sample
da
ta
from
var
i
ous
cl
asses
[
24]
,
in
this
Fi
gure
3(a)
has
“
Samp
le
i
mag
es
of
healt
hy
tomato
”
,
“
gr
a
pe
”
,
an
d
“
straw
berry
le
aves
”
a
nd
Fig
ure
3(b)
has
sam
ple
im
ages
of
diseased
le
a
ve
s.
Dee
p
le
a
rn
i
ng
m
odel
s
re
quire
e
xte
ns
ive
dataset
s
to
f
unct
ion
ef
fecti
vely
.
Be
fore
impleme
nting
enh
a
ncem
ents
li
ke
ro
ta
ti
on,
z
oomin
g,
a
nd
s
hiftin
g,
the
im
ages
ar
e
resize
d
to
224
×
224
pix
el
s
us
in
g
Ker
as'
I
mage
data
ge
ne
rator
[
25]
.
T
he
augme
ntati
on
te
ch
nique
se
r
ves
tw
o
pr
ima
r
y
pu
rposes:
e
xpan
ding
the
dataset
a
nd
addressi
ng
ov
erf
it
ti
ng
durin
g
trai
ning.
T
he
dataset
is
div
i
de
d
int
o
trai
ni
ng
an
d
validat
io
n
set
s
in
a
n
80:2
0
rat
io
for
e
xperim
entat
ion
.
S
pe
ci
fical
ly,
t
he
vali
dation
set
co
nt
ai
ns
481
ima
ge
s,
w
hile
the
tr
ai
nin
g
set
co
m
pr
ise
s
1,922 ima
ges
.
(a)
(b)
Figure
3
.
S
am
pl
e d
at
a f
rom
va
rio
us
classe
s
(
a
)
“
sam
ple ima
ge
s of
healt
hy to
mato
”
,
“
grape
”
, and
“
straw
be
r
ry
le
aves
”
a
nd (
b)
sam
ple image
s of
diseased
leaves
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
Toma
t
o pla
nt di
sease pre
dicti
on system
wi
th
a new fra
mew
ork
SSMAN
…
(
Sa
r
av
anan M
adde
ri Siv
alinga
m
)
945
T
a
b
l
e
1.
T
o
m
a
t
o
p
l
a
n
t
d
i
s
e
a
s
e
d
a
t
a
s
e
t
s
u
mm
e
r
y
Categ
o
ry
Clas
s
Tr
ain
in
g
s
am
p
les
Testin
g
s
am
p
les
Tomato (
Health
y
)
1
646
162
Tomato (
Late
Blig
h
t)
2
631
158
Tomato (
Yello
w_
Leaf
_
Cu
rl_Viru
s)
3
645
161
2.2
.
Pr
oposed
pre
-
tra
ine
d SSM
AN
fr
ame
wo
r
k for
leaf
disease cl
as
sifi
cat
i
on
The
Alex
Net
arch
it
ect
ure,
des
pite
bein
g
a
strai
gh
t
f
orward
a
nd
eff
ic
ie
nt
m
od
el
for
sce
ne
cl
assifi
cat
ion
,
has
li
mit
at
ions
.
It
str
uggles
with
t
he
“
non
-
tran
sp
a
ren
c
y
of
inter
mediat
e
la
ye
rs
”
an
d
faces
chall
enges
i
n
ha
nd
li
ng
m
ulti
-
scal
e
themat
ic
s
cenes
i
n
cl
assi
f
ic
at
ion
ta
s
ks
[26]
.
T
his p
r
oces
s
entai
ls
e
mp
l
oyin
g
the
pr
e
-
trai
ne
d
Ale
xN
et
a
rc
hi
te
ct
ur
e,
wh
ic
h
ori
gin
at
es
f
rom
trai
ni
ng
AlexN
et
on
e
xte
ns
ive
nat
ur
al
i
mage
dataset
s
[27
]
.
Be
nef
it
ing
fro
m
a
nalo
gous
s
emanti
c
sce
ne
inf
or
mati
on
,
t
he
use
s
Ale
xN
et
arc
hi
te
ct
ur
e
a
chieves
swift
a
nd
sat
isf
act
ory
resu
lt
s
f
or
sce
ne
ima
ge
s
[
28]
.
T
o
acc
ommo
date
t
he
mu
lt
i
-
scal
e
c
omplexit
y
of
pa
rtic
ula
r
semanti
c
sce
ne
s,
the
Alex
Net
arc
hitec
tu
re
inc
orp
or
at
e
s
the
se
manti
c
scene
mod
ule
(SS
M
)
,
a
pote
nt
mu
lt
i
-
scal
e
po
oling
operati
on.
Im
pro
ving
th
e
represe
ntati
on
of
i
nterme
di
at
e
la
yer
in
f
ormat
ion
within
the
pre
-
trai
ned
Alex
N
et
arch
it
ect
ur
e
no
t
on
l
y
bo
ost
s
cl
assifi
cat
ion
pe
rforma
nce
by
st
rengthe
ni
ng
the
rob
us
t
ness
of
netw
ork
weig
ht
s
but
al
so
en
h
ances
tr
ans
pa
r
ency
th
rou
ghout
inter
mediat
e
la
yer
s
.
Util
iz
i
ng
the
be
nef
it
s
of
the
SSM
strat
e
gy
and
with
t
he
obje
ct
ive
of
e
nhancin
g
re
gion
of
intere
st
cl
as
sific
at
ion
pe
rformance
th
r
ough
th
e
Alex
Net
arc
hitec
ture,
the
f
ollow
i
ng frame
w
ork
S
S
M
A
N m
od
el
pro
posed
a
nd prese
nted
i
n
Fi
gure
4.
Figure
4. Pro
pose
d
SS
MA
N
model arc
hitec
ture
This
m
od
el
i
ntegr
at
es
the
SS
M
strat
e
gy
i
nto
the
deep
le
a
r
ning
Ale
xN
et
arch
it
ect
ure
as
a
means
t
o
achieve
imp
roved
pe
rfo
rma
nc
e.
It
is
ex
pla
ined
t
hat
the
su
pe
r
visio
n
la
yer
s
are
integ
r
at
ed
as
i
nterm
ediat
e
la
yer
s
within
deep
le
ar
ning
Alex
Net
arc
hitec
ture
[29
]
.
F
ur
t
her
m
ore,
t
he
y
sta
te
d
t
hat
the
SS
M
m
odel
is
com
bin
e
d
with
the
Alex
Net
a
rch
it
ect
ure,
gra
nting
t
he
pre
-
t
raine
d
Ale
xN
e
t
arch
it
ect
ur
e
t
he
abili
ty
to
ha
nd
le
mu
lt
i
-
scal
e
in
f
ormat
ion
ef
fec
ti
vely.
I
n
a
co
mprehe
ns
ive
a
ssessme
nt,
it
was
no
te
d
that
the
SS
MAN
model,
dep
ic
te
d i
n Fi
gure
4, e
nab
le
s i
mpro
ve
d
sce
ne
classi
ficat
ion
performa
nce e
ven w
it
h l
imi
te
d
sa
mp
le
ima
ge
ry.
4.
RESU
LT
S
AND DI
SCUS
S
ION
Tom
at
o
plant
l
eaves
dataset
i
s
util
iz
ed
t
o
tr
ai
n
the
pro
pos
ed
S
S
M
A
N
fr
a
mew
ork,
an
d
i
ts
res
ults
are
con
t
rasted
with
th
os
e
of
pro
minent
pr
e
-
trai
ned
m
od
el
s
li
ke
YO
L
O
a
nd
R
-
CNN
[
30]
.
The
em
ploye
d
Alex
Net
model
co
ns
ist
s
of
pair
s
of
c
onvo
l
ution
an
d
poolin
g
la
yer
s
. Co
nv
olu
ti
on
la
yer
s
are r
es
ponsi
ble
f
or
f
il
te
ring
an
d
extracti
ng
feat
ur
es
f
rom
the
i
nput
ima
ge,
w
hile
poolin
g
la
yer
s
re
du
ce
the
netw
ork's
c
ompu
ta
ti
on
a
nd
de
cre
ase
the
in
pu
t
i
ma
ge
siz
e.
Th
e
Alex
Net
m
od
e
l
is
trai
ne
d
over
10
ep
oc
hs
,
util
iz
ing
t
he
Ad
a
m
op
ti
miz
er
a
nd
cat
egorical
cross
-
ent
rop
y
as
t
he
ch
os
e
n
lo
ss
functi
on.
Se
ns
i
ti
vity
ind
ic
at
es
the
accu
racy
with
w
hich
the
model
identifie
s
plan
ts
with
a
dise
ase,
w
hile
s
pe
ci
fici
ty
ref
le
ct
s
it
s
abili
ty
to
pr
e
dict
plants
with
ou
t
t
he
di
sease
accuratel
y.
T
h
e
s
e
r
e
s
u
l
t
s
s
h
o
w
t
h
a
t
a
m
u
l
t
i
-
c
l
a
s
s
i
f
i
c
a
t
i
o
n
o
c
c
u
r
s
w
h
i
l
e
u
s
i
n
g
t
o
m
a
t
o
l
e
a
v
e
d
a
t
a
s
e
t
.
T
o
e
v
a
l
u
a
t
e
t
h
e
i
r
m
o
d
e
l
,
t
h
e
y
u
t
i
l
i
z
e
d
a
m
u
l
t
i
-
c
l
a
s
s
c
o
n
f
u
s
i
o
n
m
a
t
r
i
x
.
B
e
f
o
r
e
e
r
r
o
r
c
o
r
r
e
c
t
i
o
n
,
e
a
c
h
i
n
p
u
t
i
m
a
g
e
c
o
u
l
d
b
e
c
l
a
s
s
i
f
i
e
d
i
n
t
o
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
940
-
948
946
o
n
e
o
f
t
h
e
‘
n
o
n
-
o
v
e
r
l
a
p
p
i
n
g
c
l
a
s
s
e
s
’
.
T
h
e
r
e
f
o
r
e
,
t
h
e
‘
c
o
n
f
u
s
i
o
n
m
a
t
r
i
x
’
f
o
r
a
s
p
e
c
i
f
i
c
c
l
a
s
s
i
i
s
a
s
s
h
o
w
n
i
n
T
a
b
l
e
2
.
T
h
e
p
e
r
f
o
r
m
a
n
c
e
p
ar
a
m
e
t
e
r
s
f
o
r
m
u
l
ti
c
l
a
s
s
c
l
a
s
s
i
f
i
c
a
t
ion
a
r
e
g
iv
en
i
n
T
ab
l
e
3.
T
a
b
l
e
2.
T
o
m
a
t
o
p
l
a
n
t
d
i
s
e
a
se
c
l
a
s
s
i
f
i
c
a
ti
o
n
a
n
d
r
e
l
a
t
e
d
c
o
n
f
u
s
i
o
n
m
a
t
r
i
x
r
e
s
u
l
t
s
S.
N
o
.
P
er
f
o
r
m
a
n
c
e
m
e
t
r
i
c
s
F
o
r
m
u
l
a
1
A
c
c
u
ra
cy
(
T
r
u
e
P
o
s
i
t
i
v
e
+
T
r
u
e
N
e
g
a
ti
v
e
)
/
(
T
r
u
e
P
o
s
i
t
i
v
e
+
T
r
u
e
N
e
g
a
t
i
v
e
+
F
a
l
s
e
P
o
s
i
t
i
v
e
+
F
a
l
s
e
N
e
g
a
t
iv
e
)
2
S
e
n
s
i
t
i
v
i
t
y
(
T
r
u
e
P
o
s
i
t
i
v
e
)
/
(
T
r
u
e
P
o
s
i
t
i
v
e
+
F
a
l
s
e
N
e
g
a
t
i
v
e
)
3
S
p
e
c
i
f
i
c
i
t
y
(
T
r
u
e
N
e
g
a
ti
v
e
)
/
(
T
r
u
e
N
e
g
a
t
i
v
e
+
F
a
l
s
e
P
o
s
i
t
i
v
e
)
4
F
1
s
c
o
r
e
2
*
[(
S
e
n
s
i
t
i
v
i
t
y
*
S
p
e
c
i
f
i
c
i
t
y)
/
(
S
e
n
s
i
ti
v
i
t
y
+
S
p
ec
i
f
i
c
i
t
y
)]
Tr
u
e Pos
itiv
e (
TP)
: Co
rr
ectly
cl
ass
ifyin
g
a
p
lan
t as dis
eased
when
it
actu
all
y
has
a
d
iseas
e.
Tr
u
e Negativ
e
(
T
N
):
Co
rr
e
ctly
classif
y
in
g
a
p
lan
t as health
y
when
it
is in
d
eed h
ealth
y
.
False Po
sitiv
e (
FP
)
: I
n
co
rr
ect
ly
classi
fyin
g
a
h
ealth
y
plan
t as dis
eased
.
False Neg
ativ
e
(F
N):
Inco
r
rectly cl
a
ss
ifyin
g
a
d
iseas
ed
plan
t as health
y
.
T
a
b
l
e
3.
P
l
a
n
t
d
i
s
e
a
s
e
c
l
a
s
s
i
f
i
c
a
tio
n
p
e
r
f
o
r
m
a
n
c
e
p
ar
a
m
e
t
e
r
s
f
o
r
m
u
l
ti
c
l
a
s
s
c
l
a
s
s
i
f
i
c
a
t
ion
Sl.
No
Perf
o
r
m
an
ce
m
et
ri
cs
1.
TP
i
=
∑
=
1
2.
TN
i
=
∑
∑
,
=
=
1
=
1
3.
TP
i
=
∑
;
≠
1
=
4.
TN
i
=
∑
,
≠
1
=
1
F
u
r
t
h
e
r
m
o
r
e
,
F
i
g
u
r
e
4
p
r
o
v
i
d
e
s
d
e
t
a
i
l
e
d
i
n
f
o
r
m
a
t
i
o
n
o
n
t
h
e
t
r
a
i
n
i
n
g
a
n
d
t
e
s
t
i
n
g
a
c
c
u
r
a
c
y
,
a
s
w
e
l
l
a
s
t
h
e
l
o
s
s
o
b
t
a
i
n
e
d
b
y
t
h
e
A
l
e
x
N
e
t
a
l
g
o
r
i
t
h
m
.
T
o
i
l
l
u
s
t
r
a
t
e
t
h
e
r
o
b
u
s
t
n
e
s
s
o
f
S
S
M
A
N
,
t
h
e
r
e
s
e
a
r
c
h
e
r
s
u
t
i
l
i
z
e
d
a
v
a
r
i
e
t
y
o
f
I
m
a
g
e
N
e
t
w
i
t
h
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
b
a
s
e
d
C
N
N
m
o
d
e
l
s
,
s
u
c
h
a
s
Y
O
L
O
a
n
d
R
-
C
N
N
,
d
u
r
i
n
g
t
h
e
t
r
a
n
s
f
e
r
l
e
a
r
n
i
n
g
p
h
a
s
e
i
n
A
l
e
x
N
e
t
.
T
h
e
o
u
t
c
o
m
e
s
a
r
e
p
r
e
s
e
n
t
e
d
i
n
T
a
b
l
e
4
.
F
i
g
u
r
e
5
h
a
s
t
h
e
g
r
a
p
h
i
c
a
l
r
e
p
r
e
s
e
n
t
a
t
i
o
n
f
o
r
p
e
r
f
o
r
m
a
n
c
e
c
o
m
p
a
r
a
t
i
v
e
s
t
u
d
y
b
e
t
w
e
e
n
p
r
o
p
o
s
e
d
S
S
M
A
N
a
n
d
e
x
i
s
t
i
n
g
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
a
l
g
o
r
i
t
h
m
s
.
I
n
t
h
i
s
w
e
c
a
n
s
e
e
t
h
a
t
S
S
M
A
N
p
e
r
f
o
r
m
s
b
e
t
t
e
r
i
n
a
c
c
u
r
a
c
y
,
s
e
n
s
i
t
i
v
i
t
y
,
s
p
e
c
i
f
i
c
i
t
y
a
n
d
F
1
-
s
c
o
r
e
p
a
r
a
m
e
t
e
r
s
t
h
a
n
o
t
h
e
r
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
a
l
g
o
r
i
t
h
m
s
.
A
c
c
o
r
d
i
n
g
t
o
F
i
g
u
r
e
6
,
t
h
e
S
S
M
A
N
f
r
a
m
e
w
o
r
k
,
w
h
i
c
h
i
n
c
o
r
p
o
r
a
t
e
s
t
h
e
A
l
e
x
N
e
t
a
l
g
o
r
i
t
h
m
,
a
c
h
i
e
v
e
d
t
h
e
h
i
g
h
a
c
c
u
r
a
c
y
o
f
9
8
.
3
0
%
,
w
i
t
h
a
s
e
n
s
i
t
i
v
i
t
y
o
f
9
7
.
5
6
%
a
n
d
a
s
p
e
c
i
f
i
c
i
t
y
o
f
9
6
.
7
1
%
.
T
a
b
l
e
4.
Co
mpa
rison
betwee
n
p
rop
os
ed
SSMAN a
nd e
xis
ti
ng
YO
L
O
an
d
R
-
C
NN m
od
el
s and r
el
at
e
d resu
lt
s
Deep learnin
g
m
o
d
el
Lear
n
in
g
a
ccura
cy
Lear
n
in
g
s
en
sitiv
ity
Lear
n
in
g
s
p
ecificit
y
Lear
n
in
g
F1
-
s
co
re
YOLO
9
7
.33
9
6
.33
9
5
.33
9
6
.07
R
-
CNN
9
6
.22
9
6
.11
9
5
.01
9
5
.55
SSMAN
9
8
.30
9
7
.56
9
6
.71
9
7
.13
F
i
g
u
r
e
5
.
Re
su
l
ts
o
f
a
cc
u
ra
c
y
a
n
d
l
o
ss
f
o
r
t
r
a
i
n
i
ng
a
nd
t
e
s
t
i
ng
d
a
t
a
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
Toma
t
o pla
nt di
sease pre
dicti
on system
wi
th
a new fra
mew
ork
SSMAN
…
(
Sa
r
av
anan M
adde
ri Siv
alinga
m
)
947
F
i
g
u
r
e
6
.
C
o
m
p
ar
i
s
o
n
of
p
e
r
f
o
r
m
a
n
c
e
o
f
“
p
r
o
p
o
s
e
d
S
S
M
A
N
M
od
e
l
”
w
i
t
h
e
x
i
s
ti
ng
(
YO
L
O
a
n
d
R
-
C
N
N
m
o
d
e
l)
5.
CONCL
US
I
O
N
The
st
udy
int
rod
uces
SSM
AN,
a
deep
l
earn
i
ng
a
rc
hitec
ture
that
ut
il
iz
es
a
la
yer
-
base
d
cl
ass
decomp
os
it
io
n
te
ch
nique
t
o
cl
assify
the
for
t
om
at
o
le
ave
dataset
.
T
o
validat
e
t
he
SSM
AN
m
odel
,
we
employe
d
var
i
ou
s
pr
e
-
trai
ne
d
CN
N
m
odel
s
su
c
h
as
Y
OL
O
an
d
R
-
CN
N.
SSMA
N
util
iz
i
ng
Alex
Net
at
ta
ine
d
the
hi
gh
est
ac
cur
ac
y
of
98.
30%.
Durin
g
te
sti
ng
,
perfor
mance
est
imat
ion
metri
cs
in
cl
ud
in
g
“
acc
ur
acy
”
,
“
pr
eci
sio
n
”
,
“
r
ecal
l
”
,
an
d
“
F1
-
sc
ore
”
wer
e
empl
oyed
f
or
the
model
e
va
luati
on
.
T
his
disco
very
facil
it
at
es
early
detect
io
n
of
pla
nt
dise
ases,
t
her
e
by
reducin
g
cr
op
loss
a
nd
t
he
sp
rea
d
of
dise
ases.
Ulti
mate
ly
,
t
he
model'
s
dep
l
oyment
on po
rtab
le
d
evices
aims
to
e
nh
a
nce
ef
f
ic
ie
ncy
further
.
REFERE
NCE
S
[1]
J.
Liu
an
d
X.
W
a
n
g
,
“
Tomato
d
isea
se
o
b
ject
d
etectio
n
m
eth
o
d
co
m
b
in
in
g
p
rior
k
n
o
wled
g
e
atten
tio
n
m
echan
ism
an
d
m
u
ltiscale
features,
”
Fro
n
tier
s in
P
la
n
t Scien
ce
,
v
o
l.
1
4
,
2
0
2
3
,
d
o
i:
1
0
.33
8
9
/fpls
.20
2
3
.12
5
5
1
1
9
.
[2]
A.
Ka
m
ran
et
a
l.
,
“
Ins
ig
h
ts
in
to
th
e
g
en
etic
v
ariabilit
y
an
d
ev
o
lu
tio
n
ary
d
y
n
am
ics
o
f
to
m
ato
sp
o
tted
wilt
o
rtho
to
sp
o
v
irus
in
Ch
in
a,
”
BMC G
en
o
mics
,
v
o
l.
2
5
,
n
o
.
1
,
2
0
2
4
,
d
o
i: 10
.1
1
8
6
/s1
2
8
6
4
-
0
2
3
-
0
9
9
5
1
-
9.
[3]
M.
Jeg
u
iri
m
an
d
A.
Zorp
as,
“
To
m
ato
p
rocess
in
g
b
y
-
p
rod
u
cts:
su
sta
in
ab
le
ap
p
licatio
n
s,
”
To
ma
to
Pr
o
ce
ss
in
g
b
y
-
Pro
d
u
cts
:
S
u
sta
in
a
b
le A
p
p
lica
tio
n
s
,
p
p
.
1
–
4
2
4
,
2
0
2
2
,
d
o
i: 1
0
.10
1
6
/C2
0
1
9
-
0
-
0
4
7
0
1
-
2.
[4]
T.
B.
Torr
es
et
a
l.
,
“
Cau
sal
ag
en
ts
o
f
St
em
p
h
y
liu
m
-
in
d
u
ced
foliar
d
isea
ses
o
f
to
m
ato
es
a
n
d
o
th
er
So
lan
ace
ae
h
o
sts
in
B
razil
,
”
Jo
u
rn
a
l of App
lied
Micr
o
b
io
lo
g
y
,
v
o
l
.
1
3
5
,
n
o
.
3
,
2
0
2
4
,
d
o
i: 10
.1093
/jam
b
io
/lx
ae03
8
.
[5]
M.
L
i
et
a
l.
,
“
Hig
h
-
p
erf
o
rman
ce
p
la
n
t
p
est
an
d
d
iseas
e
d
etectio
n
b
ased
o
n
m
o
d
el
en
semble
with
in
cept
io
n
m
o
d
u
le
an
d
clu
ster
alg
o
rithm,
”
Plan
ts
,
v
o
l.
1
2
,
n
o
.
1
,
2
0
2
3
,
d
o
i: 10
.33
9
0
/p
la
n
ts1
2
0
1
0
2
0
0
.
[6]
V.
Bin
n
ar
an
d
S.
S
h
arm
a,
“
Plan
t
leaf
d
iseas
es
d
etectio
n
u
sin
g
d
eep
learnin
g
alg
o
rithms,
”
Lec
tu
re
No
tes
in
Elec
t
rica
l
Eng
in
eerin
g
,
v
o
l.
9
4
6
,
p
p
.
2
1
7
–
2
2
8
,
2
0
2
3
,
d
o
i: 10
.
1
0
0
7
/
9
7
8
-
9
8
1
-
19
-
5868
-
7
_
1
7
.
[7]
D.
Sh
ah
,
V
.
Tr
iv
e
d
i,
V.
Sh
eth
,
A
.
Sh
ah
,
an
d
U
.
Ch
a
u
h
an
,
“
Res
TS:
res
id
u
al
d
eep
in
terpr
etab
le
a
rchitectu
re
for
p
lan
t
d
iseas
e
d
etectio
n
,
”
Info
rma
tio
n
P
ro
cess
in
g
i
n
A
g
ricu
ltu
re
,
v
o
l.
9
,
n
o
.
2
,
p
p
.
2
1
2
–
2
2
3
,
2
0
2
2
,
d
o
i: 1
0
.1
0
1
6
/j.inp
a.2021
.06.
0
0
1
.
[8]
J.
A.
Pan
d
ian
,
V
.
D.
Ku
m
ar,
O.
G
em
an
,
M.
Hn
atiu
c
,
M.
Ari
f
,
an
d
K.
Kan
ch
an
ad
ev
i,
“
Plan
t
d
iseas
e
d
etectio
n
u
sin
g
d
eep
co
n
v
o
lu
ti
o
n
al neu
r
al netwo
rk,
”
App
lied
Scien
ces (Sw
itzer
la
n
d
)
,
v
o
l.
1
2
,
n
o
.
1
4
,
2
0
2
2
,
d
o
i: 1
0
.33
9
0
/ap
p
1
2
1
4
6
9
8
2
.
[9]
S.
V
allab
h
ajo
sy
u
la
,
V
.
Sistl
a,
an
d
V.
K.
K.
Ko
lli
,
“
Tr
a
n
sfer
le
arnin
g
-
b
ased
d
eep
en
semble
n
eu
ral
n
etwo
rk
for
p
lan
t
lea
f
d
iseas
e
d
etectio
n
,
”
Jo
u
rn
a
l of Plan
t Disea
ses
an
d
P
ro
tectio
n
,
v
o
l.
1
2
9
,
n
o
.
3
,
p
p
.
5
4
5
–
5
5
8
,
2
0
2
2
,
d
o
i: 10
.10
0
7
/s
4
1
3
4
8
-
021
-
0
0
4
6
5
-
8.
[10
]
A.
Kh
ak
i
m
o
v
,
I.
Salak
h
u
td
in
o
v
,
A.
Om
o
lik
o
v
,
an
d
S
.
Utag
an
o
v
,
“
T
rad
itio
n
al
an
d
cu
r
rent
-
p
ros
p
ectiv
e
m
eth
o
d
s
o
f
ag
ricultu
ra
l
p
lan
t
d
iseas
es
d
e
tectio
n
:
a
review,
”
IOP
Co
n
feren
ce
S
eries
:
Ear
th
a
n
d
Envir
o
n
men
ta
l
S
cien
ce
,
v
o
l.
9
5
1
,
n
o
.
1
,
2
0
2
2
,
d
o
i: 10
.1088
/1
7
5
5
-
1
3
1
5
/
9
5
1
/
1
/0
1
2
0
0
2
.
[11
]
P.
Vasav
i,
A
.
Pu
n
ith
a,
an
d
T.
V.
N.
Rao
,
“
Cro
p
leaf
d
iseas
e
d
etectio
n
an
d
class
ification
u
sin
g
m
achi
n
e
learnin
g
an
d
d
eep
learnin
g
alg
o
rithms
b
y
v
isu
al
sy
m
p
to
m
s:
a
r
ev
iew,
”
Inter
n
a
tio
n
a
l
Jo
u
rn
a
l
o
f
El
ectri
ca
l
a
n
d
Co
mp
u
ter
Eng
in
eerin
g
(I
JEC
E)
,
v
o
l.
1
2
,
n
o
.
2
,
p
p
.
2
0
7
9
–
2
0
8
6
,
2
0
2
2
,
d
o
i: 10
.1159
1
/i
jece.v1
2
i2
.pp
2
0
7
9
-
2086.
[12
]
J.
An
d
rew,
J.
Eun
ice,
D.
E.
Po
p
esc
u
,
M.
K.
Ch
o
wd
ar
y
,
an
d
J.
Hem
an
th
,
“
Deep
learnin
g
-
b
ased
leaf
d
iseas
e
d
etectio
n
in
crop
s
u
sin
g
im
ag
es for a
g
ricultu
ral
ap
p
licat
io
n
s,
”
Agr
o
n
o
my
,
v
o
l.
1
2
,
n
o
.
1
0
,
2
0
2
2
,
d
o
i: 10
.3390
/a
g
ron
o
m
y
1
2
1
0
2
3
9
5
.
[13
]
A.
Pan
d
ian
J,
K.
Kan
ch
an
ad
ev
i,
N.
R.
Rajalak
sh
m
i,
an
d
G.ar
u
lk
u
m
ar
an
,
“
An
im
p
rov
ed
d
eep
resid
u
al
co
n
v
o
lu
ti
o
n
al
n
eu
ral
n
etwo
rk
for
p
lan
t
leaf
d
iseas
e
d
etectio
n
,
”
Co
mp
u
ta
tio
n
a
l
Int
ellig
en
ce
a
n
d
Neu
ro
scien
ce
,
v
o
l.
2
0
2
2
,
2
0
2
2
,
d
o
i: 10
.1155
/2
0
2
2
/
5
1
0
2
2
9
0
.
[14
]
P.
D.
Kalw
ad
,
S.
G.
Kan
ak
ar
ad
d
i,
T.
P
reeti,
S.
Ichal
ak
aranji,
S.
Sali
m
ath
,
an
d
S.
N
ay
ak
,
“
Ap
p
le
lea
f
d
ise
ase
d
etectio
n
an
d
an
aly
sis
u
sin
g
d
ee
p
learnin
g
tech
n
iq
u
e,
”
Lectu
re
No
tes
in
Netw
o
rks
a
n
d
S
ystems
,
v
o
l.
1
9
1
,
p
p
.
8
0
3
–
8
1
4
,
2
0
2
2
,
d
o
i:
1
0
.10
0
7
/
9
7
8
-
981
-
16
-
0
7
3
9
-
4
_
7
6
.
[15
]
R.
An
ith
a
an
d
A.
Bazila
Ban
u
,
“
L
eaf
d
iseas
e
d
etecti
o
n
u
sin
g
d
eep
lea
rnin
g
,
”
Jo
u
rn
a
l
o
f
Artificia
l
Intellig
en
ce
a
n
d
Ca
p
su
le
Netw
o
rks
,
v
o
l.
4
,
n
o
.
2
,
p
p
.
9
9
–
1
1
0
,
2
0
2
2
,
d
o
i: 10
.3
6
5
4
8
/jaicn
.20
2
2
.2.0
0
2
.
[16
]
P.
Jo
h
ri,
N.
Parash
ar,
U.
Sax
en
a,
an
d
K.
Pu
sh
p
an
jali,
“
Plan
t
d
iseas
e
d
etec
tio
n
u
sin
g
d
eep
learnin
g
m
o
d
el,
”
in
2
0
2
2
Inter
n
a
ti
o
n
a
l
Co
n
feren
ce
o
n
Ma
ch
in
e
Lea
rn
in
g
,
Big
Da
ta
,
Clo
u
d
a
n
d
Par
a
llel
Co
mp
u
tin
g
,
COM
-
IT
-
C
ON
2
0
2
2
,
2
0
2
2
,
p
p
.
4
5
6
–
4
6
2
,
d
o
i: 10
.1
1
0
9
/COM
-
IT
-
CON5
4
6
0
1
.2022
.98
5
0
6
6
7
.
[17
]
S.
Jad
h
av
an
d
A.
M.
Lal,
“
Multi
-
cl
ass
p
lan
t
leaf
d
ise
ase
d
etectio
n
u
sin
g
a
d
eep
co
n
v
o
lu
t
io
n
al
n
eu
ral
n
etw
o
rk,
”
Inter
n
a
tio
n
a
l
90
92
94
96
98
100
Accu
rac
y
Sensi
ti
vit
y
Specifi
c
i
t
y
F1 Scor
e
In
terms
of %
Machin
e
Learning
Perform
ance
Metrics
Yolo
R-CNN
SSMAN
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
940
-
948
948
Jo
u
rn
a
l of Info
rmation
Sys
tem Mo
d
elin
g
an
d
Desig
n
,
v
o
l
.
1
3
,
n
o
.
1
,
p
p
.
1
–
1
4
,
2
0
2
2
,
d
o
i: 10
.4
0
1
8
/ijismd.3
1
5
1
2
6
.
[18
]
R.
Bo
ra,
D
.
Parasa
r,
an
d
S.
Ch
arhate,
“
Plan
t
leaf
d
iseas
e
d
etectio
n
u
sin
g
d
eep
learnin
g
:
a
review,
”
in
2
0
2
2
IE
E
E
7
th
Inter
n
a
tio
n
a
l
co
n
feren
ce for
Co
n
verg
en
ce in Techn
o
lo
g
y,
I2C
T 20
2
2
,
2
0
2
2
,
p
p
.
3
4
–
3
7
,
d
o
i:
1
0
.1109
/I2C
T54
2
9
1
.20
2
2
.98
2
4
9
2
5
.
[19
]
S.
Ash
win
k
u
m
ar,
S.
Rajag
o
p
al,
V.
Manima
ran,
an
d
B.
Jeg
ajo
th
i,
“
Au
to
m
ated
p
lan
t
leaf
d
i
seas
e
d
etectio
n
an
d
class
ification
u
sing
o
p
tim
al
Mob
ileN
et
b
ased
co
n
v
o
lu
tio
n
al
n
eu
ral
n
et
wo
rks
,
”
Ma
teria
ls
To
d
a
y:
Pro
ceed
in
g
s
,
v
o
l.
5
1
,
p
p
.
4
8
0
–
4
8
7
,
2
0
2
1
,
d
o
i: 10
.1016
/j.m
atp
r.
2
0
2
1
.05
.5
8
4
.
[20
]
P.
Bed
i
an
d
P.
Go
le,
“
Plan
t
d
iseas
e
d
etectio
n
u
sin
g
h
y
b
rid
m
o
d
el
b
ased
o
n
co
n
v
o
l
u
tio
n
al
au
to
en
co
d
er
an
d
co
n
v
o
lu
ti
o
n
al
n
eu
ral
n
etwo
rk,
”
Artificia
l I
n
tellig
en
ce in Ag
ricu
ltu
re
,
v
o
l.
5
,
p
p
.
9
0
–
1
0
1
,
2
0
2
1
,
d
o
i: 10
.10
1
6
/j.aiia.
2
0
2
1
.05
.0
0
2
.
[21
]
M.
E
.
H
.
Ch
o
wd
h
u
ry
et
a
l.
,
“
Au
to
m
atic
an
d
r
eliab
le
leaf
d
iseas
e
d
etec
tio
n
u
sing
d
eep
learnin
g
tech
n
iq
u
es,
”
Agr
iEng
in
eerin
g
,
v
o
l.
3
,
n
o
.
2
,
p
p
.
2
9
4
–
3
1
2
,
2
0
2
1
,
d
o
i:
10
.33
9
0
/ag
rien
g
in
eering
3
0
2
0
0
2
0
.
[22
]
S.
M.
Siv
a
lin
g
am
an
d
L.
D
.
Bad
a
b
ag
n
i,
“
Enco
u
ragin
g
h
y
g
ien
e
p
erm
a
n
en
ce
in
to
m
ato
l
eaf
an
d
ap
p
ly
in
g
m
achi
n
e
le
arnin
g
tech
n
iq
u
es,
”
Ind
o
n
esia
n
Jo
u
rnal
o
f
Electrica
l
Eng
in
eerin
g
a
n
d
Co
mp
u
ter
S
cien
ce
(I
J
EE
CS
)
,
v
o
l.
3
3
,
n
o
.
1
,
p
p
.
3
4
3
–
3
4
9
,
2
0
2
4
,
d
o
i: 10
.1159
1
/ijee
cs.v
3
3
.i1
.pp
3
4
3
-
3
4
9
.
[23
]
K.
Su
d
h
a
Ran
i
an
d
B.
Priya
M
ad
h
u
ri,
“
Plan
t
leaf
d
ise
ase
d
etectio
n
u
sin
g
m
a
ch
in
e
lea
rnin
g
tech
n
iq
u
es,
”
Lec
tu
re
No
tes
o
n
Da
ta
Eng
in
eerin
g
and Co
mmu
n
ica
tio
n
s Te
ch
n
o
lo
g
ies
,
v
o
l.
5
8
,
p
p
.
5
1
1
–
5
1
8
,
2
0
2
1
,
d
o
i: 10
.10
0
7
/9
7
8
-
981
-
15
-
9
6
4
7
-
6
_
4
0
.
[24
]
J.
Liu
an
d
X
.
W
an
g
,
“
Plan
t
d
iseas
es
an
d
p
ests
d
etectio
n
b
ased
o
n
d
eep
learnin
g
:
a
review
,
”
Plan
t
Meth
o
d
s
,
v
o
l.
1
7
,
n
o
.
1
,
2
0
2
1
,
d
o
i: 10
.1186
/s13007
-
021
-
0
0
7
2
2
-
9.
[25
]
V.
K.
Vish
n
o
i,
K
.
Ku
m
ar,
an
d
B
.
Ku
m
ar
,
“
Plan
t
d
is
ease
d
etect
io
n
u
sin
g
co
m
p
u
tatio
n
al
in
tellig
en
ce
an
d
i
m
ag
e
p
rocess
in
g
,
”
Jo
u
rn
a
l of Plan
t Disea
ses
an
d
P
ro
tectio
n
,
v
o
l.
1
2
8
,
n
o
.
1
,
p
p
.
1
9
–
5
3
,
2
0
2
1
,
d
o
i: 10
.10
0
7
/s
4
1
3
4
8
-
0
2
0
-
0
0
3
6
8
-
0.
[26
]
A.
Ab
b
as,
S
.
Jain
,
M
.
Go
u
r,
an
d
S.
Van
k
u
d
o
t
h
u
,
“
To
m
ato
p
lan
t
d
iseas
e
d
etectio
n
u
sin
g
tr
an
sfer
l
earnin
g
wit
h
C
-
GA
N
sy
n
th
eti
c
im
ag
es,
”
Co
mp
u
te
rs
an
d
E
lectro
n
ics in
A
g
ricu
ltu
re
,
v
o
l.
1
8
7
,
2
0
2
1
,
d
o
i: 1
0
.10
1
6
/j.compag
.2021
.10
6
2
7
9
.
[27
]
V.
Sin
g
h
,
N.
Sh
a
r
m
a,
an
d
S
.
Sin
g
h
,
“
A
r
ev
iew
o
f
i
m
ag
in
g
tech
n
iq
u
es
for
p
lan
t
d
iseas
e
d
etectio
n
,
”
Artifi
ci
a
l
Intellig
en
ce
in
Agr
icu
ltu
re
,
v
o
l.
4
,
p
p
.
2
2
9
–
2
4
2
,
2
0
2
0
,
d
o
i: 10
.10
1
6
/j.aiia.
2
0
2
0
.10
.002.
[28
]
M.
Ji,
L.
Zhan
g
,
an
d
Q.
W
u
,
“
Au
t
o
m
atic
g
rape
le
af
d
iseas
es
id
en
tification
v
ia
Un
ited
Mod
el
b
ased
o
n
m
u
ltip
le
co
n
v
o
lu
ti
o
n
a
l
n
eu
ral
n
etwo
rks
,
”
Info
rma
tio
n
P
ro
cess
in
g
in Agr
icu
ltu
r
e
,
v
o
l.
7
,
n
o
.
3
,
pp
.
4
1
8
–
4
2
6
,
2
0
2
0
,
d
o
i: 10
.10
1
6
/j.inp
a.20
1
9
.1
0
.00
3
.
[29
]
P.
Sh
ar
m
a,
Y.
P.
S.
Ber
wal,
an
d
W
.
Gh
ai,
“
P
erf
o
r
m
an
ce
an
aly
sis
o
f
d
ee
p
learnin
g
CNN
m
o
d
els
for
d
iseas
e
d
etectio
n
in
p
lan
ts
u
sin
g
im
ag
e
seg
m
en
tatio
n
,
”
In
fo
rma
tio
n
Pro
cess
in
g
in
Agr
icu
l
tu
re
,
v
o
l.
7
,
n
o
.
4
,
p
p
.
5
6
6
–
5
7
4
,
2
0
2
0
,
d
o
i: 10
.1016
/j.inp
a
.20
1
9
.11
.0
0
1
.
[30
]
S.
D.
Kh
irade
an
d
A.
B.
Patil
,
“
Pla
n
t
d
iseas
e
d
etectio
n
u
sin
g
im
ag
e
p
rocess
in
g
an
d
m
achi
n
e
learnin
g
alg
o
rithm,
”
Jo
u
rn
a
l
o
f
Xidia
n
Univer
sity
,
v
o
l.
1
4
,
n
o
.
7
,
2
0
2
0
,
d
o
i: 10
.37
8
9
6
/jx
u
1
4
.7/012
.
BIOGR
AP
HI
ES OF
A
UTH
ORS
Saravanan
Madderi
Sivali
nga
m
is
profe
ss
or
in
the
Depa
rt
me
nt
of
Co
mput
e
r
Scie
nc
e
and
En
gine
er
ing,
Save
e
tha
Schoo
l
o
f
E
ngine
er
ing,
Save
et
ha
Insti
tut
e
of
Medic
a
l
and
Te
chn
ic
a
l
Sci
en
ce
s,
Ch
enna
i
,
I
ndia
.
H
avi
ng
a
tot
a
l
of
25
y
e
ars
of
academi
c
experie
n
ce,
inc
ludi
ng
2
y
ea
r
s
of
industry
exp
eri
en
ce
and
h
e
is
an
ac
co
mpl
ishe
d
rese
arc
h
er
and
a
ca
d
em
i
c
in
the
fi
el
d
of
co
mput
er
sc
ie
n
ce
and
engi
n
ee
r
ing
.
He
has
aut
hor
ed
2
Books,
ed
i
te
d
4
p
ape
rs,
publi
shed
1
58
p
ape
rs
in
r
efe
r
eed
interna
ti
ona
l
j
ourna
l
and
pre
s
ent
ed
26
papers
in
ref
ereed
conf
ere
n
ce
pro
c
ee
dings
and
16
pa
t
ent
s
publ
ished
.
He
was
a
lso
giv
en
21
m
aj
or
inv
ited
cont
ributions
an
d/or
t
ec
hni
cal
r
e
ports
abstract
s
a
nd/or
pap
ers
r
ead.
Hav
ing
cl
oud
founda
t
ion
ce
rt
ifi
c
ate
and
d
at
a
ana
ly
ti
cs
IB
M
Cognos
C
ert
i
fic
a
te
.
Sin
ce
20
17
he
h
as
serv
e
d
a
t
Hara
ma
ya
Univer
sity,
E
a
st
Afric
a
as
a
p
rofe
ss
or
in
th
e
Scho
ol
of
Comput
ing
for
two
y
ea
rs.
D
r.
Sar
ava
n
an
is
a
m
em
be
r
of
IE
EE
,
IST
E
a
nd
IE
T
as
we
ll
as
Student
Br
a
nch
Coun
ci
l
lor
of
IE
EE
and
Innova
ti
on
Amb
assador
of
Institution’s
innov
at
i
on
Council
(IIC
),
SIM
ATS.
Dr
.
Sara
v
ana
n’s
exp
ertise
l
ie
s
in
artificia
l
in
te
l
ligence
and
cl
oud
-
base
d
t
ec
hno
lo
gie
s.
He
has
be
en
re
cognize
d
for
his
groundbr
ea
king
work
in
deve
lop
ing
a
le
s
s
cost
c
abi
ne
t
d
yei
ng
pro
ce
ss
u
sing
proc
ess
mi
ning
t
ec
hniqu
e
for
thi
s
dev
el
o
ped
a
new
a
lgor
it
h
m
c
al
l
ed
“
Li
n
kRule
Miner
”
.
D
r.
Sara
van
an
le
ads
a
dyn
am
i
c
rese
arc
h
g
roup
foc
used
on
adv
anc
ing
art
if
ic
i
al
int
ellig
enc
e
-
b
ase
d
produc
ts.
His
la
b's innovat
ive
rese
a
rch
has
bee
n
publi
shed
i
n
leading
pe
er
-
re
vie
wed
journals.
He
al
so
ha
s
the
best
fa
culty
and
r
ese
ar
che
r
awa
rd
from
na
tional
and
in
te
rn
a
ti
onal
soc
ie
t
ie
s.
He
a
ctivel
y
me
ntors
gr
aduat
e
studen
ts
and
c
oll
abor
at
es
with
industry
p
art
n
e
rs
to
br
idge
the
gap
be
twee
n
ac
ad
em
i
a and
pr
ac
t
ic
a
l appl
i
ca
t
io
ns.
He can
be
co
nta
c
te
d
at e
m
ail:
sara
nen
adu@
g
ma
il.c
o
m.
Lak
sh
mi
Devi
Bad
abagn
i
r
ec
e
ive
d
the
B
.
T
ec
h
.
degr
ee
in
c
omput
er
s
ci
en
ce
and
engi
ne
eri
ng
fro
m
Vaa
gd
evi
En
gine
er
ing
and
T
ec
hnology
and
Scie
nc
es,
JN
TU
A
Univer
sity
,
Ananta
pur,
And
hra
Prad
esh
in
2
013
and
M.
Te
ch
degr
ee
in
co
mp
ute
r
scie
n
ce
and
eng
ine
er
ing
from
PV
KK
Ins
ti
tutions
,
JN
TUA,
AP
in
2015
.
C
urre
ntl
y
she
is
an
assistan
t
pro
fessor
at
th
e
D
epa
rtment
o
f
Comput
er
S
cienc
e
and
E
ngi
nee
ring
in
SO
ET
,
Sri
Padm
ava
t
i
Mahi
la
Visw
avi
dyal
ay
a
m,
Ti
rup
at
i
,
AP
.
Her
rese
arc
h
in
t
ere
sts inc
lud
e
d
i
sea
se
d
et
e
ction
a
nd
pre
d
ic
t
ion
in
l
ea
fs
usin
g
machine
l
e
arn
ing
te
ch
niq
ues
.
She
ca
n
be
contac
t
e
d
at
e
ma
i
l
la
kshmide
v
i580
3@gma
il.c
o
m.
Evaluation Warning : The document was created with Spire.PDF for Python.