TELKOM
NIKA
, Vol.13, No
.2, June 20
15
, pp. 730 ~ 7
3
8
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i2.1468
730
Re
cei
v
ed
Jan
uary 26, 201
5
;
Revi
sed Ma
rch 2
9
, 2015;
Acce
pted April 15, 2015
Resear
ch on Beef Skeletal Mat
u
rity Determination
Based on Shape Description and Neural Network
Xiang
y
an Me
ng*, Yumiao Ren, Haixian Pan
Coll
eg
e of Elec
tronic Informati
on Eng
i
n
eeri
n
g
,
Xi’a
n T
e
chnol
ogic
a
l Un
iversit
y
, 71
003
2, Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:
w
u
lizu
@
1
26.
com
A
b
st
r
a
ct
Physio
l
og
ical
matur
i
ty is an
i
m
p
o
rtant i
ndic
a
tor for
be
ef q
uality. In traditi
ona
l metho
d
, the
maturit
y
grad
e is deter
mi
ne
d by subj
ectively
ev
alu
a
t
ing the de
gree
of cartilage os
si
ficatio
n
at the tips of the dorsa
l
spin
e of the thoracic vertebr
a
e
. T
h
is
paper
uses the co
mp
uter vision to
r
epl
ace the arti
ficial
metho
d
for
extracting
o
b
je
ct (cartila
ge
an
d b
one)
re
gio
n
s
. Hu
inva
r
i
ant
mo
me
nts of
o
b
ject r
egi
on
w
e
re c
a
lcu
l
ate
d
as
the reg
i
on
al s
hap
e char
acte
ristic para
m
ete
r
s. A
trained
Hopfi
e
ld
neur
a
l
netw
o
rk mod
e
l w
a
s use
d
for
re
co
gn
i
z
ing
carti
l
a
ge
a
n
d
bo
ne
a
r
ea
in
th
o
r
aci
c
vertebr
ae i
m
a
ge
bas
ed o
n
min
i
mu
m Eu
clide
an
dista
n
c
e
.
T
he result sh
o
w
ed that the a
ccuracy of net
w
o
rk re
cogniti
o
n
for cartila
ge
and
bon
e reg
i
o
n
w
a
s 92.75
% an
d
87.68
%, resp
e
c
tively. F
o
r aut
omatica
lly
mat
u
rity pre
d
ictio
n
, the accur
a
cy of pred
iction w
a
s 86
%. Algor
ith
m
prop
osed
in
thi
s
pap
er pr
ove
d
the i
m
age
d
e
s
c
riptio
n
a
nd ne
ural netw
o
rk mode
lin
g
w
a
s
a
n
effective meth
od
for extracting i
m
a
ge featur
e regi
ons.
Key
w
ords
:
B
eef Skeletal
Maturity, Im
age Pro
c
e
ssi
n
g
, Invaria
n
t Mom
ents, Neu
r
al Network
1. Introduc
tion
Standardization of me
at p
r
odu
cts can f
a
cilit
ate m
a
rketing an
d me
rch
andi
sin
g
. For the
majority of
b
eef cattle
sla
ughtered
in t
he
Un
ited
States,
ca
rcass value
ha
s th
ree
dete
r
mini
ng
factors: weig
ht, evaluation
of intram
uscular fat
a
nd physiol
ogical maturity,
and
estimate of
the
percentage yi
eld of sal
able meat product [1],[2
]. The European Beef Carcass Grading sy
stem
take
s into a
c
count ca
rca
ss
confo
r
matio
n
and fattn
e
ss [
3
]-[5]. The drawb
a
cks of visual in
sp
ectio
n
are i
n
con
s
ist
ence an
d variations, b
e
ca
use
gradin
g
results m
a
y
differ a
c
cordi
ng to
subj
ect
i
ve
experie
nce of
each expe
rt
meat
grade
r.
So the b
eef
pro
c
e
ssi
ng
e
n
terp
rises i
s
hard
to p
r
ovi
d
e
the con
s
u
m
ers with me
at produ
ct
s
of co
nsi
s
tent quali
t
y. For det
ect
i
on of quality degree of be
ef
carca
s
s, researche
s
i
n
th
is a
r
ea
be
ga
n in th
e e
a
rl
y 80s.
Many
peo
ple
have
studi
ed
on t
he
detectio
n
of b
eef marbling,
muscle
colo
r
and fat
color by
comp
uter vision
[6]-[8]. Some
previo
us
resea
r
ch sh
o
w
that only few peo
ple
had carried
out the re
se
arch on im
a
ge autom
ate
d
segm
entation
algorith
m
for cartila
ge a
n
d
bone
are
a
s
and the
auto
m
atic ide
n
tification of skel
e
t
al
maturity g
r
ad
e. Yield q
uali
t
y assessm
e
nt by imag
e
pro
c
e
ssi
ng
st
arted
with th
e work by
Ha
tem
[9],[10], he studie
d
the
segm
entation
me
thod b
a
s
ed
on HS
L
and CIEL
a
b
col
o
r
spa
c
e
transfo
rmatio
n, and discu
s
sed imag
e recogniti
o
n
method of skeletal matu
rity. Lean yie
l
d
estimation
sy
stem
su
ch a
s
CVS sy
stem
of USA, ha
s been
used i
n
actu
al p
r
od
uction [1
1]. L
i
u
Muhua,
et al.[
12] u
s
ed
Oht
a
color
syste
m
for
autom
a
t
ed segme
n
ta
tion of the
cartilage a
nd
bo
ne
area
s in the thora
c
i
c
verte
b
rae im
age
s.
2. Metho
d
s
Beef maturity is determi
ned by the degree of
ca
rtilage o
ssifi
cation in the thora
c
ic
vertebrae. Th
ere are five
maturity grad
es whi
c
h
reflect the degre
e
of cartilage
ossificatio
n
, from
A to E (se
e
from the Fi
gure
1).
Cartilage area
i
n
the thoraci
c
verteb
ra
e
sho
w
s si
gn
s of
ossificatio
n
. For ‘A’ and ‘B’ maturity, there is le
ss evid
ence of ossifi
cation, and th
en the ca
rtila
ge
become
s
p
r
o
g
re
ssively ossified with ag
e
until
it ap
p
e
a
rs
a
s
b
one
(
s
u
c
h
as
E-m
a
turity
). It’s v
e
ry
difficult
to evaluate
th
e de
gree
of ca
rtil
age ossifi
cation b
e
cau
s
e
of the
differe
nce
in
color,
size
and
sha
pe of
ca
rtilage
and
bone
area. B
u
t these
pa
ra
meters can
chara
c
te
rize th
e ca
rtilage
an
d
bone a
r
ea fro
m
arou
nd. Firstly, image segmentatio
n wa
s ca
rri
ed o
n
to isolate th
e ca
rtilage an
d
bone f
r
om v
e
rteb
ra im
ag
e. In this
pa
per,
colo
r a
n
d
shape
de
scripto
r
s
were
sel
e
cte
d
a
s
the
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Research on
Beef Skeletal
Maturity Dete
rm
ination Based on Sh
ape .... (Xiangyan Meng)
731
indicators for segme
n
tatio
n
of cartilag
e
and bon
e area, an
d Hopfield netwo
rk was u
s
ed
to
automatic in
d
e
x and re
cog
n
ition of targe
t
area.
2.1. Segmentation o
f
car
tilage and bo
ne area
To ide
n
tify the d
egree
of ossification
pre
c
isely, bo
th ca
rtilage
area
an
d b
o
ne a
r
ea
sho
u
ld b
e
co
nsid
ere
d
a
s
i
m
porta
nt obje
c
ts. Befo
re
ex
tracting
the fe
ature
s
of in
di
cators,
cartila
ge
and bo
ne a
r
ea sh
ould
be
sepa
rate
d from the thora
c
ic im
age, re
spe
c
tively. Sometime
s, it is
difficult to separate the
obje
c
t are
a
f
r
om
su
rro
un
ding a
r
ea
b
e
ca
use of si
milarity of color.
Espe
cially ca
rtilage h
a
s
si
milar
colo
r with fat, and bone h
a
s th
e simila
r col
o
r with m
u
scle.
Beside
s, the
ca
rtilage
co
lor fo
r
same
maturity ma
y be differen
t
among
ani
mals
rai
s
e
d
on
different di
et [9]. So a g
ood
se
gment
ation effe
ct can’t be
obtai
ned
with th
e
col
o
r th
re
sh
old
segm
entation
method. Accordin
g to this cha
r
a
c
teri
stic, colo
r an
d sha
pe of ca
rt
ilage area an
d
bone a
r
ea a
r
e sele
cted a
s
the dictator o
f
degree of o
s
sificatio
n
.
(1) Sele
ction
and calculatio
n of shap
e pa
ramete
rs
An effective sha
pe d
e
scri
ptor is
a key co
mp
one
nt of multimedia
conte
n
t de
scriptio
n,
sin
c
e
sha
pe i
s
a fu
ndam
e
n
tal prope
rty of an o
b
je
ct. The inva
riant
moment
of
cartila
ge
regi
on
wa
s sel
e
cte
d
to describ
e the sha
pe ch
ara
c
teri
st
ics.
The two
-
dim
ensi
onal mo
ments of a g
r
ay
function
f(x
,
y)
through o
r
d
e
r
(j+
k
)
is defin
ed as,
dxdy
y
x
f
y
x
m
k
j
jk
)
,
(
j,k=0,1,2,…
(1)
Whe
r
e,
j+
k
is the o
r
de
r of
moment. Be
cause
colo
r fe
ature
s
al
one
were n
o
t sufficient to
isolate
the
ca
rtilage
and
b
o
ne
regio
n
, ma
ny re
gion
s
wit
h
such
colo
r f
eature
s
re
sult
ed. In
ord
e
r to
isolate
the ta
rget re
gion
a
c
curately, sh
a
pe p
a
ra
meter we
re
pro
p
o
s
ed to d
e
scrib
e its
sha
pe. T
h
e
image
sh
ould
be t
r
an
sform
ed into
bin
a
ry image
fo
r p
r
ocessin
g
. Th
e central mo
ment of
regi
o
n
wa
s cal
c
ul
ate
d
as follo
wing
,
dxdy
y
x
f
y
y
x
x
k
j
jk
)
,
(
)
(
)
(
(2)
Normali
z
ing
central
mome
nt,
)
(
00
jk
jk
)
1
2
(
k
j
,Where,
μ
00
is z
e
ro
-o
rde
r
centr
a
l
moment. In term
s of ce
ntral moment
s, even mome
nts define
d
by Hu(196
2), wh
ich a
r
e invari
ant
to object scal
e, position, a
nd orie
ntation
,
are given as,
02
20
1
(3)
2
11
2
02
20
2
4
)
(
(4)
2
03
21
2
12
30
3
)
3
(
)
3
(
(5)
2
03
21
2
12
30
4
)
(
)
(
(6)
22
2
2
5
3
0
1
2
3
01
2
3
01
2
2
1
0
3
2
10
3
2
1
0
3
3
0
1
2
2
1
0
3
(
3
)(
)[(
)
3
(
)
]
(
3
)
(
)[3
(
)
(
)
]
(7)
)
)(
(
4
]
)
(
)
)[(
(
03
21
12
30
11
2
03
21
2
12
30
02
20
6
(8)
22
2
2
7
2
1
0
3
3
01
2
3
01
2
2
1
0
3
1
2
3
0
2
1
0
3
3
01
2
2
1
0
3
(3
)
(
)
[
(
)
3
(
)
]
(3
)(
)[3
(
)
(
)
]
(9)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 2, June 20
15 : 730 – 73
8
732
(2) Seg
m
enta
t
ion method
s of objective region
To red
u
ce the comp
utatio
nal com
p
lexity fo
r neural n
e
twork recog
n
ition and co
mputation
of invariant m
o
ment, noise
s and b
a
ckg
r
ound
sho
u
ld
be rem
o
ved i
n
the image.
1) Pre
-
segme
n
tation ba
sed
on colo
r
Acco
rdi
ng to the cha
r
a
c
teri
stics of imag
e (s
e
en from
Figure 2.(a)), sep
a
ratin
g
op
eration
will be
con
d
u
cted to
ca
rtilage a
r
ea
a
nd bo
ne a
r
e
a
. The first step in color
wa
s de
sign
e
d
to
sep
a
rate
the
obje
c
tive re
gi
on. Fo
r
ca
rtilage
area,
in
whi
c
h th
e g
r
a
y
scal
e
valu
e
of R(re
d) >20
0
,
and average
value and
sta
ndard devia
ti
on of R,G(gre
en), and B(bl
ue) fra
m
e we
re cal
c
ul
ated.
2) Hol
e
filling and filtering
In order to m
a
tch the tem
p
late, hole filling
operation
was
conducted on object region. T
o
redu
ce
the
calcul
ation of
sub
s
e
que
nt label p
r
o
c
e
ssi
ng an
d inva
ri
ant mome
nt
para
m
eters,
some
non-ca
rtilage
regi
on with small are
a
should be e
liminated from
image. In th
is pa
pe
r, re
g
i
on
prelimi
nary selectio
n wa
s operatio
n wa
s co
ndu
ct
ed by filtering and re
gio
n
labelin
g. Disk
filter(diam
e
ter=1 pixel)
wa
s sele
cted for
smoothi
ng filtering, be
ca
use it made little influence o
n
sha
pe of
re
gi
on. Region
al
scre
enin
g
o
peratio
n
wa
s to stati
s
tic t
he regio
n
wit
h
big
area, a
n
d
obje
c
tive regi
on (a
rea
>80
0
pixels)
we
re pre
s
e
r
ved.
(3)Se
g
me
ntation of bone a
r
ea
Acco
rdi
ng to
the ch
aracte
ri
stics of b
one
ar
ea, se
parating
op
eratio
n will
be co
ndu
cted
to
cartila
ge a
r
e
a
and b
one a
r
ea. The first
step in ba
ckgrou
nd remo
val was
de
signed to
sep
a
rate
the regio
n
, in whi
c
h the
grayscal
e value of
R(red
)
>200, an
d averag
e valu
e and sta
n
d
a
rd
deviation
of R,G(gre
en), and
B(blue
)
frame we
re calc
ul
ated. As
sho
w
n i
n
Fig
u
re 2.
(b), th
e
rude
segm
entation
of cartila
ge a
r
ea
wa
s cond
ucted.
Cartila
ge area a
nd
bone
are
a
we
re sepa
rate
d
by
colo
r ch
ar
act
e
rist
i
cs
from backg
rou
nd.
2.2. Recog
n
ition of car
tilage and bo
n
e
area
Becau
s
e th
e
colo
r of
cart
ilage a
nd fat
tissu
e
a
r
oun
d wa
s hi
ghly
simila
r, the
effect of
threshold se
gmentation method
was bad.
But
cart
ilage re
gion
have differen
t
characte
risti
c
in
sha
pe. In the sam
e
way
,
the color o
f
bone and
meat wa
s hi
ghly simila
r, and the
sha
p
e
cha
r
a
c
teri
stic wa
s u
s
ed
to
sep
a
rate
bon
e re
gion
from
aro
und. Afte
r large n
u
mb
er of
cal
c
ulati
on
of cartilag
e
a
nd bon
e regi
o
n
invariant m
o
ment pa
ram
e
ters.
(1)
Hopfiel
d
n
e
twork mo
del
establi
s
hme
n
t
The Hopfield
neural net
work i
s
a fo
rm
of recurrent
artificial n
e
u
r
al net
wo
rk b
y
John
Hopfiel
d
. It h
a
s
different
chara
c
te
risti
c
s from
othe
r
netwo
rk in
n
e
twork struct
ure
and
lea
r
ning
method.
Hopf
ield n
e
two
r
k
can
sim
u
late
the biol
ogical
neu
ral
mem
o
ry me
ch
ani
sm. So it’s ve
ry
suitabl
e for retrieval and
reco
gnition. L
earni
ng rule
of Hopfield n
e
twork i
s
to find the co
nne
ction
weig
ht am
on
g different n
e
u
ron
s
.
He
bbi
an le
arning
rule i
s
sele
cte
d
a
s
th
e le
arning
metho
d
of
netwo
rk. T
h
e
rule
of wei
g
ht adju
s
tment
pro
c
e
dure
is that, firstly, the conn
ectio
n
s b
e
twe
en t
he
units are wei
ghted; wij is t
he wei
ght of
the co
nne
ctio
n from unit j to unit i.
j
i
j
i
m
m
w
p
n
n
j
n
i
ij
,
,
0
1
)
(
)
(
(10)
whe
r
e,
p
i
s
th
e numb
e
r of
pattern;
m
i
(n)
is the value
of neuron i in p
a
ttern nu
mbe
r
n an
d the sum
run
s
over all
pattern
s fro
m
n
=1 t
o
n
=
p
, 1
≤
i,j
≤
p.
Hopfield net
s serve a
s
con
t
ent-add
re
ssa
b
le
memory
s
y
s
t
ems
with binary thres
h
old units
. But
in
this pa
pe
r, cha
r
a
c
teri
stic paramete
r
s
are
difficult to
co
nvert to bi
na
ry vector,
so t
he in
put ve
ctor
wa
s inva
ri
ant mom
ent
para
m
eters v
e
ctor.
40 verteb
ra i
m
age
s we
re
use
d
for featu
r
es extractio
n
and neu
ral n
e
twork trai
nin
g
.
(2)Stability test
Segmente
d
cartilage
regio
n
is different
from templat
e
in regio
n
si
ze, edg
e sh
a
pe and
locatio
n
. In
orde
r to ve
ri
fy the templ
a
te can
repl
ace
the a
c
t
ual
cartila
ge
regi
on,
sev
e
ral
operation
s
fo
r template
we
re carried
out,
su
ch
a
s
r
edu
ction, ve
rtical
flip, ho
rizo
ntal flip a
nd
ed
ge
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Research on
Beef Skeletal
Maturity Dete
rm
ination Based on Sh
ape .... (Xiangyan Meng)
733
compli
catio
n
. Di
stan
ce b
e
t
ween
ori
g
in
al template
i
n
variant
mo
ment pa
ram
e
ters vecto
r
and
transfo
rme
d
template
i
n
variant
mom
e
nt
pa
ramete
rs vector wa
s cal
c
ulate
d
, and result wa
s
sho
w
n i
n
ta
ble 1. See
n
from the
ta
ble, simil
a
rity mea
s
ure b
e
t
ween
ori
g
in
al template
and
transfo
rme
d
template was high. The re
sult sh
owed that the sha
p
e
para
m
eters were
sele
ct
ed
rea
s
on
able.
Table 1. Similarity measure
betwee
n
orig
inal template
and tran
sfo
r
med template
Transform
template1
template 2
template 3
template 4
template 5
template 6
vertical
flip
0
0.009
0.0558
0.0174
0.0607
0.0026
horizontal
flip
0
0.009
0.0558
0.0174
0.0607
0.0026
Reduction
0.0175
1.00E-04
7.68E-04
5.20E-04
6.00E-04
0.0017
Edge
complicatio
n
0.0099
0.0088
0.029
0.0096
0.0129
0.0285
(3)
Regi
on re
cog
n
ition an
d
segme
n
tatio
n
Accu
rate
seg
m
entation of
the cartilag
e
area
i
s
premise of correct cl
assification of
degree of
ca
rtilage o
s
sification. Accord
ing to pi
ct
ure
s
of be
ef dorsal verte
b
ra, 6 template
s
of
cartila
ge
and
6 templ
a
tes of bon
e
wa
s
sele
cted.
And then
sh
ape
ch
ara
c
te
ristic (i
nvaria
nt
moment parameters)
of template we
re
calc
ulated.
The
Eucli
d
e
an di
stan
ce
betwe
en
sh
a
p
e
cha
r
a
c
t
e
ri
st
ic
s of
sam
p
le a
nd t
e
mplat
e
s wa
s cal
c
ul
at
e
d
as f
o
llo
wing
:
]
))
(
min[(
2
6
,...
3
,
2
,
1
n
i
n
i
M
X
norm
d
(11)
Whe
r
e,
d
i
prese
n
ts the shorte
st dista
n
ce b
e
twe
e
n
the
i
-
th
sa
mple ch
aract
e
risti
c
vecto
r
and
template c
h
arac
teris
t
ic
vec
t
or;
X
i
i
s
th
e ch
aracte
ristic vecto
r
of
measuri
ng
re
gion,
X
i
=[
x
1
, x
2
,
x
3
,…,
x
7
];
M
n
i
s
the c
h
arac
t
e
ris
t
ic
vec
t
or of template,
M
n
=[
m
1
, m
2
, m
3
,…,
m
7
].
Regi
on mat
c
hing an
d re
cog
n
ition wa
s acco
mpli
sh
ed by neu
ra
l netwo
rk. M
a
tchin
g
method i
s
mi
nimum
dista
n
c
e m
e
thod. B
e
fore
re
co
gni
tion, a di
stan
ce th
re
shol
d
wa
s
sele
cted
to
judge if the
re
gion i
s
the
ca
rtilage o
r
bo
n
e
regi
on.
If distan
ce b
e
twe
en sample
ch
ara
c
teri
stic
a
nd
template ch
a
r
acte
ri
stic is
less than thresh
old,
the region is ta
rg
et region, an
d vice versa.
Labeli
ng num
ber of targ
et region
wa
s sa
ved and
u
s
ed
to subsequ
e
n
t feature extractio
n
.
0
5
10
15
20
25
30
35
40
0
0.
5
1
1.
5
s
a
mp
l
e
n
u
mb
e
r
di
s
t
a
n
c
e
bet
w
e
e
n
c
a
rt
i
l
age
r
e
gi
on
a
n
d
t
e
m
p
l
a
t
e
c
a
r
t
ila
g
e
non
-
c
a
r
t
i
l
a
ge
di
s
t
an
c
e
t
h
r
e
s
hol
d
0
5
10
15
20
25
30
35
40
45
50
0
0.
5
1
1.
5
2
2.
5
s
a
mp
l
e
n
u
mb
e
r
di
s
t
a
n
c
e
be
t
w
ee
n bo
ne
r
e
gi
on
an
d t
e
m
p
l
a
t
e
bo
ne
no
n-
b
o
n
e
di
s
t
a
n
c
e
t
h
r
e
s
hol
d
(a)
(b)
Figure 1. Similarity measure bet
we
en se
gmented
regi
on and templ
a
te
The sh
orte
st distan
ce
bet
ween regi
on
pa
ramete
r ve
cto
r
a
nd tem
p
lat
e
pa
ram
e
ters vecto
r
wa
s
sho
w
n
in
Figu
re
1. Se
en from the
fi
gure,
we
can
indi
cate
th
at the
di
stan
ce betwe
en sev
en
invariant
mo
ment p
a
ram
e
ters of
cartila
ge a
nd te
mp
l
a
te is very
small, and
the
avera
g
e
dist
ance
is 0.0
291; o
n
the oth
e
r
hand, th
e di
stan
ce
between inva
ria
n
t mome
nt pa
rameters
of n
on-
cartila
ge regi
on and tem
p
l
a
te is big, an
d the aver
a
g
e
distan
ce i
s
0.600. According to di
stan
ce
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 2, June 20
15 : 730 – 73
8
734
cal
c
ulate
d
, the distan
ce th
reshold i
s
de
termin
e
d
as
0.1 for cartila
ge regi
on an
d 0.2 for bon
e
regio
n
(see
n
from Figure
1(a), an
d Fi
gure 1
(
b
)).
T
he ope
ration
s of distan
ce
calculation
are
accompli
sh
ed
by Hopfield
neural n
e
twork.
The o
peratio
n flow of region reco
gnition a
n
d
segm
entation
is given i
n
th
e followi
ng Fi
gure
2. Figu
re 3 sho
w
ed t
he Segm
enta
t
ion pro
c
e
s
si
ng
results, an
d
Figure 3.(a
-d
) re
pre
s
e
n
te
d the pr
e-tre
a
tment proce
ss
of
image,
Figure 3.(e,
f
)
indicated the recognitio
n
a
nd se
gmentat
ion re
sult
of cartila
ge re
gi
on. Regio
n
reco
gnition is
the
pro
c
e
ss of calcul
ating the
shorte
st dist
ance,
and se
gmentation p
r
ocess is
to l
abel the targ
et
regio
n
an
d save the labeli
ng num
ber.
Figure 4.
(a
)
and (b)
sho
w
s the final b
o
ne an
d ca
rtil
age
regio
n
se
gme
n
ted re
sult by algorithm in t
h
is pa
per.
3.
Segmenta
tio
n
result ev
aluation
In ord
e
r to
eli
m
inate the
errors from
cattle
breed
s,
all image
s were captu
r
ed
in Haoyue
Corp., Ch
an
gch
un Provin
ce,
China. T
he obj
ective
regi
on of
e
x
ample ima
g
e
wa
s m
anu
ally
segm
ented from the imag
e. Recogni
tio
n
accu
ra
cy and cla
s
sificati
on
erro
r[13] were sele
cte
d
a
s
evaluation p
a
r
amete
r
s fo
r evaluation
of
obje
c
t regio
n
segm
entation
.
%
100
f
f
f
R
S
RA
(12)
whe
r
e
R
f
i
s
n
u
mbe
r
of actu
al cartila
ge o
r
bone re
gion,
and
S
f
is the numbe
r of se
gmented
cartila
ge o
r
b
one re
gion. Region mi
scl
assifi
catio
n
erro
r is define
d
a
s
follows:
%
100
/
/
B
B
F
F
B
B
N
N
N
E
(13)
whe
r
e,
E
B
is region cl
assifi
cation e
r
ror;
N
B/F
is numbe
r of object re
g
i
on whi
c
h is
miscl
assified
into
non-obje
c
t re
gion;
N
F/B
is
numbe
r of no
n-obj
ect regio
n
whi
c
h is mi
scl
assified int
o
obje
c
t regi
o
n
;
N
B
is actu
al n
u
mbe
r
of obje
c
t regio
n
.
After recognit
i
on of neural netwo
rk, 50 v
e
rt
eb
ra imag
es (1
28 cartil
age re
gion
s
and 12
8
bone
regi
ons) we
re te
sted
for se
gment
ation eval
uati
on. Among th
em, 20 imag
es of A-m
a
turity,
10 image
s of B-maturity, 10 image
s of C-m
a
turity
an
d 10 image
s of D-matu
rity. The accuracy o
f
segm
entation
re
sult wa
s cal
c
ulate
d
by
formul
a 1
3
and
14. And
the result showed th
at the
accuracy
of
netwo
rk reco
gnition fo
r
cartilage
an
d
bon
e
regi
o
n
was 92.7
5
% and
87.6
8
%,
respe
c
tively.
The cla
s
sifica
tion error of a
l
gorithm
wa
s 14.49%.
4.
Skeletal maturit
y
prediction model
Six texture feature
s
were sele
cted, i
n
clu
d
ing
coa
r
se
ne
ss, in
contrast, de
gree of
dire
ction, line
a
rity, regula
r
ity and roug
hn
ess. These
si
x attributes are vector refle
c
ts the cartila
ge
texture features [15]. In g
eneral, high
-l
evel vis
ual fe
ature
s
of three kin
d
s
of attributes
can
be
use
d
to
capt
ure th
e textu
r
e, commo
nly use
d
in
th
e
conte
n
t ba
se
d imag
e retri
e
val, and
bet
ter
texture feature sele
ction.
(1) Coa
r
s
ene
ss
1) It rel
a
tes t
o
dista
n
ces
o
f
notable
spat
ial variation
s
of grey level
s
, that is, impli
c
itly, to the si
ze
of the primit
ive element
s forming th
e
text
ure. Th
e pro
p
o
s
ed
comp
utationa
l pro
c
ed
ure
accou
n
ts fo
r
differen
c
e
s
b
e
twee
n the
a
v
erage
si
gnal
s fo
r the
no
A
t
each pixel
(x,y), comput
e
six averag
es
for the wind
o
w
s of si
ze 2
k
×
2
k
,
k
=0,1,...,5, around the pixel.
2
)
At ea
ch
pixe
l, c
o
mpu
t
e a
b
s
o
lu
te d
i
ffe
r
e
nc
es
E
k
(x,
y
)
bet
ween
t
he p
a
irs of
n
onoverl
appi
n
g
averag
es in t
he hori
z
o
n
tal and vertical d
i
rectio
ns.
3) At each pi
xel, find the value of
k that
maximize
s the differe
nce
E
k
(x,y
)
in eit
her di
re
ction
and
set
t
he be
st
s
i
ze
S
best
(x,y
)
=2
k
.
4) Co
mpute t
he co
arsen
e
ss feature F
crs
by averagin
g
S
best
(x,y)
ove
r
the entire im
age.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Research on
Beef Skeletal
Maturity Dete
rm
ination Based on Sh
ape .... (Xiangyan Meng)
735
Figure 2. The
operatio
n flow of regio
n
re
cog
n
ition an
d
segme
n
tatio
n
(a)
(b)
(c
)
(d)
(e)
(f)
Figure 3. Segmentation p
r
o
c
e
ss of cartila
ge regi
on
(a) o
r
iginal d
r
awi
ng (b) thresh
old segm
entati
on (c) h
o
le filling (d
) filtering (e)re
c
ognition a
nd
segm
entation
(f) seg
m
ente
d
regi
o
n
(pseu
do-colo
r processing
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 2, June 20
15 : 730 – 73
8
736
(a) bon
e reg
i
on
(b) cartilag
e
regio
n
(c) target regio
n
(bone a
nd cart
ilage)
Figure 4. Re
sult of segmen
tation
Instead
of the avera
ge of
S
best
(
x,
y)
, a
n
improve
d
coarsen
e
ss fe
ature to d
eal
with
textures having
multiple coa
r
sene
ss prop
erti
e
s
i
s
a histo
g
ra
m
cha
r
a
c
teri
zi
ng the
whol
e
distrib
u
tion of
the best si
ze
s over the im
age.
(2) Contr
a
st
Contrast
mea
s
ures h
o
w grey levels
q
;
q
= 0, 1, ...,
q
ma
x
, vary in the image
g
an
d
to wh
at
extent their d
i
stributio
n is
bias
ed to bl
a
ck
or
white.
The second
-orde
r
a
nd n
o
r
mali
zed fo
urth
-
order
central
moment
s of the grey
level histogram
(empiri
c
al prob
ability distribution), that is, the
var
i
anc
e
,
σ
2
, and kurto
s
is,
α
4
, are use
d
to define the contra
st:
4
co
n
n
F
(14)
whe
r
e
4
4
4
;
ma
x
22
0
()
P
r
(
q
/
g
)
q
qm
;
ma
x
4
4
0
(
)
P
r(q
/
g
)
q
qm
and
m
is the mean grey level,
i.e.
the first order mom
ent
of the grey
le
vel probability distribution. The value
n
=0.
2
5
is
recomme
nde
d as the be
st for discrimi
na
ting the textures.
(3) Dire
ction degree
Deg
r
ee of di
rection
a
lity is measured u
s
i
ng t
he freq
u
ency di
stribut
ion of orient
e
d
local
edge
s ag
ain
s
t their dire
ctio
nal angl
es. T
he edg
e stre
ngth
e(x,y)
a
nd the directi
onal an
gle
a(x
,
y)
are
comp
ute
d
usin
g the Sobel ed
ge de
tector a
ppr
oximating the pi
xel-wi
se x- a
nd y-de
rivatives
of the image:
(x
,
y
)
0
.
5
(
(
x
,
y
)
(x
,
y
)
)
xy
e
1
(
x
,
y
)
t
a
n
(
(
x,
y
)
/
(
x,
y
)
)
yx
a
whe
r
e
∆
x
(x
,y)
and
∆
y
(x
,y
)
are the h
o
ri
zontal and ve
rtical g
r
ey le
vel differences bet
wee
n
the
neigh
bou
ring pixels,
re
spe
c
tively.
A histog
ram
H
dir
(a
)
of qua
ntized di
re
cti
on value
s
a i
s
con
s
tru
c
ted
by countin
g
numbe
rs
of the
edge
p
i
xels
with the
corre
s
p
ondin
g
di
re
ctional
angle
s
and
th
e ed
ge
stren
g
th g
r
eate
r
th
an
a predefin
ed
threshold.
Th
e hi
stogram i
s
relatively
u
n
iform fo
r im
age
s with
out
stron
g
o
r
ient
ation
and exhibits
pea
ks for hi
ghly dire
ction
a
l image
s. The deg
ree of
directio
nality relates to the
sha
r
pn
ess of the pea
ks:
peaks
pw
dir
p
peaks
dir
p
a
H
a
a
rn
F
1
2
)
(
)
(
1
(15)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Research on
Beef Skeletal
Maturity Dete
rm
ination Based on Sh
ape .... (Xiangyan Meng)
737
whe
r
e
n
p
is th
e num
be
r of
pea
ks,
a
p
is the p
o
sition
of
the pth
pea
k,
w
p
i
s
the
ra
n
ge of th
e an
gl
es
attributed to the p
th
peak (that is, the range bet
we
en valleys around the pe
ak),
r
de
note
s
a
norm
a
lizi
ng f
a
ctor rel
a
ted
to quanti
z
ing
levels of th
e
angle
s
a
, and
a
is th
e q
uan
tized di
re
ctio
nal
angle (cycli
ca
lly in modulo 180
o
).
Table 2 sho
w
s
the relati
onship
b
e
tween
the
cattle
age
an
d degree of
cartilage
ossificatio
n
. From the ta
ble
,
we can
see
that t
he deg
ree of cartilag
e
ossification
in the verteb
rae
is the
mo
st i
m
porta
nt indi
cator of
skele
t
al maturi
ty. For
”A” maturit
y
, the cartilage in the thorac
ic
vertebrae is free of ossification; and
for ”E”
m
a
turity, there
is com
p
lete
ly evidence
of
os
sif
i
cat
i
o
n
.
Predi
ction re
sults a
s
sho
w
n in Fig
u
re
5.
sho
w
the predi
ction a
c
curacy of ma
turity
grad
e.
Table 2. Rel
a
tionshi
p between thor
aci
c
vertebral ca
rtilage bo
ne an
d degree of skeletal matu
ri
ty
index
Skeletal maturit
y
A B
C
D
E
Cattle age
Less than 24
months
24-36mont
hs
36-48mont
hs
48-72mont
hs
More than 7
2
mo
nths
cartilage
ossifi
cation
Free of
ossifi
cation
Small part
of ossificat
i
on
Most ossifi
cation
Most ossifi
cation
Complete ossific
a
tion
Figure 5. Evaluation of the degr
ee of accura
cy of cartil
age
Artificial neu
ral netwo
rks
were train
ed
to pr
edic
t
the maturity grades
from the textur
e
feature
s
. Th
e
neu
ral
netwo
rk output
was the m
a
tu
rity grad
es. The
netwo
rks had
two
l
a
yers a
n
d
the log-sigm
o
i
d function
was u
s
ed i
n
e
a
ch laye
r. Th
e neu
ral net
work
wa
s trai
ned by u
s
ing
the
backward p
r
o
pagatio
n alg
o
r
ithm in
Matla
b
. The
aver
a
ge a
c
curacy
of pre
d
ictio
n
i
s
8
6
.0%. Hat
e
m
use
d
the hue
function of the HSL
colo
r system
to se
parate th
e ob
ject re
gion, a
nd the avera
ge
accuracy of
his meth
od i
s
75% [9]. The re
sult
sh
owe
d
that the pre
d
iction
accuracy of
our
method
prove
d
efficient. T
h
e predi
cted m
a
turity
grade
s we
re
com
pared with
the
grade
s fro
m
the
trained
g
r
ad
e
r
s. T
he
re
aso
nably g
ood
result
s from th
e two
very
different
set
s
of
sa
mple
s
sh
o
w
that the pro
c
edure devel
o
ped in thi
s
rese
arch i
s
fa
irly robu
st. T
he metho
d
we pro
p
o
s
ed f
o
r
cartila
ge
and
bone
segm
entation a
n
d
pre
d
ictio
n
model fo
r o
ssifi
cation
proved u
s
eful
for
determi
ning t
he maturity grade
s.
5. Conclu
sions
In this p
ape
r, we
have inv
e
stigate
d
the
seg
m
entatio
n metho
d
of
cartila
ge
and
bone
in
vertebra ima
ges. Ho
pfield
neural net
work a
nalys
i
s
wa
s perfo
rme
d
to identify the ca
rtilage
area
and bon
e area. The re
su
lt proved the
image de
sc
ription and ne
ural net
work
modelin
g is
an
effective met
hod for
extra
c
ting ima
ge f
eature
re
gion
s. The
se
gm
entation met
hod
woul
d be
the
premi
s
e of au
tomated beef
maturity gradi
ng. For
com
m
ercial pu
rpo
s
e, E-matu
rity image wa
s not
colle
cted in
enterp
r
i
s
e. So the useful
n
e
ss of
this o
b
ject re
gion
segm
entation
and re
cog
n
ition
method warrants furth
e
r rese
arche
s
. T
he re
sults
of
this wo
rk al
so contri
bute to an automat
ed
beef maturity gradi
ng sy
ste
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 2, June 20
15 : 730 – 73
8
738
Ackn
o
w
l
e
dg
ements
This
pap
er i
s
fund by the
spe
c
ial
scient
ific
re
se
arch
plan p
r
oje
c
t
of Shaanxi P
r
ovince
Educatio
n Office, NO.1
4JK
1341; Th
e au
thors
are
g
r
at
eful to Haoyu
e
Co
rp., Cha
ngchun, for t
he
assista
n
ce gi
ven with the
site for this
researc
h
.
Referen
ces
[1]
La
w
r
ence T
E
,
F
a
rro
w
RL, Z
o
l
ling
e
r BL, Spiv
e
y
KS.
T
e
chnic
a
l note: T
he Unit
ed States D
epartme
n
t o
f
Agricult
ure b
e
e
f
yie
l
d gr
ade
e
quati
on re
quir
e
s modificati
on
to reflect the c
u
rrent lo
ng
issi
mus muscle
area to hot car
c
ass
w
e
i
ght rel
a
tions
hip.
Jo
ur
nal of Ani
m
al S
c
ienc
e.
200
8; 86: 1434
–1
43
8.
[2]
Acheson R, Woerner DR, Belk KE, Engle TE, et
al. Effects of qualit
y
gra
de,
y
i
eld grade, gender, and
cattle t
y
pe o
n
the nutri
ent co
mpositi
on of r
e
tail cuts from
the be
ef loi
n
and r
oun
d.
Me
at Scienc
e
.
201
4; 96(1): 44
7.
[3]
Beriain MJ, Indura
i
n
G, Carr T
R
, et al. Contrasting a
ppr
ais
a
ls
of qu
alit
y a
nd val
ue of b
e
e
f carcasses
in Spa
i
n a
nd th
e Unite
d
State
s
.
Revue Méd.
Vét.
2013; 16
4
(
7): 337-3
42.
[4]
Craig
i
e C
R
, R
o
ss DW
, Maltin CA, et al. T
he relati
ons
hip
bet
w
e
en vi
de
o
image
an
al
ysi
s
(VIA), visual
classificati
on, a
nd sa
lea
b
l
e
me
at
y
i
el
d of sirl
oi
n an
d fill
et cuts of beef carc
as
ses differi
ng i
n
bree
d an
d
gen
der.
Liv
e
stock Scienc
e
. 2013; 15
8: 169-
178.
[5]
Craig
i
e CR, N
a
vaj
a
s EA, Purchas RW, Maltin
CA, Bünger L., Hoskin SO, Ross DW,
Morris ST,
Roe
he R. A revie
w
of the d
e
vel
opme
n
t an
d use
of vid
e
o
imag
e an
al
ysis (VIA) for b
eef carcass
eval
uatio
n as
an alt
e
rnativ
e
to the current
EUROP s
y
st
em and
other
subjectiv
e
s
ystems.
Meat
Scienc
e
. 201
2; 92: 307
–31
8.
[6]
Jie
w
e
n
Z
,
Muhua L, Hai
d
o
n
g
Z
.
Segmentation of lon
g
issi
mus dorsi
a
n
d
marblin
g in ri
be
ye im
agi
n
g
base
d
o
n
mat
h
ematica
l
mor
p
hol
og
y.
T
r
ans
a
c
tions of
the C
h
in
ese
S
o
ci
ety of Agric
u
ltura
l
Engi
neer
in
g
.
200
4;
20(1):1
4
3
-14
6
.
[7]
Muhu
a L,
Jie
w
en Z
,
H
a
id
on
g
Z
.
Segmentati
o
n of fat
an
d
lea
n
me
at
in
b
eef
imag
es b
a
se
d
on fuzz
y
C-
means cl
usteri
ng.
Non
g
ye G
ongc
he
ng Xu
e
bao/T
r
ans
actio
n
s of
the Chi
n
ese Soci
ety of Agricultur
al
Engi
neer
in
g
. 2004; 20(
2): 161
-163.
[8]
Can
nel
l R
C
, B
e
lk KE,
W
i
se J
W
, et al. On
lin
e ev
alu
a
tio
n
of
a comm
ercia
l
v
i
de
o im
ag
e
an
al
ysis
s
y
ste
m
(Comput
er Visi
on S
y
st
em) to pred
ict beef ca
rcass red meat
y
i
e
l
d an
d for augme
n
ting th
e assig
n
ment
of U SDA
y
i
eld
grad
es.
J Animal Science
. 20
02; 80(5): 1
195
-120
1.
[9]
Hatem I,
T
an J. Cartilag
e
an
d bon
e segm
e
n
tation i
n
vert
ebra im
ag
es.
Transactions of the ASAE
.
200
3; 46(5): 14
29- 14
34.
[10]
Hatem I, T
an J, Gerrard
D E.
Determin
a
tio
n
of
an
imal
ske
l
e
tal
m
a
turit
y
b
y
ima
ge proc
e
ssing
.
Meat
Scienc
e
. 200
3; 65: 999-1
0
0
4
.
[11]
Hatem I,
T
an J.
Cartilage
seg
m
e
n
tatio
n
in vertebr
a i
m
ages
. 2000 A
SAE Annual International
Meetin
g, T
e
ch
nical P
apers: E
ngi
neer
in
g Sol
u
ti
ons for a Ne
w
C
entur
y; 20
0
0
; 1: 1419-
143
0.
[12]
Muhu
a L,
Jie
w
en Z
,
Xi
ao
bo Z
.
Auto-segm
en
tation
of
carti
l
a
ge
an
d b
o
n
e
a
r
eas
in th
orac
i
c
vertebr
ae
imag
es .
T
r
ansactions of the
CSAE
. 2005; 2
1
(8): 110-
11
3.
[13] Guang
ho
ng
Z
.
Meat Scienc
e
. Chin
a Agric
u
ltu
r
al Press. Beij
i
ng. 199
9.
[14]
Z
hang
YJ, Ge
rgran
d
s JJ. S
egme
n
tatio
n
e
v
alu
a
tion
us
in
g u
l
timate m
e
asurem
ent
acc
u
rac
y
.
SPIE
.
199
2; 165
7: 44
9-46
0.
[15]
Badu
Ra
o M, Prabh
akar
a
Ra
o B, Govard
ha
n A. Co
ntent b
a
sed
im
ag
e ret
r
ieval
usi
ng
do
mina
nt colo
r
,
textur
e a
nd sh
ape.
Inter
natio
nal J
ourn
a
l
of
Engi
neer
in
g S
c
ienc
e a
nd T
e
chno
logy
.
201
1; 3(4): 2
887
-
289
6.
Evaluation Warning : The document was created with Spire.PDF for Python.