Vol. 2 No.

3 (2024)

A Review on Deep Learning-based approaches for Image Dehazing

Sanaullah Memon*
Department of Information Technology, Shaheed Benazir Bhutto University
Shaheed Benazirabad, Pakistan.
sanaullah.memon_nf@sbbusba.edu.pk
Rafaqat Hussain Arain
Institute of Computer Science, Shah Abdul Latif University Khairpur, Pakistan.
rafaqat.arain@salu.edu.pk
Ghulam Ali Mallah
Institute of Computer Science, Shah Abdul Latif University Khairpur, Pakistan.
ghulam.ali@salu.edu.pk
Sidra Rehman
Department of Computer Science, Iqra University Karachi, Pakistan.
sidra.rehman_n@iqra.edu.pk
Javeria Barkat
Department of Computer Science, Iqra University Karachi, Pakistan.
javeria.barkat@iqra.edu.pk
Muhammad Ahmad Siddiqui
Department of Computer Science, University of the Punjab, Lahore. Pakistan.
asahmadsiddiqui@gmail.com

Abstract
Images captured in unpredictable weather conditions frequently suffer from
significant degradation. The scattering and absorption of airborne particles in the
atmosphere effect on image quality such as poor visibility, low contrast, and color
distortions. The problem of image degradation is addressed by many computer
vision applications in unpredictable weather conditions as these conditions
diminish the clarity of the visual scene due to loss of image details. The learning-
based image dehazing approaches play an imperative role to eliminate haze and
enhance the quality of haze-free image. This paper presents a review of different
learning-based image dehazing approaches which employ different techniques to
approximate atmospheric light and transmission map to restore a haze-free image
with image details and color fidelity.

Keywords: Image dehazing; Image degradation; Image quality; Weather;

Transmission map; Atmospheric light.

310
Vol. 2 No. 3 (2024)

Introduction
Single image dehazing is an advanced computational technique utilized to recover
visibility and improve the excellence of hazy images [1] as shown in Fig. 1. The aim
of single image dehazing is to calculate the underlying scene radiance and
eliminate the unwanted atmospheric effect caused by scattering and absorption of
light due to particles in the atmosphere such as fog, smoke, dust [2]. The
atmospheric degradation diminishes the color saturation and contrast in the
captured images, making it difficult for automated systems and human viewers to
see important details [3]. The assessment of transmission map and the atmospheric
light is included by the dehazing process. The spatially diverse haze density and
degradation in various regions of the image is indicated by the transmission map.
The dominant light in the scene is denoted by the atmospheric light, which is
owing to dispersing of light by the tiny particles [4]. Recently, several algorithms
and methods have been suggested for single image dehazing, employing various
approaches like dark channel prior (DCP) [5], color attenuation [6], and image
fusion [7]. These methods often employ optimization algorithms, advanced image
processing techniques, and machine learning models to accomplish delightful
dehazing results. Single image dehazing has achieved important attention owing to
its practical applications in fields, including surveillance, autonomous vehicles, and
outdoor imaging, where it is important to get clear and visually attractive images
even in bad weather conditions [8]. Many research papers on image dehazing have
been presented in [9-12]. The comparison of five algorithms based on physical
scattering model for image dehazing is described in [9]. Various defogging
approaches based on enhancement and restoration are explored in [10][11].
Different visibility enhancement methods introduced for both uniform and non-
uniform fog conditions are presented in [12]. In this paper, a review is conducted
on various deep learning-based image dehazing approaches. These approaches
will enable readers to comprehend the effectiveness of each approach and
contribute to the development of advanced dehazing approaches.

Fig. 1 (a) Hazy image, (b) Haze-free image

311
Vol. 2 No. 3 (2024)

Image Dehazing Approaches

The study of different research papers related to image dehazing for last six years
is described in this section. Various researchers and scholars have proposed
different approaches to analyze transmission map and atmospheric light and made
their contributions using different techniques for image dehazing. This section is
divided into five sub sections based on the different techniques and network
structures for single image dehazing. It includes end-to-end approaches, attention
based approaches, fusion based approaches, attention and fusion based
approaches, and U-Net based approaches.

End-to-End approaches
The enhanced version of the CycleGAN framework for single image dehazing
entitled Cycle-Dehaze was presented by D. Engin et al. [13]. To enhance the
dehazing performance, several modifications to an end-to-end CycleGAN [14]
architecture are developed. The approach does not necessitate the pairing of hazy
and haze-free images for training and testing. Instead, it employs CycleGAN to
obtain the style transfer from hazy images to dehazed images. Besides, the
suggested approach does not rely on assessment of the variables related to the
atmospheric scattering model. Cycle-Dehaze improves the texture information
recovery and generates a visually superior dehazed image by integrating a
perceptual loss function into the existing CycleGAN framework as illustrated in Fig.
2. Cycle-Dehaze requires significant processing power and extensive parameter
tuning to produce haze-free images. This factor makes the approach less robust
and may require domain expertise to accomplish optimal results.

Fig. 2 Qualitative results of the Cycle-Dehaze and CycleGAN approaches on I-Haze

and O-Haze datasets. From left to right, the first column pair shows the hazy
images, the second and third column pair shows the outcome of CycleGAN and
Cycle-Dehaze approaches and the fourth column pair shows the ground truth
image.

312
Vol. 2 No. 3 (2024)

A generic model-agnostic convolutional neural network for single image dehazing

was introduced by Z. Liu et al. [15]. The network structure applies the
downsampling and upsampling operations to extract different features and
produces an estimation of the desired outcome using transformation. The network
employs paired hazy and haze-free images of ITS and OTS datasets and mean
square error (MSE) loss function to train the approach in fully supervised manner. It
assists to produce visually pleasing haze-free images. The network eliminates haze,
enhances image details and accomplishes better performance in terms of dehazing
quality. The generic model-agnostic approach facilitates network to handle
different hazy scenarios without explicitly modeling the atmospheric scattering
process.

H. H. Yang et al. [16] suggested network for single image dehazing named Y-Net.
The network combines multi-scale features, allowing for better representation of
haze-related details and context. It supports the wavelet transform to extract
structural information which assists in preserving significant image details during
the dehazing process. The wavelet SSIM loss function is utilized for training the
network where it employs a series of discrete wavelet transformations to segregate
the image into patches of varying sizes, each characterized by various frequencies
and scales as shown in Fig. 3. Y-Net is evaluated on the RESIDE dataset and
compared against existing image dehazing approaches. The experimental findings
show that the network accomplishes greater performance using both the
qualitative and quantitative metrics.

Fig. 3 (a) The process of the discrete wavelet transform. (b) The real image. (c) The
outcome obtained from applying the discrete wavelet transform twice. (d) The
ratios pertaining to various patches.

313
Vol. 2 No. 3 (2024)

Y. Shao et al. [17] suggested domain adaptation method to tackle the single image
dehazing problem where the training and testing data come from different
domains. The image translation module and two dehazing modules are comprised
by the domain adaptation structure. To establish a connection between synthetic
and real domains, the bidirectional translation network is employed effectively
enabling the translation of images between the two domains. The results obtained
from translation of two synthetic hazy images are shown in Fig. 4. Subsequently,
the images are utilized to train these two image dehazing networks before and
after translation, while enforcing a consistency constraint. The real hazy images into
the dehazing training process are integrated during this phase, utilizing the
characteristics of clear images to enhance the domain adaptively. While training
both the image translation and dehazing networks, the enhanced outcomes are
achieved by the approach.

Fig. 4 The results obtained from translation of two synthetic hazy images. From left
to right (a-b), (a) Synthetic hazy image, (b) Translated image.

A. Singh et al. [18] described single image dehazing approach which handles
various types of challenging haze scenarios such as dense haze and non-
homogeneous haze. The approach uses a back projected pyramid network (BPPN)
architecture that contains different blocks. The pyramid convolution technique is
developed to acquire spatial features of various levels. The iterative U-Net block
learns complex and distinct haze features without loss of the structural information.
The four contemporary challenging datasets of diverse haze scenarios are utilized
to optimize the performance. The network is trained employing the incorporation
of MSE loss, content loss, adversarial loss, and structural similarity loss. The
suggested approach is assessed on the challenging datasets and compared with
other dehazing approaches. Experimental findings show that the BPPN
accomplishes competitive dehazing performance across different types of haze
scenarios.

314
Vol. 2 No. 3 (2024)

Attention based approaches

Y. Lee et al. [19] proposed a novel approach for image dehazing. In this approach,
the benefits of a U-Net architecture with contextualized attention mechanisms are
combined to enhance the quality of haze-free images. The contextualized attentive
U-Net model combines the parallel dilated convolution module and the squeeze-
and-excitation modules that demonstrate outstanding performance in image
segmentation tasks. The encoder-decoder network structure captures the
contextual and attentive features of the input hazy image and reconstructs the
dehazed image. The contextualized attention mechanism enables the network to
pay attention on important image regions during the dehazing process. With the
incorporation of contextual information, the network understands the global scene
features and makes informed decisions when removing haze. RESIDE dataset is
utilized for training the proposed approach. The training process includes mean
square error and perceptual loss to calculate the inconsistency between the
predicted and ground truth haze-free images. The approach is evaluated on
synthetic and real world images and compares it with other algorithms. The
experimental findings show that the proposed approach achieves better dehazing
performance in terms of subjective and objective metrics. The contextualized
attentive U-Net effectively diminishes haze and improves the visibility of fine
details in hazy images.

An end-to-end single image dehazing network named AED-Net was proposed by S.

A. Hovhannisyan et al. [20]. The network recovers essential scene information
without depending on atmospheric scattering model, external information, or
various images of same scene. To improve the ability of network, the region-aware
modified Gamma correction (RAMGC) is integrated to refine edges and distorted
colors as shown in Fig. 5. The four loss functions such as smooth L1 loss, MS-SSIM
loss, perceptual loss, and adversarial loss are utilized to compute the numerical
disparity between the dehazed results and ground truth images. For experimental
findings, three datasets namely NH-HAZE2, I-Haze, O-Haze are utilized to train and
assess the network. The AED-Net shows promising findings in terms of image
dehazing quality, outperforming numerous existing dehazing algorithms. The
effectiveness of algorithm makes it appropriate for various real-world applications
requiring single image dehazing.

315
Vol. 2 No. 3 (2024)

Fig. 5 The efficiency of the region-aware modified Gamma correction (RAMGC)

The approach employs a combination of generative adversarial networks and an

attention mechanism for improving hazy images was proposed by Y. Ma et al. [21].
Hazy images frequently endure from diminished visibility and color distortion,
making them less visually attractive. The generative adversarial networks consist of
a generator and a discriminator network, which is utilized to address these issues.
The approach does not entail paired datasets and does not depend on
atmospheric scattering model during the haze-removing process. It integrates
channel attention and domain attention mechanisms into the generator network
that allows the model to concentrate on significant regions of the image to
improve significant image details and textures while suppressing noise and artifacts.
Dense blocks are employed to augment the depth of the network and enhance its
ability for feature extraction. The generator network recovers the background
details during dehazing process, while the discriminator network differentiates
between the generated and real clear images. Cycle-consistency loss is utilized to
reduce the discrepancy between the hazy images and their reconstructed
counterparts. The model gradually enhances its ability to dehaze hazy images
effectively by optimizing the generative adversarial network. Experimental results
show the superiority of the proposed approach in terms of both visual quality and
objective metrics as compared with several existing dehazing approaches.

An end-to-end deep learning-based approach for real-time single image dehazing

was presented by C. Y. Jeong et al. [22]. A zoomed convolution group is developed
for reducing the processing time of model without compromising the excellence of
recovered image. To improve the dehazing performance, efficient channel
attention mechanisms are incorporated in the network. The L1 loss is employed for
model training. For experimental results, the RESIDE dataset is employed to train
and evaluate the dehazing performance of model. The approach accomplishes
better dehazing results while managing real-time processing speed, making it
suitable for applications that require efficient and fast image dehazing.

316
Vol. 2 No. 3 (2024)

Fusion based approaches

The novel fusion based approach for single image dehazing was suggested by W.
Ren et al. [23]. The network leverages both the local and global information and
acquires new strategy based on fusion to improve the dehazing performance. The
network obtains three inputs from the hazy image employing encoder-decoder
structure. The encoder captures the contextual information of the input image,
which is then employed by the decoder to calculate the individual contributions of
each input towards accomplishing the ultimate deblurring outcome, resulting in a
confidence map at the pixel level. The input images are gated and merged by
employing the confidence map, resulting in the haze-free image as shown in Fig. 6.
Gated fusion network requires substantial processing power and memory resources.
It may fail to generate satisfactory outcomes in variations of weather and lighting
conditions, and haze densities.

Fig. 6 The efficiency of the gated fusion network (GFN)

The approach for image dehazing and deraining is proposed by D. Chen et al. [24],
which utilizes a smoothed dilation technique to eliminate grid artifacts caused by
the dilated convolution. The features from various levels are fused employing
gated subnetworks. The image is improved by collecting information from
neighboring regions and fusing features from various levels. Mean square error
loss function is utilized to train the network with RESIDE dehazing benchmark
which contains synthetic images. Experimental results demonstrate that GCANet
accomplishes outstanding performance in single image dehazing. This CNN based
approach still possesses some limitations. The image possesses the less contextual
information. As the dilation rate rises, the information from the nearest elements of
the convolution kernel becomes highly varied leading to grid artifacts in the haze-
free results. Furthermore, this approach is not suitable to generate highly detailed
information.

317
Vol. 2 No. 3 (2024)

The deep fusion approach for single image dehazing was introduced by Z. Deng et
al. [25], which combines several dehazing models to separate layers to improve the
quality of hazy image. It comprises three stages to produce the final dehazed
image. Initially, the attentional feature integration module is formulated to improve
the incorporation of features at diverse convolution neural network layers, and
produce attentional multi-level integrated features. Subsequently, these features
are employed to produce a haze-free output using an atmospheric scattering
model and four haze-layer separation models. These outcomes are then combined
to generate the final dehazed image. In order to access the dehazing performance,
the network is compared with various image dehazing approaches using two
synthetic and real-world benchmark datasets. Experimental findings prove that the
suggested approach accomplishes outstanding dehazing performance. It generates
dehazed results with the improved image details and diminished artifacts.

A multi-scale approach with dense feature fusion was proposed by J. Pan et al. [26]
that leverages both local and global information for effective dehazing. The
proposed approach employs two principles such as boosting and error feedback to
solve the dehazing problem. With the incorporation of boosting strategy, the
network design is effective to recover the dehazed image. To enhance the network
performance, a dense feature fusion module integrates back-projection technique
in the network. This fusion assists to capture multi-scale details and improves the
representation power of the network. Experimental findings on different datasets
exhibit that the network accomplishes good performance in terms of dehazing
quality, while comparing to state-of-the-art approaches. The network eliminates
haze while preserving image details and generates visually pleasing results. The
boosting strategy and dense feature fusion module with back-projection technique
contribute to the overall success of the proposed approach.

The U-Net architecture for image dehazing was proposed by G. Fan et al. [27]. The
proposed network structure leverages depth information to improve the dehazing
process. It combines multi-scale depth maps at various stages employing encoder-
decoder structure with skip connections. The network captures both local and
global cues, enables more precise and comprehensive dehazing while fusing depth
information at multiple scales. The negative SSIM loss function is utilized to train
the network. The synthetic image dataset, NYUv2 depth dataset and Make3D
dataset are used to verify the approach. It ensures that haze-free images preserve
both visual and depth information. The experimental findings illustrate that the
network attains greater dehazing performance by incorporating the multi-scale

318
Vol. 2 No. 3 (2024)

depth information which removes haze, improves visibility, and generates high
quality haze-free images.

Single image dehazing using multi-scale approach was proposed by Z. Chen et al.
[28], which integrates both global and local features at various scales effectively. It
improves hazy images that are suffered from color distortion, reduced contrast,
and loss of fine details. The approach comprises two feature extraction modules
and one deep fusion module. The global features are computed in the global
feature extraction module which captures the overall scene transmission and
atmospheric light. Multi-scales are considered to handle object sizes and multiple
levels of haze. A deep fusion module is utilized to combine the global and local
features through skip connections, where the local features portray the image
contents. The fusion strategy integrates the complementary information from both
types of features, improving the overall dehazing performance. To train the
network, mean square error loss function is utilized to compute the difference
between the haze-free image and ground truth image. For experimental results,
artificially synthesized foggy images are used to train and evaluate the proposed
approach. Experimental findings demonstrate that the proposed approach
accomplishes significant improvements in terms of color fidelity, visibility, and
preservation of fine details when compared to other dehazing algorithms.

J. Xu et al. [29] presented the innovative approach for single image dehazing which
integrates the transformer and convolution neural network architectures. For
improving the dehazing capability, the network captures both the global and local
features using transformer-convolution hybrid layer. The adaptive fusion
mechanism accomplishes a trainable merging of the output findings from both the
swin-transformer and the optional convolution blocks. The five subsets of the
RESIDE dataset are employed to train the network and L1 loss function is employed
to ensure the generation of visually pleasing haze-free images. The experimental
findings illustrate that the suggested approach accomplishes superior performance
compared to existing dehazing approaches. It effectively eliminates haze, enhances
image visibility, and preserves image details. Moreover, the integration of
transformer and CNN architectures provides a synergistic effect, improving the
efficiency of the dehazing approach.

Attention and Fusion based approaches

The GridDehazeNet for single image dehazing was presented by X. Liu et al. [30].
The network does not depend on atmospheric scattering model. It comprises three
modules such as pre-processing module, backbone module, and post-processing

319
Vol. 2 No. 3 (2024)

module. The pre-processing module produces relevant features in its learned

inputs, surpassing the limited potential of manually selected pre-processing
techniques as shown in Fig. 7. The structure utilized for the backbone network is
GridNet [31], which integrates a new technique to estimate the multi-scales
employing a grid network and attention mechanism. This method effectively
alleviates the common bottleneck issue faced by the conventional multi-scale
methods. The post-processing module assists in minimizing artifacts from final
output. Extensive experimental findings exhibit that the suggested approach
accomplishes greater performance in terms of dehazing quality, color fidelity,
brightness maintenance and outperforms existing methods on the large scale
synthetic dataset. Furthermore, the network sometimes generates dark artifacts in
some smooth areas.

Fig. 7 The Judgment of Hazy image, dehazed image, and multiple learned inputs

An attention-based deep learning approach named FFA-Net was suggested by X.

Qin et al. [32], that effectively eliminates haze from images by incorporating feature
fusion and attention mechanisms. The network structure comprises three
components such as Feature attention, block structure, and attention-based
feature fusion. Various features and pixels are treated unequally by feature
attention, enabling increased flexibility in handling diverse information types. The
local residual learning and feature attention are incorporated by block structure.
Through local residual learning, less significant information can be bypassed
employing several local residual connections. This enables the network to
concentrate more on significant information. The attention-based feature fusion
employs feature attention module at various levels, through which the weights of
the features are adaptively learned assigning greater importance to the significant
features. Simple L1 loss function is utilized to train the network with RESIDE
dehazing benchmark that contains synthetic hazy images. Experimental results

320
Vol. 2 No. 3 (2024)

demonstrate that FFA-Net accomplishes outstanding performance in single image

dehazing. The attention-based feature fusion allows the network to eliminate haze
while preserving image details and producing visually attractive results.

A novel approach titled “Hierarchical Feature Fusion with Mixed Convolution

Attention for Single Image Dehazing” was proposed by X. Zhang et al. [33]. The
approach comprises an end-to-end network structure with skip connections to
extract multi-level features using a feature extraction block. The mixed convolution
attention mechanism is utilized to lessen redundancy among features, adaptively
emphasize important features while squashing inappropriate information, assisting
effective feature fusion. The deep semantic loss, perceptual loss, MSE loss, and
smooth L1 loss are utilized to compute the statistical disparity between the haze-
removed results and real dehazed images. The synthetic and real-world datasets,
namely RESIDE, I-Haze, and O-Haze are utilized to train and test the approach for
experimental results. Experimental assessments on benchmark datasets exhibit that
the proposed approach accomplishes superior performance in terms of dehazing
quality and objective evaluation metrics. The dehazed images preserve important
details and produce more visually realistic results. The mixed convolution attention
and hierarchical feature fusion contribute to improving visibility and eliminating
haze efficiently, assisting the potential of the approach for single image dehazing
applications.

A novel approach named as “Multi-stream Fusion Network With Generalized

Smooth L1 Loss for Single Image Dehazing” was suggested by X. Zhu et al. [34].
The information of multi-streams is utilized by the network to improve the
dehazing process. The network structure combines different components like an
encoder-decoder structure, attention mechanisms, and skip connections, to
capture and refine significant features at various scales. A generalized smooth L1
loss function is designed for training and addressing the network dehazing
challenges. The robust and accurate dehazing results are promoted by
incorporating the advantages of smooth L1 and L2 loss functions. The synthetic
and realistic image dehazing datasets are employed to train and test the approach.
The suggested approach attains better dehazing performance employing
qualitative and quantitative evaluation metrics on benchmark datasets. It removes
haze effectively from images and improves visibility, examining its potential for
real-world applications.

A two-stage approach for single image dehazing employing an encoder-decoder

network structure was introduced by X. Li et al. [35], that leverages the Swin-

321
Vol. 2 No. 3 (2024)

Transformer model to effectively recover haze-free images from hazy inputs. The
approach comprises two stages. In the first stage, a transformer-cnn codec is
developed to extract and merge both local and global features. An inter-block
supervision mechanism reduces the loss of feature information resulting from
upsampling and downsampling processes, thereby enriching the features. In the
next stage, the local features are extracted by the original resolution block
following the process of interaction and feature fusion. Furthermore, the
combination of shallow and deep features is facilitated by the integration of fusion
attention mechanism between the stages, thereby enhancing the learning
competence of the network. The network is trained employing joint loss function.
RESIDE, I-Haze, and O-Haze benchmark datasets are employed for training and
evaluating the proposed approach. Experimental findings illustrate that the
dehazing performance of proposed approach is greater as compared to various
other approaches.

The single image dehazing approach was suggested by S. Memon et al. [36]. The
approach integrates multi-stream features at three different resolution levels. The
attention mechanism is utilized to adaptively emphasize important features while
squashing inappropriate features. Deep semantic loss, smooth L1 loss, and
perceptual loss are utilized to compute the statistical variation between the
dehazed results and real images. For experimental findings, RESIDE and
externelcvpr are employed to train and assess the approach. The suggested
approach gets improved performance in terms of qualitative and quantitative
evaluation metrics on synthetic and real-world datasets. The approach effectively
removes haze from images, improves visibility and retains images with sharp
textural and structural details.

U-Net based approaches

Pavan A et al. [37] suggested a novel approach to eliminate haze from image
named LCA-Net. The LCA-Net architecture integrates the benefits of convolutional
neural networks and autoencoders for effective dehazing. The autoencoder design
enables the network to learn a compact representation of input image, while
convolution layers allow the extraction of the significant features. An encoder-
decoder network structure of the LCA-Net squeezes the hazy image and recovers
the dehazed image from the latent representation. The network is trained on a
custom dataset employing the mean square error loss function. It accomplishes
greater performance compared to other dehazing approaches which is
demonstrated by the experimental assessments on benchmark datasets.

322
Vol. 2 No. 3 (2024)

Le-Anh Tran et al. [38]presented an approach for single image dehazing which
employs the transmission map extracted by adopting DCP as additional input to
the network. The approach employs encoder-decoder network architecture (U-Net),
spatial pyramid pooling module, and swish activation function to accomplish better
performance. The high-level features from the input hazy image are extracted and
analyzed by the encoder and an output haze-free image is generated by decoder.
To train the network, a combination of MSE loss, perceptual loss, and adversarial
loss are utilized to compute the difference between the dehazed outputs and the
equivalent haze-free images. For experimental findings, the four benchmark
datasets of hazy images such as Dense-Haze, I-Haze, O-Haze, and NH-Haze are
utilized to train and evaluate the approach. Experimental findings show that the
suggested approach enhances the visibility of hazy images, leading to improved
image quality and details as shown in Fig. 8.

Fig. 8 Visualization of the dehazing outcomes achieved on synthetic images, where

each pair comprises of a hazy image on the left and its corresponding dehazed
image on the right.

Performance Evaluation on Synthetic Image Datasets

The performance of various image dehazing approaches is evaluated on the
synthetic image datasets as shown in Table 1. The different loss functions are
employed to train the networks that examine the inconsistency between the
dehazed results and ground truth haze-free images. The quantitative results on the
indoor and outdoor images are considered to measure the efficiency of each
approach. With the incorporation of deep semantic loss in [33], the outcomes of
proposed network are impressive. The deep semantic loss assists model
optimization and enhances the dehazing performance on indoor and outdoor
synthetic images. The dehazed images preserve significant details and generate
more visually attractive results than other dehazed approaches.

323
 Vol. 2 No. 3 (2024)

Table 1. Quantitative results with loss functions of various dehazing approaches on

synthetic images
Approach Loss Functions Metrics
Mean square error Loss (MSE), Smooth L1 Loss (L1),
Indoor Outdoor
Perceptual Loss (Perc), Deep semantic Loss (DS), Structural
similarity index measure Loss (SSIM), Content Loss (Cont),
Adversarial Loss (Adver), Cycle-consistency Loss (CC)
MSE L1 Perc DS SSIM Cont Adver CC PSNR SSIM PSNR SSIM
Cycle-Dehaze [13]   18.03 0.80 19.92 0.64
GMAN [15]  27.94 0.897 26.00 0.936
Y-Net [16]  --- --- 26.61 0.947
Domain    --- --- 27.76 0.93
Adaptation [17]
BPPN [18]     22.56 0.8994 24.27 0.8919
DSEU [19]   23.57 0.917 28.31 0.955
AED-Net [20]     20.75 0.872 25.56 0.845
A-CycleGAN [21]  26.428 0.886 27.476 0.947
Real-time  35.59 0.9854 --- ---
dehazing [22]
GFN [23]   22.30 0.8800 28.29 0.9621
GCANet [24]  30.23 0.9800 --- ---
DM2F-Net [25]  34.29 0.9844 29.37 0.9464
MSBDN [26]  28.01 0.9109 27.96 0.9465
MSDFN [27]  30.88 0.9965 33.73 0.998
Multi-scale single  26.90 0.9651 21.79 0.9048
image dehazing
[28]
TCFDN [29]  37.62 0.9910 --- ---
GridDehazeNet   32.16 0.9836 30.86 0.9819
[30]
FFA-Net [32]  36.39 0.9886 33.57 0.9840
Hierarchical     35.21 0.9954 34.98 0.9920
Feature Fusion
Network [33]
MSFNet [34]  34.74 0.9895 32.10 0.9849
Two-stage single  30.84 0.9628 36.33 0.9836
image dehazing
network [35]
AMSFF-Net [36]    34.87 0.9899 32.23 0.9854
LCA-Net [37]  18.23 0.7808 23.37 0.8763
EDN-GTM [38]    22.90 0.8270 23.46 0.8198

324
Vol. 2 No. 3 (2024)

Conclusion
This paper presents a review on deep learning-based image dehazing approaches.
The performance of various approaches is assessed by evaluating the quantitative
results using loss functions on synthetic images. The hierarchical feature fusion
network accomplishes the superior performance than other dehazing approaches.
The haze-free image preserves important details and produces more visually
realistic results. The mixed convolution attention and hierarchical feature fusion
contribute to improving visibility and eliminating haze efficiently. Further, the
advancements in deep learning-based approaches have improved the quality of
haze-free images. The exploration of various network architectures, attention
mechanisms and incorporation of generative adversarial networks have led to
distinguished progress in handling intricate scenes and challenging haze
conditions. The continued research in this field embraces great promise to further
improve the performance of single image dehazing approaches.

References
[1] S. Hong, M. Kim, and M. G. Kang, “Single image dehazing via atmospheric
scattering model-based image fusion,” Signal Processing, vol. 178, 2021, doi:
10.1016/j.sigpro.2020.107798.
[2] F. Guo, X. Zhao, J. Tang, H. Peng, L. Liu, and B. Zou, “Single image dehazing
based on fusion strategy,” Neurocomputing, no. xxxx, 2019, doi:
10.1016/j.neucom.2019.09.094.
[3] H. Wu et al., “Contrastive Learning for Compact Single Image Dehazing,” Proc.
IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., vol. 1, pp. 10546–
10555, 2021, doi: 10.1109/CVPR46437.2021.01041.
[4] R. Fattal, “Dehazing using color-lines,” ACM Trans. Graph., vol. 34, no. 1, pp.
1–14, 2014, doi: 10.1145/2651362.
[5] K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel
prior,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 12, pp. 2341–2353,
2011, doi: 10.1109/TPAMI.2010.168.
[6] D. Ngo, G. D. Lee, and B. Kang, “Improved color attenuation prior for single-
image haze removal,” Appl. Sci., vol. 9, no. 19, 2019, doi: 10.3390/app9194011.
[7] Y. Wang, Z., Li, F., Cong, R., Bai, H., & Zhao, “Adaptive feature fusion network
based on boosted attention mechanism for single image dehazing,”
Multimed. Tools Appl., vol. 81, no. 8, pp. 11325–11339, 2022.
[8] M. Zheng, G. Qi, Z. Zhu, Y. Li, H. Wei, and Y. Liu, “Image Dehazing by An
Artificial Image Fusion Method based on Adaptive Structure Decomposition,”
vol. 14, no. 8, 2020, doi: 10.1109/JSEN.2020.2981719.
[9] Q. Liu, H. Zhang, M. Lin, and Y. Wu, “Research on image dehazing algorithms

325
Vol. 2 No. 3 (2024)

based on physical model,” 2011 Int. Conf. Multimed. Technol. ICMT 2011, pp.
467–470, 2011, doi: 10.1109/ICMT.2011.6003078.
[10] A. K. Tripathi and S. Mukhopadhyay, “Removal of fog from images: A review,”
IETE Tech. Rev. (Institution Electron. Telecommun. Eng. India), vol. 29, no. 2, pp.
148–156, 2012, doi: 10.4103/0256-4602.95386.
[11] M. K. Saggu and S. Singh, “A review on various haze removal techniques for
image processing.,” Int. J. Curr. Eng. Technol., vol. Vol 5, no. No 3, pp. 1500–
1505, 2015.
[12] J. P. Tarel, N. Hautière, L. Caraffa, A. Cord, H. Halmaoui, and D. Gruyer, “Vision
enhancement in homogeneous and heterogeneous fog,” IEEE Intell. Transp.
Syst. Mag., vol. 4, no. 2, pp. 6–20, 2012, doi: 10.1109/MITS.2012.2189969.
[13] D. Engin, A. Genc, and H. K. Ekenel, “Cycle-dehaze: Enhanced cyclegan for
single image dehazing,” IEEE Comput. Soc. Conf. Comput. Vis. Pattern
Recognit. Work., vol. 2018-June, pp. 938–946, 2018, doi:
10.1109/CVPRW.2018.00127.
[14] A. A. Zhu, J.Y., Park, T.; Isola, P.; Efros, “Unpaired Image-to-Image Translation
using Cycle-Consistent Adversarial Networks Jun-Yan,” Proc. IEEE Int. Conf.
Comput. Vis., pp. 183–202, 2017,
[Online]. Available: http://link.springer.com/10.1007/978-1-60327-005-2_13
[15] Z. Liu, B. Xiao, M. Alrabeiah, K. Wang, and J. Chen, “Single Image Dehazing
with a Generic Model-Agnostic Convolutional Neural Network,” IEEE Signal
Process. Lett., vol. 26, no. 6, pp. 833–837, 2019, doi:
10.1109/LSP.2019.2910403.
[16] H. H. Yang, C. H. H. Yang, and Y. C. James Tsai, “Y-Net: Multi-Scale Feature
Aggregation Network with Wavelet Structure Similarity Loss Function for
Single Image Dehazing,” ICASSP, IEEE Int. Conf. Acoust. Speech Signal Process.
- Proc., vol. 2020-May, no. 1, pp. 2628–2632, 2020, doi:
10.1109/ICASSP40776.2020.9053920.
[17] Y. Shao, L. Li, W. Ren, C. Gao, and N. Sang, “Domain adaptation for image
dehazing,” Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., pp.
2805–2814, 2020, doi: 10.1109/CVPR42600.2020.00288.
[18] A. Singh, A. Bhave, and D. K. Prasad, “Single image dehazing for a variety of
haze scenarios using back projected pyramid network,” arXiv:2008.06713v1,
pp. 1–16, 2020.
[19] Y. Lee and L. Wong, “IMAGE DEHAZING WITH CONTEXTUALIZED ATTENTIVE
U-NET Visual Processing Lab , Faculty of Computing and Informatics ,
Multimedia University , Malaysia,” pp. 2–6.
[20] S. A. Hovhannisyan, H. A. Gasparyan, S. S. Agaian, and A. Ghazaryan, “AED-
Net: A Single Image Dehazing,” IEEE Access, vol. 10, pp. 12465–12474, 2022,

326
Vol. 2 No. 3 (2024)

doi: 10.1109/ACCESS.2022.3144402.
[21] Y. Ma, J. Xu, F. Jia, W. Yan, Z. Liu, and M. Ni, “Single image dehazing using
generative adversarial networks based on an attention mechanism,” IET
Image Process., no. June 2021, pp. 1897–1907, 2022, doi: 10.1049/ipr2.12455.
[22] C. Y. Jeong, K. D. Moon, and M. Kim, “An end-to-end deep learning approach
for real-time single image dehazing,” J. Real-Time Image Process., vol. 20, no.
1, pp. 1–11, 2023, doi: 10.1007/s11554-023-01270-2.
[23] W. Ren et al., “Gated Fusion Network for Single Image Dehazing,” Proc. IEEE
Comput. Soc. Conf. Comput. Vis. Pattern Recognit., pp. 3253–3261, 2018, doi:
10.1109/CVPR.2018.00343.
[24] D. Chen et al., “Gated context aggregation network for image dehazing and
deraining,” Proc. - 2019 IEEE Winter Conf. Appl. Comput. Vision, WACV 2019,
vol. 00151, no. 1, pp. 1375–1383, 2019, doi: 10.1109/WACV.2019.00151.
[25] Z. Deng and L. Zhu, “Deep Multi-Model Fusion for Single-Image Dehazing,”
Int. Conf. Comput. Vis., pp. 2453–2462, 2019, doi: 10.1109/ICCV.2019.00254.
[26] J. Pan, L. Xiang, Z. Hu, X. Zhang, and F. W. M. Yang, “Multi-Scale Boosted
Dehazing Network with Dense Feature Fusion,” pp. 2157–2167.
[27] G. Fan, Z. Hua, and J. Li, “Multi-scale depth information fusion network for
image dehazing,” Appl. Intell., vol. 51, no. 10, pp. 7262–7280, 2021, doi:
10.1007/s10489-021-02236-2.
[28] Z. Chen, H. Zhuang, J. Han, Y. Cui, and J. Deng, “Multi-scale single image
dehazing based on the fusion of global and local features,” IET Image
Process., vol. 16, no. 8, pp. 2049–2062, 2022, doi: 10.1049/ipr2.12467.
[29] J. Xu, Z. X. Chen, H. Luo, and Z. M. Lu, “An Efficient Dehazing Algorithm Based
on the Fusion of Transformer and Convolutional Neural Network,” Sensors,
vol. 23, no. 1, pp. 1–15, 2023, doi: 10.3390/s23010043.
[30] X. Liu, Y. Ma, Z. Shi, and J. Chen, “GridDehazeNet: Attention-based multi-
scale network for image dehazing,” Proc. IEEE Int. Conf. Comput. Vis., vol.
2019-Octob, pp. 7313–7322, 2019, doi: 10.1109/ICCV.2019.00741.
[31] D. Fourure, R. Emonet, E. Fromont, D. Muselet, A. Tremeau, and C. Wolf,
“Residual conv-deconv grid network for semantic segmentation,” Br. Mach.
Vis. Conf. 2017, BMVC 2017, 2017, doi: 10.5244/c.31.181.
[32] X. Qin, Z. Wang, Y. Bai, X. Xie, and H. Jia, “FFA-Net: Feature fusion attention
network for single image dehazing,” AAAI 2020 - 34th AAAI Conf. Artif. Intell.,
pp. 11908–11915, 2020, doi: 10.1609/aaai.v34i07.6865.
[33] X. Zhang, J. Wang, T. Wang, and R. Jiang, “Hierarchical Feature Fusion with
Mixed Convolution Attention for Single Image Dehazing,” IEEE Trans. Circuits
Syst. Video Technol., vol. 8215, no. c, pp. 1–13, 2021, doi:
10.1109/TCSVT.2021.3067062.

327
Vol. 2 No. 3 (2024)

[34] X. Zhu, S. Li, Y. Gan, Y. Zhang, and B. Sun, “Multi-Stream Fusion Network with
Generalized Smooth L1Loss for Single Image Dehazing,” IEEE Trans. Image
Process., vol. 30, pp. 7620–7635, 2021, doi: 10.1109/TIP.2021.3108022.
[35] X. Li, Z. Hua, and J. Li, “Two-stage single image dehazing network using swin-
transformer,” IET Image Process., vol. 16, no. 9, pp. 2518–2534, 2022, doi:
10.1049/ipr2.12506.
[36] S. Memon, R. H. Arain, and G. A. Mallah, “AMSFF-Net: Attention-Based Multi-
Stream Feature Fusion Network for Single Image Dehazing,” J. Vis. Commun.
Image Represent., vol. 90, no. August 2022, p. 103748, 2023, doi:
10.1016/j.jvcir.2022.103748.
[37] S. S. S. Pavan A, Adithya Bennur, Mohit Gaggar, “LCA-Net : Light
Convolutional Autoencoder for Image Dehazing,” arXiv:2008.10325v1, no. x,
2020.
[38] L. A. Tran, S. Moon, and D. C. Park, “A novel encoder-decoder network with
guided transmission map for single image dehazing,” Procedia Comput. Sci.,
vol. 204, pp. 682–689, 2022, doi: 10.1016/j.procs.2022.08.082.

328