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.
9
, No
.
5
,
Octo
ber
201
9
, pp.
3495
~3
503
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v9
i
5
.
pp3495
-
35
03
3495
Journ
a
l h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Qualit
y gradin
g of soybe
an seeds u
sing image anal
ysis
Sutasinee
Jit
anan
1
,
P
awat C
himl
ek
2
1
Depa
rtment of
Com
pute
r
Scie
n
ce
and
Inform
a
tion T
e
chnol
og
y
,
Facul
t
y
of
Sci
en
ce
,
Nare
suan
Un
ive
rsit
y
,
Th
ailan
d
2
Depa
rtment of
Com
pute
r
Scie
n
ce
and
Inform
a
tion T
e
chnol
og
y
,
Facul
t
y
of
Sci
en
ce
and Technolo
g
y
,
Pibulsongkram Raj
abh
at
Univ
er
sit
y
,
Tha
i
la
nd
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Sep
5
, 2
01
8
Re
vised
A
pr
7
,
201
9
Accepte
d
Apr
20
, 201
9
I
m
age
proc
essing
and
m
ac
hine
le
arn
ing
te
chn
iq
ue
are
m
odifi
ed
t
o
use
the
qual
ity
gr
adi
ng
of
so
y
be
an
see
d
s
.
Due
to
q
ualit
y
gra
ding
is
a
v
er
y
importan
t
proc
ess
for
the
so
y
bea
n
industr
y
and
so
y
b
ea
n
f
armers.
The
r
e
ar
e
stil
l
som
e
cri
tica
l
proble
m
s
tha
t
nee
d
to
be
over
come
.
Th
er
efo
re,
t
h
e
ke
y
c
ontri
buti
ons
of
thi
s
pap
er
ar
e
first
,
a
m
et
ho
d
to
e
li
m
ina
t
e
shadow
noise
f
or
segm
ent
so
y
bea
n
se
eds
of
high
qual
ity
.
Second,
a
nove
l
appr
oac
h
for
c
olor
fea
tur
e
which
robust
for
il
lumina
ti
on
changes t
o
red
u
ce
s
proble
m
of
col
o
r
diffe
r
enc
e
.
Thi
rd,
an
appr
o
a
ch
to
discov
er
a
set
of
fe
at
ure
a
n
d
to
form
cl
assif
ie
r
m
odel
to
strengt
hen
the
d
iscri
m
ina
ti
o
n
po
wer
for
so
y
b
ea
n
cl
assifi
ca
t
ion.
T
his
stu
d
y
used
bac
kgroun
d
subtrac
ti
on
to
red
uce
shadow
appe
ari
n
g
in
t
he
ca
ptur
ed
image
and
p
ropo
sed
a
m
et
hod
to
ext
ra
ct
col
or
f
eat
ure
base
d
on
rob
ustness
f
or
il
lumination
changes
which
w
as
H
components
i
n
HS
I
m
odel
.
We
pro
posed
cl
assifi
er
m
odel
using
combinat
i
on
of
the
col
or
histogra
m
of
H
components
in
HS
I
m
odel
and
GLCM
stat
isti
cs
to
rep
rese
nt
t
he
col
or
and
te
xt
ure
fea
tur
es
to
strengt
hen
th
e
discri
m
ina
ti
on
p
owe
r
of
soy
b
ea
n
gra
ding
and
to
solve
shap
e
var
ia
n
ce
in
eac
h
so
y
bea
n
se
ed
s
cl
ass
.
SV
M
cl
assifie
rs
ar
e
ge
ner
ated
to
ide
nti
f
y
norm
al
see
ds,
purpl
e
se
eds,
gre
en
see
ds
,
wrinkl
ed
see
ds
,
and
o
the
r
see
d
t
y
pes.
W
e
conduc
t
ed
exp
eri
m
ent
s
on
a
dat
ase
t
compos
ed
of
1,
320
s
o
y
bea
n
see
ds
a
nd
6,
600
se
ed
i
m
age
s
with
var
i
es
in
br
ight
ness
le
v
el
s
.
The
expe
riment
al
r
e
sults
ac
hi
eve
d
a
cc
ura
cies
of
99.
2%,
97.
9
%
,
10
0
%,
100
%
,
98.
1
%,
and
100
%
for
ov
era
l
l
se
eds,
norm
al
see
d
s,
purple
see
ds
,
gre
en
se
eds
,
wrinkle
d
se
eds,
and
oth
er
se
eds,
resp
ec
t
ive
l
y
.
Ke
yw
or
d
s
:
Im
age classi
ficat
ion
So
ybea
n
cl
assi
ficat
ion
S
oybea
n see
d
Copyright
©
201
9
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
:
Su
ta
sinee
Jit
an
an,
Dep
a
rtm
ent o
f C
om
pu
te
r
Scie
nce a
nd Info
rm
a
ti
on
Tec
hnol
ogy,
Faculty
of S
ci
e
nce,
Nar
es
ua
n Un
i
ve
rsity
,
Ph
it
san
ulok
65
000 Thail
an
d
.
Em
a
il
: sutasi
neec@nu.ac
.th
1.
INTROD
U
CTION
So
ybea
ns
a
re
an
im
po
rtant
agr
ic
ultur
al
cr
op
that
is
widel
y
con
s
um
ed
becau
se
it
is
an
excep
ti
ona
l
so
urce
of
nutri
ents,
with
a
h
i
gh
protei
n
an
d
ver
y
high
oil
con
te
nt
[1
]
.
S
oybean
qu
al
it
y
aff
ect
s
the
pr
ic
in
g
an
d
qu
al
it
y
of
gr
ai
n
for
c
rop
ping
an
d
c
onsu
m
ption
.
S
oybea
n
diseases
grea
tl
y
red
uce
the
eco
nom
ic
val
ue
of
so
ybea
n
pro
du
ct
s
and
res
ult
i
n
eco
no
m
ic
los
ses
for
the
so
y
bean
in
dustry
and
fa
rm
ers.
Thu
s
,
the
de
velo
pm
ent
of
a
ra
pid
an
d
reli
able
m
et
h
od
of
detect
in
g
the
ap
pear
a
nc
e
qu
al
it
y
of
so
ybea
ns
is
of
gr
eat
sig
nifica
nce
to
so
ybea
n
far
m
ers
a
nd
the
s
oybe
an
i
ndus
try
[
2]
.
Var
i
ous
s
oy
bean
diseases
aff
ect
t
he
see
ds
ap
pea
ran
ce
in
te
rm
s
of
siz
e,
s
ha
pe,
and
c
olor.
Di
sease
aff
ect
e
d
so
ybea
ns
ca
n
be
pur
ple
seed
s,
gr
e
en
see
ds
,
wr
in
kled
see
ds,
a
nd
sm
a
ll
/spli
t
seeds.
F
or
the
m
os
t
par
t,
the
pro
du
ct
ivit
y
of
soy
bean
s
depen
ds
on
t
he
qual
it
y
of
grai
ns,
an
d
that
is
wh
y
qual
it
y gra
ding is a
v
e
ry
i
m
po
rtant
pr
oc
ess for
the
so
y
bean in
dustry a
nd so
y
bea
n
fa
r
m
ers.
A
gradi
ng
m
ac
hin
e is use
d
to
cl
assify
so
ybea
ns
acco
rd
i
ng to
their q
ualit
y. This m
achine can
sep
a
rate
on
ly
f
orei
gn
m
at
erial
and
s
eeds
with
non
-
sta
ndar
d
siz
es
.
H
ow
e
ve
r,
th
e
m
achine
cannot
cl
assify
lo
w
-
grade
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19
:
3495
-
3503
3496
seeds
s
uch
as
dr
ie
d
seeds
,
gr
een
seed
s,
pu
r
ple
seeds
,
an
d
con
ta
m
inants
with
the
sam
e
siz
e
as
regular
so
ybea
n
seeds.
T
her
e
fore,
sk
il
le
d
w
orker
s
are
need
e
d
to
carry
out
the
qual
it
y
gr
ad
ing
of
so
ybea
n
gr
ai
n.
Howe
ve
r,
this
appr
oach re
qu
i
res
m
any w
or
ke
rs,
i
n
a
dd
it
io
n t
o bein
g
ti
m
e
-
consum
ing
, a
nd
pro
ne
to
hum
an
e
rror.
Anothe
r
qu
al
it
y
gr
a
ding
ap
proac
h
c
on
sist
s
of
us
in
g
im
a
ge
pr
ocessin
g
te
chn
i
qu
es
a
nd
m
achin
e
le
arn
in
g
al
gori
thm
s
to
evalua
te
the
qual
it
y
of
a
gri
cultu
ral
pro
du
ct
s
a
nd
f
ood.
I
n
recent
ye
ars,
s
uc
h
m
e
tho
ds
hav
e
bec
om
e
wides
pr
ea
d
[
3
-
8
]
.
Me
batsio
n
et
al
.
hav
e
cl
assifi
ed
cereal
gr
ai
ns
,
nam
ely
;
bar
le
y,
oat,
r
ye
and
wh
eat
us
in
g
m
orp
ho
l
og
ic
al
a
nd
R
GB
c
olo
r
featur
e
with
a
n
identific
at
io
n
accuracy
of
99
.6
%
[
9
]
.
Olgu
n
et
al.
hav
e
de
vel
op
e
d
a
n
a
uto
m
at
e
d
syst
em
fo
r
w
heat
grai
n
cl
as
sific
at
ion
t
hat
can
ac
hieve
an
accu
racy
of
88.
33%
us
in
g
t
he
dens
e
scal
e
-
in
var
ia
nt
feat
ur
e
tra
nsfo
rm
(S
IF
T
)
featur
e
,
w
hich
is
eval
uated
by
a
s
upport
vecto
r
m
achine
(
SVM
)
cl
assifi
er
[
10
]
.
A
no
t
her
pro
po
se
d
a
ppr
oach
f
or
so
y
be
an
disease
re
cognit
ion
of
s
oybea
n
le
aves
is
base
d
on
the
im
age
local
descr
i
ptors
an
d
Ba
g
of
Visu
al
Wo
r
ds
,
w
hich
ar
e
m
et
ho
ds
rob
us
t
t
o
occlusi
on
[
11
]
.
Tan
et al
. [
12
]
h
a
ve
pr
opos
e
d a m
et
ho
d
usi
ng
back
pro
pa
ga
ti
on
(
B
P)
a
rtific
ia
l neural
n
et
works
to
rec
ognize
s
oybea
n
see
d
di
sea
ses
su
c
h
a
s
so
y
bean
f
r
ogey
e,
m
il
dew
ed
s
oybea
ns
,
w
or
m
-
eat
en
s
oybeans,
and
dam
aged
so
ybea
ns
,
with
an
acc
uracy
of
90%
f
or
he
te
rogen
e
ous
s
oy
be
an
seeds
w
it
h
seve
ral
diseases.
The
syst
em
’s
perform
ance
w
as
lim
i
te
d
by
c
olor
di
ff
e
ren
ce
an
d
sh
a
dow
noise
un
der
t
he
co
ndit
ion
of
na
tural
li
gh
t. T
heref
ore, this
pro
pose
d req
uires
the l
igh
t s
ource t
ha
t distri
bute
d t
he
li
gh
t e
ve
nly
.
In
a
recent
st
udy,
W
ei
-
Zh
en
et
al
.
[13]
de
ve
lop
e
d
a
m
od
e
l
to
est
i
m
at
e
t
he
pe
rce
nt
de
f
oliat
ion
of
so
ybea
n
le
ave
s
from
RGB
i
m
ages
us
in
g
t
he
le
af
a
re
a
a
nd
e
dge
pi
xel
nu
m
ber
.
M
om
in
et
al
.
[1
4]
hav
e
dev
el
op
e
d
a
m
achine
visi
on
s
yst
e
m
us
ing
al
l
co
m
po
ne
nts
of
the
HS
I
c
ol
or
feat
ur
es
fro
m
fr
on
t
a
nd
ba
ck
lit
i
m
ages
to
dete
ct
m
at
erial
s
ot
her
t
han
grai
n
(MO
Gs),
suc
h
as
sp
li
t
bea
ns
,
con
ta
m
inate
d
beans,
def
ect
iv
e
bea
ns
and
ste
m
/po
d
m
at
erial
s,
in
soy
bean
harvesti
ng.
T
heir
m
et
ho
d
achie
ved
a
c
cur
aci
es
of
96
%,
75%
,
an
d
98%
f
or
def
ect
ive
bea
ns
an
d
ste
m
s/po
ds
,
res
pecti
vel
y
and
pe
rfor
m
ance
dep
e
nds
on
bac
k
a
nd
f
r
on
t
li
ght
c
on
tr
olli
ng
of
web cam
era that aff
ect
with il
lum
in
at
ion
c
ha
ng
e
s
of
HSI c
ol
or
c
om
po
ne
nt
s.
Althou
gh
so
m
e
existi
ng
im
age
pr
ocessin
g
and
m
achine
l
earn
i
ng
te
c
hn
i
qu
e
a
re
m
od
ifi
ed
to
us
e
the
qu
al
it
y
gr
a
di
ng
of
s
oybea
n
seeds
.
T
her
e
are
sti
ll
so
m
e
crit
ic
al
pro
blem
s
that
need
to
be
ove
rco
m
e.
First,
sha
do
w
no
ise
ca
n
occur
w
he
n
cha
nging
cam
era
an
gl
e
and
the
co
ndit
ion
of
natu
r
al
li
gh
t.
This
c
onditi
on
can
re
du
ce
acc
ur
acy
of
im
ag
e
segm
entat
ion
and
it
has
a
n
eff
ect
on
detec
ti
on
of
the
bo
unda
ries
f
or
s
oy
bea
n
seed.
Seco
nd,
colo
r
dif
fere
nc
e
can
de
grade
cl
assifi
cat
ion
powe
r
of
the
qu
al
it
y
gr
a
ding
of
s
oybea
n
seeds
.
Current
res
ear
ch
of
col
or
fe
at
ur
e
e
xtracti
on
are
no
t
rob
ust
fo
r
c
ha
ng
i
ng
of
il
lum
inatio
n
beca
us
e
th
e
li
gh
t
so
urce
was
des
ign
e
d
f
or
each
dataset
.
The
refor
e
,
colo
r
feat
ur
e
s
hould
be
extracte
d
f
r
om
the
color
m
odel
and
com
po
ne
nts
th
at
are
no
t
se
nsi
ti
ve
to
il
lu
m
inance
c
hange
s
and
t
o
en
ha
nce
cl
assify
r
esults.
Thi
rd,
sh
a
pe
var
ia
nce
in
ea
ch
s
oybea
n
se
eds
cl
ass
a
re
an
ob
sta
cl
e
for
s
oybea
n
see
ds
cat
e
gorizat
ion.
F
or
this
r
easo
n,
com
bin
at
ion
of
colo
r
feat
ur
e
a
nd
oth
e
r
feat
ures
t
hat
are
no
t
base
d
on
s
hap
e
can
im
pr
ov
e
t
he
ef
fici
ency
of
th
e
cl
assifi
cat
ion
i
n
the
each
clas
s of s
oybea
n
se
eds.
To
this
e
nd,
t
his
pa
pe
r
pr
opos
e
d
a
f
ram
ewo
r
k
to
ge
ner
at
e
a
ne
w
cl
assi
fier
m
od
el
f
or
the
qual
it
y
gr
a
ding
of
s
oy
bean
seeds
w
hi
ch
ad
dr
es
ses
these
t
hree
c
ha
ll
eng
es.
O
ur
a
ppr
oach
c
onsis
te
d
of
proc
ess
that
rob
us
t
for
sh
a
how
noise
,
c
ha
ng
i
ng
of
il
lu
m
inati
on
an
d
sh
a
pe
va
riance
.
Our
in
vestig
at
ion
f
ocused
on
th
e
diseases
of
so
y
bean
see
ds
su
c
h
as
dr
ie
d
so
y
beans,
green
s
oybea
ns
,
pu
rp
l
e
so
ybea
ns
,
w
r
ink
le
d
s
oybe
a
ns,
and
sm
a
ll
/spli
t soy
beans.
2.
PROP
OSE
D
METHO
D
This
work
f
oc
us
es
on
the
re
cogniti
on
an
d
cl
assifi
cat
ion
of
t
he
qu
al
it
y
gr
a
ding
of
so
y
bean
see
ds
.
The
m
ajor
ste
ps
of
cl
assifi
c
at
ion
i
nclu
de
i
m
age
segm
entat
ion
,
see
d
c
r
opping,
feat
ure
extracti
on,
m
od
el
const
ru
ct
io
n, s
el
ect
ion
of
s
uit
able cl
assifi
er
m
od
el
, an
d ac
c
ur
acy
as
sessm
e
nt, whic
h
a
re
presente
d
in
Fig
ur
e
1.
2.1.
Ima
ge
d
ata
se
t
This
stu
dy
use
d
1,3
20
so
y
bean
s
eeds
from
the
Seed
Re
search
a
nd
De
velo
pm
e
nt
Ce
nter
a
t
Ph
isa
nu
l
ok
e
P
r
ov
i
nce.
S
pecifi
cal
ly
,
this
study
us
e
d
the
C
ha
ing
Ma
i
60
s
eed
beca
us
e
it
is
the
m
os
t
po
pula
r
so
ybea
n
see
d
in
Thail
an
d.
T
he
so
y
bea
n
sa
m
ples
con
sist
ed
of
40
0
no
r
m
al
seeds,
w
hi
le
the
infected
see
ds
com
pr
ise
d
27
5
purp
le
see
ds
,
215
gr
ee
n
see
ds
,
220
wr
in
kl
ed
seed
s,
an
d
210
oth
er
see
ds.
The
s
oybea
n
seed
sam
ples are
presented
in Fi
gu
re
2
.
The
hardw
a
re
us
e
d
f
or
t
he
a
cqu
isi
ti
on
of
i
m
ages
co
ns
ist
ed
of
a
c
olo
r
ca
m
era
(EOS
70
0D,
Ca
no
n)
with
a
zo
om
len
s
with
fo
c
al
le
ng
t
h
of
8
-
55
m
m
(EF
-
S
18
-
55
IS
STM),
li
ght
s
ource,
an
d
black
A
3
pa
pe
r
(
11.
7
x
16.
5
in)
.
The
c
a
m
era
was
m
ounted
on
t
o
a
sta
nd
a
bove
a
li
gh
t
with
a
n
11
watt
lam
p.
Each
gro
up
of
s
oy
bean
sam
ples
was
captu
red
on
black
pap
e
r
by
th
e
cam
era
at
a
distance
of
12
i
n,
a
nd
at
an
a
ngle
of
60°
to
90°
with
the
black
pa
pe
r.
A
phot
ogra
ph
of
the
black
pap
e
r
wit
hout
t
he
so
y
bean
see
ds
wa
s
capt
ur
e
d
be
fore
the
s
oy
bean
seeds
we
re
ad
ded.
Each
im
a
ge
of
the
so
yb
ean
inclu
de
d
so
ybea
n
sam
ple
s
con
ta
ini
ng
a
ppr
ox
im
at
ely
10
to
30
seeds. T
he
im
a
ges were
capt
ured
w
it
h a si
ze
of 80
0 x 600
pi
xels a
nd a r
e
sol
ution
of
72 p
i
xels/i
n.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Qualit
y g
r
adin
g
of s
oyb
e
an se
eds usin
g
i
mag
e ana
ly
sis (
Su
t
as
inee
Jit
anan)
3497
Fig
ure
1. Fr
am
ewor
k of
pro
pose
d
m
et
ho
d
(a)
(b)
(c)
(d)
(e)
Figure
2
.
S
oybe
an
sam
ples: (
a)
norm
al
seeds
, (b) pu
rp
le
se
eds
a
ff
ect
e
d by Ce
rcosp
or
a
Le
af Sp
ot,
(c)
green
seed
s
, and (
d)
w
rin
kl
ed
see
ds
(e)
ot
her seed
s
2.2.
Ima
ge
se
gmen
tation a
nd
se
e
d c
ro
ppin
g
In
the
ca
pture
d
im
age,
the
bounda
ry
of
each
so
y
bean
s
eed
was
se
gme
nted
us
in
g
ba
ckgr
ou
nd
su
bt
racti
on
[
15
]
,
wh
ic
h
is
a
n
appr
oach
us
e
d
widely
to
se
pa
rate
the
fore
groun
d
obj
ect
from
the
backg
r
ound
in
vid
e
os
ca
pture
d
by
sta
ti
c
ca
m
eras.
I
n
t
his
stu
dy,
we
a
ppli
ed
t
his
m
eth
od
t
o
sti
ll
im
age
segm
entat
ion
.
More
ov
e
r,
t
his
stud
y
us
e
d
th
e
fr
am
e
diff
e
re
nce
al
go
rithm
for
im
age
seg
m
entat
ion
bec
ause
it
is
the
s
i
m
plest
m
et
ho
d,
i
n
ad
diti
on
to
bein
g
appropr
ia
te
f
or
sti
ll
i
m
ages
and
a
ble
of
ha
nd
li
ng
li
ghti
ng
changes.
Le
t
I
be
a
captu
red
im
age
that
con
ta
ins
so
ybea
n
see
ds
,
and
the
i
ntens
it
y
fo
r
pix
el
s
i
n
I
be
de
no
te
d
as
P[
I].
T
he
n,
P[
B]
is
intensit
y
for
pix
el
s
in
t
he
ba
c
kgr
ound
im
age
subtract
ed
fro
m
the
corres
pond
i
ng
pi
xels
at
the
sam
e
po
s
it
ion
i
n
P[
I]
a
nd
com
pu
te
d
by
(
1
)
.
Fi
nally
,
P[
F]
is
t
he
re
su
lt
of
th
e
fr
am
e
diff
e
re
ncin
g
c
om
pu
ta
ti
on
a
nd
is
e
xp
resse
d
as foll
ows.
P[
F] =
P[I]
–
P
[B]
(
1)
In
eac
h
com
po
ne
nt
of
the
RGB
colo
r
spa
ce,
the
fr
am
e
diff
e
ren
ce
w
as
cal
culat
ed
betwee
n
the
backg
rou
nd
a
nd
capt
ur
e
d
im
ages.
T
he
re
s
ul
t
of
this
proc
ess
identifie
s
t
he
re
gion
of
s
oybea
n
see
ds
i
n
the
captu
red
im
ag
e.
T
his
m
et
ho
d
can
help
to
el
i
m
inate
the
pro
blem
of
a
s
hadow
ap
pea
rin
g
i
n
the
capt
ur
e
d
i
m
age.
Subseque
ntly
,
this
im
age
is
co
nv
e
rted
t
o
a
bi
nar
y
im
a
ge
by
local
ly
ada
ptive
th
re
sh
ol
ding
[16],
w
hich
com
pu
te
s
a
threshold
f
or
eac
h
pix
el
by
us
in
g
the
local
m
e
an
intensit
y
aro
un
d
the
neig
hbor
hood
of
the
pix
el
.
The
e
xistence
of
ho
le
s
i
n
the
bin
a
ry
i
m
age
ind
ic
at
es
the
e
xistence
of
no
i
se.
The
se
hole
s
are
rem
ov
e
d
by
a
m
or
phologica
l
op
e
rati
on
[17].
Then
,
the
re
gi
on
bounda
ries
of
each
s
oybea
n
seed
are
trac
ed
us
in
g
the
Moore
-
Neig
hbor
traci
ng
al
gorithm
.
A
so
y
bean
se
ed
is
represe
nt
ed
by
the
co
nn
ect
e
d
com
po
ne
nts
of
the
reg
i
on
bounda
ries.
Th
e
sp
li
t
s
oybea
n
seeds
ar
e
s
hown
in
Fig
ur
e
3.
A
ddit
ion
al
ly
,
the
s
oybe
an
s
eeds
a
re
c
r
oppe
d
t
o
cov
e
r
t
he
re
gi
on
bounda
ries
of eac
h
s
oybea
n see
d, as
sho
w
n
in
Fig
ure
4.
C
a
p
t
u
r
e
d
i
m
a
g
e
I
m
a
g
e
s
e
g
m
e
n
t
a
t
i
o
n
(
b
a
c
k
g
r
o
u
n
d
s
u
b
t
r
a
c
t
i
o
n
)
S
e
e
d
c
r
o
p
p
i
n
g
(
M
o
o
r
e
-
N
e
i
g
h
b
o
r
t
r
a
c
i
n
g
)
F
e
a
t
u
r
e
E
x
t
r
a
c
t
i
o
n
(
H
c
o
m
p
o
n
e
n
t
s
i
n
H
S
I
m
o
d
e
l
)
M
o
d
e
l
c
o
n
s
t
r
u
c
t
i
o
n
(
S
V
M
c
l
a
s
s
i
f
i
e
r
)
C
l
a
s
s
i
f
i
e
r
m
o
d
e
l
s
F
e
a
t
u
r
e
E
x
t
r
a
c
t
i
o
n
C
a
p
t
u
r
e
d
i
m
a
g
e
I
m
a
g
e
s
e
g
m
e
n
t
a
t
i
o
n
S
e
e
d
c
r
o
p
p
i
n
g
p
u
r
p
l
e
s
e
e
d
g
r
e
e
n
s
e
e
d
w
r
i
n
k
l
e
d
s
e
e
d
n
o
r
m
a
l
s
e
e
d
o
t
h
e
r
s
e
e
d
L
e
a
r
n
i
n
g
P
h
a
s
e
C
l
a
s
s
i
f
i
c
a
t
i
o
n
P
h
a
s
e
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19
:
3495
-
3503
3498
(a)
(b)
(c)
(d)
Figure
3. Im
age seg
m
entat
ion f
or
s
plit
so
y
be
an
see
ds
:
(a)
ba
ckgr
ound im
a
ge,
(
b)
ca
pture
d
im
age,
(c)
bac
kgr
ound
subtracti
on
, a
nd
(d) regi
on
bo
unda
ries of eac
h
s
oybea
n
see
d
(a)
(b)
(c)
(d)
(e)
Figure
4. Cr
opped see
ds
: (a
) norm
al
seeds
, (b)
pur
ple see
ds, (c
) gr
ee
n
see
ds,
(d) wri
n
kle
d
se
eds, an
d (e)
o
t
her seed
s
2.3.
Feature e
xt
r
ac
tion
This
stu
dy
c
om
bin
ed
colo
r
a
nd
te
xture
c
harac
te
risti
cs
to
im
pr
ov
e
t
he
ef
f
ic
ie
ncy
of
t
he
cl
assifi
cat
ion
from
pr
ob
le
m
of
s
hap
e
var
ia
nce
in
each
s
oy
bean
see
ds
cl
ass.
The
refo
re,
this
stud
y
us
e
d
the
col
or
histogram
al
gorithm
[1
8]
and
G
rey
Lev
el
Co
-
occ
urrenc
e
Ma
trix
(G
L
CM
)
[19]
fo
r
t
he
extracti
on
of
col
or
a
nd
te
xture
featur
e
s.
The
HS
I
m
od
el
is
us
e
d
for
the
colo
r
histo
gr
am
becau
se
it
is
m
or
e
cl
os
el
y
r
el
at
ed
to
how
hu
m
ans
per
cei
ve
colo
r.
The
com
po
ne
nts
of
this
m
od
el
are
hue,
sa
turati
on,
an
d
intensit
y.
The
captu
red
im
ages
are
conve
rted
f
rom
the
RGB
mo
del
to
the
H
S
I
m
od
el
us
in
g
the
colo
r
trans
form
at
ion
[
17
]
exp
r
esse
d
by
(2)
-
(4),
wh
e
re H
is
h
ue
,
S
is
sat
ur
at
io
n,
an
d
I
is
the
intensit
y
of
t
he
HS
I
m
od
el
;
R, G
, and
B
de
note
the r
e
d,
gree
n,
a
nd
blu
e
c
olo
r
of
the
RGB
m
od
el
,
resp
ect
ivel
y.
Our
ap
proa
ch
sepa
r
at
e
s
the
col
or
c
om
po
ne
nts
(
H)
fro
m
the
Saturati
on
co
m
po
nen
t
(S)
and
the
lum
inance
com
po
ne
nt
(I
)
w
hich
is
le
ss
sensiti
ve
to
il
lu
m
inati
on
change
s
.
This
m
e
tho
d
re
du
ce
d
pro
blem
of
colo
r
dif
fere
nce
in
li
gh
ti
ng
cha
ng
es
.
The
colour
featu
re
s
are
extracte
d
from
the
colo
r
com
pone
nts
(
H)
of
HS
I
c
olor
m
od
el
base
d
on
hi
stog
ram
analy
sis
wh
ic
h
ha
ve
robust
for
cha
ng
i
ng
of il
lum
inati
on.
=
−
1
{
0
.
5
[
(
−
)
+
(
−
)
]
[
(
−
)
2
+
(
−
)
(
−
)
]
1
/
2
}
(
2)
=
1
−
3
(
+
+
)
[
min
(
,
,
)
]
(
3)
=
1
3
(
+
+
)
(
4)
Figure
5
s
hows
the
col
or
s
in
the
H
(
hue
)
c
om
po
ne
nt
s
of
al
l
s
oybe
an
see
ds
in
the
dataset
.
The
dif
fer
e
nce
s
can
be
ob
s
er
ved
in
the
hue
value
of
t
he
H
SI
m
od
el
.
The
m
ax,
m
in,
an
d
m
ean
hue
valu
es
ar
e
li
ste
d
in Ta
ble
1.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Qualit
y g
r
adin
g
of s
oyb
e
an se
eds usin
g
i
mag
e ana
ly
sis (
Su
t
as
inee
Jit
anan)
3499
(a)
(b)
(c)
(d)
(e)
Figure
5. H
ue hist
ogram
s o
f HSI m
od
el
fo
r
al
l soybea
n
see
ds
i
n datase
t:
(a
) norm
al
seeds
, (b) pu
rp
le
se
eds,
(c)
green
seed
s
, (d) w
rin
kled
s
eeds,
and
(e) ot
her seed
s
Table
1.
H
ue v
al
ues of
all
im
a
ges
i
n datase
t
So
y
b
ean type
Hu
e valu
e
Max
Min
Mean
No
r
m
al
seed
s
358
0
86
Pu
rple seed
s
359
0
101
Gree
n
seed
s
357
0
79
W
rink
led
seed
s
357
0
87
Oth
er
seed
s
356
0
56
The
hu
e
val
ue
s
of
eac
h
s
oybe
an
see
d
a
re
tr
ansfo
rm
ed
to
a
colo
r
histo
gra
m
to
represe
nt
the
col
or
distrib
ution
i
n
each
i
m
age.
The
col
or
histogram
cou
nts
the
num
ber
of
pix
el
s
in
eac
h
col
or
in
the
i
m
age,
and
t
he
c
olor
s
pace
is
di
vid
e
d
int
o
a
num
ber
of
c
olo
r
bi
ns.
Gi
ven
a
col
or
s
pace
with
k
col
or
bi
ns
,
t
he
c
olo
r
histo
gr
am
of
a
n
im
age
I
is
de
fine
d
as
H
(I)
=
[
h(1),
h(2),
…h(k
)],
w
he
re
h(
i
)
is
t
he
num
ber
of
pix
el
s
with
a
colo
r
bin
i
cal
cul
at
ed
by
Eq
ua
ti
on
(
5)
.
Addi
ti
on
al
ly
,
n_
i
is
the
nu
m
ber
of
pix
el
s
with
a
colo
r
bin
i,
an
d
N
is
the num
ber
of
pix
el
s i
n
a
n
im
age.
ℎ
(
)
=
(5)
The
GLCM
is
a
sta
ti
st
ic
al
m
e
thod
of
e
xam
i
ning
te
xtures
thr
ough
analy
zi
ng
the
s
patia
l
relat
ion
s
hip
of
pix
el
s
by
c
al
culat
ing
how
of
te
n
diff
e
re
nt
com
bin
at
ion
s
of
pi
xel
with
gr
ey
le
vels
oc
cur
i
n
a
n
im
age.
The
G
LCM
m
et
hod
pe
rfor
m
s
bette
r
than
ot
her
te
xture
di
scrim
inati
on
m
et
hods
,
wh
ic
h
can
be
us
ed
t
o
so
lve
the
prob
le
m
of
sh
ape
va
rianc
e
in
each
so
yb
ean
seeds
cl
as
s
.
The
GLCM
sta
ti
sti
cs
with
reg
a
rd
to
the
t
extu
re
featur
e
s
are
ba
sed
on
the
t
wo
-
orde
r
sta
ti
sti
cal
par
am
et
ers
that
are
c
ontras
t,
energy,
c
orr
el
at
ion
,
a
nd
en
tro
py
[20].
T
he
te
xu
re
feat
ur
es
of
t
he
s
oybea
n
see
d
are
s
hown
in
Table
2.
Ha
ra
li
ck
[21]
has
pro
posed
1
4
ty
pe
s
of
GLCM
sta
ti
stics
to
de
scri
be
the
te
xture
fe
at
ur
es.
I
n
this
stud
y,
we
use
d
f
our
ty
pe
s
of
sta
ti
sti
cs,
nam
el
y,
con
t
rast,
e
ntropy,
co
rrel
at
ion
,
and
ene
r
gy,
to
extract
the
te
xt
ur
e
featur
e
s
of
so
y
bean
seed
s.
The
eq
uatio
ns
f
or
cal
culat
ing
the
se
featu
res
are
pr
ese
nte
d
in
Table
2.
Co
ntr
ast
m
easur
es
t
he
local
var
ia
ti
on
s
in
the
gra
y
-
le
vel
co
-
occ
urren
ce
m
at
rix.
Entr
op
y
m
easur
es
th
e
rand
om
ness
or
disor
der
of
the
im
age
area.
Correl
at
io
n
is
a
m
easur
e of ho
w
co
rr
el
at
ed
a
pix
el
to
it
s n
ei
ghbor ov
e
r
the en
ti
re i
m
a
ge.
Finall
y, ener
gy
m
easur
es the textu
ral
un
i
form
i
ty
o
f
t
he
im
age ar
ea.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19
:
3495
-
3503
3500
T
able
2
.
Form
ulas to
calc
ulate
textu
re feat
ures b
ase
d o
n G
LCM
Textu
re
f
eatu
re
Equ
atio
n
Co
n
trast
∑
∑
(
−
)
2
(
,
)
=
1
=
1
Entro
p
y
−
∑
∑
(
,
)
=
1
=
1
(
,
)
Co
rr
elatio
n
∑
∑
(
−
̅
)
(
−
̅
)
(
,
)
=
1
=
1
Energy
∑
∑
(
,
)
2
=
1
=
1
2.4.
Model
constr
uctio
n
A
suppo
rt
vector
m
achine
(SVM)
is
a
su
pe
rv
ise
d
le
arn
i
ng
al
go
rithm
that
is
m
os
tly
us
ed
for
vis
ual
patte
rn
re
co
gn
it
ion
an
d
im
ag
e
cl
assifi
cat
ion
[22,
23
]
.
T
he
obj
ect
ive
of
SV
M
cl
assifi
e
r
[
24
]
is
hype
r
plane
cl
assifi
er,
w
hich
determ
ines
a
n
optim
al
li
ne
to
se
par
at
e
t
he
trai
ning
set
of
the
two
cl
asses.
Fo
r
a
m
ulti
-
cl
ass
SV
M
cl
assifi
e
r
,
the
t
wo
-
cl
ass
es
sepa
rati
on
i
s
operate
d
as
one
-
a
gainst
t
o
a
ll
.
A
cl
assifi
er
m
od
el
is
con
st
ru
ct
e
d
to
assig
n
one
of
the
cl
asse
s
to
the
te
st
sam
ples.
In
this
st
ud
y,
the
SV
M
cl
assifi
er
was
us
e
d
t
o
cl
assi
fy
the
i
m
ages o
f
d
ise
ased soybea
n s
eeds
us
in
g
c
ol
or an
d
te
xtu
re
fe
at
ur
es a
nd a
poly
no
m
ia
l ker
ne
l.
3.
RESU
LT
S
A
ND AN
ALYSIS
In
this secti
on, we present the
resu
lt
s obtai
ne
d
by experim
ents u
sin
g
a col
or
histogram
.
Additi
on
al
ly
,
we
prese
nt
the
GLCM
sta
ti
stics
to
evaluate
t
he
accu
racy
of
the
pro
po
se
d
so
ybea
n
see
d
disease
cl
assifi
cat
ion
m
et
ho
d
f
or
q
ua
li
ty
gr
adin
g
.
E
ach
im
age
in
trai
nin
g
dataset
was
ad
de
d
an
d
reduced
of
bri
gh
t
le
vel
to
e
va
luate
rob
us
tness
of
l
igh
ti
ng
c
hange
s,
w
hich
co
ns
i
ste
d
of
five
br
igh
tne
ss
le
vels
.
The
re
fore,
t
he
re
are
6,6
00
i
m
ages
in d
at
aset
s
f
or
exp
e
rim
ents.
Figure
6
s
hows
so
yb
ea
n
see
ds
of f
ive
brig
htne
ss levels.
Figure
6. S
oybe
an
see
ds
of f
i
ve bri
gh
t
ness
l
evels
In
t
he
cl
assifi
c
at
ion
m
et
ho
d,
we
us
e
a
m
ulti
-
cl
ass
SV
M
c
la
ssifie
r
with
a
10
-
fo
l
d
cr
oss
validat
io
n
wh
ic
h
are
par
ti
ti
on
ed
i
nto
10
fo
l
ds
.
On
e
fo
l
d
is
us
e
d
f
or
t
est
ing
pr
ocess,
wh
il
e
the
rem
ai
nin
g
f
old
s
ar
e
us
e
d
for
trai
ni
ng
pr
ocess.
The
S
V
M
cl
assifi
er
is
exec
uted
10
t
i
m
es
and
us
es
each
fo
l
d
on
l
y
on
ce
.
T
her
e
fore,
the
acc
ur
acy
r
at
e
is
the
ave
r
age
acc
uracy
of
the
10
tim
es
execu
ti
on.
T
hi
s
stu
dy
def
i
ne
d
five
cl
ass
es
for
the
cl
assifi
cat
ion
of
s
oybea
n
se
ed
diseases:
norm
al
seeds,
pur
ple
seeds,
g
ree
n
see
ds
,
wr
i
nk
le
d
seed
s,
an
d
oth
e
r
see
ds
.
The
perf
or
m
ance
was
m
easur
e
d
i
n
te
r
m
s
of
preci
sion,
r
ecal
l,
an
d
the
F
1
-
m
easur
e
[
25]
,
and
c
om
par
ed
with
the
gro
und
-
trut
h
data
set
,
w
hich
wa
s
gen
e
rated
by
us
ers.
W
e
com
par
ed
the
color
histo
gr
am
us
ing
di
ff
e
ren
t
num
ber
s
of
col
or
bi
ns
of
the
HSI
m
od
el
a
nd
the
RGB
m
od
el
.
The
H
SI
a
r
e
cy
li
nd
rical
g
e
om
et
ries w
hose
angular
dim
ension
ranges
fro
m
0
° to
360° to
r
ep
rese
nt hue, st
arti
ng
with th
e re
d
pr
im
ary
at
0°
,
ye
ll
ow
pr
im
a
ry
at
0°,
gr
ee
n
pr
im
ary
at
120°
,
cy
an
pr
im
ary
at
180°,
bl
ue
pr
im
ary
at
240°,
m
agen
ta
pri
m
ary
at
300°
,
a
nd
t
hen
wr
a
pp
ing
back
t
o
re
d
at
360°.
T
he
refor
e
,
we
se
par
at
e
d
the
H
(Hue)
com
po
ne
nts
in
to
6
c
olo
r
bin
s
,
36
c
olor
bi
ns
,
72
c
ol
or
bi
ns
,
144
c
olo
r
bins,
288
col
or
bin
s
and
360
c
olor
bin
s
.
The
S
(S
at
ur
at
ion)
com
po
ne
nt
s
wer
e
sepa
rated
into
4
bin
s
and
t
h
e
I
(Inte
ns
it
y)
com
po
ne
nts
wer
e
se
pa
rated
into
4
bin
s.
Ea
ch
cha
nnel
of
the
RGB
m
od
el
was
separ
at
ed
into
2,
3,
4,
5,
6,
7
a
nd
8
bin
s.
T
he
be
st
colo
r
featur
e
was
c
om
bin
ed
with
the
GLCM
s
ta
ti
sti
cs.
The
cl
assifi
er
m
odel
s
we
re
dev
el
op
e
d
us
i
ng
th
e
SV
M
cl
assifi
er w
it
h a p
olyno
m
ia
l fu
nctio
n.
Table
3
s
hows
the
preci
sio
n,
r
ecal
l,
an
d
F
-
M
easur
e
f
or
s
oybean
cl
assifi
ca
ti
on
us
in
g
c
olor
feat
ur
es
of
th
e
HS
I
m
od
el
.
The
ex
per
im
e
ntal
resu
lt
s
ha
ve
fou
r
par
ts
with
dif
fer
e
nt
com
po
ne
nts
of
the
HS
I
m
od
el
.
Fr
om
the
ta
ble,
it
is
ob
s
er
ved
that
t
he
m
axi
m
u
m
F
-
m
easur
e
of
0.988
has
occurre
d
with
360
bin
s
of
H
com
pone
nts
(36
0
c
olors)
.
A
com
par
iso
n
of
the
s
oy
bean
cl
assi
ficat
ion
ac
cur
acy
of
the
RGB
c
olo
r
m
od
el
a
nd
the
H
S
I
c
olor
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Qualit
y g
r
adin
g
of s
oyb
e
an se
eds usin
g
i
mag
e ana
ly
sis (
Su
t
as
inee
Jit
anan)
3501
m
od
el
are
sho
wn
i
n
Fi
gure
7.
The
res
ults
ind
ic
at
e
that
c
ol
or
f
eat
ur
e
us
i
ng
H
c
om
po
ne
nts
of
the
HSI
m
od
el
giv
es
high
cl
as
sific
at
ion
acc
uracy
m
or
e
than
the
ov
e
rall
colo
r
featu
re
ba
sed
on
t
he
RG
B
m
od
e
l.
T
herefo
re,
the
HS
I
m
od
el
with
H
com
ponen
ts
dem
on
str
at
ed
that
robu
s
t
with
li
gh
t
chan
gi
ng
a
nd
us
e
d
lowest
dim
ension
s
of col
or
featu
r
es.
Table
3.
Res
ults o
f
s
oybea
n
cl
assifi
cat
ion
us
i
ng co
l
or f
eat
ur
es of the
HSI
m
od
el
Co
lo
r
Co
m
p
o
n
en
t
Co
lo
r
Featu
re
Precis
io
n
Recall
F
-
Measu
r
e
H
6
bin
s o
f
H
(6 co
lo
rs)
0
.81
3
0
.78
9
0
.79
1
3
6
bin
s o
f
H
(36
colo
rs)
0
.93
5
0
.93
0
.93
7
2
bin
s o
f
H
(72
colo
rs)
0
.95
9
0
.95
3
0
.95
3
1
4
4
bin
s o
f
H
(14
4
colo
rs)
0
.96
7
0
.96
1
0
.96
1
2
8
8
bin
s o
f
H
(28
8
colo
rs)
0
.98
4
0
.98
3
0
.98
3
3
6
0
bin
s o
f
H
(36
0
colo
rs)
0
.98
8
0
.98
8
0
.98
8
H and
I
6
bin
s o
f
H
,
4
bin
s o
f
I
(
2
4
colo
rs)
0
.84
6
0
.82
5
0
.82
7
3
6
bin
s o
f
H
,
4
bi
n
s o
f
I
(
1
4
4
colo
rs)
0
.93
3
0
.92
5
0
.92
6
7
2
bin
s o
f
H,
4 b
in
s o
f
I
(
2
8
8
colo
rs)
0
.95
2
0
.94
5
0
.94
6
1
4
4
bin
s o
o
f
H,
4 b
in
s o
f
I
(
5
7
6
color
s)
0
.96
3
0
.95
9
0
.95
9
2
8
8
bin
s o
f
H
,
4
bin
s o
f
I
(11
5
2
color
s)
0
.96
3
0
.96
1
0
.96
2
3
6
0
bin
s o
f
H
,
4
bin
s o
f
I
(
1
4
4
0
color
s)
0
.97
1
0
.96
9
0
.96
9
H and
S
6
bin
s o
f
H,
4 b
in
s o
f
S
(24
colo
rs)
0
.86
1
0
.84
0
.84
2
3
6
bin
s o
f
H
,
4
bi
n
s o
f
S
(14
4
colo
rs
)
0
.94
8
0
.94
3
0
.94
3
7
2
bin
s o
f
H,
4 b
in
s o
f
S
(28
8
colo
rs)
0
.96
2
0
.95
8
0
.95
8
1
4
4
bin
s o
f
H
,
4
bin
s o
f
S
(57
6
colo
r
s)
0
.96
7
0
.96
5
0
.96
5
2
8
8
bin
s o
f
H,
4 b
in
s o
f
S
(11
5
2
colo
r
s)
0
.98
6
0
.98
5
0
.98
5
3
6
0
bin
s o
f
H,
4 b
in
s o
f
S
(14
4
0
colo
r
s)
0
.98
5
0
.98
4
0
.98
4
All
6
bin
s o
f
H
,
4
bin
s o
f
S,
4
bin
s (96
c
o
lo
rs)
0
.87
5
0
.85
2
0
.85
5
3
6
bin
s o
f
H
,
4
bi
n
s o
f
S,
4 b
in
s (576 co
lo
rs)
0
.96
1
0
.95
7
0
.95
6
7
2
bin
s o
f
H,
4 b
in
s o
f
S,
4 b
in
s (11
5
2
colo
rs)
0
.96
9
0
.96
5
0
.96
5
1
4
4
bin
s o
f
H,
4 b
in
s o
f
S,
4 b
in
s (2304
c
o
lo
rs)
0
.96
8
0
.96
6
0
.96
6
2
8
8
bin
s o
f
H,
4 b
in
s o
f
S,
4 b
in
s (4608
colo
rs)
0
.98
0
.97
9
0
.97
9
3
6
0
bin
s o
f
H,
4 b
in
s o
f
S,
4 b
in
s
(5760
colo
rs)
0
.98
4
0
.98
4
0
.98
4
Figure
7. A
com
par
ison
of th
e cla
ssific
at
ion acc
uracy
of t
he
RGB c
olor m
od
el
a
nd the
HSI
c
olor m
od
el
Table
4
s
how
s
the
cl
assifi
c
at
ion
acc
ur
acy
us
in
g
t
he
tw
o
feat
ur
e
set
s
and
an
SV
M
cl
assifi
er.
F
-
m
easur
e
of
0.8
66
wa
s
ob
ta
i
ne
d
with
GLCM
base
d
te
xt
ur
e
f
eat
ur
e
w
her
eas
F
-
m
easur
e
of
0.992
was
obta
ined
with
col
or
a
nd
te
xtu
re
featu
re
s,
w
hich
was
be
st
cl
assifi
er
m
od
el
.
T
he
resu
l
t
il
lustrate
s
that
the
cl
assifi
er
base
d
on
col
or
an
d
t
extu
re
fe
at
ur
e
can
ide
ntify
s
oybea
n
see
ds
with
highest
a
ccur
acy
a
nd
i
m
pr
ov
e
cl
assif
ic
at
ion
perform
ance f
r
om
u
sing col
or f
eat
ure
on
ly
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19
:
3495
-
3503
3502
Table
4.
C
om
par
iso
n of ef
fici
ency o
f
cl
assifi
cat
ion
of d
if
fe
r
ent f
eat
ur
e
Featu
re
Precisio
n
Recall
F
-
Measu
r
e
Co
lo
r
: 36
0
bin
s o
f
H (
3
6
0
colo
rs)
0
.98
8
0
.98
8
0
.98
8
Textu
re
:
GLCM
0
.86
8
0
.86
6
0
.86
6
Co
lo
r
an
d
T
ex
tu
re
0
.99
3
0
.99
2
0
.99
2
T
he
cl
assifi
cat
ion
pe
rfor
m
ance
for
eac
h
s
oybean
cl
ass
us
i
ng
the
best
cl
as
sifie
r
m
od
el
w
as
show
n
i
n
Table
5
.
The
F
-
m
easur
e
f
or
the
cl
assifi
cat
ion
of
norm
al
seeds,
pu
rp
le
s
eeds,
green
s
eeds,
w
rin
kled
seeds
,
and
o
the
r
seed
s
was
0.9
79,
1
,
1
,
0.9
81,
an
d
1
res
pecti
vely
.
Du
e
to
t
he
dist
inct
color
d
if
fe
ren
ce of
p
urple
seeds
and
green
see
ds
,
it
m
akes
hi
gh
cl
assifi
cat
ion
ac
cu
racy.
More
ov
e
r,
o
th
er
seed
s
we
re
cl
assifi
ed
wit
h
hi
gh
accuracy
beca
us
e
t
hey
co
ns
i
ste
d
of
fro
geye
seeds
,
dr
ie
d
seeds
,
a
nd
da
m
aged
see
ds
,
wh
ic
h
hav
e
di
ff
e
ren
t
colo
rs
an
d
te
xtu
re
in
c
om
par
ison
t
o
oth
e
r
s
oybea
n
cl
asses
,
wh
il
e
norm
al
seeds
an
d
w
rink
le
d
seeds
ha
ve
a
si
m
il
ar
yellow
color.
H
ow
e
ve
r,
the
best
cl
assifi
er
m
od
el
can
be
us
ed
to
cl
assify
so
ybean
see
ds
with
a
n
aver
a
ge
acc
ura
cy
o
f
99.
2
%.
Table
5.
Res
ults o
f
s
oybea
n
cl
assifi
cat
ion
us
i
ng co
l
or f
eat
ur
e
an
d
te
xt
ur
e
fe
at
ur
e
So
y
b
ean class
Precisio
n
Recall
F
-
Measu
r
e
No
r
m
al
seed
s
1
0
.96
0
.97
9
Pu
rple seed
s
1
1
1
Gree
n
seed
s
1
1
1
W
rink
led
seed
s
0
.96
5
1
0
.98
1
Oth
er
seed
s
1
1
1
Av
erage
0
.99
3
0
.99
2
0
.99
2
4.
CONCL
US
I
O
N
We
pr
e
sente
d
a
fram
ewo
rk
f
or
so
y
be
a
n
qual
it
y
gr
adin
g
wh
ic
h
has
t
hr
e
e
m
ai
n
adv
a
ntages.
First
t
he
us
e
of
bac
kgr
ound
subtract
i
on
reduces
th
e
prob
le
m
of
a
sh
ad
ow
ap
pe
arin
g
in
the
captu
red
im
age
wh
e
n
changin
g
cam
era
an
gle
an
d
the
conditi
on
of
na
tural
li
ght
,
co
ns
e
qu
e
nt
ly
,
i
m
pr
ov
in
g
the
res
ults
o
f
seeds
segm
entat
ion
and
c
rop
ping.
S
econd
a
m
et
ho
d
is
propose
d
f
or
e
xtracti
on
colo
r
featu
re
ba
sed
on
r
obus
t
ne
ss
for
il
lu
m
inati
on
ch
ang
e
s
w
hich
is
H
com
po
ne
nts
in
HSI
m
od
el
.
Thir
d
an
a
ppr
oach
t
o
fin
d
oth
er
featu
res
to
so
lve
sh
a
pe
var
ia
nce
in
eac
h
s
oybe
an
se
e
ds
cl
ass
is
prese
nted.
Ba
sed
on
the
exp
e
rim
ental
resu
lt
s,
th
e
pro
po
s
e
d
te
chn
iq
ue
com
bin
es
a
c
olor
hi
stog
ram
of
H
com
po
ne
nts
in
HSI
m
od
el
a
nd
the
G
LCM
s
ta
ti
sti
cs
to
be
m
or
e
i
m
pr
ove
cl
assifi
cat
ion
acc
uracy
than
us
i
ng
col
or
feature
on
ly
.
F
uture
work
will
foc
us
on
desig
ni
ng
a
com
pr
ehe
ns
ive
cl
assifi
er
f
or
var
i
ou
s
s
oybea
n
see
d
ty
pe
s
a
nd
the
pro
pose
d
m
et
ho
d
can
be
c
om
bin
ed
with
a
harvester
m
achine.
ACKN
OWLE
DGE
MENTS
We ac
knowle
dge the
f
i
nan
ci
a
l su
ppor
t
of
Na
resu
a
n U
niv
e
rs
it
y.
REFERE
NCE
S
[1]
V.
Kum
ar,
et
al
.
,
“
A
c
om
par
at
iv
e
assess
m
ent
of
tot
al
ph
enol
i
c
c
onte
nt
,
fer
ri
c
re
duci
ng
-
an
ti
-
oxid
at
iv
e
power,
fre
e
rad
ical
-
sc
ave
ngi
ng
ac
t
ivi
t
y
,
vita
m
in
C
and
isofl
avone
s
content
i
n
so
y
be
an
with
var
y
ing
see
d
co
at
co
lour,”
Foo
d
Re
search
In
te
rn
ati
onal
,
vol
/i
ss
ue:
43
(
1
)
,
pp
.
323
-
328
,
2010
.
[2]
M.
Carmona,
et
al
.
,
“
Deve
lopment
and
va
li
da
tion
of
a
fungic
id
e
scoring
s
y
st
e
m
for
m
ana
gement
of
late
sea
s
on
so
y
bea
n
dise
ase
s in
Argen
ti
n
a,”
Crop P
rotection
,
vol. 70, pp. 83
-
91,
2015
.
[3]
T.
Brosnan
and
D.
W
.
Sun,
“
Im
proving
qualit
y
i
nspec
ti
on
of
foo
d
p
roduc
ts
b
y
co
m
pute
r
vision
-
a
rev
ie
w,
”
Journal
of
Food
Engi
n
eer
ing
,
vol
/i
ss
ue:
61
(
1
)
,
pp
.
3
-
16
,
2
004.
[4]
H.
Sabrol
and
S.
Kum
ar,
“
Rec
ognit
ion
of
T
om
at
o
La
t
e
Bli
ght
b
y
using
DW
T
and
Compone
nt
Anal
y
sis
,
”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ri
ng
(
IJE
CE)
,
vol
.
7
,
pp
.
194
-
199,
2017.
[5]
R.
R.
Parm
ar,
et
al
.
,
“
Unifie
d
app
roa
ch
in
food
qu
al
ity
evalua
t
ion us
ing
m
ac
hine
v
ision,
” in
Adv
an
ce
s in
Computin
g
and
Comm
unic
a
ti
ons
,
pp
.
239
-
24
8
,
2011
.
[6]
S.
Li
li
k
,
et
al
.
,
“
Digit
al
Im
age
B
ase
d
Ide
nti
f
ic
a
tion
of
Ric
e
V
ari
e
t
y
Us
ing
Im
age
Proce
ss
ing
and
Neura
l
Network,”
TEL
KOMNIKA
(Tele
communic
a
t
ion,
Comput
ing,
El
e
ct
ronics
and
Control)
,
vol. 1
6
, p
p
.
182
-
190
,
2
015.
[7]
C.
J.
Du
and
D.
W
.
Sun,
“
Le
arn
ing
te
chn
ique
s
used
in
computer
vision
for
food
qual
ity
ev
al
ua
ti
on:
A
rev
ie
w
,
”
Journal
of
Food
Engi
ne
ering
,
vol
.
72
,
pp
.
39
-
55
,
2006.
[8]
A.
Ta
nnou
che
,
et
a
l
.
,
“
A
fast
a
nd
eff
icien
t
sha
pe
desc
rip
tor
fo
r
an
adv
anced
wee
d
t
y
pe
class
ifi
cation
a
pproa
c
h,
”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
vol
.
6
,
p
p
.
1168
-
1175,
201
6.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Qualit
y g
r
adin
g
of s
oyb
e
an se
eds usin
g
i
mag
e ana
ly
sis (
Su
t
as
inee
Jit
anan)
3503
[9]
H.
K.
Meba
tsio
n,
et
al
.
,
“
Autom
at
ic
c
la
ss
ifi
c
ation
of
non
-
touching
ce
r
ea
l
gr
ains
in
digi
tal
images
using
li
m
ited
m
orphologi
ca
l
a
nd
col
or
fe
at
ur
es,
”
Comput
ers an
d
Elec
troni
cs
in
Agric
ult
ure
,
vo
l. 90, pp. 99
-
105,
2013.
[10]
M.
Olgun
,
et
al
.
,
“
W
hea
t
gr
ai
n
c
la
ss
ifi
c
at
ion
b
y
using
dense
SIF
T
feature
s
with
SV
M
cl
assifie
r,
”
Computers
and
El
e
ct
ronics
in A
gricul
ture
,
vol
.
1
22,
pp
.
185
-
190
,
2016.
[11]
R.
D.
L.
Pires
,
et
al
.
,
“
Local
desc
ript
ors
for
soy
b
ea
n
d
iseas
e
rec
ognition,
”
Computers
an
d
El
ectronics
i
n
Agric
ult
u
re
,
vo
l. 125, pp. 48
-
55,
2016.
[12]
K.
Z.
Ta
n
,
et
a
l
.
,
“
Ide
nti
fi
ca
t
ion
of
disea
ses
for
s
o
y
be
an
see
ds
b
y
computer
vision
apply
ing
BP
neur
al
ne
twork
,
”
Inte
rnational
Jo
urnal
of Agric
ul
t
ural
and
B
iol
ogi
cal
Engi
ne
ering
,
vol. 7, pp. 43
-
5
0,
2014
.
[13]
W
.
Li
ang,
et
a
l
.
,
“
Es
ti
m
at
ion
of
so
y
bea
n
leaf
area,
edg
e,
and
d
ef
oli
ation
using
c
olor
image
anal
y
sis,
”
Computers
and
Elec
troni
cs
in
Agri
cul
ture
,
v
ol.
150
,
pp
.
41
-
5
1,
Jul
2018
.
[14]
M.
A.
Mom
in,
et
al
.
,
“
Mac
hi
ne
vision
base
d
soy
b
ea
n
qu
alit
y
evalua
t
ion,”
Computers
an
d
El
ectronics
i
n
Agric
ult
ure
,
vo
l. 140, pp. 452
-
46
0,
Aug 2017
.
[15]
Y.
Bene
zeth,
e
t
al
.
,
“
Review
and
eva
luation
o
f
comm
only
-
im
ple
m
ent
ed
ba
ck
ground
subtrac
tion
al
gorit
hm
s,”
in
2008
19
th
In
t
ernati
onal
Confer
enc
e
on
Pa
tt
ern
Recogni
t
ion
,
pp
.
1
-
4
,
2008
.
[16]
S.
Z.
L
i
and
A.
Jain,
Eds.
,
“
Local
Adapt
ive
Thr
esholdi
ng,
”
in
E
ncy
c
lope
dia
o
f
Bi
ometri
cs
,
Bost
on,
MA
:
Springer
US
,
pp.
939
-
939
,
2009
.
[17]
R.
C.
Gonza
lez,
et
al
.
,
“
Digit
al I
m
age
Proce
ss
ing
Us
ing
MA
TL
AB
,”
Pe
arson
Edu
ca
t
ion, 2004.
[18]
X.
Y.
W
ang,
et
al
.
,
“
Robust
ima
ge
ret
r
ie
v
al
b
ase
d
on
col
or
h
i
sto
gra
m
of
loc
a
l
fe
at
ure
r
egi
ons
,
”
Mult
imedi
a
Tools
and
Applications
,
vol
/i
ss
ue:
49
(
2
)
,
pp
.
323
-
345
,
2
010.
[19]
S.
Pare
,
e
t
al
.
,
“
An
opti
m
al
col
o
r
image
m
ultil
ev
el
thre
sholding
t
ec
hniqu
e
using
gre
y
-
le
v
el
co
-
oc
cur
ren
c
e
m
at
r
ix,”
Ex
pert
Syste
ms
wit
h
App
licati
on
s
,
vol. 87, pp. 33
5
-
362,
Nov
2017
.
[20]
M.
A.
Ta
hir,
et
al
.
,
“
Acc
e
le
ra
ting
the
computat
ion
of
GLCM
and
Hara
lick
texture
feature
s
on
rec
onfigura
b
l
e
har
dware
,
” in
20
04
Inte
rnat
ional
Confe
renc
e
on
I
mage
Proce
ss
in
g,
2004
.
ICIP
’0
4,
v
ol
.
5
,
pp
.
285
7
-
2860
,
2004
.
[21]
R.
M.
H
ara
l
ic
k
,
et
al
.
,
“
Te
x
tur
al
fe
at
ure
s
for
i
m
age
cl
assifi
ca
t
ion,
”
I
EE
E
Tr
ansacti
ons
on
Sy
stems,
Man,
an
d
Cybe
rnet
ic
s
,
vol
/i
ss
ue:
SMC
-
3
(
6
)
,
pp
.
610
-
621
,
Nov
1973.
[22]
F.
Avila
,
et
al
.
,
“
A
m
et
hod
to
construc
t
fru
it
m
at
urity
co
lor
sca
l
es
base
d
on
support
m
ac
hin
es
fo
r
reg
ressio
n:
Applic
a
ti
on
to
o
l
ive
s a
nd
gra
p
e
s
ee
ds,”
Journal
o
f
Food
Engi
n
ee
r
ing
,
vo
l. 162, pp. 9
-
17,
Oct
2
015.
[23]
F.
Melga
ni
and
L.
Bruz
zone,
“
Cla
ss
ifi
ca
t
ion
of
h
y
per
sp
ec
tr
a
l
remote
sensin
g
images
with
support
vec
tor
m
ac
hine
s,”
IE
E
E
Tr
ansacti
ons
on
Geosci
en
ce a
nd
Re
mot
e
S
ensi
ng
,
vol
/i
ss
ue:
42
(
8
)
,
pp
.
1778
-
17
90,
Aug
2004
.
[24]
M.
K.
S.
Varm
a,
et
al
.
,
“
Pixel
-
base
d
c
la
ss
ifi
c
ation
using
suppo
rt
ve
ct
or
m
ac
hi
ne
c
la
ss
ifi
er
,
”
i
n
2016
IE
EE
6t
h
Inte
rnational
Co
nfe
renc
e
on
Advance
d
Comput
in
g
(
IACC)
,
pp.
51
-
55
,
2016
.
[25]
C.
Sam
m
ut
and G.
I. W
ebb,
“
Ency
cl
op
edi
a
of
M
ac
hin
e
L
ea
rn
ing
,
1st e
d
.
Spring
er P
ubli
shing Compan
y
,
In
c.,
2011
.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Su
tasinee
Jitanan
,
she
rec
ei
ved
her
B.
Sc.
in
Com
pute
r
Scie
nce
from
Chia
ng
Mai
Univer
sity
in
1996
and
her
M.Sc.
in
Inform
at
io
n
Te
chnol
ogy
from
King
Mongkut’s
Instit
ute
of
Te
chnol
ogy
La
dkra
bang
in
2001.
She
com
ple
te
d
Ph.D
degr
ee
in
Com
pute
r
Engi
nee
ring
from
Kaset
sart
Univer
sity
in
y
ea
r
2015
.
Curre
ntl
y
,
she
is
working
as
le
ct
ure
r
in
depa
rtment
of
Com
pute
r
Scie
nce
and
Inform
at
ion
Te
chnol
ogy
,
Facul
ty
of
Scie
nce
,
Nare
suan
Univer
sity
,
Phitsanul
ok,
Tha
il
and.
Her
rese
arc
h
int
ere
sts inc
lude
pat
te
rn
ana
ly
sis a
nd
m
ult
imedia
proc
essing.
Paw
at
Ch
iml
ek
,
he
rec
ei
ved
h
is
B.
Sc.
in
Com
pute
r
Scie
nce
from
Pibulsongkram
Raj
abha
t
Un
ive
rsity
and
his
M.Sc.
in
Inform
at
ion
Te
chnol
ogy
from
Nare
suan
Univer
sity
.
He
is
cur
ren
tl
y
le
ct
ure
r
in
depa
rtment
of C
om
pute
r
Scie
nce
and
Inform
at
ion Te
chnol
ogy
,
Facul
ty
of
Scie
nce
and
Te
chnol
ogy
,
Pibulsongkram
Raj
abha
t
Univer
sity
,
Phitsanul
ok,
Tha
il
and.
H
is
rese
arc
h
int
ere
sts
inc
lude
informati
on
ret
rie
val
.
Evaluation Warning : The document was created with Spire.PDF for Python.