I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
E
lec
t
r
ical
an
d
Com
p
u
t
e
r
E
n
gin
e
e
r
in
g
(
I
JE
CE
)
Vol.
1
4
,
No.
5
,
Oc
tober
20
2
4
,
pp
.
5185
~
5
19
9
I
S
S
N:
2088
-
8708
,
DO
I
:
10
.
11591/i
jec
e
.
v
1
4
i
5
.
pp
5
185
-
5
199
5185
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
e
c
e
.
iaes
c
or
e
.
c
om
Ph
ase
st
r
u
c
t
u
r
e
s
-
b
as
e
d
h
yb
r
id
ap
p
r
oa
c
h
e
s f
or
d
e
f
e
c
t
d
e
t
e
c
t
io
n
in
vi
al
s
Vis
h
wanat
h
a
C.
R.
1
,2
,
As
h
a
V.
1
1
R
e
s
e
a
r
c
h C
e
nt
r
e
-
D
e
p
a
r
tm
e
nt
of
M
a
s
te
r
of
C
omput
e
r
A
ppl
ic
a
ti
ons
, N
e
w
H
or
iz
on C
ol
le
ge
of
E
ngi
ne
e
r
in
g, B
e
nga
lu
r
u, A
f
f
il
ia
te
d t
o
V
is
ve
s
va
r
a
y
a
T
e
c
hnol
ogi
c
a
l
U
ni
ve
r
s
it
y, B
e
l
a
ga
vi
, K
a
r
na
ta
ka
, I
ndi
a
2
D
e
pa
r
tm
e
nt
of
M
a
s
t
e
r
of
C
omput
e
r
A
ppl
ic
a
ti
ons
, N
it
te
M
e
e
na
ks
hi
I
ns
ti
tu
te
of
T
e
c
hnol
ogy, Ye
la
h
a
nka
, B
e
ng
a
lu
r
u, K
a
r
na
ta
ka
, I
ndi
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
Apr
25,
2024
R
e
vis
e
d
J
un
7,
2024
Ac
c
e
pted
J
un
16,
2024
Q
u
a
l
i
t
y
co
n
t
r
o
l
a
n
d
a
s
s
u
ran
ce
i
n
p
h
armace
u
t
i
ca
l
v
i
a
l
man
u
fa
ct
u
ri
n
g
ar
e
p
aramo
u
n
t
t
o
en
s
u
re
d
ru
g
s
afe
t
y
a
n
d
eff
i
cacy
.
D
efe
ct
s
s
u
c
h
as
crac
k
s
,
b
u
b
b
l
es
,
b
l
ac
k
s
p
o
t
s
,
a
n
d
w
r
i
n
k
l
e
s
can
co
m
p
ro
m
i
s
e
p
ro
d
u
c
t
q
u
al
i
t
y
a
n
d
p
at
i
en
t
s
afet
y
.
T
h
i
s
s
t
u
d
y
p
ro
p
o
s
e
s
a
n
o
v
e
l
met
h
o
d
o
l
o
g
y
t
h
at
i
n
t
e
g
rat
e
s
fas
t
non
-
l
o
cal
mea
n
s
(F
N
L
M)
fi
l
t
eri
n
g
w
i
t
h
h
y
b
r
i
d
i
mag
e
p
ro
ces
s
i
n
g
t
ech
n
i
q
u
e
s
t
o
d
e
t
ect
t
h
es
e
d
efect
s
.
Prev
i
o
u
s
a
p
p
r
o
ach
e
s
h
av
e
o
ft
e
n
s
t
ru
g
g
l
e
d
w
i
t
h
s
u
b
t
l
e
an
o
ma
l
i
e
s
i
n
t
e
x
t
u
re
an
d
s
u
rface
feat
u
re
s
.
T
h
e
p
ro
p
o
s
e
d
s
o
l
u
t
i
o
n
l
ev
er
ag
e
s
p
h
a
s
e
s
t
r
u
ct
u
re
an
al
y
s
i
s
,
u
t
i
l
i
z
i
n
g
p
h
a
s
e
s
t
re
t
ch
t
ran
s
f
o
rm
(PST
)
t
o
effect
i
v
el
y
h
i
g
h
l
i
g
h
t
s
u
b
t
l
e
a
n
o
ma
l
i
e
s
b
y
ex
t
ract
i
n
g
f
eat
u
re
s
s
en
s
i
t
i
v
e
t
o
p
h
a
s
e
v
ar
i
at
i
o
n
s
.
T
h
e
s
e
feat
u
res
are
fu
r
t
h
er
ref
i
n
e
d
u
s
i
n
g
G
au
s
s
i
a
n
fi
l
t
er
i
n
g
,
w
i
t
h
O
t
s
u
t
h
re
s
h
o
l
d
i
n
g
ap
p
l
i
ed
f
o
r
p
reci
s
e
s
eg
me
n
t
a
t
i
o
n
an
d
d
efec
t
b
o
u
n
d
ar
y
i
d
e
n
t
i
fi
ca
t
i
o
n
.
Mo
rp
h
o
l
o
g
i
c
al
d
i
l
a
t
i
o
n
en
h
an
ce
s
d
e
t
ect
i
o
n
s
p
ee
d
an
d
accu
racy
,
w
h
i
l
e
reg
i
o
n
o
f
i
n
t
eres
t
(RO
I)
i
d
e
n
t
i
fi
ca
t
i
o
n
ai
d
s
i
n
l
o
cal
i
z
i
n
g
d
efect
s
an
d
faci
l
i
t
a
t
i
n
g
d
eci
s
i
o
n
-
ma
k
i
n
g
.
T
h
e
s
y
s
t
em
d
emo
n
s
t
rat
e
s
s
i
g
n
i
f
i
ca
n
t
i
mp
r
o
v
eme
n
t
s
i
n
q
u
al
i
t
y
co
n
t
r
o
l
,
ach
i
ev
i
n
g
h
i
g
h
p
erf
o
rman
c
e
met
ri
c
s
:
p
rec
i
s
i
o
n
(9
8
.
8
5
%
),
recal
l
(9
8
.
5
7
%
),
accu
racy
(9
8
.
3
6
%
),
s
p
eci
f
i
ci
t
y
(9
8
.
0
%
),
an
d
F
1
-
s
c
o
re
(9
8
.
7
1
%
).
It
al
s
o
ac
h
i
e
v
es
i
m
p
res
s
i
v
e
A
U
C
-
R
O
C
(9
8
.
1
8
%
)
an
d
A
U
C
-
PR
(9
9
.
0
8
%
)
v
al
u
es
,
d
emo
n
s
t
rat
i
n
g
i
t
s
ro
b
u
s
t
n
es
s
an
d
s
u
i
t
a
b
i
l
i
t
y
fo
r
d
efect
d
e
t
ect
i
o
n
i
n
p
h
arm
aceu
t
i
ca
l
v
i
a
l
s
.
K
e
y
w
o
r
d
s
:
De
f
e
c
t
ins
pe
c
ti
on
F
a
s
t
non
-
loca
l
mea
ns
f
il
ter
ing
Ots
u
thr
e
s
holdi
ng
P
ha
r
mac
e
uti
c
a
l
vials
P
ha
s
e
s
tr
e
tch
tr
a
ns
f
or
m
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
Vis
hwa
na
tha
C
.
R
.
R
e
s
e
a
r
c
h
C
e
ntr
e
-
De
pa
r
tm
e
nt
of
M
a
s
ter
of
C
omput
e
r
Applica
ti
ons
,
Ne
w
Hor
izon
C
oll
e
ge
of
E
nginee
r
ing,
B
e
nga
lur
u,
Af
f
i
li
a
ted
to
Vis
ve
s
va
r
a
ya
T
e
c
hnologi
c
a
l
Unive
r
s
it
y
B
e
laga
vi
-
590018,
Ka
r
na
taka
,
I
ndia
E
mail:
vis
hwa
na
thcr
@gmail.
c
om;
vis
hwa
na
tha.
c
r
@nmi
t.
a
c
.
in
1.
I
NT
RODU
C
T
I
ON
Va
r
ious
types
of
c
ontaine
r
s
a
r
e
c
omm
only
us
e
d
f
or
pha
r
mac
e
uti
c
a
l
dr
ug
pa
c
ka
ging
s
uc
h
a
s
s
yr
inges
,
vials
,
a
mpoul
e
s
,
a
nd
c
a
r
tr
idges
.
Via
ls
ha
ve
ga
ined
univer
s
a
l
a
c
c
e
ptanc
e
due
to
thei
r
ine
r
tnes
s
,
s
tabili
ty,
a
nd
de
f
ianc
e
of
c
he
mi
c
a
l
r
e
a
c
ti
ons
.
Author
it
ies
of
pha
r
mac
ope
ia
ha
ve
de
f
ined
the
s
tanda
r
ds
f
or
glas
s
c
o
ntaine
r
s
s
uit
a
ble
f
or
pha
r
mac
e
uti
c
a
l
us
e
.
T
he
y
c
a
n
be
one
of
T
ype
I
,
I
I
,
or
I
I
I
.
T
ype
I
is
made
of
ne
utr
a
l
bor
os
il
ica
te
glas
s
a
s
s
hown
in
F
igur
e
1,
T
ype
I
I
is
made
o
f
s
o
da
-
li
me
glas
s
a
nd
T
ype
I
I
I
is
made
o
f
unt
r
e
a
ted
s
oda
-
li
me
glas
s
.
E
ve
r
y
ye
a
r
ove
r
35
to
45
bil
li
on
s
uc
h
c
ontaine
r
s
a
r
e
pr
oduc
e
d
[
1]
.
Va
r
ious
f
a
c
tor
s
li
ke
us
a
bil
it
y,
s
tor
a
ge
,
a
dmi
nis
tr
a
ti
on,
s
ter
il
it
y
to
pr
e
ve
nt
c
ontam
ination,
c
ompatibi
li
ty
wi
th
the
dr
ug
f
or
mul
a
ti
on
to
pr
e
ve
nt
int
e
r
a
c
ti
ons
or
lea
c
hing,
a
nd
p
r
otec
ti
on
of
the
d
r
ug
f
r
om
e
nvir
onmenta
l
f
a
c
tor
s
li
ke
li
ght,
moi
s
t
ur
e
,
a
nd
oxyge
n
play
im
por
tant
r
oles
in
s
e
lec
ti
ng
the
type
o
f
c
ontaine
r
f
o
r
pha
r
mac
e
uti
c
a
l
us
e
.
Va
r
ious
types
of
plas
ti
c
bott
les
s
uc
h
a
s
high
-
de
ns
it
y
polyethylene
(
HD
P
E
)
bot
tl
e
s
,
polypr
opy
len
e
(
P
P
)
bott
les
,
a
nd
poly
e
thyl
e
ne
ter
e
phthala
te
(
P
E
T
)
bott
les
a
r
e
a
ls
o
in
us
e
a
s
dr
ug
c
ontaine
r
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5185
-
5
199
5186
F
igur
e
1.
P
ha
r
mac
e
uti
c
a
l
v
ial
De
f
e
c
t
de
tec
ti
on
in
pha
r
mac
e
uti
c
a
l
vials
is
c
r
it
ic
a
l
f
or
maintaining
p
r
oduc
t
in
tegr
it
y.
De
f
e
c
ts
l
ike
c
r
a
c
ks
,
blac
k
s
pots
,
wr
inkl
e
s
,
bubbles
,
pa
r
ti
c
les
,
d
e
lamination,
or
im
pu
r
it
ies
c
a
n
c
ompr
omi
s
e
the
int
e
gr
it
y
of
the
pr
oduc
t
lea
ding
to
potential
r
is
ks
f
o
r
pa
ti
e
nt
s
[
1]
.
B
r
e
a
ka
ge
s
a
r
e
ve
r
y
c
omm
on
dur
ing
the
pr
oduc
ti
on
pha
s
e
[
2]
.
De
lamination
o
f
glas
s
will
di
r
e
c
tl
y
im
pa
c
t
pa
ti
e
nt
he
a
lt
h
a
nd
ha
s
r
e
s
ult
e
d
in
the
r
e
c
a
ll
i
ng
of
the
dr
ug
p
r
oduc
t
[
3]
.
De
lamination
is
due
to
the
de
tac
hment
of
glas
s
f
lake
s
with
in
the
vial
a
nd
it
r
e
s
ult
s
in
c
ontamination
of
dr
ug
f
lui
d
potentially
a
f
f
e
c
ti
ng
i
ts
s
a
f
e
ty
a
nd
e
f
f
ica
c
y
a
nd
ne
e
ds
to
be
a
ddr
e
s
s
e
d
s
e
r
ious
ly.
T
r
a
dit
ional
methods
a
r
e
ti
me
-
c
ons
umi
ng,
of
ten
f
a
l
l
s
hor
t
in
e
f
f
icie
nc
y
in
identif
ying
s
ubtl
e
de
f
e
c
ts
,
a
nd
pr
one
to
human
e
r
r
or
ne
c
e
s
s
it
a
ti
ng
a
utom
a
ted
s
olut
ions
f
or
e
f
f
icie
nt
a
nd
a
c
c
ur
a
te
de
f
e
c
t
de
tec
ti
on
.
F
igu
r
e
2
s
hows
va
r
ious
types
of
c
os
metic
de
f
e
c
ts
on
the
vial
s
ur
f
a
c
e
.
(
a
)
(
b)
(
c
)
(
d)
(
e
)
(f)
F
igur
e
2.
Va
r
ious
de
f
e
c
ts
on
the
vial
(
a
)
br
oke
n
,
(
b
)
c
r
a
c
k
,
(
c
)
blac
k
s
pot
,
(
d)
s
c
r
a
tche
s
,
(
e
)
bubb
le
,
a
n
d
(
f
)
de
lamination
I
nte
gr
a
ti
on
of
va
r
ious
de
f
e
c
t
de
tec
ti
on
s
t
r
a
teg
ies
us
ing
ma
c
hine
vis
i
on
is
made
to
ma
ke
mac
hines
mor
e
int
e
ll
i
ge
nt
in
t
he
I
ndus
t
r
y
4
.
0
a
nd
o
nwa
r
ds
e
r
a
[
4
]
.
T
h
is
r
e
s
e
a
r
c
h
p
r
op
os
e
s
a
met
hodo
logy
c
omb
i
ning
f
a
s
t
non
-
loca
l
m
e
a
ns
(
F
NL
M
)
f
i
lt
e
r
ing
w
it
h
a
dva
nc
e
d
im
a
ge
p
r
oc
e
s
s
ing
te
c
hni
que
s
to
im
pr
o
ve
the
a
c
c
u
r
a
c
y
a
n
d
e
f
f
ica
c
y
o
f
d
e
f
e
c
t
de
tec
t
ion
i
n
v
ials
.
I
n
v
ial
ins
pe
c
ti
on,
whe
r
e
im
a
ge
s
mi
gh
t
ha
ve
va
r
yi
ng
il
lum
i
na
ti
on
or
nois
e
leve
ls
,
F
NL
M
's
r
obus
tnes
s
e
ns
u
r
e
s
c
ons
is
t
e
nt
de
f
e
c
t
de
tec
t
ion
a
c
r
os
s
d
if
f
e
r
e
nt
c
on
dit
ions
[
5]
.
P
ha
s
e
s
tr
e
tch
tr
a
ns
f
o
r
m
(
P
S
T
)
a
c
ts
a
s
a
n
ini
t
ial
s
tage
to
c
a
p
tu
r
e
poten
ti
a
l
a
noma
li
e
s
in
t
he
v
ial
im
a
ge
s
by
e
x
tr
a
c
ti
n
g
us
e
f
u
l
f
e
a
tu
r
e
s
r
e
la
ted
to
te
xtu
r
a
l
va
r
ia
ti
o
ns
a
nd
s
t
r
uc
tur
a
l
i
r
r
e
gul
a
r
i
ti
e
s
in
t
he
im
a
ge
.
P
S
T
is
ba
s
e
d
on
t
he
p
r
inc
ipl
e
s
of
d
if
f
e
r
e
nt
ial
int
e
r
f
e
r
e
nc
e
c
o
nt
r
a
s
t
m
icr
os
c
opy
.
I
t
ope
r
a
tes
i
n
t
he
f
r
e
q
ue
nc
y
dom
a
in
a
nd
e
xpl
oit
s
t
he
pha
s
e
inf
o
r
ma
ti
o
n
c
o
ntai
ne
d
wit
hin
a
n
im
a
g
e
t
o
e
n
ha
nc
e
s
ubtl
e
f
e
a
tur
e
s
tha
t
mi
gh
t
be
o
the
r
wis
e
c
ha
ll
e
ngi
ng
to
de
tec
t
[
6]
.
P
S
T
e
x
tr
a
c
ts
p
ha
s
e
-
r
e
la
ted
f
e
a
tu
r
e
s
b
y
a
n
a
lyz
ing
the
pha
s
e
c
on
gr
u
e
nc
y
a
c
r
os
s
d
if
f
e
r
e
nt
o
r
ien
tat
ions
a
nd
s
c
a
les
wit
hin
the
i
mage
.
A
f
te
r
P
S
T
f
e
a
tur
e
e
x
tr
a
c
ti
on,
s
ubs
e
que
nt
pr
oc
e
s
s
ing
s
teps
li
ke
f
il
t
e
r
ing
,
thr
e
s
holdi
ng
,
a
nd
mo
r
p
holo
gica
l
ope
r
a
t
ions
a
r
e
o
f
te
n
a
pp
li
e
d
t
o
r
e
f
ine
a
nd
loca
li
z
e
the
de
f
e
c
ts
.
F
i
lt
e
r
ing
a
ids
i
n
f
u
r
t
he
r
r
e
f
ineme
nt
s
o
tha
t
de
f
e
c
ts
that
a
r
e
s
u
btl
e
ge
t
n
oti
c
e
d
by
the
s
ys
t
e
m
.
S
e
gmen
tati
on
a
ids
in
ide
nti
f
yin
g
bounda
r
ies
t
hus
s
im
pli
f
yi
ng
i
mage
r
e
p
r
e
s
e
nta
ti
on
[
7]
.
O
ts
u
th
r
e
s
ho
ldi
ng
,
a
wide
ly
us
e
d
met
hod
f
or
i
mage
s
e
gmenta
ti
on
,
he
lps
in
d
is
ti
nguis
h
ing
f
o
r
e
g
r
oun
d
f
r
om
ba
c
kg
r
ou
nd
[
8]
.
M
o
r
pho
log
ica
l
op
e
r
a
ti
ons
l
ike
d
il
a
t
ion
a
id
in
im
pr
ovin
g
the
c
on
ti
n
uit
y
of
de
tec
ted
r
e
gi
ons
,
e
nha
nc
in
g
de
f
e
c
t
iden
ti
f
ica
t
ion
a
c
c
ur
a
c
y
.
P
r
e
vious
s
tudi
e
s
ha
ve
de
mons
tr
a
ted
the
e
f
f
e
c
ti
ve
ne
s
s
of
va
r
ious
im
a
ge
pr
oc
e
s
s
ing
tec
hnique
s
in
de
f
e
c
t
de
tec
ti
on.
T
he
us
e
of
s
uppor
t
ve
c
tor
m
a
c
hine
(
S
VM
)
,
puls
e
-
c
oupled
ne
ur
a
l
ne
twor
k
(
P
C
NN
)
,
c
onvolut
ional
ne
ur
a
l
ne
twor
ks
(
C
NN
)
,
a
nd
a
dva
nc
e
d
non
-
de
s
tr
uc
ti
ve
tes
ti
ng
(
NDT
)
a
r
e
e
f
f
e
c
ti
v
e
in
the
de
tec
ti
on
of
de
f
e
c
ts
on
glas
s
s
ur
f
a
c
e
s
[
9]
.
S
mar
t
manuf
a
c
tur
ing
ha
s
led
to
be
tt
e
r
qua
li
ty
c
ontr
ol
in
indus
tr
ies
today
thr
ough
a
r
ti
f
icia
l
vis
ion,
mac
hine
lea
r
ning
,
a
nd
s
tatis
ti
c
a
l
pr
oc
e
s
s
c
ontr
ol
[
10]
.
C
NN
s
a
r
e
quit
e
c
omm
on
a
mong
va
r
ious
de
f
e
c
t
d
e
tec
ti
on
tec
hniques
[
11
]
–
[
13]
.
De
f
e
c
ti
ve
c
a
ps
on
the
bott
les
we
r
e
f
ound
us
ing
a
li
ghtwe
ight
C
NN
model
by
Abhiji
t
e
t
al
.
[
14]
in
h
igh
-
s
pe
e
d
pr
oduc
ti
on
li
ne
s
.
Ada
m
opti
mi
z
e
r
is
int
r
oduc
e
d
in
the
s
ys
tem
to
im
pr
ove
the
pa
r
a
mete
r
s
li
ke
we
ight
s
a
nd
ke
r
ne
l
va
lu
e
s
a
long
with
a
c
r
os
s
-
e
ntr
opy
los
s
modul
e
to
incr
e
a
s
e
the
pr
oba
bil
it
y
of
de
tec
ti
on.
T
he
model
a
c
hieve
s
be
tt
e
r
a
c
c
ur
a
c
y
in
identif
ying
f
law
e
d
bott
le
c
a
ps
than
e
xis
ti
ng
met
hods
.
De
s
pit
e
c
ha
ll
e
nge
s
li
ke
s
ide
im
a
ge
c
a
ptur
e
a
nd
bott
le
c
las
s
if
ica
ti
on
is
s
ue
s
,
de
e
p
lea
r
ning
methods
e
f
f
e
c
ti
ve
ly
a
ddr
e
s
s
thes
e
is
s
ue
s
,
r
e
s
ult
ing
in
a
98%
a
c
c
ur
a
c
y
r
a
te.
A
li
mi
tation
o
f
the
s
tudy
is
it
s
f
oc
us
s
olely
on
bo
tt
le
c
a
p
identif
ica
ti
on,
a
s
f
utur
e
r
e
s
e
a
r
c
h
c
ould
e
x
pa
nd
the
model's
s
c
ope
to
de
tec
t
li
quid
f
il
l
-
leve
ls
,
f
or
e
ign
p
a
r
ti
c
les
,
a
nd
labe
l
a
li
gnment
de
f
icie
nc
ies
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
P
has
e
s
tr
uc
tur
e
s
-
bas
e
d
hy
br
id
appr
oac
he
s
for
de
fec
t
de
tec
ti
on
in
v
ial
s
(
V
is
hw
anatha
C
.
R
.
)
5187
C
NN
s
a
r
e
popular
ly
us
e
d
in
the
de
tec
ti
on
of
a
nim
a
ls
,
pe
ts
,
a
nd
other
na
tur
a
l
ob
jec
ts
.
De
f
e
c
tNe
t
is
a
s
tanda
r
d
a
ppr
oa
c
h
us
e
d
f
or
f
law
de
tec
ti
on
[
15]
.
M
obil
e
Ne
t
-
v2
-
de
n
s
e
model
wa
s
pr
opos
e
d
by
L
in
e
t
al.
[
16
]
f
or
be
tt
e
r
de
tec
ti
on
of
f
a
ult
s
in
wove
n
f
a
b
r
ics
.
I
n
t
his
c
a
s
e
,
a
s
tanda
r
d
f
unc
ti
on
a
s
ymm
e
tr
ic
los
s
f
unc
ti
on
wa
s
int
r
oduc
e
d
in
the
ne
twor
k
to
obtain
be
tt
e
r
r
e
s
ult
s
.
How
e
ve
r
,
the
s
tudy
a
c
knowle
dge
s
li
mi
tations
s
uc
h
a
s
da
tas
e
t
diver
s
it
y,
ge
ne
r
a
li
z
a
ti
on
to
other
domains
,
a
nd
potential
ove
r
f
it
ti
ng.
Ka
z
mi
e
t
al
.
[
17
]
us
e
d
s
im
ple
tec
hniques
li
ke
ha
r
d
th
r
e
s
holdi
ng
f
or
im
a
ge
binar
iza
ti
on
a
nd
identi
f
ied
de
f
e
c
ts
on
the
p
las
ti
c
bott
le
a
t
a
n
a
c
c
ur
a
c
y
r
a
te
of
95%
without
r
e
quir
ing
h
igh
c
om
putation
powe
r
or
e
xtens
ive
tr
a
ini
ng
da
ta,
making
it
highl
y
f
e
a
s
ibl
e
f
or
f
a
s
t
ins
pe
c
ti
on.
T
he
s
ys
tems
ins
pe
c
ted
f
law
s
on
the
c
a
p,
de
nt
on
the
body,
a
nd
t
he
labe
l.
How
e
ve
r
,
a
notable
we
a
kne
s
s
li
e
s
in
the
li
mi
ted
c
a
pa
bil
it
y
of
the
s
ys
tem
c
ompar
e
d
to
a
r
ti
f
icia
l
in
t
e
ll
igenc
e
(
AI
)
ba
s
e
d
a
lgor
it
hms
.
M
a
les
a
a
nd
R
a
jki
e
wic
z
[
1
8]
de
ve
loped
a
n
e
xc
lus
ive
s
ys
tem
f
o
r
pe
t
bott
le
i
ns
pe
c
ti
on
us
ing
a
de
e
p
ne
ur
a
l
ne
twor
k
.
T
he
methodology
i
nvolved
the
c
a
li
br
a
ti
on
of
a
s
ys
tem
us
ing
a
c
he
c
ke
r
boa
r
d
pa
tt
e
r
n
f
or
im
a
ge
a
c
quis
it
ion
f
oll
owe
d
by
a
n
ins
pe
c
ti
on
f
r
a
mew
or
k
us
ing
the
C
us
tom
KSM
a
ppr
o
a
c
h.
T
his
a
ppr
oa
c
h
s
igni
f
ica
ntl
y
r
e
duc
e
s
pr
e
diction
a
nd
t
r
a
i
ning
ti
mes
withou
t
s
a
c
r
if
icing
a
c
c
ur
a
c
y
but
ha
s
a
li
mi
ted
s
c
ope
of
a
ppli
c
a
ti
on
to
s
pe
c
if
ic
indus
tr
ies
a
nd
pr
o
duc
ti
on
li
ne
s
.
Z
hou
e
t
a
l.
[
19]
B
T
M
f
a
s
t
tec
hnique
to
f
ind
de
f
e
c
ts
in
the
body
of
the
glas
s
bott
le.
T
he
metho
d
us
e
d
Ots
u
thr
e
s
holdi
ng
to
s
e
gment
the
template
of
e
a
c
h
im
a
ge
,
then
binar
y
template
matc
hing
(
B
T
M
)
is
a
ppli
e
d
on
a
f
i
lt
e
r
e
d
im
a
ge
to
ob
tain
the
bott
l
e
mout
h
bounda
r
y
a
t
the
top
por
ti
on
a
nd
the
de
f
e
c
t
is
ident
if
ied.
T
he
s
tr
e
ngths
of
the
f
r
a
mew
or
k
include
it
s
a
bil
it
y
to
r
e
duc
e
c
omput
a
ti
ona
l
c
ompl
e
xit
y
th
r
ough
down
s
a
mpl
ing
a
nd
obtain
a
c
c
ur
a
te
loca
li
z
a
ti
on
r
e
s
ult
s
with
a
low
loca
li
z
a
ti
on
e
r
r
or
o
f
a
bout
1
.
5
a
nd
a
f
a
s
ter
pr
oc
e
s
s
ing
ti
me
of
a
ppr
oxi
mate
ly
4
.
3
ms
.
How
e
ve
r
,
it
s
d
r
a
wba
c
k
is
the
li
mi
ted
e
va
luation
da
tas
e
ts
,
whic
h
may
not
f
ull
y
r
e
pr
e
s
e
nt
r
e
a
l
-
wor
ld
s
c
e
na
r
ios
.
A
s
e
gmenta
ti
on
a
ppr
oa
c
h
ba
s
e
d
on
C
NN
wa
s
pr
opos
e
d
by
L
u
e
t
al.
[
20]
f
or
f
law
de
tec
ti
on
on
the
s
ur
f
a
c
e
s
.
A
ne
twor
k
M
R
D
-
Ne
t
whic
h
is
obtaine
d
us
ing
the
ne
twor
ks
mul
ti
s
c
a
le
f
e
a
tur
e
e
nha
nc
e
me
nt
f
us
ion
(
M
F
E
F
)
a
nd
r
e
ve
r
s
e
a
tt
e
nti
on
(
RA
)
wa
s
us
e
d.
T
he
M
obil
e
Ne
tV2
model
wa
s
pr
e
-
tr
a
ined
to
e
xtr
a
c
t
f
e
a
tur
e
maps
on
mul
ti
s
c
a
le
a
nd
obtaine
d
good
a
c
c
ur
a
c
y
on
s
mall
s
ize
s
a
mpl
e
s
.
Gong
e
t
al
[
21]
us
e
d
de
f
or
mable
template
matc
hing
f
or
de
f
e
c
t
loca
li
z
a
ti
on
on
the
labe
ls
of
the
bott
le.
E
ve
n
mi
nute
de
f
e
c
ts
we
r
e
de
tec
ted
by
a
pplyi
ng
a
da
pti
ve
thr
e
s
holdi
ng
on
the
labe
ls
with
C
hines
e
letter
s
.
How
e
ve
r
,
the
s
c
ope
is
li
m
it
e
d
only
to
labe
ls
of
the
bott
le
.
L
v
e
t
al.
[
22]
im
pr
ov
is
e
d
YO
L
Ov3
a
nd
us
e
d
f
ou
r
s
c
a
les
to
de
tec
t
e
ve
n
m
inut
e
objec
ts
.
A
s
pa
ti
a
l
pyr
a
mi
d
s
tr
uc
tu
r
e
wa
s
in
tr
oduc
e
d
to
f
ind
f
law
s
in
the
glas
s
with
a
n
a
c
c
ur
a
c
y
va
lue
o
f
97
.
66%
.
M
a
e
t
al.
[
23]
us
e
d
the
e
ns
e
mbl
e
lea
r
ning
(
EL
)
a
pp
r
oa
c
h
f
or
f
l
a
w
de
tec
ti
on
of
medic
ine
bott
les
.
Ac
quir
e
d
im
a
g
e
s
we
r
e
pr
e
tr
e
a
ted
to
r
e
duc
e
the
va
r
iation
e
f
f
e
c
t
c
a
us
e
d
due
to
the
c
ha
nge
in
li
ghti
ng
c
ondit
ion
.
M
ult
ipl
e
f
e
a
tur
e
s
we
r
e
e
xtr
a
c
ted
f
or
the
e
f
f
e
c
ti
ve
de
tec
ti
on
of
f
law
s
.
T
he
s
ys
tem
f
e
a
tur
e
s
a
tunnel
s
tr
uc
tur
e
that
s
e
a
ml
e
s
s
ly
int
e
gr
a
tes
int
o
e
xis
ti
ng
p
r
oc
e
s
s
e
s
a
nd
de
vice
s
,
e
ns
ur
ing
a
utom
a
ted
ins
pe
c
ti
on
wi
thout
dis
r
upti
on
.
L
e
ve
r
a
ging
a
high
-
pr
e
c
is
ion
vis
ion
de
tec
ti
on
method
,
the
s
ys
tem
e
mpl
oys
a
loca
l
ba
c
kgr
ound
dif
f
e
r
e
nc
e
a
ppr
oa
c
h
a
nd
im
a
ge
gr
a
y
-
leve
l
e
qua
li
z
a
ti
on
pr
e
pr
oc
e
s
s
ing
to
mi
ti
ga
te
il
lum
ination
e
f
f
e
c
ts
.
A
v
ir
tual
s
im
ulation
in
s
pe
c
ti
on
wa
s
pr
opos
e
d
in
[
24]
to
il
lus
tr
a
te
the
f
law
de
tec
ti
on
in
be
e
r
bott
les
.
How
e
ve
r
,
the
s
tudy
wa
s
li
mi
ted
to
vir
tually
de
mons
tr
a
ti
ng
how
f
law
s
c
a
n
be
f
ound
in
be
e
r
bott
les
,
a
nd
ther
e
f
o
r
e
it
s
pr
a
c
ti
c
a
li
ty
is
ve
r
y
li
mi
ted
in
s
c
ope
.
S
im
il
a
r
ly,
a
modeling
a
nd
a
nim
a
ti
on
a
p
pli
c
a
ti
on
wa
s
pr
opos
e
d
to
uti
li
z
e
the
ga
me
e
ngi
ne
UE
4.
Xu
a
nd
Hua
ng
[
25
]
p
r
opos
e
d
S
C
-
C
AE
Ne
t,
f
or
glas
s
de
f
e
c
t
de
tec
ti
on,
a
ddr
e
s
s
ing
c
ha
ll
e
nge
s
in
manua
l
a
nnotation
a
c
quis
it
ion.
B
y
e
mpl
oying
a
c
onv
olut
ional
a
uto
-
e
nc
ode
r
f
ol
lowe
d
by
im
a
ge
di
f
f
e
r
e
nc
e
c
omput
a
ti
on,
the
method
e
f
f
e
c
ti
ve
ly
de
tec
ts
de
f
e
c
t
r
e
gions
.
T
he
int
r
oduc
ti
on
of
s
kip
c
onne
c
ti
ons
a
nd
we
ight
e
d
s
tr
uc
tur
a
l
s
im
il
a
r
i
ty
los
s
e
nha
nc
e
s
de
tec
ti
on
pr
e
c
is
ion.
T
he
ou
tcome
s
with
S
C
-
C
AE
Ne
t
s
how
a
n
a
c
c
ur
a
c
y
of
0
.
9475,
a
long
with
low
e
s
c
a
pe
a
nd
ove
r
kil
l
r
a
tes
,
de
tec
ti
ng
de
f
e
c
ts
in
jus
t
32
ms
pe
r
im
a
ge
.
How
e
ve
r
,
it
s
potential
we
a
kne
s
s
e
s
may
includ
e
li
mi
tations
in
ha
ndli
ng
c
ompl
e
x
de
f
e
c
t
pa
tt
e
r
ns
a
nd
va
r
iabili
ty
in
glas
s
types
.
He
e
t
al.
[
26]
int
r
oduc
e
d
a
n
ins
pe
c
ti
on
s
ys
tem
high
-
a
ngle
a
nd
mul
ti
-
view
(
HA
M
V
)
to
de
tec
t
bott
le
c
a
p
de
f
e
c
ts
.
T
he
method
us
e
d
the
ba
c
kgr
ound
r
e
c
ons
tr
uc
ti
on
of
li
ne
s
tr
uc
tur
e
e
leme
nt
(
B
R
L
S
E
)
tec
hnique
whic
h
f
oc
us
e
d
on
the
u
s
e
of
li
ne
s
tr
uc
tur
e
e
leme
nts
f
or
ba
c
kgr
ound
textur
e
r
e
c
ons
tr
uc
ti
on.
Non
-
s
kid
ba
r
int
e
r
f
e
r
e
nc
e
wa
s
e
li
mi
na
ted
to
a
gr
e
a
t
e
xtent
to
de
tec
t
f
law
s
e
f
f
e
c
ti
v
e
ly.
T
he
s
ys
tem
ins
pe
c
ted
400
piec
e
s
of
c
a
ps
in
a
mi
nute
a
t
a
n
a
c
c
ur
a
c
y
of
mo
r
e
than
95
%
.
S
a
r
i
a
nd
Ula
s
[
27
]
pr
opos
e
d
a
de
e
p
lea
r
ning
mod
e
l
ba
s
e
d
on
YO
L
O
v
3
f
or
de
f
e
c
t
de
tec
ti
on
o
f
glas
s
jar
s
ur
f
a
c
e
s
.
Ac
hieving
a
pe
r
f
o
r
manc
e
of
94
.
65%
f
or
r
a
w
da
ta,
the
s
tudy
e
xplo
r
e
s
c
olor
s
pa
c
e
c
onve
r
s
ion
(
C
S
C
)
tec
hniques
,
s
pe
c
if
ica
ll
y
hue
,
s
a
tur
a
ti
on
,
v
a
lue
(
HSV
)
a
nd
C
I
E
-
L
a
b
L
uv,
f
or
im
a
ge
pr
e
pr
o
c
e
s
s
ing.
W
hil
e
the
HSV
method
s
hows
a
de
c
r
e
a
s
e
in
p
e
r
f
or
manc
e
,
the
C
I
E
-
L
a
b
L
uv
method,
pa
r
ti
c
ula
r
ly
with
a
da
pti
ve
his
togr
a
m
e
qua
li
z
a
ti
on,
e
xhibi
ts
s
igni
f
ica
nt
im
pr
ove
ments
,
with
maximum
r
e
c
a
ll
,
pr
e
c
is
ion,
a
nd
F
1
-
s
c
or
e
s
ur
pa
s
s
ing
97%
.
F
ur
ther
mor
e
,
c
ompar
a
ti
v
e
a
na
lys
is
with
F
a
s
ter
R
-
C
NN
de
mons
tr
a
te
s
Y
OL
O
v
3's
s
upe
r
ior
pe
r
f
or
manc
e
,
e
s
pe
c
ially
whe
n
c
ombi
ne
d
with
a
da
pti
ve
his
togr
a
m
e
qua
li
z
a
ti
on
in
the
C
I
E
-
L
a
b
L
uv
method.
Xie
e
t
al.
[
28]
pr
opos
e
d
F
E
-
YO
L
O
model
f
or
r
e
a
l
-
ti
me
a
ppli
c
a
ti
ons
.
T
he
YO
L
O
model
wa
s
im
pr
ovis
e
d
by
int
r
oduc
ing
a
ne
twor
k
ba
s
e
d
on
a
f
e
a
tur
e
pyr
a
mi
d
a
nd
a
pr
e
diction
box
r
e
gr
e
s
s
ion
m
ode
l.
k
-
mea
ns
+
+
wa
s
us
e
d
late
r
to
e
nha
nc
e
the
c
onve
r
g
e
nc
e
a
nd
obtain
be
tt
e
r
de
tec
ti
on
a
c
c
ur
a
c
ies
c
ompar
e
d
to
e
xis
ti
ng
YO
L
O
-
ba
s
e
d
models
.
M
a
o
e
t
al.
[
29]
pr
opos
e
d
a
n
a
dva
nc
e
d
ne
twor
k
Dy
-
YO
L
Ov5s
ba
s
e
d
on
de
e
p
lea
r
ning
f
or
f
law
s
ins
pe
c
ti
on
of
s
mar
tphone
glas
s
.
T
he
s
tudy
pr
opos
e
d
a
model
c
a
ll
e
d
dyna
mi
c
he
a
d
f
or
the
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5185
-
5
199
5188
de
tec
ti
on
of
f
law
s
dyna
mi
c
a
ll
y.
Hu
e
t
a
l.
[
30]
pr
opos
e
d
low
-
li
ght
e
nha
nc
e
ment
a
nd
mul
ti
-
s
c
a
le
f
e
a
tur
e
e
xtr
a
c
ti
on
(
LE
–
M
S
F
E
–
DD
Ne
t
)
to
de
tec
t
f
law
s
in
the
im
a
ge
s
take
n
in
dim
li
ghti
ng
c
ondit
io
ns
.
T
he
y
int
r
oduc
e
d
a
block
to
ove
r
c
ome
il
lum
ination
va
r
i
a
bil
it
y
a
long
with
the
us
e
of
the
S
E
-
F
P
model.
T
he
mul
ti
-
s
c
a
le
f
e
a
tur
e
e
xtr
a
c
ti
on
tec
hnique
a
ppli
e
d
a
f
ter
thi
s
s
tage
r
e
s
ult
e
d
in
e
f
f
e
c
ti
ve
f
law
de
tec
ti
on
with
a
n
a
ve
r
a
ge
a
c
c
ur
a
c
y
of
94.
3%
.
An
ins
pe
c
ti
on
s
ys
tem
de
ve
l
ope
d
by
W
a
ng
e
t
al.
[
11]
f
or
de
tec
ti
ng
s
c
a
le
de
f
e
c
ts
on
medic
a
l
s
yr
inges
.
B
y
pr
opos
ing
a
two
-
s
tage
f
r
a
mew
or
k
c
ompr
is
ing
a
s
c
a
le
e
xtr
a
c
ti
on
ne
twor
k
(
S
e
N
e
t)
a
nd
a
s
c
a
le
de
f
e
c
t
dis
c
r
im
inator
(
S
DD
)
,
the
s
tudy
of
f
e
r
s
a
n
innovative
a
ppr
oa
c
h
to
de
f
e
c
t
loca
li
z
a
ti
o
n.
S
e
Ne
t
e
f
f
e
c
ti
ve
ly
e
xtr
a
c
ts
the
main
s
tr
uc
tu
r
e
o
f
s
c
a
les
,
while
S
DD
de
tec
ts
a
nd
labe
ls
s
c
a
le
de
f
e
c
ts
.
T
he
s
tr
e
ngths
of
the
method
li
e
in
it
s
a
bil
it
y
to
a
c
hieve
high
de
tec
ti
on
pe
r
f
or
manc
e
with
a
s
mall
number
o
f
ne
ga
ti
ve
s
a
mpl
e
s
.
How
e
ve
r
,
the
li
mi
tation
may
include
the
ne
e
d
f
o
r
f
ur
ther
va
li
da
ti
on
a
c
r
os
s
diver
s
e
s
yr
inge
manuf
a
c
tur
ing
e
nvir
onments
a
nd
de
f
e
c
t
va
r
iations
to
e
ns
ur
e
r
obus
t
pe
r
f
or
manc
e
in
r
e
a
l
-
wor
ld
s
e
tt
ings
.
T
a
ng
e
t
al.
[
31]
us
e
d
e
xtr
a
c
ti
on
of
a
da
pti
ve
bounda
r
y
a
nd
e
s
s
e
nti
a
l
f
e
a
tu
r
e
s
,
c
r
oppe
d
the
im
a
ge
s
to
ge
t
unif
or
m
c
on
f
igur
a
ti
ons
,
a
nd
then
pr
oduc
e
d
mas
ke
d
im
a
ge
s
whic
h
a
r
e
f
e
d
in
t
o
the
de
tec
ti
on
s
ys
tem
to
de
tec
t
de
f
e
c
ts
on
the
a
lum
inum
pr
of
il
e
us
ing
mea
n
s
tr
uc
tur
a
l
s
im
il
a
r
it
y
index
me
a
s
ur
e
(
M
S
S
I
M
)
c
ompar
is
on
.
T
he
a
c
c
ur
a
c
y
of
the
s
ys
tem
r
e
a
c
he
d
97.
7%
.
Va
r
ious
a
pp
r
oa
c
he
s
we
r
e
pr
opos
e
d
to
de
tec
t
f
law
s
on
the
s
ur
f
a
c
e
.
T
he
f
e
a
tur
e
s
li
ke
c
olor
,
textur
e
,
a
nd
s
ha
pe
a
r
e
c
ons
ider
e
d
f
or
the
s
tudy
[
32]
.
M
a
ny
a
ppr
oa
c
he
s
we
r
e
pr
opos
e
d
f
or
de
f
e
c
t
de
tec
ti
on
on
s
tee
l
s
ur
f
a
c
e
s
[
33]
.
De
s
hpa
nde
e
t
al
.
[
34]
us
e
d
a
S
iame
s
e
C
NN
s
ys
tem
to
identif
y
f
law
s
on
the
s
te
e
l
s
ur
f
a
c
e
a
nd
obtaine
d
a
tes
ti
ng
a
c
c
ur
a
c
y
o
f
92
.
55%
.
C
a
o
e
t
al.
[
35]
us
e
d
the
YO
L
Ov7
model
f
or
f
law
de
tec
ti
on
on
the
c
hip
pa
c
ka
ge
.
Anc
hor
f
r
a
mes
we
r
e
c
lus
ter
e
d
us
ing
K
-
mea
ns
+
+
,
C
B
A
M
wa
s
mer
ge
d
with
R
F
B
s
tr
uc
tur
e
,
a
nd
a
s
a
r
e
s
ult
,
C
P
-
C
lus
ter
wa
s
int
r
oduc
e
d
whic
h
led
to
a
n
incr
e
a
s
e
in
the
a
c
c
ur
a
c
y
of
f
law
de
tec
ti
on.
De
e
p
lea
r
ning
models
a
r
e
in
us
e
f
o
r
the
de
tec
ti
on
of
s
u
r
f
a
c
e
de
f
e
c
ts
[
36]
,
[
37
]
.
A
de
e
p
lea
r
ning
model
b
a
s
e
d
on
R
e
s
N
e
t50
wa
s
pr
opos
e
d
by
Ya
ng
e
t
al
.
[
38
]
.
He
r
e
,
f
e
a
tur
e
e
xtr
a
c
ti
on
wa
s
done
us
ing
M
C
F
block
f
or
pe
r
f
or
manc
e
e
nha
nc
e
ment
of
the
tec
hnique.
A
s
e
gmenta
ti
on
a
nd
de
e
p
lea
r
ning
ba
s
e
d
a
ppr
oa
c
h
wa
s
a
dopted
by
T
a
be
r
nik
e
t
al
.
[
39
]
,
[
40]
f
or
c
r
a
c
k
de
tec
ti
on
on
the
s
ur
f
a
c
e
.
T
he
r
e
a
r
e
models
b
a
s
e
d
on
c
onti
nuous
incr
e
menta
l
lea
r
ning
f
or
e
nha
nc
ing
the
de
tec
ti
on
of
a
ny
ne
w
type
of
de
f
e
c
ts
.
T
he
s
e
AI
e
na
bled
incr
e
menta
l
s
ys
tems
f
a
c
il
it
a
te
lea
r
ning
e
it
he
r
by
im
pleme
nti
ng
c
las
s
or
tas
k
models
.
T
his
model
p
r
o
pos
e
s
the
c
onc
e
pt
of
li
f
e
long
lea
r
ning
ha
ving
va
r
io
us
s
tep
s
s
uc
h
a
s
da
ta
a
c
quis
it
ion,
da
ta
a
na
lys
is
,
e
xpe
r
t
int
e
r
view
s
,
da
ta
pr
e
-
pr
oc
e
s
s
ing,
da
ta
labe
li
ng,
tr
a
i
ning
the
model,
a
nd
model
de
ploym
e
nt
.
T
he
s
ys
tem
will
be
upda
ted
c
onti
nuous
ly
with
ne
w
c
a
tegor
ies
o
f
da
ta
unti
l
no
ne
w
da
ta
is
a
va
il
a
ble
[
41]
.
F
r
om
the
s
tudi
e
s
a
ddr
e
s
s
ing
bott
le
de
f
e
c
ts
to
glas
s
pr
oduc
t
qua
li
ty
c
ontr
ol
a
nd
medic
a
l
s
yr
inge
s
c
a
le
de
tec
ti
on,
f
r
om
a
da
pti
ve
bounda
r
y
e
xt
r
a
c
ti
on
f
o
r
a
lum
inum
pr
o
f
il
e
de
f
e
c
t
de
tec
ti
on
to
S
iame
s
e
C
NN
s
ys
tems
f
or
f
law
identif
ica
ti
on
on
s
tee
l
s
ur
f
a
c
e
s
,
e
a
c
h
a
pp
r
oa
c
h
br
ings
unique
s
tr
e
ngths
a
nd
innovations
to
t
he
f
ield.
F
ur
ther
mor
e
,
the
int
e
gr
a
ti
on
of
de
e
p
lea
r
ning
mo
de
ls
li
ke
YO
L
Ov7
a
nd
R
e
s
Ne
t50,
a
long
wi
th
c
onti
nuous
incr
e
menta
l
lea
r
ning
a
ppr
oa
c
he
s
,
de
mons
tr
a
tes
the
e
volvi
ng
lands
c
a
pe
of
de
f
e
c
t
de
tec
ti
on
method
ologi
e
s
.
De
s
pit
e
their
s
tr
e
ngths
in
a
c
c
ur
a
c
y
a
nd
e
f
f
ic
ienc
y,
thes
e
methods
f
a
c
e
c
ha
ll
e
nge
s
r
e
late
d
to
da
ta
r
e
quir
e
ments
,
c
omput
a
ti
ona
l
c
ompl
e
xit
y
,
a
nd
r
e
a
l
-
wor
ld
a
ppli
c
a
bil
it
y.
M
oving
f
or
wa
r
d
,
the
c
ombi
na
ti
on
of
innovative
tec
hnologi
e
s
with
tr
a
dit
ional
mac
hin
e
vis
ion
a
ppr
oa
c
he
s
holds
p
r
omi
s
e
f
o
r
e
nha
nc
in
g
de
f
e
c
t
de
tec
ti
on
s
ys
tems
a
c
r
os
s
indus
tr
ies
,
pr
ov
idi
ng
f
ur
t
he
r
va
li
da
ti
on
a
nd
r
e
f
ineme
nt
to
e
ns
ur
e
r
obus
t
pe
r
f
or
manc
e
in
diver
s
e
manuf
a
c
tur
ing
e
nvir
onments
.
T
he
objec
ti
ve
o
f
the
pr
opos
e
d
method
is
to
de
te
c
t
de
f
e
c
ts
in
pha
r
mac
e
uti
c
a
l
vials
by
a
ddr
e
s
s
ing
c
r
a
c
ks
,
bubbles
,
blac
k
s
pots
,
a
nd
wr
inkl
e
s
.
T
his
s
tudy
a
c
hieve
s
thi
s
by
int
e
gr
a
ti
ng
hybr
id
im
a
ge
pr
oc
e
s
s
ing
tec
hniques
with
P
S
T
.
T
he
s
tr
uc
tur
e
of
the
pa
pe
r
is
a
s
f
ol
lows
:
s
e
c
ti
on
2
pr
ovides
the
methodology
p
r
opos
e
d
f
or
de
f
e
c
t
identif
ica
ti
on
.
S
e
c
ti
on
3
de
tails
the
e
xpe
r
im
e
nts
c
onduc
ted
a
nd
the
r
e
s
ult
ing
output
s
.
F
inally,
s
e
c
ti
on
4
pr
e
s
e
nts
the
c
onc
lus
ions
dr
a
wn
f
r
om
the
s
tudy.
2.
M
E
T
HO
D
T
he
pr
opos
e
d
s
ys
tem
invol
ve
s
a
mul
ti
-
s
tage
pr
oc
e
s
s
uti
li
z
ing
hybr
id
im
a
ge
pr
oc
e
s
s
ing
tec
hniques
f
or
de
f
e
c
t
de
tec
ti
on
in
pha
r
mac
e
uti
c
a
l
vials
,
a
s
il
lus
tr
a
ted
in
F
igur
e
3.
T
his
mul
ti
-
s
tage
pr
oc
e
s
s
e
ns
ur
e
s
a
c
ompr
e
he
ns
ive
a
na
lys
is
by
int
e
gr
a
ti
ng
va
r
ious
a
d
va
nc
e
d
im
a
ge
pr
oc
e
s
s
ing
methods
,
e
nha
nc
ing
the
a
c
c
ur
a
c
y
a
nd
r
e
li
a
bil
it
y
o
f
de
f
e
c
t
de
tec
ti
on.
B
y
leve
r
a
gin
g
thes
e
tec
hniques
,
the
s
ys
tem
e
f
f
e
c
ti
ve
ly
identi
f
ies
a
nd
c
a
tegor
ize
s
de
f
e
c
ts
,
im
pr
oving
qua
li
ty
c
ontr
ol
mea
s
ur
e
s
in
pha
r
mac
e
uti
c
a
l
vial
pr
oduc
ti
on
.
T
he
pr
oc
e
s
s
s
tar
ts
with
the
im
a
ge
a
c
quis
it
ion
pha
s
e
whe
r
e
good
-
r
e
s
olut
ion
im
a
ge
s
of
pha
r
mac
e
uti
c
a
l
vials
a
r
e
c
a
ptur
e
d
us
ing
a
s
pe
c
ialize
d
indus
tr
ial
c
a
mer
a
.
T
he
a
c
quir
e
d
im
a
ge
s
unde
r
go
pr
e
pr
oc
e
s
s
ing
to
e
nha
nc
e
c
ontr
a
s
t,
r
e
move
nois
e
,
a
nd
pr
e
pa
r
e
them
f
or
s
ubs
e
que
nt
a
na
lys
is
.
I
n
the
pr
e
pr
oc
e
s
s
ing
s
tage
,
F
NL
M
f
il
ter
ing
is
a
ppli
e
d
to
the
a
c
quir
e
d
im
a
ge
to
r
e
du
c
e
nois
e
a
nd
e
nha
nc
e
im
a
ge
qua
li
ty,
pr
e
s
e
r
ving
s
tr
uc
tur
a
l
de
tails
.
T
he
n
Ga
us
s
ian
low
-
pa
s
s
f
il
ter
is
a
ppli
e
d
to
f
ur
ther
s
moot
he
n
the
im
a
ge
,
s
uppr
e
s
s
ing
high
-
f
r
e
que
nc
y
nois
e
.
I
n
vial
i
mage
s
,
whe
r
e
de
f
e
c
ts
mi
ght
be
s
ubtl
e
a
nd
int
e
r
twined
with
the
s
tr
uc
tur
e
,
pr
e
s
e
r
vi
ng
thes
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
P
has
e
s
tr
uc
tur
e
s
-
bas
e
d
hy
br
id
appr
oac
he
s
for
de
fec
t
de
tec
ti
on
in
v
ial
s
(
V
is
hw
anatha
C
.
R
.
)
5189
de
tails
is
c
r
uc
ial
f
or
a
c
c
ur
a
te
de
f
e
c
t
de
tec
ti
on.
T
h
e
pha
s
e
f
e
a
tur
e
s
a
r
e
e
xtr
a
c
ted
us
ing
the
P
S
T
a
lgor
it
hm
a
nd
it
is
then
s
e
gmente
d
us
ing
Ots
u
T
hr
e
s
holdi
ng
.
Ots
u
thr
e
s
holdi
ng,
a
wide
ly
us
e
d
s
e
gmenta
ti
on
metho
d,
a
ids
in
a
c
c
ur
a
tely
s
e
pa
r
a
ti
ng
de
f
e
c
ts
f
r
om
the
ba
c
kgr
ound
.
F
ur
ther
r
e
f
ineme
nt
of
de
f
e
c
t
-
r
e
late
d
f
e
a
tur
e
s
is
c
a
r
r
ied
out
thr
ough
Ga
us
s
ian
f
il
ter
ing
f
o
r
e
nha
nc
e
d
c
lar
it
y
.
M
or
phologi
c
a
l
dil
a
ti
on
is
then
a
ppli
e
d
to
e
nha
nc
e
de
f
e
c
t
r
e
gions
,
a
idi
ng
in
pr
e
c
is
e
loca
li
z
a
ti
on
a
nd
bou
nda
r
y
r
e
f
ineme
nt
.
T
he
n
the
potential
de
f
e
c
t
a
r
e
a
s
a
r
e
pinpoi
nted
by
identi
f
ying
r
e
gions
of
int
e
r
e
s
t
(
R
OI
)
.
T
he
c
ombi
na
ti
on
of
thes
e
s
teps
r
e
s
ult
s
in
the
loc
a
li
z
a
ti
on
of
de
f
e
c
ts
withi
n
the
s
pe
c
if
ied
R
OI
.
T
he
c
ontour
s
r
e
pr
e
s
e
nt
the
bounda
r
ies
of
the
de
f
e
c
ts
,
a
nd
bounding
r
e
c
tangle
s
a
r
e
dr
a
wn
a
r
ound
them
f
or
vis
ua
li
z
a
ti
on
a
nd
a
na
lys
is
.
T
he
de
c
is
ion
whe
ther
to
a
c
c
e
pt
th
e
vial
o
r
r
e
jec
t
is
made
ba
s
e
d
on
thi
s
s
tep
f
or
be
tt
e
r
qua
li
ty
c
ontr
ol
a
s
s
e
s
s
ment.
F
igur
e
3.
T
he
f
low
of
ope
r
a
ti
ons
f
or
de
f
e
c
t
ins
pe
c
ti
on
in
pha
r
mac
e
uti
c
a
l
vials
2.
1.
F
as
t
n
on
-
local
m
e
an
s
f
il
t
e
r
in
g
F
a
s
t
non
-
loca
l
mea
ns
(
F
NL
M
)
f
il
ter
ing
is
a
de
nois
ing
tec
hnique
c
omm
only
us
e
d
in
im
a
ge
pr
oc
e
s
s
ing.
F
NL
M
is
a
da
pti
ve
a
nd
c
ons
ider
s
s
im
il
a
r
it
ies
a
mong
dif
f
e
r
e
nt
im
a
ge
r
e
gions
pr
e
s
e
r
ving
s
tr
uc
tur
a
l
inf
or
mation.
I
n
vial
im
a
ge
s
,
whe
r
e
de
f
e
c
ts
mi
gh
t
manif
e
s
t
a
s
a
nomalies
a
mong
other
wis
e
unif
or
m
a
r
e
a
s
,
a
da
pti
ve
f
il
ter
ing
he
lps
in
dis
ti
nguis
hing
potential
de
f
e
c
ts
f
r
om
the
ba
c
kgr
ound.
I
t
e
f
f
e
c
ti
ve
ly
r
e
duc
e
s
nois
e
without
e
xc
e
s
s
ively
s
moot
hing
e
dge
s
or
bounda
r
ies
[
42]
.
T
his
is
be
ne
f
icia
l
in
vial
im
a
ge
s
whe
r
e
de
f
e
c
ts
mi
ght
be
loca
ted
ne
a
r
e
dge
s
or
bounda
r
ies
,
e
ns
ur
ing
that
thes
e
de
f
e
c
t
-
r
e
late
d
de
tail
s
a
r
e
r
e
taine
d.
T
he
F
NL
M
is
de
f
ined
a
s
in
(
1)
:
[
]
(
)
=
∑
(
,
)
(
)
=
(
1)
w
he
r
e
the
a
c
quir
e
d
unf
il
ter
e
d
im
a
ge
is
de
noted
b
y
whos
e
c
oor
dinate
va
lues
a
r
e
r
e
pr
e
s
e
nted
by
a
nd
.
T
he
dis
tanc
e
we
ight
is
r
e
pr
e
s
e
nted
by
(
,
)
whos
e
pixel
s
im
il
a
r
it
y
is
given
by
a
nd
with
the
c
ondit
ion
0
≤
(
,
)
≤
1
s
a
ti
s
f
ied.
T
he
dis
tanc
e
we
ight
(
,
)
is
de
f
ined
a
s
in
(
2)
:
(
,
)
=
1
(
)
−
‖
(
)
−
(
)
‖
2
,
2
2
(
2)
I
n
(
2)
,
the
we
ight
e
d
E
uc
li
de
a
n
dis
tanc
e
is
r
e
pr
e
s
e
nted
by
the
ter
m
‖
(
)
−
(
)
‖
2
,
2
.
I
n
thi
s
,
t
he
s
tanda
r
d
de
viation
is
given
by
a
nd
f
il
ter
ing
e
f
f
e
c
ti
ve
ne
s
s
is
r
e
pr
e
s
e
nted
by
the
s
moot
hing
va
r
iable
.
T
he
ter
m
(
)
is
a
c
ons
tant
va
lue
r
e
pr
e
s
e
nti
ng
a
nor
malize
d
f
a
c
tor
.
I
t
is
given
a
s
in
(
3
)
:
(
)
=
∑
−
‖
(
)
−
(
)
‖
2
,
2
2
(
3)
2.
2.
P
h
as
e
s
t
r
e
t
c
h
t
r
an
s
f
or
m
I
n
the
de
f
e
c
t
de
tec
ti
on
pr
oc
e
s
s
,
pha
s
e
s
tr
e
tch
tr
a
ns
f
or
m
(
P
S
T
)
s
e
r
ve
s
a
s
a
powe
r
f
ul
ini
ti
a
l
s
tep
in
f
e
a
tur
e
e
xtr
a
c
ti
on
.
T
he
P
S
T
is
a
c
omput
a
ti
ona
l
i
maging
tec
hnique
that
c
a
n
e
f
f
e
c
ti
ve
ly
c
a
ptur
e
a
n
d
a
na
lyze
textur
e
s
,
s
ha
pe
s
,
a
nd
s
tr
uc
tur
a
l
f
e
a
tur
e
s
withi
n
a
n
im
a
ge
.
P
S
T
is
us
e
d
f
or
e
xtr
a
c
ti
ng
f
e
a
tur
e
s
ba
s
e
d
on
pha
s
e
inf
or
mation
that
he
lps
to
identi
f
y
potential
de
f
e
c
ts
in
the
vial
im
a
ge
s
f
oc
us
ing
on
textur
a
l
va
r
iati
ons
a
nd
s
tr
uc
tur
a
l
a
nomalies
.
P
S
T
e
mphas
ize
s
identif
ying
ir
r
e
gular
it
ies
that
may
s
igni
f
y
c
r
a
c
ks
,
i
mpu
r
it
ies
,
or
bubbles
on
the
s
ur
f
a
c
e
of
the
vials
.
I
t
highl
ight
s
va
r
iations
in
the
im
a
ge
's
pha
s
e
to
identif
y
potential
d
e
f
e
c
ts
.
A
2D
pha
s
e
ke
r
ne
l
f
unc
ti
on
is
us
e
d
f
or
f
indi
ng
e
dge
-
r
e
late
d
s
p
a
ti
a
l
inf
or
mation
in
P
S
T
.
T
his
inf
or
mation
is
r
e
pr
e
s
e
nted
in
the
f
r
e
que
nc
y
domai
n.
S
uppos
e
a
vial
im
a
ge
is
in
a
higher
f
r
e
que
nc
y
domain,
then
P
S
T
a
ppli
e
s
a
high
pha
s
e
va
lue
ther
e
by
de
te
c
ti
ng
e
dge
s
ba
s
e
d
on
f
r
e
que
nc
y
-
de
pe
nde
nt
a
tt
r
ibu
tes
.
T
his
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5185
-
5
199
5190
ope
r
a
ti
on
is
r
e
pr
e
s
e
nted
a
s
in
(
4)
,
whe
r
e
a
nd
a
r
e
2D
s
pa
ti
a
l
va
lues
of
the
input
im
a
ge
de
noted
by
[
,
]
,
whos
e
pha
s
e
pr
of
il
e
output
is
de
f
ined
by
(
,
)
.
(
,
)
=
∠
〈
2
{
[
,
]
∙
[
,
]
∙
2
{
[
,
]
}
}
〉
(
4)
w
he
r
e
∠
de
notes
the
a
ngle
ope
r
a
to
r
,
2D
f
a
s
t
F
our
ier
tr
a
ns
f
or
m
(
F
F
T
)
is
de
noted
by
2
a
nd
2D
inver
s
e
f
a
s
t
F
our
ier
tr
a
ns
f
or
m
is
de
noted
by
2
,
a
nd
a
r
e
2D
f
r
e
que
nc
y
pa
r
a
mete
r
s
.
[
,
]
de
notes
the
loca
li
z
a
ti
o
n
ke
r
ne
l’
s
f
r
e
que
nc
y
r
e
s
pons
e
a
nd
the
wa
r
pe
d
p
ha
s
e
ke
r
ne
l
is
de
noted
by
[
,
]
whic
h
is
a
nonl
in
e
a
r
f
r
e
que
nc
y
-
de
pe
nde
nt
pha
s
e
de
f
ined
a
s
in
(
5)
:
[
,
]
=
∙
[
,
]
(
5)
T
he
pha
s
e
ke
r
ne
l
de
r
ivative
pr
of
il
e
is
de
noted
by
[
,
]
is
c
ons
ider
e
d
to
be
a
li
ne
a
r
f
unc
ti
on
c
onc
e
r
ning
pa
r
a
mete
r
s
a
nd
.
S
uc
h
pr
o
f
il
e
is
obtaine
d
by
us
i
ng
a
n
inver
s
e
tange
nt
f
unc
ti
on
p
r
oduc
ing
the
P
S
T
ke
r
ne
l
pha
s
e
a
s
given
in
(
6)
:
[
,
]
=
[
,
]
=
[
]
=
∙
∙
∙
−
1
(
∙
)
−
(
1
2
)
∙
ln
(
1
+
(
∙
)
2
)
∙
∙
−
1
(
∙
)
−
(
1
2
)
∙
ln
(
1
+
(
∙
)
2
)
(
6)
w
he
r
e
=
√
2
+
2
,
=
−
1
(
/
)
a
nd
de
notes
the
max
f
r
e
que
nc
y
.
T
he
ph
a
s
e
pr
of
il
e
pa
r
a
mete
r
s
wa
r
p
a
nd
s
tr
e
ngth
a
r
e
r
e
pr
e
s
e
nted
by
a
nd
r
e
s
pe
c
t
ively.
T
he
y
a
r
e
us
e
d
to
bui
ld
pha
s
e
loca
li
z
a
ti
on
ke
r
ne
ls
a
nd
he
lp
e
xtr
a
c
t
the
e
dge
inf
o
r
mation
us
ing
the
thr
e
s
hold
va
lues
ge
ne
r
a
ted.
T
he
P
S
T
-
e
xtr
a
c
ted
f
e
a
tur
e
s
r
e
late
d
to
pha
s
e
inf
or
mation
s
e
r
ve
a
s
a
ba
s
i
s
f
or
s
ubs
e
que
nt
pr
oc
e
s
s
ing
s
teps
in
the
de
f
e
c
t
de
tec
ti
on
pipeline
.
T
he
s
e
f
e
a
tur
e
s
a
r
e
of
ten
r
e
f
ined
a
nd
a
na
lyze
d
f
u
r
ther
us
ing
a
ddit
ional
im
a
ge
pr
oc
e
s
s
ing
tec
hniques
to
pr
e
c
is
e
ly
loca
li
z
e
a
nd
c
las
s
if
y
the
de
f
e
c
ts
withi
n
the
vial
im
a
ge
s
.
2.
2.
1.
Algor
it
h
m
f
or
f
e
at
u
r
e
e
xt
r
ac
t
ion
u
s
in
g
P
S
T
T
he
a
lgor
it
hm
invol
ve
s
tr
a
ns
f
or
mi
ng
the
im
a
g
e
,
c
ons
tr
uc
ti
ng
a
P
S
T
ke
r
ne
l,
e
xt
r
a
c
ti
ng
pha
s
e
f
e
a
tur
e
s
,
a
nd
pe
r
f
or
mi
ng
mor
phologi
c
a
l
ope
r
a
ti
ons
to
de
tec
t
potential
de
f
e
c
ts
ba
s
e
d
on
pha
s
e
va
r
iatio
ns
.
1)
I
nit
ializa
ti
on:
S
e
t
the
pa
r
a
mete
r
=
0.
5
.
Ge
ne
r
a
te
a
mes
h
gr
id
r
e
pr
e
s
e
nti
ng
the
im
a
ge
c
oor
dinate
s
withi
n
the
r
a
nge
of
–
to
f
or
both
the
x
a
nd
y
a
xe
s
.
2)
C
onve
r
s
ion
to
polar
c
oor
d
inate
s
:
C
onve
r
t
the
C
a
r
tes
ian
c
oor
dinate
s
,
to
polar
c
oo
r
din
a
tes
(
T
HE
T
A
,
R
HO
)
.
3)
L
oc
a
li
z
a
ti
on
ke
r
ne
l
a
ppli
c
a
ti
on
to
r
e
duc
e
nois
e
:
A
pp
l
y
a
l
oc
a
l
iza
t
io
n
ke
r
ne
l
t
o
t
he
o
r
i
gi
na
l
i
ma
ge
in
t
he
f
r
e
qu
e
n
c
y
do
ma
in
to
r
e
du
c
e
no
is
e
.
T
h
is
in
vo
lves
:
i
)
F
o
u
r
ie
r
t
r
a
ns
f
or
m
in
g
th
e
or
i
gi
na
l
im
a
g
e
I
;
ii
)
C
o
ns
t
r
uc
t
in
g
a
nd
a
pp
ly
in
g
a
n
e
xp
on
e
n
ti
a
l
f
i
lt
e
r
(
e
xp
o
)
b
a
s
e
d
o
n
f
r
e
qu
e
n
c
y
d
is
ta
nc
e
(
R
H
O
)
t
o
t
he
F
o
ur
ie
r
-
tr
a
ns
f
o
r
med
im
a
g
e
t
o
r
e
d
uc
e
no
is
e
;
i
ii
)
I
nve
r
s
e
F
ou
r
i
e
r
t
r
a
ns
f
o
r
m
in
g
t
he
f
il
te
r
e
d
i
ma
ge
to
ob
ta
in
f
il
te
r
e
d
im
a
ge
r
e
p
r
e
s
e
nt
e
d
a
s
I
ma
ge
_or
ig
_f
i
lt
e
r
e
d
.
4)
P
S
T
ke
r
ne
l
c
ons
tr
uc
ti
on
a
nd
a
ppli
c
a
ti
on:
C
ons
tr
uc
t
a
P
S
T
ke
r
ne
l
(
P
ST
_K
e
r
ne
l
)
ba
s
e
d
on
the
f
r
e
que
nc
y
dis
tanc
e
(
R
HO
)
a
nd
us
e
r
-
de
f
ined
pa
r
a
mete
r
s
(
W
ar
p_s
tr
e
ngth
a
nd
P
has
e
_s
tr
e
ngth
).
Apply
the
c
ons
tr
uc
ted
P
S
T
ke
r
ne
l
to
the
f
il
ter
e
d
im
a
ge
(
I
mage
_or
ig_f
il
ter
e
d
)
in
the
f
r
e
que
nc
y
dom
a
in
to
obtain
I
mage
_or
ig_f
il
ter
e
d_P
ST
.
5)
F
e
a
tur
e
e
xtr
a
c
ti
on:
C
a
lcula
te
the
pha
s
e
of
the
t
r
a
ns
f
or
med
im
a
ge
(
P
HI
_featur
e
s
)
obtaine
d
f
r
om
the
P
S
T
.
If
M
or
ph_fl
ag
is
0,
the
outpu
t
is
the
pha
s
e
f
e
a
tur
e
s
(
out=
P
HI
_
featur
e
s
).
If
M
or
ph_fl
ag
is
not
0,
f
ur
the
r
pr
oc
e
s
s
ing
s
teps
f
or
de
f
e
c
t
de
tec
ti
on
a
r
e
e
xe
c
uted:
−
T
hr
e
s
hold
the
pha
s
e
im
a
ge
(
P
HI
_
featur
e
s
)
to
identif
y
s
ha
r
p
tr
a
ns
it
ions
us
ing
us
e
r
-
de
f
ined
thr
e
s
holds
(
T
hr
e
s
hold_max
a
nd
T
hr
e
s
hold_min
)
a
nd
r
e
move
nois
e
f
r
om
da
r
k
a
r
e
a
s
.
−
P
e
r
f
or
m
binar
y
mor
pho
logi
c
a
l
ope
r
a
ti
ons
(
thi
n
,
pe
r
im
e
ter
e
xtr
a
c
ti
on
,
e
r
ode
)
to
e
nha
nc
e
de
f
e
c
t
f
e
a
tur
e
s
in
the
t
r
a
ns
f
or
med
im
a
ge
.
T
he
P
S
T
tec
hnique
c
a
n
be
e
f
f
e
c
ti
ve
f
or
de
tec
ti
ng
de
f
e
c
ts
in
vial
im
a
ge
s
by
r
e
duc
ing
nois
e
,
pr
e
s
e
r
vi
ng
e
dge
s
,
e
nha
nc
ing
c
ontr
a
s
t,
a
nd
pr
ov
idi
ng
a
mor
e
s
uit
a
ble
input
f
or
s
ubs
e
que
nt
im
a
ge
pr
oc
e
s
s
ing
s
teps
.
2.
3.
Gau
s
s
ian
f
il
t
e
r
in
g
Ga
us
s
ian
f
il
ter
ing
is
a
ppli
e
d
a
f
ter
P
S
T
to
e
nha
nc
e
the
e
xtr
a
c
ted
f
e
a
tur
e
s
,
f
ur
ther
highl
ight
ing
r
e
leva
nt
s
tr
uc
tur
e
s
or
de
f
e
c
ts
.
T
his
s
tage
im
pr
ove
s
the
vis
ibi
li
ty
of
de
f
e
c
ts
by
a
djus
ti
ng
the
loca
l
c
ontr
a
s
t,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
P
has
e
s
tr
uc
tur
e
s
-
bas
e
d
hy
br
id
appr
oac
he
s
for
de
fec
t
de
tec
ti
on
in
v
ial
s
(
V
is
hw
anatha
C
.
R
.
)
5191
making
s
ubtl
e
a
nomalies
mor
e
dis
c
e
r
nibl
e
.
C
ons
e
que
ntl
y,
the
im
a
ge
is
be
tt
e
r
p
r
e
pa
r
e
d
f
o
r
f
ur
ther
r
e
f
ineme
nt
in
the
de
f
e
c
t
de
tec
ti
on
pr
oc
e
s
s
,
e
ns
ur
ing
mor
e
a
c
c
ur
a
te
a
nd
r
e
li
a
ble
r
e
s
ult
s
.
2.
4.
Ot
s
u
t
h
r
e
s
h
old
in
g
Ots
u's
thr
e
s
holdi
ng
method
is
e
mpl
oye
d
to
de
t
e
r
mi
ne
opti
mal
thr
e
s
holds
r
e
s
ult
ing
in
pr
e
c
is
e
s
e
gmenta
ti
on
of
de
f
e
c
ts
f
r
om
the
ba
c
kgr
ound
f
or
be
tt
e
r
identif
ica
ti
on
of
s
ubtl
e
de
f
e
c
t
a
r
e
a
s
e
xtr
a
c
ted
us
ing
P
S
T
.
T
he
method
wor
ks
on
the
im
a
ge
a
nd
s
e
gr
e
ga
tes
it
int
o
two
c
las
s
e
s
of
pixels
with
dis
ti
nc
t
i
ntens
it
y
leve
ls
,
a
nd
it
ge
ne
r
a
tes
a
n
opti
mum
thr
e
s
hold
va
lu
e
that
mi
nim
ize
s
the
int
r
a
c
las
s
va
r
ianc
e
a
nd
maxi
mi
z
e
s
the
int
e
r
c
las
s
va
r
ianc
e
.
T
o
do
it
,
the
method
f
ir
s
t
c
o
mput
e
s
the
his
togr
a
m
of
the
input
im
a
ge
.
T
he
h
is
togr
a
m
is
then
nor
malize
d
to
ge
ne
r
a
te
pr
oba
bil
it
y
mas
s
f
unc
ti
ons
f
or
the
f
or
e
gr
ound
(
objec
t)
a
nd
ba
c
kgr
ou
nd
(
non
-
objec
t)
c
las
s
e
s
.
T
he
int
e
r
c
las
s
va
r
ianc
e
f
or
e
a
c
h
po
s
s
ibl
e
thr
e
s
hold
is
given
by
(
7)
a
s
:
2
(
)
=
0
(
)
∙
(
0
(
)
−
)
2
+
1
(
)
∙
(
1
(
)
−
)
2
(
7)
w
he
r
e
0
(
)
,
0
(
)
a
nd
1
(
)
,
1
(
)
r
e
pr
e
s
e
nts
c
umul
a
ti
ve
pr
oba
bil
it
ies
a
nd
mea
n
int
e
ns
it
ies
f
or
the
two
c
las
s
e
s
of
im
a
ge
s
.
T
he
global
mea
n
int
e
ns
it
y
of
the
e
nti
r
e
im
a
ge
is
given
by
(
8)
a
s
:
=
∑
∙
(
)
−
1
=
0
(
8)
w
he
r
e
r
e
pr
e
s
e
nts
the
number
o
f
int
e
ns
it
y
leve
ls
.
T
he
opti
mal
thr
e
s
hold
va
lue
is
ge
ne
r
a
ted
ba
s
e
d
on
th
e
maximum
va
lue
ge
ne
r
a
ted
f
or
2
(
)
is
given
by
(
9
)
a
s
:
=
2
(
)
(
9)
w
he
r
e
r
e
pr
e
s
e
nts
the
th
r
e
s
hold
va
lue
that
c
a
n
e
f
f
e
c
ti
ve
ly
s
e
gment
the
f
o
r
e
gr
ound
a
nd
ba
c
kgr
ound
in
a
wa
y
that
maximi
z
e
s
the
in
ter
c
las
s
va
r
ianc
e
.
T
his
s
tep
is
f
oll
owe
d
by
the
a
ppli
c
a
ti
on
of
Ga
us
s
ian
f
il
ter
ing
to
r
e
f
ine
de
f
e
c
t
f
e
a
tur
e
s
f
ur
the
r
.
2.
5.
M
or
p
h
ologi
c
al
d
il
at
io
n
an
d
ROI
i
d
e
n
t
if
icat
ion
T
o
inc
r
e
a
s
e
de
f
e
c
t
de
tec
ti
on
a
c
c
ur
a
c
y,
it
is
im
po
r
tant
to
e
nha
nc
e
the
c
lar
it
y
of
de
f
e
c
t
r
e
gions
a
nd
f
a
c
il
it
a
te
it
s
a
c
c
ur
a
te
loca
li
z
a
ti
on
dil
a
te
mor
pholo
gica
l
ope
r
a
ti
on
is
us
e
d.
M
or
phologi
c
a
l
dil
a
ti
on
is
typi
c
a
ll
y
pe
r
f
or
med
us
ing
a
s
tr
uc
tur
ing
e
leme
nt
(
ke
r
ne
l)
.
T
he
ba
s
ic
idea
is
to
s
li
de
the
ke
r
ne
l
ove
r
the
im
a
ge
,
a
nd
a
t
e
a
c
h
pos
it
ion,
i
f
the
ke
r
ne
l
ove
r
laps
wi
th
f
or
e
gr
o
und
(
non
-
z
e
r
o)
pixels
,
the
c
e
nter
of
the
ke
r
ne
l
is
s
e
t
to
the
f
or
e
gr
ound
in
the
r
e
s
ult
ing
im
a
ge
.
F
or
a
bina
r
y
im
a
ge
a
nd
a
s
tr
uc
tur
ing
e
leme
nt
,
the
dil
a
ti
on
ope
r
a
ti
on
⨁
is
de
f
ined
in
(
10)
.
⨁
(
,
)
=
⋃
(
,
)
∈
(
+
,
+
)
(
10)
w
he
r
e
the
pixel
c
oor
dinate
s
a
r
e
r
e
pr
e
s
e
nted
by
(
,
)
a
nd
the
c
oor
dinate
s
of
the
s
tr
uc
tu
r
ing
e
leme
nt
is
r
e
pr
e
s
e
nted
by
(
,
)
.
M
or
phologi
c
a
l
dil
a
ti
on
a
ids
in
the
identif
ica
ti
on
of
R
OI
s
ba
s
e
d
on
the
pr
oc
e
s
s
e
d
im
a
ge
,
f
oc
us
ing
on
a
r
e
a
s
s
howing
potential
de
f
e
c
ts
.
2.
6.
De
f
e
c
t
d
e
t
e
c
t
ion
an
d
d
e
c
is
ion
-
m
ak
in
g
De
f
e
c
t
de
tec
ti
on
is
done
by
na
r
r
owing
down
to
a
s
pe
c
if
ic
a
r
e
a
of
int
e
r
e
s
t
(
R
OI
)
a
nd
f
indi
ng
the
e
xa
c
t
loca
ti
ons
of
de
f
e
c
ts
withi
n
that
r
e
gion.
T
his
i
nvolves
de
ter
mi
ning
c
ontour
s
with
in
the
R
OI
a
nd
dr
a
wing
r
e
c
tangle
s
a
r
ound
the
identi
f
ied
de
f
e
c
ts
.
T
he
vis
ua
li
z
a
ti
on
include
s
a
tr
a
ns
pa
r
e
nt
r
e
c
tangle
f
or
the
R
OI
a
nd
white
r
e
c
tangle
s
a
r
ound
the
de
f
e
c
ts
.
T
his
loca
li
z
e
d
inf
or
mation
is
us
e
f
ul
f
o
r
a
na
lys
is
,
vis
ua
li
z
a
ti
on,
a
nd
f
ur
ther
de
c
is
ion
-
making
in
qua
li
ty
c
ontr
ol
or
ins
pe
c
ti
on
a
ppli
c
a
ti
ons
.
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
Da
tas
e
t
De
s
c
r
ipt
ion:
Due
to
the
s
pe
c
ialize
d
na
tur
e
of
the
r
e
s
e
a
r
c
h
domain
a
nd
the
lac
k
of
publ
icly
a
va
il
a
ble
da
tas
e
t
s
tailor
e
d
f
or
vial
de
f
e
c
t
ins
pe
c
ti
on,
a
n
in
-
hous
e
da
tas
e
t
wa
s
c
ur
a
ted.
R
e
c
ogn
izing
the
li
mi
tations
inher
e
nt
in
the
da
tas
e
t,
including
it
s
s
ize
a
nd
diver
s
it
y,
mea
s
ur
e
s
we
r
e
take
n
to
mi
ti
g
a
te
bias
thr
ough
e
xtens
ive
a
ugmenta
ti
on
(
r
otation
a
nd
s
c
a
li
ng)
a
nd
va
li
da
ti
on
pr
oc
e
dur
e
s
.
F
igu
r
e
4
il
lus
t
r
a
tes
the
s
e
tu
p
de
s
i
gne
d
f
o
r
c
a
p
tu
r
in
g
vial
im
a
ge
s
,
c
omp
r
is
ing
a
Ne
bu
la
L
E
D
ba
c
k
li
gh
t
s
e
r
vin
g
a
s
the
li
ght
s
o
ur
c
e
,
a
n
in
dus
t
r
ial
c
a
me
r
a
,
a
c
onve
y
or
,
a
nd
a
de
s
ktop
c
omp
ute
r
.
T
he
Ne
b
ula
L
E
D
ba
c
kli
ght
(
Ne
b
ula
-
80R
)
wa
s
c
hos
e
n
f
or
it
s
c
us
t
o
mi
z
a
ble
br
igh
tnes
s
leve
ls
,
pr
ovid
ing
a
da
pta
bil
it
y
t
o
va
r
ying
li
gh
ti
ng
c
o
ndi
ti
o
ns
.
F
o
r
p
r
e
c
is
e
im
a
ge
a
c
quis
i
ti
o
n
,
a
T
a
m
r
on
m
a
c
hin
e
v
is
ion
c
a
mer
a
(
5M
P
,
VC
XG
-
52M
R
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5185
-
5
199
5192
pa
ir
e
d
w
it
h
a
50
mm
B
a
u
mer
le
ns
(
M
118
F
M
50
)
wa
s
ut
il
ize
d.
T
his
c
a
me
r
a
-
lens
c
om
bina
ti
on
o
f
f
e
r
s
a
dva
nc
e
d
f
e
a
tu
r
e
s
s
uc
h
a
s
a
ut
o
-
f
oc
us
a
nd
i
mage
s
tab
il
iza
ti
o
n,
e
ns
u
r
ing
h
igh
-
qua
li
t
y
im
a
ge
s
c
r
uc
ial
f
o
r
a
c
c
ur
a
te
de
f
e
c
t
de
tec
ti
on
.
M
or
e
ove
r
,
the
s
e
tup's
ve
r
s
a
ti
li
ty
e
na
bles
s
e
a
ml
e
s
s
int
e
gr
a
ti
on
int
o
e
xis
ti
ng
mac
hine
vis
ion
s
ys
tems
,
f
a
c
il
it
a
ti
ng
a
utom
a
ti
on
a
nd
e
nha
nc
ing
p
r
oduc
ti
on
e
f
f
icie
nc
y.
T
he
de
s
ktop
c
omput
e
r
,
e
quipped
with
a
2.
8
GH
z
,
5
th
ge
ne
r
a
ti
on
c
or
e
i5
p
r
oc
e
s
s
or
,
8
GB
o
f
R
AM
,
GPU
(
Nvid
ia
R
T
X
3070
)
ope
r
a
ti
ng
a
t
a
f
r
e
que
nc
y
of
1
,
500
M
Hz
,
a
nd
1
T
B
of
s
tor
a
ge
s
pa
c
e
,
c
ompl
e
tes
the
c
onf
igu
r
a
ti
on.
T
he
im
a
ge
s
a
mpl
e
s
a
r
e
mon
oc
hr
ome
im
a
ge
s
ha
ving
r
e
a
s
ona
bly
good
r
e
s
olut
ion.
F
igur
e
4
il
lus
tr
a
tes
the
s
e
tup
de
s
igned
f
or
c
a
ptur
ing
vial
im
a
ge
s
,
c
ompr
is
ing
a
Ne
bula
L
E
D
ba
c
kli
ght
s
e
r
ving
a
s
the
li
ght
s
our
c
e
,
a
n
indus
tr
ial
c
a
mer
a
,
a
c
onve
yor
,
a
nd
a
de
s
ktop
c
omput
e
r
.
F
ig
ur
e
4(
a
)
s
hows
the
c
a
mer
a
pos
it
ioned
f
or
vial
im
a
ge
c
a
ptur
e
,
while
F
igur
e
4
(
b)
dis
plays
the
vial
plac
e
d
with
a
ba
c
kli
ght
f
or
im
a
ge
c
a
ptur
e
.
T
he
s
e
c
onf
igur
a
ti
ons
a
r
e
c
r
it
ica
l
f
or
a
c
hieving
high
-
qua
li
ty
im
a
ge
s
ne
c
e
s
s
a
r
y
f
or
a
c
c
ur
a
te
de
f
e
c
t
de
tec
ti
on.
(
a
)
(
b)
F
igur
e
4.
S
ys
tem
s
e
tup
f
or
vial
im
a
ge
c
a
ptur
e
(
a
)
c
a
mer
a
pos
it
ioned
f
or
vial
im
a
ge
c
a
ptur
e
a
nd
(
b)
v
i
a
l
plac
e
d
with
ba
c
kli
ght
f
o
r
im
a
ge
c
a
ptur
e
T
he
im
pleme
ntation
of
the
pr
opos
e
d
tec
hnique
is
done
us
ing
P
ython.
De
gr
e
e
of
s
moot
hing
(
DoS
)
f
or
NL
M
F
wa
s
s
e
t
to
25
f
or
be
tt
e
r
r
e
s
ult
s
.
T
he
im
a
ge
is
then
pr
e
pr
oc
e
s
s
e
d
us
ing
a
Ga
us
s
ian
low
pa
s
s
f
il
ter
to
f
ur
ther
s
moot
he
n
the
im
a
ge
.
T
he
s
tep
f
o
r
pha
s
e
f
e
a
tur
e
e
xtr
a
c
ti
on
invol
ve
s
ini
ti
a
li
z
ing
pa
r
a
mete
r
(
)
that
r
e
pr
e
s
e
nts
ha
lf
o
f
the
s
pa
ti
a
l
r
a
nge
us
e
d
to
c
r
e
a
te
the
mes
h
g
r
id
of
c
oor
d
inate
s
(
,
)
withi
n
a
r
a
nge
(
-
0.
5
to
0.
5
)
ba
s
e
d
on
the
s
ha
pe
of
the
inpu
t
im
a
ge
r
e
pr
e
s
e
nted
by
a
nd
it
c
ontr
ols
the
s
pa
ti
a
l
domain
of
the
ge
ne
r
a
ted
pha
s
e
ke
r
ne
ls
.
T
he
s
e
ge
ne
r
a
ted
c
oor
dinate
s
a
r
e
c
onve
r
ted
to
polar
c
oo
r
dinate
s
θ,
(
T
H
E
T
A,
R
HO
)
r
e
s
pe
c
ti
ve
ly.
A
loca
li
z
a
ti
on
ke
r
ne
l
is
c
r
e
a
ted
to
r
e
duc
e
nois
e
in
the
or
igi
na
l
im
a
ge
(
)
by
a
pplyi
ng
a
f
r
e
que
nc
y
-
domain
ope
r
a
ti
on.
I
t
c
a
lcula
tes
a
f
il
ter
that
a
c
ts
on
the
F
our
ier
-
tr
a
ns
f
or
med
im
a
ge
.
T
he
P
S
T
ke
r
ne
l
is
then
c
ons
tr
uc
ted
us
ing
the
c
a
lcula
ted
polar
c
oor
dinate
s
(
R
HO
)
a
nd
e
dge
de
tec
ti
on
pa
r
a
mete
r
s
wa
r
p
s
tr
e
ngth
(
)
a
nd
pha
s
e
s
tr
e
ng
th
(
)
.
T
he
va
lues
o
f
a
nd
a
f
f
e
c
t
e
a
c
h
f
r
e
que
nc
y’
s
ph
a
s
e
a
long
with
the
P
S
T
ke
r
ne
l
s
tr
uc
tur
e
a
nd
a
ls
o
it
s
de
r
ivative
pr
o
f
il
e
.
F
igu
r
e
5
s
hows
the
2D
pha
s
e
ke
r
ne
ls
a
nd
thei
r
c
or
r
e
s
ponding
de
r
ivative
p
r
of
il
e
s
f
or
s
ix
dif
f
e
r
e
nt
va
lues
of
W
a
nd
S
va
r
iable
s
.
F
igu
r
e
5
(
a
)
il
lus
tr
a
tes
the
2D
pha
s
e
ke
r
ne
ls
,
while
F
igur
e
5
(
b)
de
picts
the
2D
pha
s
e
de
r
ivative
pr
o
f
il
e
s
.
T
he
f
igu
r
e
s
hows
the
e
f
f
e
c
t
of
W
a
nd
S
va
r
iable
s
on
pha
s
e
ke
r
ne
l
s
tr
uc
tur
e
.
T
his
ke
r
ne
l
is
a
ppli
e
d
to
the
f
il
ter
e
d
im
a
ge
to
p
e
r
f
or
m
the
P
S
T
tr
a
ns
f
or
mation
,
yielding
a
n
im
a
g
e
that
is
pr
oc
e
s
s
e
d
us
ing
P
S
T
.
T
he
n
the
pha
s
e
a
ngle
of
e
a
c
h
pixel
is
e
xtr
a
c
ted
f
r
om
the
P
S
T
-
tr
a
ns
f
or
med
im
a
ge
.
T
o
do
thi
s
,
loca
li
z
a
ti
on
a
nd
P
S
T
tr
a
ns
f
or
mation
ne
e
ds
to
be
a
ppli
e
d.
T
his
in
f
or
mation
c
ha
r
a
c
ter
ize
s
the
ph
a
s
e
s
hif
t
int
r
oduc
e
d
by
the
P
S
T
a
nd
c
a
n
high
li
ght
s
tr
uc
tu
r
a
l
va
r
iations
in
the
i
mage
r
e
late
d
to
potential
de
f
e
c
ts
.
F
igur
e
6
s
hows
va
r
ious
s
tage
s
of
the
vial
de
f
e
c
t
de
tec
ti
on
pr
oc
e
s
s
f
or
a
s
a
mpl
e
im
a
ge
with
a
c
r
a
c
k
on
it
s
body.
T
he
pha
s
e
im
a
ge
is
ge
ne
r
a
ted
us
ing
2
D
I
F
F
T
by
c
onve
r
t
ing
the
f
r
e
que
nc
y
domain
to
th
e
s
pa
ti
a
l
domain.
T
he
f
igur
e
include
s
the
o
r
igi
na
l
im
a
ge
,
nonloca
l
mea
ns
f
il
te
r
ing,
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
us
i
ng
P
S
T
,
Ga
us
s
ian
f
il
ter
ing,
Ots
u's
thr
e
s
holdi
ng,
mor
pholo
gica
l
dil
a
ti
on,
identif
ying
R
OI
,
g
r
ound
t
r
uth
im
a
ge
,
a
nd
de
f
e
c
ts
highl
ight
e
d
,
a
s
s
hown
in
F
igur
e
s
6(
a
)
to
6
(
i)
.
T
he
s
e
s
tage
s
il
lus
tr
a
te
the
tr
a
ns
f
or
mation
of
th
e
im
a
ge
thr
ough
va
r
ious
p
r
oc
e
s
s
ing
s
teps
to
a
c
c
ur
a
tely
de
tec
t
a
nd
highl
ight
de
f
e
c
ts
.
T
he
pha
s
e
of
the
tr
a
ns
f
or
med
i
mage
is
thus
c
a
lcula
ted.
T
he
n
th
r
e
s
holdi
ng
is
a
ppli
e
d
to
identif
y
s
ha
r
p
tr
a
ns
it
ions
in
the
pha
s
e
im
a
ge
,
mar
king
po
tential
de
f
e
c
t
a
r
e
a
s
.
P
ixels
with
pha
s
e
va
lues
a
bove
the
maximum
thr
e
s
hold
va
lue
o
r
be
low
the
m
ini
mum
thr
e
s
hold
va
lue
a
r
e
mar
ke
d
a
s
potential
de
f
e
c
ts
.
Additi
ona
ll
y,
ve
r
y
da
r
k
a
r
e
a
s
in
the
or
igi
na
l
im
a
ge
a
r
e
c
ons
ider
e
d
a
s
nois
e
a
nd
a
r
e
r
e
moved.
B
inar
y
mo
r
phologi
c
a
l
op
e
r
a
ti
ons
(
thi
n,
pe
r
im
e
ter
e
xt
r
a
c
ti
on,
a
nd
e
r
os
ion)
a
r
e
pe
r
f
or
med
to
r
e
f
ine
a
nd
e
nha
nc
e
the
potential
de
f
e
c
t
a
r
e
a
s
,
r
e
s
ult
ing
in
the
P
S
T
pha
s
e
output
a
s
s
hown
in
F
igu
r
e
6(
c
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
P
has
e
s
tr
uc
tur
e
s
-
bas
e
d
hy
br
id
appr
oac
he
s
for
de
fec
t
de
tec
ti
on
in
v
ial
s
(
V
is
hw
anatha
C
.
R
.
)
5193
(
a
)
(
b)
F
igur
e
5.
Va
r
ious
va
lues
of
W
a
nd
S
a
r
e
us
e
d
f
or
o
btaining
(
a
)
2D
pha
s
e
ke
r
ne
ls
a
nd
(
b
)
2D
pha
s
e
de
r
ivative
pr
of
il
e
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
1
4
,
No.
5
,
Oc
tober
2
02
4
:
5185
-
5
199
5194
T
he
output
obtaine
d
a
f
te
r
a
ppli
c
a
ti
on
of
P
S
T
is
pos
t
pr
oc
e
s
s
e
d
a
nd
s
e
gmente
d
us
ing
Ots
u.
F
or
identif
ica
ti
on
of
s
ubtl
e
de
f
e
c
ts
,
it
is
de
laye
d
f
u
r
ther
.
A
binar
y
mas
k
(
r
oi_mas
k
)
is
then
c
r
e
a
ted
with
t
he
s
a
me
dim
e
ns
ions
a
s
the
pr
oc
e
s
s
e
d
dil
a
ted
gr
a
y
im
a
ge
.
A
white
r
e
c
tangle
is
dr
a
wn
on
the
mas
k,
r
e
pr
e
s
e
nti
ng
the
r
e
gion
of
int
e
r
e
s
t
(
R
OI
)
whe
r
e
de
f
e
c
ts
a
r
e
e
xpe
c
t
e
d
to
be
p
r
e
s
e
nt
a
s
s
hown
in
F
igu
r
e
6(
g
)
.
T
he
r
e
c
tangle
is
s
pe
c
if
ied
by
it
s
top
-
lef
t
a
nd
bott
om
-
r
ight
c
oo
r
dinate
s
(
500,
260
)
a
nd
(
1530
,
380)
,
r
e
s
pe
c
ti
ve
ly.
A
vis
ua
li
z
a
ti
on
of
the
gr
ound
tr
uth
is
c
r
e
a
ted
by
ove
r
laying
the
r
oi_mas
k
on
the
or
igi
na
l
pr
oc
e
s
s
e
d
gr
a
y
im
a
ge
(
dil
a
ted)
.
T
he
r
oi_mas
k
is
a
ppli
e
d
to
the
pr
oc
e
s
s
e
d
gr
a
y
im
a
ge
,
r
e
s
ult
ing
in
a
n
im
a
ge
c
ontaining
only
the
de
f
e
c
ts
withi
n
the
R
OI
.
F
or
e
a
c
h
c
ontour
r
e
pr
e
s
e
nt
ing
a
de
f
e
c
t
withi
n
the
R
OI
,
a
white
bounding
r
e
c
t
a
ngle
is
dr
a
wn
a
r
ound
e
a
c
h
de
f
e
c
t
to
vis
ua
ll
y
highl
ight
the
identif
ied
r
e
gions
on
the
output
_image
a
s
s
hown
in
F
igur
e
6(
i)
.
T
he
s
e
r
e
c
tangle
s
highl
ight
the
identi
f
i
e
d
de
f
e
c
ts
.
He
r
e
de
f
e
c
ts
a
r
e
identif
ied
by
f
indi
ng
c
ontour
s
withi
n
a
mas
ke
d
r
e
gion
of
the
im
a
ge
.
E
a
c
h
c
onto
ur
r
e
p
r
e
s
e
nts
the
bounda
r
y
of
a
de
f
e
c
t.
Adjus
ti
ng
the
a
lpha
va
lues
a
ll
ows
f
or
be
tt
e
r
vis
ua
li
z
a
ti
on
of
the
ove
r
la
y
a
nd
tr
a
ns
pa
r
e
nc
y.
F
igur
e
7
s
hows
the
3D
P
S
T
ke
r
ne
l
a
nd
3D
de
r
iv
a
ti
ve
pr
of
il
e
ge
ne
r
a
ted
f
o
r
the
s
a
mpl
e
vial
im
a
ge
given
in
F
igur
e
6
.
T
he
tot
a
l
ti
me
take
n
by
the
s
ys
tem
to
de
tec
t
the
de
f
e
c
t
wa
s
6.
4
ms
.
T
he
s
e
ns
it
ivi
ty
is
c
a
lcula
ted
by
c
ompar
ing
the
pha
s
e
im
a
ge
with
the
gr
ound
t
r
uth
im
a
ge
.
T
a
ble
1
s
hows
the
s
e
ns
it
ivi
ty
obtaine
d
f
or
va
r
ious
va
lues
of
W
a
nd
S
a
long
with
a
g
r
a
ph
f
or
the
s
a
mpl
e
vial
im
a
ge
.
F
r
om
T
a
ble
1,
it
c
a
n
be
obs
e
r
ve
d
that
f
or
the
given
vial
im
a
ge
with
c
r
a
c
k,
the
be
s
t
s
e
ns
it
ivi
ty
obtaine
d
is
0.
986
whe
n
the
va
lue
o
f
W
a
nd
S
is
s
e
t
to
30
a
nd
0.
3
r
e
s
pe
c
ti
ve
ly.
T
he
r
e
f
or
e
,
the
s
a
me
va
lues
of
W
a
nd
S
a
r
e
us
e
d
in
the
e
xpe
r
im
e
nt.
F
igur
e
8
s
hows
the
s
e
ns
it
ivi
ty
e
va
luation
gr
a
ph
f
or
va
r
ious
va
lues
of
W
a
nd
S
.
(
a
)
(
b)
(
c
)
(
d)
(
e
)
(f)
(
g)
(
h)
(
i)
F
igur
e
6
.
S
a
mpl
e
ou
tput
o
f
the
p
r
opos
e
d
tec
hnique
(
a
)
or
igi
na
l
im
a
ge
,
(
b)
nonloca
l
mea
ns
f
il
ter
ing,
(
c
)
f
e
a
tur
e
e
xtr
a
c
ti
on
u
s
ing
P
S
T
,
(
d)
Ga
us
s
ian
f
il
ter
ing
,
(
e
)
Ots
u's
thr
e
s
holdi
ng
,
(
f
)
mor
phologi
c
a
l
dil
a
ti
on
,
(
g)
identi
f
ying
R
OI
,
(
h)
g
r
ound
t
r
uth
i
mage
,
a
nd
(
i
)
de
f
e
c
ts
highl
ight
e
d
F
igur
e
7.
3D
P
S
T
ke
r
ne
l
a
nd
3D
de
r
ivative
pr
of
il
e
ge
ne
r
a
ted
f
or
the
s
a
mpl
e
vial
im
a
ge
Evaluation Warning : The document was created with Spire.PDF for Python.