GGADN: Guided Generative Adversarial Dehazing
Network
Zhang Jian ( 41789332@qq.com )
Wuhan Sports University
Wanjuan Song
Hubei University of Education
Research Article
Keywords: Dehazing, GAN, Guidance, Loss function
Posted Date: April 7th, 2021
DOI: https://doi.org/10.21203/rs.3.rs-386958/v1
License: This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License
Version of Record: A version of this preprint was published at Soft Computing on August 3rd, 2021. See
the published version at
https://doi.org/10.1007/s00500-021-06049-w.
Noname manuscript No.
(will be inserted by the editor)
GGADN: Guided Generative Adversarial Dehazing
Network
Jian Zhang · Wanjuan Song
Received: date / Accepted: date
Abstract Image dehazing has always been a challenging topic in image pro-
cessing. The development of deep learning methods, especially the Genera-
tive Adversarial Networks(GAN), provides a new way for image dehazing.
In recent years, many deep learning methods based on GAN have been ap-
plied to image dehazing. However, GAN has two problems in image dehazing.
Firstly, For haze image, haze not only reduces the quality of the image, but
also blurs the d et ai l s of the image. For Gan network, it is difficult for the
generator to restore the details of the whole image while removing the haze.
Secondly, GAN model is d efi n ed as a minimax problem, which we akens the
loss fu n ct i on. It is difficult to distingu i sh whether GAN is making progress in
the training process. Therefore, we propose a Guided Generati ve Adversarial
Dehazing Network(GGADN). Different from other generation adversarial net-
works, GGADN adds a guided module on the generator. The guided module
verifies the network of each layer of the generator. At the same time, the de-
tails of the map generated by each layer are strengthened. Network training
is based on the pre-trained VGG feature model and L1-regularized gradient
prior which is developed by new loss function paramet er s. From the dehazing
results of synthetic images and real images, proposed method is better than
the state-of-the-art dehazing methods.
Keywords Dehazing · GAN · Guidance · Loss func t ion
J. Zhang
College of Sport Engineering and Information Technology, Wuhan Sports University, Wuhan,
China
W. Song
College of Computer, Hubei University of Education, Wuhan, Hubei, China
E-mail: key
swj@whu.edu.cn
2 Jian Zhang, Wanjuan Song
1 Introduction
In computer vision, weather is an important factor affecting the quality of
image[1–6]. Especially for haze, floating particles in haze lead to the f ad in g
and blurri n g of pictures, and the reduction of contrast and softness. They
absorb and scatter light, resulting in serious color attenuation, poor clarity
and contrast, and poor visual effect, which has a serious impact on subse-
quent computer vision tasks[7–16]. Therefore, it is necessar y to re move haze
effectively.
In recent years, the research of image dehazing algorithm has made great
progress. At present, image dehazing research is mainly divided into two types,
feature-based method and learning-based method. The difficulty of feature-
based method lies in the feature extraction and a priori choice[17–20]. The
common dehazing features and priors are as foll ows:
– Contrast: Tan found that the contrast of haze free image was hi gh . Thus,
image dehazing is performed by maximizing the local contrast of the image.
– Dark channel priori: He found that the value of dar k ch ann el in haze free
image is close to zero, and then it can be used to estimate the transmission
image.
– The prior of color attenuation: Zhu found the relationship betwe en haze
concentration and brightness and satu rat i on through statistics. And a li n-
ear model of scene depth is created to solve the scene. Then, the haze-free
image is calculated.
Learning based dehazing algorithms can be divided into two kinds, step-
by-step learning algorithm and end-to-end learning algorithm. Step by st ep
learning algorithm is similar to the trad it i on al method, focusing on t h e pre-
diction of intermediate variables. For example, Cai [21] designed a dehazenet
by analyzi ng artificial prior features to complete the prediction of transmis-
sion image. Similarly, Ren [22] proposed a multi-scale convolutional neural
network mscnn, which can accurately predict the transmission image through
two different scale network models. End-to-end learning algorithm can realize
image dehazing simply and efficiently through the design of full convolutional
neural network. For example, consi de ri n g that the above algorith m ignores the
reasonable prediction of atmospheric light value, Li [23] integrated multiple in-
termediate variables in the at mos ph er i c scattering model into one using li ne ar
variation, and proposed AOD net to directly predict haze-free images.
In this paper, we pres ents a Guided Generative Adversarial Dehazing Net-
work(GGADN). Different from other generation adversarial networks, GGADN
adds a guided module on the generator. The guided module verifies the net-
work of each layer of the generator. At the same time, the details of the map
generated by each layer are strengthened. GGADN is trai n ed and corrected
in a synthetic fuzzy image dataset containing indoor and outdoor images.
Network training is based on the pre-trained Visual Geometry Group (VGG)
feature model and L1-regularized gradient prior which is developed by new loss
function parameters. From the dehazing results of synthetic images and real
GGADN: Guided Generative Adversarial Dehazing Network 3
images, proposed method is better than the state-of-th e-ar t dehazing methods.
And the image after dehazing is cl ear er in detail.
2 Related Works
2.1 Atmospheric Scattering Model
The purpose of image defogging is t o restore a clear image from the blurred
image corroded by haze or smoke. In the field of compu t er vision, in order
to overcome the image distortion caused by haze, McCartney [24] proposed
an atmospheric scat t er i n g model which can be used to describe the for mat i on
process of haze image [25–27]. The equation is as follow.
I(x) = J(x)t(x) + A(1 − t(x)) (1)
Where, I(x) is a haze image and J(x) is a haze-free image. A(x) is the
value of atmospher i c light, represe nting the intensity of atmospheric light.
T (x) is the transmittance, which indi cat es the part of light that is not scattered
when it reaches the imaging device through the atmospheric medium, and x
represents the pixel position. In the formula, the first term on the right is th e
direct attenuation term, which represents the reflected light of the object after
atmospheric attenuation, and the second term is the enhanced atmospheric
light obtained by atmospheric scattering.
When the composition of the atmosphe re is unifor m, that is, A(x) is con-
stant, the transmittance can be expressed as:
t(x) = e
−βd(x)
(2)
Where β is the attenuation coefficient of the atmosphere and D(x) is the
depth of the scene. From the formula, it is not difficult to fin d that the scene
depth and atmospheric light value have a great influence on the image dehazing
effect. In single image dehazing, only the h aze image is known, the atmospheric
light valu e and sce ne depth are unknown, and the assumption of uniform
atmospheric composition is not necessaril y true. Therefore, how to remove
haze effectively is a challenging problem.
2.2 Generative Adversarial Networks
Gan(Generative adversarial networks) is a deep learn in g model designed by
Goo df el l ow [28]. Gan is a deep l e arn i ng model, which is one of the most promis-
ing unsupervised learning methods on complex distribution in recent years
[29–35]. The model prod uc es good output through the game learning of two
modules in the framework: Generati ve model and Discriminative model. In the
original Gan theory, we don’t need that G and D are neural networks, we just
need to fit the corresponding generating and discriminating functions. But in
4 Jian Zhang, Wanjuan Song
practice, deep neural networks are generally used as G and D. An excellent
Gan application needs a good training method, otherwise the output may not
be ideal du e to the freedom of the neural network model.
However, the n etwork structure of Gan is unstable in the training process,
and some artifacts such as noise and color shift are often produced in the
synthetic i m age. Sohn [36] introduced conditional information into Gan. The
increase of conditional information variables ensures the stability of learning
process to a certain extent and improves the representation ability of generator,
but the running time i s too long. Deniz [37] propose an end-to-end ne twork
called Cycle -D eh aze for sin gle image de haz i ng. In order to improve the quality
of texture information recovery and produce a better visual haze free image,
the CycleGan agent is enhanced by combining cycle consistency and perceptual
loss. Du [38] also uses the end-to-end learning method to learn the mapping
from hazy image to haze free i mage directly. This de haz in g network generates
the confrontation training network through the Gan model. An adaptive loss
function is used in discriminator. A post -p r ocessing method for halo artifacts
removal using guided filtering is proposed.
Although the existing Gan network has achieved some results in haze re-
moval, there are also some problems. Firstly, Gan model is defined as a mini-
max problem, which has no loss function. It is difficult to dist i n gui s h whether
Gan is making progress in the training process. The learning process of Gan
may have collapse problem, and the generator begins to degenerate. It always
generates the sam e sample points and cannot continue learning. W he n the gen-
erating model collapses, the disc ri mi n at ion mod el will also point to the similar
direction for the similar sample points, so the training cannot continue. Sec-
ondly, For haze image, haze not only reduces the quality of the image, but also
blurs the details of the image. For Gan network, it is difficult for the generator
to restore the de tai l s of the whole image while removing the haze. Especially
for the image with complex structure, the effect of d eh azi ng is not ideal.
In view of the above problems, this paper improves the dehazing Gan and
designs a Guided Generative Adversarial Dehazing Network(GGADN).
3 The proposed method
This paper presents a Guided Generative Adversarial Dehazi n g Network(GGADN).
In GGADN the generator and discriminator architecture are modi fie d. The
synthetic data set is trained by end-to-end training neural network. The loss
function is modified by using the pre-trained VGG feature and L1-regularizat ion
gradient. Sigmoid function i s introduced to the last layer of discri mi nat or for
feature mapping. In order to carry out probability analysis, the discri mi nant
results can be n orm ali z ed to [0,1].