TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3894 ~ 39
0
1
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.4294
3894
Re
cei
v
ed Se
ptem
ber 1, 2013; Re
vi
sed
No
vem
ber 2
0
,
2013; Accep
t
ed De
cem
b
e
r
10, 2013
A Fast Beef Marbling Segmentation Algorithm Based
on Image Resampling
Bin Pang, Xiao Sun, Xin
Sun, Kunjie Chen*
Coll
eg
e of Engi
neer
ing, Na
nj
in
g Agricu
ltural
Univers
i
t
y
, N
a
n
jing
210
03
1, P. R. China
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: chenku
n
ji
en
a
u
@1
63.com
A
b
st
r
a
ct
W
i
th the
min
i
at
uri
z
a
t
i
on a
nd
p
o
rtabi
lity of o
n
li
ne d
e
tectio
n a
nd gr
adi
ng
eq
uip
m
ent, traditi
ona
l P
C
is be
in
g re
plac
ed w
i
th ARM
o
r
DSP e
m
bed
d
ed syste
m
s
in
beef
qua
lity gr
adi
ng
ind
u
stry. As the
low
ba
si
c
freque
ncy of embe
dde
d system, the traditio
nal b
eef
mar
b
li
ng se
g
m
e
n
tatio
n
method ca
n not me
e
t
requ
ire
m
e
n
ts o
f
real-ti
m
e
p
e
rformanc
e. T
he f
a
st se
g
m
entati
on alg
o
rith
m
of
be
ef marb
li
ng base
d
o
n
i
m
ag
e
resa
mpl
i
n
g
is
put forw
ard ai
mi
ng th
e dis
a
dvanta
ges
th
at the traditi
ona
l
meth
od is ti
me-cons
u
m
in
g a
n
d
does
not ap
ply
to embe
dde
d s
ystems. F
i
rst, the e
n
tropi
es of
the orig
in
al i
m
age
and r
e
sa
mplin
g i
m
a
ge w
e
re
calcul
ated acc
o
rdi
ng
to
the entropy princ
i
p
l
e
to
d
e
ter
m
in
e the i
m
age r
e
sa
mp
lin
g rate
base
d
on
ent
rop
y
constrai
nt acc
o
rdi
ng to th
e c
han
ges
of rel
a
tive in
for
m
atio
n entro
py of r
e
sa
mp
lin
g i
m
a
ge. T
h
e
n
fu
zzy c-
me
an
(F
CM) c
l
uster s
e
g
m
e
n
tation
w
a
s co
n
ducted
o
n
th
e
resa
mpl
i
n
g
i
m
age
to c
a
lcu
l
at
e the
b
eef i
m
a
g
e
seg
m
e
n
tatio
n
t
h
resh
old.
F
i
na
l
l
y, be
ef
mar
b
li
ng
are
a
is
se
g
m
e
n
ted
vi
a
mo
rpho
log
i
cal
a
n
d
l
ogic
o
perati
ons
o
n
a
se
rie
s
o
f
im
ag
e
s
. Th
e
e
x
p
e
r
im
en
ta
l
re
su
l
t
s sh
o
w
th
a
t
th
i
s
p
r
o
p
o
s
ed
a
l
go
ri
thm
to
o
k
0
.
5
7
s
o
n
a
v
e
r
age
in b
eef
marb
li
n
g
i
m
a
ge se
g
m
entatio
n u
nder
the constra
i
nts
that the loss
ra
te of relativ
e
i
n
formatio
n
entro
py
rang
ed betw
e
en 0.5-1.0
%
, w
h
ich
is
on
ly
6.43%
of th
at of the
trad
i
t
io
n
a
l
FC
M cl
uste
r se
gme
n
t
ati
o
n
alg
o
rith
m, ind
i
c
a
ting si
gnific
a
n
t
ly aug
me
nted
efficiency of se
gmentati
on.
Ke
y
w
ords
: be
ef, Image se
g
m
e
n
tatio
n
, mar
b
lin
g, calcu
l
ati
on efficie
n
cy
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
In available b
eef gradi
ng st
anda
rd
s wo
rl
dwid
e, the grade of beef
marbli
ng is d
e
termin
ed
based o
n
th
e ri
chn
e
ss of
intramu
s
cul
a
r fat in t
h
e
rib-eye
se
ction of b
eef carcass. As b
eef
marbli
ng g
r
a
de is a
u
tom
a
tically dete
r
mined u
s
in
g
comp
uter vi
sion a
nd im
age p
r
o
c
e
ssi
ng
techni
que
s,
marbli
ng ima
ge shoul
d be
first segme
n
ted from t
he ri
b-eye im
age
of beef
carca
ss
t
o
extract the q
uantized cha
r
acteri
stic val
u
e that c
an
ref
l
ect the ri
chn
e
ss deg
re
e o
f
marbling, a
nd
thereafte
r au
tomatic dete
r
minatio
n is cond
ucte
d on the beef
marblin
g g
r
ade
by mode
identificatio
n according to
the quanti
z
e
d
cha
r
a
c
te
ri
stic value. Th
erefo
r
e, the
segm
entation
of
marbli
ng fro
m
the rib-ey
e se
ction im
age of beef
carca
s
s se
rv
es a
s
the b
a
si
s of auto
m
atic
evaluation of beef marbli
n
g
grad
e, whil
e the ac
cura
cy and efficie
n
cy of marbli
ng seg
m
enta
t
ion
evidently influences the a
u
tomatic
evalu
a
tion of beef marbli
ng grad
e.
Nume
ro
us
m
e
thod
s for b
eef ma
rbling
image
seg
m
entation
h
a
ve bee
n p
r
eviously
repo
rted. F
o
r the first tim
e
, McDonal
d
and
Che
n
[1] seg
m
ente
d
the ima
g
e
of beef ri
b-eye
se
ction into f
a
t and mu
scl
e
are
a
s
by im
age p
r
o
c
e
ssi
ng, then
cal
c
ulated the tot
a
l are
a
of fat, and
obtaine
d the
relation
shi
p
b
e
twee
n fat area an
d t
he
sensory evalu
a
tion result
s
of beef qu
ality.
Shiranita et a
l
. [2] extracte
d a rectan
gul
ar bl
a
ck an
d white imag
e with 340
×21
2
pixels and 4-bit
grayscal
e fro
m
a
beef
rib
eye ima
ge, a
nd p
e
rf
o
r
med
re
gion
segm
entation
and
binary
treatm
ent
on its fat and muscle by neural netwo
rk, aiming
to
acqui
re a b
eef marblin
g image that only
contai
ned
white adipo
se
pixel and bl
ack mu
scl
e pixel. Chen
and Qin [3]
prop
osed a
beef
marbli
ng ima
ge segme
n
ta
tion metho
d
based o
n
grad
er'
s
vi
sion threshold
s
an
d auto
m
atic
thresholdi
ng.
Ja
ckm
an
et
al. [4] pro
p
o
s
ed a
me
tho
d
of autom
atic beef m
a
rblin
g segme
n
tation
according
to
the marbling
and
colo
r
ch
ara
c
teri
stics
of one
sid
e
o
f
beef, whi
c
h
wa
s a
dapte
d
to
different envi
r
onment
s of i
m
age a
c
q
u
isi
t
ion. D
ue to
complex an
d chang
eable
be
ef marblin
g, no
clea
r bou
nda
ry can be discerne
d
betwee
n
muscle
an
d
fat areas. Th
erefo
r
e, marb
ling can h
a
rdly
be preci
s
ely
segmented.
The results of Subbi
ah
et al. [5] show that fuzzy c-mean (FCM)
algorith
m
fun
c
tione
d
well i
n
the
segme
n
tation of
be
ef marbling
i
m
age
with
hi
gh
rob
u
stne
ss. O
n
this ba
si
s, Du
et al. [6] pro
posed a KF
CM algo
rithm
whi
c
h al
so
worked in
se
g
m
enting b
eef
rib-
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Fast Beef Marbli
ng Seg
m
entation Algorit
hm
Based
on Im
age Resam
p
ling (Bin
Pang)
3895
eye ima
ge i
n
to ba
ckg
r
oun
d
,
muscle
an
d
fat regi
on
s. Q
i
u et
al. [7] p
r
ese
n
ted
a fa
st modified
F
C
M
algorith
m
for beef marblin
g segm
entati
on, sug
g
e
s
ti
ng that FCM is highly effective. As to the
thresholdi
ng
seg
m
entati
on meth
od,
the s
hape
of histog
ram ap
paren
tly impacts the
segm
entation
effects. If the beef im
age hi
stog
ra
m is a
sin
g
le pe
ak
or pea
k-to
-vall
e
y
cha
r
a
c
teri
stics are un
cle
a
r,
the opti
m
al thre
shol
d can
not be
conve
r
ge
d, leading to
low
segm
entation
accu
ra
cy o
r
even
segme
n
tation fa
ilu
re. Ho
weve
r,
relatively go
o
d
segm
entati
on
effect ca
n be
obtained
by the FCM al
g
o
rithm,
re
ga
rdless of the
beef imag
e h
i
stogram. Th
us,
FCM
algo
rith
m is i
deal fo
r
beef ma
rblin
g
se
gment
atio
n. Cu
rrently, real-t
ime
and
online dete
c
tion
beef marblin
g is preferre
d. To meet the real
-time
requi
rem
ent
s of onlin
e detectio
n
, image
pro
c
e
ssi
ng
must be hi
ghly efficient
and time
-saving. Ho
we
ver, althoug
h FCM ima
g
e
s
e
g
m
e
n
t
a
t
io
n a
l
g
o
r
ithm b
a
s
e
d
o
n
p
i
xe
l
c
l
as
s
i
fic
a
tio
n
h
a
s
s
a
tis
f
a
c
to
r
y
s
e
g
m
en
ta
tio
n
e
ffe
c
t
and
high ro
bu
stn
e
ss, it canno
t meet the real-time
requ
ireme
n
ts du
e
to low efficiency an
d time-
con
s
umi
ng issue.
Re
sampli
ng i
s
a
proce
s
s o
f
transfo
rmin
g
a di
screte i
m
age
whi
c
h i
s
defined
at on
e set
of
coo
r
din
a
te lo
cation
s to
a
n
e
w
set
of
coo
r
dinate
p
o
ints. Image
re
sa
mpling
metho
d
can
be
utili
ze
d
to red
u
ce the
dimen
s
ion
a
li
ty of the origi
nal imag
e, re
serve
effectiv
e pixels,
rem
o
ve re
dund
a
n
t
pixels, and
decrea
s
e th
e amou
nt of image p
r
o
c
essing d
a
ta, thereby a
c
cele
rating im
ag
e
pro
c
e
ssi
ng. T
o
ensure the
informatio
n a
nd qualit
y of i
m
age
s, re
sa
mpling
can b
e
con
d
u
c
ted
by
entropy
co
nstraint [8-11] t
o
minimi
ze
the lo
ss
of
u
s
eful i
n
form
a
t
ion, to si
mul
t
aneou
sly lo
wer
image dime
n
s
ion
a
lity, and to decre
ase the data volu
me of image pro
c
e
ssi
ng, thus redu
cin
g
the
time requi
red
for image p
r
o
c
e
ssi
ng.
The appli
c
ati
on of embdd
ed microp
ro
cessor in b
eef
image acqui
sition and p
r
oce
s
sing
as well as
q
uality
gra
de determi
nation
enabl
es
related e
quipm
ent to be
mi
niaturi
z
ed
an
d
portabl
e, thu
s
re
alizi
ng
online d
e
tect
ion and
cla
ssifi
cation of
beef qualit
y. Neverthel
ess,
comp
ared wi
th PC, ARM and DSP microprocesso
rs a
r
e di
sa
dvantage
ou
s in the lack of
arithmeti
c
ca
pability, and
l
ongtim
e
co
nsumption i
n
i
m
age
processi
ng
an
d co
mputation of large
data volum
e
[12]. Therefore, to optimi
z
e the
exis
tin
g
be
ef imag
e
pro
c
e
s
sing
algorith
m
a
n
d
to
develop
a n
o
vel one
sui
t
able for the
ARM o
r
DSP micropro
c
e
s
sors lay
a technolo
g
i
c
al
foundatio
n for the future rese
arch on
miniaturi
z
e
d
beef quality gradi
ng sy
ste
m
to allow o
n
lin
e
detectio
n
and
gradi
ng of b
eef quality. This
study
targ
ets to analy
z
e the influen
ce of re
sampli
ng
rate on im
ag
e quality and
image
segm
entation effi
ci
ency, ba
sed
on whi
c
h
a fa
st se
gmentati
o
n
algorith
m
of
beef ma
rblin
g
image
s for
e
m
bdde
d sy
stem wa
s e
s
ta
blish
ed relyin
g on info
rmat
ion
entropy con
s
traints a
nd resampling.
2. Segmenta
tion Algorith
m
Based on
Entrop
y
Constrain
t
and Resampling
2.1. Image Preproc
essin
g
Beef imag
e
prep
ro
ce
ssin
g refers to
a
n
op
eratio
n o
f
removin
g
th
e ba
ckgroun
d of b
eef
image. T
he b
eef targ
et area after ba
ckgrou
nd
rem
o
val can
be
o
b
tained via
thre
shol
d, re
g
i
on
gro
w
th an
d
morp
holo
g
ica
l
operation
s
[13]. As this stu
d
y focuse
s o
n
de
veloping a f
a
st
segm
entation
algorithm of
beef marbling
,
the oper
atio
n of backgro
und re
moval i
s
not de
scrib
ed
herei
n.
2.2. Image Resampling
Uniformly-sp
ace
re
sam
p
li
ng can b
e
p
e
rform
ed in t
he digital im
age-fo
rmin
g
prin
ciple.
The sa
mplin
g
transfo
rmatio
n is de
scribe
d as follo
ws:
10
10
x
x
y
y
(1)
Whe
r
e
00
(,
)
x
y
is co
ordin
a
te of th
e origi
nal ima
ge pixel, and
11
(,
)
x
y
is calculate
d
pixel
c
o
or
d
i
na
te
.
01
is
image resampling rate. A
lower
indi
cate
s lo
we
r i
m
age
sam
p
le
size
after re
sam
p
l
i
ng, but more
loss
of imag
e informatio
n
and mo
re seriou
s ima
g
e
distortio
n
. The
beef gray
scal
e image
s at different re
sam
p
ling rate
s
are sho
w
n in Fi
gure 1.
Figure 1 exhi
bits that whe
n
0.5
, the image remain
s un
chang
ed; wh
e
n
0.1
, the
image suffers from si
gnifi
cant detail lo
ss; when
0.05
, th
e image is
severely disto
r
ted. To
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3894 – 39
01
3896
comp
are the cha
nge
s of image hi
stogram
at differen
t
resam
p
ling rates,
H
ist
and
H
is
t
wer
e
set
as the ori
g
in
al image hi
stogram a
nd the hist
og
ra
m of resam
p
ling image resp
ectively. The
followin
g
equ
ation ca
n be
use
d
first to n
o
rmali
z
e
H
ist
and
H
is
t
res
p
ec
tively.
mi
n
(
)
ma
x
(
)
m
i
n
(
)
Hi
s
t
Hi
s
t
Hi
s
t
H
ist
H
ist
(2)
Figure 1. Beef Grayscale I
m
age
s at Different Resam
p
ling Rates
Thereafter th
e histo
g
ra
ms of the four i
m
age
s in
Fi
g
u
re 1
are di
splayed in Fi
g
u
re 2. A
s
sho
w
n
in
Fi
gure
2,
with
ch
angi
ng
sampling
rate
, the ba
si
c
sha
pe
of re
sampli
ng i
m
age
histog
ram
s
re
mains inta
ct, but they
are subje
c
t to detail variations.
Whe
n
0.5
, the histogra
m
s
of the resam
p
ling imag
e
and the o
r
igi
nal image
al
most re
se
mbl
e
; when
0.1
, they begin to
differ obviou
s
ly; but when
0.0
5
, the differen
c
es b
e
twee
n them a
r
e extremely sig
n
ificant. The
res
u
lt
s
su
gge
st
t
hat
a
s
t
h
e
sam
p
ling
rat
e
de
cr
ea
se
s,
the differe
nce
s
b
e
twe
en th
e hi
stogram
s
of
two image
s a
r
e in
cre
a
sed, whi
c
h is m
a
in
ly attribut
ed to the loss of image info
rma
t
ion. Therefo
r
e,
to ensu
r
e th
e quality of
image, an a
ppro
p
ri
ate re
sampli
ng rat
e
is pre
r
eq
u
i
site for ima
g
e
resampli
ng to
control the lo
ss of ima
ge informatio
n.
Figure 2. Normalize
d
Hi
sto
g
ram
s
at Different Resam
p
ling Rates
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Fast Beef Marbli
ng Seg
m
entation Algorit
hm
Based
on Im
age Resam
p
ling (Bin
Pang)
3897
2.3. Image Informatio
n Entropy
Calculation
The am
ount
of image i
n
formatio
n is
gene
rally
ex
pre
s
sed
by informatio
n e
n
tropy, of
whi
c
h Shan
n
on entro
py is use
d
most
co
mmonly. Its basi
c
form is:
1
()
l
o
g
(
)
m
ii
i
HP
x
P
x
(3)
Whe
r
e
m
is n
u
m
ber
of cate
gorie
s,
x
is the
element, a
n
d
i
is the
i
category. Fo
r a
n
image with th
e size of
M
N
, its informatio
n en
tropy is defin
ed as:
1
11
00
lo
g
,
1
,
1,
,
,
0,
0,
1
,
,
T
kk
k
MN
ki
j
ij
ij
HP
P
P
k
MN
I
ij
k
k
el
se
kT
…
(
4
)
Whe
r
e
255
T
is the grayscale lev
e
l, and
k
P
meets the followi
ng conditions:
0
1
T
k
k
P
(5)
The inform
ation entro
py of the original i
m
age is
set as
1
H
, and the information ent
ropy
of image with
the samplin
g
rate of
is set
as
H
. Its relative loss of info
rmation i
s
def
ined a
s
:
12
1
1
H
H
H
(
6
)
2.4. Dete
rmination of
Re
sampling Ra
te
The resampli
ng rate that
met the info
rmation lo
ss i
n
terval of
mi
n
m
a
x
,
w
a
s
sea
r
che
d
within the ra
n
ge of the re
sa
mpling rate of
0,
1
.
Acco
rdi
ng to
the
relation
ship
bet
wee
n
imag
e
an
d
1
, it is
sup
posed th
at th
e
resampli
ng ra
te of the last step i
s
0
for th
e cu
rre
nt se
a
r
ch
of the ste
p
size an
d re
sampli
ng rate
to be
h
and
re
spe
c
tively, the followin
g
eq
uation
s
are d
e
rived:
1m
i
n
,
00
,
hk
h
(
7
)
1m
a
x
,
00
1/
2
,
hk
h
(
8
)
W
h
er
e
k
is the
pa
ramete
r
chara
c
te
rizi
ng
the chan
ge
rate of
se
arch,
whi
c
h
is the
binary
s
e
ar
ch method in cas
e
of
0.
5
k
. After the approp
riate sampling
rate
meeting the
relative
informatio
n entropy
con
s
traint
mi
n
m
a
x
,
wa
s found,
sa
mpling
imag
e was utili
zed fo
r
segm
entation
thresh
old cal
c
ulatio
n that wa
s appli
ed there
a
fter.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
0
46
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3894 – 39
01
3898
2.5. Propose
d
Image Segmenta
tion Al
gorithm
The ste
p
s of
threshold
segmentatio
n
algorith
m
based on the
criteria of inf
o
rmation
entropy a
r
e a
s
follows:
Step 1: Give the initial step si
ze in the sea
r
ch of
h
, the initial
k
value and the
initial
sampli
ng rat
e
0
1
.
A
ssign
0
h
, set th
e
sa
mpling
state
to calculate
the info
rmat
ion
entropy
1
H
of th
e origin
al bee
f grayscale i
m
age.
Step 2: Use
to sam
p
le th
e ori
g
inal
be
ef gray
scale i
m
age,
cal
c
ul
ate the info
rmation
entropy
H
of sampling i
m
ag
e, and
cal
c
ul
ate the relati
ve loss of inf
o
rmatio
n
1
according t
o
Equation (4), (5) a
nd (6
).
Step 3: If
1m
i
n
, mak
e
0
, calculat
e the re
sam
p
ling rate of th
e next step u
s
ing
Equation (7), and follo
w Step 2.
Step 4 If
1m
a
x
, mak
e
0
, cal
c
ulat
e the re
sa
mp
ling rate
of the next step
u
s
ing
Equation (8), and follo
w Step 2.
Step 5: When
1m
i
n
m
a
x
,
, end this se
arch, and return the resa
mpling rate
.
Step 6: Take
the sampli
n
g
image
with
the resampli
ng rate of
as the input d
a
taset,
and
use the
FCM
algo
rith
m to
cal
c
ulat
e the
beef
im
age
se
gment
ation th
re
shol
d who
s
e
clu
s
ter
numbe
r is 2.
Step 7: Use t
he segme
n
ta
tion thre
shol
d
obtai
ne
d in
Step 6 for th
resh
old
segm
entation
on the origi
n
al beef grayscale image, and
acquire the
beef fat and muscle regions.
2.6. Marbling Segmenta
tion
The
step
s o
f
beef ma
rbl
i
ng segm
ent
ation a
r
e a
s
follows [13]
: First, logi
cal XOR
algorith
m
wa
s con
d
u
c
ted on
the
targ
et
regi
on (F
ig
u
r
e 3
(
a
)
) an
d
the fat regio
n
(Fi
gure 3
(
b
)
)
derived from
Section 2.1 a
nd 2.4 re
sp
ectively, and th
e re
sults a
r
e
displ
a
yed in
Figure 3(c). A
fter
the omnidi
re
ctional co
rrosi
on of
Figure 3(c), sm
all areas
were re
moved on
ce
again, an
d then
the image
was exp
and
ed
omnidi
re
ctio
nally to obt
ai
n a complet
e
muscle
regi
on, as
sh
own in
Figure
3(d). As
the stru
ct
ure of
rib
-
ey
e se
ction i
m
age of b
eef
carca
s
s i
s
a
v
ailable, an
d
th
e
longi
ssim
u
s
dorsi was th
e larg
est m
u
scl
e conn
ec
t
e
d re
gion i
n
the imag
e, Figure 3(d)
wa
s
subj
ecte
d to
cavity filling, and the
large
s
t conne
cted
regio
n
wa
s re
serve
d
to
obt
a
in the m
a
sk
of
longi
ssim
u
s
dorsi regio
n
, as
sho
w
n i
n
Figu
re 3
(
e
)
. Figu
re 3
(
b
)
and
Figu
re
3(e
)
were
the
n
subj
ecte
d to logic "an
d
" op
eration
s
to acquire t
he b
eef
marblin
g regi
on, as sho
w
n
in Figure 3
(
f).
Figure 3. Segmentation of Marbli
ng fr
om
a Represent
a
tive Beef Image
2.7. Appar
a
tus and Data
Processin
g
A digital ca
m
e
ra, Di
mag
e
Z1, Minolta
Co. Ltd, wa
s u
s
ed i
n
imag
e
captu
r
e. Th
e
outpu
t
image
s we
re
store
d
in red
-
gree
n-blue fo
rmat. T
he co
mputer u
s
ed
in this study is a 2.6GHz PC
equip
ped wit
h
a 40 GB ha
rd drive a
nd 2
.
0G DDR2 of
RAM.
All image processi
ng al
gorithms
were i
m
plem
ented
with Matlab. SPSS 18 was used for
data analy
s
is.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Fast Beef Marbli
ng Seg
m
entation Algorit
hm
Based
on Im
age Resam
p
ling (Bin
Pang)
3899
3. Results a
nd Analy
s
is
3.1. Effec
t
s
of Re
sampling Ra
te on the Rela
tiv
e
Loss of In
for
m
ation
To study the
relation
shi
p
betwe
en the
relative lo
ss
of informatio
n and th
e re
sampli
n
g
rate, the
rate
of relative i
n
formatio
n lo
ss
of the b
e
e
f image
s (
160
0
1
200
pixels) at
different
resampli
ng rates
sho
w
n
in Figu
re 1
(
a) was
cal
c
ulated
within
the sam
p
lin
g rate
rang
e
of
0,
1
.
0
accordi
ng to Equation (6) (Figure 4).
Figure 4. Rel
a
tive Loss of Information at
Different Re
sampling
Rate
s
As sh
own in
Figure 4, wh
e
n
0.4<
≤
1.0,
1
is sl
owly elev
ated with d
e
crea
sing
, and
the lo
ss of im
age i
n
form
ation
wa
s g
r
ad
u
a
lly increa
se
d. Wh
en
=0.4
, the corre
s
p
ondin
g
rate o
f
relative i
n
formation l
o
ss i
s
0.8
8
%. Th
e
re
sam
p
ling
i
m
age
retai
n
s more th
an
9
9
% of the
o
r
i
g
inal
image i
n
form
ation, indi
cati
ng only mil
d
loss a
nd th
e uno
bviou
s
affected im
a
ge qu
ality at the
resampli
ng rate of 0.4.
The loss of image inform
ation
1
start to incre
a
se ra
pidly as
d
e
c
r
e
as
es
w
i
th
in
th
e
r
a
nge
o
f
0<
<
0
.4. When
=0.1, t
he relative lo
ss of info
rma
t
ion re
ache
s
5.87%, su
gg
esting
a
relat
i
vely large
lo
ss of ima
ge i
n
formatio
n. At this time, th
e hi
stogram
of
resampli
ng i
m
age u
nde
rg
oes
signifi
ca
nt deformatio
n
(Figu
r
e 2
c
), indicating t
hat the qualit
y of
resampli
ng image is
signi
ficantly affected by t
he resampling
rate.
Therefo
r
e, when re
sa
mpli
ng
image is u
s
e
d
for segm
en
tation, the resampli
ng rate
should n
o
t be lowe
r than 0.4. Otherwi
se,
there will be
seri
ou
s loss of image information,
whi
c
h may severely influence
the accuracy
of
image segm
e
n
tation.
3.2. Effec
t
s
of Re
sampling Ra
te on th
e Efficien
c
y
of Image Segmentation
To
study th
e efficie
n
cy
of imag
e
se
gment
ation
at differe
nt resam
p
ling
ra
tes, the
traditional
F
C
M
imag
e seg
m
entation alg
o
rithm wa
s
u
s
ed
for th
e m
a
rblin
g
segm
entation
of b
eef
image
s with different
values to record the
con
s
u
m
ing time
of com
puter in
se
gmentatio
n
pro
c
e
ssi
ng a
s
an indi
cat
o
r to evaluat
e the s
egm
e
n
tation effici
ency. The
chang
es of time
con
s
um
ption
of com
puter for
segm
entati
on on
beef
i
m
age
s
in Fig
u
re 1a
at different re
sam
p
ling
rates a
r
e
sho
w
n in Figu
re
5.
As pre
s
ent
e
d
in Figure
5, the FCM algorith
m
sp
end
s nea
rly 8 s on the marbli
ng
segm
entation
of the
ori
g
in
al imag
e, b
u
t with
re
du
cin
g
resamplin
g
rate, the
time
co
nsumption
of
comp
uter i
s
rapidly lowere
d. When
=0.
4
, the time consumption d
r
op
s to 1.242
s, only one-
sixth of that
whe
n
=1.0.
Re
sampli
ng
rate re
marka
b
l
y affects th
e
efficien
cy of
beef im
age
segm
entation
,
and
red
u
ci
ng the
re
sa
mpling
rate
can
si
gnifica
ntly augme
n
t the o
peration
a
l
efficien
cy of beef image se
gmentation.
In addition, the se
gmenta
t
ion threshol
ds gi
ven by
the FCM a
l
gorithm at
different
resampli
ng rates
we
re
compa
r
ed, a
n
d
wh
en
0.4
, the se
gme
n
tation threshold
s
of the
resampli
ng a
nd o
r
iginal
i
m
age
s a
r
e id
entical
at 10
3. The
re
sult
s infe
r that
whe
n
0.4
, the
segm
entation
effects re
ga
rding
the
ori
g
inal a
nd
re
sampling
imag
es
sh
ould
be
simila
r o
n
t
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3894 – 39
01
3900
basi
s
of the
traditional
F
C
M alg
o
rith
m seg
m
entat
ion. Wh
en
0.05
0.4
, the se
gment
ation
threshold
s
a
r
e sli
ghtly di
fferent, while
whe
n
0.05
, the segm
entatio
n thre
shol
ds differ
distin
ctly. The re
sult
s im
ply that the
e
ffect of
resam
p
ling
i
m
age
se
gm
entation m
a
y be
signifi
cantly d
i
fferent from that of original
one.
Figure 5. Image Segme
n
ta
tion Time at Different Resampling
Rate
s
3.3. Beef M
a
rbling Image Segmenta
tion in Case o
f
Entropy
Constrain
t
The ab
ove e
x
perime
n
ts
show th
at re
sampling
rate
exerts a
sig
n
i
ficant impa
ct
on the
beef ima
ge i
n
formatio
n lo
ss and
segm
entation e
ffici
ency. A lo
we
r resamplin
g
rate b
o
o
s
ts t
h
e
segm
entation
efficien
cy at the co
st of
aggravated l
o
ss of ima
g
e
in
formatio
n. Therefore, it is
imp
e
r
a
t
ive
to
fin
d
o
u
t
a
s
u
ita
b
l
e
re
s
a
mp
lin
g
r
a
te
th
a
t
n
o
t
on
ly c
o
n
t
r
o
ls
th
e lo
ss
o
f
image
informatio
n within an a
c
ce
ptable rang
e
to ensur
e
the
quality of im
age
segm
ent
ation, but al
so
redu
ce
d the
time co
nsu
m
ption of
compute
r
,
the
r
eby imp
r
ovi
ng the effici
ency of im
a
g
e
segm
entation
.
As is
sho
w
n
above, when
0.4
, the loss of
resampli
ng i
m
age info
rm
ation was
less
than 1
%
without signifi
cantly jeopa
rd
i
z
ed
image q
ualit
y, while the
com
puter f
o
r
segm
entation
operatio
n took only 1.24
2 s, indica
tin
g
a signifi
ca
ntly improve
d
segm
entati
on
efficien
cy. Therefo
r
e, if the loss of relati
ve inform
atio
n entro
py is constraine
d wi
thin 1% to sel
e
ct
the re
sampli
ng rate
between 0.4
-
1.
0, i
t
is possibl
e
to redu
ce th
e
comp
uter ti
me co
nsump
t
ion
and to maint
a
in the quali
t
y of image segm
entation
.
Thus, und
e
r
the co
nst
r
a
i
nts of relativ
e
informatio
n e
n
tropy lo
ss o
f
[
0
.005
,
0
.
0
1
]
, the 126
beef ima
ges
acq
u
ire
d
were se
gme
n
ted
usin
g the method
s describ
ed in se
ction
s
2.4 and
2.5
,
which were
comp
ared wit
h
the traditional
beef marbling
segme
n
tatio
n
method of
FCM
[7]. The results a
r
e su
mmari
zed in
Table 1.
Table 1. Co
m
pari
s
on of Ca
lculatio
n Efficiency bet
wee
n
Traditio
nal
FCM Algo
rith
m and
Propo
se
d Algorithm
Maximum
Minimum Mean
Standard
de
rivation
Time of FCM
algorithm (ms)
8974
7855
8532
345
Time of algorith
m
herein (ms)
722
473
570
77
Table 1
sho
w
s that thi
s
algorith
m
herein
sp
ent 0.5
7
s on ave
r
a
ge on be
ef marbli
n
g
image se
gm
entation con
s
train
ed by
the
lo
ss
ra
te
s of relative
informatio
n
entropy
ran
g
i
ng
betwe
en 0.5-1.0% (only 6.43% of
that
of the traditional FCM cl
uster seg
m
enta
t
ion algorithm
),
whi
c
h si
gnificantly raise
s
the efficien
cy of
segme
n
tation and e
n
sures the imag
e quality as wel
l
.
4. Conclusio
n
Re
sampli
ng rate exerts
a significa
nt infl
uence on the i
n
formatio
n lo
ss
and
seg
m
entation
efficien
cy of
beef ima
ge.
The im
age
segmentatio
n efficien
cy
wa
s signifi
cantly
boo
sted as the
resampli
ng
ra
te de
cre
a
sed,
and
the
relat
i
ve loss
of informatio
n o
n
ly su
btly increa
sed
when t
h
e
resampli
ng ra
te drop
ped from 1.0 to 0.4, but it t
hereafter rapidly in
cre
a
sed
with further
red
u
ct
io
n
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Fast Beef Marbli
ng Seg
m
entation Algorit
hm
Based
on Im
age Resam
p
ling (Bin
Pang)
3901
of re
sampli
n
g
rate. Bei
n
g co
nst
r
ain
e
d
by the lo
ss of relative
informatio
n
entro
py ran
g
ing
betwe
en 0.5
-
1%, the FM
C ima
ge
se
gmentatio
n
method ba
se
d
on entro
p
y
con
s
trai
nt
and
resampli
ng
prop
osed in
this study
sub
s
tant
ially
enhan
ce
d the efficien
cy
of beef image
segm
entation
and
mai
n
tai
ned th
e im
a
ge q
uality si
multaneo
usly
, whi
c
h l
a
y
a technol
ogi
cal
foundatio
n fo
r the future rese
arch o
n
miniaturi
z
e
d
beef qu
ality gradin
g
sy
ste
m
to ena
ble
online
detectio
n
and
gradin
g
of be
ef quality.
Referen
ces
[1]
McDon
a
ld T
P
, Chen YR. Se
parati
ng co
nne
cted muscl
e ti
ssues in im
ag
es of beef car
c
ass ribe
ye
s
.
Transactions of the ASAE.
1990; 33(6): 2
059
-206
5.
[2]
Shiranita K,
Hay
a
shi K, Otsubo
A, Miy
a
jima T
,
T
a
kiy
a
ma
R. Dete
rmi
nati
on of
meat qu
alit
y
b
y
im
ag
e
process
i
ng
an
d neur
al n
e
t
w
ork techni
qu
es
. IEEE In
ternation
a
l Co
nfere
n
c
e on Fuzz
y
S
y
stems. San
Antoni
o, T
X
, U
SA. 2000.
[3]
Che
n
K, Qin C. Segmentati
o
n
of
beef marbl
i
ng bas
ed o
n
vi
sion thres
hol
d.
Computers a
n
d
Electron
ics
in Agric
u
lture.
200
8; 62(2): 22
3-23
0.
[4]
Jackman P, Sun DW, Allen P. Au
tomatic
seg
m
entatio
n
of b
eef l
ong
issimu
s dorsi
musc
le
and
marb
li
n
g
b
y
an ad
apta
b
l
e
alg
o
rithm.
Meat Scienc
e.
2
009; 83(
2): 187
-194.
[5]
Subb
iah J, Ra
y N, Kranzl
e
r GA, Acton ST
.
Comp
uter
visio
n
segme
n
tatio
n
of the long
is
simus dorsi for
beef qu
al
it
y
gr
adi
ng.
T
r
ansac
tions of the AS
AE.
2004; 47(
4
)
: 1261-1
2
6
8
.
[6]
Du CJ, Su
n
DW
, Jackman
P, Allen
P. Devel
opm
ent
of a h
y
b
r
id
im
age
proc
essin
g
al
gorithm
fo
r
automatic
eval
uatio
n of intra
m
uscul
a
r fa
t content i
n
be
ef M-long
issimus
dorsi.
Meat Sc
ienc
e.
2
0
08;
80(4): 12
31-
12
37.
[7]
Qiu J, Shen M, Peng Z
,
Lian
g K, W
u
H, Sh
i J.
Beef marbling e
x
tracti
on
base
d
on mo
di
fied fuzz
y
C-
means
cluster
i
ng
alg
o
rithm.
Non
g
ye J
i
xie
X
ueb
ao/T
r
ans
ac
tions
of th
e
Ch
ines
e Soc
i
ety
of Agric
u
ltura
l
Machinery.
20
10; 41(8): 1
84-
188.
[8]
Parker JA, Ke
n
y
on
RV, T
r
oxel D. Com
pari
s
on
of i
n
terp
ol
ating m
e
tho
d
s
for imag
e resa
mplin
g.
IEEE
transactio
n
s on
Medica
l Ima
g
i
ng.
198
3; 2(1): 31-3
9
.
[9]
Pal NR. On minimum cr
oss-e
n
trop
y thresh
ol
din
g
.
Pattern Recognition.
199
6; 29(4): 57
5-5
80.
[10]
Brink AD, Pen
dock NE. Mini
mum
cross-ent
ropy
thr
e
shold
selection.
Pattern Rec
ogn
itio
n.
1996; 2
9
(1):
179-
188.
[11]
de Al
bu
quer
qu
e MP, Esqu
ef IA, Gesuald
i
M
e
llo
AR. Imag
e
thresho
l
di
ng
u
s
ing T
s
allis
ent
rop
y
.
Pattern
Reco
gniti
on L
e
tters.
2004; 25(
9): 1059-
10
65.
[12]
Bin P, Xia
o
S, Liu D, Chen
K. On-line Acq
u
isitio
n an
d Re
al-tim
e Segm
e
n
tation S
y
stem
of Beef Rib-
e
y
e Image.
T
r
a
n
sactio
ns of the Chi
nese S
o
ci
ety for Agricult
ural Mac
h
in
ery
. 2013; 44(
6): 189-1
93.
[13]
Che
n
K, Su
n
X, Qin Ch, T
ang
X. C
o
l
o
r gr
adi
n
g
of
beef fat
b
y
usin
g c
o
mpute
r
visio
n
a
n
d
su
pport v
e
ctor
machi
ne.
Co
mputers an
d Ele
c
tronics in Agr
i
culture.
20
10;
70(1): 27-
32.
Evaluation Warning : The document was created with Spire.PDF for Python.