Single image dehazing using a

multilayer perceptron

Sebastián Salazar-Colores
Ivan Cruz-Aceves
Juan-Manuel Ramos-Arreguin

Sebastián Salazar-Colores, Ivan Cruz-Aceves, Juan-Manuel Ramos-Arreguin, “Single image

dehazing using a multilayer perceptron,” J. Electron. Imaging 27(4),
043022 (2018), doi: 10.1117/1.JEI.27.4.043022.
Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Journal of Electronic Imaging 27(4), 043022 (Jul∕Aug 2018)

Single image dehazing using a multilayer perceptron

Sebastián Salazar-Colores,a Ivan Cruz-Aceves,b,* and Juan-Manuel Ramos-Arreguinc
a
Universidad Autónoma de Querétaro, Facultad de Informática, Santiago de Querétaro, Querétaro, México
b
Consejo Nacional de Ciencia y Tecnología, Centro de Investigación en Matemáticas, Guanajuato, Guanajuato, México
c
Universidad Autónoma de Querétaro, Facultad de Ingeniería, Santiago de Querétaro, Querétaro, México

Abstract. This paper presents an algorithm to improve images with hazing effects. Usually, the dehazing meth-
ods based on the dark channel prior make use of two different stages to compute the transmission map of
the input image. The stages are the transmission map estimation and a transmission map refinement.
However, the main disadvantage of these strategies is the trade-off between accurate restoration and computa-
tional time. The proposed method uses a multilayer perceptron to compute the transmission map directly from
the minimum channel and a contrast stretching technique to improve the dynamic range of the restored image.
The multilayer perceptron is trained in terms of mean squared error using a training set of 80 images. To evaluate
the restoration quality, the metrics of peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) index
are used. The experimental results have proven that the proposed method achieves superior performance in
terms of restoration quality (PSNR = 18.77, SSIM index = 0.8454) compared with nine state-of-the-art dehazing
methods. In addition, based on the average computational time achieved by the proposed method (0.52 s using
a test set of 46 images), it can be highly suitable for real-time applications. © 2018 SPIE and IS&T [DOI: 10.1117/1.JEI
.27.4.043022]

Keywords: artificial neural network; dark channel prior; defogging; image enhancement; multilayer perceptron; single image dehazing.
Paper 180180 received Mar. 5, 2018; accepted for publication Jul. 3, 2018; published online Jul. 25, 2018.

1 Introduction based on a statistic analysis of haze-free images. They use

Image enhancement is a very active research area since the the DCP to compute the transmission and restore the
image quality has a direct influence on computer vision sys- image using an atmospheric scattering model. The DCP
tems. In general, outdoor images are susceptible to different algorithm obtains accurate results in most of the cases.
atmospheric conditions, especially haze, fog, and smog. The The main drawback of DCP method is that it requires an
main effects caused by these atmospheric conditions are intensive computational processing to refine a generated
images with low contrast, distorted colors, and reduced coarse transmission, making it inadequate for many applica-
visibility. Consequently, the performance of systems, such tions where a short response time is required. In order to
as surveillance,1 target tracking,2 remote sensing,3 traffic improve the computational time of the DCP along with
control,4 and autonomous driving,5 could be highly affected. the qualitative results of the DCP, different works have
In the literature, the problem of eliminating or reducing been introduced. Pang et al.15 proposed to refine the trans-
the effects of atmospheric scattering has been widely studied. mission by applying an edge-preserving smoothing operator
The most representative strategies can be classified accord- named guided filter using a patch radius of 30 pixels.
ing to the input information such as multiple images,6,7 infra- Nishino et al.16 used a probabilistic Bayesian method that
red radiation,8 three-dimensional data,9 and the use of a combines the estimation of the albedo and depth of a
single image.10 However, the most studied approach is based scene. Gibson et al.17 proposed a variant of DCP replacing
on a single image, because additional data are not commonly the minimum operator with the median operator. Zhu et al.18
available.11 proposed a linear color attenuation based on the difference of
In the single image dehazing problem, several represen- the saturation and the brightness of the pixels. Berman et al.19
tative works have been proposed. Tarel and Hautiere12 pro- used a pixel-oriented algorithm, where pixels are clustering
posed a method based on two different filters: a median filter in the RGB space to estimate a haze-line used to recover the
and a filter for preserving edges and corners. The method haze-free image.
estimates the scattered airlight for recovering a haze-free In particular, artificial neural networks (ANN) have begun
image. Fattal13 assumed that there is a nonlocal relation to attract more attention in computer vision systems. ANN
between the shading and the signal. Since the method are computing systems inspired by biological neural net-
tries to solve a nonlinear inverse problem by computing a works trying to emulate the microstructure of the brain.
transmission map, this approach generates suitable results One of the most popular ANN classes is the multilayer per-
in regions where the shading and transmission are different. ceptron (MLP) because of its robustness and ease of
However, it does not perform well in regions where they are implementation.20 It is an ANN with at least three layers
similar. He et al.14 introduced the dark channel prior (DCP) of neurons and it commonly uses a supervised learning strat-
egy called backpropagation.21 Recently, it has been used to

*Address all correspondence to: Ivan Cruz-Aceves, E-mail: ivan.cruz@cimat

.mx 1017-9909/2018/$25.00 © 2018 SPIE and IS&T

Journal of Electronic Imaging 043022-1 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Salazar-Colores, Cruz-Aceves, and Ramos-Arreguin: Single image dehazing using a multilayer perceptron

solve different problems in image processing, such as human 2.2 Dark Channel Prior
face recognition,22 skin segmentation,23 and denoising.24 In The DCP proposed by He et al.14 is based on empirical obser-
addition, different deep learning strategies have been vations about the statistical features in the intensity of dark
recently introduced such as Gaussian progress regression25 channel for haze-free images.
and convolutional neural networks (CNN),26–28 where the The dark channel I dark ðx; yÞ for a digital image Iðx; yÞ is
dehazing task is performed in terms of an supervised learning defined as follows:
problem, which is roughly processed using RGB images
directly as input data and the transmission maps as target I dark ðx; yÞ ¼ min ½ min I C ðzÞ;
EQ-TARGET;temp:intralink-;e004;326;683 (4)
C∈R;G;B z∈Ωðx;yÞ
variables.
In this work, a method for single image dehazing using an where Ωðx; yÞ is a patch of m × m (usually m ¼ 15)14 cen-
MLP to estimate the transmission map rather than RGB tered on ðx; yÞ, I C is the color channel of I: R, G or B, and z
images as input data are proposed. In the training step of represents the index for a pixel in the domain of Ωðx; yÞ.
the MLP 144,000 samples were used, they were obtained The transmission ^tðx; yÞ is estimated as
from a subset of 80 real-world images. In order to define
dark 
the optimal architecture, several MLP topologies were tested, I ðx; yÞ
^tðx; yÞ ¼ 1 − ω ;
EQ-TARGET;temp:intralink-;e005;326;604 (5)
and the best topology was selected based on the mean A
squared error (MSE). To evaluate the performance of the pro-
where ω is a value that defines the desired level of restora-
posed method, an analysis was carried out using the peak
tion, being 1 the highest possible value.
signal-to-noise ratio (PSNR) and the similarity structural
The atmospheric light A is defined as
index measure (SSIM).29 The tests were performed in
both real-world and synthetic images. X
3

The content of the present paper is organized as follows: A ¼ max

EQ-TARGET;temp:intralink-;e006;326;534 I C f arg max ½I dark ðx; yÞg; (6)
C¼1 ðx;yÞ∈ð0.1%hwÞ
Section 2 describes the fundamentals of image dehazing
along with the fundamentals of the MLP. Section 3 explains where h and w are the height and width of I dark ðx; yÞ and C
the proposed method. The details of the computational indicates each of the RGB channels of the input image
experiments are presented in Sec. 4, and conclusions are Iðx; yÞ. To estimate the image without affectations Jðx; yÞ,
given in Sec. 5. Eq. (3) is applied.
The DCP algorithm gets accurate values of A and tðx; yÞ.
However, as it is shown in Fig. 1, the use of windows of
2 Background m × m, when ^tðx; yÞ is estimated, generates artifacts on
the edges of the restored image Jðx; yÞ. The artifacts are gen-
2.1 Atmospheric Scattering Model erated because the morphology of the windows used in the
The atmospheric scattering is a physical phenomenon where DCP does not adapt to the complex geometric boundaries
the light passing through the particles in the atmosphere is needing an additional refinement step to remove or reduce
deviated from its straight path. The formation of an image this effect.11
can be explained using the atmospheric scattering model pro-
posed by McCartney and Hall,30 and it is defined as follows: 2.3 Multilayer Perceptron
ANN are learning machines inspired in a very simplified
Iðx; yÞ ¼ Jðx; yÞtðx; yÞ þ A½1 − tðx; yÞ;
EQ-TARGET;temp:intralink-;e001;63;357 (1) mathematical model from biological neural systems.31 In
these models, each processing element (PE) is called a neu-
where Iðx; yÞ corresponds to the captured image, Jðx; yÞ is ron, and the connection between PE’s forms a network. One
the image of the original scene without affectations, A is the of the first PE developed was the perceptron that consists of
color of atmospheric light, and tðx; yÞ is called transmission, multiple inputs Xi multiplied by an scalar value known as
which can be defined in a homogeneous atmosphere as weight W ij , and a bias bj is added. The obtained result is
processed by a predefined activation function f, which
tðx; yÞ ¼ e−βdðx;yÞ ; (2) can be described as follows:

X 
EQ-TARGET;temp:intralink-;e002;63;283

where β is the scattering coefficient of the atmosphere and Yj ¼ f ðW ij  X i þ bj Þ :

EQ-TARGET;temp:intralink-;e007;326;237 (7)
i
dðx; yÞ is the scene depth.
In order to get a haze-free image, Eq. (1) can be expressed The most common activation function f used in percep-
as follows: tron is the hyperbolic tangent function tanh that is expressed
as
Iðx; yÞ − A
Jðx; yÞ ¼
EQ-TARGET;temp:intralink-;e003;63;206 þ A: (3) tanhðxÞ ¼ 2 · σð2xÞ − 1; (8)
tðx; yÞ
EQ-TARGET;temp:intralink-;e008;326;161

where σðxÞ is
It is important to point out that in Eq. (3), additionally to ex
the input image Iðx; yÞ, the transmission map tðx; yÞ and σðxÞ ¼ : (9)
1 þ ex
EQ-TARGET;temp:intralink-;e009;326;119

atmospheric light A must to be determined. The DCP method

makes possible to obtain an accurate estimation of the The training of the ANN adjusts the weight and bias
variables tðx; yÞ and A. values, in order to obtain the desired output according to

Journal of Electronic Imaging 043022-2 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Salazar-Colores, Cruz-Aceves, and Ramos-Arreguin: Single image dehazing using a multilayer perceptron

Fig. 1 Result of the DCP. (a) Input image Iðx ; y Þ, (b) dark channel map with Ω ¼ 15, (c) estimation
of Jðx ; y Þ using Ω ¼ 15, (d) dark channel map with Ω ¼ 35, and (e) estimation of Jðx ; y Þ using
Ω ¼ 35.

the inputs combinations. For solving nonlinear separation

problems many perceptrons in one or more hidden layer
topologies are connected, this network is called MLP.
Backpropagation is an efficient training algorithm for MLP,
and it is the most used in the mid-1980s to the present.
Backpropagation is a generalization of the least squares
method where the weights are updated according to the
mean square error, which is determined with a comparison
between the desired and the computed output.31 The idea is
to find the minimum error function in relation to the connec-
tion weights. The error function is computed as
1X
E¼ ð^y − ym Þ2 ; (10)
2 j∈M m
EQ-TARGET;temp:intralink-;e010;63;305

with y^ m being the desired output of m’th ym .

In the present work, an MLP is used to achieve accurate
transmission maps in order to obtain high-quality free-haze
images. This method is described in the following section.

3 Proposed Method for Image Dehazing

3.1 Proposed Method

In general, the dehazing methods based on the DCP strategy
have achieved suitable results regarding different state-
of-the-art strategies. In order to compute the transmission
map using the DCP strategy, two stages have to be
performed.32
In the first stage, a coarse transmission is estimated using
square-patches of a predefined size. In the second stage,
the coarse transmission map is improved using different
strategies such as the soft matting,14 bilateral filter,33 median
filter,17 mean-shift,34 guided filter,15 Wiener filter,35 and Fig. 2 Flowchart of the proposed method.

Journal of Electronic Imaging 043022-3 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Salazar-Colores, Cruz-Aceves, and Ramos-Arreguin: Single image dehazing using a multilayer perceptron

graph-cuts.36 The main disadvantage of these strategies is the t 0 ðx; yÞ ¼ MLP½tmin ðx; yÞ
EQ-TARGET;temp:intralink-;e013;326;752 (13)
trade-off between accurate restoration and computational
time. The main idea behind the proposed method is the strategy
The proposed method is based on the transmission map to estimate the transmission map and the introduction of an
tmin defined as additional step for improving the contrast image. In order to
improve the contrast of the recovered image, the luminance
was modified by introducing the spacecolor L  a  b and a
tmin ðx; yÞ ¼ 1 − ωI min ðx; yÞ;
EQ-TARGET;temp:intralink-;e011;63;697 (11) contrast stretching38 strategy. In Fig. 2, the procedure carried
out by the proposed method is introduced.
where the minimum channel I min ðx; yÞ is defined as
follows:37 3.2 Synthetic Ground-Truth Generation
Since the MLP is a supervised technique, a ground-truth
is required to perform the training step. In the present
I c ðx; yÞ work, the ground-truth data are computed using the soft mat-
I min ðx; yÞ ¼ min : (12)
Ac ting algorithm proposed by Levin et al.,39 which has proven
EQ-TARGET;temp:intralink-;e012;63;644

c∈ðR;G;BÞ
to be suitable for the dehazing problem,14 nevertheless, its
computational processing time and memory requirements
are both high.
The minimum channel tmin is equivalent to the dark chan-
The image soft matting procedure can be expressed as
nel when the patch size is Ω ¼ 1. follows:
tmin ðx; yÞ has an excellent spatial resolution; however, its
depth resolution computed is low because it omits the infor- I ¼ Fα þ Bð1 − αÞ;
EQ-TARGET;temp:intralink-;e014;326;528 (14)
mation about of neighboring pixels. From the computed
transmission map tmin ðx; yÞ, the application of an ANN where B and F are the background and foreground color
MLP to estimate a transmission map t 0 ðx; yÞ is introduced information, respectively, and α is the opacity of foreground.
as follows: In the work of He et al.,14 it was proved the equivalence
between the α map and the transmission map t̄ðx; yÞ.
Then, the optimal t̄ðx; yÞ can be computed by solving the
sparse linear system expressed as

ðL þ λUÞt̄ ¼ λ^t;
EQ-TARGET;temp:intralink-;e015;326;431 (15)

where the used parameters were λ ¼ 10−4 , t̄ is the optimal

transmission, ^t is the transmission computed from the
dark channel, L is called matting Laplacian matrix, and U
is the identity matrix of the same size of L.

3.3 Training Process

In order to train the MLP, a data set of 80 real-world images
different size and features was used. From every image, two
images of the same size were generated. The first one con-
tains the estimated transmission map tmin ðx; yÞ using
Eq. (11), and the second one is the transmission result t̄ðx; yÞ,
which is obtained by applying the soft-matting method
to tmin ðx; yÞ.
The training step of an MLP requires two vectors of sam-
Fig. 3 Training process for the MLP based on transmission maps.
ples: the input data and the target vectors. In this work, each
sample was acquired from a square window of size l centered

Fig. 4 Process to estimate the transmission map using the MLP.

Journal of Electronic Imaging 043022-4 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Salazar-Colores, Cruz-Aceves, and Ramos-Arreguin: Single image dehazing using a multilayer perceptron

in positions ðx; yÞ where the length for each sample is l × l. 3.4 Application of MLP
The input vector was obtained from tmin ðx; yÞ and the target As it is shown in Fig. 4, to compute the transmission map
vector from t̄ðx; yÞ. The setup to perform the training stage is t 0 ðx; yÞ using the trained MLP, an input vector is generated
shown in Fig. 3. from square-windows of tmin ðx; yÞ. Each two-dimensional

Fig. 5 Examples of the processing the proposed method: (a) input images Iðx ; y Þ, (b) initial transmission
tmin ðx; y Þ, (c) final transmission t 0 ðx ; y Þ, and (d) recovered images Jðx ; yÞ.

Fig. 6 Relation between MLP architecture and performance using the training set of images.

Journal of Electronic Imaging 043022-5 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Salazar-Colores, Cruz-Aceves, and Ramos-Arreguin: Single image dehazing using a multilayer perceptron

Fig. 7 Relation between δ sliding windows used in the application of the MLP in the image, (a) SSIM
index and (b) time processing.

Fig. 8 Transmissions and results generated using different δ sliding windows: (a) δ ¼ 1, (b) δ ¼ 2,
(c) δ ¼ 4, (d) δ ¼ 8, (e) δ ¼ 12, and (f) δ ¼ 16.

Fig. 9 Comparison of the recovered results for real-world images using the ten methods: (a) input
images, (b) Tarel and Hautiere,12 (c) He et al.,14 (d) Pang et al.,15 (e) Gibson et al.,17 (f) Zhu et al.,18
(g) Berman et al.,19 (h) Ren et al.,26 (i) Cai et al.,27 (j) Li et al.,28 and (k) proposed.

Journal of Electronic Imaging 043022-6 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Salazar-Colores, Cruz-Aceves, and Ramos-Arreguin: Single image dehazing using a multilayer perceptron

Fig. 10 Comparison of the recovered results for synthetic images by the ten methods: (a) images
with haze simulated, (b) original images, (c) Tarel and Hautiere,12 (d) He et al.,14 (e) Pang et al.,15
(f) Gibson et al.,17 (g) Zhu et al.,18 (h) Berman et al.,19 (i) Ren et al.,26 (j) Cai et al.,27 (k) Li et al.,28
and (l) proposed.

Journal of Electronic Imaging 043022-7 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Salazar-Colores, Cruz-Aceves, and Ramos-Arreguin: Single image dehazing using a multilayer perceptron

square-window of size s is converted to a one-dimensional 1 X w X h

signal of size s × s. The interval of each sampling in the MSE ¼ ½Jðx; yÞ − I o ðx; yÞ2 ; (17)
ðw  hÞ x¼1 y¼1
EQ-TARGET;temp:intralink-;e017;326;752

image is expressed by the variable δ. If δ ¼ s, the sampling

has not intersection in the pixels positions, and if δ ¼ 1 the
sampling is realized in every pixel. The value of each pixel in where I o ðx; yÞ and Jðx; yÞ are the restored and target images,
the transmission map t 0 ðx; yÞ is the minimum value of the respectively, and (w) and (h) represent the width and height
superposition of windows. of both images.
In Fig. 5, an example of the proposed algorithm in four The SSIM is a normalized metric based on a perception-
real-world images and the results of applying the proposed based model, and it is defined as
method are presented, along with the input tmin ðx; yÞ and out-
put t 0 ðx; yÞ transmission maps of the MLP are shown. The Sðx; yÞ ¼ f½lðx; yÞ; cðx; yÞ; sðx; yÞ;
EQ-TARGET;temp:intralink-;e018;326;655 (18)
parameters values are δ ¼ 16 and s ¼ 8.
where lðx; yÞ is the luminance comparison, cðx; yÞ is the
contrast comparison, and sðx; yÞ is the structure comparison.
3.5 Evaluation Metrics
In order to measure the performance of the proposed method, 4 Results and Discussion
two metrics widely used in the literature25,26,28 have been
adopted; PSNR and structural similarity index.29 The evaluation performance of the proposed method was
The PSNR is defined as assessed using a database of 126 RGB images. The database
  was divided into two subsets: a training set of 80 images and
peak value2 a test set of 46 images. The test set of images consists of
PSNR ¼ 10  log10 ;
EQ-TARGET;temp:intralink-;e016;63;550 (16) 30 real-world images and 16 synthetic images acquired
MSE
from Ref. 40 and the Middlebury Stereo Dataset.41 The
where the peak value is the biggest possible value in the haze effects on the synthetic images were simulated using
pixels of the compared images and MSE represents the the scattering model presented in Eq. (1). The value of each
mean squared error, which is determined as follows: channel of A was assigned randomly between 0.7 and 1

Table 1 Comparative analysis using the PSNR.

Tarel and He Pang Gibson Zhu Berman Ren Cai Li

Image Hautiere12 et al.14 et al.15 et al.17 et al.18 et al.19 et al.26 et al.27 et al.28 Proposed

Aloe 17.91 14.20 14.22 15.46 18.42 16.05 17.99 18.14 16.60 16.08

Baby1 19.96 13.98 14.23 22.81 23.09 17.64 23.09 22.01 25.98 17.69

Books 16.03 12.71 13.37 15.43 22.36 13.31 20.07 21.03 19.07 23.03

Bowling1 19.34 11.60 11.71 12.88 23.29 14.06 20.62 27.20 25.69 13.00

Dolls1 13.24 14.93 14.98 15.22 15.94 15.70 17.06 14.37 15.20 15.23

Moebius1 18.97 15.04 15.28 19.86 24.44 19.18 20.64 24.28 21.22 23.49

Monopoly 17.41 10.48 10.76 13.39 21.52 16.43 18.93 19.99 19.30 19.26

Reindeer 14.77 18.32 18.21 19.39 16.20 18.60 16.82 14.11 18.11 16.82

Bikes1 8.29 18.49 18.31 15.47 12.59 26.08 14.36 14.38 14.22 21.77

Church 11.76 12.51 12.69 14.00 13.77 15.96 15.01 15.24 11.03 17.04

Cones 19.38 20.16 19.45 19.03 19.98 19.77 17.60 19.46 21.43 17.50

Couch 16.12 18.44 18.60 17.41 18.89 18.47 19.24 17.89 17.02 19.00

Flower2 16.88 17.59 17.99 18.62 21.01 11.92 13.27 20.15 14.55 22.24

Mansion 16.45 19.80 19.75 16.98 18.09 17.42 17.86 17.93 15.44 22.35

Roofs2 11.58 22.86 22.73 17.51 15.42 21.00 13.95 13.70 15.98 20.77

Trees2 10.69 17.84 17.70 16.60 13.38 20.79 14.29 12.24 15.15 15.12

Average 15.55 16.19 16.25 16.88 18.65 17.65 17.55 18.26 17.87 18.77

Journal of Electronic Imaging 043022-8 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Salazar-Colores, Cruz-Aceves, and Ramos-Arreguin: Single image dehazing using a multilayer perceptron

values, and β has random values between 0.5 and 1.5 as it 4.2 Performance Analysis
was proposed by Ren et al.26 Furthermore, the computational The proposed method was compared with nine state-of-art
experiments were performed using the Matlab software, dehazing methods: Tarel and Hautiere,12 He et al. method,14
version R2016a on a computer with 3.10 GHz Intel Core Pang et al. method,15 Gibson et al.,17 Zhu et al.,18 Berman
i5-2400 and 12 GB RAM memory. et al., Ren et al.,26 Cai et al.,27 and Li et al.28 The parameters
used in each method are the proposed by every author.

4.1 Parameters Tuning 10 3

Average execution time in seconds (log )

10
To obtain the best possible performance of the proposed
method, several tests with different MLP configurations
10 2
using the training set of 80 images were realized. In Fig. 6,
the results of the most significant configurations in terms of
the MSE metric are shown. The best MLP performance is
presented by the configuration 256 to 1024, which corre- 10 1

sponds to a window size s ¼ 16.

Once the value s was determined, in order to obtain the
optimal value of δ (window sliding value), the method was 10 0 He et al. Zhu et al.

tested over the training set varying the δ value in the range
Tarel et al. Gibson et al.
Pang et al. Cai et al.

½1; ::; 16. Three aspects were considered to choose the δ value:
Ren et al. Li et al.
Berman et al. Proposed

a quantitative analysis using the index SSIM [Fig. 7(a)], a 10 -1

0.2 0.3 0.5 1 1.5 2.07
time processing analysis [Fig. 7(b)], and a qualitative analysis 106
Number of pixels
(Fig. 8). Based on the obtained results, the best value was
δ ¼ 8, representing the highest SSIM and lowest processing Fig. 11 Performance analysis in terms of computational time
time. processing.

Table 2 Comparative analysis using the SSIM index.

Tarel and He Pang Gibson Zhu Berman Ren Cai Li

Image Hautiere12 et al.14 et al.15 et al.17 et al.18 et al.19 et al.26 et al.27 et al.28 Proposed

Aloe 0.9028 0.7855 0.7976 0.8381 0.9549 0.8250 0.9473 0.9458 0.9220 0.8180

Baby1 0.9032 0.7763 0.7961 0.9403 0.9706 0.8563 0.9596 0.9655 0.9729 0.8317

Books 0.8403 0.7378 0.7715 0.8237 0.9561 0.7705 0.9254 0.9439 0.9357 0.9219

Bowling1 0.9255 0.6760 0.6902 0.7305 0.9764 0.7786 0.9496 0.9844 0.9814 0.7698

Dolls1 0.7354 0.7912 0.7737 0.8317 0.8119 0.8343 0.8413 0.7958 0.8136 0.8268

Moebius1 0.8665 0.7933 0.7940 0.9080 0.9593 0.8731 0.9322 0.9629 0.9259 0.9159

Monopoly 0.8929 0.7197 0.7308 0.8486 0.9763 0.9137 0.9562 0.9647 0.9605 0.9407

Reindeer 0.6842 0.7966 0.7805 0.8033 0.7409 0.7779 0.7632 0.7102 0.7814 0.7753

Bikes1 0.4939 0.7785 0.7560 0.6436 0.5143 0.8479 0.5963 0.5469 0.5009 0.7953

Church 0.7627 0.7692 0.7682 0.8532 0.8218 0.8832 0.8544 0.8198 0.6108 0.8889

Cones 0.8530 0.9090 0.8904 0.9186 0.9274 0.9162 0.8186 0.9247 0.8995 0.8776

Couch 0.7950 0.8450 0.8356 0.8183 0.8293 0.8620 0.8658 0.8217 0.8169 0.8245

Flower2 0.8023 0.8431 0.8591 0.8640 0.9630 0.6652 0.5919 0.9105 0.7814 0.9086

Mansion 0.8434 0.8958 0.8917 0.8424 0.8769 0.8697 0.8730 0.8362 0.6907 0.9140

Roofs2 0.6552 0.8445 0.8424 0.7806 0.6619 0.8602 0.6296 0.6156 0.6601 0.8335

Trees2 0.5462 0.6833 0.6710 0.6758 0.5800 0.7387 0.6122 0.5492 0.5877 0.6833

Average 0.7814 0.7903 0.7906 0.8201 0.8451 0.8295 0.8198 0.8311 0.8026 0.8454

Journal of Electronic Imaging 043022-9 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Salazar-Colores, Cruz-Aceves, and Ramos-Arreguin: Single image dehazing using a multilayer perceptron

For the proposed method (the source code and dehazing 6. Y. Y. Schechner, S. G. Narasimhan, and S. K. Nayar, “Instant dehazing
of images using polarization,” in Proc. of IEEE Computer Society Conf.
results of the proposed method42), the parameters are: δ ¼ 8, on Computer Vision and Pattern Recognition (CVPR), Kauai, Vol. 16,
window size s ¼ 16, and w ¼ 0.7. pp. 325–332 (2001).
7. S. Fang et al., “Image dehazing using polarization effects of objects and
To perform a visual analysis, in Fig. 9, the dehazing airlight,” Opt. Express 22, 19523–19537 (2014).
results of the proposed and comparative methods over a sub- 8. T. Fang, Z. Cao, and R. Yan, “A unified dehazing approach for infrared
set of real-world images of the test set are presented. In this images,” Proc. SPIE 8917, 89170V (2013).
9. J. Kopf et al., “Deep photo: model-based photograph enhancement and
subset, a higher contrast along with more natural intensities viewing,” ACM Trans. Graphics 27(5), 116 (2008).
is obtained by the proposed method. In addition, Fig. 10 10. C. Chengtao, Z. Qiuyu, and L. Yanhua, “A survey of image dehazing
shows the results using synthetic images. approaches,” in Control and Decision Conf. (CCDC), Qingdao, China,
Vol. 1, pp. 3964–3969 (2015).
To evaluate the restoration level, Tables 1 and 2 present 11. V. Sahu and M. Singh, “A survey paper on single image dehazing,” Int.
the PSNR and SSIM index values, respectively. The pro- J. Recent Innov. Trends Comput. Commun. 3(2), 85–88 (2007).
posed method achieves the highest average value of PSNR, 12. J.-P. Tarel and N. Hautiere, “Fast visibility restoration from a single
color or gray level image,” in IEEE Int. Conf. on Computer Vision
and the highest average in terms of the SSIM index. Finally, (ICCV), Kyoto, Japan, Vol. 1, pp. 2201–2208 (2009).
Fig. 11 shows the time processing comparison, where the 13. R. Fattal, “Single image dehazing,” ACM Trans. Graphics 27, 72 (2008).
14. K. He, J. Sun, and X. Tang, “Single image haze removal using dark chan-
proposed method obtains the best performance. nel prior,” IEEE Trans. Pattern Anal. Mach. Intell. 33(12), 2341–2353
(2010).
5 Conclusion 15. J. Pang, A. Oscar, and G. Zheng, “Improved single image dehazing
using guided filter,” in Proc. of the APSIPA Annual Summit and
In this paper, a single image dehazing method was proposed. Conf. (APSIPA ASC), Xi’an, China, Vol. 1, pp. 1–4 (2011).
This method uses an artificial neuronal network MLP to 16. K. Nishino, L. Kratz, and S. Lombardi, “Bayesian defogging,” Int. J.
Comput. Vision 98(3), 263–278 (2012).
estimate the transmission map of a haze image. To obtain 17. K. B. Gibson, D. T. Võ, and T. Q. Nguyen, “An investigation of dehaz-
the optimal MLP configuration, a training set of 80 real- ing effects on image and video coding,” IEEE Trans. Image Process.
world images was used. In the training experiments, a num- 21(2), 662–673 (2012).
18. Q. Zhu, J. Mai, and L. Shao, “A fast single image haze removal algo-
ber of hidden layers containing different number of neurons rithm using color attenuation prior,” IEEE Trans. Image Process. 24,
were tested, where the best performance in terms of the MSE 3522–3533 (2015).
was achieved using a 256- to 1024- MLP configuration with 19. D. Berman, T. Treibitz, and S. Avidan, “Non-local image dehazing,” in
IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)
an MSE = 0.000151. In order to evaluate the restoration qual- (2016).
ity of the proposed method, the SSIM index and the PSNR 20. M. Gardner and S. Dorling, “Artificial neural networks (the multilayer
were used. The experimental results have proven that the perceptron): a review of applications in the atmospheric sciences,”
Atmos. Environ. 32(14–15), 2627–2636 (1998).
proposed method achieves a superior performance than 21. D. Kriesel, A brief introduction to neural networks, Retrieved August
seven state-of-art methods in terms of restoration quality, Vol. 1(1), p. 244, http://www.dkriesel.com (2005).
22. D. Bhattacharjee et al., “Human face recognition using fuzzy multilayer
obtaining a PSNR value of 18.77 and SSIM index of 0.8454 perceptron,” Soft Comput. 14(6), 559–570 (2010).
over a test set of synthetic images. In addition, a comparative 23. S. L. Phung, A. Bouzerdoum, and D. Chai, “Skin segmentation using
analysis using the test set of 40 images between the proposed color pixel classification: Analysis and comparison,” IEEE Trans.
Pattern Anal. Mach. Intell. 27(1), 148–154 (2005).
and comparative methods reveals that the lowest computa- 24. H. C. Burger, C. J. Schuler, and S. Harmeling, “Image denoising:
tional time was obtained by the proposed method (0.52 s). can plain neural networks compete with BM3D?” in Proc. of the
Given the suitable results in terms of restoration quality IEEE Computer Society Conf. on Computer Vision and Pattern
Recognition, pp. 2392–2399 (2012).
and execution time of the proposed method with respect 25. X. Fan et al., “Two-layer Gaussian process regression with example
to the seven state-of-art dehazing methods, it can be highly selection for image dehazing,” IEEE Trans. Circuits Syst. Video
appropriate to be used in real-time systems. Technol. 27(12), 2505–2517 (2017).
26. W. Ren et al., “Single image dehazing via multi-scale convolutional
neural networks,” in Proc. of 14th European Conf. on Computer
Disclosures Vision–ECCV, Amsterdam, The Netherlands, Part II, B. Leibe et al.,
Eds., pp. 154–169, Springer International Publishing, Cham (2016).
The authors declare there is no conflict of interest. 27. B. Cai et al., “DehazeNet: an end-to-end system for single image haze
removal,” IEEE Trans. Image Process. 25(11), 5187–5198 (2016).
Acknowledgments 28. B. Li et al., “AOD-Net: all-in-one dehazing network,” in Proc. of the
IEEE Int. Conf. on Computer Vision, pp. 4780–4788 (2017).
Sebastián Salazar-Colores (CVU 477758) would like to 29. R. Dosselmann and X. Yang, “A comprehensive assessment of the struc-
thank CONACYT (Consejo Nacional de Ciencia y tural similarity index,” Signal, Image and Video Process. 5(1), 81–91
(2011).
Tecnología) for the financial support of his PhD studies 30. E. J. McCartney and F. F. Hall, “Optics of the atmosphere: scattering by
under Scholarship 285651. In addition, the authors thank molecules and particles,” Phys. Today 30(5), 76–77 (1977).
CONACYT for the project Cátedras-CONACYT No. 31. L. Fausett, “Fundamentals of neural networks,” in Int. Geoscience and
Remote Sensing Symp. (IGARSS), Vol. 1(1), pp. 1–5 (2014).
3150-3097. 32. K. Tang, J. Yang, and J. Wang, “Investigating haze-relevant features in a
learning framework for image dehazing,” in IEEE Conf. on Computer
References Vision and Pattern Recognition (CVPR), Columbus, Ohio, Vol. 1,
pp. 2995–3002 (2014).
1. Y. Cheng, W. Niu, and Z. Zhai, “Video dehazing for surveillance 33. C.-H. Yeh et al., “Haze effect removal from image via haze density esti-
unmanned aerial vehicle,” in Proc. of AIAA/IEEE Digital Avionics mation in optical model,” Opt. Express 21, 27127–27141 (2013).
Systems Conf. (2016). 34. Q. Zhu et al., “Mean shift-based single image dehazing with re-refined
2. A. L. Baruwa, A. N. Evans, and R. Wyatt, “The development of transmission map,” in IEEE Int. Conf. on Systems, Man and Cybernetics
a remote sensing system with real-time automated horizon tracking (SMC), San Diego, California Vol. 1, pp. 4058–4064 (2014).
for distance estimation at sea,” Proc. SPIE 8892,889217 (2013). 35. K. B. Gibson and T. Q. Nguyen, “Fast single image fog removal using
3. J. Long et al., “Single remote sensing image dehazing,” IEEE Geosci. the adaptive Wiener filter,” in IEEE Int. Conf. on Image Processing,
Remote Sens. Lett. 11(1), 59–63 (2014). Melbourne, Canada, Vol. 1, pp. 714–718, IEEE (2013).
4. S. Kim, S. Park, and K. Choi, “A system architecture for real time traffic 36. M. Zhu and B. He, “Dehazing via graph cut,” Opt. Eng. 56, 113105 (2017).
monitoring in foggy video,” in Korea-Japan Joint Workshop on 37. S. Jeong and S. Lee, “The single image dehazing based on efficient trans-
Frontiers of Computer Vision (FCV), Mokpo, South Korea, Vol. 1, mission estimation,” in IEEE Int. Conf. on Consumer Electronics,
pp. 1–4, (2015). pp. 376–377 (2013).
5. X. Yan et al., “Effects of foggy conditions on drivers’ speed control 38. C. C. Yang, “Image enhancement by modified contrast-stretching
behaviors at different risk levels,” Saf. Sci. 68, 275–287 (2014). manipulation,” Opt. Laser Technol. 38(3), 196–201 (2006).

Journal of Electronic Imaging 043022-10 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
 Salazar-Colores, Cruz-Aceves, and Ramos-Arreguin: Single image dehazing using a multilayer perceptron

39. A. Levin, D. Lischinski, and Y. Weiss, “A closed-form solution to Ivan Cruz-Aceves received his PhD in electrical engineering from the
natural image matting,” IEEE Trans. Pattern Anal. Mach. Intell. 30(2), Universidad de Guanajuato in 2014. He has been working at the
228–242 (2008). Mexican National Council on Science and Technology assigned to
40. R. Fattal, “Dehazing using color-lines,” ACM Trans. Graphics 34(1), the Center for Research in Mathematics in Guanajuato, Mexico,
1–14 (2014).
41. H. Hirschmuller and D. Scharstein, “Evaluation of cost functions for stereo since 2014. His research areas are biomedical signal and image
matching,” in Conf. on Computer Vision and Pattern Recognition, pp. 1–8 analysis, computational intelligence, and evolutionary computation.
(2007).
42. I. Cruz-Aceves, “Single image dehazing using a multilayer perceptron,” Juan-Manuel Ramos-Arreguin received his MS degree in electrical
2018 http://personal.cimat.mx:8181/~ivan.cruz/dehazing.html. engineering option instrumentation and digital systems from the
University of Guanajuato and his PhD in mechatronics science
Sebastián Salazar-Colores received his BS degree in computer sci- from the Centro de Ingeniera y Desarrollo Industrial (Engineering
ence from the Universidad Autónoma Benito Juárez de Oaxaca and and Industrial Development Center). Since 2009 he has been part
his MS degree in electrical engineering at Universidad de Guanajuato. of the Engineering Department at the UAQ where he works as
HE is a PhD candidate in computer science at the Universidad a researcher and lecturer. His research interest includes mecha-
Autónoma de Querétaro. His research interests are image processing tronics and embedded systems.
and computer vision.

Journal of Electronic Imaging 043022-11 Jul∕Aug 2018 • Vol. 27(4)

Downloaded From: https://www.spiedigitallibrary.org/journals/Journal-of-Electronic-Imaging on 14 Jun 2021

Terms of Use: https://www.spiedigitallibrary.org/terms-of-use