I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
, pp.
4711
~
4721
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
13
.i
4
.pp
4711
-
4721
4711
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
A
n
c
h
or
se
l
e
c
t
i
on
b
ase
d
d
e
e
p
l
e
ar
n
i
n
g t
w
o st
age
f
ab
r
i
c
d
e
f
e
c
t
l
oc
al
i
z
at
i
on
H
at
t
ar
k
i
P
oo
j
a
1
, S
h
r
id
e
va S
om
a
2
1
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
S
c
i
e
nc
e
a
nd
E
ngi
ne
e
r
i
ng,
V
i
s
ve
s
va
r
a
ya
T
e
c
hnol
ogi
c
a
l
U
ni
ve
r
s
i
t
y
, B
e
l
a
ga
vi
, I
ndi
a
2
D
e
pa
r
t
m
e
nt
of
C
om
put
e
r
S
c
i
e
nc
e
a
nd
E
ngi
ne
e
r
i
ng, P
D
A
C
ol
l
e
ge
of
E
ngi
ne
e
r
i
ng, G
ul
ba
r
ga
, I
ndi
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
N
ov
22
,
2023
R
e
vi
s
e
d
A
pr
2
,
2024
A
c
c
e
pt
e
d
A
pr
17
,
2024
Localizing
and
classifying
fabric
defects
is
a
crucial
step
in
the
quality
control
process
used
in
the
production
of
textiles.
Recently,
fabric
defect
classif
ication
and
detection
have
made
use
of
deep
learning
approaches
based
on
anchor
selection.
But
due
to
in
effectiveness
in
anchor
selection,
the
comput
ational
overhead and localiza
tion error are
higher in
these solutions.
As a solu
tion to
this
problem,
this
work
proposes
a
two
-
stage
improvised
anchor
se
lection
deep
learning
technique.
In
first
stage,
qu
aternion
fourier
transform
fre
quency
domain
analysis
along
with
super
pixel
segmentation
is
done
over
the
fabric
image
to
select
probable
defect
r
egions.
In
the
second
stage
deep
le
arning
based regression along with super pixel segme
nt comparison is done o
ver the
probable
defect
regions
localize
and
categorize
the
defect.
D
ue
to
effectivenes
s
in
selectio
n
of
probable
defect
regions
and
categorizat
ion
of
regions,
the
defect
localizatio
n
accuracy
is
increased
at
a
comparati
ve
low
computat
ional
overhead
in
the
proposed
two
stage
improvise
anchor
selection
deep
learning
technique.
Testing
against
the
irish
longitudinal
study
on
ageing
(
TILDA)
fabric
defect
detection
dataset,
the
proposed
solution
is
fo
und
to
provide
1.2%
higher
f
abric
defect
localization
accuracy
at
a
3%
lower
computat
ion overh
ead compared
to mos
t recent ex
isting
works.
K
e
y
w
o
r
d
s
:
A
nc
hor
s
e
le
c
ti
on a
ppr
oa
c
h
D
e
e
p l
e
a
r
ni
ng r
e
gr
e
s
s
io
n
F
a
br
ic
de
f
e
c
t
de
te
c
ti
on
Q
ua
te
r
ni
on F
our
ie
r
t
r
a
ns
f
or
m
S
upe
r
pi
xe
l
s
e
gm
e
nt
a
ti
on
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
H
a
tt
a
r
ki
P
ooj
a
D
e
pa
r
tm
e
nt
of
C
om
put
e
r
S
c
ie
nc
e
a
nd
E
ngi
ne
e
r
in
g,
V
is
ve
s
v
a
r
a
ya
T
e
c
hnol
ogi
c
a
l
U
ni
ve
r
s
it
y
B
e
la
ga
vi
, I
ndi
a
E
m
a
il
:
pooj
a
ha
tt
a
r
ki
11@
r
e
di
f
f
m
a
il
.c
om
1.
I
N
T
R
O
D
U
C
T
I
O
N
D
ur
in
g
th
e
te
xt
il
e
pr
oduc
ti
on
pr
oc
e
s
s
,
a
v
a
r
ie
ty
of
e
le
m
e
nt
s
in
f
lu
e
nc
e
f
a
br
ic
qu
a
li
ty
,
in
c
lu
di
ng
m
a
te
r
ia
l
qua
li
ty
, m
e
c
ha
ni
c
a
l
c
ons
id
e
r
a
ti
ons
, dye
t
ype
, ya
r
n s
iz
e
, a
nd hu
m
a
n f
a
c
to
r
s
. O
ne
of
t
he
m
os
t
im
po
r
ta
nt
qua
li
t
y
c
ont
r
ol
pha
s
e
s
in
th
e
te
xt
il
e
m
a
nuf
a
c
tu
r
in
g
pr
oc
e
s
s
is
th
e
de
t
e
c
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
of
f
a
br
ic
f
a
ul
ts
.
T
he
tr
a
di
ti
ona
l
s
tr
a
te
gy
of
us
in
g
hum
a
n
vi
s
ua
l
e
xa
m
in
a
ti
on
to
f
in
d
f
a
ul
ts
is
c
os
tl
y,
ti
m
e
-
c
ons
um
in
g,
a
nd
s
lo
w
(
12
m
e
te
r
s
pe
r
m
in
ut
e
)
.
I
t
is
c
ons
e
que
nt
ly
uns
ui
te
d
f
or
m
a
s
s
pr
oduc
ti
on.
F
a
s
t
a
s
s
e
m
bl
y
is
ne
c
e
s
s
a
r
y
in
la
r
ge
-
s
c
a
le
m
a
nuf
a
c
tu
r
in
g pr
oc
e
s
s
e
s
, a
nd f
a
ul
ts
m
us
t
be
i
de
nt
if
ie
d e
a
r
ly
on t
o a
voi
d r
e
s
our
c
e
w
a
s
te
a
nd f
ut
ur
e
pha
s
e
s
of
de
f
e
c
t
de
v
e
lo
pm
e
nt
.
E
a
r
ly
de
f
e
c
t
di
s
c
ove
r
y
boos
ts
p
r
oduc
t
va
lu
e
a
nd
r
e
duc
e
s
c
or
por
a
te
lo
s
s
e
s
[
1]
.
M
a
ny
a
ut
om
a
ti
c
s
tr
a
te
gi
e
s
f
or
de
te
c
ti
ng
f
a
br
ic
pr
obl
e
m
s
ha
v
e
be
e
n
in
v
e
s
ti
ga
te
d
in
a
tt
e
m
pt
to
a
tt
a
in
th
is
obj
e
c
ti
ve
.
E
xi
s
ti
ng
a
ut
om
a
ti
c
f
a
br
ic
f
la
w
d
e
te
c
ti
on
s
ys
te
m
s
c
a
n
be
di
vi
de
d
in
to
two
c
a
te
gor
ie
s
:
c
la
s
s
ic
a
l
a
ppr
oa
c
he
s
a
nd
m
a
c
hi
ne
le
a
r
ni
ng
te
c
hni
que
s
.
T
r
a
di
ti
ona
l
m
e
t
hods
de
te
c
t
pr
obl
e
m
s
by
e
xt
r
a
c
ti
ng
s
ta
ti
s
ti
c
a
l,
s
tr
uc
tu
r
a
l,
s
pe
c
tr
a
l,
a
nd
m
ode
l
-
ba
s
e
d
in
f
or
m
a
ti
on
a
nd
th
r
e
s
ho
ld
in
g
th
e
m
.
T
he
y
a
r
e
not
a
da
pt
a
bl
e
to
f
a
br
ic
te
xt
ur
a
l
c
ha
nge
s
, a
nd t
hr
e
s
hol
ds
or
c
ha
r
a
c
te
r
is
ti
c
s
m
us
t
be
f
in
e
-
tu
ne
d or
t
a
il
or
e
d t
o e
a
c
h f
a
br
ic
.
I
n s
hor
t,
t
he
s
e
s
tr
a
te
gi
e
s
l
a
c
k g
e
ne
r
a
li
ty
. A
s
a
r
e
s
ul
t,
m
a
c
hi
ne
l
e
a
r
ni
ng a
ppr
oa
c
he
s
a
r
e
pr
e
s
e
nt
e
d a
s
s
ol
ut
io
ns
t
o t
hi
s
pr
obl
e
m
.
T
he
y
f
a
ll
in
to
two
c
a
te
gor
ie
s
:
c
onve
nt
io
na
l
m
a
c
hi
ne
le
a
r
ni
ng
m
e
th
ods
a
nd
de
e
p
l
e
a
r
ni
ng
m
e
th
ods
.
I
n
or
de
r
to
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4711
-
4721
4712
de
te
c
t
da
m
a
g
e
d
f
a
br
ic
s
,
c
la
s
s
ic
a
l
m
a
c
hi
ne
le
a
r
ni
ng
te
c
hni
que
s
t
a
ke
a
r
a
nge
of
m
a
nu
a
ll
y
c
r
e
a
te
d
vi
s
ua
l
s
ig
na
ls
a
nd
tr
a
in
m
a
c
hi
ne
le
a
r
ni
ng
a
lg
or
it
hm
s
li
ke
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
(
S
V
M
)
,
a
r
ti
f
ic
ia
l
ne
ur
a
l
ne
twor
ks
(
ANN
)
a
s
w
e
ll
a
s
ot
he
r
s
.
R
e
c
e
nt
ly
,
f
a
br
ic
de
f
e
c
t
de
te
c
ti
on
ha
s
m
a
de
us
e
of
de
e
p
le
a
r
ni
ng
[
2
]
,
[
3]
.
B
y
us
in
g
c
onvolut
io
n
ope
r
a
ti
ons
,
de
e
p
le
a
r
ni
ng
c
l
a
s
s
if
ie
r
s
m
a
y
le
a
r
n
c
om
pl
ic
a
te
d
f
e
a
t
ur
e
s
w
it
hout
th
e
r
e
qui
r
e
m
e
nt
f
or
ha
nd
-
c
r
a
f
te
d
f
e
a
tu
r
e
s
.
O
ve
r
70
di
s
ti
nc
t
c
a
te
gor
ie
s
e
xi
s
t
f
or
f
a
br
ic
de
f
e
c
ts
.
T
he
m
a
jo
r
it
y
of
f
a
ul
ts
in
phot
os
a
r
e
qui
te
s
m
a
ll
a
nd
ha
ve
a
n
odd
a
s
pe
c
t
r
a
ti
o.
O
ne
s
ta
ge
a
nd
two
s
ta
ge
d
e
e
p
le
a
r
ni
ng
a
lg
or
it
hm
s
a
r
e
th
e
two
ty
pe
s
of
a
lg
or
it
hm
s
ut
il
iz
e
d
f
or
m
ic
r
os
c
opi
c
f
a
ul
t
de
te
c
ti
on.
O
ne
-
s
ta
ge
a
lg
or
it
hm
s
a
r
e
m
or
e
a
c
c
ur
a
te
but
ha
ve
a
f
a
s
te
r
in
s
pe
c
ti
on
s
pe
e
d
[
4]
.
T
w
o
s
ta
ge
a
lg
or
it
hm
s
ha
ve
hi
ghe
r
a
c
c
ur
a
c
y
but
ha
v
e
hi
ghe
r
r
unt
im
e
.
A
nc
hor
s
e
le
c
ti
on
ba
s
e
d
de
e
p
le
a
r
ni
ng
a
ppr
oa
c
he
s
r
e
duc
e
th
e
c
om
put
a
ti
ona
l
ove
r
he
a
d
of
two
s
ta
ge
a
lg
or
it
hm
s
by s
e
le
c
ti
ng
a
nc
hor
poi
nt
s
or
pr
oba
bl
e
de
f
e
c
ti
ve
r
e
gi
ons
a
r
ound
w
hi
c
h
de
f
e
c
t
d
e
te
c
ti
on
s
h
oul
d
w
or
k.
B
ut
th
e
s
e
m
e
th
ods
c
a
n
be
f
ur
th
e
r
im
pr
ove
d
by
s
e
le
c
ti
ng
th
e
a
nc
hor
w
it
h
h
ig
he
r
pr
oba
bi
li
ty
of
be
in
g
de
f
e
c
t
a
nd
r
e
duc
in
g
th
e
num
be
r
of
r
e
dunda
nt
a
nc
hor
s
.
S
e
le
c
ti
n
g
th
e
a
n
c
hor
s
w
it
h
hi
ghe
r
pr
oba
bi
li
ty
of
tu
r
ni
ng
de
f
e
c
t
r
e
duc
e
s
th
e
c
om
put
a
ti
on
c
om
pl
e
xi
ty
a
nd
in
c
r
e
a
s
e
s
th
e
de
f
e
c
t
lo
c
a
li
z
a
ti
on
a
c
c
ur
a
c
y.
T
hi
s
w
or
k
pr
opos
e
s
a
two
s
ta
ge
im
pr
ovi
s
e
d
a
nc
hor
s
e
le
c
ti
on
de
e
p
le
a
r
ni
ng
te
c
hni
que
a
ddr
e
s
s
in
g
th
e
pr
obl
e
m
in
de
f
e
c
t
l
oc
a
li
z
a
ti
on
a
nd
c
a
te
gor
iz
a
ti
on
a
t
a
c
om
pa
r
a
ti
ve
ly
lo
w
e
r
c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
[
5]
.
T
he
pr
opos
e
d
two
s
ta
ge
s
ol
ut
io
n,
s
e
le
c
t
th
e
m
os
t
pr
oba
bl
e
r
e
gi
ons
f
or
de
f
e
c
t
a
na
ly
s
is
in
th
e
f
ir
s
t
s
ta
ge
s
o
th
a
t
c
om
put
a
ti
ona
l
c
om
pl
e
xi
t
y
is
r
e
duc
e
d.
I
n
th
e
s
e
c
ond
s
ta
ge
,
de
e
p
le
a
r
ni
ng
ba
s
e
d
r
e
gr
e
s
s
io
n
a
na
ly
s
i
s
us
in
g
a
n
opt
im
iz
e
d
c
onvolut
io
na
l
n
e
ur
a
l
ne
twor
k
(
C
N
N
)
is
done
to
lo
c
a
li
z
e
a
nd
c
a
te
gor
iz
e
th
e
de
f
e
c
t
a
c
c
ur
a
te
ly
w
it
h
th
e
r
e
gi
ons
s
e
le
c
te
d
in
th
e
f
ir
s
t
s
ta
ge
.
F
ol
lo
w
in
g
a
r
e
th
e
nove
l
c
ont
r
ib
ut
io
ns
of
t
he
pr
opos
e
d s
ol
u
ti
on
:
−
A
nove
l
in
te
gr
a
ti
on
of
qua
te
r
ni
on
f
our
ie
r
tr
a
ns
f
or
m
ba
s
e
d
f
r
e
que
nc
y
dom
a
in
a
n
a
ly
s
is
w
it
h
s
up
e
r
pi
xe
l
s
e
gm
e
nt
a
ti
on
to
s
e
le
c
t
th
e
m
os
t
pr
oba
bl
e
de
f
e
c
t
r
e
gi
ons
in
th
e
f
a
br
ic
a
t
th
e
f
ir
s
t
s
ta
ge
is
pr
opos
e
d.
B
y
s
e
le
c
ti
ng
a
c
c
ur
a
te
de
f
e
c
t
r
e
gi
ons
,
r
e
dunda
nt
a
nd
ir
r
e
le
va
nt
r
e
gi
ons
a
r
e
is
ol
a
te
d
f
r
om
a
na
ly
s
is
c
ont
r
ib
ut
in
g
to
l
ow
e
r
f
a
ls
e
pos
it
iv
e
s
a
nd r
e
duc
e
d c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
.
−
A
nove
l
C
N
N
a
r
c
hi
te
c
tu
r
e
w
it
h
s
e
que
nc
e
of
f
ul
l
-
c
onne
c
ti
on
la
ye
r
s
w
hi
c
h
a
r
e
th
e
n
f
in
a
ll
y
br
a
nc
he
d
to
S
of
tm
a
x
c
la
s
s
if
ic
a
ti
on
a
nd
bounding
box
r
e
gr
e
s
s
io
n
to
lo
c
a
li
z
e
a
nd
c
a
te
gor
iz
e
th
e
de
f
e
c
t
in
th
e
r
e
gi
ons
s
e
le
c
te
d
in
f
ir
s
t
s
t
a
ge
i
s
pr
opos
e
d.
T
h
e
pr
opos
e
d
C
N
N
a
r
c
hi
te
c
tu
r
e
is
a
bl
e
to
le
a
r
n
m
or
e
in
tr
ic
a
te
f
e
a
tu
r
e
s
ne
c
e
s
s
a
r
y
f
or
de
f
e
c
t
lo
c
a
li
z
a
ti
on
a
nd
c
a
te
gor
iz
a
ti
on.
A
s
th
e
r
e
s
ul
t,
th
e
a
c
c
ur
a
c
y
of
de
f
e
c
t
lo
c
a
li
z
a
ti
on/
c
a
te
gor
iz
a
ti
on i
s
i
nc
r
e
a
s
e
d.
T
he
s
tr
uc
tu
r
e
of
th
is
doc
um
e
nt
is
a
s
f
ol
lo
w
s
:
th
e
c
ur
r
e
nt
te
c
hni
que
f
or
lo
c
a
li
z
in
g
f
a
br
ic
de
f
e
c
ts
is
de
s
c
r
ib
e
d
in
s
e
c
ti
on
2
.
T
he
two
-
s
ta
g
e
de
e
p
le
a
r
ni
ng
s
ol
ut
io
n
ba
s
e
d
on
a
nc
hor
in
g
is
d
e
s
c
r
ib
e
d
in
f
ul
l
in
s
e
c
ti
on
3
.
T
he
out
c
om
e
s
of
th
e
s
ugge
s
te
d
s
ol
ut
io
n
a
r
e
gi
ve
n
in
s
e
c
ti
on
4
.
T
h
e
c
on
c
lu
s
io
n
a
nd
a
r
e
a
s
f
or
a
ddi
ti
ona
l
in
ve
s
ti
ga
ti
on a
r
e
pr
e
s
e
nt
e
d
in
s
e
c
ti
on
5
.
2.
R
E
L
A
T
E
D
WORK
A
lm
e
id
a
e
t
al
.
[
6]
pr
opos
e
d
a
c
us
to
m
iz
e
d
C
N
N
a
r
c
hi
te
c
tu
r
e
a
nd
tr
a
in
e
d
w
it
h
f
a
br
ic
ne
t
da
ta
s
e
t
to
de
te
c
t
f
a
br
ic
de
f
e
c
ts
.
T
he
a
r
c
hi
te
c
tu
r
e
h
a
d
f
our
c
onvolut
io
na
l
a
nd
m
a
x
pool
la
ye
r
s
f
ol
lo
w
e
d
by
two
f
ul
ly
c
onne
c
te
d
la
y
e
r
s
.
R
e
c
ti
f
ie
d
li
ne
a
r
uni
t
(
R
e
L
U
)
w
a
s
u
s
e
d
a
s
a
c
ti
va
ti
on
f
unc
ti
on.
T
he
in
put
im
a
ge
s
a
r
e
pr
e
-
pr
oc
e
s
s
e
d
us
in
g
hi
s
to
gr
a
m
e
qua
li
z
a
ti
on
to
im
pr
ove
th
e
c
ont
r
a
s
t.
T
he
pr
e
-
pr
oc
e
s
s
e
d
im
a
ge
is
th
e
n
pa
s
s
e
d
to
C
N
N
.
T
he
m
e
th
od
w
or
ks
w
e
ll
f
or
onl
y unif
or
m
c
ol
or
s
a
nd c
a
nn
ot
w
or
k w
e
ll
f
or
c
om
pl
e
x
te
xt
ur
e
s
.
L
i
e
t
al
.
[
7]
in
te
gr
a
te
d both de
e
p f
e
a
tu
r
e
s
a
nd ha
ndc
r
a
f
te
d f
e
a
tu
r
e
s
t
o s
ol
ve
th
e
pr
obl
e
m
of
de
te
c
ti
ng de
f
e
c
ts
i
n
f
a
br
ic
w
it
h
c
om
pl
e
x
te
xt
ur
e
s
.
H
a
nd
c
r
a
f
te
r
f
e
a
tu
r
e
s
a
nd
de
e
p
f
e
a
tu
r
e
s
a
r
e
f
us
e
d
a
nd
di
m
e
n
s
io
n
r
e
duc
ti
on
is
don
e
us
in
g
pr
in
c
ip
a
l
c
om
pone
nt
a
na
ly
s
is
(
P
C
A
)
.
L
ow
r
a
nk
de
c
om
pos
it
io
n
m
ode
l
is
pr
opos
e
d
to
de
c
om
pos
e
th
e
f
e
a
tu
r
e
s
to
lo
w
r
a
nk
m
a
tr
ix
a
nd
s
p
a
r
s
e
m
a
tr
ix
r
e
pr
e
s
e
nt
in
g
th
e
ba
c
kgr
o
und
r
e
gi
on
a
nd
de
f
e
c
t
r
e
gi
on.
T
he
m
e
th
od
ha
s
hi
ghe
r
f
a
ls
e
pos
it
iv
e
s
a
nd
s
om
e
pa
tt
e
r
s
a
r
e
not
r
e
c
ogni
z
e
d
f
or
de
f
e
c
ts
.
Y
a
pi
e
t
a
l
.
[
8]
pr
opos
e
d
a
s
ig
na
tu
r
e
-
ba
s
e
d
m
e
th
od
to
de
te
c
t
f
a
br
ic
de
f
e
c
ts
.
I
m
a
ge
is
f
ir
s
t
pr
e
-
pr
oc
e
s
s
e
d
to
de
te
c
t
ba
s
ic
pa
tt
e
r
n
s
iz
e
f
or
im
a
ge
de
c
om
pos
it
io
n
in
to
bl
oc
ks
.
S
ig
na
tu
r
e
f
or
e
a
c
h
bl
oc
k
is
f
or
m
e
d
us
in
g
r
e
dunda
nt
c
ont
our
le
t
tr
a
ns
f
or
m
.
B
a
ye
s
ia
n
c
la
s
s
if
ie
r
is
tr
a
in
e
d
to
c
la
s
s
if
y
th
e
s
ig
na
t
ur
e
to
de
f
e
c
ti
ve
or
non
-
de
f
e
c
ti
ve
ba
s
e
d
on
tr
a
in
in
g
d
a
ta
s
e
t.
T
he
m
e
th
od
is
ba
s
e
d
on
non
-
ove
r
la
ppi
ng
bl
oc
k
d
e
c
om
pos
it
io
n a
nd
m
in
or
de
f
e
c
ts
a
t
bounda
r
y
of
d
e
c
om
pos
e
d
bl
oc
ks
a
r
e
not
de
te
c
te
d.
S
us
a
n
a
nd
S
ha
r
m
a
[
9]
pr
opos
e
d
an
un
s
upe
r
vi
s
e
d
m
e
th
od
f
or
f
a
br
ic
de
f
e
c
t
de
te
c
ti
on.
T
he
i
m
a
ge
i
s
s
pl
it
t
o t
e
xt
ur
a
l
pa
tc
he
s
t
hr
ough a
s
li
di
ng w
in
dow
a
ppr
oa
c
h. F
or
e
a
c
h of
t
he
pa
tc
h
e
s
, a
r
e
gul
a
r
it
y
in
de
x
ba
s
e
d
on
G
a
us
s
ia
n
ga
in
is
c
a
lc
ul
a
te
d.
T
he
r
e
gul
a
r
it
y
in
de
x
de
vi
a
ti
ng
f
r
om
th
e
s
ta
nda
r
d
de
vi
a
ti
on
a
l
l
pa
tc
he
s
a
r
e
de
t
e
c
te
d
a
s
de
f
e
c
ti
ve
pa
tc
h
e
s
.
T
he
m
e
th
od
ha
s
hi
ghe
r
f
a
ls
e
pos
it
iv
e
s
a
nd
w
or
k
s
onl
y
f
or
c
e
r
ta
in
di
s
tr
ib
ut
io
n
of
de
f
e
c
ts
.
T
ong
e
t
al
.
[
10]
pr
opos
e
d
a
de
f
e
c
t
de
te
c
ti
on
m
ode
l
ba
s
e
d
on
non
-
lo
c
a
ll
y
c
e
nt
r
a
li
z
e
d
s
pa
r
s
e
r
e
pr
e
s
e
nt
a
ti
on.
T
he
m
e
th
od
s
bui
ld
a
di
c
ti
ona
r
y
f
r
om
non
de
f
e
c
ti
ve
s
a
m
pl
e
s
.
U
s
in
g
th
e
di
c
ti
ona
r
y,
a
non
de
f
e
c
ti
ve
f
a
br
ic
pa
tt
e
r
n
is
c
ons
tr
uc
te
d
a
nd
c
om
p
a
r
e
d
to
a
f
a
br
ic
pa
tt
e
r
n
im
a
ge
to
s
pot
th
e
de
f
e
c
t.
T
he
m
e
th
od
h
a
s
hi
gh
ove
r
he
a
d a
nd w
or
ks
onl
y f
or
c
e
r
ta
in
pa
tt
e
r
ns
. Q
u
e
t
al
.
[
11]
pr
o
pos
e
d a
di
c
ti
ona
r
y
-
ba
s
e
d
f
a
br
ic
de
f
e
c
t
m
e
th
od
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
A
nc
hor
s
e
le
c
ti
on ba
s
e
d de
e
p
le
a
r
ni
ng t
w
o s
ta
ge
f
ab
r
ic
de
fe
c
t
lo
c
al
iz
at
io
n
(
H
at
ta
r
k
i
P
ooj
a
)
4713
us
in
g
dua
l
s
c
a
le
s
.
D
ic
ti
ona
r
y
i
s
le
a
nt
f
r
om
tr
a
in
in
g
im
a
ge
s
in
t
w
o
G
a
us
s
ia
n
s
c
a
le
s
.
M
a
tc
hi
ng
is
don
e
f
or
te
s
t
im
a
ge
in
two
s
c
a
le
s
to
is
ol
a
te
th
e
de
f
e
c
t.
L
ik
e
ot
he
r
di
c
ti
ona
r
y
-
ba
s
e
d
m
e
th
ods
,
th
is
m
e
th
od
w
or
ks
onl
y
f
or
c
e
r
ta
in
pa
tt
e
r
ns
.
S
hi
e
t
al
.
[
12]
p
r
opos
e
d
a
lo
w
-
r
a
nk
de
c
om
pos
it
io
n
m
ode
lt
o
s
e
gm
e
nt
th
e
de
f
e
c
ti
ve
r
e
gi
on
i
n
th
e
f
a
br
ic
. G
r
a
di
e
nt
w
e
ig
ht
m
a
tr
ix
i
s
f
ir
s
t
c
ons
tr
uc
te
d f
r
om
t
he
o
r
ig
in
a
l
im
a
ge
. T
hi
s
m
a
tr
ix
i
s
de
c
om
pos
e
d i
nt
o
lo
w
r
a
nk
pa
r
t,
s
pa
r
s
e
pa
r
t
a
nd
noi
s
e
pa
r
t.
S
a
li
e
nc
y
m
a
p
is
c
ons
t
r
u
c
te
d
f
r
om
s
pa
r
s
e
pa
r
t
a
nd
th
r
e
s
hol
de
d
to
lo
c
a
te
th
e
f
a
br
ic
de
f
e
c
t.
T
he
m
e
th
od
ha
s
hi
ghe
r
f
a
ls
e
r
a
ti
o
f
or
dot
pa
tt
e
r
ne
d
f
a
br
ic
.
H
ua
ngpe
ng
e
t
al
.
[
13]
pr
opos
e
d
an
uns
upe
r
vi
s
e
d
m
e
th
od
us
in
g
te
xt
ur
e
pr
io
r
f
or
de
te
c
ti
on
of
de
f
e
c
ts
.
A
te
xt
ur
e
pr
io
r
m
a
p
is
c
ons
tr
uc
te
d
w
it
h
hi
ghe
r
va
lu
e
in
m
a
p
in
di
c
a
ti
ng
pr
oba
bl
e
a
bnor
m
a
l
r
e
gi
ons
.
F
r
o
m
th
e
pr
io
r
m
a
p,
lo
w
r
a
nk
de
c
om
pos
it
io
n
is
done
t
o
id
e
nt
if
y
th
e
de
f
e
c
ti
ve
r
e
gi
ons
.
T
he
m
e
th
od
w
or
ks
onl
y
f
or
th
e
c
a
s
e
of
s
m
a
ll
de
f
e
c
t
s
a
nd
a
s
de
f
e
c
t
s
iz
e
in
c
r
e
a
s
e
s
,
th
e
a
c
c
ur
a
c
y
o
f
d
e
te
c
ti
on
r
e
duc
e
s
.
W
a
ng
e
t
al
.
[
14]
in
te
gr
a
te
d
m
a
nua
l
a
nd
de
e
p
le
a
r
ni
ng
lo
c
a
l
f
e
a
tu
r
e
s
to
de
te
c
t
f
a
br
ic
f
la
w
s
.
L
oc
a
l
f
e
a
tu
r
e
s
a
r
e
c
om
bi
ne
d
w
it
h
de
e
p
f
e
a
tu
r
e
s
th
a
t
w
e
r
e
e
xt
r
a
c
te
d
us
in
g
a
C
N
N
.
T
he
s
pa
r
s
e
m
a
tr
ix
is
f
ound
by
e
xe
c
ut
in
g
non
-
c
onve
x
to
ta
l
v
a
r
ia
ti
on.
A
s
a
li
e
nc
y
m
a
p
i
s
pr
oduc
e
d
f
r
om
it
.
D
e
f
e
c
t
r
e
gi
ons
a
r
e
f
ound
by
s
e
gm
e
nt
in
g
th
e
s
a
li
e
n
c
y
m
a
p.
T
hi
s
s
tr
a
te
gy
in
v
ol
ve
s
a
la
r
ge
r
c
om
put
a
ti
ona
l
bur
de
n.
S
e
ve
r
a
l
r
e
s
e
a
r
c
he
r
s
[
15
]
,
[
16]
e
m
pl
oye
d
a
m
ul
ti
s
c
a
le
C
N
N
to
id
e
nt
if
y
f
a
br
ic
f
la
w
s
.
T
o
f
in
d
th
e
f
a
ul
ts
’
,
c
lu
s
te
r
in
g
a
na
ly
s
is
i
s
pe
r
f
or
m
e
d
us
in
g
pr
e
-
known
d
e
f
e
c
t
s
iz
e
in
f
or
m
a
ti
on.
T
he
a
ppr
oa
c
h
is
not
uni
ve
r
s
a
l
f
or
a
ll
pa
tt
e
r
ns
,
but
it
doe
s
w
or
k
f
o
r
m
in
ut
e
f
la
w
s
.
M
e
i
e
t
al
.
[
17]
s
ugge
s
te
d
a
m
e
th
od
f
or
f
a
ul
t
de
te
c
ti
on
a
nd
lo
c
a
li
z
a
ti
on
ba
s
e
d
on
uns
upe
r
vi
s
e
d
le
a
r
ni
ng.
C
onvolut
io
na
l
d
e
noi
s
in
g a
ut
oe
nc
ode
r
ne
twor
ks
in
ve
s
ti
ga
t
e
f
la
w
s
a
t
e
a
c
h
r
e
s
ol
ut
io
n
w
hi
le
r
e
c
ons
tr
uc
ti
ng
th
e
pi
c
tu
r
e
a
t
s
e
ve
r
a
l
G
a
us
s
i
a
n
pyr
a
m
id
la
ye
r
s
.
T
hi
s
t
e
c
hni
que
r
e
qui
r
e
s
a
s
m
a
ll
tr
a
in
in
g
s
e
t,
but
th
e
c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
is
r
e
la
ti
ve
ly
hi
gh.
W
e
i
e
t
al
.
[
18]
us
e
d a
C
N
N
a
nd
c
om
pr
e
s
s
iv
e
s
e
ns
in
g
in
t
a
nde
m
t
o i
de
nt
if
y f
a
br
ic
f
la
w
s
.
C
om
pr
e
s
s
iv
e
s
e
ns
in
g
i
s
a
t
e
c
hni
que
f
or
e
xpa
ndi
ng t
he
t
r
a
in
in
g
s
e
t
’
s
vol
um
e
a
nd
a
ugm
e
nt
in
g
da
ta
w
it
h
ti
ny
s
a
m
pl
e
num
be
r
s
.
T
hi
s
te
c
hni
qu
e
r
a
is
e
s
C
N
N
'
s
tr
a
in
in
g
vol
um
e
,
w
hi
c
h
in
c
r
e
a
s
e
s
a
c
c
ur
a
c
y.
H
o
w
e
ve
r
,
th
e
a
c
c
ur
a
c
y
of
th
e
C
N
N
w
il
l
be
im
pa
c
te
d
if
it
is
ov
e
r
lo
a
de
d
w
it
h
tr
a
in
in
g
e
xa
m
pl
e
s
of
va
r
io
us
pa
tt
e
r
ns
.
O
uya
ng
e
t
al
.
[
19]
us
e
d
C
N
N
f
or
f
a
br
ic
de
f
e
c
t
de
te
c
ti
on.
B
a
s
e
d
on
r
a
ndom
di
s
tr
ib
ut
io
n
of
m
ot
if
in
de
f
e
c
ti
ve
f
a
br
ic
,
de
f
e
c
t
pr
oba
bi
li
ty
m
a
p
is
ge
n
e
r
a
te
d
a
nd
th
is
is
pa
s
s
e
d
a
s
in
put
to
C
N
N
f
or
s
e
gm
e
nt
a
ti
on of
de
f
e
c
ts
.
T
hi
s
m
e
th
od i
s
a
ppl
ic
a
bl
e
onl
y a
bi
gge
r
s
iz
e
de
f
e
c
t
.
J
in
g
e
t
al
.
[
20]
us
e
d
de
e
p
C
N
N
f
or
de
f
e
c
t
de
te
c
ti
on.
T
he
f
a
br
ic
im
a
ge
is
de
c
om
pos
e
d
in
to
pa
tc
he
s
a
nd
e
a
c
h
lo
c
a
l
pa
tc
h
is
la
be
le
d.
C
N
N
is
tr
a
in
e
d
w
it
h
la
b
e
le
d
p
a
tc
he
s
.
S
li
di
ng
ove
r
th
e
im
a
ge
,
e
a
c
h
pa
tc
h
w
a
s
c
la
s
s
if
ie
d
to
de
f
e
c
t
or
not
us
in
g
th
e
C
N
N
.
T
he
m
e
th
od
w
or
ks
on
ly
f
or
bi
g
s
iz
e
de
f
e
c
ts
.
W
e
i
e
t
al
.
[
21]
pr
opos
e
d
a
f
a
s
te
r
r
e
gi
ona
l
-
ba
s
e
d
c
onvolut
io
na
l
ne
twor
k
f
or
f
a
br
ic
de
f
e
c
t
d
e
te
c
ti
on.
V
is
ua
l
ge
om
e
tr
y
gr
oup
(
V
G
G
16
)
w
a
s
us
e
d
f
or
f
e
a
tu
r
e
e
xt
r
a
c
ti
on.
V
G
G
16
f
e
a
tu
r
e
s
a
r
e
e
xt
r
a
c
te
d
a
t
va
r
io
us
la
ye
r
s
a
nd
pool
i
ng
la
ye
r
is
a
da
pt
e
d
to
f
us
e
th
e
f
e
a
tu
r
e
s
.
T
he
m
e
th
od
w
or
ks
onl
y
f
or
u
ni
f
or
m
pa
tt
e
r
ns
.
Z
ha
n
g
e
t
al
.
[
22]
us
e
d
K
m
e
a
n
s
a
lg
or
it
hm
to
c
la
s
s
if
y
de
f
e
c
ti
ve
f
a
br
ic
s
.
B
a
s
e
d
on
hi
s
to
gr
a
m
,
th
e
f
a
br
ic
s
a
r
e
c
lu
s
te
r
e
d
t
o
two
c
la
s
s
of
de
f
e
c
ti
ve
a
nd
non
-
de
f
e
c
ti
ve
.
T
he
m
e
th
od
w
or
ks
o
nl
y f
or
bi
g
de
f
e
c
ts
.
P
r
io
r
i
a
nc
hor
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
P
R
A
N
-
N
e
t)
w
a
s
e
m
pl
oye
d
by
[
23]
to
de
te
c
t
f
a
br
ic
de
f
e
c
ts
.
T
he
a
ut
hor
r
e
s
e
r
ve
d
m
or
e
in
-
de
pt
h
in
f
or
m
a
ti
on
a
bout
m
in
ut
e
f
a
ul
ts
by
us
in
g
th
e
f
e
a
tu
r
e
pyr
a
m
id
ne
twor
k
(
F
P
N
)
to
p
ic
k
m
ul
ti
-
s
c
a
le
f
e
a
tu
r
e
m
a
ps
.
I
n
or
de
r
to
f
in
d
e
xt
r
e
m
e
f
la
w
s
m
or
e
pr
e
c
is
e
ly
a
nd
e
f
f
e
c
ti
ve
ly
,
th
e
a
ut
hor
s
s
ugge
s
te
d
a
m
e
th
od
f
or
ge
ne
r
a
ti
ng s
pa
r
s
e
pr
io
r
i
a
nc
hor
s
us
in
g
gr
ound
tr
ut
h
boxe
s
f
or
f
a
br
ic
de
f
e
c
ts
r
a
th
e
r
th
a
n
pe
r
m
a
ne
nt
a
nc
hor
s
.
T
o
f
in
d
f
a
br
ic
f
la
w
s
,
C
he
ng
e
t
al
.
[
24]
e
m
pl
oye
d
a
n
e
nha
n
c
e
d
ve
r
s
io
n
of
th
e
you
onl
y
lo
ok
onc
e
ve
r
s
io
n
3
(
Y
O
L
O
v3
)
a
lg
or
it
h
m
.
H
e
r
e
a
r
e
two
c
r
uc
ia
l
a
c
ti
ons
to
ta
ke
:
in
it
ia
ll
y
,
th
e
k
-
m
e
a
ns
a
lg
or
it
hm
i
s
u
s
e
d t
o i
de
nt
if
y t
he
s
iz
e
a
nd qua
nt
it
y o
f
pr
e
vi
ous
f
r
a
m
e
s
, a
nd t
he
f
a
br
ic
de
f
e
c
t
s
i
z
e
i
s
c
om
bi
ne
d
w
it
h
it
to
c
a
r
r
y
out
th
e
di
m
e
ns
io
n
c
lu
s
te
r
in
g
of
ta
r
ge
t
f
r
a
m
e
s
ba
s
e
d
on
Y
O
L
O
v3.
S
e
c
ondl
y,
th
e
hi
gh
-
le
ve
l
da
ta
is
in
te
gr
a
te
d
w
it
h
th
e
lo
w
-
le
ve
l
f
e
a
tu
r
e
s
,
a
nd
th
e
f
e
a
tu
r
e
m
a
ps
of
va
r
io
us
s
iz
e
s
a
r
e
s
uppl
e
m
e
nt
e
d
w
it
h
th
e
Y
O
L
O
d
e
te
c
ti
on
la
y
e
r
to
de
te
c
t
de
f
e
c
t
s
.
T
he
t
e
c
hni
que
is
li
m
it
e
d
to
la
tt
ic
e
a
nd
gr
a
y
c
lo
th
.
S
a
ndhya
et
al
.
[
25]
tr
a
in
e
d
A
le
xne
t
to
c
la
s
s
if
y
f
a
br
ic
im
a
ge
to
de
f
e
c
ti
ve
or
non
de
f
e
c
ti
ve
.
B
ut
th
e
a
ppr
oa
c
h
w
or
ks
onl
y
f
or
bi
gge
r
de
f
e
c
ts
.
S
e
ve
r
a
l
r
e
s
e
a
r
c
he
r
s
in
[
26
]
‒
[
28]
opt
im
iz
e
d
C
N
N
f
or
f
a
br
ic
de
f
e
c
t
de
te
c
ti
on.
V
G
G
16
m
ode
l
w
a
s
c
us
to
m
iz
e
d
f
or
num
be
r
of
c
ha
nne
ls
to
i
m
pr
ove
th
e
a
c
c
ur
a
c
y
of
de
f
e
c
t
de
te
c
ti
on.
T
he
m
e
th
od
r
e
qui
r
e
s
la
r
ge
vol
um
e
of
tr
a
in
in
g
da
ta
s
e
t
.
L
uo
e
t
al
.
[
29]
pr
opos
e
d
a
n
a
tt
e
nt
io
n
m
e
c
ha
ni
s
m
ove
r
Y
O
L
O
de
e
p
le
a
r
ni
ng
ne
twor
k
f
or
f
a
br
ic
de
f
e
c
t
de
t
e
c
ti
on.
I
n
th
is
f
a
br
ic
im
a
ge
is
pr
oc
e
s
s
e
d
a
s
w
hol
e
u
s
in
g
im
pr
ove
d
Y
O
L
O
de
e
p l
e
a
r
ni
ng mode
ls
. T
h
e
s
um
m
a
r
y of
t
he
done
i
s
pr
e
s
e
nt
e
d i
n T
a
bl
e
1.
T
hough
th
e
ne
twor
k
c
a
n
de
te
c
t
de
f
e
c
t,
it
c
a
nnot
c
a
te
gor
iz
e
th
e
de
f
e
c
t.
A
ls
o,
th
e
c
om
put
a
ti
o
n
c
om
pl
e
xi
ty
i
nc
r
e
a
s
e
s
e
xpone
nt
ia
ll
y
a
s
t
he
f
a
br
ic
di
m
e
ns
io
n i
nc
r
e
a
s
e
s
. H
ua
ng
a
nd
X
ia
ng
[
30]
pr
opos
e
d
a
C
N
N
a
r
c
hi
te
c
tu
r
e
ba
s
e
d
on
r
e
pe
a
te
d
pa
tt
e
r
n
de
te
c
ti
on
to
lo
c
a
li
z
e
f
a
br
i
c
de
f
e
c
ts
.
T
hough
th
e
m
ode
l
is
a
bl
e
to
lo
c
a
li
z
e
de
f
e
c
t
w
it
h
hi
ghe
r
a
c
c
ur
a
c
y,
it
c
a
n
lo
c
a
li
z
e
onl
y
f
or
c
e
r
ta
in
f
a
br
ic
ty
pe
s
a
nd
th
e
c
om
put
a
ti
on
c
om
pl
e
xi
ty
in
c
r
e
a
s
e
e
xpone
n
ti
a
ll
y w
it
h
i
nc
r
e
a
s
e
i
n f
a
br
ic
di
m
e
n
s
io
n. L
uo
e
t
al
.
[
31]
c
om
bi
ne
d Y
O
L
O
V
3 w
it
h de
f
or
m
a
bl
e
C
N
N
to
lo
c
a
li
z
e
f
a
br
ic
de
f
e
c
ts
.
T
he
im
a
ge
is
pr
oc
e
s
s
e
d
a
s
w
ho
le
Y
O
L
O
to
e
xt
r
a
c
t
f
e
a
tu
r
e
s
.
T
he
f
e
a
tu
r
e
s
a
r
e
th
e
n
pr
oc
e
s
s
e
d
by
de
f
or
m
a
bl
e
C
N
N
to
lo
c
a
li
z
e
th
e
de
f
e
c
t.
B
ut
a
s
th
e
f
a
br
ic
is
pr
oc
e
s
s
e
d
a
s
w
hol
e
,
th
e
c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
is
hi
ghe
r
.
F
a
ng
e
t
al
.
[
32]
pr
opos
e
d
a
C
N
N
a
r
c
hi
te
c
tu
r
e
w
it
h
a
tt
e
nt
io
n
m
e
c
h
a
ni
s
m
to
de
te
c
t
f
a
br
ic
de
f
e
c
ts
.
T
he
f
a
br
ic
im
a
ge
i
s
f
ir
s
t
f
il
le
te
d
us
in
g
f
r
e
que
nc
y
dom
a
in
f
il
te
r
in
g
a
nd
th
e
n
pr
oc
e
s
s
e
d
w
it
h
C
N
N
a
r
c
hi
te
c
tu
r
e
to
lo
c
a
li
z
e
th
e
de
f
e
c
ts
.
T
he
tr
a
in
in
g
vol
um
e
ne
e
d
f
or
de
f
e
c
t
lo
c
a
li
z
a
ti
on
ne
e
ds
to
be
ve
r
y hi
gh t
o a
c
hi
e
ve
be
tt
e
r
a
c
c
ur
a
c
y
. L
iu
e
t
al
.
[
33]
m
odi
f
ie
d t
he
Y
O
L
O
V
4 a
r
c
hi
te
c
tu
r
e
by
r
e
pl
a
c
in
g m
a
xpool
w
it
h
s
of
tp
ool
a
nd
s
e
r
ie
s
of
c
onvolut
io
na
l
la
ye
r
s
f
or
f
a
br
ic
de
f
e
c
t
de
te
c
ti
on.
T
hough
th
e
m
odi
f
ic
a
ti
on
yi
e
ld
e
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4711
-
4721
4714
6%
hi
ghe
r
a
c
c
ur
a
c
y
c
om
pa
r
e
d
to
Y
O
L
O
V
4,
th
e
c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
is
hi
ghe
r
a
s
im
a
ge
is
pr
oc
e
s
s
e
d a
s
w
hol
e
.
F
r
om
th
e
s
ur
ve
y,
it
c
a
n
be
s
e
e
n
th
a
t
m
os
t
im
por
ta
nt
p
r
obl
e
m
s
in
e
xi
s
ti
ng
w
or
ks
a
r
e
c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
,
hi
ghe
r
f
a
ls
e
pos
it
iv
e
s
a
nd
la
c
ki
ng
ge
ne
r
ic
de
f
e
c
t
de
te
c
ti
on
c
a
pa
bi
li
ty
.
T
he
pr
opos
e
d
s
ol
ut
io
n
in
th
is
w
or
k a
ddr
e
s
s
e
s
t
he
s
e
i
s
s
u
e
s
.
T
a
bl
e
1. S
ur
ve
y s
um
m
a
r
y
W
or
ks
S
ol
ut
i
on
I
s
s
ue
D
i
f
f
e
r
e
nc
e
t
o pr
opos
e
d
w
or
k
A
l
m
e
i
da
e
t
al
.
[
6]
F
a
br
i
c
i
m
a
ge
pr
oc
e
s
s
e
d
a
s
w
hol
e
w
i
t
h
m
odi
f
i
e
d
C
N
N
a
r
c
hi
t
e
c
t
ur
e
T
he
m
e
t
hod
w
or
ks
w
e
l
l
f
or
onl
y
uni
f
or
m
c
ol
or
s
a
nd
c
a
nnot
w
or
k
w
e
l
l
f
or
c
om
pl
e
x
t
e
xt
ur
e
s
T
he
pr
opos
e
d
w
or
k
s
e
l
e
c
t
s
m
os
t
pr
oba
bl
e
de
f
e
c
t
r
e
gi
ons
a
nd
pr
oc
e
s
s
e
a
c
h r
e
gi
on i
n
s
t
e
a
d
of
w
hol
e
i
m
a
ge
.
S
o
c
om
put
a
t
i
ona
l
c
om
pl
e
xi
t
y
i
s
l
ow
e
r
c
om
pa
r
e
d t
o e
xi
s
t
i
ng w
or
k
L
i
e
t
al
.
[
7]
I
nt
e
gr
a
t
e
d
bot
h
de
e
p
a
nd
ha
ndc
r
a
f
t
e
d
f
e
a
t
ur
e
s
t
o
de
t
e
c
t
f
a
br
i
c
de
f
e
c
t
s
H
a
s
hi
ghe
r
f
a
l
s
e
pos
i
t
i
ve
s
a
nd
pa
t
t
e
r
n s
pe
c
i
f
i
c
T
he
pr
opos
e
d s
ol
ut
i
on w
or
ks
f
or
a
ny pa
t
t
e
r
ns
Y
a
pi
e
t
al
.
[
8]
B
l
oc
k
ba
s
e
d
c
om
pa
r
i
s
on
u
s
i
ng
B
a
ye
s
i
a
n
c
l
a
s
s
i
f
i
e
r
m
i
nor
de
f
e
c
t
s
a
t
bounda
r
y
of
de
c
om
pos
e
d
bl
oc
ks
a
r
e
not
de
t
e
c
t
e
d
P
r
opos
e
d
s
ol
ut
i
on
doe
s
dyna
m
i
c
r
e
gi
on
-
ba
s
e
d
a
na
l
ys
i
s
i
ns
t
e
a
d
of
f
i
xe
d
bl
oc
ks
.
T
hi
s
r
e
duc
e
s
t
he
c
om
put
a
t
i
ona
l
c
om
pl
e
xi
t
y
a
nd
r
e
duc
e
s
f
a
l
s
e
pos
i
t
i
ve
s
S
us
a
n
a
nd
S
ha
r
m
a
[
9]
I
m
a
ge
s
pl
i
t
t
o f
i
xe
d s
i
z
e
pa
t
c
he
s
a
nd
pa
t
c
he
s
c
om
pa
r
e
d
u
s
i
ng
G
a
us
s
i
a
n g
a
i
n
H
a
s
hi
ghe
r
f
a
l
s
e
pos
i
t
i
ve
s
a
nd
w
or
ks
onl
y
f
or
c
e
r
t
a
i
n
di
s
t
r
i
but
i
on of
de
f
e
c
t
s
P
r
opos
e
d
s
ol
ut
i
on
i
s
i
nde
pe
nd
e
nt
of
de
f
e
c
t
di
s
t
r
i
but
i
on
T
ong
e
t
al
.
[
10]
B
ui
l
t
di
c
t
i
ona
r
y
f
r
om
non
de
f
e
c
t
i
ve
s
a
m
pl
e
s
a
nd
c
om
pa
r
i
s
on i
s
done
H
a
s
hi
ghe
r
f
a
l
s
e
pos
i
t
i
ve
s
a
nd
s
uf
f
e
r
s
f
r
om
z
e
r
o
-
da
y
pr
obl
e
m
T
he
pr
opos
e
d
s
ol
ut
i
on
i
s
not
a
f
f
e
c
t
e
d
f
r
om
z
e
r
o
-
da
y
pr
obl
e
m
.
Q
u
e
t
al
.
[
11]
D
i
c
t
i
ona
r
y
ba
s
e
d
f
a
br
i
c
de
f
e
c
t
m
e
t
hod us
i
ng
dua
l
s
c
a
l
e
s
.
I
t
w
or
ks
onl
y
f
or
c
e
r
t
a
i
n
pa
t
t
e
r
ns
T
he
pr
opos
e
d s
ol
ut
i
on i
s
pa
t
t
e
r
n i
nde
pe
nde
nt
H
ua
ngpe
ng
e
t
al
.
[
13]
T
e
xt
ur
e
ba
s
e
d a
n
a
l
ys
i
s
t
o de
t
e
c
t
de
f
e
c
t
W
or
ks
onl
y
f
or
c
e
r
t
a
i
n
de
f
e
c
t
di
m
e
ns
i
ons
F
r
e
que
nc
y
ba
s
e
d
a
na
l
ys
i
s
a
nd
w
or
ks
f
or
a
ny
de
f
e
c
t
di
m
e
ns
i
ons
O
uya
ng
e
t
al
.
[
19]
M
ot
i
f
f
ba
s
e
d
de
f
e
c
t
m
a
p
c
ons
t
r
uc
t
i
on
a
nd
pr
oc
e
s
s
i
ng
w
i
t
h C
N
N
t
o l
oc
a
l
i
z
e
de
f
e
c
t
W
or
ks
onl
y
f
or
c
e
r
t
a
i
n
de
f
e
c
t
di
m
e
ns
i
ons
F
r
e
que
nc
y
ba
s
e
d
de
f
e
c
t
m
a
p
c
ons
t
r
uc
t
i
on
a
nd
pr
oc
e
s
s
i
ng
w
i
t
h
C
N
N
t
o
w
or
k
ge
n
e
r
i
c
f
or
a
ny
de
f
e
c
t
di
m
e
ns
i
ons
S
ol
ut
i
ons
[
29]
–
[
33]
Y
O
L
O
ba
s
e
d
pr
oc
e
s
s
i
ng
of
w
hol
e
i
m
a
ge
a
t
onc
e
t
o
l
oc
a
l
i
z
e
t
he
de
f
e
c
t
C
om
put
a
t
i
ona
l
c
om
pl
e
xi
t
y
a
nd
f
a
l
s
e
pos
i
t
i
ve
s
a
r
e
hi
ghe
r
D
e
f
e
c
t
a
na
l
ys
i
s
i
s
done
onl
y
on
pr
oba
bl
e
de
f
e
c
t
r
e
gi
ons
r
e
duc
i
ng
t
he
c
om
put
a
t
i
ona
l
c
om
pl
e
xi
t
y a
nd f
a
l
s
e
pos
i
t
i
ve
s
3.
P
R
O
P
O
S
E
D
M
E
T
H
O
D
O
L
O
G
Y
F
r
om
th
e
s
ur
ve
y,
it
c
a
n
be
s
e
e
n
th
a
t
de
e
p
le
a
r
ni
ng
m
e
th
ods
pr
ovi
de
be
tt
e
r
lo
c
a
li
z
a
ti
on
of
f
a
br
ic
de
f
e
c
ts
but
t
he
ir
c
om
put
a
ti
on c
om
pl
e
xi
ty
i
s
hi
ghe
r
. A
nc
hor
ba
s
e
d a
ppr
oa
c
he
s
r
e
duc
e
d t
he
c
om
put
a
ti
on c
om
pl
e
xi
ty
of
de
e
p
le
a
r
ni
ng
a
ppr
oa
c
he
s
.
B
ut
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
a
nc
hor
-
ba
s
e
d
a
ppr
oa
c
he
s
c
a
n
be
f
ur
th
e
r
im
pr
ove
d
by
s
e
le
c
ti
ng
th
e
be
s
t
a
nc
hor
s
.
M
ot
iv
a
te
d
by
th
is
ob
s
e
r
va
ti
on,
th
is
w
or
k
pr
opos
e
s
a
de
e
p
le
a
r
ni
ng
-
ba
s
e
d
a
nc
hor
s
e
le
c
ti
on
a
ppr
oa
c
h
(
A
S
A
)
a
nd
in
te
gr
a
te
s
w
it
h
de
e
p
le
a
r
ni
ng
-
ba
s
e
d
f
a
br
ic
de
f
e
c
t
lo
c
a
li
z
a
ti
on
f
or
im
pr
ove
d
lo
c
a
li
z
a
ti
on
a
c
c
ur
a
c
y
a
nd
r
e
duc
e
d
c
om
put
a
ti
ona
l
ove
r
he
a
d.
T
he
pr
opos
e
d
w
or
k
a
dopt
s
th
e
f
a
br
ic
d
e
f
e
c
t
de
te
c
ti
on
pr
oc
e
s
s
of
P
R
A
N
-
N
e
t.
I
n
P
R
A
N
-
N
e
t,
f
e
a
tu
r
e
s
w
e
r
e
e
xt
r
a
c
te
d
a
t
va
r
io
us
s
c
a
le
s
us
in
g
R
e
s
n
e
t
-
101
w
he
r
e
a
bout
12 t
e
xt
ur
e
f
e
a
tu
r
e
s
a
r
e
e
xt
r
a
c
te
d
a
nd t
he
n
t
he
P
r
io
r
i
a
nc
hor
s
a
r
e
ge
ne
r
a
te
d
i
n e
a
c
h s
c
a
le
of
f
e
a
tu
r
e
m
a
p
a
s
de
f
e
c
t
pr
opos
a
ls
.
T
h
e
de
f
e
c
t
pr
opos
a
l
s
a
r
e
th
e
n
c
la
s
s
if
ie
d
to
de
f
e
c
t
or
no
de
f
e
c
t
by
th
e
c
la
s
s
if
ic
a
ti
on
ne
twor
k a
s
s
how
n i
n F
ig
ur
e
1. M
ul
ti
s
c
a
le
f
e
a
tu
r
e
m
a
p ge
ne
r
a
ti
on us
in
g R
e
s
ne
t
-
101 pr
opos
e
d i
n P
R
A
N
-
ne
t
is
us
e
d
in
th
is
w
or
k.
T
he
th
r
e
e
f
e
a
tu
r
e
m
a
ps
of
C
la
s
s
1,
C
la
s
s
2
,
a
nd
C
la
s
s
3
w
hi
c
h
ha
s
e
nough
in
f
or
m
a
ti
on
f
or
ti
ny de
f
e
c
t
de
te
c
ti
on i
s
us
e
d i
n t
hi
s
w
or
k.
A
lg
or
it
hm
1 s
how
s
t
h
e
de
te
c
t
de
f
e
c
t
us
in
g
A
S
A
.
F
ig
ur
e
.1 S
ta
ge
s
i
n P
R
A
N
-
N
e
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
A
nc
hor
s
e
le
c
ti
on ba
s
e
d de
e
p
le
a
r
ni
ng t
w
o s
ta
ge
f
ab
r
ic
de
fe
c
t
lo
c
al
iz
at
io
n
(
H
at
ta
r
k
i
P
ooj
a
)
4715
A
lg
or
it
hm
1
.
D
e
te
c
t
de
f
e
c
t
us
in
g
A
S
A
I
nput
F
a
br
ic
i
m
a
ge
s
(
D
e
f
e
c
ti
ve
a
nd nor
m
a
l
f
a
br
ic
)
O
ut
put
D
e
f
e
c
t
ty
pe
(
T
in
y a
nd B
ig
de
f
e
c
ts
)
S
te
p 1
ti
H
S
V
t
r
a
ns
f
or
m
(
im
a
ge
)
S
te
p 2
qv Q
D
C
T
(
ti
)
S
te
p 3
G
qv
G
a
us
s
ia
n_S
m
oot
h(
qv)
;
S
te
p 4
s
i
I
nve
r
s
e
Q
D
C
T
(
G
qv)
;
S
te
p 5
f
or
e
a
c
h s
e
gm
e
nt
i
n s
i
C
a
lc
ul
a
te
f
or
e
a
c
h s
e
gm
e
nt
If
>
0.7
K
e
e
p s
e
gm
e
nt
E
ls
e
M
a
s
k
s
e
gm
e
nt
S
te
p 6
f
s
i
E
xt
r
a
c
t
P
R
N
e
t
f
e
a
tu
r
e
s
(
s
i)
S
te
p 7
de
f
e
c
tt
ype
C
la
s
s
if
y f
s
i
us
in
g s
of
tm
a
x
S
te
p 8
r
e
tu
r
n de
f
e
c
tt
ype
P
R
A
N
-
N
e
t
s
e
le
c
te
d
a
nc
hor
s
us
in
g a
lo
c
a
ti
on
pr
e
di
c
ti
on
ne
twor
k
w
hi
c
h
ope
r
a
te
s
a
t
le
ve
l
of
pi
xe
ls
.
A
1×
1
c
onvolut
io
na
l
la
ye
r
is
us
e
d
to
c
r
e
a
te
th
e
f
a
ul
ty
s
c
or
e
m
a
p
by
s
e
m
a
nt
ic
s
e
gm
e
nt
a
ti
on.
T
h
e
de
f
e
c
t
pr
oba
bi
li
ty
m
a
p
is
th
e
n
c
r
e
a
te
d
by
tr
a
ns
f
or
m
in
g
th
e
de
f
e
c
ti
ve
s
c
or
e
s
u
s
in
g a
n
e
le
m
e
nt
-
w
is
e
d s
ig
m
oi
d
f
unc
ti
on,
a
nd
it
ha
s
th
e
s
a
m
e
di
m
e
ns
io
ns
a
s
th
e
f
e
a
tu
r
e
m
a
p.
D
e
f
e
c
t
pi
xe
ls
a
r
e
th
os
e
w
hos
e
va
lu
e
s
a
bove
th
e
th
r
e
s
hol
d
va
lu
e
Δ
L
,
w
he
r
e
Δ
L
is
th
e
pr
e
s
e
t
th
r
e
s
hol
d
of
100
f
or
de
te
r
m
in
in
g
th
e
va
lu
e
of
a
pi
xe
l
in
a
f
a
u
lt
y
r
e
gi
on.
T
hi
s
r
e
s
ul
ts
in
la
r
ge
num
be
r
of
a
nc
hor
r
e
gi
ons
in
th
e
P
R
A
N
-
N
e
t
a
nd
a
ls
o
de
f
e
c
ti
ve
pr
opos
a
ls
a
r
e
m
a
d
e
ba
s
e
d
on
f
ix
e
d
s
iz
e
w
he
r
e
s
iz
e
i
nf
or
m
a
ti
on f
or
e
a
c
h de
f
e
c
t
m
us
t
be
known pr
io
r
i.
D
e
vi
a
ti
ng
f
r
om
P
R
A
N
-
N
e
t,
th
is
w
or
k
pr
opos
e
s
an
a
nc
hor
poi
nt
a
nd
de
f
e
c
ti
ve
pr
opos
a
l
r
e
gi
on
s
e
le
c
ti
on
ba
s
e
d
on
f
r
e
que
nc
y
dom
a
in
a
n
a
ly
s
is
c
om
bi
ne
d
w
it
h
s
upe
r
pi
xe
l
-
ba
s
e
d
s
e
le
c
ti
on.
T
he
s
ugge
s
te
d
a
ppr
oa
c
h
us
e
s
in
put
phot
ogr
a
phs
th
a
t
ha
v
e
be
e
n
s
ubs
a
m
pl
e
d
to
ha
lf
of
th
e
ir
or
ig
in
a
l
s
iz
e
be
c
a
u
s
e
pr
oc
e
s
s
in
g
th
e
or
ig
in
a
l
s
iz
e
of
th
e
im
a
ge
s
w
il
l
r
e
s
ul
t
in
hi
ghe
r
c
om
put
a
ti
ona
l
c
os
ts
. T
he
s
ubs
a
m
pl
e
d
r
e
d,
gr
e
e
n,
a
nd
bl
ue
(
R
G
B
)
i
m
a
ge
i
s
s
ubj
e
c
te
d t
o t
he
H
S
I
t
r
a
ns
f
or
m
to
pr
oduc
e
hue
, s
a
tu
r
a
ti
on, a
nd i
nt
e
ns
it
y
(
H
S
I
)
i
m
a
ge
s
.
U
s
e
of
H
S
I
in
f
r
e
que
nc
y
dom
a
in
a
na
ly
s
is
is
pr
e
f
e
r
r
e
d
ove
r
R
G
B
c
ol
or
s
pa
c
e
be
c
a
us
e
of
it
i
s
gr
e
a
te
r
c
onf
or
m
it
y
w
it
h
th
e
hum
a
n
vi
s
ua
l
s
ys
te
m
.
P
r
io
r
to
be
in
g
c
onv
e
r
te
d
to
H
S
I
,
th
e
c
o
lo
r
s
in
R
G
B
a
r
e
nor
m
a
li
z
e
d
w
it
hi
n
th
e
in
te
r
va
l
0
to
1.
w
he
r
e
th
e
nor
m
a
li
z
e
d
num
be
r
s
f
or
th
e
pi
xe
l
a
r
e
R
’
,
G
’
,
a
nd
B
’
.
U
s
in
g
R
G
B
c
ol
or
im
a
ge
s
,
th
e
H
S
I
tr
a
ns
f
or
m
i
s
a
ppl
ie
d a
s
pe
r
(
5)
a
nd (
6)
.
=
+
+
(
1)
′
=
∗
255
(
2)
′
=
∗
255
(
3)
′
=
∗
255
(
4)
=
{
360
−
,
≤
,
>
(
5)
=
1
−
3
(
+
+
[
m
in
(
,
,
)
]
(
6)
=
1
3
(
+
+
)
(
7)
=
{
(
(
1
2
)
)
[
(
−
)
+
(
−
)
]
/
(
[
(
−
)
(
−
)
1
2
)
}
(
8)
(
,
)
=
(
,
)
1
+
(
,
)
2
+
(
,
(
,
)
=
(
,
)
1
+
(
,
)
2
(
,
)
=
(
,
)
1
+
(
,
)
2
(
,
)
=
(
,
)
1
+
(
,
)
2
+
(
,
(
9)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4711
-
4721
4716
w
he
r
e
R
,
G
,
a
nd
B
r
e
pr
e
s
e
nt
th
e
im
a
ge
’
s
r
e
d,
gr
e
e
n,
a
nd
bl
u
e
c
om
pone
nt
'
s
pi
x
e
l
va
lu
e
s
.
In
(
9)
pr
ovi
de
s
a
qua
te
r
ni
on r
e
pr
e
s
e
nt
a
ti
on of
t
he
H
S
I
i
m
a
ge
. w
he
r
e
t
h
e
pi
xe
l
’
s
p
os
it
io
n i
s
(
n,
m
)
a
r
e
t
h
e
pi
xe
l
’
s
hue
,
s
a
tu
r
a
ti
on,
a
nd
in
te
ns
it
y
a
t
(
n,
m
)
.
T
he
c
ondi
ti
on
de
te
r
m
in
e
s
th
e
va
lu
e
of
μ
.
T
he
pi
xe
l
a
t
(
u,
v
)
unde
r
goe
s
th
e
qua
te
r
ni
on
F
our
ie
r
t
r
a
ns
f
or
m
, w
it
h
F
1 a
nd F
2 r
e
pr
e
s
e
nt
in
g t
he
pi
xe
l'
s
f
our
i
e
r
de
r
iv
a
ti
ve
f
unc
ti
on, a
s
i
ndi
c
a
te
d by (
10)
.
(
,
)
=
1
(
,
)
+
1
(
,
)
2
(
10)
1
(
,
)
1
√
∑
∑
−
1
=
0
−
1
=
0
−
1
2
(
(
)
+
1
(
,
)
(
11)
A
n
im
a
ge
’
s
bor
de
r
s
a
nd
ot
he
r
a
br
upt
pi
xe
l
c
ha
nge
s
gr
e
a
tl
y
in
c
r
e
a
s
e
th
e
hi
gh
-
f
r
e
que
nc
y
c
om
pone
nt
of
th
e
f
our
ie
r
tr
a
ns
f
or
m
.
T
he
r
e
f
or
e
,
a
tt
e
nua
ti
ng
th
e
de
s
ig
na
te
d
r
a
n
ge
of
hi
gh
-
f
r
e
que
nc
y
e
le
m
e
nt
s
in
th
e
im
a
g
e
’
s
qua
te
r
ni
on
f
our
ie
r
tr
a
ns
f
or
m
a
ll
ow
s
f
or
f
r
e
que
nc
y
dom
a
in
s
m
oot
hi
ng
of
th
e
im
a
ge
.
T
he
qu
a
te
r
ni
on
f
our
ie
r
c
onve
r
te
d
da
ta
a
r
e
th
e
n
s
e
nt
vi
a
a
ga
us
s
ia
n
qua
te
r
ni
on
hi
gh
pa
s
s
f
il
te
r
.
In
(
12)
pr
ovi
de
s
th
e
two
-
di
m
e
ns
io
na
l
G
a
us
s
ia
n
hi
gh
pa
s
s
f
il
te
r
tr
a
ns
f
e
r
f
unc
ti
on
w
it
h
th
e
c
ut
-
o
f
f
f
r
e
que
nc
y
a
t
a
di
s
ta
nc
e
D
0.
w
he
r
e
D
(
u,
v
)
is
th
e
di
s
ta
nc
e
be
tw
e
e
n
poi
nt
(
u,
v)
a
nd
th
e
or
ig
in
of
th
e
f
r
e
que
nc
y
r
e
c
ta
ngl
e
(
M
/2
,
N
/2
)
w
he
r
e
M
a
nd
N
a
r
e
th
e
r
e
c
ta
ngl
e
’
s
di
m
e
ns
io
n
s
a
nd
M
a
nd
N
s
ta
nd
f
or
th
e
r
e
c
ta
ngl
e
’
s
le
ngt
h
a
nd
br
e
a
dt
h,
r
e
s
p
e
c
ti
ve
ly
,
a
nd
σ
i
s
a
m
e
a
s
ur
e
of
G
a
us
s
ia
n di
s
pe
r
s
io
n. I
t
is
c
om
put
e
d u
s
in
g t
he
f
or
m
ul
a
i
n (
13)
.
(
,
)
(
12)
(
,
)
=
[
(
−
2
)
2
(
13)
(
,
)
=
1
√
∑
∑
−
1
=
0
−
1
=
0
−
1
2
(
(
)
+
1
(
(
,
)
=
1
√
∑
∑
−
1
=
0
−
1
=
0
−
1
2
(
(
)
+
(
,
)
=
1
√
∑
∑
−
1
=
0
−
1
=
0
−
1
2
(
(
)
+
(
14)
C
om
pa
r
e
d
to
id
e
a
l
hi
gh
pa
s
s
f
il
te
r
s
,
th
e
out
c
om
e
s
of
G
a
us
s
ia
n
hi
gh
pa
s
s
f
il
te
r
in
g
a
r
e
s
m
oot
he
r
.
A
n
out
put
im
a
ge
is
pr
oduc
e
d
us
in
g
th
e
in
ve
r
s
e
qu
a
te
r
ni
on
tr
a
ns
f
o
r
m
a
f
te
r
G
a
us
s
ia
n
hi
gh
pa
s
s
f
il
te
r
in
g.
I
n
(
14)
de
s
c
r
ib
e
s
ho
w
to
pe
r
f
or
m
th
e
in
ve
r
s
e
tr
a
ns
f
or
m
.
T
he
f
r
e
que
nc
y
dom
a
in
m
a
p
w
it
h
hi
ghl
ig
ht
e
d
s
a
li
e
nt
r
e
gi
on
s
is
th
e
out
put
im
a
ge
th
a
t
is
pr
oduc
e
d
f
ol
lo
w
in
g
th
e
in
ve
r
s
e
tr
a
ns
f
or
m
.
S
upe
r
pi
xe
ls
a
r
e
c
ol
le
c
ti
ons
of
s
pa
ti
a
ll
y
r
e
la
te
d
pi
xe
ls
w
it
h
c
om
pa
r
a
bl
e
c
o
lo
r
or
in
te
n
s
it
y
c
ha
r
a
c
t
e
r
is
ti
c
s
.
I
m
a
ge
s
a
r
e
di
vi
de
d
i
nt
o
s
e
gm
e
nt
s
w
it
h
th
e
ir
n
a
tu
r
a
l
bor
de
r
s
r
e
t
a
i
ne
d
u
s
i
ng
s
u
pe
r
pi
xe
l
s
e
g
m
e
nt
a
ti
on.
T
he
m
o
s
t
w
id
e
ly
u
s
e
d
t
e
c
hni
qu
e
f
or
s
up
e
r
p
ix
e
l
s
e
gm
e
n
ta
t
io
n
i
n
t
hi
s
s
t
ud
y
i
s
s
im
p
le
li
ne
a
r
it
e
r
a
ti
ve
c
lu
s
t
e
r
i
ng
(
S
L
I
C
)
.
S
a
li
e
nt
a
r
e
a
s
e
gm
e
n
t
s
’
s
u
pe
r
pi
x
e
l
s
e
gm
e
n
t
s
ha
v
e
a
d
is
c
e
r
ni
bl
e
c
ol
or
c
ont
r
a
s
t
w
it
h
o
th
e
r
s
e
gm
e
nt
s
,
a
nd
t
he
ir
s
p
a
ti
a
l
di
s
tr
ib
u
ti
o
n
i
s
s
p
a
r
s
e
r
th
a
n
th
a
t
of
ot
he
r
s
e
gm
e
n
ts
.
B
a
s
e
d
o
n
t
he
s
e
t
w
o
o
b
s
e
r
va
ti
o
n
s
,
e
a
c
h
of
s
a
li
e
nt
r
e
g
i
on
i
de
nt
if
ie
d
in
f
r
e
qu
e
n
c
y
d
om
a
in
a
na
l
y
s
i
s
,
a
r
e
e
v
a
lu
a
t
e
d
i
n
t
e
r
m
s
of
c
on
tr
a
c
t
w
it
h
o
th
e
r
s
e
gm
e
nt
s
to
d
e
c
i
de
pr
ob
a
b
il
it
y
of
it
b
e
in
g
a
d
e
f
e
c
t
iv
e
r
e
g
io
n
.
T
he
pr
ob
a
b
il
it
y
of
a
s
a
li
e
n
t
r
e
gi
on
t
o
a
n
a
n
c
h
or
r
e
g
io
n
or
d
e
f
e
c
t
pr
opo
s
a
l
i
s
f
o
und
a
s
gi
ve
n i
n (
15)
.
=
1
(
15)
W
he
r
e
=
∑
(
,
)
.
|
|
−
|
|
=
1
=
∑
(
,
)
.
|
|
−
|
|
=
1
(
16)
(
,
)
=
[
|
(
,
)
(
17)
T
he
w
e
ig
ht
W
i
s
c
a
lc
ul
a
te
d
ba
s
e
d
on
th
e
a
r
e
a
of
s
upe
r
pi
xe
l
s
e
gm
e
nt
a
nd
m
c
is
th
e
s
a
li
e
nt
r
e
gi
on
th
e
s
pa
ti
a
l
s
im
il
a
r
it
y
is
gi
ve
n
a
s
s
how
n
in
(
17)
.
W
he
n
i
s
gr
e
a
t
e
r
th
a
n
th
e
th
r
e
s
hol
d,
th
e
s
a
li
e
nt
r
e
gi
on
is
c
hos
e
n
a
s
th
e
a
nc
hor
r
e
gi
on
a
f
te
r
a
c
om
pa
r
is
on
is
m
a
de
u
s
in
g
a
th
r
e
s
hol
d
of
0.
8
f
or
r
e
s
e
m
bl
a
nc
e
of
th
e
r
e
gi
on
to
th
e
pr
oba
bl
e
r
e
gi
on.
A
R
O
I
a
li
gn
la
ye
r
r
e
s
iz
e
s
de
f
e
c
t
s
ugge
s
ti
on
s
to
th
e
f
ix
e
d
s
iz
e
a
nd
th
e
f
e
a
tu
r
e
m
a
ps
th
a
t
go
w
it
h
th
e
m
.
F
ol
lo
w
in
g t
ha
t,
t
he
f
e
a
tu
r
e
s
a
r
e
pa
s
s
e
d vi
a
a
s
e
r
ie
s
of
f
ul
l
-
c
onn
e
c
ti
o
n l
a
ye
r
s
be
f
or
e
br
a
nc
hi
ng out t
o
bounding
box r
e
gr
e
s
s
io
n a
nd c
la
s
s
if
ic
a
ti
on. T
h
e
r
e
gr
e
s
s
io
n br
a
nc
h pr
e
di
c
t
s
t
he
pos
it
io
n of
t
he
de
f
e
c
t
a
f
te
r
t
he
s
e
c
ond
a
r
y
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
A
nc
hor
s
e
le
c
ti
on ba
s
e
d de
e
p
le
a
r
ni
ng t
w
o s
ta
ge
f
ab
r
ic
de
fe
c
t
lo
c
al
iz
at
io
n
(
H
at
ta
r
k
i
P
ooj
a
)
4717
lo
c
a
ti
on,
w
hi
le
th
e
c
l
a
s
s
if
ic
a
ti
on
br
a
n
c
h
pr
e
di
c
ts
th
e
ty
pe
of
th
e
d
e
f
e
c
t.
I
n
th
is
w
or
k,
th
e
c
la
s
s
if
ie
r
is
th
e
S
of
tm
a
x
c
la
s
s
if
ie
r
.
T
h
e
li
ke
li
hood
of
e
a
c
h
f
a
ul
t
a
s
w
e
ll
a
s
th
e
a
b
s
e
nc
e
of
a
ny
de
f
e
c
ts
i
s
e
s
ti
m
a
t
e
d
by
th
e
s
of
tm
a
x
c
la
s
s
if
ie
r
.
I
n
a
r
e
gr
e
s
s
io
n
s
it
ua
ti
on,
th
e
S
of
tm
a
x
c
la
s
s
if
ie
r
is
e
m
pl
oye
d.
A
s
s
um
e
th
a
t
th
e
s
of
tm
a
x
c
la
s
s
i
f
ie
r
ne
e
ds
to
e
s
ti
m
a
te
th
e
pr
oba
bi
li
ty
f
or
e
a
c
h
of
th
e
K
va
lu
e
s
unde
r
th
e
f
ol
lo
w
in
g
s
c
e
na
r
io
s
:
{
1
,
2
,
.
.
.
}
.
T
he
K
di
m
e
ns
io
na
l
ve
c
to
r
c
ont
a
in
in
g
th
e
K
e
s
ti
m
a
te
d
pr
oba
bi
li
ti
e
s
i
s
th
e
r
e
s
ul
t
of
th
e
s
of
tm
a
x
c
la
s
s
if
ie
r
.
I
n
(
18
)
pr
ovi
de
s
t
he
l
os
s
f
unc
ti
on t
ha
t
w
il
l
be
us
e
d t
o t
r
a
in
t
he
s
of
tm
a
x r
e
gr
e
s
s
io
n c
la
s
s
if
ie
r
.
=
−
[
∑
∑
1
{
(
)
=
}
l
o
g
1
−
0
=
1
(
(
)
=
|
(
)
;
]
(
18)
(
(
)
=
|
(
)
;
)
=
e
x
p
(
(
)
(
)
)
∑
=
1
(
(
)
(
)
)
(
19)
4.
E
X
P
E
R
I
M
E
N
T
A
L
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
T
he
pr
opos
e
d s
ol
ut
io
n i
s
i
m
pl
e
m
e
nt
e
d i
n
M
A
T
L
A
B
2019b a
nd t
he
pe
r
f
or
m
a
nc
e
i
s
t
e
s
te
d i
n m
a
c
hi
ne
w
it
h
c
onf
ig
ur
a
ti
on
of
:
I
nt
e
l
i5
c
e
nt
r
a
l
pr
oc
e
s
s
in
g
uni
t
(
C
P
U
)
,
8
G
B
r
a
ndom
a
c
c
e
s
s
m
e
m
or
y
(
R
A
M
)
a
nd
W
in
dow
s
10
O
S
.
A
da
ta
s
e
t
of
f
a
br
ic
de
f
e
c
t
de
f
e
c
ti
on
is
c
ons
tr
uc
te
d
w
it
h
two
da
ta
s
e
ts
of
te
xt
il
e
de
f
e
c
t
de
te
c
ti
on
[
34]
a
nd
th
e
s
ta
nda
r
d
T
I
L
D
A
da
ta
s
e
t.
T
he
da
ta
s
e
t
ha
s
im
a
ge
s
in
to
ta
l
s
ix
c
la
s
s
e
s
:
good,
c
ol
or
,
c
ut
,
hol
e
,
th
r
e
a
d,
m
e
ta
l_
c
ont
a
m
in
a
ti
on
(
M
C
)
.
T
he
da
ta
s
e
t
ha
s
m
ix
of
bot
h
ti
ny
a
n
d
bi
gge
r
de
f
e
c
ts
.
T
he
pe
r
f
or
m
a
n
c
e
is
m
e
a
s
ur
e
d
in
te
r
m
s
of
a
c
c
ur
a
c
y,
pr
e
c
is
io
n,
r
e
c
a
ll
,
lo
c
a
li
z
a
ti
on
e
r
r
or
,
a
nd
de
f
e
c
t
de
te
c
ti
on
ti
m
e
.
T
he
pe
r
f
or
m
a
nc
e
is
m
e
a
s
ur
e
d f
or
e
a
c
h c
la
s
s
of
de
f
e
c
ts
a
nd f
or
bot
h t
in
y
a
nd bigge
r
de
f
e
c
ts
. T
he
pe
r
f
or
m
a
nc
e
of
pr
opos
e
d A
S
A
i
s
c
om
pa
r
e
d
a
ga
in
s
t
P
R
A
N
-
N
e
t
[
23]
,
m
ul
ti
s
c
a
le
C
N
N
[
16]
a
nd
f
a
s
te
r
r
e
gi
on
-
ba
s
e
d
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
(
R
C
N
N
)
[
21]
.
T
he
c
om
pa
r
is
on
of
e
xi
s
ti
ng
th
r
e
e
s
ol
ut
io
ns
w
it
h
th
e
pr
opos
e
d
w
or
k
is
pr
e
s
e
nt
e
d
in
T
a
bl
e
2. F
ig
ur
e
2 s
how
s
t
h
e
c
om
pa
r
is
on of
pe
r
f
or
m
a
nc
e
i
n t
e
r
m
s
of
a
c
c
ur
a
c
y, pr
e
c
i
s
io
n
,
a
nd r
e
c
a
ll
.
T
a
bl
e
2. P
e
r
f
or
m
a
nc
e
c
om
pa
r
is
on
M
e
t
hod
A
c
c
ur
a
c
y
P
r
e
c
i
s
i
on
R
e
c
a
l
l
A
S
A
93.1
64.5
55.3
P
R
A
N
-
N
e
t
91.9
62.3
53.3
M
ul
t
i
s
c
a
l
e
-
C
N
N
89.4
58.4
51
F
a
s
t
e
r
R
C
N
N
87.1
60.2
50.9
F
ig
ur
e
2
.
C
om
pa
r
is
on of
pe
r
f
or
m
a
nc
e
i
n t
e
r
m
s
of
a
c
c
ur
a
c
y, pr
e
c
is
io
n
,
a
nd r
e
c
a
ll
T
he
pr
opos
e
d
s
ol
ut
io
n
ha
s
a
tl
e
a
s
t
1.2%
hi
ghe
r
a
c
c
ur
a
c
y
c
o
m
pa
r
e
d
to
P
R
A
N
-
N
e
t,
3.7%
hi
ghe
r
a
c
c
ur
a
c
y
c
om
pa
r
e
d
to
m
ul
ti
s
c
a
le
C
N
N
,
6%
hi
ghe
r
a
c
c
ur
a
c
y
c
o
m
pa
r
e
d
to
f
a
s
te
r
R
C
N
N
a
s
s
ho
w
n
in
F
ig
ur
e
2.
A
S
A
pe
r
f
or
m
s
be
tt
e
r
w
he
n
c
om
pa
r
e
d
to
th
e
e
xi
s
ti
ng
th
r
e
e
m
e
th
ods
be
c
a
us
e
of
th
e
f
e
a
tu
r
e
m
a
ps
th
a
t
a
r
e
e
xt
r
a
c
te
d
a
t
va
r
io
us
s
c
a
le
s
u
s
in
g
R
e
s
ne
t
-
101.
T
he
a
c
c
ur
a
c
y
h
a
s
in
c
r
e
a
s
e
d
in
th
e
pr
opo
s
e
d
s
ol
ut
io
n
du
e
to
s
e
le
c
ti
on
of
a
nc
hor
r
e
gi
ons
w
it
h
hi
ghe
r
p
r
oba
bi
li
ty
o
f
de
f
e
c
t
lo
c
a
li
z
e
d
ne
a
r
it
.
T
hi
s
r
e
duc
e
d
th
e
f
a
ls
e
pos
it
iv
e
po
s
it
iv
e
s
in
th
e
pr
opos
e
d
s
ol
ut
io
n.
T
h
e
num
be
r
of
a
nc
hor
r
e
gi
ons
f
ound
P
R
A
N
-
N
e
t
is
ve
r
y
hi
ghe
r
c
om
pa
r
e
d
to
pr
opos
e
d
s
ol
ut
io
n
a
nd
m
a
ny
of
r
e
dunda
nt
or
ir
r
e
le
va
nt
a
nc
hor
p
oi
nt
s
c
ont
r
ib
ut
e
d
to
hi
ghe
r
f
a
ls
e
pos
it
iv
e
s
.
D
ue
to
th
is
P
R
A
N
-
N
e
t
a
c
c
ur
a
c
y
is
lo
w
e
r
c
om
pa
r
e
d
to
pr
opos
e
d
s
ol
ut
io
n.
F
a
s
te
r
R
C
N
N
m
is
s
e
s
out
m
a
ny
s
m
a
ll
de
f
e
c
ts
to
s
pe
e
d
up
in
de
f
e
c
t
de
te
c
ti
on
s
ta
ge
a
nd
th
is
ha
s
r
e
duc
e
d
th
e
a
c
c
ur
a
c
y.
M
ul
ti
s
c
a
le
a
ppr
oa
c
h
do
e
s
a
na
ly
s
is
a
t
va
r
io
us
s
c
a
le
s
but
t
he
s
m
a
ll
de
f
e
c
ts
(
le
s
s
t
ha
n 1%
of
to
ta
l
im
a
ge
s
)
ge
t
m
is
s
e
d
out
.
T
he
a
ve
r
a
ge
t
im
e
Pe
r
f
o
r
ma
n
c
e
Met
r
i
c
s
Pe
r
f
o
r
ma
n
c
e
c
o
mp
a
r
i
s
o
n
o
f
d
i
f
f
e
r
e
n
t
s
t
u
d
i
e
s
A
c
c
u
r
a
c
y
P
r
ec
i
s
i
o
n
R
ec
a
l
l
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4711
-
4721
4718
to
de
te
c
t
de
f
e
c
t
is
m
e
a
s
ur
e
d
a
c
r
os
s
th
e
f
our
s
ol
ut
io
ns
a
nd
th
e
r
e
s
ul
t
is
gi
ve
n
in
T
a
bl
e
3.
F
ig
ur
e
3
s
how
s
th
e
c
om
pa
r
is
on of
de
f
e
c
t
de
te
c
ti
on i
n t
e
r
m
s
of
t
im
e
.
T
a
bl
e
3. C
om
put
a
ti
on t
im
e
c
om
pa
r
is
on
D
e
f
e
c
t
de
t
e
c
t
i
on t
i
m
e
(
m
s
)
M
e
t
hod
C
ol
or
C
ut
H
ol
e
T
hr
e
a
d
MC
A
vg
A
S
A
120
127
119
131
123
124
P
R
A
N
-
N
e
t
132
135
132
142
130
134.2
M
ul
t
i
s
c
a
l
e
-
C
N
N
161
167
162
172
164
165.2
F
a
s
t
e
r
R
C
N
N
123
130
124
137
127
128.2
F
ig
ur
e
3
.
C
om
pa
r
is
on of
de
f
e
c
t
de
te
c
ti
on i
n t
e
r
m
s
of
t
im
e
T
he
a
ve
r
a
ge
ti
m
e
ta
ke
n
f
or
de
f
e
c
t
de
te
c
ti
on
in
pr
opos
e
d
s
ol
ut
io
n
is
8%
lo
w
e
r
c
om
pa
r
e
d
t
o
P
R
A
N
-
N
e
t,
33%
lo
w
e
r
c
om
pa
r
e
d
to
m
ul
ti
s
c
a
le
C
N
N
a
nd
3%
lo
w
e
r
c
om
pa
r
e
d
to
F
a
s
te
r
R
C
N
N
a
s
s
how
n
in
F
ig
ur
e
3.
T
he
ti
m
e
ha
s
r
e
duc
e
d
in
pr
opos
e
d
s
ol
ut
io
n
due
to
s
e
le
c
ti
on
of
f
e
w
a
nc
hor
r
e
gi
ons
w
he
r
e
de
f
e
c
t
de
te
c
ti
on
is
f
oc
us
e
d.
P
R
A
N
-
N
e
t
h
a
s
hi
gh
e
r
ti
m
e
due
to
m
or
e
a
nc
hor
r
e
gi
ons
c
om
pa
r
e
d
to
pr
opos
e
d
s
ol
ut
io
n.
F
a
s
te
r
R
C
N
N
r
e
duc
e
d t
he
e
x
e
c
ut
e
d by c
om
pr
is
in
g o
n t
h
e
a
c
c
ur
a
c
y. T
he
t
e
s
t
im
a
ge
s
a
r
e
s
pl
it
t
o t
w
o c
a
te
gor
ie
s
na
m
e
ly
:
t
in
y
(
<
1%
of
to
ta
l
im
a
ge
)
a
nd
bi
g
(
>
1%
of
to
ta
l
im
a
g
e
)
a
nd
a
c
c
ur
a
c
y
is
c
om
pa
r
e
d
a
c
r
os
s
4
s
ol
ut
io
n
f
or
bot
h i
m
a
ge
c
a
te
gor
ie
s
. T
he
r
e
s
ul
ts
a
r
e
gi
ve
n i
n
T
a
bl
e
4.
T
a
bl
e
4. C
om
pa
r
is
on of
a
c
c
ur
a
c
y f
or
t
in
y a
nd big de
f
e
c
ts
A
c
c
ur
a
c
y
M
e
t
hod
T
i
ny de
f
e
c
t
s
B
i
g de
f
e
c
t
s
A
S
A
92
94.2
P
R
A
N
-
N
e
t
90.9
92.9
M
ul
t
i
s
c
a
l
e
-
C
N
N
87.8
88.1
F
a
s
t
e
r
R
C
N
N
86.2
88
T
he
a
ve
r
a
g
e
a
c
c
ur
a
c
y
f
or
ti
ny
de
f
e
c
ts
in
pr
opos
e
d
s
ol
ut
io
n
is
1.1%
hi
ghe
r
c
om
pa
r
e
d
to
P
R
A
N
-
N
e
t,
4.2%
hi
ghe
r
c
om
pa
r
e
d t
o M
ul
ti
s
c
a
le
C
N
N
a
nd 5.8%
hi
ghe
r
c
o
m
pa
r
e
d t
o F
a
s
te
r
R
C
N
N
.
T
he
a
ve
r
a
ge
a
c
c
ur
a
c
y
f
or
bi
g
de
f
e
c
ts
in
pr
opos
e
d
s
ol
ut
io
n
is
1.3%
hi
ghe
r
c
om
pa
r
e
d
to
P
R
A
N
-
N
e
t,
6.1%
hi
ghe
r
c
om
pa
r
e
d
to
M
ul
ti
s
c
a
le
C
N
N
a
nd
6.2%
hi
ghe
r
c
om
pa
r
e
d
to
F
a
s
te
r
R
C
N
N
.
T
he
pr
opos
e
d
s
ol
ut
io
n
pe
r
f
or
m
e
d
w
e
ll
f
or
t
in
y
de
f
e
c
ts
due
t
o i
ts
uni
que
s
e
le
c
ti
on of
a
nc
hor
poi
nt
s
. T
he
s
e
a
nc
hor
poi
nt
s
w
e
r
e
a
bl
e
t
o c
ove
r
e
ve
n a
t
in
y de
f
e
c
t
r
e
g
io
n. T
he
a
c
c
ur
a
c
y f
or
di
f
f
e
r
e
nt
t
ype
s
of
de
f
e
c
ts
i
s
m
e
a
s
ur
e
d i
n pr
opos
e
d s
ol
ut
io
n a
nd give
n i
n
T
a
bl
e
5.
T
he
0
20
40
60
80
100
120
140
160
180
A
S
A
P
R
A
N
-
N
E
T
M
u
l
t
i
s
c
a
l
e
C
N
N
F
a
s
t
er
R
CN
N
T
i
m
e
M
et
h
o
d
o
l
o
gy
C
om
pa
r
i
s
on
of
de
f
ect
de
t
ect
i
on
i
n
t
er
m
s
of
t
i
m
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
A
nc
hor
s
e
le
c
ti
on ba
s
e
d de
e
p
le
a
r
ni
ng t
w
o s
ta
ge
f
ab
r
ic
de
fe
c
t
lo
c
al
iz
at
io
n
(
H
at
ta
r
k
i
P
ooj
a
)
4719
hi
ghe
s
t
a
c
c
ur
a
c
y
is
a
c
hi
e
v
e
d
f
or
C
ol
or
de
f
e
c
t
in
p
r
opos
e
d
s
ol
ut
io
n
f
ol
lo
w
e
d
by
c
ut
de
f
e
c
t.
T
he
r
e
c
e
iv
e
r
ope
r
a
ti
ng c
ha
r
a
c
te
r
is
ti
c
(
R
O
C
)
pl
ot
c
om
pa
r
in
g f
our
s
ol
ut
io
ns
i
s
gi
ve
n i
n F
ig
ur
e
4.
T
a
bl
e
5. C
om
pa
r
is
on of
a
c
c
ur
a
c
y f
or
va
r
io
us
de
f
e
c
ts
D
e
f
e
c
t
A
c
c
ur
a
c
y
C
ol
or
94.5
C
ut
93.2
H
ol
e
91.1
T
hr
e
a
d
92.4
MC
90.3
F
ig
ur
e
4. A
R
O
C
c
ur
ve
c
om
pa
r
is
on
f
or
A
S
A
, P
R
A
N
-
N
e
t,
m
ul
ti
s
c
a
le
C
N
N
,
f
a
s
te
r
R
C
N
N
T
he
R
O
C
a
r
e
a
is
hi
gh
e
r
in
pr
opos
e
d
s
ol
ut
io
n
c
om
pa
r
e
d
to
e
x
is
ti
ng
w
or
ks
.
T
h
e
pr
opos
e
d
s
ol
ut
io
n
de
m
ons
tr
a
te
s
b
e
tt
e
r
tr
a
de
-
of
f
be
twe
e
n
s
e
n
s
it
iv
it
y
a
nd
s
pe
c
if
i
c
it
y
c
om
pa
r
e
d
to
e
xi
s
ti
ng
w
or
ks
.
T
he
d
e
f
e
c
t
lo
c
a
li
z
a
ti
on e
r
r
or
i
s
m
e
a
s
ur
e
d
i
n t
e
r
m
s
of
H
a
us
dor
f
f
di
s
t
a
n
c
e
(
)
.
T
hi
s
i
s
m
e
a
s
ur
e
d by
c
a
lc
ul
a
ti
ng t
he
di
f
f
e
r
e
nc
e
of
c
ont
our
s
of
pr
e
di
c
te
d
a
nd
a
c
tu
a
l
de
f
e
c
t
r
e
gi
on
a
s
gi
ve
n
in
(
20)
.
L
ow
e
r
th
e
va
lu
e
of
lo
w
e
r
is
th
e
de
f
e
c
t
lo
c
a
li
z
a
ti
on e
r
r
or
. T
he
a
ve
r
a
ge
va
lu
e
i
s
m
e
a
s
ur
e
d f
or
t
in
y a
nd big de
te
c
ts
a
nd t
he
r
e
s
ul
t
is
gi
ve
n i
n T
a
bl
e
6.
(
,
)
=
{
s
up
in
f
(
,
)
,
s
up
in
f
(
,
)
,
(
20)
T
a
bl
e
6
.
D
e
f
e
c
t
lo
c
a
li
z
a
ti
on e
r
r
or
M
e
t
hod
T
i
ny
B
i
g
A
S
A
18
43
P
R
A
N
-
N
e
t
24
57
M
ul
t
i
s
c
a
l
e
-
C
N
N
29
62
F
a
s
t
e
r
R
C
N
N
30
64
T
he
de
f
e
c
t
lo
c
a
li
z
a
ti
on
e
r
r
or
in
te
r
m
s
of
is
a
t
le
a
s
t
33%
lo
w
e
r
f
or
ti
ny
de
f
e
c
ts
a
nd
32%
lo
w
e
r
f
or
bi
g
de
f
e
c
ts
.
T
h
e
de
f
e
c
t
lo
c
a
li
z
a
ti
on
e
r
r
or
ha
s
r
e
duc
e
d
in
th
e
pr
opo
s
e
d
s
ol
ut
io
n
due
to
s
e
l
e
c
ti
on
of
m
os
t
pr
oba
bl
e
r
e
gi
ons
us
in
g
c
om
bi
ne
d
f
r
e
que
nc
y
dom
a
in
a
na
ly
s
is
w
it
h
s
up
e
r
pi
xe
l
s
e
gm
e
nt
a
ti
on
a
t
f
ir
s
t
s
ta
ge
a
nd
de
e
p
le
a
r
ni
ng
r
e
gr
e
s
s
io
n
on
th
e
s
e
le
c
te
d
r
e
gi
on
in
th
e
s
e
c
ond
s
t
a
ge
.
A
s
s
um
m
a
r
iz
e
d
f
r
om
T
a
bl
e
1,
th
e
c
om
put
in
g
c
om
pl
e
xi
ty
in
a
ppr
oa
c
h
pr
oc
e
s
s
in
g
im
a
ge
a
s
w
hol
e
is
hi
gh
e
r
c
o
m
pa
r
e
d
to
a
nc
hor
s
e
le
c
ti
on
-
ba
s
e
d
a
ppr
oa
c
he
s
.
A
S
A
r
e
duc
in
g
th
e
c
om
put
in
g
c
om
pl
e
xi
ty
by
na
r
r
ow
in
g
th
e
de
f
e
c
t
a
na
ly
s
is
to
s
e
le
c
ti
ve
r
e
gi
ons
.
B
ut
w
it
h
im
pr
ope
r
s
e
le
c
ti
on
of
p
r
oba
bl
e
r
e
gi
ons
,
th
e
y
in
c
r
e
a
s
e
d
th
e
f
a
l
s
e
pos
it
iv
e
s
a
nd
a
ls
o
in
c
r
e
a
s
e
d
th
e
c
om
put
in
g
c
om
pl
e
xi
ty
th
ough
it
is
lo
w
e
r
th
a
n
de
e
p
le
a
r
ni
ng
a
ppr
oa
c
he
s
pr
oc
e
s
s
in
g
im
a
ge
a
s
w
hol
e
.
T
hi
s
w
or
k
s
ol
ve
d
th
is
pr
obl
e
m
by
s
e
le
c
ti
ng
th
e
be
s
t
pr
oba
bl
e
r
e
gi
ons
a
nd
doi
ng
de
e
p
le
a
r
ni
ng
r
e
gr
e
s
s
io
n
a
na
ly
s
is
on
th
os
e
r
e
gi
ons
.
F
r
om
th
e
r
e
s
ul
ts
of
de
f
e
c
t
lo
c
a
li
z
a
ti
on
e
r
r
o
r
,
a
c
c
ur
a
c
y
a
nd
de
f
e
c
t
d
e
te
c
ti
on
ti
m
e
,
it
c
a
n
be
s
e
e
n
th
a
t
p
r
opos
e
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
13
, N
o.
4
,
D
e
c
e
m
be
r
2024
:
4711
-
4721
4720
s
ol
ut
io
n
ha
s
e
f
f
e
c
ti
ve
ly
s
ol
ve
d
th
e
pr
obl
e
m
s
.
A
ls
o,
th
e
pr
opos
e
d
s
ol
ut
io
n
pe
r
f
or
m
e
d
c
ons
is
te
nt
ly
f
or
di
f
f
e
r
e
nt
ty
pe
s
of
de
f
e
c
ts
a
nd de
f
e
c
t
di
m
e
ns
io
ns
.
5.
C
O
N
C
L
U
S
I
O
N
T
hi
s
w
or
k
pr
opos
e
d
a
n
A
S
A
b
a
s
e
d
two
s
ta
g
e
de
e
p
le
a
r
ni
ng
f
a
b
r
ic
de
f
e
c
t
lo
c
a
li
z
a
ti
on
te
c
hni
que
. T
he
a
nc
hor
s
w
e
r
e
s
e
le
c
te
d
w
it
h
m
a
c
hi
ne
le
a
r
ni
ng a
ppr
oa
c
h
u
s
in
g s
a
li
e
nc
y
r
e
gi
on
de
te
c
ti
on
a
nd
s
c
or
in
g
th
e
s
a
li
e
nt
r
e
gi
on f
or
a
pr
oba
bi
li
ty
o
f
de
f
e
c
t
oc
c
ur
r
e
nc
e
. D
e
e
p l
e
a
r
ni
ng l
oc
a
li
z
e
s
t
he
f
a
ul
ts
i
n t
he
s
e
le
c
te
d a
n
c
hor
r
e
gi
ons
.
T
he
pr
opos
e
d
s
ol
ut
io
n
is
a
bl
e
to
r
e
duc
e
th
e
c
om
put
a
ti
ona
l
ove
r
he
a
d
by
3%
a
nd
a
ls
o
in
c
r
e
a
s
e
s
th
e
f
a
br
ic
de
f
e
c
ti
on l
oc
a
li
z
a
ti
on by 1.2%
.
C
om
pa
r
in
g
t
he
s
ugge
s
te
d a
ppr
oa
c
h t
o t
he
c
ur
r
e
nt
m
e
th
ods
,
it
i
s
f
ound
th
a
t
th
e
la
tt
e
r
pe
r
f
or
m
s
w
or
s
e
ove
r
a
ll
in
te
r
m
s
of
a
c
c
ur
a
c
y,
pr
e
c
is
io
n,
a
nd
r
e
c
a
ll
,
a
nd
A
S
A
c
a
n
ha
ndl
e
bot
h
la
r
ge
a
nd
li
tt
le
f
la
w
s
. F
ut
ur
e
s
tu
di
e
s
w
il
l
a
s
s
e
s
s
t
h
e
s
ugge
s
te
d s
ol
ut
io
n'
s
r
e
s
il
ie
nc
e
a
g
a
in
s
t
a
r
a
nge
of
i
nt
r
ic
a
te
t
e
xt
ur
e
s
.
R
E
F
E
R
E
N
C
E
S
[
1]
J
.
B
ul
l
on,
A
.
G
.
A
r
r
i
e
t
a
,
A
.
H
.
E
nc
i
na
s
,
a
nd
A
.
Q
.
D
i
o
s
,
“
M
a
nuf
a
c
t
ur
i
ng
pr
oc
e
s
s
e
s
i
n
t
he
t
e
xt
i
l
e
i
ndus
t
r
y.
E
xpe
r
t
s
ys
t
e
m
s
f
or
f
a
br
i
c
s
pr
oduc
t
i
on,”
A
D
C
A
I
J
:
A
dv
anc
e
s
i
n
D
i
s
t
r
i
but
e
d
C
om
put
i
ng
and
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
J
our
nal
,
vol
.
6,
no.
1,
pp.
41
–
50,
J
a
n.
2017,
doi
:
10.14201/
A
D
C
A
I
J
2017614150.
[
2]
T
.
M
a
hm
ud,
J
.
S
i
kde
r
,
R
.
J
.
C
ha
km
a
,
a
nd
J
.
F
a
r
dous
h,
“
F
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
s
ys
t
e
m
,”
i
n
I
nt
e
l
L
i
ge
nt
C
om
put
i
ng
and
O
pt
i
m
i
z
at
i
on:
P
r
oc
e
e
di
ngs
of
t
h
e
3
rd
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on I
nt
e
l
l
i
ge
nt
C
o
m
put
i
ng and
O
pt
i
m
i
z
at
i
on 2020 (
I
C
O
2020)
, 2021, pp. 788
–
800
,
doi
:
10.1007/
978
-
3
-
030
-
68154
-
8_68.
[
3]
C
.
L
i
,
J
.
L
i
,
Y
.
L
i
,
L
.
H
e
,
X
.
F
u,
a
nd
J
.
C
he
n,
“
F
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
i
n
t
e
xt
i
l
e
m
a
nuf
a
c
t
ur
i
ng:
a
s
ur
ve
y
of
t
he
s
t
a
t
e
of
t
he
a
r
t
,”
Se
c
ur
i
t
y
and C
om
m
uni
c
at
i
on N
e
t
w
o
r
k
s
, pp. 1
–
13, 2021, doi
:
10.1155/
2021/
994
8808.
[
4]
C
.
C
.
-
H
o
a
nd
G
.
K
.
H
.
P
a
ng,
“
F
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
by
F
ou
r
i
e
r
a
na
l
ys
i
s
,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
I
ndus
t
r
y
A
ppl
i
c
at
i
ons
,
vol
.
36,
no. 5, pp. 1267
–
1276, 2000, doi
:
10.1109/
28.871274.
[
5]
J
. W
a
ng, K
. C
he
n, S
. Y
a
ng, C
.
C
. L
oy,
a
nd D
.
L
i
n, “
R
e
gi
on pr
opos
a
l
by
gui
de
d
a
nc
hor
i
ng,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
C
onf
e
r
e
n
c
e
on C
om
put
e
r
V
i
s
i
on and P
at
t
e
r
n R
e
c
ogni
t
i
on
, 2019, pp. 2960
–
2969
, doi
:
10.1109/
C
V
P
R
.2019.00308.
[
6]
T
. A
l
m
e
i
da
, F
. M
out
i
nho,
a
nd J
. P
.
M
.
-
C
a
r
va
l
ho, “
F
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
w
i
t
h
de
e
p l
e
a
r
ni
ng a
nd f
a
l
s
e
ne
g
a
t
i
ve
r
e
duc
t
i
on,”
I
E
E
E
A
c
c
e
s
s
, vol
. 9, pp. 81936
–
81945, 2021, doi
:
10.1109/
A
C
C
E
S
S
.2021.3086028.
[
7]
C
.
L
i
,
B
.
J
i
a
ng, Z
.
L
i
u,
Y
.
D
ong,
S
. T
a
ng,
a
nd
D
.
W
e
ng,
“
F
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
ba
s
e
d
on
de
e
p
-
ha
ndc
r
a
f
t
e
d
f
e
a
t
ur
e
a
nd
w
e
i
ght
e
d
l
ow
-
r
a
nk m
a
t
r
i
x r
e
pr
e
s
e
nt
a
t
i
on,”
J
our
nal
of
E
ngi
ne
e
r
e
d F
i
be
r
s
and F
abr
i
c
s
, vo
l
. 16, 2021, doi
:
10.1177/
15589250211008453.
[
8]
D
.
Y
a
pi
,
M
.
S
.
A
l
l
i
l
i
,
a
nd
N
.
B
a
a
z
i
z
,
“
A
ut
om
a
t
i
c
f
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
u
s
i
ng
l
e
a
r
ni
ng
-
ba
s
e
d
l
oc
a
l
t
e
xt
ur
a
l
di
s
t
r
i
but
i
ons
i
n
t
h
e
c
ont
our
l
e
t
dom
a
i
n,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
A
ut
om
at
i
on
Sc
i
e
nc
e
and
E
ngi
ne
e
r
i
ng
,
vol
.
15,
no.
3,
pp.
1014
–
1026,
J
u
l
.
2018,
doi
:
10.1109/
T
A
S
E
.2017.2696748.
[
9]
S
.
S
us
a
n
a
nd
M
.
S
ha
r
m
a
,
“
A
ut
om
a
t
i
c
t
e
xt
ur
e
de
f
e
c
t
de
t
e
c
t
i
on
us
i
ng
G
a
u
s
s
i
a
n
m
i
xt
ur
e
e
nt
r
opy
m
ode
l
i
ng,”
N
e
ur
oc
om
put
i
ng
,
vol
.
239, pp. 232
–
237, 2017, doi
:
10.1016/
j
.ne
uc
om
.2017.02.021.
[
10]
L
. T
ong, W
. K
. W
ong,
a
nd C
. K
. K
w
ong,
“
F
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on f
or
a
ppa
r
e
l
i
n
dus
t
r
y:
a
nonl
oc
a
l
s
pa
r
s
e
r
e
pr
e
s
e
nt
a
t
i
on a
ppr
oa
c
h,”
I
E
E
E
A
c
c
e
s
s
, pp. 1
–
1, 2017, doi
:
10.1109/
A
C
C
E
S
S
.2017.2667890.
[
11]
T
.
Q
u,
L
. Z
ou,
Q
. Z
ha
ng,
X
.
C
he
n,
a
nd
C
.
F
a
n,
“
D
e
f
e
c
t
de
t
e
c
t
i
on
on
t
he
f
a
b
r
i
c
w
i
t
h
c
om
pl
e
x
t
e
xt
ur
e
v
i
a
dua
l
-
s
c
a
l
e
ove
r
-
c
om
pl
e
t
e
di
c
t
i
ona
r
y,”
T
he
J
our
nal
of
T
he
T
e
x
t
i
l
e
I
ns
t
i
t
ut
e
, vol
. 107, no. 6, pp. 743
–
756, J
un. 2016, doi
:
10.1080/
00405000.2015.1061760.
[
12]
B
.
S
hi
,
J
.
L
i
a
ng,
L
.
D
i
,
C
. C
he
n,
a
nd
Z
.
H
ou,
“
F
a
br
i
c
de
f
e
c
t
d
e
t
e
c
t
i
on
vi
a
l
ow
-
r
a
nk
de
c
om
pos
i
t
i
on
w
i
t
h
gr
a
di
e
nt
i
nf
or
m
a
t
i
on,”
I
E
E
E
A
c
c
e
s
s
, vol
. 7, pp. 130423
–
130437, 2019, doi
:
10.1109/
A
C
C
E
S
S
.2019.293984
3.
[
13]
Q
.
H
ua
ngpe
ng,
H
.
Z
ha
ng,
X
.
Z
e
ng,
a
nd
W
.
H
ua
ng,
“
A
ut
om
a
t
i
c
vi
s
ua
l
d
e
f
e
c
t
de
t
e
c
t
i
on
u
s
i
ng
t
e
xt
ur
e
pr
i
or
a
nd
l
ow
-
r
a
nk
r
e
pr
e
s
e
nt
a
t
i
on,”
I
E
E
E
A
c
c
e
s
s
, vol
. 6, pp. 37965
–
37976, 2018, doi
:
10.1109/
A
C
C
E
S
S
.2018.2852663.
[
14]
J
.
W
a
ng,
C
.
L
i
,
Z
.
L
i
u,
Y
.
D
ong,
a
nd
Y
.
H
ua
ng,
“
C
om
bi
ng
d
e
e
p
a
nd
ha
nd
c
r
a
f
t
e
d
f
e
a
t
ur
e
s
f
or
N
T
V
-
N
R
P
C
A
ba
s
e
d
f
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on,”
i
n
P
at
t
e
r
n R
e
c
ogni
t
i
on and C
om
put
e
r
V
i
s
i
on
,
2019, pp. 479
–
490
, d
oi
:
10.1007/
978
-
3
-
030
-
31726
-
3_41.
[
15]
Y
.
L
i
,
W
.
Z
ha
o,
a
nd
J
.
P
a
n,
“
D
e
f
or
m
a
bl
e
pa
t
t
e
r
ne
d
f
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
w
i
t
h
f
i
s
he
r
c
r
i
t
e
r
i
on
-
ba
s
e
d
de
e
p
l
e
a
r
ni
ng,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
A
ut
om
at
i
on
Sc
i
e
nc
e
and
E
ngi
ne
e
r
i
ng
,
vol
.
14,
no.
2,
pp.
1256
–
1
264,
A
pr
.
2017,
doi
:
10.1109/
T
A
S
E
.2016.2520955.
[
16]
S
.
Z
ha
o,
L
.
Y
i
n,
J
.
Z
ha
ng,
J
.
W
a
ng,
a
nd
R
.
Z
hong,
“
R
e
a
l
‐
t
i
m
e
f
a
br
i
c
de
f
e
c
t
d
e
t
e
c
t
i
on
ba
s
e
d
on
m
ul
t
i
‐
s
c
a
l
e
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k,”
I
E
T
C
ol
l
abor
at
i
v
e
I
nt
e
l
l
i
ge
nt
M
anuf
ac
t
ur
i
ng
, vol
. 2, no. 4, pp. 189
–
1
96, D
e
c
. 2020, doi
:
10.1049/
i
e
t
-
c
i
m
.2020.0062.
[
17]
S
. M
e
i
, Y
. W
a
ng, a
nd G
. W
e
n, “
A
ut
om
a
t
i
c
f
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on w
i
t
h a
m
u
l
t
i
-
s
c
a
l
e
c
onvol
ut
i
ona
l
de
noi
s
i
ng a
ut
oe
nc
ode
r
ne
t
w
or
k
m
ode
l
,”
Se
ns
or
s
, vol
. 18, no. 4, A
pr
. 2018, doi
:
10.3390/
s
18041064.
[
18]
B
.
W
e
i
,
K
.
H
a
o,
X
.
T
a
ng,
a
nd
Y
.
D
i
ng, “
A
ne
w
m
e
t
hod
u
s
i
ng
t
he
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k
w
i
t
h
c
om
pr
e
s
s
i
ve
s
e
ns
i
ng
f
or
f
a
br
i
c
de
f
e
c
t
c
l
a
s
s
i
f
i
c
a
t
i
on
ba
s
e
d
on
s
m
a
l
l
s
a
m
pl
e
s
i
z
e
s
,”
T
e
x
t
i
l
e
R
e
s
e
a
r
c
h
J
ou
r
nal
,
vol
.
89,
no.
17,
pp.
3539
–
3555,
S
e
p.
2019,
doi
:
10.1177/
0040517518813656.
[
19]
W
. O
uya
ng, B
. X
u, J
. H
ou, a
nd X
. Y
u
a
n, “
F
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on u
s
i
ng a
c
t
i
va
t
i
on l
a
ye
r
e
m
be
dde
d c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k,
”
I
E
E
E
A
c
c
e
s
s
, vol
. 7, pp. 70130
–
70140, 2019, doi
:
10.1109/
A
C
C
E
S
S
.2019.2913620.
[
20]
J
.
J
i
ng,
H
.
M
a
,
a
nd
H
.
Z
ha
ng,
“
A
ut
om
a
t
i
c
f
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
us
i
ng
a
de
e
p
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k,”
C
ol
or
at
i
on
T
e
c
hnol
ogy
, vol
. 135, no. 3, pp. 213
–
223, J
un. 2019, doi
:
10.1111/
c
ot
e
.12394.
[
21]
B
.
W
e
i
,
K
.
H
a
o,
X
.
S
.
T
a
ng,
a
nd
L
.
R
e
n,
“
F
a
br
i
c
d
e
f
e
c
t
de
t
e
c
t
i
on
ba
s
e
d
on
f
a
s
t
e
r
R
C
N
N
,”
A
dv
anc
e
s
i
n
I
nt
e
l
l
i
ge
nt
Sy
s
t
e
m
s
an
d
C
om
put
i
ng
, vol
. 849, pp. 45
–
51, 2019, doi
:
10.1007/
978
-
3
-
319
-
99695
-
0_6.
[
22]
H
.
Z
ha
ng,
J
.
M
a
,
J
.
J
i
ng,
a
nd
P
.
L
i
,
“
F
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
m
e
t
hod
ba
s
e
d
o
n
i
m
pr
ove
d
f
a
s
t
w
e
i
ght
e
d
m
e
di
a
n
f
i
l
t
e
r
i
ng
a
nd
K
-
m
e
a
ns
,”
F
angz
hi
X
ue
bao/
J
our
nal
of
T
e
x
t
i
l
e
R
e
s
e
a
r
c
h
, vol
. 40, no. 12, pp. 50
–
5
6, 2019, doi
:
10.13475/
j
.f
z
xb.20181200407.
[
23]
P
.
P
e
ng,
Y
.
W
a
ng,
C
.
H
a
o,
Z
.
Z
hu,
T
.
L
i
u,
a
nd
W
.
Z
hou,
“
A
ut
om
a
t
i
c
f
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
m
e
t
hod
u
s
i
ng
pr
a
n
-
ne
t
,”
A
ppl
i
e
d
Sc
i
e
nc
e
s
, vol
. 10, no. 23, pp. 1
–
13, 2020, doi
:
10.3390/
a
pp10238434.
[
24]
Z
.
C
he
ng,
X
.
L
uo,
Y
.
S
hi
,
a
nd
K
.
K
i
t
a
,
“
F
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
a
l
gor
i
t
hm
ba
s
e
d
on
Y
O
L
O
v3
T
r
a
ns
f
e
r
l
e
a
r
ni
ng,”
i
n
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on F
r
ont
i
e
r
s
of
E
l
e
c
t
r
oni
c
s
, I
nf
or
m
at
i
on and
C
om
put
at
i
on T
e
c
hnol
ogi
e
s
, N
e
w
Y
or
k, U
S
A
:
A
C
M
, M
a
y
2021, pp. 1
–
7
,
doi
:
10.1145/
3474198.3478149.
[
25]
N
.
S
a
ndhya
,
N
.
M
.
S
a
s
hi
kum
a
r
,
M
.
P
r
i
ya
nka
,
S
.
M
.
W
e
ni
s
c
h,
a
nd
K
.
K
um
a
r
a
s
a
m
y,
“
A
ut
om
a
t
e
d
f
a
br
i
c
de
f
e
c
t
de
t
e
c
t
i
on
a
n
d
c
l
a
s
s
i
f
i
c
a
t
i
on:
a
de
e
p
l
e
a
r
ni
ng
a
ppr
oa
c
h,”
T
e
x
t
i
l
e
and
L
e
at
he
r
R
e
v
i
e
w
,
vol
.
4,
pp
.
315
–
335,
D
e
c
.
2021,
doi
:
10.31881/
T
L
R
.2021.24.
Evaluation Warning : The document was created with Spire.PDF for Python.