RESEARCH ARTICLE | DECEMBER 01 2017

Demonstration of acoustic source localization in air using

single pixel compressive imaging 
Jeffrey S. Rogers; Charles A. Rohde ; Matthew D. Guild; Christina J. Naify; Theodore P. Martin;
Gregory J. Orris

J. Appl. Phys. 122, 214901 (2017)

https://doi.org/10.1063/1.5003645


View Export
Online Citation

13 April 2024 13:14:24

 JOURNAL OF APPLIED PHYSICS 122, 214901 (2017)

Demonstration of acoustic source localization in air using single pixel

compressive imaging
Jeffrey S. Rogers,1 Charles A. Rohde,1 Matthew D. Guild,1 Christina J. Naify,2
Theodore P. Martin,1 and Gregory J. Orris1
1
U.S. Naval Research Laboratory, Code 7160, Washington, DC 20375, USA
2
NASA Jet Propulsion Laboratory, Pasadena, California 91109, USA
(Received 6 September 2017; accepted 9 November 2017; published online 1 December 2017)
Acoustic source localization often relies on large sensor arrays that can be electronically complex
and have large data storage requirements to process element level data. Recently, the concept of a
single-pixel-imager has garnered interest in the electromagnetics literature due to its ability to form
high quality images with a single receiver paired with shaped aperture screens that allow for the
collection of spatially orthogonal measurements. Here, we present a method for creating an
acoustic analog to the single-pixel-imager found in electromagnetics for the purpose of source
localization. Additionally, diffraction is considered to account for screen openings comparable to
the acoustic wavelength. A diffraction model is presented and incorporated into the single pixel
framework. In this paper, we explore the possibility of applying single pixel localization to
acoustic measurements. The method is experimentally validated with laboratory measurements
made in an air waveguide. https://doi.org/10.1063/1.5003645

I. INTRODUCTION transmitted, and analyzed. For a signal that is k-sparse,

meaning that there are k  MN non-zero source locations
The ability to localize acoustic and electromagnetic
being estimated, compressive sensing theory states that the
sources is a fundamental step in forming images which are
signal can be recovered with U orthogonal measurement
widely used in sonar, radar, and biomedical applications.
modes where k < U < MN.
Imaging methods have been studied extensively in order to

13 April 2024 13:14:24

improve image quality with sub-diffraction resolution and
thereby reduce sensor requirements.1–5 This type of imaging,
depicted in Fig. 1(a), is commonly achieved using a phased
detector array, often composed of many thousands transmit
and receive elements.6,7 While effective at target localiza-
tion, the large number of required detectors and point-
by-point imaging technique historically used leads to high
monetary investment, electrical complexity, and collection
of redundant spatial information. Alternatively, other imag-
ing systems require fewer sensors but rely on mechanical
scanning of the acoustic or electromagnetic field.1 These
scanning methods have the ability to perform comparably to
fully populated phased detector arrays, but can be quite time
consuming in acquiring data. Additionally, they require that
the statistics describing the acoustic or electromagnetic field
remain spatially stationary over the scanning time.
Compressive sensing has recently become an attractive
approach to source localization due to the ability to reduce
the number of measurements taken while maintaining a high
degree of accuracy.8 The rationale behind compressive sens-
ing is that most natural scenes are spatially sparse, and it is
unnecessary to scan the entire scene to localize sources since
most information received is redundant. Standard (non-com-
pressive) imaging of a scene with M  N pixels requires
taking O(MN) measurements in order to image the entire
scene.9 By using compressive sensing, U measurements may
be taken, where U is less than MN, without any loss of image
resolution as depicted in Fig. 1(b). In addition to a smaller
measurement set, the final image is inherently compressed.
This limits the amount of data which must be stored,
Although compressive sensing offers a significant advan-
tage over traditional beamforming due to the reduced file size
and fewer required measurements, a new approach called
single pixel compressive imaging has presented an even fur-
ther measurement advantage. Single pixel imaging, depicted
in Fig. 1(c), presents an approach in which a single, omnidi-
rectional detector occluded by a spatially orthogonal passive
aperture to perform compressive sensing.10–14 In the most
(a) (b)
Object Object
d d
MN Active U<MN Active
Receive Array
Receive Array
(c)
Object
d U<MN
Passive
Screens
Receive Pixel
FIG. 1. Schematic illustrating different imaging methods and their respec-
tive array geometry. (a) A M  N element array with sensor separation d
imaging an object, shown in red, emitting a signal. (b) A sparse sensing
array used in compressive imaging where the array has been reduced to U
elements and U < MN. (c) The geometry assumed in single pixel imaging
where there is a single receive element, denoted by the blue dot, and U
passive screens situated between the source and receiver.

0021-8979/2017/122(21)/214901/6/$30.00 122, 214901-1

214901-2 Rogers et al. J. Appl. Phys. 122, 214901 (2017)

straightforward implementations of this approach, originally

demonstrated in the terahertz electromagnetic frequency
range, binary opaque and transparent masks are patterned to
form orthogonal basis sets through which the spatial sampling
is generated.15–17 Beampatterns are illuminated through the
masks used to independently scan the scene. In recent years,
several additional approaches, including metamaterial-based,
frequency-dependent scanning have been demonstrated for
electromagnetic terahertz waves.11 Despite the range of
approaches and results for electromagnetic imaging, very few
studies have attempted single-pixel imaging in the acoustic
regime. In 2015, Xie et al. demonstrated a single-microphone
listening system to discriminate individual sound sources and
their locations by exploiting frequency-diverse structures,18
while in 2016 a similar technique19 was developed for sound
waves in water. Here, we consider a method for localization
that exploits the spatial diversity afforded by orthogonal
screens similar to those shown in Fig. 1(c) at a single narrow-
band frequency. The advantage to this approach is that locali-
zation and imaging can be performed at each narrowband
frequency then incoherently combined to improve the signal- FIG. 2. (top) 2D experimental geometry for the acoustic single-pixel-
to-noise ratio. imager. An object emits an acoustic signal (red) into the waveguide with a
position, h, relative to the screen’s broadside. Each screen has at least one
It is assumed in the electromagnetic literature that opening with a minimum width, 2a. A single omnidirectional receiver (blue)
screen openings are small relative to a wavelength and is depicted on the opposite side of the screen. (bottom) Experimental wave-
diffraction effects can therefore be neglected. In this manu- guide as built with height, h, and width w.
script, we present a realization of a single-pixel acoustic

13 April 2024 13:14:24

imager, which utilizes a compressive sensing technique cou-
pled to a series of spatially orthogonal acoustic screens in
order to perform source localization. To account for larger
openings, near field diffraction effects are taken considered.
A diffraction model is derived for a 2-D geometry and for-
mulated in the single pixel compressive sensing framework.
The model and its resulting inversion are verified on data
collected in an air acoustic experiment.
contiguous multiple open unit cells. For the geometry shown
in Fig. 2, to image the sound source in azimuth, the matrix A
would be comprised of the set of replica vectors representing
the phase angle /m of the radiated acoustic pressure Pm from
each mth opening in the mask given by
ImðPm Þ
tan /m ¼ : (2)
ReðPm Þ

II. ACOUSTIC SINGLE PIXEL IMAGING The equation for the pressure field in 2D space due to a rect-
For simplicity but without loss of generality, we consider angular opening of width 2a is given by20
the two dimensional waveguide depicted in Fig. 2. There is  ðmÞ  
1  2
Pm ðy; zÞ ¼ p0 ej /0 þ 2k0 f m 1þðz=f m Þ
1 
an acoustic radiating source on one side of the waveguide 2( "
and a single omnidirectional receiver element on the opposite rffiffiffiffiffiffiffiffi #
zR ym
side. In between the acoustic source and receiver is an aper-  erf j 1 þ
ture that allows for U different masks to be inserted. fm a
"rffiffiffiffiffiffiffiffi #)
The signal model used for the geometry described in zR y
Fig. 2 assumes a one dimensional mask and acoustic locali- þ erf j 1  m ; (3)
fm a
zation in azimuth given by the angle, h. Localization is per-
formed at a single narrowband frequency. Assuming that the ðmÞ
masks have small openings relative to a wavelength, the where p0 is the source pressure, /0 denotes the phase shift
received data, yU1 is a frequency domain vector given by for the mth opening, k0 is the wavenumber in the host fluid,
zR ¼ k0 a2 =2 is the Rayleigh distance, erf denotes the error
y ¼ Ax; (1) function, and
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where the sensing matrix A ¼ WU is the product of the f m ¼ y2 þ z2 ; (4)
m
matrix, ULQ (responsible for transforming the signal xQ1
into the spatial frequency domain) and the matrix WUL (rep- ym ¼ y  ym  rm sin hq ; (5)
resenting the measurement process in which U orthogonal qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
masks, containing L unit cells, are chosen). rm ¼ ðy  ym Þ2 þ z2 : (6)
The sensing matrix A is assumed to be known a priori
and can be precomputed. In this case, A must account for large Here, ym is the center of the mth opening, rm is the distance
openings compared to a wavelength due to the possibility of from the mth opening to the receive pixel, and hq 2 [–90 , 90 ]
214901-3 Rogers et al.
is the azimuthal direction discretized into Q angles. These
quantities are depicted in Fig. 2 along with the frame of refer-
ence for the coordinate system. It should be noted that the
expression given in Eq. (3) is based on a Fresnel expansion and
is only valid for small y – ym and at larger values of rm. The
propagation delay from the sound source to each opening in the
mask is given by
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðmÞ
/0 ðhq Þ ¼ k0 r 2 þ y2m  2ym r sin ðhq Þ; (7)
where r is the distance from the acoustic source to the center
of the mask.
Diffraction around the openings in the mask is
accounted for by modeling the total pressure field radiating
from each rectangular opening. This total acoustic pressure
field is illustrated in Fig. 3 and can be calculated by summing
the contributions from each opening in the mask by
FIG. 3. (a) Possible screen geometries forming W for a 16 unit cell basis.
Each row is a 1-D screen, used in the geometry described in Fig. 2. Black
squares block sound propagation and white squares are open. The first ten
masks, highlighted in yellow, were chosen for the experimental measure-
ments. The masks outlined in red were used to perform the inversion and
generate the results shown in Fig. 4. (b) Full field experimental measurement
showing diffraction pattern from mask #3 for a far-field source located at
–2 . (c) Full field diffraction pattern computed using the analytical model,
Eqs. (3)–(8), for mask #3 and a source located at –2 .
J. Appl. Phys. 122, 214901 (2017)
X
M
Ptot ðy; z; tÞ ¼ ejxt Pm ðy; zÞ: (8)
m¼1
A vector, au ¼ ½Ptot ðyr ; zr Þjh1 …Ptot ðyr ; zr ÞjhQ , is computed
for the uth mask by evaluating the above expression at the
single pixel receiver location, (yr, zr), and for a grid of
hypothesized source directions, [h1…hQ]. The matrix A is
formed by stacking the vectors corresponding to the U masks
used or AUQ ¼ ½a1 ; …; aU . It should be noted that in this
experimental setup, the acoustic source was several aperture
lengths away from the mask and the expression in Eq. (7)
could be simplified to a plane-wave model that is only
dependent on hq. For cases when acoustic sources are closer
to the nearfield of the mask, 2-D localization is possible by
computing A as a function of the source angle, hq, and its
range, r, then performing the inversion.
A severely underdetermined set of equations must be
solved to obtain x since it is assumed that U  Q. However,
if one can assume that the data are k-sparse (meaning that x
has k nonzero elements) and k < U the solution to Eq. (1) can
be obtained with a ‘1 minimization21
x^ ¼ minjjxjj1 subject to Ax ¼ y; (9)
P
where jjxjj1 denotes the ‘1 norm ( i jAxi  yi j) of x and x^ is
the solution of the unknown x.
13 April 2024 13:14:24
In order to guarantee exact recovery of x, the sensing
matrix A must satisfy the restricted isometry property (RIP),
which can be expressed as22
ð1  dk Þjjxjj2‘2  jjAxjj2‘2  ð1 þ dk Þjjxjj2‘2 ; (10)
where dk is the restricted isometry constant and jj  jj‘2 repre-
sents the Euclidean norm. The sensing matrix A satisfies the
RIP for values of dk not too close to 1. In order to achieve
this condition with a compressed sinusoidal basis given by
U; W can be a matrix of transmission “spikes” or 00 s and 10 s.
Here, W is chosen to be a set of orthogonal “spike combs”
given by a 2-Bit quantized Discrete Cosine Transform
(DCT) where the kernel is given by
(
0 cos ððu  1Þp‘=LÞ < 0
wðuÞ‘ ¼ (11)
1 cos ððu  1Þp‘=LÞ 0;
for u ¼ 1, 2,…, U masks and ‘ ¼ 1, 2,…, L unit cells. Figure
3(a) depicts the DCT kernel where each row represents one
of 16 possible masks from (11) and each column is a unit
cell. If the unit cell is black, it represents a 0 or no transmis-
sion through the unit cell and a white unit cell is a 1 or per-
fect transmission through the unit cell.
It has recently been shown that computing the RIP
directly for a matrix, A, and a given number of sources, k, is
NP-hard.23 However, for a fixed k, one can Monte Carlo
over source positions and compute a Statistical Restricted
Isometry Property (StRIP).24 By examining the probability
distribution function for the RIP parameter and evaluating it
for a 95% confidence interval, we can obtain an StRIP for
the set of masks shown in Fig. 3 of dk95 ¼ 0:79. Thus, in the
214901-4 Rogers et al. J. Appl. Phys. 122, 214901 (2017)

statistical sense, the A matrix given in Eq. (1) satisfies dk < 1 time averaging, the data were transformed to the frequency
for our choice of masks. domain via a FFT.
The full field plots in Fig. 3 were taken for two different
III. RESULTS source positions for all 10 masks to analyze the error as a
function of receiver position. The back of the waveguide was
To experimentally test our predicted results, a wave-
removed and the receiver microphone was mounted on to a
guide was constructed out of acrylic sheets and sealed on all
computer controlled x-y stage (Velmex VXM) with an
sides with RTV silicone. The waveguide geometry is illus-
extension rod. Time series waveforms were collected
trated in Fig. 2. Using a spacer layer, the waveguide height,
throughout the waveguide transmission domain and proc-
h ¼ 9 mm, was selected to be less than a half-wavelength at
essed to retrieve the instantaneous acoustic amplitude and
the design frequency of 10 kHz (k 34.3 mm). This main-
phase at equally spaced points (Dx ¼ Dy ¼ 5 mm). For the
tained a transverse planewave mode in the waveguide, simu-
full field analysis, the amplitude and phase of the acoustic
lating an infinite 2D system. Sound absorbing foam was
wave were extracted from the collected time series data with
lined along the waveguide interior to minimize edge reflec-
a fast Fourier transform of a time windowed, linear
tions and further approximate an infinite 2D system.
An open slot was left at the waveguide center, to allow
the insertion of laser cut blocking screens. The masks are
designed to have 16 unit cells each having a width of k/4, and
the binary screen patterns used are those highlighted in Fig.
3(a). Due to the masks consisting of multiple adjacent open
unit cells, diffraction effects can be observed in the measured
pressure field as observed in Fig. 3(b). These diffraction
effects can be accounted for using the model described by
Eq. (3). Qualitatively, the model shows agreement with
experimental data as depicted by comparing Fig. 3(c) with
Fig. 3(b). It should be noted that although masks were created
for a single frequency and manually exchanged in this experi-

13 April 2024 13:14:24

ment, this approach can be generalized to a broadband source
yielding several measurement modes with a single screen
similar to what has been done in electromagnetics.11
The indicated receiver in Fig. 2 was a Br€uel and Kjær
4939-A-011 microphone placed on the center axis of the
waveguide, 530 mm away from the aperture screen location.
This receiving microphone was amplified by a Br€uel and
Kjær 2690-0S4 pre-amplifier set to 1 V/Pa.
The source, a Dayton Audio 32A-8 speaker, was
mounted 530 mm away from the aperture screen, on a plate
with perpendicular translating motion, ruled in 1 mm incre-
ments. The source speaker was driven by an Agilent 33220A
function generator, using a programmed 1 ms duration 10 kHz
center frequency pulse, windowed with a Hanning envelope
function. The source was placed at seven distinct loca-
tions: 615 mm, 635 mm, 655 mm, and 675 mm above (þ)
and below (–) the central wave guide longitudinal axis, to cre-
ate the respective set of 8 azimuthal bearings, 6 2 , 64 ,
66 , 68 .
The source pulse and collection timing were synchro-
nized to allow averaging of the collected signals as well as
arrival timing data for computing the full acoustic field plots.
The synchronization was accomplished with a TTL signal
generated by the Agilent 33220 function generator. This
signal was used to initiate the data collection. The data
collection DAQ (National Instruments cRIO-9039 real time
controller) was an NI-9223 module with a 1 Msample/s data
rate. The collected waveforms were, for each of the ten aper-
tures, time windowed with a 6 ms Tukey window to mini-
mize edge reflection noise, and 512 time series samples were
FIG. 4. Retrieved x^, the direction depended power vector, for the eight
averaged together for each aperture-type/source-location source positions [(a)–(h)] 8 , 6 , 4 , 2 , –2 , –4 , –6 , –8 . With (red,
combination to maximize the signal to noise ratio. Following dashed) measured position indicated.
214901-5 Rogers et al. J. Appl. Phys. 122, 214901 (2017)

frequency modulated pulse. Note that the masks are centered
at 0 mm along the Y axis and 590 mm along the Z axis.
IV. DISCUSSION
The retrieved ‘1 normalization results following Eq. (9)
for all tested azimuths are shown in Fig. 4. A total of only
four out of the ten highlighted masks were needed to suc-
cessfully invert for the eight different source positions. The
masks used are outlined in red in Fig. 3. Note that the four
masks used to generate the results described here are not
unique and the method performs similarly for any other con-
figuration using the same number of randomly selected
masks. Plots in Fig. 4 show x^ðhÞ, the retrieved power, in dec-
ibels, as a function of angular position, h. For all measured
source positions, the maximal points in Fig. 4 correspond to
the known source positions (the red dashed line). This
reflects our successful determination of source position with
a highly under-sampled set of possible k-vectors.
The additional peaks that appear in the inversions are
due to grating lobes (or grating orders) caused by the high
degree of periodic order inherent to the DCT basis chosen to
construct the masks. For example, in Fig. 3(b), grating lobes
are observed due to the large separation between openings in
the middle of the mask. These grating lobes result in spuri-
ous peaks in the inversion output and may obscure the source
being localized as observed in Fig. 4(h). Here, a grating lobe
caused a spurious peak near positive endfire having roughly
the same magnitude as the peak corresponding to the true
source direction.
In our experimental design, the grating lobes are a result
of lengthening the aperture to obtain a higher resolution out-
put. At the receiver, the grating lobes only appear at angles
outside the physical boundary of the waveguide. Therefore,
when determining a source position, we only need to con-
sider the maximum amplitude between 68 . To better illus-
trate the localization performance, an error surface is shown
in Fig. 5. The surface represents the localization error at dif-
ferent single point receiver locations within the waveguide.
The error is computed by taking the absolute value of the
difference between the true source bearing and the bearing
corresponding to the maximum amplitude between 68 for
source positions of 2 and –6 . The figure illustrates the
effect of the grating lobes and nearfield diffraction on locali-
zation performance. At distances very close to the mask,
there is a mismatch in the diffraction model used in Eqs.
(3)–(8). The structured peaks emanating radially from the
mask in the error surface illustrate the ambiguity caused by
grating lobes from the masks. However, the localization
error is minimal at positions sufficiently far from the mask
and near the center of the waveguide (to minimize imperfect
side wall absorption).

13 April 2024 13:14:24

(a) 12

-60
10
-40

8
-20
Y (mm)

0 6

20
4

40
2
FIG. 5. Target localization error surfa-
60 ces measured in degrees plotted in the
0 ‘z’ axis. The error is computed by tak-
50 100 150 200 250 300 350 400 450 500 550 Deg. ing the absolute value of the difference
Z (mm) between the localization estimate max-
imum over a grid of receiver positions
(b) 12 within the waveguide for a source
-60
position of (a) 2 and (b) –6 . The
position and size of the single pixel
10
receiver are indicated with the gray
-40
dot.
8
-20
Y (mm)

0 6

20
4

40
2
60
0
50 100 150 200 250 300 350 400 450 500 550 Deg.
Z (mm)
214901-6 Rogers et al.
V. CONCLUSIONS
In this work, we have presented our application of com-
pressive imaging to build a single pixel acoustic camera. We
have used this technique to determine the azimuthal position
of an acoustic source, using an omni-directional receiver and
a set of analog apertures in a simplified 2D waveguide. The
experimental results have shown an under-sampled set of
four orthogonal aperture screens allowed the determination
of an acoustic source position with an accuracy of 61 for a
large portion of the acoustic waveguide. This research lays
the ground work for building more complex single pixel
imaging systems for sampling sparse acoustic targets.
ACKNOWLEDGMENTS
This work was supported by ONR.
1 (2011).
S. Zhang, L. Yin, and N. Fang, “Focusing ultrasound with an acoustic
metamaterial network,” Phys. Rev. Lett. 102(19), 194301 (2009).
2 Cummer, “Single-sensor multispeaker listening with acoustic meta-
J. Zhu, J. Christensen, J. Jung, L. Martin-Moreno, X. Yin, L. Fok, X.
Zhang, and F. Garcia-Vidal, “A holey-structured metamaterial for acoustic
deep-subwavelength imaging,” Nat. Phys. 7(1), 52–55 (2011).
3 ing estimation based on compressed sensing,” in Proceedings of the IEEE/
V. M. Garcıa-Chocano, J. Christensen, and J. Sanchez-Dehesa, “Negative
refraction and energy funneling by hyperbolic materials: An experimental
demonstration in acoustics,” Phys. Rev. Lett. 112(14), 144301 (2014).
4
N. Fang, H. Lee, C. Sun, and X. Zhang, “Sub–diffraction-limited optical
imaging with a silver superlens,” Science 308(5721), 534–537 (2005).
5 21
D. Schurig and D. Smith, “Sub-diffraction imaging with compensating
bilayers,” New J. Phys. 7(1), 162 (2005).
6
T. S. Brandes and R. H. Benson, “Sound source imaging of low-flying air-
borne targets with an acoustic camera array,” Appl. Acoust. 68(7),
752–765 (2007).
7 23
J. H. Ender, “On compressive sensing applied to radar,” Signal Process.
90(5), 1402–1414 (2010).
8
E. J. Candès and M. B. Wakin, “An introduction to compressive
sampling,” IEEE Signal Process. Mag. 25(2), 21–30 (2008).
9 24
S. Qaisar, R. M. Bilal, W. Iqbal, M. Naureen, and S. Lee, “Compressive
sensing: From theory to applications, a survey,” J. Commun. Networks 15,
443–456 (2013).
J. Appl. Phys. 122, 214901 (2017)
10
M.-J. Sun, M. P. Edgar, G. M. Gibson, B. Sun, N. Radwell, R. Lamb, and
M. J. Padgett, “Single-pixel three-dimensional imaging with time-based
depth resolution,” Nat. Commun. 7, 12010 (2015).
11
J. Hunt, T. Driscoll, A. Mrozack, G. Lipworth, M. Reynolds, D. Brady,
and D. R. Smith, “Metamaterial apertures for computational imaging,”
Science 339(6117), 310–313 (2013).
12
M. F. Duarte and Y. C. Eldar, “Structured compressed sensing: From the-
ory to applications,” IEEE Trans. Signal Process. 59, 4053–4085 (2011).
13
R. F. Marcia, Z. T. Harmany, and R. M. Willett, “Compressive coded
aperture imaging,” Proc. SPIE 7246, 72460G (2009).
14
A. Wagadarikar, R. John, R. Willett, and D. Brady, “Single disperser
design for coded aperture snapshot spectral imaging,” Appl. Opt. 47,
B44–B51 (2008).
15
W. L. Chan, K. Charan, D. Takhar, K. F. Kelly, R. G. Baraniuk, and D. M.
Mittleman, “A single-pixel terahertz imaging system based on compressed
sensing,” Appl. Phys. Lett. 93(12), 121105 (2008).
16
A. Liutkus, D. Martina, S. Popoff, G. Chardon, O. Katz, G. Lerosey, S.
Gigan, L. Daudet, and I. Carron, “Imaging with nature: Compressive
imaging using a multiply scattering medium,” Sci. Rep. 4, 5552 (2014).
17
R. M. Willett, R. F. Marcia, and J. M. Nichols, “Compressed sensing for
practical optical imaging systems: A tutorial,” Opt. Eng. 50(7), 072601
18
Y. Xie, T.-H. Tsai, A. Konneker, B.-I. Popa, D. J. Brady, and S. A.
materials,” Proc. Natl. Acad. Sci. U.S.A. 112(34), 10595–10598 (2015).
19
B. Gao, D. Gao, H. Wang, and N. Wang, “A new method of passive bear-
OES China Ocean Acoustics (COA), January (2016), pp. 1–4.
20
M. D. Guild, C. J. Naify, T. P. Martin, C. A. Rohde, and G. J. Orris,
“Superresolution through the topological shaping of sound with an acous-
tic vortex wave antenna,” preprint arXiv:1608.01887.
E. J. Candès et al., “Compressive sampling,” in Proceedings of the
International Congress Mathematicians (Madrid, Spain, 2006), Vol. 3, pp.
13 April 2024 13:14:24
1433–1452.
22
E. J. Candes and T. Tao, “Decoding by linear programming,” IEEE Trans.
Inf. Theory 51(12), 4203–4215 (2005).
A. M. Tillmann and M. E. Pfetsch, “The computational complexity of the
restricted isometry property, the nullspace property, and related concepts
in compressed sensing,” IEEE Trans. Inf. Theory 60(2), 1248–1259
(2014).
R. Calderbank, S. Howard, and S. Jafarpour, “Construction of a large class
of deterministic sensing matrices that satisfy a statistical isometry proper-
ty,” IEEE J. Sel. Top. Signal Process. 4(2), 358–374 (2010).