Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 4
,
A
ugu
st
2016
, pp
. 15
77
~
1
586
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
4.1
006
7
1
577
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Vision-b
ased Crack Identification
on the Concrete Slab Surface
Using F
u
zzy Reas
oning Ru
les and Self-Organizing
Kw
an
g B
a
ek
Ki
m
1
, H
y
un
Jun P
a
rk
2
, Doo Heon So
ng
3
1
Departem
ent
of
Com
puter Eng
i
neering
,
S
i
l
l
a
Univers
i
t
y
,
Repub
li
c of Kor
e
a
2
Department of Computer
Engin
eering
,
Pu
san National University, Repub
lic of
Korea
3
Department of Computer
Games, Yong-In Song
Dam Co
llege, R
e
public of
Korea
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Feb 4, 2016
Rev
i
sed
Jun
29,
201
6
Accepte
d
J
u
l 16, 2016
Identif
yi
ng cr
ac
ks
on the s
u
rf
ac
e of
conc
r
e
te slab structure
is important for
structure
stabi
lit
y m
a
in
ten
a
nce
.
I
n
order
to avo
i
d
subjectiv
e v
i
sual
inspect
ion
,
it is necessar
y
to develop an
automate
d identification and measu
r
ing s
y
stem
b
y
v
i
sion based m
e
thod. Al
t
hough ther
e h
a
ve b
een som
e
intellig
ent
com
puteriz
ed in
spection m
e
tho
d
s, the
y
are se
nsitive
to noise
due to
th
e
brightness contr
a
st and objects such as
form
s and joints of certa
i
n
size often
falsely
classified
as cracks. In this
paper, we propose a new fuzzy
logic based
image processing method that extracts cr
acks fr
om concrete
slab structur
e
including s
m
all
cracks
that were ofte
n neg
l
ected as noise.
We extract
candid
a
te
cr
ack
are
a
s
b
y
app
l
ying fuz
z
y
m
e
th
od with thr
e
e
c
o
lor ch
annel
values
of con
c
r
e
te s
l
ab s
t
ructu
r
e. Th
en furth
e
r refinem
e
nt
pr
oces
s
e
s
are
performed with
Self Organ
i
zin
g
Ma
p algorith
m and density
based nois
e
rem
oval proces
s
to obtain b
a
s
i
c
crack
char
act
er
is
tic a
ttr
ibutes
f
o
r further
analy
s
is.
Experimental result ver
i
fies
that th
e pro
posed method is sufficien
tly
identif
ied cr
acks
with various sizes
with high accuracy
(97.3%) among 1319
ground truth
cr
acks from 30 images.
Keyword:
Conc
rete slab
s
t
ructu
r
e
Crack
i
d
en
tificatio
n
Fuzzy logic
Noi
s
e
rem
oval
Sel
f
or
gani
zi
ng
m
a
p
Copyright ©
201
6 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Kwa
n
g
Bae
k
Kim
,
Depa
rt
em
ent
of C
o
m
put
er
En
gi
nee
r
i
n
g,
Silla Un
iv
ersit
y
,
1
4
0
Baeg
yang
-d
aero
(
B
lvd
)
70
0 b
e
on
-g
il (
R
d
)
, Sasang
-gu
,
Bu
san 469
58
,
Rep
u
b
lic of
Ko
r
e
a.
Em
ail: gbkim
@
silla.ac.kr
1.
INTRODUCTION
Crack
s in
co
ncrete slab
s h
a
v
e
h
a
rm
fu
l in
flu
e
n
ce on
th
e
to
leran
c
e, du
rab
ility, waterp
ro
ofing
,
and
appea
r
a
n
ce of the structure
thus t
h
ey should
be m
eas
u
r
ed
co
rrectly i
n
tim
e. Th
ere hav
e
bee
n
nu
m
e
rou
s
st
udi
es
usi
n
g
s
e
ns
or t
e
c
hni
qu
e [1]
or
f
r
om
t
h
e st
r
u
ct
u
r
e
he
al
t
h
m
oni
t
o
ri
n
g
(
S
HM
)
poi
nt
of
vi
e
w
by
vi
brat
i
o
n-
base
d m
e
t
hod
ol
o
g
i
e
s f
o
r
va
r
i
ous st
ruct
ures
i
n
[
2
]
-
[
6
]
an
d se
veral
re
vi
e
w
s m
a
y
su
m
m
ari
ze t
echni
qu
es an
d
their c
h
aracteri
s
tic
s in
[7
],
[8
].
Whi
l
e
m
o
st
S
H
M
m
e
t
hods
base
d
on
vi
br
at
i
on a
n
al
y
s
i
s
t
r
y
t
o
ext
r
act
gl
obal
m
odal
feat
u
r
es
as
si
gnat
u
res
o
f
s
t
ruct
u
r
al
i
n
t
e
gr
i
t
y
, No
nd
est
r
u
c
t
i
v
e Eval
uat
i
o
n
(N
DE
) m
e
t
hods
, es
peci
al
l
y
w
h
en
t
w
o
o
r
hi
g
h
er
di
m
e
nsi
onal
i
m
agi
ng m
e
t
h
o
d
s a
r
e em
pl
oy
ed, a
r
e a
b
l
e
t
o
provide
a di
re
ct characte
r
ization
of l
o
cal st
ruct
ural
dam
a
ge [9]. T
hus
, we a
r
e interested in s
u
ch an approach
for ide
n
tifying cracks i
n
co
ncrete slab structure wit
h
in
tellig
en
t im
a
g
e
p
r
o
cessi
n
g
.
In
practice, eng
i
n
e
ers larg
ely
rely o
n
v
i
su
al
in
sp
ectio
n
which
is q
u
a
litati
v
e
and
sub
j
ecti
v
e in
n
a
ture
th
at d
e
p
e
nd
s
on
th
e i
n
sp
ector’s ex
p
e
rtise [10
]
. Th
u
s
, it is m
u
ch
n
eed
ed
meth
o
d
o
l
og
y to
au
to
m
a
te id
en
tifyin
g
and a
n
alyzing
crack c
h
aracte
r
istics such as
width, leng
th an
d
d
i
rection with
i
m
ag
e proces
sing techniques
[1
1]
.
Unfortunately, there is no firm
m
a
the
m
atic
al
m
ode
l for th
e crack fi
gu
res
.
An
d the c
onc
rete structu
r
e
is expose
d
t
o
t
h
e e
x
ternal environm
ent right after th
e c
o
nstruction, c
o
ns
eque
ntly a perfect crac
k e
x
traction
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
77
–
1
586
1
578
m
e
t
hod i
s
y
e
t
t
o
be de
vel
o
p
e
d. Va
ri
o
u
s i
m
age pr
ocessi
ng t
e
c
hni
ques
suc
h
as
W
a
vel
e
t
t
r
ansf
orm
,
Fo
uri
e
r
tran
sform
[1
2
]
, ad
v
a
n
c
ed
f
iltering
[1
3
]
, ad
aptiv
e th
resho
l
d
i
n
g
[1
4
]
,
p
e
rco
l
atio
n
[15
]
, C-V m
o
d
e
l [16
]
,
fractal
di
m
e
nsi
on a
n
al
y
s
i
s
[1
7]
, a
n
d
i
n
co
r
p
o
r
at
i
n
g
st
at
i
s
t
i
cal
i
n
ference
[1
8]
f
o
r
va
ri
o
u
s
goal
s
.
On
e
of th
e
practical d
i
fficu
lties in
d
e
v
e
lo
p
i
ng
au
to
m
a
tic crack
id
entificatio
n
to
o
l
with
im
ag
e
pr
ocessi
ng t
e
c
hni
que i
s
rem
ovi
n
g
n
o
i
s
e ef
fe
ct
i
v
el
y
and acc
ur
ately. Es
pecially for concre
te slab
stru
cture, it is
m
u
ch harder t
h
an that
of
verti
cal
m
a
terials like walls in
t
h
at
fre
que
ntly, form
s and joi
n
ts a
r
e falsely ident
i
fied
as crac
ks
[
19]
. E
v
en
t
h
e
t
r
a
c
ks
of
wat
e
r l
eaki
n
g c
o
ul
d
be m
i
si
dent
i
f
i
e
d as
crac
ks
s
i
nce
oft
e
n
t
i
m
es, t
h
e
bri
ght
ness c
o
nt
rast
i
s
not
e
n
ou
gh
t
o
di
scri
m
i
nat
e
suc
h
ob
ject
s aut
o
m
a
ti
cal
ly.
Thu
s
,
p
r
ev
iou
s
ly we app
lied
in
tellig
en
t b
i
n
a
rizatio
n
p
r
o
c
edu
r
es and
im
ag
e restoratio
n
t
r
eat
m
e
n
t
to
reduce s
u
c
h
fa
lse positives a
s
shown
i
n
[11],[20],[21].
Howe
ve
r, in pra
c
tice, it is found that s
u
ch
m
e
thods
h
a
v
e
d
i
fficu
lties wh
en
th
ere is n
o
clear distin
ctio
n
in
brigh
t
n
e
ss
b
e
tween
th
e crack
an
d
t
h
e surface and
sen
s
itiv
e to th
e ou
tdo
o
r ligh
t
s.
Thu
s
, i
n
th
is
pap
e
r, we propo
se a m
e
th
o
d
to
ov
erco
m
e
o
r
at least mitig
ate su
ch weakn
e
sses. Th
e
hi
g
h
l
i
ght
s
o
f
ou
r
new m
e
t
hod a
r
e e
x
t
r
act
i
ng c
a
n
d
i
d
at
e c
r
ack a
r
eas
wi
t
h
f
u
zzy
rea
s
o
n
i
n
g w
h
i
c
h ha
s be
e
n
applied t
o
m
a
ny engi
neeri
n
g areas success
f
ully [22] by
gi
ving R, G, B
channel val
u
es
of concrete surface
i
nde
pen
d
e
n
t
l
y
and
rem
ovi
n
g
noi
se
f
r
om
t
h
o
s
e can
di
dat
e
s
by
Sel
f
O
r
gani
zi
ng M
a
p
(S
O
M
) [
2
3]
.
W
i
t
h
suc
h
treatm
e
nt, minute noises
(less
than 1c
m
long)
that we
re not re
m
oved
be
fore are s
u
cces
sf
ully discriminated
by
the de
nsity dist
inction
bet
w
ee
n
the
norm
a
l s
u
rface a
n
d the
crack.
2.
IDENTIFYING CRACKS
FROM CONCRETE
SLAB
SURFACE
IMAGE
The
o
v
eral
l
di
agram
of t
h
e
pr
op
ose
d
c
r
ack
e
x
t
r
act
i
o
n m
e
t
hod
i
s
as s
h
ow
n
i
n
Fi
g
u
r
e
1.
Fig
u
re
1
.
Al
g
o
rith
m
o
u
tlin
e
2.
1.
Local Smooth
i
ng
In o
r
der t
o
e
n
hance t
h
e
bri
g
ht
ness c
ont
rast
, we use l
o
cal
sm
oot
hi
n
g
t
echni
que
whi
c
h
di
vi
des t
h
e
i
m
age into m
u
ltiple bloc
ks a
n
d a
pply sm
oot
hing
fu
nction Ski defi
ned
as form
ula
(1)
to each bloc
k.
)
1
(
1
)
(
)
(
,
1
0
)
(
)
(
1
0
1
0
0
L
S
f
X
P
X
T
X
n
n
X
P
X
T
S
i
k
i
HE
L
j
j
i
X
L
i
k
k
j
i
i
j
k
j
j
i
X
k
i
i
k
(
1
)
whe
r
e
n
i
den
o
t
e
s
t
h
e num
ber of pi
xel
s
i
n
t
h
e
i
th
bl
oc
k a
nd
n
k
i
den
o
t
e
s t
h
e b
r
i
g
ht
ness
val
u
e
of
k
th
b
r
i
g
ht
est
pi
xel
in
th
e
i
th
bl
oc
k
and
T
i
(
X
k
) is th
e cu
m
u
lated
su
m
o
f
h
i
stog
ram
P
X
i
(
X
j
) fo
r e
ach o
f
L
b
l
o
c
ks. Th
e
resu
lt of th
is
pr
ocess
i
s
t
o
o
b
t
a
i
n
e
nha
nce
d
n
o
rm
al
i
zed bri
ght
ness
val
u
e
a
m
ong
L
bl
oc
ks
as defi
ne
d
i
n
f
o
rm
ul
a
(1
).
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Visio
n
-ba
s
ed
C
r
a
ck
Id
en
tifica
tio
n
on
th
e Concrete S
l
ab
Su
rfa
ce
Using
Fu
zzy .... (Kwa
ng
Ba
ek Kim)
1
579
M
i
nut
e crac
ks
are si
m
i
l
a
r t
o
n
o
i
s
es i
n
bri
ght
ness.
I
n
o
r
de
r t
o
di
st
i
n
g
u
i
s
h t
h
em
, we di
vi
d
e
t
h
e o
r
i
g
i
n
al
im
age t
o
sm
all ob
ject
s an
d ap
pl
y
l
o
cal
sm
oot
hi
ng m
e
t
hod a
s
fol
l
o
wi
n
g
an
d t
h
e res
u
l
t
i
s
as sho
w
n i
n
Fi
gu
re
2
(b
).
(a)
Orig
in
al imag
e
(b
) After lo
cal m
o
o
t
h
i
ng
Fi
gu
re
2.
Local
sm
oot
hi
n
g
ef
f
ect
2.
2.
Extr
actin
g
C
a
ndida
te
Cr
ack
Are
a
s w
i
th
F
u
zz
y meth
o
d
The
ori
g
inal concrete im
age may have low
cont
rast
suc
h
t
h
at the cracks
and a
d
jace
nt noise area
ha
ve
sim
i
l
a
r bri
ght
ness ra
nge
. U
s
i
ng t
h
i
s
cha
r
act
eri
s
t
i
c
, we di
vi
de l
o
cal
l
y
sm
oot
he
d im
age i
n
t
o
m
a
ny
sm
al
l
random
object
s and c
o
m
pute avera
g
e
gray value. Then
f
o
r t
h
e a
r
ea
havi
ng
bel
o
w a
v
e
r
age
gray
val
u
e
,
we
appl
y
fuzzy
m
e
t
h
o
d
t
o
R
,
G
,
B
cha
n
nel
i
n
f
o
rm
at
i
on wi
t
h
co
rres
p
on
di
n
g
m
e
m
b
ershi
p
f
unct
i
o
ns
de
fi
n
e
d as
Fi
gu
re 3
.
M
e
m
b
ers
h
i
p
fu
nct
i
o
n ra
nge
fo
r Fi
g
u
re
1 i
s
defi
ne
d as Tabl
e
1. T
h
e n
o
t
a
t
i
on l
i
k
e R
(G, B
)
i
n
T
a
bl
e
1
denotes t
h
e a
v
erage
col
o
r c
h
annel
valu
e in ob
j
e
c
t
ar
e
a
in R,
G,
B ch
ann
e
l in
r
e
sp
e
c
tiv
e
l
y.
Fo
r
ex
amp
l
e
,
in
Tab
l
e 3
,
v
a
riable
V3
is d
e
termin
ed
as;
avera
g
e(R
cha
nnel
value
)
×
(3
/4) in R ch
ann
e
l
a
v
e
r
ag
e
(
G
ch
an
n
e
l
v
a
lu
e)
×
(3
/4) in G ch
ann
e
l
avera
g
e(B
cha
nnel
value
)
×
(3
/4) in B ch
ann
e
l
with
in
[0
, 25
5
]
an
d th
at
v
a
riab
le no
tatio
n v3
is
o
n
t
h
e
x
-
ax
is of Fi
g
u
re
3 (a
)-
(c)
in
res
p
ectively
.
By
apply
i
n
g
m
e
m
b
ers
h
i
p
fu
nct
i
o
n
as sh
ow
n
i
n
Fi
gu
re 3, we ha
ve fuz
z
y
sym
bol
R
1
-
R
4, G1
-
G
4
,
B
1
-B
4 wi
t
h
resp
ect
t
o
t
h
e
b
r
i
g
ht
nes
s
val
u
e
o
f
t
h
e
pi
xel
.
(a) R
m
e
mbers
h
ip f
u
nction
(
b
)
G
ne
m
b
ershi
p
function
(c) B
m
e
m
b
ership
f
unction
Fi
gu
re
3.
M
e
m
b
ers
h
i
p
f
u
nct
i
o
ns
of
R
,
G, B
C
h
an
nel
-
fi
r
s
t
part
Tabl
e
1. M
e
m
b
ershi
p
fu
nct
i
o
n
ra
nge
f
o
r
Fi
g
u
r
e
3
v1
0
v2
R(G, B
)
/ 2
v3
R(
G,
B
)
* (
3
/ 4)
v4
R(
G,
B
)
* (
5
/ 4)
v5
R(
G,
B
)
* (
6
/ 4)
v6
255
R(G, B
)
=
average
R(G, B
)
in object a
r
ea
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
77
–
1
586
1
580
The
n
we ca
n e
x
tract candidat
e
crack a
r
eas by the follo
wing
fu
zzy reaso
n
in
g
ru
les.
Ag
ain
,
th
is set
of
ru
les is symme
tric with resp
ect to
th
e
c
o
l
o
r
chan
nel
R
,
G,
and
B
t
hus
w
e
de
n
o
t
e
ra
n
g
e
vari
a
b
l
e
R
1
,
G
1
,
an
d
B1
wit
h
resp
ect to
th
e co
l
o
r ch
ann
e
l as
R1(G
1, B1)
f
o
r
notatio
n
a
l conv
enien
ce.
IF X is R1(G1,
B1) and Y
is R1
(G1, B1)
then
W is C1
IF X is R3(G3,
B3) and Y
is R1
(G1, B1)
then
W is C2
IF X is R1(G1,
B1) and Y
is R2
(G2, B2)
then
W is C1
IF X is R3(G3,
B3) and Y
is R2
(G2, B2)
then
W is C3
IF X is R1(G1,
B1) and Y
is R3
(G3, B3)
then
W is C2
IF X is R3(G3,
B3) and Y
is R3
(G3, B3)
then
W is C3
IF X is R1(G1,
B1) and Y
is R4
(G4, B4)
then
W is C3
IF X is R3(G3,
B3) and Y
is R4
(G4, B4)
then
W is C4
IF X is R2(G2,
B2) and Y
is R1
(G1, B1)
then
W is C1
IF X is R4(G4,
B4) and Y
is R1
(G1, B1)
then
W is C2
IF X is R2(G2,
B2) and Y
is R2
(G2, B2)
then
W is C2
IF X is R4(G4,
B4) and Y
is R2
(G2, B2)
then
W is C3
IF X is R2(G2,
B2) and Y
is R3
(G3, B3)
then
W is C2
IF X is R4(G4,
B4) and Y
is R3
(G3, B3)
then
W is C4
IF X is R2(G2,
B2) and Y
is R4
(G4, B4)
then
W is C3
IF X is R4(G4,
B4) and Y
is R4
(G4, B4)
then
W is C4
Fu
zzy
r
e
aso
n
i
n
g r
u
le
(1
)
W
i
t
h
abov
e fuzzy reason
ing
ru
les, w
e
h
a
v
e
th
e qu
alitativ
e ran
g
e
v
a
riab
le C1
to
C4
. Th
en
, th
e second
fuzzy
m
e
m
b
ershi
p
f
unct
i
o
n
d
e
fi
ne
d as Fi
gu
r
e
4 i
s
use
d
t
o
obt
ai
n
t
h
e
fi
na
l
m
e
m
b
ershi
p
deg
r
ee.
Fo
r e
x
am
pl
e,
each
G c
h
a
nne
l value
is gi
ve
n to the
m
e
m
b
ershi
p
func
tion de
fine
d in Figure
3 to c
o
m
pute the
m
e
m
b
e
r
shi
p
degree. T
h
e
n
the reas
oning rule (1
) is applied with Max-Min
m
e
thod.
T
h
en t
h
e second
m
e
m
b
ership function
defi
ned as
Fi
g
u
re
4 i
s
ap
pl
i
e
d t
o
det
e
rm
i
n
e t
h
e m
e
m
b
ershi
p
de
gree a
n
d i
t
i
s
defuzzi
fi
ed
by
cent
e
r
of
g
r
avi
t
y
m
e
t
hod a
s
f
o
r
m
ul
a (2).
Fi
gu
re
4.
M
e
m
b
ers
h
i
p
f
u
nct
i
o
n
- sec
o
n
d
pa
rt
)
(
)
(
X
u
X
X
u
W
(
2
)
The
n
the
decisi
on rule for ca
ndidate
a
r
ea
of
c
r
ack
i
s
de
fi
ne
d
as Tabl
e
2.
Table
2. C
r
iteria for c
r
ac
k ca
ndidates
0 < W < 2
Candidate Cr
acks
2 < W < 4
4 < W < 6
6 < W < 8
Noise Area
2.
3.
Further
Re
finement
by Self
Organiz
i
ng Map
Algorithm
R
G
B
col
o
r
i
n
f
o
rm
at
i
on m
a
y
not
be
su
ffi
ci
e
n
t
l
y
st
ro
ng
t
o
di
st
i
n
g
u
i
s
h
cra
c
ks
fr
om
noi
se
s. T
hus
w
e
appl
y
sel
f
-
o
r
g
a
n
i
z
i
n
g
m
a
p al
g
o
ri
t
h
m
on
t
h
e i
m
age aft
e
r
ap
p
l
y
i
ng f
u
zzy
m
e
t
h
o
d
s
h
ow
n as
Fi
gu
re
5.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Visio
n
-ba
s
ed
C
r
a
ck
Id
en
tifica
tio
n
on
th
e Concrete S
l
ab
Su
rfa
ce
Using
Fu
zzy .... (Kwa
ng
Ba
ek Kim)
1
581
Figure
5. Im
age after a
p
plying
fuzzy m
e
thod
SOM
i
s
a
n
u
n
s
upe
r
v
i
s
ed
ne
ur
al
net
w
or
k l
e
a
r
ni
n
g
algorithm that m
i
mics the
cha
r
acteristics of
hum
an
cerebral corte
x
and has been
success
f
ully applied to m
a
ny engi
neeri
ng
a
p
plications
[23]. The perform
a
nce of
SO
M learn
i
ng
is k
n
o
w
n
as b
e
in
g
inf
l
u
e
n
c
ed
b
y
th
e typ
e
o
f
lear
n
i
ng
r
a
d
i
u
s
sh
own
as Figu
r
e
6
and
w
e
adopt
the recta
ngle t
y
pe in t
h
is
paper.
(a
)
One
di
m
e
nsi
o
nal
l
ear
ni
ng
ra
di
us
(b
) R
ect
an
gl
e t
y
pe l
ear
ni
n
g
ra
di
us
Fi
gu
re
6.
Lear
ni
n
g
t
y
pes
o
f
S
O
M
Fi
gu
re
7.
A
p
pl
y
i
ng S
O
M
al
g
o
ri
t
h
m
From
t
h
e l
o
cal
l
y
sm
oot
he
d i
m
age, we
a
ppl
y
avera
g
e
g
r
ay
val
u
e
o
f
3×
3
m
a
sk sh
o
w
n
a
s
Fi
g
u
re
7
t
o
SOM
l
ear
ni
n
g
and
t
h
e
o
u
t
p
ut
i
s
com
put
ed
b
y
fo
rm
ul
a (3
)
wi
t
h
c
o
n
n
ect
i
o
n st
re
n
g
t
h
s
c
o
nt
r
o
l
l
e
d
by
f
o
r
m
ula
(4
).
i
i
ji
X
W
j
D
2
)
(
)
(
(3)
)
(
1
k
ji
i
k
ji
k
ji
W
x
a
W
W
(4)
whe
r
e
D
de
not
es t
h
e si
m
i
l
a
ri
ty
,
X
d
e
no
tes the p
a
ttern
,
W
is
th
e conn
ection
streng
th
and
α
is th
e learn
i
ng
rate.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
77
–
1
586
1
582
The sam
e
pat
t
e
rns
o
f
3×
3 m
a
sk an
d a
v
e
r
age
gray
va
lues a
r
e
use
d
to
recognize any m
i
nute candidate
cracks
fr
om
the im
age aft
e
r fuzzy
m
e
t
hod sh
ow
n as F
i
gu
re 8. T
h
e basel
i
n
es i
n
F
i
gu
re 8 are re
m
ove
d
afterwa
r
ds.
Fi
gu
re
8.
A
f
t
e
r
ap
pl
y
i
ng
SO
M
2.
4.
Noise
Rem
o
va
l by
Densi
t
y
I
n
fo
rma
tio
n
C
r
acks
ha
ve l
o
wer
de
nsi
t
y
an
d l
o
wer
b
r
i
g
ht
ness,
hi
ghe
r l
e
ngt
h/
wi
dt
h rat
e
t
h
an t
h
at
o
f
ra
nd
om
noi
se
as
sh
o
w
n
i
n
Fi
gu
re 9.
Th
us
, we use fo
rm
ul
a
(
5
)
a
f
t
e
r
a
p
pl
y
i
ng Grass
f
i
r
e
al
go
ri
t
h
m
[21]
.
(a) C
r
ack
(b
) Noise
Fi
gu
re
9.
De
ns
i
t
y
di
ffere
nces
u
x
ob
A
A
A
f
(
5
)
whe
r
e
Ax
,
A
u
den
o
t
e
s t
h
e
wi
dt
h a
n
d hei
ght
of
rect
an
gl
e c
i
rcum
scri
bed
wi
t
h
ob
ject
s e
x
t
r
act
ed
by
Gr
assfi
r
e
alg
o
rith
m
in
resp
ectiv
ely and
A
ob
de
not
es t
h
e
n
u
m
b
er o
f
pi
xels in e
x
tracted object.
The
n
we a
ppl
y
t
h
e
fi
nal
deci
s
i
on
r
u
l
e
sh
o
w
n
as Ta
bl
e
3 a
n
d t
h
e
res
u
l
t
i
s
l
i
ke as s
h
ow
n i
n
Fi
gu
re
1
0
.
The ex
pe
ri
m
e
n
t
al
t
h
resh
ol
d 0
.
3 wa
s o
b
t
a
i
n
e
d
fr
om
pri
o
r
o
b
ser
v
at
i
o
n o
f
10
ran
d
o
m
images n
o
t
use
d
i
n
t
h
i
s
expe
rim
e
nt that the cha
r
acteri
s
tic coefficie
n
t f of
fo
rm
ul
a (5)
has ce
rt
ai
n t
e
nde
ncy
.
Si
nce
crack
s ha
ve
hi
ghe
r
l
e
ngt
h/
wi
dt
h r
a
t
e
, t
h
at
t
e
ndency
wi
t
h
res
p
ect
t
o
the object labelling Grass
f
ire algorithm
can eas
ily be
fo
rm
ulated.
Table 3. Final crack
decisi
on rule
f >= 0.
3
Noise
f < 0.
3
Crack
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Visio
n
-ba
s
ed
C
r
a
ck
Id
en
tifica
tio
n
on
th
e Concrete S
l
ab
Su
rfa
ce
Using
Fu
zzy .... (Kwa
ng
Ba
ek Kim)
1
583
Figure 10. Fina
l
crack
ext
r
action
For the a
n
alysis, we
com
pute
the length,
width, a
nd
t
h
e a
n
gle of t
h
e ide
n
tified c
r
ack
with res
p
ect t
o
t
h
e m
e
t
hod
use
d
i
n
[
2
4]
.
3.
E
X
PERI
M
N
ET AN
D
A
N
A
L
YSIS
The
pr
o
p
o
s
ed
m
e
t
hod
was
i
m
pl
em
ent
e
d w
i
t
h
M
i
cros
oft
Vi
sual
st
udi
o
20
0
8
a
n
d
e
xpe
ri
m
e
nt
s were
p
e
rform
e
d
o
n
IBM-co
m
p
atib
le PC with
In
t
e
l i5
3
.
0
GHz CPU an
d 4GB RAM. Th
irty Dig
ital i
m
ag
es of
conc
ret
e
s
u
r
f
ac
e t
a
ken
wi
t
h
a
C
a
no
n
3
5
0
D
di
gi
t
a
l
cam
era of
8
00×
6
0
0
si
ze
were
use
d
i
n
t
h
e ex
peri
m
e
nt
s.
In
ou
r p
r
e
v
i
o
us
at
t
e
m
p
t
[21]
, t
h
e g
r
ay
val
u
e
was u
s
ed i
n
cr
ack ext
r
act
i
on
as i
s
. Th
us, t
h
e
m
e
t
hod
wa
s
sensitive t
o
the infl
uence
of the
outdoor li
ght
or m
a
y face with environm
ent like low
bri
ght
ness c
ont
rast
betwee
n concrete surface a
n
d crack candida
tes. Howe
ver,
the propose
d
m
e
thod uses R, G, B values a
s
col
o
r
i
n
f
o
rm
at
i
on an
d a
ppl
i
e
s
fuzzy
m
e
t
hod
an
d
S
O
M
i
n
n
o
i
s
e
re
m
oval
.
(a
)
Ori
g
inal I
m
age
(
b
)
Pre
v
ious
1 [
21]
(c
)
Pre
v
ious
2
[24]
(
d
)
Pr
opose
d m
e
thod
Fi
gu
re
1
1
. C
o
m
p
ari
s
on
of
t
h
e p
r
o
p
o
sed
m
e
tho
d
a
n
d
pre
v
i
o
us m
e
t
hods
Th
ere is ano
t
her related
app
r
o
ach with
sligh
tly sd
ff
eren
t
p
u
rp
o
s
es. In
[25
]
, it tries to
reco
gn
ize fi
ve
crack patterns –
horiz
o
ntal,
vertical,
left d
i
ag
on
al,
righ
t d
i
ag
on
al, an
d
und
irection
a
l –
with
b
ack
prop
ag
atio
n
n
e
ur
al n
e
two
r
k in
co
nj
un
ction w
ith
in
ag
e processin
g
t
echni
que
s. Duri
ng t
h
e proces
s, it is supposed to
extract
t
h
e crac
ks
b
u
t
as sh
o
w
n
i
n
Fi
gu
re
1
1
(
d
)
,
t
h
at
m
e
t
hod i
s
es
peci
al
l
y
weak
f
o
r
“u
n
d
i
r
ect
i
o
n
a
l
”
cracks
i
n
t
h
at
t
h
e
syste
m
is p
r
one to
recogn
ize
false po
sitiv
e
no
is
es as crack
s
d
u
r t
o
low i
n
ten
s
ity con
t
rast.
We believe t
h
at such im
provem
ent gives
us m
o
re accurate crack ide
n
tification res
u
lt shown a
s
Figure 12 as a com
p
arative exam
pl
e in th
at the propose
d m
e
thod is m
o
re accurate and disc
rim
i
native in
ex
tracting
m
i
n
u
t
e crack
s
.
Howev
e
r, i
f
th
e co
n
c
rete su
rf
ace h
a
s
relativ
ely lo
n
g
(>1
c
m
)
furrows or
filth
s, t
h
e
pr
o
pose
d
m
e
t
hod
fai
l
s
t
o
ext
r
act
cr
acks
c
o
rrectly as shown
in Figure
12.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
77
–
1
586
1
584
Figure 12. False
identification
exam
ple
Fi
gu
re
1
3
dem
onst
r
at
es t
h
e
l
a
bel
l
e
d c
r
ac
k
re
sul
t
i
n
orde
r t
o
analyze s
o
m
e
characte
r
istics of crac
ks.
Suc
h
c
h
aract
e
r
i
s
t
i
c
s - wi
dt
h,
l
e
ngt
h,
di
rect
i
o
n
- are
s
u
m
m
a
ri
zed i
n
Tabl
e
4
.
(a)
O
r
iginal im
age
(
b
) C
r
ac
k ide
n
tified
Fig
u
re
13
. C
r
ack
id
en
tification
ex
am
p
l
e
Table
4. C
r
ack cha
r
acteristics for
Figure
17
Crack
#
Leng
th(c
m)
Width(
c
m
)
Angle(
˚
)
Crack
#
Leng
th(c
m)
Width(
c
m
)
Angle(
˚
)
1
11.
127
0.
276
24.
23
14
4.
533
0.
409
48.
01
2
14.
654
0.
743
56.
25
15
1.
919
0.
231
40.
49
3
3.
723
0.
323
12.
97
16
3.
954
0.
212
49.
57
4
6.
211
0.
319
20.
21
17
2.
910
0.
301
50.
74
5
4.
214
0.
218
22.
88
18
4.
507
0.
499
39.
23
6
6.
035
0.
349
56.
08
19
1.
618
0.
248
31.
94
7
3.
727
0.
228
64.
54
20
2.
163
0.
229
23.
82
8
3.
651
0.
361
53.
13
21
4.
319
0.
369
63.
54
9
3.
286
0.
249
46.
54
22
3.
158
0.
368
25.
55
10
3.
722
0.
408
37.
20
23
3.
009
0.
320
40.
42
11
5.
952
0.
410
49.
25
24
4.
449
0.
425
57.
26
12
4.
654
0.
336
61.
86
25
2.
998
0.
521
40.
90
13
3.
873
0.
409
25.
03
26
1.
989
0.
229
28.
41
Fo
r all 30
i
m
ag
es u
s
ed
in
th
i
s
ex
p
e
rim
e
n
t
,
we id
entifie
d 97.3% of crac
ks
as summarize
d in Table
5
whe
r
e t
h
e
gr
ou
nd
t
r
ut
h c
r
ack
s
are
veri
fi
ed
by
fi
el
d e
n
gi
nee
r
.
Tab
l
e
5
.
C
r
ack id
en
tification
statistics
I
m
ages
# of
Crack
s
Identifie
d
Accuracy
30
1319
1283
97.
3%
Fro
m
literatu
re rev
i
ew, th
is resu
lt is b
e
tter th
an
u
s
ing
o
t
h
e
r variou
s imag
e pro
cessing tech
n
i
qu
es
[1
5]
whi
c
h re
po
rt
ed t
h
e e
r
r
o
r i
d
e
n
t
i
f
i
cat
i
o
n rat
e
o
f
3
.
5
6
-
8
.
9
5
%
fr
om
di
ffe
rent
i
m
age
set
usi
n
g 5
di
f
f
ere
n
t
im
age processi
ng techniques.
4.
CO
NCL
USI
O
N
In t
h
i
s
pa
per,
we p
r
o
p
o
se a
new m
e
t
hod t
o
ext
r
act
an
d a
n
al
y
ze cracks o
n
co
nc
ret
e
sl
ab
st
ruct
u
r
e
by
in
tellig
en
t i
m
a
g
e pro
c
essing
tech
n
i
q
u
e
s.
While p
r
ev
i
o
u
s
m
e
th
od
s use gray v
a
lu
e of th
e i
m
ag
e d
i
rectly, we u
s
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Visio
n
-ba
s
ed
C
r
a
ck
Id
en
tifica
tio
n
on
th
e Concrete S
l
ab
Su
rfa
ce
Using
Fu
zzy .... (Kwa
ng
Ba
ek Kim)
1
585
R, G, B cha
nnel values as c
o
lor inform
ation a
n
d ap
ply fuzzy reas
oni
ng and SO
M alg
o
rith
m
in
ex
tractin
g
candi
date crac
ks a
n
d rem
oving noises.
Wit
h
those ca
reful treatm
e
nts, we ca
n s
u
cces
sfully extra
c
t minute
cracks w
h
i
c
h were oft
e
n
i
g
n
o
re
d
i
n
pre
v
i
o
us
st
u
d
i
e
s.
T
h
en som
e
chara
c
teristics of cracks s
u
ch as l
e
ngt
h,
wi
dt
h
,
a
n
d
di
re
ct
i
on c
oul
d
be
easi
l
y
anal
y
zed.
Wh
ile t
h
e aim of th
is research
was con
f
i
n
ed
to
e
x
tract cracks acc
urately and c
o
m
putes som
e
basic
ch
aracteristics o
f
cracks, we ex
p
ect th
at m
o
re in
tellig
en
t and
u
s
efu
l
to
o
l
s t
h
at can
an
alyze th
e prog
ressio
n
o
f
crack
s and
d
i
rectio
n
s
o
f
crackin
g
fo
r m
o
re i
n
tellig
en
t m
a
in
ten
a
n
c
e of con
c
rete stru
ctu
r
e i
n
fu
t
u
re
research
.
REFERE
NC
ES
[1]
R. Rajn
i,
et al.
,
“Complementary
Split Ring
Res
onato
r b
a
sed Sensor for Crack
Detection
,
”
Intern
ational Journal of
Electrica
l
and
C
o
mputer Engin
e
ering
, vo
l/issue:
5(5), pp
. 1012-1
017, 2015
.
[2]
J.
P.
Sa
wy
e
r
,
et al.
, “Structural
damage detectio
n and
identification using fuzzy
logic,”
AIAA jo
urnal
, vol/issue:
38(12), pp
. 2328
-2335, 2000
.
[3]
N. Bakhar
y
,
et
al
.
, “
S
truc
ture d
a
m
a
ge det
ect
ion
using neural
net
w
ork with m
u
lti-
stage substructur
i
ng,”
Advan
ces
i
n
Structural Engin
eering
, vol/issue: 13(1), pp. 95-1
10, 2010
.
[4]
P
.
Been
a,
et al.
,
“
S
tructural
da
m
a
ge det
e
c
tion
using fuzz
y
cog
n
itive
m
a
ps and
Hebbian
l
earni
ng,”
Applied
S
o
ft
Computing,
vo
l/issue: 11(1), pp.
1014-1020, 201
1.
[5]
D.
G.
Aggelis,
et al.
, “
NDT ap
proach for
char
act
eriz
ation
of s
ubs
urface
crack
s
in concre
te
,”
Construction an
d
Building
Mat
e
ri
als
, vol/issue: 25
(7), pp
. 3089-30
97, 2011
.
[6]
D.
A.
Tibaduiza,
et al.
, “Damage classification in
stru
ctural health monitoring using princip
a
l co
mponent analy
s
is
and self-org
anizing maps,”
S
t
ructural Control and
Health
Monitoring
, vol/issue: 20
(10), pp
. 1303-1
316, 2013
.
[7]
R. P. Day
a
l,
et
a
l
.
, “Analy
sis of s
m
art crack detection
methodolog
ies in v
a
rious str
u
ctures,”
Journa
l of Eng
i
neering
and Technolog
y
Research
, vol/is
sue: 3(5), pp
. 13
9-147, 2011
.
[8]
B. Gunes,
et a
l
.
, “
S
tructu
ral
Heal
th Monitor
i
n
g
and Dam
a
ge
A
sse
ss
me
nt Pa
rt II: Appl
i
c
a
tion
of the
Dam
a
ge
Locating Vector
(DLV) Method
to th
e ASCE
B
e
nchmark Structure Ex
pe
ri
me
ntal
Da
ta
,”
International Journal
of
Physica
l
Sc
ien
c
e
s
, vol/issue: 7(9)
, pp
. 1509-1515
, 2012.
[9]
Z. Chen
,
et al.
,
“Image-based framework for concrete su
rfac
e
cr
ack m
onitoring
and quantification,”
Ad
vances in
Civil
Eng
i
neer
in
g
, pp
. 1-18
, 201
0.
[10]
M. Dilena
,
et al.
, “Dy
n
amic testing of a damaged bridge,”
Mech
anical Systems and Signal Processing
, vol/issue
:
25(1), pp
. 1485-
1507, 2011
.
[11]
K.
B.
Kim,
et
a
l
.
, “Recognition
of con
c
rete sur
f
ace cr
ack
s usin
g the ART1-b
as
ed RBF networ
k,”
Advances in
Neural Networks-ISNN 2006, Springer Berlin H
e
idelberg
, pp. 669
-675, 2006
.
[12]
I. Abdel-Q
a
der
,
et al.
, “Analy
s
i
s of edg
e
-detection techn
i
ques
for
cra
c
k ident
i
f
ica
tion in
brid
ges,”
Journal o
f
Computing in
Civil Engin
eering
, vol/issue: 17(4), pp.
255-263
,
20
03.
[13]
T. Nis
h
ik
awa,
et al.
, “
C
oncre
te
c
r
ack d
e
te
ct
ion b
y
m
u
lt
iple
seque
ntial
im
age
fil
t
er
ing,”
Computer
Aided
Civil an
d
Infrastructure Engineering
, vol/issue: 27(1), pp.
29-47, 2012
.
[14]
Y. Fujita
,
et a
l
.
, “A robust automatic cr
ack d
e
tection
method fr
om noisy
c
oncr
e
te surfaces,”
Ma
chine Vision
an
d
Applica
tions
, vo
l/issue: 22
(2), pp
. 245-254
, 2011
.
[15]
T. Yamaguch
i
,
et al.
, “
F
as
t cr
ack
dete
ct
ion m
e
tho
d
for larg
e-s
i
z
e
concre
te s
u
rfa
ce
im
ages
us
ing p
e
rcol
ation-b
a
s
e
d
im
age proc
es
s
i
n
g
,”
Ma
chin
e
Vision and App
lica
t
ions
, vol/issue: 2
1
(5), pp
. 797-80
9, 2010
.
[16]
G
.
Li
,
et al.
, “Image-based
Meth
od for Concr
e
te
Bridge Crack Detection,”
Journal of Information
&
C
o
mputational
Scien
c
e
, vol/issue: 10(8)
, pp
. 222
9-2236, 2013
.
[17]
A. F
a
rhidzad
eh,
et al.
, “A vision-based techn
i
q
u
e fo
r da
ma
ge
a
sse
ssme
n
t of r
e
info
rced concr
e
te structures,”
In
SPIE Smart Stru
ctures and Materials+
Nondestructive Evaluatio
n and Health Mo
nitoring, In
terna
tional Society fo
r
Optics and
Photonics
, pp
. 90642
H-90642H, 2014.
[18]
P.
Pra
s
a
nna
,
et al.
, “Computer-vision based
crack
d
e
te
ct
ion
and ana
l
y
s
is,”
In SPIE Smart Structures an
d
Materials+ Nondestructive Eva
l
uation and Health Monitoring
(pp. 834542-834542). International Society
fo
r
Optics and
Photonics, 2012
.
[19]
B. Y.
Lee,
et al.
, “Surface Cr
ack Evaluation M
e
t
hod in
Concr
e
te Stru
ctures,”
Journal of th
e K
o
rean Society fo
r
Nondestructive Testing
, pp
. 173-
182, 2007
.
[20]
Y. W. Woo,
et al.
, “
F
CM
-bas
e
d
RBF
network in iden
tif
ying
concre
te s
l
ab s
u
rface
cra
c
ks
,”
I
n
tell
igent
Syste
m
s
Design and App
l
ications
, ISDA’0
8. Eighth
Intern
ational Con
f
eren
ce on
. I
E
EE
, vol. 1
,
pp
. 664-669
, 2008.
[21]
K.
B.
Kim,
et al.
, “
A
nal
y
s
i
s
o
f
Concret
e
S
l
ab
S
u
rface Cra
c
ks
Us
ing ART2-bas
ed H
y
br
id Ne
ural Network,
”
Information-An I
n
ternational
Interdisciplinary Jo
urnal
, vol/issue:
16(1B), pp
. 777-
782, 2013
.
[22]
V. Godbole, “Performance
analy
s
is of
clusterin
g
protocol using
fuzzy
log
i
c for
wireless sensor network,”
IAES
International Jo
urnal of
Artif
icia
l Intelligence (
I
J-AI)
, vol/issue: 1
1
(3), pp
. 103-11
1, 2012
.
[23]
T. Kohonen
,
et al.
, “Engineerin
g applications o
f
the self-org
anizing map,”
Pr
oceed
ings
of the
IEEE
, vol/issue:
84(10), pp
. 1358
-1384, 1996
.
[24]
K.
B.
Kim,
et
a
l
.
, “
Vis
i
on bas
e
d Crack
Ident
i
fi
cat
ion and
Anal
ys
is
on
the
S
u
rface
of Con
c
ret
e
S
l
ab S
t
ru
ctures
,”
Information-An I
n
ternational
Interdisciplinary Jo
urnal
, vol/issue:
18(6A), pp. 238
1-2386, 2015
.
[25]
B. Y. Lee,
et a
l
.
, “A Techniqu
e for Pattern R
ecogni
tion of Concrete Surface Cracks,”
Journ
a
l of the Korea
Concrete
Institu
t
e
, vo
l/issue:17(3
)
, pp
. 369-374
, 2
006.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
15
77
–
1
586
1
586
BIOGRAP
HI
ES OF
AUTH
ORS
Kwang Baek Ki
m
.
Kwang Baek
Kim
receiv
e
d hi
s
M
.
S
.
and P
h
.D. degr
ees
from
t
h
e Departm
e
n
t
of Computer Science, Pusan Na
tional University
,
Busan, Korea,
in
1993 and 1999, respectiv
ely
.
From 1997 to the present, he
is a professor at
th
e Department o
f
Comput
er Engineering
,
Silla
Univers
i
t
y
, Kore
a. He is
curren
t
l
y
an as
s
o
c
i
at
e ed
itor for J
ournal o
f
Intell
igen
ce an
d Inform
ation
S
y
s
t
em
s
and The Open Com
puter S
c
ienc
e J
ournal (US
A
). His
res
earch in
ter
e
s
t
s
include fu
z
z
y
neural network
and appl
ications, bioinformatics
,
and image proces
sing.
H
y
un Jun Park
.
He receiv
ed his
M.S. degr
ees fro
m the Dep
a
rtment of Computer
Science, Pusan
National Univer
sity
, Busan, Ko
rea,
in 2009. Fr
om
2009 to the present, h
e
is a Ph.D. course
student at th
e
Department of
Computer Engineering
,
Pusan
National Univer
sity
, Korea. His
res
earch
in
teres
t
s
includ
e
com
puter v
i
s
i
on,
im
ag
e processing, neur
al n
e
twork
and
applications.
Doo Heon Song. Doo Heon Song received
a B
.
S.
de
gree in Statistics & Computer
Science from
S
e
oul Nationa
l
Univers
i
t
y
and
M
.
S
.
degree
Com
puter S
c
ien
ce from
the Korea Advanc
ed
Institute of Sci
e
nce and T
echno
log
y
in 1983
. He
received his Ph.D. Cert
ificat
e in Com
puter
Science from the University
of California in 19
94. Form 1983-
1986, he was a research scientist
at the Kore
a Ins
t
itute of S
c
i
e
nc
e
and Techno
log
y
.
He has been a
professor at the
Department of
Computer Games, Yong-in Songdam College, Korea,
sin
ce 1997
. His research
in
terests include
ITS, m
achine
le
arning,
arti
fic
i
al
intel
lig
ence
,
medical image pr
ocessing,
cogn
itive,
and game
intel
ligen
ce
.
Evaluation Warning : The document was created with Spire.PDF for Python.