Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 3
,
Ju
n
e
201
6, p
p
. 1
168
~ 11
75
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
3.9
978
1
168
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
A Fast and Efficient Shape Desc
riptor for an Ad
vanced W
e
ed
Type Classifica
ti
on Approach
Adil Tannouc
he, K
h
alid
Sb
ai,
Miloud
Rahmou
ne, Ami
n
e
Z
o
ubir, Rachid
Agounoune,
Ra
chid Saa
d
ani, Abdela
li
Ra
hma
n
i
Laboratoire d’Etude des M
a
tér
i
aux Avancés
et A
pplic
ations,
FS-EST, Moul
a
y
Is
m
a
il Universi
t
y
,
BP 11201, Zitou
n
e, Mekn
es, Morocco
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Ja
n
2, 2016
Rev
i
sed
Mar
11
, 20
16
Accepted
Mar 26, 2016
In weed management, the d
i
stinction be
tween monocots and dicots species is
an im
portant is
s
u
e. Inde
ed, th
e
yield is
m
u
ch hig
h
er with the
app
lic
ation of a
selec
tive tr
eatm
e
nt instead of using a
broadcast
herbicide overall the parcel.
This article pr
esents a f
a
st shape desc
riptor
desig
n
ed to d
i
stinguis
h
between
thes
e two fam
ili
es
of weeds
.
The effic
i
en
c
y
of t
h
e des
c
riptor is
evalu
a
ted
b
y
anal
yz
ing d
a
ta
with th
e p
a
ttern
recogn
itio
n process
kno
wn as th
e
discriminant f
a
ctor analy
s
is (DFA). Ex
cellent results have been
obtain
e
d in
the d
i
ffer
e
nti
a
tio
n betwe
e
n
these
two weed
speci
e
s
.
Keyword:
Machine Vision
Real-tim
e I
m
a
g
e Processing
Weed Type
Classification
Precision Agri
culture
Sha
p
e Descri
pt
or
Copyright ©
201
6 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
A
d
il Tann
ou
che,
Depa
rtem
ent of Electrical a
nd Co
m
p
u
t
er
Engin
eer
ing
,
LEM
2
A
, Hi
g
h
Sch
ool
o
f
Tec
h
nol
ogy
,
M
e
knes
,
M
o
r
o
cco, +
2
12
6
5
0
9
60
2
0
7
.
Em
ail: tannouc
he@gm
a
il.co
m
1.
INTRODUCTION
The wee
d
control is a key ele
m
en
t in s
m
art agric
u
lture. The chem
ical weed control
is widely
pract
i
ced i
n
or
der t
o
re
d
u
ce t
h
e i
n
fest
at
i
o
n rat
e
s and t
o
i
m
prove ha
rve
s
t
s
.
W
eeds a
r
e di
vi
de
d i
n
t
o
t
w
o
bi
g
fam
i
l
i
e
s:
t
h
e M
o
n
o
c
o
t
y
l
e
do
no
us fam
i
l
y
whi
c
h i
s
cha
r
act
eri
zed
by
l
ong a
nd t
h
i
n
l
eaves;
an
d t
h
e
Dico
tyled
ono
us fam
i
l
y
with
sho
r
t leav
es (see Fi
g
u
re
1
)
. In
v
i
ew
of th
is
d
i
v
e
rsity, th
e weed
co
n
t
ro
l
p
e
rform
a
n
ce is ev
en
b
e
tter
with
th
e
application
of a select
ive treatm
e
nt
instead
of
usi
n
g a single broadcast
herbicide ove
r
all the parcel.In this
re
gard,
recent resea
r
ches in com
puter
vision
ha
ve given
birth to s
e
veral
efficient techniques for dete
ction an
d /
or
cl
assi
fi
cat
i
on of
weeds
.
Tra
d
i
t
i
onal
l
y
, t
w
o
m
a
i
n
appr
oac
h
es are
use
d
:
A sp
ectral appro
ach
: th
is appro
a
ch
con
s
ists in
d
ecod
i
ng
the sp
ectral in
fo
rmatio
n
to
detect the presence
of
weed i
n
the
pa
rcels. In [1] a
nd
[2
], res
p
ectively, the aut
h
ors
use
d
th
e i
n
form
ation re
ve
aled by the
ne
ar
infra
re
d (
N
IR
)
and
the sim
p
le (RGB
)
pictures t
o
dete
ct weeds
.
In
[3],
the
aut
h
ors used t
h
e
UV
fluoresce
n
ce s
p
ectrum
.
In [4]
,
the au
t
h
ors a
n
alyzed the
hypers
pectral imag
es to
properl
y
detect and se
lect
t
h
e wee
d
s
.
Des
p
i
t
e
hi
s
per
f
o
r
m
ance, t
h
i
s
t
e
c
hni
que
re
q
u
i
r
e
s
an
ex
pe
nsi
v
e
equi
pm
ent
.
A sp
atial ap
pro
a
ch
: th
is tech
n
i
q
u
e
fo
cu
ses o
n
th
e d
i
st
ribu
tio
n
o
f
weed
s in
th
e parcel o
r
on
their
m
o
rph
o
lo
gical fo
rm
s [5]
in o
r
der t
o
ide
n
tify
them
. In
[6
] and
[7
],
th
e au
th
or
s
d
e
tected we
eds by
obse
rvi
ng
their presence i
n
the see
d
line spacing
.
In
[8
], th
e au
tho
r
s exa
m
in
ed
th
e sh
ap
e of weed
s b
y
u
s
ing
th
e seven
m
o
m
e
nt
s of
H
u
[
9
]
an
d si
x s
h
ape
desc
ri
pt
o
r
s t
o
ac
hi
ev
e a
better selection of
weeds
.
T
h
e results are
ve
ry
satisf
acto
r
y i
n
sp
ite of
th
e processin
g
ti
m
e
o
f
abo
u
t
t
w
o fr
ames p
e
r
second
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
A Fast and
Efficien
t
Sh
ap
e
Descrip
t
o
r
fo
r
an
Ad
va
n
c
ed
Weed
Typ
e
Cl
a
ssifica
tio
n .... (Ad
il Tann
ou
ch
e)
1
169
In
t
h
is stud
y,
we presen
t a new inn
o
v
a
tiv
e, fast
and efficient approac
h
fo
r t
h
e selection of
weeds
speci
es.
Ou
r a
p
p
r
oach
base
d
essent
i
a
l
l
y
by
t
h
e ap
pl
i
cat
i
on o
f
bi
na
ry
de
scri
pt
o
r
desi
g
n
e
d f
o
r
t
h
i
s
pu
r
pos
e.
This descri
ptor
called Adja
cencies Descri
ptor
ret
u
rn
s
the
num
ber
of horiz
ontal, vertical and
diagonal
adjace
nci
e
s
fo
r
a gi
ve
n 2
D
o
b
ject
.
Thi
s
t
e
c
hni
que al
l
o
w
di
st
i
n
g
u
i
s
hi
ng
bet
w
ee
n t
h
e s
h
ape
s
wi
t
h
a
r
o
u
n
d
ed
m
o
rph
o
l
o
gy
(
D
i
c
ot
)
an
d
ot
h
e
rs
wi
t
h
a l
o
n
g
t
h
i
n
m
o
r
p
h
o
l
o
gy
(M
on
oc
ot
).
Fi
gu
re
1.
W
e
e
d
sam
p
l
e
s:
m
onoc
ot
i
n
fi
r
s
t
l
i
n
e a
n
d
di
c
o
t
i
n
seco
n
d
l
i
n
e
2.
MATE
RIAL
AN
D METH
OD
The im
aging s
y
ste
m
is co
m
pose
d
of a standard RG
B cam
e
r
a. The cam
era held in vertica
l
position at
25 t
o
3
0
cm
above t
h
e re
gi
o
n
of i
n
t
e
rest
. T
h
us, t
h
e vi
si
bl
e scene c
ove
rs a
n
area
o
f
5
0
×
50 cm
2
(see Figu
re
2
)
.
Th
is set
u
p allows
u
s
t
o
o
v
e
rco
m
e th
e p
e
rsp
e
ctiv
e v
i
ew prob
lem
an
d
im
p
r
o
v
e
th
e
sp
atial
reso
l
u
tio
n.
Fi
gu
re
2.
Ac
q
u
i
si
t
i
on p
r
oc
ess:
cam
era on a
t
r
i
p
o
d
at
a
p
p
r
oxi
m
a
t
e
ly
0.3
m
hei
ght
poi
nt
i
n
g
v
e
rtically d
o
wn
ward
2.
1.
The Adjacenc
i
es Descriptor
Our re
gion-bas
ed desc
ript
or c
a
lculates the num
b
er
of ho
ri
z
ont
al
, vert
i
cal
and di
ag
o
n
al
adjace
nci
e
s
bet
w
ee
n a gi
v
e
n o
r
i
g
i
n
al
pi
xel
and t
h
ei
r adjace
nt
o
n
es
(see Figure
3). The origin
al
pixel (green
cell), is
sur
r
o
u
n
d
e
d
by
ei
ght
ot
he
r peri
phe
ral
pi
xel
s
(
y
el
l
o
w
an
d bl
u
e
cel
l
s
).
C
e
l
l
s
i
n
y
e
l
l
o
w
sh
ow
h
o
r
i
z
o
n
t
a
l
and
ve
rt
i
cal
adjace
nci
e
s a
n
d t
h
e
bl
ue
on
es
sh
ow
di
a
g
onal
l
y
adjace
ncies with respect
t
o
t
h
e orig
in
al cell.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E
V
o
l
.
6,
No
. 3,
J
u
ne 2
0
1
6
:
11
6
8
– 11
75
1
170
Fi
gu
re
3.
Desc
ri
pt
o
r
st
r
u
ct
u
r
e
The p
r
o
p
o
se
d descri
pt
o
r
cal
cul
a
t
e
s t
w
o n
u
m
bers of adja
c
e
nci
e
s bet
w
ee
n
a gi
ven o
r
i
g
i
n
al
pi
xel
(C
0)
and t
h
ei
r a
d
jac
e
nt
pi
xel
s
(C
A
,
suc
h
as
A =
1,
2,
3,
4,
5,
6
,
7,
8
)
. T
h
e fi
rst
i
s
t
h
e n
u
m
b
er
of
h
o
ri
z
ont
al
an
d
vertical adjace
ncies (N
HV
), t
h
e seco
nd i
s
t
h
e
num
ber
of
di
a
g
o
n
al
a
d
jace
nc
i
e
s (N
D
).
Th
e ad
j
acen
c
y (Ad
j
) is th
e
resul
t
of
t
h
e bi
nary
o
p
erat
o
r
XN
OR
bet
w
ee
n
t
h
e
origi
n
al c
e
ll and a
n
adja
cent cell as
foll
owi
ngs:
A
d
j
(
C
A)
= C0 (
XNO
R)
CA
(
1
)
There
b
y
,
i
n
a
bi
na
ry
i
m
age, t
h
e ad
jace
ncy
num
bers
f
o
r a
gi
ve
n
ori
g
i
n
al
pi
xel
C
(
x,y
)
a
r
e cal
cul
a
t
e
d
according t
o
the followi
ng form
ulas:
N
HV
=
A
d
j(C(
x-ste
p
, y
)) +
A
d
j(C(
x+step, y)
) +
Adj(C(x
,
y
-
step
)) +
Ad
j(C
(
x, y+step))
(2)
N
D
=
Ad
j(C(x-step
,y
-step)) + Ad
j(C(x
-
step
,y+step
)
) +
Adj(C(x
+
step
,y
-step
)
)
+A
dj(
C
(x
+step,
y+step
)) (
3
)
App
l
yin
g
th
is d
e
scrip
t
o
r
on all p
i
x
e
ls b
e
l
o
ng
ing
t
o
on
e p
a
rticu
l
ar
ob
ject, allo
w to
d
i
stin
gu
ish
b
e
tween
sh
apes with
roun
ded
and
filled
m
o
rp
ho
log
y
(Dico
t
) and
o
t
h
e
rs
sh
ap
es
with
a lo
ng
and
th
in
m
o
rph
o
l
o
gy
(
M
on
oc
ot
).
The Step
varia
b
le m
u
st then be ada
p
ted to t
h
e re
so
lu
tion
of th
e obj
ect to
d
e
scri
b
e
. Ind
e
ed
, a sm
all
v
a
lu
e
retain
s
on
ly n
o
i
se, wh
i
l
e to
o
larg
e
v
a
lu
e en
co
m
p
asses th
e obj
ect
with
ou
t defin
i
n
g
its ch
aracteristics.
Good
resu
lts are ob
tain
ed using
th
e fo
llowing em
p
i
rical fo
rm
u
l
a
:
(4
)
Whe
r
e
S is the
area (i
n
pixels
) of
th
e obj
ect to
d
e
scr
i
b
e
.
2.
2.
Implementation
In d
i
g
ital i
m
a
g
e seg
m
en
tati
o
n
app
licatio
ns, clu
s
te
ri
n
g
t
echni
que
i
s
u
s
ed t
o
se
gm
ent
re
gi
o
n
s
o
f
in
terest and
to
d
e
tect bo
rd
ers
o
f
ob
ject
s i
n
a
n
i
m
age [1
0]
.
The
gra
d
i
e
nt
m
a
gni
t
ude
an
d
co
here
nce i
s
use
d
t
o
segm
ent
fi
nge
r
p
ri
nt
im
age [1
1]
. I
n
o
u
r
ap
pl
i
cat
i
on, t
h
e
o
b
t
ai
ned i
m
age is segm
ent
e
d i
n
or
der t
o
i
s
ol
at
e t
h
e
v
e
g
e
tation
of th
e rest of th
e scen
e. Accord
i
n
g
to
[1
2
]
,
th
is
o
p
e
ration
is effectiv
ely carried
ou
t b
y
th
resho
l
d
i
ng
t
h
e i
m
age gi
ve
n
by
t
h
e
f
o
l
l
o
w
i
ng
f
o
rm
ul
a:
G
r
a
y
= r
× R + g
× G +
b
×
B
(5
)
Wi
t
h
:
r
=
−
0
.
88
4,
g
=
1
.
26
2 an
d b =
−
0.
31
1.
Whe
r
e R
,
G a
n
d B
rep
r
ese
n
t
r
e
d, g
r
ee
n an
d
bl
ue c
o
m
pone
nt
s of eac
h pi
x
e
l
.
Thu
s
, t
h
e pi
xel
s
rel
a
t
e
d
to
th
e
v
e
g
e
tatio
n are ob
tain
ed
for: Gray >
30
(see Fi
g
u
re
4).
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
A Fast and
Efficien
t
Sh
ap
e
Descrip
t
o
r
fo
r
an
Ad
va
n
c
ed
Weed
Typ
e
Cl
a
ssifica
tio
n .... (Ad
il Tann
ou
ch
e)
1
171
The ap
pl
i
cat
i
o
n o
f
desc
ri
pt
o
r
i
s
real
i
zed o
n
t
h
e segm
ent
e
d i
m
age by
assum
i
ng a [3
2
×
32]
pi
xel
s
sliding
window. T
h
en, the results retu
rne
d
by the descri
ptor are classi
fi
ed accordi
ng t
o
DFA disc
rimination
fu
nct
i
o
n (6
). A
m
a
jori
t
y
of vo
t
e
s pro
v
i
d
e
d
b
y
t
h
e sl
i
d
i
ng w
i
nd
ow
obt
ai
ns
t
h
e fi
nal
cl
assi
fi
cat
i
on of a reg
i
on i
n
the im
age.
Fi
gu
re
4.
Se
gm
ent
a
t
i
on
Ve
get
a
t
i
on /
gr
o
u
n
d
,
veget
a
t
i
o
n a
n
d
t
h
e rest
o
f
t
h
e
s
cene a
r
e r
e
p
r
es
ent
e
d
res
p
ect
i
v
el
y
by
w
h
i
t
e
an
d
b
l
ack pi
xel
s
3.
RESULTS
A
N
D
DI
SC
US
S
I
ONS
The
Local Bi
n
a
ry
Patter
n
(L
BP)
[1
3]
ins
p
i
r
es
ou
r
desc
rip
t
or.
It allo
ws
d
e
scribin
g
the
s
i
ze an
d the
m
o
rph
o
l
o
gi
cal
f
o
rm
of
wee
d
s
by
t
w
o i
n
t
e
gers
N
D
and
N
HV
. T
o
asse
s
s
and e
v
aluate
its perform
ance, we
designe
d
a dat
a
base constitut
e
d of 40 im
age
s
of the m
o
st
freque
n
t weeds. Then, each
image is shifted and
r
o
tated r
a
ndomly to
in
cr
ease th
e size
o
f
ou
r d
a
tab
a
se.
Fin
a
lly, th
ese imag
es ar
e stan
dar
d
ized
to th
e
w
o
r
k
i
ng
size [32 * 32],
For DFA m
e
thod, m
o
re
than
100 s
u
bjects a
r
e suggeste
d
,
but
according to
[14] the
ge
neral
rule
is to
h
a
v
e
a
ratio
o
f
10
subj
ects p
e
r v
a
riab
le
in
serted
i
n
th
e
an
alysis. For a to
tal o
f
86
im
a
g
es,
we u
s
ed
half o
f
t
h
e i
m
ages fo
r
l
earni
ng
an
d t
h
e
ot
he
r h
a
l
f
t
o
t
h
e
t
e
st
. T
h
e
pr
oce
d
ure
of
DF
A
was
per
f
o
rm
ed by
m
eans
o
f
XLST
AT
so
ft
ware
.
DF
A i
s
pr
oba
bl
y
t
h
e
m
o
st
freq
u
e
n
t
l
y
used
supe
rvi
s
e
d
pa
t
t
e
rn rec
o
g
n
i
t
i
on m
e
t
hod a
n
d t
h
e best
-
st
udi
e
d
o
n
e [
1
5]
. DF
A i
s
bas
e
d o
n
t
h
e
det
e
r
m
i
n
at
i
on o
f
di
scri
m
i
nant
f
u
n
c
t
i
ons,
w
h
i
c
h
m
a
xim
i
ze t
h
e rat
i
o
o
f
betwee
n-class
varia
n
ce and
minimize the ratio of
with
in-class v
a
rian
ce. As in
PCA,
th
is tech
n
i
q
u
e
is a
fact
ori
a
l
m
e
t
hod
. I
n
fact
, usi
ng t
h
i
s
m
e
t
hod,
dat
a
are se
p
a
rat
e
d i
n
a
priori
defi
ned c
l
asses. The
objective
so
ugh
t u
s
i
n
g
DFA is to
in
vestig
ate if th
e v
a
riab
les N
D
a
nd
N
HV
are
su
fficien
t
or no
t to
allo
w a
g
ood
a
post
e
ri
o
r
i
cl
assi
fi
cat
i
on
of
dat
a
i
n
t
h
ei
r
a pri
o
ri
gr
o
u
p
s
. T
h
e f
o
l
l
o
w
i
ng t
a
bl
es (Ta
b
l
e
1
an
d
2)
s
h
o
w
resp
ectiv
ely t
h
e
W
ilk
s' Lam
b
d
a
test (Rao
app
r
ox
im
a
tio
n
)
an
d th
e Bartlett's test of sp
h
e
ri
city:
Tabl
e 1.
W
i
l
k
s'
Lam
bda t
e
st
(R
ao ap
pro
x
i
m
at
i
on)
Tabl
e 2
.
Sp
h
e
ri
ci
ty
B
a
rtl
e
tt'
s
t
e
st
W
i
l
k
s' Lam
b
d
a
test in
terpretatio
n
:
L
a
m
bda 0,
448
F (
V
aleur
obser
vée)
24,
669
F (
V
aleur
cr
itique)
3,
232
DDL1
2
DDL2
40
p-
value <
0,
0001
alpha 0,
05
Khi² (Valeur
obser
vée)
55,95
Khi² (Valeur critiq
ue)
3,841
DDL
1
p-
value <
0,
0001
alpha 0,
05
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E
V
o
l
.
6,
No
. 3,
J
u
ne 2
0
1
6
:
11
6
8
– 11
75
1
172
H0:
The
m
ean vectors
of the t
w
o classes a
r
e
equal.
Ha: At least
one of the
m
ean vect
o
r
s
i
s
di
f
f
e
r
ent
fr
om
anot
her
.
Sin
ce th
e calcu
lated
p-v
a
lu
e
is less th
an
th
e sig
n
i
fi
ca
nce l
e
vel
of al
pha
= 0.
05
, o
n
e m
u
st
re
ject
t
h
e
n
u
ll
h
ypo
th
esi
s
H0
, and
retain
th
e altern
ativ
e
h
ypo
th
es
is
Ha.
Th
e
risk
of rej
ecting
the nu
ll h
ypo
th
esis H0
wh
en
it is tru
e
is less th
an
0
.
01
%.
Sph
e
ricity Bartlett
's test in
terpretatio
n
:
H0:
The
r
e i
s
n
o
si
gni
fi
ca
nt
l
y
di
ffe
re
nt
co
rrel
a
t
i
on
of
0
bet
w
een
vari
a
b
l
e
s.
Ha:
At
l
east
o
n
e
o
f
t
h
e
co
rrel
a
t
i
ons
bet
w
ee
n t
h
e
vari
a
b
l
e
s i
s
si
gni
fi
ca
nt
l
y
di
ffe
rent
fr
om
0.
Sin
ce t
h
e calcu
lated
p
-
v
a
lu
e
is less th
an
t
h
e sign
ifi
cance
level
of al
pha
= 0.05, one m
u
st re
ject t
h
e
n
u
ll
h
ypo
th
esi
s
H0
, and
retain
th
e altern
ativ
e
h
ypo
th
es
is
Ha.
Th
e
risk
of rej
ecting
the nu
ll h
ypo
th
esis H0
wh
en
it is tru
e
is less th
an
0
.
01
%.
The ca
n
oni
cal
di
scri
m
i
nant
fu
nct
i
o
n
s
ret
u
r
n
e
d
by
D
F
A
as:
F1
= 0.013
N
HV
–
0.
01
N
D
– 0.
96
(6
)
Th
is fu
nctio
n
i
s
u
s
ed
to
classi
fy th
e weed
s i
n
to
two
classes Mo
no
co
t and Dico
t (see
Figu
re
5
)
. Th
e
classificatio
n
qu
ality is sh
own b
y
th
e ROC cu
rv
e
(see
Figu
re 6).
Ap
pl
y
i
ng
DFA
o
n
dat
a
ba
se i
m
ages, a
g
o
o
d
separa
tion bet
w
een
weeds
s
p
ecies was
obtained.
(Figure
5) s
h
o
w
s
ho
w
t
h
e fi
rst
DF
A
fu
nct
i
on
di
scr
i
m
i
nat
e
am
ong cl
ust
e
rs.
DF
A
m
odel
was cr
oss-
val
i
d
at
e
d
usi
n
g
leave-one
-out
approach. An
accuracy of
94.74 % succe
ss rate in
the recognition
of dicots.
On t
h
e
other
hand,
DF
A classifier for the
m
onocots
ha
s reached
95.
83
% of the c
o
rre
c
t cl
assification. T
o
tal accura
cy is of
95.35%
succe
s
s
rate
(see Ta
bl
e 3).
Tab
l
e
3
.
C
o
nfusio
n m
a
trix
fo
r th
e
resu
lts
o
f
cro
s
s-v
a
lid
atio
n
Fr
o
m
\ T
o
Dicot
M
onocot
T
o
tal
% cor
r
ect
Dicot
18
1 19
0,
9474
M
onocot
1
23
24
0,
9583
T
o
tal
19
24
43
0,
9535
In a
d
di
t
i
on, t
h
e
m
a
i
n
adva
nt
age o
f
ou
r de
s
c
ri
pt
o
r
resi
des
i
n
i
t
s
speed
(
S
ee Tabl
e
4)
and ea
se o
f
i
m
p
l
e
m
en
tatio
n
.
In
deed
: Step
is th
e o
n
l
y p
a
ram
e
ter to
ad
ju
st.
Hen
c
e, th
e d
e
scrip
t
or p
r
esen
ts robu
stn
e
ss
agai
nst
t
h
e
bri
ght
ness c
h
a
nge
, r
o
t
a
t
i
o
n
an
d t
r
ansl
at
i
o
n.
In t
h
is table,
we ca
n clearly notice that
our des
c
ri
pt
or
ha
s a low com
p
utational cost.
In
practice, i
n
preci
si
o
n
ag
ri
c
u
l
t
u
re
, a scene i
s
never c
ove
re
d wi
t
h
ve
get
a
t
i
on t
o
10
0%. T
h
e ex
peri
m
e
nt
s were co
nd
uct
e
d o
n
Di
c
o
t
s
Mo
n
o
co
t
s
-2
-1
0
1
2
-2
-
1
0
1
2
F2
(
0
,
0
0
%
)
F
1
(
1
00,
00
%
)
(
A
xe
s
F
1
e
t
F
2
:
100,00
%)
Ba
r
y
cen
t
r
es
0
0,
2
0,
4
0,
6
0,
8
1
00
,
5
1
S
e
n
s
i
b
il
it
é
1 -
S
p
é
c
i
f
i
c
i
t
é
C
o
ur
be
R
O
C
(
A
U
C
=
0
,
9
9
8
)
Figure 5. W
e
e
d
classifica
tion accordi
n
g to c
a
nonical
di
scri
m
i
nant
fu
nct
i
o
n
s
ret
u
r
n
e
d
by
D
F
A
Fig
u
re
6
.
Th
e
ROC curv
e : sen
s
itiv
ity /
s
p
ecificit
y
re
p
ort
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
A Fast and
Efficien
t
Sh
ap
e
Descrip
t
o
r
fo
r
an
Ad
va
n
c
ed
Weed
Typ
e
Cl
a
ssifica
tio
n .... (Ad
il Tann
ou
ch
e)
1
173
a com
puter with i5 process
o
r
at 2.2
GHz
with 4GOc
tets DDR4
ram
.
Th
e ex
ecu
tion
time is less th
an
5.10
-3
s
by im
age.
Tabl
e 4. N
u
m
b
er of p
o
ssi
bl
e
c
o
m
p
ari
s
on
s de
pen
d
i
n
g o
n
t
h
e
si
ze
o
f
t
h
e
sl
i
d
i
ng wi
n
d
o
w
W
i
ndow Size
Nu
m
b
er
of co
m
p
arisons
Height W
i
dth
H L
H
×L×8
32
32
8192
640
480
2457
600
Th
is allows
our
d
e
scri
p
t
or to wo
rk
at
real t
i
m
e
wi
t
h
a st
a
nda
r
d
val
u
e
o
f
2
5
fram
e
s per
seco
n
d
.
I
n
p
r
actice, a m
a
j
o
rity o
f
vo
tes p
r
ov
id
ed
b
y
the slid
in
g
windo
w
ob
tain
s th
e fin
a
l classificatio
n
of a reg
i
on
in
the
im
age. The res
u
l
t
s
of
ou
r e
x
p
e
ri
m
e
nt
are very
pr
om
i
s
i
ng. We got a corre
ct classification rate of around 85%
on a
set of 50 i
m
ages analyzed. M
o
noc
o
tyledonous
wee
d
s
have
bee
n
clas
sified c
o
rrectly up t
o
90%
of
cases,
whi
l
e
t
h
e di
c
o
t
y
l
e
do
no
us w
eeds ha
ve bee
n
reco
g
n
i
zed
to about 80%. Figure
7 shows an exam
ple of
pr
ocessi
ng
pe
r
f
o
r
m
e
d by
ou
r
ap
pr
oac
h
, m
onoc
ot
y
l
edo
n
o
u
s
wee
d
s a
r
e m
a
rke
d
i
n
bl
ue
and
di
c
o
t
y
l
e
do
no
u
s
weeds are
m
a
rked in
re
d. T
h
i
s
figure also
s
h
ows
t
w
o e
r
r
o
rs
o
u
t
o
f
2
6
cl
ass
i
fi
cat
i
ons.
Fi
gu
re
7.
C
l
assi
fi
cat
i
on
resul
t
s:
M
o
n
o
c
o
t
y
l
e
do
n
ous
i
n
bl
ue,
di
cot
y
l
e
d
o
n
o
u
s
i
n
red
an
d cl
a
ssi
fi
cat
i
on e
r
r
o
rs
4.
CO
NCL
USI
O
N
We ac
hi
eve
d
a sy
st
em
for
di
scri
m
i
nat
i
ng bet
w
ee
n m
o
n
o
cot
y
l
e
d
o
n
o
u
s
an
d
di
cot
y
l
e
d
o
n
o
u
s
wee
d
s
speci
es. T
h
i
s
sy
st
em
i
s
based on o
u
r a
d
jace
nci
e
s desc
ri
pt
o
r
desi
g
n
e
d
fo
r t
h
i
s
pu
rp
ose. T
h
i
s
i
s
a rob
u
st
bi
na
r
y
descri
pt
o
r
, fast
and easy
t
o
u
s
e. The
resul
t
s
obt
ai
ne
d w
e
re
very
sat
i
s
fact
ory
f
o
r a
fast
execut
i
o
n t
i
m
e of t
h
e
or
der
of
2
5
f
r
am
es per secon
d
.
We pl
a
n
t
o
im
pro
v
e t
h
ese res
u
l
t
s
b
y
ot
her a
d
o
p
t
e
rs o
f
l
ear
ni
n
g
an
d
cl
assi
fi
cat
i
on
m
e
t
hods
. I
n
t
h
e l
i
ght
of t
h
ese
res
u
l
t
s
, t
h
e
p
r
op
ose
d
cl
assi
fi
cat
i
on sy
st
em
s
rep
r
ese
n
t
an
ex
cel
l
e
nt
,
fast an
d e
fficie
n
t s
h
ape
desc
ri
pto
r
f
o
r a
n
a
d
vanced wee
d
ty
pe classification a
p
proach.
REFERE
NC
ES
[1]
M. M. Siddiqi,
et
al
.
,
“
A
rea
l
t
i
m
e
speci
fic
weed
discrim
i
nati
on system
using m
u
l
ti-lev
e
l w
a
vel
e
t
decom
position
,
”
Int.
J.
Agr
i
c.
Bio
l
.
, vol/issue: 11(
5), pp
. 559-565
,
2009.
[2]
A. Tannou
che,
et
al
.
,
“
A
rea
l
tim
e effi
ci
ent m
a
na
gem
e
nt of onion
s weeds based o
n
a m
u
ltil
a
y
er
p
e
rcep
tron neur
al
networks technique,”
Int
l
J
Far
m
&
Alli Sc
i.
, vo
l/issue: 4
(
2), pp.
161-166, 2015
.
[3]
L. Long
champs,
et al
.
,
“
D
is
crim
ination of corn
, g
r
as
s
e
s
and dicot
weeds
b
y
th
eir
UV-induced flu
o
res
cenc
e
s
p
ec
tr
al
signature,”
Prec
i
s
ion Agric
, vo
l.
11, pp
. 181-197
, 2010, DOI: 10.1
007/s11119-009-9126-0.
[4]
X. Hadoux,
et a
l
., “Weeds-wheat discrimination
using h
y
perspectral imag
er
y
,
” CIGR-Ageng 2012,
International
Conference on
Agricultural Engineering
, Valencia, Spain, pp. 6, 2
012.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E
V
o
l
.
6,
No
. 3,
J
u
ne 2
0
1
6
:
11
6
8
– 11
75
1
174
[5]
H. T.
S
ogaard
,
“
W
eed clas
s
i
fi
c
a
tion b
y
ac
tiv
e s
h
ape m
odels
,
”
Biosystems Engin
eering
, vol/issue: 91(3), pp
. 271-
281, 2005
.
[6]
A. Tannouch
e
,
et
al
., “
A
fas
t
an
d effici
ent appro
ach for weed
s id
entifi
c
a
tion usin
g Haar-like f
eat
ures,”
Ameri
c
an
-
Eurasian Journa
l of Sustainable
Agriculture
, vo
l/issue: 9(4), pp. 4
4
-48, 2015
.
[7]
A. Te
ll
aech
ea
,
et al.
, “A computer v
i
sion app
r
o
ach for
weeds
id
entification
thr
ough Support V
ector
Mach
ines,”
Applied
Sof
t Co
mputing
, vo
l. 11, pp. 908-915, 20
11.
[8]
P. J. Herrera,
et
al
., “A novel approach for weed ty
pe classif
i
cation based on shape descriptors and a fuzzy
decision-making
method,”
Senso
r
s
, vol. 14, pp. 1
5304-15324, 20
14. DOI: 10
.339
0/s140815304.
[9]
M. K. Hu,
“
P
att
e
rn recogn
ition
b
y
m
o
m
e
nt
invar
i
ants,”
Proc. I
R
E
(
C
orrespondence)
, vol. 49, pp.14
28, 1961
.
[10]
J. Harikiran
,
et al.
, “
M
ultip
le
featur
e fuz
z
y
c-m
eans clust
e
ring a
l
gorithm
for segm
entati
on of m
i
croarr
a
y
im
ages
,”
Interna
tional
Journal o
f
Elect
rical and
Computer Engin
eering
, vol/issue: 5(5), pp. 1045-
1053, 2015
.
[11]
Saparudin,
et a
l
.
, “Segmentation
of fingerprin
t
image base
d on
gradien
t
magnitude and coh
e
ren
ce,”
Inte
rnat
i
o
nal
Journal of Electrical and
Computer
Eng
i
neer
ing
,
vol/issue:
5(5), p
p
. 1202-1215
, 2
015.
[12]
X. P. Burgos-Ar
tizzu,
et al.
, “
R
eal-tim
e im
ag
e proces
s
i
ng for cro
p
/weed dis
c
rim
i
n
ation in m
a
i
ze
fields
,
”
Comput.
Ele
c
tron.
Agric
.
, vol. 75, pp. 337
-346, 2011
.
[13]
T. Oj
ala
,
et al
.,
“
A
com
p
arativ
e s
t
ud
y of
textu
r
e m
eas
ures
wi
th classification
based on
featur
e distribu
tions
,”
Pattern
Recognition
, vo
l/issue: 1
9
(3), pp
. 51-59
,
1996.
[14]
J.
F.
Hair,
et
al
.,
“
M
ultivaria
te
Da
ta Ana
l
y
s
is,”
5th
ed.
,
Pren
ti
ce
‐
Hall, Upper Sadd
le
River, NJ, 1998
.
[15]
D. F. Morrison,
“Multivariate
Statisti
cal Method
s,” McGraw-Hil
l
,
Singapor
e, 2nd
edition
,
1988
.
BIOGRAP
HI
ES OF
AUTH
ORS
Adil Tanno
uc
h
e
is currently
a PhD student at th
e Labor
atoir
e
d’Etude des Mateériaux Avanceés
et Applications, Moulay
Ismail University
,
Fac
u
lty
of Scie
nce
s
in Me
kne
s,
Moroc
c
o
.
His
res
earch
int
e
res
t
s
are focus
e
d
in
m
achine vis
i
on
,
artif
ici
a
l
inte
lli
gence
and th
eirs
appli
cat
ion in
agricu
lture
.
Khalid Sbai
is a full prof
essor since 2001 in
Electroni
cs. He received his M
.
sc. Degr
ee in
Electroni
cs from Valenci
e
nne Universit
y
(Franc
e) in 1996 and his Habilitation i
n
Phy
s
i
c
s from
Moula
y
Ism
a
il
Universit
y
in
20
08. His rese
arch
in
te
res
t
s
in
clud
e S
t
ruc
t
ural
stud
ies, v
i
brational
and electronic pr
operties of
carbo
n nanotub
es
M
iloud Rahmo
une
is a full professor at Moula
y
Ism
a
il Universit
y
. He r
e
c
e
ived h
i
s Msc. Degree
in applied m
e
c
h
anics from
Universit
e
´
Mont
pe
llier 2 (France) and his Ph.D. degrees in
M
echatron
i
cs
fr
om
Univers
it M
ontpellier 2 (France) and Univ
er
site
´ Ha
ssa
n II M
ohammedia, in
1993 and 1996
respectively
.
His
resear
ch interes
t
s
includ
e structural D
y
namics,
active
control,
and s
m
art m
a
t
e
ri
als
.
Amine Z
o
ubir
is
a s
e
nior lec
t
ur
er at Univ
ers
i
t
y
M
oula
y
Is
m
a
il
,
M
o
rocco. He r
e
c
e
ived th
e M
S
c.
Degree (Mag
ister) in
fluid mechanics from th
e Un
iversity
of
Ly
on
1 (Fran
c
e) and th
e Ph.D
degree
in
M
echa
n
ics
,
Energ
e
ti
cs
,
Civil
Eng
i
neer
i
ng and Acoustics of INSA Ly
on
(France)
. He is
an act
ive r
e
s
ear
cher a
t
Therm
a
l
& M
a
teri
al Re
s
earch Unit (ad
v
anced m
a
t
e
ria
l
s
and ener
g
y
s
y
stem). His
area of r
e
sear
ch
are focused o
n
th
e
numerical modeling of
c
onvectiv
e heat transfer
and around
th
e d
i
agnosis of
ener
g
y
performance
in build
ings.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
A Fast and
Efficien
t
Sh
ap
e
Descrip
t
o
r
fo
r
an
Ad
va
n
c
ed
Weed
Typ
e
Cl
a
ssifica
tio
n .... (Ad
il Tann
ou
ch
e)
1
175
Rac
h
id Agouno
un
is
a s
e
nior l
e
cturer
at
Univer
s
i
t
y
M
oula
y
Is
m
a
il
, M
o
rocco
. H
e
re
ceiv
e
d th
e
M
S
c. Degre
e
(
M
agis
ter)
in m
e
chani
c
s
and
ene
r
geti
c s
y
s
t
em
fr
om
the Univers
i
té d
e
Lor
r
ain
e
,
Nanc
y, F
r
anc
e
a
nd the P
h
.D degree in s
c
ien
ce f
o
r engineers
fro
m
the Univers
ite´
de Lorraine
,
Nanc
y, F
r
anc
e
.
He is
an active
res
earcher a
t
Therm
a
l & M
a
t
e
ria
l
Res
earch
Unit (advanc
ed
m
a
teria
l
s
and en
erg
y
s
y
s
t
em
). Hi
s
area of res
e
arc
h
includes
Th
er
m
a
l Com
f
ort, Building Th
erm
a
l
Simulation, ren
e
wable
energ
y
an
d Porous Media.
Rachid Saadani
is a senior lecturer at University
Moulay
Is
mail, Morocco
. Was born i
n
M
o
rocco in 197
7. He re
ce
ived
t
h
e M
S
c. Degre
e
(M
agis
ter) in
the
r
m
a
l and en
erge
t
i
c s
y
s
t
em
from
the Univers
it M
a
rne La Val
l
é
e
,
P
a
ris
Es
t, P
a
ris
,
F
r
ance and t
h
e P
h
.D degree
in s
c
ience for
engine
ers
from
t
h
e Univers
ite´
Paris
Es
t, Crét
eil
,
P
a
ris
.
He is
an
act
ive res
e
a
r
cher
at Therm
a
l &
M
a
teri
al Res
e
a
r
ch Unit (advanc
ed m
a
teria
l
s
and energ
y
s
y
s
t
em
).
His
area of res
earch in
clud
es
Therm
a
l Com
f
or
t, Bu
ilding
Th
er
m
a
l Sim
u
lation
,
renewabl
e
energ
y
and Porous Me
dia.
Abdelali
Rahm
ani is
a full profes
s
o
r at Univers
i
t
y
M
oula
y
Is
m
a
il, M
o
rocco
. He receiv
e
d the
M
S
c. Degree
in
theore
tic
al ph
ys
ics
from
the Univers
i
t
e
´
M
ontpelli
er 2, F
r
anc
e
and th
e P
h
.D
degree
in Ma
teri
als scien
c
e
from
the
the Univ
ersi
té Montpe
lli
er 2
,
France
. He
is th
e Dire
ctor of
Laborator
y
of Studies of Avanced Mater
i
als
an
d
Applica
tions
).
His
area of r
e
s
earch
includ
es
Computational p
h
y
sics
and n
a
no
materials.
Evaluation Warning : The document was created with Spire.PDF for Python.