Sparsity-based Time-Frequency Analysis for

Automatic Radar Waveform Recognition

Shuimei Zhang, Ammar Ahmed, and Yimin D. Zhang
Department of Electrical and Computer Engineering
Temple University, Philadelphia, PA 19122, USA

Abstract—In this paper, we develop a novel pre-processing centration. The Wigner-Ville distribution (WVD) is commonly
algorithm to achieve effective signal denoising for improved referred to as the prototype bilinear TF distribution. However,
recognition of noisy radar signals. The algorithm is considered in the bilinear nature of the WVD renders cross-terms to appear
the instantaneous autocorrelation function domain in which time
or lag slices are converted to a Hankel matrix, and an atomic midway between true signal components in the case of non-
norm-based method is applied to mitigate the impacts of noise. linear or multi-component signals. Such cross-terms prohibit
Cross-terms are suppressed by using a time-frequency kernel, accurate analysis and interpretation of the signal instantaneous
such as the Choi-Williams distribution, and a sparsity-based frequency (IF) signatures [7]–[10]. In this case, TF kernels
reconstruction technique is utilized to obtain a high-resolution are used to mitigate the effects of cross-terms so as to readily
time-frequency distribution of the radar waveforms. Simulation
results verify the effectiveness of the proposed method. The determine the modulation type and estimate the modulation
proposed denoising algorithm for radar waveform recognition parameters. Among the many available TF kernels, the Choi-
enables a substantial increase of the overall successful recognition Williams distribution (CWD), which is based on exponential
rate from 90.24% to 97.76%. weighting in both time and lag domains, is considered as a
Index Terms—Radar waveform recognition, non-stationary favorable choice because of its simplicity and robustness to
signal, time-frequency analysis, sparse reconstruction, atomic
norm. different signal parameters [5], [6], [11].
Existing methods reported in the open literature, however,
do not provide a high-resolution TFR and render poor TFR
I. I NTRODUCTION
when the LPI radar signals are noisy. There are limited
Automatic radar waveform recognition plays an important works reporting signal denoising. In [6], the authors applied
role in electronic warfare, such as radar emitter identification a designed filter to smooth the TFR, where each point on the
and threat detection. This becomes even more critical when TFR is updated by averaging the points within the coverage of
radars adopt low probability of intercept (LPI) radar wave- a square. In [12], image morphology algorithm was utilized for
forms that differ to the traditional linear frequency modulated noise removing after converting the TFR into a binary image.
(LFM) waveforms for pulse compression [1]. Commonly In [3], signal denoising was achieved by a nonlinear wavelet
used LPI radar waveforms include those based on nonlinear transformation technique. However, the denoising capability
frequency modulation (FM) and polyphase code modulation of these methods are limited.
[2]. Examples of polyphase codes include the Frank, P1, P2, In this paper, we propose a novel pre-processing algo-
P3, and P4 codes [3], [4]. The objective of radar waveform rithm for feature enhancement of noisy radar waveforms. The
recognition is to detect and classify the received radar signals algorithm is considered in the instantaneous autocorrelation
based on the pulse compression waveform [5]. function (IAF) domain, which is related to the TFR by a one-
A number of methods have been proposed to recognize dimensional (1-D) Fourier transform with respect to lag for
radar signals with different kinds of modulations. Feature each time slice or the ambiguity function (AF) with respect
extraction and classifier design are two key procedures for to time for each lag slice. In the IAF domain, each time- or
radar waveform recognition [6]. Selection of an effective lag-domain slice is converted to a Hankel matrix to denoise
feature extraction algorithm is important to achieve a desirable the IAF entries via the atomic norm-based approach. Then,
waveform classification performance. Because of the sparsity CWD is utilized to mitigate the cross-terms and further reduce
of many non-stationary signals, including the above mentioned the residual effect of noise. Finally, sparse TFR reconstruction
LPI radar waveform signals, in the time-frequency (TF) do- techniques are used to further improve the sparsity of the
main, this domain is effective for signal characterization and yielding TF estimate. To our best knowledge, the sparsity of
feature extraction. The objective of this paper is to provide an the FM and polyphase code LPI radar signals in the TF domain
enhanced TF representation (TFR) of noisy LPI radar signals has not been considered for radar waveform recognition. The
to enable improved feature extraction. proposed method can be easily of joint use with any other
The ability to recognize the time-varying frequency vari- pre-processing techniques.
ations enables classification of different modulations [6]. In Notations : Lower-case (upper-case) bold characters are
∗ T H
order to achieve this, bilinear TF analysis is widely adopted used to denote vectors (matrices). (·) , (·) and (·) denote
for radar waveform recognition due to its high energy con- the complex conjugation, transpose and the Hermitian trans-

978-1-7281-6813-5/20/$31.00 ©2020 IEEE 548

Authorized licensed use limited to: Ozyegin Universitesi. Downloaded on January 26,2022 at 14:58:10 UTC from IEEE Xplore. Restrictions apply.
pose, respectively. Fx (·) and Fx−1 (·) represent the discrete D. Ambiguity Function
Fourier transform (DFT) and inverse DFT (IDFT) with respect The AF is obtained by applying 1-D DFT to the IAF
to x, respectively. Y = H(x, p) converts vector x to Hankel Rxx (t, τ ) with respect to t, expressed as
matrix Y with pencil parameter p, whereas x = (Y ) defines 
the inverse operation of Hankel matrix conversion. · denotes Axx (θ, τ ) = Ft [Rxx (t, τ )] = Rxx (t, τ )e−j2πθt dt, (8)
the ceiling function. t

where θ is the frequency shift, also known as Doppler. Note

II. S IGNAL M ODEL AND T IME -F REQUENCY that, unlike WVD, the AF entries are in general complex since
R EPRESENTATION the IAF is Hermitian symmetric only with respect to τ but not
A. Signal Model with respect to t.

Consider a discrete-time signal, x(t), t = 1, ..., T , III. P ROPOSED M ETHOD

A. Signal Denoising via Atomic Norm
x(t) = s(t) + n(t) = a(t)ejψ(t) + n(t), (1)
 At time instant t, we assume P frequency components
where j = (−1) is the imaginary unit, s(t) is the transmitted are present, i.e., there exist P nonzero entries in the TFR
signal with the phase law ψ(t), and n(t) is the additive white corressponding to the specific value of t. From (7), we know
Gaussian noise. that the IAF is the 1-D IDFT of the WVD with respect to
The instantaneous frequency of x(t) is expressed as: frequency f . Therefore, the IAF slice y t at time instant t can
be expressed as
1 dψ(t)
f (t) = . (2) P

2π dt
y t = Ff−1 (wt ) = cp ej2πfp τ , (9)
B. Instantaneous Auto-correlation Function p=1

The IAF of the transmitted signal s(t) and the received where cp denotes the complex amplitude of the p-th signal
signal x(t) are respectively defined as component, fp is the corresponding signal frequency, and wt
 τ ∗ τ denotes the TF slice along the f dimension at time instant t.
Rss (t, τ ) = s t + s t− , (3) We solve this signal denoising problem in the context of
2 2
Hankel matrix completion [13], [14] by forming the following
and  τ ∗ τ Hankel matrix from y t :
Rxx (t, τ ) = x t + x t− , (4) ⎡ t ⎤
2 2 y1 y2t ··· yqt2
⎢ y2t y3t · · · yqt2 +1 ⎥
where τ is the time lag. Denote vector y t as an IAF slice of ⎢ ⎥
H ty = H(y t , q1 ) = ⎢ . . .. .. ⎥ , (10)
the transmitted signal s(t) that contains all IAF entries along ⎣ .. .. . . ⎦
the τ dimension corresponding to time t. The IAF slice y t is yqt1 yqt1 +1 ··· t
yQ
Hermitian symmetric about τI = 0, where I is the index of
the center element in τ . In other words, y t has the following where q1 is the pencil parameter, which is usually set to
property: Q/2, and q2 = Q − q1 + 1.
 t H  t  The same P frequency components are shared among all
t T
yI−1 , · · · , y2t , y1t = yI+1 t
, · · · , yQ−1 , yQ , (5) the columns in H ty . The atom set for H ty can be expressed
as
where τ = [τ1 , · · · , τQ ]T denotes the lag vector, and Q is the A = {A(f, φ)|f ∈ [0, 1), φ2 = 1} , (11)
cardinality of τ .
We similarly define z t as the IAF slice of the received where A(f, φ) is an atom representing H ty and is formulated
signal x(t) along the τ dimension. Then, z t is also Hermitian as
symmetric, i.e., A(f, φ) = a(f )φH , (12)
 t H  t 
t T a(f ) = ej2πf τ ∈ Cq1 , f ∈ [0, 1), φ ∈ Cq2 with φ2 = 1.
zI−1 , · · · , z2t , z1t = zI+1 t
, · · · , zQ−1 , zQ . (6)
Note that atomic norm approach avoids the off-grid issue by
C. Wigner-Ville Distribution exploiting the continuity of the frequency set.
The atomic norm of H ty can be obtained by solving the
The DFT of the IAF Rxx (t, τ ) with respect to τ is the well- following problem [15]–[17]:
known WVD, i.e.,  
 H ty A = inf β > 0 : H ty ∈ βconv(A)
 
Wxx (t, f ) = Fτ [Rxx (t, τ )] = Rxx (t, τ )e−j2πf τ dτ. (7)   (13)
t
τ = inf |cp | : H y = |cp |a(fp , φp ) ,
Because Rxx (t, τ ) is Hermitian symmetric, the WVD of the p p

observed data is real-valued. where conv(A) is the convex hull of A.

549

Authorized licensed use limited to: Ozyegin Universitesi. Downloaded on January 26,2022 at 14:58:10 UTC from IEEE Xplore. Restrictions apply.
 B. Choli-Williams Distribution
The analysis and interpretation of the signal IF signatures
become cumbersome in the presence of cross-terms as a result
$ORQJ
of the bilinear WVD. A number of TF kernels are developed
GLUHFWLRQ to suppress cross-terms while preserving auto-terms. In this
$ORQJGLUHFWLRQ
paper, we use the CWD, which is defined for a continuous-
time signal x(t) as [20]
 ∞ ∞  
1 (t − μ)2
WCW =  exp −
Fig. 1. The noisy IAF for atomic norm-based processing.
−∞ −∞ 4πτ 2 /σ 4τ 2 /σ (18)
 τ   τ 
∗
·x t+ x t− exp(−jωτ )dtdτ,
2 2
Define H tz = H(z t , q1 ) as the Hankle matrix obtained from
the noisy IAF slice at time instant t, the signal denoising where σ > 0 is a scaling factor that controls the attenuation.
problem can be expressed as The CWD uses an exponential kernel function
 
Φ(θ, τ ) = exp −θ2 τ 2 /σ . (19)
t 1
Ĥ y = arg min H tz − H ty 2F + ηH ty A , (14)
Hy 2
t C. Obtaining TF distribution Using Sparse Reconstruction
FM and polyphase code LPI radar signals are sparsely
where η is a regularization parameter trading off between the
presented in the TF domain. To perform sparsity-based TFR
fidelity of the Hankel matrix fitting and the atomic norm.
reconstruction, we utilize the 1-D IDFT relationship between
Problem (14) can be reformulated as
the TF slice wt and the IAF slice y t at time instant t, i.e.,
1 η y t = Gf w t , ∀t,
min H tz − H ty 2F + (Tr(T (u)) + Tr(W )) (20)
H ty ,u,W 2 2
 t  (15) where Gf is a matrix performing the 1-D IDFT with respect
T (u) H y
s.t.  0, to f . As such, the sparse reconstruction problem for (20) is
(H ty )H W
described as
where Tr(·) denotes the trace operator and T (u) denotes a min ||wt ||1 s.t. y t − Gf wt = 0, ∀t. (21)
Hermitian Toeplitz matrix with u as its first column. Problem
(15) can be effectively solved via the alternating direction Many compressive sensing techniques can be used to solve
method of multipliers (ADMM) approach [18] as outlined in (21), such as the orthogonal matching pursuit (OMP) [21] and
t Bayesian sparse learning techniques [22], [23]. In this paper,
[19]. Based on the obtained Ĥ y , we can estimate the denoised
OMP is applied at each time instant because OMP allows us
IAF slice as
t to specify the sparsity at each time instant.
ŷ t = (Ĥ y ). (16)
IV. S IMULATION R ESULTS
For the underlying problem, the Hermitian symmetry prop- As an example of the LPI radar waveforms, we consider
erty of the IAF can be used to reduce the computational the P2 code as an example. P2 code waveforms have a stair-
complexity. For each time instant t, ŷ t have a Hermitian wise frequency signature and thus are difficult to be accurately
symmetric structure, given as represented in the TF domain as compared to other waveforms
 whose frequency changes straightly with the time, such as
 t ∗  ∗ T
ŷ t = ŷ1t , · · · , ŷI−1
t
, ŷIt , ŷI−1 , · · · , ŷ1t . (17) the P3 and P4 code waveforms. Therefore, preserving the
stair-wise frequency signature is critical to distinguish the
In practice, x(t) is of a finite length, which leads to a diamond underlying waveform from those whose frequency signature
shape with a time-varying width of Q = T − |T + 1 − 2t| in varies linearly with time. The phase of P2 code signal is varied
the IAF due to zero-padding. according to [1]:
When the signals are processing in batch, the denoising π
φi,j = − [2i − 1 − M ][2j − 1 − M ], (22)
procedure becomes unreliable when Q is small because of 2M
zero-padding. To avoid this issue, we denoise the entries of where i, j = 1, 2, · · · , M and M is the order. In our simula-
z t by utilizing both time and lag slices as proposed in [14]. tions, the carrier frequency is 1/4 of the sampling frequency.
As shown in Fig. 1, the area for 0.25T ≤ t ≤ 0.75T are first The cycles per phase code is 1 and the number of phase M
recovered by utilizing the time slice along the lag domain as is set as 6.
depicted above. The similar 1-D DFT relationship between the Fig. 2 shows the real-part of the P2 code radar waveform
IAF and the AF as well as the sparsity of the AF are exploited and the corresponding IAF magnitude, WVD, and CWD. No
to denoise the entries for 0 ≤ τ ≤ 0.25T by utilizing the lag noise is considered in these results. For the single-component
slices. signal with a constant amplitude, the IAF has a constant

550

Authorized licensed use limited to: Ozyegin Universitesi. Downloaded on January 26,2022 at 14:58:10 UTC from IEEE Xplore. Restrictions apply.
 0.5
waveform - real part 50 50
0.5 0.4

0.3
0 0 0

f
0.2
-0.5
0.1
-50 -50
0
20 40 60 80 100 120 20 40 60 80 100 120 20 40 60 80 100 120 20 40 60 80 100 120
t t t t
(a) (b) (a) (b)
0.5 0.5 0.5

0.4 0.4 0.4

0.3 0.3 0.3

f

f
0.2 0.2 0.2

0.1 0.1 0.1

0 0 0
20 40 60 80 100 120 20 40 60 80 100 120 20 40 60 80 100 120
t t t
(c) (d) (c)

Fig. 2. P2 code signal without the noise. (a) Real-part waveform. (b) IAF. Fig. 4. P2 code signal in the noisy environment (SNR = 0 dB). (a) IAF
(c) WVD. (d) CWD. denoised via atomic norm. (b) denoised CWD. (c) proposed TFR (atomic
norm + CWD + OMP).
waveform - real part

2
50

1
TABLE I
L IST OF THE PARAMETERS OF SIMULATED SIGNALS
0 0

Signal Waveforms Parameter Uniform Ranges

-1

-50
Sampling rate (fs ) U(1)
-2
20 40 60 80 100 120 20 40 60 80 100 120
Carrier frequency (fc ) U(1/8, 1/4)
P2 Cycles per phase (cpp) [1, 5]
t t
Frequency steps (m) [4, 8]
(a) (b)
0.5 0.5 Carrier frequency (fc ) U(1/8, 1/4)
P3, P4 Cycles per phase (cpp) [1, 5]
0.4 0.4 Frequency steps (m) [32, 70]
0.3 0.3 Initial frequency (f0 ) U(1/16, 1/8)
LFM
Bandwidth (Δf ) U(1/16, 1/8)
f

0.2 0.2
Number change [3, 6]
Costas codes
0.1 0.1 Fundamental frequency (fmin ) U(1/24, 1/20)
0 0
20 40 60 80 100 120 20 40 60 80 100 120
t t
(c) (d) recognize the auto-term TF distributions.
Fig. 4 depicts the performance of the proposed method in
Fig. 3. P2 code signal in the noisy environment (SNR = 0 dB). (a) Real-part
waveform. (b) IAF. (c) WVD. (d) CWD. the same noisy environment as in Fig. 3. Fig. 4(a) depicts the
recovered IAF via the atomic norm-based approach, which
performs denoising on the IAF entries. We notice that the
magnitude within the diamond-shaped region. We observe that recovered IAF has a much smaller number of high values
the WVD does not provide a clear signal TF signature due to than that in the noisy IAF depicted in Fig. 3(b) and is close
the cross-term effects even though no noise is present. The to the noise-free IAF as shown in Fig. 2(b). The CWD
CWD depicted in Fig. 2(d) shows a clear TF signature, while obtained from the recovered IAF is shown in Fig. 4(b). In
the TF resolution is poor. comparison to Fig. 3(d), the denoised CWD provides a much
In Fig. 3, we show the results with added noise, where the more smooth TF signature. Fig. 4(c) shows the final TFR
input signal-to-noise ratio (SNR) is 0 dB. As a result, the obtained through sparse reconstruction using OMP. It is clear
waveform magnitude varies randomly, as shown in Fig. 3(a). that the proposed method not only mitigates the effects of
The IAF becomes noisy and the magnitude does not maintain a the noise, but also provides a high-resolution TFR, thereby
constant value. Due to the existence of noise, aliases appear in significantly improving the feature representation of the radar
the WVD compared to Fig. 2(d). On the other hand, the CWD signals.
in Fig. 3(c) is distorted heavily by the noise. It is difficult to To verify the effectiveness of the proposed method in wave-

551

Authorized licensed use limited to: Ozyegin Universitesi. Downloaded on January 26,2022 at 14:58:10 UTC from IEEE Xplore. Restrictions apply.
 to 0 dB. The size of the input TFR images is 128 × 128. Two
Costas 0.944 0.004 0.000 0.016 0.036 classifiers are trained, one using the traditional CWD results
and the other one using the proposed atomic norm-based pre-
LFM 0.000 1.000 0.000 0.000 0.000
processing method in the sparse reconstruction framework
(atomic norm + CWD + OMP). Both classifiers share the
True label

0.012 0.000 0.960 0.004 0.024

same structure and hyper-parameters, which are depicted in
P2
Table II. ReLU is chosen as the activation function for each
convolution layer. Keras with Tensorflow backend is adopted
P3 0.020 0.000 0.008 0.792 0.180
for CNN implementation.
The classification results obtained using the two TFRs
0.032 0.000 0.040 0.112 0.816
P4
are shown in Fig. 5. The average accuracy of the testing
classification is 90.24% for traditional CWD-based TFRs and
s

4
97.76% for the TFRs obtained from the proposed method. In
ta

P
F
L
s
o
C

Predicted label
particular, the recognition rate between the challenging pair of
P3 and P4 has been greatly improved by the proposed atomic
(a) norm-based pre-processed method.

Costas 1.000 0.000 0.000 0.000 0.000 V. C ONCLUSION

In this paper, we proposed a novel pre-processing algorithm
LFM 0.000 1.000 0.000 0.000 0.000
for the recognition of noisy LPI radar waveforms. We exploit
True label

the atomic norm-based denoising operation in instantaneous

P2 0.004 0.000 0.960 0.016 0.020
autocorrelation function domain followed by cross-term sup-
pression using CWD. Sparsity-base reconstruction technique
P3 0.000 0.000 0.004 0.960 0.036
is employed to obtain a high-resolution TF distribution. The
effectiveness of the proposed approaches is evidently demon-
P4 0.004 0.000 0.012 0.016 0.968
strated using simulation results processed for the P2 coded
radar signal. Furthermore, the classification result using the
s

CNN demonstrates that the proposed method enables an

ta

P
F
L
s
o
C

Predicted label
improvement of the overall recognition accuracy from 90.24%
to 97.76%.
(b)

Fig. 5. Normalized confusion matrix for classification performance compar- R EFERENCES

ison. (a) CWD-based TFRs; (b) Proposed TFRs (atomic norm + CWD +
OMP) [1] P. E. Pace, Detecting and Classifying Low Probability of
Intercept Radar, 2nd ed. Norwood, MA, USA: Artech
TABLE II House, 2009.
L IST OF THE PARAMETERS OF THE CNN STRUCTURE [2] M. I. Skolnik, Introduction to Radar Systems, 2nd ed.
Layer (type) Parameter Setting New York, NY, USA: McGraw-Hill, 1981.
Batch normalization – [3] T. R. Kishore and K. D. Rao, “Automatic intrapulse mod-
Convolution 1 7 kernels, size 5 × 5, stride 1 ulation classification of advanced LPI radar waveforms,”
Maximum pooling 1 size 2 × 2, stride 2
Convolution 2 4 kernels, size 5 × 5, stride 1
IEEE Trans. Aerosp. Electron. Syst., vol. 53, no. 2, pp.
Maximum pooling 2 size 2 × 2, stride 2 901–914, April 2017.
Convolution 3 4 kernels, size 5 × 5, stride 1 [4] C. Shi, F. Wang, M. Sellathurai, and J. Zhou, “Low prob-
Maximum pooling 3 size 2 × 2, stride 2
Flatten layer – ability of intercept-based distributed MIMO radar wave-
Fully connected layer 1 100 neurons form design against barrage jamming in signal-dependent
Fully connected layer 2 5 neurons clutter and coloured noise,” IET Signal Process., vol. 13,
no. 4, pp. 415–423, May 2019.
[5] J. Lunden and V. Koivunen, “Automatic Radar Waveform
form classification, we process the results using a convolution Recognition,” IEEE J. Sel. Topics Signal Process., vol. 1,
neural network (CNN) [24], [25]. Five types of waveforms, no. 1, pp. 124–136, June 2007.
including Costas, LFM, P2, P3 and P4, are considered for [6] C. Wang, J. Wang, and X. Zhang, “Automatic radar
classification. The detailed parameter settings are shown in waveform recognition based on time-frequency analysis
Table I. U(·) denotes a uniform distribution of the frequency. and convolutional neural network,” in Proc. IEEE Int.
1, 000 samples are generated for each waveform class, from Conf. Acoust., Speech Signal Process., Mar. 2017, pp.
which 75% of the samples are utilized for training and the 2437–2441.
remaining 25% are utilized for testing. The input SNR is set

552

Authorized licensed use limited to: Ozyegin Universitesi. Downloaded on January 26,2022 at 14:58:10 UTC from IEEE Xplore. Restrictions apply.
 [7] B. Boashash (ed.), Time-Frequency Signal Analysis and [16] G. Tang, B. Bhaskar, P. Shah, and B. Recht, “Compressed
Processing: A Comprehensive Reference, 2nd ed. Aca- sensing off the grid,” IEEE Trans. Inf. Theory, vol. 59, no.
demic Press, 2015. 11, pp. 7465–7490, Nov. 2013.
[8] Y. D. Zhang, “Resilient quadratic time-frequency distribu- [17] Z. Yang and L. Xie, “On gridless sparse methods for line
tion for FM signals with gapped missing data,” in Proc. spectral estimation from complete and incomplete data,”
IEEE Radar Conf., Seattle, WA, May 2017, pp. 1765– IEEE Trans. Signal Process., vol. 63, no. 12, pp. 3139–
1769. 3153, June 2015.
[9] V. S. Amin, Y. D. Zhang, and B. Himed, “Improved [18] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eck-
instantaneous frequency estimation of multi-component stein, “Distributed optimization and statistical learning via
FM signals,” in Proc. IEEE Radar Conf., Boston, MA, the alternating direction method of multipliers,” Found.
April 2019. Trends Mach. Learn., vol. 3, no. 1, pp. 1–122, Jan. 2011.
[10] V. S. Amin, Y. D. Zhang, and B. Himed, “Sparsity-based [19] Y. Li and Y. Chi, “Off-the-grid line spectrum denoising
time-frequency representation of FM signals with burst and estimation with multiple measurement vectors,” ArXiv
missing samples,” Signal Process., vol. 155, pp. 25–43, Preprint ArXiv:1408.2242, 2014.
Feb. 2019. [20] H. I. Choi and W. J. Williams, “Improved time-frequency
[11] Q. Guo, X. Yu, and G. Ruan, “LPI radar waveform representation of multicomponent signals using exponen-
recognition based on deep convolutional neural network tial kernels,” IEEE Trans. Acoust., Speech, Signal Pro-
transfer learning,” Symmetry, vol. 11, no. 4, pp. 540, April cess., vol. 37, no. 6, pp. 862–871, June 1989.
2019. [21] J. A. Tropp and A. C. Gilbert, “Signal recovery from
[12] M. Zhang, M. Diao, L. Gao, and L. Liu “Neural networks random measurements via orthogonal matching pursuit,”
for radar waveform recognition,” Symmetry, vol. 9, no. 5, IEEE Trans. Inf. Theory, vol. 53, no. 12, pp. 4655-4666,
pp. 75, May 2017. 2007.
[13] Y. Chen and Y. Chi, “Harnessing structures in big data via [22] S. Ji, Y. Xue, and L. Carin, “Bayesian compressive
guaranteed low-rank matrix estimation: recent theory and sensing,” IEEE Trans. Signal Process., vol. 56, no. 6, pp.
fast algorithms via convex and nonconvex optimization,” 2346–2356, June 2008.
IEEE Signal Process. Mag., vol. 35, no. 4, pp. 14–31, [23] Q. Wu, Y. D. Zhang, M. G. Amin, and B. Himed,
July 2018. “Complex multitask Bayesian compressive sensing,” in
[14] S. Zhang and Y. D. Zhang, “Robust time-Frequency anal- Proc. IEEE Int. Conf. Acoust. Speech Signal Process.,
ysis of multiple FM signals with burst missing samples,” Florence, Italy, May 2014.
IEEE Signal Process. Lett., vol. 26, no. 8, pp. 1172–1176, [24] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,”
June 2019. Nature, vol. 521, no. 7553, pp. 436–444, 2015.
[15] V. Chandrasekaran, B. Recht, P. A. Parrilo, and A. S. [25] L. Zhang, F. Yang, Y. D. Zhang, and Y. J. Zhu, “Road
Willsky, “The convex geometry of linear inverse prob- crack detection with deep convolution neural network,” in
lems,” Found. Comput. Math., vol. 12, no. 6, pp. 805–849, Proc. IEEE Int. Conf. Image Process., Phoenix, AZ, Sept.
Dec. 2012. 2016.

553

Authorized licensed use limited to: Ozyegin Universitesi. Downloaded on January 26,2022 at 14:58:10 UTC from IEEE Xplore. Restrictions apply.