TELKOM
NIKA
, Vol. 11, No. 8, August 2013, pp. 42
4
4
~4
250
e-ISSN: 2087
-278X
4244
Re
cei
v
ed Fe
brua
ry 5, 201
3; Revi
se
d
May 9, 201
3; Accepted Ma
y
17, 2013
Removal of Atmospheric Particles in Poor Visibility
Outdoor Images
Yaseen
Al-Z
ubaid
y
*
, Ros
a
lina Abdul
Salam
F
a
cult
y
of Scie
nce an
d T
e
chnolo
g
y
Unvers
iti
Sains
Islam M
a
la
ysi
a
(USIM), Negeri Sem
b
i
l
an, Mal
a
ysia
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:
y
a
s
een
_d
a
w
od@
hotmai
l
.co
m
*, rosali
na@
usim.ed
u
.m
y
A
b
st
r
a
ct
T
he visi
bil
i
ty of
a sce
ne is
de
grad
ed
by w
e
a
t
her ph
en
o
m
e
n
a
such
as r
a
in
dri
z
z
l
e, fo
g a
n
d
ha
z
e
.
T
he degr
ad
atio
n of imag
e sce
ne is d
ue
the s
ubstanti
a
l pr
es
ence
of p
a
rticle
s in the at
mosp
here that scatt
er
a
n
d
ab
so
rb l
i
gh
t. As th
e l
i
g
h
t
sp
re
ad
s from
ob
j
e
ct to th
e ob
se
rve
r
, th
e col
o
r a
n
d
in
te
n
s
i
t
y is ch
an
ge
d
b
y
th
e
atmos
p
h
e
ric
p
a
rticles. In
this
rese
arch, w
e
sugg
est n
e
w
meth
ods
to
pre
c
isely
detect
ai
rlight
an
d c
o
rre
ctl
y
estimate th
e at
mos
p
h
e
ric ve
il
from i
m
ag
e tha
t
capture
d
in
b
ad w
eath
e
r. T
he resu
lt of su
g
gested
metho
d
s
w
ill be use
d
in
scattering at
mosph
e
ric
mod
e
l
to remove
atmosph
e
ric partic
l
es na
me
ly, rai
n
dri
z
z
l
e, fog a
n
d
ha
z
e
fro
m
a si
ngl
e i
m
ag
e. Therefore a h
i
g
h
e
r
visibil
i
ty imag
e w
ill be pro
d
u
c
ed.
Ke
y
w
ords
: atmos
p
h
e
ric part
i
cles; airl
ig
ht; scattering at
mo
spher
ic mode
l
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The ima
g
e
s
captu
r
ed i
n
o
u
tdoor
scen
e
s
are
u
s
ually degrade
d
by the
turbid me
dium
in
the atmosphe
re. Bad we
ather
su
ch a
s
fog and h
a
ze
redu
ce the vi
sibility and color fidelity, and
the pa
rticle
s
in atmo
sphe
re ca
use ab
sorption
and
scatterin
g
. Th
e irradia
n
ce
received by t
he
came
ra from
the scene p
o
int is attenu
ated alon
g th
e line of sig
h
t. Furtherm
o
re, the incomi
ng
light is blend
ed with the airlight [1] (ambient light
reflected into th
e line of sight by atmosph
e
ri
c
particl
es). Th
e deg
ra
ded i
m
age
s lo
se t
he
contrast
a
nd
colo
r fideli
t
y of the scen
e. In additio
n
, the
amount of de
grad
ation de
p
end
s on the d
i
stan
ce
s of the scene p
o
int
s
from the ca
mera.
Removin
g
rai
n
dri
zzl
e, fog and ha
ze
can
signifi
cantly increa
se the v
i
sibility of the scene
and co
rrect
t
he colo
r shift cau
s
e
d
by
th
e
airli
ght.
Mo
reover,
rain
drizzl
e, fog
and
ha
ze
rem
o
va
l is
critical for a
wide
range of image-rel
a
ted
applications, such as su
rveillance sy
stem
s, intelligent
vehicle
s
, sate
llite imaging, and outd
oor
obje
c
t recogn
ition system
s.
Bad visi
bility in poor weather
i
s
the m
o
st important problem
for different
applicat
ions of
comp
uter visi
on. Most aut
omatic sy
ste
m
s for
monit
o
ring, intellig
ent vehicle
s
,
outdoor o
b
j
e
ct
recognitio
n
, a
s
sume th
at the inp
u
t ima
ges have
cl
e
a
r visi
bility [2]. Unfortun
ately, this do
es no
t
happ
en all th
e time, therefore imp
r
oving
visibility is an inevitable task.
In comp
uter
vision, the at
mosp
he
ric
scattering m
o
d
e
l is u
s
u
a
lly use
d
to de
scribe the
formation
of a foggy or h
a
zy image. Almost all e
s
tabli
s
he
d metho
d
s
are ba
se
d o
n
this mo
del [
3
].
R
e
fe
re
nc
e [4
] r
e
s
t
o
r
e
s
c
o
ntr
a
s
t
o
f
image
sc
en
e
b
y
us
in
g tw
o or
mo
r
e
imag
e
s
ta
k
e
n
in un
ifo
r
m
bad weathe
r.
Ho
wever, thi
s
method requi
res
ch
angi
ng
weath
e
r
con
d
i
tions to capt
ure the im
age
s
that will be
used in
cont
rast
restoration. A method proposed
by [5] uses
the polarization
approa
ch to
enha
nce the
visibility. It i
s
ba
se
d on
the fact that
the scattere
d of airlig
ht by
su
spe
nde
d atmosp
he
ric pa
rticle
s is pa
rtially polar
ized
. In
this method, two imag
es or mo
re a
r
e
taken
thro
ug
h a p
o
lari
ze
r at differe
nt orientat
io
ns. The captu
r
ed
image
s are
analyzed,
taking
into accou
n
t the effects
of the pola
r
i
z
ation of
at
mosp
he
ric
scatterin
g
. Thi
s
process
wi
ll be
inverted to
re
cover the
ima
ge from the
e
ffect of
ha
ze.
Polari
zation
-b
ase
d
a
pproa
ch requi
re
s two
or more imag
es; it is not su
fficient to be applie
d in de
nse fog o
r
ha
ze weathe
r.
Another meth
od p
r
op
osed
by [6] re
covers the
scen
e o
f
singl
e ima
g
e
by u
s
ing
ad
ditional
informatio
n provide
d
intera
ctivity by use
r
. Th
e
use
r
provides a
ddition
al input su
ch a
s
approximatel
y dire
ction
of
scen
e d
epth,
or a
goo
d
re
gion of
colo
r fidelity.
This method
re
qui
res
addition
al informatio
n fro
m
user. Also
the resu
lt wi
ll be varied base
d
on the area that will
be
cho
s
e
n
by the use
r
. Reference [7] use
s
3D-
geom
e
t
rical mod
e
l to determin
e
the depth of the
scene
an
d
recove
r
scen
e alb
edo. It
requi
re
s
an i
n
tera
ctive re
gistratio
n
pro
c
e
s
s to
align
the
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Rem
o
val of A
t
m
o
spheri
c
Particle
s in Po
or Visibilit
y O
u
tdoor Im
age
s (Ya
s
ee
n Al-Zubaid
y
)
4245
image
with
g
eometri
c
mod
e
ls
of the
world such a
s
terrain
or buil
d
in
gs. By
regi
ste
r
ing
the im
ag
e
to these mod
e
ls, depth be
come
s availa
ble at eac
h p
i
xel. Howeve
r, the main disadva
n
tage o
f
this app
roa
c
h
is unavaila
bility of 3D model for all imag
e locatio
n
s.
Gene
rally, all previou
s
method
s req
u
ired a
dditio
nal informati
on to recover scen
e
albed
o. In practical ap
plications, it i
s
di
fficult
to use these techniq
ues,
so
su
ch
approa
che
s
a
r
e
rest
ricte
d
. Ne
w enh
an
cem
ent visibility method
s are
able to re
co
ver scen
e al
bedo by ma
king
variou
s assu
mptions a
bou
t airlight or co
lors in the
scene.
In this
re
sea
r
ch
we
sug
g
e
st a
ne
w
method to
e
s
timate the
dire
ct attenu
ation a
n
d
airlight. T
he
proposed technique
w
ill be used to
restore the visi
b
ility of the i
m
age that i
s
to
remove th
e a
t
mosph
e
ri
c p
a
rticle
s. Th
ere are th
re
e steps involve
d
:
detecting th
e set
s
of whi
t
e
area
to e
s
tim
a
te the
skylight, usin
g two
-
step
of e
s
ti
mations to in
feren
c
e th
e a
t
mosph
e
ri
c v
e
il,
and re
cove
rin
g
the image b
y
using the previous
e
s
tima
tions in atmo
sph
e
ri
c scattering m
odel.
2. Backg
rou
nd
In co
mpute
r
vision th
e
atmosp
he
ric scattering
mod
e
l is
widely
use
d
to
de
scribe
the
formation of a
fog and ha
ze
image as foll
ows [3]:
I(x
)
= Ap(x
) e
−
β
d(x
)
+ A
(
1
−
e
−
β
d(x
)
)
(1)
Whe
r
e I(x) is the obse
r
ve
d image inte
nsity, A
is th
e global atm
o
sp
heri
c
light
(skylig
ht),
and p
(
x) i
s
scene
albed
o, d(x) i
s
a
scene
depth
and
β
i
s
th
e scatterin
g
coeffici
ent of
the
atmosp
he
re. The first term
Ap(x
) e
−β
d(x) on the righ
t hand of the E
quation (1
) i
s
calle
d dire
ct
attenuation, a
nd the secon
d
term A(1
−
e
−β
d
(
x)) i
s
called airli
ght. To re
cove
r fog and h
a
ze f
r
om
attenuation a
nd infidelity color, we n
eed
to find the va
lue of A,
β
, a
nd d(x) in (1).
Re
cent
work
on si
ngle im
a
ge to recover scene
albe
d
o
impo
se
s li
mitations u
p
o
n
either
the airlight or
dire
ct attenua
tion, or on bo
th.
Tan method [2] is base
d
on two ba
si
c ob
servatio
n
s
:
first, the images that ca
ptured in
clea
r-day have
m
o
re contra
st (enh
an
ced vi
sibility) than
the
image
captured in
b
ad
we
ather;
se
co
n
d
, the
air
lig
ht in im
age
scene i
s
varia
n
c
e
and
p
r
ima
r
ily
depe
nd
on di
stan
ce
betwe
en the
obj
ect
and
the vie
w
er. G
ene
rally, there
a
r
e t
w
o di
sadva
n
ta
ge
of this
method, Firs
t, it doesn’t fully rec
o
ver the
scen
e’s ori
g
in
al colors
or
albed
o, it is
just
enha
nce the contrast of a
n
input image
; seco
nd, it
doesn’t kno
w
the actu
al value of airlight,
so
the output image tend to h
a
ve large
r
saturation tha
n
those in cle
a
r
day image
s.
Fattal [8] relie
s o
n
the
assu
mption that th
e tran
smi
s
sio
n
and
surfa
c
e
sh
ading
are l
o
cally
uncorrelate
d. He e
s
timat
e
s the o
p
tical tran
sm
issi
on in ha
zy
singl
e imag
e
.
Base on t
h
is
estimation, th
e scattered li
ght is
remove
d to enh
an
ce
the visibility of the scen
e
and recover t
he
low contrast i
n
the input i
m
age. Howe
ver, perfo
rma
n
ce of thi
s
m
e
thod g
r
eatly
depen
ds o
n
th
e
quality of the input d
a
ta. On the
oth
e
r
words,
thi
s
meth
od re
quire
s
signifi
cant info
rmat
ion
(varia
nce) in
colo
rs of the i
nput image to
work comple
tely.
Tarel
an
d Huati [9] present fast
algo
rithm
to
re
co
ver sce
ne
al
bedo
for
sin
g
le in
put
image. They
assume the
small obj
ect i
n
foggy or
h
a
zy we
ather
has lo
w saturation col
o
r. T
h
is
method
ba
se
d on
medi
an
filter. It is al
so
pro
p
o
s
e
n
e
w filter to p
r
eserve
ed
ge
and
corn
er t
o
enha
nce the
visibility restoration. Thi
s
met
hod a
ssumes th
at the atmosphe
ri
c veil mu
st be
smoothi
ng all
the time, and this assumpti
on will
maxim
i
ze the atmo
spheri
c
veil of the image.
He et al [10] pre
s
ent
s sim
p
le and effe
ctive
method to remove the fog from si
ngl
e image.
They depe
nd
on kind of statistics of outdoo
r imag
es that is free of haz
e.
Re
sult of these
statistics i
s
t
he mo
st lo
ca
l patche
s
in
haze-
free
out
door imag
es co
ntain
som
e
pixel
s
(call
e
d
“da
r
k pixel
s
”) which h
a
ve very low intensity in
at le
ast one color chann
el. According to thi
s
method, the intensity of these da
rk pixels in t
hat ch
annel is mai
n
ly cont
ribute
d
by the airlight.
Therefore, th
ese
da
rk pixe
ls
can
directly
provid
e
acc
u
r
a
te
es
tima
tion
o
f
th
e h
a
z
e
’s
tr
an
smiss
i
on
.
Usi
ng thi
s
pri
o
r info
rmation
with the h
a
ze imagin
g
mo
del, they re
co
ver input im
a
ge from
effect
of
haze and p
r
o
duce a goo
d
depth ma
p. Ho
wever, thi
s
metho
d
is i
n
valid wh
en the scen
e obj
ect
inherently
si
milar to
the
at
mosp
he
ric lig
ht over a
larg
e lo
cal
re
gion
and
n
o
sha
d
o
w i
s
cast
on
th
e
obje
c
t.
The meth
od
propo
se
d in
[3] a
s
sume
s that the
recovered
ima
g
e
sce
n
e
albe
do h
a
s
highe
r color
contrast a
nd
the depth ma
p tends to
b
e
all most sm
o
o
th except al
ong ed
ge
s wi
th
large d
epth ju
mps. It performs white b
a
l
ance and
sim
p
lifies the sca
ttering atmo
spheri
c
mo
del to
remove th
e fog an
d en
ha
nce th
e visibil
i
ty for si
ngle i
m
age. In thi
s
method, the
atmosp
he
ric
veil
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 8, August 2013: 4244 –
4250
4246
infers by u
s
in
g two step
s. First, define t
he rou
ghly e
s
timation of the atmosphe
ric veil by usi
n
g
minimal
com
pone
nt of the col
o
r-corre
c
ted ima
ge.
Secon
d
, refi
ne the
coa
r
ser e
s
timate
of
atmosp
he
ric
veil by u
s
ing
wei
ghted
le
ast
squ
a
re
s
(WLS) o
p
timi
zation
fram
e
w
ork [1
1]. T
h
is
method
refin
e
s the
co
arse atmosphe
ri
c veil ba
se
d
on WLS sm
o
o
thing. Thi
s
refining
ca
uses
some di
sto
r
tion in estimati
on of the atmosp
heri
c
veil.
The existing
literature sho
w
s t
hat there
are still pro
b
lems in det
ection of airli
ght, and
there is n
o
preci
s
e inferen
c
e to the atm
o
sp
her
i
c
veil that can hel
p
to recove
r the scene al
be
do
corre
c
tly an
d
to re
move f
og imp
a
ct f
r
om the
sin
g
l
e
imag
e. In
orde
r to
add
ress th
e a
b
o
v
e
mentione
d p
r
oblem
s, we d
e
velop
a ne
w tech
niqu
e to
estimate
ai
rlight an
d di
re
ct attenuation
to
recover the
scen
e from the
effect of rain drizzle, fog a
nd ha
ze.
3. Rese
arch
Metho
dolog
y
In ord
e
r to
ad
dre
s
s the p
r
o
b
lem of the
removal of
rai
n
dri
z
zle, fog
and h
a
ze fro
m
sin
g
le
image
and
a
c
hieve th
e re
sea
r
ch o
b
je
ctives, a
syst
e
m
atic meth
od
ology ph
ases are
define
d
as
s
h
ow
n
in
F
i
gu
r
e
1
Figure 1. Re
search Meth
od
ology
3.1. Problem Identifica
tio
n
In this pha
se,
the probl
em of the remov
a
l of
rain dri
z
zle, fog and h
a
ze i
s
investi
gated in
orde
r to imp
r
ove the visibil
i
ty of rain dri
z
zly,
foggy an
d ha
zy image
s by ad
dre
s
si
ng the p
r
obl
e
m
of low co
ntra
st and color i
n
fidelity.
3.2. Image Acquisition
In this phase,
we will use experim
ental set-u
p
ca
mera to obtain the rain dri
z
zly, foggy
and ha
zy ima
ges from different lo
cation
s and with va
ri
ous level
s
of visibility in order to ap
ply the
proposed met
hods
to rem
o
ve
the
effect of
poor weathe
r. The
captured i
m
ages
will provide input
to the suggested system
and the system will process the im
ages
to improve the visibility.
3.3. Analy
s
is
of Curr
ent T
echnique
s
In this pha
se,
the current
method
s and
techni
que
s re
lated with im
age en
han
ce
ment are
to be investig
ated by focu
sing on scatte
ring atmo
sphe
ric a
pproa
che
s
. By
examining the cu
rren
t
literature, this pha
se aim
s
to identify the limitat
ion of
the existing
method
s in o
r
de
r to ad
dre
s
s
these limitatio
ns in the prop
ose
d
method.
Anal
y
s
is of Curre
nt
Techni
q
ues
Proposed
A
pp
roach
Problem
Identification
Tes
t
i
n
g
Implementation
Ima
g
e
Evaluation
Image Acquisitio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Rem
o
val of A
t
m
o
spheri
c
Particle
s in Po
or Visibilit
y O
u
tdoor Im
age
s (Ya
s
ee
n Al-Zubaid
y
)
4247
3.4. Propose
d
Appro
ach
As a
re
sult
of investig
ating the
existi
ng m
e
thod
s
and te
ch
niqu
es
examine
d
in the
previou
s
pha
se, thi
s
pha
se aim
s
to
p
r
opo
se
metho
d
s to
a
ddress the
rest
ricti
ons in
existi
ng
method
s and
particula
rly to meet the aim and obje
c
tives of the re
se
arch.
3.5. Implementa
tion
This pha
se will
impleme
n
t the
propo
se
d
met
hod
s m
e
ntioned i
n
p
r
e
v
ious p
h
a
s
e
by usin
g
obje
c
t ori
ent
ed p
r
og
ramm
ing lan
gua
ge
su
ch
as Java. To me
et the aim
and
obje
c
tives of
the
research, different
algorithms
will
be i
m
plement
ed
to rem
o
ve the rain
dri
zzl
e, fog and
haze
effects an
d e
nhan
ce the vi
sibility of image.
3.6. Testing
In this phase
we will test the perfo
rma
n
c
e of
the sug
geste
d syste
m
. The testing of the
system
will
b
e
ba
se
d on
quality of out
put imag
e.
B
a
se
d o
n
such testin
g, the
system
will
be
further imp
r
o
v
ed if the output image
s wi
ll not
be prod
ucin
g the sati
sfacto
ry re
sul
t
s.
3.7. Ev
aluation
In this
pha
se
, the results
of the
sug
g
e
s
ted
method
s are
to
be
evaluated
comp
ared
to
the existing method
s in order to examin
e the qua
lity of output image. To
achi
eve the evaluation,
two different
methods will
be used as follows:
3.7.1. Qualitativ
e Method
the huma
n
p
e
rception
met
hod
will be u
s
ed to
com
p
are a
m
on
g th
e input ima
g
e
, output
image, a
nd th
e imag
e capt
ured
in
clea
r
weath
e
r fo
r t
he same
sce
ne. In this
me
thod we will
u
s
e
different ima
g
e
s
with different levels
of visib
ility to exam the qu
ality of images
b
a
se
d on
hum
an
perception. To achi
eve thi
s
method, few num
bers
of
users
will be invited to examine the quality
of the outp
u
t image
and
write their note
s
on
the q
u
e
s
tionn
aire
s th
at distri
buted
to them for t
h
is
purp
o
se.
3.7.2. Statisti
cal method
In this method, we will
use a histogram t
o
evaluate the images bef
or
e and after
visibility
enha
ncement
. The
sug
g
e
s
ted meth
od
wi
ll be al
so
co
mpared
with t
he e
s
tabli
s
he
d metho
d
s (i.
e
.,
Gray World a
nd Hi
stogram
Equalizatio
n) as
well as
the lates
t
res
e
arc
h
methods
.
4. Visibilit
y
Enhancement
To re
move ra
in dri
zzl
e, fog
and h
a
ze fro
m
input ima
g
e
, we u
s
e
atmosp
he
ric
scattering
model. Ou
r prop
osed me
thod con
s
i
s
ts of three ste
p
s: estimatio
n
of the value of skylight
A,
inferen
c
e
at
mosp
he
ric ve
il from o
b
served imag
e
I(x), and u
s
e
the outp
u
t of
previou
s
step
s to
recover
scen
e albed
o p(x)
of the input image.
The
s
e
step
s are d
e
fined a
s
sh
own in Figure 2
Figure 2. Visibility Enhancement Steps
Estimating Sky
light
Estimating Atmospheric
Veil
Captured
Images
Recover the scene albedo
(removing of atm
o
spheric
particles)
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 8, August 2013: 4244 –
4250
4248
4.1. Estimating Sk
y
light
Most of p
r
evious
study e
s
t
i
mates the
skylight
A from high inten
s
ity pixel in the i
m
age.
But this ap
proach is i
n
a
c
curate, b
e
cau
s
e the
hi
gh i
n
tensity pixel
may not be
the skylight, but
rathe
r
the
co
lor of
an
obj
ect in
the
scene. To
solve this proble
m
, we
sugge
st a th
re
e-step
algorith
m
to find the skylig
ht A. First, d
e
tect a
ll the pixel sets of high inten
s
ity area in wh
ol
e
image. Seco
nd, cal
c
ulate
the variance of the con
t
rast in ea
ch
set. Third, cho
o
se the high
intensity pixel
to be skylight
value A from
the set that have low value
s
of cont
ra
st.
4.2. Estimating Atmo
sph
e
ric Veil
The pa
rticle
s in the atmosphere (m
ediu
m
tr
ansmissi
on) atten
uate
the scene
ra
dian
ce.
The attenuati
on expone
ntially incre
a
se
s with depth
of sce
ne d(x). In other wo
rds, the medi
um
transmissio
n t(x) can b
e
e
x
presse
d by expone
nt
attenuation. We can
simplify the de
scriptio
n as
the following:
t(x) =
e
−
β
d(x
)
(2)
Another effect of atmo
sph
e
re
pa
rticle
s
is th
e
ad
ditio
n
of a
n
atm
o
sph
e
re
veil V
(
x). Th
e
atmosp
he
re veil V(x) can b
e
expre
s
sed
by the followi
ng:
V(x) =
1 - t(x)
(3)
Acco
rdi
ng to
He
et al [1], to e
s
timate at
mosp
he
ric ve
il, we
cho
o
se
the minim
u
m value
for e
a
ch pixe
l from
all
col
o
r
ch
annel
s.
This process will
be
ro
ug
hly estim
a
ting for atmo
sp
heri
c
veil, expresse
d as follo
wing
:
V(x) =
min I(x
)
(4)
∈
,
,
In any outd
oor im
age
s,
the levels
of contrast
among th
e
pixels give
us a
good
kno
w
le
dge of
deep scen
e. In other wo
rd
s, the contrast
level is a vital factor to ke
ep the realty of
the outdoo
r scene. Upon
this fa
ctor, we propo
se
a new alg
o
rit
h
m to enhan
ce the previou
s
roug
hly estim
a
tion. This al
gorithm
classifies ro
u
ghly atmosp
he
ric
veil to number of classe
s that
equal
s the
nu
mber
of co
ntrast level in
origin imag
e. T
h
is o
peration
pre
c
isely esti
mates th
e re
al
homog
eno
us atmosp
he
ric veil.
4.3. Recov
e
r the Scene
Albedo
Duri
ng th
e p
r
evious
step
we o
b
taine
d
t
he atmo
sp
he
ric veil
V(x); f
r
om thi
s
valu
e we
can
cal
c
ulate me
dium tran
smi
ssi
on t(x) up
o
n
(3):
t(x) =
1 - V(x)
(5)
No
w we
can
simplify the scatterin
g
mod
e
l in (1) to be
as follo
wing:
I(x) =
Ap(x) t(x) +
A V(x)
(6)
The value
s
we obtai
ned
from two p
r
e
v
ious e
s
timat
i
ons all
o
w u
s
to recove
r the scen
e
albed
o by usi
ng the followi
ng equ
ation that is derive
d
from (6):
p
Ix
A
Vx/A tx
(7)
5. Implementation
To imple
m
ent
the su
gge
ste
d
metho
d
s
an
d rem
o
ve the
effects
of su
spe
nde
d pa
rticle
s in
bad weathe
r we will p
e
rfo
r
m the followin
g
step
s:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Rem
o
val of A
t
m
o
spheri
c
Particle
s in Po
or Visibilit
y O
u
tdoor Im
age
s (Ya
s
ee
n Al-Zubaid
y
)
4249
5.1. Acquire
Image in Ba
d Wea
t
her
We will u
s
e
mounted exp
e
rime
ntal ca
mera
with
hig
h
resolution
colors to captu
r
e imag
es
in bad weath
e
r co
ndition
s.
This imag
e will be
the inp
u
t images fo
r our sugg
este
d system.
5.2. Acquire
Image in Clear Wea
t
he
r
After perform
ing step A, the mounte
d
came
ra
will b
e
in the sam
e
locatio
n
an
d same
positio
n till the bad
weath
e
r
co
ndition
s
chang
e and tu
rn
into
clea
r
weath
e
r
con
d
itions an
d thi
s
i
t
may takes fe
w h
ours. T
h
e
n
an
other im
age
will
be ta
ken
an
d thi
s
i
s
the
target i
m
age. T
h
is i
m
age
will be used i
n
the testing
step.
5.3. Pair Ima
g
es for
All Weath
e
r Phen
omenon
In this
step,
A and B
ste
p
s
will b
e
u
s
ed to
captu
r
e the p
a
ir
of the imag
es
for ea
ch
weath
e
r
phe
nomen
on
su
ch a
s
fog, h
a
ze, mist, an
d
drizzle
rai
n
.
One im
age
is ca
pture
d
in
bad
weath
e
r
(lo
w
colo
r
cont
ra
st) a
nd a
noth
e
r ima
ge i
s
capture
d
in cl
ear we
ather con
d
ition
s
(hi
g
h
colo
r co
nt
ra
st
).
5.4. Diffe
ren
t
Backg
round
In many pre
v
ious metho
d
s which e
n
han
ce
the visibility of the image that has b
een
captu
r
ed i
n
b
ad we
athe
r condition
s, the
result
may n
o
t be co
rrect
whe
n
the sce
ne obje
c
ts
are
simila
r to the airlight an
d no sha
dow i
s
ca
st
on them. To make sure that the pro
posed metho
d
s
will solve this problem, pai
r
of
images i
n
different locations
and different
backgrounds will be
captu
r
ed to compa
r
e the o
u
tput image
s with cle
a
r day
images in th
e testing ste
p
.
5.5. Different le
v
e
ls of
visibilit
y
Some enhancement meth
ods
do not
work i
n
low lev
e
l visi
bility such as dense
fog. To
make
sure that the proposed meth
ods
will work in variety levels
of
visibility, th
e pair
of images
will be taken in different l
e
vels
of visibility such
as
mist (li
ght fog), medium
and dense fog to
comp
are the output image
with cle
a
r day
images in th
e testing ste
p
.
5.6. Apply
i
ng the sug
g
e
s
ted me
thod
s
After the dat
aba
se of ima
ges
ha
s be
e
n
creat
ed
wit
h
different ty
pe of imag
es that has
been
captu
r
ed in
differe
nt co
ndition
s, backg
ro
und
s a
n
d
level
of visibilitie
s, the
estimati
ng
skylig
ht and
atmosp
he
ric
veil method
s will be
appl
ied to remov
e
the effect
s of su
spe
n
d
ed
particl
es from
input image
s and will prod
uce hi
ghe
r qu
ality images.
To
evaluate our sug
g
e
s
te
d
method
s, we w
ill apply
two main te
sts. First, the
output
images of suggest
ed sy
stem will be vi
sually compared
with
the images that we
re captured in
clea
r weathe
r co
ndition
s (targ
e
t ima
ges).
Secon
d
, the qualit
y of output image
s will
be
statistically compa
r
ed wit
h
input imag
es; it w
ill also be com
pared with re
sul
t
s of the latest
r
e
sear
ch methods
.
6. Expecte
d
Resul
t
The outp
u
t image
s shoul
d be prim
arily
similar to the
image
s that are
captu
r
ed
in a clea
r
day. In other words, after
applying the
prop
osed me
thod, the input images tha
t
are affected
by
bad weathe
r
con
d
ition
s
wil
l
be simila
r to the qua
lity of clea
r we
ather ima
g
e
s
, In addition, th
e
effects of su
spend
ed atmo
sph
e
ri
c pa
rticles will be
re
moved from the output ima
ge. Furthe
rm
ore,
highe
r quality
output image
s with hig
h
co
ntra
st a
nd fidelity colors wi
ll be prod
uce
d
.
7. Conclusio
n
Enhan
cing
a
n
y outdo
or i
m
age
s that
a
r
e
captu
r
e
d
i
n
ba
d weathe
r is ba
se
d o
n
two m
a
in
factors, ai
rlig
ht and
direct
attenuatio
n. The
co
rrect
prio
r
assum
p
tions to e
s
ti
mate the
s
e t
w
o
factors
are
th
e cl
ue
for re
covering
scen
e alb
edo
fro
m
effect
s
of
su
spe
nde
d p
a
rticle
s by u
s
ing
the scattering
atmosph
e
ri
c
model.
In this
res
e
arc
h
, two prior
ass
u
mptions
were
sugg
est
ed in orde
r to remove the
effect of
atmosp
he
ric
particl
es from
outdoor ima
ges. First,
skylight value can be dete
c
te
d from the pixels
set that
have
high
inten
s
it
y and
low va
rian
ce
co
lor contra
st. Seco
nd, the
preci
s
e
estim
a
tion
of
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 8, August 2013: 4244 –
4250
4250
the real
hom
ogen
ou
s atm
o
sp
heri
c
veil
depe
nd
s on
t
he cl
assification of ro
ugh
atmosp
he
ric
veil
into a numbe
r of levels that equal the nu
mber of color
contrast in th
e input image
.
To impleme
n
t these t
w
o
prio
r a
s
sum
p
tions, the
skylight dete
c
t
i
on metho
d
h
a
s b
een
prop
osed to
cal
c
ulate
airli
ght. We
al
so
su
gge
st the
hom
ogen
ou
s atm
o
sphe
ri
c veil
metho
d
to
cal
c
ulate di
re
ct attenuation
.
The re
sults o
f
sugge
sted
method
s will be use
d
in the scatte
ring
atmosp
he
ric
model to
recover
scen
e albed
o and
remove atm
o
sp
heri
c
pa
rti
c
le
s, namely
rain d
r
izzly, foggy and h
a
zy
effects an
d o
b
tain the true
colo
rs im
age.
Referen
ces
[1]
K He, J Sun,
X T
ang. Singl
e
Image Haze R
e
mova
l Usin
g Dark Cha
n
n
e
l
Prior.
IEEE Tra
n
sactions on
Pattern Analys
i
s
and Mach
in
e Intelli
genc
e.
20
11; 33: 1-1
3
.
[2]
R T
an. Visibi
li
t
y
i
n
b
ad
w
e
ather from
a
singl
e im
ag
e.
IEEE Com
puter Societ
y Conference on
Co
mp
uter Visi
on an
d Pattern
Recog
n
itio
n
. 2
008; 1-8.
[3]
J Yu, Q Liao.
F
a
st Sing
le Ima
ge F
og R
e
mov
a
l Usi
ng Ed
ge-
Preservi
ng S
m
oothi
ng
. Proce
edi
ngs of t
h
e
IEEE Internati
ona
l C
onfer
en
ce o
n
Aco
u
stic
s,
Speec
h, a
n
d
Si
gna
l Pr
oce
ssing, ICASSP
. 201
1; 1
245-
124
8.
[4]
SG Naras
i
mha
n
, SK N
a
yar.
Contrast
r
e
stor
ation
of
w
e
at
h
e
r d
egra
d
e
d
i
m
ages.
IEEE
T
r
ansactio
n
s o
n
Pattern Analys
i
s
and Mach
in
e Intelli
genc
e.
20
03; 25: 71
3-72
4.
[5]
YY Schechn
er
, SG Narasimhan, SK Na
ya
r.
Instant Deha
z
i
n
g
of Imag
es Usin
g Pola
ri
z
a
t
i
o
n
.
Proc
.
IEEE Conf. Computer Visi
on
and Patter
n
Re
cogn
ition. 2
001
; 1: 325-33
2.
[6]
SG Narasimhan, SK Nay
a
r.
I
n
teractive (
de)
w
eatherin
g of
an i
m
age
usi
n
g phys
i
cal
mo
dels
. in Proc.
200
3 ICCV W
o
rkshop o
n
Col
o
r and Phot
ome
t
ric Met
hods in
Comp
uter Visi
on (CPMCV). 2
003; 1-8.
[7]
J Kopf, B Neu
bert, B Chen,
M Cohe
n, D C
ohe
n-Or, O Deussen, M U
y
tt
end
ael
e, D Lis
c
hinski. D
e
e
p
Photo: Mo
de
l-
Based
Ph
otog
raph
Enh
anc
e
m
ent a
n
d
Vie
w
i
ng.
ACM T
r
ans. Grap
hics
.
20
08;
27(5):
116:1-
11
6:10.
[8]
R F
a
ttal. Singl
e imag
e de
hazi
ng.
ACM T
r
ans
actions o
n
Graphics
. 20
08; 27
: 1-9.
[9]
J T
a
rel, N Hau
t
i.
Fast visibilit
y restoratio
n from
a sin
g
l
e
co
lor or gr
ay lev
e
l i
m
a
ge.
Proc
. 2009 IEE
E
Internatio
na
l C
onfere
n
ce o
n
Comp
uter Visi
on (ICCV). 200
9; 2201-
22
08.
[10]
K He, J Sun, X T
ang. Singl
e
Image Ha
ze
Remov
a
l Usi
n
g Dark Ch
ann
el Prior.
IEEE Transactions on
Pattern Analys
i
s
and Mach
in
e Intelli
genc
e.
20
11;
33(1
2
): 234
1-23
53.
[11]
Z
F
a
rbman, R
F
a
ttal, D Lisc
hi
nski, R Sz
elisk
i
.
Edg
e
-Preser
v
ing
Dec
o
mpo
s
itions for
Multi
-
Scale T
o
n
e
and D
e
tail Ma
n
i
pul
atio
n.
ACM T
r
ansactio
n
s o
n
Graphics.
20
08; 27(3): 1-1
0
.
Evaluation Warning : The document was created with Spire.PDF for Python.