A Survey on Single Image Dehazing Approaches

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 04 | Apr 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 1384
A Survey on Single Image Dehazing Approaches
Aswathy Vishnu1, Dr. Baiju P. S.2
1PG Student, Dept. of Electronics & Communication Engineering, LBSITW, Kerala, India
2Assistant Professor, Dept. of Electronics & Communication Engineering, LBSITW, Kerala, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Absorption, and scattering caused by particles
suspended in the atmosphere results in low visibility. Image
dehazing is the process of recovering a haze-free image
from a hazy image. Single image dehazing, which aims to
recover the clear image solely from an input hazy is a
challenging ill-posed problem. Remarkable progress has
been made in recent years on single image dehazing which
has been an under-constrained challenge. This paper gives a
brief review of the existing single image dehazing
approaches.
Key Words: Image dehazing, Atmospheric Scattering
Model (ASM), Image restoration, Dark Channel Prior
(DCP), Image depth information, Airlight
1. INTRODUCTION
Haze is an atmospheric phenomenon that occurs when
suspended aerosols interact with light. It degrades the
image quality by introducing blurring effect, reducing
contrast, and creating false colors in the acquired image
resulting in low visibility. Poor visibility in outdoor haze
scenes generates significant problems for many
applications of computer vision systems including
surveillance, intelligent vehicles, object recognition, etc.
Therefore, an effective haze removal method is necessary.
The process of removing haze from a hazy image is
referred to as dehazing and it is an area of active research
and remarkable progress has been made in recent years on
single image dehazing which has been under-constrained
challenge.
2. ATMOSPHERIC SCATTERING MODEL
Image dehazing is an increasingly widespread approach to
address the degradation of images of the natural
environment by low-visibility weather, atmospheric
particles, and other phenomena. Advancements in
autonomous systems and platforms have increased the
need for low-complexity, high-performing dehazing
techniques.
An Atmospheric Scattering Model (ASM) to describe the
formation of hazy images, shown in Fig 1, can be expressed
as [1],[2],
where I is the hazy image, x is the pixel location of the
image, t(x) is the medium transmission map, J is the
dehazed image and A is the atmospheric light vector in RGB
domain. If the atmosphere condition is assumed to be
homogeneous, t(x) can be represented as
where is the medium extinction coefficient, and d(x) is
the depth between the objects and the camera. The value of
is assumed to be constant in every wavelength of light
and hence, t(x) is considered the relative depth of the scene
with a value between 0 and 1.
Fig -1: ASM for Hazy Image Formation
3. SINGLE IMAGE DEHAZING METHODS
Fattal proposed a method for estimating the optical
transmission in hazy scenes given a single input image [3].
In this method, the image was first split into regions with
constant albedo and the airlight-albedo ambiguity was
removed by introducing a constant that requires the
surface shading and medium transmission to be locally
uncorrelated. The use of robust statistics helps to cope with
complicated scenes containing different surface albedos
and the use of an implicit graphical model makes it possible
to extrapolate the solution to pixels where no reliable
estimate is available. Based on the recovered transmission
values the scene depths can be estimated. This method
showed a significant reduction of the airlight and restored
the contrasts of complex scenes. However, it does not
assume the haze layer to be smooth i.e., it permits
discontinuities in the scene depth and medium thickness.
Carr et al. proposed a single image dehazing method based
on the assumption that the neighboring pixels in an image
will have similar depths [4]. This method showed that a
priori camera geometry can be exploited to improve the
results of any statistical estimation technique. This can be
implemented as a soft constraint within an energy
minimization framework and it leads to a preference that
pixels should get further away as one scans the image from
bottom to top. This preference was fully compatible with

the α-expansion algorithm; however, it does not always
have to be true. The geometric model can be used for fog
removal purposes and also can be incorporated with
different depth estimation techniques.
Fang et al. proposed a single image dehazing method based
on segmentation [5]. The image segmentation was used to
calculate the dark channel instead of patch and then
transmission maps prior were obtained according to the
black body theory. Since the abrupt change of depth usually
happens in the image edge regions, this method reduces
the depth discontinuities within each patch and eliminates
halo artifacts to a large extent. Also, this method overcomes
the inherent deficiency of restoration model which can
obtain better dehazed results. Since the errors in the dark
channel prior remain, this method is limited to obtaining
the haze-free result to some extent when color of the scene
object is similar to the atmospheric light.
He et al. introduced a single image dehazing method based
on Dark Channel Prior (DCP) [6] and it was based on the
assumption that in most of the non-sky patches, at least
one color channel has some pixels whose intensity is very
low and close to zero and as a result, the minimum
intensity in such a patch will be considered as zero. This
method was very effective in recovering vivid colors and
low contrast objects. However, if the haze is removed
thoroughly, then the dehazed output loses the depth effect
and it seems to be unnatural. Hence, a very small amount of
haze has to be kept for distant objects. A major limitation of
this method was that since the transmission map may not
be always constant in a patch, it creates halos and block
artifacts in the output. Thus soft matting technique was
employed for better image restoration and the results of
the restored images are impressive with visual contrast.
However, this method was computationally complex
because of the calculation of DCP. Also, as the haze imaging
model assumes common transmission for all color
channels, this method may fail to recover the true scene
radiance of the distant objects and they remain bluish.
Kim et al. proposed a simple and adaptive single image
dehazing algorithm based on contrast enhancement [7].
This method first estimates the airlight in a given hazy
image based on the quad-tree subdivision and then
estimates the optimal transmission to maximize the
contrast. It provided good dehazing results at low
computational complexity since the transmission was
assumed to be a constant over an entire image. However, it
may not produce faithful results when the depth
differences between foreground objects and the
background are very large. As a result, this method was
also extended to estimate a space-varying transmission
map to dehaze an image with a complicated depth
structure more accurately. However, this optimization is
less reliable since a smaller number of pixels were
employed in the cost function formulation.
Ancuti et al. introduced a single image dehazing method
using multi-scale fusion [8]. The fusion-based technique
was based on the concept that two input images were
derived from the original input to recover the visibility for
each region of the scene in at least one of them. To blend
the information of the derived inputs effectively, weight
maps were used to filter the important features to preserve
the regions with good visibility. Finally, the Laplacian of
the inputs and Gaussian of the weights are blended in a
multi-scale fashion to reduce the artifacts. Even though this
method provided faster and more accurate results, it was
limited only to homogeneous hazy images.
Zhu et al. proposed a color attenuation prior-based method
for single image dehazing [9]. The color attenuation prior
was based on the difference between the brightness and
the saturation of the pixels within the hazy image. This
simple and powerful prior can help to create a linear model
for the scene depth of the hazy image. The bridge between
the hazy image and its corresponding depth map was built
effectively by learning the parameters of this linear model.
With the recovered depth information, the haze can be
removed from the hazy input image. This method provided
much sharper and natural results free from halo effects.
But, linear color attenuation prior was also based on
statistics that were not sensitive to the scene objects with
inherent white color.
Ren et al. introduced a multi-scale deep neural network for
single-image dehazing by learning the mapping between
hazy images and their corresponding transmission maps
[10]. The scene transmission map was first estimated by a
coarse-scale network and then refined by a fine-scale
network. Even though this method was easy to implement
and reproduce, it was less effective for nighttime hazy
images.
Singh et al. designed Gradient profile prior (GPP) to
evaluate depth map from hazy images [11]. The developed
gradient-based profile prior was able to reduce the color
and texture distortion issues. The transmission map was
improved by utilizing guided anisotropic diffusion and an
iterative learning-based image filter (GADILF). The
restoration model was improved to reduce the effect of
pixels saturation and color distortion from restored images.
This method was able to suppress visual artifacts for hazy
images and yield high-quality results with high
computational speed. However, if the hazy image is
removed completely, then the restored image looks like an
artificial image.
Zhu et al. proposed a novel fast single image dehazing
algorithm based on artificial multi-exposure image fusion
to enhance the performance and robustness of image
dehazing [12]. Based on a set of gamma-corrected
underexposed images, pixel-wise weight maps were
constructed by analyzing both global and local exposure to
guide the fusion process. The spatial dependence of

luminance of the fused image was reduced, and its color
saturation was balanced in the dehazing process. However,
it does not consider the saturation variation in a hazy
image.
Zheng et al. proposed an adaptive multiple-exposure image
fusion (AMEF) algorithm for single image dehazing [13]. An
adaptive gamma transformation was utilized for each
component (R, G, and B) of the color image, based on the
mean and standard values of each component, that is, the
global characteristics of the hazy image. A sequence of
adaptive gamma corrections were employed to extract a
collection of under-exposed multi-exposure image
sequences from a haze image. Then, the Gaussian pyramid
and Laplacian pyramid with the homomorphic filtering
algorithm was utilized to address the exposed obtainable
images, followed by a modified Laplacian filter for
calculating the contrast of the exposed accessible images.
This method provided higher contrast, richer details, and a
better visual effect in the dehazed image. However, the
linear adjustment of image saturation and the adaptive
selection of image block size resulted in increased
computational complexity.
Baiju et al. proposed an optimization framework using a
low-rank approximation to efficiently estimate the scene
transmission map for single image dehazing [14]. A low-
rank approximation technique with weighted nuclear norm
minimization was introduced to smoothen the coarse
transmission map obtained from hazy data to remove the
visual artifacts in the dehazed image. This efficient
optimization model estimates scene transmission map for
dehazed image using a single available hazy image and
achieves fast dehazing. This method does not require any
pre-trained model and was capable of retrieving good
results in less execution time.
Sahu et al. proposed a single image dehazing method based
on using a new color channel prior [15]. In this method,
atmospheric light was estimated by dividing an image into
blocks, then the score of each block was computed. The
block having the highest score was further used for
calculating the atmospheric light. A new color model was
adopted to calculate the transmission map and it was
further used for computing radiance. Although the results
were acceptable for indoor images, further research needs
to be performed to generate realistic and visually pleasing
images.
Zhang et al. introduced a single image dehazing using a
dual-path recurrent network (DPRN) [16]. The DPRN
consists of a feature extraction block, a transmission map
estimation block, a dual-path block with a parallel
interaction function, and an image reconstruction block.
Initially, the DPRN uses the feature extraction block and
the transmission map estimation block to extract features
from the hazy image. These features are then fed into the
dual-path block, which utilizes two parallel branches to
restore the basic content and details of the clear images.
The dehazed result is obtained by processing the output
features of the dual-path block by the image reconstruction
block. Even though this method produces images with clear
content and fine details, the model needs to be trained
initially with different hazy images which is a time-
consuming process.
3. CONCLUSION
Image has important applications in many fields such as
marine surveillance, environment monitoring, and so on.
The scattering effects of the atmospheric particles in the air
play a main role, resulting in contrast reduction and color
fading. As a result, the clear image is necessary. The main
advantage of single image dehazing is that a haze-free
image can be obtained from only a single available image.
This work is a summary of different single image dehazing
techniques with their advantages and limitations.
ACKNOWLEDGEMENT
We would like to thank the Director of LBSITW and the
Principal of the institution for providing the facilities and
support for our work.
REFERENCES
[1] H. Israel and F. Kasten, “Koschmieders theorie der
horizontalen ¨ sichtweite,” in Die Sichtweite im Nebel und
die Moglichkeiten ihrer ¨ kunstlichen Beeinflussung ¨ .
Springer, 1959, pp. 7–10.
[2] R. T. Tan, “Visibility in bad weather from a single image,”
in 2008 IEEE conference on computer vision and pattern
recognition. IEEE, 2008, pp. 1–8.
[3] R. Fattal, “Single image dehazing,” ACM transactions on
graphics (TOG), vol. 27, no. 3, pp. 1–9, 2008.
[4] P. Carr and R. Hartley, “Improved single image dehazing
using geometry,” in 2009 Digital Image Computing:
Techniques and Applications. IEEE, 2009, pp. 103–110.
[5] S. Fang, J. Zhan, Y. Cao, and R. Rao, “Improved single
image dehazing using segmentation,” in 2010 IEEE
International Conference on Image Processing. IEEE, 2010,
pp. 3589–3592.
[6] K. He, J. Sun, and X. Tang, “Single image haze removal
using dark channel prior,” IEEE transactions on pattern
analysis and machine intelligence, vol. 33, no. 12, pp. 2341–
2353, 2011.
[7] J.-H. Kim, J.-Y. Sim, and C.-S. Kim, “Single image dehazing
based on contrast enhancement,” in 2011 IEEE
International Conference on Acoustics, Speech and Signal
Processing (ICASSP). IEEE, 2011, pp. 1273–1276.

[8] C. O. Ancuti and C. Ancuti, “Single image dehazing by
multi-scale fusion,” IEEE Transactions on Image Processing,
vol. 22, no. 8, pp. 3271–3282, 2013.
[9] Q. Zhu, J. Mai, and L. Shao, “Single image dehazing using
color attenuation prior.” in BMVC. Citeseer, 2014.
[10] W. Ren, S. Liu, H. Zhang, J. Pan, X. Cao, and M.-H. Yang,
“Single image dehazing via multi-scale convolutional neural
networks,” in European conference on computer vision.
Springer, 2016, pp. 154– 169.
[11] D. Singh, V. Kumar, and M. Kaur, “Single image
dehazing using gradient channel prior,” Applied
Intelligence, vol. 49, no. 12, pp. 4276–4293, 2019.
[12] Z. Zhu, H. Wei, G. Hu, Y. Li, G. Qi, and N. Mazur, “A novel
fast single image dehazing algorithm based on artificial
multiexposure image fusion,” IEEE Transactions on
Instrumentation and Measurement, vol. 70, pp. 1–23, 2020.
[13] M. Zheng, G. Qi, Z. Zhu, Y. Li, H. Wei, and Y. Liu, “Image
dehazing by an artificial image fusion method based on
adaptive structure decomposition,” IEEE Sensors Journal,
vol. 20, no. 14, pp. 8062–8072, 2020.
[14] P. S. Baiju, S. L. Antony, and S. N. George, “An
intelligent framework for transmission map estimation in
image dehazing using total variation regularized low-rank
approximation,” The Visual Computer, pp. 1–16, 2021.
[15] G. Sahu, A. Seal, O. Krejcar, and A. Yazidi, “Single image
dehazing using a new color channel,” Journal of Visual
Communication and Image Representation, vol. 74, p.
103008, 2021.
[16] X. Zhang, R. Jiang, T. Wang, and W. Luo, “Single image
dehazing via dual-path recurrent network,” IEEE
Transactions on Image Processing, vol. 30, pp. 5211–5222,
2021.

A Survey on Single Image Dehazing Approaches

More Related Content

Similar to A Survey on Single Image Dehazing Approaches (20)

More from IRJET Journal (20)

Recently uploaded (20)

A Survey on Single Image Dehazing Approaches