Int
ern
at
i
onal
Journ
al of
El
e
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
11
,
No.
2
,
A
pr
il
2021
, p
p.
17
19
~
1727
IS
S
N:
20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v
11
i
2
.
pp
1719
-
17
27
1719
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om
Detectio
n of
c
it
ru
s
l
eaf
d
i
seases
u
s
ing
a
d
eep
l
earni
ng
t
ec
hn
iqu
e
Ah
med
R. Lu
aibi
,
T
ariq
M. Salm
an
,
A
bb
as
Hussein
Mi
ry
El
e
ct
ri
ca
l
Eng
in
ee
ring
Depa
r
tment,
Mus
ta
nsir
i
y
a
h
Univer
sit
y
,
Ir
a
q
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
2
9
, 20
20
Re
vised
Sep
11
,
20
20
Accepte
d
Nov 8
, 2
0
20
The
food
sec
ur
i
t
y
m
aj
or
threat
s
are
the
dise
ase
s
aff
ecte
d
in
p
lant
s
such
as
ci
trus
so
tha
t
the
ide
nt
ifica
t
io
n
in
an
ea
rli
e
r
ti
m
e
is
ver
y
important
.
Conveni
ent
m
a
la
d
y
re
cogni
t
io
n
ca
n
assist
t
he
cl
i
ent
wi
th
responding
imm
edi
at
ely
and
sketc
h
for
som
e
guar
ded
a
ct
iv
it
i
es.
Thi
s
re
cogni
t
ion
ca
n
b
e
complet
ed
witho
ut
a
hum
an
b
y
u
ti
lizing
pl
ant
lea
f
pic
ture
s
.
The
r
e
are
m
an
y
m
et
hods
emplo
y
ed
for
the
c
la
ss
i
fic
a
ti
on
and
d
etec
t
ion
in
m
ac
hi
ne
learni
ng
(ML)
m
odel
s,
but
the
combina
tion
of
inc
rea
sing
adva
nc
es
in
computer
vision
appe
ars
the
d
eep
le
arn
ing
(DL)
are
a
res
ea
rch
to
ac
hi
eve
a
gre
at
pote
ntial
in
te
rm
s
of
inc
rea
si
ng
ac
cur
a
c
y
.
In
t
his
pape
r,
two
wa
y
s
of
conve
n
ti
o
nal
neur
al
net
works
are
use
d
named
AlexNe
t
and
Res
Net
m
odel
s
with
and
without
data
augmenta
t
ion
i
n
v
o
l
v
e
s
t
h
e
p
ro
c
e
s
s
o
f
c
r
e
a
t
i
n
g
n
e
w
d
a
t
a
p
o
i
n
t
s
b
y
m
a
n
i
p
u
l
a
t
i
n
g
t
h
e
o
r
i
g
i
n
a
l
d
a
t
a
.
T
h
i
s
p
r
o
c
e
ss
i
n
c
r
e
a
s
e
s
t
h
e
n
um
b
e
r
o
f
t
r
a
i
n
i
n
g
i
m
a
g
e
s
i
n
DL
wi
t
h
o
u
t
t
h
e
n
e
e
d
t
o
a
d
d
n
e
w
p
h
o
to
s
,
i
t
w
i
l
l
a
p
p
r
o
p
r
i
a
t
e
i
n
t
h
e
c
a
s
e
o
f
s
m
a
l
l
d
a
t
a
s
e
t
s
.
A
self
-
da
ta
set
of
200
ima
ges
of
disea
s
es
and
he
alt
h
y
ci
trus
l
ea
ves
are
col
l
ec
t
ed.
The
t
rai
ned
m
odel
s
w
it
h
data
augmen
ta
ti
on
g
ive
the
best
results
with
95.
83%
and
97.
92
%
for
R
es
Net
an
d
Alex
Ne
t
respe
ctively
.
Ke
yw
or
d
s
:
Alex
Net
C
it
ru
s leaves
D
at
a au
gm
entat
ion
D
eep
lear
ning
Re
s
Net
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
:
A
hm
ed
R. L
ua
ibi
Dep
a
r
tm
ent o
f El
ect
rical
En
gi
neer
i
ng
Mustansiriy
ah
Un
i
ver
sit
y
Ba
ghda
d,
Ir
a
q
Em
a
il
:
ah
m
edr
aheem
17
5@g
m
ai
l.co
m
1.
INTROD
U
CTION
In
Ir
a
q,
cit
r
us
is o
ne o
f
the m
os
t val
uab
le
it
em
s,
the 2
01
9
s
urvey o
f
the
producti
on of cit
ru
s tre
es wa
s
com
plete
d
by
the
Direct
or
at
e
of
Agricult
ural
Stat
ist
ic
s,
a
s
urvey
co
ntaine
d
in
the
Ce
nt
ra
l
Stat
istics
Bure
au's
annual
plan
an
d
c
overin
g
fiv
e
m
ai
n
ty
pes:
or
a
nge,
sour
l
e
m
on
,
sweet
lem
on
,
m
and
ari
n
a
nd
bitt
er
or
an
ge
.
Be
cause
of
i
na
dequate ca
re a
nd lack
of
pesti
ci
de
us
a
ge, m
a
ny d
ise
ases
af
f
ect
ing
cit
r
us
tr
ees ha
ve
s
pr
ea
d,
s
uc
h
as;
Ph
yl
locnist
is
ci
trel
la
,
la
ck
of
el
em
ents,
scal
e
insect
s,
et
c.,
reali
zi
ng
t
ha
t
the
con
se
quences
of
diseas
es
hav
e
kill
ed
la
r
ge
nu
m
ber
s o
f
c
it
r
us t
rees a
nd
l
ow
pro
du
ct
ivit
y.
Wh
e
re t
he
ave
r
age
producti
vit
y of
t
he ora
ng
e
tree in
Ir
a
q
was
est
im
at
ed
at
only
13.
5
kg,
w
hich
is
a
ver
y
l
ow
a
m
ou
nt,
co
ns
i
der
i
ng
it
to
be
the
fir
st
tree
of
ci
tr
us
fruit
s
in
Ir
a
q,
and
w
he
re
the
ave
rag
e
pro
duct
ivit
y
of
ot
he
r
ci
tru
s
trees
is
si
m
il
ar
to
the
am
ou
nt
of
or
a
nge
pro
du
ct
io
n
[1]
.
In
t
his
pap
e
r
t
hr
ee
diseases
of
ci
tr
us
a
nd
healt
hy
le
aves
discu
ssed
an
d
detect
ed;
the
first
ty
pe
of
disease
is
the
Ph
yl
locnisti
s
c
it
rell
a
disease
wh
ic
h
is
a
sig
nificant
pest
of
the
w
orl
dwide
com
m
ercial
ci
trus
pro
du
ct
io
n.
In
disti
nctive
ser
pen
ti
ne
m
ines,
egg
s
a
re
la
id
on
y
oung
le
av
es
and
la
rv
ae
feed
i
ns
ide
the
le
af
ti
ssu
e,
e
ve
ntu
a
ll
y
pupating
in
a
pu
pal
cel
l
at
t
he
le
af
m
arg
in w
it
h
de
velo
pme
ntal
pe
rio
d
va
ryi
ng
f
ro
m
13
t
o
52
days
de
pendin
g
on
the
te
m
pe
ratur
e
[2]
.
T
he
second
diseas
e
is
the
l
ack
of
el
e
m
ent
disea
se
wh
ic
h
is
ha
pp
e
ne
d
because
the
sup
ply
of
su
c
h
e
lem
ents
su
ch
a
s
Zn
,
Mn
an
d
Fe
is
relat
ed
to
s
o
il
-
P
h,
def
i
ci
ency
sym
pto
m
s
of
these
th
ree
el
e
m
ents
m
a
y
al
s
o
a
pp
ea
r
c
onc
urren
tl
y
withi
n
a
can
opy
of
t
he
tree
an
d
of
t
en
c
ov
e
r
one
a
no
t
her
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.
11
, No
.
2
,
A
pr
i
l
2021
:
1719
-
1727
1720
within
a
sin
gle
le
af
[
3]
.
Finall
y,
insect
scal
e
disease
wh
ic
h
within
t
he
s
uperf
am
ily
Coccoidea
is
ref
e
rr
e
d
to
a
s
a
broad
c
ommun
it
y
of
in
sect
s,
Scal
e
insect
s
feed
in
g
on
y
oung,
de
vel
op
i
ng
ti
ps
m
ay
ca
us
e
wa
rp
e
d
f
ol
ia
ge.
Feedin
g
on
le
a
ves
can
t
urn
th
e
m
ye
l
low
an
d
plants
can
l
ook
water
-
stres
se
d.
Str
ong
i
nf
es
ta
ti
on
s
can
ca
use
the
br
a
nc
hes
a
nd s
tem
s to
die
bac
k
[2]
.
DL
is
a
f
or
m
of
ML
,
based
on
a
de
ep
ne
ural
netw
ork
wi
th
seve
ral
hidd
en
la
ye
rs.
It
is
one
of
t
he
la
te
st
exa
m
ple
s
of
resea
rch
i
nto
ML
an
d
ar
ti
fici
al
intel
li
gen
ce
(
AI
)
[4]
.
To
day,
DL
is
beco
m
ing
one
of
the
m
os
t rele
van
t i
den
ti
ficat
io
n
te
chn
i
qu
e
s.
Co
nvol
ution ne
ur
al
n
et
w
ork (
CN
N)
is
DL'
s b
asi
c
m
et
ho
d, it
incr
eases
accuracy
by
program
m
ing
a
l
arg
e
am
ou
nt
of
data
f
or
e
xtra
ct
ing
featur
e
s
and
m
ulti
ple
hid
de
n
la
ye
rs
us
i
ng
an
ML
m
od
el
[5]
.
In
[6]
K
rizhe
vs
ky
im
ple
m
e
nted
a
deep
C
NN
t
o
ide
ntify
1.2
m
i
ll
ion
im
ages
with
I
m
ageN
et
and
f
or
the
fir
st
tim
e
achieved
the
top
-
1
a
nd
to
p
-
5
er
ror
rate
in
the
I
m
age
Re
co
gn
it
ion
Com
petit
ion
,
after
wh
ic
h
the
re
se
arch
e
rs
ca
ught
the
interest
of
this
fiel
d.
D
L has
us
e
d
pla
nt
il
lness
di
a
gnos
is
and
d
et
ect
ion.
Kaur
et
al
.
[7]
Use
d
Goo
gle
Net
CNN
m
od
el
to
detect
and
cl
assify
healt
hy
and
disease
for
dif
f
eren
t
kinds
of
plants
and achie
ve 9
7.8
2%
acc
ur
acy
.
In
[
8]
Sa
hid
a
n
et
al
.
pro
po
se
d
a
le
af
rec
ogniti
on
by
us
in
g
a
conv
olu
ti
onal
neural
netw
ork
an
d
bag
of
featur
e
s
,
they
us
e
d
a
public
data
set
nam
e
d
F
olio.
T
he
e
xp
e
rim
ental
re
su
lt
s
in
dicat
e
that
ba
g
of
fea
tures
achieves
bette
r
accuracy
com
par
e
d
to b
asi
c
CNN
with
82.
03%
accu
racy
.
In
[
9]
K.
P.
Fe
ren
t
in
os
u
sed
a
n
ope
n
dataset
of
87.
848
im
ages
co
ntains
healt
hy
a
nd
disease
pla
nt
le
aves
ap
plied
to
dif
fer
e
nt
C
NN
m
et
ho
ds
nam
ed
VGG,
Ale
x
Ne
t,
G
oogle
Net
and
O
ve
rf
eat
CNN
an
d
t
he
r
esult
sho
w
that
V
GG
ha
s
the
best
value
with
0.4
7%
error
a
n
d
99.
53%
acc
ur
acy
f
or
t
he
te
ste
d
s
et
.
Xing
et
al
.
in
[
10
]
intr
od
uced
a
rec
ogni
ti
on
m
od
el
for
ci
t
ru
s
disease
an
d
pe
sts
by
us
i
ng
w
eakly
de
ns
e
co
nn
ect
e
d
c
onvoluti
on
net
work,
he
use
d
a
sel
f
-
dataset
f
or
ci
tr
us
a
nd
app
li
ed
it
to d
i
ff
e
ren
t
CN
N
m
od
el
s
t
he
ex
perim
ental
resu
lt
s
sh
ow
th
at
NIN
-
16
ac
hieve
d
91.
66%
te
st
accurac
y
wh
ic
h
was
hi
gher
t
han
t
he
S
ENet
-
16
m
od
e
l
with
88.
36%.
W
eakly
De
ns
e
Net
-
16
ha
ve
th
e
higher
acc
uracy
of
93.33%
tha
n
NIN
-
16,
V
GG
-
16
ac
hieve
d
the
sec
ond
hi
ghest
cl
assifica
t
ion
accu
racy
93%
with
the
m
os
t
com
pu
ti
ng
m
od
el
siz
e
res
our
ces
of
120.2
MB
.
The
go
al
s
of
t
his
st
ud
y
are
ide
ntifie
d
and
cl
assifi
e
d
healt
hy
le
aves
and
different
ty
pe
of
di
sease
occu
r
re
d
in
the
ci
trus
le
aves
by
us
in
g
two
m
od
el
s
of
conven
ti
onal
neura
l
netw
ork
w
hich
is
Ale
xN
et
a
nd
Re
sNet
with
data
a
ug
m
enter
a
nd
diff
e
r
ent
pa
ram
et
ers
to
ac
hieve
th
e
best
accuracy.
T
his
work is
pro
po
s
ed usi
ng
PC, C
or
e
i7, an
d
M
A
TLAB
r20
19 b.
2.
MA
TE
RIA
L
S
A
ND
METH
OD
2.1.
D
atase
t
In
this work, a data
set
o
f
20
0
i
m
ages f
or
hea
lt
hy an
d
Ph
yl
lo
cnisti
s cit
rell
a,
l
ack o
f
el
em
ent, an
d
scal
e
insect
s
disease
each
with
50
i
m
ages.
The
data
set
div
id
ed
into
70
pe
rcen
t
f
or
t
rain
ing
,
20
pe
rce
nt
for
validat
io
n,
a
nd 10 perce
nt for
trai
ning
of
t
he pr
opos
e
d
m
et
ho
d, Fi
gure
1
in
dicat
es the t
hr
e
e ty
pes
of cit
r
us l
eaf
diseases.
(a)
(b)
(c)
Figure
1.
Ty
pe
s of cit
r
us
leav
e d
ise
ases
,
(a
)
P
hyll
ocn
ist
is ci
trel
la
,
(b)
la
c
k of
el
em
ent
, (
c)
s
cal
e insect
s
The
dataset
is resized b
y form
at
(
heig
ht X
w
idth X
nu
m
ber
o
f
cha
nnel
),
f
or A
le
xNet
the size beco
m
e
(22
7x227x
3)
a
nd
(
224x
224x
3)
f
or
Re
s
Net
m
od
el
.
Then,
the
data
augm
entat
ion
appl
ie
d
fo
r
the
r
esi
zed
i
m
ages.
Alth
ou
gh
CN
N
is
ve
r
y
powe
rful,
t
he
res
ult
m
ay
be
beco
m
e
in o
ve
rf
it
ti
ng
an
d
ca
nnot ach
ie
ve
the
goal
resu
lt
s
beca
us
e
the
num
ber
of
im
ages
us
e
d
is
not
e
nough
so
it
a
rtifici
al
ly
enlar
ges
t
he
dataset
us
in
g
la
bel
-
pr
ese
r
ving
tra
ns
f
or
m
at
ion
s
[
11
]
.
D
a
t
a
a
ug
m
e
nt
at
i
on
i
nv
ol
ve
s
t
he
pr
oc
e
s
s
of
c
r
e
a
t
i
ng
ne
w
da
t
a
po
i
nt
s
by
m
a
ni
pu
l
a
t
i
ng
th
e
or
i
gi
na
l
da
t
a
.
T
hi
s
pr
oc
e
s
s
i
nc
r
e
a
s
e
s
t
he
nu
m
be
r
of
t
r
a
i
ni
ng
im
a
ge
s
i
n DL
wi
t
ho
ut
t
he
ne
e
d t
o
a
dd
ne
w
ph
ot
o
s
[5
,
6]
,
i
n
t
hi
s
w
or
k
t
he
a
u
gm
e
nt
a
t
i
on
i
s
do
n
e
by
:
-
Ra
ndom
r
eflect
ion
in
the le
ft
-
rig
ht d
irect
i
on.
-
The
width
of
horiz
on
ta
l
tra
ns
la
ti
on
ap
plied
t
o
the
in
put
im
age;
the
pix
el
scal
e
of
th
e
tra
ns
la
ti
on
distan
ce
[3
0
,
-
30
]
is
det
erm
ined.
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
Detect
ion
of ci
trus leaf
dise
ase
s u
si
ng a dee
p l
ear
ning tech
ni
qu
e (
A
hm
e
d R
. Lua
i
bi)
1721
-
Ver
ti
cal
transl
at
ion
ra
nge
a
dded
to
the
in
put
im
age;
translat
ion
dista
nc
e
is
com
pu
te
d
in
pi
xels
wit
h
a
range
of p
i
xels
[
30
,
-
30]
.
2.2.
C
N
N
art
e
chitcter
CNN
is
one
of
the
DL
arc
hitec
tures
a
nd
it
s
m
os
t
co
m
m
on
in
so
lvin
g
the
i
m
age
cl
assifi
cat
ion
pro
blem
,
it
is
t
he
m
os
t
eff
ect
i
ve
an
d
powe
rful
DL
te
chn
i
que.
CN
N'
s
are
an
ev
olu
ti
on
of
tradit
ion
al
ar
ti
fici
a
l
neural
netw
orks
(
ANN),
f
oc
us
in
g
pri
m
ari
ly
on
app
li
cat
ion
s
with
repea
ti
ng
patte
r
ns
in
var
io
us
ar
eas
of
m
od
el
li
ng
sp
ac
e,
in
a
pa
rtic
ul
ar
im
age.
Thei
r
m
ai
n
char
act
erist
ic
is
that
they
drast
ic
al
ly
re
du
ce
the
nu
m
ber
of
structu
ral
el
em
ents
(n
um
ber
of
arti
fici
al
neur
on
s
)
re
qu
i
red
as
co
m
pared
to
tradit
io
na
l
feed
f
orward
neural
netw
orks
wit
h
the
m
et
ho
dolog
y
use
d
in
t
heir
la
ye
rin
g.
CNN
is
fee
d
-
f
orward
an
d
is
a
hig
hly
infl
uen
ti
al
detect
ion
m
et
ho
d.
T
he
str
uct
ur
e
of
the
net
work
is
sim
ple
;
has
fe
we
r
tra
ining
pa
ram
et
e
rs.
C
NN
re
pr
e
sents
a
ver
y
e
f
fecti
ve
detect
ion
proc
ess.
O
n
t
he
othe
r
ha
nd,
t
he
ne
twork
m
od
el
'
s
com
plexity
and
weig
ht
num
ber
s
are
dim
inished.
Figure
2
s
how
s
the
m
ai
n
struc
ture
of
CN
N
that
con
ta
i
ns
m
ai
nly
five
layers;
the
input
la
ye
r,
conv
olu
ti
on
la
ye
r
with
act
iva
ti
on
f
unct
io
n,
poolin
g
la
ye
r,
fu
ll
y
co
nn
e
ct
ed
la
ye
r,
a
nd
fi
nally
the
softm
ax
la
ye
r
[12
-
15]
.
Figure
2. M
ai
n st
ru
ct
ur
e
of C
NN
In
t
he
c
onvolu
ti
on
la
ye
rs
t
ha
t
con
sist
of
a
series
of
c
onvoluti
onary
kernels
in
w
hich
each
neur
on
beh
a
ves
as
a
ke
rn
el
,
th
e
co
nvolu
ti
on
pro
ces
s
beco
m
es
a
correla
ti
on
proc
ess
wh
il
e
the
ke
rn
el
is
sym
metri
cal
[16]
.
The
proc
ess
of
c
onvolut
ion
has
thr
ee
pri
m
ary
adv
antages.
In
th
e
sam
e
functi
on
m
ap
the
weig
ht
sha
ri
ng
m
et
ho
d
reduce
s
the
num
ber
of
pa
ram
et
ers
and
he
nce
the
num
ber
of
ope
r
at
ion
s.
Local
c
onnecti
vity
al
lows
the
analy
sis
of
a
ss
ociat
ion
s
betw
een
ad
j
ace
nt
pix
el
s.
Last
ly
,
inv
aria
nce
to
the
obj
ect
'
s
or
igin
al
lows
to
loca
te
the
ta
rg
et
in
de
penden
t
of the
ob
je
ct
'
s p
la
ce in the
picture
[
17
]
.
The
po
olin
g
la
ye
r
is
us
ed
t
o
m
ini
m
iz
e
th
e
m
easur
em
ents
of
the
func
ti
on
m
aps
and
net
wor
k
par
am
et
ers
inc
reasin
gly.
P
oo
l
ing
la
ye
rs
a
re
t
her
e
fore
in
va
riant
in
e
ncodin
g,
si
nce
their
c
om
pu
ta
ti
on
s
ta
ke
int
o
account
ad
j
ace
nt
pix
el
s.
T
wo
m
ajo
r
ty
pes
of
poolin
g
la
ye
rs,
m
ax
pooling
la
ye
rs
and
av
erag
e
poolin
g
la
ye
rs
occur
[
18]
.
Th
e
m
os
t
us
ed
te
chn
i
qu
e
s
are
a
ver
a
ge
poolin
g
and
m
axi
m
u
m
poolin
g.
M
os
t
i
m
ple
m
entat
io
ns
us
e
m
ax
-
po
oling
because
it
can
le
ad
t
o
f
ast
er
co
nver
ge
nc
e,
pic
k
s
up
e
rior
i
nv
a
riant
f
eat
ur
es
a
nd
e
nh
a
nc
e
gen
e
rali
zat
ion
[19]
.
Fu
ll
y
connecte
d
(
FC
)
la
ye
rs
com
pr
ise
abou
t
90
per
ce
nt
of
a
CNN'
s
pa
r
a
m
et
ers.
Using
this,
the
neural
net
w
ork
is
fed
into
a
prede
fine
d
-
le
ng
t
h
vector.
We
m
ay
ei
ther
fee
d
the
vect
or
i
nt
o
a
va
riet
y
of
i
m
age
cl
assifi
cat
ion
gro
ups,
or
ta
ke
i
t
as
a
f
un
ct
io
n
vecto
r
f
or
f
ollo
w
-
up
pro
ce
ssin
g
.
T
houg
h
c
ha
ng
i
ng
the
str
uc
ture
of
the
f
ully
con
necte
d
la
ye
r
is
un
c
omm
on
,
so
m
e
e
ff
or
t
has
gone
in
to
m
ake
it
m
or
e
e
ffi
cien
t.
The
FC
la
ye
r
is
the
highe
r
-
le
ve
l
rep
re
sentat
io
n
of
the
i
nput
sign
al
,
t
he
out
pu
t
resu
lt
in
g
f
r
om
the
con
vol
ution,
act
ivati
on,
a
nd
poolin
g
la
ye
rs
pr
e
viously
ad
de
d.
T
hese
la
ye
r
s
are
no
t
s
uppo
sed
to
pro
vid
e
est
i
m
at
es
of
cl
assifi
cat
ion
.
T
he
FC
la
ye
r
is
us
e
d
at
this
sta
ge
to
identify
t
he
i
nput
pict
ur
e
a
cco
rd
i
ng
to
t
he
tra
ining
set
by
lo
ok
i
ng
at
the
fe
at
ure
s
[19
,
20]
.
Af
te
r
each
c
onvol
ution
al
la
ye
r,
the
Re
LU
act
ivati
on
la
ye
r
i
s
conve
ntio
nally
us
ed.
It
al
lows
introd
ucin
g
no
n
-
li
nea
rity
within
the n
et
w
ork.
Re
LU
was
m
or
e
co
m
pu
ta
ti
on
al
ly
e
ff
ect
ive
than
ta
nh
or
sigm
oid
functi
on
without sig
nifica
nt c
hange i
nacc
ur
a
cy
[21]
.
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.
11
, No
.
2
,
A
pr
i
l
2021
:
1719
-
1727
1722
3.
PROP
OSE
D WOR
K
AND
E
X
PERI
MEN
TAL RES
UL
TS
In
this
pa
pe
r,
the
input
18
0
im
ages
are
div
i
ded
int
o
trai
ni
ng,
validat
io
n,
and
te
st
i
m
ages.
The
im
ages
first,
resizi
ng
a
nd
trai
ning
by
on
e
of
t
he
two
CNNs
m
od
el
s
to
cl
assify
the
disease
ty
pe
of
ci
trus
le
aves
.
The
F
igure
3
s
hows
the m
ai
n
flo
w chart
of the i
m
age classi
ficat
ion wit
h data
au
gm
enter.
Figure
3.
The
fl
ow
c
ha
rt of im
age classi
ficat
ion w
it
h data
au
gm
entat
ion
u
si
ng CN
N
3.1.
Alex
Net
Alex
Kr
iz
he
vsky
is
the
creator
of
the
Alex
Net
platfo
rm
,
a
sta
te
-
of
-
the
-
a
rt
pr
e
-
trai
ne
d
CNN
[
22]
.
It
has
us
e
d
f
or
num
ero
us
c
om
par
iso
ns
in
se
ve
ral
dif
fer
e
nt
fiel
ds
.
F
or
this
reason
t
his
m
od
el
arc
hitec
tur
e
ha
s
been
u
se
d
f
or
im
age
cl
assifi
cat
ion
in
se
ver
al
diff
e
re
nt
ex
pe
rim
ents
[1
2].
It
is
a
deep
C
N
N
w
hich
is
c
onsist
s
of
twenty
-
fi
ve
la
ye
rs
includi
ng
on
e
in
put
la
ye
r,
five
co
nvolu
t
ion
la
ye
rs,
sev
en
Re
LU
la
ye
rs,
two
cr
os
s
-
ch
ann
e
l
norm
al
iz
a
ti
on
la
ye
rs,
one
S
of
t
Ma
x
la
ye
r,
an
d
finall
y
on
e
out
pu
t
la
ye
r.
T
he
recti
fied
li
near
un
it
(ReL
U)
w
hich
is t
he
nonlinea
r
act
ivati
on
f
unct
ion
th
res
ho
l
ds t
he
val
ue
of i
nput less tha
n ze
ro
a
nd se
ts t
hem
to
zero
. It
can
be
descr
i
bed m
at
he
m
at
ic
ally as fo
ll
ow
[
23]
.
(
)
=
{
≥
0
0
≤
0
(1)
3.2.
Residu
al
ne
twork archi
tectur
es
(
ResNet
)
Re
sNet is a d
e
ep
CN
N,
with
a sp
eci
al
ly
b
uilt
r
esi
du
al
str
uc
ture
that ca
n su
pport
a v
ery de
ep
net
work.
Cl
assic
deep
conv
olu
ti
on
ne
ur
al
net
works
can'
t
be
qu
it
e
la
rg
e,
e
ven
as
the
com
plexit
y
rises,
the
accuracy
decr
ease
s.
Re
s
Net
'
s
autho
r
c
onj
ect
ur
es
t
he
identit
y
m
app
in
g
is
hard
to
re
m
e
m
ber
.
The
deep
resid
ual
le
arn
i
ng
syst
e
m
is
su
ggest
ed
to
so
lve
this
prob
le
m
,
and
the
netw
ork
le
ar
ns
the
r
esi
du
al
rathe
r
than
di
rect
m
a
pp
i
ng
[24]
.
T
hi
s
m
od
e
l
w
on
t
he
I
m
a
ge
N
e
t
c
omp
e
t
i
ti
on
i
n
20
1
5.
R
e
s
N
e
t'
s
f
un
da
m
e
nt
al
br
e
a
kt
hr
o
ug
h
w
a
s
t
ha
t
it
a
l
l
ow
e
d
us
t
o
s
uc
c
e
s
s
f
ul
l
y
t
r
a
in
i
nc
r
e
di
bl
y
de
e
p
ne
u
r
a
l
ne
t
w
or
ks
w
i
t
h
15
0+
l
a
y
e
r
s
.
T
he
R
es
N
e
t
50
p
r
o
po
s
e
d
i
n
[
25
]
w
i
t
h
50
r
e
s
i
du
a
l
ne
t
w
or
k
l
a
y
e
r
s
by
H
e
e
t
al
.
T
he
he
i
gh
t
o
f
t
he
c
on
vo
l
ut
i
on
l
a
y
e
r
s
i
s
33
f
i
l
t
e
r
s
a
nd
t
hi
s
m
od
e
l
ha
s
a
n
in
p
ut
s
i
z
e
of
22
4*
22
4
[20]
.
Ea
ch
m
od
el
is
us
ed
to
trai
n
the
i
m
ages
with
S
GD
M
a
nd
m
ax
Epo
c
h
of
10
with
m
in
i
batch
siz
e
=6
and
i
niti
al
le
ar
ning
rate
1e
-
4
.
Table
1
a
nd
T
able
2
s
how
th
e
validat
ion
a
nd
te
st
accuracy
f
or
the
t
wo
m
od
el
s
in
t
he
case
of
data
a
ugm
enter
a
nd
without
,
Ale
xNet
m
od
el
ga
ve
th
e
be
st
te
st
accuracy
with
sm
a
ll
el
apsed
tim
e
co
m
par
ed
with
Re
sNet
m
od
el
.
Figure
4
an
d
Fi
gure
5
show
s
t
he
trai
ning
process
for
Al
exN
et
a
nd
Re
s
Net,
the
blu
e
l
ine
ind
ic
at
e
to
trai
ning
accu
r
acy
and
the
bl
ack
li
ne
in
dica
te
to
validat
io
n
acc
ur
acy
,
w
hile
in
the
seco
nd
sh
a
pe
the
re
d
li
ne
ind
ic
at
ed
to
tr
ai
ni
ng
l
oss
an
d
the
blac
k
li
ne
ind
ic
at
e to
v
al
i
dation l
os
s
.
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
Detect
ion
of ci
trus leaf
dise
ase
s u
si
ng a dee
p l
ear
ning tech
ni
qu
e (
A
hm
e
d R
. Lua
i
bi)
1723
Table
1
.
T
he
v
al
idati
on
a
nd
t
est
a
ccur
a
cy
fo
r
e
ac
h
m
od
el
wi
th
d
at
a
a
ugm
e
ntati
on
Test
Mod
el
Valid
atio
n
Accura
cy
Test Ac
cu
rac
y
Elaps
ed
T
i
m
e
Alex
Net
9
7
.92
%
9
0
.00
%
14
m
in
9
se
c
Res
Net
9
5
.83
%
8
5
.00
%
31
m
in
12
sec
Table
2
.
T
he
v
al
idati
on
a
nd
t
est
a
ccur
a
cy
fo
r
e
ac
h
m
od
el
wi
tho
ut
d
at
a
a
ug
m
entat
ion
Test
Mod
el
Valid
atio
n
Accura
cy
Test Ac
cu
rac
y
Elaps
ed
T
i
m
e
Alex
Net
9
5
.83
%
9
0
.00
%
13
m
in
28
sec
Res
Net
9
3
.75
%
8
0
.00
%
30
m
in
36
sec
Figure
4.
Vali
da
ti
on
a
cc
ur
acy
for Ale
xNet
and Res
Net
Alex
Net w
it
h a
ugmentat
io
n
Alex
Net A
cc
ur
acy
Re
sNet A
cc
ura
cy
Re
sNet w
it
h a
ugmentat
io
n
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.
11
, No
.
2
,
A
pr
i
l
2021
:
1719
-
1727
1724
Figure
5
.
Vali
da
ti
on
l
oss
for Ale
xNet
and R
esNet
Figure
6
(
a
)
s
hows
t
he
te
st
c
onf
us
io
n
m
at
ri
x
f
or
Ale
xN
et
with
data
au
gm
entat
ion
,
the
Ph
yl
loc
nisti
s
ci
trel
la
i
m
ages
are
predict
ed
wrong
once
a
s
scal
e
insect
.
F
igure
6
(
b
)
s
how
s
t
he
resu
lt
s
of
Ale
xN
et
c
onf
us
io
n
m
at
rix,
it
is
s
how
t
hat
the
heathy
im
ages
are
pr
e
dicte
d
wro
ng
as
la
c
k
of
el
em
ent
diseaes
a
nd
th
e
ci
trel
la
d
ise
aes
predict
ed
w
ron
g
once
as
scal
e
insect
diseaes.
T
he
bl
ue
cel
l
ind
ic
at
ed
to
the
c
orre
ct
pr
edict
io
n
a
nd
t
he
pink
on
e
in
dic
at
ed
to
the
wro
ng
pre
dicti
on
.
Figure
7
(
a
)
s
hows
t
he
te
st
co
nfusion
m
at
rix
f
or
Re
s
Net
wi
th
data
aug
m
entat
ion
,
the
healt
hy
i
m
ages
are
predic
te
d
wrong
twic
e
as
la
ck
of
el
em
ents
and
the
Ph
yl
locnisti
s
ci
trel
la
i
m
ages
pr
edict
ed
wro
ng
on
ce
as
scal
e
insect
,
Figure
7
(
b
)
sh
ow
s
that
the
heathy
i
m
ages
are
pr
e
dicte
d
wrong
three tim
es as lack
of ele
m
ent
s,
a
nd the lac
k of el
em
ent p
re
dicte
d wro
ng once as
h
eat
hy one
.
AlexNe
t Loss
ResNet
Loss
AlexNe
t
wi
th au
gme
nt
at
ion
ResNet
with
aug
me
nt
at
ion
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
Detect
ion
of ci
trus leaf
dise
ase
s u
si
ng a dee
p l
ear
ning tech
ni
qu
e (
A
hm
e
d R
. Lua
i
bi)
1725
(a)
(b)
Figure
6
.
The
test
conf
us
io
n m
at
rix
for Ale
xN
et
,
(a
)
w
it
h data au
gm
entat
ion
,
(
b)
w
it
ho
ut
d
at
a
au
gm
entat
ion
(a)
(b)
Figure
7
.
The
test
conf
us
io
n m
at
rix
for
Re
s
Net
,
(a)
w
it
h d
at
a aug
m
entat
ion
,
(b)
w
it
hout
data a
ug
m
entat
ion
4.
CONCL
US
I
O
N
In
t
his
pap
e
r
t
wo
m
od
el
s
of
deep
CN
N
na
m
ed
Alex
Net
a
nd
Re
sNet,
eac
h
m
od
el
use
d
t
o
te
st
a
set
of
i
m
ages
con
sist
of
healt
hy
an
d
dif
fer
e
nt
ty
pe
s
of
ci
tr
us
le
a
ves
disease
s
P
hyll
ocn
ist
is
ci
trel
la
,
l
ack
of
e
lem
en
t
and
scal
e
insec
ts.
The
res
ults
sh
ow
that
Alex
Net
giv
es
the
be
st
accuracy
w
it
h
data
aug
m
e
ntati
on
97.
92%
an
d
Re
sNet
gav
e
95.83%
w
hile
the
resu
lt
s
w
it
ho
ut
data
au
gm
entat
ion
giv
e
le
ss
accura
cy
with
95
.
83
%
fo
r
Alex
Net an
d
93.
75
f
or
Res
Ne
t, from
the r
esu
lt
s
we
co
nclu
de
that trai
ning
DL
ne
ural
n
et
work
m
od
el
s
o
n
m
or
e
data
will
le
ad
to
m
or
e
sk
il
lful
m
od
el
s,
an
d
a
ug
m
entat
ion
te
chn
i
qu
e
s
will
gen
e
rate
im
ag
e
var
ia
ti
ons
th
at
can
boos
t
ap
pro
pr
i
at
e
m
od
el
s'
abili
ty
to
gen
e
rali
ze
wh
at
they
ha
ve
le
ar
ned
t
o
new
im
ages.
T
he
el
apse
d
ti
m
e
for
trai
ning
show
s
that
AlexN
et
is
the
si
m
plest
structu
re
tha
n
Re
sNet
with
a
trai
ning
tim
e
of
14
m
in
9
sec.
Also
,
for
far
t
her
a
na
ly
sis
the
con
f
usi
on
m
at
rix
for
the
two
m
odel
s
done
f
or
th
e
te
st
i
m
ages
as
a
real
te
st
f
or
the
m
od
el
s.
All t
he
work is
done
with MA
TL
A
B R
2019b.
ACKN
OWLE
DGE
MENTS
Th
e
a
uthors
w
ou
l
d
li
ke
to
th
ank
M
us
ta
ns
i
riy
ah
U
niv
er
sit
y
(www.
uom
us
ta
ns
iriy
ah.
e
du.i
q)
Ba
ghda
d
-
Ir
a
q for it
s s
upport in
the
pres
ent wo
rk.
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.
11
, No
.
2
,
A
pr
i
l
2021
:
1719
-
1727
1726
REFERE
NCE
S
[1]
C
ent
ra
l
Sta
ti
sti
c
al
Organ
izati
on
,
“
Ira
q
-
Dire
ct
or
ate
of
Agr
ic
ul
tura
l
Stat
ist
ic
s
,
”
2019
.
[2]
B.
Scal
e
,
“
Scal
e
insec
ts
A
diffi
c
ult
proble
m
tha
t
ca
n
be
m
ana
ge
d,
”
Pl
an
t
Health
Australi
a
,
Ind
ustry
Bi
osec
urit
y
Pl
an
for t
he
Nur
sery
Industry
,
Pl
ant
He
al
th
Aus
tr
al
i
a, Ca
nber
r
a,
ACT,
2013
.
[3]
M.
A.
S.
Adhiwi
bawa
,
“
Detect
io
n
of
Anom
al
ie
s i
n
Cit
rus L
e
ave
s
Us
ing
Digit
al
Im
age
Proce
ss
ing a
nd
T
2
Hotelling
Multi
var
i
ate
Co
ntrol
Chart,”
20
19
Int
ernati
onal
Conf
ee
n
ce
of
A
rtif
icial
Int
el
l
igence
and
Inf
orm
a
ti
on
Technol
og
y
,
2019
,
pp
.
310
-
3
14
.
[4]
M.
R.
Minar
an
d
J.
Nahe
r,
“
Rec
ent
Advan
ce
s
i
n
Dee
p
L
ea
rni
n
g:
An
Overvi
ew
,
”
arXi
v:
1807.
0
8169
,
pp.
1
-
31,
2018.
[5]
K.
Li
,
et
al
.
,
“
Us
ing
dee
p
lear
ning
for
Im
age
-
Based
diffe
r
ent
degr
ee
s
of
gin
kgo
le
af
d
isea
s
e
class
ifi
cation
,
”
Inf
orm
ati
on
,
vo
l. 11, no. 2,
pp.
95
-
108,
2020
.
[6]
A.
Krizh
evsk
y
,
et
al.
,
“
Im
age
Ne
t
cl
assifi
ca
t
ion
with
dee
p
convo
lut
iona
l
neur
a
l
n
et
works
,
”
Com
mun
ic
ati
on
of
th
e
ACM
,
vo
l. 60, n
o.
6
,
pp
.
84
-
90
,
2017.
[7]
S.
Kaur,
et
al.
,
“
Plant
Disea
se
Cla
ss
ifi
c
at
ion
us
ing
Dee
p
L
ea
rni
ng
Google
Net
Model,
”
In
te
rna
ti
onal
Journal
o
f
Innov
ative
Te
ch
nology
and
E
xplor
ing
Engi
n
ee
ri
ng
(
IJI
TEE)
,
vol
.
8
,
no
.
9
,
pp
.
31
9
-
322,
2019
.
[8]
N.
F.
Sahid
an,
e
t
al.
,
“
Eva
lu
at
io
n
of
basic
convo
lut
ional
neur
al
n
et
work
and
b
ag
of
feature
s
for
l
ea
f
r
ec
ogni
ti
on
,
”
Indone
s
ian
J
our
nal
of
Elec
tr
ic
al
Eng
ineering
and
Comput
er
Sc
i
en
ce
(
IJEECS)
,
vol
.
14
,
no
.
1
,
pp
.
3
27
-
332,
2019
.
[9]
K.
P.
Fere
nt
inos,
“
Dee
p
learni
ng
m
odel
s
for
pla
n
t
disea
se
d
etec
t
i
on
and
dia
gnosi
s
Dee
p
learni
ng
m
odel
s
for
pla
n
t
disea
se
d
etec
t
ion
and
d
ia
gnosis,
”
Comput
ers and
El
e
ct
ron
ic
s
in
A
gric
ult
ure
,
vol
.
1
45,
pp
.
311
-
318
,
2018.
[10]
S.
Xing
,
e
t
a
l.
,
“
Cit
rus
Pests
and
Dis
ea
ses
R
ec
ogni
ti
on
Mod
el
Us
ing
W
ea
kl
y
Dense
Conne
ct
ed
Convo
lut
io
n
Network,
”
Senso
rs
,
vol. 19, no. 1
4,
pp
.
1
-
18
,
201
9.
[11]
A.
Mikołajc
z
y
k
and
M.
Groch
ows
ki,
“
Data
a
ugm
ent
at
ion
for
improving
de
e
p
learni
ng
in
i
m
age
class
ifi
c
ation
proble
m
,
”
2018
Int
ernati
onal
Int
erdisci
p
li
nary
P
hD Work
shop
(
IIP
hDW 2018
)
,
20
18
,
pp
.
117
-
122
.
[12]
A.
Y.
A.
Sal
a
m
a,
et
a
l.
,
“
Sheep
Id
ent
if
ication
Us
ing
a
H
ybrid
Dee
p
L
ea
r
ning
and
Ba
y
es
ia
n
Opt
imiza
t
io
n
Approac
h,
”
IE
E
E
A
ccess
,
vol
.
7
,
pp.
31681
-
3168
7,
2019
.
[13]
M.
Zul
q
arn
ai
n
,
et
al
.
,
“
A
compa
ra
ti
v
e
rev
ie
w
on
dee
p
learni
ng
m
odel
s
for
te
x
t
cl
assifi
ca
t
ion,”
I
ndones
ian
J
ourn
al
of
E
le
c
tr
ic
al
En
g
ine
ering
and
C
omput
er
Sci
ence
(
IJE
ECS)
,
vol
.
19,
no
.
1
,
pp
.
18
56
-
1866,
2020
.
[14]
Y.
Le
cun
,
et
a
l.
,
“
Gradi
ent
-
Base
d
Le
arn
ing
Appl
ie
d
to
Docum
ent
Rec
ognition,”
P
roce
edi
ngs
of
th
e
IEE
E
,
vol.
86
,
no.
11
,
pp
.
2278
-
2324,
1998
.
[15]
I.
Nam
at
ēvs,
“
Dee
p
Convolut
i
onal
Neura
l
Net
works
:
Struct
ure
,
Feat
ur
e
Ext
ra
ct
ion
and
Traini
ng,
”
Inf
orm
ati
o
n
Technol
ogy
and
Manag
eme
nt
S
ci
enc
e
,
vo
l. 20, no
.
1
,
pp
.
40
-
47
,
2
018.
[16]
A.
Khan,
et
al.
,
“
A
Surve
y
of
the
Recent
Arc
hit
e
ct
ure
s
of
Dee
p
Convol
ut
io
nal
Neura
l
Ne
t
works
,
”
Arti
fici
al
Inte
lligen
ce R
e
view,
vo
l. 53,
pp.
5455
-
5516
,
20
20
.
[17]
A.
Ahm
ad,
“
Obj
ec
t
Rec
ogn
it
ion
in
3D data
using
Capsule
s,
”
The
s
es,
S
y
ra
cuse
Uni
ver
sit
y
,
2018.
[18]
C.
Y.
Lee,
et
a
l.
,
“
Gen
erali
z
in
g
pooli
ng
func
t
i
ons
in
convo
lutional
neur
a
l
n
etw
orks:
Mixed,
gat
ed
,
and
tr
ee,
”
Proc
ee
d
ings
of
19th
Int
ernat
ion
al
Conf
ere
n
ce
o
n
Arti
f
ic
ia
l
Int
ell
ige
n
ce
S
tat
ist
ic
AISTATS
2016
,
vol.
51
,
2016
,
pp.
464
-
472
.
[19]
E.
Cengİl,
“
Mu
lt
iple
Cla
ss
ifi
c
ation
of
Flower
I
m
age
s
Us
ing
T
ran
sfer
Le
arn
in
g,
”
2019
Int
ern
ati
onal
Artif
icia
l
Inte
ll
ige
n
ce and
Data
Proce
ss
ing
Symp
osium
(
IDAP
)
,
2019
,
pp
.
1
-
6
.
[20]
R.
I.
B
endj
i
ll
a
l
i,
e
t
a
l.
,
“
Ill
um
ina
ti
on
-
robust
f
ac
e
r
ec
ogn
it
ion
base
d
on
de
e
p
convol
ut
iona
l
neur
al
n
et
wor
ks
arc
hi
te
c
ture
s,
”
Indone
s
ian
J
ournal
of
El
ectr
ic
a
l
Eng
ine
ering
a
nd
Comput
er
S
ci
en
ce
(
IJE
ECS
)
,
vol.
18,
no.
2,
pp.
1015
-
1027
,
2020.
[21]
A.
Sacr
é, “Effi
c
i
ent
In
te
r
ac
t
ive Annota
ti
o
n
for
C
y
tomine,
”
Univer
sité
d
e Li
èg
e, L
i
ège
,
Be
lgi
que
,
2
019.
[22]
N.
N.
A.
A.
Ha
m
id,
et
al
.
,
“
Com
par
ing
bags
of
fea
ture
s,
conv
e
nti
onal
convo
lutional
neur
al
netw
ork
and
al
exnet
for
fruit
re
cogni
t
ion
,
”
Indone
s
ian
J
o
urnal
of
El
e
ct
r
i
c
al
Eng
ineering
and
Comput
er
Sci
en
ce
(
IJE
ECS)
,
vol.
14
,
no.
1
,
pp.
333
-
339
,
20
19.
[23]
S.
Chandra
,
et
al.
,
“
Frontie
rs
in
I
nte
lligent
Com
p
uti
ng:
Th
eor
y
an
d
Applic
ations
,
”
Proce
edi
ngs
of
the
Inte
rnat
ional
Confe
renc
e
on
F
ICTA
2018
,
vo
l.
2
,
2020
.
[24]
Y.
Li
,
e
t
al
.
,
“
SA
R
Ship
Dete
ct
i
on
Based
on
Resnet
and
Tra
nsfe
r
Le
arn
ing
,
”
201
9
IEE
E
Int
ernat
i
onal
Geosci
ence
and
Re
mot
e
S
en
s
ing
Symp
osium
,
IGAR
SS,
2019
,
pp.
1188
-
1191
.
[25]
K.
He,
e
t
al.
,
“
Dee
p
residual
learni
ng
for
imag
e
rec
ogn
it
ion
,
”
I
EE
E
Conf
ere
nc
e
on
Comput
er
Vis
ion
and
Pat
t
ern
Re
cogn
it
ion
,
20
16
,
pp
.
770
-
778
.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Ah
me
d
R.
Lu
aibi
was
born
in
Baghda
d
,
Ira
q
in
1989.
He
rec
e
ive
d
h
is
B.
Sc.
d
egr
ee
i
n
Com
m
unic
at
ion
Engi
ne
eri
ng
fro
m
al
-
Mansour
Univer
sit
y
Col
lege,
B
aghda
d
,
Ir
aq
in
2015
.
His
rec
en
t
rese
a
rch
a
ct
ivit
y
is
citrus
disea
se
class
ifi
c
ation
and
area
de
t
ec
t
ion
using
image
proc
essing
.
Email
:
ahmedraheem175@
gm
ai
l.
com
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
Detect
ion
of ci
trus leaf
dise
ase
s u
si
ng a dee
p l
ear
ning tech
ni
qu
e (
A
hm
e
d R
. Lua
i
bi)
1727
Tari
q
M.
Salm
an
was
born
in
Baghda
d,
Ira
q
in
1972.
He
obta
in
ed
his
B.
Sc.
in
El
ectri
ca
l
Engi
ne
eri
ng
in
1995,
an
M.Sc.
in
Com
m
unic
at
ion
Engi
n
ee
r
in
g
in
2003
at
th
e
Univer
sit
y
of
Te
chno
log
y
,
Ira
q,
and
Ph.D.
in
Te
l
ec
om
m
unic
ation
and
Network
devi
ce
s
in
2012
at
Belaruss
ia
n
Stat
e
Univer
sit
y
of
Inform
at
ic
s
and
Radi
o
El
e
ct
ro
nic
s,
Belarus.
From
2006
to
201
2
he
worked
as
a
le
c
ture
r
in
the
El
e
ct
ri
ca
l
Engi
n
ee
ring
Fa
cul
t
y
,
at
Al
-
Mus
ta
nsiri
y
ah
Univ
ersity
,
Ira
q.
Sinc
e
the
begi
nning
of
20
18,
he
works
as
an
assistant
profe
ss
or
in
the
sam
e
F
ac
ulty
.
He
is
a
consult
ant
m
ember
of
the
Ir
aqi
Eng
ine
er
ing
union
since
201
3.
He
is
in
te
rest
e
d
in
the
subj
ect
of
wire
le
ss
and
net
work
device
s
,
vide
o
,
and
image
proc
essing
s
y
stems
.
Mus
ta
n
siri
y
ah
Univer
si
t
y
Engi
n
ee
ring
Facul
t
y
,
E
lectr
i
c
al
Engi
ne
eri
ng
Depa
rtme
nt,
Ir
a
q
.
Email
:
tjibori@g
m
ai
l.
com
Ab
ba
s
Hus
sien
Mi
r
y
He
r
ec
e
iv
ed
his
B.
Sc.
d
eg
ree
in
E
lectr
i
ca
l
Engi
neering
in
2005
from
the
Mus
ta
nsiri
y
ah
Univer
sit
y
and
his
M.Sc.
d
egr
e
e
in
cont
ro
l
and
computer
enginee
ring
in
2007
from
Baghda
d
Univer
sit
y
.
He
rec
e
ive
d
a
Ph.
D.
degr
e
e
in
2
011
in
con
trol
and
comput
er
engi
ne
eri
ng
fro
m
the
Basra
h
Univer
sit
y
,
Ira
q.
In
2007,
he
joi
n
ed
t
he
fa
cul
t
y
of En
gine
er
ing
a
t
th
e
Mus
ta
nsiri
y
ah
Univer
sit
y
in
B
aghda
d.
His
re
ce
nt
r
ese
ar
ch
a
ct
ivitie
s
are
image
pro
ce
ss
ing,
art
if
ic
i
al
intellig
enc
e
,
cont
rol
,
roboti
cs,
and
sw
arm
opti
m
iz
at
ions,
now
he
has
be
en
an
As
sist.
Prof.
at t
h
e
Mus
t
ansi
ri
y
ah
Univer
sit
y
,
Ir
aq
.
Email
:
abba
sm
ir
y
83@uom
ustansiriy
ah
.
edu
.
iq
Evaluation Warning : The document was created with Spire.PDF for Python.