TELKOM
NIKA
, Vol.12, No
.4, Dece
mbe
r
2014, pp. 89
0~8
9
6
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v12i4.437
890
Re
cei
v
ed Se
ptem
ber 12, 2014; Revi
se
d Octob
e
r 27
2014; Accept
ed No
vem
b
e
r
15, 2014
Prediction and Realization of DO in Sewage Treatment
Based on Machine Vision and BP Netural Network
Liu Liping*, Li Zhigang, Sunjinsheng
,
Liangna
Schoo
l of Information En
gi
ne
erin
g, Hebe
i
U
n
ited U
n
ivers
i
ty, T
angshan, C
O
06300
0C
hin
a
Ph./F
ax:+
86-0
315
25
972
25
*Co
rre
sp
ondi
ng autho
r, e-mail: Liulp
@
h
euu.ed
u.cn
A
b
st
r
a
ct
Dissolv
ed Oxy
gen (DO) is on
e of the most imp
o
rtant par
a
m
eters
descri
b
ing b
i
och
e
m
ic
a
l
process
in w
a
stew
ater treatment. It is usua
lly
me
asur
ed w
i
th
diss
olv
ed oxy
gen
meters, and c
u
rre
ntly galv
a
n
i
c a
n
d
pol
arogr
ap
hic
electro
des ar
e
the pred
o
m
i
n
ant metho
d
s. Expens
ive,
me
mbr
a
n
e
surfac
e inactiv
a
tio
n
, and
espec
ial
l
y ne
e
d
of clean
in
g and ca
libr
a
tin
g
very fr
equentl
y
are common
disadv
anta
g
e
s
of electrode-
type
me
asuri
ng se
n
s
ors. In our w
o
rk, a novel
met
hod for
Pre
d
icti
on an
d Re
ali
z
a
t
ion diss
olve
d
oxyge
n
bas
ed-
on
Machi
ne V
i
sio
n
a
n
d
BP N
e
tu
ral N
e
tw
ork w
a
s rese
arche
d
.
Pictures
of the
w
a
ter-body
su
rface i
n
aerati
o
n
basi
n
s ar
e c
a
p
t
ured
an
d tran
sforme
d
into
H
S
I space
d
a
ta. T
hese
d
a
ta p
l
us the
corres
p
ond
ent
measur
e
d
DO valu
es are
process
ed w
i
th
a ne
ural
netw
o
rk. Using
th
e
w
e
ll-train
ed n
e
u
ral n
e
tw
ork, a satisfied res
u
lt
for
classifyin
g diss
olve
d oxyg
en a
ccordi
ng to the
w
a
ter-body pic
t
ures has be
en
reali
z
e
d
.
Ke
y
w
ords
: dis
s
olve
d oxyg
en,
mach
in
e visio
n
, BP neural
n
e
tw
ork, sew
age treatment
1. Introduc
tion
Dissolve
d O
xygen (DO
)
is d
e
fined
a
s
the
mea
s
u
r
e of
wate
r
quality indi
ca
ting fre
e
oxygen disso
l
ved in wate
r. The qu
antity of dissol
ved
oxygen in wa
ter is typically expresse
d in
parts pe
r milli
on (ppm
) or
milligram
s
p
e
r
liter
(mg/l)
. S
i
nce
oxygen i
s
solubl
e in
water, the a
m
o
unt
of dissolved o
xygen in wate
r is in
the stat
e of dynamic
equilib
rium [1
].
The most co
mmon appli
c
ation for dissolved oxyge
n
measure
m
ent occurs in waste
water treatm
ent. Biochem
ical b
r
ea
kdo
w
n of sewag
e
is a
c
hieve
d
by bacte
ri
al attack i
n
the
pre
s
en
ce
of
oxygen. Thi
s
pro
c
e
s
s typically ta
kes
pl
ace i
n
an
ae
ration b
a
si
n
of a wa
ste
water
treatment
pla
n
t, and i
s
a
c
compli
sh
ed
b
y
aeratin
g o
r
bubbli
ng
air
(or pu
re
oxygen) th
ro
ugh t
he
wa
ste wate
r.
Maintainin
g the prope
r co
nce
n
trati
on
o
f
dissolved o
xygen in the
aeratio
n ba
si
n is
necessa
ry to
ke
ep mi
cro
o
rga
n
isms al
ive and
allo
w
b
r
ea
k do
wn of
organi
c wa
ste.
Th
ese
microorgani
sms turn o
r
g
anic
waste
s
into inorga
nic byprodu
cts, spe
c
ificall
y
, carbon
dio
x
ide,
water an
d
sludge.
Whe
n
the me
asu
r
ed
dissolve
d oxygen
d
e
crea
se
s b
e
l
ow a
de
sired
con
c
e
n
tration
,
the plant co
ntrol
sy
stem
automatically add
s air to
t
he ae
ration b
a
sin to p
r
ovi
d
e
life-su
staini
ng
oxygen fo
r t
he mi
croorga
nism
s, an
d t
o
facilitate
th
orou
gh
mixing of the
orga
nic
wa
ste. Witho
u
t enough di
ssolved oxyge
n
con
c
ent
rati
on, benefici
a
l
microo
rg
ani
sms will die while
trouble
s
o
m
e filamentou
s microbe
s
pro
liferate, cau
s
i
ng slud
ge se
ttling proble
m
s.
Conve
r
sely,
aeratio
n i
s
th
e large
s
t
sin
g
le o
perating
expen
se,
an
d oxygen
lev
e
ls
gre
a
ter th
an the
requi
red
optimum con
c
entration
s are wa
steful an
d inefficient.
There a
r
e
many differe
nt dissolve
d
oxygen
se
nso
r
te
ch
nol
ogie
s
, ea
ch
with
its
advantag
es a
nd disadvant
age
s [2]. Most continuo
us measurement
dissol
ved ox
ygen se
nsors in
today’s m
a
rketplace u
s
e
galvani
c (spontan
eou
s voltage) or elec
trolytic (applie
d
volta
ge)
measuri
ng cells. In eith
er case, the system
m
easure
s
a
n
e
l
ectri
c
curre
n
t betwee
n
two
electrode
s,
whi
c
h ispro
p
o
rtional to th
e dissol
ved
oxygen in the pro
c
e
s
s. The mo
st bori
ng
disa
dvantag
e
of oxygen
sensor te
ch
nol
ogie
s
in
co
m
m
on i
s
the
o
b
ligatory
routi
ne mai
n
tena
nce
of DO
se
nsors [3]. The
se
n
s
or shoul
d be
clea
ned,
and
/or the
elect
r
ode
s shoul
d
be repla
c
e
d
i
n
a
sho
r
t time
p
e
riod,
an
d t
hen
re
-calibrated
ca
re
full
y.
Otherwise,
the se
nsor’
s
sen
s
ibility and
measuri
ng accuracy will be declined greatly.
For
copi
ng
with this
pro
b
lem, a data
fusion
and
artificial intell
igen
ce meth
od wa
s
p
r
es
e
n
t
ed
in
th
e
lite
r
a
t
ur
e [4
]. It u
s
es
s
e
ve
ra
l ea
s
y
me
as
ur
e
d
pa
r
a
me
te
rs
, su
c
h
as
e
l
ec
tric
a
l
curre
n
t, PH,
temperature,
etc, to p
r
edi
ct indi
re
ctly dissolved
oxygen value
u
s
ing
an
artificial
neural
netwo
rk m
e
thod. In
our work, a
vision
ba
se
d m
e
thod fo
r
cla
s
sifying
DO
was
re
se
arche
d
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Predi
ction an
d Reali
z
ation
of DO in Sewage Tr
eatm
ent Based on
Machi
ne Visi
on … (Liu Li
p
i
ng)
891
[5]. It use
s
t
he
colo
r info
rmation
of
water-body i
n
the ae
ration
basi
n
a
s
th
e
analy
s
is ba
se,
imitating an experien
c
e
d
o
perato
r
’
s
ocul
ar se
ns
e. HSIspa
ce data a
r
e extra
c
ted from the wate
r-
body imag
es.
These data
plus the m
e
a
s
ured
DO val
ues a
r
e p
r
o
c
essed
with a
n
artificial n
e
u
ral
netwo
rk.
Usi
ng the
well
-traine
d
ne
ura
l
netwo
rk,
a
sat
i
sf
ie
d r
e
sult
f
o
r
cla
s
sif
y
ing di
ss
ol
v
ed
oxygen acco
rding to the water-body pi
cture
s
ha
s bee
n reali
z
ed.
The arti
cle is stru
cture
d
a
s
follows. Secti
on I is an i
n
trodu
ction.
Section II pre
s
ent
s the
technologi
cal
background and methodology. Sect
ion III describes the image
processi
ng and
colo
ur fe
ature extra
c
tion
of wate
r-b
od
y. Section
IV gives
out th
e DO
cl
assifi
cation
metho
d
and
result. Sectio
n V ends with
the con
c
lu
sio
n
.
2. Backg
rou
nd and Meth
odolog
y
Image processing i
s
a
potential
approa
ch
to
develop a
n
online
sy
stem to
aerate
cla
ssifi
cation
in
se
wage t
r
eatme
n
t
based
on
the visual
col
o
r
pro
p
e
r
ties of
water whi
c
h
stimulate
s
hu
man eyes to
distingui
sh the vari
etie
s. BP neural ne
twork (ANN) techn
o
logy is a
kind of nonli
n
ear scie
nce d
e
velope
d fro
m
1980’
s, wh
ich trie
s to simulate so
me
basi
c
attribut
es
of
people, such as self-a
dapting,
self-o
rgani
zin
g
an
d fault tolera
nce. ANN h
a
s
be
en u
s
ed
in
many
field
s
, such
a
s
mo
de identificatio
n and system
si
mulation.Inte
grating
water colo
r
featu
r
e
o
f
ANN, the p
a
per h
ope
d that artificial
neural
net
wo
rk m
odel
co
uld pe
rform
the re
sea
r
ch
of
Aerating
Cla
s
sificatio
n
well
. For all th
ese
,
the pape
r d
r
ew the foll
owi
ng research
plan
(Figu
r
e
1).
Therein ,Wat
er photo
s
are
collecte
d
in xijiao
swa
ge
of TangShan
in July 26, 2012,and DO were
measured in
situ simulta
n
e
ous.
F
i
gure
1
. T
he Frame of Rese
a
r
ch Plan
The meth
od
ology of cla
ssifying
dissolved
oxyge
n
usi
ng ima
ge processin
g
and BP
neural n
e
two
r
k i
s
sho
w
n
schemati
c
ally
in Fig.
1. A
t
the begi
nni
ng, a lot
of image
s of th
e
wa
stewater-b
ody we
re cap
t
ured a
nd ev
ery co
rr
espo
ndent me
asu
r
ed
DO value
s
we
re
re
cord
ed.
Then the ima
ges
were pro
c
e
s
sed with
MATLAB and
their colo
ur f
eature
s
were
extracted fro
m
these im
age
s. The colou
r
feature val
ues a
nd th
e
DO cl
as
sif
i
c
a
t
i
on re
sult
s ac
cor
d
ing t
o
t
h
e
measured di
ssolved oxyge
n
values
we
re taken
as
th
e data for trai
ning the ne
ural netwo
rk. After
the trai
ning
o
f
the ne
ural n
e
twork bein
g
well
do
ne, t
he
coeffici
ent
s of
the
neu
ral net
work were
determi
ned.
Finally, at th
e on
-line
ste
p
, the
colo
ur
feature
of a
new imag
e o
f
the waste
w
ater-
body is fed
to the neural
netwo
rk a
n
d
the cla
s
sification of di
ssolved oxyge
n
is e
s
timated
accordingly.
3. Image processing a
nd color fea
t
ur
e
extra
ction
3.1 Pre-pro
c
essing of
Water Image
The wate
r
i
m
age nee
ds
seve
ral ste
p
s of
pre-proce
s
sing
befo
r
e the
colo
r feature
extraction fo
r BP neural n
e
twork a
s
inp
u
t, which i
n
cl
ude ima
ge di
gital, intercep
tion of re
sea
r
ch
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 890
– 896
892
regio
n
, image
de-noi
sin
g
a
nd image
seg
m
entation.
Th
ere, the imag
e digital has
been p
r
o
c
e
s
sed
by image in
p
u
t device
(
the
digital came
ra, sojob
s
th
at use
r
s ne
e
d
to do a
r
e
interception
of
resea
r
ch regi
on, image de
-noisi
ng and i
m
age segm
e
n
tation.
There ima
g
e
s
of
wate
r t
a
ke
nby digit
a
l camerawe
re p
r
ep
ro
ce
ssed
on
Matl
abs (6.5
).
First, the
re
search
re
gion
we
re
sel
e
ct
ed (200
dpi, (200
× 2
00)) ,
t
hen ima
ge
median
filter is
desi
gne
d to
remove the
no
ise
of the im
a
ge a
nd
ke
ep t
he d
e
tails of t
he ima
ge,an
d
it ca
n p
r
o
c
e
s
s
200
×20
0
ima
ge succe
s
sfu
lly. In the phase
of
seg
m
enting a ima
ge, use the
best threshol
d
method (BT
M
) to segm
ent image
s.
3.2 Color Fe
ature Ex
tra
c
tion of Wa
ter
Image
A color
spa
c
e is a tool to
visualize, creat
e and
spe
c
ify the colo
r.Followin
g
a spe
c
ific
appli
c
ation, the col
o
r may
be sp
ecifie
d in different
wa
ys. For hum
a
n
s the colo
r is a pe
rceptio
n of
brightn
e
ss, h
ue a
nd
satu
ration. Fo
r
co
mputers
col
o
r is an
excita
tion of three
pho
sph
o
rs
(b
lue,
red, and g
r
e
en), and for a printing p
r
ess color i
s
a reflectan
c
e and ab
sorban
ce of cyan,
magenta, yell
ow an
d bla
ck inks
on the
pape
r. So,
a colo
r spa
c
e i
s
the represe
n
tation of three
attributes
use
d
to descri
b
e
a colo
r. A color
spa
c
e i
s
also a m
a
the
m
atical rep
r
e
s
entatio
n of our
perceptio
n.
Even thoug
h
RGB
colo
r
sp
ace
ca
n b
e
u
s
ed
for
pixel
cla
ssifi
cation,
it has di
sadv
antage
s
of illuminatio
n depe
nde
nce. It means t
hat if the
illumination of i
m
age
cha
n
g
e
s, the water pixel
cla
ssifi
cation
rule
s ca
n not perform
well
. Furtherm
o
re, it is not p
o
ssible to se
parate a pixe
l's
value into intensity and ch
romin
a
n
c
e. The ch
romin
a
n
c
e ca
n be used in modelin
g colo
r of water
rathe
r
than m
odelin
g its int
ensity. This
g
i
ves mo
re
rob
u
st re
present
ation for
wate
r pixels. So it
is
need
ed to tra
n
sform RGB
colo
r space to one of t
he colo
r
spa
c
es whe
r
e
the se
paratio
n
between
intensity
an
d chromin
a
n
c
e is
mo
re discri
m
inat
e. Be
ca
use
of the
co
nversi
on
bet
wee
n
RGB a
nd
HSI
colo
r spa
c
e
s
,
we use HSI
colo
r spa
c
e
to
mo
del water pixels. The
conve
r
si
on
from RGB
to
HSI colo
r sp
a
c
e is fo
rmulat
ed as follo
ws:
(1)
(2)
(3)
(4)
Whe
r
e
H i
s
hue, S an
d I are
ch
ro
minan
ce a
n
d
intensity. T
he computati
on was
perfo
rmed
by
a p
r
og
ram
d
e
velope
d u
s
i
ng MATLAB
6.5. Mean
va
lues of
R,G
and B
of
wat
e
r
resea
r
ch regi
ons a
nd H,S and I throug
h
color
spa
c
e t
r
an
sform
a
tion
was give
n in table 1 :
Table1. The
test sam
p
le
Index
DO
R
G
B
H
S
I
5036
4.8
155.741
154.3571
151.3306
62.5562
0.0193
151.7806
5037
4.8
155.1651
153.4655
150.8566
63.4016
0.0164
152.3842
5038
4.85
159.25
158.25
155.5
49.7767
0.0138
157.6667
… …
…
…
…
…
…
…
5119
2.55
152.8489
151.6211
147.843
53.0012
0.0199
149.9324
4. Establish of BP Neural
Net
w
o
r
k for Aera
ting Cla
ssifica
tion
As the colo
r feature p
a
ra
m
e
ters
com
p
ri
sing H , S and I were a
c
hiev
ed by Matlab, and in
the stu
d
y of t
he
relativity b
e
twee
n va
riet
y and
col
o
r fe
ature
are expl
ored,
indi
cate
s it i
s
fe
asi
b
le
to
]
B
G
,
180
B
G
0
)
3
arctan(
90
[
360
1
；
，
F
H
3
B
G
R
I
]
B)
,
,
min(
[
1
S
I
G
R
B
G
G
R
B
2
F
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Predi
ction an
d Reali
z
ation
of DO in Sewage Tr
eatm
ent Based on
Machi
ne Visi
on … (Liu Li
p
i
ng)
893
use A
NN in
monitori
ng of
dissolved ox
ygen in wat
e
r for aerating
cla
ssifi
cation.
Practice
pro
v
es
that ,the artificial ne
ural n
e
t
work is
suita
b
le to simulat
e
su
ch
compli
cated relation
ship
4.1 Model of
BP Neur
al Net
w
o
r
k for
Aerating
Clas
sification
(1)
Normali
z
a
t
ion Processi
ng for ANN
The valu
es
of the three
cha
r
acte
ri
stic va
riabl
e
s
a
r
e n
o
t at thesame
q
uantitative level an
d
the rang
es of
their values
are al
so dif-f
e
rent. The
s
e
differen
c
e
s
in
fluence the repre
s
e
n
tation
o
f
each cla
s
sof stitchin
g defe
c
ts. To produ
ce the mo
st efficient traini
ng,
the dataa
re preprocesse
d
before traini
ng and ne
e
d
to be pro
c
e
s
sed in
u
n
iformme
asurement. Another rea
s
on
for
prep
ro
ce
ssin
g the data ist
hat whe
n
the
input vari
abl
es of a si
gm
oid functio
n
a
r
e in the rang
eof
[0
, 1
], th
e
d
i
s
t
a
n
c
e
s
o
f
th
e
ou
tp
u
t
s ar
e
ver
y
d
i
ffe
re
n
t, w
h
ic
h isu
s
e
f
u
l
fo
r c
l
a
s
s
i
fic
a
tion
and
recognitio
n
.T
he data prep
roce
ssing i
s
ex
pre
s
sed a
s
follows:
(5)
Whe
r
e the
sy
mbol “Y
” is th
e relativ
e
cha
r
acte
ri
stic
val
ue, “X” i
s
the
absolut
e cha
r
acteri
stic
valu
e,
“Xmin” a
nd “Xmax” are th
e minimum a
nd the maxim
u
m values of
“X”.
(2) Mo
dula
r
Structu
r
e of BP Neural Net
w
ork
Theo
retical re
sea
r
ch ha
s p
r
oved that if a
n
A
NN i
n
cl
ud
es bi
ases,at l
east a S
-
style
crypti
c
layer a
nd a
linear outp
u
t layer, it ca
n
approa
ch
an
y rational fu
n
c
tion.In this
pape
r the
m
odel
use
d
hyperbo
lic tangent S-style transfe
r
func
tion b
e
tween the input
layer and hi
dden laye
r, and
the hidde
n la
yer and outp
u
t
layer adopte
d
linear tran
sfer functio
n
.
A feedforwa
rd
neu
ral
n
e
twork was develo
ped,
wbi
c
h
wa
s mad
e
u
p
of in
put
layer
，
hi
dde
n layer an
d o
u
tput layer
．
The ba
ck pro
pagatio
n algo
rithm was u
s
ed to lea
r
n a
n
d
train th
e n
eural net
work
．
The
A
N
N
mo
del stru
cture use
d
wa
s sh
own
in Figu
re
4,
in
whi
c
h
and
denote
the ma
ss m
a
trice
s
betwee
n
the i
nput la
yer an
d hi
dd
en laye
r
，
a
n
d
the
hidde
n
layer a
nd
out
put layer,
re
spect i
rely;
a
nd
denote
th
e threshold v
a
lue
of the
n
euro
n
s of
hidde
n layer
and outp
u
t layer re
spe
e
lively.
F
i
gure 2. T
he structure of BP neur
al net
w
o
rk
min
max
min
X
X
X
X
Y
1
2
1
2
1
2
1
2
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 890
– 896
894
The followi
ng
three pa
ram
e
ters
wa
s ch
ose
n
as the i
nput nod
es o
f
the neural n
e
tworks
predi
ction m
odel
：
hue
(H),Chro
mina
nc(S) and inten
s
ity(I).And
the classificatio
n
of aerating
as
sho
w
n in Fig.
4 wa
s ado
pte
d
as outp
u
ts.
After the inpu
t and output node
s were fi
xed
on,the hi
dden n
ode
s n
eed to be fixed on by
dynamic con
s
tru
c
tion
met
hod.That i
s
,
adeq
uate
n
o
des were
e
m
ployed i
n
t
he b
eginni
ng
,and
then the voi
d
nod
es
will
b
e
re
moved
u
n
til the nu
mb
er of
hidd
en
node
s
can
n
o
t be mi
nified
. At
last,eight
wa
s ad
opted
a
s
the optimal
numbe
r of
hi
dden
nod
es.
C
on
se
quently
,the stru
ctu
r
e
o
f
neural networks p
r
edi
ction
model of Aerating Cla
s
sification 3 x 8 x 3
，
as
sho
w
n
in Figure 4.
4.2 Resul
t
s a
nd Its Analy
s
is
4.2.1 Model Training
The n
e
xt co
d
e
segme
n
t is used to
train
ea
ch m
odul
e, wh
ere
ne
wff is
a M
a
tlab n
eura
l
netwo
rk toolb
o
x function
to create
afee
d-forwa
r
d ba
ck-p
rop
agatio
n
network with
the spe
c
ifi
e
d
para
m
eters.
net=n
ewff(th
resh
old,[8,3],{'
t
ansig','l
og
sig'
},'trainlm'
)
;
net.trainParam.epochs=1000;
net.trainParam.goal=0.1;
net=init(net);
net=train(net,P,T);
F
i
gure 3. Artific
i
al n
eura
l
net
w
o
rk error curve
It can he
see
n
that the trai
ning
con
s
trin
gen
cy is p
r
ef
erabl
e for the
rea
s
on th
at the targ
et
error
can b
e
rapidly achiev
ed after trai
ning 28 ep
ochs
．
The e
r
ror
curve of traini
ng network was
sho
w
n in Fi
g.5The
refore, the relatio
n
shi
p
betwe
en co
lor feature an
d aerat
ing cl
assificatio
n
can
be set up.At
the sam
e
tim
e
,the
ma
ss matric
es,
and
,and
the th
re
shol
d valu
e v
e
ctors,
and
,can b
e
o
b
tained corre
s
po
ndin
g
lyan
d were kept a
s
the neu
ral n
e
twork pa
ram
e
ters,
w
hi
ch
were sh
own in Table 2 an
d
Table 3
．
1
2
1
2
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Predi
ction an
d Reali
z
ation
of DO in Sewage Tr
eatm
ent Based on
Machi
ne Visi
on … (Liu Li
p
i
ng)
895
Table2. Ma
ss values bet
we
en input an
d hidde
n layers, and thre
shol
d value of hid
den layer
Mass v
a
lue
s
Thres
hold v
a
lue
Hue
（
H
）
Chro
mina
nc(S
)
Inten
s
it
y
(
I)
0.6518
-7.6573
8.2532
0.8182
0.2436
-0.6510
2.6207
4.0872
-6.2405
-0.6751
0.8410
-5.367
-3.9012
-3.8356
3.8281
4.6958
0.4494
1.5544
-5.8424
2.8022
0.0992
-2.3539
-0.8121
4.6878
-6.2933
1.6178
-0.7029
-1.4721
-2.9985
-3.2261
-4.8568
-4.7064
4.2.2 Resul
t
s
and discuss
ion
The verifying
result
s we
re
showed in T
able
2 after testifyingsamp
l
es we
re inp
u
tted to
the having be
en traine
d ANN model.
Table 3. The
results of validation of the
m
odel
H S
I
H.(act
ual)
M.(act
ual)
L.(ac
t
ual
)
H.
M.
L.
0.311335
0.0172
0.614487
1
0
0
1
0
0
0.292836
0.0157
0.620588
1
0
0
1
0
0
0.308533
0.0187
0.59174
1
0
0
1
0
0
0.322089
0.0092
0.607516
1
0
0
1
0
0
0.399052
0.0149
0.614892
1
0
0
1
0
0
0.238413
0.0838
0.446399
0
1
0
0
1
0
0.234573
0.0419
0.525496
0
1
0
0
1
0
0.315992
0.0166
0.525293
0
1
0
0
1
0
0.200985
0.0771
0.492889
0
1
0
0
1
0
0.181888
0.0166
0.570262
0
1
0
0
0
0
0.339193
0.0164
0.612217
0
1
0
1
0
0
0.380631
0.0154
0.643307
0
1
0
0
1
0
0.315996
0.0303
0.568026
0
0
1
0
0
1
0.437966
0.0249
0.531571
0
0
1
0
0
1
0.485452
0.0259
0.548281
0
0
1
0
0
1
0.255985
0.0297
0.514312
0
0
1
0
0
1
0.268197
0.0249
0.644771
0
0
1
0
0
1
0.281333
0.0297
0.535976
0
0
1
0
0
1
0.248457
0.0206
0.550715
0
0
1
0
0
1
0.236011
0.016
0.626797
0
0
1
0
0
1
As can b
e
seen from Ta
ble 3,
the
predictio
n erro
rs of t
he t
w
en
ty group
s
wa
s two, i
n
other
wo
rd
s the p
r
edi
cti
ng a
c
curacy
of the m
o
d
e
l is
better
than 90%
a
c
cordi
ng to
the
experim
ent. It can be con
c
l
uded that ima
ge pro
c
e
s
sin
g
techni
que combinin
g with
artificial neu
ral
netwo
rk m
e
th
od is on
e of effective mean
s for
the predi
ction of aerating cla
s
sificati
on.
5. Conclusio
n
s
The colo
r fe
ature of wate
r image in swage tr
eatm
e
n
t
used to aerating cla
s
sification ca
n
be o
b
tained
by analy
z
ing
the water i
m
a
ge u
s
in
g ima
ge p
r
o
c
e
ssi
n
g
techniq
ue.
The
colo
r fe
a
t
ure
of HSI color
space ca
n be
descr
i
bed by
the colo
r tran
sform
a
tion fro
m
RGB col
o
r
spa
c
e.
The col
o
r fea
t
ure of HSI color spa
c
e were empl
oyed
as input fact
ors, an
d the non-li
nea
r
mappin
g
mo
del was buil
t
based o
n
there l
a
tion
ship bet
wee
n
colo
r fe
ature and
ae
rati
ng
cla
ssifi
cation
of three BP
n
eural
net
works. The
neu
ral
netwo
rk mo
d
e
l sh
ows a
hi
gh a
c
curacy
of
predi
cting the
aerating
cla
s
sificatio
n
in swag
e treatme
nt.
Referen
ces
[1]
Yang
Xia
o
min
g
, Li Mi
ngh
ua
n, Yang
Pu,
et al. Diss
o
lve
d
Oxyg
en
Pre
d
ictio
n
Mod
e
l
and
its Erro
r
Revision.
C
ontr
o
l Eng
i
ne
eri
ng
of Chin
a
. 200
4; 11(2): 127-
12
9.
[2]
S. Maqu
iné
d
e
Souz
a, Y. Gr
andv
alet, T
.
Deno
eu
x.
M
onit
o
rin
g
of
a s
l
ud
ge
de
w
a
t
e
ri
ng
equ
ipme
nt b
y
imag
e classific
a
tion.
W
a
vel
e
t Appl
icatio
ns in
Industria
l Proc
essin
g
II
. 2004; 5607: 3
4
-45.
1
1
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 890
– 896
896
[3]
F
e
i Ni, Z
huan
g F
u
, QiXin
Cao, et al.
Image pr
ocess
i
ng meth
od fo
r e
y
es l
o
cati
o
n
base
d
on
segmentation tex
t
ure.
Sens
ors
and Actuators
.
2008; 1
43(2):
439-
451.
[4]
Veere
ndra S
i
n
gh, Veer
en
dra
Si
ng
h. Appl
ic
ation
of imag
e
proce
ssi
ng
an
d radi
al b
a
sis
neur
aln
e
t
w
ork
techni
qu
es for ore sortin
g an
d
ore classific
a
ti
on.
Miner
als E
ngi
neer
in
g
. 20
10; 18(1
5
): 141
2-14
20.
[5]
W
K
. W
ong, CW
M. Yuen. Stitchin
g defect d
e
tect
ion
and cl
assificati
on us
i
ng
w
a
v
e
l
e
t transform and BP
neur
al net
w
o
rk
.
Expert Systems w
i
th Appl
ic
ations
. 20
09; 3
6
(2): 384
5-3
8
5
6
.
[6]
J. Qiao, N. W
a
ng, MO. Ngadi
, et al. Predicti
ng me
ch
an
ical
properti
es of fried ch
icken
n
ugg
ets usin
g
imag
e proc
ess
i
ng
an
d ne
ura
l
net
w
o
rk tec
h
n
i
qu
es.
Jour
nal
of F
ood E
ngi
n
eeri
ng.
2
011; 79(3);
1
0
6
5
-
107
0.
[7]
Yong Z
e
ng, D
a
-Ch
eng
Li
u, Ju-Xua
n Li,
Xi
ang-Y
u
Ho
u. Groupi
ng H
eur
istics on V
ehic
l
e Sch
edu
li
n
g
Probl
em.
Base
d o
n
C
han
ge
a
b
le E
x
pe
ndit
u
r
e
Co
efficie
n
t M
ode
l.
T
E
LKOMNIKA Indo
nesi
an J
ourn
a
l
of
Electrical E
ngi
neer
ing
. 2
013;
11(1): 39
9-4
0
.
Evaluation Warning : The document was created with Spire.PDF for Python.