IRJET - Dehazing of Single Nighttime Haze Image using Superpixel Method

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 998
Dehazing Of Single Nighttime Haze Image using Superpixel Method
A. Susmi Raj1, R. Soundararajan2
1PG Scholar, Dept. of Computer Science and Engineering, SVS College of Engineering Coimbatore, Tamilnadu, India
2Assistant Professor, Dept. of Computer Science and Engineering, SVS College of Engineering
Coimbatore, Tamilnadu, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - The open air picturescapturedinharshclimate are
corrupted due to the nearness of fog, mist, rain and so on.
Pictures of scenes captured in lousy climate have destitute
contrasts and colors. This might also cause problem in
spotting the objects within the captured murky images. Dueto
murkiness there may be trouble to numerous computer vision
application because it diminishes the perceivability of the
scene. Picture de-hazing is one of the essential imperative
inspect variety in picture preparing. Cloudinessisgenuinely an
Climatic effect.
In this paper a novel super-pixel based single image haze
removal algorithm with a neural network method is proposed
for nighttime haze image. The input nighttime haze image is
first decomposed into a glow image and glow-loose nighttime
haze picture the usage of their relative smoothness. A super
pixel based approach is then brought to compute the price of
atmospheric light and dark channel for each pixel in the glow
free haze image. The transmission map is decomposed from
the darkish channel of the glow-unfastened haze photo by
means of the weighted guided photograph filter. Since super-
pixels normally adhere to the limits of gadgets well, a smaller
nearby window size may be selected. In addition to this the
gathered images (the hazy images) will input to a neural
network that will increase the quality oftheoutput imagethat
the actual image we want to get.
Key words: Nighttime image haze removal, Glow
decomposition, Morphological artifacts, Super-pixel
segmentation, Weighted guided image filtering
1. INTRODUCTION
In real life, natural phenomena (inclusive of rain, haze,snow
etc)can lead to degradation of outdoor images. Because in
these environments, massive amount of particle floating in
the atmosphere, mild propagation within the atmosphere
will be suffering from these floating particles.Inthesector of
computer vision image de-hazing has a much broader
demand. From the perspective of image processing, it may
provide pretreatment for some visual algorithms. Besides
from the practical factor of view, it plays an important role
in military system and civil system [1]. Haze free images
makes the scene look more practical and provide more
useful information. Thereforethestudiesof imagede-hazing
has critical practical importance and developmentprospect.
There were many haze removal techniques which are
primarily based on the optical model developed fordaytime
haze images. At present, the most widely bodily model is a
linear equation such as transmission and atmospheric light.
According to the version, day time de-hazingtechniquewant
to estimate the atmospheric mild and transmission map.
Many classic strategies use additional information or
multiple photographs to do away with haze from the haze
snapshot. For example Schechner et al. [2] proposed a novel
technique of using images with different polarizer
orientation to de-hazing. Lai et al. [3] proposed an
interesting technique to derive the most beneficial
transmission map at once from the sunlight hours haze
image version. In [4], a quick algorithm for single image de-
hazing is proposed based totally on linear transformation
with the aid of assuming that a linear relationship existence
inside the minimal channel between the hazy image and the
haze-free image. On top of dark channel prior, He et al.
introduced a simple method to study single image haze
elimination in [5]. The dark channel prior is based on a
commentary that most nearby patches in haze-free scene
images include a few pixels that have very low intensities in
at the least one coloration channel. Many dark channel prior
based totally haze removal algorithms were delivered since
then [7][8]. The dark channel earlier based totally methods
usually work well, but the dark channel earlier has
limitations. For example,morphological artifactsareanissue
for the darkish channel previous while the preliminary
transmission map is computed use of the dark channel
earlier [5]. A simple edge retaining decomposition based
totally framework changed into proposed for single image
haze removal in [9]. Simplified dark channel of the haze
image is decomposed into a base layer and a element layer
via the weighted guided image filter (WGIF) in [11], and the
transmission map is expected from the base layer. Image
enhancement based techniques alsoareappliedtotheimage
de-hazing, for example [12][13][14].
Even though these techniques generally carry out well for
daylight hours haze image shots, they are now not properly
ready to correct nighttime scenes due to varying imaging
conditions such as active light assets or glow effects. Due to
existence of active light assets, for example, avenue lighting
et al, and their related glow, the version of daytime haze
images are now not applicable to the midnight haze image.
Recently, several interesting methods have beenemergedto
decorate nighttime haze images [15][16][17][20]. Pei and
Lee [15] introduced a method primarily based the color
transfer processing. The colorings of a nighttime hazeimage
had been mapped to those of a daytime haze image. Then, a
dark channel prior primarily based set of rules was supplied
to estimate the transmission map. A post-processing step
become also furnished to improvetheinsufficient brightness

and coffee overall evaluation of haze-free image. Due to the
coloration transfer, even though their approach has the
dependable de-hazing quality, the shade of the complete
haze free image looks unnatural. Zhang et al [16] proposeda
new photo model which accounted for various illumination.
First, the mild intensity became estimated and enhanced to
achieve an illumination balanced result, then, they expected
the shade traits of the incident light, finally, they used the
darkish channel previous to cast off the haze.
The main contribution of this paper is a new type of haze
elimination algorithm using the concept of super-pixel.
Compared with the patch based totally methods, the super-
pixel can be used to reduce morphologic artifacts because of
the patch. This is due to the fact that the super-pixel can
adhere to boundaries of items within thehazeimage well. As
such, the radius of the WGIF may be reduced. Subsequently,
more exceptional details may be preserved. In addition, the
proposed super-pixel based totally technique has more
chance to correctly estimate the transmission maps for all
pixels I white gadgets nearby a digital camera than
algorithms in [9][17] and [25]. In other words, the proposed
method provides a technique to a challenging problem on
the estimation of transmissionmapsforpixelsin whiteitems
close by the digital camera. The rest of this paper is
organized as preliminaryknowledge,procedureofnighttime
single image haze removal, experimental results and
conclusion.
2. PRELIMINARY KNOWLEDGE
Since the WGIF and the SLIC might be implemented in the
proposed method, the relevant expertise on them are
summarized in this section.
2.1. Weighted Guided Image Filter
Denote the guidance photograph as I, the filtering output as
Z, and the input photograph as X. The pictures I and X can
be identical. Z is thought to be a linear remodel of I in a
window Ωk focused at the pixel p, and the radius of the
window Ωk is r
Z(p) = ak I(p)+bk,∀p ∈ Ωk, (1)
Wherein the price of (ak,bk) can be computed by means of
minimizing the following cost function in the window Ωk.
E(ak,bk) = ∑ p∈Ωk ((ak I(p)+bk – Xp)2 +λak
2 ), (2)
And λ is a regularization parameterpenalizinglargeak.Since,
halo artifacts may appear on some edges, an edge-
conscious weighting is introduced and incorporated into the
GIF to form the WGIF. The edge-aware weighting ΓI(k) is :
E(ak,bk) = ∑ p∈Ωk ((ak I(p)+bk − X(p))2
+ λ ΓI(k) a 2 k ), (3)
The solutions of ak and bk are given as:
ak = µI ⊙X,r (k)−µI,r(k)µX,r(k) σ 2 I,r (k)
+ λ ΓI (k) , (4)
bk = µX,r(k)−akµI,r(k) (5)
2.2 Super-pixel Segmentation Via The SLIC
A super-pixel is a small region composed by a sequence of
pixels with adjacent positions, similar shade, similar
brightness, similar texture and different characteristics.The
SLIC (Simple Linear Iterative Clustering) is a common
approach of the super pixel segmentation. This method can
phase pixels quick and simply, besides, it may perceive
barriers better. The color images are converted into a five-
dimensional feature vectors V = [l,a,b,x,y] in CIELAB color
area and XY coordinate. Where, [l,a,b] suggests the color of
pixel, [x,y] shows the placement of pixel. It is a distance
measurement standard for five-dimensional characteristic
vectors, and a neighborhood clustering procedure for photo
pixels. SLIC set of rules can generate compact and about
uniform super-pixels. Besides, t hasa highoverall evaluation
in terms of speed, contour preserving and hyperpixel shape,
and the segmentation impact is in keeping with people’s
expectation. The metric of SLIC is given as:
dlab = √ (l(p)−l(p ′ ))2 +(a(p)−a(p ′ ))2
+(b(p)−b(p ′ ))2 (6)
dxy = √ (x(p)− x(p ′ ))2 +(y(p)−y(p ′ ))2 (7)
Ds = √ ( dlab Nlab )2 +( dxy Nxy )2 (8)
where, dlab represents color distance; dxy represents spatial
distance; p and p ′ are two pixels; Nxy = √ K N , where K is
called the total number of pixels, N is the number of the
super pixels; Nlab varies with special images, however it is
normally a set value. The super-pixels normally can pick out
obstacles better. So, it is able to be expected that the concept
of super-pixel may be used to reduce the morphological
artifacts within the single image haze removal.
3. NIGHTTIME SINGLE IMAGE HAZE REMOVAL
In this section, first provide information on the model of
nighttime haze images. Based on this version, the glow is
eliminated from the input photograph therebytheglow-free
haze image is obtained. The model of the glow-loose haze
photo is similar with the version of sunlight hours haze
picture except that the atmospheric light is spatially varying
inside the nighttime haze photograph. It is thus predicted
that existing dark channel based algorithm for sunlight
hours haze removal may be prolonged to address nighttime
haze image. Based on this observation, a super-pixel based
haze elimination algorithm is offered for the glow-free
nighttime haze images.

3.1 Modelling Of A Nighttime Haze Image
Since the light source best from the sun, the sunlight hours
images are not stricken by otherlightsources. Inthedaylight
hours haze image de-hazing algorithm, the model is
generally used as;
Xc(p) = Zc(p)t(p) + Ac(1−t(p)) (9)
When a photo is captured at night, light sources especially
come from road lighting fixtures and car lighting etc. These
lights aren’t international uniform, thus; the glow is present
inside the image. The glow is an atmospheric point spread
function (APSF). Inspired by this, the complete nighttime
haze scenes by adding the glow version into the daylight
haze picture model:
Xc(p) = Zc(p)t(p)+ Ac(1−t(p))+ Aa ∗ APSF (10)
In which c∈{r,g,b} represents coloration channel index, Xc
shows discovered nighttime images. Zc represents the scene
radiance vector. T is the transmission map and indicates the
elements of light that penetrates via the haze. Ac represents
atmospheric light which is not always globally uniform any
longer. Aa indicates active light sources, that the intensity is
convolved with APSF. So that it will get the scene radiance
vector, this is, the haze free image Z, the glow (that is
Aa*APSF) are decomposed from the input photograph
(Xc(p)), and the atmospheric light Ac and the transmission
map t are estimated.
3.2 Glow Removal from the Input Image
From the unprocessed images, we can see that glow can
reduce the visibility of the photograph or may bemakea few
objects unseen. In order to attain an excellent visual image,
the glow should be eliminated from the input photograph.
For simplicity, the equation (10) can be written as:
Xc(p) = Jc(p)+Gc(p) (11)
Where, Jc(p) = Zc(p)t(p) + Ac(1-t(p)), we name it a glow free
nighttime haze image, and Gc(p) = Aa* APSF, it’s far a glow
photograph. From the equation (12), the glow elimination
can be seemed as a layer separation problem. The objective
function for glow layer separation may be described as
follow:
E(J) = ∑ p (ρ(J(p) ∗ f1,2) + λ1((X(p)− J(p)) ∗ f3)2)
s.t.0 ≤ J(p) ≤ X(p) (12)
∑Jr(p) = ∑ p Jg(p) = ∑ p Jb(p)
Where, f12 is the 2-route first spinoff filters. f3 is the second
order Laplacian filter out and the operator ∗ denotes
convolution. ρ(µ) = min(µ2,τ) is a robust characteristic
which preserves large gradients of the input photo X within
the closing nighttime haze layer [17]. For simplicity, we
outline FjL = L ∗ fi then the objective function can be
rewritten as;
E(J) = ∑ p (ρ(F1,2 J(p)) + λ1(F3 J(p)− F3 X(p))2 ) (13)
According to the method [23], with a view to pass the FjL
time period outdoor the ρ(·) function, a weightβisdelivered
to the goal function, the brandnewobjectivefunctionmaybe
written as;
E(J) = ∑ p (β(F1,2 J(p)−g1,2)2 +ρ(g1,2)
+ λ1(F3 J(p)− F3 X(p))2 ) (14)
(a) (b) (c)
(d) (e) (f)
Fig.1: (a,d) two nighttime haze pictures; (b,e) glow pix
with the aid of the method[23]; (c,f) glow photographs by
the proposed algorithm. The glow pictures produced with
the aid of the proposed set of rules are smoother.
It can be validated that the above solution is an
approximated solution at the same time as the provided
solution within the equation (16) is an precise one.
Therefore, our output is more correct than the output in
[23]. The Fig.1 indicates separated glow picture with the aid
of the algorithm in [23] and the proposed set of rules.
Clearly, the glow’s element in (c,f) is greater meticulous and
non-stop than (b,e). Therefore, the proposed glow
decomposition method is higher than the technique in
[23].Since the input image consists of a glow picture and a
glow-loose haze photo the procedureofglowdecomposition
directly impacts the compositionofglow-freeimage,thereby

(a) (b) (c) (d) (e)
(f) (g) (h) (i) (j)
Fig. 2: (a,f) two haze images; (b,g) glow images by the method [23]; (c,h) glow-free haze images by the method [23]; (d,i)
glow images by the proposed algorithm; (e,j) glow-free haze images by the proposed algorithm. There are more details in
the glow-free haze images by the proposed algorithm
affecting the high-quality of de-hazed picture. The Fig.2
shows the glow images and glow-loose images by means of
the Li’s algorithm and the proposed algorithm. It may be
visible from the Fig.2 (c)(e) that the glow may be removed
better through our algorithm, and from the Fig.2 (h)(j) that
the glow unfastened haze photographs are toward the
nature scene.
3.3 Atmospheric Light Estimation
After the glow being removed from the input image, a glow-
free nighttime haze image is finded. In order to repair the
haze-free photo, the atmospheric light Ac and the
transmission map t is need to be estimated.Sincethemodels
of a daytime haze image are almost the same besides for the
atmospheric light, it is able to be expectedthat some existing
method of daytime haze image de-hazing can be carried out
to removal the haze from the glow-free nighttime haze
image. The atmospheric light is always appeared as the
brightest shade within the daylight haze image. However,
due to the presence of the light sources, the atmospheric
light is not always global uniform inside the glow-free
nighttime haze image. According to the photograph
formation model, the picture Zc(p) is a produce of
illumination and reflectance components:
Zc(p) = Ac(p)Rc(p) (15)
where Rc(p) is the reflectance component. The glow-free
nightime haze image is then represented as:
Xc(p) = Ac(p)Rc(p)t(p) + Ac(p)(1−t(p)) (16)
the components Ac(p) and t(p) are assumedto beconstantin
a local window of p. It can then be derived that
max p ′∈Ω(p) {Xc(p ′ )} = Ac(p) max p ′∈Ω(p)
{Rc(p ′ )}t(p)+ Ac(p)(1−t(p)) (17)
Using the following maximum reflectance prior [24],
max p ′∈Ω(p) {Rc(p ′ )} = 1 (18)
it can be derived that
Ac(p) = max p ′∈Ω(p) {Xc(p ′ )}. (19)
(a) (b) (c)
Fig. 3: (a) a glow-free haze image; (b) an initial
atmospheric light image; (c) a refined atmospheric light
image. There are morphological artifacts in the initial
atmospheric light image and they are removed by the
WGIF.
On the premise of the feature ofsuper-pixel,theatmospheric
light more risk to be uniform in each super-pixel. Hence, the
glow free nighttime haze image Jc is decomposed into N
super-pixels in place of a grid of small areas in [17]. The
brightest pixel within the ith super-pixel and it is assigned to
all the pixels inside the super-pixel to make up the
preliminary atmospheric light IAc(p). Even though the
morphological artifacts are reduced the usage of the super-
pixels, there are still visible morphological artifacts. WGIF
can be followed to remove the morphological artifacts. The

WGIF can be followed to remove the morphological artifacts
in order to get the clean global atmospheric light Ac(p).
The photos to be filtered are IAc(p) and the guidance images
are the color components of the glow-free nighttime haze
images Jc(p).
(a) (b) (c)
(d) (e) (f)
Fig.4: (a),(d) are the equal glow-free haze image by way of
the proposed method; (b),(c) respectively is an
atmospheric light image, de-hazed image via the method
[17]; (e),(f) respectively is an atmospheric light image, de-
hazed image by means of the proposed method.
3.4 Transmission Map Estimation
After getting the atmospheric light, the transmission map is
estimated to recover the scene radiance. Here the conceptof
DehazeNet is used to estimate the transmission map.
DehazeNet is an give up-to-end system. It without delay
learns and estimates the mapping relations among hazy
image patches and their medium transmissions. This is
completed through special design of its deep architecture to
encompass establishedimagede-hazingprinciples.Ipropose
a novel nonlinear activation feature in DehazeNet, clled
Bilateral Rectified Linear Unit (BReLU).
BReLU is a singular activationcharacteristic thatisbeneficial
for image recovery and reconstruction. BReLU extends
Rectified Linear Unit(ReLU) and demonstrates its
importance in obtaining accurate image recuperation.
Technically BReLU uses the bilateral Restraint to lessen
search space and enhance convergence. I establish
connections between components of DehazeNet and those
assumptions/priors used in existing dehazing method, and
explain that DehazeNet improves over these strategies by
automatically gaining knowledge of all these additives from
give up to quit. Neural network (DehazeNet) net are
designed to implement four sequential operations for
Medium transmissionestimation,namely, featureextraction,
multi- scale mapping, local extremum, and nonlinear
regression are represented in Fig.5.
1) Feature Extraction: Feature extraction is performed by
the usage of the convolution layer of nueral network. The
input image can be filtered with different kinds of filters. So
we can maintain the vital info and can do away with the
unwanted details. Convolution is performed by the usage of
5 filters of length 3X5.
2) Multi- Scale Mapping: Here we can filter with different
filters having different scaleslike 4X3,4X5,4X7.Then wecan
analyze the details clearly.
3) Local Extremum: When the last two steps are
completed, we get a lot of filtered images. Local extremumis
the process of selecting the most important values from
these filtered images.
4) Nonlinear Regression: Non-linear regression isthelast
step. It is the process of selecting the most important values
and getting the output. When it does, we get a transmission
map. After getting the atmospheric light, the transmission
map is estimated to recover the scene radiance.
3.5 Recovery of the Scene Radiance
The haze-free image Zc(p) can be recovered the use of
equation (20). It is worth nothing that the value of tp is close
to zero, Zc(p)t(p) is also near to zero. In this situation, if the
haze free image is restored, the noise inside the haze free
image will be substantially amplified. Thus a lower sure tm
which was added to constrain tp can lessen the noise. The
value of the tm is decided by way of the denses of thehaze.Its
value is decided on as 0.2 if the haze is not heavy, 0.375
otherwise. The final image is restored as
Zc(p) = Jc(p)− Ac(p) max(t(p),tm) + Ac(p) (20)
It is well worth nothing that the proposed edition
mechanism of tm is coarse and pleasant of de-hazed images
may want to be advanced through a finer choice of tm with
appreciate to special haze levels.
4. Experimental Results
Here I compare proposed algorithm with the daylight hours
haze elimination algorithms in [5] and four nighttime haze
elimination algorithms in [16][17] and [24].
4.1 Comparison Among Different Haze Removal Algorithm
For comparison, the resultstheuseofdehazingalgorithmsin
[5], [16], [17], [20], [24] and proposed algorithm are shown
within the Fig.14. He’s approach in [5] is a classical set of
rules for daytime haze images. It is not alwayssurprisedthat
it is not appropriate to put off the haze from

Fig. 5 The architecture of DehazeNet. DehazeNet conceptually consists of four sequential operations (feature extraction,
multi-scale mapping, local extremum and non-linear regression).
(f) (g) (h) (i) (j)
(k) (l) (m) (n) (o)
Fig.6: (a,f,k) haze images; (b,g,l) dehazed images by the method in [5]; (c,h,m) dehazed images by method [6]; (d,I,n)
dehazed images by the method in [17]; (e,j,o) dehazed images by proposed algorithm.
nighttime haze image. Only part of haze can be removed via
this method and the glow nonetheless exists in the final
pictures. The techniques in [16] and [17] are designed for
images, there are visible halo artifacts and noise is amplified
within the de-hazed snap shots as illustrated in the Fig.6
(h),(m) glow around the mild sources. The method in [16]
was improved in [24].
Neither the algorithm in [20] nor the set of rules in [24]
addressed the glow artifacts. As such, the glow artifacts are
grater visible within the de-hazed images by way of the
algorithms in [20] and [24]. It is well worth nothing that a
bigger tm is selected in the equation (20) to keep away from
amplifying noise in the sky regions. On the other hand,
excellent details can also be smoothedthroughtheproposed
algorithm. It is desired to layout a finer tm for the proposed
set of rules. In spite of our algorithm works nicely in de-
hazing of nighttime haze image, there are still some
deficiencies. For example, the haze removal image appears
darker than the input image. Fortunately, this problem may
be addressed using a single image brightening algorithm in
[26].
The running time of the proposed algorithm is barely longer
because of segment the glow-free image into super-pixels.
The approach in [17] works better than the method [16].
Unfortunately, noise is likewise amplified and halo artifacts
nevertheless exists inside the final images.Thesetofrulesin
[17] money owned that the atmospheric light is locally
regular in a grid of small area. However, this is not
continually true. On the other hand, accordingtocapabilities
of the super-pixels, it’s far more likely that the atmospheric
mild is regular within the super-pixel.
(a) (b) (c) (d) (e)

5. CONCLUSION
This paper presented a novel super-pixel based nighttime
haze image de-hazing set of rules. The proposed set of rules
can eliminate the glow and the haze higher than thestate-of-
the art methods. Local exceptional detailsarealsopreserved
better, and the color distortion and halo artifacts are
glaringly reduced. As such, the haze-loose image is in the
direction of a nature scene. On the alternative hand, because
segmentation of glow-free nighttimehazeimagesintosuper-
pixels is required with the aid of the proposedalgorithm, the
running time and the complexity are increased. It is known
that the existing single image haze elimination algorithms
suffer from amplifying noise inside the sky region. This
problem could be addressed by means of deciding on a finer
adaptive tm inside the equation (20). This problem can be
studied in future research.
REFERENCES
[1] S. G. Narasimhan and S. K. Nayar, “Vision and the
atmosphere,” International Journal onComputerVision,
vol. 48, no. 3, pp. 233-254, Jun. 2002.
[2] Y. Y. Schechner, S. G. Narasimhan, and S. K. Nayar,
“Polarization based vision through haze,” Applied
Optics, vol. 42, no. 3, pp. 511-525, 2003.
[3] Y. H. Lai, Y. L. Chen, C. J. Chiou, and C. T. Hsu, “Single-
image dehazing via optimal transmission map under
scene priors,” IEEE Transactions on Circuit andSystems
for Video Technology, vol. 25, no. 1, pp. 1-14, Jan. 2015.
[4] W. C. Wang, X. H. Yuan, X. J. Wu, and Y. L. Liu, “FastImage
Dehazing Method Based on Linear Transformation,”
IEEE Transactions on Multimedia, vol. 19, no. 6, pp.
1142-1155, Jun. 2017.
[5] K. M. He, J. Sun, and X. O. Tang, “Guided image filtering,”
IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 35, no.6, pp. 1397-1409, Jun. 2013.
[6] F. Kou, W. H. Chen, C. Y. Wen, and Z. G. Li, “Gradient
Domain Guided Image Filtering,” IEEE Transactions on
Image Processing, vol. 24, no.11, pp. 4528-4539, Nov.
2015
[7] Z. G. Li, J. H. Zheng, W. Yao, and Z. J. Zhu, “Single image
haze Removal via a simplified dark channel,” in 2015
IEEE International Conference on Acoustics,Speechand
Signal Processing, pp. 1608-1612, Apr. 2015.
[8] Q. S. Zhu, D. Wu, Y. Q. Xie, and L. Wang, “Quick shift
segmentation guided single image haze removal
algorithm,” Proceedings of the 2014 IEEE International
Conference on Robotics and Biomimetics, pp. 113- 117,
Dec. 2014
[9] J. Zhang, Y. Cao, and Z. Wang, “Nighttime haze removal
based on a New Imaging Model,” IEEE International
Conference on Image Processing, pp. 4557-4561, Oct.
2014..
[10] C. Ancuti, C. O. Ancuti, C. De Vleeschouwer, and A. C.
Bovik, “Nighttime dehazing by fusion,” IEEE
International Conference on Image Processing, pp.
2256-2260, 2016.
[11] Z. G. Li, J. H. Zheng, Z. J. Zhu, W. Yao, and S. Q. Wu,
“Weighted guided image filtering,” IEEETransactionson
Image Processing, vol. 24, no. 1, pp. 120-129, Jan. 2015.
[12] H. T. Xu, G. T. Zhai, X. L. Wu, and X. K. Yang,
“Generalized Equalization Model for Image
Enhancement,” IEEE Transactions on Multimedia, vol.
16, no. 1, pp. 61-82, Jan. 2014.
[13] K. T. Shih and H. H. Chen, “Exploiting Perceptual
Anchoring for Color Image Enhancement,” IEEE
Transactions on Multimedia, vol. 18, no. 2, pp. 300-310,
Feb. 2016.
[14] F. Kou, Z. Wei, W. H. Chen, X. M. Wu, C. Y. Wen, and Z. G.
Li, “Intelligent detail enhancement for exposurefusion,”
IEEE Transactions on Multimedia, Aug. 2017.
[15] S. C. Pei and T. Y. Lee, “Nighttime haze removal using
color transfer pre-processing and dark channel prior,”
IEEE International Conference on Image Processing,pp.
957-960, Oct. 2012.
[16] J. Zhang, Y. Cao, and Z. Wang, “Nighttime haze removal
based on a New Imaging Model,” IEEE International
Conference on Image Processing, pp. 4557-4561, Oct.
2014.
[17] Y. Li, R. T. Tan, and M. S. Brown, “Nighttime haze
removal with glow and multiple light colors,” IEEE
International Conference on Computer Vision, pp. 226-
234, Dec. 2015
[18] Y. Li and M. S. Brown, “Single image layer separation
using relative smoothness,” IEEE Conference Computer
Vision and Pattern Recognition, 2014.
[19] J. Zhang, Y. Cao, S. Fang, Y. Kang, and C. W. Chen, “Fast
haze removal for nighttime image using the reflectance
prior,” in 2017 Conference on Computer Vision and
Pattern Recognition, Jul. 2017.
[20] K. M. He, J. Sun, and X. O. Tang, “Single image haze
removal usingdark channel prior,”IEEETransactionson
Pattern Recognition and MachineIntelligence,pp.2341-
2353, 2011.

[21] F. Liu, C. Shen, G. Liu, and I. Reld, “Learning Depth from
Single Monocular Images Using Deep Convolutional
Neural Fields,” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 38, no. 10, pp.2024-2039,
Dec. 2016.
[22] Z. G. Li and J. H. Zheng, “Single image brightening via
exposure fusion,” in2016IEEEInternational Conference
on Acoustics, Speech and Signal Processing, pp. 1756-
1760, Shanghai, China, Mar. 2016.
[23] R. Achanta, A. Shaji, K. Smith, A. Lucchi, and P. Fua, “SLIC
superpixels compared to state-of-the-art superpixel
methods,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 34, no. 11, pp. 2274-2282,
Nov. 2012.
[24] L. K. Choi, J. You, and A. C. Bovik, “Referenceless
prediction of perceptual fog density and perceptual
image defogging,” IEEE Transactions on Image
Processing, vol. 24, no. 11, pp. 3888-3901, Nov. 2015.
[25] S. G. Narasimhan and S. K. Nayar, “Shedding light on the
weather,” IEEE ConferenceComputerVisionandPattern
Recognition, 2003.s
[26] Z. G. Li and J. H. Zheng, “Single Image De-Hazing Using
Globally Guided Image Filtering,” IEEE Transactions on
Image Processing, vol. 27, no. 1, pp. 442-450, Jan. 2018.

IRJET - Dehazing of Single Nighttime Haze Image using Superpixel Method

More Related Content

What's hot (20)

Similar to IRJET - Dehazing of Single Nighttime Haze Image using Superpixel Method (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET - Dehazing of Single Nighttime Haze Image using Superpixel Method