Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
9
, No
.
6
,
Decem
ber
201
9
, p
p.
5192~
5204
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v9
i
6
.
pp5192
-
52
04
5192
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Robust featu
re ext
ractio
n metho
ds for ge
neral fis
h classifi
cation
Mutase
m K
. Alsm
adi
1
,
Mohamme
d T
ayf
ou
r
2
,
Ra
e
d
A.
Alkh
as
awne
h
3
,
Us
am
a
B
ada
w
i
4
,
Ibra
him
A
lm
ar
as
hde
h
5
,
Fir
as H
ad
d
ad
6
1,
2,4,5
Depa
rtment
of
Mana
gemen
t Inform
at
ion
S
y
s
t
ems
,
Coll
eg
e
of
Applie
d
Stud
ie
s
and
Com
m
unity
Servic
e
,
Im
am Abdu
l
rah
m
an
Bin
Faisa
l U
nive
rsit
y
,
Saud
i
Arabi
a
3,6
Depa
rtment
of
Gene
r
al
Course
s,
Coll
ege of
Ap
pli
ed
Studie
s
an
d
Com
m
unity
Se
rvic
e
,
Im
am Abdul
rah
m
an
Bin
Faisa
l U
nive
rsit
y
,
Saud
i
Arabi
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ma
r
2
, 2
01
9
Re
vised
Ju
l
6
,
201
9
Accepte
d
J
ul
17
, 2
01
9
Im
age
rec
ogniti
on
proc
ess
coul
d
be
pla
gued
by
m
an
y
probl
e
m
s
inc
ludi
ng
noise,
ov
erlap,
distort
ion,
err
o
r
s
in
th
e
out
co
m
es
of
segm
ent
at
ion
,
and
impediment
of
obje
c
ts
withi
n
the
image.
Base
d
on
fea
ture
se
le
c
ti
on
and
combinat
ion
th
e
or
y
be
twee
n
m
a
jor
ext
ra
cted
feature
s,
thi
s
stud
y
at
t
empts
to
esta
bli
sh
a
s
y
st
e
m
tha
t
coul
d
recogniz
e
fish
object
withi
n
the
image
uti
l
iz
ing
te
xtur
e,
an
chor
point
s,
and
sta
tis
ti
ca
l
m
e
asure
m
ent
s.
Th
en,
a
g
ene
ri
c
fish
cl
assifi
ca
t
ion
is
exe
cu
te
d
with
th
e
application
of
an
innovative
class
ifi
cation
eva
lu
at
ion
throu
gh
a
m
eta
-
heur
isti
c
al
gor
it
hm
kn
own
as
Mem
et
ic
Algorit
hm
(Gene
tic
Algori
thm
with
Sim
ula
t
ed
Annea
ling)
with
bac
k
-
propa
gation
al
gorit
hm
(MA
-
B
Cla
ss
ifi
er
).
H
ere
,
images
of
d
ange
rous
and
no
n
-
dange
rous
fish
are
re
cognize
d.
Im
age
s
of
dange
rous
f
ish
are
furth
er
re
c
ogniz
ed
as
Preda
tor
y
or
Pois
on
fish
family
,
where
as
famili
e
s
of
non
-
dange
rous
fish
are
cl
assifi
ed
int
o
g
ard
en
and
food
f
amil
y
.
A
total
of
24
fish
famili
es
were
used
in
te
sting
the
proposed
protot
y
pe
,
where
b
y
ea
ch
famil
y
en
compass
es
diffe
ren
t
num
be
r
of
spec
ie
s.
Th
e
proc
ess
of
c
las
sific
at
ion
was
succ
essful
l
y
under
ta
k
en
b
y
th
e
proposed
pro
to
t
y
p
e, whe
reb
y
4
00
disti
nc
t
fish
i
m
age
s were
used
in
the
expe
rimental
te
sts.
Of
the
se
fish
images,
250
were
used
for
tra
ini
ng
ph
ase
while
150
were
used
for
te
sting
phase
.
The
ba
ck
-
pr
opaga
t
ion
al
gorit
hm
and
t
he
proposed
M
A
-
B
Cla
ss
ifi
er
produc
ed
a
gen
era
l
a
c
cur
a
c
y
rec
ogni
ti
on
r
ate of 82.
25
and
90
%
respe
ct
iv
ely
.
Ke
yw
or
d
s
:
An
c
hor
points
m
easur
em
ents
Ba
ck pr
op
a
gati
on alg
or
it
hm
M
et
a
-
heurist
ic
al
gorithm
Feat
ur
es
ex
t
rac
ti
on
Stat
ist
ic
al
m
ea
su
rem
ents
Textu
re m
easur
em
ents
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Muta
sem
A
ls
m
adi,
Dep
a
rtm
ent o
f M
IS
, C
ollege
of Ap
plied
Stu
di
es an
d
Com
m
un
it
y Se
rv
ic
e,
Im
a
m
A
bd
ul
ra
hm
an
Bi
n
Fais
al
U
ni
ver
sit
y,
Al
-
D
am
m
a
m
,
Saudi A
ra
bia
.
Em
a
il
:
m
ks
al
s
m
adi@g
m
ai
l.c
om
,
m
kals
m
adi@iau.e
du.sa
1.
INTROD
U
CTION
The
tra
diti
on
al
i
m
age
recogni
ti
on
proces
s
m
os
tly
e
m
plo
yed
the
s
kill
s
and
se
ns
es
of
hum
an
wh
ic
h
has
ca
us
e
d
in
accurate
a
nd
un
s
at
isfact
or
y
rec
ogniti
on
process
.
D
ue
to
the
im
po
rt
ance
of
this
process
,
com
pu
te
rs
are
now used
f
or
re
cogniti
on
pro
cesses d
ue
to t
hei
r
le
vel
of
ac
cur
acy
a
nd
e
ff
i
ci
ency. In this
reg
a
rd,
a
nu
m
ber
of
appr
oach
es
ha
ve
been
em
plo
ye
d
for
im
age
processin
g
[1
-
8]
an
d
patte
r
n
rec
ogniti
on
[9
-
11]
.
Im
age
recog
niti
on
has
bee
n
act
ively
stud
ie
d
a
nd
se
ver
al
researc
hes
ha
ve
bee
n
c
onduc
te
d
on
this
s
ubj
ect
.
Ther
e
are
in
f
act
m
any
issues
ass
ociat
ed
with
im
age
re
cogniti
on
s
uch
as
distor
ti
on
,
obje
ct
ove
rlap
a
nd
blo
c
kad
e
in
dig
it
al
i
m
ages,
and
al
so
er
r
or
s
in
the
resu
lt
s
of
se
gm
entat
io
n
[12
-
15]
.
As
has
bee
n
show
n
i
n
pr
ese
nt
re
sear
ches,
t
he
al
re
ady
avail
able
fish
rec
ogniti
on
syst
em
s
are
sti
ll
la
cking
i
n
ce
rtai
n
areas.
Fo
r
insta
nce,
t
he
c
urren
t
syst
e
m
s
are
sti
ll
no
t
suffici
ently
able
to
detect
and
cl
assify
fi
sh
.
A
par
t
from
that,
sign
ific
a
nt
nu
m
ber
of
daily
deaths
sti
ll
happens
owin
g
to
fail
ur
e
in
m
aking
disti
nctio
n
betwee
n
fis
hes
that
are
dange
rous
a
nd
tho
se
that ar
e
not
[16
-
20]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Ro
bu
st f
e
atu
re
ext
ra
ct
ion met
hods
f
or
ge
ner
al fi
sh
cl
as
sif
ic
ation
(
Muta
se
m
Als
madi
)
5193
The
m
ai
n
con
trib
ution
of
this
work
is
to
de
velo
p
a
syst
e
m
fo
r
fish
im
age
cl
assifi
cat
ion
,
wh
e
re
this
syst
e
m
e
m
plo
ys
ancho
r
points,
te
xt
ur
e
as
well
as
st
at
ist
ic
al
m
easur
em
ents
f
or
featu
res
e
x
tr
act
ion
.
Accor
dingly
,
the
cl
assifi
cat
ion
of
fish
im
ag
es
has
bee
n
ch
os
e
n
as
the
fo
c
al
po
int
of
this
stud
y.
Fish
im
age
of
certai
n
siz
e
a
nd
f
orm
at
is
the
input
of
the
pr
opos
e
d
syst
em
,
an
d
us
in
g
the
an
ch
or
poi
nts,
te
xtu
re
an
d
sta
t
ist
ic
a
l
m
easur
em
ents,
featu
re
s
of
the
fis
h
im
ages
will
be
e
xtracted
.
T
he
n,
us
i
ng
the
MA
-
B
Cl
assif
ie
r
on
the
extracte
d
featur
e
s,
the
f
ish
in
quest
io
n
w
ould
be
cl
assifi
ed
int
o
da
ng
e
r
ou
s
a
nd
non
-
da
nger
ous
fish.
This
fish
im
ag
e
will
f
ur
t
her
be
cl
assifi
ed
as
Pr
e
dator
y
or
P
oison
ou
s
fis
h
(
if
da
nger
ous)
,
or
ga
r
den
or
f
ood
fis
h
(if non
-
da
nger
ou
s
). T
his stu
dy
w
ould
b
e
of
value
t
o
m
any aren
as
inclu
di
ng m
arine,
in
du
stry an
d
a
gr
ic
ul
ture.
The
rest
of
the
pa
per
is
orga
nized
as
fo
ll
ows:
Sect
io
n
2
discusse
d
the
li
te
ratur
e
re
view.
Sect
io
n
3
descr
i
bes
m
a
ter
ia
ls and
m
et
ho
ds
em
plo
ye
d.
Sect
ion
4
desc
r
ibes the classi
f
ie
r
arch
it
ect
ure
. S
ect
ion
5
disc
us
ses
and
eval
uates
the
obta
ine
d
r
esults
usi
ng
th
e
pro
posed
M
A
-
B
Cl
assifi
e
r
an
d
oth
e
r
cl
a
ssific
at
ion
m
eth
ods.
Sect
ion
6
c
onc
lud
es
the
pa
per
.
2.
LIT
ERATUR
E REVIE
W
Fish
recog
niti
on
is
ver
y
c
om
pl
ic
at
ed
an
d
dif
ficult
ta
sk
bu
t
is
use
f
ul
to
business
an
d
a
gr
ic
ultur
e
.
Distortio
n,
ov
erlap,
noise
,
di
stortion,
occlusio
n,
a
nd
al
s
o
er
r
or
i
n
se
gm
entat
ion
are
a
m
on
g
the
c
ha
ll
eng
es
faced
i
n
achie
ving
accu
rate
and
reli
able
fis
h
rec
o
gn
it
io
n
[2
1
-
3
4
]
.
Am
on
g
the
works
re
le
van
t
to
this
s
tud
y
is
on
e
from
Mok
ti
and
Sala
m
[35
]
.
I
n
their
w
ork,
the
a
uthor
s
app
li
ed
a
hy
br
i
d
of
Me
a
n
-
sh
ift
an
d
m
edi
an
-
c
ut
al
gorithm
s
in
their
segm
entat
ion
of
c
olor
fis
h
im
ages.
P
rio
r
to
that,
the
im
age
pre
-
proc
essin
g
te
ch
nique
was
app
li
ed
in
ord
er
to
im
pr
ove
the
im
age
be
fore
it
s
c
olo
r
sp
ace
was
tr
a
ns
f
or
m
ed
int
o
LU
V
c
olor
s
pace.
The
n,
t
o
cl
us
t
er
ar
ound
t
he
re
gion,
t
he
a
uthors
ap
plied
the
m
ean
-
sh
i
ft
segm
entat
ion
.
H
ow
e
ve
r,
since
the
segm
entat
ion
exec
uted
by
Me
an
-
s
hift
a
lgorit
hm
is
of
lo
w
le
vel,
c
ertai
n
re
gion
carries
no
se
m
antic
m
eaning
.
He
nc
e,
m
edian
-
cut
al
gorithm
can
be
use
d
as
so
l
ution
t
o
this
prob
le
m
.
Fu
rt
herm
or
e,
this
al
gorithm
has
the
ca
pa
bili
ty
to
red
uce
t
he
colo
r
de
pth
i
ns
ide
t
he
im
ag
e.
W
it
h
the
a
ppli
cat
ion
of
M
e
an
-
s
hift
a
nd
m
edian
-
cut
hy
br
id
,
the
resu
lt
s
wer
e
im
pro
ved
e
sp
eci
a
ll
y
in
te
r
m
s
regi
on
gro
up
i
ng.
Be
sides
that,
t
he
a
uthors
we
r
e
able
to
achieve
cl
earer
bor
der
s
of
se
gm
ented
re
gions
ea
sie
r.
Usi
ng
t
he
pr
opos
e
d
al
gorithm
pr
ovides
the
ad
va
ntages
of
Me
an
-
s
hift se
gm
entat
ion
m
et
ho
d.
At
t
he
sam
e
tim
e,
the
wea
kn
e
sses
as
so
ci
at
ed
with
t
he
us
e
of m
edian
-
cut
gro
up
i
ng m
et
h
od are
ev
a
de
d.
In
Alsm
adi
et
al
.
[
36
]
,
a
fi
sh
cl
assifi
cat
ion
prototype
was
pro
pose
d.
This
prot
otype
com
bin
es
betwee
n
the
fe
at
ur
es
e
xtracte
d
f
r
om
m
easurem
ents
of
sh
a
pe
an
d
siz
e
us
in
g
the
m
easur
e
m
ents
of
dista
nce
a
nd
geo
m
et
ry.
The
authors
em
pl
oyed
20
dif
fere
nt
fam
i
li
es
o
f
fish
co
ntaini
ng
diff
e
ren
t
fi
sh
ty
pes
each,
an
d
the
sam
ple
us
ed
co
ntains
a
total
of
35
0
di
ff
e
ren
t
ty
pes
of
fish
im
ages.
These
im
ages
wer
e
s
plit
into
two
dataset
s,
with
257
trai
ni
ng
i
m
ages in
o
ne d
at
aset
an
d
93 test
ing
im
ages i
n
the o
t
her
. T
he
au
th
or
s att
ai
ne
d
86%
accuracy
w
he
n
us
in
g
the
ne
ural
netw
ork
as
so
ci
at
ed
with
t
he
bac
k
-
pro
pa
gation
al
gorith
m
on
the
data
set
of
us
e
d
te
st.
Fro
m
the
resu
lt
s,
the
auth
ors
pr
ov
e
d
the
a
bili
ty
of
the
pr
opose
d
cl
assifi
er
i
n
cat
eg
or
iz
in
g
the
fish
into
it
s
c
orrect
cl
ass,
in
cat
e
gorizi
ng
the
cl
assed
fis
h
into
po
iso
n
or
no
n
-
poiso
n
fish,
and
in
cat
eg
ori
zi
ng
the pois
on and
non
-
poiso
n fis
h
int
o
it
s c
orre
ct
f
am
i
ly
.
In
Alsm
adi
et
al
.
[
18
]
,
the
extracte
d
fea
tures
f
r
om
co
lor
te
xture
m
easur
em
ents
was
us
e
d
i
n
com
bin
at
ion
.
I
n
par
ti
cula
r,
t
he
aut
hors
em
plo
ye
d
gray
le
ve
l
co
-
occ
urren
c
e
m
at
rix
(G
LC
M)
f
or
the
pro
du
ct
i
on
of
a
prot
otype
fo
r
cl
assify
in
g
fish.
I
n
te
sti
ng
the
pr
oto
t
ype,
the
auth
ors
us
e
d
20
dif
fer
e
nt
fish
fa
m
ilies
con
ta
ini
ng
di
fferent
num
ber
of
fish
ty
pes
e
ach.
Alto
gethe
r,
the
re
wer
e
610
dif
fere
nt
fish
im
ages
wer
e
use
d.
These
im
ages
wer
e
cl
assed
i
nto
tw
o
datas
et
s.
On
e
datas
et
con
ta
ins
50
0
trai
ning
i
m
a
ges
w
hile
the
oth
e
r
con
ta
in
s
110
t
est
ing
im
ages.
Neural
netw
ork
c
onnected
t
o
the
bac
k
-
pro
pag
at
io
n
al
gori
thm
was
us
ed
in
this
work,
a
nd
the
auth
ors
ac
hie
ved
84%
accu
racy
on
t
he
te
st
dataset
.
With
the
a
pp
li
cat
ion
of
t
he
pro
po
s
e
d
cl
assifi
er,
the
a
uthors
wer
e
a
bl
e
to
cat
ego
riz
e
the
fish
into
i
ts
cl
us
te
r,
cat
egorize
the
cl
ust
ered
fis
h
into
po
is
on
or no
n
-
po
is
on
on
e
s,
a
nd fu
rth
er categ
or
iz
e
the
po
is
on a
nd
non
-
poiso
n fis
h
int
o
it
s m
at
c
hing
fam
i
ly
.
In
Alsm
adi
et
al
.
[17]
a
fish
cl
assifi
cat
ion
prototype
was
prese
nted
.
This
prot
otype
com
bin
es
featur
e
s e
xtracted fr
om
co
lor
s
ign
at
ure m
easur
em
ents. H
ere
, hist
ogram
o
f
c
olor, R
GB co
l
or s
pace,
i
n
ad
di
ti
on
to
GLCM
wa
s
us
e
d.
Acc
ordi
ngly
,
the
aut
hor
s
us
e
d
a
cr
op
out
of
c
olor
sig
natu
re
f
or
dif
fe
rin
g
fam
ilies
of
fish.
Th
e
po
is
on
an
d
non
-
poiso
n
fi
sh
we
re
al
l
use
d
f
or
the
e
xtr
act
ed
col
or
si
gnat
ure
feat
ur
es
.
The
aut
hors
use
d
20
diff
e
re
nt
fam
ilies
of
fis
h
i
n
t
est
ing
t
he
pro
po
s
ed
syst
em
.
In
each
fam
il
y,
the
re
wer
e
di
ff
e
ren
t
ty
pe
s
of
fis
h.
Ov
e
rall
,
there
wer
e
610
dif
fe
ren
t
im
ages
of
fish
in
the
sa
m
ple
u
sed
an
d
these
i
m
ages
wer
e
div
ide
d
into
tw
o
dataset
s:
400
trai
ning
im
ages
in
one
dataset
an
d
21
0
te
sti
ng
im
ages
in
the
oth
er
datas
et
.
The
a
utho
r
s
us
e
d
neural
ne
tw
ork
ass
ociat
ed
with
the
bac
k
-
pro
pa
gation
a
lgorit
hm
and
i
t
yi
el
ded
84%
overall
acc
uracy
on
the
e
m
plo
ye
d
te
st
dataset
.
Th
is
stud
y
pro
ve
s
that
the
cl
assifi
er
it
pr
opose
d
al
lows
fis
h
to
be
cl
assed
in
to
it
s
corres
pondin
g
cl
us
te
r.
F
ur
t
he
rm
or
e,
the
cl
ust
ered
fis
h
co
uld
f
urt
her
be
cl
assed
into
po
is
on
or
non
-
pois
on
on
e
s,
a
nd int
o
i
ts ap
propriat
e f
a
m
ily.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
6
,
Dece
m
ber
2
01
9
:
5192
-
5204
5194
Ba
daw
i
an
d
A
ls
m
adi
[37
]
de
m
on
strat
ed
a
ge
ner
ic
fis
h
cl
assifi
cat
ion
syst
e
m
.
This
syst
em
e
m
plo
ys
a
hybri
d
m
et
aheurist
ic
al
gor
it
h
m
,
wh
ic
h
i
s
ge
netic
al
go
rithm
with
it
erated
l
ocal
se
arch,
in
a
ddit
ion
to
back
-
prop
a
gati
on
al
gorithm
(G
A
ILS
-
BPC
)
.
Cl
assifi
cat
ion
was
e
xecu
te
d
base
d
on
a
c
om
bin
at
ion
be
tween
sign
ific
a
nt
ext
racted
feat
ur
e
s
with
the
use
of
anc
hor
points
an
d
te
xt
ur
e
a
nd
sta
ti
sti
cal
m
easur
e
m
ents;
this
al
lows
fish
i
m
ages
to
be
c
la
ssed
int
o
da
nger
ous
a
nd
no
n
-
da
ng
e
rous
fa
m
ilies,
wh
il
e
da
ng
e
r
ou
s
fam
ilies
of
fish
can
be
cl
a
ssed
into
P
red
a
tory
and
P
oison
fish
fam
il
y.
Fu
rt
her
m
or
e,
f
a
m
ilies
of
non
-
dange
rous
fis
h
can
be
cl
assed
int
o
ga
rd
e
n
a
nd
foo
d
fish
fam
ily.
This
stu
dy
use
d
24
fis
h
f
a
m
ilies
in
te
s
ti
ng
th
e
proto
ty
pe,
and
the
fam
i
li
e
s
of
fis
h
em
pl
oyed
in
this
w
or
k
co
ntain
di
f
fer
e
nt
nu
m
ber
of
s
pecies
of
fish
each
.
The
re
wer
e
two
phases
,
na
m
el
y
trai
nin
g
ph
a
se
an
d
te
stin
g
ph
as
e.
Trai
ning
phase
em
plo
ye
d
220
fis
h
i
m
ages
wh
il
e
te
sti
ng
ph
a
se
em
plo
yed
100
fish
i
m
ages.
He
nce,
ov
e
rall
,
this
stud
y
us
e
d
32
0
dif
fer
e
nt
fis
h
i
m
a
ges.
The
ov
e
rall
accuracy
of r
ec
ogniti
on r
at
e
obta
ined
b
y t
he a
uthors w
he
n u
sing t
he pr
oto
t
ype w
it
h G
A
IL
S
-
BPC
was 8
0.5%.
In
Alsm
adi
et
al
.
[
9]
,
a
pro
toty
pe
that
use
s
Hy
br
id
Me
m
et
ic
Algo
rith
m
(G
eneti
c
Al
gorithm
and
Gr
eat
Del
ug
e
Local
Searc
h)
a
nd
Ba
c
k
-
Pr
opa
gatio
n
Cl
assifi
er
(HGAG
D
-
BPC
)
and
Ba
c
k
-
Pro
pag
at
i
on
Cl
assifi
er
(BP
C)
f
or
cl
assify
ing
fish
wa
s
dem
on
strat
ed
.
Her
e
,
the
aut
hors
pe
rfo
rm
ed
the
cl
assifi
cat
ion
ta
sk
fo
ll
owin
g
the
com
bin
at
ion
be
tween
the
e
xtracted
feat
ur
es
ob
ta
ine
d
f
ro
m
Po
te
ntial
Loc
al
Geo
m
et
ric
F
eat
ur
es
(P
L
GF
)
an
d
Sh
a
pe
Feat
ur
e
s.
I
n
t
his
w
ork
,
H
G
AGD
-
BPC
achieve
d
bette
r
resu
lt
s
at
96%.
Som
ehow
,
com
par
ed
to
B
PC,
H
GAGD
-
BPC
had
high
com
pu
ta
ti
on
al
tim
e,
bu
t
BPC
sh
owe
d
lowe
r
per
ce
ntage
at
86%.
These
cl
assifi
e
rs
al
low
fis
h
to
be
gro
uped
i
nto
it
s
cor
re
sp
onding
cl
us
te
r
.
Then
,
the
cl
ust
ered
fis
h
co
ul
d
be
furthe
r
cl
asse
d i
nto
po
is
on
or
non
-
poiso
n fis
h,
a
nd int
o
it
s ri
gh
tf
ul f
am
il
y.
Saye
d
et
al
.
[38]
dem
on
strat
ed
an
aut
om
at
e
d
fish
s
pecies
identific
at
io
n
syst
e
m
fo
ll
ow
i
ng
a
m
od
ifie
d
crow
searc
h
optim
iz
at
ion
al
go
rithm
in
their
work.
Me
dia
n
filt
ering
was
us
e
d
to
ge
ne
ra
te
s
m
oo
th
im
a
ge
an
d
el
i
m
inate
the
no
ise
.
T
his
re
du
ce
s
the
div
e
rsity
of
i
ntensi
ti
es
between
the
neig
hbors.
A
k
-
m
ean
cl
us
te
ring
al
gorithm
was
then
use
d
to
s
egm
ent
the
fish
i
m
age
into
m
ul
ti
ple
seg
m
ents,
w
hich
broug
ht
forth
the
feature
extracti
on
pro
cess
base
d
on
sh
a
pe
a
nd
te
xture
f
or
the
t
ask
of
cl
assifi
cat
ion
.
The
da
ta
dim
ension
al
it
y
of
the ex
tract
ed f
eat
ur
es w
a
s d
e
creased
i
n
this work, and for
t
he
pur
pose, a new
m
od
ifie
d
bi
nar
y ve
rsion of
crow
search
alg
or
it
hm
w
as app
li
ed. The classi
ficat
ion
tas
k
in
volv
ed
the
u
se
of s
upport
vecto
r m
achine and
de
ci
sio
n
trees
an
d
th
e
sp
eci
es
of
fish
wer
e
cl
assed
f
ollo
wing
ei
ther
their
cl
ass
(
e.g
.
,
Acti
nopt
erygii
and
C
hondric
ht
hyes)
or
their
order.
The
a
ut
hors
w
orke
d
on
270
im
ages
in
dif
fer
e
nt
sp
eci
es,
cl
asse
s
and
orders
on
t
he
pro
posed
sys
tem
and
report
ed
the
superi
or
it
y
of
the
pro
posed
syst
em
co
m
par
ed
t
o
oth
e
r
adv
a
nce
d
al
gor
it
h
m
s.
The
aut
hor
s
al
so
report
ed
that
the
ov
erall
fish
sp
eci
es
syst
e
m
of
identific
at
ion
yi
el
ds
10
fo
l
ds
on
a
ver
a
ge,
96%
accu
r
acy
of
cl
assifi
cat
ion
f
or
cl
as
sific
at
ion
base
d
cl
ass
a
nd
74
%
for
cl
assifi
c
at
ion
base
d on fis
h o
rd
e
r.
3.
RESEA
R
CH
MARER
IALS
AND MET
H
ODS
3.1.
Dataset
A
total
of
40
0
fis
h
im
ages
wer
e
us
e
d
i
n
this
stu
dy.
T
he
im
ages
we
re
f
ro
m
Globa
l
Inform
at
ion
Syst
e
m
(G
IS)
on
Fishes
(f
is
h
-
base
),
a
nd
t
he
y
wer
e
obta
ine
d
in
Se
ptem
ber
,
20
13.
T
he
im
ages
inclu
de
d
real
-
world
im
ages
of
fish
ca
pture
d
in
"c
ontrolle
d",
"
ou
t
-
of
-
t
he
-
wat
er"
a
nd
"i
n
-
sit
u"
set
ti
ng
s
.
For
the
"c
ontrolle
d"
i
m
ages,
they
e
nco
m
pass
fis
h
sp
eci
m
ens
in
posit
ion
w
her
e
t
heir
fins
a
re
s
pread
a
nd
the
i
m
ages
we
re
ca
ptured
in
c
on
sist
e
nt
ba
ckgr
ound
with
il
lum
inati
on
that
is
c
ontr
ol
le
d.
T
he
datas
et
em
plo
ye
d
in
t
his
w
ork
co
n
ta
in
s
three cate
gorie
s; eac
h of t
hem
contai
ns
dif
fere
nt num
ber
of
fish fam
ilies as f
ollow
i
ng
:
a.
Dange
rous
Fi
sh
Fam
i
li
es:
Ca
rch
a
rh
i
nu
s
Leucas,
Ca
rc
ha
rod
on
Ca
rc
ha
ria,
Atracto
ste
us
Sp
at
ula
and
Hydrocy
nu
s
G
oliat
h.
b.
Po
iso
n Fi
sh Fa
m
ilies:
Red Sn
app
e
r, Tri
gg
e
r, Porc
up
i
ne
a
nd Thorn.
c.
Garde
n
an
d
Food
Fam
il
ie
s:
Acestrorh
y
nch
i
da
e,
Acro
po
m
aat
idae,
Albuli
da
e,
Anom
al
op
id
ae,
Ca
esi
on
ida
e,
Dr
e
pa
nid
ae,
I
sti
ophoridae,
Lei
ognathi
dae,
Me
galop
i
dae,
Plat
yc
eph
al
id
ae,
Pr
ia
cant
hid
ae,
Sc
om
br
id
ae,
Sigan
i
dae, Si
ll
aginidae
, S
t
rom
at
ei
dae an
d
Triaca
nt
hid
ae.
3.2.
Te
xt
ure
fe
atu
res calcul
at
in
g
usin
g
GLC
M
The
com
pu
ta
ti
on
of
te
xture
f
eat
ur
es
acco
r
din
g
to
G
LCM
involve
s
seve
n
ste
ps
.
Im
age
acqu
isi
ti
on
is
the
fir
st
an
d
m
os
t
im
po
rtant
s
te
p.
T
he
sec
on
d
ste
p
in
volves
the
tra
nsfo
rm
a
ti
on
of
di
gital
fish
im
age
into
gr
ay
scal
e
i
m
age
w
her
e
by
dig
it
al
fish
im
ages
ar
e
the
on
es
co
m
m
on
ly
dealt
with
in
this
re
search
.
No
ta
bly,
fis
h
diff
e
rs
in
te
rm
s
of
s
ha
pe,
an
d
there
f
or
e,
a
crop
facti
on
is
e
m
plo
ye
d
in
order
t
o
m
anu
al
ly
deter
m
ine
the
fish
sh
a
pe
so
t
hat
error
c
ould
be
era
dicat
ed
.
T
hen
i
n
the
t
hir
d
ste
p,
a
cr
op
ou
t
of
t
he
pa
tt
ern
of
interest
,
in
this
case, it
is the sh
ape
of the
fish
s
hap
e
, is separat
ed
f
r
om
the
b
ack
gro
und.
Using this a
ppr
oach, h
i
gh qual
it
y fish
recog
niti
on
ca
n be att
ai
ne
d.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Ro
bu
st f
e
atu
re
ext
ra
ct
ion met
hods
f
or
ge
ner
al fi
sh
cl
as
sif
ic
ation
(
Muta
se
m
Als
madi
)
5195
In
the
f
ourth
s
te
p,
the
ca
ptur
ed
cr
op
is
filt
rated
from
fish
pix
el
s.
F
or
th
e
purpose
,
a
5×
5
Ga
us
sia
n
Fil
te
r
is
us
e
d.
The
filt
rated
c
rop
im
age
is
ne
xt
sp
li
t
into
bl
ock
s
with
t
he
siz
e
of
4×
4;
t
his
is
the
fifth
ste
p.
This
is
f
ollowe
d
by
c
om
pu
ta
ti
on
of
im
age
qu
al
it
y
feature
s
for
each
bl
ock
i
n
acco
r
da
nce
with
t
he
GLCM
(this
is
t
he
sixt
h
ste
p)
.
Last
ly
,
in
the
se
ven
t
h
ste
p,
t
he
at
ta
in
ed
featur
e
s
are
store
d.
I
n
br
ie
f
,
the
c
om
pu
ta
ti
on
of
fish’s text
ur
e
fe
at
ur
es
by
GL
CM
is as outl
in
ed belo
w:
Step
1:
di
gital
f
ish im
age acquirem
ent.
Step
2:
co
nver
sion o
f
im
age to
gr
ay
scale
im
age.
Step
3:
m
anu
a
l
determ
inati
o
n
of
fis
h
sh
a
pe
us
in
g
a
cr
op
facti
on
to
e
ra
dicat
e
error
s
,
wh
il
e
the
cr
op
ou
t
of
the p
at
te
r
n o
f
i
nterest
(f
is
h
s
ha
pe) is s
ub
tract
ed fr
om
the b
a
ckgr
ound.
Step
4: f
il
trat
io
n of capt
ur
e
d
c
ro
p o
ut of
fish pixels
with a
5×5
Gau
s
sia
n Fi
lt
er.
Step
5: d
i
visio
n of t
he fil
trat
ed
c
rop
im
age in
to
4×
4 bloc
ks.
Step
6:
Ca
lc
ul
at
ion
of
im
age
qu
al
it
y
featu
re
s
for
al
l
blo
c
ks
us
in
g
GLCM.
Com
pu
ta
ti
on
on
24
disc
rete
i
m
age
qu
al
it
y feat
ur
e
s f
or
th
e “fil
trat
ed
cro
p”
f
ollo
wing the four d
irect
io
ns
(hor
iz
on
ta
ll
y (9
0
0
), ver
ti
cal
ly
(
0
0
),
and
t
wo
dia
gonally
(
45
0
a
nd
135
0
)),
s
pe
ci
fical
ly
,
Averag
e
or
m
ea
n
val
ue,
Stan
dard
Dev
ia
ti
on,
Con
tra
st,
Dissim
il
arity, H
om
og
e
neity
, and
Energy.
Step
7: sto
rag
e
of the
att
ai
ned featu
res
i
n
the
d
at
aba
se.
3.3.
Shape
fe
ature
s ca
lc
ul
at
in
g
u
sing anch
or
p
oint
s
loc
at
i
on
detecti
on
In
m
easur
in
g
fish
s
hap
e
,
se
veral
anchor
point
s
nee
d
to
be
i
de
ntifie
d,
a
s
s
hown
i
n
Fi
gure
1
.
Detect
io
n
of
a
nc
hor
points
ha
s
in
dee
d
bee
n
the
inte
rest
of
m
any
since
past
se
ve
ral
ye
ars
par
t
ic
ularly
am
on
g
th
os
e
work
i
ng in
patte
rn
r
e
co
gnit
ion. Po
ints
’
detect
ion
is u
se
d
to fi
nd
a sig
nifica
nt set o
f
points
w
hich wil
l faci
li
ta
te
the
at
ta
inm
ent
of
anc
hor
m
easur
em
e
nts
f
or
patte
rn
s
of
i
nterest,
in
this
ca
s
e,
fis
h
obj
ect
.
I
n
this
stu
dy,
a
ncho
r
po
i
nt
detect
ion
is
us
ed
for
th
e
determ
inati
o
n
of
23
la
be
le
d
points
w
hich
will
facil
i
ta
te
the
determ
inatio
n
of
the
locat
ion
of
each
featu
re
in
fish
i
m
age
reco
gnit
ion.
This
is
fo
ll
ow
e
d
by
the
com
pu
ta
ti
on
of
the
ge
om
et
rical
featur
e
s
with
t
he
us
e
of
the
de
te
rm
ined
a
nc
hor
po
i
nts;
this
is
f
or
cl
assi
fyi
ng
the
fish.
W
hen
the
e
ntire
ancho
r
po
i
nts
over
th
e
fish
obj
ect
hav
e
been
det
ect
ed,
distanc
e
and
an
gle
m
easur
em
ents
are
us
ed
to
extract
the sig
nifica
nt
featur
e
s.
Figure
1
.
The
locati
ons
of
t
he a
ncho
r po
i
nt m
easur
em
ents
Sh
a
pe
m
easur
e
m
ents
inv
ol
ve
the
com
pu
ta
ti
on
of
the
ed
ge
an
d
distanc
e
m
easur
em
ents
of
the
fish
obj
ect
as
well
as
the
determ
inati
on
of
th
e
sign
ific
a
nt
ide
ntica
l
and
differin
g
pa
rts
f
or
each
fam
il
y
of
fish.
Fu
rt
her
m
or
e,
cl
assifi
cat
ion
of
gr
eat
er
acc
ur
acy
w
ould
r
esult
if
the
procedu
re
of
cl
assifi
cat
ion
e
m
plo
ys
the
m
easur
em
e
nts
of
vect
or
'
s
ang
le
s
em
plo
yi
ng
t
hr
ee
point
s
f
or
each
an
gl
e
of
the
cau
dal
fin
an
gle
a
nd
fish
head
a
ngle
[36
]
.
I
n
ad
diti
on,
ap
plica
ti
on
of
distance
m
easur
em
ents
al
lows
t
he
deter
m
inati
on
as
w
e
ll
as
extracti
on
of s
ever
al
featu
res i
nclud
i
ng the
ra
diu
s
of
fish
ey
e an
d
le
ngth
of p
ect
oral
f
i
n.
3.4.
Dist
an
ce
and
angle me
as
uri
ng
tools
Dista
nce
an
d
ang
le
m
easur
em
ents
are
us
ed
to
com
pu
te
sh
ape
feat
ur
e
s.
In
detai
l,
the
distan
c
e
m
easur
em
ents
com
pr
ise
the
di
sta
nce
bet
wee
n
21
anc
hor
points
nam
el
y:
P1
,
P
2,
P
3,
P
4,
P
5,
P6,
P
7,
P
8,
P9,
P10,
P
11,
P12,
P1
3,
P14,
P
15,
P
18,
P19,
P
20,
P21,
P
22,
and
P
23.
Ta
ble
1
.
Twe
nty
two
e
xtracted
f
eat
ur
es
from
the
deter
m
ined
anc
hor
po
i
nts.
Fi
gure
1
pr
ov
i
des
the
detai
ls.
Me
an
w
hile
,
the
a
ng
le
betwee
n
th
ree
ancho
r
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
6
,
Dece
m
ber
2
01
9
:
5192
-
5204
5196
po
i
nts
over
t
he
fish
ob
j
ect
m
a
kes
up
the
a
ngle
m
easur
em
ent
s
su
c
h
as:
ey
e
-
end
m
ou
th
a
ng
le
,
Ca
ud
al
fi
n
ang
le
and
fish
hea
d
ang
le
.
Fig
ur
e
1
pro
vi
des
the
detai
ls.
The
a
nc
hor
po
i
nts
an
d
the
featur
e
s
el
ect
ed
are
co
m
pu
te
d
util
iz
in
g
dista
nc
e
an
d
an
gle
m
easur
em
ents.
These
a
nc
hor
po
i
nts
a
nd
fe
at
ur
e
a
re
prese
nted
i
n
Ta
ble
1
a
nd
Table
2,
a
nd th
ei
r
detai
ls are
provid
e
d
i
n
the
ens
uing s
ubsec
ti
on
.
Table
1.
T
we
nt
y t
wo
featu
res c
al
culat
ed usin
g
the
d
et
e
rm
in
ed
distance m
easur
em
ent
Distan
ce.
#
An
ch
o
r
Po
in
ts Featu
res
Distan
ce.
#
An
ch
o
r
Po
in
ts Featu
res
D1
(P1 an
d
P2)
D1
2
(P13
and
P14
)
D2
(P11
and
P19
)
D1
3
(P2 an
d
P15
)
D3
(P1 an
d
P3)
D1
4
(P19
and
P10
)
D4
(P3 an
d
P5)
D1
5
(P20
and
P21
)
D5
(P5 an
d
P6)
D1
6
(P22
and
P23
)
D6
(P7 an
d
P8)
D1
7
(P5
an
d
P3)
D7
(P11
and
P12
)
D1
8
(P13
and
P2)
D8
(P12
and
P22
)
D1
9
(P4 an
d
P3)
D9
(P18
and
P13
)
D2
0
(P4 an
d
P11
)
D1
0
(P2 an
d
P5)
D2
1
(P5 an
d
P4)
D1
1
(P5 an
d
P20
)
D2
2
(P19
and
P4)
Dista
nce
m
easur
em
ents
are
integ
ral
in
patte
rn
rec
ogniti
on
par
ti
cula
rly
in
the
ta
sk
of
r
obus
t
fe
at
ur
e
s
extracti
on
to
i
m
pr
ov
e
the
ac
cur
acy
of
cl
ass
ific
at
ion
.
I
n
al
gebraic
ge
om
e
try
,
the
com
puta
ti
on
of
the
distance
‘D’
betwee
n
th
e points C=
(a
1,
b1)
and E
=
(a
2,
b2)
is e
xpr
essed
i
n form
ula 1
.
=
√
(
∆
)
2
+
(
∆
)
2
=
√
(
2
−
1
)
2
+
(
2
−
1
)
2
(1)
As
s
how
n
in
Fi
gure
1
,
t
her
e
ar
e
23
a
nchor
po
ints,
an
d
t
hese
po
i
nts
de
note
t
he
le
ngth
bet
w
een
a
ncho
r
po
i
nts
as
Table
1
is
showi
ng.
Hen
ce
,
with
th
e
app
li
cat
ion
of
the
distan
ce
m
easur
em
ent
fo
rm
ula,
a
tot
al
of
15
featur
e
s
we
re
a
tt
ai
ned
.
T
he
an
gle
betwee
n
t
w
o
vect
or
s
,
s
us
pe
nd
e
d
by
one
po
i
nt,
is
du
bb
e
d
the
s
hortest
ang
le
at
wh
ic
h
one
of
the
vecto
rs
ha
s
to
be
ro
ta
te
d
to
the
posit
ion
that
is
co
-
di
recti
on
al
with
ano
the
r
vect
or
[39]
.
The fo
rm
ula bel
ow
c
om
pu
te
s
the a
ng
le
θ
between t
w
o vectors:
cos
=
|
̅
.
̅
|
|
̅
|
.
|
̅
|
(2)
The
res
ultant
ei
gh
t
featu
res
that
com
pu
te
d
with
the
a
ngle
m
easur
em
ents
acco
rd
i
ng
the
anchor
points
highli
gh
te
d
i
n Fi
gure
1
a
re sh
own
i
n
Ta
ble
2.
Table
2
.
Sixtee
n feat
ur
e
s that
wer
e
calc
ulate
d usin
g
t
he dete
rm
ined
anc
hor po
i
nts
An
g
le.
#
An
ch
o
r
Po
in
ts Featu
res
An
g
le.
#
An
ch
o
r
Po
in
ts Featu
res
A1
(P15
,
P9
and
P16
)
A9
(P16
,
P4
and
P17
)
A2
(P9, P4 an
d
P10
)
A1
0
(P9, P15
and
P10
)
A3
(P9, P17
and
P16
)
A1
1
(P1,
P1
6
and
P17
)
A4
(P21
,
P1
6
and
P17
)
A1
2
(P20
,
P1
6
and
P17
)
A5
(P16
,
P1
5
and
P17
)
A1
3
(P23
,
P1
3
and
P1)
A6
(P20
,
P4
and
P2
1
)
A1
4
(P23
,
P1
3
and
P11
5
)
A7
(P5, P3 an
d
P16
)
A1
5
(P9, P16
and
P10
)
A8
(P5, P3 an
d
P17
)
A1
6
(P15
,
P1
0
and
P17
)
3.5.
St
ati
stic
al
me
as
ureme
nt
s
Stat
ist
ic
al
m
ea
su
rem
ents
were
carrie
d
out
util
iz
ing
the
f
eat
ur
es
e
xtract
ed
f
ro
m
i
m
ages
of
fish
belo
ng
i
ng
to
24
fish
fam
il
ie
s.
This
is
will
de
te
rm
ine
an
d
as
certai
n
th
e
sig
nificant
featu
re
s
w
hich
will
as
sist
i
n
the
at
ta
inm
ent
of
rec
ogniti
on
of
acc
ur
acy
wh
il
e
al
so
rec
ognizin
g
the
fi
sh
im
ages
into
ei
ther
da
nger
ou
s
or
non
-
da
nger
ous
fa
m
ily.
Acco
r
dingly
,
the
ou
t
com
es
of
cor
re
la
ti
on
gro
unde
d
upon
the
fea
tures
extr
act
ed
with
the
ap
plica
ti
on
of
m
easur
em
e
nts
of
a
nc
hor
po
i
nts
ca
n
be
viewe
d
in
F
ig
ur
e
2
.
T
he
sta
ti
sti
cal
ou
tc
ome
s
from
the
extracte
d
featu
res
de
m
on
strat
e
the
div
e
rse
c
orr
el
at
ion
val
ue
betwee
n
ce
r
ta
in
extracte
d
featu
res
par
ti
cula
rly
he
ad,
ey
e
an
d
ca
ud
al
an
gles
of
the
fis
h.
T
hese
featu
res
can
in
fact
be
deem
ed
as
good
featu
res
for
i
m
pr
ovin
g
the
accuracy
of
c
la
ssific
at
ion
,
wh
ic
h,
a
re
of
value
i
n
the
c
on
te
xt
of
this
work.
As
a
n
e
xam
ple,
the
da
nger
ous
fish
fam
il
ie
s
sh
ow
ne
gative
co
rr
el
at
io
n
va
lue
be
twee
n
t
he
hea
d
a
nd
e
ye
ang
le
s
.
Des
cribe
d
diff
e
re
ntly
,
for
fish
belo
nging
t
o
these
f
a
m
ilies;
the
head
an
gle
inc
reases
as
the
ey
e
ang
le
dec
reases
.
Fu
rt
her
m
or
e,
the
value
of
c
orrelat
ion
bet
ween
the
ca
udal
an
d
ey
e
a
ng
le
f
or
fish
in
the
se
fam
il
i
es
al
s
o
app
ea
rs ne
gative.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Ro
bu
st f
e
atu
re
ext
ra
ct
ion met
hods
f
or
ge
ner
al fi
sh
cl
as
sif
ic
ation
(
Muta
se
m
Als
madi
)
5197
On
t
he
oth
e
r
hand,
fis
h
bel
onging
t
o
cert
ai
n
no
n
-
dange
rous
fish
fam
i
li
es
app
ea
r
to
ha
ve
posit
ive
correla
t
ion
value
bet
ween
t
he
head
a
nd
ey
e
ang
le
.
T
his
m
eans
that
for
fish
bel
onging
to
these
fa
m
ilies;
the h
ead a
ng
le
increases as th
e ey
e ang
le
incr
eases. Ad
diti
onal
ly
, th
e v
al
ue
s o
f
co
rr
el
at
io
n
betwee
n
the
caud
al
and
ey
e
an
gles
al
so
app
ea
r
to
be
posit
ive.
As
c
an
be
c
onstrue
d
from
th
e
find
i
ng
s
,
the
resu
lt
ant
valu
es
of
correla
ti
on
of
the
e
xtracte
d
feature
s
diff
e
r
based
on
fam
ily.
The
co
rr
el
at
io
n
values
will
increas
e
the d
ist
incti
on
betwee
n
the
se
fam
i
li
es o
f
fis
h (
po
is
on, non
-
poiso
n, wild
and
food
fish fam
i
li
es).
Figure
2
.
Co
rr
e
la
ti
on
res
ults
ba
sed o
n
the
f
ea
tures
t
hat w
e
re
extracte
d u
si
ng
an
c
hor p
oin
ts
m
easur
em
ents
4.
CLASSIFIE
R
ARCHIT
EC
TURE
4.1.
Genetic
algori
th
m
A
ge
netic
al
gorithm
(G
A)
e
nc
om
passes
a
he
ur
ist
ic
ap
proa
ch
an
d
it
is
groun
ded
upon
popula
ti
on
wh
ic
h
si
m
ulate
s
the
natur
al
se
le
ct
ion
proce
dure
.
GA
create
s
valuab
le
no
ve
l
so
luti
on
s
to
chall
eng
i
ng
prob
lem
us
in
g
sam
ple
so
luti
ons
in
a
popula
ti
on.
T
he
re
are
t
hr
ee
ke
y
ph
ase
s
in
G
A
w
he
reb
y
t
he
first
ph
ase
co
ncerns
the
sel
ect
ion
t
echn
i
qu
e
to
se
le
ct
two
so
l
ution
s
f
ro
m
the
popula
ti
on
a
nd
then
rec
om
bi
ne
them
.
Re
le
van
tl
y,
sever
al
te
c
hn
i
qu
e
s
of
sel
ect
ion
wer
e
pro
po
s
ed
i
n
Mi
chalewic
z
[
16]
inclu
ding
T
our
nam
ent
Sele
ct
ion
,
Tr
uncat
io
n
Se
le
ct
ion
an
d
Rou
le
tt
e
Wh
eel
Sele
ct
ion
.
T
he
second
ph
a
s
e
include
s
the
us
e
of
a
cr
osso
ver
op
e
rato
r
to
ca
rr
y
out
the
m
at
ing
proce
ss.
Cros
s
over
is
the
ge
netic
wa
y
to
discov
e
r
novel
so
l
utions,
i.e.,
so
luti
ons
with
bette
r
value
of
fitness,
within
the
s
earch
s
pac
e.
The
thir
d
ph
ase
involves
th
e
us
e
of
a
Mut
at
ion
op
e
rato
r
f
or
di
scov
e
rin
g
the
neig
hbor
s
ol
utions.
Muta
ti
on
operato
r
is
a
local
searc
h.
Durin
g
this
ph
ase
,
the
popula
ti
on
is
updated
.
F
urt
her
m
or
e,
the
gen
e
rati
on
of
s
olu
ti
ons
with
be
tt
er
value
of
f
it
ness
will
i
m
pr
ov
e
the quali
ty
o
f
s
earch
sp
ace
[1
2].
4.1.1.
Initiali
z
at
ion
The
pro
du
ct
io
n
of
weig
hts
f
or
fee
d
-
forwar
d
arti
fici
al
ne
ural
net
work
is
us
ua
ll
y
a
rand
om
pr
ocess
.
Nonetheless
,
the
re
pr
ese
ntati
on
of
c
hrom
os
om
e
gr
eat
ly
con
t
rib
utes
to
the
su
ccess
of
gen
e
ti
c
al
gorithm
.
In
this
rese
arc
h,
a
sim
ple
pr
esentat
ion
for
chrom
os
om
e
g
rou
nd
e
d
up
on
bin
a
ry
represe
ntati
on
is
ap
plied
f
or
each
so
l
ution.
Her
e
,
the
chro
m
os
o
m
e
den
ot
es
sever
al
wei
gh
ts
(
real
valu
es)
that
are
rando
m
ly
ob
ta
ined
from
the
m
at
rices
weigh
t.
These
w
ei
ghts
are
in
fr
act
io
n
num
ber
s
f
or
m
,
the
se
fr
act
i
on
nu
m
ber
s
are
sig
ni
fied
in
the g
e
nes
w
it
hi
n
the
chr
om
os
om
e in b
ina
ry
strings.
4.1.2.
Roulette
w
hee
l sel
ection
Rou
le
tt
e
W
he
el
Sele
ct
ion
is
the
m
os
t
com
m
on
ly
us
ed
m
et
ho
d.
Ro
ul
et
te
W
heel
S
el
ect
ion
was
the
creati
o
n
of
Ba
ker
(
1987)
an
d
it
is
known
as
the
si
m
plest
sche
m
a
of
sel
ection
.
As
ex
pla
ined
in
Leu
ng
et
al
.
(
2003
),
in
t
he
ap
plica
ti
on
Ro
ul
et
te
Wh
eel
Sel
ect
ion
,
tw
o
c
hrom
os
om
es
fr
om
the
popu
la
ti
on
a
re
chosen
t
o
go
thr
ough
ge
neti
c
operati
ons
f
or
re
producti
on
us
in
g
the
spi
nn
in
g
t
he
r
ou
le
tt
e
wh
eel
m
et
hod.
Pare
nts
with
hi
gh
po
te
ntial
argua
bly
will
gen
erate
bette
r
offsprin
g,
w
hich
ha
ve
bette
r
chance
of
s
urvival
.
The
c
hrom
os
om
e w
it
h
hi
gher
v
al
ue
of
fitnes
s is li
kely
to ha
ve greater
op
port
un
it
y i
n bei
ng chose
n
as
parent
.
Durin
g
the
sel
ect
ion
pr
ocess,
stochastic
sel
ect
ion
is
m
ade
from
on
e
ge
ne
rati
on,
an
d
th
e
sel
ect
ion
beco
m
es
the
fo
un
dation
for
the
ens
uing
ge
ner
at
io
n.
T
he
ru
le
of
thu
m
b
is
that
the
fitt
e
st
on
es
ha
ve
great
e
r
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
6
,
Dece
m
ber
2
01
9
:
5192
-
5204
5198
su
r
viv
al
oppor
tun
it
y
as
oppo
sed
to
t
he
wea
ker
on
es,
j
ust
li
ke
the
real
li
f
e
sit
uation
in
na
ture.
T
he
fitt
er
ones
will
then
form
the
m
at
ing
pool
fo
r
the
f
ollo
w
ing
ge
ne
rati
on.
Howev
e
r,
the
weak
e
r
ones
st
il
l
hav
e
a
chance
a
s
they
m
a
y ca
rr
y
v
al
ua
ble
ge
netic
co
di
ng for t
he
n
e
xt g
e
n
e
rati
on
s
.
4.1.3.
Cro
s
sover
Cros
s
over is the
cor
e of g
en
et
ic
algo
rithm
, a
nd
it
d
en
otes a m
at
ing
p
r
ocess
. Th
e proces
s of cr
osso
ve
r
aim
s
at
discoverin
g
new
s
olut
ion
s
withi
n
the
search
sp
ace
.
In
a
custom
ary
cro
ssove
r
op
erator,
the
pair
ing
of
ind
ivi
du
al
s
of
the
popula
ti
on
is
done
i
n
a
r
andom
m
ann
er
.
T
wo
m
at
ing
chrom
os
om
es
are
c
ut
on
e
ti
m
e
at
the
res
pecti
ve
po
i
nts
an
d
t
he
sect
ion
s f
ollo
w
ing
t
he
cuts
a
re
exc
hange
d.
T
he
po
i
nt
of
cr
osso
ver
u
s
ually
can b
e
rand
om
l
y
sel
e
ct
ed.
Des
cribe
d
in
m
or
e
det
ai
l,
the
reco
m
bin
at
io
n
of
in
div
id
uals
gen
e
rates
ne
w
in
div
id
uals
by
ha
ving
t
he
inf
or
m
at
ion
from
two
or
m
or
e
par
e
nts
com
bin
ed.
Th
is
is
ge
ner
al
l
y
execu
te
d
t
hro
ugh
the
m
erg
ing
of
par
e
nts’
var
ia
ble
value
s.
Si
ngle
point
cr
os
s
ov
e
r
is
re
garde
d
as
the
sim
plest
cro
ss
over
m
et
hod.
As
desc
ribe
d
in
A
ns
ari
a
nd
Hou
(
1999)
,
the
outc
om
e
of
this
m
et
ho
d
is
on
e
or
tw
o
child
strin
g
th
rou
gh
the
ra
ndom
sel
ect
ion
of
c
ro
s
so
v
er
l
ocati
on
within
patte
r
n
stri
ng
le
ng
t
h.
In
the
e
xec
ution
of
sin
gl
e
point
cro
ss
over
,
a
po
int
is
ran
dom
ly
sel
ect
ed.
The
n,
the
parents
’
chrom
os
om
es
will
be
sever
ed
f
ro
m
that
po
int,
and
exch
a
nge
will
b
e m
ade to
t
he resulta
nt
sub
-
c
hrom
os
om
es.
4.1.4.
Muta
tion
In
ge
netic
al
gorithm
,
m
utatio
n
f
un
ct
io
ns
as
a
local
sea
rch
f
or
t
he
di
sco
ver
y
of
t
he
neig
hbor
so
luti
ons.
Mut
at
ion
c
om
pr
ise
s
a
ge
netic
op
e
rator
f
or
s
us
ta
inin
g
the
di
ver
s
it
y
of
gen
es
f
r
om
on
e
gen
e
ra
ti
on
of
a
popula
ti
on
of
ch
r
om
os
om
e
s
to
the
f
or
t
hc
om
ing
one.
F
ollow
i
ng
the
process
of
cr
osso
ver
,
the
res
ultant
ind
ivi
du
al
s
or
weig
hts
will
go
th
rou
gh
the
process
of
m
ut
at
ion
.
M
utati
on
usual
ly
re
places
the
gen
es
that
ar
e
lost
durin
g
the
evo
luti
on
a
ry
process
in
a
di
ff
e
ren
t
f
or
m
or
ge
ner
at
es
ne
w
ge
nes
that
wer
e
no
t
ex
pl
or
e
d
i
n
the
or
igi
nal
popula
ti
on.
With
m
utati
on
,
t
he
al
gorithm
co
uld
preve
nt
from
being
tra
pped
in
local
m
ini
m
a.
This
is
becau
se
the
po
pula
ti
on
of
chrom
os
ome
s
will
be
pr
ev
ented
f
ro
m
beco
m
ing
too
iden
ti
cal
to
on
e
an
oth
e
r
wh
ic
h
woul
d
c
ause
a
slo
w
do
wn
or
a
halt
to
the
e
voluti
on.
Acc
ordin
gly,
a
va
riable
is
c
ho
s
en
with
s
pe
ci
fied
pro
bab
il
it
y,
wh
il
e
it
s
value
is
al
te
red
us
i
ng
a
ran
dom
value.
A
non
-
unif
or
m
m
utatio
n
m
et
ho
d
is
ap
pl
ie
d
in
this
resea
rch.
This
m
et
ho
d
trans
form
s
on
e
of
t
he
ge
nes
belo
ng
i
ng
to
the
par
e
n
t
in
acco
rd
a
nc
e
wit
h
the no
n
-
unifo
r
m
p
ro
ba
bili
ty
d
ist
rib
ution.
4.1.5.
Fitness
fu
n
cti
on
The
fitn
ess
f
un
ct
ion
pr
ov
i
des
assessm
ent
on
the
pe
rfor
m
ance
of
eac
h
in
di
vidual,
an
d
it
is
base
d
on
pro
blem
.
Her
e,
the
perform
a
nce
of
eac
h
i
nd
i
vidual
is
com
pu
te
d
us
in
g
the
pe
r
centa
ge
Var
ia
nce
Ac
count
Functi
on
(VA
F)
bet
wee
n
tw
o
sig
nals.
As
highli
gh
te
d
in
Sh
et
a
(
2006)
,
the
com
pu
ta
ti
on
of
V
AF
f
ol
lows
the form
ula b
el
ow
:
V
=
1
−
v
a
riance
(
y
−
y
_
e
st
)
v
a
rian
c
e
(
y
)
∗
100%
(3)
Ba
sed
on
the
expressio
n
a
bove:
y
de
note
s
the
real
ou
t
pu
t;
y_est
denotes
the
proj
e
ct
ed
outpu
t
of
a
m
od
el
,
and
VAF
is
qu
a
ntif
ie
d
for
the
two
sign
al
s
to
ge
ne
rate
the
ou
t
put
V.
Fo
r
the
tw
o
sign
al
s
,
the
VAF
is
equ
i
valent
t
o
100%
,
a
nd
s
houl
d
they
hav
e
di
ff
e
ren
t
value
,
t
he
VAF
will
be
le
ss.
Wh
e
n
y
and
y_est
ha
ve
m
or
e
than
on
e
c
ol
um
n,
the
com
pu
ta
ti
on
of
V
A
F
is
m
ade
fo
r
each
col
um
n
in
y
and
y
_est
.
I
n
ge
ner
al
,
VAF
is
app
li
ed
in
t
he
ver
i
ficat
ion
of
the
m
od
el
’
s
accu
rateness
thr
ough
the
c
om
par
ison
of
the
real
outp
ut
wi
th
the m
od
el
’s proj
ect
e
d ou
t
put.
4.1.6.
St
oppi
n
g
cri
te
ri
on
In
G
As,
a
stoppin
g
crit
eri
on
is
ge
ner
al
l
y
sign
ifie
d
by
the
m
axi
m
u
m
nu
m
ber
of
gen
erati
on
s
.
Nonetheless
,
wh
e
n
i
deal
val
ue
of
fitness
(i
.e.
op
ti
m
al
weigh
t)
can
be
a
chieve
d,
a
sto
pp
i
ng
crit
eri
on
is
al
so
con
sidere
d
to
hav
e
bee
n
achi
eved.
H
ow
e
ve
r
,
in
this
resear
ch,
m
axi
m
u
m
nu
m
ber
of
ge
ne
rati
on
s
is
use
d
with
no consi
der
at
io
n on w
hethe
r
t
he
ideal
fitness
v
al
ue
is att
ai
ne
d or n
ot.
4.2.
Simul
at
e
d
anneali
ng al
go
ri
th
m
Si
m
ulati
ng
A
nneal
ing
(
SA)
i
ntr
oduce
d
by
Kir
kp
at
ric
k
[40]
has
great
er
r
obus
tne
ss
as
oppose
d
t
o
si
m
ple
local
search
owin
g
to
the
fact
that
it
al
so
acce
pt
s
w
or
se
s
olu
ti
on
s
with
s
om
e
pro
bab
il
it
y
[41
,
42]
.
SA
has
been
popula
rly
em
pl
oyed
in
the
so
l
ution
of
ha
r
d
com
bin
at
or
y
prob
le
m
s.
Ap
art
from
that,
the
us
e
of
SA
at
te
m
pts
to
pr
e
ve
nt
e
ntrapm
ent
in
local
optim
u
m
so
luti
on
th
r
ough
the
a
ll
ot
m
ent
of
pro
bab
il
it
ie
s
to
m
ov
e
s
t
hat
app
ea
r
to
be
deteri
or
at
in
g.
I
n
this
reg
a
rd,
SA
co
uld
acce
pt
so
luti
ons,
w
hich
,
as
oppose
d
to
pa
st
on
es
,
are
neither
bet
te
r
nor
m
uch
worse,
wh
ic
h
al
lows
escape
from
local
op
tim
u
m
and
discov
e
ry
of
the
global
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Ro
bu
st f
e
atu
re
ext
ra
ct
ion met
hods
f
or
ge
ner
al fi
sh
cl
as
sif
ic
ation
(
Muta
se
m
Als
madi
)
5199
on
e
[43
,
44]
.
The
a
uthors
i
n
[41]
pro
vid
e
d
a
gen
e
ric
SA
al
gorithm
fo
r
pro
blem
of
m
a
xim
iz
at
ion
.
SA
beg
i
ns
with
best
wei
gh
t
s
olu
ti
on
(
s)
picke
d
ou
t
fr
om
the
pool
of
popula
ti
on
within
the
gen
et
ic
al
gor
it
h
m
.
The
beg
i
nn
i
ng
tem
per
at
ur
e
is
at
1000
w
hen
the
searc
h
be
gi
ns
a
nd
t
he
fin
a
l
tem
per
at
ur
e
i
s
at
0
with
it
er
at
ion
nu
m
ber
,
#It
er
i
s
set
to
10000.
At
each
it
erati
on,
the
te
m
perat
ur
e
is
rand
om
ly
red
uce
d
th
rou
gh
t
he
creat
ion
of
a
frac
ti
on
num
ber
bet
ween
0
and
1.
Ne
xt,
a
neig
hbor
is
outl
ined
by
the
in
discrim
inate
creati
on
of
a
ra
ndom
fr
act
io
n
nu
m
be
r
be
twee
n
0
a
nd
1,
w
hich
is
inclu
ded
into
th
e
so
l
ution
val
ue
.
A
pply
ing
t
he
f
or
m
ula
expr
essed
in
Equ
at
i
on
(
3)
,
com
pu
ta
ti
on
is
m
ade
on
th
e
fitness
functi
on
value
of
the
new
nei
ghbor.
A
worse
so
lut
ion
is
receive
d
if
t
he
ind
isc
rim
inate
l
y
produce
d
num
ber
is
lower
than
e
-
δ/T,
where
δ
=
f
(s’)
-
f(s),
T=c
k
(ck
de
no
t
e
s
the
tem
per
at
ure
of
the
pr
e
s
ent
it
erati
on
nu
m
ber
)
.
The
n,
up
date
is
m
ade
on
the
pr
ese
nt
so
luti
on
(
s)
.
This
process
e
ns
ue
s
un
ti
l t
he hig
hest
nu
m
ber
of
it
erati
ons
is a
chie
ved, i.e.
, 1000 i
te
rati
ons.
4.3.
Neur
al
ne
two
rk m
od
el
Neural
net
wor
k
wit
h
BP
al
go
rithm
is
us
ed
f
or
t
rainin
g
a
nd
cl
assifi
cat
ion
pur
po
se
[
45
]
.
The
sel
ect
i
on
of
the
ne
uro
ns
nu
m
ber
f
or
t
he
input
an
d
hi
dd
e
n
la
ye
r
wa
s
gro
unde
d
up
on
t
he
ex
per
i
m
ent
per
f
orm
e
d
in
this
stud
y.
This
al
lows
the
decisi
on
on
the
a
ppr
opriat
e
num
ber
of
ne
uro
ns
f
or
the
im
pr
ove
m
ent
of
acc
uracy
of
the
cl
assifi
cat
ion
[46]
.
Me
an
wh
il
e,
as
M
A
-
B
Cl
assifi
er
will
cl
assify
24
fi
sh
fam
il
ie
s;
there
will
be
24
ne
uro
ns
within t
he ou
t
put l
ay
er.
4.4.
Memetic
al
gori
th
m
Hybr
i
dizat
ion
of
tw
o
or
m
or
e
al
gorithm
s
tog
et
her
ha
s
be
en
at
te
m
pted
by
c
ountless
sch
olars
for
the
purpose
of
i
m
pr
ovin
g
the
perform
ance
of
the
sea
rch
al
gorithm
s
[47,
48
]
.
S
uch
at
te
m
pt
is
underpinn
e
d
by
the
no
ti
on
tha
t
hybr
idiza
ti
on
al
lows
the
m
erg
i
ng
of
the
best
featu
res
f
ro
m
on
e
al
gor
it
h
m
with
tho
se
of
oth
e
rs
[
49,
50]
.
As
hi
gh
li
ghte
d
in
Mo
scat
o
in
[51]
,
m
e
m
et
ic
al
go
rith
m
is
an
au
gme
ntati
on
of
ge
netic
al
gorithm
,
excep
t
that
her
e,
a
local
search
is
app
li
ed
on
in
di
vid
uals
f
ollo
w
ing
ge
netic
operators
for
inst
ance,
si
m
ulati
ng
ann
eal
ing
, a
nd ste
epest
descen
t
a
lgorit
hm
.
Table
3
.
Ne
uro
ns
nu
m
ber
f
or
each
neural
network la
ye
r
Clas
sif
iers
NO.
Neu
ron
s in
la
y
ers
Inp
u
t .
Layer
H.
L
ay
e
r
#
1
Ou
tp
u
t.
Lay
e
r
#
3
BP
64
25
24
MA
-
B
64
25
24
Table
4.
Mem
et
ic
algo
rithm
pa
ram
et
ers
set
tin
g
Para
m
eter
Valu
e
Iter
atio
n
s n
u
m
b
er
o
f
GA
1000
Rate o
f
Cros
so
v
er
0
.09
Rate o
f
M
u
tatio
n
0
.02
Initial te
m
p
erature
1000
Fin
al te
m
p
e
rature
0
SA gen
eration
nu
m
b
er
600
Accor
ding
to
Tan
et
al
.
in
[
52]
,
the
us
e
of
local
sea
rch
al
gorithm
i
m
pr
ov
es
the
ex
plo
it
a
ti
on
process
no
t
e
xplo
rati
on
process.
Re
le
van
tl
y,
am
ong
c
ountless
sc
ho
la
rs,
m
e
m
et
i
c
al
gorithm
s
hav
e
been
ap
pli
ed
to
i
m
pr
ove
the
s
ta
nd
a
rd
ge
netic
al
gorithm
per
f
or
m
ance
-
wi
se.
N
onet
heles
s
,
am
on
g
a
num
ber
of
re
se
arch
e
rs
the
te
rm
“hybr
id”
is
us
ed
in
ste
ad
of
the
te
rm
“
m
e
m
et
ic
”
wh
e
n
ge
netic
al
go
rithm
is
com
bin
ed
with
local
search
ap
proac
h.
Acc
ordin
gly,
the
num
ber
of
ne
uro
n
s
for
e
ach
neural
net
work
la
ye
r
is
presente
d
in
Ta
ble
3.
Table
4
s
hows
the p
a
ram
et
ers
set
ti
ng
of
t
he o
f
m
e
m
e
ti
c algo
rithm
.
5.
E
X
PERI
MEN
TAL RES
UL
TS A
ND DIS
CUSSIO
N
In
previ
ou
s
st
ud
ie
s
su
c
h
as
[53]
pe
rfor
m
ed
fis
h
recog
niti
on
base
d
on
te
xture
feat
ur
es
,
the
te
xture
featur
e
s
we
re
extracte
d
on
ly
fr
om
the
fish
ve
ntral
par
t,
This
li
m
i
te
d
area
(
ven
t
ral
pa
rt)
reduces
the
ef
fecti
ve
ne
ss
of
t
he
e
xtr
act
ion
of
te
xt
ur
e
feat
ur
es
in
the
cl
assi
ficat
ion
sta
ge,
bec
aus
e
t
he
value
s
an
d
relat
ion
s
hip
s
be
tween
t
he
neighb
or
i
ng
pi
xels
of
the
fis
h
ve
ntral
par
t
te
xt
ur
e
a
re
c
onve
rg
e
d
t
o
eac
h
ot
her,
wh
e
re
this
m
a
kes
it
diff
ic
ult
for
the
cl
assifi
er
to
accu
ratel
y
reco
gniz
e
th
e
processe
d
fis
h
im
age
[1
8,
26,
54]
.
The
ta
il
of
fis
h
is
al
so
ta
ke
n
into
acco
unt
in
this
co
ntext.
So
m
e
fishes
hav
e
for
ked
ta
il
wh
il
e
oth
er
s
hav
e
rou
nd
e
d
ta
il
.
The
ta
il
sh
ape
na
ture
is
co
ns
i
de
red
i
n
this
stud
y.
Acc
ordin
g
to
[
55]
,
so
m
e
fishes
hav
e
double
e
m
arg
inate
d
s
hap
e
ta
il
,
wh
il
e
so
m
e
hav
e
l
un
at
e
s
hap
e
or
fork
ed
s
hap
e
ta
il
.
The
sp
aces
that
are
pr
esen
t
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
6
,
Dece
m
ber
2
01
9
:
5192
-
5204
5200
betwee
n
the
s
hap
e
c
ha
racter
s
of
fish
ob
j
ec
t
are
ve
ry
m
uc
h
af
fected
by
the
ext
racted
te
xture
feat
ur
e
va
lues
accor
ding
to
te
xture
m
easur
e
m
ents.
Brie
fly
exp
la
ine
d,
the
diff
e
re
nces
te
xt
ur
e
betwe
en
the
bac
kgrou
nd
i
m
age
and
the
fis
h
ob
j
ect
com
pr
ise
t
he
disti
nctio
n
betwee
n
the
fish
fam
ilies,
par
ti
cularly
betw
een
poiso
n
an
d
non
-
po
is
on f
is
h fa
m
ily acc
ordin
g t
o
te
xt
ur
e
m
ea
su
rem
ents.
Ther
e
f
or
e;
t
his
w
ork
i
nten
de
d
to
rec
over
this
li
m
i
ta
ti
o
n
by
util
iz
ing
the
dif
fer
e
nc
es
in
te
xt
ur
e
betwee
n
the
ba
ckgr
ound
i
m
age
and
the
fish
obj
ect
.
In
cert
ai
n
instances,
certai
n
fishes
possess
do
rsal
fin
an
d
adip
os
e
fin
a
nd
the
re
is
a
ga
p
(i
nter
dorsal
-
adip
os
e
s
pace)
betwee
n
the
se
two
ty
pes
of
f
in.
O
n
t
he
ot
he
r
ha
nd,
a
m
on
g
s
om
e
i
den
ti
cal
fis
hes,
t
heir
dorsal
fi
n
a
nd
a
dipose
fin
a
re
cl
os
e
t
oget
he
r
with
no
ga
p.
Table
5
sh
ows
the
dif
fer
e
nce
betwee
n
the
e
xtracted
s
ha
pe
featur
e
s
f
ro
m
po
is
on
fis
h
fa
m
ily
(Red
Sn
a
pp
e
r)
a
nd
no
n
-
po
is
on
fish
fam
i
ly
(S
com
br
idae).
A
s
can
be
see
n
in
Table
5,
in
certai
n
sit
uatio
ns
,
s
om
e
po
is
on
fish
ha
ve
s
m
al
le
r
distance
m
easur
em
ents
as
oppose
d
t
o
non
-
po
is
on
fis
h.
S
uppose
that
for
po
is
on
fis
h
obje
ct
,
the
le
ng
t
h
of
it
s
m
ou
th
is
15
.
3
pix
el
s
a
nd
t
hat
of
it
s
dor
sal
fin
is
23.
66
pix
e
ls.
Me
anwhil
e,
for
the
Sc
om
br
idae
fis
h,
t
he
l
engt
h
of
it
s
m
ou
th
is
25.33
pix
el
s
a
nd
t
hat
of
it
s
dorsal
fin
is
66.
77
pix
el
s
.
I
n
the
m
easur
em
e
nts
of
an
gle,
f
or
t
he
Scom
br
idae
fis
h,
the
c
au
dal
fi
n
an
gle
is
11
6
pix
el
s,
a
nd
f
or
the
Re
d
S
na
pper
fis
h,
the
ca
udal
fin
a
ngle
is
14
5
pix
el
s.
In
the
m
easur
em
ents
of
s
ha
pe,
for
t
he
Sc
om
br
idae
fish
obj
ect
,
it
s
co
ntour
of
le
ng
t
h
is
16
43
pi
xels
wh
il
e
that
for
Re
d
Sn
a
pper
f
ish
obj
ect
,
it
is
2625
pix
el
s.
No
ta
bly,
a
m
ong
so
m
e
fa
m
ili
es,
they
hav
e
s
i
m
i
la
r
sh
a
pe
c
har
act
e
rs.
Table
5
.
So
m
e extracted
shap
e featu
res fr
om b
oth
p
oiso
n
a
nd
non
-
poiso
n fi
sh
fam
i
li
es
Sh
ap
e Features
No
n
-
Po
iso
n
Fish
Po
iso
n
Fish
Fish
M
o
u
th
leng
th
(
P1
and
P3)
2
5
.33
1
5
.3
Distan
ce between
t
h
e r
ig
h
t
-
en
d
of
m
o
u
th
and
the start of
d
o
rsal f
in
(
P3
and
P11
)
177
104
Do
rsal Fin L
en
g
th
(
P1
1
and
P12
)
6
6
.77
2
3
.66
Cau
d
al Fin An
g
le (
P1
6
,
P1
5
and
P17
)
116
145
Leng
th
con
to
u
r
o
f
f
ish
1643
2625
Howe
ver,
eac
h
fam
ily
carries
it
s
own
trai
ts
t
hat
are
sp
eci
fic
to
s
pecies.
For
exam
ple,
f
ork
ed
ta
il
s
are
a
trai
t
to
m
a
ny
sp
eci
es.
A
s
su
ch
,
it
is
necessa
ry
to
e
m
plo
y
m
ultip
le
sh
a
pe
cha
racteri
sti
cs
in
m
akin
g
cl
assifi
cat
ion
to
m
any
diff
eri
ng
fam
ilies.
Fu
rt
her
m
or
e,
as
ind
ic
at
ed
i
n
[
56
,
57]
,
the
use
of
t
he
cha
rac
te
risti
cs
ind
ivi
du
al
ly
m
ay
no
t
offer
co
m
ple
te
identific
at
ion
.
Ra
t
her
,
the
c
har
act
eris
ti
cs
need
t
o
be
us
e
d
in
c
om
bin
at
i
on
in
order
to
pro
vid
e
ade
quat
e
inf
or
m
at
ion
in
the
cl
assifi
cation
of
num
ero
us
fam
ilies
int
o
their
co
rr
es
pond
i
ng
fam
i
li
es.
Twen
ty
four
feat
ures
wer
e
ext
rac
te
d
us
in
g
GL
CM
m
e
tho
d
ut
il
i
zi
ng
gr
ay
te
xture
m
easur
e
m
ents,
22
dista
nce
fe
at
ur
es
we
re
e
xtracted
util
iz
ing
dista
nce
m
easur
em
ents,
16
an
gle
feat
ur
es
we
re
e
xt
racted
util
iz
ing
an
gle
m
easur
em
ent
s,
an
d
2
feat
ures
we
re
extra
ct
ed
us
in
g
sta
ti
sti
cal
m
easur
e
m
ents.
The
back
-
pro
pag
at
io
n
cl
assifi
er
wa
s
a
pp
li
ed
us
i
ng
a
set
of
i
nput
f
eat
ur
es,
but
th
ere
are
iss
ues
associat
ed
with
this
cl
assifi
er
inclu
ding
ent
rap
m
ent
in
the
lo
cal
op
ti
m
a
and
lo
w
rate
of
c
onve
r
ge
nce
[
9]
.
A
s
a
so
luti
on,
a
hybri
d
m
et
a
-
heu
risti
c
al
gorithm
(MA
-
B
Cl
assifi
er)
was
pro
posed
i
n
this
stud
y.
T
he
m
e
ta
-
heuris
ti
c
al
go
rithm
s
olv
es
the
opti
m
iz
at
i
on
pro
blem
.
Fu
rt
her
m
or
e,
as
oppose
d
to
th
e
conve
ntio
nal
back
-
prop
a
gat
ion
al
go
rithm
,
m
et
a
-
heurist
ic
a
lgo
r
it
h
m
de
m
on
strat
es
high
le
ve
l
of
eff
e
ct
iv
eness
in
the
pr
e
ve
ntion
of
getti
n
g
tra
pped
i
n
the local
optim
a.
Accor
dingly
,
BP
and
M
A
-
B
Cl
assifi
ers
wer
e
us
e
d
in
te
sti
ng
t
he
ex
tract
ed
featu
re
s.
As
s
uch,
the
achie
ved
ou
tc
om
es
dem
on
st
rate
the
s
uccess
of
the
fe
at
ures
e
xtr
act
ion
a
nd
re
cogniti
on
m
eth
ods
i
n
achievin
g
high
cl
assifi
cat
ion
accuracy
as
op
po
s
ed
to
past
m
et
ho
ds.
I
n
fa
ct
,
the
best
accuracy
res
ults
wer
e
a
t
87% a
nd 95%,
wh
il
e the
wo
rs
t on
e
s
wer
e at
81% a
n
d 8
7% respecti
vely
as
shown i
n
F
i
gu
res
3
a
nd
4
.
The
re
su
lt
s
s
how
var
ia
ti
on
a
nd
t
his
is
at
tribu
te
d
to
ide
nti
cal
ness
of
s
ha
pe
an
d
te
xt
ur
e
of
m
os
t
fish
fam
i
li
es
with
on
e
a
no
t
her.
T
he
ori
gin
al
pi
xe
l
values
m
a
y
al
so
be
prese
nt
and
this
cau
se
s
identic
al
values
of
extracte
d
featu
res
w
hich
will
cause
the
c
om
ple
xit
y
of
th
e
extracte
d
fea
tures
to
increa
se.
MA
-
B
Cl
assifi
e
r
pro
po
se
d
i
n
th
is
wor
k
will
tr
ai
n
an
d
cl
assif
y
these
feat
ur
e
s.
N
otably,
s
om
e
fam
il
ie
s
of
fish
ca
rr
y
t
heir
ow
n
sp
eci
es
-
s
pecifi
c
-
trai
ts,
a
nd
t
his
facil
it
at
es
MA
-
B
Cl
ass
ifie
r
in
cl
assif
yi
ng
t
hem
.
Fo
r
insta
nce,
s
om
e
of
the
non
-
po
is
on
fish
fam
i
ly
has
an
gle
of
up
pe
r
tria
ngle
rese
m
bl
ing
oth
er
da
ng
e
r
ou
s
fish
f
a
m
ilies.
Also,
these
non
-
poiso
n
fis
hes
car
ry
so
m
e
disti
nctive
fe
at
ur
es
incl
ud
i
ng
the
sp
ace
le
ng
t
h
betw
een
t
he
rig
ht
-
e
nd
of
first
dorsal
fi
n
a
nd
the
b
e
ginnin
g
of
seco
nd
dors
al
fin,
Pelvic
fin
le
ngth
an
d
H
ead
width.
Fi
gures
3
an
d
4
present
the
outc
om
es
of
recog
niti
on
accuracy
for
each
fis
h
fa
m
ily.
Hen
ce
,
this
stu
dy
was
able
to
s
ucce
ssfu
ll
y
recog
nize
the
fam
i
li
es
of
da
ng
e
r
ou
s
fis
h
with
high
cl
as
sific
at
ion
acc
uracy
owin
g
to
their
sp
eci
es
-
sp
eci
fic
trai
ts
(d
if
fer
e
nt
sh
ape
c
om
par
ed
with
ot
he
r
fam
ily)
that
are
disti
nct
from
oth
er
no
n
-
pois
on
an
d
po
is
on
fam
i
li
es o
f
fis
h.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Ro
bu
st f
e
atu
re
ext
ra
ct
ion met
hods
f
or
ge
ner
al fi
sh
cl
as
sif
ic
ation
(
Muta
se
m
Als
madi
)
5201
As
oppose
d
to
the
conve
ntional
m
e
tho
ds,
the
extracte
d
feat
ur
es
us
in
g
the
pro
po
se
d
m
et
h
od
s
(an
c
ho
r
po
i
nts,
te
xture
and
sta
ti
sti
cal
m
easur
em
ent
s)
with
the
pr
opos
e
d
BP
an
d
MA
-
B
Cl
as
sifie
rs
sho
w
su
pe
rio
r
perform
ance
over
t
he
sta
te
of
t
he
art
m
eth
ods
us
e
d
in
[3
4
,
3
7
]
par
ti
c
ularly
with
re
sp
ect
to
recog
niti
on
accuracy
with
a
per
ce
nta
ge
of
82.
25
a
nd
90
as
sho
wn
in
F
igure
5
.
T
he
m
et
hods
of
a
nc
hor
points
a
nd
te
xtur
e
m
easur
em
ents
app
ea
r
to
be
le
ss
influ
e
nced
by
the
exp
ressi
on
of
fis
h
an
d
the
global
var
ia
ti
on
s
in
fish
obj
ect
pr
ese
nce
wit
hin
the
im
age.
Fu
rt
her
m
or
e,
MA
-
B
Cl
assifi
er
show
s
bette
r
perf
or
m
ance
wh
e
n
com
par
ed
t
o
the
co
nventio
na
l
BP
cl
assifi
er
acco
r
ding
to
the
featu
res
e
xtracted
with
GLCM,
a
ng
le
as
well
as
dist
ance
m
easur
em
ents.
Using
SA
wit
h
GA,
the
rec
ogniti
on
acc
ur
a
cy
of
MA
-
B
Cl
assifi
er
is
s
ubsta
ntial
ly
i
m
pr
ov
e
d
a
s
the w
ei
gh
ts
to be
us
e
d
in
the
proc
ess
of BP
C t
rainin
g
a
nd
processi
ng are im
pr
oved
and
op
ti
m
iz
ed.
Figure
3
.
Re
co
gn
it
io
n
acc
ur
ac
y resu
lt
s
us
in
g B
P classi
fier
Figure
4
.
Re
co
gn
it
io
n
acc
ur
ac
y resu
lt
s
us
in
g M
A
-
B cl
assifi
e
r
Figure
5
.
Com
par
is
on
gra
ph
f
or overall
acc
uracy
r
es
ults bet
ween p
r
opos
e
d B
P a
nd
MA
-
B
cl
assifi
e
r
s a
nd o
t
her co
m
par
at
ive
m
eth
ods
Evaluation Warning : The document was created with Spire.PDF for Python.