TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 13, No. 3, March 2
015,
pp. 503 ~ 51
1
DOI: 10.115
9
1
/telkomni
ka.
v
13i3.709
6
503
Re
cei
v
ed
No
vem
ber 2
8
, 2014; Re
vi
sed
De
cem
ber 3
0
,
2014; Accep
t
ed Jan
uary 1
6
, 2015
Feature Extraction and Classification for Multiple
Species of Gyrodactylu
s
Ectoparasite
Roz
n
iz
a Ali*
1,2
, Amir Hussain
2
, Mustafa
Man
1
1
School of Infor
m
atic & Appli
e
d Mathematics,
Universiti Ma
l
a
y
s
ia T
e
rengg
anu, Mal
a
ysia
2
Institute of Computin
g Sc
ien
c
e and Math
e
m
atics, Univ
er
sit
y
of Stirlin
g, UK
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: rozniza@
umt.edu.m
y
/ ral
@
cs.stir.ac.uk
A
b
st
r
a
ct
Active Sh
ap
e
Mod
e
ls (AS
M) are
ap
pli
e
d to th
e
attachment
ho
oks
of sev
e
ral
s
peci
e
s of
Gyrodactylus, i
n
clu
d
in
g the
n
o
tifiab
le p
a
tho
gen G. sal
a
ris,
to assig
n
e
a
c
h
spec
ies to its
true spec
ies t
y
pe.
ASM is use
d
a
s
a feature
extraction to
ol to s
e
lect
i
n
for
m
atio
n from
ho
ok i
m
ages th
at can
be us
ed as
in
p
u
t
data
into
trai
n
ed c
l
assifi
ers.
Lin
ear (
i
.e. L
D
A an
d K-
NN)
and
n
on-l
i
ne
ar
(i.e. MLP
a
n
d
SVM)
mo
dels
ar
e
used
to cl
assif
y
Gyrodactyl
u
s
speci
e
s. Sp
ec
ies of Gy
ro
dac
tylus, ectop
a
ra
sitic
mon
o
g
ene
tic flukes
of fis
h
,
are difficult to
discrimin
ate and i
dentify a
ccordi
ng
to morph
o
lo
gy alo
n
e
and the
i
r speciati
on curre
n
t
ly
requ
ires taxo
n
o
mic ex
pertise
. T
he current
exercis
e
sets
out to confi
d
e
n
t
ly classify
sp
e
c
ies, w
h
ich i
n
this
exa
m
p
l
e inc
l
ud
es a species w
h
ich is
a notifi
a
ble p
a
thog
en o
f
At
lantic salmo
n
, to their true class w
i
th a hig
h
degr
ee
of acc
u
racy. T
he fi
ndi
ngs fro
m
t
he c
u
rrent
exerci
s
e
de
monstrates
that i
m
p
o
rt of
ASM data
into
a
MLP class
i
fier,
outperfor
m
s s
e
vera
l oth
e
r methods
of
clas
sificatio
n
(i.e.
LDA, K-NN
an
d SVM) that w
e
r
e
assesse
d, w
i
th an avera
ge cl
assificati
on acc
u
racy of 98.7
2
%
.
Ke
y
w
ords
:
ma
rgin
al ho
oks, feature extractio
n
, gyrodactyl
u
s
,
machi
ne l
ear
nin
g
Copy
right
©
2015 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
There are o
v
er 440 de
scrib
ed spe
c
i
e
s of
Gyro
d
a
ctylus, which are typical
l
y small
(
<
1mm)
, ec
top
a
r
as
itic
w
o
rms
pr
inc
i
pall
y
infecting fish [1]. Most speci
e
s
are im
perfe
ctly kno
w
n,
with many d
e
scriptio
ns li
mited to an incom
p
lete
m
o
rph
o
logi
cal
descri
p
tion of
their attach
ment
hoo
ks. While
molecul
a
r tech
niqu
es h
a
ve, in re
ce
nt years, ma
de a vast
contributio
n to the
discrimi
nation
of on
e
spe
c
ie
s fro
m
a
n
o
ther,
spe
c
ie
s d
e
finition
s often
contin
ue to
rely o
n
morp
holo
g
ica
l
ch
aracte
risti
c
s (i.e.
attachment
hoo
k
morp
holo
g
y
and i
n
p
a
rti
c
ular the
sh
ap
e of
the si
ckle
of t
he 1
6
small
perip
he
ral m
a
rgin
al h
o
o
k
s whi
c
h
a
r
e
re
gard
e
d
as th
e key taxono
mic
feature
)
[2]. While m
o
st
speci
e
s of Gy
roda
ctylus
a
r
e
non-patho
ge
nic, cau
s
ing li
ttle harm to their
host
s
, othe
r sp
eci
e
s like Gyroda
ctylus
sal
a
ri
s M
a
lmbe
rg, 1
9
57, which i
s
an
OIE (O
ffice
Internation
a
l des Epi
z
ootie
s) - li
sted p
a
thoge
n of
Atlantic sal
m
on
can majo
r effects on
wild a
n
d
cult
ur
ed
f
i
sh.
Gy
ro
da
ct
y
l
us
sala
ri
s h
a
s
de
cima
te
d t
he juve
nile
salmon
pop
ul
ation in
ove
r
40
Norwe
g
ian ri
vers [3] is a
n
uncontroll
e
d
incr
ea
se
s in the size of the para
s
ite
population
on
resi
dent
salmon
popul
ations have
ne
ce
ssitate
d ex
treme mea
s
u
r
es su
ch as
the use
of th
e
bioci
de roten
one to
strate
gically
kill-o
u
t stret
c
he
s
of
river
system
s, in orde
r to remove the e
n
t
ire
fish and
G. salari
s pop
ulat
ion within a
river [3].
Given the impa
ct that G. sala
ris h
a
s h
ad in
Norway a
nd
else
wh
ere i
n
Scan
dinavia
and
Ru
ssi
a
[4-6], many
Europ
ean
sta
t
es in
cludi
ng
the
UK now h
a
ve mandato
r
y
surveilla
nce
prog
ramm
es
screeni
ng wild salmo
n
id
popul
ations
(e.g.
bro
w
n trout, cha
rr, g
r
ayli
ng, Atlantic
salmo
n
etc)
for the p
r
e
s
ence of notif
iable p
a
thog
ens
inclu
d
ing
G. sala
ris.
Cu
rre
nt OIE metho
dologi
es
fo
r t
he di
scrimin
a
tion of G. sal
a
ris from
oth
e
r
spe
c
ie
s of Gyroda
ctylus
that
o
c
cur on salmo
n
id
s
requi
re
co
nfirmation
of
identity by
both
morp
holo
g
ica
l
and mole
cul
a
r app
ro
ache
s, whi
c
h can
be time co
nsuming. In the
case of a type I
error, whe
r
e G.
sala
ri
s
is miside
ntified as
a
nothe
r speci
e
s and g
oes und
etect
ed
resultin
g
i
n
the
death of fish, or a type II
error,
whe
r
e
a non
-path
o
g
enic
sp
eci
e
s
is mi
scla
s
sified a
s
G.
sala
rie
s
cau
s
in
g fish t
o
be
tre
a
ted
unne
ce
ssarily
, the e
n
viron
m
ental
and
e
c
on
omic impl
ication
s
ca
n
be
con
s
id
era
b
le
[7]. For thi
s
rea
s
on, a
nd be
ca
use
of the wid
e
l
y
varying pa
thogeni
city seen
betwe
en
clo
s
ely related
speci
e
s, a
c
curate patho
gen
identificatio
n
is of p
a
ra
mo
unt impo
rtan
ce.
The difficulty
of discrimin
a
ting patho
g
enic
spe
c
ie
s
from simil
a
r
con
gen
ers, is comp
oun
de
d by
the very limit
ed n
u
mbe
r
of
morphol
ogically discrete
cha
r
a
c
teri
stics in
the
s
e
sp
ecie
s. O
w
in
g
to
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 3, March 2
015 : 503 – 5
1
1
504
these difficulties, the task of morphol
o
g
ical id
ent
ification cu
rrentl
y
relies up
on
a very limited
numbe
r
of do
main exp
e
rt
s
cap
able
of a
n
a
lysing
an
d d
e
termini
ng
sp
ecie
s. Tim
e
t
a
ke
n to i
denti
f
y
spe
c
ie
s can
be dra
m
atical
ly reduced
if the initial iden
tification of a spe
c
ime
n
as
being G. sala
ris
or G. sala
ris-like a
c
cordin
g to it morphology ca
n b
e
achieve
d
b
o
th more qui
ckly an
d more
accurately. In
the
event of
a suspe
c
ted
outbre
a
k,
the
dem
and
for i
dentificatio
n
may exceed
the
available
su
p
p
ly of suitabl
e expe
rtise
and fa
cilit
ies.
The
r
e i
s
, th
erefo
r
e, a
re
al ne
ed fo
r t
h
e
developm
ent
of rapid, accurate, semi
-automatic
/ a
u
tomatic dia
g
nosti
c tools that are able
to
confid
ently identify G. salaris in
any give
n popul
ation
of spe
c
imen
s.
The aim
s
of the cu
rre
nt study
were therefore, to explore t
he pote
n
tial use of a
n
Active
Shape M
odel
(ASM) to ext
r
act featu
r
e i
n
formation
fro
m
the attach
ment hoo
ks o
f
each
sp
eci
e
s of
Gyroda
ctylus.
Given
the
subtle diffe
ren
c
e
s
in
the
ho
ok
sh
ape
of
each
spe
c
ie
s, it is
hop
ed
that
this a
pproa
ch
move
s to
wards th
e
rapid
automated
cl
assificatio
n
of
sp
eci
e
s with
improved
rat
e
s
of co
rrect
cla
ssifi
cation
over exi
s
ting m
e
thod
s an
d
negate
s
the
curre
n
t labo
ri
ous process
of
taking m
anu
al mea
s
urem
ents which a
r
e u
s
ed to a
ssi
st expe
rts in identifying spe
c
ie
s. In
this
work, it
not o
n
ly co
ntribute
s
to th
e im
ag
e p
r
oc
essin
g
area, al
so
to
the
a
g
ri
cultu
r
e are
a
,
where
the systemati
c
and a
c
cura
te system is
provide
d
in p
r
edi
cting the
ectop
a
ra
site
of Gyroda
ctylus
spe
c
ie
s.
2. Methodol
ogical Appro
ach
2.1. Specimen Prepara
t
io
n
Specim
en
s of
Gyro
da
ctylus
(G.
cole
me
n n
= 10;
G.
derjavin
o
ide
s
n
=
25; G.
salari
s n
=
34; G. truttae n = 9)
we
re remove
d from thei
r respective salm
onid ho
sts a
nd fixed in 8
0
%
ethanol. Sub
s
eq
uently sp
ecime
n
s
were prep
are
d
for scan
ning
electron mi
croscopy (SEM
) by
transfe
rring i
ndividual, di
stilled wate
r ri
nse
d
, spe
c
i
m
ens
onto 1
3
mm diame
t
er rou
nd gl
a
ss
coverslip
s,
where
they h
a
d
their po
ste
r
ior atta
chme
nt org
an
exci
sed
u
s
ing
a
scalpel
and
the
attachme
nt h
ooks
rele
ase
d
usi
ng
a p
r
ot
eina
se-K
ba
sed dig
e
stio
n f
l
uid (i.e. 1
00
_g/ml protein
a
se
K, 75 mM T
r
i
s
-HCl,
pH 8,
10 mM E
D
TA
, 5% SDS).
Once the
hoo
ks were
free
d
from e
n
cl
osi
ng
tissu
e, the
preparation
s
were
flushed
with distille
d water,
ai
r-drie
d
,
sputte
r-coat
ed with gold
and
then examin
ed an
d ph
otogra
phe
d u
s
i
ng a
JEOL
JSM52
00 scannin
g
ele
c
tron
mi
cro
s
co
pe
operating at a
n
accele
ratin
g
voltage of 10kV.
2.2. Curren
t
Appro
ach
A numbe
r of
statistical cl
a
ssifi
cation
ba
sed
app
ro
aches
appli
ed t
o
morphol
ogi
cal d
a
ta
[8-10], and m
o
lecular-ba
se
d techni
que
s
targeting
spe
c
ific ge
nomi
c
region
s [11,
12], have be
en
develop
ed to
discriminate
the
patho
g
enic spe
c
ie
s, G. salari
s,
from
othe
r non
-p
athog
enic
spe
c
ie
s
of Gy
roda
ctylus th
at co
-o
ccur o
n
salmoni
d h
o
sts.
Whil
e e
a
ch
techniq
u
e
is abl
e to
d
e
tect
G.
sal
a
ri
s within
a pop
ulati
on
of spe
c
im
ens and
to
di
scrimin
a
te it from it
s cong
eners
with hi
gh
levels of
co
rrect
cla
ssifi
cat
i
on, the te
ch
nique
s
can
b
e
time
con
s
u
m
ing [7]. If image
re
cog
n
i
tion
softwa
r
e coul
d be develop
ed to extract key discri
min
a
tory feature
s
from the attachme
nt hoo
ks of
each sp
ecie
s, then it is a
n
ticipate
d
tha
t
the
identification pro
c
e
ss could be a
c
cele
rated
with
equivalent o
r
better rate
s o
f
correct ide
n
tification.
Several efforts have be
e
n
devoted to
the
recognit
i
on of digital
image
s, especi
a
lly
microsco
pe i
m
age
s, but
so far it i
s
still
an un
re
so
lve
d
problem [1
3, 14], due to
disto
r
tion, no
ise,
segm
entation
erro
rs,
overl
ap a
nd
occlu
s
ion
in
co
l
o
u
r
ima
g
e
s
. Re
cog
n
ition an
d
cl
assification
techni
ques have gained a lot of a
ttention
in recent years
due
to
many scientists utilising these
techni
que
s in
orde
r to enh
ance their o
w
n probl
em do
mains.
To p
r
ovide
a
potential
sol
u
tion to the
probl
em d
e
scrib
ed
above
,
image
anal
ysis i
s
explore
d
. Image analy
s
is is a field of scie
n
ce
whi
c
h allo
ws
sci
entists to explore a
com
p
lex
assortm
ent of images
a
nd effectivel
y predi
ct
structure from
the image
s autonomo
u
s
ly.
Acco
rdi
ng to
Kasturi
[15], i
m
age
analy
s
i
s
refers to
alg
o
rithm
s
a
nd t
e
ch
niqu
es th
at are ap
plied
to
image
s to
ob
tain a
com
p
u
t
er read
able
descri
p
tion f
r
om pixel
dat
a. Instea
d of
image
an
alysis,
image processing
techniq
u
e
s have
also
been
devel
op
ed. In
cont
ra
st with im
age
analysi
s
, ima
g
e
pro
c
e
ssi
ng i
n
volves the
use of el
e
c
troni
c tool
s
whi
c
h
allow th
e u
s
e
r
to d
e
fine
ch
ange
s
within
the
para
m
eters
o
f
the ele
c
tro
n
ic
signal
[16].
This
app
roa
c
h is
nee
ded t
o
be
appli
ed
to increa
se th
e
pictori
a
l information for h
u
man inte
rpretation.
One
of the examples
of im
age processi
ng is
removing the illumination from images.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Feature Extra
c
tion an
d Cla
ssifi
cation for
Multip
le S
p
e
c
ies of
Gy
ro
da
ct
ylu
s
…
(Ro
z
niza A
li)
505
Re
cog
n
isin
g spe
c
ie
s gro
u
p
from
the hoo
k
featu
r
e
s
ma
ke
s th
e
sp
eci
e
s
re
cognition
pro
c
e
s
s mo
re
accu
rate
an
d effective. F
eature
extra
c
tion is the
ke
y to both
obj
ect
seg
m
enta
t
ion
and recogniti
on, as it is to
any pattern
cl
assificati
o
n
ta
sk. Exampl
es of t
he features that mig
h
t be
of intere
st to extract in
clu
de length,
wi
dth,
sha
pe a
nd angl
e. In the manu
al
measurement
of
feature
s
, the
s
e ta
sks he
avily depen
d
on the
co
nce
n
tration
of the pe
rson ta
king t
h
e
measurement
; otherwi
se, the re
sult
of
morp
homet
ric analysi
s
will
be false. An
d of cou
r
se, the
temporal d
u
ration of th
e
manual
mea
s
ureme
n
t proc
e
s
s
i
s
sub
s
tantial. Wi
th
state-of-the
-art
comp
uter p
r
oce
s
sing te
chniqu
es, the
s
e p
r
o
c
e
s
se
s are po
ssi
b
l
e to be don
e efficiently and
effectively, and thus, will provide the pre
d
icti
on in a
sh
orter time an
d more a
c
curately.
Feature extraction i
s
e
s
sential in m
any
vision
and bio
m
etri
c ap
plicatio
n
s
. The
perfo
rman
ce
of feature
-
ba
sed fa
ce
re
cognition
algo
rithms
reli
es
heavily on th
e quality of t
he
feature extra
c
tion. Selectio
n of
a feature
extraction m
e
thod is p
r
ob
ably the singl
e most impo
rtant
factor i
n
a
c
hi
eving hig
h
re
cog
n
ition p
e
rf
orma
nce
[17,
18]. In this
study, accuracy in the featu
r
e
extraction i
s
a must, sin
c
e, the majority of
t
he Gyroda
ctylus sp
ecie
s have a
similar shap
e to
each other, e
s
pe
cially G. salari
s and G.
thymalli [7].
In hum
an
co
mmuni
cation,
sh
ape
de
scri
ption
(f
eature
s
) have
be
en
used. It i
s
o
n
e
of the
most imp
o
rta
n
t visual attri
butes
of an
o
b
ject a
nd the
first u
s
ed t
o
perfo
rm o
b
je
ct cla
s
sificati
on
and identifica
t
ion [19, 20].
Specifici
a
lly, in classifi
catio
n
and identifi
c
ation of mult
iple spe
c
ie
s o
f
Gyroda
ctylus,
sha
pe info
rm
ation ha
s b
e
e
n
used,
altho
ugh different
method
s of id
entification
h
a
s
been a
pplied
[8-10], [21].
One of th
e
obje
c
tive of this
re
sea
r
ch
is to id
entify and utili
se
an ima
ge
p
r
ocesi
ng
technology that has ability to ex
tract the noi
sy im
ages
with
simi
l
a
r
pattern representation. For
these
re
ason
s, the A
c
tive
Shape
Mo
d
e
l (ASM)
te
chniqu
e h
a
s b
een
explored
to evalu
a
te
th
e
suitability of
using it for extracting i
n
formative f
eatures of m
u
lptipl
e spec
i
e
s
of Gyrodactylus.
In
the ca
se of SEM images o
f
the fish parasite, onl
y the sha
pe feature
s
are
con
s
ide
r
ed, si
nce, it
found that th
e texture info
rmation d
o
e
s
not incr
ea
se
the acuu
ra
cy of predictin
g the spe
c
ie
s.
Shape
or
co
ntour
refe
rs to the b
oun
d
a
ry of th
e
ob
ject, an
d that
rep
r
e
s
ent
s t
he
shap
e of
the
obje
c
t.
2.3. Potential Solutions
ASMs metho
d
have
been
su
ccessfully
utilised
for u
nderstan
ding
of
factors
u
nderlyin
g
morp
holo
g
ic
and pit
c
h
rel
a
ted fun
c
tion
al variatio
ns
affecting vo
cal structu
r
e
s
and the
ai
rway in
health an
d di
sea
s
e [2
2]. In addition, the
ASM me
thod
wa
s found to
be the be
st m
e
thod that ca
n
accou
n
t for the varieties in
variation [23].
Another
su
ccessful appli
c
a
t
ion of ASM for face
re
cog
n
ition [24]. In this study, ASM wa
s
applie
d to the alignme
n
t of the face, with four
maj
o
r improvem
ents. The
s
e
are: (1
) a m
odel
combi
n
ing
a
Sobel filter
[25] and the
2D p
r
ofile
i
n
se
archin
g
for a face in
an imag
e; (2
)
appli
c
ation
of
the
Ca
nny [
26] meth
od f
o
r
edg
e e
nha
ncem
ent; (3)
use
of
a SV
M to
cla
ssify
the
landma
r
k poi
nts; and
(4
)
automatic
adj
ustment
of the 2D profile
according to
the si
ze of t
he
input imag
e.
With the i
n
tro
ductio
n
of thi
s
im
p
r
ovem
e
n
t, it has imp
r
oved th
e pro
c
e
ss
of findin
g
landma
r
ks a
n
d
thus will
sa
ve time durin
g the training
and testin
g of images.
ASM was al
so implemente
d
for extracti
ng feature
s
for plant reco
gnition ba
sed
on the
leaf sha
pe [2
3]. In this study, ASM was applied fo
r reco
gni
sing
weed spe
c
ie
s, and du
e to using
the ASM, it
wa
s found to
be possible
to not only ta
ke leaf sha
pes into a
c
count, but also the
overall ge
om
etry of the se
edling
s
.
With the stati
s
tical
sha
pe
model
s, sha
p
e
c
an b
e
ch
a
r
acte
ri
sed in t
e
rm
s of inde
pend
en
t
mode
s of va
riation. Variati
on in th
e im
age p
r
e
s
e
n
ta
tion is
a
key
point that n
e
eds to focused.
This is be
ca
u
s
e
a
singl
e
speci
e
s may
come in
a
va
ri
ation, yet still
be
pa
rt of th
e same
spe
c
i
e
s.
For exam
ple,
locatio
n
and
water te
mp
eratu
r
e
can contri
bute
to
these differences.
Altho
ugh
dep
site this variation, the o
v
erall sh
ape
of the hook
re
mains the
sa
me.
The ASM te
chni
que
pe
rmits u
s
e
r
s to con
s
tru
c
t
a ge
neral
shape
mod
e
l
whi
c
h
is
sub
s
e
que
ntly applied to all images in o
r
der to landm
ark the ima
g
e
area for ev
ery given ima
ge,
providin
g a p
a
ttern that en
cap
s
ul
ates th
e vari
ation
se
en acro
ss th
e
range
of sha
pe image
s. T
h
e
sub
s
e
que
nt ability (cla
ssifi
cation
rate)
o
f
the develop
ed mod
e
l to sep
a
rate
”im
age cl
asse
s” is in
part ba
sed o
n
the numbe
r of images used in the tr
aining set - in theory,
the greater the nu
mber
of images th
at are used in training an
d con
s
tru
c
tin
g
the models, the better the cla
ssifi
cati
on
ability of the resultant model. Given the su
cce
ss
of ASM in resolvin
g ima
ge-b
a
sed, sh
ape
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 3, March 2
015 : 503 – 5
1
1
506
recognitio
n
p
r
oblem
s
within
the biom
edi
cal sp
he
re,
th
e re
se
arch
prese
n
ted in
thi
s
chapte
r
ai
ms
to determin
e
its utility when
applied to
SEM images of
Gyroda
ctylus hooks.
The appli
c
ati
on of the ASM method to the analys
i
s
of Gyroda
ctylus attachmen
t
hooks is
pre
s
ente
d
in Figure 1. The input for the cla
ssi
fi
cation
system is the
speci
m
en im
age
s, whe
r
e a
pre
-
processin
g
step i
s
app
lied to the re
quire
d
imag
e
s
. On
ce ho
oks have b
een
pro
c
e
s
sed to
a
comm
on
orie
ntation, the A
S
M app
roa
c
h
is th
en
appl
i
ed to
extra
c
t
informative fe
ature
s
. Th
ese
feature
s
are then redu
ced
by a sub
s
eq
u
ent PCA step
to sele
ct key
feature
s
to b
e
use
d
a
s
inp
u
t
feature
s
for each of the machi
ne lea
r
ning cla
s
sification techni
q
ues. Fou
r
m
a
chi
ne lea
r
ni
ng
cla
ssifie
r
s h
a
v
e been u
s
ed
to evaluate the ASM perfo
rman
ce.
2.3.1. ASM Constr
uction
ASM were o
r
iginally deve
l
oped for the
re
co
gnition
of landmarks on medical
x-rays.
Landm
ark p
o
i
nts
can
be a
c
qui
red
by a
pplying a
sa
mp
le templ
a
te to a
”p
robl
em area
”, wh
ich
appe
ars to re
pre
s
ent a b
e
tter strate
gy over edge
ba
se
d detectio
n
a
ppro
a
che
s
[27], as any noi
se
or un
wante
d
obje
c
ts withi
n
the image can be igno
re
d in the sele
ction of the shape
conto
u
r. In
the current st
udy, the sha
pe of each a
ttachment
ho
ok imag
e is pre
s
ente
d
by a vector of the
pos
ition of eac
h landmark
,
D =
(d1;
e1; :::; dn;
en), where (d
iei) denotes the 2D image
coo
r
din
a
te of
the ith land
mark poi
nt. The shape ve
ctor of the
h
ook i
s
then
n
o
rmali
s
e
d
into a
comm
on
coo
r
dinate
syste
m
. Procru
ste
s
an
alysi
s
is
then ap
plied
in alignin
g
th
e trainin
g
set of
image
s. Thi
s
align
s
ea
ch
sh
ape
so th
at the
sum
of dista
n
ces
of ea
ch
sha
pe to
the
m
ean
is minimi
sed
.
For this p
u
rpose, one
h
o
o
k ima
ge is
selecte
d
as
an
example
initial estimat
e
of the mean
shap
e and
sc
ale
d
so that
, which mini
m
i
se
s the F.
Figure 1. The
methodolo
g
i
c
al ap
pro
a
ch use
d
in the current study
Specim
en
s o
f
Gyroda
ctylu
s
were pi
cke
d
from the
skin and fin
s
of
salm
onid
s
a
nd thei
r
attachme
nt h
ooks
rel
e
a
s
e
d
by
prote
o
lytic di
ge
sti
on. I
m
age
s
of the
sm
alle
st ho
o
k
stru
ctu
r
es,
the
margi
nal
hoo
k
sickle
s
whi
c
h
are the
ke
y to se
pa
ratin
g
spe
c
ie
s a
n
d typically
m
easure
le
ss t
han
0.007mm
in l
ength,
were
capture
d
u
s
in
g
a sca
nni
n
g
e
l
ectro
n
mi
cro
s
cope. T
he i
m
age
s
were
pre-
pro
c
e
s
sed be
fore bein
g
su
bjecte
d to an
Active Shape Model feat
ure extra
c
tio
n
step to defi
ne
110 lan
d
ma
rks a
nd to fit the model to
the training
set of hoo
k i
m
age
s. Then
, this informa
t
ion
were u
s
e
d
to train
4
cla
ssifie
r
s (K-NN, LDA,
M
L
P, SVM) and
sep
a
rate th
e four sp
eci
e
s of
Gyroda
ctylus
whi
c
h in
clud
e
s
the n
o
tifiable patho
gen,
G. sala
ris. Ab
breviation
s: K
-
NN, K Nea
r
e
s
t
Neig
hbo
rs; L
D
A,
Line
ar Discri
mina
nt Analysis
;
MLP, Multi-Layer Pe
rc
eptron; SVM, Support
Vector M
a
chi
ne.
Assu
ming s
sets of landma
r
k poi
nts Di
whi
c
h are aligned into a common
shap
e pattern
for ea
ch
spe
c
ie
s, if this
d
i
stributio
n
ca
n
be mod
e
lle
d,
then
n
e
w example
s
ca
n
be gen
erat
ed
simila
r to those in the ori
g
inal trainin
g
sets, an
d the
n
these n
e
w
sha
p
e
s
ca
n be examine
d
to
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Feature Extra
c
tion an
d Cla
ssifi
cation for
Multip
le S
p
e
c
ies of
Gy
ro
da
ct
ylu
s
…
(Ro
z
niza A
li)
507
deci
de wheth
e
r they rep
r
e
s
ent re
asona
ble example
s
. In particula
r, D = M(b
)
is u
s
ed to ge
nerate
new vecto
r
s,
whe
r
e
b i
s
a
vector of p
a
rameters
of th
e mod
e
l. If th
e di
stributio
n
para
m
eters
can
be mod
e
lled,
p(b
)
, these can then
be li
mited such
th
at the gen
era
t
ed D’
s a
r
e
si
milar to tho
s
e
in
the trainin
g
set. Similarly it shoul
d be p
o
ssi
ble to e
s
timate p(D) u
s
i
ng the mod
e
l. To simplify the
probl
em, Prin
ciple
Com
p
o
nent Analysi
s
(PCA) i
s
ap
plied, to re
du
ce the
dimen
s
ion
a
lity of the
data. PCA su
mmari
ze
s the
variation see
n
acro
ss
the
data, allowi
n
g
one to app
roximate any of
the origin
al points usi
ng a
model. The model con
s
tructed he
re was ba
sed o
n
68 SEM hook
image
s, ea
ch
with
45
point
s d
e
termi
ned
as th
e o
p
tima
l numb
e
r of l
andma
r
k p
o
in
ts to effe
ctively
cha
r
a
c
teri
se
the shap
e of each ho
ok.
The
sub
s
eq
uent PCA step red
u
ced the numbe
r of
extracted
sh
ape features, remo
ving
redu
nda
nt feature
s
an
d
retaining t
hose that best
cha
r
a
c
teri
se
morp
holo
g
ica
l
differences
betwe
en the true spe
c
ie
s o
f
Gyroda
ctylus.
2.3.2. ASM Fitting
Once the AS
M model
ha
s
been
co
nstru
c
ted, it is
imp
o
rtant to fit the define
d
mo
del to a
seri
es
of ne
w inp
u
t ima
ges to
dete
r
mine the
p
a
r
amete
r
s of
the model th
at are th
e b
e
st
descri
p
tors
of hoo
k shape.
ASM find
s t
he m
o
st
accurate
pa
ram
e
ters of the
d
e
fined
model
for
the ne
w ho
o
k
ima
g
e
s
. Th
e ASM fitting attempts to
”be
s
t fit” the
defined
mod
e
l pa
ramete
r to
each ima
ge.
Coote
s
et al. [28] explain
ed that
by a
d
justin
g ea
ch
model
pa
ra
meter from t
h
e
defined
mod
e
l will
permit an extractio
n
pattern of
t
he
ima
ge se
ries
to be cre
a
ted.
Du
rin
g
the
model
fitting, it mea
s
ures n
e
wly int
r
odu
ced im
age
s
an
d u
s
e
s
thi
s
m
odel to
corre
c
t the valu
es o
f
curre
n
t para
m
eters, leadi
ng to a better fit.
2.4. Machine
Learning Cl
assifier
s
Followi
ng AS
M and
PCA,
the d
a
ta
were
asse
sse
d
u
s
ing
four method
s
of machine
learni
ng cl
assifiers. The
s
e
are b
r
iefly de
scribe
d belo
w
.
Linear Disc
r
i
minant Ana
l
y
s
is (LDA):
LDA cla
ssifi
cation is p
e
rformed by finding a
linear
co
mbin
ation of featu
r
es wh
i
c
h
be
st ch
aracte
rise two o
r
mo
re cla
s
se
s of
obje
c
ts [29].
The
purp
o
se of LDA is to find a linear fun
c
ti
on of
.
K-Nea
r
es
t Neighbors (K-NN):
Th
e K-NN find
s the
K neare
s
t neighb
or an
d use
s
a
majority vote
to dete
r
mine
the cl
ass lab
e
l [30].
Th
e
o
b
jective i
s
to
cla
ssify a
n
un
kno
w
n
exam
ple
R, whe
r
e the
equatio
n is
.
Multi-Lay
e
r Percep
tron
(MLP):
Seve
ral layers of n
euro
n
s are d
e
sig
ned i
n
M
L
P or fee
d
forwa
r
d
neu
ral network.
Each l
a
yer i
s
compl
e
tely co
nne
cted t
o
the n
e
xt a
nd ea
ch
ne
u
r
on
cal
c
ulate
s
a transfo
rme
d
weig
hted
lin
e
a
r com
b
inat
i
on of it
s in
p
u
ts, with
the
vector of o
u
t
pu
t
activation
s from the pre
c
e
d
ing layer, th
e transpo
sed
column ve
ctor of weig
hts and a boun
ded
non-de
crea
sing no
n-lin
ea
r function
(sig
moid),
with
o
ne of the
wei
ghts a
c
ting
a
s
a traina
ble
bias
con
n
e
c
ted
to a con
s
tant
input [31
]. The
algorithm
is expressed a
s
.
Support Ve
ctor Ma
chin
e (SVM):
T
h
e goal of SVM is to produ
ce a mo
del whi
c
h
predi
cts a
sp
ecie
s
cla
s
s o
f
data in
stan
ces in
the
te
st
ing set which
are
given
on
ly the feature
s
.
SVM uses a
kernel to
ma
p the fe
ature
spa
c
e
(h
yp
erplane
) into
a
high-dimen
s
i
onal t
r
an
sformed
spa
c
e [32]. It
is expre
s
sed
as
.
For ea
ch
app
roa
c
h, a 10
-f
old cro
ss vali
dation was
u
s
ed i.e. the d
a
ta we
re divi
ded into
k
(10
)
sub
s
ets,
whe
r
e
k-1 subsets
were
use
d
fo
r train
i
ng an
d the remainin
g sub
s
et u
s
ed
as t
he
test set. This process wa
s rep
eated 1
0
times
usin
g a different test set on e
a
ch run and
the
averag
e cla
s
sificatio
n
perf
o
rma
n
ce com
puted.
3. Experimental Re
sults
Although the
attachme
nt a
pparatus
of
Gyroda
ctylus consi
s
ts
of three main el
em
ents (i.e
two larg
er ce
ntrally positio
ned an
cho
r
s or hamuli;
two conn
ectin
g
bars bet
ween
the hamuli; and,
16 pe
riph
eral
ly distribute
d
margi
nal h
o
o
ks), thi
s
st
u
d
y
set
s
out
t
o
cla
s
sif
y
sp
ecie
s b
a
s
ed
on
feature
s
extracted from th
e sickle
s of the marg
i
nal
hoo
ks only. As the study
is base
d
on
the
analysi
s
of b
i
ologi
cal stru
cture
s
, the
s
e
requi
re p
r
o
c
essing
sub
s
e
quent to cap
t
ure in o
r
de
r to
stand
ardi
se
d the positio
n a
nd format of the image.
Proce
s
sing to standardise the orientatio
n of
the imag
e is applie
d to
redu
ce
processing
tim
e
and
com
p
lexity during th
e traini
ng a
n
d
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 3, March 2
015 : 503 – 5
1
1
508
con
s
tru
c
tion
of the ASM model. Tab
l
e 1 pre
s
e
n
ts detaile
d a
c
cura
cy re
su
lts for sel
e
ct
ed
cla
ssif
i
e
r
.
Table 1. Cla
s
sificatio
n
rate
for multiple
speci
e
s of Gyroda
ctylus u
s
i
ng ASM approach
Classifier
Accu
racy
(%
)
LD
A
K-NN
MLP
SVM
97.14
±
6
.
02
93.71
±
8
.
90
98.72
±
3
.
83
91.03
±
7
.
96
Table 2. A co
nfusio
n matri
x
showi
ng the
classi
fication
of Gyroda
ctylus spe
c
imen
usin
g an L
D
A
cla
ssifie
r
. Each of G. derj
a
vinoide
s (G. der
) and G. salari
s (G. sal) have an indi
vidual
miscl
as
sif
i
c
a
t
i
on
G.
col
G.
der
G. sal
G.
tru
Pr
ecis
i
on
(%)
G.
col
G. der
G.
s
a
l
G.
tru
10
0 0
0
0
24
0 1
0 0
33
1
0 0
0
9
100
96
97.06
100
R
eca
ll
(%)
100
100
100
81.82
Table 3. Usin
g the K-NN cl
assifier, G. co
lemen (G. col
)
is man
age t
o
have full cla
ssifi
cation,
while oth
e
r speci
e
s (G. de
rjavinoid
e
s
(G. der), G. sa
laris
(G. sal
)
and G. truttae
(G. tru) re
ma
in
miscl
as
sif
i
ed
G.
col
G.
der
G. sal
G.
tru
Pr
ecis
i
on
(%)
G.
col
G. der
G.
s
a
l
G.
tru
10
0 0
0
2
22
0 1
0 0
33
1
0 1
0
8
100
88
97.06
88.89
R
eca
ll
(%)
83.33
95.6
5
100
80
Table 4. MLP
classifer p
e
rf
orm
s
well
with the
co
rre
ct
cla
ssif
i
cat
i
on
G.
colem
en (
G
.
col),
G.
derjavin
o
ide
s
(G. der) and
G. truttae (G. tru).
G.
col
G.
der
G. sal
G.
tru
Pr
ecis
i
o
n
(%)
G.
col
G. der
G.
s
a
l
G.
tru
10
0 0
0
0
25
0 0
0 0
33
1
100
100
97.0
6
Re
ca
ll
(%)
100
100
100
90
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Feature Extra
c
tion an
d Cla
ssifi
cation for
Multip
le S
p
e
c
ies of
Gy
ro
da
ct
ylu
s
…
(Ro
z
niza A
li)
509
Table 5. Usin
g SVM classif
i
er, only one
spe
c
ie
s (G.
colemen
(G. col)) h
a
s ma
na
ged to get full
c
l
as
s
i
fic
a
tion,
while the wors
t misc
la
s
s
i
fication is
G. truttae (G. tru).
G.
col
G.
der
G. sal
G.
tru
Pr
ecis
i
on
(%)
G.
col
G. der
G.
s
a
l
G.
tru
10
0 0
0
0
24
1 0
0 1
32
1
0 2
2
5
100
96
94.12
55.56
R
eca
ll
(%)
100
88.89
91.43
8
3
.
33
The L
D
A
cla
ssifie
r
, u
s
ing
110 of l
and
mark poi
nts
wa
s abl
e to
corre
c
tly cla
s
sify all
spe
c
ime
n
s
of Gyroda
ctylu
s
to their tru
e
cla
ss,
ex
cept for one
spe
c
ime
n
of G. derjavin
o
i
des
whi
c
h we
re classified as G.
truttae an
d
on
e sp
e
c
i
m
en
of G.
sa
laris ha
s bee
n ide
n
fified a
s
G.
truttae (T
able
2). T
he K-NN
cla
ssifie
r
al
so
ha
s si
m
ila
r t
r
u
e
clas
sif
i
cat
i
on
a
s
L
D
A
cla
ssif
i
e
r
(T
able
3), with a
dditi
on of on
e of the nin
e
G. tru
ttae sp
e
c
ime
n
s
wa
s misall
ocate
d
a
s
G.
sala
ris. T
he t
w
o
non-li
nea
r ap
proa
ch
es
ML
P (Table
4) a
nd SVM (Ta
b
l
e 5)
were al
so a
b
le to a
c
hieve high
ra
tes
of corre
c
t cla
ssifi
cation, b
o
th with 98.7
2
% and
91.0
3
%. Compa
r
i
ng the four classifiers, ML
P
cla
ssifie
r
is th
e leadin
g
for
cla
ssifie
r
in a
c
hievin
g the
highe
st cla
s
si
fication of the
four spe
c
ies
of
Gyroda
ctylus.
This is not surp
risi
ng, sin
c
e MLP is
a
well pe
rform
a
nce
cla
ssifie
r
in many field [33,
34].
The cu
rrent study is base
d
on a smalle
r se
t of highe
r quality SEM image
s and al
though
the avera
ge
corre
c
t cl
assi
fication is
hi
gher
(i
.e. 98.
53%) than th
at achieve
d
usin
g the L
D
A
method
appli
ed to 2
5
p
o
int
-
to-p
oint me
a
s
ureme
n
ts
m
anually extra
c
ted from lig
h
t
micrograph
s of
557
sp
eci
m
e
n
s
(i.e. 9
2
.59
%
and
98.5
3
%
) [35,
36], this
app
roa
c
h
app
ea
rs pro
m
ising
a
nd
n
o
w
will be ap
plie
d to hooks p
r
epa
re
d for light micros
co
py hopefully with equ
al or better rate
s
of
corre
c
t cla
s
si
fication. The
ASM-PCA-M
LP based
ap
proa
ch
applie
d to SEM image
s of the h
ook
sickle
s of
Gyroda
ctylus ap
pears
to
o
u
t perfo
rm ot
her metho
d
s tha
t
have b
een
tested
to id
ent
ify
and di
scrimin
a
te this sp
eci
e
s with
confid
ence.
With the
s
e
succe
ssful
re
sults fo
r extra
c
tion
and
cl
a
ssifi
cation, th
e difficultie
s f
a
ce
d by
domain
expe
rt can
be mi
ni
mised.
The
s
e
difficultie
s in
clud
e ma
nual
cla
ssifi
cation
, a tediou
s a
nd
time con
s
umi
ng pro
c
e
s
s. Another
chall
enge in t
he
manual a
p
p
r
oach, inaccu
rate point to point
measurement
s, whi
c
h resu
lt in inaccura
te spe
c
ie
s id
entification, can also be o
v
erco
me. No
w,
withthis n
e
wl
y applied co
mbination of tech
niqu
es
, d
o
main expe
rt
s use these method
s for feature
measurement
s and
spe
c
ie
s identification.
4. Conclusio
n
The current study set out to explore the
utility of a n
o
vel ASM-PCA-machine learning
cla
ssif
i
e
r
ba
s
ed ap
pro
a
c
h
i
n
cla
s
sif
y
ing
spe
c
ie
s
of
Gy
roda
ctylus which are ecto
para
s
ite
s
of
fish.
ASM applied
to 68 SEM image
s of the margin
al hoo
k
sickle
was a
b
l
e to overcom
e
the limitation
and difficultie
s in extractin
g
feature info
rmati
on fro
m
the hooks. T
he bes
t ap
proach, which use
d
a MLP meth
od of cla
s
sification, was able to
improve u
pon
the perfo
rma
n
ce of p
r
evi
ous
approa
che
s
(i.e. 98.72%
cf. 92.59%
using
an L
D
A-b
a
sed cl
assifier a
ppli
ed to manu
ally
extracted m
o
rphomet
ric d
a
ta). Future wo
rk
will
asse
ss the performa
n
ce of this m
e
thod on la
rg
er
datasets and will
expl
ore
new me
thods based on an ensemble
of cl
assifiers,
which have shown
promi
s
in
g
results, with
the
aim
s
of
providing
a
relia
b
l
e mo
del fo
r
the ide
n
tificat
i
on of
spe
c
ie
s,
inclu
d
ing the
pathog
en G. sala
ris, by no
n-expe
rts a
n
d
fish health re
sea
r
che
r
s.
Ackn
o
w
l
e
dg
ements
Gratefully a
c
kno
w
le
dge
s
the collobo
ra
tion
wo
rk b
e
t
ween
Stirlin
g University, United
Kingdom a
n
d
Anhui Uni
v
ersity, Chin
a (Sino-UK
High
er Edu
c
ation Re
se
arch Partn
e
rsh
i
p
Fundin
g
Call
). To the team of parasitol
o
g
y, Stir
ling University, thank you
for all
o
win
g
the use of
Gyroda
ctylus
dataset in this study.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 3, March 2
015 : 503 – 5
1
1
510
Referen
ces
[1]
P. D. Harris, A. P. Shinn, J.
Cabl
e, and T
.
A. Bakke, Nominal sp
eci
e
s of the genus
G
y
rod
a
ct
ylus v
.
Nordma
nn
18
32 (Mo
no
ge
ne
a: G
y
ro
dact
y
l
i
dae),
w
i
t
h
a
l
i
st of pri
n
cip
a
l
host sp
eci
e
s, S
y
stem
ati
c
Parasito
log
y
, 2
004; 59: 1
–27.
[2]
G. Malmberg,
T
he excretor
y s
y
stems a
n
d
margi
n
a
l
ho
oks as a
bas
is for the s
y
s
t
ematics of
G
y
rod
a
ct
ylus (
T
rematoda, Monog
en
ea), Arki
v for
Z
oologi S
e
ries 2-B
and
2
3
(nr. 1). 1970;
1–23
5.
[3]
T
.
A. Bakke, J. Cabl
e, an
d P. D. Harris, “T
he
biol
og
y
of g
y
r
odact
y
l
i
d mo
no
gen
ea
ns: the ”
R
ussia
n
-d
ol
l
killers”,”Adv
an
ces in Paras
i
tol
o
g
y
, vol. 64, pp
. 161–3
76, 20
0
7
.
[4]
I. Alenas, “G
yrodact
y
l
u
s sal
a
ris pa l
a
x i
svenska vatte
ndra
g
och l
a
xprob
l
ematik
en
p°a svensk
a
vastk
y
ste
n
[In S
w
e
d
i
sh],” Van
n
, vol. 199
8; 1: 135–
14
2.
[5]
E. Ieshko, B.
Berland, B. S
hulma
n, G. Bristo
w
,
a
nd I. Shurov, Pa
rasites
of Atlantic sal
m
on parr i
n
some Kar
e
li
an
rivers flo
w
i
n
g into th
e W
h
ite Se
a,
Abs
t
ract F
ourth Internati
o
n
a
l S
y
mp
osi
u
m o
f
Ichthyopar
asito
l
ogy
, 19
95; 10
0.
[6]
P. Rintamaki, G
y
rodact
ylus s
a
laris at a fish farm in norther
n F
i
nlan
d,
Parasites of F
r
eshw
ater F
i
she
s
in North-W
e
st Europ
e
(O.N.
Bauer, ed.)
, 19
89; 123
–1
30.
[7]
A. P. Shin
n, C. Col
lins, A. Ga
rc´
ı
a-V´as
qu
ez
, M. Sno
w
, G. Pala
din
i
, T
.
Lindenstrø
m
, M. Lon
gsh
a
w
,
I
.
Mat
˘
ejus
ov´a,
D. M. Stone, J.
F
.
T
u
rnbull, S.
M. Pic
on-
Cam
a
cho,
C. V´azq
uez R
i
ver
a
, R.
A. Dug
u
id, T
.
A. Mo, H. H
a
n
s
en, K. Olsta
d
, J. Ca
ble,
P. D
.
Ha
rris, R.
Ker
r
, D. Graham,
G. H. Yoon,
K. Buchm
a
n
n
,
R. Ra
yn
ard, S
.
Irving, an
d J
.
E.
Bron, Mul
t
i-centre testin
g an
d va
lid
ati
on of c
u
rrent
protoco
l
s for
G
y
rod
a
ct
ylus
salaris (M
ono
gen
ea) i
dentifi
c
ation,
Intern
a
t
iona
l Jour
nal
of Parasitol
o
gy
, 201
0; 40
:
145
5–
146
7.
[8]
S. E. McHugh,
A. P. Shinn, a
n
d
J. W
.
Ka
y
,
Di
scr
iminati
on
of G. salaris a
nd
G. th
y
m
al
li us
i
ng statistica
l
classifiers
app
li
ed to morph
o
m
e
tric data,
Parasitology
, 200
0; 121: 31
5–
32
3.
[9]
A. P. Shinn, H. Hanse
n
, L.
Bachma
nn, a
nd T
.
A. Bakke, T
he use of mo
rphom
etric
characters t
o
discrimi
nate s
pecim
ens of l
abor
ator
y-rear
ed an
d
w
i
l
d
p
opu
latio
n
s of G
y
rod
a
ct
ylus
salaris a
nd G.
th
y
m
a
lli (mo
no
gen
ea),
Foli
a Parasito
log
i
ca
, 2
004; 51: 2
39–
2
52.
[10]
A. P. Shinn, J.
W. Kay
,
and
C. Sommerv
ill
e, T
he use
of
statistical
c
l
ass
i
fier for
the
dis
c
riminati
o
n
o
f
speci
e
s of the
gen
us G
y
ro
d
a
ct
y
l
us (Mon
o
gen
ea) p
a
rasit
i
zing s
a
lmo
n
i
d
s,
Parasitol
ogy
, 2000; 1
20:
261
–2
69.
[11]
C. O. Cun
n
in
gham,
D. M.
McG
illivra
y, K.
MacKe
n
zie,
and
W
.
T
.
Me
lvin, D
i
scrimi
n
ation
bet
w
e
e
n
G
y
rod
a
ct
ylus
salaris,
G. de
rjavin
i
and
G. trutt
ae (Pl
a
t
y
helmi
nthes:
Mono
ge
nea)
us
i
ng r
e
strictio
n
fragment le
ngt
h pol
ym
orph
is
ms and an o
l
i
g
onuc
leoti
de
pr
obe
w
i
t
h
in th
e small su
bun
it ribosom
al RN
A
gen
e,
Parasito
l
ogy
, 199
5; 111
: 87–94.
[12]
H. Hanse
n
, L
.
Bachmann,
and B. T
.
A., Mi
tochondr
ia
l DNA vari
ati
on of G
y
ro
da
ct
y
l
us s
pp.
(Mono
gen
ea, G
y
rod
a
ct
yli
dae
)
pop
ulati
ons i
n
fecting
At
l
anti
c
salmo
n
, gra
y
l
i
ng,
and
rai
n
bo
w
tro
u
t i
n
Nor
w
a
y
and
S
w
e
d
en,
Parasit
o
logy
, 200
3; 33: 1471
–1
47
8.
[13]
T
.
Moyo, S. Banga
y, an
d G. Foster,
T
he iden
tifi
cation of ma
mmalia
n speci
e
s throug
h the classificati
o
n
of
ha
ir patterns
usi
ng ima
ge p
a
ttern
rec
o
g
n
iti
on,
C
o
mput
er
Graphic, V
i
rtua
l R
eal
ity, Visu
a
lisatio
n
an
d
Interaction
, 2
0
06; 177
–1
81.
[14]
C. H. L
a
i, S. S
.
Yu, H. Y.
T
s
eng,
an
d M. H
.
T
s
ai, A proto
z
oan
par
asite
extracti
on
sche
m
e for
dig
i
tal
microscop
e
im
ages,
Co
mpute
r
i
z
e
d
Me
dica
l Ima
g
i
ng an
d Graph
ics
, 201
0; 34: 122–
13
0.
[15]
R. Kasturi, L. O’Gorman,
an
d V. Govind
ar
aju, D
o
cume
nt image
an
al
ysi
s
: A primier,
S
adh
an
a, India
n
Acade
my
of S
c
ienc
es, Spec
i
a
l Issue
o
n
In
dia
n
L
ang
ua
ge
Docu
ment A
n
alysis
an
d U
n
derstan
din
g
,
200
2; 27(1): 3–
22.
[16]
C. S. Z
u
ria, R. J. M., D.
Baez-Lop
ez, and
G. E.
F
l
ores-Verda
d,
Matlab
base
d
ima
ge p
r
ocessi
ng l
a
b
exper
iments,
F
r
ontiers i
n
Edu
c
ation
, 19
98; 3
:
1255–
12
58.
[17] O. D.
T
r
ier, A.
K. Jains, and
T
.
T
a
xt, F
eature extracti
on m
e
thods fo
r char
acter recog
n
iti
on - a surve
y
,
Pattern Reco
g
n
itio
n
, 199
6; 29(4): 641
–6
62.
[18]
G. Kumar and P. K. Bhatia, A detaile
d revie
w
of featur
e extracti
on i
n
image proc
es
sing s
y
stems,
Advanc
ed C
o
mp
utin
g an
d C
o
mmunic
a
tio
n
T
e
chno
log
i
es
, 201
4;
321
–3
31
.
[19]
S. Loncaric, A surve
y
of shap
e ana
l
y
sis tech
niq
ues,
Pattern
Recog
n
itio
n
, 1
998; 31(
9): 983
–10
01.
[20]
R. d. T
o
rres, L
.
F
a
lcao,
an
d
L. d. F
.
C
o
sta, A gr
ap
h-bas
e
d
a
ppr
o
a
ch
for
multisc
a
le
sh
ape
a
nal
ys
is,
Pattern Reco
g
n
itio
n
, 200
3; 37: 1163
–1
17
4.
[21]
A. P. Shinn, D. I.
Gibson,
and C. Sommervillae, Morphomet
r
ic di
scrimi
nati
on of G
y
ro
dact
y
l
u
s sal
a
ris
malmber
g (mo
nog
en
ea) fro
m
species of
Gy
r
o
d
a
ct
yl
us
paratisi
ng bri
t
ish salmo
n
ids
using n
o
ve
l
param
eters,
Fish Diseases
, 2
001; 24: 8
3–9
7
.
[22]
N. A. Miller, J.
S.
Gregor
y
,
R. M. Aspden, P. J. Stoller
y
, a
n
d
F.
J. Gilbert, U
s
ing activ
e
sha
pe mod
e
li
n
g
base
d
o
n
fMRI
g to stud
y
mor
pho
log
i
c a
nd
p
i
tch-rela
t
ed fu
n
c
tiona
l ch
ang
e
s
affecting v
o
c
a
l structures
and the
air
w
a
y
,
Voice
, 201
4.
[23]
S_ga
ard, W
e
e
d
classific
a
tion
b
y
active sh
ap
e mode
ls,
Auto
mati
on a
nd E
m
ergi
ng T
e
ch
nol
ogi
es
, 200
5;
91(3): 27
1–
281
.
[24]
H. Le a
nd
N. Vo, F
a
ce al
ig
n
m
ent usi
ng
ac
tive sh
ape
mod
e
l a
nd s
upp
ort vector mach
in
e,
B
i
om
e
t
r
i
cs
and Bi
oinf
orma
tics
, 2012; 4(6)
: 224–2
34.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Feature Extra
c
tion an
d Cla
ssifi
cation for
Multip
le S
p
e
c
ies of
Gy
ro
da
ct
ylu
s
…
(Ro
z
niza A
li)
511
[25]
M. Salahs
ho
or
, A. Brouman
d
n
ia
, a
nd M. R
a
stgarp
our, Ap
plicati
on
of
int
e
lli
ge
nt s
y
stem
s for iran
ia
n
licens
e pl
ate re
cogn
ition,
Intelligent System
s.
IEEE
, 2014; 1–6.
[26]
K. J. Karande
and S.
N. T
a
lbar, Ca
nn
y e
d
ge d
e
tectio
n for face reco
gn
ition us
in
g ica,
Indep
en
den
t
Co
mp
one
nt An
alysis of Edg
e
Informati
o
n
for F
a
ce Reco
gniti
on. Sprin
ger
, 2
014; 21
–3
3.
[27]
R. Maini
an
d H
.
Aggar
w
a
l,
Stud
y a
nd c
o
mp
ariso
n
of vari
o
u
s ima
ge e
d
g
e
detectio
n
tech
niq
ues,
Imag
e
Processi
ng
, 20
09; 3(1): 1–
60.
[28]
T
.
F
.
Cootes,
G. J. Ed
w
a
r
d
s
,
and C.
J. T
a
ylor, Active
a
ppe
aranc
e mo
dels,
Pattern Analys
is
an
d
Machi
ne Intell
i
genc
e
, 200
1; 23(6): 681
–6
85.
[29]
J. Lu, K. N. Plat
ani
otis, and
A. N. Venetsanop
oul
os, F
a
ce re
cog
n
itio
n u
s
ing LD
A base
d
alg
o
rithms,
Neur
al Netw
or
ks
, 2002.
[30]
J. Song, Y. Hu
ang, D. Z
h
o
u
,
H. Z
ha, and C
.
L. Giles, IKNN: In
formative k-near
est
nei
g
hbor patter
n
classification,
Sprin
ger-Ver
la
g Berli
n
Hei
del
berg
, 20
07; 24
8–2
64.
[31]
V. H. C. de Al
buq
uerq
ue, A.
R. de Al
e
x
an
d
r
ia, P. C. Cortez,
and J. M. R. S.
T
r
avers, Evalu
a
tion
of
multila
ye
r perc
eptron a
nd sel
f
-organ
izin
g map ne
ural n
e
tw
o
r
k topo
log
i
e
s
appli
ed to
microstructu
r
e
segme
n
tatio
n
from metall
ogra
phic im
ages,
N
D
T
& E Internationa
l
, 200
9; 42: 644–
65
1.
[32] Y.
Xi
e,
An
i
n
troducti
on t
o
s
upp
ort vector
mac
h
i
ne
an
d i
m
p
l
e
m
e
n
tatio
n
in
R
, Proc
ee
din
g
s of t
h
e
Internatio
na
l Joint Co
nfere
n
c
e
on Ne
ural N
e
t
w
orks, 2
007.
[33]
L. P. Corde
lla
, R. Limon
g
ie
llo, a
nd C. S
ans
o
ne, Net
w
ork intrusi
on
detectio
n
b
y
a multi-sta
g
e
classificati
on s
y
stem, In: Rol
i
,
Kittler, and
W
i
nde
att (Eds.): Multiple
Cla
ssifier S
y
stem
s,
LNCS 30
77.
Sprin
ger
, 20
04
; 324–3
33.
[34]
M. F
.
Kabir, D.
L. Schm
oldt,
P. A. Araman,
M. E. Schafer,
and
S.
-M. Lee,
Class
if
yin
g
d
e
f
ects in
pal
let
stringers b
y
u
l
trason
ic scann
i
ng,
W
ood a
nd
F
i
ber Scie
nce
, 200
3; 34.
[35]
R. Ali, A. H
u
ss
ain, J. E. Bro
n
,
an
d A. P. Sh
in
n, Multi-sta
ge c
l
assifi
cati
on
of
G
y
rod
a
ct
ylus
s
peci
e
s us
ing
machi
ne le
arni
ng an
d featur
e selecti
on te
chni
ques,
Intel
lige
n
t Systems
Design a
nd
Appl
icatio
ns
,
201
1; 457
–4
62
.
[36]
——,
T
he use of asm f
eature
extracti
on a
nd
machi
ne le
arn
i
ng for
the discr
iminati
on of me
mbers of th
e
fish ectopar
asit
e gen
us G
y
ro
d
a
ct
y
l
us,
Neur
al
Informati
on Pr
ocessi
ng
, 20
11
; 7666: 45
7–4
6
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.