J Optlaseng 2020 106301
J Optlaseng 2020 106301
J Optlaseng 2020 106301
a r t i c l e i n f o a b s t r a c t
Keywords: Single-pixel imaging is a novel computational imaging scheme. It allows one to use a spatially unresolvable
Computational imaging photodetector to acquire the spatial information of objects and reconstruct images by a computational means.
Single-pixel imaging The core of single-pixel imaging is encoding the spatial information of objects into one-dimensional light signals
Image reconstruction
by using spatial light modulation and subsequently decoding the information to reconstruct the image. Such
Performance comparison
a novel imaging scheme has a potential of tacking some challenges in conventional imaging, such as imaging
with non-visible light, imaging under weak-light conditions, imaging through turbid media, etc. However, the
practicality of single-pixel imaging is limited by its imaging quality and efficiency. Thus, how to improve the
imaging quality and reduce the imaging time in single-pixel imaging is extensively explored in the field. In
this paper, we comprehensively investigate five representative and widely used single-pixel imaging methods –
computational ghost imaging, compressive sensing ghost imaging, Hadamard single-pixel imaging, Fourier single-
pixel imaging, and binary Fourier single-pixel imaging. We review the principle of these methods and compare
the performance of these methods by numerical simulations and experiments. The comparison not only reveals
the connections and differences among these methods, but also shows the advantages and disadvantages of their
own. This paper also discusses the problems that remain to be solved in single-pixel imaging and prospects for
the future development.
∗
Corresponding author.
E-mail addresses: [email protected], [email protected] (Z. Zhang).
https://doi.org/10.1016/j.optlaseng.2020.106301
Received 20 March 2020; Received in revised form 25 May 2020; Accepted 3 July 2020
0143-8166/© 2020 Elsevier Ltd. All rights reserved.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
intensity of the light modulated by the object. The advanced GI with the the advanced algorithms (such as, compressive sensing) deployed in the
use of an SLM to generate programmable speckle patterns is proposed image reconstruction process might be computationally exhausted, re-
and called computational ghost imaging (CGI) [4]. sulting in a time-consuming image reconstruction. How to improve the
Image sensors are typically silicon-based, therefore, can work at visi- image quality and reduce the image acquisition time is extensively ex-
ble waveband. In comparison with image sensors, single-pixel detectors plored in the field. We note that different sampling strategies and image
are much simpler and cheaper. Due to the miniaturization and broad reconstruction algorithms might affect the achievable image quality and
spectral range, single-pixel detectors are widely applicable to different efficiency significantly. And each strategy or algorithm has its own ad-
imaging schemes in various spectral regions. In addition, single-pixel de- vantages, limitation, and suitable applications. Thus, it is necessary to
tectors have higher quantum efficiency and a higher filling factor than take the representative and most widely used single-pixel imaging meth-
image sensors. Given such advantages, single-pixel imaging has a po- ods into comparison and look into their commons and differences. It is
tential of tackling some challenges in conventional imaging, such as also necessary to investigate their performance from various aspects and
imaging under weak-light conditions, imaging through turbid media, show what applications each method is suitable for.
etc. Particularly, single-pixel imaging is suitable for the circumstances In this paper, we comprehensively investigate five single-pixel imag-
where pixelated detectors are expensive or unavailable (such as, in- ing methods, including CGI [3,4,22], compressive sensing ghost imag-
frared [5], deep ultraviolet, X-ray [6,7], or terahertz [8-12]). For exam- ing (CSGI) [23-29], Hadamard single-pixel imaging (HSI) [16,30-35],
ple, Zhang et al. successfully realized X-ray single-pixel imaging with a Fourier single-pixel imaging (FSI) [36-40], and binary Fourier single-
much higher contrast-to-noise ratio compared to projection X-ray imag- pixel imaging (BFSI) [41,42]. We review the principle of these methods
ing at the same low-radiation dose. X-ray imaging with low-radiation and introduce the implementation of the methods in detail. We also
dose can avoid the object to be imaged being severely damaged by high compare the performance of these methods through numerical simula-
photon energy [7]. Stantchev et al. achieved real-time terahertz imaging tions and experiments. The comparison not only reveals the connections
by using a single-pixel detector [12]. In addition, single-pixel imaging and differences among these methods, but also shows the advantages
has contributed many novel ideas in optics. For example, Peng et al. and disadvantages of their own. This paper also discusses the problems
proposed a light-field microscope based on single-pixel imaging. The remained to solve in single-pixel imaging and the prospects for future
proposed microscopy can achieve refocusing without any mechanical development.
motion [13]. Zhang et al. reported an ultra-thin and multi-functional
single-pixel scanner. By using a solar panel as the single-pixel detec-
2. Principle
tor, the scanner can be as thin as only 2.48 mm [14]. Zhang et al. also
proposed a single-pixel imaging method for photovoltaic devices char-
Spatial information modulation in single-pixel imaging is referred to
acterization. The method is able to simultaneously acquire the structural
generating a sequence of structured patterns by using an SLM and using
image and the functional image of a single or multiple photovoltaic de-
the structured patterns to encode the spatial information of objects into
vices in an efficient way, which benefits in-line inspection [15]. More-
1-D light signals. Specifically, structured patterns are displayed on the
over, Zhang et al. reported a secured single-pixel broadcast imaging
SLM and the patterns are projected onto the object to be imaged through
method which can encrypt the image before it is captured [16]. Such a
a lens system. The intensity of resulting light (transmitted, reflected or
scheme can avoid information leakage in the conventional encrypt-after-
fluorescence) is collected by a single-pixel detector. The relationship
capture scheme before the information is encrypted. The novel scheme
among the reflectivity (reflected object assumed), the structured pat-
of spatial information acquisition inspired by single-pixel imaging has
tern, and the single-pixel measurement can be depicted as Eq. (1),
also been applied to a few non-imaging applications, such as real-time
fast-moving object tracking [17], real-time fast-moving object classifi- 𝑀 ∑
∑ 𝑁
cation [18], distributed fiber-optic sensing [19], and etc. Ota et al. pro- 𝐷𝑖 = 𝛼 𝑂(𝑥, 𝑦) ⋅ 𝑃𝑖 (𝑥, 𝑦), (1)
posed a cytometry based on single-pixel imaging [20]. The proposed 𝑥=1 𝑦=1
cytometry is able to separate morphologically similar cell types in an where (x, y) is the spatial coordinate, O denotes object reflectivity, Pi is
ultrahigh-speed fluorescence imaging–activated cell sorter. Zhang et al. the i th structured pattern, Di is the i th single-pixel measurement cor-
successfully demonstrated a single-pixel imaging inspired object track- responding to Pi , 𝛼 is a constant factor depending on the optoelectronic
ing method with a tracking frame rate up to 1666 frames per second response of the utilized single-pixel detector, and · denotes a point-wise
[21]. This method allows for tracking a high-speed moving object in multiplication. In Eq. (1), O(x, y) and Pi (x, y) are expressed in a discrete
real time and for a long duration. form. The size of both the object and the pattern is M × N pixels. With
However, the practicality of single-pixel imaging is limited by its the knowledge of the structured patterns {Pi }, the image reconstruction
imaging quality and imaging efficiency. The imaging quality of the algorithms of single-pixel imaging are able to recover the object image
cutting-edge single-pixel imaging methods, in terms of spatial resolution I from the single-pixel measurements {Di }. The reconstructed image I
and SNR, can hardly be compared with that of the conventional imag- is proportional to object reflectivity O. Thus, the relationship among
ing method. The achievable spatial resolution in single-pixel imaging the single-pixel measurements, the structured patterns, and the recon-
is determined by the resolution of the structured patterns generated by structed object image can be expressed as,
the utilized SLM. Currently, the element pitch of the cutting-edge SLMs
𝑀 ∑
∑ 𝑁
is at a several-micron level. The SNR of the images reconstructed by
single-pixel imaging is generally lower than that of conventional imag- 𝐷𝑖 = 𝐼 (𝑥, 𝑦) ⋅ 𝑃𝑖 (𝑥, 𝑦). (2)
𝑥=1 𝑦=1
ing, evident by that the images reconstructed by single-pixel imaging
are generally noisy especially when the images are acquired using an The size of the reconstructed image is M × N pixels. Eq. (2) implies
SLM that operates at a high refreshing rate. Apart from imaging quality, that each single-pixel measurement is an inner product of the object
imaging efficiency is also an issue that limits the practicality of single- image and the structured pattern. The inner product depicts how similar
pixel imaging. Although single-pixel imaging removes the need for pix- the object image I and the structured pattern Pi are.
elated detectors, the spatial information acquisition process is at the As Fig. 1 shows, spatial light modulation in single-pixel imaging can
expense of time. As a single-pixel detector is with only one pixel avail- be implemented in two different schemes [43], that is, the structured
able, the data throughput of the detector is very limited. As a result, in illumination scheme (Fig. 1(a)) and the structured detection scheme
the case without available prior knowledge, single-pixel imaging has to (Fig. 1(b)). For the structured illumination scheme, the SLM is placed
take a huge number of measurements to collect sufficient spatial infor- before the object and, therefore, this scheme is also termed front mod-
mation for recovering an image with fine details. On the other hand, ulation. The structured patterns displayed on the SLM are projected to
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 1. Two different spatial light modulation schemes in single-pixel imaging. (a) Structured illumination scheme, where the SLM is placed before the object and
used to modulate the illumination light field; (b) structured detection scheme, where the SLM is placed after the object and used to modulate the detection light
field.
the target object through a lens and structured illumination is gener- as to compress sparse signals in an efficient manner. Sparse represen-
ated. The back-scattered light is then collected by the single-pixel detec- tation means that a signal of length M × N can be represented with
tor. For the images reconstructed in the structured illumination scheme, only n ≪ M × N non-zero coefficients. By storing only the values and
the field of view is determined by the illumination area and the shad- locations of the non-zero coefficients, a compressed representation of
ing profile is determined by the position of the single-pixel detector the signal can be derived. In CS, however, the signal recovery is more
with respect to the object. Such an interesting phenomena is subject to involved, typically achieved using convex-optimization-based recovery
Helmholtz reciprocity [44] and also known as dual photography [45]. methods. Adopting CS in single-pixel imaging, CSGI is a framework that
For the structured detection scheme, the SLM is placed after the object allows for recovering an image with fewer single-pixel measurements
and, therefore, this scheme is also termed back modulation. The light than the pixels of the image, which is expected to tackle the low data
field to be detected by the single-pixel detector is modulated by the efficiency problem of CGI. Specifically, with n structured patterns and
SLM. For the images reconstructed in the structured detection scheme, n single-pixel measurements provided, CSGI attempts to recover the ob-
the field of view is determined by the detection components (that is, the ject image I of size M × N. From the mathematical perspective, M × N
SLM and the single-pixel detector) and the shading profile is determined unknowns are attempted to be determined by solving a set of n equa-
by the illumination source. Besides, the difference of the point spread tions, which is an under-determined problem, because the number of
function in the two schemes is comprehensively discussed in [46]. equations, n, is less than that of unknowns, M × N. To solve this prob-
The differences among the single-pixel imaging methods mainly fall lem, CSGI exploits the sparsity of natural images, which can effectively
into three categories: 1) the type of structured patterns used in spatial shrink the solution space. The sparsity of natural images is referred to
light modulation, 2) the sampling strategy (specifically, the ordering of that the gradient’s integral of a natural image is statistically low. The
structured patterns), and 3) the image reconstruction algorithm. sparsity, G, of image I can be expressed as
𝑀 ∑
∑ 𝑁
{ }
2.1. Computational ghost imaging 𝐺 (𝐼 ) = [𝐼 (𝑥, 𝑦) − 𝐼 (𝑥 − 1, 𝑦)]2 + [𝐼 (𝑥, 𝑦) − 𝐼 (𝑥, 𝑦 − 1)]2 , (4)
𝑥=2 𝑦=2
CGI uses random patterns for spatial light modulation. Typically, the
where I(x, y) is a natural image. CSGI allows using arbitrary struc-
random patterns are binary so that CGI can take the advantage of high-
tured patterns for spatial light modulation, such as random patterns,
speed binary pattern generation given by a digital micro-mirror device
Hadamard basis patterns, Fourier basis patterns, discrete cosine basis
(DMD). As Eq. (3) shows, CGI uses a correlation-based algorithm for
patterns, wavelet basis patterns patterns, etc. Random patterns are a
image reconstruction and the algorithm is iterative.
typical choice. To reconstruct the object image, CSGI attempts to search
𝑛
( 𝑛
)( 𝑛
)
1∑ 1∑ 1∑ for a candidate of I(x, y) which can minimize G in Eq. (4). To do so,
𝐼= 𝐷𝑛 − 𝐷 𝑃𝑛 − 𝑃 , (3) different algorithms have been proposed, such as l1 -magic [25], orthog-
𝑛 𝑖=1 𝑛 𝑖=1 𝑖 𝑛 𝑖=1 𝑖
onal matching [26], augmented Lagrangian an alternating direction al-
where n is the number of structured patterns (or single-pixel measure- gorithm (TVAL3) [27], and NESTA [28]. Most of the algorithms are
ments). We note that the spatial coordinate (x, y) is omitted in Eq. (3) for iterative and based on convex optimization. Here, we focus on TVAL3
simplicity. The equation implies that the reconstructed image I is equiv- as it is popularly used. TVAL3 solves the under-determined problem by
alent to the weighted sum of the structured patterns. The weights are re- using matrix operations. As Fig. 2 shows, TVAL3 models the problem in
ferred to the single-pixel measurements. As is based on a statistic model, a matrix form, mathematically expressed as
CGI typically requires far more single-pixel measurements than image
pixels (that is, n ≫ M × N) to reproduce a high-quality image. For exam- 𝐏 × 𝐈 = 𝐃, (5)
ple, Sun et al. reconstructed a 128 × 96-pixel image with 106 single-pixel where P denotes the structured patterns matrix with n rows and M × N
measurements [47]. columns, I is the object image flatten to be a 1-D vector with M × N el-
ements, and D is the single-pixel measurements vector with n elements.
2.2. Compressive sensing ghost imaging Note that the i th element in D is the single-pixel measurement corre-
sponding to the i th structured pattern Pi (that is, the i th row of matrix
In order to reduce the number of single-pixel measurements, a frame- P). As Fig. 2 illustrates, matrix P required by TVAL3 consists of all struc-
work based on compressive sensing (CS) [23] – CSGI – was proposed for tured patterns used in spatial light modulation. The principle of TVAL3
single-pixel imaging. CS is a powerful tool to recover sparse signals from is detailed in [27]. It can be seen that the main difference between CGI
a small number of random measurements of the original signal, as well and CSGI in principle is the image reconstruction algorithm.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 2. Illustration of (a) the matrix operations and (b) matrix flattening in CSGI.
2.3. Hadamard single-pixel imaging coefficients can be acquired at a higher priority. The zig-zag, the square,
and the circular sampling paths used in HSI are investigated in [31].
Except for random patterns, orthogonal patterns can be used as struc- Sampling path is also referred to pattern ordering [49,50]. There have
tured patterns for spatial light modulation in single-pixel imaging. Dif- been several pattern ordering algorithms proposed for HSI, such as Rus-
ferent from CGI, basis scan single-pixel imaging methods are based on sian doll [30], cake-cutting [33] and origami [34]. A zig-zag sampling
a deterministic model, using the basis patterns of some transformation path is shown in Fig. 3, for example. A Hadamard basis pattern 𝑃+ (𝑥, 𝑦)
for spatial light modulation. The patterns include 2-D Hadamard trans- can be obtained by applying an inverse Hadamard transform to a Dirac
form basis patterns, 2-D Fourier transform basis patterns, 2-D discrete delta function 𝛿 H (u, v),
cosine transform basis patterns, etc. There are mainly two advantages of 1[ { }]
using basis patterns. First, basis patterns are orthogonal and therefore 𝑃+ (𝑥, 𝑦) = 1 + 𝐻 −1 𝛿H (𝑢, 𝑣) , (6)
2
can reduce redundancy in data acquisition. As such, one can reconstruct where (u,v) is the coordinate in the Hadamard domain representing the
a clear image from a small number of measurements. Second, as any number of sign changes along x or y direction respectively, and 𝐻 −1 {}
image can be composed by a complete set of orthogonal basis patterns, denotes an inverse Hadamard transform and
basis scan single-pixel imaging allows for a perfect (lossless) reconstruc- {
1, 𝑢 = 𝑢0 , 𝑣 = 𝑣0
tion of an image under a noiseless situation. In basis scan single-pixel 𝛿H (𝑢, 𝑣) = . (7)
0, otherwise
imaging methods, the single-pixel measurements are equivalent to the
coefficients of the object image in the transform domain associated with The sample Hadamard basis pattern 𝑃+ shown in Fig. 3 is derived by
the utilized basis patterns. Basis scan is referred to acquiring the coef- substituting 𝑢 = 4 and 𝑣 = 3 into Eq. (6). Thus, it has 4 sign changes each
ficients in the transformation domain one by one in an order indicated row and 3 sign changes each column. Finally, a 2-D inverse Hadamard
by a specific path, also known as a sampling path. Finally, the image is transform is applied to the acquired Hadamard transform to reconstruct
reconstructed by an inverse transform. Both HSI and FSI belong to basis the object image. The mathematical details can be referred to [31].
scan single-pixel imaging. Differential HSI is an embodiment of HSI, allowing each Hadamard
HSI is based on Hadamard transform [48]. As Fig. 3 shows, coefficient H(u, v) to be acquired in a manner of differential mea-
Hadamard transform basis patterns are binary, consisting of white-and- surement. Differential HSI is conducive to suppression of noise. As il-
black blocks. HSI acquires the Hadamard transform of the object image lustrated in Fig. 3, to acquire a coefficient H(u, v), differential HSI
and reconstructs the object image by applying an inverse Hadamard takes two single-pixel measurements. The one is acquired by using
transform. The Hadamard transform of an image is actually a complete a Hadamard basis pattern 𝑃+ (𝑥, 𝑦) and the other by using its inverse
set of Hadamard coefficients. Each coefficient is the weight correspond- 𝑃− (𝑥, 𝑦) = 1 − 𝑃+ (𝑥, 𝑦). The coefficient H(u, v) is calculated by the differ-
ing to a unique Hadamard basis pattern. One can follow a sampling path ence of the two corresponding single-pixel measurements
to acquire the coefficients in a specific order so that the more important 𝐻 (𝑢, 𝑣) = 𝐷+ − 𝐷− , (8)
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
where 𝐷+ and 𝐷− are measurements corresponding to the structured coefficients is the same as that of image pixels (M × N). One can follow
patterns 𝑃+ (𝑥, 𝑦) and 𝑃− (𝑥, 𝑦), respectively. Hadamard coefficients are a sampling path to acquire the coefficients in a specific order so that
real-valued and the number of Hadamard coefficients is the same as that the more important coefficients can be acquired at a higher priority.
of image pixels. Fully sampling an N × N-pixel image using differential The circular, the diamond, and the spiral sampling paths used in FSI are
HSI takes 2N2 measurements. However, HSI is limited by that N must investigated in [31]. A circular sampling path is shown in Fig. 4, for ex-
be an integer and N, N/12 or N/20 must be a power of 2. ample. With the prior knowledge that the Fourier transform of any real-
There is another strategy to generate Hadamard basis patterns, valued image is conjugated symmetric, fully sampling an M × N-pixel
which is based on the Hadamard matrix. A Hadamard matrix H is like image using four-step phase-shifting FSI takes 2 × M × N measurements.
the structured pattern matrix P in CSGI shown in Fig. 2. The Hadamard It can be seen that the four-step phase-shifting FSI is essentially a differ-
matrix has size (N × N) × (N × N). The matrix consists of N × N Hadamard ential method of measurement. Finally, a 2-D inverse Fourier transform
basis patterns and the size of each basis pattern is N × N. To obtain a is applied to the acquired Fourier transform to reconstruct the object im-
Hadamard basis pattern from the matrix, one can extract a row or a col- age. The mathematical details can be referred to [36]. We note that FSI
umn from the Hadamard matrix and reshape it from the 1-D array to an and HSI are single-pixel imaging methods based on basis scan and the-
N-by-N matrix, which is the inverse operation of flattening depicted in oretically allow for a perfect reconstruction for any images in noiseless
Fig. 2. situations.
sampling ratio for the other four methods is defined as (n/2)/(M × N),
where n denotes the number of single-pixel measurements, M × N is re-
ferred to the size of the reconstructed image, and the division by 2 is due
to the use of the differential measurement strategy. For BFSI, the sam-
pling ratio is defined as (n/1.5)/(M × N), as it enables a fully-sampled
reconstruction with 1.5-fold as many measurements as image pixels.
In Fig. 8, we present the images reconstructed by the five single-
pixel imaging methods from the fully-sampled data (that is, sampling
Fig. 6. Illustration of Fourier basis pattern binarization. (a) Grayscale Fourier
ratio = 100%). We also use structural similarity index (SSIM) [52] to
basis pattern; (b) upsampled grayscale Fourier basis pattern; (c) binary Fourier
basis pattern.
quantitatively evaluate the reconstruction quality. As the results show,
HSI and FSI can reconstruct all the four different test images losslessly,
evident by SSIM = 1. CSGI is able to reconstruct nearly lossless images
2.5. Differential measurement – a strategy for noise reduction and from fully-sampled data, with the resulting SSIMs very close to 1. Al-
negative-valued patterns generation though BFSI is unable to reconstruct the images losslessly, the quality
of the reconstructions is satisfactory, evident by no apparent artifacts or
The differential measurement method is a typically adopted strat- noise. The SSIMs by BFSI are greater than 0.9 except for the “Text” test
egy for noise elimination and realizing negative-valued patterns gener- image. The quantization errors induced in the Fourier basis patterns bi-
ation in single-pixel imaging, which can be evident by Eqs. (8), (11), narization process is the main reason why BFSI is unable to reconstruct
and (12). Specifically, a pattern 𝑃+ with a value range of [0, 1] is used lossless images even from fully-sampled data. The quality of the images
for spatial light modulation. The resulting single-pixel measurement is reconstructed by CGI is the lowest. Although the reconstructions by CGI
denoted 𝐷+ . And then the inverse pattern 𝑃− = 1 − 𝑃+ is used for spatial appear rather noisy, the high-contrast contents in the reconstructed im-
light modulation and the resulting single-pixel measurement is denoted ages are still recognizable. Considering CGI remarkably underperforms
𝐷− . Finally, by 𝑃 = 𝑃+ − 𝑃− , a pattern with a value range of [−1, 1] is the other four methods in terms of imaging quality, we exclude CGI in
obtained, and the corresponding single-pixel measurement can be cal- the remaining noiseless simulations.
culated by 𝐷 = 𝐷+ − 𝐷− . In the second noiseless simulation, we compare the quality of images
reconstructed by CSGI, HSI, FSI, and BFSI. Particularly, the images are
3. Simulation reconstructed from under-sampled data. Fig. 9 shows the reconstruc-
tions of “Cameraman” for different image sizes and different sampling
We perform numerical simulations to compare the five single-pixel ratios. As Fig. 9 shows, the quality of reconstructed images improves as
imaging methods – CGI, CSGI, HSI, FSI, and BFSI – from the aspects of the sampling ratio increases. When the sampling ratio is low (that is,
imaging quality, imaging efficiency, robustness to noise, and algorithm sampling ratio = 10%), all the reconstructions are blurred. The recon-
complexity. Without loss of generality, we use four different test images structions by FSI or BFSI have ringing artifacts and the reconstructions
in the simulations. As the images shown in Fig. 7, each image has its by HSI have mosaic artifacts. The ringing and the mosaic artifacts are
own characteristics. “Cameraman” is a natural image with continuous due to the characteristics of the basis patterns. The reconstructions by
tone and its energy concentrates at low-frequency band in the Fourier CSGI have patch-like artifacts. When the sampling ratio is medium (that
space. “Barbara” is also a natural image and has richer high-frequency is, sampling ratio from 30% to 50%), the quality of the reconstructions
components than Cameraman. “Siemens star” and “Text” are artificial improves remarkably. For image regions dominated by slow variations
images. “Siemens star” is binary and commonly used to test the spatial (such as, the sky and the coat of the cameraman), FSI and BFSI give the
resolution of imaging systems. It consists of a pattern of bright “spokes” best results. For image regions of rapid changes (such as, the camera
on a dark background that radiate from a common center and become and the face of the cameraman), CSGI can recover the richest details
wider as they get further from it. “Text” is grayscale and consists of texts and the sharpest edges. We note that for the same sampling ratio, the
in black with a background gradually changing from dark (left) to bright reconstruction quality will be higher, if an image is in a larger size. It
(right). Such a design allows for testing the performance of imaging is because, given the same sampling ratio, more single-pixel measure-
methods in reconstructing images with a different image contrast. ments will be required, if the image to be reconstructed is larger. For
We conduct the simulations in both noiseless and noisy situations. example, an image of 64 × 64 pixels reconstructed with a sampling ra-
Please note that the strategy of differential measurement is adopted. All tio of 10% takes only 820 single-pixel measurements, but an image of
the simulations are implemented on a computer with an Intel ® XeonTM 256 × 256 pixels reconstructed with the same sampling ratio takes up to
E5–2690 v2 3.0 GHz CPU, 64 GB RAM, and MATLAB 2019a. 13,106 single-pixel measurements. The more single-pixel measurements
are taken, the more spatial information is acquired. Consequently, the
3.1. Noiseless simulation quality of the resulting reconstruction is higher.
In the previous simulation, we conclude that it is the number of
First, we investigate the imaging quality of different single-pixel single-pixel measurements that affects the reconstruction quality mostly.
imaging methods with respect to the sampling ratio. Except for BFSI, the Thus, we investigate the image reconstruction quality of different single-
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 7. Test images used in the numerical simulations (top row) and their Fourier transform (bottom row): (a) Cameraman; (b) Barbara; (c) Siemens star; (d) Text.
pixel imaging methods with respect to the number of measurements in and the quality improvement is approximately linear. The SSIM curves
this simulation. Fig. 10 shows the reconstructions of “Cameraman” for for different image sizes are highly consistent. Thus, in the following
different image sizes and different numbers of measurements. FSI and simulations, we only present the results for test images of 256 × 256
BFSI outperform HSI and CSGI in the case where the number of mea- pixels.
surements is ultra-low (that is, 200). The cameraman can be roughly re- The reconstructed images of “Barbara” are presented in Fig. 12. We
constructed by FSI or BFSI. For the same number of measurements, BFSI select two image regions for comparison. The region marked by the red
can obtain 1.3-fold as many Fourier coefficients as those by FSI. As the box is the scarf which has fine texture corresponding to high-frequency
more spatial information is acquired, BFSI achieves higher reconstruc- spatial information. The other region marked by the blue box is the face
tion quality than FSI. For HSI, FSI, and BFSI, the reconstructed images of Barbara. In this region, the tone varies slowly. Thus, this region is
look almost the same for the same number of measurements, regardless dominated by low-frequency spatial information. The selection of these
of the image size. On the contrary, the reconstruction quality of CSGI regions allows for the comparison of the single-pixel imaging methods
seems sensitive to both the number of measurements and the image size. in reconstructing high-frequency and low-frequency information.
Specifically, given the same number of single-pixel measurements and As the results shown in Fig. 12(a), when the sampling ratio is 30%,
the number of measurements is small, the larger the reconstructed image the profile of the scarf can be well reconstructed by FSI and BFSI, al-
is, the lower reconstruction quality will be. As Fig. 10 shows, the black though the detailed textures are totally lost. The reconstructions of CSGI
coat can be marginally recognized in the image of 64 × 64 pixels recon- and HSI are noisy or with obvious artifacts. When the sampling ratio is
structed from 200 measurements. However, the contents in the images 50%, CSGI outperforms the other three methods in reconstructing rapid
of 128 × 128 pixels and 256 × 256 pixels reconstructed from the same changes, evident by that it is capable of reconstructing the texture of
number of measurements are not recognizable. Such an observation is the scarf at a certain degree. HSI is better than FSI and BFSI in recover-
also applicable for the case where the number of measurements is 2,000. ing high-frequency information. As the results shown in Fig. 12(b), the
Interestingly, when the number of measurements is large (20,000), the advantage of FSI and BFSI is remarkable especially when the number
quality of the image reconstructed by CSGI is the highest. And the qual- of measurements is small. The face reconstructed by BFSI from 20,000
ity of the image of 256 × 256 pixels is even higher than that of 128 × 128 measurements looks clearer than the those reconstructed by other meth-
pixels. According to the results, we conclude that CSGI is advantageous ods. SSIM curves shown in Fig. 13 are consistent with the conclusion
in the case where the number of single-pixel measurements is large. We drawn above.
can further conclude that, for the basis scan single-pixel imaging meth- According to the results of “Cameraman” and “Barbara”, we con-
ods, the quality of reconstructed images is mainly determined by the clude that FSI and BFSI outperform HSI and CSGI in reconstructing nat-
number of measurements regardless of the image size. ural images which are dominated by slow variations, especially when
In Fig. 11, we present the SSIM versus the number of measurements the number of measurements is small. We also note that CSGI is more
(and sampling ratio) of the five methods. It can be seen that, except for advantageous in recovering the regions of rapid changes, when the sam-
the results by CGI, the quality of the reconstructions improves rapidly pling ratio is medium.
as the sampling ratio or the number of measurement increases, espe- The reconstructed images of “Siemens star” are presented in Fig. 14.
cially when the sampling ratio is low or the number measurement is The feature of “Siemens star” is that the spatial frequency of the contour
small. Such improvement slows down as the sampling ratio or the num- increases as the radius of the contour decreases. Thus, the image can be
ber of measurements increases. For the quality of images reconstructed used to evaluate the reconstruction quality with respect to spatial fre-
by CGI, it is remarkably lower than that by the remaining four methods quency. As the results show, CSGI has difficulty in nicely reconstructing
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 8. Comparison of the images reconstructed by different single-pixel imaging methods from fully-sampled data. The size of the reconstructed images is 256 × 256
pixels.
such an artificial image in the cases where the sampling ratio is 10% or by FSI or BFSI. The contrast of the low-frequency spokes reconstructed
30%. The under-sampled reconstructions of FSI or BFSI are equivalent by CSGI is the lowest, but the contrast of the high-frequency spokes by
to a low-pass result of the original image, because the circular sam- CSGI is the highest. There turns out a trade-off between low-frequency
pling path used in FSI and BFSI permits the low-frequency information spatial information recovery and high-frequency spatial information re-
to be acquired at a higher priority than the high-frequency information. covery.
Thus, in the under-sampled reconstructions, the high-frequency features Fig. 16 shows the SSIM versus sampling ratio (and the number of
are totally lost. On the contrary, in the case of under sampling, the high- measurements) of five different methods for “Siemens star”. Interest-
frequency features can be reconstructed by HSI at some degree, although ingly, when the sampling ratio is ultra-low (lower than 10%), the qual-
the contrast is relatively low. ity of the reconstruction by FSI or BFSI is the poorest, even poorer than
In addition, we choose the two circular profiles in the reconstructed CGI. When the sampling ratio ranges from ~10% to ~60%, the quality
“Siemens star” images for comparison. As Fig. 14(b) shows, the blue cir- of the reconstruction by FSI or BFSI improves significantly and becomes
cle indicates the profile corresponding to a smaller radius and the spatial the highest. However, when the sampling ratio is higher than ~60%,
frequency of the spokes is high. The red circle indicates the profile cor- the quality of the reconstruction by CSGI becomes the highest.
responding to a larger radius and the spatial frequency of the spokes is The reconstructed images of “Text” are presented in Fig. 17. We
low. Fig. 15 shows the intensity distributions of the chosen profiles for choose two image regions for comparison. The one region marked by
different single-pixel imaging methods. As the results show, the contrast the red box consists of some texts in a small font size. This region is
of the low-frequency spokes reconstructed by FSI and BFSI is high while dominated by rapid changes, corresponding to high-frequency spatial
the contrast of the high-frequency spokes is very low. The contrast of the information. The other region marked by the blue box consists of texts
low-frequency spokes reconstructed by HSI is lower than that by FSI or in different font sizes and a gradient change background. Such a de-
BFSI, and the contrast of the high-frequency spokes is higher than that sign is to simulate different levels of image contrast. In our case, the
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 9. Comparison of the "Cameraman" images reconstructed by different single-pixel imaging methods with different sampling ratios.
contrast increases from left to right. As the results shown in Fig. 17, The quality of the reconstructions by FSI, BFSI, and HSI seems indepen-
when the sampling ratio is 10%, FSI and BFSI outperform the others. In dent on image contrast. On the contrary, the reconstruction quality by
the reconstruction by CSGI, neither the characters nor the digits can be CSGI is sensitive to image contrast. Specifically, as the top right panel
identified. The digits and characters reconstructed by FSI and BFSI are in Fig. 17(b) shows, the quality of reconstruction by CSGI is higher in
clearer than those by HSI when the sampling ratio is low or medium. the high-contrast regions than that in the low-contrast regions.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 11. Quantitative comparison of reconstructed images by different single-pixel imaging methods for “Cameraman”.
Fig. 12. Comparison of the "Barbara" images reconstructed by different single-pixel imaging methods. The size of the test image is 256 × 256 pixels.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 14. Comparison of the "Siemens star" images reconstructed by different single-pixel imaging methods. The size of the test image is 256 × 256 pixels.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 15. Comparisons of (a) the low-frequency spokes and (b) the high-frequency spokes intensity profile reconstructed by different single-pixel imaging methods.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Table 1
Comparison of image reconstruction time.
Table 2
Comparison of memory space demanded.
Fig. 16. Quantitative comparison of the "Siemens star" images reconstructed by Method CGI CSGI HSI FSI BFSI
different single-pixel imaging methods.
Spatial complexity O(M × N) O((M × N)2 ) O(M × N) O(M × N) O(M × N)
Poisson noise are presented in Figs. 24-27. As the figures show, the
quality of the images reconstructed by all the five single-pixel imaging the image reconstruction time for CSGI is much longer than the other
methods degrades as the Poisson noise level increases. The deterioration methods.
According to the results, we conclude that CGI is the most computa-
rates are different for different methods. Specifically, the deterioration
rate for CGI is the lowest. HSI gives the best reconstruction under Pas- tionally efficient and CSGI is the most computationally exhausted. FSI is
sion noise condition. The results for CSGI are worse than those for BFSI more computationally efficient than HSI and, therefore, has advantages
in real-time image reconstruction.
or FSI, which is different from the Gaussian noise case. A quantitative
comparison is presented in Fig. 28. Table 2 shows the comparison of the demanded memory for the
five methods. The result shows that CSGI is rather memory demand-
ing, because the implementation of the TVAL3 algorithm requires that
3.3. Algorithm complexity the structured patterns and single-pixel measurements should be loaded
into memory all at a time. Specifically, CSGI consumes over 32GB mem-
The complexity of an algorithm is commonly evaluated from two ory in reconstructing a 256 × 256-pixel image with a sampling ratio of
aspects – temporal complexity (computation time) and spatial complex- 100%.
ity (memory consumption). Here, we compare the temporal complexity
and the spatial complexity of the five single-pixel imaging methods. The 4. Experiment
comparisons are conducted using the simulation data of 256 × 256-pixel
“Cameraman” test image. We also compare the five single-pixel imaging methods with real ex-
Table 1 shows the comparison of computation time spent in im- perimental data. The comparison is conducted by using two typically
age reconstruction. Except for CSGI, the times presented in the ta- used SLMs. The one is a digital projector and the other is a DMD. For
ble are the average of 100,000 image reconstructions. It is because the digital projector, it enables high-brightness and high-contrast illu-
Fig. 17. Comparison of the "Text" images reconstructed by different single-pixel imaging methods. The size of the test image is 256 × 256 pixels.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 19. Comparison of the "Cameraman" images reconstructed by different Fig. 20. Comparison of the "Barbara" images reconstructed by different single-
single-pixel imaging methods from single-pixel measurements contaminated pixel imaging methods from single-pixel measurements contaminated with
with Gaussian noise. The size of the images are 128 × 128 pixels. Gaussian noise. The size of the images are 128 × 128 pixels.
mination. However, the refreshing rate of the digital projector is much lution of the projector is 1920 × 1080 pixels. The projector switches
lower than the achievable refreshing rate of the DMD. In our experi- patterns at a fixed interval. A photodiode amplifier (PDA, Thorlabs
ment, we can only control the projector to switch 5 patterns per second. PDA100A-EC) is used as the single-pixel detector. The target scene con-
Accordingly, the integration time for each single-pixel measurement is sists of two objects – an enlarged and printed USAF-1951 resolution
0.2 sec. The longer integration time results in the higher SNR of the test chart and a Hello Kitty doll. When the target object is under il-
single-pixel measurements. Thus, we assume that it is a low-noise condi- lumination by the projector, the PDA collects the back-scattered light.
tion and we focus on the quality of images with respect to the number of The detected electric signals are delivered to a data acquisition board
measurements or the sampling ratios. Specifically, the quality of images (DAQ, National Instruments USB-6366 (BNC)) for analog-to-digital con-
reconstructed from under-sampled data is compared. On the contrary, version. The final digitalized data is fed to the computer for image
the refreshing rate of the utilized DMD is up to 22.7 KHz. Thus, the in- reconstruction. The structured patterns are of 256 × 256 pixels. To
tegration time of each single-pixel measurement is short, resulting in a make full use of the resolution of the projector, the patterns are up-
low SNR. In this case, the quality of reconstructed images is mainly af- sampled to 1024 × 1024 pixels by using the ‘bicubic’ interpolation
fected by noise. Consequently, we investigate on how the refreshing rate algorithm.
of the DMD affects the reconstruction quality for different single-pixel The reconstructed images are shown in Fig. 30. The sampling ratio
imaging methods. of the reconstructed images is 100%. The quality of the image recon-
structed by CGI is far worse than the quality of others, which coincides
4.1. Experiment using a digital projector for spatial light modulation with the simulation results. We note that BFSI is more efficient than the
other methods, because BFSI only takes 98,304 measurements to fully
The experimental setup is presented in Fig. 29. The digital projector sample a 256 × 256-pixel image with 25% of the measurements reduced,
(DLPDLCR4710EVM-G2) is used for spatial light modulation. The reso- in comparison with the other four methods. However, the quality of the
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 21. Comparison of the "Simens star" images reconstructed by different Fig. 22. Comparison of the "Text" images reconstructed by different single-pixel
single-pixel imaging methods from single-pixel measurements contaminated imaging methods from single-pixel measurements contaminated with Gaussian
with Gaussian noise. The size of the images are 128 × 128 pixels. noise. The size of the images is 128 × 128 pixels.
image reconstructed by BFSI is slightly poorer than those by CSGI, HSI, the single-pixel detector. The structured patterns are pre-generated and
or FSI, which is evident by that the background looks a bit noisy. stored in the RAM of the DMD controller. The initial resolution of the
Fig. 31 shows the results of under-sampled images, where we com- patterns is 256 × 256 pixels. As the resolution of the DMD is 1024 × 768
pare the images reconstructed from the same number of measurements pixels, we upsample all the patterns to 512 × 512 pixels so as to make
by different single-pixel imaging methods. When the number of mea- good use of the modulation area. We note that upsampling provides the
surements is 200, BFSI and FSI are able to reconstruct the outline of the reconstructed images with no extra resolution enhancement. The reso-
object while the other methods cannot. For the image reconstructed by lution of the reconstructed images is still 256 × 256 pixels. We also note
HSI, the scene is contaminated by the mosaic artifacts. When the num- that FSI cannot be conducted by using a DMD directly, because DMD
ber of measurements reaches 2000, the object can be well recognized in only allows for binary patterns generation. Thus, FSI is excluded in this
the images reconstructed by the three basis scan based methods, while experiment.
the image by CSGI is still blurred. When the number of measurements We use two different scenes to conduct the comparison. The one is
reaches 20,000, the imaging quality is high for all methods except CGI. a USAF-1951 resolution target which is a 2-D and binary object. Please
note that the resolution target is printed on a piece of A4 paper. The
4.2. Experiment using a DMD for spatial light modulation other scene is 3-D and complex, consisting of a pair of china toys. As
the reconstructed images shown in Fig. 33, we note that the imaging
We also conduct the experiment using a DMD for spatial light mod- quality of all methods except CGI degrades noticeably as the refreshing
ulation. The schematic set-up is presented in Fig. 32. The setup con- rate of DMD increases. The degradation of imaging quality is because
sists of a 3-watt white-light LED, a DMD (ViALUX V4395), an imaging of the shorter integral time of each single-pixel measurement caused by
lens, a target scene, a collecting lens, and a PDA (Thorlabs PDA101A) as the higher refreshing rate of the DMD.
Fig. 23. Quantitative comparisons of the four test images reconstructed by different single-pixel imaging methods from single-pixel measurements contaminated
with Gaussian noise.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 24. Comparison of the "Cameraman" images reconstructed by different Fig. 26. Comparison of the "Siemens star" images reconstructed by different
single-pixel imaging methods from single-pixel measurements contaminated single-pixel imaging methods from single-pixel measurements contaminated
with Poisson noise. The size of the images is 128 × 128 pixels. with Poisson noise. The size of the images is 128 × 128 pixels.
Fig. 25. Comparison of the "Barbara" images reconstructed by different single- Fig. 27. Comparison of the "Text" images reconstructed by different single-pixel
pixel imaging methods from single-pixel measurements contaminated with Pois- imaging methods from single-pixel measurements contaminated with Poisson
son noise. The size of the images is 128 × 128 pixels. noise. The size of the images is 128 × 128 pixels.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 28. Quantitative comparisons of the four test images reconstructed by different single-pixel imaging methods from single-pixel measurements contaminated
with Poisson noise.
Although the imaging quality of CGI is relatively low, CGI turns out
to be the most robust to noise. The quality of the image for 20,000 Hz
is not much lower than that for 50 Hz. HSI reproduces the best recon-
structions at all refreshing rates. The results by BFSI is a little bit noisy
than those by HSI. The contrast of the images reconstructed by CSGI is
relatively low. Specifically, the contrast reduces as the refreshing rate of
the DMD increases. The reduction of image contrast is due to occurrence Fig. 32. Schematic diagram of the experimental set-up where a DMD is used
of pixels with an abnormally large intensity. Overall, the experiment re- for spatial light modulation.
sults show high consistency with the results in the noisy simulations.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
Fig. 33. Comparison of the fully-sampled images reconstructed by different single-pixel imaging methods and for different referring rate of DMD. The size of the
reconstructed images is 256 × 256 pixels.
5. Discussion and conclusions great potential in quality enhancement, efficiency improvement, and
noise robustness, because the CSGI, as a framework, can employ the
In this paper, five single-pixel imaging methods are compared in state-of-the-art CS algorithm for image reconstruction. In addition, how
terms of imaging quality, imaging efficiency, robustness to noise, and to implement CSGI without matrix operation to tackle the problem of
algorithm complexity. According to the results, several conclusions can huge memory consumption is worth exploring.
be drawn. 1) When the sampling ratio is 100%, the basis scan methods
(such as, FSI, and HSI) allow for perfect reconstruction. The reconstruc- Disclosures
tion by CSGI is nearly perfect. On the contrary, CGI is not able to repro-
duce a perfect reconstruction even with over sampling data. 2) Under The authors declare that there are no conflicts of interest related to
noiseless conditions, the reconstruction quality is mainly dependent on this article.
the number of single-pixel measurements. For the basis scan methods,
given the same number of measurements, the quality of reconstructed Declaration of Competing Interest
images is independent on the size of the image size. However, for CSGI,
the larger size of the reconstructed image is, the lower reconstruction The authors declare no competing financial interests.
quality will be, when the number of measurements is small. 3) Con-
sidering Gaussian noise and Poisson noise, CGI is the most robust and CRediT authorship contribution statement
HSI can reproduce the best reconstruction under noisy conditions. 4)
FSI is good at reconstructing object images dominated by slow varia- Tian’ao Lu: Software, Investigation, Writing - original draft. Ziheng
tions while CSGI is more advantageous in recovering the regions of rapid Qiu: Software, Investigation, Writing - review & editing. Zibang Zhang:
changes. 5) CSGI is the most resource consuming, in terms of compu- Supervision, Funding acquisition, Writing - original draft. Jingang
tational time and memory consumption. 6) BFSI is the most efficient Zhong: Conceptualization.
method in terms of data acquisition and image reconstruction, but more
noise sensitive than FSI, HSI and CSGI. Funding
According to the principal analysis and results, there is always a
trade-off between the imaging quality and the imaging efficiency. The National Science Foundation of China (NSFC) (61905098 and
trade-off makes super high-resolution single-pixel imaging and imaging 61875074) and Fundamental Research Funds for the Central Univer-
a dynamic scene challenging. Currently, single-pixel imaging is a gen- sities (21618307).
eral imaging method capable of imaging all kinds of objects. It is reason-
able to explore content-adaptive methods which can make use of prior References
knowledge of target objects to optimize the structured patterns so as to
[1] Pittman TB, Shih YH, Strekalov DV, Sergienko AV. Optical imaging by
reduce the number of measurements. Adopting machine learning might means of two-photon quantum entanglement. Physical Review A 1995 Nov
become a solution to improve the imaging quality and the imaging ef- 1;52(5):R3429–32.
ficiency in content-adaptive single-pixel imaging. Additionally, how to [2] Bennink RS, Bentley SJ, Boyd RW. “Two-Photon” coincidence imaging with a clas-
sical source. Phys. Rev. Lett. 2002 Aug 26;89(11):113601.
utilize the redundancy between frames in dynamic single-pixel imaging [3] Shapiro JH. Computational ghost imaging. Physical Review A 2008 Dec
is worth exploring. Although CSGI is rather resource-consuming, it has 18;78(6):061802.
T. Lu, Z. Qiu and Z. Zhang et al. Optics and Lasers in Engineering 134 (2020) 106301
[4] Shapiro JH, Erkmen BI. Ghost imaging: from quantum to classical to computational. [29] Bian L, Suo J, Dai Q, Chen F. Experimental comparison of single-pixel imaging al-
Adv Opt Photonics 2010;2(4):405–50. gorithms. Journal of the Optical Society of America A 2018;35(1):78–87.
[5] Edgar MP, Gibson GM, Bowman RW, Sun B, Radwell N, Mitchell KJ, et al. Simulta- [30] Sun M, Meng L, Edgar MP, Padgett MJ, Radwell N. A Russian dolls ordering of the
neous real-time visible and infrared video with single-pixel detectors. Sci Rep 2015 Hadamard basis for compressive single-pixel imaging. Sci Rep 2017 Jun 14;7(1):1–7.
May 22;5:10669. [31] Zhang Z, Wang X, Zheng G, Zhong J. Hadamard single-pixel imaging versus Fourier
[6] Greenberg J, Krishnamurthy K, Brady D. Compressive single-pixel snapshot x-ray single-pixel imaging. Opt Express 2017;25(16):19619–39.
diffraction imaging. Opt Lett 2014;39(1):111–14. [32] Xiao Y, Zhou L, Chen W. Direct single-step measurement of Hadamard spectrum
[7] Zhang A, He Y, Wu L, Chen L, Wang B. Tabletop x-ray ghost imaging with ultra-low using single-pixel optical detection. IEEE Photonics Technology Letters 2019 Apr
radiation. Optica 2018;5(4):374–7. 11;31(11):845–8.
[8] Chan WL, Charan K, Takhar D, Kelly KF, Baraniuk RG, Mittleman DM. A single-pixel [33] Yu W. Super sub-Nyquist single-pixel imaging by means of cake-cutting Hadamard
terahertz imaging system based on compressed sensing. Appl Phys Lett 2008 Sept basis sort. Sensors 2019 Aug 25;19(19):4122.
23;93(12):121105. [34] Yu W, Liu Y. Single-pixel imaging with origami pattern construction. Sensors 2019
[9] She R, Liu W, Lu Y, Zhou Z, Li G. Fourier single-pixel imaging in the terahertz regime. Sept 30;19(23):5135.
Appl Phys Lett 2019 Jul 8;115(2):021101. [35] Wu H, Wang R, Li C, Chen M, Zhao G, He Z, et al. Influence of intensity fluctu-
[10] Watts CM, Shrekenhamer D, Montoya J, Lipworth G, Hunt J, Sleasman T, et al. Ter- ations on Hadamard-based computational ghost imaging. Opt Commun 2020 Jan
ahertz compressive imaging with metamaterial spatial light modulators. Nat Pho- 1;454:124490.
tonics 2014 Jun 29;8(8):605. [36] Zhang Z, Ma X, Zhong J. Single-pixel imaging by means of Fourier spectrum acqui-
[11] Stantchev RL, Sun B, Hornett SM, Hobson PA, Gibson GM, Padgett MJ, et al. Nonin- sition. Nat Commun 2015 Feb 4;6:6225.
vasive, near-field terahertz imaging of hidden objects using a single-pixel detector. [37] Bian L, Suo J, Hu X, Chen F, Dai Q. Efficient single pixel imaging in Fourier space.
Sci Adv 2016 Jun 3;2(6):e1600190. Journal of Optics 2016 Jul 14;18(8):085704.
[12] Stantchev RI, Yu X, Blu T, Pickwell-MacPherson E. Real-time terahertz imaging with [38] Zhang Z, Zhong J. Three-dimensional single-pixel imaging with far fewer measure-
a single-pixel detector. Nat Commun 2020 May 21;11:2535. ments than effective image pixels. Opt Lett 2016;41(11):2497–500.
[13] Peng J, Yao M, Cheng J, Zhang Z, Li S, Zheng G, et al. Micro-tomography via sin- [39] Jiang H, Zhu S, Zhao H, Xu B, Li X. Adaptive regional single-pixel imaging based on
gle-pixel imaging. Opt Express 2018;26(24):31094–105. the Fourier slice theorem. Opt Express 2017;25(13):15118–30.
[14] Zhang Z, Su Z, Deng Q, Ye J, Peng J, Zhong J. Lensless single-pixel imaging by [40] Meng W, Shi D, Huang J, Yuan K, Wang Y, Cu Fan. Sparse Fourier single-pixel imag-
using LCD: application to small-size and multi-functional scanner. Opt Express ing. Opt Express 2019;27(22):31490–503.
2019;27(3):3731–45. [41] Zhang Z, Wang X, Zhong J. Fast Fourier single-pixel imaging via binary illumination.
[15] Zhang Z, Yao M, Li X, Deng Q, Peng Q, Zhong J. Simultaneous functional and Sci Rep 2017 Sept 20;7(1):1–9.
structural imaging for photovoltaic devices. Solar Energy Materials and Solar Cells [42] Huang J, Shi D, Yuan K, Hu S, Wang Y. Computational-weighted Fourier single-pixel
2019;193:101–6. imaging via binary illumination. Opt Express 2018;26(13):16547–59.
[16] Zhang Z, Jiao S, Yao M, Li X, Zhong J. Secured single-pixel broadcast imaging. Opt [43] Edgar MP, Gibson GM, Padgett MJ. Principles and prospects for single-pixel imaging.
Express 2018;26(11):14578–91. Nat Photonics 2019 Dec 3;13(1):13–20.
[17] Zhang Z, Ye J, Deng Q, Zhong J. Image-free real-time detection and tracking of fast [44] Max B, Emil W. Principles of Optics: electromagnetic Theory of Propagation. Inter-
moving object using a single-pixel detector. Opt Express 2019;27(24):35394–401. ference and diffraction of light. 7th ed. Cambridge University Press; 1988.
[18] Zhang Z, Li X, Zheng S, Yao M, Zheng G, Zhong J. Image-free classification of fast– [45] Sen P, Chen B, Garg G, Marschner SR, Horowitz M, Levoy M, et al. Dual photography.
moving objects using “learned” structured illumination and single-pixel detection. In: SIGGRAPH´05: ACM SIGGRAPH 2005 Papers; Jul 2005. p. 745–55.
Opt Express 2020;28(9):13269–78. [46] Shi D, Huang J, Wang F, Cao K, Yuan K, Hu S, et al. Enhancing resolution of sin-
[19] Shu D, Zhou D, Zhou X, Peng W, Chen L, Bao X. Signal-to-noise ratio analysis of gle-pixel imaging system. Opt Rev 2015 Aug 19;22(5):802–8.
computational distributed fiber-optic sensing. Opt Express 2020;28(7):9563–71. [47] Sun B, Edgar MP, Bowman R, Vittert LE, Welsh S, Bowman A, et al. 3D computational
[20] Ota S, Horisaki R, Kawamura Y, Ugawa M, Sato I, Hashimoto K, et al. Ghost cytom- imaging with single-pixel detectors. Science 2013 May 17;340(6134):844–7.
etry. Science 2018 Jun 15;360(6394):1246–51. [48] Pratt WK, Kane J, Andrews HC. Hadamard transform image coding. Proceedings of
[21] Zhang Z, Ye J, Deng Q, Zhong J. Image-free real-time detection and tracking of fast the IEEE 1969;57(1):58–68.
moving object using a single-pixel detector. Opt Express 2019;27(24):35394–401. [49] Vaz PG, Amaral D, Ferreira LR, Morgado M, Cardoso J. Image quality of com-
[22] Sun B, Welsh SS, Edgar MP, Shapiro JH, Padgett MJ. Normalized ghost imaging. Opt pressive single-pixel imaging using different Hadamard orderings. Opt Express
Express 2012;20(15):16892–901. 2020;28(8):11666–81.
[23] Donoho DL. Compressed sensing. IEEE Transactions on Information Theory 2006 [50] Yu X, Yang F, Gao B, Ran J, Huang X. Deep Compressive Single Pixel Imag-
Apr 3;52(4):1289–306. ing by Reordering Hadamard Basis: a Comparative Study. IEEE Access 2020 Mar
[24] Katz O, Bromberg Y, Siliberberg Y. Compressive ghost imaging. Appl Phys Lett 2009 17;8:55773–84.
Sept 30;95(13):131110. [51] Floyd R, Steinberg L. An adaptive algorithm for spatial grey scale. Proc. Society for
[25] Hale E.T., Yin W., Zhang Y. A fixed-point continuation method for l1-regularized Information Display 1976;17(2):75–7.
minimization with applications to compressed sensing. CAAM Technical Report [52] Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: from error
TR07-07, Rice University 2007 Jul 7;43:44. visibility to structural similarity. IEEE Transactions on Image Processing 2014 Apr
[26] Elad M. Optimized projections for compressed sensing. IEEE Transactions on Signal 13;13(4):600–12.
Processing 2007 Nov 19;55(12):5695–702. [53] Le T, Rick C, Thomas JA. A variational approach to reconstructing images corrupted
[27] Li C. [Ph.D. dissertation]. Rice University; 2010. by Poisson noise. J Math Imaging Vis 2007 Mar 30;27(3):257–63.
[28] Becker S, Bobin J, Candès EJ. NESTA: a fast and accurate first-order method for
sparse recovery. SIAM J Imaging Sci 2011 Jan 6;4(1):1–39.