Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
23
,
No.
3
,
Septem
ber
2021
, pp.
1681
~
1688
IS
S
N: 25
02
-
4752,
DOI: 10
.11
591/ijeecs
.v
23
.i
3
.
pp
1681
-
1688
1681
Journ
al h
om
e
page
:
http:
//
ij
eecs.i
aesc
or
e.c
om
Recognit
ion of
m
ango l
eaf diseas
e usi
ng con
v
olutio
nal neur
al
network
models
:
a t
ransfe
r learnin
g approa
ch
Ad
it
ya Ra
jbo
ng
shi
1
,
Thaha
ri
m Kh
an
2
, M
d. Mahbu
bur
Rahma
n
3
An
i
k Pr
amanik
4
, Sha
h
Md
T
anvir
Siddiquee
5
,
Naray
an
R
an
j
an C
hakra
bo
r
t
y
6
1
Depa
rtment of
Com
pute
r
Scie
n
ce
and Engi
ne
ering,
Jaha
ng
irna
g
a
r
Univer
sit
y
,
Dh
aka
,
Bang
la
desh
2,4,5,6
Depa
rtment
of
Com
pute
r
Sci
enc
e
and
Engi
n
e
eri
ng,
Daffodi
l
I
nte
rna
ti
ona
l
Uni
ver
sit
y
,
Dhaka
,
Bangl
ad
esh
3
Cro
wd Re
alt
y
,
Tok
y
o
,
Jap
an
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
6
,
2021
Re
vised
Ju
l
1
8
,
2021
Accepte
d
Aug
4
,
2021
The
ac
knowled
gm
ent
of
pl
ant
disea
ses
assumes
an
ind
ispensabl
e
par
t
i
n
ta
king
in
fecti
ous
pre
vention
m
eas
ure
s
to
improve
the
qua
li
t
y
and
amount
of
har
vest
y
i
el
d
.
Mec
haniza
t
ion
of
pla
nt
d
isea
se
s
is
a
lot
adva
n
ta
geous
as
i
t
dec
re
ase
s
th
e
ch
ec
king
work
in
an
en
orm
ous
cu
l
ti
vated
area
whe
re
m
ango
is
pla
nt
ed
to
a
hug
e
extend.
L
ea
v
es
bei
ng
th
e
food
hotspot
for
pla
nt
s,
the
e
a
r
l
y
and
pre
ci
se
r
ec
o
gnit
ion
of
leaf
d
isea
ses
is
signif
i
ca
nt
.
Th
is
work
foc
used
on
grouping
and
disti
nguishing
th
e
disea
ses
of
m
ango
le
ave
s
t
hrou
gh
the
proc
ess
of
CNN
.
DenseNet
201,
I
nce
pt
ionRe
sN
etV
2,
Inc
eptionV3,
ResNet50,
ResNet152V2,
a
nd
Xce
pti
on
al
l
the
se
m
odel
s
of
CNN
with
tra
nsfer
learni
ng
te
chn
ique
s
are
u
sed
her
e
for
get
ti
ng
bet
t
er
accur
acy
from
the
ta
r
get
ed
data
set.
Im
age
a
cquis
it
ion,
image
se
gm
ent
at
ion
,
and
feature
s
ext
ra
ction
ar
e
t
h
e
steps
invol
ved
in
disea
se
detec
ti
o
n.
Diffe
r
ent
kind
s
of
leaf
dise
ase
s
which
are
conside
red
as
t
he
c
la
ss
for
th
i
s
work
such
as
ant
hr
ac
nose
,
g
al
l
m
a
chi
,
powder
y
m
il
de
w,
red
rust
are
u
sed
in
the
dataset
consisti
n
g
of
1500
images
of
disea
sed
and
al
so
health
y
m
a
ngo
le
av
es
image
data
anot
h
er
c
la
ss
is
al
so
adde
d
in
th
e
d
at
ase
t.
W
e
hav
e
a
lso
evalua
ted
the
over
a
ll
p
erf
orm
ance
m
at
ric
es
and
fo
und
tha
t
the
D
ense
Net201
out
per
form
s
by
ob
ta
ini
ng
the
highe
st
ac
cu
racy a
s 98
.
00%
th
an othe
r
m
odel
s
.
Ke
yw
or
ds:
Cl
assifi
cat
ion
Den
s
eNet
201
Ma
ngo
le
af
Neural
netw
ork
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Ad
it
ya
Raj
bongsh
i
Dep
a
rtm
ent o
f C
om
pu
te
r
Scie
nce a
nd E
ng
i
ne
erin
g
Jaha
ng
ir
na
gar
Un
i
ver
sit
y
Dh
a
ka, B
an
gla
des
h
Em
a
il
: adit
ya
ra
j
.
j
uc
se@
gm
ai
l.
com
1.
INTROD
U
CTION
Ma
ngo
is
one
of
the
ta
sty
and
m
os
t
signi
ficant
natu
ral
product
cr
ops
wh
ic
h
is
de
velo
ped
i
n
hortic
ulture
.
It
is
sent
out
to
nu
m
erous
nati
on
s
as
cr
ude
or
rea
dy
f
oods
grown
from
the
gro
und
as
ha
nd
le
d
consum
ables
li
ke
rea
dy
m
ang
o
cuts
or
s
queeze
c
rude
m
ang
o
pic
kle,
an
d
s
o
f
or
th
m
ang
o
is
plen
ti
fu
l
i
n
nu
t
rient
A
a
nd
C. D
ise
ase
in
l
eaf m
akes th
e
photosy
nth
esi
s
b
e im
ped
ed
a
nd a
ff
ect
th
e acc
ru
al
of the
pla
nt
.
U
n
b
o
l
t
i
n
g
o
f
d
i
s
e
a
s
e
c
a
n
b
e
c
o
m
p
l
e
t
e
d
b
y
f
a
r
m
e
r
s
b
y
i
n
c
e
s
s
a
n
t
c
h
e
c
k
i
n
g
o
f
t
h
e
p
l
a
n
t
l
e
a
v
e
s
.
F
o
r
l
i
t
t
l
e
-
s
c
o
p
e
r
a
n
c
h
e
r
s
,
e
a
r
l
y
r
e
c
o
g
n
i
z
a
b
l
e
p
r
o
o
f
o
f
i
l
l
n
e
s
s
i
s
a
l
o
t
o
f
c
o
n
c
e
i
v
a
b
l
e
a
n
d
r
e
a
d
y
t
o
c
o
n
t
r
o
l
t
h
e
b
u
g
s
b
y
n
a
t
u
r
a
l
p
e
s
t
i
c
i
d
e
s
o
r
b
y
t
h
e
u
t
i
l
i
z
a
t
i
o
n
o
f
n
e
g
l
i
g
i
b
l
e
m
e
a
s
u
r
e
o
f
c
o
m
p
o
u
n
d
p
e
s
t
i
c
i
d
e
s
.
F
o
r
h
u
g
e
s
c
o
p
e
r
a
n
c
h
e
r
s
r
e
g
u
l
a
r
o
b
s
e
r
v
i
n
g
a
n
d
e
a
r
l
y
d
i
s
t
i
n
g
u
i
s
h
i
n
g
p
r
o
o
f
o
f
i
l
l
n
e
s
s
i
s
u
n
i
m
a
g
i
n
a
b
l
e
a
n
d
i
t
b
r
i
n
g
s
a
b
o
u
t
a
s
e
r
i
o
u
s
e
p
i
s
o
d
e
o
f
t
h
e
s
i
c
k
n
e
s
s
a
n
d
i
r
r
i
t
a
t
i
o
n
d
e
v
e
l
o
p
m
e
n
t
w
h
i
c
h
c
a
n
'
t
b
e
c
o
n
s
t
r
a
i
n
e
d
b
y
n
a
t
u
r
a
l
m
e
t
h
o
d
s
.
I
n
t
h
i
s
c
i
r
c
u
m
s
t
a
n
c
e
,
f
a
r
m
e
r
s
a
r
e
c
o
m
p
e
l
l
e
d
t
o
u
t
i
l
i
z
e
t
h
e
n
o
x
i
o
u
s
s
y
n
t
h
e
t
i
c
s
u
b
s
t
a
n
c
e
s
t
o
t
e
r
m
i
n
a
t
e
t
h
e
d
i
s
e
a
s
e
t
o
h
o
l
d
t
h
e
h
a
r
v
e
s
t
y
i
e
l
d
.
T
h
i
s
i
s
s
u
e
c
a
n
b
e
t
a
c
k
l
e
d
v
i
a
m
e
c
h
a
n
i
z
i
n
g
t
h
e
c
h
e
c
k
i
n
g
c
y
c
l
e
b
y
t
h
e
u
t
i
l
i
z
a
t
i
o
n
o
f
c
u
t
t
i
n
g
-
e
d
g
e
p
i
c
t
u
r
e
h
a
n
d
l
i
n
g
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2
02
1
:
16
81
-
16
88
1682
m
e
t
h
o
d
s
.
T
h
e
p
r
o
p
o
s
e
d
w
o
r
k
a
i
m
s
a
r
e
m
a
k
i
n
g
a
m
o
d
e
l
i
n
v
o
l
v
e
d
i
n
d
i
s
e
a
s
e
d
e
t
e
c
t
i
o
n
w
h
i
c
h
i
s
p
e
r
f
o
r
m
e
d
b
y
u
s
i
n
g
t
h
e
t
e
c
h
n
i
q
u
e
o
f
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
(
C
N
N
)
w
h
i
c
h
i
s
a
p
a
r
t
o
f
a
r
t
i
f
i
c
i
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
(
A
N
N
)
.
V
e
r
y
f
e
w
r
e
s
e
a
r
c
h
h
a
s
b
e
e
n
p
e
r
f
o
r
m
e
d
r
e
g
a
r
d
i
n
g
t
h
e
r
e
c
o
g
n
i
t
i
o
n
o
f
m
a
n
g
o
l
e
a
f
d
i
s
e
a
s
e
.
B
u
t
t
h
e
r
e
h
a
v
e
s
o
m
e
l
i
m
i
t
a
t
i
o
n
s
a
n
d
m
o
s
t
o
f
t
h
e
a
p
p
r
o
a
c
h
e
s
w
e
r
e
u
s
e
d
t
r
a
d
i
t
i
o
n
a
l
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
.
Ba
i
et
al
.
[1
]
pr
esent
a
te
ch
niq
ue
for
bir
d
ac
knowle
dgm
ent
based
on
In
ce
ption
-
v3.
The
obj
ect
ive
of
the
LifeC
LE
F
2019
Bi
r
d
Re
cogniti
on
is
t
o
ide
ntify
a
nd
arr
a
nge
659
fled
gling
sp
ec
ie
s
inside
t
he
ga
ve
so
un
ds
ca
pe
ch
ronicl
es.
Lo
g
-
Me
l
sp
ect
rogr
a
m
s
are
re
m
ov
e
d
as
highli
ghts
and
I
nce
ption
-
v3
is
util
iz
ed
f
or
bir
d
so
un
d
rec
ognit
ion
.
D
hank
har
[2
]
propose
d
a
m
et
ho
d
for
t
he
use
of
f
eat
ur
e
e
xtracti
on
of
facial
ex
pr
e
ssions
with
a
blen
d
of
ne
ural
orga
ni
zat
ion
f
or
the
ackno
wled
gm
e
nt
of
var
io
us
f
aci
al
feeli
ng
s.
A
hum
an
m
ind
ca
n
hav
e
a
par
c
el
of
feeli
ng
s
y
et
this
pa
pe
r
m
anag
es
t
he
f
undam
ental
7
feeli
ngs.
Thi
s
pa
per
m
anag
es
t
he
excep
ti
onal
ly
eff
ect
ive
m
od
e
l
of
the
joine
d
m
od
el
s
of
VGG
16
a
nd
Re
s
Net5
0
ex
pa
nding
t
he
product
ivit
y
to
92.4%.
C.
Lin
et
al
.
[
3]
f
ocus
es
on
c
hrom
os
om
e
cl
assifi
cation
t
hroug
h
a
deep
le
arn
i
ng
m
od
el
.
Althou
gh
t
he
deep
le
ar
ning
-
base
d
I
nce
ptio
n
e
ng
i
neer
i
ng
has
yi
el
de
d
c
ut
ti
ng
-
e
dg
e
exe
cution.
In
this
way,
they
buil
d
up
a
pro
gr
am
m
ed
c
hrom
os
om
e
gr
ouping
w
hich
is
cal
le
d
CIR
-
Net
dep
e
ndent
on
I
nce
ption
-
Re
sN
et
.
Lin
et
al
.
[4
]
pro
po
se
d
a
t
ra
ns
fe
r
le
ar
ning
-
base
d
strat
egy is
pr
ese
nted
f
or
traf
fic
sig
n
ac
knowle
dgm
ent
.
It
al
so
dec
rea
ses
th
e
m
easur
e
of
pr
e
par
i
ng
i
nfor
m
at
ion
as
w
el
l
as
m
itigates
cal
culat
ion
c
os
t
uti
li
zi
ng
the
I
nce
ption
-
v3
m
od
e
l.
The
ou
tc
om
es
acc
om
pl
ish
in
this
work
a
re
up
to
99.
18
%
of
ackno
wled
gme
nt
exactnes
s
at
0.
05
le
ar
ning
rate
(no
rm
al
pr
eci
s
ion
of
99.
09
%.
Fer
reira
et
al
.
[
5]
de
fine
d
a
m
od
el
f
or
t
he
cha
racteri
z
at
ion
of
br
east
cancer
histolo
gy
pictu
res.
The
pr
e
-
ow
ned
orga
nizat
ion
of
this
pr
o
pose
d
work
is
t
o
pe
r
form
the
char
a
ct
erizat
ion
wit
h
I
nce
ption
-
Re
sn
et
V
2.
T
o
de
feat
the
a
bse
nce
of
inf
orm
at
ion
,
i
nfor
m
at
ion
e
nlar
gem
e
nt
is
perform
e
d
as
well
.
Ra
hm
an
et
al.
[6
]
pro
po
s
ed
a
no
t
her
w
ork
wh
ic
h
is
to
rec
ognize
diff
e
re
nt
local
bir
ds
of
Ba
ngla
des
h
fr
om
i
m
age
inf
or
m
at
ion
.
H
ere
m
ai
nly
fo
ur
m
et
ho
do
l
og
ie
s
to
be
spe
ci
fic
In
ce
ptio
n
v3
w
it
hout
tra
ns
fe
r
le
ar
ning,
In
ce
ptio
n
-
v3
with
trans
f
er
le
arn
in
g,
M
ob
il
eNet
wit
hout
tra
ns
fe
r
le
arn
i
ng,
an
d
Mob
il
eNet
wit
h
trans
fer
le
ar
nin
g
for
fig
ur
in
g
ou
t
how
t
o
achie
ve
the
e
rr
a
nd.
M
ukti
et
al.
[
7]
in
tro
du
ce
d
a
T
ra
ns
fe
r
Lea
rn
i
ng
-
base
d
C
NN
m
od
el
that
was
creat
ed
for
t
he
disti
nguish
i
ng
pr
oof
of
plant
di
sease
ex
act
ly
.
The
dataset
,
t
hey
ha
ve
util
iz
ed
is
com
pr
ise
d
of
7029
5
prepa
rin
g
pictures
a
nd
1757
2
ap
pro
val
i
m
ages
ho
l
ding
38
dis
ti
nct
cl
asses
of
plant
l
eaves
pictures
.
T
hey
hav
e
fo
c
us
e
d
pr
e
dom
inantly
on
th
e
Re
sNe
t50
netw
ork,
a
well
-
know
n
C
NN
m
od
el
as
a
pr
e
-
pr
e
par
e
d
m
od
e
l i
n
Tra
nsfer
L
earn
i
ng. Sharm
in
et
a
l.
[
8] pr
opose
d fis
h determ
inati
on
in fr
eshwat
er.
Anothe
r
m
or
e
appr
oach
is
al
so
inclu
de
d
he
rew
it
h
as
set
ti
ng
up
4
ty
pes
of
feat
ur
es
th
e
n
gray
scal
e
i
m
age
create
d
from
the
colo
r
i
m
age.
A
gain
t
hro
ugh
this
his
togram
m
od
el
wh
ic
h
is
a
ppea
red
duri
ng
the
tim
e
of
c
onve
rsion
perform
ed
seg
m
entat
ion
.
F
or
recogn
iz
in
g
t
he
fis
h
3
cl
ass
ifie
rs
are
use
d
in
this
m
od
el
a
m
on
g
them
SV
M
per
f
or
m
s
well
.
Pr
akas
h
et
al
.
[9
]
pro
pose
d
an
oth
e
r
w
ork
on
var
i
ou
s
infecti
ons
w
hich
is
respo
ns
ible
for
the
i
nf
lu
enc
e
of
m
ang
o
na
tural
pr
oduct
pro
fita
bili
ty
.
This
pap
e
r
a
ddit
ion
al
ly
re
presents
pr
eca
utio
n
wa
ys
and
so
l
utio
ns
f
or
in
fecti
on.
T
he
fr
am
ewo
r
k
dem
on
strat
es
in
this
pape
r
ha
ve
t
he
cap
aci
ty
to
identify
le
af
-
ba
sed
diseases
howe
ve
r
there
a
re
m
any
sign
ific
ant
infecti
ons
that
influ
ence
the
eff
ic
ie
ncy
of
the
m
ang
o
c
rop.
S
hukla
et
al.
[
10]
propose
d
one
w
ork
on
I
m
age
processi
n
g
wh
ic
h
is
one
of
the
f
undam
ental
strides
f
or
hold
ing
in
novatio
n
in
a
picture
tha
t
m
igh
t
be
de
ba
sed
beca
us
e
of
a
few
com
m
otion
s
or
beca
use
of
diff
e
re
nt cau
se
s.
This p
a
per
is
an
essen
ce o
f a
f
ew
Im
age P
r
eprocessi
ng
procedu
res
that ca
n
be
util
iz
ed.
Lei
et
al
.
[
11]
pro
po
s
ed
one
wor
k
on
Se
gm
entat
ion
of
im
ages
wh
ic
h
is
sig
nific
ant
f
or
picki
ng
appr
opriat
e
te
rr
it
ory
of
interest
i
n
a
picture
.
T
his
is
crit
ic
al
to
le
ss
en
handlin
g
ti
m
e
that
m
igh
t
be
nee
ded
to
de
al
with
rathe
r
whole
picture
.
T
his
pa
per
sho
ws
a
ne
w
s
e
gm
entat
i
on
strat
egy
cal
le
d
f
ast
a
nd
robu
st
f
uzzy
c
-
m
eans
cal
c
ulati
on
.
Th
e
m
agn
ific
ence
of
this
cal
cula
ti
on
is
that
it
conseq
ue
ntly
picks
a
bu
nch
that
con
ta
ins
il
lness.
Likewise,
it
neg
at
es
in
dicat
ing
pref
ere
nce
s
ov
e
r
cust
oma
ry
FCM
cal
c
ulati
on
.
Pate
l
et
al.
[
12]
pro
po
s
ed
on
e
wor
k
that
e
m
ph
asi
zes
ev
ery
featur
e
of
the
i
m
age
to
be
separ
at
ed
.
Highli
gh
ts
that
can
be
extricat
ed
f
ro
m
the
picture
m
ay
be
of
dif
fer
e
nt
so
rts
li
ke
w
orl
dw
ide
or
nearby.
I
n
that,
the
re
are
subcat
e
gories
li
ke
su
r
face
-
base
d
or
colo
r
-
base
d
feat
ur
es
.
Ha
rali
ck
et
al.
[
13]
pro
pose
d
on
e
w
ork
th
at
portrays
s
om
e
eff
ect
ively
proces
s
-
a
ble
te
xtural
featur
e
s
de
penden
t
on
gray
ton
e
s
patia
l
con
diti
ons
an
d
s
hows
their
a
ppli
cat
ion
in
cat
egory
-
ide
ntific
at
ion
unde
rtakin
gs
of
three
var
io
us
ty
pes
of
pictu
re
inf
or
m
at
ion
,
two
sorts
of
c
ho
ic
e
gui
delin
es.
These
ou
tc
om
es
dem
on
strat
e
th
at
the
ef
fecti
ve
proces
s
a
ble
te
xtural
inclu
des
li
kely
have
overall
a
ppr
opriat
eness
f
or
a
wi
de
assor
tm
ent
of
picture
gro
up
i
ng
a
pp
li
cat
io
ns.
H
u
et
al.
[
14
]
pro
pose
d
ano
t
her
m
et
ho
d
that
is
a
featur
e
in
MATLAB
-
bas
ed
f
ram
ewo
r
k
is
by
util
iz
ing
gr
ay
le
vel
co
-
e
ve
nt
m
at
ri
x
(
GLCM
)
.
MATLAB
has
an
i
n
-
fabrica
te
d
cap
aci
ty
fo
r
draw
ing
GLCM.
T
his
pa
per
giv
e
s
conditi
ons
to
com
pu
ti
ng
diff
e
ren
t
s
urface
-
bas
e
d
highli
gh
ts
f
rom
the
GLCM
gr
i
d.
Padm
aja
et
al.
[15
]
propose
d
on
e
m
et
h
od
w
hich
is
on
featu
re
e
xtract
ion
,
f
or
eff
ect
ual
disea
ses
disco
ver
y
it
is
i
m
po
rtant
to
m
ake
a
le
gitim
at
e
deter
m
in
at
ion
of
highli
gh
ts
that
best
dep
ic
t
s
the
gi
ven
sic
kness
is
si
gn
i
ficant.
T
his
paper
deli
neates
var
i
ou
s
hi
gh
li
gh
ts
determ
inat
ion
cal
culat
io
ns
a
nd
highli
gh
ts
scor
ing
tec
hn
i
ques.
The
m
ai
n
fo
c
us
of
this
w
ork
is
to
recog
nize
the
diseases
of
m
ang
o
le
a
ves
with
s
om
e
prom
inent
m
od
el
s
of
CN
N.
T
her
e
a
re
f
our
ty
pes
of
di
seases
that
ar
e
us
ed
for
rec
ogniti
on
w
hich
are
al
so
def
i
ned
a
s
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Reco
gn
it
io
n of
man
go leaf
dis
ease
us
in
g
c
on
volutio
na
l
ne
ural netw
or
k
model
s…
(
Adity
a Rajb
ongshi
)
1683
cl
asses
into
th
e
m
od
el
al
ong
with
the
set
of
healt
hy
le
aves
.
De
ns
e
Net2
01,
I
ncep
ti
onRe
s
NetV
2,
In
ce
ptio
nV
3,
Re
sNet5
0,
Re
s
Net1
52V
2,
X
c
eption
al
l
thes
e
m
od
el
s
are
us
e
d
he
re
an
d
then
ou
r
res
ul
t
is
weigh
ed
on
t
he
pr
i
nciple
of
se
ven
perf
or
m
ance
m
at
rices
su
ch
as
accu
racy
,
F1
sc
or
e
,
pre
ci
sion
,
se
ns
it
iv
it
y,
sp
eci
fici
ty
,
FN
R,
FPR. T
he
c
onc
ise
traci
ng of
our
stu
dy is
giv
e
n belo
w,
w
hich
n
a
rr
at
es t
h
e
w
ho
le
stu
dy prec
ise
ly
:
Firstl
y creat
e a
n
a
uto
m
at
ed
m
et
hod for
recog
nizing t
he dise
ase o
f
m
ang
o
le
aves.
T
h
i
s
p
r
o
p
o
s
e
d
s
t
u
d
y
g
i
v
e
s
b
e
t
t
e
r
a
c
c
u
r
a
c
y
f
o
r
m
a
k
i
n
g
a
d
e
p
t
h
o
b
s
e
r
v
a
t
i
o
n
a
m
o
n
g
t
h
e
v
a
r
i
o
u
s
m
o
d
e
l
o
f
C
N
N
.
The
acc
ur
acy
of
this pro
po
se
d
syst
e
m
is
good
e
nough
t
ha
n
oth
e
r
m
et
ho
ds
b
ecau
se
of
t
he
i
m
ple
m
entat
io
n
of
trans
fer
lea
rn
i
ng am
on
g al
l t
he
classe
s.
This
pro
po
se
d
w
ork
is
de
bunke
d
in
s
uc
h
a
n
e
ff
ect
ive
way
wh
ic
h
m
ake
it
com
petit
ive
than
ot
her
dem
on
strat
ed
m
od
el
f
or m
ang
o l
eaves
d
ise
a
ses d
et
ect
io
n.
2.
RESEA
R
CH MET
HO
D
In
t
his
par
t,
th
e
pro
posed
w
ork
is
des
cribe
d
in
a
f
ull
m
ann
er
li
ke
data
preprocessi
ng,
de
scriptio
n
of
m
od
el
,
trai
ning,
a
nd
te
sti
ng
of
t
he
data
set
,
cl
assifi
cat
ions,
an
d
al
s
o
t
he
desc
riptio
n
of
the
dataset
.
B
el
ow
Figure
1
il
lustr
at
e the pr
ocedu
re
of
t
he pr
opose
d wor
k.
Figure
1.
Im
ple
m
entat
ion
pro
cedure
of the
pro
posed
wo
rk
2.1.
Dataset
acqu
isi
ti
on
To
de
velo
p
the
first
dataset
,
t
he
whole
m
ang
o
le
af
was
shoo
t.
A
r
ound
1000
pictu
res
w
ere
ass
um
ed
con
t
ro
l
over
a
four
te
e
n
-
day.
Pict
ur
es
of
diseases
we
re
ta
ke
n
us
in
g
a
co
uple
of
ge
no
ty
pe
s
of
m
ang
o
to
f
urnis
h
the
profo
und
l
earn
i
ng
m
od
el
with
the
total
ra
ng
e
of
in
dicat
ion
s
f
o
r
ev
ery
il
lness.
Th
e
pictu
res
we
r
e
th
e
n
m
et
a
m
or
phos
e
d
by
the
m
ea
ns
of
data
au
gm
entat
ion
to
m
ake
the
aux
il
ia
ry
dataset
.
These
dataset
s
wer
e
screene
d
to
ta
r
get
m
od
el
eff
i
ci
ency
with
w
ho
le
le
af
i
m
ages,
ye
t
a
few
i
m
ages
con
tra
ste
d
with
m
or
e
edite
d
le
aves.
The
ce
ntral
sp
ec
ulati
on
was
that
th
e
pictures
of
e
dited
le
aves
w
ou
l
d
upgra
de
m
od
el
eff
ic
ie
ncy
to
appr
opriat
el
y
per
cei
ve
disea
se
as
the
i
nform
at
ion
al
colle
ct
ion
was
m
or
e
prom
inent.
Her
e
Fig
ur
e
2
show
s
so
m
e d
at
a set f
ro
m
5
diff
e
ren
t
classe
s
of
t
he data
set
.
Figure
2.
Vis
ua
li
zat
ion
of
dataset
(
a)
H
eal
th
y l
eaf (b) P
owder
y m
il
dew
(
c
)
A
nt
hr
ac
nose
(d)
Re
d
r
us
t
(e) Gall
m
achi
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2
02
1
:
16
81
-
16
88
1684
2.2.
Data
pre
processin
g
Data
prep
r
oce
ssing
ref
e
rs
t
o
the
tran
sf
or
m
at
ion
s
of
t
he
r
aw
data
befo
r
e
the
dataset
i
s
try
ing
t
o
process
with
a
ny
oth
e
r
al
gori
thm
.
W
he
n
the
raw
dataset
is
try
ing
to
proce
ss
with
a
ny
othe
r
al
gorithm
in
CN
N
then
this
will
le
ad
to
a
bad
cl
assifi
cat
ion
r
esult
[16].
Dat
a
aug
m
entat
ion
is
perform
e
d
in
this
re
gard
f
or
reducin
g
the
ba
d
cl
assifi
cat
io
n
pro
blem
.
Diver
sit
y
of
t
rain
ing
data
an
d
to
incre
ase
the
a
m
ou
nt
of
data
Data
aug
m
entat
ion
[
17
]
is
us
e
d.
Cl
ass
i
m
balance
pro
blem
s
in
the
data
set
can
al
so
be
s
olv
e
d
to
so
m
e
extent
with
data
au
gm
entat
ion
.
Af
te
r
pe
r
form
ing
the
a
ug
m
entat
ion
t
her
e
a
re
al
m
os
t
500
im
ages
for
eac
h
ta
r
ge
t
cl
as
s
wh
ic
h
inc
rease
s
the
am
ou
nt
of
data
set
to
a
huge
e
xten
d.
A
fter
the
a
ugm
e
ntati
on
,
the
data
set
is
read
y
t
o
s
plit
for
trai
ning a
nd test
in
g purpo
ses.
We
us
e t
he
r
at
io
of 80%
-
20% fo
r
trai
n a
nd test
operati
on in
t
he data
s
et
.
2.3.
Model
d
escri
pt
ion
The
first
la
ye
r
of
a
CN
N
m
od
el
is
c
onvoluti
on
al
la
ye
r
wh
ic
h
is
co
ns
i
der
as
a
n
in
put
la
ye
r.
T
he
ou
t
pu
t
of CN
N
can be
de
no
te
d
m
at
he
m
at
ic
a
l
ly
as f
ollo
ws
:
=
f (
∑
−
1
∗
+
x
∈
M
y
)
(1)
Wh
e
re,
C
y
rep
resen
ts
the
set
of
outp
ut
fea
ture
m
aps,
M
y
rep
re
sents
th
e
set
of
input
m
aps,
q
xy
represe
nts
the
ker
nel
f
or
c
onvoluti
on,
r
y
r
epr
ese
nts
the
bias
te
rm
.
Be
l
ow
F
i
gure
3
dep
ic
t
s
the
w
orki
ng
proce
dure
for C
NN
i
n
t
his pr
opos
e
d wor
k.
Figure
3.
Wo
r
ki
ng
proc
e
dure
of CN
N
in
m
a
ngo
disease
det
ect
ion
Ag
ai
n
f
or
the
batch
no
rm
aliz
at
ion
,
the
la
ye
r
is
us
ually
pe
rfor
m
ed
betwe
en
the
co
nvol
ut
i
on
al
la
ye
r
and
t
he
Re
LU
la
ye
r.
Re
LU
la
ye
r
is
an
act
ivati
on
la
ye
r
that
is
m
ai
nly
us
ed
for
ad
ding
s
ome
non
-
li
ne
arit
y
int
o
the
netw
ork
by
chan
gi
ng
al
l
the
neg
at
ive
act
ivati
on
valu
es
to
0.
It
sp
e
eds
up
furthe
r
m
or
e,
di
m
inis
hes
th
e
sensiti
vity
of
the
m
od
el
.
In
this
la
ye
r,
the
act
ivati
on
of
each
la
ye
r
is
stan
da
r
dized
by
s
ub
t
racti
ng
the
sm
a
ll
er
than
m
ini
-
batch
m
eans
and
s
epar
at
in
g
by
t
he
m
ini
-
batch
sta
nd
a
rd
de
viati
on
.
T
his
is
tr
ai
le
d
by
m
ov
ing
t
he
input by a
n o
ffset
β and
a
fter
ward scali
ng it
by a
factor ɣ
.
The
m
at
he
m
at
i
c re
pr
ese
ntati
on
of
the
batch
norm
a
li
za
ti
on
ou
t
pu
t i
s
g
i
ven b
el
ow
,
= CN
ɣ, β
(a
i
)
≡ ɣ
a
i
+
C
(2)
a
i
is t
he n
or
m
a
l
iz
at
ion
of acti
va
ti
on
wh
ic
h
is
giv
e
n by the
eq
uation
,
a
i
=
+
√
2
+
∈
(3)
is t
he
m
ini batch m
ean
2
is m
i
ni b
at
c
h varia
nc
e
Ag
ai
n
the
re
al
ways
a
s
oft
m
a
x
la
ye
r
is
pres
ent
with
outp
ut
la
ye
r.
S
of
tm
ax
la
ye
r
al
wa
ys
pro
du
ce
ou
t
pu
t
base
d o
n
pro
bab
il
it
y. The
m
at
he
m
at
i
cal
eq
uati
on of
softm
ax
is giv
en belo
w
,
P (
C
x
| m
,
θ
)
=
P
(
m
,
θ
|
C
x
)
P
(
C
x
)
∑
P
(
m
,
θ
|
C
y
)
P
(
C
y
)
=
1
(4)
Exce
pt
f
or
this
tradit
io
nal
wa
y
of
CN
N
t
his
pro
posed
w
or
k
is
c
onstr
ucted
on
the
basis
of
6
well
-
known
m
od
el
s
for
the
im
age
data
set
.
Den
se
Net2
01,
I
nce
ptio
nResNetV
2,
In
c
eptionV
3,
Re
sNet5
0,
Re
sNet1
52V
2,
Xcep
ti
on
al
l
these
m
od
el
are
us
ed
he
re
in
t
his
w
ork.
Fo
r
trai
ning
purpos
es,
we
us
e
d
an
ou
t
pu
t
la
ye
r
of
1000
neur
on
s
.
De
pe
nd
i
ng
on
the
cl
ass
we
ha
ve
al
so
ad
de
d
the
outp
ut
la
ye
r
of
5
ne
uro
ns
as
w
e
hav
e
on
ly
5
cl
asses
into o
ur d
at
aset
[18].
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Reco
gn
it
io
n of
man
go leaf
dis
ease
us
in
g
c
on
volutio
na
l
ne
ural netw
or
k
model
s…
(
Adity
a Rajb
ongshi
)
1685
2.4
.
Tr
aining
a
n
d
te
stin
g
Fo
r
t
rainin
g
pu
rposes,
we
are
try
ing
to
trai
n
our
data
base
d
on
100
e
po
c
hs
as
we
are
try
ing
that
10
0
tim
es
tho
se
le
arn
i
ng
Algorithm
s
will
wo
r
k
throu
gh
the
e
ntire
trai
ning
data
set
.
Af
te
r
trai
nin
g
the
t
erm
of
te
sti
ng
is c
om
i
ng for gett
in
g
t
he
acc
ur
acy
.
In
this pr
opos
e
d work we
are
tr
yi
ng
to
test
the
n
e
wly t
raine
d m
od
e
l
with
the
te
st
da
ta
set
.
Af
te
r
t
est
ing
the
im
a
ges
data
,
5×
5
c
onf
us
io
n
m
at
ri
ces
are
f
or
m
ed
for
each
m
od
e
l.
The
n
we have
con
ve
rted
t
he 5×5
in bina
ry form
at
w
hic
h
is
pr
e
se
nted
i
n
T
a
ble
1.
Table
1.
C
onf
usi
on m
at
rices a
s b
i
nar
y
form
at
Metho
d
Clas
s
TP
FP
FN
TN
Den
seNet2
0
1
An
th
racno
se
59
1
1
239
Gall M
achi
58
1
2
239
Health
y
Leaf
59
2
1
238
Po
wd
ery
Mildew
59
1
1
239
Red
Ru
st
59
1
1
239
Incep
tio
n
V3
An
th
racno
se
56
1
4
239
Gall M
achi
57
2
3
238
Health
y
Leaf
59
3
1
237
Po
wd
ery
Mildew
59
4
1
237
Red
Ru
st
59
1
1
239
Res
Net5
0
An
th
racno
se
59
1
1
239
Gall M
achi
56
1
2
239
Health
y
Leaf
59
5
1
235
Po
wd
ery
Mildew
58
1
2
239
Red
Ru
st
59
1
1
239
Res
Net1
5
2
V2
An
th
racno
se
57
2
3
238
Gall M
achi
54
2
6
238
Health
y
Leaf
58
3
2
237
Po
wd
ery
Mildew
59
8
1
232
Red
Ru
st
56
1
4
239
Xcepti
o
n
An
th
racno
se
59
1
1
239
Gall M
achi
57
1
3
239
Health
y
Leaf
59
3
1
237
Po
wd
ery
Mildew
59
1
1
239
Red
Ru
st
59
1
1
239
Incep
tio
n
Res
NetV2
An
th
racno
se
58
1
2
239
Gall M
achi
56
2
4
238
Health
y
Leaf
58
4
2
236
Po
wd
ery
Mildew
59
2
1
238
Red
Ru
st
59
1
1
239
3.
Results
and
D
isc
ussion
Fo
r
e
xp
e
rim
e
ntal
analy
sis
of
each
m
od
e
l,
we
ha
ve
ut
il
iz
ed
seve
n
pe
rfor
m
ance
m
at
rices.
T
he
fo
ll
owin
g
e
qu
at
ion
s
a
re
us
e
d
to
eval
uate
t
he
se
ve
n
perform
ance
m
at
ric
es
su
c
h
a
s
acc
ur
acy
[19],
Pre
ci
sion
[20], F
1
sc
ore
[21],
se
ns
it
ivit
y
[
2
2],
s
pecific
it
y
[2
3],
FN
R
[
24
]
,
FP
R
[
2
5].
Accuracy
=
(
+
+
+
+
)
×1
00%
(5)
Pr
eci
sio
n
=
(
+
)
×100%
(6)
Sens
it
ivit
y =
(
+
)
×100%
(7)
F1
-
Score =
(
2
×
(
Pr
e
c
ision
×
S
e
n
si
tivity
(
Pr
e
c
isio
n
+
S
e
n
sit
ivity
)
)
×10
0%
(8)
Sp
eci
fici
ty
=
(
+
)
×100%
(9)
FN
R =
(
+
)
×100%
(10)
FPR =
(
+
)
×100%
(11)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2
02
1
:
16
81
-
16
88
1686
Table
2
e
xhibi
ts
the
pe
rfor
m
ance
e
valuati
on
m
at
rices
fo
r
the
De
ns
e
Net201
m
od
el
f
or
cl
ass
wise
(
ant
hr
ac
nose,
gall
m
achi,
healt
hy
le
af,
powd
e
ry
m
il
dew
,
and
red
ru
st
).
W
e
fin
d
from
Table
2
t
hat
th
e
accuracy
is
98.
00%
f
or
the
identific
at
io
n
of
diseases
of
m
ango
le
a
ves.
F1sco
re,
accuracy,
sen
s
it
ivit
y,
sp
eci
fici
ty
,
FNR
,
and
FPR
ha
ve
al
so
bee
n
m
easur
e
d.
T
he
hi
gh
est
F1
sc
or
e
fo
r
po
wd
e
ry
m
il
dew
,
red
r
ust
,
was
98.33%
.
The
highest
pr
eci
si
on
,
an
d
sensi
vity
fo
un
d
for
anthr
ac
nose,
pow
der
y
m
i
ldew,
an
d
re
d
r
us
t
,
is
98.33%
.
T
he
highest
s
pecifi
ci
ty
,
ob
se
rv
e
d
for
a
nthracn
os
e,
gall
m
ac
hi,
pow
der
y
m
il
dew
,
re
d
r
us
t
,
is
99.58%
.Th
e
l
owest
FN
R
a
nd
FPR
,
f
ound
f
or
anth
racnose,
pow
de
ry
m
il
dew
,
re
d
ru
st
,
is
1.67%
an
d
0
.
42%
as
well
.
Table
3
ex
hib
i
ts
the
per
f
or
m
ance
eval
uatio
n
m
a
tric
es
fo
r
the
In
ce
ptionR
esNetV
2
m
o
del
for
cl
ass
wise.
We
fin
d
from
Table
3
t
hat
the
acc
ura
cy
is
96.67%
f
or
t
he
i
den
ti
fic
at
ion
of
diseas
es
of
m
ang
o
le
aves
us
in
g
In
ce
ptio
nResNet
V2
m
od
el
.
T
he
high
est
F1
sc
or
e
a
nd
preci
sio
n
f
or
re
d
rust
,
wa
s
98.
33%.T
he
highest
sensiti
vity
,
ob
serv
e
d
for
gal
l
m
achi,
pow
der
y
m
il
dew
,
and
re
d
ru
st
i
s
98.
33%.
T
he
highest
s
pec
ific
it
y,
ob
s
er
ved
f
or
a
nthracn
os
e,
a
nd
red
r
us
t
is
99
.58%.T
he
lo
we
st
FN
R,
f
ound
for
pow
der
y
m
il
dew
,
an
d
re
d
Rust
is 1.6
7%.
For
he
al
thy l
eaf the
lowe
r
FPR
is
0.1
7%.
Table
2.
Cl
ass
wise
perform
a
nce
e
valuati
on
m
at
rices fo
r
Dense
Net2
01
Den
seNet2
0
1
Clas
s
Accurac
y
F1
Precisio
n
Sen
sitiv
ity
Sp
ecif
icity
FNR
FPR
An
th
racno
se
9
8
.00
%
9
8
.31
%
9
8
.33
%
9
8
.33
%
9
9
.58
%
1
.67
%
0
.42
%
Gall M
achi
9
7
.48
%
9
8
.30
%
9
6
.67
%
9
9
.58
%
3
.33
%
0
.42
%
Health
y
Leaf
9
7
.52
%
9
6
.72
%
9
8
.33
%
9
9
.17
%
1
.67
%
0
.83
%
Po
wd
ery
Mildew
9
8
.33
%
9
8
.33
%
9
8
.33
%
9
9
.58
%
1
.67
%
0
.42
%
Red
Ru
st
9
8
.33
%
9
8
.33
%
9
8
.33
%
9
9
.58
%
1
.67
%
0
.42
%
Table
3.
Cl
ass
wise
perform
a
nce
e
valuati
on
m
at
rices fo
r
In
cepti
on
Re
s
Net
V2
Incep
tio
n
Res
NetV2
Clas
s
Accurac
y
F1
p
recisio
n
Sen
sitiv
ity
Sp
ecif
icity
FNR
FPR
An
th
racno
se
9
6
.67
%
9
7
.48
%
9
8
.31
%
9
6
.67
%
9
9
.58
%
3
.33
%
0
.42
%
Gall M
achi
9
4
.92
%
9
6
.55
%
9
3
.33
%
9
9
.17
%
6
.67
%
0
.83
%
Health
y
Leaf
9
5
.08
%
9
3
.55
%
9
6
.67
%
9
8
.33
%
3
.33
%
0
.17
%
Po
wd
ery
Mildew
9
7
.52
%
9
6
.72
%
9
8
.33
%
9
9
.17
%
1
.67
%
0
.83
%
Red
Ru
st
9
8
.33
%
9
8
.33
%
9
8
.33
%
9
9
.58
%
1
.67
%
0
.42
%
Table
4
ex
hib
it
s
the
pe
r
form
a
nce
e
valuati
on
m
at
rices
fo
r
t
he
In
ce
ptio
nV3
m
od
el
fo
r
cl
as
s
wise.
We
fin
d
f
ro
m
Tab
le
4
that
the
accuracy
is
96.67%
f
or
the
i
den
ti
ficat
io
n
of
diseases
of
m
ang
o
le
aves
us
in
g
In
ce
ptio
nV3
m
od
el
.
F1
sc
ore,
accuracy,
se
ns
i
ti
vity
,
sp
eci
fici
ty
,
FN
R,
an
d
F
PR
hav
e
al
s
o
be
en
m
easur
ed.
The
highest
F
1
sco
re
f
or
red
rust
was
98.
3
3%.
The
highest
pr
eci
sion
f
ound
for
re
d
r
us
t
is
98.33%
.Th
e
hi
gh
e
st
sensiti
vity
obs
erv
e
d
for
healt
hy
le
af,
pow
der
y
m
il
dew
,
and
re
d
ru
st
i
s
98.
33%.
T
he
highest
s
pec
ific
it
y,
ob
s
er
ved
f
or
An
t
hr
ac
nose,
and
Re
d
Rust
is
99
.58%
.The
lowest
FN
R
,
found
f
or
he
al
thy
le
af,
po
w
de
ry
m
il
dew
, and re
d ru
st,
is
1.67%. F
or
ant
hr
ac
no
s
e
the
lo
wer
FPR is
0.42
%
.
Table
4.
Cl
ass
wise
perform
a
nce e
valuati
on
m
at
rices fo
r
In
cepti
onV
3
Incep
tio
n
V3
Clas
s
Accurac
y
F1
Precisio
n
Sen
sitiv
ity
Sp
ecif
icity
FNR
FPR
An
th
racno
se
9
6
.67
%
9
5
.73
%
9
8
.25
%
9
3
.33
%
9
9
.58
%
6
.67
%
0
.42
%
Gall M
achi
9
5
.80
%
9
6
.61
%
9
5
.00
%
9
9
.17
%
5
.00
%
0
.83
%
Health
y
Leaf
9
6
.72
%
9
5
.16
%
9
8
.33
%
9
8
.75
%
1
.67
%
1
.35
%
Po
wd
ery
Mildew
9
6
.72
%
9
5
.16
%
9
8
.33
%
9
8
.75
%
1
.67
%
1
.25
%
Red
Ru
st
9
8
.33
%
9
8
.33
%
9
8
.33
%
9
9
.58
%
1
.67
%
0
.42
%
Table
5
ex
hib
i
ts
the
per
f
or
m
ance
evaluati
on
m
a
tric
es
fo
r
the
Re
sNet50
m
od
el
fo
r
cl
a
ss
wise.
W
e
fin
d
f
ro
m
Tab
le
5
that
the
accuracy
is
97.00%
f
or
the
i
den
ti
ficat
io
n
of
diseases
of
m
ang
o
le
aves
us
in
g
Ra
sNet5
0
m
odel
.
F1sco
re,
ac
cur
acy
,
sen
sit
ivit
y,
sp
eci
fici
t
y,
FN
R,
an
d
F
PR
ha
ve
al
s
o
been
m
easur
ed
.
T
he
highest
F1
sc
ore
an
d
pr
ece
sion
f
or
anth
rac
no
s
e,
an
d
re
d
r
us
t
was
98.
33
%.
The
hi
gh
e
s
t
sensiti
vity
,
ob
ser
ve
d
for
a
nthracn
ose
,
gall
m
achi,
healt
hy
le
af
,
an
d
red
r
us
t
is
98.
33%.
T
he
highest
s
pe
ci
fici
ty
ob
ser
ve
d
for
anth
racnose,
ga
ll
m
achi,
pow
der
y
m
il
dew
,
and
re
d
r
us
t
is
99.58%
.
T
he
lowest
F
NR
f
ound
for
a
nthra
cnose,
healt
hy
le
af,
a
nd
r
ed
ru
s
t
is
1.6
7%.
F
or
a
nthracn
os
e, g
al
l
m
achi,
pow
de
ry
m
il
dew
,
an
d
re
d
r
us
t
the
lowe
r
FPR
is 0.4
2%.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Reco
gn
it
io
n of
man
go leaf
dis
ease
us
in
g
c
on
volutio
na
l
ne
ural netw
or
k
model
s…
(
Adity
a Rajb
ongshi
)
1687
Table
5.
Cl
ass
wise
perform
a
nce
e
valuati
on
m
at
rices fo
r
R
esNet5
0
Res
Net5
0
Clas
s
Accurac
y
F1
Precisio
n
Sen
sitiv
ity
Sp
ecif
icity
FNR
FPR
An
th
racno
se
9
7
.00
%
9
8
.33
%
9
8
.33
%
9
8
.33
%
9
9
.58
%
1
.67
%
0
.42
%
Gall M
achi
9
5
.73
%
9
8
.25
%
9
3
.33
%
9
9
.58
%
6
.67
%
0
.42
%
Health
y
Leaf
9
5
.16
%
9
2
.19
%
9
8
.33
%
9
7
.92
%
1
.67
%
2
.08
%
Po
wd
ery
Mildew
9
7
.48
%
9
8
.31
%
9
6
.67
%
9
9
.58
%
3
.33
%
0
.42
%
Red
Ru
st
9
8
.33
%
9
8
.33
%
9
8
.33
%
9
9
.58
%
1
.67
%
0
.42
%
Table
6
ex
hib
i
ts
the
pe
rfor
m
ance
e
valuati
on
m
at
rices
fo
r
the
Re
sNet
152V2
m
od
el
f
or
cl
ass
wise
.
We
fi
nd
f
r
om
Table
6
t
hat
th
e
accuracy
is
94.
67%
f
or
the
i
den
ti
ficat
io
n
of
diseases of
m
ango
le
ave
s.
F
1s
c
or
e
,
accuracy,
sensi
ti
vity
,
sp
eci
fici
ty
,
FN
R,
a
nd
FPR
ha
ve
al
s
o
bee
n
m
easur
ed.
T
he
highest
F1
sc
ore
f
or
he
al
thy
le
af
,
was
95.
87%.
The
highe
st
pr
e
ci
si
on,
found
f
or
re
d
rust
is
98.
25%.
The
highest
se
ns
it
ivit
y
obser
ved
f
or
pow
der
y
m
il
dew
,
a
nd
red
r
us
t
is
98.
33%.
T
he
hi
gh
est
sp
eci
fici
ty
ob
se
rv
e
d
f
or
re
d
rust
is
99.58%
.
T
he
l
ow
est
FN
R
fou
nd
f
or powde
ry m
i
ldew,
a
nd
red r
ust
is 1.6
7%.
For
re
d
r
us
t
,
the
lo
wer FPR is
0.4
2%.
Table
7
e
xhibit
s
the p
e
rfor
m
ance
e
valuati
on
m
at
rices
for
th
e
Xce
ptio
n
m
od
el
f
or
cl
ass
wi
se. W
e find
from
Table
7
t
hat
the
acc
urac
y
is
97.67%
f
or
the
i
den
ti
fica
ti
on
of
disease
s
of
m
ang
o
le
aves
us
in
g
Xce
ption
m
od
el
.
F1
sco
r
e,
accuracy,
se
ns
it
ivit
y,
sp
eci
fici
ty
,
FN
R,
and
F
PR
ha
ve
al
so
bee
n
m
ea
su
re
d.
T
he
hi
ghest
F
1
scor
e
,
pr
eci
sio
n
a
nd
s
ensiti
vi
ty
fo
r
a
nthrac
no
s
e,
pow
der
y
m
il
dew
,
an
d
red
r
us
t
was
98.
33%.
T
he
hig
he
st
sp
eci
fici
ty
,
obs
erv
e
d
f
or
ant
hracnose
,
gall
m
achi,
po
wd
e
ry
m
il
dew
,
an
d
re
d
r
us
t
is
99.
58
%.Th
e
l
ow
e
st
F
NR,
fou
nd
f
or
a
nthracn
os
e,
healt
hy
le
af,
po
wd
e
r
y
m
il
dew
,
an
d
red
r
us
t
is
1.67%.
F
or
a
nthracn
os
e,
gall
m
achi,
pow
der
y m
il
dew
, a
nd
red r
us
t
, th
e l
ow
e
r FP
R i
s 0
.
42%.
Table
6.
Cl
ass
wise
perform
a
nce e
valuati
on
m
at
rices fo
r
R
esNet1
52V
2
Res
Net1
5
2
V2
Clas
s
Accurac
y
F1
Precisio
n
Sen
sitiv
ity
Sp
ecif
icity
FNR
FPR
An
th
racno
se
9
4
.67
%
9
5
.80
%
9
6
.61
%
9
5
.00
%
9
9
.17
%
5
.00
%
0
.83
%
Gall M
achi
9
3
.10
%
9
6
.43
%
9
0
.00
%
9
9
.17
%
1
0
.00
%
0
.83
%
Health
y
Leaf
9
5
.87
%
9
5
.08
%
9
6
.67
%
9
8
.75
%
3
.33
%
1
.25
%
Po
wd
ery
Mildew
9
2
.91
%
8
8
.06
%
9
8
.33
%
9
6
.67
%
1
.67
%
3
.33
%
Red
Ru
st
9
5
.73
%
9
8
.25
%
9
3
.33
%
9
9
.58
%
6
.67
%
0
.42
%
Table
7.
Cl
ass
wise
perform
a
nce e
valuati
on
m
at
rices fo
r
Xce
ption
Xcepti
o
n
C
lass
Ac
c
u
rac
y
F1
P
rec
is
ion
S
e
n
si
t
iv
i
ty
S
p
e
c
if
icity
F
NR
FPR
An
th
racno
se
9
7
.67
%
9
8
.33
%
9
8
.33
%
9
8
.33
%
9
9
.58
%
1
.67
%
0
.42
%
Gall M
achi
9
6
.61
%
9
8
.28
%
9
5
.00
%
9
9
.58
%
5
.00
%
0
.42
%
Health
y
Leaf
9
6
.72
%
9
5
.16
%
9
8
.33
%
9
8
.75
%
1
.67
%
1
.25
%
Po
wd
ery
Mildew
9
8
.33
%
9
8
.33
%
9
8
.33
%
9
9
.58
%
1
.67
%
0
.42
%
Red
Ru
st
9
8
.33
%
9
8
.33
%
9
8
.33
%
9
9
.58
%
1
.67
%
0
.42
%
4.
CONCL
US
I
O
N
The
disco
ve
ries
of
this
resea
r
ch
dem
on
strat
e
that
disease
disti
nguish
i
ng
pro
of
from
the
i
m
age
with
the
co
nvol
utio
nal
ne
ur
al
net
work
is
a
s
olid
strat
e
gy
f
or
high
e
xactness
autom
at
ed
dis
ti
ng
uis
hing
proof
of
m
ang
o
disease.
In
this
pro
po
s
ed
w
ork,
six
C
NN
m
od
el
s
are
util
iz
ed
fo
r
fiv
e
un
iq
ue
cl
asse
s
of
m
ang
o
dis
ease.
Im
age
processi
ng
te
c
hniq
ues
hav
e
bee
n
us
e
d
f
or
the
a
ug
m
entat
ion
of
the
i
m
age
data.
T
he
pre
-
trai
ne
d
tr
ansf
e
r
le
arn
in
g
te
ch
ni
qu
e
has
bee
n
util
iz
ed
her
e
.
Af
te
r
c
om
pletio
n
of
trai
ning
and
te
sti
ng
the
i
m
a
ge
data,
a
5×5
conf
us
io
n
m
atr
ix
has
bee
n
ge
ner
at
e
d.
Ba
se
d
on
the
eval
ua
ti
on
of
seve
n
per
f
orm
ance
m
et
rics,
the
hig
he
st
accuracy
is
98.
00%
fou
nd
f
or
the
Den
se
Net2
01
m
od
el
.
The
m
ai
n
fo
cus
s
hould
be
pointe
d
towa
rd
set
tl
ing
few
disad
va
ntages
in
the
pro
pose
d
m
et
ho
dolo
gy
al
so
inc
rease
so
m
e
m
or
e
de
visin
g
m
et
ho
ds
in
this
w
ork.
So
i
n
the
fu
t
ur
e,
we
will
try
to
fo
cus
on
m
or
e
diseases
with
thi
s
CNN
te
ch
niq
ue
to
get
bette
r
accur
acy
for
tho
s
e
ta
rg
et
ed
d
ise
as
es.
REFERE
NCE
S
[1]
J.
Bai
,
B.
W
ang
,
C.
Chen,
J.
C
hen,
and
Zhong
-
Hu
a
Fu
.
,
"Inception
-
v3
Based
Method
of
Li
fe
CLE
F
2019
Bird
Rec
ognition,"
In
CL
EF
(
Working
Note
s)
,
2019.
[2]
P.
Dhankha
r,
"R
esNet
-
50
and
V
GG
-
16
for
rec
ogniz
ing
f
ac
i
al
e
m
oti
ons,"
Inte
rn
ati
onal
Journal
of
Innov
ati
ons
i
n
Engi
ne
ering
and
Technol
og
y
(
IJI
ET)
,
vol.
13
,
no
.
4,
pp
.
126
-
130
,
2019
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2
02
1
:
16
81
-
16
88
1688
[3]
C.
Li
n
,
G.
Zh
ao, Z
.
Yang
,
A.
Yin
,
and
X.
W
ang
,
"Cir
-
net
:
Autom
a
ti
c cl
assifi
ca
t
ion of human
chr
om
osom
e
base
d
on
inc
ep
ti
on
-
resne
t
arc
hi
te
c
ture,"
IE
EE
/A
CM
Tr
ansacti
ons
on
Computati
onal
Bi
o
log
y
and
Bi
oin
formatic
s,
2020
,
doi
:
10.
1109/T
CBB
.
2020.
3003445.
[4]
C.
Li
n,
L.
L
i,
W
.
Luo,
K.
C.
P.
W
ang,
and
J.
Guo
,
"Tra
nsfer
l
ea
rn
ing
base
d
tra
ff
ic
sign
rec
ognition
using
inc
eption
-
v3
m
odel
,
"
Pe
r
iodi
ca
Polytech
nic
a
Tr
anspor
tat
ion
Eng
ineeri
ng,
vol
.
47
,
no
.
3,
pp.
242
-
2
50,
2019,
doi
:
10.
3311/PP
tr.
11
480.
[5]
C.
A.
Ferre
ir
a,
T.
Me
lo,
P.
Sou
sa,
M.
Ine
s
Me
y
er
,
and
A.
Ca
m
pil
ho
,
"Classif
ic
a
ti
on
of
bre
ast
ca
n
ce
r
histol
o
g
y
images
through
tra
nsfer
learni
ng
using
a
pre
-
tra
i
ned
in
ce
pt
ion
r
e
snet
v2,
"
In
201
8
Inte
rnat
ional
Confe
renc
e
Ima
ge
Anal
ysis
and
R
e
cogni
ti
on
,
2018
,
pp.
763
-
770
,
doi
:
10.
1007
/978
-
3
-
319
-
93000
-
8_86.
[6]
M.
M.
R
ahman,
et
a
l.
,
"
Re
cogni
t
i
on
of
Lo
cal
Bird
s of
Banglade
sh
using Mobil
eNe
t
and
In
ce
p
ti
on
-
v
3,
"
Inte
rnational
Jo
urnal
of Adv
an
c
ed
Computer
S
cienc
e
and
Appl
i
c
ati
ons
,
vo
l
.
11,
n
o.
8
,
pp
.
309
-
31
6,
2020
,
doi
:
10.
14569/IJACS
A.2020.
0110840
.
[7]
I.
Z.
Mukti,
an
d
D.
Biswas,
"
Tra
nsfer
learni
n
g
base
d
pla
nt
disea
ses
det
e
ct
io
n
using
ResNet50.
"
In
2019
4th
Inte
rnational
Co
nfe
renc
e
on
El
e
ct
rical
Informati
on
and
Comm
un
ic
ati
on
Techno
lo
gy
(
EICT
)
,
2019
,
pp.
1
-
6,
IEEE,
doi:
10
.
1109/E
I
CT48899.
2019.
9
068805.
[8]
I.
Sharm
in,
N.
F.
Islam,
I
.
Jah
an,
T.
A.
Jo
y
e
,
and
Md.
Riaz
ur
Rahman
,
"M
ac
hin
e
vision
b
ase
d
lo
ca
l
fish
rec
ogni
ti
on,
"
SN
Applied
S
cienc
e
s,
vol. 1, no. 12,
pp.
1
-
12
,
2019
,
doi
:
10
.
1007/s42
452
-
019
-
1568
-
z.
[9]
O.
Praka
sh
and
A.
Mis
ra,
"
Im
po
rta
nt
disea
s
es
of
m
ango
and
the
ir
eff
ec
t
on
produc
ti
on
,
"
Bi
ol
Me
moirs
,
vol.
18,
pp
.
39
-
55,
1992
.
[10]
K.
N.
Shukla
et
al.
,
"
A
rev
ie
w
on
image
enhance
m
ent
techniq
ues,
"
Inte
rnatio
nal
Journal
of
Engi
ne
ering
and
Appl
ie
d
Comput
er
Scienc
e
(
IJE
A
CS)
,
vol.
2
,
no
.
7
,
pp
.
232
-
235
,
2
017
.
[11]
T.
Lei
,
X.
Jia
,
Y.
Zha
ng,
L
.
He
,
and
H.
Meng
.
,
"
Signifi
ca
nt
l
y
f
ast
and
robust
fuz
z
y
c
-
m
ea
ns
clus
te
ring
al
gor
it
h
m
base
d
on
m
orph
ologi
c
al
rec
onstr
uct
ion
and
m
embership
fil
t
eri
ng
,
"
IEE
E
Tr
ansacti
ons
on
Fuzzy
S
yste
ms
,
vol.
26
,
no.
5
,
pp
.
3027
-
3041,
2018
,
doi
: 10.1109/
TFUZZ
.
2018.
2796074
.
[12]
J.
M.
Pate
l
and
N.
C.
Gam
it
,
"
A
rev
ie
w
on
fea
ture
ex
tracti
on
te
chni
qu
es
in
cont
ent
base
d
image
ret
r
ie
v
al
,
"
i
n
Inte
rnational
C
onfe
renc
e
on
Wirel
ess
Comm
unic
ati
ons
,
Signal
Proce
ss
ing
and
Net
working
(
WiSP
NET)
,
IEEE,
2016,
pp
.
2259
-
2263
.
[13]
R.
M.
Hara
li
ck
,
et
al.
,
"
T
ext
ur
al
feature
s
for
image
cl
assificat
ion,
"
IEE
E
Tr
ansacti
ons
on
systems,
man,
and
cy
berne
ti
cs
,
no
.
6,
pp
.
610
-
621
,
1973,
doi
:
10
.
11
09/T
SM
C.
1973.
4309314.
[14]
Y.
Hu,
Chun
-
x
ia
Zha
o
,
and
Hong
-
nan
W
ang,
"
Dir
ec
t
iona
l
an
aly
sis
of
t
ext
ur
e
imag
es
usinggra
y
le
v
el
co
-
oc
cur
ren
c
e
m
at
rix,
"
in
IE
EE
Pac
ific
-
Asia
workshop
on
computat
ional
int
e
ll
ig
enc
e
and
industrial
applicati
on
,
vol.
2,
I
EE
E
,
pp
.
277
-
281,
2008
,
doi:
10
.
1109/PACIIA
.
2008.
279.
[15]
D.
L.
Padm
aja
a
nd
B.
Vishnuvar
dhan,
"
Com
par
a
ti
ve
stud
y
of
fe
a
ture
subs
et
se
le
c
ti
on
m
et
hods
for
dimensional
i
t
y
red
uction
on
scie
nti
fi
c
data,
"
in
IEE
E
6th
Inte
rn
ati
onal
Confe
re
nce
on
Adv
anc
e
d
Computing
(
I
ACC)
,
IEE
E
,
20
16,
pp.
31
-
34
,
doi
:
1
0.
1109/IACC.
20
16.
16.
[16]
V.
Agarwal
,
"Resea
rch
on
dat
a
pre
pro
c
essing
and
c
a
te
gori
za
t
ion
t
e
chni
que
for
s
m
art
phone
rev
i
ew
ana
l
y
sis,"
Inte
rn
ati
onal
Journal
of
Comput
e
r
Appl
ic
a
ti
ons
,
vol.
131
,
n
o.
4,
pp.
30
-
36
,
2015
,
doi
:
10.
5120/i
j
ca
201
5907309.
[17]
A.
Mikoła
j
czy
k
,
and
M.
Groch
ows
ki,
"D
at
a
au
gm
ent
at
ion
for
improving
dee
p
le
arn
in
g
in
image
c
la
ss
ifi
c
at
io
n
proble
m
,
"
In
20
18
int
ernati
onal
int
erdisci
p
li
nary
PhD
workshop
(
IIP
hDW
)
,
pp.
117
-
122.
I
EE
E
,
2018,
doi
:
10.
1109/IIPHD
W
.
2018.
8
388338.
[18]
S.
Yu,
C.
Xu
,
and
S.
Jia,
"
Convolut
ional
neur
al
n
etw
orks
for
h
y
per
sp
ec
tr
al
image
cl
assifi
ca
t
ion,
"
Neurocomputi
ng
,
vol
.
219
,
pp
.
88
-
98,
2017
,
doi
:
1
0.
1016/j.ne
u
com.2016.
09
.
010.
[19]
A.
A.
Biswas,
A.
Majumder,
A.
Raj
bongshi,
and
Md.
M.
Rahman
,
"Rec
ogni
ti
on
of
Loc
a
l
Bi
rds
using
Diffe
ren
t
CNN
Archi
te
ct
u
res
with
Tra
nsfe
r
Le
arn
ing,
"
In
2021
Inte
rnatio
nal
Confe
ren
ce
on
Computer
Comm
unic
ati
on
a
nd
Informatic
s (
ICCCI)
,
2021,
pp
.
1
-
6,
IE
EE,
do
i:
1
0.
1109/ICCCI50
826.
2021.
94026
86.
[20]
M.
Hos
sin,
and
M.
N.
Sula
iman
,
"A
rev
ie
w
on
e
val
ua
ti
on
m
et
r
ics
for
dat
a
class
ifi
cation
ev
al
u
at
io
ns,"
Inte
rnati
ona
l
Journal
of
Data
Mini
ng
&
Knowl
edge
Manag
eme
nt
Proc
ess,
vol
.
5,
no
.
2
,
2015
,
d
oi:
10
.
5121/ijdk
p.
2015.
5201
.
[21]
A.
Majumder,
Al
A.
Biswas,
Md.
M.
Rahman,
and
A.
Raj
b
ongshi
.,
"Lo
ca
l
Freshw
at
er
Fis
h
Rec
ognition
Us
ing
Diffe
ren
t
CNN
Archi
tectur
es
wi
th
Tra
nsfer
Lear
ning,
"
Inte
rnati
onal
Journal
on
Adv
ance
d
Sc
ience
,
Eng
ineering
and
Information T
ec
hnology
,
vo
l.
11,
no.
3,
pp.
10
78
-
1083,
2021
,
doi:
10
.
18517/ij
ase
it.11.
3
.
14134
.
[22]
N.
Gu
pta
,
et
al
.
,
"A
c
cur
acy
s
e
nsiti
vity
and
sp
ec
if
ic
i
t
y
m
ea
sur
ement
of
v
ari
o
us
cl
assificat
ion
te
chn
ique
s
on
hea
l
thc
ar
e
d
ata,
"
IOSR Journal
of
Computer
Eng
i
nee
ring (
IOSR
-
J
CE)
,
vol. 11, no. 5, pp. 70
-
73,
20
13
.
[23]
C
.
S
a
m
m
u
t
,
a
n
d
G
e
o
f
f
r
e
y
I
.
“
W
e
b
b
,
e
d
s
.
E
n
c
y
c
l
o
p
e
d
i
a
o
f
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
,
”
S
p
r
i
n
g
e
r
S
c
i
e
n
c
e
&
B
u
s
i
n
e
s
s
M
e
d
i
a
,
2
0
1
1
.
[24]
A.
Raj
bongshi,
T.
Sarke
r
,
Md.
M.
Aham
ad,
Md.
And
M.
Rahman,
"Ros
e
Disea
ses
Rec
ognit
ion
using
Mobile
Net
,
"
In
2020
4th
Inter
nati
onal
Sympo
sium
on
Mult
id
isci
pli
nary
Studies
and
Innov
a
ti
v
e
Technol
ogi
es
(
ISMSIT)
,
pp.
1
-
7.
IEE
E
,
2020
,
do
i: 10.
1109/ISMS
IT50672.
2020.
92
54420.
[25]
K.
M.
Ghori
,
M.
Im
ran
,
R
.
A
.
Abbasi,
A.
Ul
la
h,
and
A
.
Na
waz
,
"P
erf
orm
a
nce
ana
l
y
sis
of
m
ac
hine
learni
ng
cl
assifi
ers
for
non
-
te
chn
ic
a
l
loss
det
e
ct
ion
,
"
Jour
nal
of
Ambi
ent
I
nte
lligen
ce
and
Hum
anized
Com
puti
ng,
pp.
1
-
1
6
,
2020,
doi
:
10
.
10
07/s12652
-
019
-
01649
-
9.
Evaluation Warning : The document was created with Spire.PDF for Python.