TELKOM
NIKA
, Vol.14, No
.2, June 20
16
, pp. 692~6
9
8
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v14i1.3461
692
Re
cei
v
ed
Jan
uary 5, 2016;
Re
vised Ap
ril
21, 2016; Accepte
d
May 6
,
2016
Features Deletion on Multiple Objects Recognition
Samuel Alv
i
n Hutama, Sapta
d
i Nugr
oho, Darma
w
a
n
Utomo*
Electron
ic and
Comp
uter Engi
neer
ing F
a
c
u
lt
y, Sat
y
a W
a
ca
na Chr
i
stian U
n
iversit
y
,
Jala
n Dip
on
eg
oro 52 – 6
0
Sal
a
tiga 5
0
7
11, (+
622
98) 3
118
84
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: du88
@
y
ah
oo
.com
A
b
st
r
a
ct
T
h
is pap
er pr
esents a
multi
p
le o
b
jects re
c
ogn
ition
met
hod us
ing th
e
SURF
and the SIF
T
alg
o
rith
m. Bot
h
al
gorith
m
s
are
use
d
for fi
nd
in
g featur
es by detectin
g
k
e
yp
oints and
extra
c
ting descri
p
to
rs
on ev
ery obj
e
c
t. T
he rando
mi
z
e
d KD-T
re
e alg
o
rith
m is
then us
ed for
match
i
n
g
thos
e descri
p
tors. T
h
e
prop
osed
met
hod
is de
leti
o
n
of certai
n fe
atures afte
r a
n
ob
ject h
a
s
bee
n reg
i
stere
d
an
d rep
e
titio
n
of
successful rec
ogn
ition. T
he
meth
od is
exp
e
cted to rec
o
g
n
i
z
e
a
ll of the r
egister
ed o
b
jec
t
s w
h
ich are sh
ow
n
in an
i
m
ag
e. A
series of tests
is don
e in
ord
e
r to un
dersta
nd the c
haract
e
ristic of
the re
cogn
i
z
a
b
le o
b
j
e
ct
and th
e
meth
o
d
cap
abi
lity to
do the r
e
co
g
n
itio
n. T
he tes
t
results show
the accur
a
cy
of the pr
opos
ed
meth
od is 9
7
%
using SU
RF
and 88.7
%
usi
n
g SIF
T
.
Ke
y
w
ords
: mu
ltiple
obj
ect recogn
ition, SURF
, SIFT
, randomi
z
e
d
KD-T
re
e
Copy
right
©
2016 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Obje
ct recog
n
ition
can
be
accompli
sh
ed
usi
ng fe
ature
-
ba
sed
alg
o
rit
h
ms. T
h
e
pro
c
e
s
s is
divided into t
w
o p
a
rts, feat
ure d
e
tectio
n
and featu
r
e
matchin
g
. Fe
ature d
e
tectio
n is a p
r
o
c
e
s
s of
detectin
g
ke
ypoints of a
n
obje
c
t and
extracting
t
hem into d
e
s
cripto
rs
so
that they ca
n be
matche
d. Thi
s
process ca
n be solve
d
usin
g many
algorithms
,
suc
h
as
SURF [1], SIFT
[
2
], or
Zerni
k
e mo
m
ents [3]. There are two p
hases of
feat
ure dete
c
tion
, training ph
ase a
nd testi
ng
pha
se [4]. All obje
c
ts feat
ure
s
d
e
tecte
d
in t
he traini
ng ph
ase a
r
e
saved i
n
d
a
taba
se. Fe
atu
r
e
matchin
g
is
a proce
s
s of
comp
ari
ng fe
ature
s
det
ect
ed in the t
r
ai
ning p
h
a
s
e
with the feat
ure
s
detecte
d in the testing ph
a
s
e. Obje
ct re
cog
n
it
ion a
c
cura
cy depe
nd
s on the d
e
te
cted keypoint
s,
extracted d
e
scripto
r
s, and the matchi
ng
pro
c
e
ss [5].
Multiple obje
c
t recognitio
n
can be a
c
hieve
d
usin
g many method
s, su
ch as u
s
i
n
g
different obje
c
t detecto
rs simultan
eou
sl
y [6],
segme
n
ting image
usin
g the SLIC sup
e
rpixe
l
s
method [7], sliding win
d
o
w
method, or a hiera
r
ch
ical
pyramid stru
cture [8] whi
c
h almost wo
rks
the same
way as th
e
sli
d
ing
wind
ow method.
An
other
differe
nt method
such
as in
crowd
cou
n
ting al
so
can b
e
u
s
ed
by applying
Linea
r Interp
olation for
ca
mera di
sto
r
tio
n
calib
ratio
n
[9],
while multipl
e
brand
s in im
age
s ca
n be
done u
s
in
g sli
d
ing wi
ndo
w method [10].
The m
o
st
problem
while
impleme
n
ting
slidi
ng
win
d
o
w i
s
that th
e recognition
time
s
become
slo
w
er afte
r ea
ch
image
divisio
n
. In this
pap
er, we p
r
op
o
s
e a
sim
p
le
r
method
whi
c
h is
done by just
deleting featu
r
es of recogni
zed o
b
ject
s in
one image p
r
oce
s
sing an
d
then repe
atin
g
the recognitio
n
based on th
e keypoi
nts wit
hout any rep
eated imag
e segm
entation
s
.
From th
e o
r
ig
inal SIFT o
r
SURF,
we
on
ly c
an
get
ke
ypoints
of on
e imag
e. To f
i
nd on
e
inquiry o
b
je
ct (I) in o
ne im
age (S
), both
image
s n
e
e
d
to be p
r
o
c
e
s
sed to yield
keypoi
nts. Th
en
by matchi
ng I
-
keyp
oints to
the S-keypoi
nts u
s
in
g
Ra
ndomi
z
ed
KD-tree
[11], si
milarity bet
ween
two obje
c
ts
can be rea
c
he
d. From the l
a
st proces
sin
g
, it turns ba
ck one of thre
e re
sults n
a
m
e
ly
a true recogn
ition, false re
cog
n
ition, or
not f
ound. If there
are t
w
o
obje
c
ts I in S, both SIFT and
SURF
only shows o
ne m
a
tche
d. To
solve this
p
r
o
b
lem, we
pro
pose a te
chn
i
que to recog
n
ize
multiple obje
c
ts that fast and also
can id
ent
ify multiple identics obje
c
ts that lay on S.
2. Rese
arch
Metho
d
An obje
c
t is reco
gni
zed if the features
matc
hin
g
p
r
o
c
e
ss
gives a
true recogniti
on re
sult
.
In this
experi
m
ent, forty o
b
ject
s fro
m
v
a
riou
s
produ
ct bran
ds with
different
ch
a
r
acte
ri
stics
such
as
size,
sha
p
e
, and patte
rn differen
c
e
are u
s
e
d
in
this te
st. Every brand im
ag
e (I) is
pro
c
e
s
sed
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Features
Del
e
tion on Multi
p
le Obje
cts
Reco
gnition (S
am
uel Alvin Hutam
a
)
693
one
by on
e
usin
g SIFT
a
nd S
URF
alg
o
rithm
and
their dete
c
ted
and
extra
c
te
d keypoint
s
are
saved
in
dat
aba
se. To
id
entify multiple obj
ect
s
in
an ima
ge
(S), every I-key
points is mat
c
he
d
with S-keyp
oi
nts. Th
e follo
wing
sectio
ns de
scribe
ho
w the
multipl
e
obj
ect
s
recognition
s
wit
hout
and with feat
ure
s
deletio
n.
2.1. Multiple Object
Recognition
Multiple different-obj
ect
(n
on ide
n
tics)
reco
gni
tion
sh
own i
n
Figu
re 1 is
a p
r
ocess that
only one
obj
ect of the
sa
me brand i
s
recogni
ze
d
regardle
ss
of their o
c
curre
n
ce
s in
an in
put
image. Ho
we
ver if in one image (S
) con
t
ains non id
e
n
tics b
r
an
ds,
almost all of the bra
n
d
s
will
be
found. This i
s
one of the ch
ara
c
teri
stics
of
the original
SIFT and SURF algo
rithm
s
.
Figure 1. Multiple different
-obje
c
ts re
co
g
n
ition
To identify multiple identi
c
s and
non id
entics
obje
c
ts, we pro
p
o
s
e
an idea a
s
shown in
Figure 2, that
is by del
etin
g keyp
oints l
o
cate
d
in the
found-obje
c
t
area
after first loop searchi
ng.
The process is then repeat
ed by feature matching
again until no feature is left. T
he next secti
o
n
will discu
s
s how thi
s
ide
a
wo
rks fast
and witho
u
t repeatin
g re
cog
n
ition pro
c
e
ssi
ng from
the
begin
n
ing.
Figure 2. Multiple identi
c
s-o
b
ject
s re
cog
n
i
tion
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 692 – 69
8
694
2.2. Featur
es
deletion
From S
e
ctio
n 2.1 h
a
ve
been
de
scrib
ed that
m
u
ltiple
obj
ect
s
reco
gnition co
nsi
s
ts of
multiple different-obj
ect
s
recognition and multiple
iden
tics-obje
c
ts
reco
gnition. In
orde
r to reali
z
e
the multiple identics-obje
c
ts re
cognitio
n
proces
s, the
followin
g
step
s are n
eed
ed
to be done.
1.
Object
s which will b
e
re
cogni
zed
are
regi
ste
r
ed
by saving th
e d
e
tected
keyp
oints a
nd th
e
extracted d
e
s
cripto
rs in the databa
se.
Figur
e 3 sh
ows the key
points of a chosen obje
c
t
regi
stered in
databa
se.
Figure 3. The
detected
key
points of the
obje
c
t
regi
ste
r
ed in data
b
a
s
e are re
pre
s
ented in dot
s
2.
Feature
s
of an inp
u
t ima
ge can
be fo
und by
dete
c
ting keyp
oint
s an
d extra
c
t
i
ng de
scri
ptors
usin
g SURF or SIFT. Figu
re 4 shows a
n
exampl
e of the input ima
ge.
The presented imag
e i
s
made ne
gative just to make it clear. Th
e image
co
nsists of two id
entics obje
c
ts which n
eed
to be re
cogni
zed.
Figure 4. Two
identics obje
c
ts in the inp
u
t image.
3.
Descri
ptors i
n
the in
put im
age a
r
e
co
m
pare
d
to
th
e
descri
p
tors of
the regi
stere
d
obje
c
ts.
Th
e
original SIFT/SURF al
gorit
hm will only recogni
ze one object regardles
s how m
any times the
pro
c
e
s
s is re
peated. Fi
gu
re 5
sho
w
s th
at only t
he
se
con
d
o
b
ject i
s
recogni
ze
d
becau
se m
o
st
of the
keyp
oi
nts of
this ob
ject a
r
e
mat
c
hed
with
the keypoi
nts of the
ide
n
tics obje
c
t
in
the
databa
se.
Figure 5. The
process only
reco
gni
ze
s the
se
co
nd on
ject, even if it
is rep
eated
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Features
Del
e
tion on Multi
p
le Obje
cts
Reco
gnition (S
am
uel Alvin Hutam
a
)
695
4.
Figure 6
sho
w
s the
regi
o
n
of intere
st (ROI)
marke
d
in white a
r
ea. Features
locate
d in this
R
O
I belong to the
s
e
c
o
nd objec
t. Ther
efor
e, they
w
ill be deleted s
i
nc
e they have been
matche
d in the first-lo
op re
cog
n
ition pro
c
e
ss.
Figure 6. The
keypoint
s col
l
ection of the
se
con
d
obje
c
t are locate
d in ROI and
su
rro
und
ed
within the whi
t
e area.
5.
The features in the ROI ar
e deleted by
nullifying the
descri
p
tors. It can
be
seen
from Figure
7
that there
are
no featu
r
e
s
i
n
the
ROI. T
hen, by r
epe
ating ste
p
three, the first o
b
ject i
s
finally
recogni
ze
d. Figure 8
sho
w
s that after de
le
ting feature
s
and repe
ating the third st
ep.
Figure 7. The
ROI without feature
s
Figure 8. The
first obje
c
t is finally reco
gni
zed.
6. Steps 3-5 a
r
e repe
ated u
n
til there is
n
o
recogni
ze
d obje
c
t in the input image.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 692 – 69
8
696
3. Results a
nd Analy
s
is
The obj
ect
chara
c
te
risti
c
s and the
syst
em pe
rform
a
nce te
sts
are
run o
n
a la
ptop u
s
ing
Intel® Core™
i5-420
0M CP
U @ 2.50
GHz Pro
c
e
s
sor a
nd 2.00 GB o
f
RAM.
3.1. Objects
Char
acteristi
cs
Obje
ct
ch
ara
c
teri
stics can be
cl
assified
as
size, sh
ap
e, and p
a
ttern of the o
b
je
ct. These
cha
r
a
c
teri
stics give diffe
rent num
ber
of detec
te
d
keypoi
nts. Di
fferent num
b
e
r of d
e
tect
ed
keypoi
nts
will
affect the fe
ature
s
mat
c
h
i
ng re
sult. A
true recogniti
on is
a
condi
tion wh
ere
a
n
obje
c
t in the testing p
h
a
s
e
is re
cog
n
ized
base
d
on the
identical o
b
je
ct in the traini
ng pha
se.
The 10
regi
stered o
b
je
cts sho
w
n in
T
able 1 a
r
e
chosen ba
se
d
on their
differen
c
e i
n
cha
r
a
c
t
e
ri
st
ic
s.
Th
ey are t
e
sted
30
tim
e
s i
n
o
r
d
e
r t
o
find th
e mi
nimum
numb
e
r
of keypoin
t
s
need
ed to obt
ain the true reco
gnition rate above 80%.
Table 1. Obje
cts with diffe
rent cha
r
a
c
teri
stics give different nu
mbe
r
of keypoints
O
b
ject Number
O
b
jects
Number of
detect
ed ke
y
points
SURF
SIFT
1 Chocolatos
135
348
2 Pejo
y
323
1085
3 Semn
y
light
14
186
4 Y
o
44
248
5 Tango
75
229
6 Raspicam
133
506
7 Formula
174
304
8 Wonderland
533
1700
9 Gofruit
59
488
10 Ger
y
saluut
39
190
Figure 9. Tru
e
recognitio
n
on 10 cho
s
en
object
s
usi
n
g
SURF an
d SIFT algorithm
s
Figure 9 sho
w
s that obj
e
c
t numbe
r 4 has a tru
e
reco
gnition ra
te above 80
%. From
Table 1, this
obje
c
t has 4
4
keypoi
nts. This
con
c
lud
e
s that the succe
ss fa
ctor is not limited
o
n
the numb
e
r
of keypoint
s.
For exampl
e
,
object num
ber o
ne ha
s
135 dete
c
ted
keypoint
s using
SURF
but th
e re
cog
n
ition
result is b
e
l
o
w tha
n
50%
. It can be u
nderstoo
d th
at external fa
ctor
su
ch a
s
lig
ht reflectio
n
on
the obje
c
t al
so affect the true re
co
gnitio
n
rate. Fig
u
re
10 sho
w
s th
at
some
p
a
rt
of the
obje
c
t n
u
mbe
r
o
n
e
is hi
dde
n
b
e
cause of
light
refle
c
tion, th
erefo
r
e
ma
ki
n
g
feature
s
dete
c
tion ina
c
cu
rate. In gene
ral, the
re
cog
n
ition re
sult
above 80%
can be
rea
c
h
e
d
if
the keypoi
nts are mo
re tha
n
44 and o
b
je
ct not giving too much refle
c
tion.
3.2. The Method Perform
a
nce
From th
e 40
regi
stered o
b
j
ects, 1
0
of t
hem a
r
e
ran
domly put in
a sin
g
le ima
g
e
for 3
0
times test. Therefo
r
e, total
numbe
r of objects in all image
s is 300
(
). The result can be see
n
i
n
Table 2.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Features
Del
e
tion on Multi
p
le Obje
cts
Reco
gnition (S
am
uel Alvin Hutam
a
)
697
Figure 10. Ob
ject numb
e
r o
n
e
Table 2. Re
sult of the method pe
rform
a
nce te
st
Number
of trul
y
r
e
cognized
objects
Number of capt
u
r
ing an image using
SURF
Number of capt
u
r
ing an image using
SIFT
(
)
(ti
m
es
)
(ti
m
es
)
6 0
0
1
6
7 1
7
3
21
8 1
8
7
56
9 4
36
7
63
10 24
240
12
120
Total 30
291
30
266
True Re
cog
n
iti
o
n (
)
= 97
%
= 88.7
%
Misdete
c
tio
n
(
)
3%
1
1
.
3%
TR and M
D
i
n
Table 2 a
r
e
calculated u
s
ing the followi
ng equ
ation.
∑
100%
(1)
∑
100%
(2)
100%
%
(3)
In the Table 2, 10 obje
c
ts (
) ca
n be re
cog
n
ized well 24 (
) an
d 1
2
(
)
time
s
u
s
ing
SURF
an
d
SIFT algo
rith
m co
nsecutively. Mo
reov
er, from
30
0
obje
c
ts, S
U
RF give
s
97
%
su
ccess re
co
gnition. This i
s
8.3% better than SIFT.
Figure 11. Multiple obje
c
t recognitio
n
u
s
ing featu
r
e
s
deletion meth
od
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 692 – 69
8
698
Figure 11
shows the
re
sult of
multi
p
le obje
c
t re
cog
n
ition
s
u
s
ing fe
ature
s
deletion
method. T
o
make
the
re
sult cle
a
r, th
e i
m
age i
s
sh
o
w
n i
n
n
egativ
e, the bl
ack-b
o
x re
ctangl
es are
befall on th
e
origin
al recta
ngle results,
and the t
w
o
ellips
are
add
ed to sho
w
the un
re
cog
n
i
z
ed
obje
c
ts.
The eight
re
co
gni
zed obje
c
ts can
be se
en within
the bol
d
bla
ck re
cta
ngle
s
.
Whil
e two
obje
c
ts
can n
o
t be re
co
gni
zed
be
cau
s
e
of reflectio
n
s
as in
obje
c
t n
u
mbe
r
on
e a
nd sm
all num
be
r
of keypoint
s as in obj
ect n
u
mbe
r
three.
4. Conclusio
n
Multiple obje
c
ts
re
co
gniti
on can
be
reali
z
ed
u
s
in
g the
features
deletio
n
method.
Features in t
he testing p
h
a
se a
r
e
com
pare
d
to a
ll feature
s
in th
e training p
h
ase. Fe
ature
s
in
recogni
ze
d o
b
ject
are d
e
l
e
ted b
e
fore t
he
re
co
g
n
ition p
r
o
c
e
s
s is re
peate
d
. Hi
gher n
u
mbe
r
of
keypoi
nts
will
gives hig
h
e
r
true
recogni
tion rate.
External
fa
ctor
su
ch
as light
refle
c
tion
al
so
affec
t
the true rec
o
gnition rate. The t
e
s
t
res
u
lt
s in
this expe
ri
ment sh
ow t
hat an obj
ect is
recogni
za
ble
if there i
s
a fixed part
of the
obje
c
t
whi
c
h h
a
s
more th
an 4
4
keyp
oints.
The
accuracy
of t
he p
r
op
osed
method
ba
se
d 40
registe
r
ed o
b
je
cts
an
d 30
time
s te
st is 97%
u
s
i
ng
SURF a
nd 88
.7% using SIFT.
Referen
ces
[1]
H Ba
y
,
A Ess,
T
T
u
y
t
ela
a
rs, LV Gool. SU
R
F
: Speede
d U
p
Ro
bust F
eat
ures.
Co
mpute
r
Vision
an
d
Imag
e Un
derst
and
ing
. 2
008;
110(
3): 346-
35
9.
[2]
DG Lo
w
e
. Di
stinctive Ima
g
e
F
eatur
es fr
om Scal
e-Inva
riant K
e
ypo
i
nt
s.
Internatio
na
l Jour
na
l of
Co
mp
uter Visi
on
. 200
4; 60(2)
: 91-110.
[3]
Saptad
i Nu
gro
ho, Darma
w
a
n
Utomo. Rotati
on In
vari
ant In
de
xi
ng F
o
r Ima
ge Usi
ng Z
e
rn
i
k
e Moments
and R
–
T
r
ee.
TELKOMNIKA Telec
o
mmun
icat
ion C
o
m
puti
ng
Electron
ics an
d Contro
l.
201
1; 9(2): 335-
340.
[4]
M Murai, S T
ade. Mu
lti-vie
obj
ect recog
n
it
ion
usin
g F
e
a
t
ure detecti
on
& descri
p
tor
alg
o
rithms.
Internatio
na
l Journ
a
l of Eng
i
n
eeri
ng,
Educ
ati
on an
d T
e
chn
o
l
ogy (ARDIJEE
T
)
. 2015; 3(2).
[5]
PM Panch
a
l, SR Panch
a
l,
SK S
hah. A C
o
mparis
on
of SIFT
and SURF
.
Internation
a
l Jour
na
l of
Innovativ
e Res
earch i
n
Co
mp
uter and
C
o
mmu
n
ic
ation En
gin
eeri
ng (IJIRCCE)
. 201
3; 1(2): 323-3
27.
[6]
K Khur
ana,
R
A
w
a
s
thi. T
e
ch
niq
ues for
Obj
e
ct Re
c
ogn
itio
n in
Imag
es a
nd Mu
lti-Obj
e
c
t
Detectio
n.
Internatio
na
l J
ourn
a
l
of Adv
ance
d
R
e
se
ar
ch i
n
C
o
mp
uter En
gin
eer
ing
& T
e
ch
no
log
y
(IJARCET
)
.
201
3; 2(4): 138
3-13
88.
[7]
M Lop
ez-de-
la
-Call
e
j
a
, T
Nagai, M Attami
mi, MN
Mi
yat
a
ke, HP Me
an
a. Object Det
e
ction
Usin
g
SURF
and Su
p
e
rpi
x
els.
Journ
a
l of Softw
are Engi
neer
in
g an
d Appl
icatio
ns
. 201
3; 6: 511-5
18.
[8]
D Schmitt, N M
c
Coy
.
Object
Classification and Localizat
ion
Using SURF Descriptors.
Cite
s
eer
. 201
1.
[9]
H Z
h
a
ng, H
Gao. L
a
rge
Cro
w
d
B
a
s
ed
on Impr
o
v
ed S
URF
Algorit
hm.
TELKOMNIKA
T
e
leco
mmunic
a
tion C
o
mputi
n
g Electron
ics a
nd Co
ntrol
. 20
14; 12(4): 8
65-
874.
[10]
M Skocz
y
las.
Automatic R
e
c
ogn
ition
of Mu
l
t
iple Br
ands
in
Images o
n
M
obil
e
D
e
vices.
Advanc
es in
Co
mp
uter Scie
nce Res
earch
.
201
3; 10: 81-9
7
.
[11]
Silp
a-Ana
n
C, Hartle
y R.
Opti
mised
KD
-trees for fast im
ag
e descriptor matching.
Co
mp
uter Vis
i
on
and Patter
n
Re
cogn
ition (CVP
R)
. 2008.
Evaluation Warning : The document was created with Spire.PDF for Python.