1. Introduction
A Wireless Multimedia Sensor Network (WMSN) consists of optical sensor nodes deployed to an area of interest and at least one data sink in different topologies [
1]. The nodes study various areas and send data to at least one sink. Multihop routing can be part of the link, raising the requirement of a high compression ratio.
These WMSNs have to function in energyconstrained environments that demand novel compression schemes to lessen the bandwidth utilisation and computational complexity [
2]. Energy conservation is one of the three energy management methods exploited in Wireless Sensor Networks (WSN); the others are energy transfer and energy harvesting [
3]. However, the large data transfers for WMSN make energy conservation the greatest tool. Compressed Sensing (CS) was introduced by Pudlewski et al. [
4] as a means to overcome this. The CS mechanism is appropriate for WMSN owing to its low complexity, high compression rate, and robustness to transmission errors [
5].
Even though there have been other approaches to complexity reduction, such as sparsity transforms [
5,
6], they come at an intractable price of image quality [
7].
Traditionally, each CS measurement maps an image onto an unrepeated measurement vector [
8]. However, this imposes a sizable memory footprint when the signal is significant, such as in highresolution images. In [
9], Gan introduced BlockCompressed Sensing (BCS) to break up the image into blocks and use the same measurement vector for every block, thus vastly minimizing the footprint.
Bajwa et al. [
10] significantly reduced sensing and storage complexity by proving the efficacy of Toeplitz and circulant sensing matrices. With these, just the first row needs random entries, while the other rows are transformations of the first. These matrices are called semideterministic, and matrices with no random entries are fully deterministic.
Some researchers have been looking at optimisationbased matrices that minimise the mutual coherence with the sparsity transform [
11,
12,
13,
14,
15,
16], but these matrices cannot be constructed on WMSN nodes. More recent approaches have been training the sensing matrix based on a signal dataset [
17,
18,
19], but similarly, this is not ideal for WMSN nodes.
Arjoune et al. [
20] contrasted different types of sensing matrices with great detail. They investigated using experiments on 1D signals on a central processing unit (CPU)powered computer. However, these time measurements (and possible energy consumption) are not characteristic of WMSNs using microcontroller units (MCUs).
Determinsistic sensing matrices have outperformed dense ones in many studies [
21,
22,
23,
24,
25,
26,
27,
28,
29,
30]. Semideterministic matrices are being displaced in popularity by fully deterministic ones, particularly matrices based on chaotic sequences [
24,
27,
28].
1.1. Motivation
Sensing complexity has yet to be analysed or compared with previous work by most researchers. Storage complexity has been studied more, but it is not a good proxy for sensing complexity. Energy consumption has been studied even less. A sensing matrix is needed that can meet the need for energy efficiency and recovery performance and compare favourably with current approaches.
1.2. Contribution and Organisation
In this paper, we analyse, evaluate, and compared the leading deterministic measurement matrices. We propose a new measurement matrix that addresses the identified shortcomings and compare it with the leading matrices in terms of its energy efficiency for construction and sensing as well as recovery performance.
We review the leading deterministic matrices regarding image quality and computational complexity.
We identify the best techniques for designing deterministic sensing matrices.
We evaluate the effect of random numbers and block size on recovery performance and energy consumption.
We propose a sensing matrix based on random sample positions.
In
Section 2, research output on Sensing Matrices is reviewed. The theoretical background is introduced in
Section 3. In
Section 4, the new sensing matrix is proposed. Experiments are detailed in
Section 5, and the results are presented in
Section 6. The results are discussed in
Section 7, and the conclusion is presented in
Section 8.
2. Related Work
There has been considerable work on developing sensing matrices for addressing the computational complexity and reconstruction quality of CS. In the next subsections, the current semideterministic and fully deterministic approaches are reviewed. Amongst semideterministic matrices, the Binary Permuted Block Diagonal (BPBD) had the most promise because of its high sparsity. The other matrices had lower sparsity and complex computations. The Deterministic Binary Block Diagonal matrix (DBBD) had the highest sparsity amongst fully deterministic matrices, which is based on the BPBD. The semideterministic matrices tend to have complex computations, while the fully deterministic matrices have low sparsity.
2.1. SemiDeterministic Matrices
He et al. [
31] proposed the BPBD to address hardware implementation and sensing efficiency. The matrix is binary and significantly sparse, which substantially simplifies hardware implementation and reduces sensing computations. This matrix requires random numbers to permute the columns during construction, which adds complexity depending on the source of the random numbers. The researchers contrasted the BPBD with other matrices with experiments on MATLAB. They found that their measurement matrix gives comparable recovery performance to scrambled block Hadamard, scrambled Fourier, partial Noiselet, and Gaussian and that their measurement matrix outperforms the other matrices at a sampling rate of 10%.
In [
23], two memoryefficient measurement matrices were proposed. The first one is called the Combination matrix, and it is the Kronecker product of the Gaussian and Toeplitz measurement matrices. This approach reduces the effective number of required random entries, but the multiplication increases the construction complexity. The second matrix is called the Hybrid matrix, and it is generated through a compound of Toeplitz and Binary matrices. The authors evaluated their matrices using experiments on the TelosB platform. They observed that their matrices performed better than the Gaussian matrix based on energy consumption and recovery performance. The Hybrid performed better than the Combination matrix.
A combination of Toeplitz, Hankel, and circulant matrices was proposed in [
26]. First, the Toeplitz matrix is constructed from Gaussian random entries. The Hankel and circulant matrices are constructed from transformations of the Toeplitz matrix. Finally, the sensing matrix is constructed by summing all three matrices. These researchers compared their matrix to the Gaussian matrix on the TelosB platform. The authors found that their matrix outperformed the Gaussian matrix on recovery performance and energy consumption.
2.2. Fully Deterministic Matrices
Ravelomanantsoa et al. [
30] proposed a simple deterministic measurement matrix that facilitates hardware implementation. This sensing matrix is named the DBBD and simplifies the BPBD by eliminating random permutations. The authors tested their proposed measurement matrix on a 1D electrocardiogram (ECG) and an electromyogram (EMG) on an MSPEXP430G2 LaunchPad development board. They experimented to assess their proposed matrix against the Gaussian and BPBD matrices. They exploited the DCT and their proposed OMPvariant reconstruction algorithm. Their measurement matrix performed better than both the Gaussian and BPBD in terms of recovery accuracy measured in SNR and computational complexity.
In [
24], the authors exploited the simplicity of the Logistic mapping chaotic systems to develop the incoherent rotated chaotic (IRC) sensing matrix. They ignored the first 1000 elements of the sequence and downsampled the subsequent elements using an interval of
d to improve the randomness. This means that to generate
t sampled values, a chaotic sequence of length
$l=1000+dt$ must be generated. The IRC matrix requires only
n factors that are rotated for each row, which improves storage compared with other chaotic matrices. The authors added an incoherence factor
$\eta $, which is multiplied for each rotation and can be separated into a matrix
$\omega $. However, the multiple multiplications from the incoherence factor add computational complexity to the generation process. The matrix was compared to the Gaussian matrix using MATLAB experiments. The authors found that their matrix outperforms the Gaussian one, but it is a challenge to select the incoherence factor, which has a meaningful impact on the reconstruction performance.
Gan et al. [
27] presented a measurement matrix that exploits bipolar chaotic sequences to simplify storage and multiplication. However, the sequences are generated from the Chebyshev chaotic system that entails the
$cos$ function, which is expensive to implement in MCU. A threshold value is applied to the sequence to make a bipolar matrix. The researchers contrasted their matrix to the Gaussian, Bernoulli, improved Hadamard, and dense Chebyshev matrices through experiments. Their matrix performed comparably to the other matrices and performed the best at low sampling rates.
Sun et al. [
28] presented the ChaosBernoulli Block Circulant Matrix to conserve transmission capacity. They produced a pseudorandom sequence by choosing the initial value and sampling interval. The sequence is derived from a nonlinear Hybrid Chaotic map where the Logistic map and Tent map are combined. The authors ignored the first 1000 values of the sequence to improve the randomness. They also sampled the chaotic sequence at an interval to improve the independence of the sampled values. The sampled sequence is then mapped through the sign function. The authors built a blockcirculant matrix from a chaotic sequence to reduce the storage and implementation requirements. The final measurement matrix was obtained by randomly selecting
M rows from the block circulant matrix. The matrix was compared to the Gaussian, Bernoulli, Hybrid chaotic, and Gaussian circulant matrices using numerical experiments. The authors found that their matrix outperformed the other matrices.
2.3. Summary
The relative performance of the different matrices is presented in
Table 1. The complexity was assigned values of Low, Medium, or High. The reconstruction accuracy was judged based on whether the matrix had higher image quality than the Gaussian matrix.
3. Preliminaries
This section gives a short background on the compressive sensing and WSN used for the development of the proposed scheme.
3.1. Compressed Sensing
The CS framework allows for the accurate recovery of data captured in a vector of length
N from only
$M\ll N$ samples. CS depends on sparsity and incoherence to ensure signal recovery [
32]. Sparsity is the quality of how few nonzero entries belong to a vector representing the signal. A signal can also be made sparse by applying a transformation,
$\mathrm{\Psi}$, to a domain where the signal can be sparse. Incoherence is the quality of low correlation between the sparsifying transform and the sensing matrix,
$\mathrm{\Phi}$ [
32]. The incoherence can be measured through mutual coherence, shown in Equation (
1), where lower values mean more incoherence.
The mutual coherence is the maximum normalised similarity between the row vectors of $\mathrm{\Phi}$, ${\varphi}_{i}$ and the column vectors of $\mathrm{\Psi}$, ${\psi}_{j}$.
Once the signal has been compressively sensed in
M measurements, it can be recovered in its entirety using numerical optimisation algorithms. The algorithms attempt to recover an unknown signal
Z from a few measurements. The signal to be compressed
$x\in {R}^{N}$ can be expressed as
where
$\mathrm{\Psi}$ is the sparsity transform and
s is the
N length vector of the captured signal. The signal
x is sparse, with a sparsity of
K. When the measurement matrix is applied to
x, we obtain a vector
y.
The robustness of CS depends on the restricted isometry property (RIP) of the sensing matrix. The RIP requires columns of
$\mathrm{\Phi}$ to be close to orthogonal. Random sensing matrices meet the RIP criteria with high probability on the condition that Equation (
4) [
32] holds. The RIP is also important in ensuring that CS can gracefully deal with additive noise.
3.2. Peak SignaltoNoise Ratio
Image quality is measured using the PeakSignalto Noise Ratio (PSNR). This metric has been shown to be robust at low sampling rates and high distortion [
33]. This was a popular objective measure of image quality. PSNR is measured in decibels (dB) and raises in value with image quality; Equation
5 is the mathematical formulation [
34].
The PSNR has an opposite relationship to the mean square error (MSE), which is the cumulative squared error between the reference ${\rho}_{r}(i,j)$ and the processed image ${\rho}_{p}(i,j)$, and L is the dynamic range of the pixel value.
3.3. Wireless Multimedia Sensor Networks
There have been many WMSN node designs, ranging from 8 MHz using the ATmega128L to 624 MHz using the XScale PXA270. In [
35], the authors classify these WMSN nodes into Low and High performance. The microcontroller unit (MCU) is an integral part of the sensor node and is responsible for a significant amount of its total energy consumption [
36]. The focus of this study will be on lowperformance nodes, in particular, the TelosB mote, for energy analysis. This mote has a Texas Instrument MSP430 MCU that runs at 8 MHz with 10 KB of RAM and 1 MB of an external flash. The mote has two AA batteries, each with a capacity of around 2850 mAh.
The choice of sensing matrix design affects the required number of random numbers for generating the measurement matrix, as well as the number of matrix multiplications during sensing. The Bernoulli matrix with
$\pm 1$ entries was favoured for ease of implementation [
37] by eliminating the multiplications. Another popular matrix is the Binary Sparse Random matrix, which has a small
d number of nonzero entries per column, thus eliminating most addition operations [
38].
Two choices exist for obtaining random entries for a measurement matrix, retrieval from memory or random generation. Retrieval from memory has the drawback of requiring large storage space, and nonvolatile memory requires significant energy to operate [
39]. The random generation of measurement matrix columns has considerable energy consumption from complex mathematical functions such as
$log$ or
$sqrt$ [
38]. Deterministic measurement matrices have been identified as a means to ease hardware requirements by obviating random number generation and/or large memory storage [
30].
Energy Consumption
It has been observed that the current drawn by 8 to 32bit MCUs deviates very slightly during active operation [
40]. It is, therefore, sufficient to measure the time length of an operation to estimate relative energy consumption, as seen in Equation (
6).
The time length
${t}_{op}$ can be estimated using the clock frequency
f of the MCU and the number of instruction cycles required for the operation
${N}_{ic}$, as shown in Equation (
7).
The MSP430 does not have a hardware multiplier, and thus, the implementation of complex mathematical operations requires multiple instruction cycles. In [
41], the authors discuss efficient algorithms to implement multiplication and division on the MCU. In
Table 2, the instruction cycle cost of different arithmetic and logical operations on the MSP430 MCU are listed [
41,
42]. In the table, each mathematical operation is listed and the instruction cycle cost is presented. All the arithmetic and logical operations require one instruction cycle except
$multiply$ and
$divide$. These arithmetic operations require 29 and 22 cycles, respectively.
4. Proposed Sensing Matrix
In this section, an energyefficient and highfidelity sensing matrix for WMSN is proposed. Fully deterministic matrices commonly use chaotic sequences to replace random numbers, which give similar sparsity to the Bernoulli matrix. Another successful fully deterministic matrix is the DBBD, which has a static matrix but still requires
$NM$ addition operations during sensing. Semideterministic matrices have had good sparsity, with the PCI requiring no
$add$ operations during sensing. The PCI also has low mutual coherence with most sparsity transforms [
43]. The weakness of the PCI matrix is that it requires expensive random numbers, which is similar to other semideterministic matrices. This shortcoming is addressed by replacing the random numbers with a chaotic sequence. The construction complexity is further improved by replacing random permutation with random sample positions, which also reduces time complexity.
4.1. Deterministic Partial Canonical Identity Matrix
The PCI matrix is traditionally constructed by randomly choosing
M rows from an
$N\times N$ identity matrix. The PCI has been used in [
43,
44,
45], but in these studies, the matrix computational complexity and recovery performance have not been investigated. The construction of the PCI is where the energy consumption, and recovery performance are determined.
A popular approach to improving the recovery performance of measurement matrices is the minimisation of mutual coherence with a sparsity basis of
$\mathrm{\Psi}$ [
11,
13,
14,
15,
16]. However, for the DPCI, mutual coherence is a constant,
$\sqrt{2}$ [
43].
4.2. Random Number Generation
Random number generation is an important factor in reducing computational complexity and recovery performance. Different randomnumbergeneration algorithms were evaluated to find the most suitable ones in terms of simplicity and recovery performance.
4.2.1. Linear Congruent Generator
The Linear Congruential Generator (LCG) is one of the bestknown and oldest algorithms for generating pseudorandom numbers. This generator is simple to implement and has low storage complexity. The algorithm exploits recurrence to generate random numbers shown in Equation (
8).
Here,
X is the state of the algorithm, and
x is a random number between 0 and 1.0. The variables
a,
c, and
m are the multiplier, increment, and modulus, respectively. The choice of values for these variables determines the performance of the algorithm. The values used were
a = 16,807,
$c=0$, and
m = 2,147,483,647 from [
46]. The resulting energy consumption is shown in Equation (
9) in terms of instruction cycles.
4.2.2. Logistic Map
The Logistic map is one of the simplest chaotic systems. This system is a suitable random number generator [
47]. The equation of the system is shown in Equation (
10). The
$\mu $ is the system parameter, and each value of
n is
${x}_{n}\in [0,1]$. A system parameter of
$\mu =4.0$ and an initial state of
${x}_{0}=0.4$ were used. The energy consumption for each number in the sequence is formulated in Equation (
11) and measured in instruction cycles.
4.2.3. Tent Map
The Tent map is an iterative piecewise linear difference equation with chaotic behaviour and a tentlike graph. This map is computationally efficient and has a uniform distribution of random numbers [
48]. The equation is shown in Equation (
12). The
$\mu $ is the system parameter, and each value of
n is
${x}_{n}\in [0,1]$. The system parameter
$\mu =1.99$ and initial state
${x}_{0}=0.4$ were used. The energy consumption for each number in the sequence is formulated in Equation (
13) in instruction cycles.
4.3. Complexity Optimisation
The computational complexity of constructing the PCI was optimised by first exploiting deterministic random numbers and also by replacing random permutation with random sample positions. This optimisation reduced the computational and time complexity.
4.3.1. Random Permutation
The Fisher–Yates Shuffle (FYS) algorithm was used to randomly permute the columns of the PCI matrix. The algorithm has two attractive properties: permutations are unbiased, and the shuffling has a linear time complexity
$\mathcal{O}\left(n\right)$ [
49]. The algorithm was used with values from the Tent map chaotic sequence as described in Algorithm 1. The energy consumption of each iteration of the algorithm is formulated using instruction cycles in Equation (
14).
Algorithm 1 Random Permutation. 
 1:
procedure Shuffle(A)  2:
${X}_{0}$ = $0.400$  3:
U = $1.99$  4:
for index i from $N1$ to 1 in A do  5:
if ${X}_{i1}<0.5$ then  6:
${X}_{i}=U{X}_{i1}$  7:
else  8:
${X}_{i}=U(1{X}_{i1})$  9:
end if  10:
j = (i− 1)${X}_{i}$ + 1  11:
Swap $A\left[i\right]$ and $A\left[j\right]$  12:
end for  13:
Return A  14:
end procedure

4.3.2. Random Sample Position
The fundamental characteristic of the PCI matrix is that there is only one nonzero entry per row. This fact is exploited to simplify the construction of the DPCI. The random sample position algorithm consists of generating a random column position for each row; see Algorithm 2.
Algorithm 2 Random Sample Position. 
 1:
procedure Sample(A)  2:
${X}_{0}$ = $0.400$  3:
for row i from 1 to M in A do  4:
${X}_{i}=random\left({X}_{i1}\right)$  5:
j = $N\times {X}_{i}$  6:
$A[i,j]$ = $1.0$  7:
end for  8:
Return A  9:
end procedure

The new construction maintains the orthogonal relationship between the columns, where the nonzero row entries do not coincide between columns. This arrangement maintains the RIP properties of the original shuffling mechanism. The new construction is more sensitive to the quality of the random number generator to distribute the nonzero entries evenly across columns. The randomness of the distribution will also be sensitive to the block size N through the sample size M.
The energy consumption of the algorithm is formulated in Equation (
15) in instruction cycles. The permutation requires
M iterations to approximate the shuffling of
N columns. Each iteration entails only one complex computation, that of a Tent random number. This eliminates most of the computational complexity from shuffling. The algorithm also reduces the required random numbers and iterations from
$N1$ to
M.
This algorithm was tested with all three random number generators, LCG, Logistic, and Tent. Finally, it was compared with the traditional implementation using random permutation; see
Figure 1 and
Figure 2. The traditional implementation with FYS performed the best in terms of recovery quality, but the LCG and Tent had a competitive performance. The Tent implementation was chosen for further evaluation because of the lower computational complexity.
5. Experimental Evaluation
MATLAB was used for compressed sensing experiments on a Windows 10 computer with 16 GB RAM and an Intel Core i7 CPU. In order to make the results representative, many images were utilised during the experiments. A selection of images from the collection of greyscale images from the Waterloo Repertoire GreySet2 [
50] was exploited in 512 × 512 sizes in order to assess image quality. The collection consists of 11 images of a wide range of subjects such as people, landscapes, animals, objects and posters (with text and subfigures); the dataset is detailed in
Table 3. The image was divided into blocks of 8 × 8, 16 × 16 and 32 × 32 pixels. Every block was made sparse using DCT and sampled using different sensing matrices. The images were recovered using blockbased CS sampling and smoothed projected Landweber (BCSSPL) method [
51]. Experiments were conducted to assess each sensing matrix at a low sampling rate of 10% to evaluate the matrices at the most challenging and energyefficient level. Each experiment was repeated multiple times to account for statistical variations using five trials. The recovery accuracy was measured using the average image quality in the form of PSNR. The energy consumption was evaluated using analytical methods.
6. Results
The results of the experiments are presented in this section. The recovery performance is first presented using image quality. Finally, the energy consumption is analysed.
6.1. Recovery Performance
The quantitative recovery performance of the different matrices is presented using image quality in
Table 4,
Table 5 and
Table 6. These tables cover disparate block sizes, 8 × 8, 16 × 16, and 32 × 32. The PSNR for each recovered image is listed for each sensing matrix, and the average over all the images is listed on the last row of each table.
In
Table 4, the highest image quality was from the Zelda file, with 32.19 dB compressed with the DBBD matrix. The lowest image quality was on the Library file, with 15.19 dB compressed using the DPCI matrix. The DBBD had the highest average performance at 24.42 dB, followed by the BPBD at 24.01 dB. The DPCI, at 22.94 dB, had the third best, and finally, the Gaussian matrix was the lowest at 21.84 dB.
Table 5 shows that the highest image quality from the Zelda file, with 33.36 dB compressed using the DBBD matrix. The lowest image quality was found in the Library file, with 15.65 dB compressed using both the DPCI and Gaussian matrices. The DBBD had the best average performance of 25.28 dB; the second was the BPBD with 24.03 dB. The Gaussian was third with 23.53 dB, and the DPCI had 23.35 dB.
Table 6 shows that the highest image quality was obtained from the Zelda file, with 32.71 dB compressed with the DBBD matrix. The lowest image quality was found in the Library file, with 14.91 dB compressed using the DPCI matrix. The DBBD had the highest average performance with 24.73 dB, followed by the BPBD with 23.96 dB. The Gaussian had the thirdbest performance with 22.94 dB, and the DPCI was last with 22.81 dB.
The qualitative difference in performance of each of the sensing matrices is shown in
Figure 3,
Figure 4 and
Figure 5 through a few examples. Four images that were compressed using the Gaussian, DBBD, BPBD. and DPCI are shown in each figure.
Figure 3 shows Boat images compressed using a block size of 8 × 8. In the figure, the DBBD had the sharpest detail but had significant pixelation artefacts. The DPCI had the secondbest detail, where small details such as the ship rigging are visible. However, the DPCI had pixelation artefacts. The BPBD had the thirdbest detail, where sections of the ship rigging are still visible. The Gaussian had no pixelation artefacts but had the lowest detail. These results are corroborated quantitatively in
Table 4.
Figure 4 shows the GoldHill2 images compressed using a block size of 16 × 16. The DBBD had the best visual detail but had some colour distortions. The BPBD and DPCI had very similar performances. The Gaussian had the least detail but had no distortions. These results are similar to the quantitative results in
Table 5.
Figure 5 shows Mandrill images compressed using a block size of 32. The DBBD had the best visual detail but had visible lines across the image. The DPCI and Gaussian had similar performance with the secondmost detail. The BPBD had blocking artefacts and had the worst performance. The results are reflected in
Table 6, where the performance numbers are close, with only the DBBD standing out.
6.2. Energy Consumption
The energy consumption of the different measurement matrices was compared using the MSP430x1xx Family MCU as a reference. The most energyefficient operating region of the MCU is 1MHz, with a voltage supply of 2.2 V. In the active mode, the MCU current draw is 220
$\mathsf{\mu}$A. The instruction cycle cost of mathematical operations is translated into energy consumption in
Table 7. Multiplication and division are the most expensive operations, while the other operations are the same and lower in energy consumption.
The construction of the DPCI requires
M Tent sequence numbers. The DPCI uses no multiplications or additions during sensing, unlike the BPBD and DBBD, which use
$NM$ additions. The construction and sensing costs are shown in
Table 8 for block sizes 8, 16, and 32, all at a 10% sampling rate. The DPCI only consumes energy for construction, while the BPBD and DBBD only consume energy for sensing. The total energy consumed for sensing the image is unaffected by block size and depends only on the size of the image. In this case, it was 114.2
$\mathsf{\mu}$J.
7. Discussion
The images had different levels of image quality after recovery. What is interesting is that three images had much lower performance, france.tiff, library.tiff, and mountain.tiff. These three images had an image quality below 20 dB for all the sensing matrices and block sizes. The first image is a poster with text and a highcontrast background. The second image is another poster with three subfigures, text, and a white background. The last image is of a mountain range with high contrast.
The DBBD had the best quantitative and qualitative performance for all the block sizes. The matrix suffered from some distortion but had the greatest detail in all the images. The BPBD had the secondhighest quantitative performance for all the block sizes, but the qualitative performance was only sometimes secondbest. The DPCI had the thirdbest quantitative results for the block size of 8 × 8. However, it had the worst performance for the bigger sizes. The DPCI had the second and thirdbest qualitative performance for some of the images. The Gaussian had the least detail and worst qualitative performance for most of the images but had no distortions.
The block size had different effects on each sensing matrix. The Gaussian had the lowest performance at the block size of 8 × 8 but had improved performance for 16 × 16 and 32 × 32. The BPBD had similar recovery performance for all the block sizes, while the DBBD and DPCI had peak performance at 16 × 16. The DPCI had the worst quantitative performance at 32 × 32, and it performed particularly severely for challenging images, with a relative dip in performance of almost 1 dB compared to the Gaussian. This performance drop may be caused by the statistical properties of the distribution of the nonzero values being affected by the sample size, N, but this needs further study.
The DPCI consumes energy for construction, unlike the DBBD, but this energy can be reduced by decreasing the block size. The DPCI consumes no energy for sensing, unlike the other matrices. Sensing energy is important because it determines the operating energy consumption of the node, and depending on the spectral and spatial resolution of the sensor, as well as the type and duration of the events being monitored, these sensing costs can accumulate exponentially.
8. Conclusions
Chaotic sequences can be used to address the shortcomings of semideterministic matrices to make fully deterministic matrices that are both energyefficient and high in fidelity. The DPCI improves the computational complexity of the PCI by first replacing the random numbers with chaotic sequences and, secondly, by replacing random permutations with random sample positions. The DPCI is outperformed by the DBBD and BPBD regarding recovery performance but offers a significant computational advantage. The poorer recovery accuracy of the DPCI might be because it is much more sparse and takes fewer samples but this needs further investigation. The DPCI offers an attractive tradeoff between recovery performance and energy efficiency that energysensitive applications can exploit. Some images are challenging for sensing matrices and need further investigation. Future research should focus on improving the recovery performance of the DPCI by investigating the sample positions that lead to better performance using training on an image dataset. These results can then be used to generalise across datasets and develop a universal sensing matrix.
Author Contributions
Conceptualization, V.S.; methodology, V.S.; software, V.S.; validation, V.S.; formal analysis, V.S.; investigation, V.S.; resources, V.S.; data curation, V.S.; writing—original draft preparation, V.S.; writing—review and editing, V.S. and A.A.M.; visualization, V.S.; supervision, A.A.M.; project administration, V.S.; funding acquisition, V.S. All authors have read and agreed to the published version of the manuscript.
Funding
This research received no external funding.
Institutional Review Board Statement
Not applicable.
Informed Consent Statement
Not applicable.
Data Availability Statement
There were no datasets created during this study and all relevant datasets are already publicly available.
Acknowledgments
The authors would like to thank all the individuals and organisations that have supported this research.
Conflicts of Interest
The authors declare no conflict of interest.
References
 Wang, P.; Dai, R.; Akyildiz, I.F. A Spatial CorrelationBased Image Compression Framework for Wireless Multimedia Sensor Networks. IEEE Trans. Multimed. 2011, 13, 388–401. [Google Scholar] [CrossRef]
 Akyildiz, I.F.; Melodia, T.; Chowdhury, K.R. A survey on wireless multimedia sensor networks. Comput. Netw. 2007, 51, 921–960. [Google Scholar] [CrossRef]
 Engmann, F.; Katsriku, F.A.; Abdulai, J.D.; AduManu, K.S.; Banaseka, F.K. Prolonging the lifetime of wireless sensor networks: A review of current techniques. Wirel. Commun. Mob. Comput. 2018, 2018, 8035065. [Google Scholar] [CrossRef]
 Pudlewski, S.; Prasanna, A.; Melodia, T. CompressedSensingEnabled Video Streaming for Wireless Multimedia Sensor Networks. IEEE Trans. Mob. Comput. 2012, 11, 1060–1072. [Google Scholar] [CrossRef]
 Angayarkanni, V.; Radha, S. Design of Bandwidth Efficient Compressed Sensing Based Prediction Measurement Encoder for Video Transmission in Wireless Sensor Networks. Wirel. Pers. Commun. 2016, 88, 553–573. [Google Scholar] [CrossRef]
 Banerjee, R.; Bit, S.D. Lowoverhead video compression combining partial discrete cosine transform and compressed sensing in WMSNs. Wirel. Netw. 2019, 25, 5113–5135. [Google Scholar] [CrossRef]
 Skosana, V.; AbuMahfouz, A.M. EnergyEfficient Sensing Matrices for Wireless Multimedia Sensor Networks: A Review. In Proceedings of the 2021 International Conference on Forthcoming Networks and Sustainability in AIoT Era (FoNeSAIoT), Nicosia, Turkey, 27–28 December 2021; pp. 51–56. [Google Scholar] [CrossRef]
 Zhang, J.; Xiang, Q.; Yin, Y.; Chen, C.; Luo, X. Adaptive compressed sensing for wireless image sensor networks. Multimed. Tools Appl. 2017, 76, 4227–4242. [Google Scholar] [CrossRef]
 Gan, L. Block Compressed Sensing of Natural Images. In Proceedings of the 2007 15th International Conference on Digital Signal Processing, Cardiff, UK, 1–4 July 2007; pp. 403–406. [Google Scholar] [CrossRef]
 Bajwa, W.U.; Haupt, J.D.; Raz, G.M.; Wright, S.J.; Nowak, R.D. ToeplitzStructured Compressed Sensing Matrices. In Proceedings of the 2007 IEEE/SP 14th Workshop on Statistical Signal Processing, Madison, WI, USA, 26–29 August 2007; pp. 294–298. [Google Scholar] [CrossRef]
 Elad, M. Optimized Projections for Compressed Sensing. IEEE Trans. Signal Process. 2007, 55, 5695–5702. [Google Scholar] [CrossRef]
 DuarteCarvajalino, J.M.; Sapiro, G. Learning to Sense Sparse Signals: Simultaneous Sensing Matrix and Sparsifying Dictionary Optimization. IEEE Trans. Image Process. 2009, 18, 1395–1408. [Google Scholar] [CrossRef] [PubMed]
 Abolghasemi, V.; Ferdowsi, S.; Sanei, S. A gradientbased alternating minimization approach for optimization of the measurement matrix in compressive sensing. Signal Process. 2012, 92, 999–1009. [Google Scholar] [CrossRef]
 Huang, H.; Makur, A. Optimized measurement matrix for compressive sensing. Sampl. Theory Signal Image Process. 2013, 12, 71–86. [Google Scholar] [CrossRef]
 Yan, W.; Wang, Q.; Shen, Y. ShrinkageBased Alternating Projection Algorithm for Efficient Measurement Matrix Construction in Compressive Sensing. IEEE Trans. Instrum. Meas. 2014, 63, 1073–1084. [Google Scholar] [CrossRef]
 Wei, Z.; Zhang, J.; Xu, Z.; Liu, Y.; Okarma, K. Measurement matrix optimization via mutual coherence minimization for compressively sensed signals reconstruction. Math. Probl. Eng. 2020, 2020, 7979606. [Google Scholar] [CrossRef]
 Baldassarre, L.; Li, Y.H.; Scarlett, J.; Gözcü, B.; Bogunovic, I.; Cevher, V. LearningBased Compressive Subsampling. IEEE J. Sel. Top. Signal Process. 2016, 10, 809–822. [Google Scholar] [CrossRef]
 Ahmed, I.; Khan, A.; Ahmad, N.; Ali, H. Speech signal recovery using block sparse bayesian learning. Arab. J. Sci. Eng. 2020, 45, 1567–1579. [Google Scholar] [CrossRef]
 Ahmed, I.; Khan, A. Learning based speech compressive subsampling. Multimed. Tools Appl. 2023, 82, 15327–15343. [Google Scholar] [CrossRef]
 Arjoune, Y.; Kaabouch, N.; El Ghazi, H.; Tamtaoui, A. A performance comparison of measurement matrices in compressive sensing. Int. J. Commun. Syst. 2018, 31, e3576. [Google Scholar] [CrossRef]
 Sun, J.; Wang, S.; Dong, Y. Sparse block circulant matrices for compressed sensing. IET Commun. 2013, 7, 1412–1418. [Google Scholar] [CrossRef]
 Su, X.; Hongpeng, Y.; Yi, C.; Yushu, X.; Xue, T. An Improved Toeplitz Measurement Matrix for Compressive Sensing. Int. J. Distrib. Sens. Netw. 2014, 10, 846757. [Google Scholar] [CrossRef]
 Nandhini, S.A.; Sankararajan, R.; Rajendiran, K. Video compressed sensing framework for wireless multimedia sensor networks using a combination of multiple matrices. Comput. Electr. Eng. 2015, 44, 51–66. [Google Scholar] [CrossRef]
 Yao, S.; Wang, T.; Shen, W.; Shaoming, P.; Chong, Y. Research of incoherence rotated chaotic measurement matrix in compressed sensing. Multimed. Tools Appl. 2017, 76, 17699–17717. [Google Scholar] [CrossRef]
 Hong, T.; Li, X.; Zhu, Z.; Li, Q. Optimized structured sparse sensing matrices for compressive sensing. Signal Process. 2019, 159, 119–129. [Google Scholar] [CrossRef]
 Nandhini, S.A.; Radha, S.; Nirmala, P.; Kishore, R. Compressive sensing for images using a variant of Toeplitz matrix for wireless sensor networks. J. RealTime Image Process. 2019, 16, 1525–1540. [Google Scholar] [CrossRef]
 Gan, H.; Xiao, S.; Zhang, T.; Liu, F. Bipolar measurement matrix using chaotic sequence. Commun. Nonlinear Sci. Numer. Simul. 2019, 72, 139–151. [Google Scholar] [CrossRef]
 Sun, Y.; Han, G.; Huang, L.; Wang, S.; Xiang, J. Construction of Block Circulant Measurement Matrix Based on Hybrid Chaos: Bernoulli Sequences. In Proceedings of the 2020 4th International Conference on Digital Signal Processing (ICDSP 2020), Chengdu, China, 19–21 June 2020; Association for Computing Machinery: New York, NY, USA, 2020; pp. 1–6. [Google Scholar] [CrossRef]
 Hanumanthu, S. Universal Measurement Matrix Design for Sparse and CoSparse Signal Recovery. Turk. J. Comput. Math. Educ. (TURCOMAT) 2021, 12, 404–411. [Google Scholar]
 Ravelomanantsoa, A.; Rabah, H.; Rouane, A. Compressed Sensing: A Simple Deterministic Measurement Matrix and a Fast Recovery Algorithm. IEEE Trans. Instrum. Meas. 2015, 64, 3405–3413. [Google Scholar] [CrossRef]
 He, Z.; Ogawa, T.; Haseyama, M. The simplest measurement matrix for compressed sensing of natural images. In Proceedings of the 2010 IEEE International Conference on Image Processing, Hong Kong, China, 26–29 September 2010; pp. 4301–4304. [Google Scholar] [CrossRef]
 Candes, E.J.; Wakin, M.B. An Introduction To Compressive Sampling. IEEE Signal Process. Mag. 2008, 25, 21–30. [Google Scholar] [CrossRef]
 Skosana, V.; AbuMahfouz, A.M. Performance Comparison of Video Encoding at Low Sampling Rates. In Proceedings of the 2021 International Symposium on Networks, Computers and Communications (ISNCC), Dubai, United Arab Emirates, 31 October–2 November 2021; pp. 1–7. [Google Scholar] [CrossRef]
 Araar, C.; Ghanemi, S.; Benmohammed, M.; Atoui, H. Pruned improved eightpoint approximate DCT for image encoding in visual sensor networks requiring only ten additions. J. RealTime Image Process. 2020, 17, 1597–1608. [Google Scholar] [CrossRef]
 Shi, H.; Hou, K.M.; Diao, X.; Xing, L.; Li, J.J.; De Vaulx, C. A Wireless Multimedia Sensor Network Platform for Environmental Event Detection Dedicated to Precision Agriculture. arXiv 2018, arXiv:1806.03237. [Google Scholar] [CrossRef]
 Singh, D.; Sai Prashanth, S.; Kundu, S.; Pal, A. Lowpower microcontroller for wireless sensor networks. In Proceedings of the TENCON 2009—2009 IEEE Region 10 Conference, Singapore, 23–26 January 2009; pp. 1–6. [Google Scholar] [CrossRef]
 Chen, F.; Chandrakasan, A.P.; Stojanovic, V.M. Design and Analysis of a HardwareEfficient Compressed Sensing Architecture for Data Compression in Wireless Sensors. IEEE J. SolidState Circuits 2012, 47, 744–756. [Google Scholar] [CrossRef]
 Caione, C.; Brunelli, D.; Benini, L. Compressive Sensing Optimization for Signal Ensembles in WSNs. IEEE Trans. Ind. Inform. 2014, 10, 382–392. [Google Scholar] [CrossRef]
 Jones, M.; Scott, J. The Energy Efficiency of 8bit Lowpower Microcontrollers. In Proceedings of the 18th Electronics New Zealand Conference (ENZCON 2011), Massey University, Palmerston North, New Zealand, 21–22 November 2011; pp. 87–90. [Google Scholar]
 Ruberg, P.; Liiv, E.; Lass, K.; Ellervee, P. Data Analysis for Embedded Software Performance and Energy Consumption Estimation. In Proceedings of the 2019 IEEE 2nd Ukraine Conference on Electrical and Computer Engineering (UKRCON), Lviv, Ukraine, 2–6 July 2019; pp. 928–933. [Google Scholar] [CrossRef]
 MSP430x5xx MSP430x6xx Family User’s Guide Rev. Q; Texas Instruments: Dallas, TX, USA, 2018.
 Efficient multiplication and division using MSP430. In Technical Report SLAA329A; Texas Instruments: Dallas, TX, USA, 2006.
 Jain, N.; Gupta, A.; Bohara, V.A. PCIMDR: Missing Data Recovery in Wireless Sensor Networks Using Partial Canonical Identity Matrix. IEEE Wirel. Commun. Lett. 2019, 8, 673–676. [Google Scholar] [CrossRef]
 Ansari, N.; Gupta, A. Image Reconstruction Using Matched Wavelet Estimated From Data Sensed Compressively Using Partial Canonical Identity Matrix. IEEE Trans. Image Process. 2017, 26, 3680–3695. [Google Scholar] [CrossRef] [PubMed]
 Jain, N.; Bohara, V.A.; Gupta, A. PCIMF: Partial Canonical Identity and Matrix Factorization Framework for Channel Estimation in mmWave Massive MIMO Systems. IEEE Open J. Signal Process. 2020, 1, 135–145. [Google Scholar] [CrossRef]
 Lewis, P.A.W.; Goodman, A.S.; Miller, J.M. A pseudorandom number generator for the System/360. IBM Syst. J. 1969, 8, 136–146. [Google Scholar] [CrossRef]
 Phatak, S.C.; Rao, S.S. Logistic map: A possible randomnumber generator. Phys. Rev. E 1995, 51, 3670–3678. [Google Scholar] [CrossRef]
 Demidova, L.A.; Gorchakov, A.V. A Study of Chaotic Maps Producing Symmetric Distributions in the Fish School Search Optimization Algorithm with Exponential Step Decay. Symmetry 2020, 12, 784. [Google Scholar] [CrossRef]
 Juniawan, F.P.; Pradana, H.A.; Sylfania, D.Y. Performance comparison of Linear Congruent method and FisherYates Shuffle for data randomization. In Proceedings of the Journal of Physics: Conference Series, Palembang, Indonesia, 26–27 November 2018; IOP Publishing: Bristol, UK, 2019; Volume 1196, p. 012035. [Google Scholar]
 Waterloo Repertoire GreySet2; University of Waterloo: Waterloo, ON, Canada, 2008.
 Mun, S.; Fowler, J.E. Block compressed sensing of images using directional transforms. In Proceedings of the 2009 16th IEEE International Conference on Image Processing (ICIP), Cairo, Egypt, 7–10 November 2009; pp. 3021–3024. [Google Scholar] [CrossRef]
 Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. 
© 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).