Two Simple Unfolded Residual Networks for Single Image Dehazing
Bartomeu Garau
1 a
, Joan Duran
1,2 b
and Catalina Sbert
1,2 c
1
Institute of Applied Computing and Community Code, Universitat de les Illes Balears (UIB), Edifici Complexe d’R+D,
Cra. de Valldemossa km 7.4, E-07122 Palma, Spain
2
Dept. of Mathematics and Computer Science, UIB, Cra. de Valldemossa km 7.5, E-07122 Palma, Spain
Keywords:
Image Dehazing, Deep Learning, Unfolding, Residual Network, Channel Attention, Variational Methods.
Abstract:
Haze is an environmental factor that impairs visibility for outdoor imaging systems, presenting challenges for
computer vision tasks. In this paper, we propose two novel approaches that combine the classical dark channel
prior with variational formulations to construct an energy functional for single-image dehazing. The proposed
functional is minimized using a proximal gradient descent scheme, which is unfolded into two different net-
works: one built with residual blocks and the other with residual channel attention blocks. Both methods
provide straightforward yet effective solutions for dehazing, achieving competitive results with simple and
interpretable architectures.
1 INTRODUCTION
The rapidly increasing use and demand for efficient
outdoor imaging systems have brought issues like de-
hazing to the forefront of image processing. Outdoor
images are often affected by atmospheric conditions
such as haze, smoke, rain or snow. In particular, haze
reduces visibility, giving scenes a gray tone and low-
ering contrast. Tackling these issues is highly relevant
for a wide range of applications, including surveil-
lance, autonomous systems, and remote sensing.
Haze is an environmental phenomenon, caused by
the scattering of light as it travels through the at-
mosphere, where airborne particles distort the light.
Moreover, the degradation depends on both the depth
of the scene and the haze density. This makes dehaz-
ing a particularly challenging problem.
Various strategies for image dehazing have been
explored in the literature (Wang and Yuan, 2017; Guo
et al., 2022; Jackson et al., 2024). Some methods
address the problem as an enhancement task, using
techniques such as histogram equalization (Jun and
Rong, 2013; Thanh et al., 2019) or the Retinex theory
(Zhou and Zhou, 2013; Galdran et al., 2018). Other
methods leverage the physical principles underlying
hazy scenes (McCartney, 1977). The resulting mod-
els can be approached in different ways, including di-
a
https://orcid.org/0009-0008-3439-8316
b
https://orcid.org/0000-0003-0043-1663
c
https://orcid.org/0000-0003-1219-4474
rect computation (Tan, 2008; He et al., 2011) or vari-
ational techniques (Fang et al., 2014; Galdran et al.,
2015; Liu et al., 2022).
With the rapid growth of artificial intelligence,
numerous dehazing methods involving deep learning
networks have emerged (Cai et al., 2016; Qin et al.,
2019; Lei et al., 2024). Some of these methods in-
clude unfolding architectures (Yang and Sun, 2018;
Fang et al., 2024), which combine the strengths of
model-based and data-driven learning approaches.
In this paper, we propose two simple model-based
deep unfolded approaches to variational image dehaz-
ing. Our proposals are based on the dark channel prior
(He et al., 2011) to estimate the main components of a
hazy image, specifically the transmission map and the
atmospheric light of the scene. We introduce a simple
variational formulation to obtain the haze-free image
as the minimizer of an energy functional. The mini-
mization of this energy is performed using a proximal
gradient descent algorithm, in which the proximal op-
erators are replaced by residual networks.
The rest of the paper is organized as follows. In
Section 2, we review the related work on image de-
hazing. Section 3 introduces the two proposed models
and, in Section 4, we discuss their implementations
and compare them with state-of-the-art approaches.
Section 5 conducts an ablation study to justify the
configurations of our architectures. Finally, conclu-
sions are drawn in Section 6.
516
Garau, B., Duran, J. and Sbert, C.
Two Simple Unfolded Residual Networks for Single Image Dehazing.
DOI: 10.5220/0013181400003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 3: VISAPP, pages
516-523
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

2 RELATED WORK
Physical principles can be used to describe the atmo-
spheric scattering that generates haze. In this context,
McCartney (McCartney, 1977) posits that a hazy im-
age is formed through the combined effects of light
attenuation and air-light scattering. This leads to the
following expression:
I(x) = J(x)t(x) + A(1 −t(x)), (1)
where I is the hazy image, J is the haze-free image,
t is the transmission map (the proportion of the clear
image that reaches the camera), and A is the atmo-
spheric light of the scene. Usually, the transmission is
related to the depth map of the scene. Estimating J,
A, and t from I is a highly ill-posed inverse problem.
To solve the decomposition problem arising from
(1), one either requires additional information or must
rely on some prior assumptions. Tan et al. (Tan, 2008)
assume that hazy images exhibit lower contrast and
that the variation of the air light is a smooth func-
tion of the distance. Fattal et al. (Fattal, 2008) sepa-
rate surface shading from the transmission map. He et
al. (He et al., 2011) introduce the dark channel prior,
which states that in most local patches of haze-free
outdoor images there are pixels with very low inten-
sities in at least one color channel.
In the variational framework, the dehazed image is
obtained as the minimizer of an energy functional that
incorporates both data-fidelity terms, which measure
the deviation from prescribed constraints involving
the hazy image, and regularization terms, which as-
sess the smoothness of the solution. Fang et al. (Fang
et al., 2014) pioneered a variational formulation for
image dehazing, where the energy functional en-
forces total variation to regularize the depth map and
weighted total variation for the dehazed image. The
authors rely on the dark channel prior to estimate an
initial transmission. Since then, several variants have
been proposed. For example, Lei et al. (Jin et al.,
2024) apply total generalized variation to the depth,
while Liu et al. (Liu et al., 2018) introduce nonlo-
cal regularization to refine the transmission map, sup-
press unwanted artifacts, and preserve image details.
In (Stipeti
´
c and Lon
ˇ
cari
´
c, 2022), the authors propose
a smooth variational formulation of the dark channel
prior that reaches a minimum when the reconstructed
image satisfies the prior.
Other variational approaches avoid (1) and exploit
alternative formation models. In this context, Galdran
et al. (Galdran et al., 2015) propose an energy that
maximizes the average contrast of the image, which
is further studied in (Galdran et al., 2017) and applied
for image fusion. On the other hand, Liu et al. (Liu
et al., 2022) decompose the hazy image as a linear
combination of structure, detail, noise and glow, and
use different regularization terms for each of these
components.
Recently, the growing popularity of deep learn-
ing architectures has lead to an increase in dehazing
methods. This trend began with (Cai et al., 2016),
which improved the estimation of the transmission
map using an end-to-end convolutional neural net-
work (CNN). In this framework, some architectures
lack physical basis and rely on artificially generated
pairs of hazy and ground-truth images (Qu et al.,
2019; Qin et al., 2019). However, there are also de-
hazing networks based on the models and priors dis-
cussed previously, such as histogram correction (Chi
et al., 2020), the Retinex theory (Li et al., 2021; Lei
et al., 2024), and the dark channel prior (Zhang and
Patel, 2018; Golts et al., 2020).
The use of formation models makes variational
methods robust to distortions, but their performance
is limited by rigid priors. Conversely, data-driven
learning approaches can easily learn natural priors,
but are less flexible and interpretable. Deep unfold-
ing networks combine the strengths of both. The gen-
eral idea involves unfolding the steps of the optimiza-
tion algorithm into a deep learning framework. These
networks can be based on transformers (Song et al.,
2023), pyramid structures (Xiao et al., 2024) or clas-
sical optimization algorithms (Yang and Sun, 2018;
Fang et al., 2024). In (Yang and Sun, 2018), the au-
thors introduce an energy functional with a novel dark
channel regularization term and subsequently unfold
a proximal point algorithm into deep CNN structures.
More recent architectures like (Fang et al., 2024) use
(1) without assuming the dark channel prior. How-
ever, the resulting algorithm is unfolded into a coop-
erative network that increases in complexity.
3 PROPOSED MODELS
Based on the haze formation model (1), we want to re-
cover J, A, and t from a single hazy image I. To ad-
dress the ill-posed nature of such a problem, we will
use the dark channel prior (He et al., 2011) to estimate
t and A, that is,
J
dark
= min
c∈{R,G,B}

min
y∈w(x)
J
c
(y)

→ 0,
where w(x) is a patch of pixels centered at x. Then,
we will estimate a rough transmission map as
˜
t
0
(x) = 1 − ν min
c∈{R,G,B}

min
y∈w(x)

I
c
(y)
A
c

, (2)
where ν is a constant set empirically to ν = 0.95. To
compute (2), we will first estimate A by taking the
Two Simple Unfolded Residual Networks for Single Image Dehazing
517

···
Basic Block
(a)
·
·
·
(b)
Conv 3x3
Conv 3x3
ReLU
Basic Block
Basic Block
Basic Block
(c)
Conv 3x3
ReLU ReLU
Conv 3x3
(d)
Figure 1: (a) Overall architecture of the unfolded formulation of (7). (b) Preprocessing block. (c) Residual Network (ResNet)
architecture. (d) Basic block residual architecture.
···
Basic
Block
···
(a)
Conv 3x3
Conv 3x3
CA Layer
Conv 3x3
ReLU
(b)
Avg Pool
Conv 3x3
ReLU
Conv 3x3
Sigmoid
(c)
Figure 2: (a) Overall architecture of the unfolded formulation of (8). The preprocessing block, the ResNet block, the basic
block are the same ones featured in Figures 1b, 1c, 1d, respectively. (b) Residual channel attention architecture for g. (c)
Channel attention (CA) layer.
mean of the top 0.1% brightest pixels of J
dark
on each
channel, as done in (He et al., 2011). Once we have
˜
t
0
, we will apply a guided filter (He et al., 2013) to
obtain the initial transmission map t
0
.
Following (Fang et al., 2014), we can rewrite each
channel of (1) as
A
c
− I
c
= t(A
c
− J
c
).
By linearizing the model, we get
log(A
c
− I
c
) = logt + log(A
c
− J
c
).
When the atmosphere is homogeneous, the transmis-
sion can be approximated by
t(x) = e
−ηd(x)
, (3)
where η > 0 describes the scattering of the medium
and d is the depth map of the scene. Using (3), and
setting f
c
=
1
η
log(A
c
− I
c
) and g
c
=
1
η
log(A
c
− J
c
),
we end up with g
c
= f
c
+ d or, in vectorial form,
g = f + d, (4)
where we denote g = {g
R
, g
G
, g
B
}, f = { f
R
, f
G
, f
B
},
and d = {d, d, d}.
3.1 Variational Formulation
We will estimate d and J from (4) as the minimizers
of an energy functional of the form
E(g, d) := R(g, d) + F(g, d),
where R and F consist of the regularization and fi-
delity terms, respectively. On the one hand, we
choose different regularizers for g and d:
R(g, d) := R
1
(g) + λR
2
(d),
where λ > 0 is a trade-off parameter. On the other
hand, we consider the following fidelity terms:
F(g, d) :=
α
2
∥g − f − d∥
2
2
+
γ
2
∥d − d
0
∥
2
2
,
where d
0
= −logt
0
and α, γ > 0. Therefore, we aim
to solve the following minimization problem:
min
g,d
{
R
1
(g) + λR
2
(d) + F(g, d)
}
. (5)
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
518

Since F is differentiable, if we assume R
1
and R
2
to be proper, convex and lower semicontiuous func-
tionals, we can solve (5) using the proximal gradi-
ent descent algorithm (Chambolle and Pock, 2016).
Therefore, the sequence of iterates {(g
k
, d
k
)} con-
verging to the solution of the minimization problem
(5) is given by
(
g
k+1
= prox
τR
1
(G
k
),
d
k+1
= prox
σλR
2
(D
k
),
(6)
where τ, σ > 0 are the step-size parameters, and
G
k
:= g
k
− τ∇
g
F(g
k
, d
k
) = (1 − τα)g
k
+ τα(f + d
k
),
D
k
:= d
k
− σλ∇
d
F(g
k
, d
k
)
= (1 − 3σα − σγ)d
k
+ ασ
∑
c
(g
k
c
− f
c
) + σγd
0
.
3.2 Unfolded Formulation
If we consider R
1
and R
2
to be two generic regulariz-
ers that are proper, convex and lower semicontiuous,
we can unfold (6) and replace the proximal operators
by learning-based networks. Therefore, (6) becomes
(
g
k+1
= ResNet
k
(G
k
),
d
k+1
= ResNet
k
(D
k
).
(7)
The hyperparameters λ, α, γ, τ and σ are learned
throughout the training phase and shared across all
stages (that is, the number of iterations of the opti-
mization algorithm). However, the residual networks
do not share weights between stages.
The overall structure of the network is illustrated
in Figure 1a. Figure 1b displays the initialization
stage, where
˜
t
0
is filtered with the Guided Image Fil-
tering Block (GIF Block) introduced in (Yang and
Sun, 2018). This block is fixed and not learned during
the training phase. In this way, we obtain the trans-
mission map t
0
, which is used to set the initialization
variables d
0
and g
0
using the relations (3)-(4). Fig-
ure 1c shows each stage of the residual network used
to compute (7), built with the basic blocks depicted in
Figure 1d. From now on, this network will be referred
to as URNet (Unfolded Residual Network).
We will now propose an alternative architecture.
The nonlocal theory for image processing is used to
capture self-similarities across different patches of an
image to smooth them out. Since g contains the
haze-free image, we want to regularize patches with
the same amount of haze in a similar way. Now,
as channel attention modules mimic the behaviour of
nonlocal regularization terms (Pereira-S
´
anchez et al.,
2024), we propose to substitute the ResNet used to
compute g
k
by a residual channel attention network
(RCANet). Thus, we can now unfold (6) as
(
g
k+1
= RCANet
k
(G
k
),
d
k+1
= ResNet
k
(D
k
).
(8)
Again, the residual networks do not share weights be-
tween stages and the hyperparameters are randomly
generated and learned during the training phase. The
structure of the new network can be seen in Figure
2a. The preprocessing block, the basic blocks are the
same as in Figures 1b and 1d, respectively. In Figure
2b, we can see the residual channel attention network
(RCANet) used to compute g
k
. The residual network
used to compute d
k
has the same structure as the one
presented in Figure 1c. Figure 2c shows the last layer
of the RCANet, the channel attention layer. From now
on, this network will be referred to as URCANet (Un-
folded Residual Channel Attention Network).
4 EXPERIMENTAL RESULTS
For the performance evaluation, we will use the
RESIDE-Standard dataset (Li et al., 2019). We have
selected the SOTS-outdoor set, which comprises 500
pairs of outdoor hazy images and their correspond-
ing ground truths. These pairs have been divided into
70% for training, 15% for validation, and 15% for
testing.
We compare our dehazing models with various
landmark and state-of-the-art methods. Specifically,
we compare with He’s dark channel prior (DCP) (He
et al., 2011), since we use their estimations for t
0
and A; Fang et al.’s variational model (Fang et al.,
2014), since our variational framework is based on
it; two physical-model-based networks, DehazeNet
(Cai et al., 2016) and AODNet (Li et al., 2017); and
two straightforward hazy-to-clear networks, FFA-Net
(Qin et al., 2019) and ConvIR (Cui et al., 2024).
AODNet, FFA-Net and ConvIR have been down-
loaded from their respective GitHub repositories,
while the other models have been implemented in Py-
torch from scratch. All methods, including ours, have
been trained during 1000 epochs using an ADAM op-
timizer with a learning rate of 10
−5
. For more details
about the configurations of the two proposed unfolded
networks, we refer to the ablation study in Section 5.
Since ground truths are available, the metrics
used for objective evaluation are Peak Signal-to-Noise
Ratio (PSNR), Structural Similarity Index Measure
(SSIM), and Spectral Angle Mapper (SAM).
Table 1 displays the average PSNR, SSIM, and
SAM values of each method on the testing set. The
proposed models yield the best results in terms of
Two Simple Unfolded Residual Networks for Single Image Dehazing
519

(a) Ground truth
PSNR / SSIM / SAM
(b) Hazy image
16.71 / 0.93 / 0.03
(c) DCP
11.92 / 0.75 / 0.08
(d) Fang et al.
11.95 / 0.82 / 0.04
(e) DehazeNet
19.64 / 0.89 / 0.05
(f) AODNet
24.98 / 0.95 / 0.02
(g) FFA-Net
11.81 / 0.75 / 0.08
(h) ConvIR
24.85 / 0.96 / 0.04
(i) URNet (Ours)
23.02 / 0.95 / 0.02
(j) URCANet (Ours)
24.04 / 0.96 / 0.02
Figure 3: Visual comparison of dehazing methods on an image of the testing set. DCP and Fang et al.’s effectively remove the
haze, but tend to darken the image and produce artifacts around the edges. DehazeNet and ConvIR yield a clear output but
fail to correctly balance the colors of the scene. AODNEt and our models provide the best visual results, with ours offering
superior color recovery.
Table 1: Quantitative comparison of various dehazing meth-
ods on the testing set of SOTS-outdoor dataset. We high-
light in blue the best result, and red the second best.
Method PSNR ↑ SSIM ↑ SAM ↓
DCP 11.863 0.785 0.075
Fang et al. 15.746 0.824 0.076
DehazeNet 16.382 0.864 0.067
AODNet 21.955 0.874 0.059
FFA-Net 11.221 0.689 0.087
ConvIR 22.370 0.903 0.113
URNet (Ours) 22.045 0.906 0.058
URCANet (Ours) 21.921 0.912 0.058
SSIM and SAM, while our URNet ranks second best
in terms of PSNR, just behind ConvIR. However, as
illustrated in Figures 3 and 4, the dehazed images pro-
vided by ConvIR are undersaturated. We also observe
that DCP and Fang et al.’s methods tend to darken
the images and introduce artifacts around the edges.
Among the deep learning models, all except FFA-Net
effectively remove the haze. However, our two meth-
ods tend to improve color recovery. DehazeNet of-
ten oversaturates the images, returning warmer col-
ors. Conversely, FFA-Net struggles significantly, in-
troducing additional haze. This issue is likely due to
being trained on a small dataset and suffering from
overfitting. This highlights the importance of robust
training protocols in achieving effective methods.
Finally, we also test the quality of the estimations
on real life images from the LIVE Image Defogging
Database (Choi et al., 2015), as shown in Figure 5.
We can see that the results inherit the qualities and
problems of the synthetic image testing. He’s, DCP
and Dehazenet remove efectively the haze, but over-
saturate the sky regions. Fang et al. removes the haze
but generates a strong halo around the edges. Both
AODNet and ConvIR remove the haze, but inherit the
color of the hazy scene. Our networks combine the
strengths of the DCP and neural networks, resulting
in a fully dehazed image with a better color balance,
even though it also saturates the sky because of the
violation of the DCP.
5 ABLATION STUDY
To optimize the configuration of our model and val-
idate each component’s contribution to performance,
we conduct several ablation studies. These focus on
the network structure, hyperparameter impact, the ne-
cessity of preprocessing and postprocessing blocks,
and the choice of loss function.
Concerning the structure of the URNet, we have
chosen to use a ResNet because the proximal operator
of a proper, lower semicontinuous and convex func-
tion R can also be defined as a resolvent operator:
prox
τR
(·) = (Id + τ∂R)
−1
(·). (9)
Then, we have trained the model varying the num-
ber of blocks of the ResNet architecture, stages and
features chosen. After this study, we have settled for
3 stages, 64 features and 3 basic blocks. With this
setting, our model has 1.3M parameters. With the
same configuration and stages, we have then com-
puted g with residual channel attention blocks instead
of residual blocks and trained the URCANet. With
this setting, the model has 2M parameters.
After the main structure of the network has been
chosen, we study possible pre and postprocessing
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
520

(a) Ground truth
PSNR / SSIM / SAM
(b) Hazy image
14.62 / 0.87 / 0.05
(c) DCP
18.07 / 0.86 / 0.17
(d) Fang et al.
17.47 / 0.87 / 0.05
(e) DehazeNet
21.32 / 0.93 / 0.11
(f) AODNet
24.26 / 0.93 / 0.03
(g) FFA-Net
15.41 / 0.89 / 0.03
(h) ConvIR
24.13 / 0.93 / 0.07
(i) URNet (Ours)
23.35 / 0.93 / 0.03
(j) URCANet (Ours)
23.79 / 0.94 / 0.06
Figure 4: Visual comparison of dehazing methods on an image of the testing set. DehazeNet and AODNet produce a haze-
free images with a warmer tone, while ConvIR results in undersaturated colors. In contrast, our models provide a better color
balance, with URNet being the most faithful and URCANet exhibiting a slightly cooler tone.
(a) Hazy image (b) DCP (c) Fang et al. (d) DehazeNet
(e) AODNet (f) FFA-Net (g) ConvIR (h) URNet (Ours) (i) URCANet (Ours)
Figure 5: Visual comparison of dehazing methods on an real life image. DCP and Dehazenet remove efectively the haze, but
oversaturate the sky, as the dark channel prior hypothesis are violated. Fang et al. removes the haze but generates a strong halo
around the building. AODNet and ConvIR both remove the haze, but not completely. Both URNet and URCANet remove the
haze and the results have a better color balance, even though it saturates the sky because of the violation of the DCP.
blocks. For the preprocessing, we mainly compare the
guided filter used by He et al. in (He et al., 2011) and
the GIF Block in (Yang and Sun, 2018) to refine the
transmission map. We have realised that He’s guided
filter does not effectively refine the borders of the im-
age, as can be seen in Figure 6, so we have chosen the
GIF block to filter t
0
. For the postprocessing, we have
studied how different blocks affect the output of the
algorithm. The candidates tested are: a residual block
(RESB), the same one depicted in Figure 1d, as a de-
noising tool; a channel attention block (CAB) (Woo
et al., 2018) to focus on recovering correctly the col-
ors; a spatial attention block (SAB) (Woo et al., 2018)
to address any possible problem resulting on d’s esti-
mation; and no block at all. Among them, the residual
block yielded the best results in terms of PSNR and
SSIM, as can be seen in Table 2.
Table 2: Comparison of PSNR values obtained from differ-
ent postprocessing blocks during the first 100 epochs.
RESB CAB SAB No block
PSNR ↑ 21.97 20.83 20.87 20.86
SSIM ↑ 0.902 0.902 0.901 0.902
SAM ↓ 0.075 0.067 0.067 0.066
Last, we discuss the loss function. Let J be the
recovered image and GT the ground truth. Our first
idea was to use either the L
1
or the MSE, that is,
L
1
(GT , J) = ∥GT − J∥
1
, (10)
or
MSE(J, GT ) = ∥GT − J∥
2
2
, (11)
Two Simple Unfolded Residual Networks for Single Image Dehazing
521

(a) (b) (c)
Figure 6: (a) Hazy image (b) Output using a GIF Block
(c) Output using He’s guided filter. Here we see how the
borders on (c) are not fully refined, with a darker frame ap-
pearing on the borders of the image.
respectively. We found that the L
1
norm was a better
choice, as it did not smooth the edges like the MSE
and preserved color more effectively. However, nei-
ther of them recovered correctly the edges, causing
the appearance of halos around the objects. Then, we
(a) (b) (c)
Figure 7: (a) Hazy image (b) Output using (12) as loss (c)
Output using MSE instead of L
1
in (12). We see how (b)
recovers a sharper image with a correct color balance, while
(c) is a bit undersaturated.
added a weighted sum of the loss at each stage of the
unfolded algorithm. Again, L
1
performs better than
MSE (see Figure 7). In the end, the final loss is set to
L (GT , J , {J
i
}
N−1
i=1
) = L
1
(GT , J)
+
ω
N
N−1
∑
i=1
L
1
(GT , J
i
),
(12)
where N is the total number of stages and ω is a con-
stant. After different tests, we set ω = 0.3 as it bal-
ances edge preservation and overall image quality.
6 CONCLUSIONS
In this paper, we have proposed two simple unfolded
residual networks for single-image dehazing. In both
cases, we have designed an energy functional to be
minimized via proximal gradient descent. On one
hand, this gives us a solid mathematical foundation
and a clear interpretation of all the variables involved
in the problem. However, the derivation of this func-
tional involved imposing some restrictive priors on
the fidelity terms, such as d being close to d
0
. This
could compromise the results when such hypothesis
are violated or, for instance, if t
0
is not accurately es-
timated. However, the unfolding process addresses
some of these problems, which can be seen comparing
Fang’s classical variational model with ours in Fig-
ures 3-5.
The results demonstrate that laying a robust math-
ematical framework not only aids in understanding
the modeling process but also facilitates the develop-
ment of efficient, interpretable neural networks that
perform comparably to state-of-the-art methods. Al-
though many models prioritize performance over in-
terpretability, our approach shows that sometimes tak-
ing a step back to lay a solid foundation can result in
simpler and more effective solutions.
ACKNOWLEDGMENTS
This work is part of the MoMaLIP
project PID2021-125711OB-I00 funded by
MCIN/AEI/10.13039/501100011033 and the Euro-
pean Union NextGeneration EU/PRTR.
REFERENCES
Cai, B., Xu, X., Jia, K., Qing, C., and Tao, D. (2016). De-
hazenet: An end-to-end system for single image haze
removal. IEEE Transactions on Image Processing,
25(11):5187–5198.
Chambolle, A. and Pock, T. (2016). An introduction to
continuous optimization for imaging. Acta Numerica,
25:161–319.
Chi, J., Li, M., Meng, Z., Fan, Y., Zeng, X., and Jing,
M. (2020). Single image dehazing using a novel his-
togram tranformation network. In 2020 IEEE Inter-
national Symposium on Circuits and Systems (ISCAS),
pages 1–5.
Choi, L. K., You, J., and Bovik, A. C. (2015). Reference-
less prediction of perceptual fog density and percep-
tual image defogging. IEEE Transactions on Image
Processing, 24(11):3888–3901.
Cui, Y., Ren, W., Cao, X., and Knoll, A. (2024). Revi-
talizing convolutional network for image restoration.
IEEE Transactions on Pattern Analysis and Machine
Intelligence, pages 1–16.
Fang, C., He, C., Xiao, F., Zhang, Y., Tang, L., Zhang, Y.,
Li, K., and Li, X. (2024). Real-world image dehazing
with coherence-based label generator and cooperative
unfolding network. arXiv preprint arXiv:2406.07966.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
522

Fang, F., Li, F., and Zeng, T. (2014). Single image dehaz-
ing and denoising: A fast variational approach. SIAM
Journal on Imaging Sciences, 7(2):969–996.
Fattal, R. (2008). Single image dehazing. ACM Trans.
Graph., 27(3):1–9.
Galdran, A., Bria, A., Alvarez-Gila, A., Vazquez-Corral, J.,
and Bertalm
´
ıo, M. (2018). On the duality between
retinex and image dehazing. In 2018 IEEE/CVF Con-
ference on Computer Vision and Pattern Recognition,
pages 8212–8221.
Galdran, A., Vazquez-Corral, J., Pardo, D., and Bertalm
´
ıo,
M. (2015). Enhanced variational image dehazing.
SIAM Journal on Imaging Sciences, 8(3):1519–1546.
Galdran, A., Vazquez-Corral, J., Pardo, D., and Bertalm
´
ıo,
M. (2017). Fusion-based variational image dehazing.
IEEE Signal Processing Letters, 24(2):151–155.
Golts, A., Freedman, D., and Elad, M. (2020). Unsu-
pervised single image dehazing using dark channel
prior loss. IEEE Transactions on Image Processing,
29:2692–2701.
Guo, X., Yang, Y., Wang, C., and Ma, J. (2022). Image
dehazing via enhancement, restoration, and fusion: A
survey. Information Fusion, 86-87:146–170.
He, K., Sun, J., and Tang, X. (2011). Single image haze
removal using dark channel prior. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence,
33(12):2341–2353.
He, K., Sun, J., and Tang, X. (2013). Guided image filtering.
IEEE Transactions on Pattern Analysis and Machine
Intelligence, 35(6):1397–1409.
Jackson, J., Agyekum, K. O., kwabena Sarpong, Ukwuoma,
C., Patamia, R., and Qin, Z. (2024). Hazy to hazy
free: A comprehensive survey of multi-image, single-
image, and cnn-based algorithms for dehazing. Com-
puter Science Review, 54:100669.
Jin, Z., Ma, Y., Min, L., and Zheng, M. (2024). Variational
image dehazing with a novel underwater dark channel
prior. Inverse Problems and Imaging.
Jun, W. and Rong, Z. (2013). Image defogging algorithm
of single color image based on wavelet transform and
histogram equalization. Applied Mathematical Sci-
ences, 7:3913–3921.
Lei, L., Cai, Z.-F., and Fan, Y.-L. (2024). Single image
dehazing enhancement based on retinal mechanism.
Multimedia Tools and Applications, 83(21):61083–
61101.
Li, B., Peng, X., Wang, Z., Xu, J., and Feng, D. (2017).
Aod-net: All-in-one dehazing network. In Proceed-
ings of the IEEE international conference on com-
puter vision, pages 4770–4778.
Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., and
Wang, Z. (2019). Benchmarking single-image dehaz-
ing and beyond. IEEE Transactions on Image Pro-
cessing, 28(1):492–505.
Li, P., Tian, J., Tang, Y., Wang, G., and Wu, C. (2021). Deep
retinex network for single image dehazing. IEEE
Transactions on Image Processing, 30:1100–1115.
Liu, Q., Gao, X., He, L., and Lu, W. (2018). Single im-
age dehazing with depth-aware non-local total varia-
tion regularization. IEEE Transactions on Image Pro-
cessing, 27(10):5178–5191.
Liu, Y., Yan, Z., Wu, A., Ye, T., and Li, Y. (2022). Nighttime
image dehazing based on variational decomposition
model. In 2022 IEEE/CVF Conference on Computer
Vision and Pattern Recognition Workshops (CVPRW),
pages 639–648.
McCartney, E. J. (1977). Optics of the atmosphere: Scat-
tering by molecules and particles. Physics Bulletin,
28(11):521.
Pereira-S
´
anchez, I., Sans, E., Navarro, J., and Duran, J.
(2024). Multi-head attention residual unfolded net-
work for model-based pansharpening. arXiv preprint
arXiv:2409.02675.
Qin, X., Wang, Z., Bai, Y., Xie, X., and Jia, H. (2019). Ffa-
net: Feature fusion attention network for single image
dehazing. CoRR, abs/1911.07559.
Qu, Y., Chen, Y., Huang, J., and Xie, Y. (2019). Enhanced
pix2pix dehazing network. In 2019 IEEE/CVF Con-
ference on Computer Vision and Pattern Recognition
(CVPR), pages 8152–8160.
Song, Y., He, Z., Qian, H., and Du, X. (2023). Vision trans-
formers for single image dehazing. IEEE Transactions
on Image Processing, 32:1927–1941.
Stipeti
´
c, V. and Lon
ˇ
cari
´
c, S. (2022). Variational formulation
of dark channel prior for single image dehazing. J.
Math. Imaging Vis., 64(8):845–854.
Tan, R. T. (2008). Visibility in bad weather from a single
image. In 2008 IEEE Conference on Computer Vision
and Pattern Recognition, pages 1–8.
Thanh, L. T., Thanh, D. N. H., Hue, N. M., and Prasath,
V. B. S. (2019). Single image dehazing based on
adaptive histogram equalization and linearization of
gamma correction. In 2019 25th Asia-Pacific Confer-
ence on Communications (APCC), pages 36–40.
Wang, W. and Yuan, X. (2017). Recent advances in image
dehazing. IEEE/CAA Journal of Automatica Sinica,
4(3):410–436.
Woo, S., Park, J., Lee, J.-Y., and Kweon, I. S. (2018). Cbam:
Convolutional block attention module.
Xiao, B., Zheng, Z., Zhuang, Y., Lyu, C., and Jia, X. (2024).
Single uhd image dehazing via interpretable pyramid
network. Signal Processing, 214:109225.
Yang, D. and Sun, J. (2018). Proximal dehaze-net: A
prior learning-based deep network for single image
dehazing. In Ferrari, V., Hebert, M., Sminchisescu,
C., and Weiss, Y., editors, Computer Vision – ECCV
2018, pages 729–746, Cham. Springer International
Publishing.
Zhang, H. and Patel, V. M. (2018). Densely connected
pyramid dehazing network. In 2018 IEEE/CVF Con-
ference on Computer Vision and Pattern Recognition,
pages 3194–3203.
Zhou, J. and Zhou, F. (2013). Single image dehazing
motivated by retinex theory. In 2013 2nd Inter-
national Symposium on Instrumentation and Mea-
surement, Sensor Network and Automation (IMSNA),
pages 243–247.
Two Simple Unfolded Residual Networks for Single Image Dehazing
523