08039246

SPECIAL SECTION ON ANALYSIS AND SYNTHESIS OF LARGE-SCALE SYSTEMS
Received July 7, 2017, accepted August 28, 2017, date of publication September 18, 2017, date of current version February 14, 2018.
Digital Object Identifier 10.1109/ACCESS.2017.2753240
A Holistic Approach to Reconstruct Data in Ocean
Sensor Network Using Compression Sensing
HUAFENG WU1 , (Senior Member, IEEE), MENG SUO1, JUN WANG2, (Senior Member, IEEE),
PRASANT MOHAPATRA3, (Fellow, IEEE), AND JUNKUO CAO4
1Merchant Marine College, Shanghai Maritime University, Shanghai 201306, China
2University of Central Florida, Orlando, FL 32836, USA
3University of California, Davis, CA 95616, USA
4Hainan Normal University, Haikou 571158, China
Corresponding authors: Huafeng Wu (hfwu@shmtu.edu.cn) and Jun Wang (Jun.Wang@ucf.edu)
ABSTRACT In the complex marine environment, a large-scale wireless sensor network (WSN) is often
deployed to resolve the sparsity issue of the signal and to enforce an accurate reconstruction of the signal
by upgrading the transmission efficiency. To best implement, such a WSN, we develop a holistic method by
considering both raw signal processing and signal reconstruction factors: a node re-ordering scheme based
on compression sensing and an improved sparse adaptive tracking algorithm. First, the sensor nodes are
reordered at the sink node to improve the sparsity of the compression sensing algorithm in the discrete cosine
transformation or Fourier transform domain. After that, we adopt the matching test to estimate sparse degree
Kis. At last, we develop a sparse degree adaptive matching tracking framework step-by-step to calculate the
approximation of sparsity, and ultimately converge to a precise reconstruction of the signal. In this paper,
we employ MATLAB to simulate the algorithm and conduct comprehensive tests. The experimental results
show that the proposed method can effectively reduce the sparsity of the signal and deliver an accurate
reconstruction of the signal especially in the case of unknown sparsity.
INDEX TERMS Compression sensing, sparsity, ocean sensor network, data reconstruction.
I. INTRODUCTION
Recent years have seen the ocean sensing network being
widely deployed in marine environmental monitoring, mar-
itime search and rescue in light of its flexibility. Due to the
dynamic characteristics of the ocean climate, the ups and
downs feature of waves and many other natural factors, sensor
node often consumes lots of additional energy to repeatedly
communicate with others. As a consequence, it is neces-
sary to take into account the energy efficiency of the ocean
sensing network building block–sensor node, which directly
determines the lifespan of nodes and thus its dependability.
To improve node energy efficiency, a compression sensing
technique [1] has been developed to reduce the total amount
of data to be transmitted on the sensor node, which appears to
play an important role in the future marine environment mon-
itoring. It may be noted that the effectiveness of compression
sensing depends on the compression performance of the data.
The better the compression performance of the original data,
the less the amount of data needed to collect and thus transmit.
Unfortunately, it could be very challenging to effectively
apply compression sensing. In the past, the sensor nodes were
arranged according to the label number, and collected at the
Sink node for forwarding. Such a straightforward protocol
has low efficiency.
In this paper, we employ a holistic method by working at
two levels of Ocean Sensor Network: raw signal processing
and signal reconstruction. We develop a new reordering algo-
rithm which resorts the sensor nodes at the Sink node. The
idea is to enhance the sparsity of the signal via reducing the
number of measurements needed for signal reconstruction,
consequently, resulting in a low compression sampling rate.
A low compression sampling rate avoids much unnecessary
communication traffic and therefore leads to a better network
bandwidth utilization, less energy consumption and longer
lifespan of the entire wireless sensor networks. It may be
noted that our proposed node re-ordering method is only
re-ordering at the Sink node instead of in the whole WSN,
where sensor nodes do not engage with the sorting procedure
at all. In this case, in addition to reconstructing the original
signal from the compressed sample data, the node is also
280
2169-3536
2017 IEEE. Translations and content mining are permitted for academic research only.
Personal use is also permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
VOLUME 6, 2018

H. Wu et al.: Holistic Approach to Reconstruct Data in Ocean Sensor Network
responsible for calculating the optimal order in the transform
domain. Without loss of generality, we assume that the sensor
node performs well in terms of computation and storage
involved in any relevant sorting, sampling and reconstruction
algorithms. It may be also noted that the position and order
of the nodes are two different concepts.
Reconstruction algorithm plays a very critical role in com-
pression perception because the key here is to exploit the most
accurate measurement data to reconstruct the original data.
With regards to the signal reconstruction problem, numerous
researchers have developed recognized solutions. The combi-
nation of optimization algorithm bears strict sampling struc-
ture constraint to the operation of high efficiency. Convex
optimization algorithms, such as base tracking [2], involve
heavy calculation and computation overhead even at a faster
measuring speed with high precision in comparison to oth-
ers. Greedy algorithm runs fast and performs sampling with
high efficiency. Example greedy algorithms include match-
ing pursuit, orthogonal matching pursuit [3], block matching
pursuit [4], the regularized orthogonal matching pursuit [5],
compressed sampling matching pursuit [6] and subspace
tracking [7], the sparse degree of adaptive matching
pursuit, etc.
To adapt to dynamics characteristics of ocean WSN,
we develop an extended sparse adaptive matching track-
ing algorithm based on aforementioned Greedy algorithm.
By combining backtracking filtering and adaptive thinking,
we allow a large number of iterative computations and large
steps be engaged in the previous algorithm under large sparse
degree conditions (CAMP), and develop a new adaptive vari-
able step-size algorithm (CAMP).
II. PROBLEM RESTATEMENT
In the compression sensing model, the sparse correlation
expression is m > C · S · log N · µ2( , ). Where m denotes
the signal to accurately recover the number of sensor nodes
to be measured, C is constant (C > 1). S represents the
sparsity of signal, N means dimension size of compressed
signal, µ denotes the correlation of measurement matrix
and transform domain (In theory, the lower the correlation,
the better compression effect).
Without losing the generality, we use orthogonal gaus-
sian random matrix as matrix , while domain is DCT
transform domain. This algorithm is also applied to differ-
ent orthogonal matrix and transform domain. For example,
can be wavelet transform domain or the Fourier transform
domain. At first, the signal X is sorted by the sensor node
label, and make the signal X in the domain carry a better
sparsity. For example, the real-life temperature data of a
selected area in East China sea are used for demo in this
paper.
Because the length of the signal (N) is much larger than
the observations (M), if we want to restore an N-dimension
signal, we need to solve the underdetermined equations
Y = X. Roughly, the equations end up with an infinite
number of solutions. However, as discussed before, the sig-
nal X can be represented if signal is sparse. This means
the problem is fundamentally different now. In principle,
the observation matrix with characteristics of RIP for the
recovery of signals in an observed value M precisely pro-
vides a guarantee. In order to clarify the problem of sig-
nal reconstruction, we first define the P norm of vector
x = {x1, x2, · · · xn}:
X p =
N
i=1
|xi|p
1/p
(1)
When p = 0 is called 0 norm, meaning the number of
elements is not zero in X. As a result, if the signal X is
a sparse representation, the solution to the system of unde-
termined equations Y = X transforms to the problem of
solving the minimum norm 0, as shown in Equation (1).
However, the condition for solving the minimum norm tells
that the linear combination of species not equaling zero may
be listed in order to obtain the optimal solution. It can be
seen that the solution of type (1) is an NP-hard problem.
Donoho and Wu Fei-Yun found that [8], [9], using a simpler
l1 optimization problem instead of norm 0 will get the same
solution:
ˆx = arg min x 1 s.t Y = X (2)
The convex optimization problem cannot be solved directly.
The current solution to this problem is mainly based on
greedy iterative algorithm, which is simple and easy to
operate. Matching pursuits(MP) and Orthogonal matching
pursuit(OMP) algorithms are often used to solve the con-
vex optimization problems. Due to their low efficiency,
researchers develop extended versions – Stagewise OMP
(StOMP) and Regularized OMP (ROMP) algorithms. Later,
with the development of the retreat screening idea, the Sub-
space pursuit (SP) and Compressive Sampling (CoSaMP)
algorithms [10] have been developed. They reduce the spar-
sity but leave the prior information sparse degree K become
an indispensable part.
III. THE NODE REORDERING MODEL BASED
ON GREEDY ALGORITHM
From the DCT domain to the unitary inverse projection
vector of the time domain, some sampling vectors can be
obtained. If we find that these sampling vectors are basi-
cally matched with one of the basis vectors, we can predict
an optimal ordering so that the energy of the signal in the
DCT or DFT domain is concentrated on a small number of
Fourier coefficients. This article will set as DCT domain,
hence the inverse projection vector of sampling domain is
−1 at this time. Each base vector can find its optimal per-
mutation, then N base vectors will have N sorting methods.
A sort of the most matching basis vector is selected from the
n sorties.
In this model, f is the discrete space vector measured by
sensor nodes. At first, the signal f is only sorted through the
sensor node id. The purpose of this article is to find a way
of ordering to make the signal more sparse in the domain .
VOLUME 6, 2018 281

For each base vector f , the following matrix is constructed as
described above, as type (3):
N×N =








| f1 − ψ1| | f1 − ψ2| · · · | f1 − ψN |
| f2 − ψ1| | f2 − ψ2| · · · | f3 − ψN |
·
·
·
| fN − ψ1| | fN − ψ2| · · · | fN − ψN |








(3)
In this algorithm, we select the set of f with the smallest
difference in each row, and calculate the minimum sum of the
N signals according to this matching method. The selected
N signals are not on the same row nor the same column.
It is noted that our aim is near solution rather than optimal
solution. The results show that the near-optimal solution
could still make the signal compression be significantly
improved.
Next, we introduce the greedy algorithm used in this paper
to find the optimal sorting way of N signals elements. This
algorithm finds the preferred sorting method for the selected
signal samples, where S express the unit vector matrix
in sparse domain (fixed), f is a discrete signal. It is impor-
tant to note f vector and vector used in this algorithm
needs to be standardized, which is with the same number.
Standardization is to reduce unnecessary error because of the
difference between signal and base vector, otherwise the com-
pressibility would be affected. The pseudo code of the greedy
optimal algorithm proposed in this paper is illustrated as
follows:
Algorithm 1 Greedy Optimal Algorithm
1: for fi = f1 · · · fN
2: for ψj = ψ1 · · · ψN
3: end
4: end
5: A ← Ø, B ← Ø, δ ← 0
6: for m = 1 · · · N
7: select i,j min from
fi ∈ {f1 , f2 · · · fN } and
ψj ∈ {ψ1 , ψ2 · · · ψN }
8: A ← A ∪ {fi} , B ← B ∪ ψj , δ = δ + i,j min
9: end
The algorithm first constructs the matrix . After n oper-
ations, the greedy algorithm is used to select optimal signals
which are neither in the same row nor in the same column.
At this time, the sum of the signals is also minimum. The
1-4 rows of the algorithm build the matrix , the fifth row
initializes the row set and the column set respectively. The
6-8 steps constitute the core part of the algorithm, and
realize the greedy optimism. In each iteration opera-
tion, the smallest signal elements in the ith row and
the jth column are selected as the optimal solution,
where the selection range of fi and fj is not in the
sets A and B.
IV. IMPROVING CURRENT SPARSE ADAPTIVE
MATCHING PURSUIT ALGORITHM
The basic idea of the matching pursuit algorithm is, in every
iteration, choosing the best matching signal of atoms from
the measurement matrix to support set F which is used to
reconstruct signal and calculate the residual r between the
original signal. Eventually, we work through the residual
by calculating choices and matching atoms to update to the
support set. After the iteration process completes, the recon-
struction of the signal is obtained.
If we want to get the size of the initial support set, we need
to first estimate the sparse degree of signal. Next, we use the
idea of variable phase step to accurately estimate the sparse
degree of signal by a gradual fashion. Finally, the signal will
be reconstructed accurately.
A. SPARSE ADAPTIVE MATCHING PURSUIT ALGORITHM
Do et. al. has improved the Sparsity adative matching pursuit
(SAMP) algorithm [11], in a way that, as long as the observa-
tion matrix and sparse signal meet the RIP conditions, signals
can be reconstructed both precisely and quickly, and avoid
prior information (sparsity). The SAMP algorithm approxi-
mates the sparsity K step by step, which can implement the
exact reconstruction of the sparse signal f under the premise
of unknown sparse degree. Given a certain iteration process,
firstly, according to type (4) to calculate the correlation coef-
ficient µ of individual atoms and residual r in measurement
matrix ; then, choose the largest amplitude which size-index
values correspond to the index set S. At this point the support
set F remains unchanged.
µ = {µi |µi = | rk−1, ϕi | , i = 1, 2, · · · , N } (4)
We then merge the index set S and the previous iteration of
the support set F, get the candidate set C, and calculate the
inner product of the atom by type (4). The current iteration
of the support set F is composed of the index which has the
biggest size-absolute values. If the current surplus energy is
less than the previous iteration surplus energy, continue to
iteration; On the contrary, go into the next phase until the
surplus energy is less than the set threshold. After a certain
number of iterations to get support F, according to index
values in support set, we take out F which is made of
the corresponding atoms in measurement matrix. The sig-
nal reconstruction can be realized by using the least-square
method. At the same time, we update the surplus energy
below:
ˆf = arg min y − F f 2 (5)
rnew = y − F ˆf (6)
The basic steps of SAMP algorithm:
Input: m-dimensional measurement vector y, M∗N mea-
surement matrix, the phase step step = 0;
Output: the sparse degree of signal f is similar to ˆf ;
1. Initialization: allowance r = y, the length of support
set: size = step, stage: stage = 1, the number of iterations:
282 VOLUME 6, 2018

K = 1, the index value set: S = Ø, the candidate set: C = Ø,
support set: F = Ø;
2. If r 2 ≤ ε, the iteration will be terminated, then atoms
which have got are used to put in type (5) to reconstruct signal
lastly; Otherwise enter step 4;
3. Calculate the correlation coefficient, and get the corre-
sponding size-index of maximum from the set of correlation
coefficient by type (4), which is index set S;
4. Put index set and support set together and get the can-
didate set C = F ∪ S, calculate the correlation coefficient of
corresponding atoms of index in C and margin by type(4), and
take out corresponding size-index of maximum to get Fnew.
At last, update margin by type (6).
5. If rnew 2 ≥ r 2, update the iterative phase stage =
stage + 1, update the support set length, turn to Step 2;
Otherwise update the candidate set F = Fnew, update the
margin r = rnew, update the number of iterations K = K +1,
go to Step 2.
As you can see from Step 5, the step size of each step is
constant, and the size of sparse K is the size of the final size.
Obviously, when step value is bigger, algorithm efficiency is
higher, but the final sparse degree K precision is significantly
reduced. When the step value is smaller, the accuracy of
sparse degree is higher, but the number of iterations will
increase, and thus algorithm execution efficiency is low. The
initial value of sparse degree in this algorithm is the step in
initial stage. It may be noted that the initial value cannot
improve the efficiency of the algorithm [12], because the
reconstruction precision and reconstruction efficiency of the
algorithm are constrained by the step size and the initial
sparsity estimate.
B. SPARSE DEGREE ESTIMATION METHOD
SAMP algorithm does not estimate specific sparse degree
of signal, while the estimated sparse degree of initial signal
will play a crucial role in the overall algorithm. In this paper,
we develop a method that initial signal’s sparse degree can be
estimated, and finally get a collection of atoms. We set F as
a real support set of y, and use sup(·) to represent the number
of elements in the collection.
The initial margin is y, and the ith element of the set
can be obtained by using Equation(3), where the index set
of the maximum value of the former K0(1 ≤ K0 ≤ N)
(N is the signal length) is recorded as F0, then sup(F0) = K0.
The method is proved in reference [13]: K0 is set as initial
value, if T
F0
y ≤ 1−δ
1+δ y 2, we need to increase K0 in turn
until the inequality is not established, K0 is the sparse degree
at this time, initial margin r can be obtained at the same time.
C. THE IDEA OF CHANGING STEP
At the unknown condition of sparse degree in SAMP algo-
rithm, it is appropriate to choose step = 1 as phase step
length, but the reconstruction efficiency will drop rapidly.
Because the estimate value at initial stage is much less than
the real sparse degree, a big stage step is chosen to increase
the execution efficiency. When the estimated sparse degree
increases to a certain value, and the sparse degree is close
to the actual sparse degree, we choose a small stage step
at this time. This method can improve the accuracy of the
reconstruction. The idea of changing the step size not only
improves the precision of the algorithm, but also improves
the execution efficiency.
In the literature [14], it is pointed out that when the support
set size has not reached the degree of sparse K, the energy
difference of the reconstructed signal in two adjacent phases
is decreasing and then gradually decreases with the increasing
number of stages. Afterwards, signal reconstruction basically
completes. This ensures that the iteration loop terminates
when the energy difference of the adjacent stages triggers.
FIGURE 1. The CAMP algorithm model.
The algorithm process of CAMP model is depicted
in Fig.1. The basic steps of CAMP algorithm is described as
follows:
Input: m-dimensional measurement vector y, M∗N mea-
surement matrix, the phase step length step = 0;
Output: the sparse degree of signal f is similar to ˆf ;
1. Initialization: the degree of sparse K, the length of
support set: size = step, support set: F = Ø, the initial size
of support set is step;
2. Using the Formula (4) to calculate the correlation coef-
ficient µ, and from which the maximum value corresponding
to the K-index value stored in the index set F;
3. If T
F y ≤ 1−δk
1+δk
y 2, then K0 = K0 + step, go to
Step 2; If T
F y ≤ 0.5 1−δk
1+δK
y 2, then step = 2 · step ,
K0 = K0 + step, go to Step 2;
VOLUME 6, 2018 283

4. Initialize the margin r = y − F
T
F y;
5. Initialization stage, stage = 1; The number of iterations,
K = 1; the size of support set, size = K0; The set of index
value, S = Ø;
6. Using the Formula (4) to calculate the correlation coeffi-
cient, and from µ put the size of the maximum corresponding
to the index value stored in the index set S;
7. Put index set and support set altogether, then,
Fnew = F ∪ S, calculate the correlation coefficient of the
corresponding atoms of index and margin by formula (4);
extract the size-index value corresponding to the maximum
value stored in Fnew. Estimate the signal f by formula (5),
and update margin by formula (6);
8. If ˆfnew − ˆf ≤ ε1, go to Step 9;otherwise go to Step 10;
9. If ˆfnew − ˆf ≤ ε2, stop the iteration; otherwise go to
step 11;
10. If rnew 2 ≥ r 2, stage = stage + 1, step = step + 1,
go to step 6; otherwise F = Fnew, r = rnew, K = K +1, go to
step 6;
11. If rnew 2 ≥ r 2, stage = stage + 1, step = step,
size = size + step, go to step 6; otherwise step = 1
2 step,
r = rnew, K = K + 1, go to step 6.
Step (1) to step (3) have estimated the degree of sparse,
the variable value is initialed at step (4) and step (5), step (6)
and step (7) use regularization to filter and update the atoms,
step (8) to step (11) control the time when iteration stop and
improve the accuracy of algorithm by threshold value ε1, ε2.
In the CAMP algorithm, the least squares method is used
in Step 7 to estimate the signal as the main computational
process.
V. SIMULATION EXPERIMENTS AND
PERFORMANCE ANALYSIS
In order to test the node reordering model based on greedy
algorithm and the performance improvement of the sparse
degree of adaptive matching pursuit algorithm, we have
developed experimental testbed using Matlab and carried out
the corresponding simulation. In the simulation run, the origi-
nal signal collected 829 ocean temperature signals in different
periods. Discrete cosine transform domain and Gaussian ran-
dom matrix are used as the sparse transformation domain and
measurement matrix respectively.
A. NODE RE-ORDERING SIMULATION BASED
ON GREEDY ALGORITHM
After simulation, we compare the original signal with sorted
signal in DCT domain, the results of simulation are illustrated
in Fig.2 and Fig.3. Fig.2 shows the original discrete signal f of
size 829 composed of the sensed temperature values accord-
ing to permutation of samples ordered by the original id of the
SNs. Fig.3 depicts the DCT transform of the actual vector f ,
because the signal is synthetically generated, the amplitude
of the signal in the figures are without unit.
After reordering the sensor nodes, the outcome of our
sorting method on the sensor nodes is shown in Fig.3 above.
FIGURE 2. Original signal.
FIGURE 3. The signal after sorted.
FIGURE 4. The original signal and the sorted signal in the DCT field.
The simulation results (Fig.4) show that the sparsity of the
signal becomes better after being projected to the DCT
domain.
B. SIMULATION OF IMPROVED SPARSE DEGREE
ADAPTIVE MATCHING TRACKING ALGORITHM
We simulate our improved algorithm at two aspects: sig-
nal reconstruction performance and error analysis. In this
experiment, we configure the step size step = 4, parameter
δk = 0.3.
1) SIGNAL RECONSTRUCTION PERFORMANCE SIMULATION
In order to verify the correctness of the reconstruction of
the CAMP algorithm, the reconstructed signals used the
CAMP reconstruction algorithm and original signal are sim-
ulated respectively, as shown in Fig.5 and Fig.6, and we
also compare the four signals respectively reconstructed by
CAMP, OMP, SAMP, SP with original signals (Fig.7). We can
observe the error range of CAMP algorithm is basically
284 VOLUME 6, 2018

FIGURE 5. Original signal.
FIGURE 6. Reconstructed signal.
FIGURE 7. Four algorithms error comparison diagram.
controlled within ±0.05, CAMP is able to recover signal
completely and its outcome matches well with the original
signal.
2) THE ERROR ANALYSIS
In this paper, the Normalized Mean Absolute Error(NMAE)
is used to measure the distortion of data reconstruction. The
formula is as follows:
NMAE =
n
i=1
fi − ˆfi
n
i=1
|fi|
(7)
Where fi is the data collected by the N nodes in the
ith sample, and ˆfi is the data reconstructed by the fusion
center.
As shown in Fig. 8, in all cases, it is clear that the error
rate of signal reconstruction decreases as M/N increases. The
results show that the error of CAMP algorithm is lower than
FIGURE 8. The algorithm error and compression ratio diagram.
that of other algorithms, such as SP, SAMP, OMP, and the per-
formance is better than other algorithms, and the compression
ratio is higher than that of other algorithms. At the same time,
the error is gradually reduced. Therefore, when the sparsity
becomes better, the CAMP algorithm can effectively reduce
the error and improve the accuracy.
VI. SUMMARY
We propose a re-ordering algorithm based on compression
sensing and an extended sparse adaptive tracking algorithm
to address the efficiency issue of large-scale wireless sensor
network in the marine environment. Our re-ordering algo-
rithm dramatically improves the signal sparsity property in
DCT or Fourier Transform Domain. Hence, the sparse degree
estimation is obtained by an atomic matching test method,
and then the sparse degree approximation is obtained step
by step in the sparse degree of adaptive matching tracking
frame, and finally the exact reconstruction of the signal is
accomplished. Experiments show that the node reordering
algorithm can reduce the sparsity of the signal in the DCT
domain. In addition, our extended sparse adaptive matching
tracking algorithm can restore the original signal more accu-
rately than existing schemes when the sparsity is unknown.
In the complex marine environment with large-scale wireless
sensor network applications, both performance and efficiency
achieved by our proposed methods seem to be better than
similar algorithms.
REFERENCES
[1] N. Jing, W. Bi, Z. Hu, and L. Wang, ‘‘A survey on dynamic compressed
sensing,’’ Acta Autom. Sinica, vol. 41, no. 1, pp. 22–37, 2015.
[2] S. S. Chen, D. L. Donoho, and M. A. Saunders, ‘‘Atomic decomposition
by basis pursuit,’’ SIAM Rev., vol. 43, no. 1, pp. 129–159, 2001.
[3] J. A. Tropp and A. C. Gilbert, ‘‘Signal recovery from random measure-
ments via orthogonal matching pursuit,’’ IEEE Trans. Inf. Theory, vol. 53,
no. 12, pp. 4655–4666, Dec. 2007.
[4] D. L. Donoho, Y. Tsaig, I. Drori, and J.-L. Starck, ‘‘Sparse solution
of underdetermined systems of linear equations by stagewise orthog-
onal matching pursuit,’’ IEEE Trans. Inf. Theory, vol. 58, no. 2,
pp. 1094–1121, Feb. 2012.
[5] D. Needell and R. Vershynin, ‘‘Uniform uncertainty principle and signal
recovery via regularized orthogonal matching pursuit,’’ Found. Comput.
Math., vol. 9, no. 3, pp. 317–334, Jun. 2009.
[6] D. Needell and J. A. Tropp, ‘‘CoSaMP: Iterative signal recovery from
incomplete and inaccurate samples,’’ Appl. Comput. Harmon. Anal.,
vol. 26, no. 3, pp. 301–321, 2009.
VOLUME 6, 2018 285

[7] W. Dai and O. Milenkovic, ‘‘Iterative subspace pursuit decoding of
weighted Euclidean superimposed codes,’’ in Proc. 5th Int. Symp. Turbo
Codes Rel. Topics, 2008, pp. 402–407.
[8] F. Y. Wu, Y.-H. Zhou, and F. Tong, ‘‘A fast sparse signal recovery algo-
rithm based on approximate l0 norm and hybrid optimization,’’ Zidonghua
Xuebao/Acta Autom. Sinica, vol. 40, no. 10, pp. 2145–2150, 2014.
[9] D. L. Donoho, M. Elad, and V. N. Temlyakov, ‘‘Stable recovery of sparse
overcomplete representations in the presence of noise,’’ IEEE Trans. Inf.
Theory, vol. 52, no. 1, pp. 6–18, Jan. 2006.
[10] E. J. Candès and T. Tao, ‘‘Decoding by linear programming,’’ IEEE Trans.
Inf. Theory, vol. 51, no. 12, pp. 4203–4215, Dec. 2005.
[11] T. T. Do, L. Gan, N. Nguyen, and T. D. Tran, ‘‘Sparsity adaptive matching
pursuit algorithm for practical compressed sensing,’’ in Proc. Asilomar
Conf. Syst. Comput., 2009, pp. 581–587.
[12] Y.-F. Shen, J.-T. Li, Z.-M. Zhu, Y.-D. Zhang, and F. Dai, ‘‘Image
reconstruction algorithm of compressed sensing based on nonlocal sim-
ilarity model,’’ Zidonghua Xuebao/Acta Autom. Sinica, vol. 41, no. 2,
pp. 261–272, 2015.
[13] Y. Cheng, F. Wei, F. Hui, Y. Tao, and H. U. Bo, ‘‘A sparsity adaptive
subspace pursuit algorithm for compressive sampling,’’ Acta Electron.
Sinica, vol. 38, no. 8, pp. 1914–1917, 2010.
[14] R. Gao, R. Zhao, and S. Hu, ‘‘Variable step size adaptive matching pursuit
algorithm for image reconstruction based on compressive sensing,’’ Acta
Opt. Sinica, vol. 30, no. 6, pp. 1639–1644, 2010.
HUAFENG WU (SM’16) received the bachelor’s
degree in navigation technology (junior prepara-
tory classes) from Jimei University in 1997, the
master’s degree in traffic information engineering
and control from Dalian Maritime University in
2004, and the Ph.D. degree in computer applica-
tion technology from Fudan University in 2008.
He is currently the Deputy Director of the Human
Resource Division, a Full Professor, and a Ph.D.
Supervisor with the Merchant Marine College,
Shanghai Maritime University. From 2008 to 2009, he conducted post-
doctoral research at Carnegie Mellon University. He was a Visiting Scholar
with Shanghai Jiao Tong University from 2012 to 2013. He serves as an
Associate Editor for the IEEE COMPUTER COMMUNICATIONS and the Journal of
Cyber-Physical Systems, and an Executive Member of the Shanghai Institute
of Electronics and a Shanghai Shuguang Scholar endorsed by the Shanghai
Government.
MENG SUO received the master’s degree in traf-
fic and transportation engineering from Shanghai
Maritime University in 2017. The objective of her
research is compression sensing in marine wireless
sensor network. She took part in two National
Natural Science Foundation Projects, such as the
3-D dynamic cooperative localization mechanism
of marine sensor networks based on wave shadow-
ing effect models and the study on the mechanism
of dynamic topology and temporal and spatial ran-
dom coverage of wireless sensor network in maritime search and rescue,
respectively.
JUN WANG is currently an Associate Pro-
fessor with the Department of Electrical and
Computer Engineering and the Director of the
Computer Architecture and Storage Systems Lab-
oratory, University of Central Florida, Orlando,
FL, USA. He has been serving as the Technical
Committee Chair of Computer Engineering Pro-
gram since 2012. He has held visiting profes-
sor positions at the Data Storage Institute, Singa-
pore, and ASTAR. He was a recipient of the U.S.
National Science Foundation Early Career Award 2009 and the U.S. Depart-
ment of Energy Early Career Principal Investigator Award 2005. He has
co-chaired technical programs in numerous computer systems conferences,
including the 10th IEEE International Conference on Networking, Architec-
ture, and Storage (NAS 2015), and 1st International Workshop on Storage
and I/O Virtualization, Performance, Energy, Evaluation and Dependability
(SPEED 2008) held together at HPCA. He was the General Executive Chair
at the IEEE CyberSciTech 2017. He has served as an Associate Editor
for the IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS and the
IEEE TRANSACTIONS ON CLOUD COMPUTING, and on the editorial board for the
International Journal of Parallel, Emergent and Distributed Systems.
PRASANT MOHAPATRA is currently a Professor
with the Department of Computer Science and
serves as the Dean and Vice-Provost of Graduate at
the University of California at Davis, Davis. He is
the former Endowed Chair with the Department of
Computer Science. He has held visiting professor
positions at AT&T, Intel Corporation, Panasonic
Technologies, the Institute of Infocomm Research,
Singapore, the National ICT Australia, the Univer-
sity of Padova, Italy, the Korea Advanced Institute
of Science and Technology, and Yonsei University, South Korea. He was
an Editor-in-Chief of the IEEE TRANSACTIONS ON MOBILE COMPUTING, and
has served on the editorial boards of the IEEE TRANSACTIONS ON COMPUTERS,
the IEEE TRANSACTIONS ON MOBILE COMPUTING, the IEEE TRANSACTIONS ON
PARALLEL AND DISTRIBUTED SYSTEMS, ACM WINET, and Ad Hoc Networks. He
has been on the program/organizational committees of several international
conferences.
JUNKUO CAO received the Ph.D. degree from
the Department of Computer Science and Engi-
neering, Fudan University, in 2009. He is currently
an Associate Professor with the School of Infor-
mation Science and Technology, Hainan Normal
University. His research interests include intelli-
gent information processing with a special focus
on natural language processing and Chinese infor-
mation processing. His research work has been
published in many famous conferences and jour-
nals, such as Acta Automatica Sinica, PR&AI, and AIRS. He also teaches
several undergraduate and graduate courses, including database principles
and applications, program design foundations I and II, operating systems,
and artificial intelligence and advanced database technology.
286 VOLUME 6, 2018

08039246

More Related Content

What's hot (20)

Similar to 08039246 (20)

Recently uploaded (20)

08039246