TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 12, Decembe
r
2014, pp. 82
6
8
~ 827
7
DOI: 10.115
9
1
/telkomni
ka.
v
12i12.65
68
8268
Re
cei
v
ed Au
gust 16, 20
14
; Revi
sed O
c
t
ober 1
2
, 201
4; Acce
pted
Octob
e
r 29, 2
014
A Difference-Based Feature Description Method of
Image T
a
rget
Gao Qiang,
Yang Wu, Ya
ng Hongy
e*
Schoo
l of Elect
r
ical a
nd Electr
onic En
gin
eer
i
ng, North Ch
in
a Electric Po
w
e
r Univ
ersit
y
,
Baod
ing
071
00
3, Hebe
i Provi
n
ce, Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:
y
h
yfe
ng@
hot
mail.com
A
b
st
r
a
ct
T
h
is pap
er pro
pose
d
a new
meth
od of feat
ure de
scri
p
tio
n
for target recogn
ition a
nd
match
i
n
g
.
F
i
rstly, a meth
od of calc
ulati
ng the d
i
ffere
nce w
a
s defin
ed. T
he gray
valu
e matrix o
f
an imag
e w
a
s
converte
d to
a
differe
nce
val
ue
matrix. T
h
e
n
the
d
i
fferenc
e va
lue,
sha
p
e
,
ang
le
an
d
other fe
atures
of
a
regi
on a
nd t
h
e
combi
ned
fea
t
ures betw
e
e
n
regi
ons w
e
re
descri
bed. F
i
n
a
lly, the
metho
d
w
a
s ap
pli
ed
to
ide
n
tify traffic signs. Exp
e
ri
ments show
e
d
that the
prop
os
ed
meth
od c
a
n repr
esent
multipl
e
feat
ures
of
imag
e such as
the gray differences, the sha
pe cha
nges, a
nd so on. T
h
roug
h theor
etic
al an
d simul
a
ti
o
n
ana
lysis, ev
en
un
der r
o
tatio
n
, shift or
sca
le tra
n
sf
or
mati
on, n
e
w
featur
es d
e
scripti
on
metho
d
stil
l c
a
n
correctly recogni
z
e the target.
Ke
y
w
ords
:
diff
erenc
e, me
mb
ershi
p
, regio
n
, combi
ned fe
atures descr
iptio
n
, traffic signs
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Overall, th
e i
m
age f
eature
s
can
be
divi
ded i
n
to
gl
ob
al an
d lo
cal
feature
s
two
categori
e
s.
Global featu
r
es loo
k
o
b
je
cts as a
whol
e
one. Each
fe
ature
ve
ctor contai
ns
all p
a
rts (or
eve
n
all
of the pixels) information reflecting the
overa
ll prope
rties of the image. No
w there a
r
e a la
rge
numbe
r of im
age de
scripti
on and
recog
n
ition method
s ba
sed o
n
gl
obal features
[1-4]. But these
global fe
atures d
on’t hav
e cle
a
rly mat
hematical
def
initions. Th
erefore,
the extracted
features
and i
m
age
s d
on’t have
bidi
rectio
nal
uniq
uene
ss. In
ad
dition, the
s
e
feature
s
can
only ap
ply to
a
certai
n type of image.
Comp
ared with the globa
l feature
s
, lo
cal
featu
r
e
s
focu
se
s
on extraction of
detail
feature
s
an
d
have a ri
ch
n
u
mbe
r
in the
image, t
he
co
rrel
a
tion
s bet
wee
n
the feat
ure
s
a
r
e
sma
ll.
For occlusion case, the detection and
matching
of
other features will not be affected by
the
disa
ppe
ara
n
ce of some fe
ature
s
. Beca
use of th
e
s
e
advantage
s,
studie
s
of lo
cal featu
r
e
s
are
very active; a large n
u
mb
er of method
s are
p
r
op
osed [5-8]. But many of these metho
d
s
use
feature
points to rep
r
e
s
ent
the image
tha
t
don’t
have a
c
tual p
h
ysi
c
al
meanin
g
s. T
he amo
unts o
f
cal
c
ulatio
n a
nd featu
r
e
po
ints a
r
e l
a
rg
e
,
and th
e
pe
rf
orma
nce of
real-time
is
po
or. So
a nu
m
ber
of improved
method
s are need
ed [9-1
1
]
.
It has sho
w
n
that a variet
y of feature
des
cri
p
tion m
e
thod
s have
their p
r
o
s
an
d co
ns.
This pap
er m
a
inly intro
d
u
c
es
a the
o
ry
o
f
diffe
ren
c
e
measure for f
eature
de
scri
ption. Th
en
b
u
ild
a new vector descriptor to repr
esent the image. Meanwhile, the
f
easi
b
ility and effectiveness of
this de
scriptio
n method wa
s verified.
2. Image Fea
t
ure Desc
ription
Meth
od
Bas
e
d on th
e Differen
ce
2.1. Definitio
n
of Differe
n
ce Meas
ure
Formula
Note: Let
1
0
,
u
be
the memb
ership de
gre
e
th
at an obj
ect
belon
gs to a
grad
e,
D
be
the differen
c
e
betwee
n
the obje
c
t and th
e grad
e. Rule
s between
u
and
D
are as
follows
:
a)
u
f
D
, The difference (
D
) is a function of the
membe
r
ship
degree (
u
).
b) The diffe
re
nce (
D
) is m
o
n
o
tone de
crea
sing fun
c
tion
of the membe
r
shi
p
deg
re
e (
u
).
The memb
ership d
egree that an obje
c
t belon
gs to
a
grad
e is big
g
e
r; the differe
nce b
e
twe
en
the
obje
c
t and th
e grad
e is sm
aller, and vice versa.
c
)
If
y
u
x
u
y
,
x
u
,
y
D
x
D
y
x
D
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Differen
c
e-Based F
eature De
scripti
on
Method of Image Ta
rget (Gao Qia
ng)
8269
d) Attribute
s
of
D
include fu
zzi
ne
ss, ad
ditivity,
monotonicity and no
n-ne
gativene
ss.
Lemma: [12]
Let
)
,
[
x
1
,
)
x
(
f
is a real
function of
x
,
whic
h satis
f
ies
following:
a)
0
)
x
(
f
;
b)
)
x
(
f
is the strictly monotone decrea
s
in
g function ,
)
y
(
f
)
x
(
f
y
x
;
c)
)
y
(
f
)
x
(
f
)
y
x
(
f
.
Then:
x
log
c
)
x
(
f
a
(1)
Definition: Le
t
1
0
,
u
be the mem
bership
deg
ree of a pa
rt
icular g
r
ad
e (e
xcellent, goo
d,
mode
rate, etc.),
,
D
0
be the di
fference bet
ween the o
b
je
ct and its grad
e
，
)
R
c
(
c
be the
c
oeffic
i
ent,
)
a
(
a
1
be the ba
se n
u
mbe
r
. The relation
ship b
e
twee
n difference and m
e
mbershi
p
value is
:
u
log
c
u
log
c
D
a
a
1
(2)
Whe
n
the co
efficient
1
c
, the base numb
e
r
10
a
, the unit of th
e differen
c
e i
s
“step”:
u
lg
u
lg
D
1
(3)
2.2. Gra
y
Ma
trix Tran
sfor
mation
The
sen
s
itivity of the h
u
m
an eye
to g
r
ay le
vel i
s
n
on-lin
ea
r. Wh
en the
gray
value i
s
relatively low,
the resolution is very strong. When
the
differen
c
e of
gray levels i
s
big to a cert
ain
extent, the human eye
ca
n easily di
stin
guish them
[1
3]. This pap
e
r
uses the dif
f
eren
ce
con
c
ep
t
based on fu
zzy membe
r
sh
ip to chan
ge
the gray valu
e matrix of image into a
differen
c
e val
u
e
matrix. By doing this, exten
d
s lo
w g
r
ay a
r
ea a
nd
com
p
re
sses
high
gray a
r
ea
s, which m
ean
s t
he
unde
rsta
ndin
g
and ide
n
tification of hum
an eye to image.
Ho
w to dete
r
mine the m
e
mbershi
p
fun
c
tion i
s
a
key
issue. Acco
rding to the
m
onotoni
c
relation
shi
p
b
e
twee
n difference value a
nd membe
r
ship deg
ree, we choo
se Z
-
type membe
r
ship
degree fun
c
tion [14], name
l
y:
c
x
c
x
))
c
x
(
a
(
)
x
(
u
b
1
1
1
0
b
,
0
a
(4)
In this pape
r,
0
3
1
c
,
b
,
a
.
We tran
sfo
r
m
the image gray value matrix
)
y
,
x
(
I
to the diffe
ren
c
e value
matrix
)
y
,
x
(
D
:
0
1
0
1
1
3
x
x
)
y
,
x
(
I
log
)
y
,
x
(
D
(5)
The differe
nce value of ea
ch pixel in im
age (exce
p
t boun
dary poi
nts) i
s
co
mpa
r
ed
with
that of eight pixels of the
neigh
borhoo
d
.
Put the
direction of this p
o
int cha
nge
s
to the maximum
differen
c
e val
ue p
o
int a
s
t
he di
re
ction
of the poi
nt. Then th
e ima
ge di
re
ction
matrix
)
y
,
x
(
A
is
c
o
ns
tituted.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 12, Decem
ber 20
14 : 8268 – 82
77
8270
Figure 1. Det
e
rmini
ng a direction
2.3. Image Segmentation
This m
e
thod
is for g
r
ay im
age
s, so first
need
to d
o
g
r
ay image
proce
s
sing. Ne
xt is the
image regio
n
segme
n
tatio
n
. This pa
per used a m
e
th
od of regio
n
gro
w
ing. Th
e
basi
c
idea i
s
to
set the pixel
s
having
simil
a
r p
r
op
ertie
s
together to
fo
rm region
s.
Different
with
the usual g
r
ay-
based thresh
old, here we
use the
differen
c
e val
u
e a
s
th
e
c
r
iterio
n
.
T
r
an
s
f
orm th
e
r
a
ng
e o
f
differen
c
e val
ue matrix of the imag
e to 0~2
55 an
d m
a
ke the
Figu
re 2. The figu
re sho
w
s tha
t
,
after tra
n
sfo
r
ming to
the
differen
c
e
value, the
di
fference b
e
twe
e
n
lo
w-g
r
ay
pi
xels of
imag
e is
enlarged. Th
e se
gmentati
on re
sult
s
of these pa
rts
wi
ll be more cl
e
a
r an
d detail
ed than the
g
r
ay
value se
gme
n
tation.
Figure 2. Con
v
ersio
n
of gra
y
value and d
i
fference valu
e
2.4. Regiona
l Feature
s
Descriptio
n
of Image
After ea
ch im
age i
s
divid
e
d
into
regi
ons, the f
eatures of ea
ch
regi
o
n
can b
e
o
b
ta
ined. A
singl
e feature tend
s to cause
erro
rs
and affe
ct su
bse
que
nt image mat
c
hin
g
and
re
cog
n
ition.
This pap
er co
mbined
the
m
ean
of ima
ge
rep
r
e
s
ent
ma
trix and
the
shape
coeffici
ent, co
nstitut
e
d
a new regio
n
a
l feature
s
vector, nam
ely:
Modulu
s
:
)
y
,
x
(
D
N
S
R
1
(6)
Dire
ction a
ngl
e:
)
y
,
x
(
A
N
O
1
(7)
Whe
r
e,
N
is the numbe
r of all pixels in th
e regio
n
,
S
is the sh
ape
coe
fficient:
2
1
4
E
E
S
(8)
E
is the area of
region,
1
E
is the perim
eter o
f
region.
Good l
o
cal feature
s
sho
u
l
d have vari
e
t
y of
prope
rties. Th
e featu
r
es corre
s
po
nding to
the image obt
ained by the same o
b
je
ct or scen
e at
different viewin
g angle
s
sh
o
u
ld be the sa
me.
0
100
200
300
0
100
200
300
x
(
gr
ay
v
a
l
u
e)
y
(
di
f
f
er
en
c
e
v
a
l
ue)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Differen
c
e-Based F
eature De
scripti
on
Method of Image Ta
rget (Gao Qia
ng)
8271
C
u
rr
en
tly u
s
ed
in
va
r
i
an
ce
is
: tr
a
n
s
l
a
t
io
n in
va
rian
ce,
rotation invari
ance an
d scale invari
an
ce.
The followi
ng
formula
s
are
derived to verify the invariance of the ne
w regi
onal d
e
s
cripto
r.
Let
'
B
be the
region
B
tran
sforme
d by tra
n
slatio
n, rotat
i
on an
d scali
ng.
)
k
(
k
0
is the
amount of scaling,
is the angle of rotat
i
on,
x
T
and
y
T
are
respectively the shift amo
unts of
the x-axis an
d y-axis di
re
ction.
'
B
)
'
y
,
'
x
(
is th
e
point
B
)
y
,
x
(
tra
n
sf
orme
d by t
r
a
n
slatio
n,
rotation an
d scalin
g.
y
,
x
I
'
y
,
'
x
I
B
'
B
,
y
,
x
D
'
y
,
'
x
D
B
'
B
. The relations
h
ip
as
follows
:
1
0
0
1
1
kTy
kTx
cos
k
sin
k
sin
k
cos
k
,
y
,
x
,
y
,
x
'
'
[15]
(9)
Namely
：
)
Tx
sin
y
cos
x
(
k
'
x
，
)
Ty
cos
y
sin
x
(
k
'
y
1) Strike the area
s (
E
,
E
'
) and
perim
eters (
1
E
,
'
E
1
) of two region
s, re
spe
c
tively:
regionB
regionB
dxdy
)
y
,
x
(
D
)
y
,
x
(
D
E
(10)
outlineB
outlineB
dxdy
)
y
,
x
(
D
)
y
,
x
(
D
E
1
(11)
E
k
)]
Ty
cos
y
sin
x
(
k
[
d
)]
Tx
sin
y
cos
x
(
k
[
d
)
y
,
x
(
D
)
'
y
,
'
x
(
D
'
E
regionB
'
regionB
2
(12
)
1
2
1
E
k
E
'
(13)
The feature vector of the
s
e
two regio
n
s
were
O
,
R
and
'
O
,
'
R
. Then:
2
1
2
1
4
1
4
E
E
N
)
Y
,
X
(
D
N
E
E
D
S
R
(14)
R
E
k
E
k
N
'
E
'
E
N
'
D
'
S
'
R
2
1
2
2
2
1
4
4
(15)
The mo
dulu
s
values of two
regio
n
s
are th
e sam
e
. This
indicates that
the modul
us
of this
descri
p
tor h
a
s
the invaria
n
c
e of tran
slati
on, rotation a
nd scalin
g.
2) Let
)
y
,
x
(
A
be the
dire
ction
matrix of the o
r
igi
nal regio
n
,
)
'
y
,
'
x
(
'
A
be the o
n
e
s
of
the
regio
n
after tran
slation, ro
tation and scaling. Let
)
y
,
x
(
0
0
be
the center o
f
a 3*3 pixel unit o
f
origin
al regi
o
n
,
)
y
,
x
(
be the poin
t
that has maximum differe
nce value of the 9 points (
0
0
x
x
),
the dire
ction
angle of the center is:
0
0
x
x
y
y
tan
rc
a
)
y
,
a(x
0
0
(16)
The point
)
y
,
x
(
0
0
is chang
ed to
)
'
y
,
'
x
(
0
0
after trans
f
ormation:
)
Tx
sin
y
cos
x
(
k
'
x
0
0
0
，
)
Ty
cos
y
sin
x
(
k
'
y
0
0
0
Becau
s
e th
e
differen
c
e v
a
lue of e
a
ch
point is
un
cha
nge
d, onl
y the coo
r
di
nate is
cha
nge
d, so
that the maximum differen
c
e value
point is cha
nged from
)
y
,
x
(
to
)
'
y
,
'
x
(
,
cal
c
ulate the
dire
ction an
gl
e of the cente
r
point:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 12, Decem
ber 20
14 : 8268 – 82
77
8272
sin
)
y
y
(
cos
)
x
x
(
cos
)
y
y
(
sin
)
x
x
(
arctan
)
'
y
,
'
x
(
'
a
0
0
0
0
0
0
(17)
It can
be
se
en th
at
0
0
0
0
y
,
x
a
'
y
,
'
x
a'
, that m
ean
s the
dire
ction
m
a
trix
y
,
x
A
'
y
,
'
x
A
B
'
B
, The relatio
n
shi
p
bet
wee
n
the two
direction
matrix
es o
n
ly rel
a
ted with th
e
rotation a
ngl
e. So the di
rectio
n of th
e vector
de
scrib
ed the t
w
o regio
n
s
has i
n
varia
n
c
e of
transl
a
tion an
d scaling, but
doesn’t have
rotation invariance.
2.5. Combined Fea
t
ures
Des
c
ription
of the Image
If a target is constituted by
a plurality of r
egion
s, a se
ri
es of two-dim
ensi
onal vect
ors
will
be used to d
e
scrib
e
its fe
ature
s
. Whil
e
the rela
tion
ships b
e
twe
e
n
each
of these regi
on
s are
in
the ra
nge
of
feature
s
of th
e targ
et, they also
ne
ed to
be
con
s
id
ered. By
feature extra
c
tion,
we
can
use a fe
a
t
ure p
o
int at t
he cente
r
of t
he regi
o
n
to
descri
be
ea
ch re
gion.
Con
nect the
featu
r
e
points
of ea
ch adja
c
e
n
t re
gion of ta
rget
to structu
r
e
triangul
ar
me
she
s
. Th
e
si
de len
g
ths a
nd
angle
s
of ea
ch triangle
can
represent the relation
shi
p
betwee
n
regi
ons.
Let a simple triangl
e co
nsi
s
ting of three point
s be a
n
example to verify the perfo
rman
ce.
Origin
al targ
et con
s
ist
s
of three reg
i
ons, thei
r centroid
s a
r
e
)
y
,
x
(
1
1
,
)
y
,
x
(
2
2
and
)
y
,
x
(
3
3
.
Con
n
e
c
ting the three p
o
i
n
ts to stru
cture a tr
ian
g
u
lar, its sid
e
lengths a
nd angle
s
a
r
e:
C
,
B
,
A
,
l
,
l
,
l
3
2
1
. After transformation, th
e ce
ntroid
s
are
cha
nge
d
to
)
'
y
,
'
x
(
1
1
,
)
'
y
,
'
x
(
2
2
and
)
'
y
,
'
x
(
3
3
, Each value corre
s
p
ond
s to the triangle
is:
'
C
,
'
B
,
'
A
,
'
l
,
'
l
,
'
l
3
2
1
.
2
2
1
2
2
1
1
y
y
x
x
l
(18)
1
2
2
1
2
2
1
1
kl
)
'
y
'
y
(
)
'
x
'
x
(
'
l
(19)
Similarly available:
2
2
kl
'
l
,
3
3
kl
'
l
.
2
1
2
3
2
2
2
1
2
l
l
l
l
l
arccos
A
(20)
A
'
l
'
l
'
l
'
l
'
l
arccos
'
A
2
1
2
3
2
2
2
1
2
(21)
Similarly available:
B
'
B
,
C
'
C
.
It can be seen that, the side le
ngt
hs of triang
ular have transl
a
tion an
d rotation
invarian
ce,
b
u
t will va
ry with ch
ang
es i
n
scal
e, but i
f
norm
a
lized,
they will
be
unchan
ged; t
he
angle
s
of tria
ngle have inv
a
rian
ce of tra
n
slatio
n, rotation and
scale
.
The
side l
e
n
g
ths a
nd
ang
les of tri
angl
e
s
st
ru
cture
d
by feature
po
ints, the mo
d
u
lus
and
dire
ction
angl
es
of ea
ch
fe
ature
point
s a
r
e all
the
i
m
p
o
rtant featu
r
e
s
of th
e ta
rge
t. They nee
d
to
be com
b
ine
d
to describ
e the feature
s
.
3. Image Rec
ognition Pro
cess o
f
this
Paper
The ima
ge reco
gnition
proce
s
s in thi
s
pape
r can
be divide
d in
to three m
a
j
o
r
steps:
Firstly, create
a differen
c
e
vector mat
r
ix of im
age (i.e
., the difference matrix an
d the dire
ctio
n
matrix). Divid
e
the image regio
n
s a
c
co
rding to
the differen
c
e val
ue and extra
c
t the features.
Secon
d
ly, feature
s
of each regio
n
and
combin
ed
feature
s
of them are de
scri
bed. If the ta
rget
can be rep
r
e
s
ented
by a
si
ngle regio
n
,
j
u
st
a step
of feature
poi
nts
matchin
g
can
be u
s
e
d
to fi
nd
the targ
et in
the imag
e. If the target i
s
con
s
ti
tuted
by a plu
r
alit
y of regio
n
s,
con
n
e
c
ting t
h
e
feature
point
s matched
wit
h
the ta
rget
to st
ru
cture
th
e trian
gula
r
mesh
es in th
e test i
m
age.
If the
side len
g
ths
and angl
es m
a
trixes of tria
ngle me
sh ca
n be matche
d with the target too, it will be
proved that th
e target can b
e
found in the
test image.
The followi
ng
three vecto
r
matrixes a
r
e
stru
ctured for feature
s
descriptio
n
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Differen
c
e-Based F
eature De
scripti
on
Method of Image Ta
rget (Gao Qia
ng)
8273
n
n
O
R
O
R
O
R
M
2
2
1
1
1
3
2
1
23
22
21
13
12
11
2
m
m
m
l
l
l
l
l
l
l
l
l
M
m
m
m
C
B
A
C
B
A
C
B
A
M
2
2
2
1
1
1
3
Eac
h
row
of
1
M
are ve
ctor val
ues
(i.e., the modulu
s
a
nd
angle
s
)
of the feature poi
nt for
each regi
on. Each row of
2
M
are the thre
e norm
a
lized
side len
g
ths of a triangle. Each ro
w of
3
M
are the thre
e
angle
s
of the triangle.
4. Experimental Re
sults
and An
aly
s
is of Tra
ffic S
i
gns Rec
ogn
ition
In orde
r to verify the corre
c
tness and vali
dity
of this method, we u
s
e a seri
es of t
r
affic
sign ima
g
e
s
for targ
et re
co
gnition, and a
nalyz
e its p
e
rforman
c
e by
experim
ental
results.
4.1. Recog
n
ition of a Single Regional
Target
Figure 3. Simple traffic sig
n
s
Figure 3
(
a
)
a
r
e
som
e
exa
m
ples of
sim
p
le traffic
sig
n
s. Be
cau
s
e
they are
all
easy to
disting
u
ish from the backgrou
nd, so th
e pret
reatme
nt process a
nd ba
ckgro
u
nd regi
on are
not
con
s
id
ere
d
. It can be seen
that every ta
rget to
be re
cog
n
ized co
n
s
ist
s
of only a regio
n
. Take
one of
the arrow sig
n
to be
explain
ed
i
n
detail.
Comp
are t
he five i
m
age
s from
(b) to
(f),
(c)
a
n
d
(d)
can be
se
en as obtai
ne
d by rotation of (b), so
thei
r feature
s
mo
dulu
s
are alm
o
st sam
e
, onl
y
the angle
s
ch
ange. (e) loo
ks the
sam
e
with (b
), but
its si
ze i
s
bigg
er, it can be
seen a
s
obtain
e
d
by scali
ng of (b). We ca
n see their featu
r
es from the followin
g
sp
ecific values:
68
75
03
21684
1
.
.
M
b
45
298
65
21752
1
.
.
M
c
03
210
03
21265
1
.
.
M
d
19
68
67
21345
1
.
.
M
e
74
320
04
21401
1
.
.
M
f
Select the
ap
prop
riate th
re
shol
d of m
o
d
u
lus (in
this
ca
se i
s
5
0
0
)
. It can
be fo
und th
at
the five targe
t
regio
n
s
are
the sa
me traf
fic sig
n
; only
the angl
es
are ch
ang
ed
wi
th rotation.
We
can
get all t
he feature value
s
of traffic si
gn
s
in Fi
gure
3(a) u
n
der the
sa
m
e
ope
ration,
thus
achi
eving re
cognition.
Let Figu
re
4(b) b
e
an
ima
ge to b
e
ide
n
t
ified. It can
be divide
d int
o
thre
e regio
n
s afte
r
removin
g
sm
all regio
n
s. Specifi
c
features valu
e
s
are
as follows. Accordi
ng to the modul
us
and
angle
s
value
s
, the straight
arrow (th
e
se
con
d
regi
on
) can b
e
re
cog
n
ize
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 12, Decem
ber 20
14 : 8268 – 82
77
8274
Figure 4. traffic sig
n
s
comb
ined multiple
regio
n
s
37
92
79
18279
21
83
66
21862
99
80
86
18119
1
.
.
.
.
.
.
M
4.2. Recog
n
ition of Tar
g
e
t
Combined Multiple
Regions
If Figure.5
(
a
)
is taken
as a
wh
ole ta
rget co
n
s
tituted by
thre
e
small
re
gion
s, (b) is
obtaine
d by rotating 90
de
gree
s
of (a
),
(c) i
s
a
dou
b
l
e expan
sio
n
of (a
). Th
en
the combi
n
e
d
feature
s
d
e
scriptio
n of th
e imag
e me
n
t
ioned in
2.5
will b
e
n
eed
ed. Get the
feature
matrix
es
throug
h the a
bove step
s. (d), (e
) and (f
) are t
he
corre
s
po
ndin
g
tria
ngula
r
me
she
s
re
spe
c
tively.
37
92
79
18279
21
83
66
21862
99
80
86
18119
1
.
.
.
.
.
.
M
a
25
0
50
0
25
0
2
.
.
.
M
a
18
0
79
2
18
0
3
.
.
.
M
a
05
172
38
18120
77
181
78
18279
56
168
66
21862
1
.
.
.
.
.
.
M
b
50
0
25
0
25
0
2
.
.
.
M
b
18
0
18
0
79
2
3
.
.
.
M
b
01
90
40
18219
97
88
39
22655
38
90
07
17456
1
.
.
.
.
.
.
M
c
25
0
50
0
25
0
2
.
.
.
M
c
17
0
80
2
17
0
3
.
.
.
M
c
Figure 5(b
)
a
nd (c) are compa
r
ed
with (a).
The o
r
der of the feature poi
nts i
n
(b) i
s
cha
nge
d, but the modul
us
of each featu
r
e poi
nt
are
a
l
most the
sa
me, and the
angle i
s
chan
ged
about 90 de
g
r
ee
s. While t
he angle
s
an
d side len
g
th
s of each tri
a
ngle structu
r
e
d
by the feature
points a
r
e
co
nstant. All th
e featu
r
e val
ues
of (c)
are
the
sam
e
as (a
). Th
e exp
e
rime
ntal re
sults
are
ju
st like t
he the
o
retica
l de
rivation
o
f
cha
p
ter 2.
Combi
n
e
the
s
e
co
ncl
u
si
o
n
s, the
cha
n
ged
target
can
al
so
be
re
cog
n
i
zed
in th
e te
st ima
ge
as
l
ong as
the a
ppro
p
ri
ate
ju
dgment criteri
a
is
sele
ct
ed.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Differen
c
e-Based F
eature De
scripti
on
Method of Image Ta
rget (Gao Qia
ng)
8275
(a)
(b)
(c
)
(d)
(e)
(f)
Figure 5. Re
search the inv
a
rian
ce of fea
t
ures d
e
script
i
on
Similarly, we can recogni
ze all the traffic si
g
n
s in th
e Figure 4(a
)
with feature
s
value
s
and combin
e
d
feature
s
. In many act
ual ca
se
s,
the mea
s
u
r
e
d
targets
wil
l
have different
variation
s
of size, locatio
n
, etc. Chan
ge
d tar
get al
so
need to be
re
cog
n
ized. Th
erefo
r
e, anal
ysis
the invarian
ce of this description metho
d
is necessa
ry. Figure 6(a
)
and (b) a
r
e t
w
o si
gn imag
es
shot on
the
actual
ro
ad. Featur
es mat
r
ixes
and
me
she
s
can
be
obtaine
d thro
ugh a
serie
s
o
f
pro
c
e
ssi
ng a
s
belo
w
. It ca
n be found th
e two imag
es are the sam
e
sign afte
r matchin
g
. We can
recogni
ze th
e
sign
ea
sily as lon
g
a
s
the
spe
c
ie
s of
p
r
e-e
s
tabli
s
he
d
template lib
rary is e
nou
gh
to
c
o
mplete.
(a)
(b)
(c
)
(d)
Figure 6. A practical examp
l
e
100
150
200
25
0
30
0
140
145
150
155
160
140
145
150
155
160
0
100
200
300
20
0
30
0
400
50
0
60
0
28
0
29
0
30
0
31
0
32
0
50
100
150
200
50
100
150
200
250
40
60
80
10
0
12
0
20
40
60
80
10
0
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 12, Decem
ber 20
14 : 8268 – 82
77
8276
61
191
23
23743
33
211
30
13427
16
173
09
10193
89
218
62
17976
92
177
65
8872
27
171
75
2606
84
203
77
7883
09
221
12
17745
15
195
60
20515
56
186
33
13695
87
173
53
10230
37
177
65
2619
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
M
a
12
0
46
0
42
0
35
0
35
0
30
0
41
0
46
0
13
0
39
0
18
0
43
0
19
0
43
0
38
0
21
0
41
0
38
0
16
0
42
0
42
0
46
0
44
0
10
0
25
0
37
0
38
0
32
0
18
0
50
0
25
0
25
0
50
0
40
0
20
0
40
0
50
0
18
0
32
0
42
0
13
0
45
0
30
0
43
0
17
0
39
0
42
0
19
0
29
0
50
0
21
0
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
M
a
08
1
80
1
27
0
88
0
13
1
13
1
28
0
75
1
11
1
59
1
44
0
11
1
10
1
57
1
47
0
12
1
48
1
53
0
32
1
44
1
38
0
22
0
28
1
64
1
27
1
21
1
66
0
90
2
09
0
16
0
03
3
05
0
06
0
32
1
50
0
32
1
06
0
04
0
04
3
58
1
31
0
26
1
40
0
60
1
14
1
47
0
54
1
13
1
01
0
11
3
02
0
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
M
a
79
186
78
23151
19
198
17
13695
73
174
83
10303
00
225
08
17724
42
181
35
8913
60
173
33
2760
85
182
86
7608
00
195
72
17950
39
178
80
20244
11
143
47
13487
15
169
17
10315
80
169
70
2506
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
M
b
40
0
15
0
45
0
35
0
34
0
31
0
40
0
45
0
15
0
37
0
20
0
43
0
20
0
42
0
38
0
22
0
42
0
36
0
44
0
36
0
20
0
27
0
36
0
37
0
13
0
47
0
40
0
33
0
17
0
50
0
27
0
23
0
50
0
40
0
20
0
40
0
42
0
14
0
44
0
50
0
22
0
28
0
16
0
40
0
44
0
50
0
17
0
33
0
38
0
41
0
21
0
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
M
b
79
1
32
0
04
1
92
0
10
1
12
1
32
0
80
1
02
1
62
1
49
0
04
1
12
1
54
1
49
0
02
1
54
1
57
0
45
0
93
0
75
1
20
1
18
1
76
0
88
0
02
2
25
0
89
2
08
0
17
0
97
2
08
0
09
0
37
1
49
0
28
1
57
1
31
0
26
1
01
0
01
0
12
3
59
1
17
1
38
0
05
0
03
0
06
3
54
0
46
1
15
1
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
M
b
5. Conclusio
n
This p
ape
r
mainly defin
ed a meth
o
d
of cal
c
ul
ating the diffe
ren
c
e b
a
sed
on the
kno
w
le
dge of
fuzzy
m
e
mb
ership and a new metho
d
of
feature
de
scription
for t
a
rget
re
cog
n
i
t
ion
and m
a
tchin
g
.
Both the fea
t
ures of ea
ch
image
regi
on
and the
rel
a
tionship b
e
twe
en region
s
ca
n
be described.
Theoreti
c
al
analysi
s
and
experim
ent
s
verified the feasi
b
ility and effectiveness of
this de
scriptio
n method, an
d it has a better ro
bu
stne
ss.
Referen
ces
[1]
Lin
de O, Lin
d
e
berg T
.
Compo
s
ed comp
le
x-c
ue histo
g
rams:
An investi
gatio
n of
the inform
ation co
nten
t
in rece
ptive fi
eld b
a
se
d im
age
descri
p
tor
s
for object r
e
cog
n
itio
n.
Co
mp
uter Visi
on
and I
m
a
g
e
Und
e
rstand
in
g
. 2011; 1
16(4):
538-
560.
[2]
Claus
i DA, Z
hao YP. Grey lev
e
l co-oc
c
urrenc
e inte
grated a
l
gor
ithm (GLC IA): A superio
r
computati
o
n
a
l
method to
determi
ne co-
o
ccu
rre
nce pr
oba
bil
i
t
y
te
xtu
r
e features.
Co
mp
uters &
Geosciences.
200
3; 29(7): 83
7-85
0.
[3]
W
ang Y, Li
an
g D, W
ang B.
Dau
bech
i
es
w
a
vel
e
t constr
uction
usin
g h
o
motop
y
m
e
th
od.
Ch
ines
e
Journ
a
l
of
Elec
tronics
. 200
7; 16(1): 93-
96.
[4]
Claus
i DA, Hu
ang D. Desi
g
n
-bas
ed te
xtur
e feat
ure fusi
on usi
ng Gab
o
r filters and
co-occurr
ence
prob
abi
lities.
IEEE Transactions on
I
m
age Processing
. 20
05; 14(7): 9
25-
936.
[5]
Moravec HP. Procee
din
g
s o
f
the
6th
Inter
nat
io
nal
Joi
n
t
Confer
ence
o
n
artificia
l
Inte
lli
genc
e.
Visual
ma
pp
ing by
a robot rover
. 1
9
7
9
; 599- 60
1.
[6]
Harris
C, Step
hens
M. Proce
edi
ngs
of the
4
t
h Alve
y V
i
si
on
Conf
erenc
e.
A
co
mbi
n
e
d
cor
ner a
n
d
ed
g
e
detector.
19
88;
147-1
51.
[7]
Smith SM, Brad
y
M. SUSAN-
a
ne
w
a
ppr
oac
h to lo
w
lev
e
l i
m
age pr
ocessi
ng.
Internati
o
n
a
l Jour
nal o
f
Co
mp
uter Visi
on
. 199
7; 23(1)
: 45-78.
[8]
Lo
w
e
DG. Di
stinctive ima
g
e
featur
es fro
m
scale-i
nvari
ant ke
y p
o
int
s
.
Internation
a
l
Journ
a
l
of
Co
mp
uter Visi
on
. 200
4; 60(2)
: 91-110.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Differen
c
e-Based F
eature De
scripti
on
Method of Image Ta
rget (Gao Qia
ng)
8277
[9]
Ke Y, Sukth
a
n
k
ar R. Proc
ee
din
g
s of th
e C
onf
ere
n
ce
on
Comp
uter Visi
on a
n
d
Patter
n
Rec
o
g
n
itio
n
.
PCA-SIF
T
:
a more disti
n
ctive r
epres
entati
on for loca
l imag
e descri
p
tors.
20
04; 511-
51
7.
[10]
Herb
ert B, An
dreas E, T
i
nn
e T
,
et
al. Speed
ed-u
p
ro
bu
st features (S
URF
).
Co
mp
uter Visi
on
an
d
Imag
e Un
derst
and
ing.
2
008;
110(
3): 346-
35
9.
[11]
Z
hu YX, C
h
e
ng S, Stanko
v
ic V,
et al. Image reg
i
stration us
in
g
BP-SIFT
.
Journal of Vis
u
a
l
Co
mmun
icati
o
n and I
m
ag
e R
epres
entati
on.
201
3; 4(24): 44
8-45
7.
[12]
Saff EB, Snide
r
AD. F
undam
entals
of compl
e
x
an
al
ys
is
w
i
t
h
ap
plic
ations t
o
en
gin
eeri
ng,
scienc
e, an
d
mathematics. L
ond
on: Prentic
e Hall. 2
0
0
3
.
[13]
Hao Y
X
, Che
n
XJ, Z
han
g BZ
.
Photometr
y
. Be
iji
ng: Bei
jin
g
Normal U
n
iv
ersit
y
Press. 19
9
7
.
[14]
Z
hang
X,
Xia
o
XL,
Xu GY.
Determin
a
tio
n
and
an
al
ysis
o
f
fuzz
y
m
e
mb
ershi
p
for SV
M.
Journa
l of
Imag
e an
d Gra
phics
. 20
06; 8(
11): 118
8-1
192
.
[15]
Z
hang
XQ, Guo MM,
T
ang Y, et al. Ne
w
g
e
o
metric featur
e
shap
e descri
p
tor.
Computer Engi
neer
in
g
and Ap
plic
atio
ns
. 2007; 4
3
(2
9): 90-92.
Evaluation Warning : The document was created with Spire.PDF for Python.