TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 8, August 201
4, pp. 6173 ~ 6180
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.538
7
6173
Re
cei
v
ed
De
cem
ber 1
8
, 2013; Re
vi
sed
April 7, 2014;
Accept
ed Ap
ril 20, 2014
Image S
e
gmentation of Adhering Bars Based on
Improved Concavity Points Searching Method
Guohua Liu*
, Bingle Liu,
Qiujie Yuan, Zhenhui Hua
ng
Schoo
l of Mechan
ical En
gi
ne
erin
g, T
i
anjin Pol
y
t
e
ch
nic Un
iv
ersit
y
, T
i
anji
n
3
003
87, Ch
ina
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: guoh
ua
lium
a
i
l
@16
3
.com
A
b
st
r
a
ct
It is difficult to track, count an
d separ
ate the
bars movin
g
a
t
a high sp
ee
d
on pro
ductio
n
line for
their overl
ap
und
er occlus
i
on. T
her
efore,
it is necessary to establ
i
s
h a relia
ble,
practical spl
i
tting
mec
h
a
n
is
m for
the adh
ered
b
a
rs. T
h
is paper
propos
ed a n
e
w
solution to th
e prob
le
m of b
a
rs adh
esio
n: the
pla
ne
array c
a
mer
a
w
a
s
utili
z
e
d
to
acq
u
ir
e
the i
m
ag
es of
movin
g
bars
so as t
o
rec
o
g
n
i
z
e
t
he c
entr
o
id
coord
i
nates
of
the b
a
rs en
ds
and
c
o
mp
ute
their ar
ea w
i
th
a Blo
b
a
l
g
o
rithm, tw
o ge
o
m
etric par
a
m
ete
r
s
w
e
re utili
z
e
d t
o
detect a
d
h
e
r
ed b
a
rs, an
d the pr
esenc
e o
f
adher
ed
bars
w
a
s analy
z
e
d
accord
ing to t
h
e
convex
hu
ll. F
o
r the
ad
here
d
bars, th
e se
g
m
e
n
tatio
n
p
o
i
n
ts w
e
re searc
h
ed
usin
g sca
n
n
in
g
meth
od
b
y
a
series
of th
e r
u
les
to d
e
ter
m
ine
the
opti
m
al
seg
m
entati
o
n
lin
e. T
h
e
pr
o
pose
d
met
hod
can
seg
m
ent
the
adh
ered
bars
effectively w
i
th
match
ed c
onc
avity po
in
ts. T
he ex
peri
m
ent
al res
u
lts sho
w
that the met
h
o
d
can w
e
l
l
s
e
g
m
ent a
n
d
co
unt
bars
movin
g
at a
hi
gh s
p
e
e
d
on
pro
ductio
n
l
i
ne, w
i
th
the c
ounti
n
g
accur
a
cy
near to 10
0% a
nd the reco
gn
i
z
i
n
g ti
me i
n
mil
liseco
nd.
Ke
y
w
ords
: ad
hesi
on, Blob a
l
gorit
hm, co
nca
v
ity analys
is, concav
e po
ints, seg
m
e
n
tatio
n
, bars
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
With the p
o
p
u
lari
zation
of negative d
e
viation
rolli
ng t
e
ch
nolo
g
y, to ensure th
e e
c
on
omic
intere
sts
of rolling mill
s, it is ne
ce
ssa
r
y to cou
n
t and
sep
a
rate
the
steel b
a
rs p
r
eci
s
ely that the
bars of
high
-spe
ed p
r
o
d
u
c
tion lin
es ne
ed to b
e
reli
ably dete
c
ted
,
locate
d an
d
determined
the
mutual po
sitional rel
a
tion
ship. The
r
efo
r
e, Enterp
rises pay mo
re
and more attention to the
method that can qui
ckly an
d accurately ac
hi
eve auto
m
atic count
in
g of steel bars.
Ho
wever, in
actual p
r
od
uction, due to p
r
odu
ction p
r
o
c
e
ss a
nd co
mplicate
d
pro
ductio
n
environ
ment of
steel ba
rs,
ther
e
is
usu
a
lly existing
severe n
o
ise
in the
acqui
red i
m
age
fo
r a
cou
p
le of rea
s
on
s: the ligh
t
chan
ging fre
quently
, the
motion ima
g
e
blurring
cau
s
ed by the qui
ck
movement of
steel ba
rs, the dark o
r
blu
e
se
ctio
n
s
for oxidation. Especi
a
lly, the image di
storti
on
cau
s
e
d
by th
e mon
o
cular i
m
age
acqui
sition will g
ene
rate o
c
clu
s
io
n and
adh
esi
on ph
eno
me
no
n
on the
ed
ge
of ima
g
e
s
.
In additio
n
,
misalig
ned
b
a
rs
will
cau
s
e the
sa
me
probl
em. All
the
probl
em
s will
make it difficu
lt to track, co
unt
and sepa
rate the steel
bars on p
r
od
uction lin
e [1].
After image t
h
re
shol
d p
r
o
c
essing
and
bi
nary p
r
o
c
e
ssi
ng, the ad
he
sive targets are in the
same
co
nne
cted domai
n b
e
ca
use the target
s have
t
he same
col
o
r an
d texture feature
s
. So,
these targets will be treated as on
e target and e
r
ro
rs a
r
e ea
sy to occur in th
e sub
s
eq
uen
t
pro
c
e
ssi
ng.
More
over, th
ere
is almo
st
no g
r
ay
difference o
r
gray
gradi
ent diffe
ren
c
e
amo
n
g
the
adhe
sive ov
al target
s, which m
a
kes i
t
difficult to
identify different regi
on
s. Therefore, it
is
necessa
ry to sep
a
rate th
e
adhe
ring
bars into in
di
vidual go
als via
a se
rie
s
of proce
s
sing,
whi
c
h
is the adh
esi
on target
seg
m
entation [2].
Image
se
gme
n
tation of
ad
herin
g
steel
b
a
rs o
r
oth
e
r o
v
al obje
c
ts is the p
r
e
c
o
ndi
tion for
sub
s
e
que
nt image a
nalysis an
d featu
r
e extra
c
tion
. On the p
r
oblem of a
d
herin
g target
s
segm
entation
,
many sch
o
lars h
a
ve
prop
osed
different meth
o
d
s,
such a
s
mathemati
c
al
morp
holo
g
y operation
s
, improve
d
wat
e
rshed tran
sform, and a
c
tive conto
u
r tracki
ng. T
h
e
traditional
wa
tersh
ed
se
gm
entation m
e
th
od [3] is
pro
n
e
to cause ov
er-se
g
mentati
on result, an
d
the accuracy
of object extraction i
s
not high.
The re
se
arch
for the imag
e segm
entati
on of
adhe
re
d steel ba
rs
or othe
r oval
object
s
has o
b
taine
d
some a
c
h
i
evements at
pres
ent. The method
of image se
gmentation
an
d
recognitio
n
b
a
se
d on
the
oval assu
mpt
i
on ha
s
bee
n
pro
p
o
s
ed i
n
referen
c
e [4],
com
b
ine
d
wi
th
the mathemat
ical mo
rphol
o
g
y, it has solved t
he pro
b
le
m of the adhered
steel ba
rs, and reali
z
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 617
3 –
6180
6174
the countin
g
of ba
rs,
but
with lo
w a
c
cura
cy. Wu
se
parate
d
steel
ba
rs from
th
e ba
ckg
r
ou
n
d
by
mean
s of
acti
ve co
ntour m
odel, a
nd
he
realized
th
e segmentatio
n and co
untin
g
of
the steel b
a
rs
by usi
ng im
age
se
gmen
tation, edg
e
extractio
n
and m
a
them
atical m
o
rph
o
logy [5]. Z
hang
prop
osed th
e
method
which ba
sed
on
template
matching a
nd va
ri
able th
re
shol
d segme
n
tation
for dete
c
tion
and po
sition
ing of bars,
to some
ext
ent, which can overcom
e
the bar
se
ction
adhe
sio
n
, occlu
s
ion, oxida
t
ion and othe
r interfere
n
ce [6].
For th
e p
r
obl
em of the
a
dhered
ba
rs
in the im
age
, referen
c
e [
7
] took the f
o
llowin
g
action
s: preproce
s
sing the
colle
cted o
r
ig
inal ba
r
imag
e; then pro
c
e
ssi
ng ea
ch
b
a
r in the ima
g
e
into a singl
e pixel by corrosio
n
and thi
nning al
gorith
m
, which
can
realize the segmentatio
n of
adhe
red b
a
rs prelimina
r
il
y; in the end, using
th
e
algorithm of
shri
nkage a
nd dete
c
ting
the
endp
oint of the line, to re
alize the on
e to one co
rre
spond
en
ce bet
wee
n
singl
e pixel and rela
ted
steel ba
r on the origi
nal im
age [7]. However, th
is met
hod can not solve the occlu
s
ion p
r
obl
em.
A method for
image segm
e
n
tation ba
sed
on local bi
na
ry pattern an
d regio
n
com
petition
wa
s propo
se
d in the referen
c
e [8]. In this meth
o
d
, the regi
o
n
com
petitio
n algo
rithm
wa
s
improve
d
, a
new
regi
on
competition hy
pothe
sis
wa
s introdu
ce
d a
c
cordi
ng to th
e ch
ara
c
te
rist
ics
of the bar im
age segme
n
tation, and an
iterative
algorithm was p
r
opo
se
d, in whi
c
h the en
erg
y
can be co
nverge
d
to
the
local minimu
m.
But
the
calcul
ation of t
h
is m
e
thod i
s
so
com
p
licated
that this meth
od is not suitable for real-ti
m
e cou
n
ting.
Referen
c
e [9] uses the a
r
e
a
method to sep
a
rate a
n
d
count the ad
here
d
steel b
a
rs, its
theory is to
calcul
ate the a
v
erage
are
a
of the bar
s section at first, then to cal
c
u
l
ate the num
ber
of bars
acco
rding to the area of
each co
nne
cted dom
ain. Erro
r in this metho
d
wi
ll be large.
Referen
c
e [1
0] takes adv
antage of the impr
oved
watersh
ed al
gorithm to h
andle the
adhe
ring
ba
rs in th
e imag
e. Firstly, the
image i
s
p
r
e
t
reated
before usi
ng
wate
rsh
ed
algo
rithm;
and th
en, the
method
of g
r
adient
ope
rat
o
r a
nd m
a
the
m
atical
morp
hology m
u
st
be u
s
e
d
to av
oid
over-se
g
ment
ation. Finally, watershe
d algorithm
ba
sed on dista
n
c
e tran
sfo
r
m
a
tion is used
to
reali
z
e the se
gmentation of
adheri
ng ba
rs. Ho
wever, t
h
is metho
d
g
enerates
con
s
ide
r
abl
e error
for occlu
s
ion
bars imag
e.
Referen
c
e [1
1] pret
reate
d
the ba
rs secti
ons
in th
e im
age
acquired
in re
al time
b
y
use
of
graying,
sm
o
o
thing, a
nd
bi
nari
z
ation,
an
d p
r
op
osed
a
metho
d
of
co
unting
oval o
b
ject
s b
a
sed
on
regio
n
g
r
o
w
i
ng metho
d
a
nd line
a
r n
o
tation metho
d
.
This meth
o
d
ca
n a
c
hiev
e the pu
rp
ose of
cou
n
ting an
d sep
a
ratin
g
the adhe
ring b
a
r
s.
The ab
ove segmentatio
n
method fo
r
adhe
r
ed
bars imag
e can
not meet real-time
requi
rem
ents not only on the accu
ra
cy but also on
t
he pro
c
e
s
sin
g
spee
d. Over-se
g
mentati
on,
exce
ssive op
eration or po
or
a
c
curacy
d
oes not
suit for hig
h
-sp
e
e
d
and
real
-time processin
g
on
bars producti
on line. In th
i
s
paper,
we
will study on this issue
in order to resolv
e the problem of
adhe
ring
ba
rs
seg
m
entati
on. First,
we
can
obtai
n a
n
a
c
curate
separation l
o
cation b
e
twe
e
n
the
bars; an
d thi
s
will lay
solid
foundatio
n fo
r target
tra
c
ki
ng, counting
and
sep
a
ratin
g
ba
rs by servo
sy
st
em.
2. Image Preproces
sing
of Steel Bars
In this pa
per, images
are
colle
cted th
roug
h the B
a
sle
r
pla
ne
array cam
e
ra
, who
s
e
resolution i
s
1296
×9
66, frame rate is 3
0
fps, and coll
ection field is 700mm
×
200
mm. One ima
g
e
can b
e
acqui
red a
nd processed
within
30ms. Th
e a
c
qui
red im
ag
e of bars in the highli
ght b
l
ue
LED light so
u
r
ce i
s
shown in Figure 1.
Figure 1. Image of Bars
(th
e
tenth frame
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Im
age Segmentation of Adheri
ng Bars
Based o
n
Impro
v
ed
Con
c
avity Point
s
…
(Guoh
ua Liu
)
6175
There is u
s
u
a
lly severe n
o
ise in the i
m
age colle
cted from the
came
ra for a
coupl
e of
r
e
as
o
n
s
:
F
i
rs
t, s
h
ap
es
o
f
ba
r
s
s
e
c
t
io
ns
a
r
e d
i
ffe
r
e
nt; se
con
d
,
field
environ
ment
i
s
com
p
lex;
thi
r
d,
light ch
ang
es freque
ntly. Therefore, im
age p
r
ep
ro
ce
ssi
ng i
s
ne
ce
ssary in thi
s
pape
r, and th
e
main mea
n
s
are ima
ge thresh
old segm
entation an
d filtering.
First, processing
of imag
e thre
shol
d
segm
entation
must be
d
one to the
colle
cted
image
s to
eli
m
inate the
in
fluence of
ba
ckgro
und
[12
]. Then, a
s
su
me that th
e i
m
age
re
gion
is
)
,
(
y
x
f
, and sele
ct the app
rop
r
iat
e
T as the th
reshold to ob
tain binary i
m
age throug
h image
binari
z
atio
n. The bin
a
ry im
age regio
n
is
)
,
(
y
x
f
. As is
sho
w
n
in Figu
re 1, section
s
of ba
rs differ
from the
ba
ckgroun
d o
b
viously
due to
l
i
ght irradiat
io
n, thus T
will
be
set a
s
average
g
r
ay val
u
e
for the whol
e image.
Aiming at the
noise jammi
ng p
r
oble
m
s
in the ima
ge
after bin
a
ri
za
tion, median
filtering
can
re
move t
he n
o
ise [13]
. Binary me
di
an filter i
s
ap
plied fo
r in
th
is p
ape
r by
u
s
ing
4
×
4 m
a
sk.
Figure 2 is th
e image after
pretreatment.
Figure 2. Pretreatme
nt of the Image
3. Blob Anal
y
s
is
Blob analy
s
is can p
r
ovide
the informati
on
of the im
age ab
out Blob num
ber, l
o
catio
n
,
sha
pe, and o
r
ientation, etc;
it can also p
r
ovi
de the geo
metric top
o
lo
gy of the related Blob.
In this paper,
the bars sections ca
n be extr
acted q
u
i
ckly and a
ccurately by usi
ng Blob
algorith
m
, an
d the length, width, are
a
, and ce
nt
roid
coordi
nate of each Blob ca
n be cal
c
ul
ated.
After pretreat
ment, the ta
rget region
h
a
s
be
com
e
an in
dep
end
ent conn
ecte
d do
main. T
h
i
s
pape
r extra
c
t
s
the Blob
of bars sectio
ns of
each fram
e image th
ro
ugh ei
ght nei
ghbo
rho
od Bl
ob
analysi
s
, and
cal
c
ulate
s
the
centroi
d
of the target re
gio
n
.
3.1. Blob Re
cognition
The bin
a
ry i
m
age i
s
)
,
(
y
x
BI
(
BI
for short
)
afte
r pret
reatme
n
t, and it can
compl
e
tely be
divided into
several sub
-
region
s
)
,
(
y
x
BI
i
, these su
b-regio
n
s are form
e
d
by m Blo
b
and
a
backg
rou
nd [14].
The
sh
ape
chara
c
te
risti
c
of the Blo
b
region
is de
scribed
by
usi
n
g the
mom
e
n
t
[15], for
one Blob ima
g
e
)
,
(
y
x
I
, its
)
(
q
p
moment is
as
follows
:
)
,
(
)
,
(
,
)
,
(
y
x
I
y
x
q
p
q
p
y
x
y
x
I
M
(
1
)
As is sh
own in formula (1),
)
,
(
y
x
is interior p
o
int or bou
nd
ary point of Blob domai
n, for
the bin
a
ry im
age, the
valu
e of
)
,
(
y
x
I
is 1
(
In
si
de Blob
) o
r
0
(Out
side
Blo
b
). Mo
reove
r
,
mome
nt
seq
uen
ce
s
q
p
M
,
is only identified by
)
,
(
y
x
I
.
Particul
arly, 0 moment of the image is a
s
follows:
)
,
(
)
,
(
0
,
0
)
,
(
y
x
I
y
x
y
x
I
M
(
2
)
Formul
a
(2
) i
s
the
a
r
ea
of
Blob regio
n
i
n
pixel
s
. Excl
ude i
n
terfe
r
e
n
ce
region
of
pseud
o
bars Blob u
s
i
ng feature hi
stogram meth
od acco
rdin
g to the Blob re
gion features,
and co
unt
0
,
0
M
of each Blob.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 617
3 –
6180
6176
3.2. Blob Ma
rk
In fact,
whe
n
Blob a
nalysi
s
is carried
out
on
the
bar im
age
s, the
r
e i
s
no
nee
d to
a
nalysi
s
the whole
im
age; in
this
pape
r, we
set the inte
re
st region
in
orde
r to
si
gn
ificantly re
du
ce
cal
c
ulatio
n. Centroid
(
x
,
y
) of the Blob re
gio
n
can b
e
obta
i
ned by formu
l
a (3).
)
,
(
)
,
(
0
,
0
1
,
0
0
,
0
0
,
1
M
M
M
M
y
x
(3)
Mark the cen
t
roid an
d se
ri
al numbe
r of
each Blob re
gion from left
to right. Figu
re 3 is
the analysi
s
result of Blob image. Tabl
e 2 sho
w
s para
m
eter value
s
of each Blob.
Figure 3. Analysis Result of Bars Blob
Table 1.
Blob Parameter Values
(pixels
)
No.
A
r
ea
Centr
o
id a
b
sci
ssa
Centr
o
id ordin
a
te
0
3216
80
64
1
1590
180
65
2
1519
242
73
4.
Concav
e Points Extr
a
c
tion and Se
gmenta
tion
of Adh
e
ring
Bars
Each i
nde
pe
ndent Blo
b
regio
n
of b
a
r
s i
m
age
is identified
a
fter image
thre
shol
d
segm
entation
,
denoi
sing
a
nd Blob a
nal
ysis. Howeve
r, som
e
of th
e target
s a
r
e
one
con
n
e
c
ted
regio
n
s
comp
ose
d
of several bars su
ch
as the Bl
ob1
in Figure 3. So it is nece
s
sary to segme
n
t
it into indepe
ndent an
d no
n-sti
ck
bars i
n
orde
r to en
sure accu
ra
cy of counting
results.
There a
r
e two ba
sic chara
c
teri
stics in
convexity adh
esio
n ta
rget: First, the
ra
dial widt
h
of conn
ecte
d
place i
s
sma
ller than the t
a
rget di
amet
er. Seco
nd, a
pair of con
c
ave points i
s
on
both
side
s
of targ
et diam
e
t
er. Based
on
this
ch
aracte
ristic,
we u
s
e
the
seg
m
ent
ation al
gorith
m
based on b
o
u
ndary con
c
av
e points
sea
r
chin
g and b
o
unda
ry tracki
ng [16].
Two
problem
s m
u
st
be
sol
v
ed to
seg
m
e
n
t the a
dhe
ri
ng regio
n
; th
e proble
m
s a
r
e
whe
r
e
to segm
ent the adh
erin
g
regio
n
and
h
o
w to segm
ent the adh
e
r
ing regio
n
. Referen
c
e [1
7]
prop
osed
an i
n
tuitive meth
od for re
gion
segm
entation
.
This m
e
thod mainly
inc
l
u
d
e
s
thr
e
e s
t
eps
:
first, finding t
he dee
pe
st concave point
s by co
ncave analysi
s
to
be ca
ndid
a
te se
gmentati
o
n
points; secon
d
, sele
cting the se
gmenta
t
ion route;
third, sele
cting
the best se
g
m
entation lin
e.
This
regio
n
segmentatio
n
method b
a
se
d on con
c
av
e
analysi
s
is
widely use
d
in
cell
s and
grai
ns
segm
entation
,
but the process of finding
segme
n
tatio
n
points i
s
co
mplicate
d
.
In this
pap
er,
we
mu
st d
e
termin
e
wheth
e
r the
ba
rs a
r
e a
dhe
sio
n
s or not
at first. If the
bars
are
a
d
h
e
sive
we
mu
st dete
c
t
and
sm
ooth
th
e
edge
an
d fin
d
the
con
c
av
e poi
nts; th
e
n
segm
ent the
steel
ba
rs i
m
age th
ro
ug
h segme
n
tation lin
e d
e
termined
by a
pair
of
con
c
ave
points. Now t
he adh
erin
g b
a
rs a
r
e
segm
ented pe
rfectl
y.
4.1. Adhesio
n
s Tes
t
For ea
ch Blo
b
regi
on getti
ng from Blob
analysi
s
, we
can find the
d
i
fferent si
ze b
e
twee
n
external co
nvex
hull and
e
x
ternal re
ctan
gle
whe
n
o
b
serving
adh
esi
ons an
d n
on-adhe
sio
n
s Blo
b
regio
n
. Thi
s
provide
s
a
b
a
si
s fo
r d
e
termining t
he
e
x
istence of
a
dheri
ng
ba
rs.
The
r
efo
r
e, t
h
is
pape
r introd
u
c
e
s
the asp
e
ct ratio an
d
area ratio a
s
morpholo
g
i
c
al pa
ramet
e
rs to esta
blish
discrimi
nant
model of adh
ering b
a
rs.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Im
age Segmentation of Adheri
ng Bars
Based o
n
Impro
v
ed
Con
c
avity Point
s
…
(Guoh
ua Liu
)
6177
The region
o
f
interest i
s
detecte
d and
expre
s
sed
with Blob
)
,
(
y
x
B
, and its extern
al
convex h
u
ll a
nd extern
al re
ctangl
e
)
,
(
y
x
R
is dra
w
n a
s
sho
w
n
in Figu
re 4 a
n
d
Figu
re 5. T
he wi
dth
and hei
ght of
)
,
(
y
x
R
are
)
)(
,
(
width
y
x
R
and
)
)(
,
(
height
y
x
R
.
Figure 4. Blob Convex Hul
l
Fi
gure 5. Blob External Re
ctangl
e
Whe
n
the Bl
ob target is t
oo lon
g
o
r
too wi
de, ba
rs a
r
e
con
s
id
ered
to be
a
dhe
sive.
Aspe
ct ratio can be calcula
t
ed by minimum exte
rnal rectan
gle, the formula i
s
as
follows.
)
)(
,
(
)
)(
,
(
height
y
x
R
width
y
x
R
K
WH
(4)
The ratio of B
l
ob regi
on a
r
ea
Bar
Area
to convex hull are
a
Convex
Area
is
the larges
t when the
bars a
r
e
no
n
-
adh
esive. S
o
, wh
en th
e
area
ratio ex
cee
d
s a th
re
shol
d, we
ca
n come
to th
e
con
c
lu
sio
n
that the bars are adhe
sive, a
nd the are
a
ratio formula i
s
as follo
ws:
Convex
Bar
A
Area
Area
K
(5)
For non
-a
dh
esio
ns
ba
rs,
the
A
K
is l
a
rg
e; but fo
r
adhe
sion
s
ba
rs, t
he
A
K
is obviously
small. The
r
ef
ore, an empi
rical thre
sh
old
can
be u
s
ed
to determine
wheth
e
r there are adh
esi
o
ns
bars o
r
not.
More
over, th
e adh
erin
g b
a
rs u
s
ually
a
ppea
r
in
th
e hori
z
ontal direction, so ba
rs
adhe
sio
n
s ap
pears wh
en
WH
K
is very l
a
rge.
In this pa
per, we
use the
threshold
1
WHT
K
and
2
WHT
K
to determi
ne
the existen
c
e of b
a
rs
adhe
sio
n
s.
I
n
co
ncl
u
sio
n
,
discrimin
a
n
t
model can
be
descri
bed a
s
follows:
adhesions
non
others
adhesions
K
K
K
and
K
K
WHT
WH
WHT
T
A
2
1
Category
4.2. Concav
e Points Extra
c
tion
Within the view field, the arrang
ement
of
bars along
the directio
n
of chain mo
vement
may be n
o
n
-
adhe
sio
n
s, t
w
o b
a
rs a
d
h
e
sio
n
s
and
multiple a
dhe
sion
s. Non
-
a
dhe
sion
s b
a
rs
can
be e
a
sily id
e
n
tified after
p
r
ep
ro
ce
ssi
ng;
two
bars
adh
esio
ns and
m
u
ltiple ad
he
si
ons can al
so
be
segm
ented
b
y
using m
a
th
ematical
morpholo
g
y meth
od if adh
esi
o
ns a
r
e n
o
t se
riou
s. Ho
wev
e
r,
partial o
c
clu
s
ion
cau
s
e
d
by field distortion an
d misalig
nment
of bars
will
lead to se
ri
ous
adhe
sio
n
s. It is not ea
sy
to segm
ent
this kin
d
of
adhe
sion
s
by mathemat
ical mo
rph
o
l
ogy
method. To
determi
ne ad
hesi
o
n
s
, we must fi
nd the
con
c
ave poi
nts by
smoot
hing processi
ng
and ma
ke the
segme
n
tatio
n
line.
In this p
ape
r,
we
assu
me t
hat the b
a
rs
se
ction
s
hav
e convex ed
g
e
s
after p
r
etreatment,
the co
ncave
points i
n
the
conta
c
t po
siti
on of tw
o
ba
rs can n
o
t be
cove
red
by other
bars, a
nd
there is n
o
ca
vity in
the sin
g
le bar, an
d its sh
ape i
s
clo
s
e to a ci
rcle
or an oval.
We put fo
rward the m
e
thod of con
c
ave poi
nt ex
traction
acco
rding to a
d
h
e
sio
n
s
situation.
l
d
an
d
s
d
are m
a
jor and
mino
r
axes
convex
hull of Blo
b
regi
on, resp
ectively. To
improve the
searchin
g sp
e
ed, sea
r
ching
onl
y happe
n
s
in the mino
r axis dire
ction
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 617
3 –
6180
6178
Standard 1: For the Blob
regio
n
of adh
ering b
a
rs wit
h
obvious ext
e
rnal
conto
u
r, its
A
K
and
WH
K
are
calculate first, then the
co
ncave point
s
can b
e
foun
d
if
A
K
and
WH
K
meet the
adhe
sio
n
test
ing co
ndition
s mentioned in
4.1.
We draw the
circl
e
with center at the centroid (
x
,
y
), and ra
diu
s
equal to
i
r
, which
meet the con
d
ition
s
i
s
d
r
d
75
.
0
25
.
0
. The arithmetic
co
ntinuation
it
self
is then
ca
rri
ed out by
dra
w
ing a
se
ries of ci
rcle
s
)
(
2
r
L
, …
)
(
i
r
L
, …
)
(
n
r
L
about this poi
nt, and toleran
c
e i
s
10 pixels.
Standard 3: As the r
adi
us increa
ses,
)
(
i
r
L
and Blob regi
on co
ntours
must interse
c
t at
the first point.
This p
o
int is one of the
possibl
e con
c
av
e p
o
ints
and
sho
u
ld
be sto
r
ed i
n
an array
)}
,
(
{
1
j
i
g
, and
i
、
j
are
hori
z
ontal a
n
d
vertical coo
r
dinat
e
s
. The
n
the next point will be found,
and the di
sta
n
ce b
e
twe
e
n
previou
s
poi
nt and this p
o
int is D; if
s
d
D
25
.
0
, radiation
conti
nue
s
and this p
o
in
t is in-pha
se
con
c
ave poi
nt and sh
oul
d be stored i
n
an array
)}
,
(
{
1
j
i
g
. With
the
contin
uing ra
diation,
if
s
d
D
25
.
0
, this p
o
int is he
teroge
neo
us
con
c
ave
poin
t
and
sho
u
ld
be
store
d
in het
erog
ene
ou
s con
c
ave poi
n
t
array
)}
,
(
{
2
j
i
g
, and
i
、
j
are ho
rizo
ntal and verti
c
al
coo
r
din
a
tes.
This meth
od is rep
eated u
n
til the
)
(
n
r
L
is ra
diated. Segm
entation line
may exist
betwe
en in-p
hase co
ncave points a
nd
hetero
gen
eo
us con
c
ave p
o
ints.
Standard 4: Cal
c
ulate the
distan
ce bet
wee
n
ea
ch p
o
int in
)}
,
(
{
1
j
i
g
and
)}
,
(
{
2
j
i
g
, and
the lines
can
not exist betwee
n
the in-ph
a
se
co
n
c
ave poi
nts
or the hete
r
o
gene
ou
s co
n
c
ave
points in th
e
array, cal
c
ul
a
t
e and
find th
e poi
nt satisf
ying
min
D
in form
ula (6),
and
then th
e two
points a
r
e the
desirable
co
ncave p
o
ints.
)
(
)}}
,
(
{
)},
,
(
{{
))
,
(
),
,
(
((
)
(
)}
(
)
1
(
),
0
(
min{
2
1
2
1
min
contour
s
j
i
g
j
i
g
j
i
g
j
i
g
D
i
D
i
D
D
D
D
(
6)
)
(
contour
s
is Blo
b
regi
o
n
contou
r in
formul
a (6
).
Figure 6
sho
w
s the
extracting process
of con
c
ave p
o
ints.
4.3. Dete
rmine the Segm
enta
tion Lin
e
Figure 6(a
)
. Con
c
ave p
o
in
ts extractio
n
Figure 6(b
)
. Con
c
ave p
o
in
ts extractio
n
Figure 6(c). Concave point
s ex
tra
c
tion
Figure 6(d
)
. Determinin
g the se
gmentat
ion line
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Im
age Segmentation of Adheri
ng Bars
Based o
n
Impro
v
ed
Con
c
avity Point
s
…
(Guoh
ua Liu
)
6179
Line
s will
not
exist bet
wee
n
in-pha
se
c
oncave poi
nt
and h
e
teroge
neou
s
con
c
a
v
e point
whe
n
we det
ermin
e
the se
gmentation li
nes. The
n
ma
ke a straight line betwe
en the two co
nca
v
e
points
whi
c
h
we g
e
t to me
et the followi
ng sta
nda
rd
s, and this
straight line i
s
the segme
n
ta
tion
line.
Standard 1: T
he middl
e pixel of
the strai
ght line conn
ecting t
he t
w
o con
c
ave
po
ints mu
st
be in
side
the
target, in
ot
her
wo
rd
s, al
l its pixel
s
b
e
long
to Blo
b
re
gion.
(Pi
x
els o
u
tsid
e
the
regio
n
edg
e a
r
e not in co
nsideratio
n).
Standard 2:
Subdom
ain
s
satisfy the
n
on-a
dhe
sio
n
s stan
da
rd
m
entione
d in
4.1 after
segm
entation
.
Thus, the a
dheri
ng ba
rs are segm
e
n
ted co
mplet
e
ly, and we
can d
e
term
ine the
segm
entation
position and geomet
ric
inf
o
rmatio
n.
Th
e pro
c
e
s
s of determi
ning t
he se
gme
n
tation
line
ab
is sh
owe
d
in Figure 6(d).
4.4. Set up Parameters
In the stan
d
a
rd
s of ad
he
sion
s te
st an
d co
ncave p
o
ints extra
c
ti
on, som
e
pa
ramete
rs
need to
be
set up, that i
s
T
K
,
1
WHT
K
and
2
WHT
K
. Gene
ral
l
y, these p
a
ra
meters a
r
e
o
b
tained
by a l
o
t
of training. The param
ete
r
s are finally
determined
throug
h a lot of experiments,
93
.
0
T
K
,
45
.
1
1
WHT
K
,
87
.
1
2
WHT
K
.
5. Segmenta
tion Re
sults
and An
aly
s
is of Adh
e
rin
g
Bars
For the
video
seq
uen
ce
s o
b
tained f
r
om
bars p
r
od
ucti
on line, th
e first
step i
s
to
extract
an imag
e of a
dheri
ng b
a
rs,
the se
co
nd
step is to
ex
tra
c
t Blob
con
c
a
v
e points, the
third
step i
s
to
segm
entation
the ad
he
ring
bars.
Comp
uter
can
com
p
lete all
process
within 3
0
m
s fo
r a f
r
a
m
e
image, so it can
satisfy th
e spe
ed requ
ireme
n
ts
of 2
5
~3
0f/s. MATLAB prog
ram
is used in th
is
pape
r.
Figure 7. Segmentation Re
sult of
Adheri
ng Bars (the
10th frame
)
Figure 8(a
)
. Image of adh
e
r
ing ba
rs (the
25th
frame)
Figure 8(b
)
. Segmentatio
n result of adh
ering
bars (the 2
5
th frame
)
Figure 9(a
)
. Image of adh
e
r
ing ba
rs (the
65th
frame)
Figure 9(b
)
. Segmentatio
n result of adh
ering
bars (the 6
5
th frame
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 617
3 –
6180
6180
The
b
a
rs seg
m
entation re
sult can be seen cl
ea
rly i
n
Figu
re
7, the ad
he
ring
bars a
r
e
segm
ented
completely. Figure 8
and F
i
gure 9
sho
w
the two othe
r adhe
rin
g
b
a
rs im
age
s a
nd
their segm
ent
ation re
sult
s. Tabl
e 2
sh
ows the
recognit
i
on time.
The
segm
entatio
n re
sults
sh
o
w
that the algori
t
hm we u
s
ed
can a
c
hi
eve desi
r
ed
seg
m
entation.
Table 2.
Re
cognition Tim
e
Frame nu
mber
10 25 65
Processin
g
ti
m
e
/ms
22 20 18
6. Conclusio
n
In view of the bar blo
c
k ad
hesi
on p
r
oble
m
s
on the
produ
ction line,
this pap
er p
r
opo
se
s
a method b
a
s
ed o
n
impro
v
ed Blob con
c
ave poi
nts
a
nalysi
s
to detect and m
a
tch the con
c
av
e
points. We
propo
se correspon
ding
segme
n
tation stand
ard
to achieve
the automatic
segm
entation
of adhe
red
bars. The ex
perim
ental re
sults
sho
w
th
at the algo
rithm ca
n find the
con
c
ave poi
n
t
s and dete
r
mine the se
g
m
entation lin
es
a
c
curately
without over-se
g
me
ntatio
n
.
The a
c
cura
cy, real-time a
nd ro
bu
stne
ss of the
alg
o
r
ithm lay a solid found
atio
n for autom
a
t
ic
tracking, cou
n
ting
and sep
a
rating steel
bars on hig
h
-spe
ed produ
ction line.
Referen
ces
[1]
Davo
d Ag
hlma
ndi, K
a
rim F
a
e
z
. Automatic S
egme
n
ta
tio
n
of
Glottal S
pace
from Vid
eo Im
ages
Base
d
on Math
emati
c
al Mor
pho
lo
g
y
and t
he h
o
ugh T
r
ansfor
m
.
Internatio
n
a
l Jo
urna
l of
Electrical
an
d
Co
mp
uter Engi
neer
ing (IJECE
)
. 2012; 2(2): 2
23-2
30.
[2]
Jun Sun, Ya
n W
ang, Xiao
ho
ng W
u
, Xi
aod
o
ng Z
han
g. A Ne
w
Ima
ge Se
g
m
entatio
n Alg
o
r
ithm and It’s
Appl
icatio
n
in
lettuce
o
b
je
ct segme
n
tati
on.
T
E
LKOM
NIKA Indo
nes
ian
Jour
na
l
of Electric
al
Engi
neer
in
g.
2012; 10(
3): 557
-563.
[3]
GH W
ang,
XH Z
hou, L
L
W
e
i. Image S
egme
n
tatio
n
B
a
sed
on W
a
te
rshed A
l
gor
ith
m
.
Computer
Simulation
. 20
09; 26(5): 2
55-
258.
[4]
H Che
n
, JZ
W
ang, GM Yao.
Pattern Rec
o
gniti
on Bas
ed
on Circ
u
l
a
r Se
gmentati
on for
Cou
n
ting
of
Boun
de
d Bar Steel.
Journ
a
l of
Engin
eeri
ng G
r
aph
ics
. 200
4; 25(4): 98-
10
3.
[5]
RQ W
u
, W
M
W
ang, JH Y
a
n
g
, HG F
u
. Aut
o
matic d
e
tectio
n of co
ntour t
o
ngu
e im
age
ba
sed o
n
active
contour mo
del.
Journa
l of Co
mp
uter App
lica
t
ions
. 200
8; 28
(S1): 217-2
19.
[6]
D Z
hang, Z
Xie
,
JS Ai. Resear
ch on re
bar se
ction p
o
sitio
n
in
g metho
d
in o
n
line r
e
b
a
r cou
n
t
ing s
y
stem.
Chin
ese Jo
urn
a
l of Scientific I
n
strument
. 20
1
0
; 31(5): 11
73-
117
8.
[7]
XY
Liu, B W
u
,
Z
Q
Su, J Guo, KL F
ang.
Met
hod for A
u
tomatic Co
unti
ng of Ste
e
l B
a
rs Base
d o
n
Mathematic
al Morph
o
lo
g
y
.
C
ontrol En
gi
neer
ing of Ch
in
a
. 2013; 20(
2): 286
-288.
[8]
YS Z
hang, YL
F
u
, Z
G
Jiang, W
Z
Li. Automatic Co
untin
g
S
y
stem of Ste
e
l Bars.
Jo
urn
a
l of Iron
an
d
Steel Res
earc
h
. 2006; 1
8
(9): 1-4.
[9]
SD Lu
o, X Y
a
n
g
, S Sha, W
J
Jia. Steel B
a
r S
p
littin
g
S
y
stem
Based
on V
i
su
al F
e
e
dback
C
ontrol.
Mini-
m
i
cr
o System
s
. 2005; 26(
1): 2000-
200
4.
[10]
Z
Q
Ma, SS Yi
n, CY
Li, YC
Z
heng. Impr
ov
ed V
ehic
l
e Oc
clusio
n S
egm
e
n
tation
Meth
o
d
Bas
ed
on
F
eature Poi
n
ts Combi
nati
on.
Journ
a
l of Co
mputatio
na
l Informati
on Syste
m
s
. 2013; 9(1): 3
73-3
80.
[11]
H W
ang, G F
a
n. Recog
n
itio
n
and T
r
acking Me
thod Bas
e
d
on Image of Real-T
ime an
d
High-S
pee
d
Bar.
Journa
l of Northe
astern U
n
iversity(N
atur
al Scie
nce).
20
11; 32(2): 2
81-
284.
[12]
MX N
i
x
o
n
, AS
Agua
do. F
eatu
r
e E
x
traction
&
Image
Pr
oces
sing for
Com
p
u
t
er Visio
n
, Aca
demic Pr
ess,
UK. 2012.
[13]
Cha
nho
Jun
g
,
Cha
n
g
i
ck Ki
m, Seou
ng W
an C
h
a
e
, Suk
j
oo
ng Oh.
Un
superv
i
sed
se
gmentati
on
of
overl
app
ed n
u
c
lei usi
ng b
a
y
e
s
ian cl
assificati
on.
IEEE Transactions o
n
Bi
omed
ical En
gi
neer
ing
. 20
10;
57(1
2
): 282
5-2
832.
[14]
OS Ahmad, J
Deb
a
y
le, N Gherra
s, B Presles, G Févotte, J
Pinoli.
Rec
o
g
n
i
z
i
n
g over
lap
p
ed particl
es
duri
ng a crystalli
z
a
t
i
o
n
proc
ess from in si
tu
video i
m
a
g
e
s for meas
ur
ing the
i
r si
z
e
distrib
u
tions
.
Procee
din
g
s of
the SPIE, F
r
a
n
ce. 201
1.
[15]
L Ji. Intellig
ent
splittin
g
in the
chromos
o
me d
o
mai
n
.
Pattern Recognition
. 1
989; 22(
5): 519
-532.
[16]
B Han, ST
W
ang.
An aut
omatic counting system
for ov
iposited eggs based on digital image
process
i
ng
. IC
EMI 2005: C
o
nferenc
e Proc
eed
ings
of T
h
e Seve
nth Internati
ona
l Co
nferenc
e o
n
Electron
ic Mea
s
ureme
n
t & Instruments. 200
5
;
6: 539-54
3.
[17]
S Kumar, SH
Ong, S R
ang
a
nath,
T
C
Ong,
FT
Che
w
. A r
u
l
e
-bas
ed
ap
pro
a
ch for
rob
u
st
clump s
p
litti
ng.
Pattern Reco
g
n
itio
n
. 200
6; 39(6): 108
8-1
0
9
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.