Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
10
,
No.
3
,
June
201
8
,
pp.
1098
~
1105
IS
S
N:
25
02
-
4752
, DO
I: 10
.11
591/
ijeecs
.
v
10
.i
3
.pp
1098
-
1105
1098
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Compre
hensiv
e Pin
ea
pp
le Segm
en
tatio
n
Tec
hn
iqu
es
with
Intellig
en
t Co
n
voluti
onal Neu
ra
l
Net
work
Muhamm
ad
Az
mi
A
hmed
Na
w
awi
,
F
at
i
mah
Sh
am
I
s
mail
,
H
az
li
na
Selam
at
Facul
t
y
of Electr
ic
a
l
Eng
ineeri
ng
,
Univer
si
ti T
ekn
ologi
Ma
lay
sia
,
81310
Skudai, Johor,
Mal
a
y
sia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n
15
, 2
01
8
Re
vised
Ma
r
12
, 2
01
8
Accepte
d
Ma
r
2
8
, 201
8
Thi
s
pape
r
prop
oses
an
int
el
li
g
e
nt
segm
ent
at
ion
te
chn
ique
for
pi
nea
ppl
e
fruit
using
Convolut
i
onal
Neur
al
N
et
work
(CNN
).
Casca
d
e
Obje
ct
Det
ector
(COD
)
m
et
hod
is
used
to
det
e
ct
the
posit
ion
of
the
pineappl
e
from
the
ca
ptur
ed
image
b
y
ret
urn
ing
the
bounding
b
ox
aro
und
the
det
e
ct
in
g
pine
app
le
.
Im
age
bac
kg
round
such
as
ground,
sk
y
and
othe
r
unwant
e
d
obje
c
ts
have
b
ee
n
removed
u
sing
Hue
val
u
e
,
Adapti
v
e
Red
and
Blu
e
Chrom
at
ic
Map
(ARB)
and
Norm
al
iz
ed
Dif
fer
enc
e
Ind
ex
(ND
I)
m
et
hods.
How
eve
r,
the
A
RB
and
ND
I
m
e
thods
are
stil
l
pr
oduci
ng
m
iscl
as
sifie
d
err
or
and
the
edge
is
not
reall
y
sm
ooth.
In
thi
s
c
ase
Te
m
pla
t
e
Matc
h
ing
Method
(TMM)
has
bee
n
implemente
d
f
or
image
enha
n
ce
m
ent
proc
ess.
Final
l
y
,
an
int
ellige
n
t
CNN
is
dev
el
op
ed
as
a
d
ecision
m
ake
r
to
selec
t
th
e
best
segm
e
nta
ti
on
image
oupu
t
from
ARB
and
ND
I.
The
r
esult
s
obt
a
ine
d
show
tha
t
the
propose
d
int
el
l
ige
nt
m
ethod
has
succ
ess
fully
v
eri
fi
ed
th
e
fruit
from
the
b
ac
kground
with
high
a
cc
ura
c
y
as
compare
d
t
o
the c
onv
ent
ion
al
m
et
hod
.
Ke
yw
or
d
s
:
Ca
scade
obj
ect
d
et
ect
or
Conv
olu
ti
onal
neural
netw
ork
Pineap
ple se
gm
entat
ion
Tem
plate
m
at
c
hing m
et
ho
d
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Fatim
ah
Sh
am
I
sm
ail,
Faculty
of Elec
tric
al
Engineer
ing
,
Un
i
ver
sit
i Te
knol
og
i M
al
ay
sia
,
81310, S
kudai,
Johor,
Mal
ay
sia
.
Em
a
il
: fatim
ahs@
utm
.
m
y
1.
INTROD
U
CTION
Pineap
ple
is
a
tro
pical
pla
nt
from
Brom
el
i
acea
e
fam
ily
and
one
of
the
m
os
t
con
s
umpti
on
f
ru
it
s
arou
nd
t
he
world
a
fter
ban
a
na
a
nd
ci
tr
us
.
The
gove
rn
m
ent
un
der
Ea
st
Coast
Eco
nom
ic
s
Re
gion
(E
CER
)
pro
gr
am
has
al
locat
ed
7400
hectares
la
nd
f
or
pin
ea
pp
le
c
ulti
vation
in
orde
r
to
inc
reas
e
the
pro
du
ct
i
on
of
pin
ea
pp
le
[
1].
Pineap
ple
gro
ws
unde
r
s
warm
c
lim
a
te
,
op
t
i
m
u
m
water
a
nd
no
i
nf
est
at
i
on.
Howe
ver,
i
t
ta
kes
about
24
to
39
m
on
ths
to
yi
el
d
fruit
befor
e
it
can
be
harvested
.
More
over
,
to
identif
y
the
har
ve
sti
ng
tim
e
beco
m
e
a
qu
it
e
chall
eng
i
ng.
This
is
becau
se
the
qu
al
it
y
and
sweetness
of
the
pin
ea
pple
are
de
pende
nt
on
th
e
tim
e
of
harvesti
ng
,
if
t
oo
ea
rly
will
cause
pi
neapple
to
bec
om
e
le
ss
sweet
,
m
eanw
hile,
ha
rv
est
it
too
la
te
will
cause
pin
ea
pple
to
beco
m
e
too
j
uicy
[
2]
.
Ther
e
fore,
pin
ea
pp
le
m
us
t
be
ha
rv
e
ste
d
be
f
or
e
it
be
com
e
ov
e
rm
at
ur
ed.
Currentl
y,
m
ac
hin
e
visi
on
sys
tem
can
pr
ov
i
de
a
huge
ad
va
ntage
in
hel
pin
g
fa
rm
ers
to
identify
the
fruit
s’
m
at
ur
it
y
an
d
the
optim
al
tim
e
fo
r
ha
r
vestin
g.
Howe
ver,
the
m
ai
n
pro
blem
that
lim
it
s
the
capabi
li
t
y
of
m
achine v
isi
on
syst
e
m
is
i
m
a
ge
recog
niti
on
process and
ob
j
ect
d
et
ect
ion
e
sp
eci
al
ly
in
detect
ing
the p
ine
app
le
i
m
age.
Ther
e
a
re
m
a
ny
research
es
ha
ve
bee
n
co
nducte
d
to
address
this
pro
blem
.
Moh
am
m
ad
et
al
.
[2
]
us
e
d
an
im
age
pr
oc
essing
te
c
hn
i
que
to
cat
eg
or
i
ze
N36
pi
neapple.
By
us
in
g
bin
a
ry
el
li
ps
e
m
ask
tog
et
he
r
with
m
or
phologica
l
norm
al
iz
ed
RGB
to
filt
er
ou
t
the
backgro
und,
the
re
searc
h
s
howe
d
a
prom
os
ing
resu
lt
.
Howe
ver,
the
us
e
d
of
bin
a
ry
el
li
ps
e
m
ask
has
it
s
own
dr
a
wb
ac
k
w
her
e
t
he
se
gm
ented
r
egio
n
did
no
t
i
nclu
de
the
entire
reg
i
on
of
intere
st.
The
m
ini
m
um
s
ymm
et
rica
l
distance
m
eth
od
has
bee
n
pro
po
se
d
by
[3
]
to
ov
e
rc
om
e
the
pro
blem
m
e
nti
on
e
d
i
n
a
bove.
The
a
utho
r
m
a
nag
e
d
t
o
se
gme
nt
the
e
ntire
r
egio
n
of
i
ntere
st
from
the b
ac
kgr
ound. H
ow
e
ve
r,
t
he
prop
os
ed
m
eth
od
only
work
s u
si
ng sym
m
e
tric
al
sh
ap
e
pine
app
le
only
.
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
Compre
hensi
ve Pi
ne
apple
Se
gm
e
nta
ti
on
Te
chn
i
qu
e
s
wi
th
I
ntell
igent
.
..
(
Mu
hamma
d
Az
mi A
hm
e
d
N
aw
awi
)
1099
Kam
aru
ddin
et
al
.
[
4]
ha
ve
c
ho
s
en
Otsu’s
Me
thod
as
the
ir
i
m
age
segm
entat
ion
t
ec
hn
i
qu
e
be
fore
app
ly
in
g
m
or
phol
og
ic
al
proc
ess.
Otsu’s
Me
thod
ass
um
ing
the
im
age
con
t
ai
ns
tw
o
cl
ass
e
s
of
pix
el
s
f
ollow
i
ng
bi
-
m
od
al
histo
gr
am
(f
ore
ground
an
d
backg
r
ound)
a
nd
it
ca
lc
ulate
s
the
opt
i
m
u
m
threshold
se
par
at
in
g
t
he
two
cl
asses.
Pr
a
bha
and
K
um
ar
[5
]
al
so
us
e
d
th
re
sh
ol
d
m
et
ho
d
t
o
se
gm
ent
the
ban
a
na
from
the
black
bac
kgr
ound.
Howe
ver,
thre
sh
ol
d
m
et
ho
d
on
ly
w
orks
e
ffi
ci
ently
if
the
i
m
age
con
ta
ins
two
c
olors
on
ly
.
So
,
the
acc
ur
ac
y
will
gr
a
du
al
ly
decr
ease
i
f
the
i
m
age
con
ta
in
s
m
or
e
tha
n
two
c
olors.
T
he
n
the
us
ed
of
g
-
r
gray
i
m
age
was
pro
po
se
d by
[6]
to
rem
ov
e t
he
b
ac
kgr
ound fr
om
the i
m
age.
Ba
sed
on
the
li
te
ratur
e
re
view,
it
was
found
that
to
extra
ct
fr
uit
from
back
gr
ound
at
t
he
plantat
io
n
le
vel
is
ver
y
diff
ic
ult
an
d
a
fe
w
facto
rs
nee
d
to
be
co
ns
ide
r
ed.
Fi
rstly
,
the
al
go
rithm
m
us
t
be
rob
us
t
for
both
i
m
m
at
ur
e
an
d
m
at
ur
e
fruit
s.
Th
us
,
m
os
t
of
t
he
resea
rc
her
s
on
ly
f
oc
us
in
g
on
m
at
ur
e
f
ru
it
because
the
in
te
ns
ity
betwee
n
m
at
ur
e
fr
uit
an
d
bac
kgr
ound
are
easi
ly
to
disti
ng
uish
e
d.
On
the
oth
er
hand,
th
e
intensit
y
bet
wee
n
i
m
m
at
ur
e
fruit
and
bac
kgr
ou
nd
is
al
m
os
t
t
he
sam
e.
Ther
efore,
to
a
pp
l
y
the
al
go
rith
m
on
i
m
m
a
ture
fr
uit
cannot
be
us
e
d
with
ou
t
m
od
ify
in
g
the
pa
ram
et
ers.
Secon
dly,
the
al
go
rithm
a
lso
m
us
t
be
robu
st
a
gainst
natu
ral
il
lum
inati
on
.
T
hir
dly,
the
f
ru
it
in
pl
an
ta
ti
on
m
os
tl
y
cl
us
te
r
to
gether
or
occl
us
i
on
with
oth
e
r
fruit
s,
le
aves
or
br
a
nc
hes.
T
her
e
f
ore,
to
i
m
pr
ove
the
rob
us
tness
of
the
al
gorit
hm
,
research
e
r
s
hav
e
c
om
bin
ed
the
process
of
c
olo
r
segm
entat
ion
with
oth
e
r
te
chn
i
qu
e
s
s
uch
as
featu
re
e
xtr
act
ion
,
e
dges
de
te
ct
ion
,
blob
-
base
d
segm
entat
ion
a
nd o
t
her
s
[7]
-
[
10
]
.
2.
RESEA
R
CH MET
HO
D
Im
age
segm
ent
at
ion
is
a
ver
y
i
m
po
rtant
pr
oc
ess
to
rem
ov
e
the
un
wan
te
d
ba
ckgr
ound
fro
m
the
m
ai
n
obj
ect
in one
i
m
age.
I
n
this
c
ase
is
pin
ea
pple
fruit
,
in o
r
der
to
ens
ure
that t
he
outp
ut
im
a
ge
ha
s
hi
gh
ac
cur
acy
.
The
process
st
arts
by
the
data
acqu
isi
ti
on,
al
l
the
i
m
ages
go
ing
to
Ca
scade
Object
Detect
or
(CO
D)
t
o
c
onfirm
the
fruit
’s
l
oca
ti
on
be
f
or
e
going
t
o
H
ue
Val
ue
Se
gm
entat
i
on
(
HV
S
)
proc
ess.
T
he
ove
ra
ll
pr
oce
dure
for
this
process
is
s
ho
wn
i
n
Fi
gure
1.
T
he
refor
e
,
to
m
ake
su
re
t
he
un
wan
te
d
ba
ckgr
ound
is
s
uccess
fu
ll
y
re
m
ov
ed
from
the
i
m
age,
this
wo
r
k
ha
s
dev
el
ope
d
a
Con
voluti
onal
Neu
ral
Net
work
(C
NN)
as
a
searching
en
gi
ne
to
evaluate
the
re
su
lt
s
from
Ad
aptive
Re
d
an
d
Bl
ue
Chrom
ati
c
Ma
p
(A
RB
)
and
Norm
al
iz
e
d
Dif
fer
e
nce
I
nd
e
x
(NDI).
Figure
1.
The
ov
e
rall
pro
c
ess
2.1. D
ata Acq
uisi
tion
Data
aqu
isi
ti
on
fo
r
im
age
has
been
ca
ptured
durin
g
day
tim
e
with
var
io
us
il
lu
m
inati
on
con
diti
ons
at
pin
ea
pp
le
plant
at
ion
at Pe
ka
n
Nan
a
s,
P
onti
an
, J
oh
or. T
he
ca
ptured
im
ages
hav
e
b
ee
n proc
essed
a
nd f
i
nally
all
these
datas
are
us
ed
to
de
velo
p
an
al
gorit
hm
fo
r
pin
ea
pple
i
m
age
segm
ent
at
ion
te
ch
niqu
e.
In
this
wor
k,
MD2
ty
pe
pin
ea
pple
is
sel
ect
ed
ba
sed
on
t
he
s
uggestio
n
f
ro
m
Ma
la
ysi
a
Pine
app
le
Ind
us
try
Boar
d
(MP
IB
),
wh
ic
h
on
e
of
the
m
os
t
popula
r
an
d
highest
dem
and
i
n
Ma
la
ysi
a.
Fig
ur
e
2
s
hows
t
he
sam
ple
i
m
ages
us
ed
in
this
pro
j
ect
.
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,
Vol
.
10
, N
o.
3
,
June
201
8
:
1089
–
1105
1100
Figure
2. Sam
ple i
m
ages w
it
h diff
e
re
nt stage
of m
at
ur
it
y and
backg
rou
nd
2.2. Alg
orit
h
m
Figure
3
sho
w
s
the
flow
c
har
t
of
the
pro
pose
d
al
gorithm
fo
r
segm
entat
ion
pr
oce
ss
an
d
de
te
ct
ing
the
pin
ea
pp
le
obje
ct
.
Since
the
i
m
age
us
ed
in
this
pro
j
ect
consi
sts
of
pin
ea
pp
le
a
nd
backg
rou
nd,
i
m
age
segm
entat
ion
pro
ces
s is im
ple
m
ented
to
rem
ov
e
an
y
unwa
nt
ed
bac
kgr
ound
from
the f
oc
us
im
age.
Figure
3. Flo
w
char
t
for
t
he pr
opos
e
d
al
gorithm
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
Compre
hensi
ve Pi
ne
apple
Se
gm
e
nta
ti
on
Te
chn
i
qu
e
s
wi
th
I
ntell
igent
.
..
(
Mu
hamma
d
Az
mi A
hm
e
d
N
aw
awi
)
1101
The
pro
posed
al
gorithm
sta
rts
with
data
acq
uisit
ion
,
by
ac
qu
i
rin
g
i
m
age
from
the
ca
m
e
ra.
Since
the
siz
e
of
t
he
cap
ture
d
im
age
is
too
la
r
ge,
t
his
pro
j
ect
has
c
onduct
ed
a
f
e
w
te
sts
to
fi
nd
ou
t
s
uitable
s
cal
e
to
resize
the
dim
e
ns
io
n
of
the
i
m
age.
By
reducing
the
dim
ensi
on
of
the
im
age,
the
com
pu
ta
ti
on
al
tim
e
has
bee
n
reduce
d.
A
fter
that,
the
pro
pose
d
al
gorithm
beg
i
ns
to
rem
ov
e
t
he
bac
kgr
ound
in
the
im
age
by
doin
g
i
m
age
segm
en
ta
ti
on
.
Firstl
y,
Ca
scad
e
O
bj
ect
Detec
tor
(CO
D)
is
im
ple
m
ented
to
detect
the
l
oca
ti
on
of
t
he
pine
app
le
in
the
i
m
age.
By
retur
ni
ng
t
he
boun
ding
box
a
rou
nd
the
detect
ing
pi
ne
app
le
,
t
he
pr
opos
e
d
al
gorit
hm
wil
l
rem
ov
e
al
l
i
m
age
(
pix
el
)
ou
tsi
de
this
boundi
ng
box
.
By
do
i
ng
t
his,
th
e
al
gorithm
ca
n
f
ur
t
her
redu
ce
the
com
pu
ta
ti
on
al
tim
e and
inc
re
asi
ng the acc
ur
acy
f
or the
n
e
xt
p
r
ocess
.
Seco
nd
ly
,
t
he
Hu
e
in
Hu
e
,
Saturati
on
a
nd
Value
(
HSV)
colo
r
s
pace
i
s
extracte
d
a
nd
by
set
ti
ng
thres
ho
l
d
valu
e
ob
ta
ine
d
fro
m
the
exp
eri
m
ental
te
st
to
detect
gr
ou
nd
an
d
sk
y.
Si
nce
the
furthe
r
i
m
age
segm
entat
ion
is
quit
e
dif
ficul
t,
it
is
necessa
r
y
to
rem
ov
e
a
ny
backg
rou
nd
that
is
easi
ly
to
disti
nguis
hed
from
the
f
ru
it
.
The
Hu
e
val
ue
for
gro
und
(ora
nge)
a
nd
sk
y
(wh
it
e)
are
easi
ly
diff
e
re
ntia
te
d
f
ro
m
oth
er
col
ors.
T
he
thres
ho
l
d
valu
e
has
bee
n
sel
e
ct
ed
ba
sed
on
t
hese
pro
per
ti
es
.
A
ny
value
le
ss
tha
n
th
res
hold
val
ue
is
rem
ov
e
d
(p
i
xel
intensit
y
beco
m
es
0)
,
m
eanw
hile,
the
value
gr
eat
er
or
eq
ual
th
an
thres
hold
va
lue
is
m
a
intai
ned.
In
i
m
age
segm
ent
at
ion
,
th
ere
a
re
of
te
n
pro
blem
wh
e
re
s
om
e
i
m
ages
only
w
ork
well
us
in
g
certai
n
m
et
ho
d
an
d
oth
e
r
im
ages
usi
ng
di
ff
e
ren
t
m
et
ho
d.
T
her
e
fore,
if
t
he
sam
e
im
age
is
us
e
d
from
diff
e
re
nt
m
et
ho
d,
the
interest
obj
ect
ca
nnot
be
se
gm
ented
com
plete
ly
fr
om
the
bac
kgr
ound
a
nd
vice
ve
rse.
Th
us
,
thi
s
will
be
res
ulti
ng
i
n
low
acc
ur
acy
a
lgorit
hm
if
the
m
ul
ti
ple
i
m
age
s
are
bei
ng
u
se
d.
T
her
e
f
or
e,
t
o
ove
rco
m
e
this
pro
blem
,
Ad
aptive
Re
d
an
d
Bl
ue
Chrom
atic
Map
(A
RB
)
an
d
Norm
al
iz
ed
Di
ff
e
ren
ce
I
ndex
(N
D
I)
at
the
sam
e
tim
e
hav
e
bee
n
i
m
p
lem
ented.
The
thr
esh
old
value
obta
ine
d
fr
om
the
experim
ental
te
st
is
us
ed
in
ARB
to
segm
ent
the
po
te
ntial
pin
ea
pp
le
fruit
from
the
le
aves.
Sim
i
la
r
to
the
H
VS
,
a
ny
ARB
val
ue
wh
ic
h
le
ss
tha
n
th
reshold
va
lue
is
rem
ov
ed
(
pix
el
intensit
y
becom
es
0)
f
ro
m
the
im
age.
This
value
is
c
on
si
de
red
as
le
a
ves.
O
n
t
he
oth
e
r
hand
,
ARB
valu
e
f
or
fruit
is
gr
eat
e
r
or
e
qual
tha
n
thres
ho
l
d
valu
e.
A
fter
im
ple
m
ent
ARB,
the
resu
lt
sti
ll
sh
ows
th
e
m
isc
la
ssifie
d
pix
el
s
and
t
he
e
dg
e
of
t
he
det
ect
ing
pi
neapp
le
is
no
t
sm
oo
th.
This
is
be
c
ause
ARB
value
f
or
backg
rou
nd
es
pecial
ly
fo
r
da
rk
e
r
le
aves
ha
ving
al
m
os
t
th
e
sa
m
e
value
with
the
f
ru
it
.
Ther
e
f
or
e,
t
o
r
e
m
ov
e
m
isc
la
ssifie
d
error
a
nd
sm
oo
th
the
e
dg
e
of
the
detect
in
g
pin
ea
pp
le
,
thi
s
pro
j
ect
has
pro
po
se
d
the
us
e
of
Tem
plate
Mat
chin
g
Me
th
od
(TMM
)
f
or
f
urt
her
segm
entat
ion
process.
The
m
ajo
r
a
xi
s
an
d
m
ino
r
a
xis
ar
e
cal
culat
ed
base
d
on the r
es
ulti
ng
resu
lt
fro
m
the A
RB
to
fin
d
the r
i
gh
t t
em
plate
, which
th
en
is use
d
to r
e
m
ov
e
any m
isc
la
ssifi
ed pixels a
nd s
m
oo
th the e
dg
e of the
d
et
ect
ing pi
neapple
[
7].
The
im
ple
m
en
ta
ti
on
of
AR
B
and
NDI
will
produce
two
se
gm
entation
s
im
age,
there
fore,
t
he
Conv
olu
ti
onal
Neural
Netw
ork
(CN
N)
is
im
plem
ented
in
t
his
al
gorit
hm
as
decisi
on
m
a
ker.
T
he
pur
pose
of
i
m
ple
m
entat
io
n
of
C
N
N
is
to
sel
ect
the
best
segm
entat
ion
i
m
ag
e
from
ARB
and
N
DI.
Firstl
y,
the
C
N
N
will
determ
ine
the
segm
entat
ion
im
age
pro
du
ce
d
f
ro
m
ARB
and
N
DI
w
heth
er
the
obj
ect
is
a
pin
ea
pple
or
no
t.
I
f
the
se
gm
entat
i
on
im
age
is
a
pin
ea
pp
le
,
the
al
gorithm
will
acce
pt
the
seg
m
entat
ion
im
a
ge.
Me
a
nwhile
,
if
t
he
seg
m
entat
ion
im
age
is
no
t
a
pin
eap
ple,
the
al
gorithm
will
rej
ect
it
.
Ho
we
ve
r,
if
both
se
gme
ntati
on
im
ages
are
cl
assifi
ed
as
a
pin
ea
pp
le
,
the
CNN
will
go
t
o
the
sec
ond
ste
p
by
usi
ng
th
e
weig
ht
pro
duced
wh
e
n
cl
ass
ify
the
i
m
age to
d
et
er
m
ine the h
ig
he
st weig
ht b
e
tw
een A
RB
a
nd
NDI.
Fi
nally
, th
e syst
em
w
il
l
acce
pt only
the
o
ut
pu
t
segm
entat
ion
im
age w
it
h
t
he hig
hest
weig
ht.
3.
RESU
LT
S
A
ND AN
ALYSIS
This
sect
io
n
presents
the
res
ul
t
and
t
he
a
naly
sis
of
t
he
im
age
acc
ur
acy
f
ro
m
var
io
us
se
gm
entat
ion
sta
ges
s
uch
as
Ca
scade
Objec
t
Detect
or
(CO
D)
t
o
detect
lo
cat
ion
of
t
he
pi
neapple,
Hu
e
Value
Segm
entat
ion
(HVS
)
to
rem
ov
e
groun
d
a
nd
sk
y,
Ad
a
ptive
Re
d
an
d
Bl
ue C
hrom
atic Ma
p
(
ARB)
and
N
or
m
al
iz
ed
Difference
Inde
x
(
NDI)
a
r
e
for
le
aves
re
m
ov
al
,
Te
m
plate
Ma
tc
hin
g
M
et
hod
(
TMM
)
for
se
g
m
entat
i
on
e
nhan
cem
e
nt
an
d
edg
e
sm
oo
thin
g
an
d
C
onvolu
ti
on
al
Ne
ur
al
Netw
ork
(C
N
N)
ha
ve
be
en
us
e
d
as
decisi
on
m
aker
t
o
se
le
ct
the
best se
gm
entation
im
age.
3.1. C
as
ca
de
Obj
ec
t Detec
t
or (C
OD)
Firstl
y,
a
Ca
sc
ade
Object
Det
ect
or
(
CO
D)
ha
s
bee
n
sel
ect
e
d
to
detect
t
he
locat
io
n
of
t
he
pin
ea
pple
from
an
captured
im
age.
Re
fer
to
Fig
ure
4,
t
he
CO
D
ret
urn
s
the
boun
ding
box
ar
ound
th
e
detect
ing
p
in
eapp
le
and
rem
ov
e
al
l
pix
el
s
outsi
de
this
boundi
ng
box.
By
im
ple
m
ented
CO
D,
pin
ea
pp
le
ca
n
be
locat
ed
a
ny
wh
e
re
in
the
im
age
an
d
this
m
et
ho
d
has
increa
se
d
t
he
acc
ur
acy
as
s
how
n
i
n
Figures
4(a)
a
nd
4(b)
beca
use
the
backg
rou
nd of
the sele
ct
ed
im
age
has bee
n d
ecreased
.
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,
Vol
.
10
, N
o.
3
,
June
201
8
:
1089
–
1105
1102
(a)
(b)
Figure
4. Re
su
l
t afte
r
a
pp
ly
in
g C
OD (le
ft
=
or
i
gin
al
im
age,
ri
gh
t
=
resu
lt
a
fter
COD
)
3.2.
Hue
V
alu
e Segm
ent
ati
on (
HVS)
The
seco
nd
st
ep
is
t
he
sel
ect
ion
of
Hu
e
pa
ram
et
er.
The
acc
ur
acy
will
increase
d
due
t
o
t
he
backg
rou
nd
in
the
im
age
has
bee
n
decr
ea
se
d.
Re
fer
to
Fig
ur
e
5,
the
gro
und
ha
s
been
suc
cessf
ully
rem
ov
e
d
from
the
i
m
ag
e
an
d
ot
her
obje
ct
s
are
m
ai
nta
ined.
Finall
y,
it
has
been
f
ound
t
hat
the
outpu
t
im
age
sti
ll
can
be
consi
der
e
d
a
s a
ccepta
ble,
ev
e
n
th
ough a
f
e
w
interest
pix
el
s
hav
e
b
ee
n rem
ov
e
d.
(a)
(b)
Figure
5. Re
su
l
t afte
r
a
pp
ly
in
g H
VS
(left
=
res
ult after C
O
D,
rig
ht
=
res
ult aft
er
HVS)
3.3. Ad
aptive
Red an
d Blue
Ch
r
omatic
M
ap
(ARB)
and
Norm
aliz
ed D
iffere
nce I
n
dex (NDI
)
The
t
hird
ste
p
is
to
rem
ov
e
le
aves
from
t
he
im
age
as
s
how
n
in
Fig
ur
e
6
us
i
ng
AR
B
an
d
N
D
I
m
et
ho
ds.
It
ca
n
be
see
n
that,
the
ARB
a
nd
NDI
hav
e
suc
cessf
ully
rem
ov
e
d
t
he
le
ave
s
from
the
i
m
age
a
s
sh
ow
n
in
Fig
ures
6(b)
a
nd
6(d
)
re
sp
ect
ivel
y,
w
hile
m
a
intai
ned
t
he
f
r
uit
area
a
nd
sh
a
pe
.
I
n
a
dd
it
io
n,
it
has
been
f
ound
t
ha
t
the
ou
t
pu
t
im
ages
i
n
Fi
gures
6(b
)
a
nd
6(d)
sti
ll
can
be
co
nsi
der
e
d
as
acce
ptable,
eve
n
t
houg
h
a few interest
pi
xels h
a
ve bee
n
rem
ov
e
d.
H
oweve
r,
w
hen usi
ng
im
age in
Figure
6(
a
),
t
he
A
RB
has
c
onsidere
d
al
l
pix
el
s
in
t
he
i
m
age
as
bac
kgr
ound.
T
his
is
because
the
ARB’s
value
f
or
t
he
al
l
pi
xels
in
the
im
age
is
le
ss
than
t
hr
es
hold
values
. Ho
we
ve
r
f
r
om
N
DI
m
et
hod,
t
he resul
t i
s m
uch
b
et
te
r
as s
how
n
i
n
F
igure
6(
c
).
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
Compre
hensi
ve Pi
ne
apple
Se
gm
e
nta
ti
on
Te
chn
i
qu
e
s
wi
th
I
ntell
igent
.
..
(
Mu
hamma
d
Az
mi A
hm
e
d
N
aw
awi
)
1103
(a)
(b)
(c)
(d)
Figure
6. (a
),
(
b) Result
afte
r a
pp
ly
in
g ARB
.
(
c)
, (d
)
Re
s
ult after
a
pp
ly
in
g NDI
3.4.
Te
m
pla
te
Matchin
g Met
ho
d
(
T
MM)
The
res
ult
fr
om
ARB
and
ND
I
se
gm
entat
i
on
m
et
ho
ds
ha
ve
a
few
m
isc
l
assifi
ed
pi
xels
and
the
e
dge
of
the
detect
in
g
pin
ea
pple
is
no
t
sm
oo
th.
T
he
refor
e
,
this
proj
ect
ap
plies
add
it
inal
i
m
age
enh
a
ncem
ent
pr
oces
s
for
furthe
r
se
gm
entat
ion
us
in
g
Tem
plate
Ma
tc
hin
g
Me
th
od.
The
di
ff
e
rent
in
s
hap
e
a
nd
siz
e
of
te
m
plate
ha
ve
been
c
on
st
ru
ct
ed
i
n
a
dv
a
nce
d
an
d
st
or
e
in
th
e
data
base.
Th
e
rig
ht
te
m
plate
is
sel
ect
ed
ba
sed
on
the
m
ajo
r
axi
s
and
m
ino
r
axis
.
The
n,
t
he
pine
app
le
s
h
a
pe
is
const
r
ucted
to
pro
du
ce
a
sm
oo
t
h
e
dg
e
a
nd
al
l
the
pix
el
s
outsi
de
this sha
pe
a
re e
lim
inate
d
as
show
n
i
n
Fi
gure
7.
(a)
(b)
Figure
7. Re
su
l
t afte
r
a
pp
ly
in
g t
e
m
plate
m
at
c
hing
(a)
Im
m
a
t
ur
e
p
i
neapple
(
b) Mat
ur
e
pin
ea
pp
le
3.5
.
C
onvo
lu
ti
onal Ne
ural
N
etwork
(CN
N)
The
im
ple
m
ent
at
ion
s
of
ARB
and
NDI
ha
ve
produce
d
tw
o
segm
entat
ion
im
ages.
Re
fer
t
o
Fig
ur
e
8,
Conv
olu
ti
onal
Neural
Netw
ork
(CN
N)
ha
s
s
uccess
fu
ll
y
sel
ect
ed
the
be
st
segm
entat
ion
i
m
age
bet
ween
ARB
and
N
DI
.
T
he
i
m
ple
m
entat
i
on
of
C
NN
i
n
the
al
gorith
m
has
increas
ed
the
ef
fici
ency
an
d
accu
r
acy
by
al
lowing th
e al
gorithm
to
ha
ve
m
any
m
et
ho
ds
t
o detec
t pin
eapp
le
.
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,
Vol
.
10
, N
o.
3
,
June
201
8
:
1089
–
1105
1104
(a)
(b)
(c)
Figure
8. Re
su
l
t afte
r
a
pp
ly
in
g C
NN (a
)
Re
s
ul
t fr
om
A
RB
(b
)
Re
su
lt
f
r
om
N
DI
(c
)
Re
s
ult from
CNN
4.
CONCL
US
I
O
N
The
im
age
segm
entat
ion
an
d
detect
ion
al
gorithm
s
hav
e
bee
n
prese
nte
d
to
detect
pi
neapple
f
r
uit
base
d
on
col
or
m
at
ur
it
y
ind
ic
es.
The
resu
lt
s
found
s
how
t
he
us
e
d
of
Ca
scade
O
bject
Detect
or
(CO
D)
was
ver
y
ef
fecti
ve
and
has
good
po
te
ntial
in
detect
ing
the
locat
ion
of
the
pi
neapple.
T
he
use
of
H
ue
val
ue
with
Ad
a
ptive
Re
d
and
Bl
ue
C
hro
m
at
ic
Ma
p
(AR
B)
an
d
Norm
al
iz
ed
Dif
fer
e
nc
e
Inde
x
(NDI)
we
re
ver
y
e
ffec
ti
ve
for
bac
kgr
ound
rem
ov
al
pro
cess
from
the
fo
c
us
im
age.
The
im
ple
m
en
ta
ti
on
of
Tem
plate
Ma
tc
h
ing
Me
thod
(TMM
)
has
im
pr
ov
e
d
the
accuracy
of
th
e
al
gorithm
by
detect
ing
th
e
sh
a
pe
of
t
he
pi
neapple
a
nd
the
n
el
i
m
inate
d
othe
r
pix
el
s
with
a
sm
oo
th
e
dge.
Furthe
rm
or
e,
t
he
im
ple
m
entat
ion
of
Conv
olu
ti
onal
Ne
ur
a
l
Netw
ork
(C
N
N)
as
decisi
on
m
aker
has
increase
d
t
he
e
ff
ic
ie
ncy
of
th
e
al
go
rithm
and
at
the
sam
e
tim
e
increase
d
the
a
ccur
acy
in
se
gm
enting
the
f
r
uit
fr
om
the
back
gr
ound.
F
ur
t
her
re
searc
h
ne
eds
to
be
done
us
in
g
intel
li
gen
t o
ptim
iz
at
ion
techni
qu
e
f
or
opti
m
al
m
at
ur
it
y i
den
ti
ficat
ion
.
ACKN
OWLE
DGE
MENTS
The
a
ut
hors
would
li
ke
to
than
k
f
or
th
e
su
pp
or
t
gi
ve
n
to
this
pr
oject
by
Mi
nistry
of
Higher
Ed
ucati
on (
M
OH
E
)
a
nd
Un
i
ver
sit
i Te
knologi M
al
ay
sia
(
U
TM),
unde
r G
UP
gr
a
nt:
QJ1300
00.25
23.13
H57.
REFERE
NCE
S
[1]
J.
I.
As
nor,
et
a
l
.
,
“
Pinea
ppl
e
M
at
urity
Re
cogni
t
ion
Us
ing
RGB
Ext
ra
ct
ion
,
”
Inter
nati
onal
Journ
al
of
El
e
ct
ri
cal
,
Computer,
En
erge
tic
and
Comm
unic
ati
on
Eng
in
ee
ring,
vol
.
7
,
p
p.
597
-
600
,
201
3.
[2]
S.
Moham
m
ad,
et
al
.
,
“
Cla
ss
ific
at
ion
of
Fresh
N
36
Pinea
ppl
e
Cr
op
Us
ing
Im
age
Proce
s
sing
Te
ch
nique
,
”
Adv
a
nced
Mate
rals R
ese
arch,
vo
l. 418
-
420,
pp.
1739
-
1743,
2011.
[3]
N.
C.
Za
n
,
et
al
.
,
“ROI
Segm
ent
ati
on
Us
ing
Mini
mum
Symmet
rical
Edg
es
Distance
:
A
St
udy
on
Josapin
e
Pi
neapple,
”
in
I
nte
rna
ti
ona
l
Con
fer
ence
on
Artif
i
ci
a
l
Intelli
g
ence
in
Com
pute
r
Sci
enc
e
and
ICT
.
A
ICS
2013.
2013,
pp.
130
-
137
.
[4]
S.
Kam
aru
ddin,
et
al
.
,
“
Canne
d
P
ine
app
le
Gradi
n
g
Us
ing
Pixel
s Colour
Ext
r
ac
t
ion,”
in
Int
ernati
ona
l
Confe
ren
ce
on
Arti
ficial Int
el
l
ig
enc
e
in
Compute
r Sc
ie
n
ce and
IC
T.
AICS
2013
.
2
013,
pp
.
217
-
22
6.
[5]
D.
S.
Prabh
a
an
d
J.
S.
Kum
ar,
“
As
sessment
of
Bana
na
Frui
t
Mat
urity
B
y
Im
age
Proce
ss
ing
Te
ch
nique
,
”
Journal
o
f
Food,
S
cienc
e
a
nd
Technol
og
y,
vol.
52
,
pp
.
1316
-
1327,
2015
.
[6]
X.
Li
m
ing
and
Z.
Yanc
h
ao,
“
Autom
at
ed
Straw
ber
r
y
Gr
adi
ng
Sy
stem
Based
on
Im
age
Proce
ss
ing,
”
Comput
ers
and
El
e
ct
ronics
in A
gricul
ture
,
vol
.
7
1S,
pp.
S32
-
S39,
2010.
[7]
M.
Azm
i
and
F.
S.
Ism
ai
l,
“
Sim
u
la
ti
on
and
Segm
ent
a
ti
on
Techni
q
ues
For
Crop
M
at
urity
Ide
n
ti
fi
cation
of
Pinea
pp
l
e
Fruit”
.
C
o
m
m
u
n
i
c
a
t
i
o
n
s
i
n
C
o
m
p
u
t
e
r
a
n
d
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
c
e
,
M
o
d
e
l
i
n
g
-
D
e
s
i
g
n
a
n
d
S
i
m
u
l
a
t
i
o
n
o
f
S
y
s
t
e
m
s
,
v
o
l
.
7
5
1
,
p
p
.
3
-
1
1
,
2
0
1
7
.
[8]
P.
Hui,
et
al
.
,
“Edge
Detect
ion
of
Gr
owing
Cit
rus
Based
on
Sel
f
-
adapti
v
e
Canny
Operator,”
in
2011
Inte
rna
t
ional
Confer
ence
on
Com
pute
r
Distribut
ed
Control
a
nd
Inte
ll
ig
ent
E
nvironmenta
l
Monitori
ng
.
2011
IE
EE,
2011,
pp.
342
-
345.
[9]
K.
Yam
amoto,
e
t
al.,
“
On
Plant
Dete
c
ti
on
of
Int
ac
t
Tomato
Frui
ts
Us
ing
Im
age
Anal
y
sis
And
M
ac
hin
e
L
ea
rn
ing
Methods,
”
Op
en
Acce
ss
Sensor,
vol.
14
,
pp
.
1219
1
-
12206,
2014
.
[10]
C.
Zha
o
,
e
t
al
.
,
“
Immature
Gree
n
Cit
rus
Detect
i
on
Based
on
Co
lou
r
Feat
ur
e
An
d
Sum
of
Abs
ol
ute
Tr
ansform
ed
Diffe
ren
c
e
(SA
T
D)
Us
ing
Colour
Im
age
s
in
the
C
it
rus
Grove,”
Co
mputers
and
Ele
ct
ronics
in
Agri
c
ult
ure,
vol
.
124
,
pp.
243
-
253
,
20
16.
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
Compre
hensi
ve Pi
ne
apple
Se
gm
e
nta
ti
on
Te
chn
i
qu
e
s
wi
th
I
ntell
igent
.
..
(
Mu
hamma
d
Az
mi A
hm
e
d
N
aw
awi
)
1105
BIOGR
AP
HI
ES OF
A
UTH
ORS
Muham
m
ad
Azm
i
Ahm
ed
Nawawi
re
ce
iv
ed
his
BEng.
(
Hons
.
)
i
n
El
e
ct
r
ic
a
l
(Me
cha
tron
ic
s)
from
Univer
siti
T
ekn
ologi
Malay
si
a
(UTM)
in
2015.
He
is
cur
ren
t
l
y
MP
hil
.
stud
en
t
at
Univer
si
ti
Te
knologi
Ma
lay
sia
(UTM).
H
is
cur
ren
t
rese
a
rch
int
er
ests
inc
lude
vision
s
y
stem,
image
proc
essing
and
a
rti
fi
ci
a
l
in
te
l
li
g
e
nce
.
Fati
m
ah
Sham
Ism
ai
l
has
m
ore
t
han
20
y
e
ars
ex
per
ie
n
ce
in
a
rea
of
Control,
Instr
um
ent
at
ion
and
Optimiza
ti
o
n
E
ngine
er
ing
sinc
e
joi
ning
Univer
siti
T
eknol
og
i
Malay
s
ia
(UTM)
in
1992
.
Sh
e
rec
e
ive
d
h
er
B
Sc.
(Hons
.
)
(198
9)
in
Ph
y
si
c
fr
om
Univer
siti
Keba
ngsa
an
Ma
lay
s
ia,
Master
s
Degre
e
(1992)
a
nd
Ph.D.
(2011)
from
UTM.
Her
m
ai
n
rese
arc
h
int
er
ests
inc
lude
opti
m
iz
at
ion
using
AI
te
chni
ques,
Proce
ss
C
ontrol
and
Instrum
ent
at
ion
desi
gn.
Curre
ntly
,
s
he
is
conduc
ti
ng
rese
arc
h
es
on
deve
lopment
o
pti
m
iz
ation
al
g
orit
hm
fo
r
m
u
lt
i
-
obj
ective
pr
oble
m
s,
pla
nt
opti
m
iz
ation
d
esign,
and
f
aul
t
de
t
ec
t
ion
and
di
agn
osis.
Dr.
Haz
li
na
Sel
amat
is
an
As
sociate
Prof
essor
and
a
Profess
i
onal
Engi
ne
er
a
nd
have
be
en
le
c
turi
ng
at
th
e
El
e
ct
ri
ca
l
Engi
n
ee
ring
Facu
lty
,
Univer
siti
T
ekn
ologi
Malay
si
a
(UTM)
since
the
y
e
ar
2000
.
She
gra
dua
te
d
fro
m
Im
per
ia
l
Col
le
ge
o
f
Sci
enc
e
,
T
ec
hnolog
y
a
nd
Medic
in
e
,
Univer
sit
y
of
L
ondon
in
El
ec
tr
i
ca
l
and
Elec
tron
ic
s
Engi
neering
in
1998.
Dr.
Haz
li
n
a
Sela
m
at
obta
in
ed
her
M
Eng.
and
Ph.D.
in
Elec
tr
ical
En
gine
er
ing
from
Univer
siti
Te
kn
ologi
Mal
a
y
s
ia
(UTM)
in
2000
and
2007
respe
ctively
.
Sin
ce
201
4,
she
is
the
Dire
ct
or
of
the
C
ent
r
e
for
Artifi
c
ia
l
Inte
lligen
ce
and
Roboti
cs
(CAIRO
)
UTM.
Her
princ
ip
al
ar
ea
s
of
int
er
est
are
a
dapt
iv
e
cont
ro
l,
onli
ne
s
y
s
te
m
id
ent
ifica
ti
on
and
appl
i
ca
t
ion
of
c
ontrol
to
th
e
hig
h
-
orde
r
and
nonl
ine
ar
s
y
s
te
m
s.
Her
cur
ren
t
res
ea
rch
works
are
in
the
ar
ea
of
el
e
ct
ron
ic
control
unit
d
esign
for
aut
om
oti
ve
appl
i
ca
t
ions,
cr
owd
m
odel
ing,
cont
rol
and
sim
ula
ti
on
for
saf
e
r
buil
ding
desig
n
and
ene
rg
y
opti
m
i
za
t
ion.
Evaluation Warning : The document was created with Spire.PDF for Python.