Int. J. Electron. Commun.

(AEÜ) 82 (2017) 321–326

Contents lists available at ScienceDirect

Int. J. Electron. Commun. (AEÜ)

journal homepage: www.elsevier.com/locate/aeue

Regular paper

Measurement matrix optimization based on incoherent unit norm tight MARK

frame
⁎
Rahim Entezari , Alijabbar Rashidi
Electrical and Electronic Engineering University Complex (EEEUC), Malek-e-Ashtar University of Technology (MUT), Tehran, Iran

A R T I C L E I N F O A B S T R A C T

Keywords: This paper considers the problem of measurement matrix optimization for compressed sensing (CS) in which the
Measurement matrix optimization dictionary is assumed to be given, such that it leads to an effective sensing matrix. Due to important properties of
Incoherent frame equiangular tight frames (ETFs) to achieve Welch bound equality, the measurement matrix optimization based
Unit norm tight frame on ETF has received considerable attention and many algorithms have been proposed for this aim. These
Partial Fourier matrix
methods produce sensing matrix with low mutual coherence based on initializing the measurement matrix with
Compressed sensing (CS)
random Gaussian ensembles. This paper, use incoherent unit norm tight frame (UNTF) as an important frame
with the aim of low mutual coherence and proposes a new method to construction a measurement matrix of any
dimension while measurement matrix initialized by partial Fourier matrix. Simulation results show that the
obtained measurement matrix effectively reduces the mutual coherence of sensing matrix and has a fast con-
vergence to Welch bound compared with other methods.

1. Introduction matrices, have a nice universal property for any basis matrix and re-
sulting random sensing matrices satisfy the RIP with overwhelming
Compressed sensing (CS) is a technique in signal processing to probability and m = O(k log(n/ k ) ) measurements [2]. Now, it is pos-
sample sparse or compressible signals at a sub-Nyquist rate [1,2]. In sible to find a unique solution by the following optimization problem:
recent years, CS has been widely studied in areas of radar imaging [3], s = min‖x|0 , s.t. y = Fs. (2)
image and video processing [4], wireless communication [5], coding
theory [6] and etc. Let a signal x of length n has a sparse representation where ‖. ‖0 is the ℓ0 norm. The ℓ0 norm is non-convex and finding a
in a known domain Ψn × n: xn × 1 = Ψn × n·sn × 1; where Ψ is called basis solution for (2) is NP-hard [8]. However, there are a variety of re-
(dictionary) matrix. In this case, s is a k-sparse signal; i.e., it has at most construction algorithms to solve this problem such as Greedy, optimi-
k nonzero elements. By making m reduced-dimensional measurements zation and statistical methods.
(k < m < n ), we have ym × 1 = Φm × n ·xn × 1, where Φm × n is referred to as the The construction of the measurement matrix with respect to dic-
measurement matrix. Finally the CS equation can be formed as: tionary matrix is one of the main challenge in compressed sensing and
ym × 1 = Φm × n ·Ψn × n·sn × 1 there are many research to show the effectiveness of non-random
(1)
measurement matrices to random measurement matrices [9–12]. The
The product F ≜ Φ. Ψ is referred to as sensing matrix. The Eq. (1) is an random matrices generated by various distributions, have some draw-
under-determined equation and generally has an infinite set of solu- backs [13]:
tions. However, under certain conditions on sensing matrix, it is pos-
sible to find a unique solution. The best condition on sensing matrix in
presence of noise for stable recovery of k-sparse signals is the restricted
• Require high storage space due to store and process each realization
of random matrices.
isometry property (RIP) [7] and is a sufficient condition on sensing
matrices to recovery guarantee. However, it is difficult to prove the
• Satisfy the RIP with overwhelming probability, which is no guar-
antee to recover all sparse signals.
sensing matrix satisfies the RIP, because require a combinatorial search
over all sub-matrices and therefore certifying RIP for an arbitrary ma-
• High computational complexity to reconstruction when compared to
non-random sensing matrices.
trix is NP-hard problem. However, it has been shown that many random
measurement matrices such as Gaussian, Bernoulli and partial Fourier A way to start designing a good measurement matrix is frame

⁎
Corresponding author.
E-mail addresses: [email protected] (R. Entezari), [email protected] (A. Rashidi).

http://dx.doi.org/10.1016/j.aeue.2017.09.015
Received 12 June 2017; Accepted 21 September 2017
1434-8411/ © 2017 Elsevier GmbH. All rights reserved.
R. Entezari, A. Rashidi Int. J. Electron. Commun. (AEÜ) 82 (2017) 321–326

theory. The frame theory, has many different types of application in- mutual coherence satisfy the RIP condition. However, the lower bound
cluding, compressed sensing [14], sparse approximation [15], quantum of μ(F ) is known as Welch bound (WB):
information theory [16], coding and communications [17]. Therefore, n−m
in this paper, we focus on measurement matrix optimization based on μWB =
m(n−1) (5)
frame theory while reconstruction process is performed by SL0 [18]
algorithm,1 because it is fast and robustness to noise. This bound is achievable for ETF [17]. For this reason, ETFs are
This paper is organized as follows: Section 2 describes some basic sometimes called maximum Welch-Bound-Equality (WBE) sequences.
definitions related to frames along with a summary of the previous In frame design theory, an important factor is frame dimension, so-
works on incoherent frames design. Section 3 describes the proposed called redundancy. Unfortunately, the ETFs are not existing for all frame
m(m + 1)
method for designing the measurement matrix based on incoherent dimensions and only exist for n ⩽ 2
in real case and n ⩽ m2 in
UNTF. Simulation and numerical results are presented in Section 4 and complex case [22]. There are only a few general construction methods
finally, the conclusion is given in Section 5. of ETFs and the ETF design is very difficult [23]. Therefore, the frame
design with low mutual coherence was considered. These frames are
2. Frame theory background referred to as incoherent frames. For the first time, the idea of the
measurement matrix optimization with respect to dictionary matrix,
In this section, the concept of frames as a key point of the best such that it leads to an effective sensing matrix was introduced by Elad
measurement matrix optimization is introduced. A frame is set of n in [12]. Their algorithm tries to minimize the t-averaged mutual co-
vectors {fi }n in Hilbert space m , if ∀ f ∈ m : herence between Φ and Ψ . However, it needs many iterations to achieve
i=1
n better performance. From formulation of Gram matrix and mutual co-
2 2 2 herence, it was concluded that the zero off-diagonal elements of the
α f 2
⩽ ∑ f ,fi ⩽β f 2
i=1 (3) Gram matrix is desired. Based on this idea, Duarte-Carvajalino and
Sapiro [24] and abolghasemi et al. [25], making the Gram matrix as
where α and β are lower and upper bound of frames respectively, with
closely as possible to identity matrix by following minimization pro-
0 < α ⩽ β < ∞. The frame synthesis operator, is defined as a matrix with
blem:
columns the frame vectors {fi }n that is, F = [f1 ,…,fn ]. Usually, frame
i=1
2
means its synthesis operator. When α = β , the frame is calledα -tight min G−I ∞
F (6)
frame. If fi = 1 for all i, we have unit norm tight frame (UNTF) and
2
exists only for α = n/ m [19]. where ‖. ‖∞ is infinity norm. However, this case only occur for m = n ,
which is not happened in CS framework. In [26], another minimization
Definition ([14]). A matrices Fm × n = [f1 ,f2 ,…,fn ]with rows f1 ,̂ f2 ,̂ …,fm̂ is problem has been proposed for compressed sensing radar (CSR) as
said to be an equiangular tight frame (ETF) in a Hilbert space , if these follows:
conditions are satisfied:
∼ 2
min G−G
• The columns have unit norm:
F (7)
F
∼
f = 1; i = 1,2,…,n . where ‖. ‖F is the Frobenius norm and G is a diagonal matrix with the
• The
i
rows are orthogonal and have equal-norm: auto-correlation of each column of F. However, this method is based on

f = n/ m ; k = 1,2,…,m . identity matrix too and diagonal elements of the Gram matrix do not
• The inner products between distinct columns are equal in modulus:
k
have a significant impact on mutual coherence.

? ? fi ,f j = μWB ; i ≠ j,i,j = 1,2,…,n .

3. The proposed method: measurement matrix optimization
For any arbitrary frame Fm × n , the mutual coherence [20] is defined
as follows, which is an effective tool for recovery guarantees of a sen- In this section, we try to design a measurement matrix such that F is
sing matrix. as close as possible to an incoherent UNTF. To solve above drawbacks, a
minimization problem based on ETF is proposed in [9]. The object is to
fi ,f j find an equivalent sensing matrix which is as close as possible to an
μ(F ) ≜ max
i≠j ETF. Then, the problem of ETF design reduces to finding nearest matrix
fi . fj (4) to ETF by:
where fi and f j are different columns of F. Indeed, mutual coherence is a min G−GETF 2
GETF ∈ Λn F (8)
factor of similarity between the frame columns. The Gram matrix of the
frame Fm × n is defined as G = F T F . The diagonal elements of Gram where Λn is the convex set of the Gram matrix of all m × n ETFs [27]:
matrix show the squared ℓ2 norms of columns F and off-diagonal ele-
ments show the inner products between distinct columns. The Gram Λn = ⎧G ∈ n × n: G = GT , diag(G ) = 1, max gij ⩽ μWB ⎫
matrix is symmetric, positive semidefinite and it has rank m. If the ⎨
⎩ i≠j ⎬
⎭ (9)
columns of F have been normalized, then the mutual coherence can be where gij is the elements of Gram matrix G and diag(G ) refers to the
viewed as the largest absolute value of the off-diagonal entry of the
main diagonal of G. We use (8) as minimization problem and try to
Gram matrix. Therefore the Gram matrix of ETF (GETF ) has the diagonal
construct incoherent UNTF based on matrix nearness problem by the
elements and the absolute values of the off-diagonal elements equal to
following theorem.
one and μWB , respectively. The following lemma explains the relation
between RIP and coherence: Theorem ([27]). Suppose that the m × n matrix F has singular value
decomposition U ΣV ∗. With respect to the Frobenius norm, a nearest α -tight
Lemma ([21]). Let Fm × n is a matrix with unit norm columns and coherence
frame to F is given by α ·UV ∗ . Assume in addition that F has full row-rank.
μ . Then Fm × n satisfies the RIP of order k with δk ⩽ (k−1)μ whenever
Then α ·UV ∗ is the unique α -tight frame closest to F. Moreover, one may
k < 1 + 1/ μ .
compute UV ∗ using the formula (FF ∗)−1/2F .
Therefore the matrices with normalized columns and with low
It is noted that for α = n/ m , nearest n/ m -unit norm tight frame to
arbitrary frame is achievable. According to above, the procedure for
1
For SL0, we have used the MATLAB code available at: http://ee.sharif.edu/~SLzero/. measurement matrix optimization has been summarized in Algorithm

322
R. Entezari, A. Rashidi Int. J. Electron. Commun. (AEÜ) 82 (2017) 321–326

1. Gram matrix. By spectral decomposition of Gram matrix, we have

Firstly, we initialize Φ0 as a nearest unit norm tight frame to a Gk = USUT = (U S ). (U S )T and then the squared-root of the Gram
random matrices by above theorem. The random matrices such as matrix is extract by Sk = S UT so Gk = SkT Sk . According to above the-
Bernoulli and partial Fourier are better than the Gaussian matrices in orem, we find nearest unit norm tight frame to Sk and obtain
some cases. For example, the advantage of Bernoulli matrix over S′k = n/ m (SkSk∗)−1/2 Sk 2 and finally, the measurement matrix update
random Gaussian matrix is reduction of the hardware complexity and by minimizing the error S′k −Φk . Ψ 2F .
suitability for practical compressed sensing. The random Gaussian and
Algorithm 1. Proposed algorithm.
Bernoulli ensembles do not possess any structure and have no fast
Matrix-vector multiplication [28], but in partial Fourier matrix, fast
Fourier transform may be used to compute Matrix-vector multiplication

with O(nlogn) compared to O(knlog(n/ k ) ) for Gaussian matrices. Thus

these ensembles are not appropriate for large scale problems. Also, the 4. Simulation and numerical results
number of required measurements is suboptimal in partial Fourier
matrix m = O(k log5n) compared to m = O(k log(n/ k ) ) for Gaussian In this section, we evaluate the proposed method with other opti-
matrices [29]. It is noted that, all of the previous methods are listed mization methods including a random Gaussian matrix, Elad’s algo-
above, use random Gaussian matrix as initializing measurement matrix, rithm [12], Xu’s algorithm [9], Tsiligianni’s algorithm [11]. It is noted
but in our method, partial Fourier matrix has been proposed as alter- that, there are some frame design such as [30] to achieve the best
natives to random Gaussian and Bernoulli matrices, which entries are mutual coherence, but in this paper, we focus on frame design in which
generated by selecting the m rows at random from the n × n discrete the dictionary is assumed to be given. The initial dictionary is random
Fourier matrix and scaling the columns to have unit norm. Gaussian matrix which every entry is i.i.d with zero mean and unit
Then, the process of optimized measurement matrix repeats for Iter variance and dimension 80 × 120 .
times. Gram matrix compute by G = FkT Fk with elements gij . To solve In Fig. 1, the mutual coherence over the iterations for frame di-
(8), we use a shrinkage function on the Gram matrix to project the mensions 25 × 120 and 15 × 120 is depicted. It is seen from Fig. 1, the
elements on Λn to producing GP : proposed method has significant improvement in mutual coherence
after a few iterations. Also, in dimension 15 × 120 , the method in [11],
⎧1 i=j converge to the proposed method, but the proposed method has a fast
⎪
gij =
⎨
sign ( )
gij
. μ WB
μWB < abs (gij ) < 1 convergence to Welch bound only a few iterations (under 20).
Fig. 2, illustrates the histogram of the absolute off-diagonal entries
⎪ gij abs (gij ) < μWB
⎩ (10) of Gram matrix. It can be observed from Fig. 2 that the proposed
method has a shift towards the left of the histogram and therefore
It is noted that the shrink function which is used in [11], is ap- outperforms the other methods. The optimal value (red bar) is related
proximate the lower bound of mutual coherence for n ≫ m . Then we
to Welch bound.
update the Gram matrix by the current solution and previous solution Fig. 3, shows the mutual coherence over the number of measure-
through a weighting process. The process of weighting involves making
ment; m = 15: 5: 60 . It can be seen that the proposed method has a
a decision between the two states of Gram matrix (projected and pre- better reduction than other methods in terms of the number of mea-
vious Gram matrix) to compensate the transformation which is used in
surements. Due to near results to Welch bound in high measurement,
previous steps. In this step, we search the best condition for Gram we compare mutual coherence values reached by the algorithms for
matrix. Therefore, we update the Gram matrix by GP and previous Gram sensing matrix with dimensions m × 120 and different values of m
matrix by: (Table 1). It can be observed that our proposed method has lower
Gk = αGP + (1−α )Gk − 1 0<α<1 (11) mutual coherence than other methods. Also, Table 2 shows the mutual
coherence values for sensing matrix with dimensions 15 × n and dif-
This step size α is selected by search for the best result for a range of ferent values of n. These results show that our proposed method out-
0 < α < 1 (producing Gk ). performs the compared algorithms in terms of mutual coherence.
The shrink function which is used, causes the Gram matrix to be-
come full-rank. Thus, we used singular value decomposition (SVD) to
Gram matrix to reduce rank to be equal to m. Let USVT is SVD of the 2
The symbol (∗) indicates conjugate transposition of matrices.

323
R. Entezari, A. Rashidi Int. J. Electron. Commun. (AEÜ) 82 (2017) 321–326

0.6 0.6
ELAD Elad
Xu 0.55 Xu
LZYCB
0.5 Tsiligianni
Tsiligianni 0.5
Proposed Proposed
WB 0.45 WB

Mutual Coherence
0.4
0.4

0.3 0.35

0.3
0.2 0.25

0.2
0.1
20 40 60 80 100 0.15
Iteration
0.1
15 20 25 30 35 40 45 50 55 60
Number of measurements

1 Fig. 3. Mutual coherence as a function of the number of measurements, where

ELAD n = 120,p = 80,m = 15: 5: 60 and 100 iteration.
0.9 Xu
LZYCB
Tsiligianni
0.8 Proposed
WB Table 1
0.7
Mutual coherence values by Elad [12], LZYCB [10], Tsiligianni [11] and proposed
method for sensing matrix F ∈ m × 120 with different values of m and 100 iteration.
0.6
m Elad LZYCB Tsiligianni Proposed Welch bound
0.5
15 0.5750 0.7860 0.3607 0.3274 0.2425
0.4
20 0.4804 0.7044 0.3140 0.2695 0.2050
25 0.3072 0.6529 0.2723 0.2285 0.1787
0.3
30 0.2618 0.5845 0.2378 0.1999 0.1588
35 0.2273 0.5490 0.2126 0.1787 0.1429
0.2
20 40 60 80 100 40 0.2043 0.5080 0.1913 0.1625 0.1296
Iteration 45 0.1831 0.5243 0.1749 0.1522 0.1183
50 0.1683 0.3935 0.1636 0.1417 0.1085
55 0.1549 0.4006 0.1527 0.1368 0.0997
Fig. 1. Mutual coherence as a function of the iteration, where n = 120, p = 80 and 100 60 0.1363 0.3804 0.1466 0.1338 0.0917
iteration.

5000 Random
Optimal Table 2
Mutual coherence values by Elad [12], LZYCB [10], Tsiligianni [11] and proposed
method for sensing matrix F ∈ 15 × n with different values of n and 100 iteration.
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
n Elad LZYCB Tsiligianni Proposed Welch bound
5000 Proposed
Optimal 80 0.4412 0.7909 0.3110 0.2669 0.2342
85 0.4814 0.7519 0.3235 0.2757 0.2357
90 0.4920 0.8217 0.3336 0.2830 0.2370
0 95 0.5194 0.7594 0.3462 0.2920 0.2382
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
100 0.5324 0.7599 0.3483 0.3006 0.2392
5000 Elad 105 0.5516 0.7906 0.3454 0.3066 0.2402
Optimal 110 0.5682 0.7884 0.3620 0.3150 0.2410
115 0.5718 0.8157 0.3554 0.3223 0.2418
120 0.5718 0.7932 0.3579 0.3294 0.2425
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

5000 Xu Fig. 4, shows the relative errors as a function of sparsity order

Optimal 3 < k < 10 by SL0 reconstruction algorithm, for a fixed number of
measurement m = 25.
0 Fig. 5, shows the relative errors as a function of the number of
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
measurement m = 16: 2: 30 by SL0 reconstruction algorithm and for a
5000 Tsiligianni fixed sparsity order k = 4 . It is noted that vanishing graph in Fig. 5 is
Optimal related to a zero error rate.
In other scenario, we evaluate the performance of the proposed
0 measurement matrix in terms of recovery percentages and reconstruc-
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Mutual Coherence
tion SNR. In these simulations (Figs. 6–8), we use OMP algorithm as
reconstruction algorithm [31] and the k-sparse signals sn × 1are gener-
Fig. 2. The histogram of the absolute off-diagonal entries of G before and after optimi- ated by standard normal distribution (0,1) with i.i.d. elements. Then
zation, where n = 120, p = 80, m = 25 and 100 iteration.

324
R. Entezari, A. Rashidi Int. J. Electron. Commun. (AEÜ) 82 (2017) 321–326

0 35
10
Proposed
30 Random

25

Reconstruction SNR (dB)

−1
10
Relative Errors

20

15

−2 10
10
Elad
5
Xu
Tsiligianni
Proposed 0
Random
−3
10 −5
3 4 5 6 7 8 9 10 2 4 6 8 10 12 14 16 18 20
Sparsity order Sparsity Order (k)
Fig. 4. Relative errors as a function of the sparsity order, where n = 120,p = 80,m = 25 Fig. 7. The reconstruction SNR for different sparsity order 1 < k < 20 using OMP for the
and 150 iteration. noisy measurements with the SNR of 15 dB where the dimension of the random Gaussian
matrix and proposed matrix is 25 × 120 .
0
10
100
Proposed
90 Random

−1
80
10
Reconstruction SNR (dB)
Relative Errors

70

60

50
−2
10 40
Elad
Xu 30
Tsiligianni
Proposed 20
Random
−3
10 10
16 18 20 22 24 26 28 30
Number of measurements
10 20 30 40 50 60 70 80 90 100
Fig. 5. Relative errors as a function of the number of measurements using SL0, where Input SNR (dB)
n = 120,p = 80,m = 16: 2: 30 and 150 iteration for fixed sparsity order k = 4 .
Fig. 8. The reconstruction SNR of a 5-sparse signal from its noisy measurement using
OMP for various input SNRs where the dimension of the random Gaussian matrix and
proposed matrix is 25 × 120 .

n−k of its elements set to zero and the location of k non-zero elements
100 (support) is chosen uniformly at random. The sensing matrices Fm × n are
Proposed generated according to their structures and the measurements by
90 Random ym × 1 = Fm × n·sn × 1. For each sensing matrix, the sampling and re-
80 construction process are averaged over 1000 different runs.
The perfect recovery percentage is the ratio between the numbers of
70
Recovery Percentage

exactly reconstructed signals to the number of all k-sparse signals. Fig. 6

60 shows the perfect recovery percentage (SNR rec ⩾ 100 dB) of noiseless k-
sparse signals at different sparsity order; increases from 1 to 12. The
50 proposed sensing matrix is 25 × 120 and the coherence of this matrix is
40 0.17. According to theoretical results, k-sparse signals can be perfectly
recovered for k < 6.59 . As can be seen in Fig. 6, the proposed matrix has
30 better performance than random Gaussian matrix.
20 In order to evaluate the effect of noise in reconstruction process, we
illustrate some simulation. For a signal sn × 1, suppose sn∗× 1 is the re-
10 constructed signal. The SNR of the reconstructed signal is referred to as
0 the reconstruction SNR and is defined as follows:
2 4 6 8 10 12
Sparsity order (k) s 2 ⎞
SNR(s ) = 20·log10⎜⎛ ∗ ⎟dB
Fig. 6. The recovery percentage for different sparsity order 1 < k < 12 using OMP, where ⎝ s−s 2 ⎠ (12)
the measurements are noiseless and the dimension of the random Gaussian matrix and
proposed matrix is 25 × 120 .
Fig. 7 demonstrates the reconstruction SNR of 1000 k-sparse signals

325
R. Entezari, A. Rashidi Int. J. Electron. Commun. (AEÜ) 82 (2017) 321–326

coding based on compressed sensing. AEU-Int J Electron Commun 2017.

of size 120 × 1 with sparsity order 1 ⩽ k ⩽ 20 . The measurements are
[7] Candes EJ, Tao T. Decoding by linear programming. IEEE Trans Inf Theory
mixed by Additive White Gaussian Noise (AWGN) with the input SNR at 2005;51(12):4203–15.
15 dB for each sparsity order. The results show that the proposed matrix [8] Ge D, Jiang X, Ye Y. A note on the complexity of lp minimization. Math Programm
2011;129(2):285–99.
outperforms than random Gaussian matrix. [9] Xu J, Pi Y, Cao Z. Optimized projection matrix for compressive sensing. EURASIP J Adv
Also, Fig. 8 shows the reconstruction SNRs of 5-sparse signals when Signal Process 2010;2010(1):560349.
[10] Li G, Zhu Z, Yang D, Chang L, Bai H. On projection matrix optimization for compressive
the measurements are mixed by AWGN with input SNRs ranging from 0 sensing systems. IEEE Trans Signal Process 2013;61(11):2887–98.
to 100 dB. The curves are close at low input SNR, but at higher input [11] Tsiligianni EV, Kondi LP, Katsaggelos AK. Construction of incoherent unit norm tight
SNR, Fig. 8, shows an improvement in reconstruction SNR for proposed frames with application to compressed sensing. IEEE Trans Inf Theory
2014;60(4):2319–30.
matrix. [12] Elad M. Optimized projections for compressed sensing. IEEE Trans Signal Process
2007;55(12):5695–702.
[13] Calderbank R, Howard S, Jafarpour S. Construction of a large class of deterministic
5. Conclusion sensing matrices that satisfy a statistical isometry property. IEEE J Sel Top Signal Process
2010;4(2):358–74.
[14] Bandeira AS, Fickus M, Mixon DG, Wong P. The road to deterministic matrices with the
In this paper, we focus on measurement matrix optimization based
restricted isometry property. J Fourier Anal Appl 2013;19(6):1123–49.
on incoherent UNTF with respect to compressed sensing. We tried to [15] Tropp JA. Greed is good: algorithmic results for sparse approximation. IEEE Trans Inf
minimize the mutual coherence between the measurement and dic- Theory 2004;50(10):2231–42.
[16] Scott AJ. Tight informationally complete quantum measurements. J Phys A: Math Gen
tionary matrix based on the features of the incoherent frames, matrix 2006;39(43):13507.
nearness problem and partial Fourier matrix. Simulation results show [17] Strohmer T, Heath RW. Grassmannian frames with applications to coding and commu-
nication. Appl Comput Harmonic Anal 2003;14(3):257–75.
that the proposed method has better performance than Elad’s, Xu’s, [18] Mohimani GH, Babaie-Zadeh M, Jutten C. Complex-valued sparse representation based on
LZYCB’s and Tsiligianni’s method in terms of mutual coherence and smoothed l 0 norm. 2008 IEEE international conference on acoustics, speech and signal
relative errors. Also, our method confirm that the performance of sen- processing. IEEE; 2008. p. 3881–4.
[19] Christensen O. An introduction to frames and Riesz bases vol. 7. Springer; 2003.
sing matrix generated by the measurement matrix optimization is better [20] Donoho DL, Huo X. Uncertainty principles and ideal atomic decomposition. IEEE Trans
than random Gaussian matrix in terms of reconstruction the noiseless Inf Theory 2001;47(7):2845–62.
[21] Cai TT, Xu G, Zhang J. On recovery of sparse signals via l1 minimization. IEEE Trans Inf
and noisy signals and perfect recovery percentages exploited the effi- Theory 2009;55(7):3388–97.
ciency of proposed method. [22] Sustik MA, Tropp JA, Dhillon IS, Heath RW. On the existence of equiangular tight frames.
Linear Algebra Appl 2007;426(2-3):619–35.
[23] Fickus M, Mixon DG. Tables of the existence of equiangular tight frames; 2015, arXiv
Acknowledgment preprint arXiv:<1504.00253>.
[24] Duarte-Carvajalino JM, Sapiro G. Learning to sense sparse signals: simultaneous sensing
matrix and sparsifying dictionary optimization. IEEE Trans Image Process
The authors would like to thank the anonymous reviewers for their 2009;18(7):1395–408.
helpful and constructive comments. [25] Abolghasemi V, Ferdowsi S, Makkiabadi B, Sanei S. On optimization of the measurement
matrix for compressive sensing. 18th European signal processing conference, 2010. IEEE;
2010. p. 427–31.
References [26] Zhang J, Zhu D, Zhang G. Adaptive compressed sensing radar oriented toward cognitive
detection in dynamic sparse target scene. IEEE Trans Signal Process 2012;60(4):1718–29.
[27] Tropp JA, Dhillon IS, Heath RW, Strohmer T. Designing structured tight frames via an
[1] Donoho DL. Compressed sensing. IEEE Trans Inf Theory 2006;52(4):1289–306. alternating projection method. IEEE Trans Inf Theory 2005;51(1):188–209.
[2] Candès EJ, Romberg J, Tao T. Robust uncertainty principles: exact signal reconstruction [28] Ailon N, Rauhut H. Fast and rip-optimal transforms. Discrete Comput Geomet
from highly incomplete frequency information. IEEE Trans Inf Theory 2014;52(4):780–98.
2006;52(2):489–509. [29] Jafarpour S. Deterministic compressed sensing. Princeton University; 2011.
[3] Muqaibel AH. Efficient reformulation of image reconstruction with compressive sensing. [30] Sadeghi M, Babaie-Zadeh M. Incoherent unit-norm frame design via an alternating
AEU-Int J Electron Commun 2017;76:46–51. minimization penalty method. IEEE Signal Process Lett 2017;24(1):32–6.
[4] John FGP, Sankararajan R. Efficient reconstruction of compressively sensed images and [31] Pati YC, Rezaiifar R, Krishnaprasad PS. Orthogonal matching pursuit: recursive function
videos using non-iterative method. AEU-Int J Electron Commun 2017;73:89–97. approximation with applications to wavelet decomposition. 1993 conference record of
[5] Xie H, Andrieux G, Wang Y, Feng S, Yu Z. A novel threshold based compressed channel the twenty-seventh Asilomar conference on signals, systems and computers, 1993. IEEE;
sensing in OFDM system. AEU-Int J Electron Commun 2017;77:149–55. 1993. p. 40–4.
[6] Xie D, Peng H, Li L, Yang Y. An efficient privacy-preserving scheme for secure network

326