IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 28, NO.

5, MAY 2019 2357

A Fast Image Dehazing Algorithm Using

Morphological Reconstruction
Sebastian Salazar-Colores, Eduardo Cabal-Yepez , Member, IEEE,
Juan M. Ramos-Arreguin, Senior Member, IEEE, Guillermo Botella,
Luis M. Ledesma-Carrillo, Student Member, IEEE,
and Sergio Ledesma , Member, IEEE

Abstract— Outdoor images are used in a vast number of appli- conditions and is caused by floating particles such as water
cations, such as surveillance, remote sensing, and autonomous drops or any other aerosol that reflects the light, diffusing it
navigation. The greatest issue with these types of images is the in the air and reducing the visibility of details. Vision systems
effect of environmental pollution: haze, smog, and fog originating
from suspended particles in the air, such as dust, carbon, applied to remote sensing, surveillance, and autonomous navi-
and water drops, which cause degradation to the image. The gation generally use input images under adverse weather con-
elimination of this type of degradation is essential for the input ditions; therefore, the results produced depend on the quality
of computer vision systems. Most of the state-of-the-art research of the image received as input. Atmospheric scattering adds
in dehazing algorithms is focused on improving the estimation nonlinear and data-dependent noise to the captured outdoor
of transmission maps, which are also known as depth maps.
The transmission maps are relevant because they have a direct image, which makes image restoration a very difficult process;
relation to the quality of the image restoration. In this paper, consequently, several research works have focused on dimin-
a novel restoration algorithm is proposed using a single image to ishing the haze effects in images captured using vision systems
reduce the environmental pollution effects, and it is based on the by designing and applying dehazing algorithms [1], which can
dark channel prior and the use of morphological reconstruction be divided into those requiring additional information from
for fast computing of transmission maps. The obtained exper-
imental results are evaluated and compared qualitatively and the scene (for instance, in [2], a method for haze removal that
quantitatively with other dehazing algorithms using the metrics utilizes the correlations between hazy images and haze-free
of the peak signal-to-noise ratio and structural similarity index; images as external information is presented) and those using
based on these metrics, it is found that the proposed algorithm one single image (in [3], a study and evaluation of existing
has improved performance compared with recently introduced single-image dehazing algorithms, using a large-scale bench-
approaches.
mark of synthetic and real-world hazy images, is presented).
Index Terms— Single-image dehazing, image enhancement, The algorithms that use a single image to diminish the effects
morphological operations, dark channel prior. of adverse atmospheric conditions are the most studied these
days because they are useful for practical applications, such
I. I NTRODUCTION as autonomous vehicle navigation, surveillance and remote
sensing [4], [5]. In this regard, dark channel prior (DCP),
O UTDOOR images are exposed to adverse weather condi-
tions such as haze, fog or smog, which cause effects such
as scene darkening, contrast degradation, and color change,
which consists of estimating the transmission map of an image
by estimating the depth of each element in it, has been used
among others. Haze is among the most common atmospheric in combination with other algorithms for fast computation of
accurate transmission maps [6], at the cost of long computation
Manuscript received February 11, 2018; revised July 20, 2018, times compromising its utilization on online vision systems.
September 1, 2018, and October 15, 2018; accepted November 23, 2018.
Date of publication December 7, 2018; date of current version January 30, Therefore, recent research has focused on the computation
2019. This work was supported in part by the National Council on Sci- speed of transmission maps, aiming to preserve the image
ence and Technology (CONACYT), Mexico, under Scholarship 285651, quality.
and by DAIP - U. de Gto. under Research Project 209, Convocatoria
Institucional de Investigacion Cientifica 2018. The associate editor coordi- In this work, a novel fast algorithm is proposed for
nating the review of this manuscript and approving it for publication was lessening haze effects using morphological reconstruction to
Prof. Khan M. Iftekharuddin. (Corresponding author: Eduardo Cabal-Yepez.) preserve important structures in all its stages and refine
S. Salazar-Colores and J. M. Ramos-Arreguin are with the Faculty of
Informatics, Universidad Autónoma de Querétaro, Queretaro 76230, Mexico the transmission-map computation of single images captured
(e-mail: [email protected]; [email protected]). by vision systems, making it feasible for use in online
E. Cabal-Yepez, L. M. Ledesma-Carrillo, and S. Ledesma are with the applications.
Division de Ingenierias del Campus Irapuato-Salamanca, Universidad de
Guanajuato, Guanajuato 38944, Mexico (e-mail: [email protected]; The remainder of the manuscript is organized as follows.
[email protected]; [email protected]). Section II provides a brief literature review of related work.
G. Botella is with the Departamento de Arquitectura de Computadores Section III introduces an overview of the theoretical back-
y Automatica, Universidad Complutense de Madrid, 28040 Madrid, Spain
(e-mail: [email protected]). ground utilized in this work. Section IV describes the proposed
Digital Object Identifier 10.1109/TIP.2018.2885490 technique and its implementation in depth. Obtained results are
1057-7149 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on October 01,2023 at 20:42:37 UTC from IEEE Xplore. Restrictions apply.
2358 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 28, NO. 5, MAY 2019

presented in Section V and compared against those obtained

from six different dehazing algorithms. Finally, conclusions
are given in Section VI.

II. R ELATED W ORK

As described before, dehazing algorithms can be divided
into two groups. Among those requiring additional information
are the following. In [7], a dehazing method is proposed
using a visible image with a near-infrared image of the same Fig. 1. Atmospheric scattering model.
scene to obtain a dehazed color image. In [8], a system
for browsing, enhancing, and manipulating casual outdoor In [21], a segmentation algorithm called mean shift filtering is
photographs is introduced by combining them with preex- proposed for refining the transmission map generated by DCP
isting georeferenced digital terrain and urban models. Other to improve the reconstructed image. In [22], a fusion-based
works estimate the captured-scene depth utilizing multiple transmission estimation is introduced, combining the fusion
images taken under different weather conditions [9], [10]. weighting scheme and the atmospheric light computed from
Some algorithms remove haze by taking advantage of light the Gaussian-based dark channel method. In [23], a linear
polarization [11], [12]. However, their main disadvantage is color-attenuation prior is proposed based on the difference
that the requirements are not always available or are difficult between the pixel brightness and saturation in hazy images.
to comply with. For example, the infrared algorithm requires a In [24], a pixel-based algorithm relying on a nonlocal prior
modified camera to get the infrared data. It is worth mention- and the assumption that a haze-free image can be faithfully
ing that algorithms requiring multiple images are difficult to represented with just a few hundred distinct colors is proposed.
implement for online applications [9], whereas algorithms that From the reviewed work in recent literature, it is clear that
use images with different levels of polarization are sensitive present research on diminishing haze effects in images focuses
to vibrations or movement [10]. on improving the speed of transmission map computation
Algorithms using a single image to diminish haze effects are for the original DCP technique [16], [25], without losing
among the most relevant research. In [13], haze is removed the image quality [15]; however, these algorithms have only
by maximizing local contrast in the analyzed image. It was achieved partial improvements [1], [5]. The contribution of
noticed that a haze-free image should have higher contrast than this work is a novel fast algorithm for reducing haze effects,
a hazed image. In [14], independent component analysis (ICA) using morphological reconstruction in all its stages to refine
is used to address haze by estimating the albedo in the captured the transmission map computation of single images captured
scene and the propagation medium, whereas transmission by a vision system, making it feasible for online applications.
and shading of the surface are correlated locally. In [15], The proposed method is based on the DCP technique and to
a transmission map is obtained by analyzing the optical model the best of the authors’ knowledge, the introduced approach,
and recasting the initial transmission map under an additional which drastically reduces processing time, has never been used
boundary prior. In [16], DCP is proposed for estimating the for this purpose. In this regard, the technique proposed in
image transmission map, which consists of estimating each this paper is at least two orders of magnitude faster than
element’s depth in the image, and it is essential for restoring the original DCP algorithm. Furthermore, obtained qualita-
it. The presented results are satisfactory; nonetheless, the trans- tive and quantitative results using the Peak Signal-to-Noise
mission map generates halos around the restored-image edges. Ratio (PSNR) and the Structural Similarity (SSIM) index show
To improve their algorithm, the authors introduced soft matting that the proposed method achieves better results than other
for polishing the transmission map, sacrificing computation recently introduced state-of-the-art algorithms for this subject.
time and jeopardizing its utilization in fast image processing
systems for online applications. Consequently, DCP has been III. BACKGROUND
used in combination with other algorithms, looking for fast In this section, concepts such as atmospheric scattering,
computation of accurate transmission maps [6]. For example, DCP and morphological reconstruction are presented as the
in [17], atmospheric scattering and DCP theories are used framework of the proposed technique for diminishing haze in
for single-image dehazing. A transmission map is estimated an image.
using a fast average filter, the region-projection method is
adopted to obtain the atmospheric light, and image color
compensation is implemented using the Weber-Fechner Law. A. Atmospheric Scattering Model
In [18], a median filter is included for removing halos from Fig. 1 depicts how a scene might be affected by envi-
an image. In [19], a bilateral filter and a transmission filter ronmental pollution producing haze in the captured image.
are used for removing noise while preserving the edges, Its corresponding mathematical model is described by (1) [26].
and the transmission map is refined to recover the scene.
I (χ) = J (χ)t (χ) + A (1 − t (χ)) (1)
In [20], a correction of white balance is introduced in DCP to
improve the image quality by obtaining two components: the In (1), I (χ) is the observed intensity in each channel, R,
reflected light in the image and the light from the environment. G and B (RGB), of the pixel χ = (x, y) in the captured scene

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on October 01,2023 at 20:42:37 UTC from IEEE Xplore. Restrictions apply.
SALAZAR-COLORES et al.: FAST IMAGE DEHAZING ALGORITHM USING MORPHOLOGICAL RECONSTRUCTION 2359

Fig. 2. Dehazing algorithm proposed in [16].

by a camera. J (χ) is the vector intensity of the scene’s original work, a methodology to dehaze an image is proposed by
area in the real world that corresponds to the pixel χ = (x, y). combining DCP with morphological reconstruction to improve
A is the color vector of the global atmospheric light. t (χ) is the computation speed for the transmission map estimation of
called the transmission, which describes the portion of light an image.
that is not scattered or absorbed and reaches the camera. Under
a homogeneous atmospheric condition, the transmission t (χ) C. Morphological Reconstruction
can be expressed by (2).
The morphological reconstruction is based on mathematical
t (χ) = e−β d(χ) (2) morphology operations, and it is used to simplify images
and preserve main features as object shapes. The proposed
where β is the atmosphere scattering coefficient and d(χ) is algorithm uses the concepts of opening and closing for
the depth of the element χ; therefore, a map of transmission is image reconstruction, which are based on techniques such
proportional to a depth map. The main difficulty in obtaining as grayscale erosion, grayscale dilation, geodesic erosion and
J (χ) rests upon the fact that, when an image is captured, dilation, and reconstruction by erosion and dilation, consider-
A and t (χ) in (1) are unknown. Therefore, the use of priors ing a grayscale image I, a flat structuring element S with an
and some assumptions are necessary to find an approximation arbitrary shape formed by pixels s, and the position (x, y) of
of J (χ). The original algorithm, which is based on DCP and the pixel χ in the image I.
was proposed in [16], makes it possible to conduct an accurate
estimation of A and t (χ), but at a high computational cost.
Hence, it is necessary to look for an alternative technique able D. Grayscale Erosion
to provide satisfactory results with less processing time. The erosion ε B of I by S is defined as the minimum value
of the image in the matching region with s when the source
B. Dark Channel Prior (DCP) is focused at the position χ = (x, y) [27], as described in (5)
DCP is a statistical observation of what happens in images [ε B (I )] (χ) = min I (χ + s) (5)
s∈S
that do not present the problem of haze and are acquired in
external environments. For most of the patches in an image
E. Grayscale Dilation
that do not represent the sky, there is at least a low-intensity
pixel in each color channel (R, G or B). Thus, the minimal The dilation δ B of I by S is defined as the maximum value
intensity of one pixel in such patches has a value near 0 [16]. of the image in the window defined by the origin of S when
For an image I (χ), the dark channel I dark (χ) is defined S is at χ = (x, y), and it is expressed by (6)
by (3) [16]. [δ B (I )](χ) = max I (χ + s) (6)
  s∈S
 c 
I dark
(χ) = min min I (y) (3)
C∈R,G,B y∈(χ) F. Morphological Reconstruction
A morphological reconstruction is expressed in terms of
where (χ) is the patch centered in χ, I C is the color
geodesic dilation and geodesic erosion.
channel C (i.e., R, G, or B) from I , and y is the pixel contained
in (χ).
Hence, DCP is expressed by (4). G. Geodesic Dilation
In grayscale images, the size-1 geodesic dilation δg(1) of I
I dark (χ) → 0 (4)
(marker image) with respect to F (mask image) is defined by
In [16], the algorithm shown in Fig. 2 is proposed to
δg(1) (I ) = δ (1) (I ) ∧ F (7)
dehaze an image I, of height h and width w pixels, using
DCP. The algorithm produces excellent results, but its main where I and F are the same size, the intensity relation I ≤ F
drawback is its computational complexity; therefore, in this holds for all pixels in the images, and ∧ is the operator

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on October 01,2023 at 20:42:37 UTC from IEEE Xplore. Restrictions apply.
2360 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 28, NO. 5, MAY 2019

Fig. 3. Flowchart of the proposed algorithm.

(n)
minimum. Hence, the n-size geodesic dilation δg is as given K. Opening and Closing by Reconstruction
in (8). Opening and closing by reconstruction restore the forms
of objects that the structuring element exceeds after each
δg(n) (I ) = δg(1) [δg(n−1) (I )] (8) process. The reconstruction precision depends on the similarity
(0) of the structuring element to the shape of the objects. The
with δG (I ) = I . n-size opening by reconstruction of an image I is described
by (13). Analogously, closing by reconstruction is defined in
equation (14).
H. Geodesic Erosion
(1) γ R(n) (I ) = R δI [ε(n) (I )] (13)
In grayscale images, the size-1 geodesic erosion εg of I
(n)
(marker image) with respect to G (the mask image) is φ R (I ) = R εI [δ (n) (I )] (14)
defined by
IV. T HE P ROPOSED A LGORITHM
εg(1) (I ) = ε(1) (I ) ∨ G (9) The dark channel described in (3) can be expressed regard-
ing a morphological erosion as:
where I and F are the same size, the intensity relation   
I ≥ F holds for all pixels in the images, and ∨ is the operator
I dark
(χ) = ε S min I C (y) (χ) (15)
maximum. The n-size geodesic erosion εg(n) is shown in (10) C∈R,G,B

Based on (15), a morphological method that can yield an

εg(n) (I ) = εg(1) [εg(n−1) (I )], where εg(0) (I ) = I. (10) image without haze or halos is proposed as depicted in Fig. 3,
and it is described as follows:
Considering I (χ) as an RGB image of height h and width w
I. Reconstruction by Dilation
pixels, where every pixel position is given by χ = (x, y),
Reconstruction by dilation is obtained through the geodesic the dark channel I dark (χ) is computed from I (χ) according
dilation of marker I on the mask F, iterating until stability is to (15), with the atmospheric light A given by
reached, and it is denoted by Rgδ (I ) as shown in (11). ⎛ ⎞
3 
A = max I C ⎝arg max I dark (χ) ⎠ (16)
Rgδ (I ) = Rg(i) (I ) (11) C=1 χ∈(0.1%∗h∗w)
(i)
Stability is reached when δg (I ) = δg
(i+1)
(I ). The image I (χ) is normalized concerning the atmospheric
light A to obtain dark-channel values between zero and one,
as given by
J. Reconstruction by Erosion I (χ)
I N (χ) = (17)
Reconstruction by erosion is obtained through the geodesic A
erosion of marker I on the mask G, iterating until stability is The minimal channel is defined as:

reached. It is denoted by Rgε as shown in (12).
I min (χ) = min I NC (y) (18)
C∈R,G,B
Rgε (I ) = εg(i) (I ) (12) The initial transmission map is computed as:

Stability is achieved when εg(i) (I ) = εg(i+1) (I ). t1 (χ) = 1 − I min (χ) (19)

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on October 01,2023 at 20:42:37 UTC from IEEE Xplore. Restrictions apply.
SALAZAR-COLORES et al.: FAST IMAGE DEHAZING ALGORITHM USING MORPHOLOGICAL RECONSTRUCTION 2361

Fig. 4. Morphological process.

Fig. 4 shows each stage of the proposed morphologic

process used for refining the initial transmission, which uses
a square structuring element S.
In the first stage, a closing by reconstruction is
performed as:
Fig. 5. Examples of transmission maps generated by the proposed algorithm,
t2 (χ) = φ 1R (t1 (χ)) (20) where (a) (c) are the input images and (b) (d) are the respective transmission
maps.
This operation removes small dark elements over the image
structuring element.
An opening by reconstruction is carried out later as: from [28] and [29]. A 15-pixel, square structuring element B
t3 (χ) = γ R1 (t2 (χ)) (21) was used during the experimentation. The proposed algo-
rithm was implemented on an Intel i5-3320 microprocessor
This procedure deletes small objects that are clearer than at 2.6 GHz utilizing MATLAB.
the environment, and its size is smaller than that of S. These
objects are saved by A. Qualitative Comparison
R(χ) = t1 (χ) − t3 (χ) (22) Fig. 6 shows real-world outside images affected by haze that
are processed through different state-of-the-art approaches,
To recover the range of values for the original transmission, such as He et al. [15] (2017), (c) Tarel et al. [18] (2009),
the intervals of t3 (χ) are changed to the intervals of t1 (χ), and (d) Gibson et al. [30] (2013), (e) Kim et al [31] (2013),
the result is saved in t3 (χ), as shown by: (f) Zhu et al. [23] (2015), (g) Berman et al. [24] (2016),

max(t1 (χ)) − min(t1 (χ)) and the proposed methodology in this work, to carry out
t3 (χ) = t3 (χ) − min(t3 (χ)) a qualitative comparison of the dehazed images obtained
max(t3 (χ)) − min(t3 (χ))
+ min(t1 (χ)) (23) through each. He et al.’s technique and the proposed method
produce an adequate image restoration, qualitatively, since
The refined transmission tmor f (χ) is recovered through: no visible defects are observed. The obtained results from
tmor f (χ) = t3 (χ) + R(χ) (24) applying Tarel et al.’s approach show some artifacts in the
restored images (Fig. 6c). The color is visibly affected in
Fig. 5 shows two examples of transmission maps tmor f (χ) the output images (Fig. 6d) from applying Gibson et al.’s
produced by the proposed algorithm. technique. The Kim et al. and Zhu et al. schemes present
Finally, the transmission map tmor f (χ) and the atmospheric slight halos in their recovered images (Fig. 6e and Fig. 6f,
light A are applied to the scattering model to retrieve the respectively). Finally, Berman et al.’s method generates leaves
ImageJ(χ) without haze, as shown by with nonhomogeneous color in the obtained images (Fig. 6g).
I (χ) − A
J (χ) = +A (25)
tmor f (χ) B. Quantitative Comparison
Fig. 7 shows eleven images used for numerical analysis of
V. E XPERIMENTAL R ESULTS the proposed method’s performance compared to other state-
The proposed algorithm’s effectiveness is assessed in of-the-art approaches in the reviewed literature. These images
this section through experimentation. The obtained results have different characteristics, and they are contaminated with
are compared qualitatively and quantitatively against those haze simulated by using the model of atmospheric scatter-
from different approaches in recent literature. The images ing considering the atmospheric light vector [0.92, 0.95, 1]
used for validating the proposed technique were taken as in [29]; they are then treated during experimentation

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on October 01,2023 at 20:42:37 UTC from IEEE Xplore. Restrictions apply.
2362 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 28, NO. 5, MAY 2019

Fig. 6. Comparison of recovered outside real-world scenes without haze. (a) Input images, the results from (b) He et al. [15], (c) Tarel et al. [18],
(d) Gibson et al. [30], (e) Kim et al [31], (f) Zhu et al. [23], (g) Berman et al. [24], and (h) proposed algorithm.

TABLE I
O BTAINED RESULTS FROM PSNR ANALYSIS

with the algorithms of He et al. [15], Tarel et al. [18], where M AX 2I H F (χ) is the squared maximum-possible pixel
Gibson et al. [30], Kim et al. [31], Zhu et al. [23], and value of the image I H F (χ), and the corresponding MSE is
Berman et al. [24], as well as the approach proposed in this given as:
work.
w h
The metrics used for quantitatively evaluating the pro- 1
MSE = (J (χ) − I H F (χ))2 (27)
posed algorithm’s performance against that of previous (w × h)
x=1 y=1
approaches in the reviewed literature are the peak signal-
to-noise ratio (PSNR) and the structural similarity (SSIM) with χ representing the pixel position (x, y) in the image, with
index (27) [32]. width and height w and h, respectively. The higher the PSNR
PSNR is the proportion between the maximum-possible value, the better the image approximation J (χ).
power of an image and the power of corrupting noise The SSIM index is used to measure the similarity between
that affects the fidelity of its representation. Given a two images, and it considers three aspects in restored images:
haze-free image I H F (χ) and the corresponding mean square lighting l(χ), contrast c(χ), and structure s(χ).
error (MSE) between I H F (χ) and its restored approxima-
tion J (χ), the PSNR is defined by SSIM(χ) = f (l(χ), c(χ), s(χ)) (28)
 
M AX 2I H F (χ) The SSIM index is a decimal value between −1 and 1.
P S N R = 10 log10 (26) SSIM = 1 only when two images with identical sets of data
MSE
are compared.

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on October 01,2023 at 20:42:37 UTC from IEEE Xplore. Restrictions apply.
SALAZAR-COLORES et al.: FAST IMAGE DEHAZING ALGORITHM USING MORPHOLOGICAL RECONSTRUCTION 2363

Fig. 7. Comparison of recovered scenes without haze. (a) Ground-truth images, (b) input images, the results from (c) He et al. [15], (d) Tarel et al. [18],
(e) Gibson et al. [30], (f) Kim et al. [31], (g) Zhu et al. [23], (h) Berman et al. [24], and (i) the proposed algorithm.

Table I and Table II show the PSNR and SSIM index four orders of magnitude. In this regard, it is important to point
metrics, respectively, obtained from each applied approach, out that, although the approach in He et al. [15] obtains higher
and they quantitatively demonstrate the performance superi- restoration metrics (PSNR and SSIM), the method introduced
ority of the proposed technique against previously introduced in this work is at least three orders of magnitude faster than it.
approaches for image dehazing restoration. Alternately, the restoration obtained from the introduced tech-
nique is quantitatively superior (PSNR y SSIM) and at least
one order of magnitude faster than the image restoration from
C. Time Performance Analysis all remaining approaches in the reviewed literature.
Table III presents a time-consumption comparison of the Table IV presents a peak memory-consumption comparison
proposed method against state-of-the-art algorithms in the of the proposed algorithm against those same approaches
reviewed literature. This table includes information regarding in Table III. From Table IV, it is worth noting that the proposed
the treated image’s size in pixels. A 2.6-GHz, 64-bit Intel approach has lower memory utilization than all the other
Core i5-3320 with 12 GB of RAM was used under the considered techniques, except that from Gibson et al. [30],
Windows 10 operating system to implement the algorithms for some cases. However, the proposed method surpasses
in MATLAB 2018a. From this table, it is worth noting that that from Gibson et al. with regard to image reconstruction
the proposed approach achieves faster processing time than performance and time consumption by at least one order of
any other technique in the reviewed literature, from one up to magnitude.

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on October 01,2023 at 20:42:37 UTC from IEEE Xplore. Restrictions apply.
2364 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 28, NO. 5, MAY 2019

TABLE II
O BTAINED RESULTS FROM SSIM ANALYSIS

TABLE III dehazing algorithms. In this regard, most of the recently

C OMPUTATIONAL TIME ANALYSIS ( SECONDS ). proposed algorithms combine DCP with different techniques
that look for fast computation of accurate transmission maps,
aiming to preserve the image quality at the cost of long
computation times. Therefore, in this work, a novel fast algo-
rithm for lessening haze effects using DCP and a newly intro-
duced approach was compared qualitatively and quantitatively
against state-of-the-art techniques; morphological reconstruc-
tion is utilized to preserve important structures of the image in
all stages. The reviewed literature about the subject involves
image reconstruction performance (utilizing PSNR and SSIM
index), computation time and memory utilization. From the
obtained results, the performance superiority of the proposed
method for diminishing haze effects during image reconstruc-
tion is clear, as is its high-speed processing time that surpasses
TABLE IV all other techniques in the reviewed literature from one to four
M EMORY UTILIZATION (M EGABYTES ) orders of magnitude with less memory utilization. Therefore,
the proposed methodology for image dehazing introduced in
this work offers a fast, high-performance dehazing technique
suitable for online vision system applications.
Future work will be directed to improve the proposed
algorithm’s performance in handling sky regions due to DCP
limitations.

R EFERENCES
[1] C. Chengtao, Z. Qiuyu, and L. Yanhua, “A survey of image dehazing
approaches,” in Proc. Control Decis. Conf. (CCDC), Qingdao, China,
2015, pp. 3964–3969.
[2] F. Yuan and H. Huang, “Image haze removal via reference retrieval and
scene prior,” IEEE Trans. Image Process., vol. 27, no. 9, pp. 4395–4409,
Sep. 2018.
[3] B. Li et al., “Benchmarking single-image dehazing and beyond,” IEEE
VI. C ONCLUSION Trans. Image Process., vol. 28, no. 1, pp. 492–505, Jan. 2019.
[4] K. B. Gibson, D. T. Vo, and T. Q. Nguyen, “An investigation of dehazing
Outdoor images are exposed to adverse weather conditions effects on image and video coding,” IEEE Trans. Image Process., vol. 21,
such as haze, which reduces the visibility of details in a no. 2, pp. 662–673, Feb. 2012.
[5] V. Sahu and M. Singh, “A survey paper on single image dehazing,” Int.
captured scene. Hence, several research works have focused on J. Recent Innov. Trends Comput. Commun., vol. 3, no. 2, pp. 85–88,
diminishing haze effects in images by designing and applying Feb. 2015.

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on October 01,2023 at 20:42:37 UTC from IEEE Xplore. Restrictions apply.
SALAZAR-COLORES et al.: FAST IMAGE DEHAZING ALGORITHM USING MORPHOLOGICAL RECONSTRUCTION 2365

[6] Y.-H. Lai, Y.-L. Chen, C.-J. Chiou, and C.-T. Hsu, “Single-image [31] J.-H. Kim, W.-D. Jang, J.-Y. Sim, and C.-S. Kim, “Optimized contrast
dehazing via optimal transmission map under scene priors,” IEEE Trans. enhancement for real-time image and video dehazing,” J. Vis. Commun.
Circuit Syst. Video Technol., vol. 25, no. 1, pp. 1–14, Jan. 2015. Image Represent., vol. 24, no. 3, pp. 410–425, Apr. 2013.
[7] L. Schaul, C. Fredembach, and S. Süsstrunk, “Color image dehazing [32] W. Wang, W. Li, Q. Wang, and M. Qi, “Multiscale single image dehazing
using the near-infrared,” in Proc. 16th IEEE Int. Conf. Image Process. based on adaptive wavelet fusion,” Math. Problems Eng., vol. 2015,
(ICIP), Cairo, Egypt, Nov. 2009, pp. 1629–1632. 2015, Art. no. 131082, doi: 10.1155/2015/131082.
[8] J. Kopf et al., “Deep photo: Model-based photograph enhancement
and viewing,” ACM Trans. Graph., vol. 27, no. 5, pp. 116-1–116-10,
Dec. 2008.
[9] P. Carr and R. Hartley, “Improved single image dehazing using geome-
try,” in Proc. Digit. Image Comput., Techn. Appl. (DICTA), Melbourne, Sebastian Salazar-Colores received the B.S.
VIC, Australia, 2009, pp. 103–110. degree in computer science from the Universidad
[10] Y. Y. Schechner, S. G. Narasimhan, and S. K. Nayar, “Instant dehazing Autónoma Benito Juárez de Oaxaca and the M.S.
of images using polarization,” in Proc. IEEE Comput. Soc. Conf. degree in electrical engineering from the Univer-
Comput. Vis. Pattern Recognit. (CVPR), Kauai, HI, USA, Dec. 2001, sity of Guanajuato. He is currently pursuing the
pp. I-325–I-332. Ph.D. degree with the Universidad Autónoma de
[11] Q. Liu, H. Zhang, M. Lin, and Y. Wu, “Research on image dehazing Querétaro. His research interests include image
algorithms based on physical model,” in Proc. Int. Conf. Multimedia processing and computer vision.
Technol. (ICMT), Hangzhou, China, 2011, pp. 467–470.
[12] X. Wang, X. Jin, G. Xu, and X. Xu, “A multi-scale decomposition based
haze removal algorithm,” in Proc. Int. Conf. Remote Sens., Environ.
Transp. Eng. (RSETE), Nanjing, China, 2012, pp. 1–4.
[13] R. T. Tan, “Visibility in bad weather from a single image,” in Proc.
IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Anchorage, AK,
USA, Jun. 2008, pp. 1–8. Eduardo Cabal-Yepez (M’09) received the M.Eng.
[14] R. Fattal, “Single image dehazing,” in Proc. ACM SIGGRAPH ACM degree from the Facultad de Ingenieria Mecanica
Trans. Graph. (TOG), vol. 27, no. 3, Aug. 2008, pp. 72:1–72:9. Electrica y Electronica, Universidad de Guanajuato,
[15] L. He, J. Zhao, N. Zheng, and D. Bi, “Haze removal using the difference- Mexico, in 2001, and the Ph.D. degree from the
structure-preservation prior,” IEEE Trans. Image Process., vol. 26, no. 3, University of Sussex, U.K., in 2007. In 2008,
pp. 1063–1075, Mar. 2017. he joined the Division de Ingenierias del Cam-
[16] K. He, J. Sun, and X. Tang, “Single image haze removal using dark pus Irapuato-Salamanca de la Universidad de
channel prior,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 12, Guanajuato, where he is currently a Titular Pro-
pp. 2341–2353, Dec. 2011. fessor. He is also a National Researcher with the
[17] W. Wang, F. Chang, T. Ji, and X. Wu, “A fast single-image dehazing Consejo Nacional de Ciencia y Tecnologia, Mexico.
method based on a physical model and gray projection,” IEEE Access, He has authored more than 50 papers in international
vol. 6, pp. 5641–5653, 2018, doi: 10.1109/ACCESS.2018.2794340. journals and conferences. His current research interests are digital image
[18] J.-P. Tarel and N. Hautière, “Fast visibility restoration from a single and signal processing, artificial intelligence, robotics, smart sensors, real-time
color or gray level image,” in Proc. 12th IEEE Int. Conf. Comput. processing, mechatronics, FPGAs, and embedded systems.
Vis. (ICCV), Kyoto, Japan, Sep./Oct. 2009, pp. 2201–2208.
[19] N. Gundawar and V. B. Baru, “Improved single image dehazing by
fusion,” Int. J. Res. Eng. Technol., vol. 3, no. 5, pp. 432–437, May 2014.
[20] R. He, Z. Wang, H. Xiong, and D. D. Feng, “Single image dehazing with
white balance correction and image decomposition,” in Proc. Int. Conf. Juan M. Ramos-Arreguin (SM’17) received the
Digit. Image Comput. Techn. Appl. (DICTA), Fremantle, WA, Australia, M.S. degree in electrical engineering option instru-
2012, pp. 1–7. mentation and digital systems from the University
[21] X. Zhu, Y. Li, and Y. Qiao, “Fast single image dehazing through of Guanajuato and the Ph.D. degree in mecha-
edge-guided interpolated filter,” in Proc. 14th Int. Conf. Mach. Vis. tronics science from the Centro de Ingeniería y
Appl. (MVA), Tokyo, Japan, 2015, pp. 443–446. Desarrollo Industrial. Since 2009, he has been with
[22] J. M. Guo, J.-Y. Syue, V. R. Radzicki, and H. Lee, “An efficient the Informatics Department, Universidad Autónoma
fusion-based defogging,” IEEE Trans. Image Process., vol. 26, no. 9, de Querétaro, where he works as a Researcher and a
pp. 4217–4228, Sep. 2017. Lecturer. His research interests include mechatronics
[23] Q. Zhu, J. Mai, and L. Shao, “A fast single image haze removal algorithm and embedded systems.
using color attenuation prior,” IEEE Trans. Image Process., vol. 24,
no. 11, pp. 3522–3533, Nov. 2015.
[24] D. Berman, T. Treibitz, and S. Avidan, “Non-local image dehazing,” in
Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Las Vegas,
NV, USA, Jun. 2016, pp. 1674–1682.
[25] T. M. Bui and W. Kim, “Single image dehazing using color ellipsoid Guillermo Botella received the M.A.Sc. degree
prior,” IEEE Trans. Image Process., vol. 27, no. 2, pp. 999–1009, in physics, the M.A.Sc. degree in electronic engi-
Feb. 2018. neering, and the Ph.D. degree from the Univer-
[26] H. Zhang, Q. Liu, F. Yang, and Y. Wu, “Single image dehazing sity of Granada, Spain, in 1998, 2001, and 2007,
respectively. He was a Research Fellow funded by
combining physics model based and non-physics model based methods,”
J. Comput. Inf. Syst., vol. 9, no. 4, pp. 1623–1631, Feb. 2013. EU working with the Department of Architecture
and Computer Technology, Universidad de Granada,
[27] P. Soille, Morphological Image Analysis: Principles and Applications,
Spain, and also with the Vision Research Labora-
2nd ed. New York, NY, USA: Springer-Verlag, 2004.
tory, University College London, U.K. After that,
[28] M. Sulami, I. Glatzer, R. Fattal, and M. Werman, “Automatic recovery of he joined as an Assistant Professor the Department
the atmospheric light in hazy images,” in Proc. IEEE Int. Conf. Comput. of Computer Architecture and Automation, Univer-
Photogr. (ICCP), Santa Clara, CA, USA, May 2014, pp. 1–11. sidad Complutense de Madrid, Spain. He was a Visiting Professor with the
[29] R. Fattal, “Dehazing using color-lines,” ACM Trans. Graph., vol. 34, Department of Electrical and Computer Engineering, Florida State University,
no. 1, Nov. 2014, Art. no. 13. Tallahassee, USA, from 2008 to 2012. He has authored more than 60 papers in
[30] K. B. Gibson and T. Q. Nguyen, “Fast single image fog international journals and conferences. His current research interests include
removal using the adaptive wiener filter,” in Proc. IEEE Int. digital signal processing for very large scale integration (VLSI), FPGAs,
Conf. Image Process. (ICIP), Melbourne, VIC, Australia, Sep. 2013, vision algorithms, and IP protection of VLSI- and FPGA-based systems.
pp. 714–718. He serves regularly as a reviewer for several IEEE journals.

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on October 01,2023 at 20:42:37 UTC from IEEE Xplore. Restrictions apply.
2366 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 28, NO. 5, MAY 2019

Luis M. Ledesma-Carrillo (S’13) received the Sergio Ledesma (M’13) received the M.S. degree
M.Eng. and Ph.D. degrees (Hons.) from the Uni- from the University of Guanajuato, while working on
versity of Guanajuato, Mexico, in 2013 and 2017, the Setup of Internet in Mexico and the Ph.D. degree
respectively. In 2017, he joined the Division de from the Stevens Institute of Technology, Hoboken,
Ingenierias del Campus Irapuato-Salamanca de la NJ, USA, in 2001. After graduating, he worked
Universidad de Guanajuato, where he is cur- for Barclays Global Investors, USA, as part of the
rently an Associate Professor. His fields of inter- IT–HR Group. He has worked as a software engineer
est include digital signal and image processing on for several years. He is currently the Creator of the
field-programmable gate arrays for applications in Software Neural Lab and Wintempla. He is also a
robotic vision and optics. Research Professor with the Universidad de Gua-
najuato, Mexico. His areas of interests are artificial
intelligence and software engineering.

Authorized licensed use limited to: Universitatea Tehnica din Cluj-Napoca. Downloaded on October 01,2023 at 20:42:37 UTC from IEEE Xplore. Restrictions apply.