Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
e
eri
ng
(IJ
E
C
E)
Vo
l.
10
, N
o.
4
,
Aug
us
t
2020
, p
p.
4390
~
4399
IS
S
N:
20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v
10
i
4
.
pp4390
-
43
99
4390
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Textur
e classific
ation o
f fa
bric d
efects
usin
g ma
chine l
earnin
g
Ya
s
sine B
en
S
alem,
M
oham
ed Naceur
Ab
delkri
m
Engi
ne
eri
ng
S
ch
ool
of
Gab
es,
Univer
sit
y
of
Gab
es,
Tun
isia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
N
ov
25
, 201
9
Re
vised
Feb 2
9
,
2020
Ac
cepte
d
Ma
r
9
, 2
020
I
n
t
h
i
s
p
a
p
e
r
,
a
n
o
v
e
l
a
l
g
o
r
i
t
h
m
f
o
r
a
u
t
o
m
a
t
i
c
f
a
b
r
i
c
d
e
f
e
c
t
c
l
a
s
s
i
f
i
c
a
t
i
o
n
w
a
s
p
r
o
p
o
s
e
d
,
b
a
s
e
d
o
n
t
h
e
c
o
m
b
i
n
a
t
i
o
n
o
f
a
t
e
x
t
u
r
e
a
n
a
l
y
s
i
s
m
e
t
h
o
d
a
n
d
a
s
u
p
p
o
r
t
v
e
c
t
o
r
m
a
c
h
i
n
e
S
V
M
.
T
h
r
e
e
t
e
x
t
u
r
e
m
e
t
h
o
d
s
w
e
r
e
u
s
e
d
a
n
d
c
o
m
p
a
r
e
d
,
G
L
C
M
,
L
B
P
,
a
n
d
L
P
Q
.
T
h
e
y
w
e
r
e
c
o
m
b
i
n
e
d
w
i
t
h
S
V
M
’
s
c
l
a
s
s
i
f
i
e
r
.
T
h
e
s
y
s
t
e
m
h
a
s
b
e
e
n
t
e
s
t
e
d
u
s
i
n
g
T
I
L
D
A
d
a
t
a
b
a
s
e
.
A
c
o
m
p
a
r
a
t
i
v
e
s
t
u
d
y
o
f
t
h
e
p
e
r
f
o
r
m
a
n
c
e
a
n
d
t
h
e
r
u
n
n
i
n
g
t
i
m
e
o
f
t
h
e
t
h
r
e
e
m
e
t
h
o
d
s
w
a
s
c
a
r
r
i
e
d
o
u
t
.
T
h
e
o
b
t
a
i
n
e
d
r
e
s
u
l
t
s
a
r
e
i
n
t
e
r
e
s
t
i
n
g
a
n
d
s
h
o
w
t
h
a
t
L
B
P
i
s
t
h
e
b
e
s
t
m
e
t
h
o
d
f
o
r
r
e
c
o
g
n
i
t
i
o
n
a
n
d
c
l
a
s
s
i
f
i
c
a
t
i
o
n
a
n
d
i
t
p
r
o
v
e
s
t
h
a
t
t
h
e
S
V
M
i
s
a
s
u
i
t
a
b
l
e
c
l
a
s
s
i
f
i
e
r
f
o
r
s
u
c
h
p
r
o
b
l
e
m
s
.
W
e
d
e
m
o
n
s
t
r
a
t
e
t
h
a
t
s
o
m
e
d
e
f
e
c
t
s
a
r
e
e
a
s
i
e
r
t
o
c
l
a
s
s
i
f
y
t
h
a
n
o
t
h
e
r
s
.
Ke
yw
or
d
s
:
Im
age p
r
ocessi
ng
Ma
chine
le
a
rn
i
ng
Patt
ern
recog
ni
ti
on
Textu
re a
naly
sis
Wove
n fab
ric
def
ect
s
Copyright
©
202
0
Instit
ute of
Ad
v
ance
d
Engi
ne
eri
ng
and
Sc
ie
n
ce
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Yassine
Ben
S
al
e
m
,
Nati
on
al
En
gine
erin
g
S
c
hool
of G
a
bes,
Un
i
ver
sit
y o
f Gabe
s,
Road
O
m
er Ib
n
El
kh
at
ta
b, Z
r
ig, Uni
ver
sit
y
of G
a
bes, T
un
i
sia
County
60
29, T
unisi
a, T
N.
Em
a
il
:
ben
sal
em
y73@
ya
ho
o.fr
1.
INTROD
U
CTION
In
orde
r
to
m
ai
ntain
the
hig
h
qual
it
y
stan
da
r
ds
est
abli
sh
e
d
f
or
t
he
cl
oth
in
g
in
dus
try
,
Texti
le
m
anu
fact
ur
e
rs
h
ave
to
m
on
it
or
the
qu
al
it
y
of
their
products.
Ther
ef
ore,
te
xt
il
e
qu
al
it
y
cont
ro
l
is
a
key
factor
for
the
i
ncr
eas
e
of
c
om
petit
i
ven
e
ss
of
thei
r
com
pan
ie
s.
T
extil
es
fau
lt
s
ha
ve
tra
diti
on
al
ly
been
detect
ed
by
hu
m
an
vis
ual
insp
ect
io
n;
sk
i
ll
ed
wor
ker
c
ou
l
d
at
m
o
st
detect
and
dis
ti
ng
uis
h
70
%
of
al
l
fa
br
ic
de
fects.
The
n,
s
om
e
def
ect
s
are
cl
assi
fied
as
re
par
a
bl
e
becau
se
the
y
can
be
proce
ssed
t
hroug
h
s
om
e
us
ual
ope
rati
ons
by
w
orkers
an
d
will
le
ave
no
trace
after
t
hat.
I
n
the
oth
e
r
ha
nd,
th
os
e
ir
re
par
a
ble
de
fects
hav
e
a
great
im
pact
on
t
he
qu
al
it
y
of
cl
oth
wh
ic
h
shou
l
d
be
el
im
inate
d
in
pr
oductio
n
a
nd
la
te
r
pic
ked
out
as
m
any
as
possible.
That’s
w
hy
we
sho
uld
w
ork
on
an
a
uto
m
at
i
cvisual
i
nspect
ion
a
nd
c
on
cei
ve
a
syst
em
to
su
bs
ti
tute
t
he
rol
e
of
te
xtil
e w
orke
rs
and s
urpas
s
th
ei
r
abili
ty
in
s
pe
ed
a
nd acc
ur
a
cy
.
In
t
he
past,
a
lot
of
wor
k
has
bee
n
done
focuse
d
on
autom
at
ic
fab
r
ic
def
ect
dete
ct
ion
[1
-
4]
.
Fabri
c
de
fect
c
la
ssific
at
ion
ha
s
not
ye
t
at
tract
ed
a
wide
re
search
interest
.
Also,
it
is
ve
r
y
rar
e
t
hat
we
hav
e
base
d
on
th
e
t
extu
re
of
ti
ssue
in
t
he
a
uto
m
at
ic
visu
al
i
ns
pe
ct
ion
visu
al
iz
at
ion
des
pite
the
a
pp
e
ara
nce
of
ne
w
te
xtu
re m
et
hods i
n
rec
ent yea
r
s su
c
h
as
Feat
ures E
xtracted
fro
m
Co
-
occ
urr
ence Mat
rix
GLC
M [5
-
7],
or
Local
Bi
nar
y patt
er
n LB
P [8,
9]
or local
phase
qua
ntiza
ti
on
L
PQ
[10].
The
ob
j
ect
cl
assifi
cat
ion
pro
blem
is
div
ide
d
into
t
wo
cat
e
gories
:
featur
e
ext
racti
on
a
nd
cl
assifi
cat
ion
.
Fo
r
the
cl
assifi
cat
ion
,
m
os
t
re
searche
rs
us
e
d
the
ne
ur
al
net
work
as
a
cl
ass
ifie
r,
bu
t
it
is
pro
ve
d
that
it
has
sh
own
weak
perf
or
m
ances
an
d
a
m
on
g
al
l,
it
has
a
too
-
l
ong
r
unning
ti
m
e
[1
1].
All
the
res
earc
h
works
are
ba
s
ed
on
thr
ee
a
ppr
oach
es:
spe
ct
ral,
sta
ti
sti
c
al
,
an
d
m
od
el
-
base
d.
T
he
st
at
ist
ic
al
app
r
oa
ches,
te
xtu
re
featu
re
s
extracte
d
from
co
-
occ
urre
nce
m
at
rix,
m
ean
an
d
sta
nd
ard
dev
ia
ti
ons
of
s
ub
blo
c
ks
[12],
autoc
orrelat
ion
of
im
ages
or
su
b
-
im
ages
[13],
an
d
Ka
r
hum
enLo
e
ve
tra
nsfo
rm
hav
e
be
en
use
d
for
the
fabric
def
ect
s
detect
ion
[
14
]
.
In
[
15]
,
Bodnar
ovaus
e
the
norm
al
ized
cr
os
s
-
co
rr
el
at
ion
f
unct
ions
for
def
ect
s
detect
io
n
of
te
xtil
e
fabr
ic
s.
Ma
n
y
m
od
el
-
base
d
te
c
hn
i
qu
e
s
for
f
abr
ic
de
fect
de
te
ct
ion
we
reexi
ste
d,
for
exa
m
ple,
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
Text
ur
e cl
as
sif
ic
ation of f
abric
d
ef
ect
s
us
in
g ma
c
hin
e
lear
ni
ng
(
Ya
ssi
ne
Be
n Sa
le
m
)
4391
Coh
e
n
a
nd
al
[
16
]
us
e
d
a
Ma
rkov
ra
ndom
fiel
d
(MRF)
m
od
el
f
or
de
fect
fa
br
ic
s
inspe
ct
ion.
Ch
en
a
nd
Jai
n
[
17
]
use
d
a
structu
ral
ap
proac
h
to
de
te
ct
def
ect
s
in
im
ages
ba
sed
on
te
xtu
re.
I
n
[
18]
they
i
m
ple
m
ented
a
n
MRF
-
base
d
m
et
hod.
F
i
n
a
l
l
y
,
T
a
j
e
r
i
p
o
u
r
h
a
s
u
s
e
d
l
o
c
a
l
b
i
n
a
r
y
p
a
t
t
e
r
n
s
(
L
B
P
)
,
t
h
i
s
m
e
t
h
o
d
i
s
k
n
o
w
n
t
o
b
e
a
h
i
g
h
l
y
d
i
s
c
r
i
m
i
n
a
t
i
v
e
t
e
x
t
u
r
e
o
p
e
r
a
t
o
r
[
1
9
]
.
I
n
t
h
i
s
p
a
p
e
r
,
w
e
b
u
i
l
d
u
p
o
n
p
r
e
v
i
o
u
s
w
o
r
k
t
o
e
s
t
a
b
l
i
s
h
a
c
o
m
p
a
r
a
t
i
v
e
s
t
u
d
y
of
t
h
r
e
e
m
e
t
h
o
d
s
b
a
s
e
d
o
n
t
e
x
t
u
r
e
:
L
B
P
m
e
t
h
o
d
,
t
h
e
c
o
-
o
c
c
u
r
r
e
n
c
e
m
a
t
r
i
x
m
e
t
h
o
d
,
a
n
d
L
P
Q
m
e
t
h
o
d
.
W
e
c
h
o
o
s
e
t
h
e
S
V
M
c
l
a
s
s
i
f
i
e
r
t
o
b
e
c
o
m
b
i
n
e
d
w
i
t
h
t
e
x
t
u
r
e
m
o
d
e
l
s
;
i
t
i
s
k
n
o
w
n
t
o
b
e
s
u
i
t
a
b
l
e
i
n
t
h
i
s
c
a
s
e
[
2
0
,
2
1
]
.
2.
RESEA
R
CH MET
HO
D
In
m
any
ap
plica
ti
on
s,es
pe
ci
al
ly
in
industrial
su
r
face
in
sp
ect
ion
,te
xture
a
na
ly
sis
is
an
i
m
po
rta
nt
issue
in
im
age
pr
oc
e
ssing,
m
edical
i
m
aging
[5
]
,
con
te
nt
-
based
i
m
age
retrieval
rem
ote
sensing
[
8],
a
nd
doc
um
ent
segm
entat
ion
,
et
c.
Since
the
acqu
isi
ti
on
of
i
m
ages,
or
ie
nt
at
ion
can
be
a
pr
oble
m
.
In
the
past,
s
om
e
te
xtu
re
analy
sis
m
et
ho
ds
ha
ve
not
co
ns
ide
red
this
,
bu
t
rece
ntly
an
increasin
g
am
ount
of
at
te
nti
on
has
bee
n
gi
ven
to
inv
a
riant
te
xt
ure
a
naly
sis
tha
t’s
w
hy
m
any
m
et
ho
ds
for
r
otati
on
in
var
ia
nce
ha
ve
been
pro
posed
.
We
note
m
any
ap
pr
oac
hes
f
or
te
xtur
e
analy
sis,
w
hich
can
be
cl
assifi
ed:
sta
t
ist
ic
,
structur
a
l,
m
od
el
-
base
d
an
d
trans
form
-
based
m
et
ho
ds
F
i
gure
1
.
W
e
h
a
v
e
e
l
a
b
o
r
a
t
e
d
o
n
a
n
a
l
g
o
r
i
t
h
m
b
a
s
e
d
o
n
t
h
e
o
r
g
a
n
i
g
r
a
m
p
r
e
s
e
n
t
e
d
i
n
F
i
g
u
r
e
2
.
T
h
e
t
r
a
i
n
a
n
d
t
h
e
t
e
s
t
f
e
a
t
u
r
e
v
e
c
t
o
r
s
o
b
t
a
i
n
e
d
a
r
e
i
m
p
l
e
m
e
n
t
e
d
i
n
t
h
e
s
u
p
p
o
r
t
v
e
c
t
o
r
m
a
c
h
i
n
e
c
l
a
s
s
i
f
i
e
r
.
Figure
1.
Text
ur
e
an
al
ysi
s m
et
hods
Figure
2
.
Cl
assifi
cat
ion
alg
ori
thm
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
20
88
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
4390
-
4399
4392
2
.1
.
LB
P me
t
ho
d
Ti
m
oO
j
al
a
has
introd
uced
th
e
first
LBP
op
erator
in
2002,
wh
ic
h
is
e
valuated
us
a
rob
us
t
m
ean
of
the
te
xture
des
cripti
on
of
im
a
ges
[22].
In
L
BP,
the
pixe
ls
are
la
bels
by
t
hr
es
holdin
g
t
he
3×
3
neig
hbor
hood
of
each
pixe
l
with
the
center
va
lue,
the
res
ult
is
a
bin
ary
num
ber
.
Finall
y,
the
histo
gr
am
of
the
la
bels
will
be
enter as
a te
xture
descr
i
pto
r
, se
e Fig
u
re
3.
Figure
3. T
he
histo
gr
am
o
f
t
he
labels
L
at
er
,
t
his
LB
P
was
e
xten
de
d
by
us
in
g
nei
ghbour
hoods
f
or
m
any
siz
es
[2
2
]
.
If
we
use
a
bi
-
li
near
ly
interp
olati
ng
a
nd
t
he
ci
rc
ular
neig
hbo
urh
oods
of
the
pix
e
l
values,
we
c
an
al
low
a
ny
nu
m
ber
of
pixe
ls
in
the
nei
ghbo
urh
ood
a
nd
ra
diu
s
.
So
we
obta
in
a
ne
w
exte
ns
i
on
of
the
LBP
cal
le
d
unifo
rm
patte
rns.
T
he
LBP
op
e
rato
r
is
cal
le
d
un
i
form
if
it
con
ta
ins
at
m
os
t
two
transiti
on
s
(b
it
wi
se)
from
0
to
1
or
vice
vers
a
wh
e
n
the b
i
nar
y st
rin
g
is c
onside
red ci
rcu
la
r
.
A
is
t
he
histo
gr
am
o
f
la
bele
d
im
age
M(
x, y)
,
it
can be
def
i
ned as
:
(1)
w
he
re
n
:
nu
m
ber
of
la
bels
pro
du
ce
d by the
L
BP
,
a
nd
I
(2)
A
is
the
hist
ogram
and
repres
ent
t
he
distri
buti
on
of
t
he
l
oc
al
m
ic
r
o
-
patte
rn
s
of
t
he
te
xtu
re
im
age.
(
R,
P)
are
the t
w
o param
e
te
rs
in t
his h
is
togram
,
P
is t
he
points
(
sam
pling)
on a
circl
e
r
adi
us
R
see
F
ig
ure
4
.
T
he fe
at
ur
es
vecto
r
of
t
he hi
stogram
fo
rm
ed
will
b
e t
he
i
nput v
ect
or for
th
e SV
M cl
as
sific
at
ion
ste
p
.
Figure
4.
Th
re
e circula
r neig
hborh
oods
:
( 1,
6), (1,
8), (2, 1
6)
2.2.
GL
CM
meth
od
T
he
Gr
ay
-
Lev
el
Co
-
Occurre
nce
Ma
trix
(
G
LCM
)
is
the
se
cond
te
xtu
r
e
a
naly
sis
m
e
tho
d
te
ste
d
in
t
his
work.
It
base
d
on
an
e
xtracti
on
of
the
te
xtur
e
m
easur
ed
fro
m
an
i
m
age.
T
her
e
are
t
hree
par
am
et
ers
va
r
ie
d
in
this
m
et
ho
d:
the
distance
D
wh
e
r
e
D
is
the
distance
betwe
en
tw
o
pix
el
s,
the
an
gle
θ
,
it
is
the
an
gle
be
twee
n
tw
o
directi
on
s
,
and
the
offset
[22,
23
].
For
θ,
Ther
e
are
f
our
directi
on
s:
θ
=
0
°
,
θ
=
45
°,θ
=
90
°
an
d
θ
=
135
°
sh
ow
n
i
n
Fi
gur
e
5
.
We
de
fine
the
offset
par
a
m
et
er:of
fset=
(
0
D,
-
D
D,
-
D
0,
-
D
–
D).
I
n
o
u
r
w
ork
,
w
e
n
o
t
e
t
h
a
t
w
e
d
e
t
e
c
t
a
s
e
c
o
n
d
p
a
r
a
m
e
t
e
r
t
h
a
t
a
f
f
e
c
t
s
t
h
e
r
e
s
u
l
t
o
f
c
l
a
s
s
i
f
i
c
a
t
i
o
n
i
t
i
s
t
h
e
n
u
m
b
e
r
o
f
g
r
a
y
-
l
e
v
e
l
s
:
t
h
e
N
u
m
l
e
v
e
l
s
.
W
e
w
i
l
l
d
e
m
o
n
s
t
r
a
t
e
t
h
a
t
t
h
i
s
p
a
r
a
m
e
t
e
r
i
s
v
e
r
y
i
n
f
l
u
e
n
t
i
a
l
i
n
t
h
e
c
l
a
s
s
i
f
i
c
a
t
i
o
n
r
e
s
u
l
t
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
Text
ur
e cl
as
sif
ic
ation of f
abric
d
ef
ect
s
us
in
g ma
c
hin
e
lear
ni
ng
(
Ya
ssi
ne
Be
n Sa
le
m
)
4393
Figure
5. GLC
M par
am
et
ers
All t
he values
of G
LCM
M
mn
sh
oul
d be
norm
al
iz
ed.
T
han GL
CM
is ex
pr
es
sed by:
(3)
The
ori
gin
of
GLCM
is
the
pap
er
[
23
]
,
Haral
ic
k
propose
d
14
dif
fe
ren
t
sta
ti
sti
cal
feature
s
in
order
t
o
est
i
m
at
e
the
si
m
i
la
rity
in
th
e
te
xtu
re
im
a
ge
.
We
ha
ve
te
ste
d
al
l
the
14
featu
res
an
d
we
c
on
c
l
uded
tha
t
the
m
os
t
sign
if
ic
ant
on
es
a
re
4
featu
res
e
xpr
essed
i
n
T
a
ble
1,
a
nd
the
n
w
e
f
or
m
the
G
L
CM
feature
s
ve
ct
or
.
We
co
ns
i
der
an
im
age
wit
h
(
N×N)
dim
ensio
n,
wh
e
re
m,n
f
or
m
ed
the
co
ordina
te
s
of
the
G
LCM
co
-
occ
urren
ce
m
at
rix,
an
d
C
mn
is t
he
c
orres
pondin
g
el
em
ent o
f
(
m,n)
.
Table
1 Harali
ck’
s
p
a
ram
et
ers
Featu
res
Fo
r
m
u
la
Co
n
trast
1
0
1
0
2
.
)
(
N
m
N
n
mn
C
n
m
C
ON
Co
rr
elatio
n
mn
N
m
N
n
y
x
y
x
C
CO
R
1
0
1
0
)
1
)(
1
(
Ho
m
o
g
en
eit
y
1
0
1
0
1
N
m
N
n
mn
n
m
C
E
N
T
An
g
u
lar
Seco
n
d
Mo
m
en
t
1
0
1
0
2
N
m
N
n
mn
C
A
S
M
2.
3
.
LPQ
met
ho
d
The
thir
d
m
e
th
od
is
(LPQ
)
L
ocal
Ph
ase
Q
ua
ntiza
ti
on
.
It
is
base
d
on
local
i
ze
of
the
phase
inf
or
m
at
ion
extracte
d
us
in
g
the
2
-
D
D
FT
or
in
othe
r
wa
y
we
can
say
i
t
us
es
a
Fo
uri
er
trans
form
in
a
sh
ort
-
te
rm
(S
TFT
)
com
pu
te
d
ove
r
M
-
by
-
M
nei
ghbo
rho
od
rectangula
r
.
I
t
was
us
e
d
on
the
blu
r
in
va
riance
pro
per
ty
o
f
the
Four
ie
r
ph
a
se s
pectr
um
[
10
]
. x
is
the
pix
el
posit
ion
of the im
age
f
(
x) d
e
fine
d by:
(4)
Wh
ic
h
is
the
basis
vect
or
of
the
2
-
D
DF
T
and
u
is
the
f
reque
ncy,
an
d
is
ano
the
r
ve
ct
or
co
ntaini
ng
(M×M
)
i
m
age
sam
ples
fr
om
.
As
it
can
be
noti
ced
from
(4
),
an
eff
ic
ie
nt
w
ay
of
im
ple
m
e
nting
t
he
STF
T
is
to use
2
-
D
c
onvo
l
utions:
f
or
all
u.
(5)
C
o
m
p
u
t
a
t
i
o
n
c
a
n
b
e
p
e
r
f
o
r
m
e
d
u
s
i
n
g
1
-
D
c
o
n
v
o
l
u
t
i
o
n
s
f
o
r
t
h
e
r
o
w
s
a
n
d
c
o
l
u
m
n
s
s
u
c
c
e
s
s
i
v
e
l
y
s
i
n
c
e
t
h
e
b
a
s
i
c
f
u
n
c
t
i
o
n
s
a
r
e
s
e
p
a
r
a
b
l
e
,
I
n
L
P
Q
o
n
l
y
f
o
u
r
c
o
m
p
l
e
x
c
o
e
f
f
i
c
i
e
n
t
s
a
r
e
c
o
n
s
i
d
e
r
e
d
,
c
o
r
r
e
s
p
o
n
d
i
n
g
t
o
2
-
D
f
r
e
q
u
e
n
c
i
e
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
20
88
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
4390
-
4399
4394
(6)
a
is a
scala
r fre
qu
e
ncy
belo
w
t
he first zer
o
c
r
os
sin
g of
that
sat
isf
ie
s the c
onditi
on
:
∠
G
(
u) =
∠
F
(
u) for
al
l
H
(
u)
≥
0
l
et
a
nd
(7)
Im
{.} an
dRe{.
}return im
agin
ary an
d real
pa
rts of a
co
m
plex nu
m
ber
,
r
es
pe
ct
ively
.
The
c
orres
pondin
g 8
-
by
-
tran
sform
at
ion
m
a
t
rix
is
:
The
n
(8)
Fo
r
eac
h
pix
el
po
sit
io
n
t
his r
e
su
lt
s in
a
v
ect
or:
(9)
We
m
ark
the
sign
s
of
t
he
r
eal
in
F(x)
a
nd
im
aginar
y
in
F(x
)
an
d
w
e
record
t
he
ph
a
se
inf
orm
a
ti
on
in
the F
ourier c
oe
ff
ic
ie
nts
by
us
i
ng a sim
ple scal
ar qua
ntiza
ti
on
.
w
he
re
g
j
is t
he j
c
om
po
ne
nt
of the
ve
ct
or:
G(
x) = [
Re
{
F(x)}
, I
m
{F(x)}]
q
j
are
the
integ
er
values bet
w
een
0
a
nd
25
5 usin
g
L
PQ
:
(10)
Finall
y we
ob
t
ai
n
a
featu
re
ve
ct
or
i
n
put f
or t
he
cl
assifi
cat
io
n ph
a
se.
2.
4. Cl
as
sific
ati
on ph
as
e
We
us
e
m
achine
le
ar
ning
an
d
pr
act
ic
al
ly
the
m
ulti
-
cl
ass
SV
M
as
a
cl
a
ssifie
r
for
t
his
stu
dy
[
24]
.
The
m
ai
n
idea
of
The
S
VM
cl
assifi
er
is
to
ob
ta
in
a
f
unct
ion
f(
x
)
wh
ic
h
determ
ines
the
optim
al
hyperplan
e
that
separ
e
t
he
diff
e
re
nt
cl
asses’
zo
nes
.
W
e
ob
ta
in
t
his
hy
pe
rp
la
ne
w
he
n
it
m
anag
ed
to
s
epar
at
e
tw
o
cl
asses
,
or
m
or
e,
of
in
put
data
po
i
nts
s
how
n
in
Fig
ure
6
.
M
re
pr
e
se
nts
the
distanc
e
from
the
hype
rp
la
ne
to
t
he
cl
os
est
po
i
nt
f
or
both
cl
asses
of
data
po
i
nts.
In
orde
r
to
m
axim
iz
e
t
he
m
arg
in
M,
we
c
on
si
der
and
the
c
onditi
ons
of
the
pro
blem
will
b
e
wr
it
te
n:
(11)
w
he
re:
w
:
a
vecto
r def
ining t
he b
oundary,
n
:
num
ber
of
i
nput
of d
at
a
of
SV
M,
b
:
a scal
ar
thre
sh
ol
d value
.
:
input data
poi
nt,
T
he o
pti
m
al
h
yper
plane
pro
bl
e
m
o
f SVM,
f(
x
)
:
(12)
:
support
vecto
r havi
ng no
n
-
z
ero La
gra
ng
e
m
ul
ti
plier
and
it
sh
ould
be
us
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
Text
ur
e cl
as
sif
ic
ation of f
abric
d
ef
ect
s
us
in
g ma
c
hin
e
lear
ni
ng
(
Ya
ssi
ne
Be
n Sa
le
m
)
4395
Fig
ure
6.
O
ptim
al
sep
arati
on
hype
rp
la
ne of SVM
In
this
case
w
e
can
no
te
that
there
are
m
or
e
than
two
ty
pe
s
of
m
ulti
-
cl
a
ss
cl
assifi
ers
in
this
SVM
al
gorithm
.
The
first
is
cal
le
d
on
e
ag
a
inst
al
l
and
the
oth
e
r
is
cal
le
d
on
e
again
st
one
.
We
ha
ve
com
par
e
d
the
pe
rfor
m
ance
of
the
t
wo
ty
pes.
T
he
ty
pe
on
e
agai
ns
t
on
e
need
s
t
o
co
nst
ru
ct
one
cl
assi
fier
f
or
t
wo
a
rbi
trary
cl
asses,
i.e.
k(
k
-
1)
/2
cl
assifi
ers
al
l
to
gether.
F
or
O
ne
ag
ain
st
all
trai
ns
k
cl
assifi
ers
(
k
is
the
nu
m
ber
of
cl
asses).
I
n
tha
t
case
,
the
cl
as
sifie
r
trie
s
to
separ
at
e
the
cl
ass
i
fr
om
the
rest.
Then
the
two
-
cl
ass
cl
assifi
ers
evaluate
t
he
i
nput
data
a
nd
vote
on
their
cl
asses.
In
th
is
stud
y,
we
ha
ve
wor
k
ed
w
it
h
a
to
olbo
x
nam
ed
‘
sim
pl
eSVM
’
in
Mat
la
b
[
25
]
.
We ca
n
c
oncl
ude that i
t i
s a
si
m
ple an
d
ef
fici
ent
to
olbox
.
3.
DA
T
ABA
SE
In
19
95,
the
Tech
nische
U
niversi
tt
Ha
m
burg
el
ab
or
at
e
d
a
database
sp
e
ci
al
te
xtil
e
te
xtu
re
im
ages
nam
edTI
LD
A
[26]
,
w
hich
co
ns
ist
s
of
8
dif
f
eren
t
cl
asses
of
w
ov
e
n
fa
br
ic
.
T
he
hardest
c
ase
in
this
database
is
that
the
or
ie
nt
at
ion
of
the
te
xtil
e
is
no
t
known.
It
is
an
i
m
po
rtant
te
st
of
the
r
obust
ne
ss
of
our
ap
proac
h.
In
our
stu
dy,
we
ha
ve
inclu
ded
a
ne
w
cl
as
s
nam
ed
“No
Def
ect
”.
T
he
r
esults
show
th
at
this
cl
ass
reg
ar
de
d
as
th
e b
e
st avail
a
ble one.
I
t
pr
e
s
ents m
any for
m
s r
el
at
ed
to the tex
t
ur
es
ana
ly
sis and
rec
ogniti
on
Fig
ur
e
7
.
Figure
7
.
4 w
oven
f
a
br
ic
de
fe
ct
s class
4.
RESU
LT
S
AND A
N
ALYSIS
We h
a
ve
ch
os
e
n
to use 5
cl
ass
es o
f wove
n
fa
br
ic
d
e
fects
, “
Ho
le
”, “
Kink”,
Mi
ssing
w
e
ft”, “o
il
sati
n”
,
and
“
No
de
fect
”
;
each
cl
ass
c
on
ta
in
s
100
im
ages
:
20
im
ages
fo
r
trai
ni
ng
and
80
f
or
the
t
est
.
In
total
,
w
e
hav
e
us
e
d 500 s
am
ples f
r
om
the
TI
LDA da
ta
base
see Table
2
.
T
he
sam
ples are
picking
ra
ndom
ly
.
I
n
o
r
d
e
r
t
o
c
o
m
p
a
r
e
t
h
e
t
h
r
e
e
p
r
o
p
o
s
e
d
m
e
t
h
o
d
s
i
n
r
e
s
o
l
v
i
n
g
t
h
e
p
r
o
b
l
e
m
o
f
c
l
a
s
s
i
f
i
c
a
t
i
o
n
o
f
w
o
v
e
n
f
a
b
r
i
c
im
a
g
e
s
,
w
e
h
a
v
e
s
t
a
r
t
e
d
w
i
t
h
t
h
e
G
L
C
M
o
n
e
.
I
n
F
i
g
u
r
e
8
,
w
e
p
r
e
s
e
n
t
t
h
e
p
e
r
f
o
r
m
a
n
c
e
s
o
f
c
l
a
s
s
i
f
i
c
a
t
i
o
n
us
in
g
the
f
ollo
wing
GLCM
pa
ram
et
er
s
(c
ontrast
,
energy,
hom
og
eneit
y,
c
orrelat
ion
)
an
d
vary
ing
the
offset
value
s.
The
res
ults
of
cl
assifi
cat
ion
presente
d
in
F
ig
ur
e
8
s
hows
th
at
the
con
trast
giv
es
a
bette
r
cl
assifi
cat
ion
r
esult
with
92%
c
ompare
d
to
oth
e
r
par
am
et
ers:
ho
m
og
eneit
y,
energy,
an
d
c
orrelat
ion
s.
With
the
GLCM
m
et
hod
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
20
88
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
4390
-
4399
4396
we
rem
ark
t
hat
we
obta
ined
the
best
pe
rce
ntage
of
t
he
s
ucc
essfu
l
cl
assi
fied
im
ages
is
9
8.2
5%.
It
is
ob
t
ai
ne
d
with
the
co
uple
(
offset
,
N
um
le
vel
s
)
=
(
6,
3),
a
nd
the
r
unni
ng
tim
e
is
about
44.
84
sec
onds
a
s
sho
wed
in
Figure
9
and T
a
ble 3.
Table
2
.
Desc
riptor f
eat
ur
es
Featu
res
N
u
m
b
e
r
o
f
trainin
g
f
eatu
res
Nu
m
b
e
r
o
f
testin
g
f
eatu
res
No
def
ect (
ND
)
20
80
Ho
le
20
80
Kin
k
20
80
Oil satin
(
OS)
20
80
Missin
g
wef
t(M
W
)
20
80
Figure
8.
Cl
assifi
cat
ion
perfor
m
ance of th
e
GLC
M par
am
et
ers
a
nd
va
rio
us
values
of
offset
s
an
d
N
um
le
vel
s
Figure
9
.
Com
par
is
on of clas
sific
at
ion
perf
orm
ance b
et
wee
n
the
th
ree
m
eth
ods
Table
3.
Cl
assi
ficat
ion
pe
rform
ance of th
e
GLC
M m
e
tho
d
GLCM
(level, of
f
set)
Clas
s”OS”
Clas
s”MW
”
Clas
s”Ho
le”
Clas
s”k
in
k
”
Clas
s”ND”
Ti
m
e
(seso
n
d
)
Res
u
lt
GLCM
(2,3
)
95
9
6
.25
8
3
.75
7
8
.75
100
4
5
.35
9
0
.75
GLCM
(4,3
)
100
95
9
6
.25
90
100
5
3
.36
9
6
.25
GLCM
(6,3
)
100
9
8
.75
95
9
7
.5
100
4
4
.84
9
8
.25
GLCM
(2,5
)
9
8
.75
9
6
.25
9
1
.25
3
6
.75
100
6
9
.35
90
GLCM
(4,5
)
6
8
.75
100
8
3
.75
50
15
6
9
.88
6
3
.5
W
it
h
the
LBP
m
et
ho
d,
we
va
ried
tw
o
pa
ra
m
et
ers
(R,
P
)
and
we
obta
in
ed
99.75
%
as
the
best
res
ult
per
ce
ntage
il
lu
s
trat
ed
in
T
abl
e
4.
We
note
that
this
resu
lt
i
s
obta
ined
wit
h
LBP
m
et
hod
an
d
wit
h
the
couple
(1,
16)
a
nd
in
a
r
unning
ti
m
e
eq
ual
to
474.4
9
sec
onds.
We
can
c
on
cl
ude
t
hat
the
pe
rcen
t
age
of
this
m
eth
od
i
s
higher
tha
n
the
G
LCM
wh
il
e t
he di
sad
va
ntag
e it
is long
run t
i
m
e sh
own
i
n
Fig
ure
10
.
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
Text
ur
e cl
as
sif
ic
ation of f
abric
d
ef
ect
s
us
in
g ma
c
hin
e
lear
ni
ng
(
Ya
ssi
ne
Be
n Sa
le
m
)
4397
Table
4.
Cl
assi
ficat
ion
pe
rform
ance of th
e
L
BP m
et
ho
d
LBP(R,
P)
Clas
s”OS”
Clas
s”MW
”
Clas
s”Ho
le”
Clas
s”k
in
k
”
Clas
s”ND”
Ti
m
e
(
ses
o
n
d
)
Res
u
lt
LBP(1,4
)
9
3
.75
9
2
.5
9
6
.25
95
100
1
1
.69
9
5
.5
LBP(2,4
)
7
8
.75
9
8
.75
95
9
1
.25
100
1
0
.7
9
2
.75
LBP(1,8
)
100
9
8
.75
9
1
.25
9
6
.25
100
1
6
.18
9
7
.25
LBP(5,8
)
6
7
.5
100
90
8
1
.25
100
1
5
.7
8
7
.75
LBP(1,1
6
)
9
9
.75
100
100
9
8
.75
100
4
7
4
.49
9
9
.75
Figure
10
.
C
om
par
is
on
of ru
nn
i
ng tim
e
of
t
he
th
ree
m
et
hods
W
it
h
t
he
LP
Q
m
et
ho
d
,
we
obta
in
us
a
resu
lt
with
t
he
c
om
bin
at
ion
of
the
f
our
par
am
et
ers
(
win_siz
e
,
decor,
f
req
ste
m
,
and
m
od
e)
=
(3,1,2
,h)
a
nd
the
res
ult
is
a
bout
90.
5%
s
ho
wn
in
Ta
ble
5
.
This
m
et
ho
d
s
ee
m
s
to
be
le
ss
acc
ur
at
e
than
t
he
tw
o
oth
e
r
m
et
ho
ds.
The
detai
le
d
pe
rce
ntage
of
co
rr
ect
ly
cl
assifi
cat
ion
s
for
ea
ch
m
et
ho
d
is
il
lus
trat
ed
in
Fi
gur
e
7.
We
no
ti
ce
that
the
im
age
s
with
no
defe
ct
s
‘ND’
a
re
e
asi
er
to
cl
assif
y
than
the
ot
her
s
.
F
urt
her
m
or
e,
the
m
et
ho
d
LBP
pro
vid
es
a
good
cl
assi
ficat
ion
that
reac
hed
98.
75%
wit
h
a
sh
ort
tim
e 1
6.
18s
. W
e can sa
y t
hat t
he GLCM
m
eth
od ca
n give
a
good classi
fica
ti
o
n rate
but
with a lon
ger tim
e.
The
LP
Q
m
eth
od
is
cha
racteri
zed
by
it
s
sh
ort
run
ning
ti
m
e
and
it
s
cl
a
ssific
at
ion
rate
lower
th
a
n
the
ot
her
t
ec
hniq
ues
pr
eci
sel
y
of
t
he
"H
ole
"
cl
ass.
A
noth
er
im
po
rtant
r
esult
co
ncerni
ng
the
S
VM
c
la
ssifie
r
al
gorithm
is
determ
ined;
m
or
e
det
ai
le
d
res
ults
are
pr
ese
nt
ed
in
F
igure
11.
The
ap
pro
ach
"1
vs
1"
inc
reases
the
cl
assifi
cat
i
on
rate
m
or
e
than
the
“
1vsal
l”
.
For
ex
am
ple
,
the
LBP
m
e
t
hod
with
par
a
m
et
ers
(1
,
16
)
le
ads
t
o
99.75%
, whe
re
as w
it
h t
he
a
pp
ro
ac
h
“
1vsal
l”
we ob
ta
ine
d 9
9.2
5
%
.
Table
5.
Cl
assi
ficat
ion
pe
rform
ance of th
e
L
PQ
m
et
ho
d
LPQ
(wins
ize,
d
ecor,f
reqs
te
m
,
m
o
d
e)
Clas
s
”OS”
Clas
s
”M
W
”
Clas
s
”Ho
le”
Clas
s
”k
in
k
”
Clas
s
”ND”
Ti
m
e
(se
so
n
d
)
Res
u
lt
LPQ(
3
,1,1
,h)
8
8
.7
5
100
80
9
2
.5
100
2
3
.03
8
6
.25
LPQ(
3
,1,2
,h)
9
6
.25
100
5
1
.25
95
100
2
2
.36
8
8
.5
LPQ(
3
,1,3
,h)
90
100
5
2
.5
95
100
2
3
.25
8
7
.5
LPQ(
9
,1,1
,h)
9
2
.5
100
4
8
.75
8
7
.5
9
8
.75
2
3
.26
8
5
.5
LPQ(
9
,1,2
,h)
90
100
4
8
.75
8
8
.75
100
2
2
.81
8
5
.5
LPQ
(9,1
,3,h
)
100
50
50
8
8
.75
100
2
3
.66
8
6
.25
Figure
11.
C
om
par
ison
of p
e
rfor
m
ance b
et
ween t
w
o
m
ult
ic
la
ss “1
vs
1” a
nd “1vs
all
”
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
20
88
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
4390
-
4399
4398
5.
CONCL
US
I
O
N
We
ha
ve
st
udie
d
the
prob
le
m
of
aut
om
ati
c
cl
assifi
cat
ion
of
the
w
ov
e
n
fa
br
ic
def
ec
ts
base
d
on
te
xtu
re
of
t
he
i
m
ages
.
T
he
presente
d
wor
k
co
ns
ist
s
on
a
com
par
at
ive
s
tud
y
of
th
ree
diff
e
re
nt
ap
pro
aches:
GLCM,
LBP
and
LP
Q.
Using
t
he
G
LCM
m
et
ho
d,
a
fter
stu
dyin
g
the
diff
e
re
nt
sta
ti
sti
cal
featur
e
s,
we
c
on
cl
ud
e
that
only
4
are
the
m
os
t
sign
i
ficant
Co
ntrast
,
H
om
og
eneit
y,
co
rr
el
at
i
on,
a
nd
ene
r
gy.
T
he
bes
t
on
e
is
the
co
nt
rast.
W
it
h
t
his
m
et
ho
d,
we
re
m
ark
that
t
he
best
perform
ance
res
ult
ofcl
as
sific
at
ion
is
a
bout
of
98.25%
with
a
44
.
84
seco
nds
runn
i
ng
ti
m
e
.
W
it
h
the
LBP
m
et
ho
d,
bas
ed
on
the
pro
per
ty
of
local
i
m
age
te
xtu
res
a
nd
th
ei
r
o
cc
urre
nce
histo
gram
,
we
obta
ine
d
97.25
%
as
t
he
bes
t
resu
lt
pe
rcent
age
with
a
r
unni
ng
tim
e abo
ut
16.
18 sec
onds
only
.
The
thir
d
m
et
ho
d
LPQ
is
bas
ed
on
local
fr
e
qu
e
ncy
coe
ff
ic
ie
nt;
it
sh
ow
e
d
no
t
bad
res
ults
with
88
.5%
perform
ance
and
22.
36
sec
onds.
T
he
e
sta
b
li
sh
ed
st
ud
y
pro
ves
t
hat
S
V
M
is
a
su
it
a
ble
cl
assifi
er
for
su
c
h
pro
blem
s.
A
TILDA
databas
e
is
us
e
d
to
te
st
our
al
gorit
hm
.
So
m
e
def
e
ct
s
wh
ic
h
are
easi
er
to
cl
assi
fy
than
oth
e
rs,
for
exa
m
ple,
un
relat
ed
co
r
pu
s
,
hole
,
and
oil
sat
in
a
re
easi
er
to
cl
a
ssify.
I
n
co
ncl
us
io
n,
we
rem
ark
t
hat
LBP
m
e
tho
d
i
s
the
best
m
e
thod
f
or
t
his
pro
blem
of
reco
gnit
ion
a
nd
cl
assifi
cat
ion
wove
n
de
fect
s,
it
is
char
act
e
rized a
lso
by go
od ru
nn
i
ng tim
e
REFERE
NCE
S
[1]
Kum
ar,
Aja
y
,
“
Com
pute
r
-
vision
-
base
d
fab
ri
c
de
fec
t
d
et
e
ction:
A
surve
y
,
”
IE
E
E
Tr
ans.
Ind.
E
le
c
tron,
vol
.
55
,
no.
1
,
pp
.
348
-
3
63,
2008
.
[2]
Xie,
Xi
anghua
,
“
A
rev
ie
w
of
r
ec
en
t
adv
anc
es
in
surfac
e
def
ect
detec
t
ion
usin
g
te
xtur
e
an
aly
s
is
te
chn
ique
s
,
”
ELCVIA
: E
lectr
onic
le
t
te
rs
on
c
omputer
vi
sion
a
nd
image
ana
ly
s
is
,
vol
.
7
,
no
.
3
,
pp.
1
-
22
,
2008
.
[3]
Murino
V.,
Bice
go
M.
and
Ross
i
I.
A,
“
Stat
isti
c
al
cl
assifi
cation
of
raw
te
xt
il
e
de
fec
ts
,
”
Proceedi
ngs
of
ICIP’
O4
,
vol.
4
,
pp
.
311
-
3
14,
2004
.
[4]
Tsai
I
.
S.,
Li
n
C
.
H.,
and
Li
n
J.
-
J
,
“
Appl
y
ing
an
art
if
ic
i
al
n
eur
al
net
work
to
p
at
t
e
rn
rec
ogn
it
ion
in
fab
ri
c
def
ec
ts
,
”
Text
ile
Re
search
J
ournal,
vol
.
65
,
no
.
3
,
pp
.
123
-
130,
1995
.
[5]
Y
a
h
i
a
S
.
,
B
e
n
S
a
l
e
m
Y
.
,
a
n
d
A
b
d
e
l
k
r
i
m
M
.
N
.
,
“
M
u
l
t
i
p
l
e
s
c
l
e
r
o
s
i
s
l
e
s
i
o
n
s
d
e
t
e
c
t
i
o
n
f
r
o
m
n
o
i
s
y
m
a
g
n
e
t
i
c
r
e
s
o
n
a
n
c
e
b
r
a
i
n
i
m
a
g
e
s
t
i
s
s
u
e
,”
1
5
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
M
u
l
t
i
-
C
o
n
f
e
r
e
n
c
e
o
n
S
y
s
t
e
m
s
,
S
i
g
n
a
l
s
a
n
d
D
e
v
i
c
e
s
,
p
p
.
2
4
0
-
2
4
5
,
2
0
1
8
.
[6]
Yahia
S.
,
Ben
Sale
m
Y.,
and
Abdelkr
im
M.
N.,
“
Anal
y
sis
o
f
3D
te
xtu
res
b
ase
d
on
f
ea
tur
es
ext
r
ac
t
ion
”.
15
th
Inte
rnational
M
ult
i
-
Conf
ere
nc
e on Sy
stems,
Sign
als an
d
Dev
ices,
2018.
[7]
Nanni
L.,
Lu
m
ini
A.,
Brah
nam
S.,
“
S
urve
y
on
LBP
b
ase
d
te
x
ture
d
e
script
ors
for
i
m
age
proc
essin
g,
”
Ex
pert
Syste
ms
w
it
h
Appl
i
cat
ion
s,
vol. 39, no. 3, p
p.
3634
-
3641
,
2
012.
[8]
Chai
re
t
R.
,
B
en
Sale
m
Y.,
and
Aoun
M.,
“
Feat
ur
es
ext
r
ac
t
ion
an
d
la
nd
cov
er
c
las
sific
at
ion
using
Senti
n
el
2
da
ta
,
”
19th
Inte
rnat
ion
al
Confe
ren
ce
o
n
Sci
en
ce
s
and
Techni
ques
of
Aut
omatic
Co
n
t
rol
and
Computer
Engi
ne
ering,
pp.
497
-
500
,
20
19.
[9]
Ben
Sale
m
Y.
,
and
Sale
m
Nasr
i,
“
W
oven
fab
r
i
c
def
ec
ts
d
et
e
ction
base
d
on
texture
cl
assifi
ca
t
i
on
al
gori
thm
,
”
8th
Int
ernati
onal
Multi
-
Confe
ren
ce
on
Syst
ems,
S
ignal
s and
De
vic
es,
2011
.
[10]
Ojansivu
V.,
an
d
Heikki
la
J.
,
“
Blur
insensit
iv
e
te
xtur
e
class
ifi
cation
using
local
phase
quant
i
z
a
tion
,”
Int
ernati
on
al
conf
ere
n
ce on
i
mage
and
signal
,
pp
.
236
-
243
,
2
008.
[11]
Ben
Sale
m
Y.,
Nasri
S.,
and
T
ourkiR.
,
“
Cl
assific
a
t
ion
of
ti
ss
ues
b
y
n
eur
a
l
netw
ork
,”
Proc
ee
di
ngs
of
the
I
EEE
Inte
rnational
Co
nfe
renc
e
on
Elec
tronic
s,
C
ircui
ts,
and
Syst
ems,
20
05.
[12]
Zha
ng
X.
F.
and
Brese
e
R.
R.
,
“
Fabric
d
efect
d
e
te
c
ti
on
and
Cl
as
sific
a
ti
on
using
i
m
age
an
aly
s
is
,
”
Text
ile
Re
search
Journal,
vo
l.
65,
no.
1,
pp.
1
-
9,
1
995.
[13]
W
ood
E.
J.,
“
Appl
y
ing
fouri
er
and
associate
d
t
ran
sform
s
to
pat
te
rn
ch
aract
er
iza
ti
on
in
te
xt
il
es
,
”
Text
i
le
R
ese
arc
h
J
ournal,
vo
l. 60, no. 4, pp. 212
-
2
20,
1990
.
[14]
Uns
er
M.
and
A
de
F.,
“
Feat
ure
e
xtra
c
ti
on
and
de
ci
sion
proc
edur
e
for
au
tomate
d
i
nspec
ti
on
of
te
x
t
ure
d
m
at
eri
a
ls,”
Pat
te
rn
Recogni
t
ion
Letters
,
vo
l.
2,
no
.
2
,
pp
.
181
-
191,
1984
.
[15]
B
o
d
n
a
r
o
v
a
A
.
,
B
e
n
n
a
m
o
u
n
M
.
,
a
n
d
K
u
b
i
k
K
.
K
.
,
“
D
e
f
e
c
t
d
e
t
e
c
t
i
o
n
i
n
t
e
x
t
i
l
e
m
a
t
e
r
i
a
l
s
b
a
s
e
d
o
n
a
s
p
e
c
t
s
o
f
t
h
e
H
V
S
,
”
P
r
o
c
e
e
d
i
n
g
s
o
f
I
E
E
E
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
S
y
s
t
e
m
s
,
M
a
n
a
n
d
C
y
b
e
r
n
e
t
i
c
s
,
v
o
l
.
5
,
p
p
.
4
4
2
3
-
4
4
2
8
,
1
9
9
8
.
[16]
Cohen
F.
S.,
Fan
Z.
,
and
Atta
li
S.
,
“
Autom
at
ed
inspec
t
i
on
of
te
xti
le
f
abr
ic
s
using
textura
l
m
odel
s,”
IEE
E
Tr
ansacti
o
ns on
Pattern
An
aly
sis and
Ma
ch
ine
In
te
l
li
gen
ce,
vol.
13
,
no
.
8
,
pp
.
803
-
808,
1991.
[17]
Chen
J.
and
Jai
n
A.
K.,
“
A
str
uct
ura
l
appr
oa
c
h
to
ide
nt
if
y
d
e
fec
ts
in
t
ext
ur
e
d
images
,”
Pro
ce
ed
ings
of
IEEE
Inte
rnational
Co
nfe
renc
e
on
Syst
ems,
Man,
and
C
ybe
rnet
ic
s,
vol.
1,
pp
.
29
-
32
,
19
88.
[18]
Atal
a
y
A.,
“
Autom
at
ed
def
ect
in
spec
ti
on
of
te
xt
i
le
fab
ri
cs
using
m
ac
hine
vision
t
ec
hniqu
es,
”
M.
S
.
the
sis,
Bogazici
Unive
rs
it
y
,
1995
.
[19]
Ta
j
eri
pour
F.,
K
abi
r
E.,
and
She
ikhi
A.,
“
Fabric
def
ect
detec
t
ion
using
m
odifi
ed
loc
al
bin
ar
y
p
at
t
e
rns
,”
EURA
SI
P
Journal
on
Adv
a
nce
s
in
Signal P
roce
ss
ing,
vol
.
2
008,
pp
.
1
-
12
,
2
007.
[20]
Burges,
C
.
J.
C.
,
“
A
tut
or
ia
l
on
support
ve
ct
or
m
a
chi
nes
for
pa
tt
ern
re
cogni
t
ion
,
”
Data
Min
ing
and
Knowle
dg
e
Di
scov
ery
,
vol
.
2,
no
.
2
,
pp
.
121
-
167
,
1998
.
[21]
Ben
Sale
m
Y.,
and
Nasri
S.,
“
A
compara
ti
ve
stud
y
of
ro
ta
t
i
on
-
inva
ri
ant
t
ex
ture
c
la
ss
ifi
c
at
i
on
using
SV
M
,”
3rd Int
ernati
ona
l
Conf
ere
nce on Signals,
C
ircui
ts
and
Syst
ems,
20
09.
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
Text
ur
e cl
as
sif
ic
ation of f
abric
d
ef
ect
s
us
in
g ma
c
hin
e
lear
ni
ng
(
Ya
ssi
ne
Be
n Sa
le
m
)
4399
[22]
Ojal
a
,
T.,
Pie
ti
k
äi
nen
,
M.,
and
Mäe
npää,
T
.
,
“
Multi
resolution
gra
y
-
sca
l
e
and
r
ota
ti
on
inva
ri
ant
te
xture
class
ifi
c
a
ti
on
with
local
bin
ar
y
pa
tterns,
”
I
EEE
Tr
ansacti
ons
on
Pa
tt
ern
Analysis
and
Mac
hi
ne
Intelli
g
ence,
vol.
24,
no
.
7
,
pp.
971
-
986
,
20
02.
[23]
Hara
lick,
R
.
M.,
Shanm
ugam,
and
K.,
Dinstei
n
,
I
.
,
“
Te
x
tura
l
f
eat
ure
s
for
image
c
la
ss
ifi
c
at
ion
,”
I
E
EE
Tr
ansacti
ons
on
systems,
man,
and
c
ybe
rne
ti
cs
,
vol. 3, no. 6,
pp.
610
-
621,
19
73.
[24]
W
ang
Z.,
Xue
X
.
,
“
Support
Vec
t
or
Mac
hin
esApplicat
ions,
”
Sprin
ger
,
2014
.
[25]
Gaë
lle,
L.,
C
anu
,
S., Vishw
ana
th
an
S.V.N.
,
Ale
x
ande
r, S.
J.
,
and
Chat
topa
dh
y
a
y
,
M.,
“
Boît
e
à
outils
SV
M sim
ple
et
rap
ide
,
”
Revue
d
'int
ellige
n
ce artific
i
el
l
e,
vol
.
19
,
no.
4
-
5
,
pp
.
741
-
767,
2005.
[26]
W
orkgroup
on
T
ext
ure
Anal
y
sis
of
DF
G,
TIL
DA
te
x
ti
l
e
t
ext
ur
e
d
at
ab
ase
,
1996
htt
p://lm
b.
infor
m
at
ik.
uni
-
fr
ei
bu
rg.
de/
r
ese
ar
ch/
df
g
-
te
xtur
e/
t
il
da
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Yas
sin
e
ben
S
ale
m
was
born
in
Gabe
s
(
Tun
isia
)
on
1973
.
He
bega
n
teac
h
ing
at
Tuni
si
a
Univer
sit
y
in
1999
in
the
«
H
ighe
r
Instit
ut
e
of
Te
chnol
og
ic
a
l
Studie
s
of
So
uss
e
-
ISETSo»
.
He
rec
ei
v
ed
the
Stat
e
Doctor
a
te
in
2011
at
the
Univer
sit
y
o
f
Sfax.
Since
2
013
he
has
bee
n
an
As
socia
te
Profes
sor
in
computer
scie
nc
e
at
the
“
Nati
ona
l
School
of
Engi
n
ee
rs
of
Gabe
s
–
ENIG”,
Univer
sit
y
of
Gabe
s,
He
is
the
m
ember
of
the
Resea
rch
L
abor
at
or
y
«
Modeli
ng,
Anal
y
s
is
and
Control
S
y
st
ems
-
MA
CS
-
LR16E
S22
(
ww
w.m
ac
s
.
tn
).
He
has
dir
e
ct
ed
m
ore
tha
n
50
Engi
ne
eri
ng
proje
c
ts a
nd
m
aste
r and
3
the
s
es.
Mohame
d
Nac
eur
Ab
delkri
m
was
born
in
G
abe
s
(Tun
isia
)
o
n
1958.
H
e
b
eg
an
t
eachi
ng
a
t
Tuni
sia
Univer
si
t
y
in
1981
in
the
«
Nati
onal
School
of
Engi
nee
rs
of
Tuni
s
–
ENIT
»
.
He
rec
ei
ved
the
Sta
te
Do
ct
o
rat
e
in
2003
.
S
inc
e
200
3
he
h
as
bee
n
a
Profe
ss
or
in
Autom
at
ic
Con
trol
at
the
“
Nati
ona
l
School
of
Engi
n
e
ers
of
Gabe
s
–
ENIG”,
Univer
sit
y
of
Gabe
s
.
He
is
the
Hea
d
of
the
Rese
arc
h
La
bora
tor
y
«
Modeli
ng,
Ana
l
y
sis
and
Con
tr
ol
S
y
stems
-
MA
CS
-
LR16E
S22
(
ww
w.m
ac
s.tn
)
with
about
100
r
ese
arc
h
ers.
He
h
as
m
ore
tha
n
50
publi
c
at
ions
and
m
ore
tha
n
200
comm
unic
at
ions
.
He
has
di
recte
d
m
ore
th
an
50
T
hesis
Evaluation Warning : The document was created with Spire.PDF for Python.