See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/343516631

Sea Clutter Suppression for Radar PPI Images Based on SCS-GAN

Article  in  IEEE Geoscience and Remote Sensing Letters · August 2020

DOI: 10.1109/LGRS.2020.3012523

CITATIONS READS
2 626

5 authors, including:

Xiaolong Chen Jian Guan

Naval Aviation University Nankai University
116 PUBLICATIONS   1,330 CITATIONS    110 PUBLICATIONS   1,386 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

FDA-MIMO radar signal processing View project

Sea clutter suppression via high-resolution sparse representation View project

All content following this page was uploaded by Xiaolong Chen on 18 August 2020.

The user has requested enhancement of the downloaded file.

 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1

Sea Clutter Suppression for Radar PPI Images

Based on SCS-GAN
Xiaoqian Mou, Xiaolong Chen , Member, IEEE, Jian Guan, Yunlong Dong, and Ningbo Liu

Abstract— The problem of strong sea clutter, e.g., sea spikes, In recent years, with the development and application of
may bring in low signal-to-clutter ratio (SCR) and cause great machine learning and deep learning [6], new methods for
interference to radar marine target detection. However, the sea clutter suppression have been proposed from the perspective
clutter suppression ability of current algorithms is limited with of image processing [7]. Zhang et al. [9] proposed a denoising
poor generalization under complex marine environment. In this method for image based on the denoising convolutional neural
letter, a novel sea clutter suppression generative adversarial
network (DnCNN) [8] and a blind denoising model based on
network (SCS-GAN) is designed and employed for marine radar
plan-position indicator (PPI) images detection. The SCS-GAN is FFDNet, which greatly improved the level of noise suppression
based on residual networks and attention module, which includes and have become a classic algorithm in the field of image
residual attention generator (RAG) and sea clutter discriminator noise suppression, but for sea clutter that is more complex
(SCD). In order to expand the data sets and improve generaliza- than noise, the above methods cannot achieve the ideal clutter
tion ability, clutter-free data set A, simulated sea clutter data set suppression effect. In general, the current methods by means
B (containing five types of sea clutter distributions), and actual of signal processing and image processing have promoted
sea clutter data set C are constructed by means of simulation the development of the clutter suppression greatly; however,
and acquisition of real radar returns. At last, the parameter, i.e., there are still some problems such as complex models and
clutter suppression ratio (CSR) is designed for evaluating the poor generalization for various backgrounds, which limits the
sea clutter suppression performances of the proposed method
improvement of sea clutter suppression.
and other denoising and clutter suppression methods includ-
ing CBM3D, denoising convolutional neural network (DnCNN), Generative adversarial network (GAN) is a new branch
FFDNet, and Pix2pix. After testing with actual data, it is proved of deep learning. Proposed by Ian Goodfellow in 2014,
that the SCS-GAN has faster clutter removal speed, stronger GAN has attracted widespread attention, which has been used
generalization ability, and at the same time marine targets in in data enhancement, image denoising [10], super-resolution
images are remained completely. reconstruction, and so on. GAN consists of generator and
Index Terms— Generative adversarial network (GAN), radar
discriminator. The generator is used to generate images, and
plan-position indicator (PPI) images, sea clutter radar data sets, the discriminator examines images to determine whether they
sea clutter suppression. are real or fake. At present, some researchers employ GANs
based on Pix2pix for image denoising, and have achieved good
results [11], but there are also deficiencies. On the one hand,
I. I NTRODUCTION the existing GAN-based denoising methods are not generalized
well and have high requirements for the completeness of the

S EA clutter is the radar backscattered echo of sea surface,

and its intensity is affected by many factors such as
sea state, rain and snow, radar transmitting frequency, and
data set. For example, the generator reconstructs the input
image, while the discriminator is used to distinguish whether
the generated image is the target image, which is actually a
polarization mode [1]. Under high sea state, the sea clutter will one-to-one mapping between the input image and the target
exhibit strong amplitudes, e.g., sea spikes, and the signal-to- image. The limited mapping range leads to the limited gen-
clutter ratio (SCR) will degrade seriously, which may result in eralization ability of the model. Therefore, when the test data
serious interference for marine target detection [2]. Most of the and the training data are insufficient and different, the test
traditional sea clutter suppression methods are from the view result will become unsatisfactory, and the actual application
of the signal processing, e.g., covariance matrix estimation and situation cannot guarantee that the data set covers all situations
filtering [3], [4], time domain accumulation, frequency domain and types of application domain images. On the other hand,
cancellation processing [5], and so on. However, in case of when the clutter is strong, the clutter points, clutter blocks,
complex and diverse environment and marine targets, clutter and shadows will exist in the generated declutter images.
modeling and model parameter estimation are difficult result- Therefore, its clutter suppression performance cannot satisfy
ing in poor detection performance. actual requirements and the reconstructed target is incomplete
in the generated images.
Manuscript received June 4, 2020; revised July 5, 2020; accepted In this letter, a novel sea clutter suppression GAN
July 24, 2020. This work was supported in part by the National Natural (SCS-GAN) is proposed and employed for marine radar plan-
Science Foundation of China under Grant U1933135, Grant 61931021, Grant position indicator (PPI) images. It includes residual attention
61871391, and Grant 61871392, in part by the National Defense Science
Foundation under Grant 2102024, and in part by the Key Research and Devel- generator (RAG) and sea clutter discriminator (SCD). The sea
opment Program of Shandong under Grant 2019GSF111004. (Corresponding clutter image is reconstructed into the clutter-free image by
author: Xiaolong Chen.) the RAG, and the SCD is used for adversarial training with
The authors are with the Radar Target Detection Research Group, Naval the RAG to improve RAG’s clutter suppression performance.
Aviation University, Yantai 264001, China (e-mail: [email protected]).
Color versions of one or more of the figures in this letter are available Compared with the existing GAN-based denoising method,
online at http://ieeexplore.ieee.org. it has better clutter suppression and generalization abilities,
Digital Object Identifier 10.1109/LGRS.2020.3012523 and the target can be completely retained after clutter
1545-598X © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: SHENZHEN UNIVERSITY. Downloaded on August 18,2020 at 13:17:13 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

suppression. Section II presents the method and network

structure of SCS-GAN. In Section III, three sea clutter
data sets, simulated and real sea clutter under different sea
states, are used to train and test SCS-GAN. The last section
concludes the letter and presents future directions.

II. S EA C LUTTER S UPPRESSION M ODEL

BASED ON SCS-GAN
A. Structure and Principle of the SCS-GAN Model
GAN is composed of generator and discriminator. The
discriminator divides the input images into two classes (real
or fake) through supervised learning, while the generator is
trained to generate images to fool the discriminator. Through
the confrontation training of the two networks, the perfor-
mance of GAN can be improved. In this letter, the proposed
SCS-GAN consists of RAG and SCD, in which RAG can
suppress the sea clutter in the image to generate clutter-free
image, SCD can discriminate the authenticity of the generated
image and enhance the generator’s ability to suppress sea
clutter through confrontation training. The objective function
is described in the following equations:
Fig. 1. Structure diagram of SCS-GAN model.
G ∗ = arg min L GAN (G) + λg L 1 (G) (1)
G
D ∗ = arg min L GAN (D) + λd L 2 (D) (2)
D

where
 
L GAN (D) = E x,y log(1 − D(x, y))2
  Fig. 2. Overall structure diagram of BARM.
+ E x log(D(x, G(x)))2 (3)
L 1 (G) = E x,y [ y − G(x) 1 ] (4)
L 2 (D) = −E x,y [min(0, D(x, y) − 1)
+ min(0, −1 − D(G(x), y))] (5)
 
L GAN (G) = E x log(1 − D(x, G(x)))2 (6)
and G ∗ is the optimal generator, D ∗ is the optimal discrimi-
nator, arg min D and arg minG are the solution function repre-
sentation of the discriminator loss function with the smallest
value and the minimum value of the generator loss function,
respectively, L GAN (G) and L GAN (D) are the adversarial loss
functions of generator and discriminator, L 1 (G) is a content
loss function reflecting the difference between the image
generated by the generator and the real image, L 2 (G) is the
classification loss function of the discriminator for the image
by the generator and the real image, λg , λd are the adjustment Fig. 3. Specific structure diagram of BARM.
parameter, x is the input image tensor, y is the true image
tensor corresponding to x, E means mathematical expectation, the integrity of the feature information. The attention modules
||y-G(x)||1 is the l1 norm of the real image and the generated can achieve reasonable distribution of model weights through
image. training, and pay attention to focusing on the key information
Fig. 1 shows the structure of RAG and SCD. The core in the image, while ignoring the unimportant information.
module of the entire clutter suppression task in RAG are basic 2) LN-TC: The LN-TC algorithm can extract the key target
attention residual module (BARM) and layer normalization- information in the image and enhance it to ensure that the
transfer and connection (LN-TC), and the core module of the marine target is completely retained while completing the sea
SCD is basic attention residual module (BDM). clutter suppression task. The formula of LN [12] is described
1) BARM: It consists of two parts (see Figs. 2 and 3): in detail as follows, and the formula of TC is (9):
residual network module and attention module. Among them,
1  n
H
the residual network module consists of convolution mod-
ule (CM) and residual connections, the attention modules are μn = x (7)
H i=1 i
composed of channel attention module (CAM) and spatial 

attention module (SAM). The residual network can construct a 1  H
 n 2
new identity map by learning the residual function to improve σn =  x i − μn + ε (8)
the feature extraction efficiency of the neural network and H i=1

Authorized licensed use limited to: SHENZHEN UNIVERSITY. Downloaded on August 18,2020 at 13:17:13 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MOU et al.: SEA CLUTTER SUPPRESSION FOR RADAR PPI IMAGES BASED ON SCS-GAN 3

to collect radar azimuth-range echo data under 0–1 sea

state level on good weather days; ② using navigation
radar to collect diverse radar azimuth-range echo data
under the second, third, and fourth level sea states.
2) Data processing: For ① data, two types of processing
are performed. One of them is to convert the echo
data to images, crop and filter the image, and finally
clutter-free images A can be obtained. The other is to
convert radar returns into images after superimposing
simulated sea clutter, and simulated sea clutter images
B can be obtained after cropping. For ② data, convert
radar returns into images to obtain real sea clutter.
3) Construction of sea clutter data sets: Construct sea
clutter training data sets A and B and test data set C,
of which clutter-free image data set A and simulated sea
clutter image data set B constitute a paired training data
Fig. 4. Algorithm flowchart. set, and real sea clutter test data set C is used as the
actual test data set.
4) Establishment of SCS-GAN model.
5) Model training and optimization. The generator with
the clutter suppression ratio (CSR) higher than the
preset threshold η is selected as the clutter suppression
model (RAG) for the optimal training model.
6) Use marine radar to collect real sea surface echo data.
7) Convert these echo data to actual sea clutter images.
8) Input the actual sea clutter images to the RAG-based
sea clutter suppression model and finally sea clutter
suppression is completed.

III. E XPERIMENTAL R ESULTS AND A NALYSIS

A. Construction of Sea Clutter Image Data Sets

In this letter, three data sets are constructed: data set A, data
Fig. 5. Training and testing for SCS-GAN. set B, and data set C. The image conversion model based on
GAN requires a pair of paired data sets for training, but the
acquisition of clutter-free data in the same scene corresponding
where μn is the mean of the nth layer feature matrix, σn is to the actual sea clutter data is too difficult, so the actual
the standard deviation of the nth layer feature matrix, n is the sea clutter data cannot be used for the training of the model.
nth layer of feature tensor, H is the number of neurons in this Based on the above consideration, radar data of low sea state
layer, xni is the i th input neuron in this layer, ε is a small under good weather are collected, and they are constructed to
constant introduced to prevent σn from being 0, which is set the clutter-free data set A. Simulated sea clutter data set B is
to 10−5 built by superimposing the simulated sea clutter on the low


x −μ sea state data with real marine targets. In order to improve the
Y =γ · f +β (9)
σ algorithm’s generalization ability, five types of sea clutter with
different parameters are simulated in data set B, i.e., Rayleigh
where Y is the new feature tensor output, x is the input distribution, Lognormal distribution, Weibull distribution, K
feature tensor, μ, σ are the collection of μn and σn , γ , β distribution, and alpha stable distribution, so that the GAN can
are the standard deviation and mean extracted from LN, f () fully learn the essential characteristics of sea clutter. The real
is a functional representation of the entire BARM. sea clutter data set C is composed of real sea clutter under
3) BDM: The BDM consists of a convolution layer, batch- different sea states and they are converted into images and
norm (BN), and Leaky Rectified Linear Unit (ReLU). cropped, which is shown in Fig. 6. The detailed content of
sea clutter dataset is shown in Table I.
B. SCS-GAN-Based Sea Clutter Suppression The details of the data collection process are as follows:
for Radar Images the JRC navigation radar antenna transmits electromagnetic
waves to obtain the echo of the civil ship, which is collected
The sea clutter suppression algorithm based on SCS-GAN by the radar signal collection device, and output to the
includes two stages: offline training stage 1)–5) and testing terminal display through signal processing to obtain the radar
stage 6)–8). The flowchart is shown in Fig. 4, and the training images. The model of the JRC navigation radar used is JMA-
and testing process of SCS-GAN is shown in Fig. 5. 5312, the transmission power is 10 kW, and the transmission
1) Data collection: Two types of radar data collection frequency is 9410 ± 30 MHz. The antenna was installed at
experiments were conducted by Japan Radio Com- a height of 80.2 m above sea level with a rotation speed
pany (JRC) navigation radar: ① using navigation radar of 24 rpm.

Authorized licensed use limited to: SHENZHEN UNIVERSITY. Downloaded on August 18,2020 at 13:17:13 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

TABLE I
D ESCRIPTION OF S EA C LUTTER D ATA S ET

Fig. 8. Sea clutter suppression results under different sea states. (a) Before
suppression (from data set C). (b) After suppression.

Fig. 6. PPI image under high sea state before cropping.

Fig. 9. Clutter suppression effect of different algorithms for level 4 sea state.
(a) Sea clutter image. (b) CBM3D. (c) DnCNN-B. (d) FFDNet. (e) Pix2pix.
(f) SCS-GAN.

evaluate the degree of image restoration, IS and FID evaluate

Fig. 7. Clutter suppression of actual images with different epochs. (a) Actual the generated images based on their quality and diversity,
image 100, 200, 450 epoch. (b) Actual image 100, 200, 450 epoch. while there are only marine targets and clutter in the images.
Therefore, they are all not appropriate to evaluate the sea
B. Model Training and Experimental Results clutter suppression. Therefore, based on the original index of
radar sea clutter suppression and image translation of GAN,
The computer configuration is as follows: Intel Core i7-
a new evaluation index, i.e., CSR, is established in this letter.
8750H processor, 16 GB memory, GeForce GTX 1050Ti
For real radar sea clutter images, the SCR of the original
graphics card, 6G video memory. The system environment
image cannot be obtained since we cannot accurately estimate
is Windows and the framework is Pytorch. The experimental
the amplitude of the sea clutter in the images. The core idea
parameter settings are as follows: the gradient descent algo-
of CSR is to reflect the degree of clutter suppression by the
rithm is the Adam, the initial learning rate is set to 0.0002,
ratio of the average power of the removed clutter to the power
and the number of iterations epochs is set to 500.
of the original target. A higher ratio indicates a higher degree
After 500 epochs iterative training, the trained SCS-GAN
of clutter suppression. The value of the CSR is less than 0 and
model is obtained. Use the test data set C (actual data) to test
negative. The larger the value and the closer to 0, the better the
and verify the trained SCS-GAN generator model and get the
clutter suppression effect is. When the CSR of the model test
optimal RAG model in 450th epoch. The PPI image under high
result is higher than the threshold η (which is an adjustable
sea state before cropping is shown in Fig. 6, we cropped a sea
threshold, the value range is −1∼−0.5dB), it can indicate that
clutter image containing a single target and a sea clutter image
the model has achieved a good clutter suppression effect. The
containing multiple targets. The clutter suppression results
formula is as follows:
under high sea states from different epoch are shown in Fig. 7. 
The clutter suppression results under different sea states are CSR = 10 log (C)2 /S12 (10)
shown in Fig. 8. It can be found that the SCS-GAN model has
1 
E F
strong generalization ability under different sea states, and can C = (T (i, j ) − Q(i, j )) (11)
achieve good clutter suppression. E F i=1 j =1

1 
N
C. Analysis and Evaluation of Experimental Results S1 = Rk Rk ∈ max N T (i, j ) (12)
N k=1
In the field of image translation based on generative con-
frontation networks, peak signal-to-noise ratio (PSNR), struc- where C is the average clutter amplitude removed by the
tural similarity (SSIM), inception score (IS), fréchet inception image, S1 is the average amplitude of the target in the original
distance (FID) [13], etc., are the evaluation indicators of image, T, Q are the original image and the image after sea
image. However, PSNR and SSIM were originally used to clutter suppression, T (i , j ), Q(i , j ) are the pixel values of

Authorized licensed use limited to: SHENZHEN UNIVERSITY. Downloaded on August 18,2020 at 13:17:13 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

MOU et al.: SEA CLUTTER SUPPRESSION FOR RADAR PPI IMAGES BASED ON SCS-GAN 5

TABLE II good generalization ability, but also ensure the integrity of the
AVERAGE CSR U NDER D IFFERENT S EA S TATES (U NIT: dB) target, which proves its effectiveness and superiority.
In summary, after testing with actual data and comparison
with existing denoising and clutter algorithms, it is proved
that the SCS-GAN has faster clutter removal speed, stronger
generalization ability, more complete marine targets after
suppression and its clutter suppression effect is several times
TABLE III that of some existing algorithms.
AVERAGE CSR R ESULT OF D ATA S ET C T EST (U NIT: dB)
IV. C ONCLUSION
In this letter, we proposed a method of sea clutter sup-
pression for radar images, and a new sea clutter suppression
GAN model (SCS-GAN) is designed. Experiments based on
actual data under different sea states proved that the SCS-
GAN has stronger clutter suppression performance and better
generalization ability compared with the existing denoising
the original image and the image after decluttering, E, F are and decluttering models, and can well guarantee the integrity
the length and width of the image, max N T (i , j ) is the first of marine targets in the image. In the future, we will try
N pixels with the highest amplitude, Rk is any one of the N to integrate the clutter suppression network and the target
highest amplitude points of the original image, and N is set detection network to build an integrated model of clutter
to 100 in this letter. suppression and target detection.
After the data set C is tested by the SCS-GAN, the test
results are evaluated based on the CSR indicators, which ACKNOWLEDGMENT
is shown in Table II. As can be seen from Table II and The authors would like to thank the anonymous reviewers
Fig. 7, with the increase of iterative epochs, the model’s for valuable comments and suggestions.
clutter suppression ability is constantly improving. After the
training of the first 100 epochs, the model has been optimized, R EFERENCES
though there are still some clutter on the edge of images. [1] X. Chen, W. Huang, C. Zhao, and Y. Tian, “Rain detection from X-band
After optimization of 500 epochs, the SCS-GAN model can marine radar images: A support vector machine-based approach,” IEEE
well suppress the sea clutter while ensuring the integrity of Trans. Geosci. Remote Sens., vol. 58, no. 3, pp. 2115–2123, Mar. 2020.
[2] G. V. Trunk and S. F. George, “Detection of targets in non-Gaussian
the target and can achieve good suppression performance for sea clutter,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-6, no. 5,
different background clutter, i.e., different levels of sea state. pp. 620–628, Sep. 1970.
[3] A. De Maio, L. Pallotta, J. Li, and P. Stoica, “Loading factor estimation
under affine constraints on the covariance eigenvalues with application
to radar target detection,” IEEE Trans. Aerosp. Electron. Syst., vol. 55,
D. Comparison With the Existing Algorithms no. 3, pp. 1269–1283, Jun. 2019.
[4] A. Aubry, A. De Maio, and L. Pallotta, “A geometric approach to
In order to further verify the effectiveness of the proposed covariance matrix estimation and its applications to radar problems,”
algorithm, it is compared with some popular algorithms in IEEE Trans. Signal Process., vol. 66, no. 4, pp. 907–922, Feb. 2018.
terms of sea clutter removal performance and computing [5] X. Yu, X. Chen, Y. Huang, and J. Guan, “Fast detection method
time, which are the classic denoising algorithms BM3D and for low-observable maneuvering target via robust sparse fractional
DnCNN, FFDNet and a kind of GAN used for style transfer, Fourier transform,” IEEE Geosci. Remote Sens. Lett., vol. 17, no. 6,
pp. 978–982, Jun. 2020.
i.e., Pix2pix. The real data set C is used for the test. The [6] N. Sun and X. Chen, “Detection and classification of maritime tar-
comparison between the evaluation results and the time taken get with micro-motion based on CNNs,” J. Radars, vol. 7, no. 5,
for each image to declutter is shown in Table III, and the pp. 565–574, 2018.
comparison of the results is shown in Fig. 9. The CBM3D’s [7] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Color image
denoising via sparse 3D collaborative filtering with grouping constraint
clutter suppression performance decreases with the increase of in luminance-chrominance space,” in Proc. IEEE Int. Conf. Image
sea states, which is unsuitable for high sea state clutter, and Process., vol. 1, Sep. 2007, pp. I-313–I-316.
the algorithm has high complexity. The clutter suppression [8] K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, “Beyond a
performance of the remaining algorithms increases with the Gaussian denoiser: Residual learning of deep CNN for image denoising,”
increase of sea states, and it takes no more than 1 s for IEEE Trans. Image Process., vol. 26, no. 7, pp. 3142–3155, Jul. 2017.
[9] K. Zhang, W. Zuo, and L. Zhang, “FFDNet: Toward a fast and flexible
each image to remove clutter. Among them, the SCS-GAN solution for CNN-based image denoising,” IEEE Trans. Image Process.,
algorithm performs the best clutter suppression capability with vol. 27, no. 9, pp. 4608–4622, Sep. 2018.
the shortest time compared with the other algorithms. [10] J. Zhang, M. Guo, and J. Fan, “A novel generative adversarial net for
It can be seen from Fig. 9 that the CBM3D’s decluttering calligraphic tablet images denoising,” Multimedia Tools Appl., vol. 79,
nos. 1–2, pp. 119–140, Jan. 2020.
effect is not obvious, and the targets in the image are also sup- [11] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation
pressed while suppressing clutter; the DnCNN-B introduces with conditional adversarial networks,” in Proc. IEEE Conf. Comput. Vis.
new type of noise, and the integrity of the target is destroyed; Pattern Recognit. (CVPR), Jul. 2017, pp. 5967–5976.
the FFDNet caused the contour deformation and blur of the [12] J. Lei Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” 2016,
target in the image; the Pix2pix caused the partial deletion of arXiv:1607.06450. [Online]. Available: http://arxiv.org/abs/1607.06450
[13] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli,
the target and the presence of clutter blocks and black spots “Image quality assessment: From error visibility to structural sim-
on the edge of the image; while the SCS-GAN can not only ilarity,” IEEE Trans. Image Process., vol. 13, no. 4, pp. 600–612,
effectively suppress sea clutter under different sea states with Apr. 2004.

Authorized licensed use limited to: SHENZHEN UNIVERSITY. Downloaded on August 18,2020 at 13:17:13 UTC from IEEE Xplore. Restrictions apply.
View publication stats