Turk J Elec Eng & Comp Sci

() : –

c TÜBİTAK
doi:10.3906/elk-

A joint image de-hazing and segmentation model

Haider ALI1 ,Awal SHER1 ,Nosheen ZIKRIA2 ,Lavdie RADA3∗

1
Department of Mathematics, UOP, Peshawar, Pakistan,
2
Jinnah College for Women, UOP, Peshawar, Pakistan,
3
Biomedical Engineering Department, Bahcesehir University, Istanbul, Turkey,

Received: .201 • Accepted/Published Online: .201 • Final Version: ..201

Abstract: Objects and their features identification in hazy or foggy weather conditions has been of an interest in the
last decades. Improving the image visualization by removing the weather influence factors towards an easy image post-
processing, such as objects detection, induces benefits on human assistance systems. In this paper, we propose a novel
variational model which will be capable to jointly segment and de-haze a given image. The proposed model incorporates
the atmospheric veil estimation and the locally computed denoising constrained surfaces into a level set function by
performing a robust and efficient image de-hazing and segmentation scheme for both grey and color outdoor images. The
proposed model not only shows efficient segmentation of objects in foggy images, by outperforming state-of-art methods
but also produces dehazed objects results in the same time.
Keywords. Active contours, image de-hazing, variational approach, segmentation, Euler-Lagrange equation.

1. Introduction
Despite the increasing ubiquity of digital photography, the images of the outdoor scene have low visibility due to
fog or haze and consecutively recovering the original colours of the natural scenes is a hard problem. Augmenting
of the depth information from images captured with ordinary cameras is still a long-standing unsolved problem in
image processing and computer vision. In the last decades, several methods which moderately solve the problem
has been proposed [1–4]. Among the earliest de-hazing approaches, one can distinguish the polarization filters
techniques [5, 6], in which multiple images were required to remove fog from a given scene. Those models
have limited applications as most of the imputed data are with only one single haze image. Moreover, this
technique does not guarantee absolute removal of fog or haze. To overcome the disadvantage of multiple images
in defogging methods, Tan et al. [7] observed that the fog-free images have higher contrast in comparison with
foggy images. As a straightforward outcome of his observation, he introduced contrast maximization method.
Although this method requires one single image as algorithm input, the restored images of this method often
will have halo artifacts and over-stretches due to the fact that the method does not conform to physical effects.
Further, He et al. [4] proposed an image prior-dark channel model to remove haze from a single input image.
The method is based on a key observation that most local patches in outdoor haze-free images contain some
pixels whose intensity is very low in at least one colour channel. Using this prior with the haze imaging model,
one can directly estimate the thickness of the haze and recover high-quality haze-free images. To remove haze

∗ Correspondence: Correspondence: lavdie.rada@eng.bau.edu.tr

1
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

He et al. [4] proposed the following formula:

I(x)dark = min ( min I c (y)), (1)

y∈Ω0 (x) c∈(r,g,b)

where I c (x) is a color channel of the given image I(x) definite in the domain Ω ∈ R2 , and Ω0 (x) is a local
patch with center x. In most of non-sky local image areas, we have,

I(x)dark = min ( min I c (y)) = 0. (2)

y∈Ω0 (x) c∈(r,g,b)

However, the output images of the method may have colour tone modification due to imprecise estimation
of air-light. Considering the degradation of image quality of outdoor scenes as a process effected by attenuation
and the air-light [8], occurred due to the scattering of atmospheric particles [9], this phenomenon can be restored
by reversing Koschmieder’s law [1] which is given with the formula:

I(x) = I0 (x)e−kd(x) + I∞ (x)(1 − e−kd(x) ) (3)

where I(x) is a given image with fog, I0 (x) is a fog-free image, k is the defunctness coefficient of the fog and d(x)
is the distance of the scene point from the photographic equipment at pixel x and I∞ (x) is the sky intensity.
For homogenous fog, both I∞ (x) and k are global constants over the whole image while for heterogenous fog
it may vary locally. Representing the airlight as A(x) the formula in equation (3) can be written as
 A(x) 
I(x) = I0 (x) 1 − + A(x). (4)
I∞

To improve this model, Cho et al. [1] proposed the minimization of the following functional:
Z Z
2
E(A(x)) = (H(W (x) − A(x))dx + λ ϕ(||∇A(x)||)dx (5)
Ω Ω

where Ω is the domain of the given image, A(x) is air-light or atmospheric veil, W (x) = min(I(x)) is the
minimum components of I(x) for each pixel (image of the whiteness within the observed image), H is the
Heaviside function, λ is a regularization parameter which balances the influence between two terms of (4), and
ϕ represents an edge-preserving regularization function (assumed to be an even function in C 2 (R)) . This model
outperformed the existing image de-hazing algorithms, by not only showing good results for images restoration
in heavy fog but at the same time by not producing halo effects. To improve the performance of Cho et al.
[1] model, several research have been introduced in the recent years [10–13]. However, these models and other
similar models [2–4] may not work properly in multiple image de-hazing cases.
On the other hand, segmentation of a given image refers to the process of locating objects and boundaries (lines,
curves, etc.) in images. The most popular and successful methods for image segmentation are active contour
approaches. Those techniques can be classified into two groups: edge-based [14–18] and region-based [19–21]
models. The choice of these models depends on results acceptability for every image and problem on each
specific segmentation application. In comparison with edge-based methods, the region-based active contour
frameworks are well known for their correctness in segmenting images with or without noise with or without
gradient. The Chan-Vese (CV) model is one of the most popular region-based active contour models [20] which

2
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

is obtained due to the minimization of the following functional:

Z Z
F CV (c1 , c2 , C) = µ length(C) + λ1 |I(x) − c1 |2 dx + λ2 |I(x) − c2 |2 dx,
inside(C) outside(C)

where λ1 , λ2 are positive constants and c1 , c2 are average intensities inside and outside of the contour C
respectively. Representing the contour C with zero level set, i.e.C = {x ∈ Ω|φ(x) = 0}, and minimizing the
above energy functional with respect to φ , c1 and c2 we obtain the following variational formulation:
   
∇φ 2 2
δ (φ) µ∇. − λ1 (I(x) − c1 ) + λ2 (I(x) − c2 ) = 0, (6)
|∇φ|
R R
ΩR
I(x)H(φ(x))dx I(x)(1 − H(φ(x)))dx
c1 = and c2 = ΩR , (7)
Ω
H(φ(x))dx Ω
(1 − H(φ(x)))dx

The drawback of the CV model is its deficiently of segmenting images with inhomogeneous intensity, including
fog and haze. To improve the CV model, Li et al [22] proposed a local binary fitting (LBF) energy model,
which is capable of segmenting image with intensity inhomogeneity. The LBF energy embedding the variational
model is given by:

F LBF (C, f1 , f2 ) = µlength(C) + λ1 e1 + λ2 e2 (8)

with
Z
e1 = Kσ (x − y)|I(y) − f1 (x)|2 H(φ(y)dy (9)
Z
e2 = Kσ (x − y)|I(y) − f2 (x)|2 (1 − H(φ(y))dy

2 2
where φ is a level set function, H(φ) is the Heaviside function, Kσ (x) = 1
(2π)n/2 σ 2
.e−|x| /2σ is the Gaussian
Kernel, and f1 , and f2 are two smooth functions that approximate the local image intensities inside and outside
the contour C , respectively. Minimization of above equation with respect to φ , f1 , and f2 leads to the following
equations:
   
∇φ
δ (φ) µ∇. − (λ1 e1 − λ2 e2 ) = 0, (10)
|∇φ|
Kσ (x) ∗ I(x)H (φ(x)) Kσ (x) ∗ I(x)(1 − H (φ(x)))
f1 (x) = , and f2 (x) = . (11)
Kσ (x) ∗ H (φ(x)) Kσ (x) ∗ (1 − H (φ(x)))

Similar work was lately introduced by Zhang et al [23]. The proposed model incorporates a local image fitting
(LIF) term for segmentation. Their proposed energy functional is as follows:
Z
1
E LIF (φ) = µlength(C) + |I(x) − I00 (x)|2 dx (12)
2 Ω

where
I00 (x) =m1 H (φ(x)) + m2 (1 − H (φ(x)) (13)

3
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

represent local computed image and m1 and m2 are the local averages given by:
\
m1 =mean(z0 ∈ ({x ∈ Ω : φ(x) < 0} Wσ (x))) (14)
\
m2 =mean(z0 ∈ ({x ∈ Ω : φ(x) > 0} Wσ (x))), (15)

with Wσ (x) is a rectangular window function. The window function can be chosen in different ways e.g. a
truncated Gaussian window or a constant window. In can be easily noticed that in case Wσ (x) is a truncated
Gaussian window then m1 and m2 are equal to f1 and f2 in equation (10).
Another attempt to enhance the CV model is by Wu et al. [24] who proposed the following energy
functional:

(I(x) − c1 )2 (I(x) − c2 )2
Z Z
2
F (φ) = λ 2 (φ + 1) dx + (φ − 1)2 dx, (16)
c1 c22

where c1 , c2 are constant and λ > 0 . The proposed variational segmentation model is convex with unique
global solution for its energy functional within the interval [−1, 1] .
Furthermore, Zhang et al. [23] combining the region base idea introduced by Chan-Vese [20] with geodesic
model idea and formulated the following minimisation functional in terms of level set formulation:

∂φ(x)
= α(SP F )|∇φ(x)|, in Ω
∂t

φ(t = 0; x) = φ0 (x), in Ω,

where SP F denotes the signed pressure force function, α is a positive constant, and φ0 is the initial level set
function. Akram et al. [25] modified Zhang model for images with intensity inhomogeneity by replacing the SPF
function with a local-SPF function. Although Akram et al. [25] improved the performance of the region-based
active contour models by efficiently stoping the contours at weak or blurred edges, the model is not designed
to handle images having multiple intensity objects. To improve local-SPF model, Ali et al. [26] proposed a
combination of SPF function with the generalized average concept in a new segmentation framework. Their
model is given as following:
∂φ(x)
= α(GSP F )|∇φ(x)|, in Ω
∂t

φ(t = 0; x) = φ0 (x), in Ω.

The model shows significant improvements over the [25] model for segmentation of multiregion images with
inhomogeneous intensities. To the best of authors knowledge, there is no single designed model to jointly
segment and dehaze images with fog or haze.
In this work, we propose a novel variational model which will be capable to jointly segment and de-haze
a given image at the same time. Our model is mainly based on estimating the atmospheric veil and level set
function for efficient image de-hazing and segmentation respectively. Most of the recent image segmentation
models may not efficiently segment foggy images. The proposed technique is robust for both grey and colour
outdoor images. The experimental tests of our proposed method on colour and grey images will validate that
our formulation is far better than the latest segmentation techniques available in the literature.

4
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

2. The Proposed Model

Let I(x) denote the given image with fog factor, over a domain Ω , I : Ω ⊂ R2 −→ R , and its statistically
locally computed approximated image If (x) as follows:

If (x) = f1 (x)H (φ(x)) + f2 (x)(1 − H (φ(x)) (17)

where f1 (x) and f2 (x) are smooth functions similarly obtained as in equation (10). Both terms will statistically
contain significant information under haze and fog condition in contrary with global fitting terms such as c1
and c2 for CV model.
In order to jointly de-haze and segment a given image we propose the following energy functional:
Z Z
2
E(φ(x), A(x)) = |∇A(x)|dx + λ H(φ(x))(W (x) − A(x))dx+
Ω Ω
Z
µ δ(φ(x))|∇φ(x)|dx+
Ω
Z Z
IA (x)
λ1 (If (x) − IA (x))2 dx + (1 − λ1 ) (logIf (x) + )dx (18)
Ω Ω If (x)

where µ, λ1 and λ are positive parameters. The first and the second term of equation (18) represent the
physical properties of the atmospheric veil. The atmospheric veil A(x) is an increasing function of the scene
point distance d(x, y), whereas A(x) is positive and higher than the minimum of the components of I(x) , [1].
This imposes the smoothness of the air-light A(x) everywhere into the image domain, positivity A(x) ≥ 0 ,
and A(x) < W (x) , where W (x) = min(I(x)) . The other terms in equation (18) are variational segmentation
fitting and smoothness terms for the curve fitting where IA (x) is given by:

I(x) − A(x)
IA (x) = A(x)
. (19)
1− I∞ (x)

In other words, the first two terms of the proposed energy function satisfy edge-preserving regularization
property while adequately remove fog whereas the last three terms will evolve the curve φ towards objects
boundary. Minimizing the energy functional E(φ(x), A(x)) with respect to φ(x) , we will get the following
equation:
∇φ(x)
µδ(φ(x))∇( ) + 2λ1 (IA (x) − If (x))(−f1 (x)δ(φ(x)) + f2 (x)δ(φ(x)))
|∇φ(x)|
1 IA (x)
+(1 − λ1 )( (f1 (x)δ(φ(x)) − f2 (x)δ(φ(x))) + 2 (f1 (x)δ(φ(x)) − f2 (x)δ(φ(x))) = 0. (20)
If (x) If (x)

Equation (20) can be rewritten

∇φ(x) h 1 IA i
µδ(φ(x))∇( ) + (f1 (x)δ(φ(x)) − f2 (x)δ(φ(x))) − 2λ1 (IA − If ) + (1 − λ1 )( + 2 ) = 0
|∇φ(x)| If If

Using the time marching method to solve the above equation we get the following evolution equation:
∂φ(x) h ∇φ(x) (1 − λ1 ) i
= δ(φ(x)) µ∇( ) + (f1 (x) − f2 (x))[ (I f + IA ) − 2λ 1 (IA − If )] . (21)
∂t |∇φ(x)| If2

5
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

In similar way, minimizing the energy functional E(φ(x), A(x)) , with respect to A(x), we derive the following
equation
 ∇A  2λ1 I  A  (1 − λ1 ) I
∇ − λ2 δ(φ)(W − A) + A
( − 1) I − A − If (1 − ) − A
( − 1) = 0.
|∇A| 3
(1 − I∞ ) I∞ I∞ 2
If (1 − I∞ ) I∞

Rewriting the above equation we obtain:

 ∇A  ( I − 1) h 2λ1  A  1 − λ1 i
∇ − λ2 δ(φ)(W − A) + I∞ A 2 A
I − A − If (1 − ) − =0
|∇A| (1 − I∞ ) 1 − I∞ I∞ If

and its evolution equation

∂A  ∇A  ( I − 1) h 2λ1  A  1 − λ1 i
=∇ − λ2 δ(φ)(W − A) + I∞ A 2 I − A − If (1 − ) − (22)
∂t |∇A| (1 − I∞ ) 1 − IA∞ I∞ If

2.1. Numerical implementation

Rewriting shortly equation (21) we have

φ(x)
= K(φ(x)),
∂t

where K(φ(x)) is the right hand side in (21). Considering Ω a two dimensional space and using the finite
differences for the discretization of equation (21) we obtain:

φn+1 n
i,j − φi,j
= K(φni,j ).
∆t

In other words,

h (1 − λ1 ) i
φn+1 n n
i,j = φi,j + ∆tδ(φ ) µκφ + (f1 − f2 )[ 2 (If + IA ) − 2λ1 (IA − If )] , (23)
If

where κφ represents the curvature and is computed according to the following formula

∇φ φxx φ2y − 2φxy φx φy + φyy φ2x

κφ = div( )= (24)
|∇φ| (φ2x + φ2y )3/2

where φx , φy , φxx , φyy and φxy are computed as follows:

1 1
φx = (φi+1,j − φi−1,j ), φy = (φi,j+1 − φi,j−1 ),
2h 2h
1 1
φxx = 2 (φi+1,j + φi−1,j − 2φi,j ), φyy = 2 (φi,j+1 + φi,j−1 − 2φi,j ),
h h
1
φxy = 2 (φi+1,j+1 − φi−1,j+1 − φi+1,j−1 + φi−1,j−1 ) (25)
h

where h is the grid point.

6
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

In similar way Eq. (22) can be discretised as follows:

An+1 n
i,j − Ai,j
= R(Ani,j )
∆t
∇A
where R(Ani,j ) is the right hand side in (22) and κA = div( |∇A| ) can be written in a similar form as follows:

∇A Axx A2y − 2Axy Ax Ay + Ayy A2x

κA = div( )= (26)
|∇A| (A2x + A2y )3/2

where Ax , Ay , Axx , Ayy and Axy are computed as:

1 1
Ax = (Ai+1,j − Ai−1,j ), Ay = (Ai,j+1 − Ai,j−1 ),
2h 2h
1 1
Axx = (Ai+1,j + Ai−1,j − 2Ai,j ), Ayy = 2 (Ai,j+1 + Ai,j−1 − 2Ai,j ),
h2 h
1
Axy = 2 (Ai+1,j+1 − Ai−1,j+1 − Ai+1,j−1 + Ai−1,j−1 ) (27)
h


n+1
Ai,j = Ani,j + ∆t κA − λ2 δ(φn )(W − Ani,j )

( II∞ − 1) h 2λ1  Ani,j  1 − λ1 i

+ An An
I − Ani,j − If (1 − ) − (28)
(1 − i,j
)2 1 − i,j I∞ If
I∞ I∞

In the following we show some basic steps of our algorithm.

Algorithm 1 Joint Segmentation De-hazing Method Algorithm: (φk (x), Ak (x)) ←

JDS(φ(k−1) , A(k−1) , µ, λ1 , λ2 , , maxiter , tol).
1. Initialize the level function φ(x) by φ0 (x) and A0 by A0 .

A0 (x) = βI(x), 0 < β < 1,

and the level set function φ(x) with φ0 (x)


−ρ if x ∈ Ω − ∂Ω

φ0 = φ(x, t = 0) = 0 if x ∈ ∂Ω

ρ if x ∈ Ω − Ω


where ρ > 0 is a constant, Ω is image domain and ∂Ω is the boundary of Ω .

2. Calculate f1k (x) and f2k (x) using equation (10).
3. Update φ(x)k equation(21) using its discretisation (23).
4. Update A(x)k equation (22) using its discretisation (23).
5. Check weather the solution is stationary. If not, repeat from step 2.

3. Experimental Results
In this section, we display the results obtained from our model for foggy images of various scene and show that
the model can easily handle dehazing and segmentation by outperforming other existing models. Simulation and

7
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

performance evaluation of the proposed algorithm is accomplished through various clean and foggy images on
MATLAB platform in Windows 8.1 environment, Intel core i3 environment on a 1.70 GHz personal computer
with 4.00 GB installed memory (RAM). For both grey and color images, we fixed the parameters λ1 = 0.01 ,
λ2 = 100 , µ = 0.02 and maxiter = 100 . To achieve better results the parameters can be tuned depending on
the fog/ haze level, or alternatively, optimized scheme can be found.
Dehazing comparison with existing models: In the first test set we show the capability of the
proposed model in fog removing. Figure 1 shows defogging results of the proposed method applied in different
images and compared with state of art Ancuti et al. [27] and Cho et al. [1] models. From the experimental
results, in this figure we witness a robust performance in restoring the given image and clearly observed the
correctness of the proposed method in terms of dehazing outdoor images. The proposed algorithm shows clear
details of fog free scene objects as compared to other methods in images with intense fog. The obtained results
are with acceptable visual quality without sky region or less sky region images.
Segmentation accuracy of the proposed method for clean images and its comparison with
state of art models: In order to show the segmentation performance of the proposed model we test the model
on grey and color clean images. Figure 2 and 3 (last columns) show the obtained results of the proposed method
in comparison with segmentation performance of SPF [28] and GSPF [26] models. As would be expected, the
application of LBF model to the same set of images gives similar results to the proposed method, not being
repetitively. From the experimental results we notice that our method performers well in segmenting grey and
vector valued images whereas the other methods have similar results. The experiments carried in this test
set are followed by experiments of images containing fog and its comparative quantitative analysis on coupled
de-haze and segmentation.
Segmentation accuracy for fog/ haze images and its comparison with state of art segmen-
tation models: This test set shows the ability of our model to remove fog meanwhile segmenting the object
of the given image. The results obtained in Figure 4 and Figure 5, show that the proposed model successfully
segmented the natural fog images (last column Figure 4 and 5). The experiments are carried out on different
number of scenes containing low and high level of fog/ haze. In these figures we simultaneously compare the
performance of the proposed method, in terms of segmentation, with state-of-art models such as LBF [22] and
Wu [24] models. The incorporation of the physical properties of the atmospheric veil into the proposed vari-
ational model and the representation of the given image as a piecewise continues function clearly makes the
proposed method to outperform the existing models which can not cope such cases. From figure 4 and 5, second
and third column we notice that LBF [22] and Wu [24] models will give poor results or even wrong results if the
scene of the image contains relatively high level of fog and the foreground objects fall into intensity variation.
Quantitative Analysis: To show the segmentation accuracy of the proposed model we use the Jaccard
Similarity (JS) coefficient for quantitative comparison of different models, which is defined as the ratio of the
junction of the areas and the union. Let R1 and R2 denotes segmentation result and ground truth (GT)
respectively, then JS coefficient is computed in the following way:
T
R1 R2
JS(R1 , R2 ) = S .
R1 R2

Clearly, the values of JS coefficients are JS ∈ [0, 1] . The accuracy of a model is better if JS coefficient is
closer to one and the vice-versa if JS is closer to zero. Figure 6 compares the JS coefficients obtained by
LBF [22], Wu models [24], and the proposed method obtained over time, presented in the second, third and

8
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

(m) (n) (o) (p)

Figure 1: The defogging performance of the proposed model in comparison with Ancuti et al. [27] and Cho et
al. [1] models. The first column is the given image, the second column shows the output results of Ancuti [27]
model, the third column the output results of Cho et al. [1] model, and the forth column shows the results of
the proposed model.

9
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

Figure 2: The first column is the given grey image, the second column shows the local-SPF results [28], the
third column shows GSPF results [26], and the forth column shows the result of our method.

10
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

(m) (n) (o) (p)

Figure 3: The first column is the given vector valued image, the second column shows the local-SPF results
[28], the third column shows GSPF results [26], and the forth column shows the result of our method.

11
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

(a) Initial Contour (b) LBF Results (c) Wu Results (d) Proposed Method Results

Figure 4: The performance of the proposed model in comparison with LBF (iteration = 3000 , λ1 = 1 , λ2 = 1 )
and Wu (iteration = 5000 ) models on images with fog. .

12
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

(a) Initial Contour (b) LBF Results (c) Wu Results (d) Proposed Method Results

Figure 5: The performance of the proposed model in comparison with LBF (iteration = 3000 , λ1 = 1 , λ2 = 1 )
and Wu (iteration = 5000 ) models on images with fog. .

13
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

the last column, respectively. We easily observes that the proposed model is performing better in comparison
with the other models. The proposed algorithm produces higher JS values as compared to other state of the art
methods for most of the images experimented. Higher values of JS shows better accuracy which indicates a good
performance of the proposed model. Moreover, the graphs showing the JS coefficients reveal robustness and
stability of the performance of the proposed model whereas the other methods cannot cope with such images.
This is to be expected as none of the methods mentioned above or other methods in literature take fog/haze
into account.

4. Conclusions
In this paper, we presented an efficient variational segmentation under a dehazing image reconstruction in
a single combined model. The proposed model jointly segments and de-hazes a given image by combining
active contour on region based for images segmentation with a variational level set formulation. To recover
the boundary of the desired object we introduce local binary fitting energy terms as well as edge preserving
regularization to utilize accurate air-light map. The model is applicable for both grey and color images.
Experimental results show that our new approach performs well in the comparison with other existing classical
models. Moreover, the proposed model has a simple structure and easily it can be adapted for special cases
such as inhomogeneous fog/ haze.

14
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

0.22 0.9 1

0.8
0.21 0.9

0.7
0.2 0.8
0.6

0.19 0.5 0.7

0.18 0.4 0.6

0.3
0.17 0.5
0.2

0.16 0.4
0.1

0.15 0 0.3
0 200 400 600 800 1000 1200 1400 1600 1800 2000 0 50 100 150 200 250 300 350 400 450 500 0 5 10 15 20 25 30 35 40 45 50

(a) Initial Contour (b) LBF Result (JS=0.2184) (c) Wu Result (JS=0.3721) (d) Our Result (JS=0.9072)

0.45 0.7 0.75

0.7
0.4 0.6
0.65

0.35 0.5 0.6

0.55
0.3 0.4
0.5
0.25 0.3
0.45

0.2 0.2 0.4

0.35
0.15 0.1
0.3

0.1 0 0.25
0 200 400 600 800 1000 1200 1400 1600 1800 2000 0 50 100 150 200 250 300 350 400 450 500 0 5 10 15 20 25 30 35 40 45 50

(e) Initial Contour (f) LBF Result (JS=0.2360) (g) Wu Result (JS=0.6203) (h) Our Result (JS=0.7454)

0.07 0.9 0.9

0.8 0.8
0.068

0.7
0.7
0.066
0.6
0.6
0.064 0.5
0.5
0.062 0.4
0.4
0.3
0.06
0.3
0.2

0.058 0.2
0.1

0.056 0 0.1
0 500 1000 1500 0 50 100 150 200 250 300 350 400 450 500 0 10 20 30 40 50 60 70 80 90 100

(i) Initial Contour (j) LBF Result (JS=0.0720) (k) Wu Result (JS=0.7976) (l) Our Result (JS=0.8734)

0.21 0.18 0.8

0.2 0.75
0.16
0.7
0.19
0.14
0.65
0.18
0.12
0.6
0.17
0.1 0.55
0.16
0.5
0.08
0.15
0.45
0.06
0.14
0.4

0.13 0.04
0.35

0.12 0.02 0.3

0 500 1000 1500 0 50 100 150 200 250 300 350 400 450 500 0 10 20 30 40 50 60 70 80 90 100

(m) Initial Contour (n) LBF Result (JS=0.1344) (o) Wu Result (JS=0.1504) (p) Our Result (JS=0.7179)

Figure 6: Quantitative Analysis using JS coefficients for comperison between LBF, Wu methods and
the proposed method.

15
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

References

[1] W. H. Cho, I. S. Na, S. C. Seo, S. K. Kim, S. Y. Park, Single image defogging method using variational approach
for edge-preserving regularization, World Academy of Science, Engineering and Technology, International Journal
of Computer, Electrical, Automation, Control and Information Engineering 7 (2013) 829–833.
[2] S. Zhang, C. S. C. Jian Yao, Single image dehazing using fixed points and nearest-neighbor regularization, et al.
(Eds.): ACCV 2016 Workshops, Part I, LNCS 10116 (2017) 18–33.
[3] L. Li, W. Feng, J. Zhang, Contrast enhancement based single image dehazing via tv-l1minimization, in: 2014 IEEE
International Conference on Multimedia and Expo (ICME), 2014, pp. 1–6. doi:10.1109/ICME.2014.6890277.
[4] K. He, J. Sun, X. Tang, Single image haze removal using dark channel prior, IEEE Transactions on Pattern Analysis
and Machine Intelligence 33 (2011) 2341–2353. doi:10.1109/TPAMI.2010.168.
[5] S. Shwartz, E. Namer, Y. Y. Schechner, Blind haze separation, in: 2006 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR’06), Vol. 2, 2006, pp. 1984–1991. doi:10.1109/CVPR.2006.71.
[6] Y. Y. Schechner, S. G. Narasimhan, S. K. Nayar, Instant dehazing of images using polarization 1 (2001) I–325–I–332.
doi:10.1109/CVPR.2001.990493.
[7] R. Tan, Visibility in bad weather from a single image, IEEE Computer Society Conference on Computer Vision
and Pattern Recognition (CVPR) 1 (2008) 1–8.
[8] M. I. Anwar, A. Khosla, Vision enhancement through single image fog removal, Engineering Science and Technology,
an International Journal 20 (2017) 1075–1083.
[9] S. S. Saggu, M. K., A review on various haze removal techniques for image processing., International journal of
current engineering and technology. 5 (2015) 1500–1505.
[10] R. Fattal, Dehazing using color-lines, ACM Trans. Graph. 34 (1) (2014) 13:1–13:14.
[11] Q. Zhu, J. Mai, L. Shao, A fast single image haze removal algorithm using color attenuation prior, IEEE Transactions
on Image Processing 24 (11) (2015) 3522–3533. doi:10.1109/TIP.2015.2446191.
[12] N. H. Kaplan, K. K. Ayten, A. Dumlu, Single image dehazing based on multiscale product prior and application
to vision control, Signal, Image and Video Processing 11 (8) (2017) 1389–1396. doi:10.1007/s11760-017-1097-4.
[13] K. C. Q. Bolun Cai, Xiangmin Xu, D. Tao, Dehazenet: An end-to-end system for single image haze removal, Trans.
Img. Proc. 25.
[14] V. Caselles, R. Kimmel, G. Sapiro, Geodesic active contours, International Journal of Computer Vision 22 (1997)
61–79.
[15] R. Malladi, J. A. Sethian, B. C. Vemuri, Shape modeling with front propagation: a level set approach, IEEE
Transactions on Pattern Analysis and Machine Intelligence 17 (1995) 158–175.
[16] R. Kimmel, A. Amir, A. M. Bruckstein, Finding shortest paths on surfaces using level sets propagation, IEEE
Transactions on Pattern Analysis and Machine Intelligence 17 (1995) 635–640. doi:10.1109/34.387512.
[17] C. Li, C. Xu, C. Gui, M. D. Fox, Level set evolution without re-initialization: a new variational formulation 1
(2005) 430–436. doi:10.1109/CVPR.2005.213.
[18] M. Kass, A. Witkin, D. Terzopoulos, Snakes: Active contour models, International Journal of Computer Vision 1
(1988) 321–331. doi:10.1007/BF00133570.
[19] N. Paragios, R. Deriche, Geodesic active regions and level set methods for supervised texture segmentation,
International Journal of Computer Vision 46 (2002) 223–247. doi:https://doi.org/10.1023/A:1014080923068.
[20] T. F. Chan, L. A. Vese, Active contours without edges, IEEE Transactions on Image Processing 10 (2001) 266–277.
[21] A. Tsai, A. Yezzi, A. S. Willsky, Curve evolution implementation of the mumford-shah functional for image
segmentation, denoising, interpolation, and magnification, IEEE Transactions on Image Processing 10 (2001) 1169–
1186. doi:10.1109/83.935033.

16
 AUTHOR and AUTHOR/Turk J Elec Eng & Comp Sci

[22] C. Li, C. Y. Kao, J. C. Gore, Z. Ding, Implicit active contours driven by local binary fitting energy (2007) 1–
7doi:10.1109/CVPR.2007.383014.
[23] K. Zhang, H. Song, L. Zhang, Active contours driven by local image fitting energy, Pattern Recogn. 43 (4) (2010)
1199–1206.
[24] Y. Wu, C. He, A convex variational level set model for image segmentation, Signal Processing 106 (2015) 123–133.
[25] F. Akram, Segmentation of regions of interest using active contours with spf function., Computational and mathe-
matical methods in medicine.
[26] H. Ali, N. Badshah, K. Chen, G. A. Khan, N. Zikria, Multiphase segmentation based on new signed pressure force
functions and one level set function, Turkish Journal of Electrical Engineering and Computer Science 25 (2017)
2943 – 2955.
[27] C. O. Ancuti, C. Ancuti, Single image dehazing by multi-scale fusion, IEEE Transactions on Image Processing
22 (8) (2013) 3271–3282. doi:10.1109/TIP.2013.2262284.
[28] K. Zhang, L. Zhang, H. Song, W. Zhou, Active contours with selective local or global segmentation: A new
formulation and level set method, Image and Vision Computing 28 (4) (2010) 668 – 676.

17