Recovering Raindrop Removal Images under Heavy Rain

Kosuke Matsumoto, Fumihiko Sakaue and Jun Sato

Nagoya Institute of Technology, Japan

{matsumoto@cv., sakaue@, junsato@}nitech.ac.jp

Keywords:

Raindrop Removal, Heavy Rain, GAN, B-spline.

Abstract:

In this paper, we propose a new method for removing raindrops in images under heavy rain. When we drive in

heavy rain, the raindrops attached to the windshield form a ﬁlm and our visibility degrades drastically. In such

situations, the existing raindrop removal methods cannot recover clear images, since these methods assume

that the background scene is visible through the gap between the raindrops, which does not happen anymore in

heavy rain. Thus, we in this paper propose a new method for recovering raindrop removal images under heavy

rain from sequential images by using conditional GAN. The results of our experiments on real images and

synthetic images show that the proposed method outperforms the state-of-the-art raindrop removal method.

1 INTRODUCTION

When driving in heavy rain, the risk of an accident

goes up since a large amount of raindrops adhere

to the windshield and hinder visibility. In recent

years, the realization of autonomous vehicles is ex-

pected, but even in the autonomous vehicles that use

in-vehicle cameras, keeping visibility in the rain is a

big problem, as with human drivers.

Thus, in recent years, some methods for removing

raindrops from in-vehicle camera images have been

proposed. However, these methods assume that the

amount of raindrops attached to the windshield is rel-

atively small hence the background scene can be ob-

served through the gaps between the raindrops. Thus,

in the case of heavy rain, when raindrops spread like

a ﬁlm over the entire windshield, these existing meth-

ods cannot generate raindrop removal images appro-

priately.

Therefore, in this paper,we propose a new method

for generating raindrop removal images properly as

shown in Fig. 1 (b) even when raindrops adhere to the

entire glass surface as shown in Fig. 1 (a).

In our method, the raindrops are not considered

as occluding objects but are considered as transparent

objects that refract incident light in various directions

and distort the observed images. Our method is based

on end-to-end learning, where the input is a set of se-

quential images distorted by heavy rain and the output

is a clear image with no distortion. We use a condi-

tional GAN framework for training a generator efﬁ-

ciently. By using the property of sequential images

(a) image under heavy rain (b) result of our method

Figure 1: Input image observed under heavy rain and rain-

drop removal image obtained from our method.

under heavy rain, where the background scene does

not change drastically while the non-uniform rain-

drops on the windshield randomly change the light

paths to the viewpoint of the camera, our network can

recover a clear undistorted image of the background

scene from a set of sequential images distorted by

heavy rain.

For training our network, we need pairs of rain-

drop and undistorted images. However, under actual

heavy rain, it is very difﬁcult to obtain correspond-

ing images with and without raindrops. Thus, we in

this research synthesize training dataset of sequential

images captured under heavy rain, and train our net-

work by using the synthetic heavy rain dataset. We

show that our network trained on the synthetic heavy

rain dataset outperforms the state-of-the-art raindrop

removal method proposed by Qian (Qian et al., 2018)

in our real image experiments.

128

Matsumoto, K., Sakaue, F. and Sato, J.

Recovering Raindrop Removal Images under Heavy Rain.

DOI: 10.5220/0009325601280136

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 4: VISAPP, pages

128-136

ISBN: 978-989-758-402-2; ISSN: 2184-4321

2 RELATED WORK

There are two major causes of poor visibility in rainy

weather. The ﬁrst one is rain streaks that block light

passing through 3D space, and the second one is rain-

drops attached to the windshield that refract light and

distort images.

For removing rain streaks in images, many authors

proposed image dehazing methods based on rain char-

acteristics (Garg and Nayar, 2004; N. and N., 2008;

Chen and Hsu, 2013; Santhaseelan and Asari, 2014),

sparse coding (Luo et al., 2015) and deep neural net-

works (Fu et al., 2017; Li et al., 2017; Yang et al.,

2017). However, these methods cannot recover im-

ages which are distorted by raindrops on the wind-

shield.

For removing raindrops attached on windows,

multiple cameras were often used to obtain the scene

information that is hidden by raindrops and invisible

at a certain viewpoint (Yamashita et al., 2005; Mat-

sui et al., 2014). The sequential images were also

used for obtaining the scene information hidden by

raindrops at a certain time instant (Yamashita et al.,

2009; Nomoto et al., 2011; You et al., 2016). More

recently, the deep learning technique is used for re-

covering raindrop removal images from a single cam-

era view (Qian et al., 2018).

Although these raindrop removal methods work

efﬁciently under light rain, they no longer work prop-

erly under heavy rain. This is because these methods

assume that the background scene can be observed

through the gaps between the raindrops, that is no

longer the case in heavy rain.

Thus, we in this paper consider the raindrops not

as occluding objects but as refractive media that dis-

tort the observed images, and propose a method for

recovering distorted images under heavy rain.

3 RAINDROP REMOVAL USING

GENERATIVE ADVERSARIAL

NETWORK

In the heavy rain, the entire image is covered with

raindrops, and the background scene cannot be ob-

served through the gaps between the raindrops. Thus,

in this research, we recover undistorted background

scene by using the entire image which is distorted by

unknown non-uniform water ﬁlm.

We assume that the 3D shape of the non-uniform

water ﬁlm changes over time, and we can observe

a set of sequential images in which the background

scene is distorted by the changing non-uniform wa-

Figure 2: Generative adversarial network (GAN) for gen-

erating raindrop removal images. Generator G generates

a raindrop removal image ˆy

from a set of sequential dis-

tortion images X and noise z. Discriminator D learns to

discriminate a false pair ( ˆy

, X) and a true pair (y

, X), and

generator G is trained so that it minimizes correct answer of

discriminator.

ter ﬁlm. We use such distorted sequential images for

recovering the undistorted back ground scene image.

Although the image distortion is dynamic and non-

uniform, we think the distorted images include the

background scene information, and hence the undis-

torted backgroundscene image can be recoveredfrom

the distorted sequential images.

For recovering undistorted scene images from the

distorted sequential images, we use conditional GAN

(Isola et al., 2017). The network structure of our con-

ditional GAN is shown in Fig. 2. The generator G

is a 16-layer convolution-deconvolution network (U-

Net) (Ronneberger et al., 2015) and the discriminator

D is a 5-layer convolution network.

The input X of the generator is a set of distorted

sequential images up to the current time t as follows:

X = {x

, x

t−1

, ·· · , x

t−T +1

} (1)

where, T is the number of time instants in the set of

sequential images, X. The generator G generates an

undistorted image ˆy

at time t from the set of distorted

sequential images X and a random noise vector z as

follows:

ˆy

= G(X, z) (2)

The ground truth of the undistorted image is denoted

as y

. Unlike the standard conditional GAN, our gen-

erator G generates a single undistorted image G(X, z)

from multiple images X.

The discriminator D is trained so that a ground

truth pair {X, y

} is determined to be true and a fake

pair {X, G(X, z)} is determined to be false. The net-

work is trained, so that the discriminator maximizes

the rate of correct judgments and the generator min-

imizes it. Thus, the training of our conditional GAN

can be described as follows:

∗

= argmin

max

cGAN

(G, D) + λL

(G) (3)

Recovering Raindrop Removal Images under Heavy Rain

129

Figure 3: B-spline surface. Red points show control points

of B-spline surface.

Figure 4: Representation of distortion by using a B-spline

surface. The magnitude of distortion at pixel (x, y) is repre-

sented by the z coordinate of the B-spline surface at (x, y).

where, L

cGAN

is the following adversarial loss:

cGAN

(G, D) = E

X,y

∼p

data

(X,y

)

[logD(X, y

)]

X∼p

data

(X),z∼p

(z)

[log(1− D(X, G(X, z)))] (4)

and L

is an L

loss as follows:

(G) = E

X,y

∼p

data

(X,y

),z∼p

(z)

− G(X, z)k

(5)

As training progresses, generator G will generate

images that make it difﬁcult for the discriminator D to

determine authenticity. In other words, a trained gen-

erator can generate raindrop removal images that are

natural to humans from a series of distorted sequential

images.

4 DATASET

Since the accuracy of deep learning depends on the

dataset, it is important to collect many data for train-

ing. However, it is very difﬁcult to obtain various

scene data with and without various raindrops, in par-

ticular under heavy rain. Thus, in this research, we

synthesize images that are distorted by various heavy

rains adhere on a windshield and build a dataset for

training.

(a) Original images

(b) Raindrop images

Figure 5: Example of raindrop images synthesized by B-

spline surfaces.

(a) T = 2, small distortion

(b) T = 3, medium distortion

Figure 6: Example images in our dataset. (a) shows an orig-

inal image and two sequential raindrop images in the case

of T = 2. (b) shows those in the case of T = 3 and (c) shows

those in the case of T = 4 respectively. The magnitude of

distortion is small in (a), and those in (b) and (c) are medium

and large respectively.

4.1 Representation of Raindrop

Distortions

Since we consider situations in which heavy rain gen-

erate non-uniform water ﬁlm on a windshield, the dis-

tortion caused by the non-uniform water ﬁlm is rep-

resented by using a B-spline surface. By using the B-

spline surface, the surface shape can be controlled by

a limited number of control points efﬁciently. Fig. 3

shows an example of B-spline surface which is gen-

erated from the control points shown in red. The B-

spline surface S(u, v)is deﬁned by using K×K control

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

130

points P

(i = 1, ·· · , K; j = 1, ··· K) as follows:

S(u, v) =

∑

i=1

∑

j=1

(u)N

(v)P

(6)

where, u and v denote parameters in horizontal and

vertical axes, and P is the degree of B-spline.

(u) and N

(v) are called blending functions

and represent the inﬂuence of the control points. In

the case of a B-spline surface, B-spline basis func-

tion is used as the blending function, which is de-

ﬁned by the position of the knot sequence x

(i =

1, ·· · , K + P+ 1), as follows:

(u) =

(

1 (x

≤ u ≤ x

i+1

)

0 (otherwise)

(7)

(u) =

u− x

i+k

− x

K−1

(u)

i+k+1

− u

i+k+1

− x

i+1

K−1

i+1

(u) (8)

In this paper, image distortion due to raindrop wa-

ter surface is represented by the B-spline surface.

4.2 Raindrop Image Synthesis

In case of heavy rain, the entire image is heavily dis-

torted by raindrops. In order to generate a distorted

image from an undistorted image, image transforma-

tion is performed by using B-spline surfaces. In our

B-spline surfaces, x axis represents the horizontal di-

rection of the image, and the y axis represents the ver-

tical direction of the image respectively. The z axis

of the B-spline surface represents the displacement

of the image point (x, y) caused by the raindrops as

shown in Fig. 4. Since we have displacement in x

and y axes, the displacement is represented by two

independent B-spline surfaces, S

(x, y) and S

(x, y).

Thus, the point coordinates (x, y) of the original im-

age is transformed into (x

′

, y

′

) in the distorted image

as follows:



′

= x+ x

(x, y)

′

= y+ y

(x, y)

(9)

where, x

(x, y) and y

(x, y) denote the displacement

of a point (x, y) in x and y axes, and these are rep-

resented by the z coordinate of S

(x, y) and S

(x, y)

respectively.

By varying the control points of the B-spline sur-

face, it is possible to represent distortions caused by

various raindrops. Thus, we control the z coordinates

of the K × K control points of S

and S

. More specif-

ically, the image distortion is controlled by a vec-

tor p = [z

, ·· · , z

, z

, ·· · , z

]

⊤

, which consists of

the z coordinates of the K × K control points of S

and

. We call it a water surface parameter.

Some example images generated by using the B-

spline surface are shown in Fig. 5.

4.3 Sequential Raindrop Image Dataset

and Training

By using the method described in section 4.2, we gen-

erated a dataset of sequential raindrop images. The

sequential images in Cityscapes dataset (Cordts et al.,

2016) were distorted by changing the water surface

parameter p randomly. The number of time instants

T in sequential images was set to 2, 3 and 4 for eval-

uating the relationship between the number of time

instants T and the accuracy of recovered raindrop re-

moval images.

Furthermore, in order to evaluate the relationship

between the magnitude of distortion and the accuracy

of undistorted image recovery, the dataset was gen-

erated changing the magnitude of image distortion in

three patterns of small, medium and large. Thus, the

sequential raindrop image dataset was generated with

3 different numbers of time instants and 3 different

patterns of distortion, and a total of 3× 3 = 9 datasets

were created. Each dataset consists of 325 training

sequences and 50 test sequences. Some examples of

nine datasets are shown in Fig. 6.

The network explained in section 3 was trained

by using the sequential raindrop image datasets. We

trained the network with 1000 epoch. Then, the

trained generator was used for generating undistorted

images from a set of distorted sequential images.

5 EXPERIMENTS

5.1 Synthetic Image Experiments

We next show the experimental results obtained from

the proposed method. We ﬁrst show results from syn-

thetic image experiments.

Fig. 7 shows results under small image distortions,

where (a) shows ground truth raindrop removal im-

ages and (b) shows input raindrop images distorted

by heavy rain. Fig. 7 (c), (d) and (e) show raindrop

removal images obtained from the proposed method

in the case of T = 2, T = 3 and T = 4 respectively.

It can be conﬁrmed that the input images in (b) have

distortions whereas the recovered images in (c), (d)

and (e) have no distortion.

Fig. 8 shows results under medium distortions. As

shown in Fig. 8 (b), we have relatively large distor-

Recovering Raindrop Removal Images under Heavy Rain

131

(a) ground truth (b) input image (c) T = 2 (d) T = 3 (e) T = 4

Figure 7: Results under small image distortions. (a) shows ground truth images, (b) shows test input images which are

distorted by raindrops, (c), (d) and (e) show raindrop removal images obtained from our method in the case of T = 2, T = 3

and T = 4 respectively.

(a) ground truth (b) input image (c) T = 2 (d) T = 3 (e) T = 4

Figure 8: Results under medium image distortions. See the caption of Fig. 7 for detail explanations.

(a) ground truth (b) input image (c) T = 2 (d) T = 3 (e) T = 4

Figure 9: Results under large image distortions. See the caption of Fig. 7 for detail explanations.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

132

(a) ground truth (b) input image (c) T = 2 (d) T = 3 (e) T = 4 (f) A

Net

Figure 10: Comparison with the existing method. Columns (a) to (e) are the same as before, and column (f) is the result of the

conventional method A

Net. The ﬁrst row shows the generated images, and the second row shows the results of segmentation

obtained from the standard segmentation network.

Table 1: Segmentation error of the proposed method (T = 2,

T = 3, T = 4) and the existing method (A

Net) under small,

medium and large image distortions.

T = 2 T = 3 T = 4 A

Net

small 0.0803 0.0776 0.0748 0.1286

medium 0.1456 0.1407 0.1293 0.1773

large 0.1851 0.1508 0.1456 0.2266

Table 2: LPIPS of the proposed method (T = 2, T = 3, T =

4) and the existing method (A

Net) under small, medium

and large image distortions.

T = 2 T = 3 T = 4 A

Net

small 0.0322 0.0291 0.0268 0.1990

medium 0.1181 0.1145 0.0927 0.1992

large 0.1179 0.1118 0.1294 0.2312

tions in input raindrop images. Nonetheless, the pro-

posed method reduced the image distortion drastically

as shown in Fig. 8 (c), (d) and (e). Also, we ﬁnd that

less distorted images were generated as we increase

the number of time instants T in the image sequence.

Fig. 9 shows results under large distortions. As

shown in Fig. 9 (b), it is almost impossible to rec-

ognize original vehicles and buildings in the input

images. Nonetheless, the proposed method recov-

ered fairly good road scene images with vehicles and

buildings. Again, image quality increases as we in-

crease the number of time instants T.

5.2 Comparison with Conventional

Methods

We next compare our method with the existing state-

of-the-art raindrop removal method. In this exper-

iment, we compared our method with A

Net (Qian

et al., 2018) proposed by Qian et al. Their method

also uses deep neural network for removing raindrops

in images. However, their method assumes that the

raindrops in images are not so many, and the back-

ground scene is occluded only partially in images.

We tested these two methods under heavy rain situ-

ations. The ﬁrst row of Fig. 10 shows the raindrop

removal images generated from the proposed method

and A

Net under heavy rain situations. As shown in

these images, A

Net cannot recover raindrop removal

images properly, while the proposed method provides

us fairly good results.

For comparing the accuracy of raindrop removal

images numerically, we computed LPIPS (Zhang

et al., 2018), which can measure semantic similar-

ity of two images. Unlike the traditional error met-

ric, such as RMSE and SSIM, it has been conﬁrmed

by many authors that LPIPS provides us image sim-

ilarity that matches the human sense. We computed

LPIPS between the ground truth images and raindrop

removal images generated by our method and A

Net.

The results are shown in Table 2. As shown in this

table, LPIPS of our method decreases as we increase

the number of time instants used in our method. We

can also ﬁnd that our method provides us much better

raindrop removal images than A

Net.

For evaluating the visibility of objects in the re-

covered images, we also evaluated our method by per-

forming the semantic segmentation on the recovered

images using a pretrained semantic segmentation net-

work. The semantic segmentation network used in

this evaluation is pix2pix (Isola et al., 2017). For

comparing the segmentation accuracy numerically,

the segmentation error was computed by normalizing

the total number of wrongly segmented pixels by the

number of total pixels N

in an image multiplied with

Recovering Raindrop Removal Images under Heavy Rain

133

(a) input images

(b) results from the proposed method

Net

Figure 11: Results from real image experiment. (a) shows images observed under heavy rain. (b) and (c) show raindrop

removal images obtained from images in (a) by using the proposed method and A

Net respectively. The visibility in (b) is

better than that in (a) and (c). For example, the white lane markers and the road guardrails are heavily distorted in (a) and (c),

but they look more accurate in (b).

the number of images N

as follows:

error =

· N

∑

i=1

∑

j=1

wrong(i, j) (10)

wrong(i, j) =







0 :

pixel j of image i is

segmented correctly

1 : others

(11)

The second row in Fig. 10 show the results of se-

mantic segmentation performed on raindrop removal

images recovered from our method and A

Net under

large image distortions. As shown in Fig. 10 (f), the

segmentation result of A

Net is very different from

that of the ground truth in (a). However, that of the

proposed method is very close to the ground truth as

shown in Fig. 10 (c), (d) and (e). Moreover, it can

be seen that the quality of the raindrop removal image

improves as we increase the number of time instants

used in the proposed method.

The improvement of accuracy in the proposed

method can be seen more clear in Table 1, which

shows the segmentation error deﬁned in Eq. (10). As

we can see in this table, the proposed method outper-

forms A

Net in all cases of small, medium and large

distortions. We can also ﬁnd that the accuracy im-

proves as we increase the number of time instants in

our method.

From these results, we ﬁnd that the proposed

method is very efﬁcient for recovering images which

are distorted by heavy rain.

5.3 Real Image Experiments

Up to now, we tested our method by using synthetic

heavy rain images. For evaluating the efﬁciency of

our method, we next apply our method to real heavy

rain images.

Fig. 11 (a) shows input images of various road

scenes observed under heavy rain. As we can see

in these images, the road scenes are heavily distorted

due to the rain water ﬁlm, and driving with these im-

ages is very difﬁcult. Fig. 11 (b) shows raindrop re-

moval images obtained from the proposed method,

and Fig. 11 (c) shows those obtained from A

Net.

From these results, we ﬁnd that both methods can-

not remove image distortions perfectly, but the pro-

posed method provides us better views of the road

scene than A

Net.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

134

Since the ground truth undistorted images of these

scenes are not available, we cannot evaluate the accu-

racy of these generated images numerically. However,

we can still ﬁnd that the visibility of the scene in (b) is

better than that in (a) and (c). For example, the white

lane markers and the road guardrails are heavily dis-

torted in (a) and (c), but they look more accurate in

(b).

Although, our method outperforms the existing

state-of-the-art method, we also ﬁnd that our method

is not perfect, and we need more improvements. In

particular, the heavy rain model used to generate the

training dataset needs to be improved. A more ac-

curate heavy rain model would lead to more accurate

raindrop removal.

6 CONCLUSIONS

In this paper, we proposed a new method for remov-

ing image distortion caused by raindrops under heavy

rain.

In heavy rain, raindrops form a non-uniform ﬁlm

on the windshield, and the visibility for a driver de-

grades drastically. The existing raindrop removal

methods cannot recover clear images in such situ-

ations, since these methods assume that the back-

ground scene is visible through the gap between the

raindrops, which does not happen anymore in heavy

rain. Thus, we in this paper proposed a new method

for recovering raindrops removal images from the se-

ries of distorted images. The results of our exper-

iments show that the proposed method outperforms

the state-of-the-art raindrop removal method in heavy

rain situations.

The proposed method is promising, but challenges

remain. In our proposed method, image degradation

due to raindrops is considered. However, in actual

heavy rains, image degradation due to rain streaks

also exists, so it is desirable to expand to a method

that improves both of these degradations.

Furthermore, it is also important to use real heavy

rain images to train the network for improving the ac-

curacy. Since the ground truth undistorted images are

not available under heavy rain, we need to consider

unsupervised leaning in the raindrop removal frame-

work.

REFERENCES

Chen, Y. and Hsu, C. (2013). A generalized low-rank ap-

pearance model for spatio-temporally correlated rain

streaks. In Proc. of International Conference on Com-

puter Vision, pages 1968–1975.

Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler,

M., Benenson, R., Franke, U., Roth, S., and Schiele,

B. (2016). The cityscapes dataset for semantic urban

scene understanding. In Proc. of the IEEE Conference

on Computer Vision and Pattern Recognition (CVPR).

Fu, X., Huang, J., Ding, X., Liao, Y., and Paisley, J. (2017).

Clearing the skies: A deep network architecture for

single-image rain removal. IEEE Transactions on Im-

age Processing, 26(6):2944–2956.

Garg, K. and Nayar, S. (2004). Detection and removal of

rain from videos. In Proc. Conference on Computer

Vision and Pattern Recognition (CVPR), volume 1.

Isola, P., Zhu, J.-Y., Zhou, T., and Efros, A. A. (2017).

Image-to-image translation with conditional adversar-

ial networks. In Proceedings of the IEEE Conference

on Computer Vision and Pattern Recognition (CVPR),

pages 1125–1134.

Li, Y., Tan, R., Guo, X., Lu, J., and Brown, M. (2017).

Single image rain streak separation using layer priors.

IEEE Transactions on Image Processing.

Luo, Y., Xu, Y., and Ji, H. (2015). Removing rain from

a single image via discriminative sparse coding. In

Proc. of International Conference on Computer Vision

(ICCV), pages 3397–3405.

Matsui, T., Sakaue, F., and Sato, J. (2014). Raindrop re-

moval by using camera array system. In IEEE 17th In-

ternational Conference on Intelligent Transportation

Systems (ITSC), pages 2249–2250. IEEE.

N., B. and N., L. (2008). Using the shape characteristics of

rain to identify and remove rain from video. In Proc.

of Joint International Workshops on Statistical Tech-

niques in Pattern Recognition and Structural and Syn-

tactic Pattern Recognition, volume LNCS5342, pages

pp 451–458.

Nomoto, K., Sakaue, F., and Sato, J. (2011). Raindrop com-

plement based on epipolar geometry and spatiotem-

poral patches. In Proc. of International Conference

on Computer Vision Theory and Applications, pages

175–180.

Qian, R., Tan, R. T., Yang, W., Su, J., and Liu, J. (2018).

Attentive generative adversarial network for raindrop

removal from a single image. In The IEEE Conference

on Computer Vision and Pattern Recognition (CVPR).

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. In International Conference on Medical

image computing and computer-assisted intervention,

pages 234–241. Springer.

Santhaseelan, V. and Asari, V. (2014). Utilizing local phase

information to remove rain from video. International

Journal of Computer Vision, pages 1–19.

Yamashita, A., Fukuchi, I., and Kaneko, T. (2009). Noises

removal from image sequences acquired with moving

camera by estimating camera motion from spatiotem-

poral information. In Proc. of International Confer-

ence on Intelligent Robots and Systems, pages 3794–

3801.

Recovering Raindrop Removal Images under Heavy Rain

135

Yamashita, A., Tanaka, Y., and Kaneko, T. (2005). Re-

moval of adherent waterdrops from images acquired

with stereo camera. In Proc. of International Confer-

ence on Intelligent Robots and Systems, pages 400–

405.

Yang, W., Tan, R., Feng, J., Liu, J., Guo, Z., and Yan, S.

(2017). Deep joint rain detection and removal from

a single image. In Proc. of the IEEE Conference on

Computer Vision and Pattern Recognition.

You, S., Tan, R., Kawakami, R., Mukaigawa, Y., and

Ikeuchi, K. (2016). Adherent raindrop model-

ing, detectionand removal in video. IEEE transac-

tions on pattern analysis and machine intelligence,

38(9):1721–1733.

Zhang, R., Isola, P., Efros, A. A., Shechtman, E., and Wang,

O. (2018). The unreasonable effectiveness of deep

features as a perceptual metric. In Proc. of the IEEE

Conference on Computer Vision and Pattern Recogni-

tion (CVPR).

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

136