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.
4277
~4
286
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v9
i
5
.
p
p4
277
-
42
86
4277
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
A
fuz
zy syst
em for d
etect
ion and
classific
ation o
f te
xtile def
ec
t
s
to ensur
e the qu
ality of f
abric pr
oducti
on
Iman Su
bhi
Moham
med,
I
sra
a Moham
med A
lh
am
dani
Depa
rtment
o
f
C
om
pute
r
Scie
n
ce,
Univer
si
t
y
of
Mos
ul
,
Ira
q
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Des
16
, 201
8
Re
vised
A
pr
16
, 2
01
9
Accepte
d
Apr
27
, 20
19
The
ai
m
of
thi
s
rese
arc
h
foc
us
es
on
construc
t
a
computer
ized
sy
st
em
for
te
xtile
d
efect
s
det
e
ct
ion
.
Th
e
s
y
stem
m
erg
es
bet
wee
n
image
proc
essing
m
et
hods,
stat
ist
i
ca
l
m
et
hods
in
addi
ti
on
to
the
Inte
lligen
t
tech
nique
s
via
Neura
l
Network
and
Fuzz
y
Logic.
Gabor
fil
t
ers
were
used
to
ide
nti
f
y
edge
s
and
to
highl
igh
t
def
ec
t
ive
a
rea
s
in
fab
ric
imag
es,
the
n
to
t
rai
n
the
neur
a
l
net
work
on
stat
i
stic
al
and
geom
et
r
y
f
eature
s
der
ive
d
from
fab
ric
images
to
form
the
spec
ial
neur
al
n
et
wor
k
disti
nguish
an
d
cl
assif
y
def
e
c
ts
int
o
the
fourte
en
cate
go
rie
s,
which
are
the
m
ost
comm
on
def
ec
ts
in
the
t
ext
i
l
e
fac
tor
y
.
The
pro
posed
work
inc
lude
s
two
phase
s
.
The
first
phase
is
to
det
ect
the
def
ects
in
f
abr
ic
s.
The
sec
ond
phase
is
th
e
class
ifi
c
at
ion
phase
of
th
e
def
ect.
At
the
def
ect
det
e
ction
stage
,
a
Discr
et
e
Cosine
Tr
a
nsfer
(DCT)
conve
rts
th
e
ima
ges
to
the
f
req
ue
nc
y
dom
ai
n.
Im
a
ge
feature
s
th
en
dr
awn
and
int
roduc
e
th
em
to
the
E
lman
Neura
l
Ne
twork
to
d
et
e
ct
the
exi
sten
ce
o
f
def
ects.
In
the
class
ifi
cation
stag
e,
th
e
images
ar
e
conve
r
te
d
to
th
e
fre
que
n
c
y
dom
ai
n
b
y
the
Gabor
fil
t
er
and
the
n
the
image
fea
tu
res
ar
e
ex
tra
c
te
d
and
insert
ed
int
o
th
e
bac
k
pr
opaga
t
i
on
net
work
to
cl
assif
y
th
e
fab
ri
c
def
ects
in
those
images.
Fuzz
y
logi
c
is
th
en
app
li
ed
to
n
eur
al
net
work
o
utput
s
an
d
int
erf
ere
n
ce
va
l
ues
are
used
in
fuz
z
y
log
ic
to
i
ncr
ea
se
f
inal
dis
cri
m
ina
t
ion.
W
e
eva
luate
a
d
i
stinc
ti
on
r
at
e
of
91.
4286%
.
After
appl
y
ing
th
e
fuz
z
y
logi
c
to
neur
al ne
twork
o
utput
; the
discr
i
m
ina
ti
on
r
at
e
wa
s ra
ised to
97
.
14
28%
Ke
yw
or
d
s
:
El
m
an
ne
ur
al
ne
twork
Fabri
c d
e
fect
de
te
ct
ion
Fu
zzy
L
ogic
Gabo
r
filt
er
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights
reserv
ed
.
Corres
pond
in
g
Aut
h
or
:
Isr
aa
Mo
ham
med Alham
dan
i,
Dep
a
rtm
ent o
f C
om
pu
te
r
Scie
nces,
Mosu
l
U
niv
e
rs
it
y,
Mosu
l,
Iraq
.
Em
a
il
: esra
m
k6
5@
gm
ail.co
m
1.
INTROD
U
CTION
Qu
al
it
y
con
tr
ol
of
industrial
pro
duct
s
has
be
com
e
increasin
gly
i
m
po
rtant.
Qu
al
it
y
con
tr
ol
is
def
ined
as
"k
nowled
ge
of
c
onf
or
m
it
y
of
pro
du
ct
s
t
o
sta
nd
a
rd
sp
eci
f
ic
at
ion
s
an
d
know
le
dg
e
of
the
causes
of
de
viati
on
and
necessa
ry
procedu
res".
The
te
xtil
e
ind
us
t
ry
is
an
industry
that
r
equ
i
res
qual
it
y
to
m
eet
co
ns
um
e
r
dem
and
[1
]
.
The
m
ai
n
iss
ue
he
re
is
how
fa
br
ic
scre
enin
g
i
m
pr
ove
s
producti
on
qu
al
it
y.
Unde
r
wh
at
ci
rcu
m
sta
nces
?
The
process
of
check
i
ng
the f
ab
rics
to
this
day
is
a
final
st
ep
befo
re
ha
nding
the
fa
br
ic
s to
the
store
a
nd
the
n
m
ark
et
ing
th
e
m
.
This
pr
oc
ess
is
sti
ll
m
a
nu
al
by
a
gro
up
of
s
kill
ed
sta
ff
with
the
utm
os
t
accuracy
from
60%
to
70%
only
[2
]
.
To
ac
hieve
t
his,
t
he
m
od
ern
te
xtil
e
industry
has
de
velo
ped
hi
gh
-
sp
ee
d
m
achines to
pr
oduce
high
-
qu
al
it
y fabr
i
cs in
the s
hortest
po
ssible t
i
m
e [3
,
4].
The
propose
d
wor
k
f
ocuses
on
so
l
ving
the
pr
ob
le
m
of
te
xtil
e
def
ec
ts
detect
ion
m
anu
al
ly
by
autom
at
e
te
xtile
def
ect
s
via
a
com
pu
te
rize
d
syst
em
.
The
com
pu
te
rized
syst
e
m
is
req
uire
d
to
c
hec
k
these
def
ect
s
in
reli
abili
t
y
and
flexibili
ty
m
ann
er.
T
his
syst
e
m
m
erg
es
the
im
age
processi
ng
m
et
ho
ds,
sta
ti
sti
cal
m
et
ho
ds,
intel
li
gen
t
te
ch
niqu
es
via
f
uzzy
lo
gic
an
d
ne
ural
netw
ork
f
or
re
cogniti
on.
A
s
well
as
to
buil
d
a
li
ve
te
xtil
e
databas
e
im
age
du
e
t
o
the
el
im
inate
d
e
xisti
ng
te
xt
il
e
database
a
nd
the
dif
ficul
ty
to
get
t
he
e
xisti
ng
on
ce
.
The
pro
po
s
ed
work
de
te
rm
ined
the
de
fecti
ve
areas
in
fabric
i
m
ages.
The
ge
om
et
rical
and
sta
ti
sti
ca
l
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
:
4
2
7
7
-
4
2
8
6
4278
featur
e
s
we
re
der
i
ved
t
o
trai
n
the
neural
ne
tworks
us
e
d.
The
propose
d
work
c
onstr
uct
ed
with
t
wo
pha
ses.
The
fi
rst
phase
is
to
detect
th
e
def
ect
s
w
hile
the
seco
nd
cl
a
ssify
them
.
W
e
go
t
91.42
86%
then
a
fter
a
pply
ing
fu
zzy
l
og
ic
t
o
neural
netw
ork
outp
ut,
t
he
re
cogniti
on
rate
was
raised
to
98.14
28%.
Re
st
of
this
resea
rch
i
s
structu
re
d
as
t
he
relat
ed
wor
k
in
this
sect
io
n
the
relat
ed
te
xtil
e
def
ect
s
s
yst
e
m
and
m
eth
od
were
pr
es
ented
,
m
et
ho
ds
of
fa
bri
c
te
st
with
va
rio
us
kind.
A
descr
i
ption
of
Ar
ti
fici
al
I
ntell
igent
Tec
hniq
ue
s
us
e
d
in
t
his
work
al
so
Ga
bor
F
il
te
r.
Anothe
r
sect
ion
deal
with
the
propose
d
work
al
gorit
hm
,
database
co
nfi
gurati
on
.
Fabri
c
def
ect
s
sta
ge
we
re
th
e
def
ect
s
is
de
no
te
d,
fa
bri
c
de
fects
cl
assifi
cat
ion
,
r
esults
wh
ic
h
are
fou
nd
a
nd
discuss
i
on,
pe
rfor
m
ance
eff
i
ci
ency
of
the
pro
po
se
d
w
or
k,
co
ncl
us
io
n,
ind
ee
d,
re
fere
nces.
T
he
ne
w
idea
adap
te
d
on
thi
s
s
yst
em
ov
er
oth
e
r
relat
e
d
work
is
the
hy
per
act
ive
bet
w
een
fu
zzy
l
og
i
c
with
ne
ur
al
netw
ork
that affe
ct
s the
resu
lt
s i
n
this
work.
2.
RELATE
D
W
ORK
In
orde
r
to
e
nh
ance
the
qual
it
y
con
tr
ol
of
th
e
te
xtil
e
produ
ct
ion
we
co
ns
t
ru
ct
a
c
om
pu
te
rized
syst
em
for
def
ect
s
det
ect
ion
a
nd
cl
as
sific
at
ion
.
The
co
ns
tr
uction
of
s
uch
a
powerfu
l
syst
em
require
d
deep
stu
dy
on
the
oth
er
s
m
eth
ods
bel
ong
t
o
the
te
xtil
e
det
ect
ion
t
o
el
im
i
nate
their
dr
a
wb
ac
k
an
d
e
xc
eed
it
in
our
s
yst
e
m
.
Re
view of
a
num
ber
o
f
t
he rel
at
ed
w
ork
to
te
xtil
e d
e
fects
d
e
te
ct
ion
is
dissc
us
se
d belo
w:
In
20
12,
resea
rch
e
rs
A
nand
H.
K
ulk
a
rn
i
a
nd
S
heetal
B.
Pati
l
[2
]
int
rod
uced
a
patte
r
n
of
detect
io
n
and
ide
ntific
at
ion
of
de
fects
of
ti
ssu
e
ba
s
ed
on
G
ray
Level
Co
-
occ
urren
ce
Ma
trix
(
GLCM)
as
we
ll
as
a
pro
bab
il
ist
ic
ne
ur
al
netw
ork.
The
overall
rat
e
of
s
uccess
of
ti
ssu
e
sel
ect
io
n
was
96.6%
a
nd
the
su
cces
s
rate
of
def
ect
detect
io
n
was
91.
1%
[
5].
Mi
chael
K.
N
ng
et
al
.
an
al
yz
ed
pa
per
a
nd
ti
ssu
e
patte
rn
s
of
t
he
patte
rn
e
d
fabrics
in
sev
eral
ty
pe.
Th
ey
exa
m
ined
and
photog
raphed
de
fects
in
these
fa
br
ic
s
us
in
g
the
conv
e
x
op
ti
m
iz
ation
al
gorithm
their
r
ecognit
ion
rate
ran
ge
d
bet
we
en
(94.9%
-
99.6
%
)
[6
]
.
I
n
20
15,
Ali
Ra
bh
i
et
al
.
reli
ed
on
local
ho
m
og
en
ei
ty
and
neural
net
work
t
o
judge
def
ect
s
in
ti
ss
ue
i
m
ages
after
ap
plyi
ng
DC
T
and
extr
act
in
g
e
nergy
c
har
act
erist
ic
s
from
i
m
ages[
7
]
,
with
a
de
te
ct
ion
r
at
e
of
97.35%
[
8
]
.
Ma
le
k
use
F
F
T
an
d
cro
ss
co
rr
el
at
i
on
m
et
ho
d
si
m
ul
ta
neo
usl
y
to
detect
14
ty
pe
of
def
ect
s
with
acc
urat
e
res
ults
with
100%
recog
niti
on
rat
e [
9].
M.
Handm
and
ulu
et
.al.
Go
t
70
%
r
at
e
wi
th
25
te
xtil
e
def
ect
an
d
no
rm
al
te
xtil
e
i
m
ages
[10].
As
prese
nted
by
P.Y
.
K
um
bhar
et
al
.
in
a
new
m
et
ho
d
wa
s
based
on
the
gen
et
ic
al
gorit
hm
,
the
extracti
on
of
certai
n
geo
m
et
rical
cha
racteri
sti
cs
al
so
t
he
use
of
SV
M
as
a
cl
assifi
cat
ion
of
ti
ssu
e
def
e
ct
s.
T
he
resear
cher
s
su
ccee
ded
in
cl
assify
ing
90.0% of
the d
e
fect
s [
11]
.
Soo and
Tae b
ased on Wavelet
transf
or
m
at
ion
w
it
h
GMM
they
get
acc
urat
e
resu
lt
s
[12
]
.
In
the
ye
ar
2017,
resea
rcher
S
un
il
Ba
ngare
et
al
.
s
ugge
ste
d
a
m
et
hod
f
or
detect
ing
ti
ssue
de
fects
us
in
g
RGB,
HSV
te
chn
i
qu
e
s
a
nd
im
age
proces
sing
te
ch
niques,
an
d
the
acc
uracy
of
detect
ion
was
96.15%
[
13
]
.
I
n
the
ye
a
r
20
18,
S
hua
ng
Me
i
et
.al.
us
i
ng
G
aussian
Pyram
id'
s
autom
atic
no
ise
reducti
on
netw
ork. The
accu
r
acy
o
btain
ed
of the
f
inal
res
ul
ts reache
d
m
or
e than 8
0.0%
[14
]
.
3.
FABRI
C
TE
S
T MET
HO
DS
The
featu
res
extracti
on
m
eth
ods
a
re
cl
as
sifie
d
int
o
th
r
ee
cat
egories
they
are.
The
sta
ti
sti
cal
appr
oach
es
t
ha
t
rely
on
sta
ti
sti
cal
beh
avior
i
n
areas
w
it
ho
ut
def
ect
s
.
Su
c
h
as
sta
ti
sti
cal
m
o
m
ent
and
cro
ss
-
c
orrelat
ion
[15
-
17]
.
Th
e
sp
ect
ral
ap
proach
e
s
that
are
us
ed
w
hen
sta
ti
sti
cal
m
et
ho
ds
are
una
ble
to
detect
the
de
fects
tha
t
app
ea
r
w
he
n
subtl
e
and
ve
ry
subtl
e
trans
form
ation
s
occ
ur.
F
or
e
xam
ple,
the
Ga
bor
f
il
te
rs,
FFT
a
nd
wa
ve
le
ts
transfo
rm
[15
-
17]
.
Mo
de
ls
base
d
a
ppr
oa
ches,
this
m
od
el
is
s
uitable
for
f
ab
ric
im
ag
es
that
con
ta
in
rand
om
su
rf
ace
va
riat
ion
s
as
well
as
rando
m
or
hand
-
patte
r
ne
d
fab
rics
.
The
Gau
s
s
Ma
rko
v
Ra
ndo
m
Fiel
d
m
od
el
is
on
e
of t
he
m
eth
ods i
n
t
his f
ie
ld [1
3,
16, 1
7]
4.
AR
TIF
ICIAL
IN
EL
LIGE
N
T T
ECH
NIQUES
The
s
pee
d
with
wh
ic
h
the
c
om
pu
te
r
is
in
it
s
m
a
the
m
atical
op
e
rati
ons
co
uld
be
us
e
d
in
m
any
non
-
m
at
he
m
a
tical
ta
sk
s
as
we
ll
.
Neu
ral
netw
orks
hav
e
bee
n
us
ed
as
a
n
int
el
li
gen
t
te
chn
i
qu
e
for
the
det
ect
ion
and cla
ssific
at
ion o
f defects
.
The follo
wing
N
e
ural
Net
w
ork were
u
se
d
i
n t
his work.
4.1.
Ne
ural
n
etwork
back p
ropaga
tion
(B
PNN)
This
is
base
d
on
t
rainin
g
in
the
erro
r
co
rr
e
ct
ion
patte
r
n.
The
sig
nal
f
r
om
the
ou
t
pu
t
t
o
the
in
put
is
re
-
in
voke
d
as
an
le
ar
ning
sta
ge,
duri
ng
wh
i
ch
t
he
netw
ork
wei
gh
ts
are
ca
lc
ulate
d,
t
he
c
ha
ng
e
s
a
re
cal
c
ulate
d,
and
the
Me
an
Sq
ua
re
Erro
r(M
SE),
w
hich
resu
lt
s
from
t
he
dif
fer
e
nce
betwee
n
the
act
ual
ou
tp
ut
and
th
e
desire
d ou
t
pu
t,
and
us
es t
he
e
rror t
o
c
ha
ng
e
the
weig
hts to
grad
ually
r
ed
uc
e the e
rror
[
18
]
.
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
A fuzzy
system
for
detect
ion
and cl
as
sif
ic
atio
n of tex
ti
le
d
ef
ect
s to
e
ns
ure
th
e
…
(
Iman
Subhi M
ohamme
d
)
4279
4.2.
El
man
n
e
twork
I
s
on
e
of
t
he
re
current
ne
ur
al
netw
orks
it
is
c
on
st
ru
ct
e
d
fro
m
two
-
la
ye
r
wi
th
a
bac
k
propagati
on
fe
e
d
that
was
re
fle
ct
ed
in
the
fir
st
la
ye
r
re
-
in
put
to
the
first
la
ye
r
input.
This
feat
ur
e
gi
ves
the
netw
ork
the
po
s
sibil
it
y
to
r
each
t
he
op
ti
m
al
so
luti
on
wit
h
a
spe
ed
in
tra
ining
a
nd
wei
ght
c
on
tr
ol.
O
ne
of
t
he
a
dv
a
nta
ges
of
the
Elm
an
net
work
is
t
hat
it
has
a
str
ong
dynam
ic
m
e
m
or
y,
so
it
is
us
e
d
i
n
dif
fe
ren
t
fiel
ds
su
c
h
a
s
cl
assifi
cat
ion
, pre
dicti
on, dyn
a
m
ic
iden
ti
ficat
ion
syst
em
s,
and quali
ta
ti
ve c
on
t
ro
l
[19].
4.
3
.
Fuz
z
y
ne
ural ne
twor
k
T
he
ne
ural
net
work
ca
n
be
c
om
bin
ed
with
the
log
ic
ca
us
e
d
by
a
num
ber
of
m
et
ho
ds
,
i
nclu
ding
the
processi
ng
of da
ta
enteri
ng
th
e
ne
ur
al
net
work.
T
he
us
e
of
t
he
ne
ur
al
n
et
w
ork
t
o
m
od
ify
t
he
par
am
et
ers
of
the
orga
nic
f
un
ct
i
on
s
in
t
he
lo
gi
c,
or
the
ne
ural
netw
ork
outp
uts
be
f
or
e
the
c
la
ssific
at
ion
decisi
on
is
m
ade.
Use
the
th
reshold
val
ue
in
the
rati
ng
deci
sion
process;
this
sign
i
ficantl
y
aff
ect
s
the
r
at
ing
rati
o
f
or
eac
h
cat
egory
an
d
overall
syst
em
.
We
m
ay
ob
ta
in
the
hi
gh
est
r
at
ing
of
t
he
fi
r
st
def
ect
at
a
c
ertai
n
th
res
ho
l
d
li
m
it
bu
t
not
the
ap
pro
pr
ia
te
lim
it
t
hat
giv
es
the
hi
gh
est
rati
ng
of
the
def
ect
.
Th
us
,
a
ne
ur
al
ne
twork
wa
s
pro
po
s
e
d
in
w
hich
t
he
t
hr
es
hold
ro
le
was
a
bo
li
s
hed
and
re
placed
by
the
log
ic
of
the
rea
sone
d
lo
gic
cal
le
d
th
e
FPB
N
.
The
outp
ut
of
the
neural
ne
twork
was
use
d
as
i
nput
to
the
rea
sonin
g
log
ic
t
hat
wa
s
base
d
on
s
pe
ci
fic
la
ws
[
20]
.
5.
GABO
R
F
ILT
ER
The
Ga
bor
filt
ers
m
i
m
ic
the
hum
an
br
ai
n'
s
abili
ty
to
re
cognize
t
he
fa
br
ic
[21]
t
he
i
m
ages
were
segm
ented
int
o
m
any
scal
ed
i
m
ages,
eac
h
con
ta
ini
ng
va
r
yi
ng
den
sit
ie
s
on
a
narrow
ra
ng
e
of
fr
e
qu
e
nc
y
and
directi
on.
Ga
bor
filt
ers
ca
n
be
de
fine
d
as
"
a
set
of
wa
velet
s,
with
eac
h
wav
el
et
ca
ptur
ing
pow
er
at
a
give
n
fr
e
qu
e
ncy
an
d
directi
on
"
,
so
Gabor
'
s
two
-
dim
ension
al
eff
i
ci
ency
in
op
ti
m
iz
ing
the
local
info
rm
at
ion
of
the
i
m
age,
as
well
as its acc
essibi
li
ty
[
22
,
23].
6.
THE
PROPO
SED W
ORK
ALGO
RITH
M
The
pro
po
se
d
work
al
gorith
m
in
ge
ner
al
i
nclu
des
tw
o
sta
ges.
The
first
sta
ge
is
the
di
scov
e
ry
of
def
ect
s
a
nd
th
ei
r
cl
assifi
cat
i
on
i
nto
tw
o
de
fecti
ve
an
d
non
-
def
ect
ive
c
at
egories.
The
second
sta
ge
is
th
e
cl
assifi
cat
ion
of
these
de
fects
into
f
ourteen
t
ypes.
I
n
the
fi
r
st
ph
ase
,
the
im
ages
wer
e
pre
-
proce
ssed
by
resize
the
i
m
age.
Then
im
ages
wer
e
treat
e
d
with
m
ulti
ple
Gabor
filt
ers
to
highli
gh
t
t
he
exis
ti
ng
def
ect
s
.
The
res
ulti
ng
i
m
age
is
segm
ented
a
nd
the
good
sta
ti
sti
cal
char
act
e
risti
cs
for
eac
h
im
age
are
e
xtracte
d
t
o
f
orm
the
in
pu
t
m
at
rix.
T
his
m
at
rix
was
e
ntere
d
to
the
Elm
an
Neural
Net
work
th
at
was
us
e
d
to
classi
fy
the
te
xt
il
e
to
non
-
de
fecti
ve
and d
e
fecti
ve
t
extil
e. Figure
1
show
s
the
step
s of the
f
ir
st st
age.
Figure
1
.
The
texti
le
d
efects
det
ect
ion
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
:
4
2
7
7
-
4
2
8
6
4280
At
the
sec
ond
sta
ge,
the
defe
ct
ive
te
xtil
e
i
m
ages
f
r
om
the
first
sta
ge
a
re
processe
d
the
n,
the
de
fects
wer
e
dete
rm
in
ed.
Thei
r
a
ppr
opriat
e
sta
ti
sti
cal
featu
res
a
r
e
extracte
d,
an
d
the
in
put
m
a
trix
is
pr
e
pa
re
d
as
an
input
to
BPN
N.
T
he
hy
br
i
dizat
ion
betwe
en
Ba
ck
Propa
gation
Ne
ur
a
l
Networ
k
an
d
the
f
uzzy
l
og
ic
i
s
perform
ed
to
obta
in
th
e
propo
sed
Fu
zzy
Ba
c
k
P
r
op
a
gatio
n
Neural
Netw
ork
(F
BP
NN).
I
n
FBPN
N
t
he
outp
ut
is
m
ade
fu
zzy
and
re
pr
e
sente
d
as
in
pu
t
t
o
t
he
f
uzzy
lo
gic.
Fu
zzy
lo
gic
de
pends
on
a
s
pe
ci
fic
la
w
to
m
a
ke
the
final
cl
assifi
cat
ion
decisi
on
an
d
cl
assify
t
he
de
fects
if
it
is
w
it
hin
the
us
ed
f
ourteen
cat
eg
ori
es.
Fi
gure
2
s
how
these step
s.
Figure
.
2
.
De
fe
ct
s classi
ficat
ion
sta
ge
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
A fuzzy
system
for
detect
ion
and cl
as
sif
ic
atio
n of tex
ti
le
d
ef
ect
s to
e
ns
ure
th
e
…
(
Iman
Subhi M
ohamme
d
)
4281
The pr
opose
d
work de
ta
il
s is d
escri
bed as
fol
lows
:
6
.
1.
D
atabase
confi
gura
tion
The
fabric
im
a
ges
wer
e
c
ollec
te
d
li
fe
by
a
di
gital
ca
m
era
(sam
su
ng
N
V
l
ens
a
nd
8
m
ega
pix
el
)
f
ro
m
a
local
te
xtil
e
factor
y.
Als
o,
sam
ples
of
fa
bri
cs
f
ro
m
the
local
m
ark
et
.
I
m
ages
we
re
ca
ptured
by
t
he
s
cann
e
r
ty
pe
(H
P
)
an
d
accuracy
(
300
dp
i
).
W
e
c
ollec
t
def
ect
ive
an
d
non
-
def
ect
iv
e
te
xtil
e
i
m
ages
to
be
us
e
d
f
or
our
pro
po
se
d
w
ork
or
any
ot
her
s
work.
Fig
ur
e
3
sh
ows
sam
ples
of
data
base
te
xtil
e
i
m
ages.
Five
hund
red
s
ixty
RGB
i
m
ages
wer
e
c
ollec
te
d
and
sa
ve
d
on
a
com
pr
essed
form
(P
NG).
We
prefe
r
this
from
on
m
any
i
m
age
form
at
in
or
de
r
to
keep
the
de
ta
il
s
of
the
i
m
age
des
pite
com
pr
ession.
T
w
enty
hundred
e
igh
ty
i
m
age
of
the
m
wer
e
non
-
de
fe
ct
ive
i
m
ages
and
280
we
re
of
t
he
14
def
e
ct
ty
pes.
Twe
nty
i
m
age
per
four
te
e
n
de
fe
ct
s
was
colle
ct
ed
as a
texti
le
d
at
abase
.
Figure
3. Sam
ple f
r
om
v
isual
database
6
.
2
.
Te
xti
le
d
efects
sta
ge
At
this
sta
ge,
the
56
0
im
ages
are
proce
sse
d.
Ne
ur
al
net
works
wer
e
use
d
to
disti
nguish
betwee
n
def
ect
ive
and t
he non
-
de
fecti
ve fab
ric. T
his
process
is eval
uated
a
s foll
ow
s:
6
.
2.
1.
Initi
al
processin
g
The
te
xtil
e
i
m
age
is
entere
d
into
the
syst
em
.
It
is
con
ve
rted
to
a
gray
scal
e
in
order
to
el
i
m
inate
com
plex
cal
cu
la
ti
on
s
as
well
as
to
facil
it
ie
s
the
deal
with
the
im
age
detai
ls.
The
n
this
i
m
age
is
treat
ed
wit
h
Gabo
r
filt
er t
he
u
se
of this
f
il
te
r
re
flect
s
bett
er r
e
su
lt
s a
nd desc
ribe su
per
i
or f
eat
ur
es
.
6.2.2.
Im
ag
e
c
rops
and
feat
ures e
xt
r
act
i
on
The
res
ulti
ng
im
age
from
Gab
or
is
segm
ented
into
9
ti
le
s
of
e
qu
al
siz
e
a
s
show
n
in
F
i
gure
4.
F
or
eac
h
ti
le
,
featur
es
w
ere
extracte
d
r
epr
ese
nted
by
The
Ac
coe
ff
i
ci
ent
(co
r
ne
r)
,
Me
an
and
m
edian
wer
e
cal
cu
la
te
d.
These
val
ues
wer
e
norm
al
ized
in
a
vecto
r
represe
nting
i
m
age
char
act
er
ist
ic
s
and
the
f
or
m
at
ion
of
thi
s
vector
is entere
d
int
o t
he
ne
ural
n
et
work th
rou
gh the
def
ect
detec
ti
on
sta
ge.
Figure
4. Ni
ne t
il
es o
f
the
defec
t
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
:
4
2
7
7
-
4
2
8
6
4282
6.2.3. F
abri
c
d
efect detec
tion st
ag
e
The
Elm
an
Neu
ral
Net
wor
k
is
app
li
ed
to
identify
an
d
de
te
ct
wh
ere
th
e
te
xtil
e
is
de
fect
or
non
-
def
ect
ive
. Th
is
n
et
wor
k
co
ns
ist
s o
f
an
in
put l
ay
er in
cl
ud
in
g 21
in
pu
t
neurons,
a
nd
on
e
hidden
la
ye
r
co
ns
i
sti
ng
of
a
bout
10
processin
g
neur
on.
I
n
eac
h
of
t
hese
la
ye
rs,
t
he
re
is
a
la
ye
r
of
interc
onnecti
on
s
t
hat
co
nne
ct
each
la
ye
r
to
the
ne
xt
la
ye
r,
i
n
w
hi
ch
the
weig
hts
of
eac
h
inter
f
ace
are
a
dju
ste
d.
T
he
wei
gh
t
is
li
nk
e
d
to
ea
ch
pair
of
ne
uro
ns
.
T
he
w
ei
ghts
re
pr
ese
nt
a
weig
ht
vect
or
(w1,
w2,
wn),
Wh
ere
the
weig
ht
associat
ed
wi
th
the
connecti
on
be
tween
a
n
in
put
ne
uro
n
an
d
a
processin
g
el
em
ent
or
betwee
n
tw
o
treatm
ent
el
e
m
ents.
The
neuron
al
s
o
co
ntains
a
t
hresh
old
value
(
con
ti
nue
d
act
ivati
on)
that
re
gu
la
te
s
the
pro
bab
il
it
y
of
act
ivati
on
and
lim
it
s
the
outp
ut
of
t
he
cel
l
w
he
re
it
m
akes
th
e
outp
ut
with
in
the
fiel
d
[
0,
1]
or
with
in
the
fiel
d
[
1,
1].
Th
e
weig
hts
asso
ci
at
ed
with
ne
ur
onal
input
de
te
rm
ine
the
prob
a
bili
ty
of
act
ivati
on
of
the
ne
ur
on
accor
ding t
o
(
1) [1
5]:
(1)
Wh
e
re
Xi r
e
pr
esents in
put ne
uro
n
a
nd
W
i
-
ve
ct
or
weig
hts.
The
t
hr
es
hold
regulat
es
the
r
esp
on
se
of
t
he
ne
uro
n
to
fall
within
a
ce
rta
in
ra
nge
of
pre
determ
ined
values
as s
how
n
in
(2) [
24]
:
(2)
The
(2)
s
hows
the
y
ou
t
pu
t
of
t
he
ne
uron
as
a
f
unct
ion
of
f
act
ivati
on
of
the
su
m
of
n
+
1
of
t
he
balance
d
in
pu
t
,
wh
e
re
n
+
1
corres
ponds
to
n
of
i
nco
m
ing
sign
al
s.
T
he
ty
pe
of
act
ivati
on
functi
on
use
d
the
n
cal
culat
es the a
ct
ual outp
ut
va
lue, a
nd h
e
re t
he
th
res
ho
l
d functi
on is
us
ed
as in e
quat
ion (
3) [24
]
:
(3)
The
de
fect
det
ect
ion
phase
a
lso
goes
thr
ou
gh
tw
o
sta
ges
the
first
is
the
networ
k
trai
ning
in
the
char
act
e
risti
cs
of
eac
h
im
age
and
it
s
ex
pecte
d
outp
ut.
T
he
ou
t
pu
t
us
ed
t
o
dev
el
op
the
ne
twork
e
ff
ic
ie
ncy
to
te
st
and
detect
the
rest
im
ages
in
or
der
t
o
fi
nd
out
w
hethe
r
the
te
ste
d
im
ages
is
f
or
de
f
ect
ive
te
xtil
e
or
not.
The
sec
ond
st
age
is
t
he
te
st
phase
via
te
s
ti
ng
th
e
im
ages
that
no
t
t
rained
to
detect
if
these
im
age
s
are
def
ect
ive
or no
t.
The net
w
ork i
s then
trai
n
ed on
t
he
feat
ur
es
this p
r
ocess
le
arn
t
he
net
wor
k
to
be
able t
o detec
t
if
the
te
xtil
e
i
m
age
ei
ther
def
ect
ive
te
xtil
e
or
no
n
-
def
ect
i
ve.
The
detect
ion
re
su
lt
s
sho
w
that
the
ou
tp
ut
is
"0"
if
the
te
xtil
e
is
fr
ee
of
de
fect
s
or
the
ou
t
pu
t
is
"1"
if
the
te
xtil
e
i
s
def
ect
ive.
Fig
ur
e
5
de
scribe
this
process
.
The
trai
ning
m
at
rix
has
bee
n
config
ur
e
d
f
rom
42
0
to
21
0
te
xtil
e
i
m
ages
without
de
fect
and
the
oth
e
r
210
has
a d
e
fect an
d
a
targ
et
m
at
rix
co
nf
i
gurati
on tha
t corres
ponds t
o
it
.
Figure
5.
Elm
a
n Neural
Netw
or
k
for defects
detect
ion
n
i
i
i
W
X
f
y
1
n
i
i
i
W
X
SU
M
1
0
0
0
1
1
1
n
i
i
i
n
i
i
i
W
X
if
W
X
if
x
f
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
A fuzzy
system
for
detect
ion
and cl
as
sif
ic
atio
n of tex
ti
le
d
ef
ect
s to
e
ns
ure
th
e
…
(
Iman
Subhi M
ohamme
d
)
4283
6.3.
F
ab
ri
c
de
fects cl
as
sific
ati
on st
age
At
this
sta
ge,
on
ly
210
defe
ct
ive
te
xtil
e
im
ages
proce
ss
ed
a
nd
cl
assifi
ed
by
the
Ba
c
k
P
ropa
gatio
n
Neural
Netw
ork
int
o four
te
e
n defects a
s
discuss
e
d belo
w:
6.3.1. Isol
at
e
t
he de
fect are
a (re
gio
n
of in
t
erest
ROI)
Af
te
r
rea
ding
and
scal
in
g
th
e
de
fecti
ve
t
e
xtil
e
i
m
age.
T
his
im
ag
e
is
treat
ed
with
G
abor
filt
er
.
The
n
the
res
ultant
gray
im
age
is
bin
a
rized
w
her
e
"0"
i
nd
ic
a
te
s
"black"
pixe
l
and
"1"
in
dicat
es
"wh
it
e"
pix
e
l
via
fixe
d
t
hr
e
sh
ol
d
[
25
]
.
T
he
de
fect
are
a
then
el
i
m
inate
d
and
disti
nguis
hed
this
process
de
no
te
s
the
autom
at
ic
cutti
ng
al
go
rithm
of
the
de
fect
area.
The
ou
tc
om
e
of
these
processes
is
a
bin
ary
im
age
with
a
descr
i
bed (R
OI).
6.3.2
.
Fea
tu
re
s
ex
tra
c
tio
n
and de
fects cl
a
ssific
at
i
on
The
total
siz
e
of
the
def
ect
a
rea
is
cal
culat
ed
an
d
co
ns
id
ered
as
a
ge
om
et
ric
featur
e
of
the
im
age.
The
pe
rim
et
er
of
the
de
fect
area
is
then
determ
ined
aft
er
Ca
nny
O
pe
r
at
ors
are
a
ppli
ed
to
the
im
age.
The
de
fect
im
age
is
segm
ented
to
ni
ne
ti
le
s,
the
n
t
heir
featur
e
s
wer
e
extracte
d
re
pre
sented
by
Me
an
a
nd
Sk
e
wn
es
s
is
cal
culat
ed
as
in
(4
)
[26].
T
he
char
act
erist
ic
s
of
the
def
ect
i
m
ages
are
arra
ng
e
d
in
a
tw
o
-
row
m
at
rix
wh
ose
nu
m
ber
of
row
s
rep
re
sent
the
nu
m
ber
of
im
ages
an
d
the
num
ber
of
col
um
ns
rep
resen
ti
ng
the
nu
m
ber
of
at
tr
ibu
te
s.
Th
e
tot
al
featu
res
is
(
2x9=
18)
a
nd
one
for
Ar
ea
s
o
the
t
otal
feat
ur
es
to
each
de
fect
te
xtil
e i
m
age is equal 1
9. T
his m
a
trix is e
nter
ed
int
o
the
neural net
work f
or cla
ssific
at
ion
of d
e
fect t
ypes.
(4)
The
def
ect
s
a
r
e
then
cl
assifi
ed
into
14
cat
egories
usi
ng
the
Ba
ck
P
rop
agati
on
Ne
ur
al
Netw
o
r
k
that
it
i
s
trai
ned an
d
t
he
n
te
ste
d.
6.3.3
.
Neur
al
network
t
r
ain
ing
The
net
work
is
trai
ned
on
the
char
act
erist
ic
vecto
r
of
th
e
de
fecti
ve
i
m
ages.
The
pu
rpose
of
le
arn
i
ng
the
net
work
is
that
to
e
nfor
c
e
the
net
wor
k
to
incre
a
se
it
s
acce
ssibil
it
y
to
cl
assify
def
ec
ts.
Each
i
nput
vecto
r
consi
sts
of
19
values
represe
nting
t
he
m
ean
and
t
he
s
kewn
ess
of
each
cel
l
in
ad
diti
on
t
o
the
siz
e
of
the
def
ect
area. T
hese
f
ea
tures
a
re
norm
al
iz
ed
to t
he high
e
st val
ue of
t
he nine ti
le
s.
6.3.4
.
Neur
al
network
for
d
efect
cl
as
sific
ati
on
Fo
ll
ow
the
sa
m
e
ste
ps
as
the
init
ia
l
pr
ocess
and
featu
re
s
e
xt
racti
on
to
obta
in
the
cha
racte
risti
c
vecto
r
of
the
in
pu
t
im
age.
Lat
er,
this
vecto
r
was
entered
int
o
the
ne
ur
al
net
work
to
ide
ntify
the
def
ect
ty
pe
an
d
r
et
urn
to
w
hi
ch
cl
ass
it
belongs.
Durin
g
the
trai
nin
g
process
,
par
a
m
et
er
and
optim
al
weigh
ts
storing
on
trai
ni
ng
phase.
The
n
thes
e
values
we
re
resto
red
f
or
te
sti
ng
phase
to
te
st
the
i
m
ag
es
wh
ic
h
non
-
trai
ne
d
on the
netw
ork
.
7.
RESU
LT
S
AND DI
SCUS
S
ION
In
order
to
get
the
best
perform
ance
of
our
syst
e
m
m
any
t
ries
wer
e
us
ed
with
m
any
NN
.
A
num
ber
of
im
ages
wer
e
sel
ect
ed
fo
r
trai
ning
with
m
ini
m
u
m
err
or
a
nd
to
reac
h
the
op
ti
m
a
l
weight
.
The
sam
e
i
mages
al
so
us
e
d
f
or
t
est
ing
phase
to
check
the
po
wer
of
the
ne
ural
netw
ork
us
ed.
T
he
res
ults
gaine
d
co
rr
es
ponds
to
each
neural
ne
twork
c
onfi
gure
d
the
po
we
rful
an
d
a
ppr
opriat
e
net
wor
k
ch
oice.
A
nother
s
am
ple
fr
om
the
i
m
age
databas
e
that
was
no
t
us
ed
in
the
tr
ai
nin
g
phase
wer
e
us
ed
for
the
te
sti
ng
ph
ase.
To
m
easur
e
the
perform
ance
of
our
syst
em
t
wo
qu
al
it
y
m
e
asur
e
s
nam
el
y,
true
value
an
d
error
value
wer
e
use
d
the
y
are
def
i
ned as
(5) a
nd (6) res
pect
ively
[
27]
.
TR_
In ord
e
r V =
A
_V (
i)/E
_V (
i)
*1
00%
(5)
ER_V =
(10
0
-
TR_V)
*100%
(6)
Wh
e
re
A
_V r
e
pr
ese
nt the
nu
m
ber
o
f
c
orrec
tl
y cl
assifi
ed
im
age,
E
_V r
e
presents
the e
ntire
database im
age.
3
1
3
1
N
i
i
X
X
N
s
k
e
w
n
e
s
s
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
:
4
2
7
7
-
4
2
8
6
4284
7.1.
El
man
ne
ural ne
twor
k
This
netw
ork
wer
e
us
e
d
f
or
de
fect
de
te
ct
ion
it
was
trai
ned
on
42
0
i
m
ages,
incl
ud
i
ng
21
0
non
-
def
ect
ive
te
xtil
e
i
m
age
and
anthe
r
210
te
xti
le
def
ect
ive
i
m
age
be
fore.
A
new
sam
ple
of
140
i
m
ages
from
the
colle
ct
ed
visu
a
l
database
non
-
trai
ne
d
be
fore
wer
e
form
ed
t
o
be
trai
ne
d
in
the
El
m
an
Neu
ral
Netw
ork
.
Thes
e
140
im
ages
include
70
of
non
-
def
ect
ive
t
extil
e
i
m
age
and
70
de
fecti
ve
te
xtil
e
i
m
a
ge.
T
his
sam
ple
wa
s
entere
d
into
E
LMAN
netw
ork
f
or
defec
t
det
ect
ion
.
T
he
re
s
ults
of
t
he
trai
ned
420
im
age
and
non
-
trai
ne
d
17
0
i
m
ages
wer
e
de
scribe
d
as
show
n
in
T
a
ble
1.
T
he
res
ults
for
the
trai
ne
d
210
i
m
ages
are
eq
ual
to
100%
f
or
def
ect
ive
a
nd
non
-
de
fecti
ve
i
m
age.
W
hile
the
rate
f
or
th
e
70
def
ect
ive
i
m
age
app
r
ox
i
m
at
e
to
98.57
%
an
d
100%
f
or
t
he
non
-
de
fecti
ve
70
im
ages.
In
add
it
io
n
,
t
he
t
otal
res
ults
are
ab
ou
t
100%
of
the
210
im
ag
e
an
d
99.28%
of
the
140
im
ages.
These
res
ults
at
al
l
ref
le
ct
the
s
up
e
rio
r
res
ults
and
the
a
ppr
opriat
e
sel
ect
ion
of
the
neural
netw
ork
.
Table
1.
Re
s
ults o
f
trai
ned an
d non
-
trai
ne
d
i
m
ages on
t
he
E
l
m
an
NN
70
n
o
t tr
ain
ed
I
m
a
g
e
2
1
0
tr
ain
ed
I
m
ag
e
Textile
t
y
p
e
9
8
.57
1
4
%
100%
Def
ect textile
100%
100%
No
n
def
ect textile
9
9
.28
5
7
%
100%
Total p
erce
n
tag
e
6
3
.07
8
5
0
:0
0
:4
6
Ti
m
e
ellips
e in sec
o
n
d
7.2
.
BPN
N
and
FBP
NN
neu
ral ne
twork
t
e
sting st
age
Wer
e
ap
plied
on
feat
ur
e
vect
or
duri
ng
the
te
st
phase
t
o
cl
assify
the
de
fe
ct
s
ty
pe
of
t
he
70
im
ages.
At
the
te
st
ph
ase
the
BPNN
trai
n
f
or
fi
ve
im
ages
per
14
def
ect
s
each
th
e
accuracy
of
t
he
defec
t
cl
assifi
cat
ion
rati
o
was
91.
4286%
with
er
r
or
rati
o
8.5
71
4%.
In
a
ddit
ion
,
FBP
N
N
tra
in
the
sam
e
i
m
ages
as
BPNN
it
s
cl
assifi
cat
i
on
r
at
io
was
a
bout
97.14
28%
wit
h
er
r
or
rate
4.2
857%
overall
t
his
im
age
set
.
The
resu
lt
s
s
ho
wn
in
T
able
2
co
nfi
gure
the
po
we
r
of
our
pr
opos
e
d
ne
ur
al
ne
twork
denote
d
by
FBP
NN
ov
e
r
BPN
N
f
or
bo
t
h
resu
lt
ant
pa
ra
m
et
ers.
Tables
2.
Res
ul
ts for test
stag
e b
y B
P
N
N
a
nd FBP
N
N
Accurac
y
of
FBPN
N
Accurac
y
of
BPN
N
Def
ect T
y
p
e
Er
ror
r
atio
Accurac
y
Er
ror
r
atio
Accurac
y
0%
100%
0%
100%
Def
ect1
(Brok
en
end
s)
0%
100%
0%
100%
Def
ect2
(
Bro
k
en
pattern
)
0%
100%
0%
100%
Def
ect3
(
co
lo
red f
leck
s)
20%
100%
40%
80%
Def
ect4
(
Cu
t
&
tea
r)
0%
100%
0%
100%
Def
ect5
(Dou
b
le end
)
0%
100%
0%
100%
Def
ect6
(f
lo
at)
0%
100%
0%
100%
Def
ect7
(ho
le)
0%
100%
20%
80%
Def
ect8
(Kno
ts)
20%
80%
40%
60%
Def
ect9
((
Missin
g
p
ick
s)
0%
100%
0%
100%
Def
ect1
0
((
Reed
M)
0%
100%
0%
100%
Def
ect1
1
(slu
b
)
20%
80%
20%
80%
Def
ect1
2
(stain
)
0%
100%
0%
100%
Def
ect1
3
(Untri
m
m
)
0%
100%
0%
100%
Def
ect1
4
(W
ef
t Ba
r)
4
.28
5
7
%
9
7
.14
2
8
%
8
.57
1
4
%
9
1
.42
8
6
%
Percentag
e
ratio
7.3.
BPN
N
and
FBP
NN
neu
ral ne
twork
t
r
aining
sta
ge
A
t t
he
trai
ni
ng
sta
ge
70 im
age f
eat
ures v
ect
or w
e
re traine
d
t
heir
re
su
lt
s is s
how
n
in
T
a
ble
3.
F
ourtee
n
def
ect
s
cl
assifi
cat
ion
acc
ur
ac
y
was
com
pu
t
ed
at
BPN
N
a
nd
FBPN
N
th
ei
r
res
ults
are
99.52
38%
an
d
100%
resp
ect
ively
.
T
his
rati
o
was
a
ff
ect
ed
by
defec
t12
,
tha
t
it
is
accuracy
rati
o
appr
ox
im
at
e
s
to
93.
4%
for
BPN
N
bu
t
100%
fo
r
F
BPNN.
Tables
3.
Res
ul
ts for traini
ng
sta
ge by B
PNN a
nd FBP
NN
Def
ect T
y
p
e
Accurac
y
/
BPNN
Accurac
y
/
FBPNN
Def
ect1
(Brok
en
end
s)
100%
100%
Def
ect2
(Brok
en
pattern
)
100%
100%
Def
ect3
(colo
red f
leck
s)
100%
100%
Def
ect4
(
Cu
t & tea
r)
100%
100%
Def
ect5
(Dou
b
le end
)
100%
100%
Def
ect6
(f
lo
at)
100%
100%
Def
ect7
(ho
le)
100%
100%
Def
ect8
(
Kn
o
ts)
100%
100%
Def
ect T
y
p
e
Accurac
y
/
BPNN
Accurac
y
/
FBPNN
Def
ect9
((
Missin
g
p
ick
s)
100%
100%
Def
ect1
0
(
(
Reed
M
arks
)
100%
100%
Def
ect1
1
(slu
b
)
100%
100%
Def
ect1
2
(stain
)
9
3
.33
3
%
100%
Def
ect1
3
(Untri
m
m
ed
Loo
se th
reads
)
100%
100%
Def
ect1
4
(W
ef
t Ba
r)
100%
100%
Percentag
e
ratio
9
9
.52
3
8
%
100%
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
A fuzzy
system
for
detect
ion
and cl
as
sif
ic
atio
n of tex
ti
le
d
ef
ect
s to
e
ns
ure
th
e
…
(
Iman
Subhi M
ohamme
d
)
4285
8.
PERFO
R
MANC
E E
F
FIC
I
ENC
Y
O
F T
HE PR
OPO
S
ED W
ORK
In
or
der
to
m
e
asur
e
the
e
nh
a
ncem
ent
of
our
pro
po
se
d
w
ork
.
W
e
com
pare
it
with
oth
er
works
in
the
sam
e
fiel
d.
Desp
it
e
the
diff
e
r
ence
in
vis
ual
database
siz
e
and
the
te
ch
ni
qu
e
s
use
d
bet
ween
our
syst
e
m
an
d
oth
e
r
relat
ed
work.
Ta
ble
4
sh
ows
t
he
sup
erior
it
y
of
o
ur
work
t
hat
reac
hed
t
o
97.
1428
%
accuracy
rate
ove
r
oth
e
rs wor
k w
hich
it
is
ra
ng
e
d betwee
n
m
ini
m
u
m
v
al
ue 80%
and a
bout
96%
of m
axi
m
u
m
v
al
ue
.
Table
4.
C
om
par
iso
n wit
h rela
te
d
w
ork
Sy
ste
m
s
Metho
d
Visu
al Database
Accurac
y
Rate
ov
er
all
DB
An
an
d
H.
et
a
l.
(20
1
2
)
[
1
]
GLCM+
Pr
o
b
ab
ilit
y
Neu
ral
N
etwo
rk
1
5
0
textile i
m
ag
e
3
0
textile i
m
ag
e f
r
ee of
def
ects
1
2
0
textile i
m
ag
e
with
4 ty
p
e of
def
ects
9
1
.1%
Ali
Jav
ed
et
al
.
(20
1
3
)
[
8
]
Reg
u
lar
Ban
d
7
8
textile i
m
ag
e
3
9
textile i
m
ag
e f
r
ee of
def
ects
3
9
textile i
m
ag
e
w
ith
d
ef
ects
96%
P.Y.
Ku
m
b
h
ar
et a
l.
(20
1
6
)
[
1
1
]
Gen
etic algo
rith
m
+
SVM
--
-
--
-
--
90%
Su
n
il Ban
g
are
et.
a
l
(20
1
7
)
[
1
3
]
I
m
ag
e
Proces
sin
g
Techn
iq
u
e
5
2
textile i
m
ag
e
2
6
textile i
m
ag
e f
r
ee of
def
ects
2
6
textile i
m
ag
e
w
ith
def
ects
9
6
.15
%
Sh
u
an
g
M
ei et.
a
l.
(20
1
8
)
[
1
4
]
Encry
p
tio
n
+ Gaus
sian
Py
ra
m
id
s
TE
L
DA datab
ase
80%
Prop
o
sed
App
roach
(20
1
8
)
Gab
o
r
f
ilter+
Fu
zz
y
Back
Pr
o
p
ag
atio
n
+
El
m
an
neu
ral
netw
o
rk
5
6
0
textile i
m
ag
e
2
8
0
textile i
m
ag
e f
ree
o
f
def
ects
2
2
0
i
m
ag
e
with
14
ty
p
e of
def
ects
9
7
.14
2
8
%
9.
CONCL
US
I
O
N
The
pro
posed
work
s
how
t
ha
t
the
us
e
of
spe
ct
ral
filt
ers
s
uc
h
as
a
Gabo
r
filt
er
has
great
ly
helpe
d
to
ob
ta
in
good
sta
ti
sti
cal
feature
s
fo
r
the
s
uc
cess
of
the
de
te
ct
ion
proces
s.
In
ad
diti
on,
El
m
an
and
BPN
N
netw
orks
pro
ve
d
thei
r
ef
fici
ency
th
rou
gh
th
ei
r
res
ul
ta
nt
r
a
ti
os
.
T
hey
s
how
99.28
57
%
a
t
the
de
fect
detect
ion
sta
ge
an
d
91.
4286%
i
n
the
de
fects
cl
assifi
ca
ti
on
sta
ge
.
As
well
as
the
m
erg
e
betwee
n
f
uz
zy
log
ic
wit
h
a
Ba
ck
Pr
opa
gatio
n
Neural
Net
work
t
hat
pro
duce
the
Fu
zzy
Ba
ck
Pro
pa
gation
Ne
ur
al
Netw
ork
F
BPNN.
Th
e
outc
om
e fr
om
FBPNN in
creased
the
pro
portio
n of
dete
ct
ion
si
gn
ific
a
ntly
to
97.
1428
%.
REFERE
NCE
S
[1]
Anand
H.
Kulka
rni
and
She
et
a
l
B.
Pati
l
,
"A
utomate
d
Ga
rm
ent
Id
ent
ifica
ti
on
and
Defe
ct
Detect
io
n
Model
Based
on
Te
xtur
e
Fea
tures
and
PN
N
,
"
In
te
rnational
Jour
nal
of
Latest
Tr
ends
in
Engi
ne
e
ring
and
Techn
ology
(
IJLTET)
,
v
ol.
1
,
no
.
2
,
201
2
.
[2]
R.
Peng
Pai,
"
Autom
at
ed
Def
ec
t
Det
ection
fo
r
Te
xtile
Fabri
c
s
Us
ing
Gabor
W
ave
le
t
Ne
tworks
,
"
Diss
ert
at
io
n,
The
HK
U Schol
ars,
Th
e
Univ
ers
ity
of
Hong Kon
g,
ht
tp:
/
/hdl
.
h
an
dle
.
n
et
/10722
/51
685.
[Inte
rn
et
]
,
2
006.
[3]
A
.
P
.
S
.
Sawhne
y
,
J.
B
.
Pric
e
an
d
Cal
amar
i
,
"A
Succe
ss
ful
W
eaving
Trial
with
a
Size
-
Fre
e
Cot
to
n
W
arp
,
"
Southe
rn
Regi
onal
R
esra
r
ch
Center,
Agri
cul
tur
al
Rese
arc
h
Servic
e
,
New
Orlea
ns,
India
n
Journal
of
F
ibre
&
Text
il
e
Re
search
,
v
ol
.
2
9,
pp
.
117
-
121
,
2004.
[4]
Ris
hav
D.,
Moni
ka
A
.
,
"Identificati
on
Of
D
efe
c
ts
In
Te
x
ti
l
e
B
ase
d
On
Stat
ist
ic
a
l
Anal
y
sis
Of
DC
T
Coeff
ic
i
ent
s
O
f
Te
xtile
Im
age
s
,
"
Aarhat
Mu
lt
id
isci
pli
nary
Int
ernati
onal Educ
a
ti
on
Re
search
Journ
al
(
AMI
ERJ)
,
2012.
[5]
Jian
y
un
Ni
,
Jing
Luo,
Za
ip
ing
Chen
and
Enze
ng
Dong,
"A
M
ult
i
Resolution
Method
for
Dete
cting
Defe
c
ts
in
Fabric
Im
age
s
Resea
rch
"
,
Jou
rnal
of
Appl
i
e
d
Sci
en
ce
s,
En
gine
ering
and
Technol
ogy
,
M
axwe
ll
Scien
tifi
c
Or
ganizati
on
,
v
ol.
5
,
no
.
5,
2013.
[6]
Micha
e
l
K.
Ng,
Henr
y
Y.
T
.
Ngan,
Xiaomin
g
Yuan,
and
W
enxi
ng
Zha
ng
,
"P
at
te
rne
d
Fabr
i
c
Inspec
t
ion
an
d
Visuali
z
at
ion
b
y
the
Method
of
I
m
age
Dec
om
positi
on
,
"
I
EEE
Tr
ans
act
ion
o
n
Au
to
mation
Scienc
e
a
nd
Engi
ne
ering
,
v
ol.
1
1
,
n
o.
3
,
pp
.
943
-
947
,
2014
.
[7]
Ali
Jave
d,
Mirza
Ahs
an
Ulla
h,
and
Aziz
-
ur
-
Reh
m
an,
"Com
par
at
ive
Anal
y
s
is
of
Diffe
ren
t
Fabri
c
Defe
ct
s
Detect
io
n
Te
chn
ique
s
,
"
Im
age,
Gr
aphi
cs
a
nd
Signal P
roc
e
ss
ing
Journal
, v
ol.
1
,
pp
.
40
-
45
,
2013.
[8]
Ali
Rebhi,
Iss
a
m
Benmham
m
e
d,
Sabeur
Abid
,
and
Farha
t
F
nai
e
ch,
"F
abr
ic
Defe
c
t
De
te
c
tion
Us
ing
Local
Hom
ogene
ity
A
naly
s
is a
nd
Neur
al
Ne
twork,
"
Hi
ndawi
Pub
li
shin
g
Corpor
ati
on,
J
ournal
of Phot
o
nic
s
,
vo
l.
9,
201
5.
[9]
Male
k
Abdel
Sal
am,
"O
nli
ne
Fabric
Inspec
t
ion
b
y
Im
age
Proce
ss
i
ng
Te
chnol
og
y
,
"
Ph.D.
The
sis su
bm
it
te
d
to
Haute
Alsac
e
Univ
ersity
,
2013
.
[10]
M Hanmandlu
,
Sujat
a
Dash
And
D K Choudhur
,
"F
abr
ic
Im
age Def
ect
De
te
c
ti
on
b
y
using
GLCM
and
ROS
ET
TA
,
"
HIT
Tr
ansacti
ons
,
v
ol. 3,
n
o.
9,
2
008
[11]
P.
Y.
Kum
bhar
,
Te
j
aswini
Mathp
at
i
,
Rohini
Kam
ara
ddi
and
Nam
rat
a
Ks
hirsaga
r
,
"Te
xtile
Fabri
c
Defe
c
ts
Dete
c
ti
o
n
and
Sorting
Us
i
ng
Im
age
Proce
s
sing
,
"
Inte
rnat
io
nal
Journal
For
Re
search
In
E
merging
Scienc
e
And
Techno
log
y,
v
ol.
3
,
no
.
3
,
201
6.
[12]
Soo
Chang
Ki
m
and
Ta
e
JinKang
,
"Te
xture
Cl
assific
a
ti
on
and
Segm
ent
at
ion
Us
ing
W
ave
le
t
Packe
t
Fram
e
a
n
d
Gauss
ia
n
Mixtur
e
Model
,
"
Pa
ttern R
e
c
ognition
,
v
ol
.
40
,
no
.
4
,
200
7.
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
:
4
2
7
7
-
4
2
8
6
4286
[13]
Sunil
L
.
Bang
ar
e,
N
.
B.
Dhawa
s,
V.
S.
Ta
w
are,
S.
K
.
Digh
e,
P.
S.
Bagmar
e,
"Im
ple
m
ent
at
ion
of
Fabri
c
Fau
lt
Dete
c
ti
on
S
y
ste
m
Us
ing
Im
age
Proce
ss
ing
,
"
In
te
rnational
Jour
nal
of
Re
s
earc
h
in
Advent
Te
c
hnology
,
v
ol
.
5
,
n
o.
6
,
2017
.
[14]
Shuang
Mei,
Yudan
W
ang,
and
Guojun
We
nm
,
"A
utomati
c
Fabric
Defe
c
t
Dete
c
ti
on
wit
h
a
Multi
-
Scale
Convolut
ional
D
enoi
sing Aut
o
en
code
r
N
et
work
Model
,"
S
ensors
,
2018
.
[15]
Aja
y
Kum
ar
an
d
Hauz
Khas,
"Com
pute
r
Visi
on
-
base
d
Fabr
ic
Defe
ct
Det
ec
t
i
on:
A
Surve
y
,
"
Depa
r
tment
of
El
e
ct
ri
ca
l
Engi
n
ee
ring
Indi
an
Ins
ti
tute
of
T
ec
hnol
og
y
,
2004
.
[16]
Maha
ja
n
P.
M.
,
J.
T.
Maha
ja
n
and
Patil
P.
M,
"A
Revi
ew
of
Autom
at
ic
Fabr
i
c
Defe
ct
De
tection
Techni
ques
,
"
Adv
anc
es
in
Co
mputati
onal
Re
s
earc
h
,
v
o
l.
1
,
no
.
2,
pp
.
18
-
29
,
20
04
.
[17]
Pri
y
ank
a
M.
Sh
anbha
g,
Manish
P.
Deshm
ukh
an
d
Shekhar
R.
Su
ral
ka
,
"O
ver
v
iew
:
Methods
Of
Autom
at
ic
Fabr
i
c
Defe
ct Detection
,
"
Global
Journa
l
Of
Engi
ne
ering
,
Design
&
Technol
ogy
,
v
ol
.
1
,
n
o.
2
,
pp
.
42
-
46
,
2012.
[18]
Tom Mitc
he
ll,
Mac
hine Learni
ng
,
publ
ished
b
y
McGraw Hill,
1
997.
[19]
Zhi
qia
ng
Zh
ang,
Shangce
Gao,
Gang
Yang,
Fangji
a
Li
and
Zheng
Ta
ng,
"A
n
A
lgori
the
m
of
Supervised
Le
arn
in
g
for
El
m
an
Neur
al
Network
,
"
In
te
rnational
Journal
of
Innov
at
i
ve
Computing
,
Information
and
Control
ICIC,
Inte
rna
ti
ona
l,
v
o
l.
5
,
n
o
.
10(A)
,
p
p.
299
7
–
3011
,
2
009.
[20]
MathWorks
Inc
.
,
"F
uzzy
Logi
c
T
oolbox™
",
Us
er’
s Guide
R2014
a
,
2014
.
[21]
S.
Ariva
zha
g
an,
L.
Gane
san
,
a
nd
S.
Bama,
"F
aul
t
Segm
ent
a
ti
on
in
Fabri
c
Im
age
s
Us
ing
Gabor
W
ave
let
Tra
nsform
,
"
Jou
rnal
of
Ma
chi
n
e Vi
sion a
nd
Appli
cat
ions,
v
ol
.
16
,
n
o.
6
,
pp
.
356
–
3
63
,
2006
.
[22]
Dengsheng
Zhang,
A
y
lwin
W
ong,
Maria
Indr
a
wan,
Guojun
L
u,
"Content
-
B
ase
d
Im
age
Ret
r
ieval
Us
ing
Gabo
r
Te
xtur
e
Fe
at
ure
s
,
"
Proc
ee
ding
of
Fi
rs
t
IE
EE
Pacif
ic
-
R
im
Confe
ren
ce
on
Mul
ti
mead
ia
,
pp
.
392
-
395
,
2000.
[23]
Om
ar
Sulta
n
Al
-
Kadi,
"A
gabor
Filt
er
Te
x
ture
Anal
y
s
is
A
f
or
Histopat
holo
gic
a
l
Bra
in
Tu
m
our
Sub
T
y
p
e
Discriminat
ion
,
"
ISESCO
JST
jo
urnal
,
v
ol
.
12,
no
.
22
,
2017
.
[24]
Aja
y
Kum
ar,
"N
eur
al
Network
Based
Dete
ct
io
n
of
Loc
al
Te
xt
il
e
Defe
ct
s
,
"
The
Journal
Of
Pat
te
rn
Re
cogn
it
io
n,
The
Pattern
R
ecogniti
on
S
ociet
y
,
vol.
36,
pp.
164
5
–
1659,
2003
.
[25]
Ma
y
a
R.
Gupt
a
,
Nath
aniel
P.
Jac
obson
and
E
ric
K.
G
arc
i
a,
"
OCR
Bina
rization
and
Im
age
P
re
-
Proce
ss
ing
fo
r
Sear
chi
ng
Histor
ic
a
l
Docum
ent
s,
"
The
Journal
Of
The
Pa
tt
ern
Rec
ognit
ion
,
v
ol
.
40
,
pp
.
389
–
397,
20
07.
[26]
David
P.
Doane
and
Lori
E.
Seward,
"M
ea
suring
S
kewne
ss:
A
Forgotte
n
Stat
isti
c
?
,
"
Jour
nal
of
Stat
isti
cs
Educ
ati
on
,
O
akland
Univer
si
t
y
,
v
ol.
19
,
n
o.
2
,
20
11.
[27]
S.
G.
HO
GG
A
R,
Math
emati
cs
of
Digi
tal
Im
ages
Creati
on
,
Compr
ession,
Re
storation
,
R
e
cogni
ti
on
,
USA
:
Cambridge
Univ
ersit
y
Press
,
Ne
w York,
2006
Evaluation Warning : The document was created with Spire.PDF for Python.